From: Georgi Gerganov Date: Sun, 9 Nov 2025 20:01:21 +0000 (+0200) Subject: sync : llama.cpp X-Git-Tag: upstream/1.8.3~340 X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=a1867e0dad0b21b35afa43fc815dae60c9a139d6;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp sync : llama.cpp --- diff --git a/examples/talk-llama/CMakeLists.txt b/examples/talk-llama/CMakeLists.txt index 182114c2..deeab482 100644 --- a/examples/talk-llama/CMakeLists.txt +++ b/examples/talk-llama/CMakeLists.txt @@ -2,6 +2,8 @@ if (WHISPER_SDL2) set(CMAKE_CXX_STANDARD 17) set(CMAKE_CXX_STANDARD_REQUIRED ON) + file(GLOB SRC_MODELS models/*.cpp) + set(TARGET whisper-talk-llama) add_executable(${TARGET} talk-llama.cpp llama.cpp @@ -29,7 +31,8 @@ if (WHISPER_SDL2) llama-sampling.cpp llama-vocab.cpp unicode.cpp - unicode-data.cpp) + unicode-data.cpp + ${SRC_MODELS}) target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS}) target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT}) diff --git a/examples/talk-llama/llama-arch.cpp b/examples/talk-llama/llama-arch.cpp index 8ca769c5..b7642b56 100644 --- a/examples/talk-llama/llama-arch.cpp +++ b/examples/talk-llama/llama-arch.cpp @@ -32,6 +32,8 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_QWEN2VL, "qwen2vl" }, { LLM_ARCH_QWEN3, "qwen3" }, { LLM_ARCH_QWEN3MOE, "qwen3moe" }, + { LLM_ARCH_QWEN3VL, "qwen3vl" }, + { LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" }, { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PHI3, "phi3" }, { LLM_ARCH_PHIMOE, "phimoe" }, @@ -103,6 +105,9 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_SEED_OSS, "seed_oss" }, { LLM_ARCH_GROVEMOE, "grovemoe" }, { LLM_ARCH_APERTUS, "apertus" }, + { LLM_ARCH_MINIMAX_M2, "minimax-m2" }, + { LLM_ARCH_COGVLM, "cogvlm" }, + { LLM_ARCH_PANGU_EMBED, "pangu-embedded" }, { LLM_ARCH_UNKNOWN, "(unknown)" }, }; @@ -145,6 +150,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_EXPERTS_PER_GROUP, "%s.experts_per_group" }, { LLM_KV_MOE_EVERY_N_LAYERS, "%s.moe_every_n_layers" }, { LLM_KV_NEXTN_PREDICT_LAYERS, "%s.nextn_predict_layers" }, + { LLM_KV_NUM_DEEPSTACK_LAYERS, "%s.n_deepstack_layers" }, { LLM_KV_POOLING_TYPE, "%s.pooling_type" }, { LLM_KV_LOGIT_SCALE, "%s.logit_scale" }, { LLM_KV_DECODER_START_TOKEN_ID, "%s.decoder_start_token_id" }, @@ -779,6 +785,45 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, }, }, + { + LLM_ARCH_QWEN3VL, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_QWEN3VLMOE, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + }, + }, { LLM_ARCH_PHI2, { @@ -2312,6 +2357,64 @@ static const std::map> LLM_TENSOR_N { LLM_TENSOR_FFN_UP_CHEXPS, "blk.%d.ffn_up_chexps" }, }, }, + { + LLM_ARCH_MINIMAX_M2, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_ATTN_Q_NORM, "blk.%d.attn_q_norm" }, + { LLM_TENSOR_ATTN_K_NORM, "blk.%d.attn_k_norm" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE_INP, "blk.%d.ffn_gate_inp" }, + { LLM_TENSOR_FFN_GATE_EXPS, "blk.%d.ffn_gate_exps" }, + { LLM_TENSOR_FFN_DOWN_EXPS, "blk.%d.ffn_down_exps" }, + { LLM_TENSOR_FFN_UP_EXPS, "blk.%d.ffn_up_exps" }, + { LLM_TENSOR_FFN_EXP_PROBS_B, "blk.%d.exp_probs_b" }, + }, + }, + { + LLM_ARCH_PANGU_EMBED, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, + { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, + { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, + { + LLM_ARCH_COGVLM, + { + { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, + { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, + { LLM_TENSOR_OUTPUT, "output" }, + { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, + { LLM_TENSOR_ATTN_QKV, "blk.%d.attn_qkv" }, + { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, + { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, + { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, + { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + { LLM_TENSOR_VISEXP_ATTN_QKV, "blk.%d.vis_attn_qkv" }, + { LLM_TENSOR_VISEXP_ATTN_OUT, "blk.%d.vis_attn_output" }, + { LLM_TENSOR_VISEXP_FFN_GATE, "blk.%d.vis_gate" }, + { LLM_TENSOR_VISEXP_FFN_DOWN, "blk.%d.vis_down" }, + { LLM_TENSOR_VISEXP_FFN_UP, "blk.%d.vis_up" }, + }, + }, { LLM_ARCH_UNKNOWN, { @@ -2488,6 +2591,11 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_SHORTCONV_CONV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_SSM_CONV}}, {LLM_TENSOR_SHORTCONV_INPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SHORTCONV_OUTPROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_VISEXP_ATTN_QKV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_VISEXP_ATTN_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_VISEXP_FFN_GATE, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_VISEXP_FFN_DOWN, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_VISEXP_FFN_UP, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, // NextN/MTP tensors are currently ignored (reserved for future MTP support) // These tensors only exist in the last layer(s) and are treated as output tensors {LLM_TENSOR_NEXTN_EH_PROJ, {LLM_TENSOR_LAYER_OUTPUT, GGML_OP_MUL_MAT}}, diff --git a/examples/talk-llama/llama-arch.h b/examples/talk-llama/llama-arch.h index dea725c1..a769dd1e 100644 --- a/examples/talk-llama/llama-arch.h +++ b/examples/talk-llama/llama-arch.h @@ -36,6 +36,8 @@ enum llm_arch { LLM_ARCH_QWEN2VL, LLM_ARCH_QWEN3, LLM_ARCH_QWEN3MOE, + LLM_ARCH_QWEN3VL, + LLM_ARCH_QWEN3VLMOE, LLM_ARCH_PHI2, LLM_ARCH_PHI3, LLM_ARCH_PHIMOE, @@ -107,6 +109,9 @@ enum llm_arch { LLM_ARCH_SEED_OSS, LLM_ARCH_GROVEMOE, LLM_ARCH_APERTUS, + LLM_ARCH_MINIMAX_M2, + LLM_ARCH_COGVLM, + LLM_ARCH_PANGU_EMBED, LLM_ARCH_UNKNOWN, }; @@ -149,6 +154,7 @@ enum llm_kv { LLM_KV_EXPERTS_PER_GROUP, LLM_KV_MOE_EVERY_N_LAYERS, LLM_KV_NEXTN_PREDICT_LAYERS, + LLM_KV_NUM_DEEPSTACK_LAYERS, LLM_KV_POOLING_TYPE, LLM_KV_LOGIT_SCALE, LLM_KV_DECODER_START_TOKEN_ID, @@ -455,6 +461,11 @@ enum llm_tensor { LLM_TENSOR_SHORTCONV_CONV, LLM_TENSOR_SHORTCONV_INPROJ, LLM_TENSOR_SHORTCONV_OUTPROJ, + LLM_TENSOR_VISEXP_ATTN_QKV, + LLM_TENSOR_VISEXP_ATTN_OUT, + LLM_TENSOR_VISEXP_FFN_GATE, + LLM_TENSOR_VISEXP_FFN_DOWN, + LLM_TENSOR_VISEXP_FFN_UP, LLM_TENSOR_NEXTN_EH_PROJ, LLM_TENSOR_NEXTN_EMBED_TOKENS, LLM_TENSOR_NEXTN_ENORM, diff --git a/examples/talk-llama/llama-batch.cpp b/examples/talk-llama/llama-batch.cpp index 55d89eca..86a1a4ba 100644 --- a/examples/talk-llama/llama-batch.cpp +++ b/examples/talk-llama/llama-batch.cpp @@ -215,6 +215,7 @@ bool llama_batch_allocr::init( /*.n_seq_tokens =*/ (uint32_t) 1, /*.n_seqs =*/ (uint32_t) batch.n_tokens, /*.n_seqs_unq =*/ (uint32_t) this->seq_id_unq.size(), + /*.n_pos =*/ n_pos_per_embd, /*.token =*/ batch.token, /*.embd =*/ batch.embd, /*.pos =*/ batch.pos, @@ -251,46 +252,72 @@ bool llama_batch_allocr::init( // consistency checks // - for (uint32_t s = 0; s < n_seq_max; ++s) { - if (seq_pos[s].empty()) { - continue; + if (n_pos_per_embd > 1) { + // M-RoPE case: allow position to "jump" forward only (non-continuous positions are allowed) + for (uint32_t s = 0; s < n_seq_max; ++s) { + if (seq_pos[s].empty()) { + continue; + } + + const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1; + + if (batch.token) { + if (p0 >= 0 && p0 >= seq_pos_min(s)) { + LLAMA_LOG_ERROR( + "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n" + " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n" + " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n" + " for M-RoPE, it is required that the position satisfies: X < Y\n", + __func__, s, s, p0, s, seq_pos_min(s)); + + return false; + } + } else { + // embedding inputs can have overlapping positions + if (p0 >= 0 && p0 > seq_pos_min(s)) { + LLAMA_LOG_ERROR( + "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n" + " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n" + " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n" + " for M-RoPE, it is required that the position satisfies: X <= Y\n", + __func__, s, s, p0, s, seq_pos_min(s)); + + return false; + } + } } + } else { + for (uint32_t s = 0; s < n_seq_max; ++s) { + if (seq_pos[s].empty()) { + continue; + } - const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1; + const llama_pos p0 = memory ? memory->seq_pos_max(s) : -1; - if (p0 >= 0) { - bool ok = true; + if (p0 >= 0) { + bool ok = true; - if (batch.token) { if (seq_pos_min(s) != p0 + 1) { ok = false; } - } else { - assert(batch.embd); - // for embeddings (typically used as vision input), we allow them to have repeating positions - // ref: https://github.com/ggml-org/llama.cpp/issues/13694#issuecomment-2983871762 - if (seq_pos_min(s) != p0 && seq_pos_min(s) != p0 + 1) { - ok = false; + if (!ok) { + LLAMA_LOG_ERROR( + "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n" + " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n" + " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n" + " it is required that the sequence positions remain consecutive: Y = X + 1\n", + __func__, s, s, p0, s, seq_pos_min(s)); + + return false; } } - if (!ok) { - LLAMA_LOG_ERROR( - "%s: the tokens of sequence %d in the input batch have inconsistent sequence positions:\n" - " - the last position stored in the memory module of the context (i.e. the KV cache) for sequence %d is X = %d\n" - " - the tokens for sequence %d in the input batch have a starting position of Y = %d\n" - " it is required that the sequence positions remain consecutive: Y = X + 1\n", - __func__, s, s, p0, s, seq_pos_min(s)); - + if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) { + LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s); return false; } } - - if (seq_pos_max(s) - seq_pos_min(s) + 1 > (int) seq_pos[s].size()) { - LLAMA_LOG_ERROR("%s: sequence %d positions are not continuous\n", __func__, s); - return false; - } } if (memory) { @@ -389,6 +416,7 @@ llama_ubatch llama_batch_allocr::ubatch_reserve(uint32_t n_seq_tokens, uint32_t /*.n_seq_tokens =*/ n_seq_tokens, /*.n_seqs =*/ n_seqs, /*.n_seqs_unq =*/ n_seqs, + /*.n_pos =*/ n_pos_per_embd, /*.token =*/ udata->token.data(), /*.embd =*/ nullptr, @@ -655,10 +683,8 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector & idxs, u auto udata = std::make_shared(); - const int32_t n_pos_cur = batch.embd ? n_pos_per_embd : 1; - const int64_t n_embd_all = batch.embd ? (int64_t) n_tokens*n_embd : 0; - const int64_t n_pos_all = (int64_t) n_tokens*n_pos_cur; + const int64_t n_pos_all = (int64_t) n_tokens*n_pos_per_embd; udata->token .resize(n_tokens); udata->embd .resize(n_embd_all); @@ -680,8 +706,13 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector & idxs, u memcpy(udata->embd.data() + i*n_embd, batch.embd + (int64_t) idxs[i]*n_embd, n_embd*sizeof(float)); } - for (int j = 0; j < n_pos_cur; ++j) { - udata->pos[j*n_tokens + i] = batch.pos[j*batch.n_tokens + idxs[i]]; + for (size_t j = 0; j < (size_t)n_pos_per_embd; ++j) { + // if we are using M-RoPE + // if the current batch is text, we need to broadcast the same position across all RoPE sections + // otherwise, the input batch is image embeddings, we copy the positions as-is + // if we are not using M-RoPE, there is only one position per token (this loop runs only once) + size_t src_off = batch.token ? 0 : j*batch.n_tokens; + udata->pos[j*n_tokens + i] = batch.pos[src_off + idxs[i]]; } udata->n_seq_id[i] = batch.n_seq_id[idxs[i]]; @@ -710,6 +741,7 @@ llama_ubatch llama_batch_allocr::ubatch_add(const std::vector & idxs, u /*.n_seq_tokens =*/ n_tokens/n_seqs, /*.n_seqs =*/ n_seqs, /*.n_seqs_unq =*/ (uint32_t) udata->seq_id_unq.size(), + /*.n_pos =*/ n_pos_per_embd, /*.token =*/ batch.token ? udata->token.data() : nullptr, /*.embd =*/ batch.embd ? udata->embd.data() : nullptr, diff --git a/examples/talk-llama/llama-batch.h b/examples/talk-llama/llama-batch.h index 0dc8cebd..209cf369 100644 --- a/examples/talk-llama/llama-batch.h +++ b/examples/talk-llama/llama-batch.h @@ -17,6 +17,16 @@ struct llama_ubatch { return b_equal_seqs != 0; } + // typical for M-RoPE cases: + // 0 - sequantial position of the tokens/embeddings in the sequence + // 1 - y position in the image + // 2 - x position in the image + // 3 - other + bool is_pos_2d() const { + // TODO @ngxson : we may need to check for model arch when more models use >1 positions + return n_pos >= 3; + } + uint32_t b_equal_seqs; // note: this is a boolean, but we use an int32_t for alignment // otherwise address sanitizer complains // TODO: whole_seqs for embeddings? @@ -25,6 +35,7 @@ struct llama_ubatch { uint32_t n_seq_tokens; // tokens per sequence set uint32_t n_seqs; // sequence sets in the ubatch uint32_t n_seqs_unq; // unique sequence ids in the ubatch + uint32_t n_pos; // number of position inputs for each token/embedding // seq_id_unq: unique sequence ids in the ubatch // seq_idx: indices of the unique sequence ids in the ubatch in [0, n_seqs_unq) @@ -33,7 +44,7 @@ struct llama_ubatch { // // size | idx | val llama_token * token; // [n_tokens] | i | id, token float * embd; // [n_embd, n_tokens] | i | embd - llama_pos * pos; // [n_tokens] | i | pos + llama_pos * pos; // [n_tokens*n_pos] | i | pos int32_t * n_seq_id; // [n_tokens] | i | - llama_seq_id ** seq_id; // [n_tokens] | s | s0, s1, seq_id llama_seq_id * seq_id_unq; // [n_seqs_unq] | s | seq_id diff --git a/examples/talk-llama/llama-chat.cpp b/examples/talk-llama/llama-chat.cpp index 0285006d..fc6a6223 100644 --- a/examples/talk-llama/llama-chat.cpp +++ b/examples/talk-llama/llama-chat.cpp @@ -73,6 +73,7 @@ static const std::map LLM_CHAT_TEMPLATES = { { "kimi-k2", LLM_CHAT_TEMPLATE_KIMI_K2 }, { "seed_oss", LLM_CHAT_TEMPLATE_SEED_OSS }, { "grok-2", LLM_CHAT_TEMPLATE_GROK_2 }, + { "pangu-embedded", LLM_CHAT_TEMPLATE_PANGU_EMBED }, }; llm_chat_template llm_chat_template_from_str(const std::string & name) { @@ -213,6 +214,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) { return LLM_CHAT_TEMPLATE_SEED_OSS; } else if (tmpl_contains("'Assistant: ' + message['content'] + '<|separator|>")) { return LLM_CHAT_TEMPLATE_GROK_2; + } else if (tmpl_contains(LU8("[unused9]系统：[unused10]"))) { + return LLM_CHAT_TEMPLATE_PANGU_EMBED; } return LLM_CHAT_TEMPLATE_UNKNOWN; } @@ -813,6 +816,35 @@ int32_t llm_chat_apply_template( if (add_ass) { ss << "Assistant:"; } + }else if (tmpl == LLM_CHAT_TEMPLATE_PANGU_EMBED) { + // [unused9]系统：xxx[unused10] + // [unused9]用户：xxx[unused10] + // [unused9]助手：xxx[unused10] + // ... + for (size_t i = 0; i < chat.size(); ++i) { + const auto & msg = chat[i]; + const std::string & role = msg->role; + const std::string & content = msg->content; + + if (i == 0 && role != "system") { + ss << "[unused9]系统：[unused10]"; + } + + if (role == "system") { + ss << "[unused9]系统：" << content << "[unused10]"; + } else if (role == "user") { + ss << "[unused9]用户：" << content << "[unused10]"; + } else if (role == "assistant") { + ss << "[unused9]助手：" << content << "[unused10]"; + } else if (role == "tool") { + ss << "[unused9]工具：" << content << "[unused10]"; + } else if (role == "function") { + ss << "[unused9]方法：" << content << "[unused10]"; + } + } + if (add_ass) { + ss << "[unused9]助手："; + } } else { // template not supported return -1; diff --git a/examples/talk-llama/llama-chat.h b/examples/talk-llama/llama-chat.h index da1b7c47..684efb4d 100644 --- a/examples/talk-llama/llama-chat.h +++ b/examples/talk-llama/llama-chat.h @@ -53,6 +53,7 @@ enum llm_chat_template { LLM_CHAT_TEMPLATE_KIMI_K2, LLM_CHAT_TEMPLATE_SEED_OSS, LLM_CHAT_TEMPLATE_GROK_2, + LLM_CHAT_TEMPLATE_PANGU_EMBED, LLM_CHAT_TEMPLATE_UNKNOWN, }; diff --git a/examples/talk-llama/llama-context.cpp b/examples/talk-llama/llama-context.cpp index bd348bca..70a3ec62 100644 --- a/examples/talk-llama/llama-context.cpp +++ b/examples/talk-llama/llama-context.cpp @@ -21,6 +21,8 @@ llama_context::llama_context( llama_context_params params) : model(model), balloc(std::make_unique(model.hparams.n_pos_per_embd())) { + // TODO warning when creating llama_context with awkward ctx size that is not a power of 2, + // may need to be backend-dependent LLAMA_LOG_INFO("%s: constructing llama_context\n", __func__); t_start_us = model.t_start_us; @@ -112,11 +114,28 @@ llama_context::llama_context( } } - const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; + // ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732 + cparams.n_ctx = GGML_PAD(cparams.n_ctx, 256); + + if (cparams.kv_unified) { + cparams.n_ctx_seq = cparams.n_ctx; + } else { + cparams.n_ctx_seq = cparams.n_ctx / cparams.n_seq_max; + cparams.n_ctx_seq = GGML_PAD(cparams.n_ctx_seq, 256); + + if (cparams.n_ctx_seq == 0) { + throw std::runtime_error("n_ctx_seq == 0"); + } + + if (cparams.n_ctx != cparams.n_ctx_seq * cparams.n_seq_max) { + cparams.n_ctx = cparams.n_ctx_seq * cparams.n_seq_max; + LLAMA_LOG_WARN("%s: n_ctx is not divisible by n_seq_max - rounding down to %u\n", __func__, cparams.n_ctx); + } + } LLAMA_LOG_INFO("%s: n_seq_max = %u\n", __func__, cparams.n_seq_max); LLAMA_LOG_INFO("%s: n_ctx = %u\n", __func__, cparams.n_ctx); - LLAMA_LOG_INFO("%s: n_ctx_per_seq = %u\n", __func__, n_ctx_per_seq); + LLAMA_LOG_INFO("%s: n_ctx_seq = %u\n", __func__, cparams.n_ctx_seq); LLAMA_LOG_INFO("%s: n_batch = %u\n", __func__, cparams.n_batch); LLAMA_LOG_INFO("%s: n_ubatch = %u\n", __func__, cparams.n_ubatch); LLAMA_LOG_INFO("%s: causal_attn = %d\n", __func__, cparams.causal_attn); @@ -125,14 +144,14 @@ llama_context::llama_context( LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); - if (n_ctx_per_seq < hparams.n_ctx_train) { - LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n", - __func__, n_ctx_per_seq, hparams.n_ctx_train); + if (cparams.n_ctx_seq < hparams.n_ctx_train) { + LLAMA_LOG_WARN("%s: n_ctx_seq (%u) < n_ctx_train (%u) -- the full capacity of the model will not be utilized\n", + __func__, cparams.n_ctx_seq, hparams.n_ctx_train); } - if (n_ctx_per_seq > hparams.n_ctx_train) { - LLAMA_LOG_WARN("%s: n_ctx_per_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", - __func__, n_ctx_per_seq, hparams.n_ctx_train); + if (cparams.n_ctx_seq > hparams.n_ctx_train) { + LLAMA_LOG_WARN("%s: n_ctx_seq (%u) > n_ctx_train (%u) -- possible training context overflow\n", + __func__, cparams.n_ctx_seq, hparams.n_ctx_train); } if (!hparams.vocab_only) { @@ -268,9 +287,7 @@ llama_context::llama_context( if (pipeline_parallel) { LLAMA_LOG_INFO("%s: pipeline parallelism enabled (n_copies=%d)\n", __func__, ggml_backend_sched_get_n_copies(sched.get())); } - } - if (!hparams.vocab_only) { llama_memory_context_ptr mctx; if (memory) { LLAMA_LOG_DEBUG("%s: reserving full memory module\n", __func__); @@ -343,7 +360,14 @@ llama_context::llama_context( { auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get()); if (!gf) { - throw std::runtime_error("failed to allocate compute pp buffers"); + if (pipeline_parallel) { + LLAMA_LOG_WARN("%s: compute buffer allocation failed, retrying without pipeline parallelism\n", __func__); + sched.reset(ggml_backend_sched_new(backend_ptrs.data(), backend_buft.data(), backend_ptrs.size(), max_nodes, false, cparams.op_offload)); + gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get()); + } + if (!gf) { + throw std::runtime_error("failed to allocate compute pp buffers"); + } } n_splits_pp = ggml_backend_sched_get_n_splits(sched.get()); @@ -448,8 +472,8 @@ uint32_t llama_context::n_ctx() const { return cparams.n_ctx; } -uint32_t llama_context::n_ctx_per_seq() const { - return cparams.n_ctx / cparams.n_seq_max; +uint32_t llama_context::n_ctx_seq() const { + return cparams.n_ctx_seq; } uint32_t llama_context::n_batch() const { @@ -803,7 +827,7 @@ int llama_context::encode(const llama_batch & batch_inp) { const auto & hparams = model.hparams; - const int64_t n_embd = hparams.n_embd; + const int64_t n_embd = hparams.n_embd_inp(); const int64_t n_vocab = model.vocab.n_tokens(); // note: during encode, we always pass the full sequence starting from pos = 0 @@ -972,7 +996,7 @@ int llama_context::decode(const llama_batch & batch_inp) { const auto & hparams = model.hparams; const int64_t n_vocab = vocab.n_tokens(); - const int64_t n_embd = hparams.n_embd; + const int64_t n_embd = hparams.n_embd_inp(); // when computing embeddings, all tokens are output const bool output_all = cparams.embeddings; @@ -2130,7 +2154,7 @@ void llama_context::opt_epoch_iter( batch.logits [pos_batch] = true; } - if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd, cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) { + if (!balloc->init(batch, model.vocab, nullptr, model.hparams.n_embd_inp(), cparams.kv_unified ? LLAMA_MAX_SEQ : cparams.n_seq_max, true)) { LLAMA_LOG_ERROR("%s: failed to initialize batch\n", __func__); return; } @@ -2378,6 +2402,10 @@ uint32_t llama_n_ctx(const llama_context * ctx) { return ctx->n_ctx(); } +uint32_t llama_n_ctx_seq(const llama_context * ctx) { + return ctx->n_ctx_seq(); +} + uint32_t llama_n_batch(const llama_context * ctx) { return ctx->n_batch(); } diff --git a/examples/talk-llama/llama-context.h b/examples/talk-llama/llama-context.h index ed6d82cb..20cbd789 100644 --- a/examples/talk-llama/llama-context.h +++ b/examples/talk-llama/llama-context.h @@ -43,11 +43,11 @@ struct llama_context { ggml_backend_sched_t get_sched() const; - uint32_t n_ctx() const; - uint32_t n_ctx_per_seq() const; - uint32_t n_batch() const; - uint32_t n_ubatch() const; - uint32_t n_seq_max() const; + uint32_t n_ctx() const; + uint32_t n_ctx_seq() const; + uint32_t n_batch() const; + uint32_t n_ubatch() const; + uint32_t n_seq_max() const; uint32_t n_threads() const; uint32_t n_threads_batch() const; diff --git a/examples/talk-llama/llama-cparams.h b/examples/talk-llama/llama-cparams.h index eae7b839..fcef8fa9 100644 --- a/examples/talk-llama/llama-cparams.h +++ b/examples/talk-llama/llama-cparams.h @@ -8,6 +8,7 @@ struct llama_cparams { uint32_t n_ctx; // context size used during inference + uint32_t n_ctx_seq; // context for a single sequence uint32_t n_batch; uint32_t n_ubatch; uint32_t n_seq_max; diff --git a/examples/talk-llama/llama-graph.cpp b/examples/talk-llama/llama-graph.cpp index 41fa6894..b199e946 100644 --- a/examples/talk-llama/llama-graph.cpp +++ b/examples/talk-llama/llama-graph.cpp @@ -810,6 +810,9 @@ ggml_tensor * llm_graph_context::build_ffn( GGML_ABORT("fatal error"); } + //expand here so that we can fuse ffn gate + ggml_build_forward_expand(gf, cur); + if (gate && type_gate == LLM_FFN_PAR) { cur = ggml_mul(ctx0, cur, tmp); cb(cur, "ffn_gate_par", il); @@ -1006,10 +1009,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn( ggml_tensor * weights_sum = ggml_sum_rows(ctx0, weights); // [1, n_tokens] cb(weights_sum, "ffn_moe_weights_sum", il); - if (arch == LLM_ARCH_BAILINGMOE2) { - weights_sum = ggml_scale_bias(ctx0, weights_sum, 1.0, 1e-20); - cb(weights_sum, "ffn_moe_weights_sum_biased", il); - } + // Avoid division by zero, clamp to smallest number representable by F16 + weights_sum = ggml_clamp(ctx0, weights_sum, 6.103515625e-5, INFINITY); + cb(weights_sum, "ffn_moe_weights_sum_clamped", il); weights = ggml_div(ctx0, weights, weights_sum); // [n_expert_used, n_tokens] cb(weights, "ffn_moe_weights_norm", il); @@ -1091,6 +1093,9 @@ ggml_tensor * llm_graph_context::build_moe_ffn( GGML_ABORT("fatal error"); } + //expand here so that we can fuse ffn gate + ggml_build_forward_expand(gf, cur); + experts = build_lora_mm_id(down_exps, cur, selected_experts); // [n_embd, n_expert_used, n_tokens] cb(experts, "ffn_moe_down", il); @@ -1137,7 +1142,7 @@ ggml_tensor * llm_graph_context::build_moe_ffn( // input embeddings with optional lora ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const { - const int64_t n_embd = hparams.n_embd; + const int64_t n_embd = hparams.n_embd_inp(); auto inp = std::make_unique(); @@ -1274,7 +1279,7 @@ ggml_tensor * llm_graph_context::build_inp_cross_embd() const { // return cur; //} - const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd; + const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp(); const auto n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train; cur = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_enc); @@ -2030,7 +2035,7 @@ int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buck if (bidirectional) { relative_bucket += (relative_position > 0) * n_buckets; - relative_position = abs(relative_position); + relative_position = std::abs(relative_position); } else { relative_position = -std::min(relative_position, 0); } diff --git a/examples/talk-llama/llama-hparams.cpp b/examples/talk-llama/llama-hparams.cpp index db65d69e..8cdbaf69 100644 --- a/examples/talk-llama/llama-hparams.cpp +++ b/examples/talk-llama/llama-hparams.cpp @@ -60,6 +60,16 @@ uint32_t llama_hparams::n_gqa(uint32_t il) const { return n_head/n_head_kv; } +uint32_t llama_hparams::n_embd_inp() const { + uint32_t n_embd_inp = n_embd; + + if (n_deepstack_layers > 0) { + n_embd_inp += n_embd * n_deepstack_layers; + } + + return n_embd_inp; +} + uint32_t llama_hparams::n_embd_k_gqa(uint32_t il) const { const uint32_t n_head_kv = this->n_head_kv(il); @@ -148,7 +158,7 @@ bool llama_hparams::is_recurrent(uint32_t il) const { } uint32_t llama_hparams::n_pos_per_embd() const { - return rope_type == LLAMA_ROPE_TYPE_MROPE ? 4 : 1; + return rope_type == LLAMA_ROPE_TYPE_MROPE || rope_type == LLAMA_ROPE_TYPE_IMROPE ? 4 : 1; } bool llama_hparams::is_swa(uint32_t il) const { diff --git a/examples/talk-llama/llama-hparams.h b/examples/talk-llama/llama-hparams.h index 6fcf91b7..9203af83 100644 --- a/examples/talk-llama/llama-hparams.h +++ b/examples/talk-llama/llama-hparams.h @@ -183,6 +183,9 @@ struct llama_hparams { std::array xielu_beta; std::array xielu_eps; + // qwen3vl deepstack + uint32_t n_deepstack_layers = 0; + // needed by encoder-decoder models (e.g. T5, FLAN-T5) // ref: https://github.com/ggerganov/llama.cpp/pull/8141 llama_token dec_start_token_id = LLAMA_TOKEN_NULL; @@ -224,6 +227,9 @@ struct llama_hparams { uint32_t n_gqa(uint32_t il = 0) const; + // dimension of main + auxiliary input embeddings + uint32_t n_embd_inp() const; + // dimension of key embeddings across all k-v heads uint32_t n_embd_k_gqa(uint32_t il = 0) const; diff --git a/examples/talk-llama/llama-kv-cache-iswa.cpp b/examples/talk-llama/llama-kv-cache-iswa.cpp index facba1d0..3a34102a 100644 --- a/examples/talk-llama/llama-kv-cache-iswa.cpp +++ b/examples/talk-llama/llama-kv-cache-iswa.cpp @@ -45,7 +45,9 @@ llama_kv_cache_iswa::llama_kv_cache_iswa( const uint32_t size_base = kv_size; - uint32_t size_swa = std::min(size_base, GGML_PAD(hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch, n_pad)); + // note: the SWA cache is always padded to 256 for performance + // https://github.com/ggml-org/llama.cpp/issues/17037 + uint32_t size_swa = GGML_PAD(std::min(size_base, hparams.n_swa*(unified ? n_seq_max : 1) + n_ubatch), 256); // when using full-size SWA cache, we set the SWA cache size to be equal to the base cache size if (swa_full) { diff --git a/examples/talk-llama/llama-kv-cache.cpp b/examples/talk-llama/llama-kv-cache.cpp index 736693e1..e26385a1 100644 --- a/examples/talk-llama/llama-kv-cache.cpp +++ b/examples/talk-llama/llama-kv-cache.cpp @@ -8,6 +8,7 @@ #include #include #include +#include #include #include #include @@ -37,8 +38,15 @@ llama_kv_cache::llama_kv_cache( const uint32_t n_layer_kv = hparams.n_layer_kv(); + // define a comparator for the buft -> ctx map to ensure that the order is well-defined: + struct ggml_backend_buft_comparator { + bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const { + return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0; + } + }; + std::map ctx_map; + // create a context for each buffer type - std::map ctx_map; auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { auto it = ctx_map.find(buft); if (it == ctx_map.end()) { @@ -53,13 +61,12 @@ llama_kv_cache::llama_kv_cache( return nullptr; } - ctx_map[buft] = ctx; - ctxs.emplace_back(ctx); + ctx_map.emplace(buft, ctx); return ctx; } - return it->second; + return it->second.get(); }; GGML_ASSERT(n_stream == 1 || n_stream == n_seq_max); @@ -167,11 +174,8 @@ llama_kv_cache::llama_kv_cache( } // allocate tensors and initialize the buffers to avoid NaNs in the padding - for (auto it : ctx_map) { - auto * buft = it.first; - auto * ctx = it.second; - - ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); + for (auto & [buft, ctx] : ctx_map) { + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); if (!buf) { throw std::runtime_error("failed to allocate buffer for kv cache"); } @@ -179,7 +183,7 @@ llama_kv_cache::llama_kv_cache( LLAMA_LOG_INFO("%s: %10s KV buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); ggml_backend_buffer_clear(buf, 0); - bufs.emplace_back(buf); + ctxs_bufs.emplace_back(std::move(ctx), buf); } { @@ -203,7 +207,7 @@ void llama_kv_cache::clear(bool data) { } if (data) { - for (auto & buf : bufs) { + for (auto & [_, buf] : ctxs_bufs) { ggml_backend_buffer_clear(buf.get(), 0); } } @@ -334,6 +338,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll llama_pos pos = v_cells[s0].pos_get(i); llama_pos shift = v_cells[s0].get_shift(i); + llama_kv_cell_ext ext = v_cells[s0].ext_get(i); + if (shift != 0) { pos -= shift; assert(pos >= 0); @@ -345,6 +351,8 @@ void llama_kv_cache::seq_cp(llama_seq_id seq_id_src, llama_seq_id seq_id_dst, ll if (shift != 0) { v_cells[s1].pos_add(i, shift); } + + v_cells[s1].ext_set(i, ext); } } @@ -379,6 +387,7 @@ void llama_kv_cache::seq_keep(llama_seq_id seq_id) { void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos shift) { GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()); + GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_add() is only supported for n_pos_per_embd() == 1"); auto & cells = v_cells[seq_to_stream[seq_id]]; auto & head = v_heads[seq_to_stream[seq_id]]; @@ -423,6 +432,7 @@ void llama_kv_cache::seq_add(llama_seq_id seq_id, llama_pos p0, llama_pos p1, ll void llama_kv_cache::seq_div(llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) { GGML_ASSERT(seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()); + GGML_ASSERT(hparams.n_pos_per_embd() == 1 && "seq_div() is only supported for n_pos_per_embd() == 1"); auto & cells = v_cells[seq_to_stream[seq_id]]; @@ -472,8 +482,8 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const { std::map llama_kv_cache::memory_breakdown() const { std::map ret; - for (const ggml_backend_buffer_ptr & buf_ptr : bufs) { - ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get()); + for (const auto & [_, buf] : ctxs_bufs) { + ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get()); } return ret; } @@ -896,6 +906,14 @@ void llama_kv_cache::apply_ubatch(const slot_info & sinfo, const llama_ubatch & cells.pos_set(idx, ubatch.pos[i]); + if (ubatch.is_pos_2d()) { + llama_kv_cell_ext ext { + /*.x =*/ ubatch.pos[i + ubatch.n_tokens*2], + /*.y =*/ ubatch.pos[i + ubatch.n_tokens], + }; + cells.ext_set(idx, ext); + } + for (int32_t s = 0; s < ubatch.n_seq_id[i]; s++) { cells.seq_add(idx, ubatch.seq_id[i][s]); } @@ -957,10 +975,14 @@ bool llama_kv_cache::get_has_shift() const { uint32_t llama_kv_cache::get_n_kv(const slot_info & sinfo) const { uint32_t result = 0; + // pad the n_kv value so that the graph remains constant across batches and can be reused + // note: this also helps some backends with performance (f.ex https://github.com/ggml-org/llama.cpp/pull/16812#issuecomment-3455112220) + const uint32_t n_pad_cur = std::max(n_pad, 256u); + for (uint32_t s = 0; s < sinfo.n_stream(); ++s) { const auto & cells = v_cells[sinfo.strm[s]]; - result = std::max(std::min(cells.size(), std::max(n_pad, GGML_PAD(cells.used_max_p1(), n_pad))), result); + result = std::max(std::min(cells.size(), std::max(n_pad_cur, GGML_PAD(cells.used_max_p1(), n_pad_cur))), result); } return result; @@ -1239,6 +1261,11 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u const llama_pos p1 = ubatch->pos[i]; + // for M-RoPE + const bool is_2d = ubatch->is_pos_2d(); + const llama_pos p1_x = is_2d ? ubatch->pos[i + ubatch->n_tokens*2] : 0; + const llama_pos p1_y = is_2d ? ubatch->pos[i + ubatch->n_tokens] : 0; + const uint64_t idst = n_kv*(h*n_stream*n_tps_pad + s*n_tps_pad + ii); for (uint32_t j = 0; j < n_kv; ++j) { @@ -1258,6 +1285,14 @@ void llama_kv_cache::set_input_kq_mask(ggml_tensor * dst, const llama_ubatch * u continue; } + // M-RoPE causal mask + if (causal_attn && is_2d && p0 == p1) { + const auto & p0_ext = cells.ext_get(j); + if (p0_ext.is_2d_gt(p1_x, p1_y)) { + continue; + } + } + // apply SWA if any if (is_masked_swa(p0, p1)) { continue; @@ -1298,7 +1333,7 @@ void llama_kv_cache::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch size_t llama_kv_cache::total_size() const { size_t size = 0; - for (const auto & buf : bufs) { + for (const auto & [_, buf] : ctxs_bufs) { size += ggml_backend_buffer_get_size(buf.get()); } @@ -1340,7 +1375,7 @@ ggml_tensor * llama_kv_cache::build_rope_shift( const auto & yarn_beta_slow = cparams.yarn_beta_slow; const auto & n_rot = hparams.n_rot; - const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE + const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE // @ngxson : this is a workaround // for M-RoPE, we want to rotate the whole vector when doing KV shift // a normal RoPE should work, we just need to use the correct ordering @@ -1551,6 +1586,9 @@ void llama_kv_cache::state_write_meta(llama_io_write_i & io, const cell_ranges_t io.write(&pos, sizeof(pos)); io.write(&n_seq_id, sizeof(n_seq_id)); + // TODO: we also need to save llama_kv_cell_ext when apply_ubatch() support loading it + // see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350 + for (const auto & seq_id : seq_ids) { io.write(&seq_id, sizeof(seq_id)); } @@ -1696,6 +1734,8 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32 return false; } + // TODO: we cannot yet restore llama_kv_cell_ext as the apply_ubatch() does not support it yet + // see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350 apply_ubatch(sinfo, ubatch); const auto head_cur = sinfo.head(); @@ -2010,8 +2050,3 @@ void llama_kv_cache_context::set_input_kq_mask(ggml_tensor * dst, const llama_ub void llama_kv_cache_context::set_input_pos_bucket(ggml_tensor * dst, const llama_ubatch * ubatch) const { kv->set_input_pos_bucket(dst, ubatch); } - -uint32_t llama_kv_cache::get_padding(const llama_cparams & cparams) { - // the FA kernels require padding to avoid extra runtime boundary checks - return cparams.flash_attn ? 256u : 32u; -} diff --git a/examples/talk-llama/llama-kv-cache.h b/examples/talk-llama/llama-kv-cache.h index 85f0663d..bf7821c0 100644 --- a/examples/talk-llama/llama-kv-cache.h +++ b/examples/talk-llama/llama-kv-cache.h @@ -19,8 +19,6 @@ struct llama_context; class llama_kv_cache : public llama_memory_i { public: - static uint32_t get_padding(const llama_cparams & cparams); - struct stream_copy_info { bool empty() const { assert(ssrc.size() == sdst.size()); @@ -217,8 +215,8 @@ private: // this is the SWA type of the cache - not to be confused with the model SWA type const llama_swa_type swa_type = LLAMA_SWA_TYPE_NONE; - std::vector ctxs; - std::vector bufs; + // ggml contexts for the KV cache along with the allocated backend buffers: + std::vector> ctxs_bufs; // the current index from where we start searching for a free slot in the ring buffer of KV cells (see find_slot()) // note: this is not part of the KV state and it's only used to speed-up the find_slot() method diff --git a/examples/talk-llama/llama-kv-cells.h b/examples/talk-llama/llama-kv-cells.h index 8f6bf014..10063bf4 100644 --- a/examples/talk-llama/llama-kv-cells.h +++ b/examples/talk-llama/llama-kv-cells.h @@ -5,9 +5,27 @@ #include #include -#include -#include +#include #include +#include +#include + +struct llama_kv_cell_ext { + // 2D spatial positions, typically used for M-RoPE + llama_pos x = 0; + llama_pos y = 0; + + // return true if the current 2D spatial position is greater than other + bool is_2d_gt(llama_pos ox, llama_pos oy) const { + return (y > oy) || (y == oy && x > ox); + } + + void reset() { + static_assert(std::is_trivially_copyable_v); + + memset(this, 0, sizeof(*this)); + } +}; // meta information about KV cells that can be part of multiple sequences at the same time // TODO: add unit tests @@ -16,6 +34,7 @@ public: void reset() { for (uint32_t i = 0; i < pos.size(); ++i) { pos[i] = -1; + ext[i].reset(); shift[i] = 0; seq[i].reset(); } @@ -43,6 +62,7 @@ public: void resize(uint32_t n) { pos.resize(n); + ext.resize(n); shift.resize(n); seq.resize(n); @@ -108,6 +128,7 @@ public: const auto idx = i + j; res.pos[j] = pos[idx]; + res.ext[j] = ext[idx]; res.seq[j] = seq[idx]; assert(shift[idx] == 0); @@ -126,6 +147,7 @@ public: const auto idx = idxs[j]; res.pos[j] = pos[idx]; + res.ext[j] = ext[idx]; res.seq[j] = seq[idx]; assert(shift[idx] == 0); @@ -154,6 +176,7 @@ public: } pos[idx] = other.pos[j]; + ext[idx] = other.ext[j]; seq[idx] = other.seq[j]; if (pos[idx] != -1) { @@ -184,6 +207,7 @@ public: } pos[idx] = other.pos[j]; + ext[idx] = other.ext[j]; seq[idx] = other.seq[j]; if (pos[idx] != -1) { @@ -203,6 +227,7 @@ public: seq[i].reset(); pos[i] = -1; + ext[i].reset(); shift[i] = 0; used.erase(i); @@ -221,6 +246,7 @@ public: if (seq[i].none()) { pos[i] = -1; + ext[i].reset(); shift[i] = 0; used.erase(i); @@ -250,6 +276,7 @@ public: seq[i].reset(); pos[i] = -1; + ext[i].reset(); shift[i] = 0; used.erase(i); @@ -340,6 +367,13 @@ public: return pos[i]; } + const llama_kv_cell_ext & ext_get(uint32_t i) const { + assert(i < pos.size()); + assert(pos[i] != -1); + + return ext[i]; + } + // note: call only if the cell is not empty llama_pos get_shift(uint32_t i) const { assert(i < pos.size()); @@ -368,6 +402,11 @@ public: used.insert(i); } + void ext_set(uint32_t i, llama_kv_cell_ext p) { + assert(i < ext.size()); + ext[i] = p; + } + // pos[i] = pos[i] + d // sets "has_shift" to true // note: call only if the cell is not empty @@ -424,6 +463,9 @@ private: std::vector pos; + // stores extra info per cell + std::vector ext; + // this array accumulates any applied shifts to the pos array since the last reset_shift() call // this is used to queue multiple updates to the pos array, which in the end can be applied in one go: // diff --git a/examples/talk-llama/llama-memory-recurrent.cpp b/examples/talk-llama/llama-memory-recurrent.cpp index d67f5a5f..276e1697 100644 --- a/examples/talk-llama/llama-memory-recurrent.cpp +++ b/examples/talk-llama/llama-memory-recurrent.cpp @@ -7,6 +7,7 @@ #include #include +#include #include #include #include @@ -32,8 +33,15 @@ llama_memory_recurrent::llama_memory_recurrent( cells.clear(); cells.resize(mem_size); + // define a comparator for the buft -> ctx map to ensure that the order is well-defined: + struct ggml_backend_buft_comparator { + bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const { + return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0; + } + }; + std::map ctx_map; + // create a context for each buffer type - std::map ctx_map; auto ctx_for_buft = [&](ggml_backend_buffer_type_t buft) -> ggml_context * { auto it = ctx_map.find(buft); if (it == ctx_map.end()) { @@ -48,13 +56,12 @@ llama_memory_recurrent::llama_memory_recurrent( return nullptr; } - ctx_map[buft] = ctx; - ctxs.emplace_back(ctx); + ctx_map.emplace(buft, ctx); return ctx; } - return it->second; + return it->second.get(); }; r_l.resize(n_layer); @@ -93,17 +100,14 @@ llama_memory_recurrent::llama_memory_recurrent( } // allocate tensors and initialize the buffers to avoid NaNs in the padding - for (auto it : ctx_map) { - auto * buft = it.first; - auto * ctx = it.second; - - ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); + for (auto & [buft, ctx] : ctx_map) { + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); if (!buf) { throw std::runtime_error("failed to allocate buffer for rs cache"); } ggml_backend_buffer_clear(buf, 0); LLAMA_LOG_INFO("%s: %10s RS buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf), ggml_backend_buffer_get_size(buf)/1024.0/1024.0); - bufs.emplace_back(buf); + ctxs_bufs.emplace_back(std::move(ctx), buf); } { @@ -129,7 +133,7 @@ void llama_memory_recurrent::clear(bool data) { used = 0; if (data) { - for (auto & buf : bufs) { + for (auto & [_, buf] : ctxs_bufs) { ggml_backend_buffer_clear(buf.get(), 0); } } @@ -364,8 +368,8 @@ llama_pos llama_memory_recurrent::seq_pos_max(llama_seq_id seq_id) const { std::map llama_memory_recurrent::memory_breakdown() const { std::map ret; - for (const ggml_backend_buffer_ptr & buf_ptr : bufs) { - ret[ggml_backend_buffer_get_type(buf_ptr.get())] += ggml_backend_buffer_get_size(buf_ptr.get()); + for (const auto & [_, buf] : ctxs_bufs) { + ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get()); } return ret; } @@ -662,7 +666,7 @@ bool llama_memory_recurrent::get_can_shift() const { size_t llama_memory_recurrent::total_size() const { size_t size = 0; - for (const auto & buf : bufs) { + for (const auto & [_, buf] : ctxs_bufs) { size += ggml_backend_buffer_get_size(buf.get()); } diff --git a/examples/talk-llama/llama-memory-recurrent.h b/examples/talk-llama/llama-memory-recurrent.h index 077c6e3c..47f01d73 100644 --- a/examples/talk-llama/llama-memory-recurrent.h +++ b/examples/talk-llama/llama-memory-recurrent.h @@ -109,8 +109,8 @@ private: const uint32_t n_seq_max = 1; - std::vector ctxs; - std::vector bufs; + // ggml contexts for the KV cache along with the allocated backend buffers: + std::vector> ctxs_bufs; size_t total_size() const; diff --git a/examples/talk-llama/llama-model.cpp b/examples/talk-llama/llama-model.cpp index e4609963..829f1e3c 100644 --- a/examples/talk-llama/llama-model.cpp +++ b/examples/talk-llama/llama-model.cpp @@ -13,9 +13,10 @@ #include "ggml-cpp.h" +#include "models/models.h" + #include #include -#include #include #include #include @@ -121,6 +122,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_30B_A3B: return "30B.A3B"; case LLM_TYPE_100B_A6B: return "100B.A6B"; case LLM_TYPE_106B_A12B: return "106B.A12B"; + case LLM_TYPE_230B_A10B: return "230B.A10B"; case LLM_TYPE_235B_A22B: return "235B.A22B"; case LLM_TYPE_300B_A47B: return "300B.A47B"; case LLM_TYPE_355B_A32B: return "355B.A32B"; @@ -274,8 +276,8 @@ static bool weight_buft_supported(const llama_hparams & hparams, ggml_tensor * w } break; case GGML_OP_IM2COL: { - const int n_embd = hparams.n_embd; - ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd, w->ne[1], 1, 1); + const int n_embd_inp = hparams.n_embd_inp(); + ggml_tensor * b = ggml_new_tensor_4d(ctx, GGML_TYPE_F32, n_embd_inp, w->ne[1], 1, 1); op_tensor = ggml_im2col(ctx, w, b, 1, 0, 0, 0, 1, 0, false, GGML_TYPE_F16); } break; case GGML_OP_SCALE: @@ -404,6 +406,19 @@ static buft_list_t make_gpu_buft_list(ggml_backend_dev_t dev, llama_split_mode s // add the device default buffer type buft_list.emplace_back(dev, ggml_backend_dev_buffer_type(dev)); + // add the device extra buffer type (if any) + ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(dev); + auto ggml_backend_dev_get_extra_bufts_fn = (ggml_backend_dev_get_extra_bufts_t) + ggml_backend_reg_get_proc_address(reg, "ggml_backend_dev_get_extra_bufts"); + + if (ggml_backend_dev_get_extra_bufts_fn) { + ggml_backend_buffer_type_t * extra_bufts = ggml_backend_dev_get_extra_bufts_fn(dev); + while (extra_bufts && *extra_bufts) { + buft_list.emplace_back(dev, *extra_bufts); + ++extra_bufts; + } + } + return buft_list; } @@ -425,7 +440,7 @@ struct llama_model::impl { llama_mlocks mlock_mmaps; // contexts where the model tensors metadata is stored as well ass the corresponding buffers: - std::vector> ctxs_bufs; + std::vector>> ctxs_bufs; buft_list_t cpu_buft_list; std::map gpu_buft_list; @@ -1013,10 +1028,34 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_QWEN3VL: + { + ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false); + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 28: type = LLM_TYPE_1_7B; break; + case 36: type = hparams.n_embd == 2560 ? LLM_TYPE_4B : LLM_TYPE_8B; break; + case 64: type = LLM_TYPE_32B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_QWEN3MOE: { ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 48: type = LLM_TYPE_30B_A3B; break; + case 94: type = LLM_TYPE_235B_A22B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_QWEN3VLMOE: + { + ml.get_key(LLM_KV_NUM_DEEPSTACK_LAYERS, hparams.n_deepstack_layers, false); + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false); ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_layer) { case 48: type = LLM_TYPE_30B_A3B; break; @@ -1856,7 +1895,8 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); switch (hparams.n_embd) { - case 1536: type = LLM_TYPE_7B_A1B; break; + case 768: type = LLM_TYPE_350M; break; + case 1536: type = (hparams.n_embd == 2048 ? LLM_TYPE_7B_A1B : LLM_TYPE_1B); break; case 2048: case 2560: type = LLM_TYPE_3B; break; case 4096: type = LLM_TYPE_32B; break; default: type = LLM_TYPE_UNKNOWN; @@ -2112,6 +2152,34 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_MINIMAX_M2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false); + + switch (hparams.n_layer) { + case 62: type = LLM_TYPE_230B_A10B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_COGVLM: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 32: type = LLM_TYPE_13B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_PANGU_EMBED: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + switch (hparams.n_layer) { + case 26: type = LLM_TYPE_1B; break; // openPangu-Embedded-1B-V1.1 + case 34: type = LLM_TYPE_7B; break; // openPangu-Embedded-7B-V1.1 + default: type = LLM_TYPE_UNKNOWN; + } + } break; default: throw std::runtime_error("unsupported model architecture"); } @@ -2219,7 +2287,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { // define a comparator for the buft -> ctx map to ensure that the order is well-defined: struct ggml_backend_buft_comparator { bool operator()(const ggml_backend_buffer_type_t & lhs, const ggml_backend_buffer_type_t & rhs) const { - return ggml_backend_buft_name(lhs) < ggml_backend_buft_name(rhs); + return strcmp(ggml_backend_buft_name(lhs), ggml_backend_buft_name(rhs)) < 0; } }; std::map ctx_map; @@ -3265,6 +3333,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } } break; case LLM_ARCH_QWEN3: + case LLM_ARCH_QWEN3VL: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -3299,6 +3368,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } } break; case LLM_ARCH_QWEN3MOE: + case LLM_ARCH_QWEN3VLMOE: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -6124,6 +6194,114 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED); } } break; + case LLM_ARCH_MINIMAX_M2: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0); + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_gqa }, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_gqa }, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0); + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k * n_head}, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_k_gqa}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0); + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff, n_embd, n_expert}, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0); + layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0); + } + } break; + case LLM_ARCH_COGVLM: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd_head_k * n_head * 3}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.visexp_attn_wqkv = create_tensor(tn(LLM_TENSOR_VISEXP_ATTN_QKV, "weight", i), {n_embd, n_embd_head_k * n_head * 3}, 0); + layer.visexp_attn_wo = create_tensor(tn(LLM_TENSOR_VISEXP_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + + layer.visexp_ffn_gate = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.visexp_ffn_down = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.visexp_ffn_up = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } + } break; + case LLM_ARCH_PANGU_EMBED: + { + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); + + // output + output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + + // if output is NULL, init from the input tok embed + if (output == NULL) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0); + + // weight tensors + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0); + layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0); + layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0); + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0); + + // bias tensors + layer.bq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd_head_k * n_head}, 0); + layer.bk = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd_gqa}, 0); + layer.bv = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd_gqa}, 0); + layer.bo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0); + + if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) { + layer.rope_long = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + } else { + layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0)); + } + + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, 0); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0); + } + } break; default: throw std::runtime_error("unknown architecture"); } @@ -6173,7 +6351,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) { bool buffer_from_host_ptr_supported = props.caps.buffer_from_host_ptr; bool