From: Georgi Gerganov Date: Sun, 15 Feb 2026 17:43:28 +0000 (+0200) Subject: talk-llama : sync llama.cpp X-Git-Tag: upstream/1.8.3+155 X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=1411e37786b6a08981e61de15e5dd6abd2c08062;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp talk-llama : sync llama.cpp --- diff --git a/examples/talk-llama/llama-arch.cpp b/examples/talk-llama/llama-arch.cpp index bd78f1e5..416c1746 100644 --- a/examples/talk-llama/llama-arch.cpp +++ b/examples/talk-llama/llama-arch.cpp @@ -37,6 +37,8 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_QWEN3NEXT, "qwen3next" }, { LLM_ARCH_QWEN3VL, "qwen3vl" }, { LLM_ARCH_QWEN3VLMOE, "qwen3vlmoe" }, + { LLM_ARCH_QWEN35, "qwen35" }, + { LLM_ARCH_QWEN35MOE, "qwen35moe" }, { LLM_ARCH_PHI2, "phi2" }, { LLM_ARCH_PHI3, "phi3" }, { LLM_ARCH_PHIMOE, "phimoe" }, @@ -72,6 +74,7 @@ static const std::map LLM_ARCH_NAMES = { { LLM_ARCH_CHATGLM, "chatglm" }, { LLM_ARCH_GLM4, "glm4" }, { LLM_ARCH_GLM4_MOE, "glm4moe" }, + { LLM_ARCH_GLM_DSA, "glm-dsa" }, { LLM_ARCH_BITNET, "bitnet" }, { LLM_ARCH_T5, "t5" }, { LLM_ARCH_T5ENCODER, "t5encoder" }, @@ -195,6 +198,7 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_EMBEDDING_SCALE, "%s.embedding_scale" }, { LLM_KV_TOKEN_SHIFT_COUNT, "%s.token_shift_count" }, { LLM_KV_INTERLEAVE_MOE_LAYER_STEP, "%s.interleave_moe_layer_step" }, + { LLM_KV_FULL_ATTENTION_INTERVAL, "%s.full_attention_interval" }, { LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" }, { LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" }, @@ -222,6 +226,9 @@ static const std::map LLM_KV_NAMES = { { LLM_KV_ATTENTION_TEMPERATURE_SCALE, "%s.attention.temperature_scale" }, { LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" }, { LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" }, + { LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, "%s.attention.indexer.head_count" }, + { LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, "%s.attention.indexer.key_length" }, + { LLM_KV_ATTENTION_INDEXER_TOP_K, "%s.attention.indexer.top_k" }, { LLM_KV_ROPE_DIMENSION_COUNT, "%s.rope.dimension_count" }, { LLM_KV_ROPE_DIMENSION_SECTIONS, "%s.rope.dimension_sections" }, @@ -366,6 +373,7 @@ static const std::map LLM_TENSOR_NAMES = { { LLM_TENSOR_SSM_CONV1D, "blk.%d.ssm_conv1d" }, { LLM_TENSOR_SSM_DT, "blk.%d.ssm_dt" }, { LLM_TENSOR_SSM_BETA_ALPHA, "blk.%d.ssm_ba" }, + { LLM_TENSOR_SSM_ALPHA, "blk.%d.ssm_alpha" }, { LLM_TENSOR_SSM_IN, "blk.%d.ssm_in" }, { LLM_TENSOR_SSM_NORM, "blk.%d.ssm_norm" }, { LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" }, @@ -512,6 +520,10 @@ static const std::map LLM_TENSOR_NAMES = { { 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_TENSOR_INDEXER_K_NORM, "blk.%d.indexer.k_norm" }, + { LLM_TENSOR_INDEXER_PROJ, "blk.%d.indexer.proj" }, + { LLM_TENSOR_INDEXER_ATTN_K, "blk.%d.indexer.attn_k" }, + { LLM_TENSOR_INDEXER_ATTN_Q_B, "blk.