is_default_buft = buft == ggml_backend_dev_buffer_type(dev); - ggml_backend_buffer_t buf = nullptr; + std::vector bufs; if (ml.use_mmap && use_mmap_buffer && buffer_from_host_ptr_supported && is_default_buft) { for (uint32_t idx = 0; idx < ml.files.size(); idx++) { // only the mmap region containing the tensors in the model is mapped to the backend buffer @@ -6186,15 +6364,16 @@ bool llama_model::load_tensors(llama_model_loader & ml) { continue; } const size_t max_size = ggml_get_max_tensor_size(ctx); - buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size); + ggml_backend_buffer_t buf = ggml_backend_dev_buffer_from_host_ptr(dev, (char *) addr + first, last - first, max_size); if (buf == nullptr) { throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); } + bufs.emplace_back(buf); buf_map.emplace(idx, buf); } } else { - buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); + ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx, buft); if (buf == nullptr) { throw std::runtime_error(format("unable to allocate %s buffer", ggml_backend_buft_name(buft))); } @@ -6204,11 +6383,12 @@ bool llama_model::load_tensors(llama_model_loader & ml) { mlock_buf->init (ggml_backend_buffer_get_base(buf)); mlock_buf->grow_to(ggml_backend_buffer_get_size(buf)); } + bufs.emplace_back(buf); for (uint32_t idx = 0; idx < ml.files.size(); idx++) { buf_map.emplace(idx, buf); } } - pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), buf); + pimpl->ctxs_bufs.emplace_back(std::move(ctx_ptr), std::move(bufs)); for (auto & buf : buf_map) { // indicate that this buffer contains weights @@ -6234,8 +6414,11 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } // print memory requirements per buffer type - for (auto & [_, buf] : pimpl->ctxs_bufs) { - LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n", __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0); + for (auto & [_, bufs] : pimpl->ctxs_bufs) { + for (auto & buf: bufs) { + LLAMA_LOG_INFO("%s: %12s model buffer size = %8.2f MiB\n", + __func__, ggml_backend_buffer_name(buf.get()), ggml_backend_buffer_get_size(buf.get()) / 1024.0 / 1024.0); + } } // populate tensors_by_name @@ -6287,8 +6470,10 @@ size_t llama_model::n_devices() const { std::map llama_model::memory_breakdown() const { std::map ret; - for (const auto & [_, buf] : pimpl->ctxs_bufs) { - ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get()); + for (const auto & [_, bufs] : pimpl->ctxs_bufs) { + for (const auto & buf : bufs) { + ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get()); + } } return ret; } @@ -6336,6 +6521,7 @@ void llama_model::print_info() const { if (!hparams.vocab_only) { LLAMA_LOG_INFO("%s: n_ctx_train = %u\n", __func__, hparams.n_ctx_train); LLAMA_LOG_INFO("%s: n_embd = %u\n", __func__, hparams.n_embd); + LLAMA_LOG_INFO("%s: n_embd_inp = %u\n", __func__, hparams.n_embd_inp()); LLAMA_LOG_INFO("%s: n_layer = %u\n", __func__, hparams.n_layer); LLAMA_LOG_INFO("%s: n_head = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head(il); }, hparams.n_layer).c_str()); LLAMA_LOG_INFO("%s: n_head_kv = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_head_kv(il); }, hparams.n_layer).c_str()); @@ -6356,6 +6542,8 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_ff = %s\n", __func__, print_f([&](uint32_t il) { return hparams.n_ff(il); }, hparams.n_layer).c_str()); LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert); LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used); + LLAMA_LOG_INFO("%s: n_expert_groups = %d\n", __func__, hparams.n_expert_groups); + LLAMA_LOG_INFO("%s: n_group_used = %d\n", __func__, hparams.n_group_used); LLAMA_LOG_INFO("%s: causal attn = %d\n", __func__, hparams.causal_attn); LLAMA_LOG_INFO("%s: pooling type = %d\n", __func__, hparams.pooling_type); LLAMA_LOG_INFO("%s: rope type = %d\n", __func__, hparams.rope_type); @@ -6364,6 +6552,10 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: freq_scale_train = %g\n", __func__, hparams.rope_freq_scale_train); LLAMA_LOG_INFO("%s: n_ctx_orig_yarn = %u\n", __func__, hparams.n_ctx_orig_yarn); LLAMA_LOG_INFO("%s: rope_finetuned = %s\n", __func__, hparams.rope_finetuned ? "yes" : "unknown"); + // MRoPE (Multi-axis Rotary Position Embedding) sections + if (const auto & s = hparams.rope_sections; s[0] || s[1] || s[2] || s[3]) { + LLAMA_LOG_INFO("%s: mrope sections = [%d, %d, %d, %d]\n", __func__, s[0], s[1], s[2], s[3]); + } if (!classifier_labels.empty()) { LLAMA_LOG_INFO("%s: n_cls_out = %u\n", __func__, hparams.n_cls_out); @@ -6429,7 +6621,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp); } - if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE) { + if (arch == LLM_ARCH_QWEN3MOE || arch == LLM_ARCH_OPENAI_MOE || arch == LLM_ARCH_QWEN3VLMOE) { LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp); } @@ -6456,8 +6648,6 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: n_ff_exp = %d\n", __func__, hparams.n_ff_exp); LLAMA_LOG_INFO("%s: n_ff_shexp = %d\n", __func__, hparams.n_ff_shexp); LLAMA_LOG_INFO("%s: n_expert_shared = %d\n", __func__, hparams.n_expert_shared); - LLAMA_LOG_INFO("%s: n_expert_groups = %d\n", __func__, hparams.n_expert_groups); - LLAMA_LOG_INFO("%s: n_group_used = %d\n", __func__, hparams.n_group_used); LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); LLAMA_LOG_INFO("%s: expert_weights_norm = %d\n", __func__, hparams.expert_weights_norm); LLAMA_LOG_INFO("%s: expert_gating_func = %s\n", __func__, llama_expert_gating_func_name((llama_expert_gating_func_type) hparams.expert_gating_func)); @@ -6562,13206 +6752,142 @@ float llama_model::get_rope_freq_scale(const llama_cparams & cparams, int il) co } ggml_tensor * llama_model::get_rope_factors(const llama_cparams & cparams, int il) const { - const uint32_t n_ctx_per_seq = cparams.n_ctx / cparams.n_seq_max; + const uint32_t n_ctx_seq = cparams.n_ctx_seq; // choose long/short freq factors based on the context size if (layers[il].rope_freqs != nullptr) { return layers[il].rope_freqs; } - if (n_ctx_per_seq > hparams.n_ctx_orig_yarn) { + if (n_ctx_seq > hparams.n_ctx_orig_yarn) { return layers[il].rope_long; } return layers[il].rope_short; } -struct llm_build_llama : public llm_graph_context { - llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); +llama_memory_i * llama_model::create_memory(const llama_memory_params & params, const llama_cparams & cparams) const { + llama_memory_i * res; - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + switch (arch) { + // Models that need specific instantiation should be handled in the + // switch statement + case LLM_ARCH_BERT: + case LLM_ARCH_JINA_BERT_V2: + case LLM_ARCH_JINA_BERT_V3: + case LLM_ARCH_NOMIC_BERT: + case LLM_ARCH_NOMIC_BERT_MOE: + case LLM_ARCH_NEO_BERT: + case LLM_ARCH_WAVTOKENIZER_DEC: + case LLM_ARCH_GEMMA_EMBEDDING: + case LLM_ARCH_DREAM: + case LLM_ARCH_LLADA: + case LLM_ARCH_LLADA_MOE: + { + res = nullptr; + } break; + // Models that need standard caching should rely on recurrent/hybrid + // checks + default: + { + if (llm_arch_is_recurrent(arch)) { + res = new llama_memory_recurrent( + *this, + GGML_TYPE_F32, + GGML_TYPE_F32, + cparams.offload_kqv, + std::max((uint32_t) 1, cparams.n_seq_max), + cparams.n_seq_max, + nullptr); + } else if (llm_arch_is_hybrid(arch)) { - ggml_tensor * inp_out_ids = build_inp_out_ids(); + // The main difference between hybrid architectures is the + // layer filters, so pick the right one here + llama_memory_hybrid::layer_filter_cb filter_attn = nullptr; + llama_memory_hybrid::layer_filter_cb filter_recr = nullptr; + if (arch == LLM_ARCH_FALCON_H1) { + filter_attn = [&](int32_t) { return true; }; + filter_recr = [&](int32_t) { return true; }; + } else if (arch == LLM_ARCH_NEMOTRON_H) { + filter_attn = [&](int32_t il) { + return !hparams.is_recurrent(il) && hparams.n_ff(il) == 0; + }; + filter_recr = [&](int32_t il) { + return hparams.is_recurrent(il) && hparams.n_ff(il) == 0; + }; + } - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; + res = new llama_memory_hybrid( + /* model */ *this, + /* attn_type_k */ params.type_k, + /* attn_type_v */ params.type_v, + /* attn_v_trans */ !cparams.flash_attn, + /* attn_kv_size */ cparams.n_ctx, + /* attn_n_pad */ 1, + /* attn_n_swa */ hparams.n_swa, + /* attn_swa_type */ hparams.swa_type, + /* recurrent_type_k */ GGML_TYPE_F32, + /* recurrent_type_v */ GGML_TYPE_F32, + /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max), + /* n_seq_max */ cparams.n_seq_max, + /* offload */ cparams.offload_kqv, + /* unified */ cparams.kv_unified, + /* filter_attn */ std::move(filter_attn), + /* filter_recr */ std::move(filter_recr)); + } else { + llama_memory_i::layer_reuse_cb reuse = nullptr; - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); + if (arch == LLM_ARCH_GEMMA3N) { + reuse = [&](int32_t il) { + if (il >= (int32_t) hparams.n_layer_kv_from_start) { + return (int32_t) hparams.n_layer_kv_from_start - (hparams.is_swa(il) ? 2 : 1); + } - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } + return -1; + }; + } - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } + if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) { + GGML_ASSERT(hparams.is_swa_any()); - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } + res = new llama_kv_cache_iswa( + *this, + params.type_k, + params.type_v, + !cparams.flash_attn, + cparams.offload_kqv, + params.swa_full, + cparams.kv_unified, + cparams.n_ctx_seq, + cparams.n_seq_max, + cparams.n_ubatch, + 1, + nullptr, + reuse); + } else { + GGML_ASSERT(!hparams.is_swa_any()); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - if (hparams.use_kq_norm) { - // Llama4TextL2Norm - Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps); - Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps); - cb(Qcur, "Qcur_normed", il); - cb(Kcur, "Kcur_normed", il); + res = new llama_kv_cache( + *this, + params.type_k, + params.type_v, + !cparams.flash_attn, + cparams.offload_kqv, + cparams.kv_unified, + cparams.n_ctx_seq, + cparams.n_seq_max, + 1, + hparams.n_swa, + hparams.swa_type, + nullptr, + nullptr); + } } - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); } + } - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network (non-MoE) - if (model.layers[il].ffn_gate_inp == nullptr) { - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(cur, "ffn_moe_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_llama_iswa : public llm_graph_context { - llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - // temperature tuning - ggml_tensor * inp_attn_scale = nullptr; - inp_attn_scale = build_inp_attn_scale(); - - auto * inp_attn = build_attn_inp_kv_iswa(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - const bool use_rope = hparams.n_no_rope_layer_step > 0 && - (il + 1) % hparams.n_no_rope_layer_step != 0; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - if (use_rope) { - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } else if (inp_attn_scale) { - Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - if (use_rope && hparams.use_kq_norm) { - // Llama4TextL2Norm - Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps); - Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps); - cb(Qcur, "Qcur_normed", il); - cb(Kcur, "Kcur_normed", il); - } - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network (non-MoE) - if (model.layers[il].ffn_gate_inp == nullptr) { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - ggml_tensor * ffn_inp_normed = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, - il); - - // Shared experts - ggml_tensor * shexp_out = build_ffn(ffn_inp_normed, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(shexp_out, "ffn_moe_shexp", il); - - cur = ggml_add(ctx0, moe_out, shexp_out); - cb(cur, "ffn_moe_out_merged", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_deci : public llm_graph_context { - llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - const int64_t n_head_kv = hparams.n_head_kv(il); - const int64_t n_head = hparams.n_head(il); - const int64_t n_ff = hparams.n_ff(il); - - if (n_head == 0) { - // attention-free layer of Llama-3_1-Nemotron-51B - cur = inpL; - } else { - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - } - - if (n_head > 0 && n_head_kv == 0) { - // "linear attention" of Llama-3_1-Nemotron-51B - cur = build_lora_mm(model.layers[il].wo, cur); - cb(cur, "wo", il); - } else if (n_head > 0) { - // self-attention - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B - if (n_ff == 0) { - continue; - } - - // modified to support attention-free layer of Llama-3_1-Nemotron-51B - ggml_tensor * ffn_inp = cur; - if (n_head > 0) { - ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - } - - // feed-forward network - if (model.layers[il].ffn_gate_inp == nullptr) { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_baichuan : public llm_graph_context { - llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr; - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - switch (model.type) { - case LLM_TYPE_7B: - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - break; - case LLM_TYPE_13B: - break; - default: - GGML_ABORT("fatal error"); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_xverse : public llm_graph_context { - llm_build_xverse(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_falcon : public llm_graph_context { - llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * attn_norm; - - attn_norm = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(attn_norm, "attn_norm", il); - - // self-attention - { - if (model.layers[il].attn_norm_2) { - // Falcon-40B - cur = build_norm(inpL, - model.layers[il].attn_norm_2, - model.layers[il].attn_norm_2_b, - LLM_NORM, il); - cb(cur, "attn_norm_2", il); - } else { - cur = attn_norm; - } - - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - // using mode = 2 for neox mode - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids); - } - - ggml_tensor * ffn_inp = cur; - - // feed forward - { - cur = build_ffn(attn_norm, // !! use the attn norm, not the result - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = ggml_add(ctx0, cur, inpL); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - // norm - cur = build_norm(cur, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_grok : public llm_graph_context { - llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - cur = build_norm(cur, - model.layers[il].attn_out_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_out_norm", il); - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // MoE branch - ggml_tensor * moe_out = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_GELU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - - if (model.layers[il].ffn_up) { - ggml_tensor * ffn_out = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_PAR, il); - cb(ffn_out, "ffn_out", il); - - cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2); - cb(cur, "ffn_out", il); - } else { - cur = moe_out; - } - - cur = build_norm(cur, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_post_norm", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cur = ggml_scale(ctx0, cur, hparams.f_logit_scale); - - // final logit soft-capping - if (hparams.f_final_logit_softcapping) { - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); - cur = ggml_tanh(ctx0, cur); - cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); - } - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_dbrx : public llm_graph_context { - llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = nullptr; - ggml_tensor * Kcur = nullptr; - ggml_tensor * Vcur = nullptr; - - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(cur, "wqkv_clamped", il); - - Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].attn_out_norm, NULL, - LLM_NORM, il); - cb(cur, "attn_out_norm", il); - - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(cur, "ffn_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_starcoder : public llm_graph_context { - llm_build_starcoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); - cb(pos, "pos_embd", -1); - - inpL = ggml_add(ctx0, inpL, pos); - cb(inpL, "inpL", -1); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - cur = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // add the input - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = build_norm(inpL, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_refact : public llm_graph_context { - llm_build_refact(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_bert : public llm_graph_context { - llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - ggml_tensor * inp_pos = nullptr; - - if (model.arch != LLM_ARCH_JINA_BERT_V2) { - inp_pos = build_inp_pos(); - } - - // construct input embeddings (token, type, position) - inpL = build_inp_embd(model.tok_embd); - - // token types are hardcoded to zero ("Sentence A") - if (model.type_embd) { - ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0); - inpL = ggml_add(ctx0, inpL, type_row0); - } - if (model.arch == LLM_ARCH_BERT) { - inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL); - } - cb(inpL, "inp_embd", -1); - - // embed layer norm - inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); - cb(inpL, "inp_norm", -1); - - auto * inp_attn = build_attn_inp_no_cache(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * cur = inpL; - - { - ggml_tensor * Qcur; - ggml_tensor * Kcur; - ggml_tensor * Vcur; - - // self-attention - if (model.layers[il].wqkv) { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - if (model.layers[il].bqkv) { - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - } - - Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - } else { - Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); - Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); - Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - } - - if (model.layers[il].attn_q_norm) { - Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head*n_head, n_tokens); - - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - } - - if (model.layers[il].attn_k_norm) { - Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head*n_head_kv, n_tokens); - - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, il); - - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - } - - // RoPE - if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE || model.arch == LLM_ARCH_JINA_BERT_V3) { - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "kqv_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // re-add the layer input - cur = ggml_add(ctx0, cur, inpL); - - // attention layer norm - cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il); - - if (model.layers[il].attn_norm_2 != nullptr) { - cur = ggml_add(ctx0, cur, inpL); // re-add the layer input - cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il); - } - - ggml_tensor * ffn_inp = cur; - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) { - // MoE branch - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - nullptr, - model.layers[il].ffn_down_exps, - nullptr, - hparams.n_expert, - hparams.n_expert_used, - LLM_FFN_GELU, - false, false, - 0.0f, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il); - cb(cur, "ffn_moe_out", il); - } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE || model.arch == LLM_ARCH_JINA_BERT_V3) { - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - } else if (model.arch == LLM_ARCH_JINA_BERT_V2) { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - // attentions bypass the intermediate layer - cur = ggml_add(ctx0, cur, ffn_inp); - - // output layer norm - cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cb(cur, "result_embd", -1); - res->t_embd = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_neo_bert : public llm_graph_context { - llm_build_neo_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - ggml_tensor * inp_pos = build_inp_pos(); - - // construct input embeddings (token, type, position) - inpL = build_inp_embd(model.tok_embd); - cb(inpL, "inp_embd", -1); - - auto * inp_attn = build_attn_inp_no_cache(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * cur = inpL; - - // pre-norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - - { - ggml_tensor * Qcur; - ggml_tensor * Kcur; - ggml_tensor * Vcur; - - // self-attention - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - // RoPE - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, nullptr, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "kqv_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // re-add the layer input - cur = ggml_add(ctx0, cur, inpL); - - ggml_tensor * ffn_inp = cur; - cb(ffn_inp, "ffn_inp", il); - - // pre-norm - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - cur = build_ffn(cur, - model.layers[il].ffn_up, - NULL, NULL, NULL, NULL, NULL, - model.layers[il].ffn_down, - NULL, NULL, NULL, - LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); - - // attentions bypass the intermediate layer - cur = ggml_add(ctx0, cur, ffn_inp); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm_enc, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_embd", -1); - res->t_embd = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_bloom : public llm_graph_context { - llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - auto * inp_attn = build_attn_inp_kv(); - - inpL = build_norm(inpL, - model.tok_norm, - model.tok_norm_b, - LLM_NORM, -1); - cb(inpL, "inp_norm", -1); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - cur = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // Add the input - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = build_norm(inpL, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_mpt : public llm_graph_context { - llm_build_mpt(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * pos; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - auto * inp_attn = build_attn_inp_kv(); - - if (model.pos_embd) { - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); - cb(pos, "pos_embd", -1); - - inpL = ggml_add(ctx0, inpL, pos); - cb(inpL, "inpL", -1); - } - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * attn_norm; - - attn_norm = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(attn_norm, "attn_norm", il); - - // self-attention - { - cur = attn_norm; - - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - if (model.layers[il].bqkv){ - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - } - - if (hparams.f_clamp_kqv > 0.0f) { - cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(cur, "wqkv_clamped", il); - } - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - // Q/K Layernorm - if (model.layers[il].attn_q_norm) { - Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head*n_head, n_tokens); - Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head*n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, il); - - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // Add the input - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // feed forward - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - model.layers[il].ffn_act, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_stablelm : public llm_graph_context { - llm_build_stablelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - ggml_tensor * inpSA = cur; - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - if (model.layers[il].attn_q_norm) { - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, - NULL, - LLM_NORM, il); - cb(Qcur, "Qcur", il); - } - - if (model.layers[il].attn_k_norm) { - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, - NULL, - LLM_NORM, il); - cb(Kcur, "Kcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - if (model.layers[il].ffn_norm) { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - } else { - // parallel residual - cur = inpSA; - } - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_qwen : public llm_graph_context { - llm_build_qwen(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 2*sizeof(float)*(n_embd)); - - // using mode = 2 for neox mode - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward forward - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_qwen2 : public llm_graph_context { - llm_build_qwen2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - if (model.output_b != nullptr) { - cur = ggml_add(ctx0, cur, model.output_b); - } - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_dream : public llm_graph_context { - llm_build_dream(const llama_model & model, const llm_graph_params & params) : - llm_graph_context(params) { - //copied from qwen2 - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_no_cache(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_llada : public llm_graph_context { - llm_build_llada(const llama_model & model, const llm_graph_params & params) : - llm_graph_context(params) { - // LLaDA is similar to LLaMA but uses non-causal attention for diffusion - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - // Non-causal attention for diffusion - auto * inp_attn = build_attn_inp_no_cache(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute separate Q, K, V projections without bias, matching LLaDALlamaBlock - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_qwen2vl : public llm_graph_context { - llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - int sections[4]; - std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_multi( - ctx0, Qcur, inp_pos, nullptr, - n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_multi( - ctx0, Kcur, inp_pos, nullptr, - n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_qwen2moe : public llm_graph_context { - llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = - build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - - // FFN shared expert - { - ggml_tensor * cur_gate_inp = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); - cb(cur_gate_inp, "ffn_shexp_gate_inp", il); - - // sigmoid - ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp); - cb(cur_gate, "ffn_shexp_gate", il); - - ggml_tensor * cur_ffn = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur_ffn, "ffn_shexp", il); - - ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate); - cb(ffn_shexp_out, "ffn_shexp_out", il); - - moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out); - cb(moe_out, "ffn_out", il); - - cur = moe_out; - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_qwen3 : public llm_graph_context { - llm_build_qwen3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_qwen3moe : public llm_graph_context { - llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = - build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - cur = moe_out; - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_phi2 : public llm_graph_context { - llm_build_phi2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * attn_norm_output; - ggml_tensor * ffn_output; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - attn_norm_output = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(attn_norm_output, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = nullptr; - ggml_tensor * Kcur = nullptr; - ggml_tensor * Vcur = nullptr; - - if (model.layers[il].wqkv) { - cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - } else { - Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); - Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); - Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - } - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - // with phi2, we scale the Q to avoid precision issues - // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66 - Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head))); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); - } - - // FF - { - ffn_output = build_ffn(attn_norm_output, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(ffn_output, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_output); - cur = ggml_add(ctx0, cur, inpL); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = build_norm(inpL, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - cb(cur, "result_output_no_bias", -1); - - cur = ggml_add(ctx0, cur, model.output_b); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -template -struct llm_build_phi3 : public llm_graph_context { - llm_build_phi3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - using inp_attn_type = std::conditional_t; - inp_attn_type * inp_attn = nullptr; - - if constexpr (iswa) { - inp_attn = build_attn_inp_kv_iswa(); - } else { - inp_attn = build_attn_inp_kv(); - } - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - auto * residual = inpL; - - // self-attention - { - // rope freq factors for 128k context - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - ggml_tensor* attn_norm_output = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM_RMS, il); - cb(attn_norm_output, "attn_norm", il); - - ggml_tensor * Qcur = nullptr; - ggml_tensor * Kcur = nullptr; - ggml_tensor * Vcur = nullptr; - - if (model.layers[il].wqkv) { - cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); - cb(cur, "wqkv", il); - - Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 0 * sizeof(float) * (n_embd)); - Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd)); - Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)); - } else { - Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); - Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); - Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - } - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); - cb(Qcur, "Qcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - residual = ggml_get_rows(ctx0, residual, inp_out_ids); - } - - cur = ggml_add(ctx0, cur, residual); - residual = cur; - - cur = build_norm(cur, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - if (model.layers[il].ffn_gate_inp == nullptr) { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(cur, "ffn_moe_out", il); - } - - cur = ggml_add(ctx0, residual, cur); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = build_norm(inpL, - model.output_norm, - model.output_norm_b, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - if (model.output_b != nullptr) { - cb(cur, "result_output_no_bias", -1); - cur = ggml_add(ctx0, cur, model.output_b); - } - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_plamo : public llm_graph_context { - llm_build_plamo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - ggml_tensor * sa_inp = cur; - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - ggml_tensor * sa_out = cur; - - cur = sa_inp; - - // feed-forward network - { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, sa_out); - cur = ggml_add(ctx0, cur, inpL); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_gpt2 : public llm_graph_context { - llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * pos; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); - cb(pos, "pos_embd", -1); - - inpL = ggml_add(ctx0, inpL, pos); - cb(inpL, "inpL", -1); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - cur = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // add the input - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = build_norm(inpL, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_codeshell : public llm_graph_context { - llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - cur = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // add the input - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = build_norm(inpL, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_orion : public llm_graph_context { - llm_build_orion(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - // if (model.layers[il].bq) { - // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - // cb(Qcur, "Qcur", il); - // } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - // if (model.layers[il].bk) { - // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - // cb(Kcur, "Kcur", il); - // } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - // if (model.layers[il].bv) { - // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - // cb(Vcur, "Vcur", il); - // } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_internlm2 : public llm_graph_context { - llm_build_internlm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_minicpm3 : public llm_graph_context { - llm_build_minicpm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - //TODO: if the model varies, these parameters need to be read from the model - const int64_t n_embd_base = 256; - const float scale_embd = 12.0f; - const float scale_depth = 1.4f; - const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k)); - - const uint32_t n_embd_head_qk_rope = hparams.n_rot; - const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; - const uint32_t kv_lora_rank = hparams.n_lora_kv; - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // scale the input embeddings - inpL = ggml_scale(ctx0, inpL, scale_embd); - cb(inpL, "inp_scaled", -1); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - ggml_tensor * q = NULL; - // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} - q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); - cb(q, "q", il); - - q = build_norm(q, - model.layers[il].attn_q_a_norm, NULL, - LLM_NORM_RMS, il); - cb(q, "q", il); - - // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} - q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); - cb(q, "q", il); - - // split into {n_head * n_embd_head_qk_nope, n_tokens} - ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), - 0); - cb(q_nope, "q_nope", il); - - // and {n_head * n_embd_head_qk_rope, n_tokens} - ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), - ggml_row_size(q->type, n_embd_head_qk_nope)); - cb(q_pe, "q_pe", il); - - // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} - ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); - cb(kv_pe_compresseed, "kv_pe_compresseed", il); - - // split into {kv_lora_rank, n_tokens} - ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, - kv_pe_compresseed->nb[1], - 0); - cb(kv_compressed, "kv_compressed", il); - - // and {n_embd_head_qk_rope, n_tokens} - ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, - kv_pe_compresseed->nb[1], - kv_pe_compresseed->nb[1], - ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); - cb(k_pe, "k_pe", il); - - kv_compressed = build_norm(kv_compressed, - model.layers[il].attn_kv_a_norm, NULL, - LLM_NORM_RMS, il); - cb(kv_compressed, "kv_compressed", il); - - // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} - ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); - cb(kv, "kv", il); - - // split into {n_head * n_embd_head_qk_nope, n_tokens} - ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), - ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), - 0); - cb(k_nope, "k_nope", il); - - // and {n_head * n_embd_head_v, n_tokens} - ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), - ggml_row_size(kv->type, (n_embd_head_qk_nope))); - cb(v_states, "v_states", il); - - v_states = ggml_cont(ctx0, v_states); - cb(v_states, "v_states", il); - - q_pe = ggml_rope_ext( - ctx0, q_pe, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(q_pe, "q_pe", il); - - // shared RoPE key - k_pe = ggml_rope_ext( - ctx0, k_pe, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(k_pe, "k_pe", il); - - ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); - cb(q_states, "q_states", il); - - ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); - cb(k_states, "k_states", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // scale_res - scale the hidden states for residual connection - const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct? - cur = ggml_scale(ctx0, cur, scale_res); - cb(cur, "hidden_scaled", il); - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - // scale the hidden states for residual connection - cur = ggml_scale(ctx0, cur, scale_res); - cb(cur, "hidden_scaled_ffn", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head scaling - const float scale_lmhead = float(n_embd_base)/float(n_embd); - cur = ggml_scale(ctx0, cur, scale_lmhead); - cb(cur, "lmhead_scaling", -1); - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_gemma : public llm_graph_context { - llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); - cb(inpL, "inp_scaled", -1); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); - cb(Qcur, "Qcur_scaled", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); - cb(sa_out, "sa_out", il); - - cur = build_norm(sa_out, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, sa_out); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_gemma2_iswa : public llm_graph_context { - llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_k; - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); - cb(inpL, "inp_scaled", -1); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv_iswa(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - cur = build_norm(cur, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); - cb(sa_out, "sa_out", il); - - cur = build_norm(sa_out, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = build_norm(cur, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, -1); - cb(cur, "ffn_post_norm", -1); - - cur = ggml_add(ctx0, cur, sa_out); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - // final logit soft-capping - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); - cur = ggml_tanh(ctx0, cur); - cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_gemma3_iswa : public llm_graph_context { - llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_k; - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) - if (ubatch.token) { - inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); - cb(inpL, "inp_scaled", -1); - } - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - // TODO: is causal == true correct? might need some changes - auto * inp_attn = build_attn_inp_kv_iswa(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const float freq_base_l = model.get_rope_freq_base (cparams, il); - const float freq_scale_l = model.get_rope_freq_scale(cparams, il); - - // norm - cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, - ext_factor, attn_factor, beta_fast, beta_slow); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315 - Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - cur = build_norm(cur, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); - cb(sa_out, "sa_out", il); - - cur = build_norm(sa_out, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = build_norm(cur, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, -1); - cb(cur, "ffn_post_norm", -1); - - cur = ggml_add(ctx0, cur, sa_out); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_gemma3n_iswa : public llm_graph_context { - const llama_model & model; - - const int64_t n_embd_head; - const int64_t n_embd_altup; - const int64_t n_altup; - const int i_altup_act; - const int n_layer_sparsity = 10; // number of layers using activation sparsity - const float f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95) - - llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) - : llm_graph_context(params), - model(model), - n_embd_head(model.hparams.n_embd_head_k), - n_embd_altup(model.hparams.n_embd_altup), - n_altup(model.hparams.n_altup), - i_altup_act(model.hparams.i_altup_act) { - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) - if (ubatch.token) { - inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); - cb(inpL, "inp_scaled", -1); - } - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - // TODO: is causal == true correct? might need some changes - auto * inp_attn = build_attn_inp_kv_iswa(); - - // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer] - ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs()); - - // inpL now has only 1 altup, project it to the rest of the altups - // these "added" altups will be concat to the last dim of inpL - { - ggml_tensor * target_magnitude = calc_magnitude(inpL); - ggml_tensor * inp_repeated = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1); - ggml_tensor * altup_added = ggml_mul_mat(ctx0, model.altup_proj, inp_repeated); // shape: [n_embd, n_tokens, n_altup - 1] - ggml_tensor * new_magnitude = calc_magnitude(altup_added); - altup_added = ggml_div(ctx0, - ggml_mul(ctx0, altup_added, target_magnitude), - new_magnitude); - inpL = ggml_concat(ctx0, inpL, altup_added, 2); // shape: [n_embd, n_tokens, n_altup] - cb(inpL, "inp_stacked", -1); - } - - // inpL now has shape: [n_embd, n_tokens, n_altup] - // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer] - - for (int il = 0; il < n_layer; ++il) { - // this block is made to be closely resemble Gemma3p5DecoderLayer on python code - const float freq_base_l = model.get_rope_freq_base (cparams, il); - const float freq_scale_l = model.get_rope_freq_scale(cparams, il); - - ggml_tensor * cur = inpL; // [n_embd, n_tokens, n_altup] - ggml_tensor * predictions = altup_predict(cur, il); // [n_embd, n_tokens, n_altup] - - // predicted value will go through self-attention and laurel - ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act); // [n_embd, n_tokens] - cur = active_prediction; - cb(cur, "active_prediction", il); - - // norm - cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // laurel - ggml_tensor * laurel_out = laurel(cur, il); // [n_embd, n_tokens] - - // self-attention - if (hparams.has_kv(il)) { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps); - - cb(Qcur, "Qcur_normed", il); - cb(Kcur, "Kcur_normed", il); - cb(Vcur, "Vcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, - ext_factor, attn_factor, beta_fast, beta_slow); - - cb(Qcur, "Qcur_pos", il); - cb(Kcur, "Kcur_pos", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, hparams.f_attention_scale, il); - } else { - // reuse KV cache of earlier layers - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, - ext_factor, attn_factor, beta_fast, beta_slow); - cb(Qcur, "Qcur_pos", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il); - } - - cur = build_norm(cur, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - cur = ggml_add(ctx0, cur, active_prediction); // [n_embd, n_tokens] - cb(cur, "attn_gated", il); - - ggml_tensor * attn_laurel = ggml_scale(ctx0, - ggml_add(ctx0, cur, laurel_out), - 1.0f / sqrtf(2.0f)); // [n_embd, n_tokens] - cb(attn_laurel, "attn_laurel", il); - - cur = build_norm(attn_laurel, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - { - ggml_tensor * up_proj = build_lora_mm(model.layers[il].ffn_up, cur); - ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur); - - if (il < n_layer_sparsity) { - // apply activation sparsity - gate_proj = gaussian_topk(gate_proj); - } - gate_proj = ggml_gelu(ctx0, gate_proj); - - cur = ggml_mul(ctx0, up_proj, gate_proj); - cur = build_lora_mm(model.layers[il].ffn_down, cur); - cb(cur, "ffn_out", il); - } - - cur = build_norm(cur, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, -1); - cb(cur, "ffn_post_norm", il); - - ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel); // [n_embd, n_tokens] - cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il); - - ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il); // [n_embd, n_tokens, n_altup] - - ggml_tensor * first_prediction; // [n_embd, n_tokens] - { - first_prediction = view_2d_slice(corrected, i_altup_act); // [n_embd, n_tokens] - first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale); - first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction); - first_prediction = ggml_gelu(ctx0, first_prediction); // [n_embd_altup, n_tokens] - cb(first_prediction, "first_prediction_gated", il); - ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il); // [n_embd_altup, n_tokens] - first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer); // [n_embd_altup, n_tokens] - cb(first_prediction, "first_prediction_scaled", il); - - first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction); // [n_embd, n_tokens] - first_prediction = build_norm(first_prediction, - model.layers[il].per_layer_post_norm, NULL, - LLM_NORM_RMS, il); - cb(first_prediction, "first_prediction_out", il); - } - - // equivalent to python code: corrected_predictions[1:] += first_prediction - { - ggml_tensor * slice_first = view_2d_slice(corrected, 0); - ggml_tensor * slice_rest = ggml_view_3d(ctx0, corrected, n_embd, n_tokens, n_altup - 1, - ggml_row_size(corrected->type, n_embd), - ggml_row_size(corrected->type, n_embd*n_tokens), - n_embd*n_tokens*ggml_element_size(corrected)); - ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction); // [n_embd, n_tokens, n_altup - 1] - corrected = ggml_concat(ctx0, slice_first, tmp, 2); // [n_embd, n_tokens, n_altup] - } - - cur = corrected; // [n_embd, n_tokens, n_altup] - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; // [n_embd, n_tokens, n_altup] - - // cur now has multiple altup(s), we want to merge them back to 1 altup - { - ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act)); // [n_embd, n_tokens] - // do a view to skip the first slice (active altup) - ggml_tensor * alt_slice = ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1, - ggml_row_size(cur->type, n_embd), - ggml_row_size(cur->type, n_embd*n_tokens), - n_embd*n_tokens*ggml_element_size(cur)); - ggml_tensor * altup_unembd = ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice); // shape: [n_embd, n_tokens, n_altup - 1] - ggml_tensor * new_magnitude = calc_magnitude(altup_unembd); - altup_unembd = ggml_div(ctx0, - ggml_mul(ctx0, altup_unembd, target_magnitude), - new_magnitude); - cb(altup_unembd, "altup_unembd", -1); - - // equivalent to torch.mean(hidden_states, dim=0) - cur = view_2d_slice(cur, 0); // [n_embd, n_tokens] - for (int i = 0; i < n_altup - 1; ++i) { - cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i)); - } - cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup)); // [n_embd, n_tokens] - cb(cur, "unembd_merged", -1); - } - - // cur now has shape: [n_embd, n_tokens] - - // TODO: move this to right after the last KV layer - { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - } - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - { - // final logit soft-capping - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); - cur = ggml_tanh(ctx0, cur); - cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); - } - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } - - ggml_tensor * calc_magnitude(ggml_tensor * x) { - return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x))); - } - - // get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim - ggml_tensor * view_2d_slice(ggml_tensor * x, int idx) { - GGML_ASSERT(idx < (int)x->ne[2]); - return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], - ggml_row_size(x->type, x->ne[0]), - idx * x->ne[0] * x->ne[1] * ggml_element_size(x)); - } - - // equivalent to get_per_layer_inputs() in python code - // output shape: [n_embd_altup, n_layer, n_tokens] - ggml_tensor * get_per_layer_inputs() { - auto inp = std::make_unique(); - ggml_tensor * inp_per_layer; - if (ubatch.token) { - inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens); - ggml_set_input(inp->tokens); - res->t_tokens = inp->tokens; - inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens); - inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens); - inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float)n_embd_altup)); - cb(inp_per_layer, "inp_per_layer_selected", -1); - } else { - GGML_ABORT("TODO: support embd input"); - } - res->add_input(std::move(inp)); - return inp_per_layer; - } - - // equivalent to project_per_layer_inputs() in python code - // this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim - // output shape: [n_embd_altup, n_tokens, n_layer] - ggml_tensor * project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) { - const float per_layer_projection_scale = 1.0f / sqrtf((float)n_embd); - const float per_layer_input_scale = 1.0f / sqrtf(2.0f); - - ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds); - per_layer_proj = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale); - per_layer_proj = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens); - per_layer_proj = build_norm(per_layer_proj, - model.per_layer_proj_norm, NULL, - LLM_NORM_RMS, -1); // [n_embd_altup, n_layer, n_tokens] - cb(per_layer_proj, "per_layer_proj", -1); - - inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj); - inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale); - cb(inp_per_layer, "inp_per_layer", -1); - - // permute to shape: [n_embd_altup, n_tokens, n_layer] - inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3)); - return inp_per_layer; - } - - // input cur shape: [n_altup, n_tokens] - // output shape: [n_altup, n_tokens] - ggml_tensor * laurel(ggml_tensor * cur, int il) { - ggml_tensor * tmp = cur; - tmp = build_lora_mm(model.layers[il].laurel_l, tmp); - tmp = build_lora_mm(model.layers[il].laurel_r, tmp); - tmp = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il); - tmp = ggml_add(ctx0, tmp, cur); - cb(tmp, "laurel_out", il); - return tmp; - } - - // input x shape: [n_embd, n_tokens] - // output shape: [n_embd, n_tokens] - ggml_tensor * gaussian_topk(ggml_tensor * x) { - ggml_tensor * mean = ggml_mean(ctx0, x); - ggml_tensor * std = ggml_sqrt(ctx0, ggml_scale(ctx0, - ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))), - 1.0f / (float)(x->ne[0] - 1) - )); - ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul)); - return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x)); - } - - // - // altup functions - // - - // equivalent to compute_router_modalities() in python code - // input x shape: [n_embd, n_tokens] - // output shape: [n_altup, n_tokens] - ggml_tensor * altup_compute_router_modalities(ggml_tensor * x, int il) { - ggml_tensor * router_inputs = build_norm(x, - model.layers[il].altup_router_norm, NULL, - LLM_NORM_RMS, il); - - // router_input_scale - router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float)n_embd); - - ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs); - return ggml_tanh(ctx0, output); // [n_altup, n_tokens] - } - - // input cur shape: [n_embd, n_tokens, n_altup] - // output shape: [n_embd, n_tokens, n_altup] - ggml_tensor * altup_predict(ggml_tensor * cur, int il) { - ggml_tensor * activated = view_2d_slice(cur, i_altup_act); // [n_embd, n_tokens] - ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens] - cb(modalities, "modalities", il); - - ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities); - cb(all_coefs, "all_coefs", il); - // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor) - all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens); - - // permute to [n_altup, n_embd, n_tokens] - ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3)); - ggml_tensor * predictions = ggml_mul_mat(ctx0, cur_permuted, all_coefs); // [n_altup, n_embd, n_tokens] - - // final shape must be the same as cur: [n_embd, n_tokens, n_altup] - predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3)); - predictions = ggml_add(ctx0, predictions, cur); - cb(predictions, "predictions", il); - - return predictions; - } - - // input predictions shape: [n_embd, n_tokens, n_altup] - // input activated shape: [n_embd, n_tokens] - // output shape: [n_embd, n_tokens, n_altup] - ggml_tensor * altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) { - ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens] - cb(modalities, "modalities", il); - - ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act); - ggml_tensor * innovation = ggml_sub(ctx0, activated, active_prediction); // [n_embd, n_tokens] - cb(innovation, "innovation", il); - - ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities); // [n_altup, n_tokens] - all_coefs = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f); // + 1.0 - cb(all_coefs, "all_coefs", il); - all_coefs = ggml_transpose(ctx0, all_coefs); // [n_tokens, n_altup] - all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup); // [1, n_tokens, n_altup] - - innovation = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1); - ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs); // [n_embd, n_tokens, n_altup] - corrected = ggml_add(ctx0, corrected, predictions); // [n_embd, n_tokens, n_altup] - cb(corrected, "corrected", il); - - return corrected; - } -}; - -struct llm_build_gemma_embedding : public llm_graph_context { - llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_k; - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) - if (ubatch.token) { - inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); - cb(inpL, "inp_scaled", -1); - } - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_no_cache(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const float freq_base_l = model.get_rope_freq_base (cparams, il); - const float freq_scale_l = model.get_rope_freq_scale(cparams, il); - - // norm - cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, - ext_factor, attn_factor, beta_fast, beta_slow); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315 - Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - cur = build_norm(cur, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); - cb(sa_out, "sa_out", il); - - cur = build_norm(sa_out, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = build_norm(cur, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, -1); - cb(cur, "ffn_post_norm", -1); - - cur = ggml_add(ctx0, cur, sa_out); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -// TODO: move up next to build_starcoder -struct llm_build_starcoder2 : public llm_graph_context { - llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_graph_context_mamba : public llm_graph_context { - llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {} - - ggml_tensor * build_mamba_layer( - llm_graph_input_rs * inp, - ggml_tensor * cur, - const llama_model & model, - const llama_ubatch & ubatch, - int il) { - - const auto * mctx_cur = inp->mctx; - - const auto kv_head = mctx_cur->get_head(); - - const auto & layer = model.layers[il]; - - const int64_t d_conv = hparams.ssm_d_conv; - const int64_t d_inner = hparams.ssm_d_inner; - const int64_t d_state = hparams.ssm_d_state; - const int64_t dt_rank = hparams.ssm_dt_rank; - const int64_t n_head = d_inner; - const int64_t head_dim = 1; - const int64_t n_seqs = ubatch.n_seqs; - // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers) - const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms; - - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - - GGML_ASSERT(n_seqs != 0); - GGML_ASSERT(ubatch.equal_seqs()); - GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); - - ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); - ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); - - ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); - conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs); - - // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} - cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); - - // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} - ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur); - // split the above in two - // => {d_inner, n_seq_tokens, n_seqs} - ggml_tensor * x = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0); - ggml_tensor * z = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner*ggml_element_size(xz)); - - // conv - { - // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs} - ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0); - - // copy last (d_conv - 1) columns back into the state cache - ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0])); - - ggml_build_forward_expand(gf, - ggml_cpy(ctx0, last_conv, - ggml_view_1d(ctx0, conv_states_all, - (d_conv - 1)*(d_inner)*(n_seqs), - kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all)))); - - // 1D convolution - // The equivalent is to make a self-overlapping view of conv_x - // over d_conv columns at each stride in the 3rd dimension, - // then element-wise multiply that with the conv1d weight, - // then sum the elements of each row, - // (the last two steps are a dot product over rows (also doable with mul_mat)) - // then permute away the ne[0] dimension, - // and then you're left with the resulting x tensor. - // For simultaneous sequences, all sequences need to have the same length. - x = ggml_ssm_conv(ctx0, conv_x, layer.ssm_conv1d); - - // bias - x = ggml_add(ctx0, x, layer.ssm_conv1d_b); - - x = ggml_silu(ctx0, x); - } - - // ssm - { - // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} - ggml_tensor * x_db = build_lora_mm(layer.ssm_x, x); - // split - ggml_tensor * dt = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0); - ggml_tensor * B = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*dt_rank); - ggml_tensor * C = ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state*x_db->nb[0], x_db->nb[1], x_db->nb[2], ggml_element_size(x_db)*(dt_rank+d_state)); - - // Some Mamba variants (e.g. FalconMamba, Jamba) apply RMS norm in B, C & Dt layers - if (ssm_dt_b_c_rms || (layer.ssm_dt_norm && layer.ssm_b_norm && layer.ssm_c_norm)) { - dt = build_norm(dt, layer.ssm_dt_norm, NULL, LLM_NORM_RMS, il); - B = build_norm(B, layer.ssm_b_norm, NULL, LLM_NORM_RMS, il); - C = build_norm(C, layer.ssm_c_norm, NULL, LLM_NORM_RMS, il); - } - - // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} - dt = build_lora_mm(layer.ssm_dt, dt); - dt = ggml_add(ctx0, dt, layer.ssm_dt_b); - - cur = x; - x = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs); - - ggml_tensor * A = layer.ssm_a; - - // use the states and the indices provided by build_recurrent_state - // (this is necessary in order to properly use the states before they are overwritten, - // while avoiding to make unnecessary copies of the states) - auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) { - ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size()); - - // Custom operator to optimize the parallel associative scan - // as described in the Annex D of the Mamba paper. - // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} - return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids); - }; - - ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows); - - // store last states - ggml_build_forward_expand(gf, - ggml_cpy(ctx0, - ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]*x->ne[3]), - ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all)))); - - ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 0); - - // TODO: skip computing output earlier for unused tokens - - y = ggml_add(ctx0, y, ggml_mul(ctx0, cur, layer.ssm_d)); - y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y); - - // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} - cur = build_lora_mm(layer.ssm_out, y); - } - - // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} - cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); - - return cur; - } - - ggml_tensor * build_mamba2_layer( - llm_graph_input_rs * inp, - ggml_tensor * cur, - const llama_model & model, - const llama_ubatch & ubatch, - int il) const { - - const auto * mctx_cur = inp->mctx; - - const auto kv_head = mctx_cur->get_head(); - - const int64_t d_conv = hparams.ssm_d_conv; - const int64_t d_inner = hparams.ssm_d_inner; - const int64_t d_state = hparams.ssm_d_state; - const int64_t n_head = hparams.ssm_dt_rank; - const int64_t head_dim = d_inner / n_head; - const int64_t n_group = hparams.ssm_n_group; - const int64_t n_seqs = ubatch.n_seqs; - - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - - GGML_ASSERT(n_seqs != 0); - GGML_ASSERT(ubatch.equal_seqs()); - GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); - - ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); - ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); - - ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); - conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs); - - // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} - cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); - - // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads - - // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs} - ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur); - - // split the above in three - ggml_tensor * z = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*zxBCdt->nb[0], zxBCdt->nb[1], zxBCdt->nb[2], 0); - ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2*n_group*d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], d_inner*ggml_element_size(zxBCdt)); - ggml_tensor * dt = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], (2*d_inner + 2*n_group*d_state)*ggml_element_size(zxBCdt)); - - // conv - { - // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs} - ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 0); - - // copy last (d_conv - 1) columns back into the state cache - ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs, conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0])); - - ggml_build_forward_expand(gf, - ggml_cpy(ctx0, last_conv, - ggml_view_1d(ctx0, conv_states_all, - (d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs), - kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all)))); - - // 1D convolution - // The equivalent is to make a self-overlapping view of conv_x - // over d_conv columns at each stride in the 3rd dimension, - // then element-wise multiply that with the conv1d weight, - // then sum the elements of each row, - // (the last two steps are a dot product over rows (also doable with mul_mat)) - // then permute away the ne[0] dimension, - // and then you're left with the resulting x tensor. - // For simultaneous sequences, all sequences need to have the same length. - xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d); - - // bias - xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b); - - xBC = ggml_silu(ctx0, xBC); - } - - // ssm - { - // These correspond to V K Q in SSM/attention duality - ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim*xBC->nb[0], xBC->nb[1], xBC->nb[2], 0); - ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], d_inner*ggml_element_size(xBC)); - ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state*xBC->nb[0], xBC->nb[1], xBC->nb[2], (d_inner + n_group*d_state)*ggml_element_size(xBC)); - - // {n_head, n_seq_tokens, n_seqs} - dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b); - - ggml_tensor * A = model.layers[il].ssm_a; - - // use the states and the indices provided by build_recurrent_state - // (this is necessary in order to properly use the states before they are overwritten, - // while avoiding to make unnecessary copies of the states) - auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) { - ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size()); - - // TODO: use semistructured matrices to implement state-space duality - // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} - return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids); - }; - - ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows); - - // store last states - ggml_build_forward_expand(gf, - ggml_cpy(ctx0, - ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, ggml_nelements(x)*x->nb[0]), - ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs, kv_head*d_state*d_inner*ggml_element_size(ssm_states_all)))); - - ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head*x->nb[1], n_seq_tokens*n_head*x->nb[1], 0); - - // TODO: skip computing output earlier for unused tokens - - y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d)); - cb(y, "mamba2_y_add_d", il); - y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y); - - // grouped RMS norm - if (model.layers[il].ssm_norm) { - y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs); - y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il); - } - - y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs); - - // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} - cur = build_lora_mm(model.layers[il].ssm_out, y); - } - - // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} - cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); - cb(cur, "mamba_out", il); - - return cur; - } -}; - -struct llm_build_mamba : public llm_graph_context_mamba { - llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) { - ggml_tensor * cur; - ggml_tensor * inpL; - - // {n_embd, n_tokens} - inpL = build_inp_embd(model.tok_embd); - - auto * rs_inp = build_rs_inp(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - if (model.arch == LLM_ARCH_MAMBA2) { - cur = build_mamba2_layer(rs_inp, cur, model, ubatch, il); - } else { - cur = build_mamba_layer(rs_inp, cur, model, ubatch, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // residual - cur = ggml_add(ctx0, cur, inpL); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - // final rmsnorm - cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } - -}; - -struct llm_build_jamba : public llm_graph_context_mamba { - llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - ggml_tensor * cur; - ggml_tensor * inpL; - - // {n_embd, n_tokens} - inpL = build_inp_embd(model.tok_embd); - - auto * inp_hybrid = build_inp_mem_hybrid(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const int64_t n_head_kv = hparams.n_head_kv(il); - - cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - if (n_head_kv == 0) { - cur = build_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il); - } else { - // Attention - - struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - // No RoPE :) - cur = build_attn(inp_hybrid->get_attn(), - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // residual - struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur); - cb(cur, "ffn_inp", il); - - cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network - if (model.layers[il].ffn_gate_inp == nullptr) { - // FFN - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(cur, "ffn_moe_out", il); - } - - // residual - cur = ggml_add(ctx0, ffn_inp, cur); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - // final rmsnorm - cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_command_r : public llm_graph_context { - llm_build_command_r(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - const float f_logit_scale = hparams.f_logit_scale; - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - ggml_tensor * ffn_inp = cur; - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - if (model.layers[il].attn_q_norm) { - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, - NULL, - LLM_NORM, il); - cb(Qcur, "Qcur", il); - } - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - if (model.layers[il].attn_k_norm) { - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, - NULL, - LLM_NORM, il); - cb(Kcur, "Kcur", il); - } - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); - } - - ggml_tensor * attn_out = cur; - - // feed-forward network - { - cur = build_ffn(ffn_inp, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - // add together residual + FFN + self-attention - cur = ggml_add(ctx0, cur, inpL); - cur = ggml_add(ctx0, cur, attn_out); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - if (f_logit_scale) { - cur = ggml_scale(ctx0, cur, f_logit_scale); - } - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_cohere2_iswa : public llm_graph_context { - llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - const float f_logit_scale = hparams.f_logit_scale; - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv_iswa(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const bool is_swa = hparams.is_swa(il); - - // norm - cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); - cb(cur, "attn_norm", il); - ggml_tensor * ffn_inp = cur; - - // self-attention - { - // rope freq factors for 128k context - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - if (is_swa) { - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); - } - - ggml_tensor * attn_out = cur; - - // feed-forward network - { - cur = build_ffn(ffn_inp, model.layers[il].ffn_up, NULL, NULL, model.layers[il].ffn_gate, - NULL, NULL, model.layers[il].ffn_down, NULL, NULL, NULL, LLM_FFN_SILU, LLM_FFN_PAR, - il); - cb(cur, "ffn_out", il); - } - - // add together residual + FFN + self-attention - cur = ggml_add(ctx0, cur, inpL); - cur = ggml_add(ctx0, cur, attn_out); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - if (f_logit_scale) { - cur = ggml_scale(ctx0, cur, f_logit_scale); - } - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -// ref: https://allenai.org/olmo -// based on the original build_llama() function, changes: -// * non-parametric layer norm -// * clamp qkv -// * removed bias -// * removed MoE -struct llm_build_olmo : public llm_graph_context { - llm_build_olmo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - NULL, NULL, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (hparams.f_clamp_kqv > 0.0f) { - Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (hparams.f_clamp_kqv > 0.0f) { - Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (hparams.f_clamp_kqv > 0.0f) { - Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, nullptr, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - NULL, NULL, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - NULL, NULL, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -template -struct llm_build_olmo2 : public llm_graph_context { - llm_build_olmo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - using inp_attn_type = std::conditional_t; - inp_attn_type * inp_attn = nullptr; - - if constexpr (iswa) { - inp_attn = build_attn_inp_kv_iswa(); - } else { - inp_attn = build_attn_inp_kv(); - } - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = inpL; - - // self_attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, - LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, - LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - const bool is_swa = hparams.is_swa(il); - - if (is_swa) { - // For sliding window layers, Olmo3 use regular rope with no yarn rope scaling. - // This is achieved here by setting freq_scale and attn_factor to 1. - // We also set ext_factor to 0 to avoid a few unnecessary computations. - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, 1.0, - 0.0, 1.0, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, 1.0, - 0.0, 1.0, beta_fast, beta_slow - ); - } else { - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - cur = build_norm(cur, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_ffn(ffn_inp, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = build_norm(cur, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, -1); - cb(cur, "ffn_post_norm", -1); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -// based on the build_qwen2moe() function, changes: -// * removed shared experts -// * removed bias -// * added q, k norm -struct llm_build_olmoe : public llm_graph_context { - llm_build_olmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, - LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, - LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(cur, "ffn_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_llada_moe : public llm_graph_context { - llm_build_llada_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_no_cache(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(cur, "ffn_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_openelm : public llm_graph_context { - llm_build_openelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const int64_t n_head = hparams.n_head(il); - const int64_t n_head_kv = hparams.n_head_kv(il); - const int64_t n_head_qkv = 2*n_head_kv + n_head; - - cur = inpL; - ggml_tensor * residual = cur; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv))); - cb(Vcur, "Vcur", il); - - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, NULL, - LLM_NORM_RMS, il); - cb(Qcur, "Qcur", il); - - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, NULL, - LLM_NORM_RMS, il); - cb(Kcur, "Kcur", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, NULL, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, NULL, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Qcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - residual = ggml_get_rows(ctx0, residual, inp_out_ids); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - inpL = cur; - } - - cur = inpL; - - // norm - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_gptneox : public llm_graph_context { - llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - cur = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // ffn - if (hparams.use_par_res) { - // attention and ffn are computed in parallel - // x = x + attn(ln1(x)) + ffn(ln2(x)) - - ggml_tensor * attn_out = cur; - - cur = build_norm(inpL, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, inpL); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, attn_out); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } else { - // attention and ffn are computed sequentially - // x = x + attn(ln1(x)) - // x = x + ffn(ln2(x)) - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - } - - cur = build_norm(inpL, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_arctic : public llm_graph_context { - llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp); - cb(ffn_out, "ffn_out", il); - - // MoE - cur = build_norm(inpSA, - model.layers[il].ffn_norm_exps, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm_exps", il); - - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(cur, "ffn_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_out); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_deepseek : public llm_graph_context { - llm_build_deepseek(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - if ((uint32_t) il < hparams.n_layer_dense_lead) { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - ggml_tensor * moe_out = - build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, false, - false, hparams.expert_weights_scale, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - - // FFN shared expert - { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_deepseek2 : public llm_graph_context { - llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - bool is_lite = (hparams.n_layer == 27); - - const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0); - - // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA - const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k; - const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v; - - const int64_t n_embd_head_qk_rope = hparams.n_rot; - const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope; - - const uint32_t kv_lora_rank = hparams.n_lora_kv; - - // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly. - // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. - const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale)); - const float kq_scale = 1.0f*mscale*mscale/sqrtf(float(n_embd_head_k)); - const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)); - - ggml_tensor * cur; - ggml_tensor * inpL; - - // {n_embd, n_tokens} - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - ggml_tensor * q = NULL; - if (!is_lite) { - q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); - cb(q, "q", il); - - q = build_norm(q, - model.layers[il].attn_q_a_norm, nullptr, - LLM_NORM_RMS, il); - cb(q, "q", il); - - q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); - cb(q, "q", il); - } else { - q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - cb(q, "q", il); - } - - // split into {n_embd_head_qk_nope, n_head, n_tokens} - ggml_tensor * q_nope = ggml_view_3d(ctx0, q, - n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(q->type, n_embd_head_k), - ggml_row_size(q->type, n_embd_head_k) * n_head, - 0); - cb(q_nope, "q_nope", il); - - // and {n_embd_head_qk_rope, n_head, n_tokens} - ggml_tensor * q_pe = ggml_view_3d(ctx0, q, - n_embd_head_qk_rope, n_head, n_tokens, - ggml_row_size(q->type, n_embd_head_k), - ggml_row_size(q->type, n_embd_head_k) * n_head, - ggml_row_size(q->type, n_embd_head_qk_nope)); - cb(q_pe, "q_pe", il); - - ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); - cb(kv_cmpr_pe, "kv_cmpr_pe", il); - - // split into {kv_lora_rank, n_tokens} - ggml_tensor * kv_cmpr = ggml_view_2d(ctx0, kv_cmpr_pe, - kv_lora_rank, n_tokens, - ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), - 0); - cb(kv_cmpr, "kv_cmpr", il); - - // and {n_embd_head_qk_rope, 1, n_tokens} - ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, - n_embd_head_qk_rope, 1, n_tokens, - ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), - ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), - ggml_row_size(kv_cmpr_pe->type, kv_lora_rank)); - cb(k_pe, "k_pe", il); - - q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(q_pe, "q_pe", il); - - k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(k_pe, "k_pe", il); - - kv_cmpr = build_norm(kv_cmpr, - model.layers[il].attn_kv_a_norm, nullptr, - LLM_NORM_RMS, il); - cb(kv_cmpr, "kv_cmpr", il); - - if (is_mla) { - // {n_embd_head_qk_nope, n_tokens, n_head} - q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3); - cb(q_nope, "q_nope_perm", il); - - // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head} - ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope); - cb(q_nope_absorbed, "q_nope_absorbed", il); - - // {kv_lora_rank, n_head, n_tokens} - q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3); - cb(q_nope_absorbed, "q_nope_absorbed_perm", il); - - // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens} - // note: rope must go first for in-place context shifting in build_rope_shift() - ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0); - cb(Qcur, "Qcur", il); - - kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens); - cb(kv_cmpr, "kv_cmpr_reshape", il); - - // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens} - ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0); - cb(Kcur, "Kcur", il); - - // {kv_lora_rank, 1, n_tokens} - ggml_tensor * Vcur = kv_cmpr; - cb(Vcur, "Vcur", il); - - // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group) - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il); - } else { - ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr); - cb(kv, "kv", il); - - // split into {n_embd_head_qk_nope, n_head, n_tokens} - ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, - n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v), - ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, - 0); - cb(k_nope, "k_nope_view", il); - - // and {n_embd_head_v, n_head, n_tokens} - ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, - n_embd_head_v, n_head, n_tokens, - ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v), - ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, - ggml_row_size(kv->type, n_embd_head_qk_nope)); - cb(Vcur, "Vcur_view", il); - - Vcur = ggml_cont(ctx0, Vcur); - cb(Vcur, "Vcur_cont", il); - - // note: rope must go first for in-place context shifting in build_rope_shift() - ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0); - cb(Kcur, "Kcur", il); - - // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups) - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - } - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - if ((uint32_t) il < hparams.n_layer_dense_lead) { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - ggml_tensor * moe_out = - build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - model.layers[il].ffn_exp_probs_b, - n_expert, n_expert_used, - LLM_FFN_SILU, hparams.expert_weights_norm, - true, hparams.expert_weights_scale, - (llama_expert_gating_func_type) hparams.expert_gating_func, - il); - cb(moe_out, "ffn_moe_out", il); - - // FFN shared expert - { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = ggml_mul_mat(ctx0, model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_bitnet : public llm_graph_context { - llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - if (model.layers[il].wq_scale) { - Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale); - } - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - // B1.K - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - if (model.layers[il].wk_scale) { - Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale); - } - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - // B1.V - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - if (model.layers[il].wv_scale) { - Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale); - } - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - NULL, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - - cur = build_norm(cur, - model.layers[il].attn_sub_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_sub_norm", il); - - cur = build_lora_mm(model.layers[il].wo, cur); - if (model.layers[il].wo_scale) { - cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale); - } - if (model.layers[il].bo) { - cur = ggml_add(ctx0, cur, model.layers[il].bo); - } - cb(cur, "attn_o_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward forward - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_scale, - model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale, - NULL, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_sub_out", il); - - cur = build_norm(cur, - model.layers[il].ffn_sub_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_sub_norm", il); - - cur = build_lora_mm(model.layers[il].ffn_down, cur); - if (model.layers[il].ffn_down_scale) { - cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale); - } - cb(cur, "ffn_down", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - // FIXME: do not use model.tok_embd directly, duplicate as model.output - cur = build_lora_mm(model.tok_embd, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_t5_enc : public llm_graph_context { - llm_build_t5_enc(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc(); - - auto * inp_attn = build_attn_inp_no_cache(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm_enc, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; - ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b); - - cur = build_attn(inp_attn, - model.layers[il].wo_enc, nullptr, - Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il); - cb(cur, "kqv_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm_enc, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // T5 uses relu, flan-T5 uses gelu-gated - cur = build_ffn(cur, - model.layers[il].ffn_up_enc, NULL, NULL, - model.layers[il].ffn_gate_enc, NULL, NULL, - model.layers[il].ffn_down_enc, NULL, NULL, - NULL, - model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, - model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, - il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cb(cur, "result_embd", -1); - - cur = build_norm(cur, - model.output_norm_enc, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_t5_dec : public llm_graph_context { - llm_build_t5_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - //const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - ggml_tensor * embd_enc = build_inp_cross_embd(); - ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec(); - - const int64_t n_outputs_enc = embd_enc->ne[1]; - - auto * inp_attn_self = build_attn_inp_kv(); - auto * inp_attn_cross = build_attn_inp_cross(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - const int64_t dec_n_layer = hparams.dec_n_layer; - - for (int il = 0; il < dec_n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; - ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b); - - cur = build_attn(inp_attn_self, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il); - cb(cur, "kqv_out", il); - } - - cur = ggml_add(ctx0, cur, inpSA); - cb(cur, "cross_inp", il); - - ggml_tensor * inpCA = cur; - - // norm - cur = build_norm(cur, - model.layers[il].attn_norm_cross, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm_cross", il); - - // cross-attention - { - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc); - - cur = build_attn(inp_attn_cross, - model.layers[il].wo_cross, nullptr, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); - cb(cur, "kqv_out", il); - - //ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); - //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); - - //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - //cb(kq, "kq", il); - - //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); - //cb(kq, "kq_soft_max_ext", il); - - //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); - //cb(v, "v", il); - - //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); - //cb(kqv, "kqv", il); - - //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); - //cb(kqv_merged, "kqv_merged", il); - - //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); - //cb(cur, "kqv_merged_cont", il); - - //ggml_build_forward_expand(gf, cur); - - //cur = build_lora_mm(model.layers[il].wo_cross, cur); - //cb(cur, "kqv_out", il); - } - - if (il == dec_n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // T5 uses relu, flan-T5 uses gelu-gated - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_RELU, - model.layers[il].ffn_gate ? LLM_FFN_PAR : LLM_FFN_SEQ, - il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cb(cur, "result_embd", -1); - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_jais : public llm_graph_context { - llm_build_jais(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - cur = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*cur->nb[0]*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*cur->nb[0]*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa)); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); - } - - // add the input - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - inpL = ggml_add(ctx0, cur, ffn_inp); - cb(inpL, "l_out", il); - } - - cur = build_norm(inpL, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_chatglm : public llm_graph_context { - llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = build_norm(inpL, - model.layers[il].attn_norm, - NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = nullptr; - ggml_tensor * Kcur = nullptr; - ggml_tensor * Vcur = nullptr; - - if (model.layers[il].wqkv == nullptr) { - Qcur = build_lora_mm(model.layers[il].wq, cur); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - } - Kcur = build_lora_mm(model.layers[il].wk, cur); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - } - Vcur = build_lora_mm(model.layers[il].wv, cur); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - } - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - } else { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - if (model.layers[il].bqkv) { - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - } - Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - } - - //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor); - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // Add the input - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - - } - - inpL = ggml_add(ctx0, cur, ffn_inp); - cb(inpL, "l_out", il); - } - - cur = build_norm(inpL, - model.output_norm, - NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_glm4 : public llm_graph_context { - llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // Pre-attention norm - cur = build_norm(inpL, - model.layers[il].attn_norm, - NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = nullptr; - ggml_tensor * Kcur = nullptr; - ggml_tensor * Vcur = nullptr; - - if (model.layers[il].wqkv == nullptr) { - Qcur = build_lora_mm(model.layers[il].wq, cur); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - } - Kcur = build_lora_mm(model.layers[il].wk, cur); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - } - Vcur = build_lora_mm(model.layers[il].wv, cur); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - } - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - } else { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - if (model.layers[il].bqkv) { - cur = ggml_add(ctx0, cur, model.layers[il].bqkv); - cb(cur, "bqkv", il); - } - Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - } - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // Post-attention norm (new!) - cur = build_norm(cur, - model.layers[il].attn_post_norm, - NULL, - LLM_NORM_RMS, il); - cb(cur, "post_attn_norm", il); - - // Add the input (residual connection after post-attention norm) - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // FF - { - // Pre-MLP norm - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // MLP - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - - // Post-MLP norm - cur = build_norm(cur, - model.layers[il].ffn_post_norm, - NULL, - LLM_NORM_RMS, il); - cb(cur, "post_mlp_norm", il); - } - - // Add residual connection after post-MLP norm - inpL = ggml_add(ctx0, cur, ffn_inp); - cb(inpL, "l_out", il); - } - - // Final norm - cur = build_norm(inpL, - model.output_norm, - NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // Output projection - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_glm4_moe : public llm_graph_context { - llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - // Only process up to last layer (skip final NextN layer) - // Final layer tensors are loaded but not processed in forward pass - const int n_transformer_layers = n_layer - hparams.nextn_predict_layers; - for (int il = 0; il < n_transformer_layers; ++il) { - ggml_tensor * inpSA = inpL; - - // Pre-attention norm - cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - } - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - } - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - } - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - // Apply Q/K norm if available (GLM-4.5 355B variant) - if (model.layers[il].attn_q_norm) { - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - } - if (model.layers[il].attn_k_norm) { - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - } - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_transformer_layers - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // Post-attention norm - cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "post_attn_norm", il); - - // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense) - if (static_cast(il) < hparams.n_layer_dense_lead) { - // Dense FFN layer - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - // Process routed experts using existing MoE infrastructure - ggml_tensor * routed_out = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - model.layers[il].ffn_exp_probs_b, - n_expert, n_expert_used, - LLM_FFN_SILU, hparams.expert_weights_norm, - true, hparams.expert_weights_scale, - (llama_expert_gating_func_type) hparams.expert_gating_func, - il); - cb(routed_out, "ffn_moe_out", il); - - // Process shared expert on original input - ggml_tensor * shared_out = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(shared_out, "ffn_shexp_out", il); - - // Final output: routed_output + shared_output - cur = ggml_add(ctx0, routed_out, shared_out); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_nemotron : public llm_graph_context { - llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - //GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, - model.layers[il].attn_norm_b, - LLM_NORM, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, - model.layers[il].ffn_norm_b, - LLM_NORM, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, model.output_norm_b, - LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_nemotron_h : public llm_graph_context_mamba { - llm_build_nemotron_h( - const llama_model & model, - const llm_graph_params & params) : - llm_graph_context_mamba(params) { - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - ggml_build_forward_expand(gf, inpL); - - auto * inp = build_inp_mem_hybrid(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - if (hparams.is_recurrent(il)) { - // ssm layer // - cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il); - } else if (hparams.n_ff(il) == 0) { - // attention layer // - cur = build_attention_layer(cur, inp->get_attn(), model, n_embd_head, il); - } else { - cur = build_ffn_layer(cur, model, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // add residual - cur = ggml_add(ctx0, cur, inpSA); - cb(cur, "nemotron_h_block_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } - - ggml_tensor * build_attention_layer( - ggml_tensor * cur, - llm_graph_input_attn_kv * inp_attn, - const llama_model & model, - const int64_t n_embd_head, - const int il) { - - // compute Q and K and (optionally) RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - return cur; - } - - ggml_tensor * build_ffn_layer( - ggml_tensor * cur, - const llama_model & model, - const int il) { - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_RELU_SQR, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - return cur; - } -}; - -struct llm_build_exaone : public llm_graph_context { - llm_build_exaone(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -template -struct llm_build_exaone4 : public llm_graph_context { - llm_build_exaone4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_k; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_v); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - using inp_attn_type = std::conditional_t; - inp_attn_type * inp_attn = nullptr; - - if constexpr (iswa) { - inp_attn = build_attn_inp_kv_iswa(); - } else { - inp_attn = build_attn_inp_kv(); - } - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // use RoPE for SWA layers or non-SWA models - const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE; - - cur = inpL; - - // self-attention - { - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - cb(Kcur, "Kcur_normed", il); - - if (use_rope) { - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - cur = build_norm(cur, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_ffn(ffn_inp, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = build_norm(cur, - model.layers[il].ffn_post_norm, NULL, - LLM_NORM_RMS, -1); - cb(cur, "ffn_post_norm", -1); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_rwkv6_base : public llm_graph_context { - const llama_model & model; - - llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) { - } - - ggml_tensor * build_rwkv6_channel_mix( - const llama_layer * layer, - ggml_tensor * cur, - ggml_tensor * x_prev, - llm_arch arch) const { - ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); - switch (arch) { - case LLM_ARCH_RWKV6: - { - ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur); - ggml_tensor * xr = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_r), cur); - - ggml_tensor * r = ggml_sigmoid(ctx0, build_lora_mm(layer->channel_mix_receptance, xr)); - ggml_tensor * k = ggml_sqr( - ctx0, - ggml_relu( - ctx0, - build_lora_mm(layer->channel_mix_key, xk) - ) - ); - cur = ggml_mul(ctx0, r, build_lora_mm(layer->channel_mix_value, k)); - } break; - default: - GGML_ABORT("fatal error"); - } - - return cur; - } - - ggml_tensor * build_rwkv6_time_mix( - llm_graph_input_rs * inp, - ggml_tensor * cur, - ggml_tensor * x_prev, - const llama_ubatch & ubatch, - int il) const { - const auto * mctx_cur = static_cast(mctx); - - const auto n_tokens = ubatch.n_tokens; - const auto n_seqs = ubatch.n_seqs; - const auto n_seq_tokens = ubatch.n_seq_tokens; - const auto n_embd = hparams.n_embd; - const auto head_size = hparams.wkv_head_size; - const auto n_head = n_embd / head_size; - const auto n_head_kv = hparams.n_head_kv(il); - - const auto kv_head = mctx_cur->get_head(); - - const auto & layer = model.layers[il]; - - bool is_qrwkv = layer.time_mix_first == nullptr; - - ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); - - sx = ggml_reshape_2d(ctx0, sx, n_embd, n_tokens); - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - - ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_x), cur); - - xxx = ggml_reshape_4d( - ctx0, - ggml_tanh( - ctx0, - ggml_mul_mat(ctx0, layer.time_mix_w1, xxx) - ), - layer.time_mix_w1->ne[1] / 5, 1, 5, n_tokens - ); - - xxx = ggml_cont(ctx0, ggml_permute(ctx0, xxx, 0, 1, 3, 2)); - - xxx = ggml_mul_mat( - ctx0, - ggml_reshape_4d( - ctx0, - layer.time_mix_w2, - layer.time_mix_w2->ne[0], layer.time_mix_w2->ne[1], 1, 5 - ), - xxx - ); - - ggml_tensor *xw, *xk, *xv, *xr, *xg; - if (layer.time_mix_lerp_fused) { - // fusing these weights makes some performance improvement - sx = ggml_reshape_3d(ctx0, sx, n_embd, 1, n_tokens); - cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens); - xxx = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xxx, layer.time_mix_lerp_fused), sx), cur); - xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); - xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); - xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); - xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); - xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); - } else { - // for backward compatibility - xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); - xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); - xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); - xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); - xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); - - xw = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xw, layer.time_mix_lerp_w), sx), cur); - xk = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xk, layer.time_mix_lerp_k), sx), cur); - xv = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xv, layer.time_mix_lerp_v), sx), cur); - xr = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xr, layer.time_mix_lerp_r), sx), cur); - xg = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xg, layer.time_mix_lerp_g), sx), cur); - } - - ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr); - ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk); - ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv); - if (layer.time_mix_receptance_b) { - r = ggml_add(ctx0, r, layer.time_mix_receptance_b); - } - if (layer.time_mix_key_b) { - k = ggml_add(ctx0, k, layer.time_mix_key_b); - } - if (layer.time_mix_value_b) { - v = ggml_add(ctx0, v, layer.time_mix_value_b); - } - - ggml_tensor * g = build_lora_mm(layer.time_mix_gate, xg); - if (is_qrwkv) { - g = ggml_sigmoid(ctx0, g); - } else { - g = ggml_silu(ctx0, g); - } - - if (n_head_kv != 0 && n_head_kv != n_head) { - GGML_ASSERT(n_head % n_head_kv == 0); - k = ggml_reshape_4d(ctx0, k, head_size, 1, n_head_kv, n_tokens); - v = ggml_reshape_4d(ctx0, v, head_size, 1, n_head_kv, n_tokens); - ggml_tensor * tmp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, head_size, n_head / n_head_kv, n_head_kv, n_tokens); - k = ggml_repeat(ctx0, k, tmp); - v = ggml_repeat(ctx0, v, tmp); - } - - k = ggml_reshape_3d(ctx0, k, head_size, n_head, n_tokens); - v = ggml_reshape_3d(ctx0, v, head_size, n_head, n_tokens); - r = ggml_reshape_3d(ctx0, r, head_size, n_head, n_tokens); - - ggml_tensor * w = ggml_mul_mat( - ctx0, - layer.time_mix_decay_w2, - ggml_tanh( - ctx0, - ggml_mul_mat(ctx0, layer.time_mix_decay_w1, xw) - ) - ); - - w = ggml_add(ctx0, w, layer.time_mix_decay); - w = ggml_exp(ctx0, ggml_neg(ctx0, ggml_exp(ctx0, w))); - w = ggml_reshape_3d(ctx0, w, head_size, n_head, n_tokens); - - if (is_qrwkv) { - // k = k * (1 - w) - k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w)); - } - - ggml_tensor * wkv_state = build_rs( - inp, mctx_cur->get_s_l(il), - hparams.n_embd_s(), n_seqs); - - ggml_tensor * wkv_output; - if (is_qrwkv) { - wkv_output = ggml_gated_linear_attn(ctx0, k, v, r, w, wkv_state, pow(head_size, -0.5f)); - } else { - wkv_output = ggml_rwkv_wkv6(ctx0, k, v, r, layer.time_mix_first, w, wkv_state); - } - cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0); - wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); - - ggml_build_forward_expand( - gf, - ggml_cpy( - ctx0, - wkv_state, - ggml_view_1d( - ctx0, - mctx_cur->get_s_l(il), - hparams.n_embd_s() * n_seqs, - hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il)) - ) - ) - ); - - if (!is_qrwkv) { - // group norm with head_count groups - cur = ggml_reshape_3d(ctx0, cur, n_embd / n_head, n_head, n_tokens); - cur = ggml_norm(ctx0, cur, 64e-5f); - - // Convert back to regular vectors. - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b); - } else { - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - } - - cur = ggml_mul(ctx0, cur, g); - cur = build_lora_mm(layer.time_mix_output, cur); - - return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs); - } -}; - -struct llm_build_rwkv6 : public llm_build_rwkv6_base { - llm_build_rwkv6(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv6_base(model, params) { - GGML_ASSERT(hparams.token_shift_count == 2); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); - - auto * rs_inp = build_rs_inp(); - - const auto n_embd = hparams.n_embd; - const auto n_seq_tokens = ubatch.n_seq_tokens; - const auto n_seqs = ubatch.n_seqs; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const llama_layer * layer = &model.layers[il]; - inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); - - ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il); - - ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); - ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift)); - - ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il); - cb(att_norm, "attn_norm", il); - - ggml_tensor * x_prev = ggml_concat( - ctx0, - att_shift, - ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), - 1 - ); - - cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il); - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il); - cb(ffn_norm, "ffn_norm", il); - - x_prev = ggml_concat( - ctx0, - ffn_shift, - ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), - 1 - ); - - token_shift = ggml_concat(ctx0, - ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)), - ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(ffn_norm)), - 1 - ); - ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - - ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); - ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); - x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - - if (il == n_layer - 1 && inp_out_ids) { - ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); - ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); - x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - } - - cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6); - cur = ggml_add(ctx0, cur, ffn_inp); - - if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) { - cur = ggml_scale(ctx0, cur, 0.5F); - } - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -// ref: https://huggingface.co/recursal/QRWKV6-32B-Instruct-Preview-v0.1/blob/main/modeling_rwkv6qwen2.py -struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { - llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv6_base(model, params) { - GGML_ASSERT(n_embd == hparams.n_embd_r()); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - auto * rs_inp = build_rs_inp(); - - const auto n_embd = hparams.n_embd; - const auto n_seq_tokens = ubatch.n_seq_tokens; - const auto n_seqs = ubatch.n_seqs; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const llama_layer * layer = &model.layers[il]; - inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); - - ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il); - - ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il); - cb(att_norm, "attn_norm", il); - - ggml_tensor * x_prev = ggml_concat( - ctx0, - token_shift, - ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), - 1 - ); - - cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il); - - token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)); - ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); - } - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_rwkv7_base : public llm_graph_context { - const llama_model & model; - - llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) { - } - - ggml_tensor * build_rwkv7_channel_mix( - const llama_layer * layer, - ggml_tensor * cur, - ggml_tensor * x_prev, - llm_arch arch) const { - ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); - switch (arch) { - case LLM_ARCH_RWKV7: - { - ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur); - - ggml_tensor * k = ggml_sqr( - ctx0, - ggml_relu( - ctx0, - build_lora_mm(layer->channel_mix_key, xk) - ) - ); - - cur = build_lora_mm(layer->channel_mix_value, k); - } break; - default: - GGML_ABORT("fatal error"); - } - - return cur; - } - - ggml_tensor * build_rwkv7_time_mix( - llm_graph_input_rs * inp, - ggml_tensor * cur, - ggml_tensor * x_prev, - ggml_tensor *& first_layer_value, - const llama_ubatch & ubatch, - int il) const { - const auto * mctx_cur = static_cast(mctx); - - const auto n_tokens = ubatch.n_tokens; - const auto n_seqs = ubatch.n_seqs; - const auto n_embd = hparams.n_embd; - const auto head_size = hparams.wkv_head_size; - const auto head_count = n_embd / head_size; - const auto n_seq_tokens = ubatch.n_seq_tokens; - - const auto kv_head = mctx_cur->get_head(); - - const auto & layer = model.layers[il]; - - bool has_gating = layer.time_mix_g1 && layer.time_mix_g2; - - ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); - ggml_tensor * dummy = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_embd, n_seq_tokens, n_seqs, has_gating ? 