%d.indexer.attn_q_b" }, }; static std::set llm_get_tensor_names(llm_arch arch) { @@ -968,7 +980,6 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_ATTN_OUT, LLM_TENSOR_ATTN_QKV, LLM_TENSOR_ATTN_GATE, - LLM_TENSOR_FFN_NORM, LLM_TENSOR_FFN_GATE_INP, LLM_TENSOR_FFN_GATE_EXPS, LLM_TENSOR_FFN_DOWN_EXPS, @@ -985,6 +996,63 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_SSM_NORM, LLM_TENSOR_SSM_OUT, }; + case LLM_ARCH_QWEN35: + return { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_POST_NORM, + LLM_TENSOR_ATTN_Q, + LLM_TENSOR_ATTN_Q_NORM, + LLM_TENSOR_ATTN_K, + LLM_TENSOR_ATTN_K_NORM, + LLM_TENSOR_ATTN_V, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_ATTN_QKV, + LLM_TENSOR_ATTN_GATE, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_UP, + LLM_TENSOR_SSM_A_NOSCAN, + LLM_TENSOR_SSM_CONV1D, + LLM_TENSOR_SSM_DT, + LLM_TENSOR_SSM_BETA, + LLM_TENSOR_SSM_ALPHA, + LLM_TENSOR_SSM_NORM, + LLM_TENSOR_SSM_OUT, + }; + case LLM_ARCH_QWEN35MOE: + return { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_POST_NORM, + LLM_TENSOR_ATTN_Q, + LLM_TENSOR_ATTN_Q_NORM, + LLM_TENSOR_ATTN_K, + LLM_TENSOR_ATTN_K_NORM, + LLM_TENSOR_ATTN_V, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_ATTN_QKV, + LLM_TENSOR_ATTN_GATE, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_EXPS, + LLM_TENSOR_FFN_DOWN_EXPS, + LLM_TENSOR_FFN_UP_EXPS, + LLM_TENSOR_FFN_GATE_INP_SHEXP, + LLM_TENSOR_FFN_GATE_SHEXP, + LLM_TENSOR_FFN_DOWN_SHEXP, + LLM_TENSOR_FFN_UP_SHEXP, + LLM_TENSOR_SSM_A_NOSCAN, + LLM_TENSOR_SSM_CONV1D, + LLM_TENSOR_SSM_DT, + LLM_TENSOR_SSM_BETA, + LLM_TENSOR_SSM_ALPHA, + LLM_TENSOR_SSM_NORM, + LLM_TENSOR_SSM_OUT, + }; case LLM_ARCH_QWEN3VL: case LLM_ARCH_CHAMELEON: case LLM_ARCH_HUNYUAN_DENSE: @@ -1597,6 +1665,46 @@ static std::set llm_get_tensor_names(llm_arch arch) { LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, }; + case LLM_ARCH_GLM_DSA: + return { + LLM_TENSOR_TOKEN_EMBD, + LLM_TENSOR_OUTPUT_NORM, + LLM_TENSOR_OUTPUT, + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_Q_A_NORM, + LLM_TENSOR_ATTN_KV_A_NORM, + LLM_TENSOR_ATTN_Q, + LLM_TENSOR_ATTN_Q_A, + LLM_TENSOR_ATTN_Q_B, + LLM_TENSOR_ATTN_KV_A_MQA, + LLM_TENSOR_ATTN_KV_B, + LLM_TENSOR_ATTN_K_B, + LLM_TENSOR_ATTN_V_B, + LLM_TENSOR_ATTN_OUT, + LLM_TENSOR_FFN_NORM, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_UP, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_EXPS, + LLM_TENSOR_FFN_DOWN_EXPS, + LLM_TENSOR_FFN_UP_EXPS, + LLM_TENSOR_FFN_GATE_INP_SHEXP, + LLM_TENSOR_FFN_GATE_SHEXP, + LLM_TENSOR_FFN_DOWN_SHEXP, + LLM_TENSOR_FFN_UP_SHEXP, + LLM_TENSOR_FFN_EXP_PROBS_B, + LLM_TENSOR_INDEXER_K_NORM, + LLM_TENSOR_INDEXER_PROJ, + LLM_TENSOR_INDEXER_ATTN_K, + LLM_TENSOR_INDEXER_ATTN_Q_B, + LLM_TENSOR_NEXTN_EH_PROJ, + LLM_TENSOR_NEXTN_EMBED_TOKENS, + LLM_TENSOR_NEXTN_ENORM, + LLM_TENSOR_NEXTN_HNORM, + LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, + LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, + }; case LLM_ARCH_BITNET: return { LLM_TENSOR_TOKEN_EMBD, @@ -2456,6 +2564,7 @@ static const std::map LLM_TENSOR_INFOS = { {LLM_TENSOR_SSM_X, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SSM_DT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SSM_OUT, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_SSM_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_SSM_BETA_ALPHA, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, {LLM_TENSOR_TIME_MIX_W2, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, @@ -2582,6 +2691,10 @@ static const std::map LLM_TENSOR_INFOS = { {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}}, + {LLM_TENSOR_INDEXER_K_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}}, + {LLM_TENSOR_INDEXER_PROJ, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_INDEXER_ATTN_K, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL_MAT}}, + {LLM_TENSOR_INDEXER_ATTN_Q_B, {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}}, @@ -2675,6 +2788,8 @@ bool llm_arch_is_hybrid(const llm_arch & arch) { case LLM_ARCH_NEMOTRON_H_MOE: case LLM_ARCH_QWEN3NEXT: case LLM_ARCH_KIMI_LINEAR: + case LLM_ARCH_QWEN35: + case LLM_ARCH_QWEN35MOE: return true; default: return false; diff --git a/examples/talk-llama/llama-arch.h b/examples/talk-llama/llama-arch.h index e8263369..52194437 100644 --- a/examples/talk-llama/llama-arch.h +++ b/examples/talk-llama/llama-arch.h @@ -41,6 +41,8 @@ enum llm_arch { LLM_ARCH_QWEN3NEXT, LLM_ARCH_QWEN3VL, LLM_ARCH_QWEN3VLMOE, + LLM_ARCH_QWEN35, + LLM_ARCH_QWEN35MOE, LLM_ARCH_PHI2, LLM_ARCH_PHI3, LLM_ARCH_PHIMOE, @@ -76,6 +78,7 @@ enum llm_arch { LLM_ARCH_CHATGLM, LLM_ARCH_GLM4, LLM_ARCH_GLM4_MOE, + LLM_ARCH_GLM_DSA, LLM_ARCH_BITNET, LLM_ARCH_T5, LLM_ARCH_T5ENCODER, @@ -199,6 +202,7 @@ enum llm_kv { LLM_KV_EMBEDDING_SCALE, LLM_KV_TOKEN_SHIFT_COUNT, LLM_KV_INTERLEAVE_MOE_LAYER_STEP, + LLM_KV_FULL_ATTENTION_INTERVAL, LLM_KV_ATTENTION_HEAD_COUNT, LLM_KV_ATTENTION_HEAD_COUNT_KV, @@ -226,6 +230,9 @@ enum llm_kv { LLM_KV_ATTENTION_TEMPERATURE_SCALE, LLM_KV_ATTENTION_KEY_LENGTH_MLA, LLM_KV_ATTENTION_VALUE_LENGTH_MLA, + LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, + LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, + LLM_KV_ATTENTION_INDEXER_TOP_K, LLM_KV_ROPE_DIMENSION_COUNT, LLM_KV_ROPE_DIMENSION_SECTIONS, @@ -404,13 +411,14 @@ enum llm_tensor { LLM_TENSOR_SSM_NORM, LLM_TENSOR_SSM_OUT, LLM_TENSOR_SSM_BETA_ALPHA, // qwen3next + LLM_TENSOR_SSM_ALPHA, // qwen3.5 // Kimi Linear KDA (using SSM_ prefix for consistency) LLM_TENSOR_SSM_CONV1D_Q, // kimi: Q conv1d weight LLM_TENSOR_SSM_CONV1D_K, // kimi: K conv1d weight LLM_TENSOR_SSM_CONV1D_V, // kimi: V conv1d weight LLM_TENSOR_SSM_F_A, // kimi: forget gate projection A LLM_TENSOR_SSM_F_B, // kimi: forget gate projection B - LLM_TENSOR_SSM_BETA, // kimi: beta mixing coefficient + LLM_TENSOR_SSM_BETA, // kimi: beta mixing coefficient and qwen3.5 LLM_TENSOR_SSM_G_A, // kimi: output gate projection A LLM_TENSOR_SSM_G_B, // kimi: output gate projection B LLM_TENSOR_TIME_MIX_W0, @@ -513,6 +521,10 @@ enum llm_tensor { LLM_TENSOR_VISEXP_FFN_GATE, LLM_TENSOR_VISEXP_FFN_DOWN, LLM_TENSOR_VISEXP_FFN_UP, + LLM_TENSOR_INDEXER_K_NORM, + LLM_TENSOR_INDEXER_PROJ, + LLM_TENSOR_INDEXER_ATTN_K, + LLM_TENSOR_INDEXER_ATTN_Q_B, LLM_TENSOR_NEXTN_EH_PROJ, LLM_TENSOR_NEXTN_EMBED_TOKENS, LLM_TENSOR_NEXTN_ENORM, diff --git a/examples/talk-llama/llama-context.cpp b/examples/talk-llama/llama-context.cpp index a6df893a..99035b6c 100644 --- a/examples/talk-llama/llama-context.cpp +++ b/examples/talk-llama/llama-context.cpp @@ -677,7 +677,7 @@ enum llama_pooling_type llama_context::pooling_type() const { float * llama_context::get_logits() { output_reorder(); - return logits; + return logits.data; } int64_t llama_context::output_resolve_row(int32_t i) const { @@ -715,7 +715,7 @@ float * llama_context::get_logits_ith(int32_t i) { output_reorder(); try { - if (logits == nullptr) { + if (logits.data == nullptr) { throw std::runtime_error("no logits"); } @@ -739,7 +739,7 @@ float * llama_context::get_logits_ith(int32_t i) { throw std::runtime_error(format("corrupt output buffer (j=%" PRId64 ", n_outputs=%d)", j, n_outputs)); } - return