6 : 5); - sx = ggml_repeat(ctx0, sx, dummy); - - ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_fused), cur); - - ggml_tensor * xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); - ggml_tensor * xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); - ggml_tensor * xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); - ggml_tensor * xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); - ggml_tensor * xa = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); - ggml_tensor * xg = has_gating ? ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 5 * sizeof(float)) : nullptr; - - ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr); - ggml_tensor * w = ggml_add( - ctx0, - ggml_mul_mat(ctx0, layer.time_mix_w2, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_w1, xw))), - layer.time_mix_w0 - ); - w = ggml_exp(ctx0, ggml_scale(ctx0, ggml_sigmoid(ctx0, w), -0.606531)); - - ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk); - ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv); - if (first_layer_value == nullptr) { - first_layer_value = v; - } else { - // Add the first layer value as a residual connection. - v = ggml_add(ctx0, v, - ggml_mul(ctx0, - ggml_sub(ctx0, first_layer_value, v), - ggml_sigmoid(ctx0, ggml_add(ctx0, - ggml_mul_mat(ctx0, layer.time_mix_v2, ggml_mul_mat(ctx0, layer.time_mix_v1, xv)), - layer.time_mix_v0 - ) - ) - ) - ); - } - - ggml_tensor * g = nullptr; - if (layer.time_mix_g1 && layer.time_mix_g2) { - g = ggml_mul_mat(ctx0, layer.time_mix_g2, ggml_sigmoid(ctx0, ggml_mul_mat(ctx0, layer.time_mix_g1, xg))); - } - - ggml_tensor * a = ggml_sigmoid(ctx0, - ggml_add( - ctx0, - ggml_mul_mat(ctx0, layer.time_mix_a2, ggml_mul_mat(ctx0, layer.time_mix_a1, xa)), - layer.time_mix_a0 - ) - ); - - ggml_tensor * kk = ggml_reshape_3d(ctx0, ggml_mul(ctx0, k, layer.time_mix_k_k), head_size, head_count, n_tokens); - kk = ggml_l2_norm(ctx0, kk, 1e-12); - - ggml_tensor * ka = ggml_mul(ctx0, k, layer.time_mix_k_a); - k = ggml_add(ctx0, k, ggml_sub(ctx0, ggml_mul(ctx0, a, ka), ka)); - - r = ggml_reshape_3d(ctx0, r, head_size, head_count, n_tokens); - w = ggml_reshape_3d(ctx0, w, head_size, head_count, n_tokens); - k = ggml_reshape_3d(ctx0, k, head_size, head_count, n_tokens); - v = ggml_reshape_3d(ctx0, v, head_size, head_count, n_tokens); - a = ggml_reshape_3d(ctx0, a, head_size, head_count, n_tokens); - - ggml_tensor * wkv_state = build_rs( - inp, mctx_cur->get_s_l(il), - hparams.n_embd_s(), n_seqs); - - ggml_tensor * wkv_output = ggml_rwkv_wkv7(ctx0, r, w, k, v, ggml_neg(ctx0, kk), ggml_mul(ctx0, kk, a), wkv_state); - cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0); - wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); - - ggml_build_forward_expand( - gf, - ggml_cpy( - ctx0, - wkv_state, - ggml_view_1d( - ctx0, - mctx_cur->get_s_l(il), - hparams.n_embd_s() * n_seqs, - hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il)) - ) - ) - ); - - if (layer.time_mix_ln && layer.time_mix_ln_b) { - // group norm with head_count groups - cur = ggml_reshape_3d(ctx0, cur, n_embd / head_count, head_count, n_tokens); - cur = ggml_norm(ctx0, cur, 64e-5f); - - // Convert back to regular vectors. - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b); - } else { - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - } - - ggml_tensor * rk = ggml_sum_rows(ctx0, - ggml_mul(ctx0, ggml_mul(ctx0, k, r), ggml_reshape_2d(ctx0, layer.time_mix_r_k, head_size, head_count))); - cur = ggml_add(ctx0, cur, ggml_reshape_2d(ctx0, ggml_mul(ctx0, v, rk), n_embd, n_tokens)); - - if (has_gating) { - cur = ggml_mul(ctx0, cur, g); - } - cur = build_lora_mm(layer.time_mix_output, cur); - - return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs); - } -}; - -struct llm_build_rwkv7 : public llm_build_rwkv7_base { - llm_build_rwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) { - GGML_ASSERT(hparams.token_shift_count == 2); - - ggml_tensor * cur; - ggml_tensor * inpL; - ggml_tensor * v_first = nullptr; - - inpL = build_inp_embd(model.tok_embd); - inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); - - auto * rs_inp = build_rs_inp(); - - const auto n_embd = hparams.n_embd; - const auto n_seq_tokens = ubatch.n_seq_tokens; - const auto n_seqs = ubatch.n_seqs; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const llama_layer * layer = &model.layers[il]; - inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); - - ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il); - - ggml_tensor * att_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); - ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], n_embd * ggml_element_size(token_shift)); - - ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il); - cb(att_norm, "attn_norm", il); - - ggml_tensor * x_prev = ggml_concat( - ctx0, - att_shift, - ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), - 1 - ); - - cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il); - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il); - cb(ffn_norm, "ffn_norm", il); - - x_prev = ggml_concat( - ctx0, - ffn_shift, - ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), - 1 - ); - - token_shift = ggml_concat(ctx0, - ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)), - ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(ffn_norm)), - 1 - ); - ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - - ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); - ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); - x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); - - if (il == n_layer - 1 && inp_out_ids) { - ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); - ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); - x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); - } - - cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7); - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - - -struct llm_build_arwkv7 : public llm_build_rwkv7_base { - llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) { - GGML_ASSERT(n_embd == hparams.n_embd_r()); - - ggml_tensor * cur; - ggml_tensor * inpL; - ggml_tensor * v_first = nullptr; - - inpL = build_inp_embd(model.tok_embd); - - auto * rs_inp = build_rs_inp(); - - const auto n_embd = hparams.n_embd; - const auto n_seq_tokens = ubatch.n_seq_tokens; - const auto n_seqs = ubatch.n_seqs; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const llama_layer * layer = &model.layers[il]; - inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); - - ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il); - - ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il); - cb(att_norm, "attn_norm", il); - - ggml_tensor * x_prev = ggml_concat( - ctx0, - token_shift, - ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), - 1 - ); - - cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il); - - token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)); - ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); - cb(ffn_inp, "ffn_inp", il); - - cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); - ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); - } - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_granite : public llm_graph_context { - llm_build_granite( - const llama_model & model, - const llm_graph_params & params) - : llm_graph_context(params) { - - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - built only if rope enabled - ggml_tensor * inp_pos = nullptr; - if (hparams.rope_finetuned) { - inp_pos = build_inp_pos(); - } - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - cur = build_attention_layer( - cur, inp_pos, inp_attn, - model, n_embd_head, il); - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // ffn - cur = build_layer_ffn(cur, inpSA, model, il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - // For Granite architectures - scale logits - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } - - ggml_tensor * build_attention_layer( - ggml_tensor * cur, - ggml_tensor * inp_pos, - llm_graph_input_attn_kv * inp_attn, - const llama_model & model, - const int64_t n_embd_head, - const int il) { - - // compute Q and K and (optionally) RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); - - const bool use_rope = hparams.rope_finetuned; - if (use_rope) { - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - return cur; - } - - ggml_tensor * build_layer_ffn( - ggml_tensor * cur, - ggml_tensor * inpSA, - const llama_model & model, - const int il) { - - // For Granite architectures - scale residual - if (hparams.f_residual_scale) { - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - } - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network (non-MoE) - if (model.layers[il].ffn_gate_inp == nullptr) { - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - } else { - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - - // For Granite MoE Shared - if (hparams.n_ff_shexp > 0) { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } else { - cur = moe_out; - } - } - - // For Granite architectures - scale residual - if (hparams.f_residual_scale) { - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - } - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - return cur; - } -}; - -struct llm_build_granite_hybrid : public llm_graph_context_mamba { - llm_build_granite_hybrid( - const llama_model & model, - const llm_graph_params & params) : - llm_graph_context_mamba(params) { - - const int64_t n_embd_head = hparams.n_embd_head_v; - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - auto * inp = build_inp_mem_hybrid(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - // Positional embeddings populated if rope enabled - ggml_tensor * inp_pos = nullptr; - if (hparams.rope_finetuned) { - inp_pos = build_inp_pos(); - } - - for (int il = 0; il < n_layer; ++il) { - struct ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - if (hparams.is_recurrent(il)) { - // ssm layer // - cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il); - } else { - // attention layer // - cur = build_attention_layer( - cur, inp_pos, inp->get_attn(), model, - n_embd_head, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - // ffn - cur = build_layer_ffn(cur, inpSA, model, il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - // For Granite architectures - scale logits - if (hparams.f_logit_scale) { - cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); - } - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } - - ggml_tensor * build_attention_layer( - ggml_tensor * cur, - ggml_tensor * inp_pos, - llm_graph_input_attn_kv * inp_attn, - const llama_model & model, - const int64_t n_embd_head, - const int il) { - - // compute Q and K and (optionally) RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); - - const bool use_rope = hparams.rope_finetuned; - if (use_rope) { - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - return cur; - } - - ggml_tensor * build_layer_ffn( - ggml_tensor * cur, - ggml_tensor * inpSA, - const llama_model & model, - const int il) { - - // For Granite architectures - scale residual - if (hparams.f_residual_scale) { - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - } - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network (non-MoE) - if (model.layers[il].ffn_gate_inp == nullptr) { - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - } else { - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - - // For Granite MoE Shared - if (hparams.n_ff_shexp > 0) { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } else { - cur = moe_out; - } - } - - // For Granite architectures - scale residual - if (hparams.f_residual_scale) { - cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); - } - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - return cur; - } -}; - -// ref: https://github.com/facebookresearch/chameleon -// based on the original build_llama() function, changes: -// * qk-norm -// * swin-norm -// * removed bias -// * removed MoE -struct llm_build_chameleon : public llm_graph_context { - llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - if (hparams.swin_norm) { - cur = inpL; - } else { - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - } - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - if (model.layers[il].attn_q_norm) { - Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens, - ggml_element_size(Qcur) * n_embd_head, - ggml_element_size(Qcur) * n_embd_head * n_head, - 0); - cb(Qcur, "Qcur", il); - - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, - model.layers[il].attn_q_norm_b, - LLM_NORM, il); - cb(Qcur, "Qcur", il); - } - - if (model.layers[il].attn_k_norm) { - Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens, - ggml_element_size(Kcur) * n_embd_head, - ggml_element_size(Kcur) * n_embd_head * n_head_kv, - 0); - cb(Kcur, "Kcur", il); - - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, - model.layers[il].attn_k_norm_b, - LLM_NORM, il); - cb(Kcur, "Kcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, nullptr, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - if (hparams.swin_norm) { - cur = build_norm(cur, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - if (!hparams.swin_norm) { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - } - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - if (hparams.swin_norm) { - cur = build_norm(cur, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - cb(cur, "result_output_with_img_logits", -1); - - // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs. - // Needs to be removed once image outputs are supported. - int img_token_end_idx = 8196; - int img_token_start_idx = 4; - int num_img_tokens = img_token_end_idx - img_token_start_idx; - // creates 1d tensor of size num_img_tokens and values -FLT_MAX, - // which ensures that text token values are always at least larger than image token values - ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens); - img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX); - cb(img_logits, "img_logits", -1); - - cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_wavtokenizer_dec : public llm_graph_context { - llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, inpL)); - - cur = ggml_conv_1d_ph(ctx0, model.conv1d, cur, 1, 1); - cur = ggml_add(ctx0, cur, model.conv1d_b); - - // posnet - for (uint32_t il = 0; il < hparams.posnet.n_layer; ++il) { - const auto & layer = model.layers[il].posnet; - - inpL = cur; - - switch (il) { - case 0: - case 1: - case 3: - case 4: - { - cur = build_norm(cur, - layer.norm1, - layer.norm1_b, - LLM_NORM_GROUP, 0); - - cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); - - cur = ggml_conv_1d_ph(ctx0, layer.conv1, cur, 1, 1); - cur = ggml_add(ctx0, cur, layer.conv1_b); - - cur = build_norm(cur, - layer.norm2, - layer.norm2_b, - LLM_NORM_GROUP, 0); - - cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); - - cur = ggml_conv_1d_ph(ctx0, layer.conv2, cur, 1, 1); - cur = ggml_add(ctx0, cur, layer.conv2_b); - - cur = ggml_add(ctx0, cur, inpL); - } break; - case 2: - { - cur = build_norm(cur, - layer.attn_norm, - layer.attn_norm_b, - LLM_NORM_GROUP, 0); - - ggml_tensor * q; - ggml_tensor * k; - ggml_tensor * v; - - q = ggml_conv_1d_ph(ctx0, layer.attn_q, cur, 1, 1); - k = ggml_conv_1d_ph(ctx0, layer.attn_k, cur, 1, 1); - v = ggml_conv_1d_ph(ctx0, layer.attn_v, cur, 1, 1); - - q = ggml_add(ctx0, q, layer.attn_q_b); - k = ggml_add(ctx0, k, layer.attn_k_b); - v = ggml_add(ctx0, v, layer.attn_v_b); - - q = ggml_cont(ctx0, ggml_transpose(ctx0, q)); - k = ggml_cont(ctx0, ggml_transpose(ctx0, k)); - - ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); - - kq = ggml_soft_max_ext(ctx0, kq, nullptr, 1.0f/sqrtf(float(hparams.posnet.n_embd)), 0.0f); - - cur = ggml_mul_mat(ctx0, kq, v); - - cur = ggml_conv_1d_ph(ctx0, layer.attn_o, cur, 1, 1); - cur = ggml_add(ctx0, cur, layer.attn_o_b); - - cur = ggml_add(ctx0, cur, inpL); - } break; - case 5: - { - cur = build_norm(cur, - layer.norm, - layer.norm_b, - LLM_NORM_GROUP, 0); - } break; - default: GGML_ABORT("unknown posnet layer"); - }; - } - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - cur = build_norm(cur, - model.tok_norm, - model.tok_norm_b, - LLM_NORM, -1); - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - inpL = cur; - - // convnext - for (uint32_t il = 0; il < hparams.convnext.n_layer; ++il) { - const auto & layer = model.layers[il].convnext; - - cur = inpL; - - cur = ggml_conv_1d_dw_ph(ctx0, layer.dw, cur, 1, 1); - cur = ggml_add(ctx0, cur, layer.dw_b); - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - cur = build_norm(cur, - layer.norm, - layer.norm_b, - LLM_NORM, -1); - - cur = build_ffn(cur, - layer.pw1, layer.pw1_b, NULL, - NULL, NULL, NULL, - layer.pw2, layer.pw2_b, NULL, - NULL, - LLM_FFN_GELU, LLM_FFN_SEQ, il); - - cur = ggml_mul(ctx0, cur, layer.gamma); - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - inpL = ggml_add(ctx0, cur, inpL); - } - - cur = inpL; - - cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); - - cur = build_norm(cur, - model.output_norm, - model.output_norm_b, - LLM_NORM, -1); - - // lm_head - cur = build_lora_mm(model.output, cur); - - cur = ggml_add(ctx0, cur, model.output_b); - - cb(cur, "result_embd", -1); - res->t_embd = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_plm : public llm_graph_context { - llm_build_plm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k)); - - const uint32_t n_embd_head_qk_rope = hparams.n_rot; - const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; - const uint32_t kv_lora_rank = hparams.n_lora_kv; - - ggml_tensor * cur; - ggml_tensor * inpL; - - // {n_embd, n_tokens} - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - ggml_tensor * q = NULL; - q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); - cb(q, "q", il); - - // split into {n_head * n_embd_head_qk_nope, n_tokens} - ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), - 0); - cb(q_nope, "q_nope", il); - - // and {n_head * n_embd_head_qk_rope, n_tokens} - ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, - ggml_row_size(q->type, hparams.n_embd_head_k), - ggml_row_size(q->type, hparams.n_embd_head_k * n_head), - ggml_row_size(q->type, n_embd_head_qk_nope)); - cb(q_pe, "q_pe", il); - - // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} - ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); - cb(kv_pe_compresseed, "kv_pe_compresseed", il); - - // split into {kv_lora_rank, n_tokens} - ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, - kv_pe_compresseed->nb[1], - 0); - cb(kv_compressed, "kv_compressed", il); - - // and {n_embd_head_qk_rope, n_tokens} - ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, - kv_pe_compresseed->nb[1], - kv_pe_compresseed->nb[1], - ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); - cb(k_pe, "k_pe", il); - - kv_compressed = build_norm(kv_compressed, - model.layers[il].attn_kv_a_norm, NULL, - LLM_NORM_RMS, il); - cb(kv_compressed, "kv_compressed", il); - - // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} - ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); - cb(kv, "kv", il); - - // split into {n_head * n_embd_head_qk_nope, n_tokens} - ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, - ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), - ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), - 0); - cb(k_nope, "k_nope", il); - - // and {n_head * n_embd_head_v, n_tokens} - ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), - ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), - ggml_row_size(kv->type, (n_embd_head_qk_nope))); - cb(v_states, "v_states", il); - - v_states = ggml_cont(ctx0, v_states); - cb(v_states, "v_states", il); - - v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, - ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), - 0); - cb(v_states, "v_states", il); - - q_pe = ggml_rope_ext( - ctx0, q_pe, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(q_pe, "q_pe", il); - - // shared RoPE key - k_pe = ggml_rope_ext( - ctx0, k_pe, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - cb(k_pe, "k_pe", il); - - ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); - cb(q_states, "q_states", il); - - ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); - cb(k_states, "k_states", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_bailingmoe : public llm_graph_context { - llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = - build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, hparams.expert_weights_norm, - false, hparams.expert_weights_scale, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - - // FFN shared expert - { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_bailingmoe2 : public llm_graph_context { - llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - const int n_transformer_layers = n_layer - hparams.nextn_predict_layers; - for (int il = 0; il < n_transformer_layers; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - cur = build_lora_mm(model.layers[il].wqkv, cur); - cb(cur, "wqkv", il); - - ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_transformer_layers - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA); - cb(sa_out, "sa_out", il); - - // MoE branch - cur = build_norm(sa_out, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - if (static_cast(il) < hparams.n_layer_dense_lead) { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - ggml_tensor * moe_out = - build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - model.layers[il].ffn_exp_probs_b, - n_expert, n_expert_used, - LLM_FFN_SILU, hparams.expert_weights_norm, - true, hparams.expert_weights_scale, - (llama_expert_gating_func_type) hparams.expert_gating_func, - il); - cb(moe_out, "ffn_moe_out", il); - - { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } - } - - cur = ggml_add(ctx0, cur, sa_out); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_dots1 : public llm_graph_context { - llm_build_dots1(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - if ((uint32_t) il < hparams.n_layer_dense_lead) { - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - ggml_tensor * moe_out = - build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - model.layers[il].ffn_exp_probs_b, - n_expert, n_expert_used, - LLM_FFN_SILU, hparams.expert_weights_norm, - true, hparams.expert_weights_scale, - (llama_expert_gating_func_type) hparams.expert_gating_func, - il); - cb(moe_out, "ffn_moe_out", il); - - { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - cb(cur, "ffn_out", il); - } - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_ernie4_5 : public llm_graph_context { - llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - { - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - } - - // self-attention - { - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_ernie4_5_moe : public llm_graph_context { - llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0"); - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - // norm - { - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - } - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - bool is_moe_layer = static_cast(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0; - - if (!is_moe_layer) { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } else { - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * moe_out = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - model.layers[il].ffn_exp_probs_b, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(moe_out, "ffn_moe_out", il); - - // Shared expert (if present) - if (hparams.n_ff_shexp > 0) { - ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(ffn_shexp, "ffn_shexp", il); - - cur = ggml_add(ctx0, moe_out, ffn_shexp); - } else { - cur = moe_out; - } - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_falcon_h1 : public llm_graph_context_mamba { - llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - // Build the inputs in the recurrent & kv cache - auto * inp = build_inp_mem_hybrid(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur-post-rope", il); - cb(Kcur, "Kcur-post-rope", il); - cb(Vcur, "Vcur-post-rope", il); - - ggml_tensor * attn_out = build_attn(inp->get_attn(), - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(attn_out, "attn_out", il); - - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - // Mamba2 layer - cb(cur, "ssm_in", il); - - ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il); - cb(ssm_out, "ssm_out", il); - - // // Aggregation - cur = ggml_add(ctx0, attn_out, ssm_out); - inpSA = ggml_add(ctx0, cur, inpSA); - cb(cur, "layer_out", il); - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = inpSA; - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, inpSA); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_plamo2 : public llm_graph_context_mamba { - llm_build_plamo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) { - ggml_tensor * cur; - ggml_tensor * inpL; - - // {n_embd, n_tokens} - inpL = build_inp_embd(model.tok_embd); - cb(inpL, "embedding_output", -1); - - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_hybrid = build_inp_mem_hybrid(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * residual = inpL; - - // ggml_graph_add_node(gf, model.layers[il].attn_norm); - // cb(model.layers[il].attn_norm, "attn_norm", il); - - // pre_mixer_norm - cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - - // check if this layer is Mamba or Attention - bool is_mamba_layer = hparams.is_recurrent(il); - - if (is_mamba_layer) { - // PLaMo-2 Mamba layer - cur = build_plamo2_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il); - } else { - // PLaMo-2 Attention layer - cur = build_plamo2_attn_layer(inp_hybrid->get_attn(), inp_pos, cur, model, il); - } - - // post_mixer_norm - cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - // residual connection - cur = ggml_add(ctx0, cur, residual); - cb(cur, "attn_residual", il); - residual = cur; - - // pre-ffn norm - cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "ffn_pre_norm", il); - - // feed-forward network - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - - // post ffn norm - cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "ffn_post_norm", il); - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - residual = ggml_get_rows(ctx0, residual, inp_out_ids); - } - - // residual connection - cur = ggml_add(ctx0, cur, residual); - cb(cur, "ffn_residual", il); - - inpL = cur; - } - - cur = inpL; - - // final norm - cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = build_lora_mm(model.output, cur); - cb(cur, "result_output", -1); - - // Explicitly mark as output tensor to ensure proper backend assignment - ggml_set_output(cur); - - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } - -private: - ggml_tensor * build_plamo2_attn_layer( - llm_graph_input_attn_kv * inp, - ggml_tensor * inp_pos, - ggml_tensor * cur, - const llama_model & model, - int il) { - - // self-attention - { - // PLaMo-2 uses combined QKV tensor - ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur); - cb(qkv, "wqkv", il); - - // split QKV tensor into Q, K, V - const int64_t n_embd_head_q = hparams.n_embd_head_k; - const int64_t n_embd_head_k = hparams.n_embd_head_k; - const int64_t n_embd_head_v = hparams.n_embd_head_v; - int32_t n_head = hparams.n_head(il); - int32_t n_head_kv = hparams.n_head_kv(il); - - const int64_t q_offset = 0; - const int64_t k_offset = n_embd_head_q * n_head; - const int64_t v_offset = k_offset + n_embd_head_k * n_head_kv; - - ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head, n_tokens, n_embd_head_q * sizeof(float), qkv->nb[1], q_offset * ggml_element_size(qkv)); - ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv, n_tokens, n_embd_head_k * sizeof(float), qkv->nb[1], k_offset * ggml_element_size(qkv)); - ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv, n_tokens, n_embd_head_v * sizeof(float), qkv->nb[1], v_offset * ggml_element_size(qkv)); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cur = build_attn(inp, - model.layers[il].wo, NULL, - Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head_v)), il); - } - - cb(cur, "attn_out", il); - - return cur; - } - - ggml_tensor * build_plamo2_mamba_layer( - llm_graph_input_rs * inp, - ggml_tensor * cur, - const llama_model & model, - const llama_ubatch & ubatch, - int il) { - - const auto * mctx_cur = inp->mctx; - - const auto kv_head = mctx_cur->get_head(); - - const int64_t d_conv = hparams.ssm_d_conv; - const int64_t d_inner = hparams.ssm_d_inner; - const int64_t d_state = hparams.ssm_d_state; - const int64_t n_heads = hparams.ssm_dt_rank; - const int64_t head_dim = d_inner / n_heads; - const int64_t n_group = hparams.ssm_n_group; - const int64_t n_seqs = ubatch.n_seqs; - - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - - GGML_ASSERT(n_seqs != 0); - GGML_ASSERT(ubatch.equal_seqs()); - GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); - - ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); - ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); - - ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); - conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2*n_group*d_state, n_seqs); - - // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} - cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); - - // in_proj: {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} - ggml_tensor * zx = build_lora_mm(model.layers[il].ssm_in, cur); - cb(zx, "mamba_in_proj", il); - // {8192, 5, 1, 1} -> {8192, 1, 5, 1} - zx = ggml_permute(ctx0, zx, 0, 2, 1, 3); - zx = ggml_cont_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs); - cb(zx, "mamba_in_proj_out", il); - - // split into z and x - // => {head_dim * n_heads, n_seq_tokens, n_seqs} - ggml_tensor * x = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], head_dim*ggml_element_size(zx)); - x = ggml_cont_3d(ctx0, x, head_dim * n_heads, n_seq_tokens, n_seqs); - // x = ggml_permute(ctx0, x, 0, 2, 1, 3); - cb(x, "mamba_x_split", il); - - ggml_tensor * z = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], 0); - cb(z, "mamba_z_split", il); - - // conv1d - { - // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs} - ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0); - cb(conv_x, "mamba_conv1d_input", il); - - // copy last (d_conv - 1) columns back into the state cache - ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, - conv_x->nb[1], conv_x->nb[2], n_seq_tokens*(conv_x->nb[0])); - - ggml_build_forward_expand(gf, - ggml_cpy(ctx0, last_conv, - ggml_view_1d(ctx0, conv_states_all, - (d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs), - kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all)))); - cb(conv_states_all, "mamba_conv1d_state", il); - - // 1D convolution - x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d); - cb(x, "mamba_conv1d", il); - - x = ggml_silu(ctx0, x); - cb(x, "mamba_conv1d_silu", il); - } - - // SSM - { - // bcdt_proj: {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} - ggml_tensor * x_bcdt = build_lora_mm(model.layers[il].ssm_x, x); - cb(x_bcdt, "mamba_bcdt_proj", il); - - // split into dt, B, C - const int64_t dt_dim = std::max(64, int(hparams.n_embd / 16)); - ggml_tensor * B = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], 0); - ggml_tensor * C = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], ggml_element_size(x_bcdt)*d_state); - ggml_tensor * dt = ggml_view_3d(ctx0, x_bcdt, dt_dim, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], ggml_element_size(x_bcdt)*(2*d_state)); - cb(B, "mamba_B_raw", il); - cb(C, "mamba_C_raw", il); - cb(dt, "mamba_dt_raw", il); - - // Apply RMS norm to dt, B, C (PLaMo-2 specific) - B = build_norm(B, model.layers[il].ssm_b_norm, NULL, LLM_NORM_RMS, il); - C = build_norm(C, model.layers[il].ssm_c_norm, NULL, LLM_NORM_RMS, il); - dt = build_norm(dt, model.layers[il].ssm_dt_norm, NULL, LLM_NORM_RMS, il); - cb(B, "mamba_B_normed", il); - cb(C, "mamba_C_normed", il); - cb(dt, "mamba_dt_normed", il); - - // dt_proj: {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} - dt = build_lora_mm(model.layers[il].ssm_dt, dt); - dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b); - cb(dt, "mamba_dt_proj", il); - - ggml_tensor * A = ggml_reshape_2d(ctx0, model.layers[il].ssm_a, 1, n_heads); - cb(A, "mamba_A", il); - - x = ggml_view_4d(ctx0, x, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x), head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0); - B = ggml_view_4d(ctx0, B, d_state, 1, n_seq_tokens, n_seqs, d_state * B->nb[0], B->nb[1], B->nb[2], 0); - C = ggml_view_4d(ctx0, C, d_state, 1, n_seq_tokens, n_seqs, d_state * C->nb[0], C->nb[1], C->nb[2], 0); - - // use the states and the indices provided by build_recurrent_state - // (this is necessary in order to properly use the states before they are overwritten, - // while avoiding to make unnecessary copies of the states) - auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) { - ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_heads, mctx_cur->get_size()); - - // Custom operator to optimize the parallel associative scan - // as described in the Annex D of the Mamba paper. - // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} - return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids); - }; - - ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows); - cb(y_ssm, "mamba_ssm_scan", il); - - // store last states - ggml_build_forward_expand(gf, - ggml_cpy(ctx0, - ggml_view_1d(ctx0, y_ssm, n_heads*head_dim*d_state*n_seqs, n_heads*head_dim*n_seq_tokens*n_seqs*ggml_element_size(y_ssm)), - ggml_view_1d(ctx0, ssm_states_all, n_heads*head_dim*d_state*n_seqs, kv_head*n_seqs*n_heads*head_dim*d_state*ggml_element_size(ssm_states_all)))); - cb(ssm_states_all, "mamba_ssm_states", il); - - ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x), head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0); - cb(y, "mamba_y_view", il); - - // Add D parameter and apply gating with z - // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs} - ggml_tensor * D = ggml_reshape_2d(ctx0, model.layers[il].ssm_d, 1, n_heads); - y = ggml_add(ctx0, y, ggml_mul(ctx0, x, D)); - cb(y, "mamba_y_add_d", il); - - y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y); - cb(y, "mamba_y_swiglu_z", il); - - // out_proj: {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} - y = ggml_view_3d(ctx0, y, head_dim * n_heads, n_seq_tokens, n_seqs, y->nb[2], y->nb[3], 0); - cur = build_lora_mm(model.layers[il].ssm_out, y); - cb(cur, "mamba_out_proj", il); - } - - // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} - cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); - cb(cur, "mamba_out", il); - - return cur; - } -}; - -struct llm_build_arcee : public llm_graph_context { - llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - // ARCEE uses relu^2 instead of silu - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - NULL, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_hunyuan_moe : public llm_graph_context { - llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, nullptr, - LLM_NORM_RMS, il); - cb(Kcur, "Kcur_norm", il); - - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, nullptr, - LLM_NORM_RMS, il); - cb(Qcur, "Qcur_norm", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // feed-forward network (non-MoE) - ggml_tensor * cur_mlp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, - model.layers[il].ffn_gate_shexp, NULL, NULL, - model.layers[il].ffn_down_shexp, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur_mlp, "ffn_mlp", il); - - // MoE branch - ggml_tensor * cur_moe = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, - true, // norm_topk_prob - false, - 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il); - cb(cur_moe, "ffn_moe_out", il); - - ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp); - cb(ffn_out, "ffn_out", il); - - cur = ggml_add(ctx0, ffn_out, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_hunyuan_dense : public llm_graph_context { - llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - // self-attention - { - // rope freq factors for llama3; may return nullptr for llama2 and other models - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = build_norm(Kcur, - model.layers[il].attn_k_norm, nullptr, - LLM_NORM_RMS, il); - cb(Kcur, "Kcur_norm", il); - - Qcur = build_norm(Qcur, - model.layers[il].attn_q_norm, nullptr, - LLM_NORM_RMS, il); - cb(Qcur, "Qcur_norm", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - // feed-forward network (non-MoE) - ggml_tensor * cur_mlp = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur_mlp, "ffn_out", il); - - cur = ggml_add(ctx0, cur_mlp, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - // lm_head - cur = build_lora_mm(model.output, cur); - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_smollm3 : public llm_graph_context { - llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - if (use_rope) { - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, - model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, - model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_openai_moe_iswa : public llm_graph_context { - llm_build_openai_moe_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv_iswa(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, nullptr, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il); - - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1) { - // skip computing output for unused tokens - ggml_tensor * inp_out_ids = build_inp_out_ids(); - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - cur = ffn_inp; - cur = build_norm(cur, - model.layers[il].attn_post_norm, nullptr, - LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - // MoE branch - cur = build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, model.layers[il].ffn_gate_inp_b, - model.layers[il].ffn_up_exps, model.layers[il].ffn_up_exps_b, - model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b, - model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SWIGLU_OAI_MOE, false, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT, - il); - cb(cur, "ffn_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_lfm2 : public llm_graph_context { - const llama_model & model; - - llm_build_lfm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params), model(model) { - - ggml_tensor * cur = build_inp_embd(model.tok_embd); - cb(cur, "model.embed_tokens", -1); - - ggml_tensor * inp_pos = build_inp_pos(); - auto * inp_hybrid = build_inp_mem_hybrid(); - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - const bool is_moe_layer = il >= static_cast(hparams.n_layer_dense_lead); - - auto * prev_cur = cur; - cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "model.layers.{}.operator_norm", il); - - cur = hparams.is_recurrent(il) ? - build_shortconv_block(cur, inp_hybrid->get_recr(), il) : - build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il) ; - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids); - } - - cur = ggml_add(ctx0, prev_cur, cur); - - auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); - cb(ffn_norm_out, "model.layers.{}.ffn_norm", il); - - ggml_tensor * ffn_out = is_moe_layer ? - build_moe_feed_forward(ffn_norm_out, il) : - build_dense_feed_forward(ffn_norm_out, il); - cb(ffn_norm_out, "model.layers.{}.ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_out); - } - - cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1); - cb(cur, "model.embedding_norm", -1); - res->t_embd = cur; - - cur = build_lora_mm(model.output, cur); - cb(cur, "lm_head", -1); - - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } - - ggml_tensor * build_moe_feed_forward(ggml_tensor * cur, - int il) const { - return build_moe_ffn(cur, - model.layers[il].ffn_gate_inp, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - model.layers[il].ffn_exp_probs_b, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - static_cast(hparams.expert_gating_func), - il); - } - - ggml_tensor * build_dense_feed_forward(ggml_tensor * cur, - int il) const { - GGML_ASSERT(!model.layers[il].ffn_up_b); - GGML_ASSERT(!model.layers[il].ffn_gate_b); - GGML_ASSERT(!model.layers[il].ffn_down_b); - return build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - } - - ggml_tensor * build_attn_block(ggml_tensor * cur, - ggml_tensor * inp_pos, - llm_graph_input_attn_kv * inp_attn, - int il) const { - GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il)); - auto const n_embd_head = hparams.n_embd_head_v; - auto const n_head_kv = hparams.n_head_kv(il); - - auto * q = build_lora_mm(model.layers[il].wq, cur); - cb(q, "model.layers.{}.self_attn.q_proj", il); - auto * k = build_lora_mm(model.layers[il].wk, cur); - cb(k, "model.layers.{}.self_attn.k_proj", il); - auto * v = build_lora_mm(model.layers[il].wv, cur); - cb(v, "model.layers.{}.self_attn.v_proj", il); - - q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens); - k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens); - v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens); - - // qk norm - q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(q, "model.layers.{}.self_attn.q_layernorm", il); - k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(k, "model.layers.{}.self_attn.k_layernorm", il); - - // RoPE - q = ggml_rope_ext( - ctx0, q, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - k = ggml_rope_ext( - ctx0, k, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cur = build_attn(inp_attn, model.layers[il].wo, NULL, - q, k, v, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - - cb(cur, "model.layers.{}.self_attn.out_proj", il); - - return cur; - } - - ggml_tensor * build_shortconv_block(ggml_tensor * cur, - llm_graph_input_rs * inp_recr, - int il) { - const auto * mctx_cur = static_cast(mctx)->get_recr(); - const uint32_t kv_head = mctx_cur->get_head(); - const int64_t n_seq_tokens = ubatch.n_seq_tokens; - const int64_t n_seqs = ubatch.n_seqs; - GGML_ASSERT(n_seqs != 0); - GGML_ASSERT(ubatch.equal_seqs()); - GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); - - GGML_ASSERT(hparams.n_shortconv_l_cache > 1); - const uint32_t d_conv = hparams.n_shortconv_l_cache - 1; - - // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} - cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); - - auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur); - cb(bcx, "model.layers.{}.conv.in_proj", il); - - constexpr auto n_chunks = 3; - GGML_ASSERT(bcx->ne[0] % n_chunks == 0); - auto const chunk_size = bcx->ne[0] / n_chunks; - auto * b = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], 0*chunk_size*ggml_element_size(bcx)); - auto * c = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], 1*chunk_size*ggml_element_size(bcx)); - auto * x = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], 2*chunk_size*ggml_element_size(bcx)); - - auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x)); - - // read conv state - auto * conv_state = mctx_cur->get_r_l(il); - auto * conv_rs = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs); - auto * conv = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs); - - bx = ggml_concat(ctx0, conv, bx, 0); - GGML_ASSERT(bx->ne[0] > conv->ne[0]); - - // last d_conv columns is a new conv state - auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2], (bx->ne[0] - conv->ne[0])*ggml_element_size(bx)); - GGML_ASSERT(ggml_are_same_shape(conv, new_conv)); - - // write new conv conv state - ggml_build_forward_expand( - gf, - ggml_cpy( - ctx0, - new_conv, - ggml_view_1d( - ctx0, - conv_state, - ggml_nelements(new_conv), - kv_head*d_conv*n_embd*ggml_element_size(new_conv) - ) - ) - ); - - auto * conv_kernel = model.layers[il].shortconv.conv; - auto * conv_out = ggml_ssm_conv(ctx0, bx, conv_kernel); - cb(conv_out, "model.layers.{}.conv.conv", il); - - auto * y = ggml_mul(ctx0, c, conv_out); - y = build_lora_mm(model.layers[il].shortconv.out_proj, y); - cb(y, "model.layers.{}.conv.out_proj", il); - // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} - y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs); - - return y; - } -}; - -struct llm_build_seed_oss : public llm_graph_context { - llm_build_seed_oss(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - if (model.layers[il].bq) { - Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); - cb(Qcur, "Qcur", il); - } - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - if (model.layers[il].bk) { - Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); - cb(Kcur, "Kcur", il); - } - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - if (model.layers[il].bv) { - Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); - cb(Vcur, "Vcur", il); - } - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network - cur = build_norm(ffn_inp, - model.layers[il].attn_post_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_post_norm", il); - - cur = build_ffn(cur, - model.layers[il].ffn_up, NULL, NULL, - model.layers[il].ffn_gate, NULL, NULL, - model.layers[il].ffn_down, NULL, NULL, - NULL, - LLM_FFN_SILU, LLM_FFN_PAR, il); - cb(cur, "ffn_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -template -struct llm_build_smallthinker : public llm_graph_context{ - llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){ - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - using inp_attn_type = std::conditional_t; - inp_attn_type * inp_attn = nullptr; - - if constexpr (iswa) { - inp_attn = build_attn_inp_kv_iswa(); - } else { - inp_attn = build_attn_inp_kv(); - } - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - ggml_tensor * probs = nullptr; - - probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL); // [n_expert, n_tokens] - cb(probs, "ffn_moe_logits", il); - - // norm - cur = build_norm(inpL,model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - if (hparams.n_no_rope_layer_step == n_layer || il % hparams.n_no_rope_layer_step != 0) { - Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - - Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow); - } - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - probs = ggml_get_rows(ctx0, probs, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * ffn_out = - build_moe_ffn(cur, - nullptr, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_RELU, true, - false, 0.0, - static_cast(hparams.expert_gating_func), - il, probs); - - cb(ffn_out, "ffn_out", il); - cur = ffn_out; - - cur = ggml_add(ctx0, cur, ffn_inp); - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); - cb(cur, "result_norm", -1); - - // lm_head - cur = build_lora_mm(model.output, cur); - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_grovemoe : public llm_graph_context { - llm_build_grovemoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - const int64_t n_chunk_expert = n_expert / hparams.n_group_experts; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - // inp_pos - contains the positions - ggml_tensor * inp_pos = build_inp_pos(); - - auto * inp_attn = build_attn_inp_kv(); - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - // norm - cur = build_norm(inpL, - model.layers[il].attn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self_attention - { - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, nullptr, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur", il); - cb(Kcur, "Kcur", il); - cb(Vcur, "Vcur", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // MoE branch - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, NULL, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, cur); // [n_expert, n_tokens] - cb(probs, "ffn_moe_logits", il); - - ggml_tensor * moe_out = - build_moe_ffn(cur, - nullptr, - model.layers[il].ffn_up_exps, - model.layers[il].ffn_gate_exps, - model.layers[il].ffn_down_exps, - nullptr, - n_expert, n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il, probs); - cb(moe_out, "ffn_moe_out", il); - cur = moe_out; - - // TODO: Only do the expert selection and weights once - moe_out = - build_moe_ffn(cur, - nullptr, - model.layers[il].ffn_up_chexps, - model.layers[il].ffn_gate_chexps, - model.layers[il].ffn_down_chexps, - nullptr, - n_chunk_expert, n_expert_used > n_chunk_expert ? n_chunk_expert : n_expert_used, - LLM_FFN_SILU, true, - false, 0.0, - LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, - il, probs); - cb(moe_out, "ffn_adj_moe_out", il); - - cur = ggml_add(ctx0, cur, ggml_scale(ctx0, moe_out, hparams.expert_group_scale)); - cb(cur, "ffn_final_moe_out", il); - - cur = ggml_add(ctx0, cur, ffn_inp); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, NULL, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -struct llm_build_apertus : public llm_graph_context { - llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { - const int64_t n_embd_head = hparams.n_embd_head_v; - - GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); - GGML_ASSERT(n_embd_head == hparams.n_rot); - - ggml_tensor * cur; - ggml_tensor * inpL; - - inpL = build_inp_embd(model.tok_embd); - - ggml_tensor * inp_pos = build_inp_pos(); - auto * inp_attn = build_attn_inp_kv(); - - const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; - - ggml_tensor * inp_out_ids = build_inp_out_ids(); - - for (int il = 0; il < n_layer; ++il) { - ggml_tensor * inpSA = inpL; - - cur = build_norm(inpL, - model.layers[il].attn_norm, nullptr, - LLM_NORM_RMS, il); - cb(cur, "attn_norm", il); - - // self-attention - { - ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); - - // compute Q and K and RoPE them - ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); - cb(Qcur, "Qcur", il); - - ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); - cb(Kcur, "Kcur", il); - - ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); - cb(Vcur, "Vcur", il); - - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", il); - - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); - - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - - Qcur = ggml_rope_ext( - ctx0, Qcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - Kcur = ggml_rope_ext( - ctx0, Kcur, inp_pos, rope_factors, - n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, - ext_factor, attn_factor, beta_fast, beta_slow - ); - - cb(Qcur, "Qcur_pos", il); - cb(Kcur, "Kcur_pos", il); - cb(Vcur, "Vcur_pos", il); - - cur = build_attn(inp_attn, - model.layers[il].wo, model.layers[il].bo, - Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); - cb(cur, "attn_out", il); - } - - if (il == n_layer - 1 && inp_out_ids) { - cur = ggml_get_rows(ctx0, cur, inp_out_ids); - inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); - } - - ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); - cb(ffn_inp, "ffn_inp", il); - - // feed-forward network with xIELU activation - { - cur = build_norm(ffn_inp, - model.layers[il].ffn_norm, nullptr, - LLM_NORM_RMS, il); - cb(cur, "ffn_norm", il); - - // Up projection - ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur); - cb(up, "ffn_up", il); - - float alpha_n_val = hparams.xielu_alpha_n[il]; - float alpha_p_val = hparams.xielu_alpha_p[il]; - float beta_val = hparams.xielu_beta[il]; - float eps_val = hparams.xielu_eps[il]; - - // Apply xIELU activation - ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val); - cb(activated, "ffn_xielu", il); - - // Down projection - cur = build_lora_mm(model.layers[il].ffn_down, activated); - cb(cur, "ffn_down", il); - } - - cur = ggml_add(ctx0, cur, ffn_inp); - cb(cur, "ffn_out", il); - - cur = build_cvec(cur, il); - cb(cur, "l_out", il); - - // input for next layer - inpL = cur; - } - - cur = inpL; - - cur = build_norm(cur, - model.output_norm, nullptr, - LLM_NORM_RMS, -1); - - cb(cur, "result_norm", -1); - res->t_embd = cur; - - // lm_head - cur = build_lora_mm(model.output, cur); - - cb(cur, "result_output", -1); - res->t_logits = cur; - - ggml_build_forward_expand(gf, cur); - } -}; - -llama_memory_i * llama_model::create_memory(const llama_memory_params & params, llama_cparams & cparams) const { - llama_memory_i * res; - - switch (arch) { - // Models that need specific instantiation should be handled in the - // switch statement - case LLM_ARCH_BERT: - case LLM_ARCH_JINA_BERT_V2: - case LLM_ARCH_JINA_BERT_V3: - case LLM_ARCH_NOMIC_BERT: - case LLM_ARCH_NOMIC_BERT_MOE: - case LLM_ARCH_NEO_BERT: - case LLM_ARCH_WAVTOKENIZER_DEC: - case LLM_ARCH_GEMMA_EMBEDDING: - case LLM_ARCH_DREAM: - case LLM_ARCH_LLADA: - case LLM_ARCH_LLADA_MOE: - { - res = nullptr; - } break; - // Models that need standard caching should rely on recurrent/hybrid - // checks - default: - { - if (llm_arch_is_recurrent(arch)) { - res = new llama_memory_recurrent( - *this, - GGML_TYPE_F32, - GGML_TYPE_F32, - cparams.offload_kqv, - std::max((uint32_t) 1, cparams.n_seq_max), - cparams.n_seq_max, - nullptr); - } else if (llm_arch_is_hybrid(arch)) { - - // The main difference between hybrid architectures is the - // layer filters, so pick the right one here - llama_memory_hybrid::layer_filter_cb filter_attn = nullptr; - llama_memory_hybrid::layer_filter_cb filter_recr = nullptr; - if (arch == LLM_ARCH_FALCON_H1) { - filter_attn = [&](int32_t) { return true; }; - filter_recr = [&](int32_t) { return true; }; - } else if (arch == LLM_ARCH_NEMOTRON_H) { - filter_attn = [&](int32_t il) { - return !hparams.is_recurrent(il) && hparams.n_ff(il) == 0; - }; - filter_recr = [&](int32_t il) { - return hparams.is_recurrent(il) && hparams.n_ff(il) == 0; - }; - } - - const auto padding = llama_kv_cache::get_padding(cparams); - - cparams.n_ctx = GGML_PAD(cparams.n_ctx, padding); - - res = new llama_memory_hybrid( - /* model */ *this, - /* attn_type_k */ params.type_k, - /* attn_type_v */ params.type_v, - /* attn_v_trans */ !cparams.flash_attn, - /* attn_kv_size */ cparams.n_ctx, - /* attn_n_pad */ padding, - /* attn_n_swa */ hparams.n_swa, - /* attn_swa_type */ hparams.swa_type, - /* recurrent_type_k */ GGML_TYPE_F32, - /* recurrent_type_v */ GGML_TYPE_F32, - /* recurrent_kv_size */ std::max((uint32_t) 1, cparams.n_seq_max), - /* n_seq_max */ cparams.n_seq_max, - /* offload */ cparams.offload_kqv, - /* unified */ cparams.kv_unified, - /* filter_attn */ std::move(filter_attn), - /* filter_recr */ std::move(filter_recr)); - } else { - const auto padding = llama_kv_cache::get_padding(cparams); - - uint32_t n_ctx_per_stream = cparams.n_ctx; - - if (!cparams.kv_unified) { - n_ctx_per_stream = (cparams.n_ctx + cparams.n_seq_max - 1)/cparams.n_seq_max; - n_ctx_per_stream = GGML_PAD(n_ctx_per_stream, padding); - - cparams.n_ctx = n_ctx_per_stream*cparams.n_seq_max; - } else { - n_ctx_per_stream = GGML_PAD(n_ctx_per_stream, padding); - - cparams.n_ctx = n_ctx_per_stream; - } - - LLAMA_LOG_DEBUG("%s: n_ctx = %u (padded)\n", __func__, cparams.n_ctx); - - llama_memory_i::layer_reuse_cb reuse = nullptr; - - if (arch == LLM_ARCH_GEMMA3N) { - reuse = [&](int32_t il) { - if (il >= (int32_t) hparams.n_layer_kv_from_start) { - return (int32_t) hparams.n_layer_kv_from_start - (hparams.is_swa(il) ? 2 : 1); - } - - return -1; - }; - } - - if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) { - GGML_ASSERT(hparams.is_swa_any()); - - res = new llama_kv_cache_iswa( - *this, - params.type_k, - params.type_v, - !cparams.flash_attn, - cparams.offload_kqv, - params.swa_full, - cparams.kv_unified, - n_ctx_per_stream, - cparams.n_seq_max, - cparams.n_ubatch, - padding, - nullptr, - reuse); - } else { - GGML_ASSERT(!hparams.is_swa_any()); - - res = new llama_kv_cache( - *this, - params.type_k, - params.type_v, - !cparams.flash_attn, - cparams.offload_kqv, - cparams.kv_unified, - n_ctx_per_stream, - cparams.n_seq_max, - padding, - hparams.n_swa, - hparams.swa_type, - nullptr, - nullptr); - } - } - } - } - - return res; -} + return res; +} ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { std::unique_ptr llm; @@ -19866,6 +6992,14 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_QWEN3VL: + { + llm = std::make_unique(*this, params); + } break; + case LLM_ARCH_QWEN3VLMOE: + { + llm = std::make_unique(*this, params); + } break; case LLM_ARCH_PHI2: { llm = std::make_unique(*this, params); @@ -20158,6 +7292,18 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_MINIMAX_M2: + { + llm = std::make_unique(*this, params); + } break; + case LLM_ARCH_COGVLM: + { + llm = std::make_unique(*this, params); + } break; + case LLM_ARCH_PANGU_EMBED: + { + llm = std::make_unique(*this, params); + }break; default: GGML_ABORT("fatal error"); } @@ -20221,6 +7367,10 @@ int32_t llama_model_n_embd(const llama_model * model) { return model->hparams.n_embd; } +int32_t llama_model_n_embd_inp(const llama_model * model) { + return model->hparams.n_embd_inp(); +} + int32_t llama_model_n_layer(const llama_model * model) { return model->hparams.n_layer; } @@ -20375,10 +7525,16 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_SEED_OSS: case LLM_ARCH_GROVEMOE: case LLM_ARCH_APERTUS: + case LLM_ARCH_MINIMAX_M2: + case LLM_ARCH_COGVLM: + case LLM_ARCH_PANGU_EMBED: return LLAMA_ROPE_TYPE_NEOX; case LLM_ARCH_QWEN2VL: return LLAMA_ROPE_TYPE_MROPE; + case LLM_ARCH_QWEN3VL: + case LLM_ARCH_QWEN3VLMOE: + return LLAMA_ROPE_TYPE_IMROPE; // all model arches should be listed explicitly here case LLM_ARCH_UNKNOWN: diff --git a/examples/talk-llama/llama-model.h b/examples/talk-llama/llama-model.h index 248f8541..71ff148e 100644 --- a/examples/talk-llama/llama-model.h +++ b/examples/talk-llama/llama-model.h @@ -114,6 +114,7 @@ enum llm_type { LLM_TYPE_30B_A3B, LLM_TYPE_100B_A6B, LLM_TYPE_106B_A12B, // GLM-4.5-Air + LLM_TYPE_230B_A10B, // Minimax M2 LLM_TYPE_235B_A22B, LLM_TYPE_300B_A47B, // Ernie MoE big LLM_TYPE_355B_A32B, // GLM-4.5 @@ -384,6 +385,13 @@ struct llama_layer { // openai-moe struct ggml_tensor * attn_sinks = nullptr; + // cogvlm + struct ggml_tensor * visexp_attn_wqkv = nullptr; + struct ggml_tensor * visexp_attn_wo = nullptr; + struct ggml_tensor * visexp_ffn_gate = nullptr; + struct ggml_tensor * visexp_ffn_down = nullptr; + struct ggml_tensor * visexp_ffn_up = nullptr; + // xIELU activation parameters for Apertus struct ggml_tensor * ffn_act_alpha_n = nullptr; struct ggml_tensor * ffn_act_alpha_p = nullptr; @@ -500,9 +508,8 @@ struct llama_model { ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const; - // note: can mutate `cparams` // TODO: move this to new llm_arch_model_i interface - llama_memory_i * create_memory(const llama_memory_params & params, llama_cparams & cparams) const; + llama_memory_i * create_memory(const llama_memory_params & params, const llama_cparams & cparams) const; // TODO: move this to new llm_arch_model_i interface ggml_cgraph * build_graph(const llm_graph_params & params) const; diff --git a/examples/talk-llama/llama-quant.cpp b/examples/talk-llama/llama-quant.cpp index 6dd40412..a56b2626 100644 --- a/examples/talk-llama/llama-quant.cpp +++ b/examples/talk-llama/llama-quant.cpp @@ -653,7 +653,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std:: gguf_set_val_f32(ctx_out.get(), o.key, o.val_f64); } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_INT) { // Setting type to UINT32. See https://github.com/ggml-org/llama.cpp/pull/14182 for context - gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)abs(o.val_i64)); + gguf_set_val_u32(ctx_out.get(), o.key, (uint32_t)std::abs(o.val_i64)); } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_BOOL) { gguf_set_val_bool(ctx_out.get(), o.key, o.val_bool); } else if (o.tag == LLAMA_KV_OVERRIDE_TYPE_STR) { diff --git a/examples/talk-llama/llama-vocab.cpp b/examples/talk-llama/llama-vocab.cpp index 639fecbd..735c5d54 100644 --- a/examples/talk-llama/llama-vocab.cpp +++ b/examples/talk-llama/llama-vocab.cpp @@ -401,6 +401,7 @@ struct llm_tokenizer_bpe : llm_tokenizer { }; break; case LLAMA_VOCAB_PRE_TYPE_GPT4O: + case LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2: regex_exprs = { // original regex from tokenizer.json // "[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]*[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]+(?i:'s|'t|'re|'ve|'m|'ll|'d)?|[^\\r\\n\\p{L}\\p{N}]?[\\p{Lu}\\p{Lt}\\p{Lm}\\p{Lo}\\p{M}]+[\\p{Ll}\\p{Lm}\\p{Lo}\\p{M}]*(?i:'s|'t|'re|'ve|'m|'ll|'d)?|\\p{N}{1,3}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n/]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", @@ -1992,6 +1993,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "grok-2") { pre_type = LLAMA_VOCAB_PRE_TYPE_GROK_2; clean_spaces = false; + } else if ( + tokenizer_pre == "minimax-m2") { + pre_type = LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2; + clean_spaces = false; } else { throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str())); } diff --git a/examples/talk-llama/llama-vocab.h b/examples/talk-llama/llama-vocab.h index 5e468675..1194ec47 100644 --- a/examples/talk-llama/llama-vocab.h +++ b/examples/talk-llama/llama-vocab.h @@ -49,6 +49,7 @@ enum llama_vocab_pre_type { LLAMA_VOCAB_PRE_TYPE_HUNYUAN_DENSE = 38, LLAMA_VOCAB_PRE_TYPE_GROK_2 = 39, LLAMA_VOCAB_PRE_TYPE_GRANITE_DOCLING = 40, + LLAMA_VOCAB_PRE_TYPE_MINIMAX_M2 = 41, }; struct LLM_KV; diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h index a0a660bf..8547226f 100644 --- a/examples/talk-llama/llama.h +++ b/examples/talk-llama/llama.h @@ -83,6 +83,7 @@ extern "C" { LLAMA_ROPE_TYPE_NORM = 0, LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX, LLAMA_ROPE_TYPE_MROPE = GGML_ROPE_TYPE_MROPE, + LLAMA_ROPE_TYPE_IMROPE = GGML_ROPE_TYPE_IMROPE, LLAMA_ROPE_TYPE_VISION = GGML_ROPE_TYPE_VISION, }; @@ -460,7 +461,11 @@ extern "C" { LLAMA_API bool llama_supports_gpu_offload(void); LLAMA_API bool llama_supports_rpc (void); + // NOTE: After creating a llama_context, it is recommended to query the actual values using these functions + // In some cases the requested values via llama_context_params may differ from the actual values used by the context + // ref: https://github.com/ggml-org/llama.cpp/pull/17046#discussion_r2503085732 LLAMA_API uint32_t llama_n_ctx (const struct llama_context * ctx); + LLAMA_API uint32_t llama_n_ctx_seq (const struct llama_context * ctx); LLAMA_API uint32_t llama_n_batch (const struct llama_context * ctx); LLAMA_API uint32_t llama_n_ubatch (const struct llama_context * ctx); LLAMA_API uint32_t llama_n_seq_max (const struct llama_context * ctx); @@ -481,6 +486,7 @@ extern "C" { LLAMA_API int32_t llama_model_n_ctx_train(const struct llama_model * model); LLAMA_API int32_t llama_model_n_embd (const struct llama_model * model); + LLAMA_API int32_t llama_model_n_embd_inp (const struct llama_model * model); LLAMA_API int32_t llama_model_n_layer (const struct llama_model * model); LLAMA_API int32_t llama_model_n_head (const struct llama_model * model); LLAMA_API int32_t llama_model_n_head_kv (const struct llama_model * model); @@ -584,7 +590,7 @@ extern "C" { LLAMA_API int32_t llama_adapter_meta_val_str_by_index(const struct llama_adapter_lora * adapter, int32_t i, char * buf, size_t buf_size); // Manually free a LoRA adapter - // Note: loaded adapters will be free when the associated model is deleted + // NOTE: loaded adapters will be free when the associated model is deleted LLAMA_API void llama_adapter_lora_free(struct llama_adapter_lora * adapter); // Get the invocation tokens if the current lora is an alora @@ -1110,8 +1116,6 @@ extern "C" { // // sample from the logits of the last token in the batch // const llama_token id = llama_sampler_sample(smpl, ctx, -1); // - // // accepting the token updates the internal state of certain samplers (e.g. grammar, repetition, etc.) - // llama_sampler_accept(smpl, id); // ... // } // diff --git a/examples/talk-llama/models/apertus.cpp b/examples/talk-llama/models/apertus.cpp new file mode 100644 index 00000000..9af19c1b --- /dev/null +++ b/examples/talk-llama/models/apertus.cpp @@ -0,0 +1,125 @@ +#include "models.h" + + + +llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * inp_pos = build_inp_pos(); + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = + hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur_pos", il); + cb(Kcur, "Kcur_pos", il); + cb(Vcur, "Vcur_pos", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network with xIELU activation + { + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, nullptr, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // Up projection + ggml_tensor * up = build_lora_mm(model.layers[il].ffn_up, cur); + cb(up, "ffn_up", il); + + float alpha_n_val = hparams.xielu_alpha_n[il]; + float alpha_p_val = hparams.xielu_alpha_p[il]; + float beta_val = hparams.xielu_beta[il]; + float eps_val = hparams.xielu_eps[il]; + + // Apply xIELU activation + ggml_tensor * activated = ggml_xielu(ctx0, up, alpha_n_val, alpha_p_val, beta_val, eps_val); + cb(activated, "ffn_xielu", il); + + // Down projection + cur = build_lora_mm(model.layers[il].ffn_down, activated); + cb(cur, "ffn_down", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, nullptr, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/arcee.cpp b/examples/talk-llama/models/arcee.cpp new file mode 100644 index 00000000..aa6167db --- /dev/null +++ b/examples/talk-llama/models/arcee.cpp @@ -0,0 +1,135 @@ +#include "models.h" + + +llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + // ARCEE uses relu^2 instead of silu + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/arctic.cpp b/examples/talk-llama/models/arctic.cpp new file mode 100644 index 00000000..e8f028a7 --- /dev/null +++ b/examples/talk-llama/models/arctic.cpp @@ -0,0 +1,138 @@ +#include "models.h" + + +llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + ggml_tensor * ffn_out = ggml_add(ctx0, cur, ffn_inp); + cb(ffn_out, "ffn_out", il); + + // MoE + cur = build_norm(inpSA, + model.layers[il].ffn_norm_exps, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm_exps", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_out); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/arwkv7.cpp b/examples/talk-llama/models/arwkv7.cpp new file mode 100644 index 00000000..107a3bef --- /dev/null +++ b/examples/talk-llama/models/arwkv7.cpp @@ -0,0 +1,86 @@ +#include "models.h" + + +llm_build_arwkv7::llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) { + GGML_ASSERT(n_embd == hparams.n_embd_r()); + + ggml_tensor * cur; + ggml_tensor * inpL; + ggml_tensor * v_first = nullptr; + + inpL = build_inp_embd(model.tok_embd); + + auto * rs_inp = build_rs_inp(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + + ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, + token_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il); + + token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + } + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/baichuan.cpp b/examples/talk-llama/models/baichuan.cpp new file mode 100644 index 00000000..c04b0c98 --- /dev/null +++ b/examples/talk-llama/models/baichuan.cpp @@ -0,0 +1,122 @@ +#include "models.h" + + +llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr; + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + switch (model.type) { + case LLM_TYPE_7B: + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + break; + case LLM_TYPE_13B: + break; + default: + GGML_ABORT("fatal error"); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/bailingmoe.cpp b/examples/talk-llama/models/bailingmoe.cpp new file mode 100644 index 00000000..ed56b9c4 --- /dev/null +++ b/examples/talk-llama/models/bailingmoe.cpp @@ -0,0 +1,144 @@ +#include "models.h" + + +llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + false, hparams.expert_weights_scale, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/bailingmoe2.cpp b/examples/talk-llama/models/bailingmoe2.cpp new file mode 100644 index 00000000..fbf7b210 --- /dev/null +++ b/examples/talk-llama/models/bailingmoe2.cpp @@ -0,0 +1,135 @@ +#include "models.h" + + + +llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + const int n_transformer_layers = n_layer - hparams.nextn_predict_layers; + for (int il = 0; il < n_transformer_layers; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 0 * sizeof(float) * (n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 1 * sizeof(float) * (n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + + if (il == n_transformer_layers - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpSA); + cb(sa_out, "sa_out", il); + + // MoE branch + cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if (static_cast(il) < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + true, hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(moe_out, "ffn_moe_out", il); + + { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/bert.cpp b/examples/talk-llama/models/bert.cpp new file mode 100644 index 00000000..3274fa3b --- /dev/null +++ b/examples/talk-llama/models/bert.cpp @@ -0,0 +1,176 @@ +#include "models.h" + + + +llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + ggml_tensor * inp_pos = nullptr; + + if (model.arch != LLM_ARCH_JINA_BERT_V2) { + inp_pos = build_inp_pos(); + } + + // construct input embeddings (token, type, position) + inpL = build_inp_embd(model.tok_embd); + + // token types are hardcoded to zero ("Sentence A") + if (model.type_embd) { + ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0); + inpL = ggml_add(ctx0, inpL, type_row0); + } + if (model.arch == LLM_ARCH_BERT) { + inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL); + } + cb(inpL, "inp_embd", -1); + + // embed layer norm + inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); + cb(inpL, "inp_norm", -1); + + auto * inp_attn = build_attn_inp_no_cache(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * cur = inpL; + + { + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; + + // self-attention + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1], + 0 * sizeof(float) * (n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 1 * sizeof(float) * (n_embd)); + Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + } + + if (model.layers[il].attn_q_norm) { + Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + } + + if (model.layers[il].attn_k_norm) { + Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il); + + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + } + + // RoPE + if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE || + model.arch == LLM_ARCH_JINA_BERT_V3) { + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // re-add the layer input + cur = ggml_add(ctx0, cur, inpL); + + // attention layer norm + cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il); + + if (model.layers[il].attn_norm_2 != nullptr) { + cur = ggml_add(ctx0, cur, inpL); // re-add the layer input + cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il); + } + + ggml_tensor * ffn_inp = cur; + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) { + // MoE branch + cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, nullptr, + model.layers[il].ffn_down_exps, nullptr, hparams.n_expert, hparams.n_expert_used, + LLM_FFN_GELU, false, false, 0.0f, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il); + cb(cur, "ffn_moe_out", il); + } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE || + model.arch == LLM_ARCH_JINA_BERT_V3) { + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } else if (model.arch == LLM_ARCH_JINA_BERT_V2) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL, + model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_GEGLU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + // attentions bypass the intermediate layer + cur = ggml_add(ctx0, cur, ffn_inp); + + // output layer norm + cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cb(cur, "result_embd", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/bitnet.cpp b/examples/talk-llama/models/bitnet.cpp new file mode 100644 index 00000000..331a3f11 --- /dev/null +++ b/examples/talk-llama/models/bitnet.cpp @@ -0,0 +1,160 @@ +#include "models.h" + + +llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].wq_scale) { + Qcur = ggml_mul(ctx0, Qcur, model.layers[il].wq_scale); + } + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + // B1.K + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].wk_scale) { + Kcur = ggml_mul(ctx0, Kcur, model.layers[il].wk_scale); + } + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + // B1.V + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].wv_scale) { + Vcur = ggml_mul(ctx0, Vcur, model.layers[il].wv_scale); + } + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + NULL, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + + cur = build_norm(cur, + model.layers[il].attn_sub_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_sub_norm", il); + + cur = build_lora_mm(model.layers[il].wo, cur); + if (model.layers[il].wo_scale) { + cur = ggml_mul(ctx0, cur, model.layers[il].wo_scale); + } + if (model.layers[il].bo) { + cur = ggml_add(ctx0, cur, model.layers[il].bo); + } + cb(cur, "attn_out", il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward forward + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, model.layers[il].ffn_up_scale, + model.layers[il].ffn_gate, NULL, model.layers[il].ffn_gate_scale, + NULL, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_sub_out", il); + + cur = build_norm(cur, + model.layers[il].ffn_sub_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_sub_norm", il); + + cur = build_lora_mm(model.layers[il].ffn_down, cur); + if (model.layers[il].ffn_down_scale) { + cur = ggml_mul(ctx0, cur, model.layers[il].ffn_down_scale); + } + cb(cur, "ffn_down", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + // FIXME: do not use model.tok_embd directly, duplicate as model.output + cur = build_lora_mm(model.tok_embd, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/bloom.cpp b/examples/talk-llama/models/bloom.cpp new file mode 100644 index 00000000..2c552d1d --- /dev/null +++ b/examples/talk-llama/models/bloom.cpp @@ -0,0 +1,101 @@ +#include "models.h" + +llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv(); + + inpL = build_norm(inpL, + model.tok_norm, + model.tok_norm_b, + LLM_NORM, -1); + cb(inpL, "inp_norm", -1); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/chameleon.cpp b/examples/talk-llama/models/chameleon.cpp new file mode 100644 index 00000000..184511ae --- /dev/null +++ b/examples/talk-llama/models/chameleon.cpp @@ -0,0 +1,178 @@ +#include "models.h" + +#include + +llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + if (hparams.swin_norm) { + cur = inpL; + } else { + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + if (model.layers[il].attn_q_norm) { + Qcur = ggml_view_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur) * n_embd_head, + ggml_element_size(Qcur) * n_embd_head * n_head, + 0); + cb(Qcur, "Qcur", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + model.layers[il].attn_q_norm_b, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + } + + if (model.layers[il].attn_k_norm) { + Kcur = ggml_view_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens, + ggml_element_size(Kcur) * n_embd_head, + ggml_element_size(Kcur) * n_embd_head * n_head_kv, + 0); + cb(Kcur, "Kcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + model.layers[il].attn_k_norm_b, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + if (!hparams.swin_norm) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + } + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + if (hparams.swin_norm) { + cur = build_norm(cur, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output_with_img_logits", -1); + + // TODO: this suppresses the output of image tokens, which is required to enable text-only outputs. + // Needs to be removed once image outputs are supported. + int img_token_end_idx = 8196; + int img_token_start_idx = 4; + int num_img_tokens = img_token_end_idx - img_token_start_idx; + // creates 1d tensor of size num_img_tokens and values -FLT_MAX, + // which ensures that text token values are always at least larger than image token values + ggml_tensor * img_logits = ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, num_img_tokens); + img_logits = ggml_clamp(ctx0, img_logits, -FLT_MAX, -FLT_MAX); + cb(img_logits, "img_logits", -1); + + cur = ggml_set_1d(ctx0, cur, img_logits, ggml_element_size(cur) * img_token_start_idx); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/chatglm.cpp b/examples/talk-llama/models/chatglm.cpp new file mode 100644 index 00000000..2685d4fb --- /dev/null +++ b/examples/talk-llama/models/chatglm.cpp @@ -0,0 +1,132 @@ +#include "models.h" + + +llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv == nullptr) { + Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + } + Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + } + Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + } else { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + } + + //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + } + + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + + cur = build_norm(inpL, + model.output_norm, + NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/codeshell.cpp b/examples/talk-llama/models/codeshell.cpp new file mode 100644 index 00000000..0b3bdbff --- /dev/null +++ b/examples/talk-llama/models/codeshell.cpp @@ -0,0 +1,111 @@ +#include "models.h" + +llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/cogvlm.cpp b/examples/talk-llama/models/cogvlm.cpp new file mode 100644 index 00000000..edf0d142 --- /dev/null +++ b/examples/talk-llama/models/cogvlm.cpp @@ -0,0 +1,100 @@ +#include "models.h" + +llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor *inpL, *cur; + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + // check ubatch to see if we have input tokens (text) + // or an input embedding vector (image) + bool is_text; + if (ubatch.token) { + is_text = true; + } else { + is_text = false; + } + + for (int il = 0; il < n_layer; ++il) { + // get either the text or image weight tensors + ggml_tensor *wqkv, *wo; + ggml_tensor *ffn_gate, *ffn_down, *ffn_up; + + if (is_text) { + wqkv = model.layers[il].wqkv; + wo = model.layers[il].wo; + ffn_gate = model.layers[il].ffn_gate; + ffn_down = model.layers[il].ffn_down; + ffn_up = model.layers[il].ffn_up; + } else { + wqkv = model.layers[il].visexp_attn_wqkv; + wo = model.layers[il].visexp_attn_wo; + ffn_gate = model.layers[il].visexp_ffn_gate; + ffn_down = model.layers[il].visexp_ffn_down; + ffn_up = model.layers[il].visexp_ffn_up; + } + + ggml_tensor * inpSA = inpL; + cur = build_norm(inpSA, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + + // build self attention + { + ggml_tensor * qkv = build_lora_mm(wqkv, cur); + + // split qkv into Q, K, V along the first dimension + ggml_tensor * Qcur = + ggml_view_3d(ctx0, qkv, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), qkv->nb[1], 0); + ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + qkv->nb[1], n_embd * ggml_element_size(qkv)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + qkv->nb[1], 2 * n_embd * ggml_element_size(qkv)); + + Qcur = ggml_rope(ctx0, Qcur, inp_pos, n_embd_head, rope_type); + Kcur = ggml_rope(ctx0, Kcur, inp_pos, n_embd_head, rope_type); + + cur = build_attn(inp_attn, + wo, nullptr, + Qcur, Kcur, Vcur, + nullptr, nullptr, nullptr, + kq_scale, il); + cb(cur, "attn_out", il); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + ffn_up, NULL, NULL, + ffn_gate, NULL, NULL, + ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output", -1); + res->t_logits = cur; + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/cohere2-iswa.cpp b/examples/talk-llama/models/cohere2-iswa.cpp new file mode 100644 index 00000000..b18aa8c4 --- /dev/null +++ b/examples/talk-llama/models/cohere2-iswa.cpp @@ -0,0 +1,131 @@ +#include "models.h" + +llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + const float f_logit_scale = hparams.f_logit_scale; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_iswa(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const bool is_swa = hparams.is_swa(il); + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); + cb(cur, "attn_norm", il); + ggml_tensor * ffn_inp = cur; + + // self-attention + { + // rope freq factors for 128k context + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (is_swa) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + } + + ggml_tensor * attn_out = cur; + + // feed-forward network + { + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + // add together residual + FFN + self-attention + cur = ggml_add(ctx0, cur, inpL); + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + if (f_logit_scale) { + cur = ggml_scale(ctx0, cur, f_logit_scale); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/command-r.cpp b/examples/talk-llama/models/command-r.cpp new file mode 100644 index 00000000..4d3b643b --- /dev/null +++ b/examples/talk-llama/models/command-r.cpp @@ -0,0 +1,122 @@ +#include "models.h" + + + +llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + const float f_logit_scale = hparams.f_logit_scale; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); + cb(cur, "attn_norm", il); + + ggml_tensor * ffn_inp = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM, il); + cb(Qcur, "Qcur", il); + } + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + } + ggml_tensor * attn_out = cur; + + // feed-forward network + { + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + // add together residual + FFN + self-attention + cur = ggml_add(ctx0, cur, inpL); + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + if (f_logit_scale) { + cur = ggml_scale(ctx0, cur, f_logit_scale); + } + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/dbrx.cpp b/examples/talk-llama/models/dbrx.cpp new file mode 100644 index 00000000..6d2a0ebf --- /dev/null +++ b/examples/talk-llama/models/dbrx.cpp @@ -0,0 +1,123 @@ +#include "models.h" + + +llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", il); + + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].attn_out_norm, NULL, + LLM_NORM, il); + cb(cur, "attn_out_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/deci.cpp b/examples/talk-llama/models/deci.cpp new file mode 100644 index 00000000..7410a3a4 --- /dev/null +++ b/examples/talk-llama/models/deci.cpp @@ -0,0 +1,135 @@ +#include "models.h" + + + +llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = + hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_head = hparams.n_head(il); + const int64_t n_ff = hparams.n_ff(il); + + if (n_head == 0) { + // attention-free layer of Llama-3_1-Nemotron-51B + cur = inpL; + } else { + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + if (n_head > 0 && n_head_kv == 0) { + // "linear attention" of Llama-3_1-Nemotron-51B + cur = build_lora_mm(model.layers[il].wo, cur); + cb(cur, "wo", il); + } else if (n_head > 0) { + // self-attention + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + // FFN-free layer of Llama-3_1-Nemotron-Ultra-253B + if (n_ff == 0) { + continue; + } + // modified to support attention-free layer of Llama-3_1-Nemotron-51B + ggml_tensor * ffn_inp = cur; + if (n_head > 0) { + ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + } + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/deepseek.cpp b/examples/talk-llama/models/deepseek.cpp new file mode 100644 index 00000000..17866c0d --- /dev/null +++ b/examples/talk-llama/models/deepseek.cpp @@ -0,0 +1,144 @@ +#include "models.h" + + + +llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = + hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, hparams.expert_weights_scale, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/deepseek2.cpp b/examples/talk-llama/models/deepseek2.cpp new file mode 100644 index 00000000..68f72f72 --- /dev/null +++ b/examples/talk-llama/models/deepseek2.cpp @@ -0,0 +1,236 @@ +#include "models.h" + + + +llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + bool is_lite = (hparams.n_layer == 27); + + const bool is_mla = (hparams.n_embd_head_k_mla != 0 && hparams.n_embd_head_v_mla != 0); + + // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA + const int64_t n_embd_head_k = is_mla ? hparams.n_embd_head_k_mla : hparams.n_embd_head_k; + const int64_t n_embd_head_v = is_mla ? hparams.n_embd_head_v_mla : hparams.n_embd_head_v; + + const int64_t n_embd_head_qk_rope = hparams.n_rot; + const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope; + + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly. + // See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation. + const float mscale = attn_factor * (1.0f + hparams.rope_yarn_log_mul * logf(1.0f / freq_scale)); + const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k)); + const float attn_factor = 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale)); + + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + if (!is_lite) { + q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = build_norm(q, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il); + cb(q, "q", il); + + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + } else { + q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(q, "q", il); + } + // split into {n_embd_head_qk_nope, n_head, n_tokens} + ggml_tensor * q_nope = + ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k), + ggml_row_size(q->type, n_embd_head_k) * n_head, 0); + cb(q_nope, "q_nope", il); + + // and {n_embd_head_qk_rope, n_head, n_tokens} + ggml_tensor * q_pe = ggml_view_3d( + ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k), + ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_cmpr_pe, "kv_cmpr_pe", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_cmpr = + ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens, + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0); + cb(kv_cmpr, "kv_cmpr", il); + + // and {n_embd_head_qk_rope, 1, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens, + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), + ggml_row_size(kv_cmpr_pe->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(q_pe, "q_pe", il); + + k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(k_pe, "k_pe", il); + + kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il); + cb(kv_cmpr, "kv_cmpr", il); + + if (is_mla) { + // {n_embd_head_qk_nope, n_tokens, n_head} + q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3); + cb(q_nope, "q_nope_perm", il); + + // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head} + ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope); + cb(q_nope_absorbed, "q_nope_absorbed", il); + + // {kv_lora_rank, n_head, n_tokens} + q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3); + cb(q_nope_absorbed, "q_nope_absorbed_perm", il); + + // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens} + // note: rope must go first for in-place context shifting in build_rope_shift() + ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope_absorbed, 0); + cb(Qcur, "Qcur", il); + + kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens); + cb(kv_cmpr, "kv_cmpr_reshape", il); + + // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens} + ggml_tensor * Kcur = ggml_concat(ctx0, k_pe, kv_cmpr, 0); + cb(Kcur, "Kcur", il); + + // {kv_lora_rank, 1, n_tokens} + ggml_tensor * Vcur = kv_cmpr; + cb(Vcur, "Vcur", il); + + // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group) + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il); + } else { + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr); + cb(kv, "kv", il); + + // split into {n_embd_head_qk_nope, n_head, n_tokens} + ggml_tensor * k_nope = + ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v), + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, 0); + cb(k_nope, "k_nope_view", il); + + // and {n_embd_head_v, n_head, n_tokens} + ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v), + ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, + ggml_row_size(kv->type, n_embd_head_qk_nope)); + cb(Vcur, "Vcur_view", il); + + Vcur = ggml_cont(ctx0, Vcur); + cb(Vcur, "Vcur_cont", il); + + // note: rope must go first for in-place context shifting in build_rope_shift() + ggml_tensor * Qcur = ggml_concat(ctx0, q_pe, q_nope, 0); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = ggml_concat(ctx0, ggml_repeat(ctx0, k_pe, q_pe), k_nope, 0); + cb(Kcur, "Kcur", il); + + // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups) + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + } + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + true, hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = ggml_mul_mat(ctx0, model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/dots1.cpp b/examples/talk-llama/models/dots1.cpp new file mode 100644 index 00000000..09c36f82 --- /dev/null +++ b/examples/talk-llama/models/dots1.cpp @@ -0,0 +1,134 @@ +#include "models.h" + + + +llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + if ((uint32_t) il < hparams.n_layer_dense_lead) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + true, hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(moe_out, "ffn_moe_out", il); + + { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/dream.cpp b/examples/talk-llama/models/dream.cpp new file mode 100644 index 00000000..2aafbae1 --- /dev/null +++ b/examples/talk-llama/models/dream.cpp @@ -0,0 +1,105 @@ +#include "models.h" + + + +llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + //copied from qwen2 + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_no_cache(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/ernie4-5-moe.cpp b/examples/talk-llama/models/ernie4-5-moe.cpp new file mode 100644 index 00000000..0d96d14e --- /dev/null +++ b/examples/talk-llama/models/ernie4-5-moe.cpp @@ -0,0 +1,150 @@ +#include "models.h" + + + +llm_build_ernie4_5_moe::llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + GGML_ASSERT(hparams.n_moe_layer_step > 0 && "Ernie 4.5 MoE requires n_moe_layer_step > 0"); + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + // norm + { + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + cb(cur, "attn_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + bool is_moe_layer = + static_cast(il) >= hparams.n_layer_dense_lead && (il + 1) % hparams.n_moe_layer_step == 0; + + if (!is_moe_layer) { + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // Shared expert (if present) + if (hparams.n_ff_shexp > 0) { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + } else { + cur = moe_out; + } + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/ernie4-5.cpp b/examples/talk-llama/models/ernie4-5.cpp new file mode 100644 index 00000000..99962af1 --- /dev/null +++ b/examples/talk-llama/models/ernie4-5.cpp @@ -0,0 +1,111 @@ +#include "models.h" + + + +llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + { + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + } + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/exaone.cpp b/examples/talk-llama/models/exaone.cpp new file mode 100644 index 00000000..62602b28 --- /dev/null +++ b/examples/talk-llama/models/exaone.cpp @@ -0,0 +1,114 @@ +#include "models.h" + + + +llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/exaone4.cpp b/examples/talk-llama/models/exaone4.cpp new file mode 100644 index 