logits + j*model.vocab.n_tokens(); + return logits.data + j*model.vocab.n_tokens(); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what()); #ifndef NDEBUG @@ -753,11 +753,11 @@ float * llama_context::get_logits_ith(int32_t i) { float * llama_context::get_embeddings() { output_reorder(); - return embd; + return embd.data; } llama_token * llama_context::get_sampled_tokens() const{ - return sampling.sampled; + return sampling.sampled.data; } float * llama_context::get_embeddings_ith(int32_t i) { @@ -766,7 +766,7 @@ float * llama_context::get_embeddings_ith(int32_t i) { output_reorder(); try { - if (embd == nullptr) { + if (embd.data == nullptr) { throw std::runtime_error("no embeddings"); } @@ -791,7 +791,7 @@ float * llama_context::get_embeddings_ith(int32_t i) { } const uint32_t n_embd_out = model.hparams.n_embd_out(); - return embd + j*n_embd_out; + return embd.data + j*n_embd_out; } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: invalid embeddings id %d, reason: %s\n", __func__, i, err.what()); #ifndef NDEBUG @@ -814,14 +814,14 @@ float * llama_context::get_embeddings_seq(llama_seq_id seq_id) { llama_token llama_context::get_sampled_token_ith(int32_t idx) { output_reorder(); - if (sampling.sampled == nullptr) { + if (!sampling.sampled.has_data()) { return LLAMA_TOKEN_NULL; } try { const int64_t row = output_resolve_row(idx); - GGML_ASSERT(row < (int64_t) sampling.sampled_size); - return sampling.sampled[row]; + GGML_ASSERT(row < (int64_t) sampling.sampled.size); + return sampling.sampled.data[row]; } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: invalid backend sampled token id %d, reason: %s\n", __func__, idx, err.what()); return LLAMA_TOKEN_NULL; @@ -831,7 +831,7 @@ llama_token llama_context::get_sampled_token_ith(int32_t idx) { float * llama_context::get_sampled_probs_ith(int32_t idx) { output_reorder(); - if (sampling.probs == nullptr) { + if (!sampling.probs.has_data()) { return nullptr; } @@ -840,7 +840,7 @@ float * llama_context::get_sampled_probs_ith(int32_t idx) { if ((size_t) row >= sampling.probs_count.size() || sampling.probs_count[row] == 0) { return nullptr; } - return sampling.probs + row*model.vocab.n_tokens(); + return sampling.probs.data + row*model.vocab.n_tokens(); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: invalid backend sampled probs id %d, reason: %s\n", __func__, idx, err.what()); return nullptr; @@ -850,7 +850,7 @@ float * llama_context::get_sampled_probs_ith(int32_t idx) { float * llama_context::get_sampled_logits_ith(int32_t idx) { output_reorder(); - if (sampling.logits == nullptr) { + if (!sampling.logits.has_data()) { return nullptr; } @@ -859,7 +859,7 @@ float * llama_context::get_sampled_logits_ith(int32_t idx) { if ((size_t) row >= sampling.logits_count.size() || sampling.logits_count[row] == 0) { return nullptr; } - return sampling.logits + row*model.vocab.n_tokens(); + return sampling.logits.data + row*model.vocab.n_tokens(); } catch (const std::exception & err) { LLAMA_LOG_ERROR("%s: invalid backend sampled logits id %d, reason: %s\n", __func__, idx, err.what()); return nullptr; @@ -871,13 +871,14 @@ const llama_token * llama_context::get_sampled_candidates_ith(int32_t idx) { try { const int64_t row = output_resolve_row(idx); - if (sampling.candidates != nullptr && + if (sampling.candidates.has_data() && (size_t) row < sampling.candidates_count.size() && sampling.candidates_count[row] > 0) { - return sampling.candidates + row*model.vocab.n_tokens(); + return sampling.candidates.data + row*model.vocab.n_tokens(); } } catch (const std::exception & err) { // fallback to full vocab list + GGML_UNUSED(err); } return sampling.token_ids_full_vocab.data(); @@ -886,7 +887,7 @@ const llama_token * llama_context::get_sampled_candidates_ith(int32_t idx) { size_t llama_context::get_sampled_candidates_count(int32_t idx) { output_reorder(); - if (sampling.candidates == nullptr) { + if (!sampling.candidates.has_data()) { return 0; } @@ -905,7 +906,7 @@ size_t llama_context::get_sampled_candidates_count(int32_t idx) { size_t llama_context::get_sampled_logits_count(int32_t idx) { output_reorder(); - if (sampling.logits == nullptr) { + if (!sampling.logits.has_data()) { return model.vocab.n_tokens(); } @@ -924,7 +925,7 @@ size_t llama_context::get_sampled_logits_count(int32_t idx) { size_t llama_context::get_sampled_probs_count(int32_t idx) { output_reorder(); - if (sampling.probs == nullptr) { + if (!sampling.probs.has_data()) { return 0; } @@ -1057,51 +1058,43 @@ bool llama_context::set_sampler(llama_seq_id seq_id, llama_sampler * sampler) { return true; } -void llama_context::set_adapter_lora( - llama_adapter_lora * adapter, - float scale) { - LLAMA_LOG_DEBUG("%s: adapter = %p, scale = %f\n", __func__, (void *) adapter, scale); +void llama_context::set_adapters_lora(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) { + LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters); - if (auto it = loras.find(adapter); it != loras.end()) { - if (it->second == scale) { - return; - } + if (adapters_lora_are_same(adapters, n_adapters, scales)) { + return; } - loras[adapter] = scale; + loras.clear(); + + for (size_t i = 0; i < n_adapters; i ++) { + if (scales[i] != 0.0f) { + loras[adapters[i]] = scales[i]; + } + } sched_need_reserve = true; } -bool llama_context::rm_adapter_lora( - llama_adapter_lora * adapter) { - LLAMA_LOG_DEBUG("%s: adapter = %p\n", __func__, (void *) adapter); - - auto it = loras.find(adapter); - if (it != loras.end()) { - loras.erase(it); - - sched_need_reserve = true; +bool llama_context::adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales) { + LLAMA_LOG_DEBUG("%s: adapters = %p\n", __func__, (void *) adapters); - return true; + if (n_adapters != loras.size()) { + return false; } - return false; -} - -void llama_context::clear_adapter_lora() { - LLAMA_LOG_DEBUG("%s: call\n", __func__); + for (size_t i = 0; i < n_adapters; i ++) { + auto it = loras.find(adapters[i]); - if (loras.empty()) { - return; + if (it == loras.end() || it->second != scales[i]) { + return false; + } } - loras.clear(); - - sched_need_reserve = true; + return true; } -bool llama_context::apply_adapter_cvec( +bool llama_context::set_adapter_cvec( const float * data, size_t len, int32_t n_embd, @@ -1254,16 +1247,16 @@ int llama_context::encode(const llama_batch & batch_inp) { auto * t_embd = res->get_embd_pooled() ? res->get_embd_pooled() : res->get_embd(); // extract logits - if (logits && t_logits) { + if (logits.data && t_logits) { ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits); GGML_ASSERT(backend_res != nullptr); - GGML_ASSERT(logits != nullptr); + GGML_ASSERT(logits.data != nullptr); - ggml_backend_tensor_get_async(backend_res, t_logits, logits, 0, n_tokens*n_vocab*sizeof(float)); + ggml_backend_tensor_get_async(backend_res, t_logits, logits.data, 0, n_tokens*n_vocab*sizeof(float)); } // extract embeddings - if (embd && t_embd) { + if (embd.data && t_embd) { ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd); GGML_ASSERT(backend_embd != nullptr); @@ -1271,11 +1264,11 @@ int llama_context::encode(const llama_batch & batch_inp) { case LLAMA_POOLING_TYPE_NONE: { // extract token embeddings - GGML_ASSERT(embd != nullptr); + GGML_ASSERT(embd.data != nullptr); const uint32_t n_embd_out = hparams.n_embd_out(); - GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd_size); - ggml_backend_tensor_get_async(backend_embd, t_embd, embd, 0, n_tokens*n_embd_out*sizeof(float)); + GGML_ASSERT(n_tokens*n_embd_out <= (int64_t) embd.size); + ggml_backend_tensor_get_async(backend_embd, t_embd, embd.data, 0, n_tokens*n_embd_out*sizeof(float)); } break; case LLAMA_POOLING_TYPE_MEAN: case LLAMA_POOLING_TYPE_CLS: @@ -1323,7 +1316,7 @@ int llama_context::encode(const llama_batch & batch_inp) { cross.n_embd = t_embd->ne[0]; cross.n_enc = t_embd->ne[1]; cross.v_embd.resize(cross.n_embd*cross.n_enc); - memcpy(cross.v_embd.data(), embd, ggml_nbytes(t_embd)); + memcpy(cross.v_embd.data(), embd.data, ggml_nbytes(t_embd)); const auto & batch = balloc->get_batch(); @@ -1363,11 +1356,10 @@ static std::map build_seq_to_output_row(const llama_ubat static void copy_tensor_async_ints( const std::map & tensor_map, - llama_token * sampled, - size_t sampled_size, + const buffer_view & sampled, const std::map & seq_to_row, ggml_backend_sched_t sched) { - if (sampled == nullptr) { + if (!sampled.has_data()) { return; } @@ -1378,23 +1370,23 @@ static void copy_tensor_async_ints( } const uint32_t row = it->second; - GGML_ASSERT(row < sampled_size); + GGML_ASSERT(row < sampled.size); GGML_ASSERT(ggml_is_contiguous(tensor) && "sampled tokens tensor must be contiguous for async copy"); ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor); - ggml_backend_tensor_get_async(backend, tensor, sampled + row, 0, sizeof(sampled[row])); + ggml_backend_tensor_get_async(backend, tensor, sampled.data + row, 0, sizeof(sampled.data[row])); } } static void copy_tensor_async_floats( const std::map & tensor_map, - float * dst, + const buffer_view & dst, size_t stride, std::vector & counts, const std::map & seq_to_row, ggml_backend_sched_t sched) { - if (dst == nullptr) { + if (!dst.has_data()) { return; } @@ -1410,7 +1402,7 @@ static void copy_tensor_async_floats( GGML_ASSERT(ggml_is_contiguous(tensor) && "logits/probs tensor must be contiguous for async copy"); ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor); - float * row_ptr = dst + (size_t) row * stride; + float * row_ptr = dst.data + (size_t) row * stride; ggml_backend_tensor_get_async(backend, tensor, row_ptr, 0, ggml_nbytes(tensor)); // Update the actual number of logits/probabilities that were written for this row. @@ -1420,12 +1412,12 @@ static void copy_tensor_async_floats( static void copy_tensor_async_candidates( const std::map & tensor_map, - llama_token * dst, + const buffer_view & dst, size_t stride, std::vector & counts, const std::map & seq_to_row, ggml_backend_sched_t sched) { - if (dst == nullptr) { + if (!dst.has_data()) { return; } @@ -1441,7 +1433,7 @@ static void copy_tensor_async_candidates( GGML_ASSERT(ggml_is_contiguous(tensor) && "candidates tensor must be contiguous for async copy"); ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(sched, tensor); - llama_token * row_ptr = dst + (size_t) row * stride; + llama_token * row_ptr = dst.data + (size_t) row * stride; ggml_backend_tensor_get_async(backend, tensor, row_ptr, 0, ggml_nbytes(tensor)); // Update the actual number of candidates that were written. @@ -1671,22 +1663,22 @@ int llama_context::decode(const llama_batch & batch_inp) { } // extract logits - if (logits && t_logits && n_outputs > 0 && needs_raw_logits(ubatch, sampling.samplers)) { + if (logits.data && t_logits && n_outputs > 0 && needs_raw_logits(ubatch, sampling.samplers)) { ggml_backend_t backend_res = ggml_backend_sched_get_tensor_backend(sched.get(), t_logits); GGML_ASSERT(backend_res != nullptr); - GGML_ASSERT(logits != nullptr); + GGML_ASSERT(logits.data != nullptr); - float * logits_out = logits + n_outputs_prev*n_vocab; + float * logits_out = logits.data + n_outputs_prev*n_vocab; if (n_outputs) { GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all); - GGML_ASSERT((n_outputs_prev + n_outputs)*n_vocab <= (int64_t) logits_size); + GGML_ASSERT((n_outputs_prev + n_outputs)*n_vocab <= (int64_t) logits.size); ggml_backend_tensor_get_async(backend_res, t_logits, logits_out, 0, n_outputs*n_vocab*sizeof(float)); } } // extract embeddings - if (embd && t_embd && n_outputs > 0) { + if (embd.data && t_embd && n_outputs > 0) { ggml_backend_t backend_embd = ggml_backend_sched_get_tensor_backend(sched.get(), t_embd); GGML_ASSERT(backend_embd != nullptr); @@ -1694,13 +1686,13 @@ int llama_context::decode(const llama_batch & batch_inp) { case LLAMA_POOLING_TYPE_NONE: { // extract token embeddings - GGML_ASSERT(embd != nullptr); + GGML_ASSERT(embd.data != nullptr); const uint32_t n_embd_out = hparams.n_embd_out(); - float * embd_out = embd + n_outputs_prev*n_embd_out; + float * embd_out = embd.data + n_outputs_prev*n_embd_out; if (n_outputs) { GGML_ASSERT( n_outputs_prev + n_outputs <= n_outputs_all); - GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd_out <= (int64_t) embd_size); + GGML_ASSERT((n_outputs_prev + n_outputs)*n_embd_out <= (int64_t) embd.size); ggml_backend_tensor_get_async(backend_embd, t_embd, embd_out, 0, n_outputs*n_embd_out*sizeof(float)); } } break; @@ -1747,7 +1739,7 @@ int llama_context::decode(const llama_batch & batch_inp) { const auto stride = n_vocab; // async copy the sampling data from the backend to the host - copy_tensor_async_ints(res->t_sampled, sampling.sampled, sampling.sampled_size, seq_to_output_row, sched.get()); + copy_tensor_async_ints(res->t_sampled, sampling.sampled, seq_to_output_row, sched.get()); copy_tensor_async_floats (res->t_sampled_logits, sampling.logits, stride, sampling.logits_count, seq_to_output_row, sched.get()); copy_tensor_async_floats (res->t_sampled_probs, sampling.probs, stride, sampling.probs_count, seq_to_output_row, sched.get()); @@ -1818,7 +1810,6 @@ int llama_context::decode(const llama_batch & batch_inp) { // uint32_t llama_context::output_reserve(int32_t n_outputs) { - const auto & hparams = model.hparams; const auto & vocab = model.vocab; @@ -1841,19 +1832,14 