00000000..8b7e3dc0 --- /dev/null +++ b/examples/talk-llama/models/exaone4.cpp @@ -0,0 +1,123 @@ +#include "models.h" + + +template +llm_build_exaone4::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_k; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_v); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + using inp_attn_type = std::conditional_t; + inp_attn_type * inp_attn = nullptr; + + if constexpr (iswa) { + inp_attn = build_attn_inp_kv_iswa(); + } else { + inp_attn = build_attn_inp_kv(); + } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // use RoPE for SWA layers or non-SWA models + const bool use_rope = hparams.is_swa(il) || hparams.swa_type == LLAMA_SWA_TYPE_NONE; + + cur = inpL; + + // self-attention + { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + cb(Kcur, "Kcur_normed", il); + + if (use_rope) { + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, + freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, + freq_scale, ext_factor, attn_factor, beta_fast, beta_slow); + } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + cb(cur, "attn_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +// Explicit template instantiations +template struct llm_build_exaone4; +template struct llm_build_exaone4; diff --git a/examples/talk-llama/models/falcon-h1.cpp b/examples/talk-llama/models/falcon-h1.cpp new file mode 100644 index 00000000..b641a094 --- /dev/null +++ b/examples/talk-llama/models/falcon-h1.cpp @@ -0,0 +1,113 @@ +#include "models.h" + + + +llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // Build the inputs in the recurrent & kv cache + auto * inp = build_inp_mem_hybrid(); + + const float kq_scale = + hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, hparams.rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur-post-rope", il); + cb(Kcur, "Kcur-post-rope", il); + cb(Vcur, "Vcur-post-rope", il); + + ggml_tensor * attn_out = build_attn(inp->get_attn(), + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(attn_out, "attn_out", il); + + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + // Mamba2 layer + cb(cur, "ssm_in", il); + + ggml_tensor * ssm_out = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il); + cb(ssm_out, "ssm_out", il); + + // // Aggregation + cur = ggml_add(ctx0, attn_out, ssm_out); + inpSA = ggml_add(ctx0, cur, inpSA); + cb(cur, "layer_out", il); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = inpSA; + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, inpSA); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/falcon.cpp b/examples/talk-llama/models/falcon.cpp new file mode 100644 index 00000000..db1ccdb5 --- /dev/null +++ b/examples/talk-llama/models/falcon.cpp @@ -0,0 +1,120 @@ +#include "models.h" + + +llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * attn_norm; + + attn_norm = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(attn_norm, "attn_norm", il); + + // self-attention + { + if (model.layers[il].attn_norm_2) { + // Falcon-40B + cur = build_norm(inpL, + model.layers[il].attn_norm_2, + model.layers[il].attn_norm_2_b, + LLM_NORM, il); + cb(cur, "attn_norm_2", il); + } else { + cur = attn_norm; + } + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + + // using mode = 2 for neox mode + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + attn_norm = ggml_get_rows(ctx0, attn_norm, inp_out_ids); + } + + ggml_tensor * ffn_inp = cur; + + // feed forward + { + cur = build_ffn(attn_norm, // !! use the attn norm, not the result + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + // norm + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/gemma-embedding.cpp b/examples/talk-llama/models/gemma-embedding.cpp new file mode 100644 index 00000000..90a98f7a --- /dev/null +++ b/examples/talk-llama/models/gemma-embedding.cpp @@ -0,0 +1,120 @@ +#include "models.h" + + + +llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_k; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) + if (ubatch.token) { + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + } + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_no_cache(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const float freq_base_l = model.get_rope_freq_base(cparams, il); + const float freq_scale_l = model.get_rope_freq_scale(cparams, il); + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315 + Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale); + + cur = + build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + + cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/gemma.cpp b/examples/talk-llama/models/gemma.cpp new file mode 100644 index 00000000..4893d9af --- /dev/null +++ b/examples/talk-llama/models/gemma.cpp @@ -0,0 +1,112 @@ +#include "models.h" + + +llm_build_gemma::llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); + cb(Qcur, "Qcur_scaled", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/gemma2-iswa.cpp b/examples/talk-llama/models/gemma2-iswa.cpp new file mode 100644 index 00000000..9cc59a53 --- /dev/null +++ b/examples/talk-llama/models/gemma2-iswa.cpp @@ -0,0 +1,125 @@ +#include "models.h" + +llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_k; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_iswa(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // final logit soft-capping + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); + cur = ggml_tanh(ctx0, cur); + cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/gemma3-iswa.cpp b/examples/talk-llama/models/gemma3-iswa.cpp new file mode 100644 index 00000000..839ff6d3 --- /dev/null +++ b/examples/talk-llama/models/gemma3-iswa.cpp @@ -0,0 +1,131 @@ +#include "models.h" + +llm_build_gemma3_iswa::llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_k; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) + if (ubatch.token) { + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + } + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // TODO: is causal == true correct? might need some changes + auto * inp_attn = build_attn_inp_kv_iswa(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const float freq_base_l = model.get_rope_freq_base (cparams, il); + const float freq_scale_l = model.get_rope_freq_scale(cparams, il); + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/model.py#L315 + Qcur = ggml_scale(ctx0, Qcur, hparams.f_attention_scale); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); + cb(sa_out, "sa_out", il); + + cur = build_norm(sa_out, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, sa_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/gemma3n-iswa.cpp b/examples/talk-llama/models/gemma3n-iswa.cpp new file mode 100644 index 00000000..a0bdd6a1 --- /dev/null +++ b/examples/talk-llama/models/gemma3n-iswa.cpp @@ -0,0 +1,377 @@ +#include "models.h" + + + +llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params), + model(model), + n_embd_head(model.hparams.n_embd_head_k), + n_embd_altup(model.hparams.n_embd_altup), + n_altup(model.hparams.n_altup), + i_altup_act(model.hparams.i_altup_act) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // important: do not normalize weights for raw embeddings input (i.e. encoded image emdeddings) + if (ubatch.token) { + inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); + cb(inpL, "inp_scaled", -1); + } + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // TODO: is causal == true correct? might need some changes + auto * inp_attn = build_attn_inp_kv_iswa(); + + // inp_per_layer shape: [n_embd_altup, n_tokens, n_layer] + ggml_tensor * inp_per_layer = project_per_layer_inputs(inpL, get_per_layer_inputs()); + + // inpL now has only 1 altup, project it to the rest of the altups + // these "added" altups will be concat to the last dim of inpL + { + ggml_tensor * target_magnitude = calc_magnitude(inpL); + ggml_tensor * inp_repeated = ggml_repeat_4d(ctx0, inpL, n_embd, n_tokens, n_altup - 1, 1); + ggml_tensor * altup_added = + ggml_mul_mat(ctx0, model.altup_proj, inp_repeated); // shape: [n_embd, n_tokens, n_altup - 1] + ggml_tensor * new_magnitude = calc_magnitude(altup_added); + altup_added = ggml_div(ctx0, ggml_mul(ctx0, altup_added, target_magnitude), new_magnitude); + inpL = ggml_concat(ctx0, inpL, altup_added, 2); // shape: [n_embd, n_tokens, n_altup] + cb(inpL, "inp_stacked", -1); + } + // inpL now has shape: [n_embd, n_tokens, n_altup] + // inp_per_layer now has shape: [n_embd_altup, n_tokens, n_layer] + + for (int il = 0; il < n_layer; ++il) { + // this block is made to be closely resemble Gemma3p5DecoderLayer on python code + const float freq_base_l = model.get_rope_freq_base(cparams, il); + const float freq_scale_l = model.get_rope_freq_scale(cparams, il); + + ggml_tensor * cur = inpL; // [n_embd, n_tokens, n_altup] + ggml_tensor * predictions = altup_predict(cur, il); // [n_embd, n_tokens, n_altup] + + // predicted value will go through self-attention and laurel + ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act); // [n_embd, n_tokens] + cur = active_prediction; + cb(cur, "active_prediction", il); + + // norm + cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // laurel + ggml_tensor * laurel_out = laurel(cur, il); // [n_embd, n_tokens] + + // self-attention + if (hparams.has_kv(il)) { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + Vcur = ggml_rms_norm(ctx0, Vcur, hparams.f_norm_rms_eps); + + cb(Qcur, "Qcur_normed", il); + cb(Kcur, "Kcur_normed", il); + cb(Vcur, "Vcur_normed", il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur_pos", il); + cb(Kcur, "Kcur_pos", il); + + cur = build_attn(inp_attn, model.layers[il].wo, + NULL, Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, + hparams.f_attention_scale, il); + } else { + // reuse KV cache of earlier layers + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l, + ext_factor, attn_factor, beta_fast, beta_slow); + cb(Qcur, "Qcur_pos", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, nullptr, nullptr, nullptr, nullptr, nullptr, hparams.f_attention_scale, il); + } + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + cur = ggml_add(ctx0, cur, active_prediction); // [n_embd, n_tokens] + cb(cur, "attn_gated", il); + + ggml_tensor * attn_laurel = ggml_scale(ctx0, ggml_add(ctx0, cur, laurel_out), + 1.0f / sqrtf(2.0f)); // [n_embd, n_tokens] + cb(attn_laurel, "attn_laurel", il); + + cur = build_norm(attn_laurel, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + { + ggml_tensor * up_proj = build_lora_mm(model.layers[il].ffn_up, cur); + ggml_tensor * gate_proj = build_lora_mm(model.layers[il].ffn_gate, cur); + + if (il < n_layer_sparsity) { + // apply activation sparsity + gate_proj = gaussian_topk(gate_proj); + } + gate_proj = ggml_gelu(ctx0, gate_proj); + + cur = ggml_mul(ctx0, up_proj, gate_proj); + cur = build_lora_mm(model.layers[il].ffn_down, cur); + cb(cur, "ffn_out", il); + } + cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", il); + + ggml_tensor * attn_ffw_laurel_gated = ggml_add(ctx0, cur, attn_laurel); // [n_embd, n_tokens] + cb(attn_ffw_laurel_gated, "attn_ffw_laurel_gated", il); + + ggml_tensor * corrected = altup_correct(predictions, attn_ffw_laurel_gated, il); // [n_embd, n_tokens, n_altup] + + ggml_tensor * first_prediction; // [n_embd, n_tokens] + { + first_prediction = view_2d_slice(corrected, i_altup_act); // [n_embd, n_tokens] + first_prediction = ggml_mul(ctx0, first_prediction, model.layers[il].altup_correct_scale); + first_prediction = build_lora_mm(model.layers[il].per_layer_inp_gate, first_prediction); + first_prediction = ggml_gelu(ctx0, first_prediction); // [n_embd_altup, n_tokens] + cb(first_prediction, "first_prediction_gated", il); + ggml_tensor * inp_this_layer = view_2d_slice(inp_per_layer, il); // [n_embd_altup, n_tokens] + first_prediction = ggml_mul(ctx0, first_prediction, inp_this_layer); // [n_embd_altup, n_tokens] + cb(first_prediction, "first_prediction_scaled", il); + + first_prediction = build_lora_mm(model.layers[il].per_layer_proj, first_prediction); // [n_embd, n_tokens] + first_prediction = + build_norm(first_prediction, model.layers[il].per_layer_post_norm, NULL, LLM_NORM_RMS, il); + cb(first_prediction, "first_prediction_out", il); + } + // equivalent to python code: corrected_predictions[1:] += first_prediction + { + ggml_tensor * slice_first = view_2d_slice(corrected, 0); + ggml_tensor * slice_rest = ggml_view_3d( + ctx0, corrected, n_embd, n_tokens, n_altup - 1, ggml_row_size(corrected->type, n_embd), + ggml_row_size(corrected->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(corrected)); + ggml_tensor * tmp = ggml_add(ctx0, slice_rest, first_prediction); // [n_embd, n_tokens, n_altup - 1] + corrected = ggml_concat(ctx0, slice_first, tmp, 2); // [n_embd, n_tokens, n_altup] + } + cur = corrected; // [n_embd, n_tokens, n_altup] + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; // [n_embd, n_tokens, n_altup] + + // cur now has multiple altup(s), we want to merge them back to 1 altup + { + ggml_tensor * target_magnitude = calc_magnitude(view_2d_slice(cur, i_altup_act)); // [n_embd, n_tokens] + // do a view to skip the first slice (active altup) + ggml_tensor * alt_slice = + ggml_view_3d(ctx0, cur, n_embd, n_tokens, n_altup - 1, ggml_row_size(cur->type, n_embd), + ggml_row_size(cur->type, n_embd * n_tokens), n_embd * n_tokens * ggml_element_size(cur)); + ggml_tensor * altup_unembd = + ggml_mul_mat(ctx0, model.altup_unembd_proj, alt_slice); // shape: [n_embd, n_tokens, n_altup - 1] + ggml_tensor * new_magnitude = calc_magnitude(altup_unembd); + altup_unembd = ggml_div(ctx0, ggml_mul(ctx0, altup_unembd, target_magnitude), new_magnitude); + cb(altup_unembd, "altup_unembd", -1); + + // equivalent to torch.mean(hidden_states, dim=0) + cur = view_2d_slice(cur, 0); // [n_embd, n_tokens] + for (int i = 0; i < n_altup - 1; ++i) { + cur = ggml_add(ctx0, cur, view_2d_slice(altup_unembd, i)); + } + cur = ggml_scale(ctx0, cur, 1.0f / float(n_altup)); // [n_embd, n_tokens] + cb(cur, "unembd_merged", -1); + } + // cur now has shape: [n_embd, n_tokens] + + // TODO: move this to right after the last KV layer + { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + { + // final logit soft-capping + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); + cur = ggml_tanh(ctx0, cur); + cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); + } + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +ggml_tensor * llm_build_gemma3n_iswa::calc_magnitude(ggml_tensor * x) { + return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x))); +} + +// get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim +ggml_tensor * llm_build_gemma3n_iswa::view_2d_slice(ggml_tensor * x, int idx) { + GGML_ASSERT(idx < (int) x->ne[2]); + return ggml_view_2d(ctx0, x, x->ne[0], x->ne[1], ggml_row_size(x->type, x->ne[0]), + idx * x->ne[0] * x->ne[1] * ggml_element_size(x)); +} + +// equivalent to get_per_layer_inputs() in python code +// output shape: [n_embd_altup, n_layer, n_tokens] +ggml_tensor * llm_build_gemma3n_iswa::get_per_layer_inputs() { + auto inp = std::make_unique(); + ggml_tensor * inp_per_layer; + if (ubatch.token) { + inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens); + ggml_set_input(inp->tokens); + res->t_tokens = inp->tokens; + inp_per_layer = ggml_get_rows(ctx0, model.tok_embd_per_layer, inp->tokens); + inp_per_layer = ggml_reshape_3d(ctx0, inp_per_layer, n_embd_altup, n_layer, n_tokens); + inp_per_layer = ggml_scale(ctx0, inp_per_layer, sqrtf((float) n_embd_altup)); + cb(inp_per_layer, "inp_per_layer_selected", -1); + } else { + GGML_ABORT("TODO: support embd input"); + } + res->add_input(std::move(inp)); + return inp_per_layer; +} + +// equivalent to project_per_layer_inputs() in python code +// this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim +// output shape: [n_embd_altup, n_tokens, n_layer] +ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer) { + const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd); + const float per_layer_input_scale = 1.0f / sqrtf(2.0f); + + ggml_tensor * per_layer_proj = ggml_mul_mat(ctx0, model.per_layer_model_proj, inputs_embeds); + per_layer_proj = ggml_scale(ctx0, per_layer_proj, per_layer_projection_scale); + per_layer_proj = ggml_reshape_3d(ctx0, per_layer_proj, n_embd_altup, n_layer, n_tokens); + per_layer_proj = build_norm(per_layer_proj, model.per_layer_proj_norm, NULL, LLM_NORM_RMS, + -1); // [n_embd_altup, n_layer, n_tokens] + cb(per_layer_proj, "per_layer_proj", -1); + + inp_per_layer = ggml_add(ctx0, inp_per_layer, per_layer_proj); + inp_per_layer = ggml_scale(ctx0, inp_per_layer, per_layer_input_scale); + cb(inp_per_layer, "inp_per_layer", -1); + + // permute to shape: [n_embd_altup, n_tokens, n_layer] + inp_per_layer = ggml_cont(ctx0, ggml_permute(ctx0, inp_per_layer, 0, 2, 1, 3)); + return inp_per_layer; +} + +// input cur shape: [n_altup, n_tokens] +// output shape: [n_altup, n_tokens] +ggml_tensor * llm_build_gemma3n_iswa::laurel(ggml_tensor * cur, int il) { + ggml_tensor * tmp = cur; + tmp = build_lora_mm(model.layers[il].laurel_l, tmp); + tmp = build_lora_mm(model.layers[il].laurel_r, tmp); + tmp = build_norm(tmp, model.layers[il].laurel_post_norm, NULL, LLM_NORM_RMS, il); + tmp = ggml_add(ctx0, tmp, cur); + cb(tmp, "laurel_out", il); + return tmp; +} + +// input x shape: [n_embd, n_tokens] +// output shape: [n_embd, n_tokens] +ggml_tensor * llm_build_gemma3n_iswa::gaussian_topk(ggml_tensor * x) { + ggml_tensor * mean = ggml_mean(ctx0, x); + ggml_tensor * std = ggml_sqrt(ctx0, ggml_scale(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))), + 1.0f / (float) (x->ne[0] - 1))); + ggml_tensor * cutoff_x = ggml_add(ctx0, mean, ggml_scale(ctx0, std, f_sparsity_std_mul)); + return ggml_relu(ctx0, ggml_sub(ctx0, x, cutoff_x)); +} + +// +// altup functions +// + +// equivalent to compute_router_modalities() in python code +// input x shape: [n_embd, n_tokens] +// output shape: [n_altup, n_tokens] +ggml_tensor * llm_build_gemma3n_iswa::altup_compute_router_modalities(ggml_tensor * x, int il) { + ggml_tensor * router_inputs = build_norm(x, model.layers[il].altup_router_norm, NULL, LLM_NORM_RMS, il); + + // router_input_scale + router_inputs = ggml_scale(ctx0, router_inputs, 1.0f / (float) n_embd); + + ggml_tensor * output = ggml_mul_mat(ctx0, model.layers[il].altup_router, router_inputs); + return ggml_tanh(ctx0, output); // [n_altup, n_tokens] +} + +// input cur shape: [n_embd, n_tokens, n_altup] +// output shape: [n_embd, n_tokens, n_altup] +ggml_tensor * llm_build_gemma3n_iswa::altup_predict(ggml_tensor * cur, int il) { + ggml_tensor * activated = view_2d_slice(cur, i_altup_act); // [n_embd, n_tokens] + ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens] + cb(modalities, "modalities", il); + + ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_predict_coef, modalities); + cb(all_coefs, "all_coefs", il); + // first dim now having n_altup^2 elements, we reshape it to 2D (so we end up with 3D tensor) + all_coefs = ggml_reshape_3d(ctx0, all_coefs, n_altup, n_altup, n_tokens); + + // permute to [n_altup, n_embd, n_tokens] + ggml_tensor * cur_permuted = ggml_cont(ctx0, ggml_permute(ctx0, cur, 1, 2, 0, 3)); + ggml_tensor * predictions = ggml_mul_mat(ctx0, cur_permuted, all_coefs); // [n_altup, n_embd, n_tokens] + + // final shape must be the same as cur: [n_embd, n_tokens, n_altup] + predictions = ggml_cont(ctx0, ggml_permute(ctx0, predictions, 0, 2, 1, 3)); + predictions = ggml_add(ctx0, predictions, cur); + cb(predictions, "predictions", il); + + return predictions; +} + +// input predictions shape: [n_embd, n_tokens, n_altup] +// input activated shape: [n_embd, n_tokens] +// output shape: [n_embd, n_tokens, n_altup] +ggml_tensor * llm_build_gemma3n_iswa::altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) { + ggml_tensor * modalities = altup_compute_router_modalities(activated, il); // [n_altup, n_tokens] + cb(modalities, "modalities", il); + + ggml_tensor * active_prediction = view_2d_slice(predictions, i_altup_act); + ggml_tensor * innovation = ggml_sub(ctx0, activated, active_prediction); // [n_embd, n_tokens] + cb(innovation, "innovation", il); + + ggml_tensor * all_coefs = build_lora_mm(model.layers[il].altup_correct_coef, modalities); // [n_altup, n_tokens] + all_coefs = ggml_scale_bias(ctx0, all_coefs, 1.0f, 1.0f); // + 1.0 + cb(all_coefs, "all_coefs", il); + all_coefs = ggml_transpose(ctx0, all_coefs); // [n_tokens, n_altup] + all_coefs = ggml_cont_3d(ctx0, all_coefs, 1, n_tokens, n_altup); // [1, n_tokens, n_altup] + + innovation = ggml_repeat_4d(ctx0, innovation, n_embd, n_tokens, n_altup, 1); + ggml_tensor * corrected = ggml_mul(ctx0, innovation, all_coefs); // [n_embd, n_tokens, n_altup] + corrected = ggml_add(ctx0, corrected, predictions); // [n_embd, n_tokens, n_altup] + cb(corrected, "corrected", il); + + return corrected; +} diff --git a/examples/talk-llama/models/glm4-moe.cpp b/examples/talk-llama/models/glm4-moe.cpp new file mode 100644 index 00000000..33ee7070 --- /dev/null +++ b/examples/talk-llama/models/glm4-moe.cpp @@ -0,0 +1,153 @@ +#include "models.h" + +llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + // Only process up to last layer (skip final NextN layer) + // Final layer tensors are loaded but not processed in forward pass + const int n_transformer_layers = n_layer - hparams.nextn_predict_layers; + for (int il = 0; il < n_transformer_layers; ++il) { + ggml_tensor * inpSA = inpL; + + // Pre-attention norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + } + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + } + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + } + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // Apply Q/K norm if available (GLM-4.5 355B variant) + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + } + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + } + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_transformer_layers - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // Post-attention norm + cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "post_attn_norm", il); + + // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense) + if (static_cast(il) < hparams.n_layer_dense_lead) { + // Dense FFN layer + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // Process routed experts using existing MoE infrastructure + ggml_tensor * routed_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, hparams.expert_weights_norm, + true, hparams.expert_weights_scale, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(routed_out, "ffn_moe_out", il); + + // Process shared expert on original input + ggml_tensor * shared_out = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(shared_out, "ffn_shexp_out", il); + + // Final output: routed_output + shared_output + cur = ggml_add(ctx0, routed_out, shared_out); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/glm4.cpp b/examples/talk-llama/models/glm4.cpp new file mode 100644 index 00000000..f789b282 --- /dev/null +++ b/examples/talk-llama/models/glm4.cpp @@ -0,0 +1,127 @@ +#include "models.h" + + + +llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // Pre-attention norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv == nullptr) { + Qcur = build_lora_mm(model.layers[il].wq, cur); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + } + Kcur = build_lora_mm(model.layers[il].wk, cur); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + } + Vcur = build_lora_mm(model.layers[il].wv, cur); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + } else { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1], + 0 * sizeof(float) * (n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 1 * sizeof(float) * (n_embd)); + Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)); + } + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + // Post-attention norm (new!) + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "post_attn_norm", il); + + // Add the input (residual connection after post-attention norm) + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + // Pre-MLP norm + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // MLP + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + // Post-MLP norm + cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "post_mlp_norm", il); + } + // Add residual connection after post-MLP norm + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + // Final norm + cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // Output projection + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/gpt2.cpp b/examples/talk-llama/models/gpt2.cpp new file mode 100644 index 00000000..60761c8e --- /dev/null +++ b/examples/talk-llama/models/gpt2.cpp @@ -0,0 +1,105 @@ +#include "models.h" + +llm_build_gpt2::llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * pos; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/gptneox.cpp b/examples/talk-llama/models/gptneox.cpp new file mode 100644 index 00000000..2151b14e --- /dev/null +++ b/examples/talk-llama/models/gptneox.cpp @@ -0,0 +1,144 @@ +#include "models.h" + + +llm_build_gptneox::llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // ffn + if (hparams.use_par_res) { + // attention and ffn are computed in parallel + // x = x + attn(ln1(x)) + ffn(ln2(x)) + + ggml_tensor * attn_out = cur; + + cur = build_norm(inpL, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, inpL); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, attn_out); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } else { + // attention and ffn are computed sequentially + // x = x + attn(ln1(x)) + // x = x + ffn(ln2(x)) + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + } + + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/granite-hybrid.cpp b/examples/talk-llama/models/granite-hybrid.cpp new file mode 100644 index 00000000..f6ca4c17 --- /dev/null +++ b/examples/talk-llama/models/granite-hybrid.cpp @@ -0,0 +1,196 @@ +#include "models.h" + + +llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + // Positional embeddings populated if rope enabled + ggml_tensor * inp_pos = nullptr; + if (hparams.rope_finetuned) { + inp_pos = build_inp_pos(); + } + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + if (hparams.is_recurrent(il)) { + // ssm layer // + cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il); + } else { + // attention layer // + cur = build_attention_layer(cur, inp_pos, inp->get_attn(), model, n_embd_head, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // ffn + cur = build_layer_ffn(cur, inpSA, model, il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // For Granite architectures - scale logits + if (hparams.f_logit_scale) { + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + } + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +ggml_tensor * llm_build_granite_hybrid::build_attention_layer(ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv * inp_attn, + const llama_model & model, + const int64_t n_embd_head, + const int il) { + // compute Q and K and (optionally) RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + + const bool use_rope = hparams.rope_finetuned; + if (use_rope) { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + const float kq_scale = + hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + return cur; +} + +ggml_tensor * llm_build_granite_hybrid::build_layer_ffn(ggml_tensor * cur, + ggml_tensor * inpSA, + const llama_model & model, + const int il) { + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + } else { + // MoE branch + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // For Granite MoE Shared + if (hparams.n_ff_shexp > 0) { + ggml_tensor * ffn_shexp = + build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + } + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + return cur; +} diff --git a/examples/talk-llama/models/granite.cpp b/examples/talk-llama/models/granite.cpp new file mode 100644 index 00000000..18748e9c --- /dev/null +++ b/examples/talk-llama/models/granite.cpp @@ -0,0 +1,211 @@ +#include "models.h" + + +llm_build_granite::llm_build_granite( + const llama_model & model, + const llm_graph_params & params) + : llm_graph_context(params) { + + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - built only if rope enabled + ggml_tensor * inp_pos = nullptr; + if (hparams.rope_finetuned) { + inp_pos = build_inp_pos(); + } + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + cur = build_attention_layer( + cur, inp_pos, inp_attn, + model, n_embd_head, il); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + // ffn + cur = build_layer_ffn(cur, inpSA, model, il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + // For Granite architectures - scale logits + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +ggml_tensor * llm_build_granite::build_attention_layer( + ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv * inp_attn, + const llama_model & model, + const int64_t n_embd_head, + const int il) { + + // compute Q and K and (optionally) RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + + const bool use_rope = hparams.rope_finetuned; + if (use_rope) { + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + return cur; +} + +ggml_tensor * llm_build_granite::build_layer_ffn( + ggml_tensor * cur, + ggml_tensor * inpSA, + const llama_model & model, + const int il) { + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + } else { + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // For Granite MoE Shared + if (hparams.n_ff_shexp > 0) { + ggml_tensor * ffn_shexp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(ffn_shexp, "ffn_shexp", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + } + + // For Granite architectures - scale residual + if (hparams.f_residual_scale) { + cur = ggml_scale(ctx0, cur, hparams.f_residual_scale); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + return cur; +} diff --git a/examples/talk-llama/models/graph-context-mamba.cpp b/examples/talk-llama/models/graph-context-mamba.cpp new file mode 100644 index 00000000..b9a363b3 --- /dev/null +++ b/examples/talk-llama/models/graph-context-mamba.cpp @@ -0,0 +1,283 @@ +#include "models.h" + +llm_graph_context_mamba::llm_graph_context_mamba(const llm_graph_params & params) : llm_graph_context(params) {} + +ggml_tensor * llm_graph_context_mamba::build_mamba_layer(llm_graph_input_rs * inp, + ggml_tensor * cur, + const llama_model & model, + const llama_ubatch & ubatch, + int il) { + const auto * mctx_cur = inp->mctx; + + const auto kv_head = mctx_cur->get_head(); + + const auto & layer = model.layers[il]; + + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t dt_rank = hparams.ssm_dt_rank; + const int64_t n_head = d_inner; + const int64_t head_dim = 1; + const int64_t n_seqs = ubatch.n_seqs; + // Some variants of Mamba arch (e.g. FalconMamba do apply layer norm on B and Dt layers) + const bool ssm_dt_b_c_rms = hparams.ssm_dt_b_c_rms; + + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); + conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner, n_seqs); + + // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); + + // {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} + ggml_tensor * xz = build_lora_mm(layer.ssm_in, cur); + // split the above in two + // => {d_inner, n_seq_tokens, n_seqs} + ggml_tensor * x = ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], 0); + ggml_tensor * z = + ggml_view_3d(ctx0, xz, d_inner, xz->ne[1], xz->ne[2], xz->nb[1], xz->nb[2], d_inner * ggml_element_size(xz)); + + // conv + { + // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs} + ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0); + + // copy last (d_conv - 1) columns back into the state cache + ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], + n_seq_tokens * (conv_x->nb[0])); + + ggml_build_forward_expand( + gf, ggml_cpy(ctx0, last_conv, + ggml_view_1d(ctx0, conv_states_all, (d_conv - 1) * (d_inner) * (n_seqs), + kv_head * (d_conv - 1) * (d_inner) *ggml_element_size(conv_states_all)))); + + // 1D convolution + // The equivalent is to make a self-overlapping view of conv_x + // over d_conv columns at each stride in the 3rd dimension, + // then element-wise multiply that with the conv1d weight, + // then sum the elements of each row, + // (the last two steps are a dot product over rows (also doable with mul_mat)) + // then permute away the ne[0] dimension, + // and then you're left with the resulting x tensor. + // For simultaneous sequences, all sequences need to have the same length. + x = ggml_ssm_conv(ctx0, conv_x, layer.ssm_conv1d); + + // bias + x = ggml_add(ctx0, x, layer.ssm_conv1d_b); + + x = ggml_silu(ctx0, x); + } + + // ssm + { + // {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} + ggml_tensor * x_db = build_lora_mm(layer.ssm_x, x); + // split + ggml_tensor * dt = ggml_view_3d(ctx0, x_db, dt_rank, n_seq_tokens, n_seqs, x_db->nb[1], x_db->nb[2], 0); + ggml_tensor * B = + ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state * x_db->nb[0], x_db->nb[1], + x_db->nb[2], ggml_element_size(x_db) * dt_rank); + ggml_tensor * C = + ggml_view_4d(ctx0, x_db, d_state, /* n_group */ 1, n_seq_tokens, n_seqs, d_state * x_db->nb[0], x_db->nb[1], + x_db->nb[2], ggml_element_size(x_db) * (dt_rank + d_state)); + + // Some Mamba variants (e.g. FalconMamba, Jamba) apply RMS norm in B, C & Dt layers + if (ssm_dt_b_c_rms || (layer.ssm_dt_norm && layer.ssm_b_norm && layer.ssm_c_norm)) { + dt = build_norm(dt, layer.ssm_dt_norm, NULL, LLM_NORM_RMS, il); + B = build_norm(B, layer.ssm_b_norm, NULL, LLM_NORM_RMS, il); + C = build_norm(C, layer.ssm_c_norm, NULL, LLM_NORM_RMS, il); + } + + // {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} + dt = build_lora_mm(layer.ssm_dt, dt); + dt = ggml_add(ctx0, dt, layer.ssm_dt_b); + + cur = x; + x = ggml_reshape_4d(ctx0, x, head_dim, n_head, n_seq_tokens, n_seqs); + + ggml_tensor * A = layer.ssm_a; + + // use the states and the indices provided by build_recurrent_state + // (this is necessary in order to properly use the states before they are overwritten, + // while avoiding to make unnecessary copies of the states) + auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) { + ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size()); + + // Custom operator to optimize the parallel associative scan + // as described in the Annex D of the Mamba paper. + // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} + return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids); + }; + + ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows); + + // store last states + ggml_build_forward_expand( + gf, ggml_cpy(ctx0, ggml_view_1d(ctx0, y_ssm, d_state * d_inner * n_seqs, x->nb[3] * x->ne[3]), + ggml_view_1d(ctx0, ssm_states_all, d_state * d_inner * n_seqs, + kv_head * d_state * d_inner * ggml_element_size(ssm_states_all)))); + + ggml_tensor * y = ggml_view_3d(ctx0, y_ssm, d_inner, n_seq_tokens, n_seqs, x->nb[2], x->nb[3], 0); + + // TODO: skip computing output earlier for unused tokens + + y = ggml_add(ctx0, y, ggml_mul(ctx0, cur, layer.ssm_d)); + y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y); + + // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} + cur = build_lora_mm(layer.ssm_out, y); + } + + // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); + + return cur; +} + +ggml_tensor * llm_graph_context_mamba::build_mamba2_layer(llm_graph_input_rs * inp, + ggml_tensor * cur, + const llama_model & model, + const llama_ubatch & ubatch, + int il) const { + const auto * mctx_cur = inp->mctx; + + const auto kv_head = mctx_cur->get_head(); + + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t n_head = hparams.ssm_dt_rank; + const int64_t head_dim = d_inner / n_head; + const int64_t n_group = hparams.ssm_n_group; + const int64_t n_seqs = ubatch.n_seqs; + + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); + conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs); + + // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); + + // d_in_proj = 2 * self.d_inner + 2 * self.ngroups * self.d_state + self.nheads + + // {n_embd, d_in_proj} @ {n_embd, n_seq_tokens, n_seqs} => {d_in_proj, n_seq_tokens, n_seqs} + ggml_tensor * zxBCdt = build_lora_mm(model.layers[il].ssm_in, cur); + + // split the above in three + ggml_tensor * z = ggml_view_4d(ctx0, zxBCdt, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * zxBCdt->nb[0], + zxBCdt->nb[1], zxBCdt->nb[2], 0); + ggml_tensor * xBC = ggml_view_3d(ctx0, zxBCdt, d_inner + 2 * n_group * d_state, n_seq_tokens, n_seqs, zxBCdt->nb[1], + zxBCdt->nb[2], d_inner * ggml_element_size(zxBCdt)); + ggml_tensor * dt = ggml_view_3d(ctx0, zxBCdt, n_head, n_seq_tokens, n_seqs, zxBCdt->nb[1], zxBCdt->nb[2], + (2 * d_inner + 2 * n_group * d_state) * ggml_element_size(zxBCdt)); + + // conv + { + // => {d_conv - 1 + n_seq_tokens, d_inner + 2*n_group*d_state, n_seqs} + ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, xBC), 0); + + // copy last (d_conv - 1) columns back into the state cache + ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs, + conv_x->nb[1], conv_x->nb[2], n_seq_tokens * (conv_x->nb[0])); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv, + ggml_view_1d(ctx0, conv_states_all, + (d_conv - 1) * (d_inner + 2 * n_group * d_state) * (n_seqs), + kv_head * (d_conv - 1) * (d_inner + 2 * n_group * d_state) * + ggml_element_size(conv_states_all)))); + + // 1D convolution + // The equivalent is to make a self-overlapping view of conv_x + // over d_conv columns at each stride in the 3rd dimension, + // then element-wise multiply that with the conv1d weight, + // then sum the elements of each row, + // (the last two steps are a dot product over rows (also doable with mul_mat)) + // then permute away the ne[0] dimension, + // and then you're left with the resulting x tensor. + // For simultaneous sequences, all sequences need to have the same length. + xBC = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d); + + // bias + xBC = ggml_add(ctx0, xBC, model.layers[il].ssm_conv1d_b); + + xBC = ggml_silu(ctx0, xBC); + } + + // ssm + { + // These correspond to V K Q in SSM/attention duality + ggml_tensor * x = ggml_view_4d(ctx0, xBC, head_dim, n_head, n_seq_tokens, n_seqs, head_dim * xBC->nb[0], + xBC->nb[1], xBC->nb[2], 0); + ggml_tensor * B = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state * xBC->nb[0], + xBC->nb[1], xBC->nb[2], d_inner * ggml_element_size(xBC)); + ggml_tensor * C = ggml_view_4d(ctx0, xBC, d_state, n_group, n_seq_tokens, n_seqs, d_state * xBC->nb[0], + xBC->nb[1], xBC->nb[2], (d_inner + n_group * d_state) * ggml_element_size(xBC)); + + // {n_head, n_seq_tokens, n_seqs} + dt = ggml_add(ctx0, ggml_cont(ctx0, dt), model.layers[il].ssm_dt_b); + + ggml_tensor * A = model.layers[il].ssm_a; + + // use the states and the indices provided by build_recurrent_state + // (this is necessary in order to properly use the states before they are overwritten, + // while avoiding to make unnecessary copies of the states) + auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) { + ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_head, mctx_cur->get_size()); + + // TODO: use semistructured matrices to implement state-space duality + // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} + return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids); + }; + + ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows); + + // store last states + ggml_build_forward_expand( + gf, ggml_cpy(ctx0, ggml_view_1d(ctx0, y_ssm, d_state * d_inner * n_seqs, ggml_nelements(x) * x->nb[0]), + ggml_view_1d(ctx0, ssm_states_all, d_state * d_inner * n_seqs, + kv_head * d_state * d_inner * ggml_element_size(ssm_states_all)))); + + ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_head, n_seq_tokens, n_seqs, x->nb[1], n_head * x->nb[1], + n_seq_tokens * n_head * x->nb[1], 0); + + // TODO: skip computing output earlier for unused tokens + + y = ggml_add(ctx0, y, ggml_mul(ctx0, x, model.layers[il].ssm_d)); + cb(y, "mamba2_y_add_d", il); + y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y); + + // grouped RMS norm + if (model.layers[il].ssm_norm) { + y = ggml_reshape_4d(ctx0, y, d_inner / n_group, n_group, n_seq_tokens, n_seqs); + y = build_norm(y, model.layers[il].ssm_norm, NULL, LLM_NORM_RMS, il); + } + + y = ggml_reshape_3d(ctx0, y, d_inner, n_seq_tokens, n_seqs); + + // {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} + cur = build_lora_mm(model.layers[il].ssm_out, y); + } + + // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); + cb(cur, "mamba_out", il); + + return cur; +} diff --git a/examples/talk-llama/models/grok.cpp b/examples/talk-llama/models/grok.cpp new file mode 100644 index 00000000..3c54dfee --- /dev/null +++ b/examples/talk-llama/models/grok.cpp @@ -0,0 +1,159 @@ +#include "models.h" + +llm_build_grok::llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + cur = build_norm(cur, + model.layers[il].attn_out_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_out_norm", il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // MoE branch + ggml_tensor * moe_out = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_GELU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + if (model.layers[il].ffn_up) { + ggml_tensor * ffn_out = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_PAR, il); + cb(ffn_out, "ffn_out", il); + + cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_post_norm", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cur = ggml_scale(ctx0, cur, hparams.f_logit_scale); + + // final logit soft-capping + if (hparams.f_final_logit_softcapping) { + cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping); + cur = ggml_tanh(ctx0, cur); + cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping); + } + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/grovemoe.cpp b/examples/talk-llama/models/grovemoe.cpp new file mode 100644 index 00000000..56b6db9a --- /dev/null +++ b/examples/talk-llama/models/grovemoe.cpp @@ -0,0 +1,141 @@ +#include "models.h" + + + +llm_build_grovemoe::llm_build_grovemoe(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_chunk_expert = n_expert / hparams.n_group_experts; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * probs = build_lora_mm(model.layers[il].ffn_gate_inp, cur); // [n_expert, n_tokens] + cb(probs, "ffn_moe_logits", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + nullptr, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il, + probs); + cb(moe_out, "ffn_moe_out", il); + cur = moe_out; + + // TODO: Only do the expert selection and weights once + moe_out = build_moe_ffn(cur, + nullptr, + model.layers[il].ffn_up_chexps, + model.layers[il].ffn_gate_chexps, + model.layers[il].ffn_down_chexps, + nullptr, + n_chunk_expert, n_expert_used > n_chunk_expert ? n_chunk_expert : n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il, + probs); + cb(moe_out, "ffn_adj_moe_out", il); + + cur = ggml_add(ctx0, cur, ggml_scale(ctx0, moe_out, hparams.expert_group_scale)); + cb(cur, "ffn_final_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/hunyuan-dense.cpp b/examples/talk-llama/models/hunyuan-dense.cpp new file mode 100644 index 00000000..7d5dcc78 --- /dev/null +++ b/examples/talk-llama/models/hunyuan-dense.cpp @@ -0,0 +1,132 @@ +#include "models.h" + +llm_build_hunyuan_dense::llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, nullptr, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_norm", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, nullptr, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_norm", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + // feed-forward network (non-MoE) + ggml_tensor * cur_mlp = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur_mlp, "ffn_out", il); + + cur = ggml_add(ctx0, cur_mlp, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/hunyuan-moe.cpp b/examples/talk-llama/models/hunyuan-moe.cpp new file mode 100644 index 00000000..77e39de5 --- /dev/null +++ b/examples/talk-llama/models/hunyuan-moe.cpp @@ -0,0 +1,154 @@ +#include "models.h" + +llm_build_hunyuan_moe::llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = 1.0f / sqrtf(float(n_embd_head)); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, nullptr, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_norm", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, nullptr, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_norm", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network (non-MoE) + ggml_tensor * cur_mlp = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur_mlp, "ffn_mlp", il); + + // MoE branch + ggml_tensor * cur_moe = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, + true, // norm_topk_prob + false, + 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur_moe, "ffn_moe_out", il); + + ggml_tensor * ffn_out = ggml_add(ctx0, cur_moe, cur_mlp); + cb(ffn_out, "ffn_out", il); + + cur = ggml_add(ctx0, ffn_out, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/internlm2.cpp b/examples/talk-llama/models/internlm2.cpp new file mode 100644 index 00000000..387e8211 --- /dev/null +++ b/examples/talk-llama/models/internlm2.cpp @@ -0,0 +1,120 @@ +#include "models.h" + +llm_build_internlm2::llm_build_internlm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/jais.cpp b/examples/talk-llama/models/jais.cpp new file mode 100644 index 00000000..3e3376e6 --- /dev/null +++ b/examples/talk-llama/models/jais.cpp @@ -0,0 +1,86 @@ +#include "models.h" + +llm_build_jais::llm_build_jais(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*cur->nb[0]*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*cur->nb[0]*(n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*cur->nb[0]*(n_embd + n_embd_gqa)); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/float(n_embd_head), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + inpL = ggml_add(ctx0, cur, ffn_inp); + cb(inpL, "l_out", il); + } + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/jamba.cpp b/examples/talk-llama/models/jamba.cpp new file mode 100644 index 00000000..a0187772 --- /dev/null +++ b/examples/talk-llama/models/jamba.cpp @@ -0,0 +1,106 @@ +#include "models.h" + +llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + auto * inp_hybrid = build_inp_mem_hybrid(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const int64_t n_head_kv = hparams.n_head_kv(il); + + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + if (n_head_kv == 0) { + cur = build_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il); + } else { + // Attention + + struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // No RoPE :) + cur = build_attn(inp_hybrid->get_attn(), + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + // residual + struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur); + cb(cur, "ffn_inp", il); + + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + // FFN + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + } + // residual + cur = ggml_add(ctx0, ffn_inp, cur); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + // final rmsnorm + cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/lfm2.cpp b/examples/talk-llama/models/lfm2.cpp new file mode 100644 index 00000000..ca06bacd --- /dev/null +++ b/examples/talk-llama/models/lfm2.cpp @@ -0,0 +1,173 @@ +#include "models.h" + +#include "../llama-memory-hybrid.h" + + +llm_build_lfm2::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params), + model(model) { + ggml_tensor * cur = build_inp_embd(model.tok_embd); + cb(cur, "model.embed_tokens", -1); + + ggml_tensor * inp_pos = build_inp_pos(); + auto * inp_hybrid = build_inp_mem_hybrid(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const bool is_moe_layer = il >= static_cast(hparams.n_layer_dense_lead); + + auto * prev_cur = cur; + cur = build_norm(cur, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "model.layers.{}.operator_norm", il); + + cur = hparams.is_recurrent(il) ? build_shortconv_block(cur, inp_hybrid->get_recr(), il) : + build_attn_block(cur, inp_pos, inp_hybrid->get_attn(), il); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + prev_cur = ggml_get_rows(ctx0, prev_cur, inp_out_ids); + } + + cur = ggml_add(ctx0, prev_cur, cur); + + auto * ffn_norm_out = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(ffn_norm_out, "model.layers.{}.ffn_norm", il); + + ggml_tensor * ffn_out = + is_moe_layer ? build_moe_feed_forward(ffn_norm_out, il) : build_dense_feed_forward(ffn_norm_out, il); + cb(ffn_norm_out, "model.layers.{}.ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_out); + } + + cur = build_norm(cur, model.tok_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "model.embedding_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + cb(cur, "lm_head", -1); + + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +ggml_tensor * llm_build_lfm2::build_moe_feed_forward(ggml_tensor * cur, int il) const { + return build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, n_expert, n_expert_used, LLM_FFN_SILU, true, false, 0.0, + static_cast(hparams.expert_gating_func), il); +} + +ggml_tensor * llm_build_lfm2::build_dense_feed_forward(ggml_tensor * cur, int il) const { + GGML_ASSERT(!model.layers[il].ffn_up_b); + GGML_ASSERT(!model.layers[il].ffn_gate_b); + GGML_ASSERT(!model.layers[il].ffn_down_b); + return build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); +} + +ggml_tensor * llm_build_lfm2::build_attn_block(ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv * inp_attn, + int il) const { + GGML_ASSERT(hparams.n_embd_v_gqa(il) == hparams.n_embd_k_gqa(il)); + const auto n_embd_head = hparams.n_embd_head_v; + const auto n_head_kv = hparams.n_head_kv(il); + + auto * q = build_lora_mm(model.layers[il].wq, cur); + cb(q, "model.layers.{}.self_attn.q_proj", il); + auto * k = build_lora_mm(model.layers[il].wk, cur); + cb(k, "model.layers.{}.self_attn.k_proj", il); + auto * v = build_lora_mm(model.layers[il].wv, cur); + cb(v, "model.layers.{}.self_attn.v_proj", il); + + q = ggml_reshape_3d(ctx0, q, n_embd_head, n_head, n_tokens); + k = ggml_reshape_3d(ctx0, k, n_embd_head, n_head_kv, n_tokens); + v = ggml_reshape_3d(ctx0, v, n_embd_head, n_head_kv, n_tokens); + + // qk norm + q = build_norm(q, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(q, "model.layers.{}.self_attn.q_layernorm", il); + k = build_norm(k, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(k, "model.layers.{}.self_attn.k_layernorm", il); + + // RoPE + q = ggml_rope_ext(ctx0, q, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, + attn_factor, beta_fast, beta_slow); + k = ggml_rope_ext(ctx0, k, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, + attn_factor, beta_fast, beta_slow); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + q, k, v, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + + cb(cur, "model.layers.{}.self_attn.out_proj", il); + + return cur; +} + +ggml_tensor * llm_build_lfm2::build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il) { + const auto * mctx_cur = static_cast(mctx)->get_recr(); + const uint32_t kv_head = mctx_cur->get_head(); + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + const int64_t n_seqs = ubatch.n_seqs; + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + GGML_ASSERT(hparams.n_shortconv_l_cache > 1); + const uint32_t d_conv = hparams.n_shortconv_l_cache - 1; + + // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); + + auto * bcx = build_lora_mm(model.layers[il].shortconv.in_proj, cur); + cb(bcx, "model.layers.{}.conv.in_proj", il); + + constexpr auto n_chunks = 3; + GGML_ASSERT(bcx->ne[0] % n_chunks == 0); + const auto chunk_size = bcx->ne[0] / n_chunks; + auto * b = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], + 0 * chunk_size * ggml_element_size(bcx)); + auto * c = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], + 1 * chunk_size * ggml_element_size(bcx)); + auto * x = ggml_view_3d(ctx0, bcx, chunk_size, bcx->ne[1], bcx->ne[2], bcx->nb[1], bcx->nb[2], + 2 * chunk_size * ggml_element_size(bcx)); + + auto * bx = ggml_transpose(ctx0, ggml_mul(ctx0, b, x)); + + // read conv state + auto * conv_state = mctx_cur->get_r_l(il); + auto * conv_rs = build_rs(inp_recr, conv_state, hparams.n_embd_r(), n_seqs); + auto * conv = ggml_reshape_3d(ctx0, conv_rs, d_conv, hparams.n_embd, n_seqs); + + bx = ggml_concat(ctx0, conv, bx, 0); + GGML_ASSERT(bx->ne[0] > conv->ne[0]); + + // last d_conv columns is a new conv state + auto * new_conv = ggml_view_3d(ctx0, bx, conv->ne[0], bx->ne[1], bx->ne[2], bx->nb[1], bx->nb[2], + (bx->ne[0] - conv->ne[0]) * ggml_element_size(bx)); + GGML_ASSERT(ggml_are_same_shape(conv, new_conv)); + + // write new conv conv state + ggml_build_forward_expand(gf, ggml_cpy(ctx0, new_conv, + ggml_view_1d(ctx0, conv_state, ggml_nelements(new_conv), + kv_head * d_conv * n_embd * ggml_element_size(new_conv)))); + + auto * conv_kernel = model.layers[il].shortconv.conv; + auto * conv_out = ggml_ssm_conv(ctx0, bx, conv_kernel); + cb(conv_out, "model.layers.{}.conv.conv", il); + + auto * y = ggml_mul(ctx0, c, conv_out); + y = build_lora_mm(model.layers[il].shortconv.out_proj, y); + cb(y, "model.layers.{}.conv.out_proj", il); + // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} + y = ggml_reshape_2d(ctx0, y, y->ne[0], n_seq_tokens * n_seqs); + + return y; +} diff --git a/examples/talk-llama/models/llada-moe.cpp b/examples/talk-llama/models/llada-moe.cpp new file mode 100644 index 00000000..5f64686f --- /dev/null +++ b/examples/talk-llama/models/llada-moe.cpp @@ -0,0 +1,122 @@ +#include "models.h" + +llm_build_llada_moe::llm_build_llada_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_no_cache(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/llada.cpp b/examples/talk-llama/models/llada.cpp new file mode 100644 index 00000000..85703366 --- /dev/null +++ b/examples/talk-llama/models/llada.cpp @@ -0,0 +1,99 @@ +#include "models.h" + +llm_build_llada::llm_build_llada(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + // LLaDA is similar to LLaMA but uses non-causal attention for diffusion + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // Non-causal attention for diffusion + auto * inp_attn = build_attn_inp_no_cache(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute separate Q, K, V projections without bias, matching LLaDALlamaBlock + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/llama-iswa.cpp b/examples/talk-llama/models/llama-iswa.cpp new file mode 100644 index 00000000..03f80616 --- /dev/null +++ b/examples/talk-llama/models/llama-iswa.cpp @@ -0,0 +1,174 @@ +#include "models.h" + +llm_build_llama_iswa::llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + // temperature tuning + ggml_tensor * inp_attn_scale = nullptr; + inp_attn_scale = build_inp_attn_scale(); + + auto * inp_attn = build_attn_inp_kv_iswa(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + const bool use_rope = hparams.n_no_rope_layer_step > 0 && + (il + 1) % hparams.n_no_rope_layer_step != 0; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (use_rope) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } else if (inp_attn_scale) { + Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale); + } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + if (use_rope && hparams.use_kq_norm) { + // Llama4TextL2Norm + Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps); + Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps); + cb(Qcur, "Qcur_normed", il); + cb(Kcur, "Kcur_normed", il); + } + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + ggml_tensor * ffn_inp_normed = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = build_moe_ffn(ffn_inp_normed, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID, + il); + + // Shared experts + ggml_tensor * shexp_out = build_ffn(ffn_inp_normed, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(shexp_out, "ffn_moe_shexp", il); + + cur = ggml_add(ctx0, moe_out, shexp_out); + cb(cur, "ffn_moe_out_merged", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/llama.cpp b/examples/talk-llama/models/llama.cpp new file mode 100644 index 00000000..ab7fd5d0 --- /dev/null +++ b/examples/talk-llama/models/llama.cpp @@ -0,0 +1,155 @@ +#include "models.h" + +llm_build_llama::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // rope freq factors for llama3; may return nullptr for llama2 and other models + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + if (hparams.use_kq_norm) { + // Llama4TextL2Norm + Qcur = ggml_rms_norm(ctx0, Qcur, hparams.f_norm_rms_eps); + Kcur = ggml_rms_norm(ctx0, Kcur, hparams.f_norm_rms_eps); + cb(Qcur, "Qcur_normed", il); + cb(Kcur, "Kcur_normed", il); + } + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network (non-MoE) + if (model.layers[il].ffn_gate_inp == nullptr) { + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/mamba.cpp b/examples/talk-llama/models/mamba.cpp new file mode 100644 index 00000000..46819613 --- /dev/null +++ b/examples/talk-llama/models/mamba.cpp @@ -0,0 +1,55 @@ +#include "models.h" + + +llm_build_mamba::llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + auto * rs_inp = build_rs_inp(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + if (model.arch == LLM_ARCH_MAMBA2) { + cur = build_mamba2_layer(rs_inp, cur, model, ubatch, il); + } else { + cur = build_mamba_layer(rs_inp, cur, model, ubatch, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // residual + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + // final rmsnorm + cur = build_norm(inpL, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + diff --git a/examples/talk-llama/models/minicpm3.cpp b/examples/talk-llama/models/minicpm3.cpp new file mode 100644 index 00000000..f374a9fd --- /dev/null +++ b/examples/talk-llama/models/minicpm3.cpp @@ -0,0 +1,199 @@ +#include "models.h" + +llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + //TODO: if the model varies, these parameters need to be read from the model + const int64_t n_embd_base = 256; + const float scale_embd = 12.0f; + const float scale_depth = 1.4f; + const float kq_scale = 1.0f / sqrtf(float(hparams.n_embd_head_k)); + + const uint32_t n_embd_head_qk_rope = hparams.n_rot; + const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // scale the input embeddings + inpL = ggml_scale(ctx0, inpL, scale_embd); + cb(inpL, "inp_scaled", -1); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + // {n_embd, q_lora_rank} * {n_embd, n_tokens} -> {q_lora_rank, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur); + cb(q, "q", il); + + q = build_norm(q, + model.layers[il].attn_q_a_norm, NULL, + LLM_NORM_RMS, il); + cb(q, "q", il); + + // {q_lora_rank, n_head * hparams.n_embd_head_k} * {q_lora_rank, n_tokens} -> {n_head * hparams.n_embd_head_k, n_tokens} + q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q); + cb(q, "q", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + kv_compressed = build_norm(kv_compressed, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + // scale_res - scale the hidden states for residual connection + const float scale_res = scale_depth/sqrtf(float(n_layer)); // TODO: is this correct? + cur = ggml_scale(ctx0, cur, scale_res); + cb(cur, "hidden_scaled", il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + // scale the hidden states for residual connection + cur = ggml_scale(ctx0, cur, scale_res); + cb(cur, "hidden_scaled_ffn", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head scaling + const float scale_lmhead = float(n_embd_base)/float(n_embd); + cur = ggml_scale(ctx0, cur, scale_lmhead); + cb(cur, "lmhead_scaling", -1); + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/minimax-m2.cpp b/examples/talk-llama/models/minimax-m2.cpp new file mode 100644 index 00000000..f7001bad --- /dev/null +++ b/examples/talk-llama/models/minimax-m2.cpp @@ -0,0 +1,124 @@ + +#include "models.h" + +llm_build_minimax_m2::llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + // GGML_ASSERT(n_embd_head == hparams.n_rot); this is wrong in case of minimax, head_dim = 128, n_rot = 64 + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * inp_pos = build_inp_pos(); + auto inp_attn = build_attn_inp_kv(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = inpL; + + // self_attention + { + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + model.layers[il].ffn_exp_probs_b, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + (llama_expert_gating_func_type) hparams.expert_gating_func, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/models.h b/examples/talk-llama/models/models.h new file mode 100644 index 00000000..2fffb382 --- /dev/null +++ b/examples/talk-llama/models/models.h @@ -0,0 +1,481 @@ +#pragma once + +#include "../llama-model.h" +#include "../llama-graph.h" +#include "../llama-memory-recurrent.h" + +#include + +struct llm_graph_context_mamba : public llm_graph_context { + llm_graph_context_mamba(const llm_graph_params & params); + + virtual ~llm_graph_context_mamba() = default; + + ggml_tensor * build_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il); + ggml_tensor * build_mamba2_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il) const; + +}; + +// Base class for RWKV-related models +struct llm_build_rwkv6_base : public llm_graph_context { + const llama_model & model; + + llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params); + + virtual ~llm_build_rwkv6_base() = default; + + ggml_tensor * build_rwkv6_channel_mix(const llama_layer * layer, + ggml_tensor * cur, + ggml_tensor * x_prev, + llm_arch arch) const; + + ggml_tensor * build_rwkv6_time_mix(llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * x_prev, + const llama_ubatch & ubatch, + int il) const; +}; + +// Base class for RWKV7-related models +struct llm_build_rwkv7_base : public llm_graph_context { + const llama_model & model; + + llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params); + + virtual ~llm_build_rwkv7_base() = default; + + // RWKV7-specific graph building methods + ggml_tensor * build_rwkv7_channel_mix(const llama_layer * layer, + ggml_tensor * cur, + ggml_tensor * x_prev, + llm_arch arch) const; + ggml_tensor * build_rwkv7_time_mix(llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * x_prev, + ggml_tensor *& first_layer_value, + const llama_ubatch & ubatch, + int il) const; +}; + +struct llm_build_apertus : public llm_graph_context { + llm_build_apertus(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_arcee : public llm_graph_context { + llm_build_arcee(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_arctic : public llm_graph_context { + llm_build_arctic(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_arwkv7 : public llm_build_rwkv7_base { + llm_build_arwkv7(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_baichuan : public llm_graph_context { + llm_build_baichuan(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_bailingmoe2 : public llm_graph_context { + llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_bailingmoe : public llm_graph_context { + llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_bert : public llm_graph_context { + llm_build_bert(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_bitnet : public llm_graph_context { + llm_build_bitnet(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_bloom : public llm_graph_context { + llm_build_bloom(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_chameleon : public llm_graph_context { + llm_build_chameleon(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_chatglm : public llm_graph_context { + llm_build_chatglm(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_codeshell : public llm_graph_context { + llm_build_codeshell(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_cogvlm : public llm_graph_context { + llm_build_cogvlm(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_cohere2_iswa : public llm_graph_context { + llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_command_r : public llm_graph_context { + llm_build_command_r(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_dbrx : public llm_graph_context { + llm_build_dbrx(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_deci : public llm_graph_context { + llm_build_deci(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_deepseek2 : public llm_graph_context { + llm_build_deepseek2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_deepseek : public llm_graph_context { + llm_build_deepseek(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_dots1 : public llm_graph_context { + llm_build_dots1(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_dream : public llm_graph_context { + llm_build_dream(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_ernie4_5 : public llm_graph_context { + llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_ernie4_5_moe : public llm_graph_context { + llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params); +}; + +template +struct llm_build_exaone4 : public llm_graph_context { + llm_build_exaone4(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_exaone : public llm_graph_context { + llm_build_exaone(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_falcon : public llm_graph_context { + llm_build_falcon(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_falcon_h1 : public llm_graph_context_mamba { + llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_gemma2_iswa : public llm_graph_context { + llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_gemma3_iswa : public llm_graph_context { + llm_build_gemma3_iswa(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_gemma3n_iswa : public llm_graph_context { + const llama_model & model; + + const int64_t n_embd_head; + const int64_t n_embd_altup; + const int64_t n_altup; + const int i_altup_act; + const int n_layer_sparsity = 10; // number of layers using activation sparsity + const float f_sparsity_std_mul = 1.6448533535003662f; // std_multiplier = normal_dist.icdf(0.95) + + llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params); + ggml_tensor * calc_magnitude(ggml_tensor * x); + ggml_tensor * view_2d_slice(ggml_tensor * x, int idx); + ggml_tensor * get_per_layer_inputs(); + ggml_tensor * project_per_layer_inputs(ggml_tensor * inputs_embeds, ggml_tensor * inp_per_layer); + ggml_tensor * gaussian_topk(ggml_tensor * x); + ggml_tensor * altup_compute_router_modalities(ggml_tensor * x, int il); + ggml_tensor * altup_predict(ggml_tensor * cur, int il); + ggml_tensor * laurel(ggml_tensor * cur, int il); + ggml_tensor * altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il); +}; + +struct llm_build_gemma_embedding : public llm_graph_context { + llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_gemma : public llm_graph_context { + llm_build_gemma(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_glm4 : public llm_graph_context { + llm_build_glm4(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_glm4_moe : public llm_graph_context { + llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_gpt2 : public llm_graph_context { + llm_build_gpt2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_gptneox : public llm_graph_context { + llm_build_gptneox(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_granite : public llm_graph_context { + llm_build_granite(const llama_model & model, const llm_graph_params & params); + +private: + ggml_tensor * build_attention_layer( + ggml_tensor * cur, + ggml_tensor * inp_pos, + llm_graph_input_attn_kv * inp_attn, + const llama_model & model, + const int64_t n_embd_head, + const int il); + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + ggml_tensor * inpSA, + const llama_model & model, + const int il); +}; + +struct llm_build_granite_hybrid : public llm_graph_context_mamba { + llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params); + ggml_tensor * build_layer_ffn(ggml_tensor * cur, ggml_tensor * inpSA, const llama_model & model, const int il); + ggml_tensor * build_attention_layer(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn, + const llama_model & model,const int64_t n_embd_head, const int il); +}; + +struct llm_build_grok : public llm_graph_context { + llm_build_grok(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_grovemoe : public llm_graph_context { + llm_build_grovemoe(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_hunyuan_dense : public llm_graph_context { + llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_hunyuan_moe : public llm_graph_context { + llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_internlm2 : public llm_graph_context { + llm_build_internlm2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_jais : public llm_graph_context { + llm_build_jais(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_jamba : public llm_graph_context_mamba { + llm_build_jamba(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_lfm2 : public llm_graph_context { + const llama_model & model; + + llm_build_lfm2(const llama_model & model, const llm_graph_params & params); + ggml_tensor * build_moe_feed_forward(ggml_tensor * cur, int il) const; + ggml_tensor * build_dense_feed_forward(ggml_tensor * cur, int il) const; + ggml_tensor * build_attn_block(ggml_tensor * cur, ggml_tensor * inp_pos, llm_graph_input_attn_kv * inp_attn, int il) const; + ggml_tensor * build_shortconv_block(ggml_tensor * cur, llm_graph_input_rs * inp_recr, int il); + +}; + +struct llm_build_llada : public llm_graph_context { + llm_build_llada(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_llada_moe : public llm_graph_context { + llm_build_llada_moe(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_llama : public llm_graph_context { + llm_build_llama(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_llama_iswa : public llm_graph_context { + llm_build_llama_iswa(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_mamba : public llm_graph_context_mamba { + llm_build_mamba(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_minicpm3 : public llm_graph_context { + llm_build_minicpm3(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_minimax_m2 : public llm_graph_context { + llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_mpt : public llm_graph_context { + llm_build_mpt(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_nemotron : public llm_graph_context { + llm_build_nemotron(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_nemotron_h : public llm_graph_context_mamba { + llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params); + ggml_tensor * build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il); + ggml_tensor * build_attention_layer(ggml_tensor * cur, llm_graph_input_attn_kv * inp_attn, + const llama_model & model, const int64_t n_embd_head, const int il); +}; + +struct llm_build_neo_bert : public llm_graph_context { + llm_build_neo_bert(const llama_model & model, const llm_graph_params & params); +}; + +template +struct llm_build_olmo2 : public llm_graph_context { + llm_build_olmo2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_olmoe : public llm_graph_context { + llm_build_olmoe(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_olmo : public llm_graph_context { + llm_build_olmo(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_openai_moe_iswa : public llm_graph_context { + llm_build_openai_moe_iswa(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_openelm : public llm_graph_context { + llm_build_openelm(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_orion : public llm_graph_context { + llm_build_orion(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_pangu_embedded : public llm_graph_context { + llm_build_pangu_embedded(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_phi2 : public llm_graph_context { + llm_build_phi2(const llama_model & model, const llm_graph_params & params); +}; + +template +struct llm_build_phi3 : public llm_graph_context { + llm_build_phi3(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_plamo2 : public llm_graph_context_mamba { + llm_build_plamo2(const llama_model & model, const llm_graph_params & params); + private: + ggml_tensor * build_plamo2_mamba_layer(llm_graph_input_rs * inp, ggml_tensor * cur, const llama_model & model, const llama_ubatch & ubatch, int il); + ggml_tensor * build_plamo2_attn_layer(llm_graph_input_attn_kv * inp, ggml_tensor * inp_pos, ggml_tensor * cur, + const llama_model & model, int il); +}; + +struct llm_build_plamo : public llm_graph_context { + llm_build_plamo(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_plm : public llm_graph_context { + llm_build_plm(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_qwen2 : public llm_graph_context { + llm_build_qwen2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_qwen2moe : public llm_graph_context { + llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_qwen2vl : public llm_graph_context { + llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_qwen3 : public llm_graph_context { + llm_build_qwen3(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_qwen3moe : public llm_graph_context { + llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_qwen3vl : public llm_graph_context { + llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_qwen3vlmoe : public llm_graph_context { + llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params); +}; + + +struct llm_build_qwen : public llm_graph_context { + llm_build_qwen(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_refact : public llm_graph_context { + llm_build_refact(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_rwkv6 : public llm_build_rwkv6_base { + llm_build_rwkv6(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_rwkv6qwen2 : public llm_build_rwkv6_base { + llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_rwkv7 : public llm_build_rwkv7_base { + llm_build_rwkv7(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_seed_oss : public llm_graph_context { + llm_build_seed_oss(const llama_model & model, const llm_graph_params & params); +}; + +template +struct llm_build_smallthinker : public llm_graph_context { + llm_build_smallthinker(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_smollm3 : public llm_graph_context { + llm_build_smollm3(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_stablelm : public llm_graph_context { + llm_build_stablelm(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_starcoder2 : public llm_graph_context { + llm_build_starcoder2(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_starcoder : public llm_graph_context { + llm_build_starcoder(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_t5_dec : public llm_graph_context { + llm_build_t5_dec(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_t5_enc : public llm_graph_context { + llm_build_t5_enc(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_wavtokenizer_dec : public llm_graph_context { + llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params); +}; + +struct llm_build_xverse : public llm_graph_context { + llm_build_xverse(const llama_model & model, const llm_graph_params & params); +}; diff --git a/examples/talk-llama/models/mpt.cpp b/examples/talk-llama/models/mpt.cpp new file mode 100644 index 00000000..2328e027 --- /dev/null +++ b/examples/talk-llama/models/mpt.cpp @@ -0,0 +1,126 @@ +#include "models.h" + + + +llm_build_mpt::llm_build_mpt(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * pos; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv(); + + if (model.pos_embd) { + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + } + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * attn_norm; + + attn_norm = build_norm(inpL, model.layers[il].attn_norm, model.layers[il].attn_norm_b, LLM_NORM, il); + cb(attn_norm, "attn_norm", il); + + // self-attention + { + cur = attn_norm; + + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + if (model.layers[il].bqkv) { + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + } + + if (hparams.f_clamp_kqv > 0.0f) { + cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(cur, "wqkv_clamped", il); + } + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 0 * sizeof(float) * (n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 1 * sizeof(float) * (n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), + cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)); + + // Q/K Layernorm + if (model.layers[il].attn_q_norm) { + Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens); + Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + } + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + + // Add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // feed forward + { + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, LLM_NORM, il); + cb(cur, "ffn_norm", il); + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + model.layers[il].ffn_act, LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/nemotron-h.cpp b/examples/talk-llama/models/nemotron-h.cpp new file mode 100644 index 00000000..54143488 --- /dev/null +++ b/examples/talk-llama/models/nemotron-h.cpp @@ -0,0 +1,121 @@ +#include "models.h" + + + +llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + ggml_build_forward_expand(gf, inpL); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + if (hparams.is_recurrent(il)) { + // ssm layer // + cur = build_mamba2_layer(inp->get_recr(), cur, model, ubatch, il); + } else if (hparams.n_ff(il) == 0) { + // attention layer // + cur = build_attention_layer(cur, inp->get_attn(), model, n_embd_head, il); + } else { + cur = build_ffn_layer(cur, model, il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + // add residual + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "nemotron_h_block_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor * cur, + llm_graph_input_attn_kv * inp_attn, + const llama_model & model, + const int64_t n_embd_head, + const int il) { + // compute Q and K and (optionally) RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + const float kq_scale = + hparams.f_attention_scale == 0.0f ? 1.0f / sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + return cur; +} + +ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, const int il) { + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, LLM_FFN_RELU_SQR, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + return cur; +} diff --git a/examples/talk-llama/models/nemotron.cpp b/examples/talk-llama/models/nemotron.cpp new file mode 100644 index 00000000..fcead041 --- /dev/null +++ b/examples/talk-llama/models/nemotron.cpp @@ -0,0 +1,122 @@ +#include "models.h" + +llm_build_nemotron::llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + //GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/neo-bert.cpp b/examples/talk-llama/models/neo-bert.cpp new file mode 100644 index 00000000..7c32bfca --- /dev/null +++ b/examples/talk-llama/models/neo-bert.cpp @@ -0,0 +1,104 @@ +#include "models.h" + +llm_build_neo_bert::llm_build_neo_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + ggml_tensor * inp_pos = build_inp_pos(); + + // construct input embeddings (token, type, position) + inpL = build_inp_embd(model.tok_embd); + cb(inpL, "inp_embd", -1); + + auto * inp_attn = build_attn_inp_no_cache(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * cur = inpL; + + // pre-norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + + { + ggml_tensor * Qcur; + ggml_tensor * Kcur; + ggml_tensor * Vcur; + + // self-attention + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + + // RoPE + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + cb(cur, "kqv_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + // re-add the layer input + cur = ggml_add(ctx0, cur, inpL); + + ggml_tensor * ffn_inp = cur; + cb(ffn_inp, "ffn_inp", il); + + // pre-norm + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + cur = build_ffn(cur, + model.layers[il].ffn_up, + NULL, NULL, NULL, NULL, NULL, + model.layers[il].ffn_down, + NULL, NULL, NULL, + LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + + // attentions bypass the intermediate layer + cur = ggml_add(ctx0, cur, ffn_inp); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm_enc, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_embd", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/olmo.cpp b/examples/talk-llama/models/olmo.cpp new file mode 100644 index 00000000..bbd623f1 --- /dev/null +++ b/examples/talk-llama/models/olmo.cpp @@ -0,0 +1,121 @@ +#include "models.h" + +llm_build_olmo::llm_build_olmo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + NULL, NULL, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Qcur = ggml_clamp(ctx0, Qcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Kcur = ggml_clamp(ctx0, Kcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (hparams.f_clamp_kqv > 0.0f) { + Vcur = ggml_clamp(ctx0, Vcur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + NULL, NULL, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + NULL, NULL, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/olmo2.cpp b/examples/talk-llama/models/olmo2.cpp new file mode 100644 index 00000000..713552da --- /dev/null +++ b/examples/talk-llama/models/olmo2.cpp @@ -0,0 +1,150 @@ +#include "models.h" + +template +llm_build_olmo2::llm_build_olmo2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + using inp_attn_type = std::conditional_t; + inp_attn_type * inp_attn = nullptr; + + if constexpr (iswa) { + inp_attn = build_attn_inp_kv_iswa(); + } else { + inp_attn = build_attn_inp_kv(); + } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = inpL; + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + const bool is_swa = hparams.is_swa(il); + + if (is_swa) { + // For sliding window layers, Olmo3 use regular rope with no yarn rope scaling. + // This is achieved here by setting freq_scale and attn_factor to 1. + // We also set ext_factor to 0 to avoid a few unnecessary computations. + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, 1.0, + 0.0, 1.0, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, 1.0, + 0.0, 1.0, beta_fast, beta_slow + ); + } else { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + cur = build_norm(cur, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_ffn(ffn_inp, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = build_norm(cur, + model.layers[il].ffn_post_norm, NULL, + LLM_NORM_RMS, -1); + cb(cur, "ffn_post_norm", -1); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +// Explicit template instantiations +template struct llm_build_olmo2; +template struct llm_build_olmo2; diff --git a/examples/talk-llama/models/olmoe.cpp b/examples/talk-llama/models/olmoe.cpp new file mode 100644 index 00000000..b8b6988f --- /dev/null +++ b/examples/talk-llama/models/olmoe.cpp @@ -0,0 +1,124 @@ +#include "models.h" + +llm_build_olmoe::llm_build_olmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/openai-moe-iswa.cpp b/examples/talk-llama/models/openai-moe-iswa.cpp new file mode 100644 index 00000000..3c0c0eec --- /dev/null +++ b/examples/talk-llama/models/openai-moe-iswa.cpp @@ -0,0 +1,123 @@ +#include "models.h" + +llm_build_openai_moe_iswa::llm_build_openai_moe_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv_iswa(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, nullptr, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, model.layers[il].attn_sinks, nullptr, 1.0f/sqrtf(float(n_rot)), il); + + cb(cur, "attn_out", il); + } + if (il == n_layer - 1) { + // skip computing output for unused tokens + ggml_tensor * inp_out_ids = build_inp_out_ids(); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = ffn_inp; + cur = build_norm(cur, + model.layers[il].attn_post_norm, nullptr, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + // MoE branch + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, model.layers[il].ffn_gate_inp_b, + model.layers[il].ffn_up_exps, model.layers[il].ffn_up_exps_b, + model.layers[il].ffn_gate_exps, model.layers[il].ffn_gate_exps_b, + model.layers[il].ffn_down_exps, model.layers[il].ffn_down_exps_b, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SWIGLU_OAI_MOE, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT, + il); + cb(cur, "ffn_moe_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/openelm.cpp b/examples/talk-llama/models/openelm.cpp new file mode 100644 index 00000000..ee46a337 --- /dev/null +++ b/examples/talk-llama/models/openelm.cpp @@ -0,0 +1,124 @@ +#include "models.h" + +llm_build_openelm::llm_build_openelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const int64_t n_head = hparams.n_head(il); + const int64_t n_head_kv = hparams.n_head_kv(il); + const int64_t n_head_qkv = 2*n_head_kv + n_head; + + cur = inpL; + ggml_tensor * residual = cur; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_reshape_3d(ctx0, cur, n_embd_head_k, n_head_qkv, n_tokens); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, cur->nb[1], cur->nb[2], 0); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*n_head); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, cur->nb[1], cur->nb[2], cur->nb[1]*(n_head+n_head_kv))); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, NULL, + LLM_NORM_RMS, il); + cb(Qcur, "Qcur", il); + + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, NULL, + LLM_NORM_RMS, il); + cb(Kcur, "Kcur", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, NULL, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, NULL, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Qcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, residual, cur); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + inpL = cur; + } + cur = inpL; + + // norm + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/orion.cpp b/examples/talk-llama/models/orion.cpp new file mode 100644 index 00000000..bb02273b --- /dev/null +++ b/examples/talk-llama/models/orion.cpp @@ -0,0 +1,123 @@ +#include "models.h" + +llm_build_orion::llm_build_orion(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + // if (model.layers[il].bq) { + // Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + // cb(Qcur, "Qcur", il); + // } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + // if (model.layers[il].bk) { + // Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + // cb(Kcur, "Kcur", il); + // } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + // if (model.layers[il].bv) { + // Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + // cb(Vcur, "Vcur", il); + // } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/pangu-embedded.cpp b/examples/talk-llama/models/pangu-embedded.cpp new file mode 100644 index 00000000..664572a5 --- /dev/null +++ b/examples/talk-llama/models/pangu-embedded.cpp @@ -0,0 +1,121 @@ +#include "models.h" + + +llm_build_pangu_embedded::llm_build_pangu_embedded(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + if (model.output_b != nullptr) { + cur = ggml_add(ctx0, cur, model.output_b); + } + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/phi2.cpp b/examples/talk-llama/models/phi2.cpp new file mode 100644 index 00000000..22dbf610 --- /dev/null +++ b/examples/talk-llama/models/phi2.cpp @@ -0,0 +1,121 @@ +#include "models.h" + + +llm_build_phi2::llm_build_phi2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * attn_norm_output; + ggml_tensor * ffn_output; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + attn_norm_output = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(attn_norm_output, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + } else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + } + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + // with phi2, we scale the Q to avoid precision issues + // ref: https://github.com/ml-explore/mlx-examples/blob/08e862336ade809bc37d1035f94b359e7d1a5152/phi2/phi2.py#L64-L66 + Qcur = ggml_scale(ctx0, Qcur, 1.0f/sqrtf(float(n_embd_head))); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + attn_norm_output = ggml_get_rows(ctx0, attn_norm_output, inp_out_ids); + } + // FF + { + ffn_output = build_ffn(attn_norm_output, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(ffn_output, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_output); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output_no_bias", -1); + + cur = ggml_add(ctx0, cur, model.output_b); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/phi3.cpp b/examples/talk-llama/models/phi3.cpp new file mode 100644 index 00000000..c8e5da33 --- /dev/null +++ b/examples/talk-llama/models/phi3.cpp @@ -0,0 +1,152 @@ +#include "models.h" + +template +llm_build_phi3::llm_build_phi3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + using inp_attn_type = std::conditional_t; + inp_attn_type * inp_attn = nullptr; + + if constexpr (iswa) { + inp_attn = build_attn_inp_kv_iswa(); + } else { + inp_attn = build_attn_inp_kv(); + } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + auto * residual = inpL; + + // self-attention + { + // rope freq factors for 128k context + ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); + + ggml_tensor* attn_norm_output = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM_RMS, il); + cb(attn_norm_output, "attn_norm", il); + + ggml_tensor * Qcur = nullptr; + ggml_tensor * Kcur = nullptr; + ggml_tensor * Vcur = nullptr; + + if (model.layers[il].wqkv) { + cur = build_lora_mm(model.layers[il].wqkv, attn_norm_output); + cb(cur, "wqkv", il); + + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 0 * sizeof(float) * (n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd)); + Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)); + } + else { + Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, attn_norm_output), model.layers[il].bq); + Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, attn_norm_output), model.layers[il].bk); + Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, attn_norm_output), model.layers[il].bv); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + } + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, rope_factors, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head))); + cb(Qcur, "Qcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + } + cur = ggml_add(ctx0, cur, residual); + residual = cur; + + cur = build_norm(cur, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // feed-forward network + if (model.layers[il].ffn_gate_inp == nullptr) { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } else { + // MoE branch + cur = build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(cur, "ffn_moe_out", il); + } + cur = ggml_add(ctx0, residual, cur); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + if (model.output_b != nullptr) { + cb(cur, "result_output_no_bias", -1); + cur = ggml_add(ctx0, cur, model.output_b); + } + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +// Explicit template instantiations +template struct llm_build_phi3; +template struct llm_build_phi3; diff --git a/examples/talk-llama/models/plamo.cpp b/examples/talk-llama/models/plamo.cpp new file mode 100644 index 00000000..04ff709f --- /dev/null +++ b/examples/talk-llama/models/plamo.cpp @@ -0,0 +1,110 @@ +#include "models.h" + +llm_build_plamo::llm_build_plamo(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + ggml_tensor * sa_inp = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_embd_head, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + sa_inp = ggml_get_rows(ctx0, sa_inp, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + ggml_tensor * sa_out = cur; + + cur = sa_inp; + + // feed-forward network + { + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, sa_out); + cur = ggml_add(ctx0, cur, inpL); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/plamo2.cpp b/examples/talk-llama/models/plamo2.cpp new file mode 100644 index 00000000..31115a08 --- /dev/null +++ b/examples/talk-llama/models/plamo2.cpp @@ -0,0 +1,316 @@ +#include "models.h" + +llm_build_plamo2::llm_build_plamo2(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + cb(inpL, "embedding_output", -1); + + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_hybrid = build_inp_mem_hybrid(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * residual = inpL; + + // ggml_graph_add_node(gf, model.layers[il].attn_norm); + // cb(model.layers[il].attn_norm, "attn_norm", il); + + // pre_mixer_norm + cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + + // check if this layer is Mamba or Attention + bool is_mamba_layer = hparams.is_recurrent(il); + + if (is_mamba_layer) { + // PLaMo-2 Mamba layer + cur = build_plamo2_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il); + } else { + // PLaMo-2 Attention layer + cur = build_plamo2_attn_layer(inp_hybrid->get_attn(), inp_pos, cur, model, il); + } + + // post_mixer_norm + cur = build_norm(cur, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + // residual connection + cur = ggml_add(ctx0, cur, residual); + cb(cur, "attn_residual", il); + residual = cur; + + // pre-ffn norm + cur = build_norm(cur, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_pre_norm", il); + + // feed-forward network + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, LLM_FFN_SWIGLU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + // post ffn norm + cur = build_norm(cur, model.layers[il].ffn_post_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_post_norm", il); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + residual = ggml_get_rows(ctx0, residual, inp_out_ids); + } + + // residual connection + cur = ggml_add(ctx0, cur, residual); + cb(cur, "ffn_residual", il); + + inpL = cur; + } + + cur = inpL; + + // final norm + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "result_norm", -1); + + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output", -1); + + // Explicitly mark as output tensor to ensure proper backend assignment + ggml_set_output(cur); + + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +ggml_tensor * llm_build_plamo2::build_plamo2_attn_layer(llm_graph_input_attn_kv * inp, + ggml_tensor * inp_pos, + ggml_tensor * cur, + const llama_model & model, + int il) { + // self-attention + { + // PLaMo-2 uses combined QKV tensor + ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur); + cb(qkv, "wqkv", il); + + // split QKV tensor into Q, K, V + const int64_t n_embd_head_q = hparams.n_embd_head_k; + const int64_t n_embd_head_k = hparams.n_embd_head_k; + const int64_t n_embd_head_v = hparams.n_embd_head_v; + int32_t n_head = hparams.n_head(il); + int32_t n_head_kv = hparams.n_head_kv(il); + + const int64_t q_offset = 0; + const int64_t k_offset = n_embd_head_q * n_head; + const int64_t v_offset = k_offset + n_embd_head_k * n_head_kv; + + ggml_tensor * Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_q, n_head, n_tokens, n_embd_head_q * sizeof(float), + qkv->nb[1], q_offset * ggml_element_size(qkv)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv, n_tokens, n_embd_head_k * sizeof(float), + qkv->nb[1], k_offset * ggml_element_size(qkv)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv, n_tokens, n_embd_head_v * sizeof(float), + qkv->nb[1], v_offset * ggml_element_size(qkv)); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + cur = build_attn(inp, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f / sqrtf(float(n_embd_head_v)), il); + } + + cb(cur, "attn_out", il); + + return cur; +} + +ggml_tensor * llm_build_plamo2::build_plamo2_mamba_layer(llm_graph_input_rs * inp, + ggml_tensor * cur, + const llama_model & model, + const llama_ubatch & ubatch, + int il) { + const auto * mctx_cur = inp->mctx; + + const auto kv_head = mctx_cur->get_head(); + + const int64_t d_conv = hparams.ssm_d_conv; + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t d_state = hparams.ssm_d_state; + const int64_t n_heads = hparams.ssm_dt_rank; + const int64_t head_dim = d_inner / n_heads; + const int64_t n_group = hparams.ssm_n_group; + const int64_t n_seqs = ubatch.n_seqs; + + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + ggml_tensor * conv = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); + conv = ggml_reshape_3d(ctx0, conv, d_conv - 1, d_inner + 2 * n_group * d_state, n_seqs); + + // {n_embd, n_tokens} => {n_embd, n_seq_tokens, n_seqs} + cur = ggml_reshape_3d(ctx0, cur, cur->ne[0], n_seq_tokens, n_seqs); + + // in_proj: {n_embd, 2*d_inner} @ {n_embd, n_seq_tokens, n_seqs} => {2*d_inner, n_seq_tokens, n_seqs} + ggml_tensor * zx = build_lora_mm(model.layers[il].ssm_in, cur); + cb(zx, "mamba_in_proj", il); + // {8192, 5, 1, 1} -> {8192, 1, 5, 1} + zx = ggml_permute(ctx0, zx, 0, 2, 1, 3); + zx = ggml_cont_4d(ctx0, zx, head_dim * 2, n_heads, n_seq_tokens, n_seqs); + cb(zx, "mamba_in_proj_out", il); + + // split into z and x + // => {head_dim * n_heads, n_seq_tokens, n_seqs} + ggml_tensor * x = ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], + head_dim * ggml_element_size(zx)); + x = ggml_cont_3d(ctx0, x, head_dim * n_heads, n_seq_tokens, n_seqs); + // x = ggml_permute(ctx0, x, 0, 2, 1, 3); + cb(x, "mamba_x_split", il); + + ggml_tensor * z = + ggml_view_4d(ctx0, zx, head_dim, n_heads, n_seq_tokens, n_seqs, zx->nb[1], zx->nb[2], zx->nb[3], 0); + cb(z, "mamba_z_split", il); + + // conv1d + { + // => {d_conv - 1 + n_seq_tokens, d_inner, n_seqs} + ggml_tensor * conv_x = ggml_concat(ctx0, conv, ggml_transpose(ctx0, x), 0); + cb(conv_x, "mamba_conv1d_input", il); + + // copy last (d_conv - 1) columns back into the state cache + ggml_tensor * last_conv = ggml_view_3d(ctx0, conv_x, d_conv - 1, d_inner, n_seqs, conv_x->nb[1], conv_x->nb[2], + n_seq_tokens * (conv_x->nb[0])); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv, + ggml_view_1d(ctx0, conv_states_all, + (d_conv - 1) * (d_inner + 2 * n_group * d_state) * (n_seqs), + kv_head * (d_conv - 1) * (d_inner + 2 * n_group * d_state) * + ggml_element_size(conv_states_all)))); + cb(conv_states_all, "mamba_conv1d_state", il); + + // 1D convolution + x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d); + cb(x, "mamba_conv1d", il); + + x = ggml_silu(ctx0, x); + cb(x, "mamba_conv1d_silu", il); + } + + // SSM + { + // bcdt_proj: {d_inner, dt_rank + 2*d_state} @ {d_inner, n_seq_tokens, n_seqs} => {dt_rank + 2*d_state, n_seq_tokens, n_seqs} + ggml_tensor * x_bcdt = build_lora_mm(model.layers[il].ssm_x, x); + cb(x_bcdt, "mamba_bcdt_proj", il); + + // split into dt, B, C + const int64_t dt_dim = std::max(64, int(hparams.n_embd / 16)); + ggml_tensor * B = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], 0); + ggml_tensor * C = ggml_view_3d(ctx0, x_bcdt, d_state, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], + ggml_element_size(x_bcdt) * d_state); + ggml_tensor * dt = ggml_view_3d(ctx0, x_bcdt, dt_dim, n_seq_tokens, n_seqs, x_bcdt->nb[1], x_bcdt->nb[2], + ggml_element_size(x_bcdt) * (2 * d_state)); + cb(B, "mamba_B_raw", il); + cb(C, "mamba_C_raw", il); + cb(dt, "mamba_dt_raw", il); + + // Apply RMS norm to dt, B, C (PLaMo-2 specific) + B = build_norm(B, model.layers[il].ssm_b_norm, NULL, LLM_NORM_RMS, il); + C = build_norm(C, model.layers[il].ssm_c_norm, NULL, LLM_NORM_RMS, il); + dt = build_norm(dt, model.layers[il].ssm_dt_norm, NULL, LLM_NORM_RMS, il); + cb(B, "mamba_B_normed", il); + cb(C, "mamba_C_normed", il); + cb(dt, "mamba_dt_normed", il); + + // dt_proj: {dt_rank, d_inner} @ {dt_rank, n_seq_tokens, n_seqs} => {d_inner, n_seq_tokens, n_seqs} + dt = build_lora_mm(model.layers[il].ssm_dt, dt); + dt = ggml_add(ctx0, dt, model.layers[il].ssm_dt_b); + cb(dt, "mamba_dt_proj", il); + + ggml_tensor * A = ggml_reshape_2d(ctx0, model.layers[il].ssm_a, 1, n_heads); + cb(A, "mamba_A", il); + + x = ggml_view_4d(ctx0, x, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x), + head_dim * n_heads * ggml_element_size(x), + head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0); + B = ggml_view_4d(ctx0, B, d_state, 1, n_seq_tokens, n_seqs, d_state * B->nb[0], B->nb[1], B->nb[2], 0); + C = ggml_view_4d(ctx0, C, d_state, 1, n_seq_tokens, n_seqs, d_state * C->nb[0], C->nb[1], C->nb[2], 0); + + // use the states and the indices provided by build_recurrent_state + // (this is necessary in order to properly use the states before they are overwritten, + // while avoiding to make unnecessary copies of the states) + auto get_ssm_rows = [&](ggml_context * ctx, ggml_tensor * states, ggml_tensor * ids) { + ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_heads, mctx_cur->get_size()); + + // Custom operator to optimize the parallel associative scan + // as described in the Annex D of the Mamba paper. + // => {d_inner, n_seq_tokens, n_seqs} and {d_state, d_inner, n_seqs} + return ggml_ssm_scan(ctx, ssm, x, dt, A, B, C, ids); + }; + + ggml_tensor * y_ssm = build_rs(inp, ssm_states_all, hparams.n_embd_s(), ubatch.n_seqs, get_ssm_rows); + cb(y_ssm, "mamba_ssm_scan", il); + + // store last states + ggml_build_forward_expand( + gf, ggml_cpy( + ctx0, + ggml_view_1d(ctx0, y_ssm, n_heads * head_dim * d_state * n_seqs, + n_heads * head_dim * n_seq_tokens * n_seqs * ggml_element_size(y_ssm)), + ggml_view_1d(ctx0, ssm_states_all, n_heads * head_dim * d_state * n_seqs, + kv_head * n_seqs * n_heads * head_dim * d_state * ggml_element_size(ssm_states_all)))); + cb(ssm_states_all, "mamba_ssm_states", il); + + ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_heads, n_seq_tokens, n_seqs, + head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x), + head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0); + cb(y, "mamba_y_view", il); + + // Add D parameter and apply gating with z + // {d_inner, n_seq_tokens, n_seqs} * {d_inner} => {d_inner, n_seq_tokens, n_seqs} + ggml_tensor * D = ggml_reshape_2d(ctx0, model.layers[il].ssm_d, 1, n_heads); + y = ggml_add(ctx0, y, ggml_mul(ctx0, x, D)); + cb(y, "mamba_y_add_d", il); + + y = ggml_swiglu_split(ctx0, ggml_cont(ctx0, z), y); + cb(y, "mamba_y_swiglu_z", il); + + // out_proj: {d_inner, n_embd} @ {d_inner, n_seq_tokens, n_seqs} => {n_embd, n_seq_tokens, n_seqs} + y = ggml_view_3d(ctx0, y, head_dim * n_heads, n_seq_tokens, n_seqs, y->nb[2], y->nb[3], 0); + cur = build_lora_mm(model.layers[il].ssm_out, y); + cb(cur, "mamba_out_proj", il); + } + + // {n_embd, n_seq_tokens, n_seqs} => {n_embd, n_tokens} + cur = ggml_reshape_2d(ctx0, cur, cur->ne[0], n_seq_tokens * n_seqs); + cb(cur, "mamba_out", il); + + return cur; +} diff --git a/examples/talk-llama/models/plm.cpp b/examples/talk-llama/models/plm.cpp new file mode 100644 index 00000000..481cbba6 --- /dev/null +++ b/examples/talk-llama/models/plm.cpp @@ -0,0 +1,168 @@ +#include "models.h" + +llm_build_plm::llm_build_plm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const float kq_scale = 1.0f/sqrtf(float(hparams.n_embd_head_k)); + + const uint32_t n_embd_head_qk_rope = hparams.n_rot; + const uint32_t n_embd_head_qk_nope = hparams.n_embd_head_k - hparams.n_rot; + const uint32_t kv_lora_rank = hparams.n_lora_kv; + + ggml_tensor * cur; + ggml_tensor * inpL; + + // {n_embd, n_tokens} + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + ggml_tensor * q = NULL; + q = ggml_mul_mat(ctx0, model.layers[il].wq, cur); + cb(q, "q", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + 0); + cb(q_nope, "q_nope", il); + + // and {n_head * n_embd_head_qk_rope, n_tokens} + ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, + ggml_row_size(q->type, hparams.n_embd_head_k), + ggml_row_size(q->type, hparams.n_embd_head_k * n_head), + ggml_row_size(q->type, n_embd_head_qk_nope)); + cb(q_pe, "q_pe", il); + + // {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens} + ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur); + cb(kv_pe_compresseed, "kv_pe_compresseed", il); + + // split into {kv_lora_rank, n_tokens} + ggml_tensor * kv_compressed = ggml_view_2d(ctx0, kv_pe_compresseed, kv_lora_rank, n_tokens, + kv_pe_compresseed->nb[1], + 0); + cb(kv_compressed, "kv_compressed", il); + + // and {n_embd_head_qk_rope, n_tokens} + ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_pe_compresseed, n_embd_head_qk_rope, 1, n_tokens, + kv_pe_compresseed->nb[1], + kv_pe_compresseed->nb[1], + ggml_row_size(kv_pe_compresseed->type, kv_lora_rank)); + cb(k_pe, "k_pe", il); + + kv_compressed = build_norm(kv_compressed, + model.layers[il].attn_kv_a_norm, NULL, + LLM_NORM_RMS, il); + cb(kv_compressed, "kv_compressed", il); + + // {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)} * {kv_lora_rank, n_tokens} -> {n_head * (n_embd_head_qk_nope + n_embd_head_v), n_tokens} + ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_compressed); + cb(kv, "kv", il); + + // split into {n_head * n_embd_head_qk_nope, n_tokens} + ggml_tensor * k_nope = ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens, + ggml_row_size(kv->type, n_embd_head_qk_nope + hparams.n_embd_head_v), + ggml_row_size(kv->type, n_head * (n_embd_head_qk_nope + hparams.n_embd_head_v)), + 0); + cb(k_nope, "k_nope", il); + + // and {n_head * n_embd_head_v, n_tokens} + ggml_tensor * v_states = ggml_view_3d(ctx0, kv, hparams.n_embd_head_v, n_head, n_tokens, + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)), + ggml_row_size(kv->type, (n_embd_head_qk_nope + hparams.n_embd_head_v)*n_head), + ggml_row_size(kv->type, (n_embd_head_qk_nope))); + cb(v_states, "v_states", il); + + v_states = ggml_cont(ctx0, v_states); + cb(v_states, "v_states", il); + + v_states = ggml_view_2d(ctx0, v_states, hparams.n_embd_head_v * n_head, n_tokens, + ggml_row_size(kv->type, hparams.n_embd_head_v * n_head), + 0); + cb(v_states, "v_states", il); + + q_pe = ggml_rope_ext( + ctx0, q_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(q_pe, "q_pe", il); + + // shared RoPE key + k_pe = ggml_rope_ext( + ctx0, k_pe, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + cb(k_pe, "k_pe", il); + + ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0); + cb(q_states, "q_states", il); + + ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0); + cb(k_states, "k_states", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + q_states, k_states, v_states, nullptr, nullptr, nullptr, kq_scale, il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_RELU_SQR, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/qwen.cpp b/examples/talk-llama/models/qwen.cpp new file mode 100644 index 00000000..31fd9b73 --- /dev/null +++ b/examples/talk-llama/models/qwen.cpp @@ -0,0 +1,108 @@ +#include "models.h" + + +llm_build_qwen::llm_build_qwen(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 2*sizeof(float)*(n_embd)); + + // using mode = 2 for neox mode + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward forward + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/qwen2.cpp b/examples/talk-llama/models/qwen2.cpp new file mode 100644 index 00000000..587a9324 --- /dev/null +++ b/examples/talk-llama/models/qwen2.cpp @@ -0,0 +1,117 @@ +#include "models.h" + +llm_build_qwen2::llm_build_qwen2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + if (model.output_b != nullptr) { + cur = ggml_add(ctx0, cur, model.output_b); + } + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/qwen2moe.cpp b/examples/talk-llama/models/qwen2moe.cpp new file mode 100644 index 00000000..49142b71 --- /dev/null +++ b/examples/talk-llama/models/qwen2moe.cpp @@ -0,0 +1,151 @@ +#include "models.h" + +llm_build_qwen2moe::llm_build_qwen2moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, false, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + + // FFN shared expert + { + ggml_tensor * cur_gate_inp = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); + cb(cur_gate_inp, "ffn_shexp_gate_inp", il); + + // sigmoid + ggml_tensor * cur_gate = ggml_div(ctx0, ggml_silu(ctx0, cur_gate_inp), cur_gate_inp); + cb(cur_gate, "ffn_shexp_gate", il); + + ggml_tensor * cur_ffn = build_ffn(cur, + model.layers[il].ffn_up_shexp, NULL, NULL, + model.layers[il].ffn_gate_shexp, NULL, NULL, + model.layers[il].ffn_down_shexp, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur_ffn, "ffn_shexp", il); + + ggml_tensor * ffn_shexp_out = ggml_mul(ctx0, cur_ffn, cur_gate); + cb(ffn_shexp_out, "ffn_shexp_out", il); + + moe_out = ggml_add(ctx0, moe_out, ffn_shexp_out); + cb(moe_out, "ffn_out", il); + + cur = moe_out; + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/qwen2vl.cpp b/examples/talk-llama/models/qwen2vl.cpp new file mode 100644 index 00000000..9be38675 --- /dev/null +++ b/examples/talk-llama/models/qwen2vl.cpp @@ -0,0 +1,117 @@ +#include "models.h" + +llm_build_qwen2vl::llm_build_qwen2vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_multi( + ctx0, Qcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_multi( + ctx0, Kcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/qwen3.cpp b/examples/talk-llama/models/qwen3.cpp new file mode 100644 index 00000000..a5cfffa5 --- /dev/null +++ b/examples/talk-llama/models/qwen3.cpp @@ -0,0 +1,117 @@ +#include "models.h" + +llm_build_qwen3::llm_build_qwen3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/qwen3moe.cpp b/examples/talk-llama/models/qwen3moe.cpp new file mode 100644 index 00000000..888534fb --- /dev/null +++ b/examples/talk-llama/models/qwen3moe.cpp @@ -0,0 +1,124 @@ +#include "models.h" + +llm_build_qwen3moe::llm_build_qwen3moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + cur = moe_out; + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/qwen3vl-moe.cpp b/examples/talk-llama/models/qwen3vl-moe.cpp new file mode 100644 index 00000000..f72f80a8 --- /dev/null +++ b/examples/talk-llama/models/qwen3vl-moe.cpp @@ -0,0 +1,149 @@ +#include "models.h" + +llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const size_t n_deepstack_layers = hparams.n_deepstack_layers; + const int64_t n_embd = hparams.n_embd; + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + std::vector deepstack_features(n_deepstack_layers, nullptr); + + if (ubatch.embd) { + // Image input: split main embd and deepstack embds + ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0); + for (size_t i = 0; i < n_deepstack_layers; i++) { + deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float)); + } + inpL = inpL_main; + } + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_multi( + ctx0, Qcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_multi( + ctx0, Kcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * moe_out = + build_moe_ffn(cur, + model.layers[il].ffn_gate_inp, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_SILU, true, + false, 0.0, + LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, + il); + cb(moe_out, "ffn_moe_out", il); + cur = moe_out; + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + if (ubatch.embd && (size_t)il < n_deepstack_layers) { + cur = ggml_add(ctx0, cur, deepstack_features[il]); + cb(cur, "deepstack_out", il); + } + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + diff --git a/examples/talk-llama/models/qwen3vl.cpp b/examples/talk-llama/models/qwen3vl.cpp new file mode 100644 index 00000000..0bae5223 --- /dev/null +++ b/examples/talk-llama/models/qwen3vl.cpp @@ -0,0 +1,141 @@ +#include "models.h" + +llm_build_qwen3vl::llm_build_qwen3vl(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const size_t n_deepstack_layers = hparams.n_deepstack_layers; + const int64_t n_embd = hparams.n_embd; + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + std::vector deepstack_features(n_deepstack_layers, nullptr); + + if (ubatch.embd) { + // Image input: split main embd and deepstack embds + ggml_tensor * inpL_main = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], 0); + for (size_t i = 0; i < n_deepstack_layers; i++) { + deepstack_features[i] = ggml_view_2d(ctx0, inpL, n_embd, n_tokens, inpL->nb[1], (i + 1) * n_embd * sizeof(float)); + } + inpL = inpL_main; + } + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); + + Qcur = ggml_rope_multi( + ctx0, Qcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + Kcur = ggml_rope_multi( + ctx0, Kcur, inp_pos, nullptr, + n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + if (ubatch.embd && (size_t)il < n_deepstack_layers) { + cur = ggml_add(ctx0, cur, deepstack_features[il]); + cb(cur, "deepstack_out", il); + } + + // input for next layer + inpL = cur; + } + + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/refact.cpp b/examples/talk-llama/models/refact.cpp new file mode 100644 index 00000000..ff5eb284 --- /dev/null +++ b/examples/talk-llama/models/refact.cpp @@ -0,0 +1,94 @@ +#include "models.h" + +llm_build_refact::llm_build_refact(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/rwkv6-base.cpp b/examples/talk-llama/models/rwkv6-base.cpp new file mode 100644 index 00000000..7beed2da --- /dev/null +++ b/examples/talk-llama/models/rwkv6-base.cpp @@ -0,0 +1,162 @@ +#include "models.h" + +llm_build_rwkv6_base::llm_build_rwkv6_base(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params), + model(model) {} + +ggml_tensor * llm_build_rwkv6_base::build_rwkv6_channel_mix(const llama_layer * layer, + ggml_tensor * cur, + ggml_tensor * x_prev, + llm_arch arch) const { + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + switch (arch) { + case LLM_ARCH_RWKV6: + { + ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur); + ggml_tensor * xr = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_r), cur); + + ggml_tensor * r = ggml_sigmoid(ctx0, build_lora_mm(layer->channel_mix_receptance, xr)); + ggml_tensor * k = ggml_sqr(ctx0, ggml_relu(ctx0, build_lora_mm(layer->channel_mix_key, xk))); + cur = ggml_mul(ctx0, r, build_lora_mm(layer->channel_mix_value, k)); + } + break; + default: + GGML_ABORT("fatal error"); + } + return cur; +} + +ggml_tensor * llm_build_rwkv6_base::build_rwkv6_time_mix(llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * x_prev, + const llama_ubatch & ubatch, + int il) const { + const auto * mctx_cur = static_cast(mctx); + + const auto n_tokens = ubatch.n_tokens; + const auto n_seqs = ubatch.n_seqs; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_embd = hparams.n_embd; + const auto head_size = hparams.wkv_head_size; + const auto n_head = n_embd / head_size; + const auto n_head_kv = hparams.n_head_kv(il); + + const auto kv_head = mctx_cur->get_head(); + + const auto & layer = model.layers[il]; + + bool is_qrwkv = layer.time_mix_first == nullptr; + + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + + sx = ggml_reshape_2d(ctx0, sx, n_embd, n_tokens); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + + ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_x), cur); + + xxx = ggml_reshape_4d(ctx0, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_w1, xxx)), + layer.time_mix_w1->ne[1] / 5, 1, 5, n_tokens); + + xxx = ggml_cont(ctx0, ggml_permute(ctx0, xxx, 0, 1, 3, 2)); + + xxx = ggml_mul_mat( + ctx0, ggml_reshape_4d(ctx0, layer.time_mix_w2, layer.time_mix_w2->ne[0], layer.time_mix_w2->ne[1], 1, 5), xxx); + + ggml_tensor *xw, *xk, *xv, *xr, *xg; + if (layer.time_mix_lerp_fused) { + // fusing these weights makes some performance improvement + sx = ggml_reshape_3d(ctx0, sx, n_embd, 1, n_tokens); + cur = ggml_reshape_3d(ctx0, cur, n_embd, 1, n_tokens); + xxx = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xxx, layer.time_mix_lerp_fused), sx), cur); + xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); + xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + } else { + // for backward compatibility + xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); + xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + xg = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + + xw = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xw, layer.time_mix_lerp_w), sx), cur); + xk = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xk, layer.time_mix_lerp_k), sx), cur); + xv = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xv, layer.time_mix_lerp_v), sx), cur); + xr = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xr, layer.time_mix_lerp_r), sx), cur); + xg = ggml_add(ctx0, ggml_mul(ctx0, ggml_add(ctx0, xg, layer.time_mix_lerp_g), sx), cur); + } + ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr); + ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk); + ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv); + if (layer.time_mix_receptance_b) { + r = ggml_add(ctx0, r, layer.time_mix_receptance_b); + } + if (layer.time_mix_key_b) { + k = ggml_add(ctx0, k, layer.time_mix_key_b); + } + if (layer.time_mix_value_b) { + v = ggml_add(ctx0, v, layer.time_mix_value_b); + } + ggml_tensor * g = build_lora_mm(layer.time_mix_gate, xg); + if (is_qrwkv) { + g = ggml_sigmoid(ctx0, g); + } else { + g = ggml_silu(ctx0, g); + } + if (n_head_kv != 0 && n_head_kv != n_head) { + GGML_ASSERT(n_head % n_head_kv == 0); + k = ggml_reshape_4d(ctx0, k, head_size, 1, n_head_kv, n_tokens); + v = ggml_reshape_4d(ctx0, v, head_size, 1, n_head_kv, n_tokens); + ggml_tensor * tmp = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, head_size, n_head / n_head_kv, n_head_kv, n_tokens); + k = ggml_repeat(ctx0, k, tmp); + v = ggml_repeat(ctx0, v, tmp); + } + k = ggml_reshape_3d(ctx0, k, head_size, n_head, n_tokens); + v = ggml_reshape_3d(ctx0, v, head_size, n_head, n_tokens); + r = ggml_reshape_3d(ctx0, r, head_size, n_head, n_tokens); + + ggml_tensor * w = + ggml_mul_mat(ctx0, layer.time_mix_decay_w2, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_decay_w1, xw))); + + w = ggml_add(ctx0, w, layer.time_mix_decay); + w = ggml_exp(ctx0, ggml_neg(ctx0, ggml_exp(ctx0, w))); + w = ggml_reshape_3d(ctx0, w, head_size, n_head, n_tokens); + + if (is_qrwkv) { + // k = k * (1 - w) + k = ggml_sub(ctx0, k, ggml_mul(ctx0, k, w)); + } + ggml_tensor * wkv_state = build_rs(inp, mctx_cur->get_s_l(il), hparams.n_embd_s(), n_seqs); + + ggml_tensor * wkv_output; + if (is_qrwkv) { + wkv_output = ggml_gated_linear_attn(ctx0, k, v, r, w, wkv_state, pow(head_size, -0.5f)); + } else { + wkv_output = ggml_rwkv_wkv6(ctx0, k, v, r, layer.time_mix_first, w, wkv_state); + } + cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0); + wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); + + ggml_build_forward_expand( + gf, ggml_cpy(ctx0, wkv_state, + ggml_view_1d(ctx0, mctx_cur->get_s_l(il), hparams.n_embd_s() * n_seqs, + hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il))))); + + if (!is_qrwkv) { + // group norm with head_count groups + cur = ggml_reshape_3d(ctx0, cur, n_embd / n_head, n_head, n_tokens); + cur = ggml_norm(ctx0, cur, 64e-5f); + + // Convert back to regular vectors. + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b); + } else { + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + } + cur = ggml_mul(ctx0, cur, g); + cur = build_lora_mm(layer.time_mix_output, cur); + + return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs); +} diff --git a/examples/talk-llama/models/rwkv6.cpp b/examples/talk-llama/models/rwkv6.cpp new file mode 100644 index 00000000..15453fbf --- /dev/null +++ b/examples/talk-llama/models/rwkv6.cpp @@ -0,0 +1,94 @@ +#include "models.h" + +llm_build_rwkv6::llm_build_rwkv6(const llama_model & model, const llm_graph_params & params) : + llm_build_rwkv6_base(model, params) { + GGML_ASSERT(hparams.token_shift_count == 2); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); + + auto * rs_inp = build_rs_inp(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + + ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il); + + ggml_tensor * att_shift = + ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); + ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], + token_shift->nb[2], n_embd * ggml_element_size(token_shift)); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, att_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), 1); + + cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il); + cb(ffn_norm, "ffn_norm", il); + + x_prev = ggml_concat( + ctx0, ffn_shift, + ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), 1); + + token_shift = ggml_concat(ctx0, + ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], + (n_seq_tokens - 1) * n_embd * ggml_element_size(att_norm)), + ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], + (n_seq_tokens - 1) * n_embd * ggml_element_size(ffn_norm)), + 1); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); + x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); + x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + } + cur = build_rwkv6_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV6); + cur = ggml_add(ctx0, cur, ffn_inp); + + if (hparams.rescale_every_n_layers != 0 && (il + 1) % hparams.rescale_every_n_layers == 0) { + cur = ggml_scale(ctx0, cur, 0.5F); + } + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/rwkv6qwen2.cpp b/examples/talk-llama/models/rwkv6qwen2.cpp new file mode 100644 index 00000000..e84e5973 --- /dev/null +++ b/examples/talk-llama/models/rwkv6qwen2.cpp @@ -0,0 +1,86 @@ +#include "models.h" + +llm_build_rwkv6qwen2::llm_build_rwkv6qwen2(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv6_base(model, params) { + GGML_ASSERT(n_embd == hparams.n_embd_r()); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + auto * rs_inp = build_rs_inp(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + + ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM_RMS, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, + token_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), + 1 + ); + + cur = build_rwkv6_time_mix(rs_inp, att_norm, x_prev, ubatch, il); + + token_shift = ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], (n_seq_tokens-1)*n_embd*ggml_element_size(att_norm)); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + } + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + + cur = inpL; + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/rwkv7-base.cpp b/examples/talk-llama/models/rwkv7-base.cpp new file mode 100644 index 00000000..cda44653 --- /dev/null +++ b/examples/talk-llama/models/rwkv7-base.cpp @@ -0,0 +1,135 @@ +#include "models.h" + +llm_build_rwkv7_base::llm_build_rwkv7_base(const llama_model & model, const llm_graph_params & params) : + llm_graph_context(params), + model(model) {} + +ggml_tensor * llm_build_rwkv7_base::build_rwkv7_channel_mix(const llama_layer * layer, + ggml_tensor * cur, + ggml_tensor * x_prev, + llm_arch arch) const { + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + switch (arch) { + case LLM_ARCH_RWKV7: + { + ggml_tensor * xk = ggml_add(ctx0, ggml_mul(ctx0, sx, layer->channel_mix_lerp_k), cur); + + ggml_tensor * k = ggml_sqr(ctx0, ggml_relu(ctx0, build_lora_mm(layer->channel_mix_key, xk))); + + cur = build_lora_mm(layer->channel_mix_value, k); + } + break; + default: + GGML_ABORT("fatal error"); + } + return cur; +} + +ggml_tensor * llm_build_rwkv7_base::build_rwkv7_time_mix(llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * x_prev, + ggml_tensor *& first_layer_value, + const llama_ubatch & ubatch, + int il) const { + const auto * mctx_cur = static_cast(mctx); + + const auto n_tokens = ubatch.n_tokens; + const auto n_seqs = ubatch.n_seqs; + const auto n_embd = hparams.n_embd; + const auto head_size = hparams.wkv_head_size; + const auto head_count = n_embd / head_size; + const auto n_seq_tokens = ubatch.n_seq_tokens; + + const auto kv_head = mctx_cur->get_head(); + + const auto & layer = model.layers[il]; + + bool has_gating = layer.time_mix_g1 && layer.time_mix_g2; + + ggml_tensor * sx = ggml_sub(ctx0, x_prev, cur); + ggml_tensor * dummy = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, n_embd, n_seq_tokens, n_seqs, has_gating ? 6 : 5); + sx = ggml_repeat(ctx0, sx, dummy); + + ggml_tensor * xxx = ggml_add(ctx0, ggml_mul(ctx0, sx, layer.time_mix_lerp_fused), cur); + + ggml_tensor * xr = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], 0); + ggml_tensor * xw = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * sizeof(float)); + ggml_tensor * xk = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 2 * sizeof(float)); + ggml_tensor * xv = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 3 * sizeof(float)); + ggml_tensor * xa = ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 4 * sizeof(float)); + ggml_tensor * xg = + has_gating ? ggml_view_2d(ctx0, xxx, n_embd, n_tokens, xxx->nb[1], n_embd * n_tokens * 5 * sizeof(float)) : + nullptr; + + ggml_tensor * r = build_lora_mm(layer.time_mix_receptance, xr); + ggml_tensor * w = ggml_add( + ctx0, ggml_mul_mat(ctx0, layer.time_mix_w2, ggml_tanh(ctx0, ggml_mul_mat(ctx0, layer.time_mix_w1, xw))), + layer.time_mix_w0); + w = ggml_exp(ctx0, ggml_scale(ctx0, ggml_sigmoid(ctx0, w), -0.606531)); + + ggml_tensor * k = build_lora_mm(layer.time_mix_key, xk); + ggml_tensor * v = build_lora_mm(layer.time_mix_value, xv); + if (first_layer_value == nullptr) { + first_layer_value = v; + } else { + // Add the first layer value as a residual connection. + v = ggml_add(ctx0, v, + ggml_mul(ctx0, ggml_sub(ctx0, first_layer_value, v), + ggml_sigmoid(ctx0, ggml_add(ctx0, + ggml_mul_mat(ctx0, layer.time_mix_v2, + ggml_mul_mat(ctx0, layer.time_mix_v1, xv)), + layer.time_mix_v0)))); + } + ggml_tensor * g = nullptr; + if (layer.time_mix_g1 && layer.time_mix_g2) { + g = ggml_mul_mat(ctx0, layer.time_mix_g2, ggml_sigmoid(ctx0, ggml_mul_mat(ctx0, layer.time_mix_g1, xg))); + } + ggml_tensor * a = ggml_sigmoid( + ctx0, ggml_add(ctx0, ggml_mul_mat(ctx0, layer.time_mix_a2, ggml_mul_mat(ctx0, layer.time_mix_a1, xa)), + layer.time_mix_a0)); + + ggml_tensor * kk = ggml_reshape_3d(ctx0, ggml_mul(ctx0, k, layer.time_mix_k_k), head_size, head_count, n_tokens); + kk = ggml_l2_norm(ctx0, kk, 1e-12); + + ggml_tensor * ka = ggml_mul(ctx0, k, layer.time_mix_k_a); + k = ggml_add(ctx0, k, ggml_sub(ctx0, ggml_mul(ctx0, a, ka), ka)); + + r = ggml_reshape_3d(ctx0, r, head_size, head_count, n_tokens); + w = ggml_reshape_3d(ctx0, w, head_size, head_count, n_tokens); + k = ggml_reshape_3d(ctx0, k, head_size, head_count, n_tokens); + v = ggml_reshape_3d(ctx0, v, head_size, head_count, n_tokens); + a = ggml_reshape_3d(ctx0, a, head_size, head_count, n_tokens); + + ggml_tensor * wkv_state = build_rs(inp, mctx_cur->get_s_l(il), hparams.n_embd_s(), n_seqs); + + ggml_tensor * wkv_output = ggml_rwkv_wkv7(ctx0, r, w, k, v, ggml_neg(ctx0, kk), ggml_mul(ctx0, kk, a), wkv_state); + cur = ggml_view_1d(ctx0, wkv_output, n_embd * n_tokens, 0); + wkv_state = ggml_view_1d(ctx0, wkv_output, n_embd * head_size * n_seqs, n_embd * n_tokens * sizeof(float)); + + ggml_build_forward_expand( + gf, ggml_cpy(ctx0, wkv_state, + ggml_view_1d(ctx0, mctx_cur->get_s_l(il), hparams.n_embd_s() * n_seqs, + hparams.n_embd_s() * kv_head * ggml_element_size(mctx_cur->get_s_l(il))))); + + if (layer.time_mix_ln && layer.time_mix_ln_b) { + // group norm with head_count groups + cur = ggml_reshape_3d(ctx0, cur, n_embd / head_count, head_count, n_tokens); + cur = ggml_norm(ctx0, cur, 64e-5f); + + // Convert back to regular vectors. + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + cur = ggml_add(ctx0, ggml_mul(ctx0, cur, layer.time_mix_ln), layer.time_mix_ln_b); + } else { + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_tokens); + } + ggml_tensor * rk = ggml_sum_rows( + ctx0, ggml_mul(ctx0, ggml_mul(ctx0, k, r), ggml_reshape_2d(ctx0, layer.time_mix_r_k, head_size, head_count))); + cur = ggml_add(ctx0, cur, ggml_reshape_2d(ctx0, ggml_mul(ctx0, v, rk), n_embd, n_tokens)); + + if (has_gating) { + cur = ggml_mul(ctx0, cur, g); + } + cur = build_lora_mm(layer.time_mix_output, cur); + + return ggml_reshape_3d(ctx0, cur, n_embd, n_seq_tokens, n_seqs); +} diff --git a/examples/talk-llama/models/rwkv7.cpp b/examples/talk-llama/models/rwkv7.cpp new file mode 100644 index 00000000..5caf6553 --- /dev/null +++ b/examples/talk-llama/models/rwkv7.cpp @@ -0,0 +1,90 @@ +#include "models.h" + +llm_build_rwkv7::llm_build_rwkv7(const llama_model & model, const llm_graph_params & params) : + llm_build_rwkv7_base(model, params) { + GGML_ASSERT(hparams.token_shift_count == 2); + + ggml_tensor * cur; + ggml_tensor * inpL; + ggml_tensor * v_first = nullptr; + + inpL = build_inp_embd(model.tok_embd); + inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1); + + auto * rs_inp = build_rs_inp(); + + const auto n_embd = hparams.n_embd; + const auto n_seq_tokens = ubatch.n_seq_tokens; + const auto n_seqs = ubatch.n_seqs; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + const llama_layer * layer = &model.layers[il]; + inpL = ggml_reshape_3d(ctx0, inpL, n_embd, n_seq_tokens, n_seqs); + + ggml_tensor * token_shift = build_rwkv_token_shift_load(rs_inp, ubatch, il); + + ggml_tensor * att_shift = + ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], token_shift->nb[2], 0); + ggml_tensor * ffn_shift = ggml_view_3d(ctx0, token_shift, n_embd, 1, n_seqs, token_shift->nb[1], + token_shift->nb[2], n_embd * ggml_element_size(token_shift)); + + ggml_tensor * att_norm = build_norm(inpL, layer->attn_norm, layer->attn_norm_b, LLM_NORM, il); + cb(att_norm, "attn_norm", il); + + ggml_tensor * x_prev = ggml_concat( + ctx0, att_shift, + ggml_view_3d(ctx0, att_norm, n_embd, n_seq_tokens - 1, n_seqs, att_norm->nb[1], att_norm->nb[2], 0), 1); + + cur = build_rwkv7_time_mix(rs_inp, att_norm, x_prev, v_first, ubatch, il); + + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + ggml_tensor * ffn_norm = build_norm(ffn_inp, layer->attn_norm_2, layer->attn_norm_2_b, LLM_NORM, il); + cb(ffn_norm, "ffn_norm", il); + + x_prev = ggml_concat( + ctx0, ffn_shift, + ggml_view_3d(ctx0, ffn_norm, n_embd, n_seq_tokens - 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], 0), 1); + + token_shift = ggml_concat(ctx0, + ggml_view_3d(ctx0, att_norm, n_embd, 1, n_seqs, att_norm->nb[1], att_norm->nb[2], + (n_seq_tokens - 1) * n_embd * ggml_element_size(att_norm)), + ggml_view_3d(ctx0, ffn_norm, n_embd, 1, n_seqs, ffn_norm->nb[1], ffn_norm->nb[2], + (n_seq_tokens - 1) * n_embd * ggml_element_size(ffn_norm)), + 1); + ggml_build_forward_expand(gf, build_rwkv_token_shift_store(token_shift, ubatch, il)); + + ffn_inp = ggml_reshape_2d(ctx0, ffn_inp, n_embd, n_tokens); + ffn_norm = ggml_reshape_2d(ctx0, ffn_norm, n_embd, n_tokens); + x_prev = ggml_reshape_2d(ctx0, x_prev, n_embd, n_tokens); + + if (il == n_layer - 1 && inp_out_ids) { + ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); + ffn_norm = ggml_get_rows(ctx0, ffn_norm, inp_out_ids); + x_prev = ggml_get_rows(ctx0, x_prev, inp_out_ids); + } + cur = build_rwkv7_channel_mix(layer, ffn_norm, x_prev, LLM_ARCH_RWKV7); + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + cur = build_norm(cur, model.output_norm, model.output_norm_b, LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/seed-oss.cpp b/examples/talk-llama/models/seed-oss.cpp new file mode 100644 index 00000000..0dc33c50 --- /dev/null +++ b/examples/talk-llama/models/seed-oss.cpp @@ -0,0 +1,124 @@ +#include "models.h" + +llm_build_seed_oss::llm_build_seed_oss(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + cur = build_norm(ffn_inp, + model.layers[il].attn_post_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_post_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/smallthinker.cpp b/examples/talk-llama/models/smallthinker.cpp new file mode 100644 index 00000000..277eec29 --- /dev/null +++ b/examples/talk-llama/models/smallthinker.cpp @@ -0,0 +1,120 @@ +#include "models.h" + +template +llm_build_smallthinker::llm_build_smallthinker(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params){ + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + using inp_attn_type = std::conditional_t; + inp_attn_type * inp_attn = nullptr; + + if constexpr (iswa) { + inp_attn = build_attn_inp_kv_iswa(); + } else { + inp_attn = build_attn_inp_kv(); + } + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + ggml_tensor * probs = nullptr; + + probs = build_lora_mm(model.layers[il].ffn_gate_inp, inpL); // [n_expert, n_tokens] + cb(probs, "ffn_moe_logits", il); + + // norm + cur = build_norm(inpL,model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self_attention + { + // compute Q and K and RoPE them + struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (hparams.n_no_rope_layer_step == n_layer || il % hparams.n_no_rope_layer_step != 0) { + Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + + Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow); + } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + probs = ggml_get_rows(ctx0, probs, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // MoE branch + cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + ggml_tensor * ffn_out = + build_moe_ffn(cur, + nullptr, + model.layers[il].ffn_up_exps, + model.layers[il].ffn_gate_exps, + model.layers[il].ffn_down_exps, + nullptr, + n_expert, n_expert_used, + LLM_FFN_RELU, true, + false, 0.0, + static_cast(hparams.expert_gating_func), + il, probs); + + cb(ffn_out, "ffn_out", il); + cur = ffn_out; + + cur = ggml_add(ctx0, cur, ffn_inp); + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} + +// Explicit template instantiations +template struct llm_build_smallthinker; +template struct llm_build_smallthinker; diff --git a/examples/talk-llama/models/smollm3.cpp b/examples/talk-llama/models/smollm3.cpp new file mode 100644 index 00000000..97c30dee --- /dev/null +++ b/examples/talk-llama/models/smollm3.cpp @@ -0,0 +1,128 @@ +#include "models.h" + +llm_build_smollm3::llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (use_rope) { + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + } + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/stablelm.cpp b/examples/talk-llama/models/stablelm.cpp new file mode 100644 index 00000000..bed1915c --- /dev/null +++ b/examples/talk-llama/models/stablelm.cpp @@ -0,0 +1,146 @@ +#include "models.h" + +llm_build_stablelm::llm_build_stablelm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + ggml_tensor * inpSA = cur; + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + if (model.layers[il].attn_q_norm) { + Qcur = build_norm(Qcur, + model.layers[il].attn_q_norm, + NULL, + LLM_NORM, il); + cb(Qcur, "Qcur", il); + } + if (model.layers[il].attn_k_norm) { + Kcur = build_norm(Kcur, + model.layers[il].attn_k_norm, + NULL, + LLM_NORM, il); + cb(Kcur, "Kcur", il); + } + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + if (model.layers[il].ffn_norm) { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + } else { + // parallel residual + cur = inpSA; + } + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/starcoder.cpp b/examples/talk-llama/models/starcoder.cpp new file mode 100644 index 00000000..e197af4a --- /dev/null +++ b/examples/talk-llama/models/starcoder.cpp @@ -0,0 +1,100 @@ +#include "models.h" + +llm_build_starcoder::llm_build_starcoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * pos = ggml_get_rows(ctx0, model.pos_embd, inp_pos); + cb(pos, "pos_embd", -1); + + inpL = ggml_add(ctx0, inpL, pos); + cb(inpL, "inpL", -1); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + cur = build_norm(inpL, + model.layers[il].attn_norm, + model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + cur = build_lora_mm(model.layers[il].wqkv, cur); + cb(cur, "wqkv", il); + + cur = ggml_add(ctx0, cur, model.layers[il].bqkv); + cb(cur, "bqkv", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); + ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); + ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); + } + // add the input + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL); + cb(ffn_inp, "ffn_inp", il); + + // FF + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, + model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = build_norm(inpL, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/starcoder2.cpp b/examples/talk-llama/models/starcoder2.cpp new file mode 100644 index 00000000..e40ef2cb --- /dev/null +++ b/examples/talk-llama/models/starcoder2.cpp @@ -0,0 +1,121 @@ +#include "models.h" + +llm_build_starcoder2::llm_build_starcoder2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, model.layers[il].attn_norm_b, + LLM_NORM, il); + cb(cur, "attn_norm", il); + + // self-attention + { + // compute Q and K and RoPE them + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + if (model.layers[il].bq) { + Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); + cb(Qcur, "Qcur", il); + } + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + if (model.layers[il].bk) { + Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); + cb(Kcur, "Kcur", il); + } + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + if (model.layers[il].bv) { + Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); + cb(Vcur, "Vcur", il); + } + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, model.layers[il].ffn_norm_b, + LLM_NORM, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL, + NULL, NULL, NULL, + model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + cb(cur, "ffn_out", il); + + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, + model.output_norm, model.output_norm_b, + LLM_NORM, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/t5-dec.cpp b/examples/talk-llama/models/t5-dec.cpp new file mode 100644 index 00000000..297e450d --- /dev/null +++ b/examples/talk-llama/models/t5-dec.cpp @@ -0,0 +1,166 @@ +#include "models.h" + +llm_build_t5_dec::llm_build_t5_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + //const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * embd_enc = build_inp_cross_embd(); + ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec(); + + const int64_t n_outputs_enc = embd_enc->ne[1]; + + auto * inp_attn_self = build_attn_inp_kv(); + auto * inp_attn_cross = build_attn_inp_cross(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + const int64_t dec_n_layer = hparams.dec_n_layer; + + for (int il = 0; il < dec_n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b; + ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b); + + cur = build_attn(inp_attn_self, + model.layers[il].wo, model.layers[il].bo, + Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il); + cb(cur, "kqv_out", il); + } + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "cross_inp", il); + + ggml_tensor * inpCA = cur; + + // norm + cur = build_norm(cur, + model.layers[il].attn_norm_cross, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm_cross", il); + + // cross-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc); + + cur = build_attn(inp_attn_cross, + model.layers[il].wo_cross, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il); + cb(cur, "kqv_out", il); + + //ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3); + //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3)); + + //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + //cb(kq, "kq", il); + + //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias); + //cb(kq, "kq_soft_max_ext", il); + + //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc))); + //cb(v, "v", il); + + //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq); + //cb(kqv, "kqv", il); + + //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3); + //cb(kqv_merged, "kqv_merged", il); + + //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens); + //cb(cur, "kqv_merged_cont", il); + + //ggml_build_forward_expand(gf, cur); + + //cur = build_lora_mm(model.layers[il].wo_cross, cur); + //cb(cur, "kqv_out", il); + } + if (il == dec_n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate ? LLM_FFN_PAR : LLM_FFN_SEQ, + il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + cb(cur, "result_embd", -1); + + cur = build_norm(cur, + model.output_norm, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/t5-enc.cpp b/examples/talk-llama/models/t5-enc.cpp new file mode 100644 index 00000000..70e1d80d --- /dev/null +++ b/examples/talk-llama/models/t5-enc.cpp @@ -0,0 +1,96 @@ +#include "models.h" + +llm_build_t5_enc::llm_build_t5_enc(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc(); + + auto * inp_attn = build_attn_inp_no_cache(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + // norm + cur = build_norm(inpL, + model.layers[il].attn_norm_enc, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; + ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b); + + cur = build_attn(inp_attn, + model.layers[il].wo_enc, nullptr, + Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il); + cb(cur, "kqv_out", il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm_enc, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + // T5 uses relu, flan-T5 uses gelu-gated + cur = build_ffn(cur, + model.layers[il].ffn_up_enc, NULL, NULL, + model.layers[il].ffn_gate_enc, NULL, NULL, + model.layers[il].ffn_down_enc, NULL, NULL, + NULL, + model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, + model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, + il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + cb(cur, "ffn_out", il); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + cb(cur, "result_embd", -1); + + cur = build_norm(cur, + model.output_norm_enc, NULL, + LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/wavtokenizer-dec.cpp b/examples/talk-llama/models/wavtokenizer-dec.cpp new file mode 100644 index 00000000..537a0d41 --- /dev/null +++ b/examples/talk-llama/models/wavtokenizer-dec.cpp @@ -0,0 +1,149 @@ +#include "models.h" + +llm_build_wavtokenizer_dec::llm_build_wavtokenizer_dec(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, inpL)); + + cur = ggml_conv_1d_ph(ctx0, model.conv1d, cur, 1, 1); + cur = ggml_add(ctx0, cur, model.conv1d_b); + + // posnet + for (uint32_t il = 0; il < hparams.posnet.n_layer; ++il) { + const auto & layer = model.layers[il].posnet; + + inpL = cur; + + switch (il) { + case 0: + case 1: + case 3: + case 4: + { + cur = build_norm(cur, + layer.norm1, + layer.norm1_b, + LLM_NORM_GROUP, 0); + + cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); + + cur = ggml_conv_1d_ph(ctx0, layer.conv1, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.conv1_b); + + cur = build_norm(cur, + layer.norm2, + layer.norm2_b, + LLM_NORM_GROUP, 0); + + cur = ggml_mul(ctx0, ggml_sigmoid(ctx0, cur), cur); + + cur = ggml_conv_1d_ph(ctx0, layer.conv2, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.conv2_b); + + cur = ggml_add(ctx0, cur, inpL); + } break; + case 2: + { + cur = build_norm(cur, + layer.attn_norm, + layer.attn_norm_b, + LLM_NORM_GROUP, 0); + + ggml_tensor * q; + ggml_tensor * k; + ggml_tensor * v; + + q = ggml_conv_1d_ph(ctx0, layer.attn_q, cur, 1, 1); + k = ggml_conv_1d_ph(ctx0, layer.attn_k, cur, 1, 1); + v = ggml_conv_1d_ph(ctx0, layer.attn_v, cur, 1, 1); + + q = ggml_add(ctx0, q, layer.attn_q_b); + k = ggml_add(ctx0, k, layer.attn_k_b); + v = ggml_add(ctx0, v, layer.attn_v_b); + + q = ggml_cont(ctx0, ggml_transpose(ctx0, q)); + k = ggml_cont(ctx0, ggml_transpose(ctx0, k)); + + ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + + kq = ggml_soft_max_ext(ctx0, kq, nullptr, 1.0f/sqrtf(float(hparams.posnet.n_embd)), 0.0f); + + cur = ggml_mul_mat(ctx0, kq, v); + + cur = ggml_conv_1d_ph(ctx0, layer.attn_o, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.attn_o_b); + + cur = ggml_add(ctx0, cur, inpL); + } break; + case 5: + { + cur = build_norm(cur, + layer.norm, + layer.norm_b, + LLM_NORM_GROUP, 0); + } break; + default: GGML_ABORT("unknown posnet layer"); + }; + } + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + model.tok_norm, + model.tok_norm_b, + LLM_NORM, -1); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + inpL = cur; + + // convnext + for (uint32_t il = 0; il < hparams.convnext.n_layer; ++il) { + const auto & layer = model.layers[il].convnext; + + cur = inpL; + + cur = ggml_conv_1d_dw_ph(ctx0, layer.dw, cur, 1, 1); + cur = ggml_add(ctx0, cur, layer.dw_b); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + layer.norm, + layer.norm_b, + LLM_NORM, -1); + + cur = build_ffn(cur, + layer.pw1, layer.pw1_b, NULL, + NULL, NULL, NULL, + layer.pw2, layer.pw2_b, NULL, + NULL, + LLM_FFN_GELU, LLM_FFN_SEQ, il); + + cur = ggml_mul(ctx0, cur, layer.gamma); + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + inpL = ggml_add(ctx0, cur, inpL); + } + cur = inpL; + + cur = ggml_cont(ctx0, ggml_transpose(ctx0, cur)); + + cur = build_norm(cur, + model.output_norm, + model.output_norm_b, + LLM_NORM, -1); + + // lm_head + cur = build_lora_mm(model.output, cur); + + cur = ggml_add(ctx0, cur, model.output_b); + + cb(cur, "result_embd", -1); + res->t_embd = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/examples/talk-llama/models/xverse.cpp b/examples/talk-llama/models/xverse.cpp new file mode 100644 index 00000000..364797dd --- /dev/null +++ b/examples/talk-llama/models/xverse.cpp @@ -0,0 +1,108 @@ +#include "models.h" + +llm_build_xverse::llm_build_xverse(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + GGML_ASSERT(n_embd_head == hparams.n_rot); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + // inp_pos - contains the positions + ggml_tensor * inp_pos = build_inp_pos(); + + auto * inp_attn = build_attn_inp_kv(); + + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + for (int il = 0; il < n_layer; ++il) { + ggml_tensor * inpSA = inpL; + + cur = build_norm(inpL, + model.layers[il].attn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "attn_norm", il); + + // self-attention + { + ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); + cb(Qcur, "Qcur", il); + + ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); + cb(Kcur, "Kcur", il); + + ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); + cb(Vcur, "Vcur", il); + + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + Qcur = ggml_rope_ext( + ctx0, Qcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + Kcur = ggml_rope_ext( + ctx0, Kcur, inp_pos, nullptr, + n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, + ext_factor, attn_factor, beta_fast, beta_slow + ); + + cb(Qcur, "Qcur", il); + cb(Kcur, "Kcur", il); + cb(Vcur, "Vcur", il); + + cur = build_attn(inp_attn, + model.layers[il].wo, NULL, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); + } + if (il == n_layer - 1 && inp_out_ids) { + cur = ggml_get_rows(ctx0, cur, inp_out_ids); + inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); + } + ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); + cb(ffn_inp, "ffn_inp", il); + + // feed-forward network + { + cur = build_norm(ffn_inp, + model.layers[il].ffn_norm, NULL, + LLM_NORM_RMS, il); + cb(cur, "ffn_norm", il); + + cur = build_ffn(cur, + model.layers[il].ffn_up, NULL, NULL, + model.layers[il].ffn_gate, NULL, NULL, + model.layers[il].ffn_down, NULL, NULL, + NULL, + LLM_FFN_SILU, LLM_FFN_PAR, il); + cb(cur, "ffn_out", il); + } + cur = ggml_add(ctx0, cur, ffn_inp); + + cur = build_cvec(cur, il); + cb(cur, "l_out", il); + + // input for next layer + inpL = cur; + } + cur = inpL; + + cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1); + + cb(cur, "result_norm", -1); + res->t_embd = cur; + + // lm_head + cur = build_lora_mm(model.output, cur); + + cb(cur, "result_output", -1); + res->t_logits = cur; + + ggml_build_forward_expand(gf, cur); +} diff --git a/scripts/sync-llama.sh b/scripts/sync-llama.sh index d5450bdd..b5428030 100755 --- a/scripts/sync-llama.sh +++ b/scripts/sync-llama.sh @@ -4,6 +4,7 @@ cp -rpv ../llama.cpp/include/llama.h ./examples/talk-llama/llama.h cp -rpv ../llama.cpp/src/llama*.cpp ./examples/talk-llama/ cp -rpv ../llama.cpp/src/llama*.h ./examples/talk-llama/ +cp -rpv ../llama.cpp/src/models/* ./examples/talk-llama/models/ cp -rpv ../llama.cpp/src/unicode.h ./examples/talk-llama/unicode.h cp -rpv ../llama.cpp/src/unicode.cpp ./examples/talk-llama/unicode.cpp cp -rpv ../llama.cpp/src/unicode-data.h ./examples/talk-llama/unicode-data.h