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) { size_t backend_float_count = 0; size_t backend_token_count = 0; - logits_size = has_logits ? n_vocab*n_outputs_max : 0; - embd_size = has_embd ? n_embd_out*n_outputs_max : 0; + logits.size = has_logits ? n_vocab*n_outputs_max : 0; + embd.size = has_embd ? n_embd_out*n_outputs_max : 0; // Allocate backend sampling output buffers if there are backend samplers configured. const bool has_sampling = !sampling.samplers.empty(); if (has_sampling) { - sampling.logits_size = n_vocab*n_outputs_max; - sampling.probs_size = n_vocab*n_outputs_max; - sampling.sampled_size = n_outputs_max; - sampling.candidates_size = n_vocab*n_outputs_max; - - backend_float_count = sampling.logits_size + sampling.probs_size; - backend_token_count = sampling.sampled_size + sampling.candidates_size; + backend_float_count = 2 * n_vocab * n_outputs_max; // logits + probs + backend_token_count = (1 + n_vocab) * n_outputs_max; // sampled + candidates } if (output_ids.empty()) { @@ -1863,7 +1849,7 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) { const size_t prev_size = buf_output ? ggml_backend_buffer_get_size(buf_output.get()) : 0; const size_t new_size = - (logits_size + embd_size + backend_float_count) * sizeof(float) + + (logits.size + embd.size + backend_float_count) * sizeof(float) + ( backend_token_count) * sizeof(llama_token); // alloc only when more than the current capacity is required @@ -1878,8 +1864,8 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) { // TODO: not needed? buf_output = nullptr; - logits = nullptr; - embd = nullptr; + logits.data = nullptr; + embd.data = nullptr; } auto * buft = ggml_backend_cpu_buffer_type(); @@ -1898,35 +1884,27 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) { float * output_base = (float *) ggml_backend_buffer_get_base(buf_output.get()); - logits = nullptr; - embd = nullptr; - size_t offset = 0; uint8_t * base = (uint8_t *) output_base; - logits = has_logits ? output_base : nullptr; - offset += logits_size * sizeof(float); - - embd = has_embd ? (float *) (base + offset) : nullptr; - offset += embd_size * sizeof(float); + logits = has_logits ? buffer_view{output_base, logits.size} : buffer_view{nullptr, 0}; + offset += logits.size * sizeof(float); - sampling.logits = nullptr; - sampling.probs = nullptr; - sampling.sampled = nullptr; - sampling.candidates = nullptr; + embd = has_embd ? buffer_view{(float *) (base + offset), embd.size} : buffer_view{nullptr, 0}; + offset += embd.size * sizeof(float); if (has_sampling) { - sampling.logits = (float *) (base + offset); - offset += sampling.logits_size * sizeof(float); + sampling.logits = {(float *) (base + offset), (size_t)(n_vocab*n_outputs_max)}; + offset += sampling.logits.size * sizeof(float); - sampling.probs = (float *) (base + offset); - offset += sampling.probs_size * sizeof(float); + sampling.probs = {(float *) (base + offset), (size_t)(n_vocab*n_outputs_max)}; + offset += sampling.probs.size * sizeof(float); - sampling.sampled = (llama_token *) (base + offset); - offset += sampling.sampled_size * sizeof(llama_token); + sampling.sampled = {(llama_token *) (base + offset), (size_t)n_outputs_max}; + offset += sampling.sampled.size * sizeof(llama_token); - sampling.candidates = (llama_token *) (base + offset); - offset += sampling.candidates_size * sizeof(llama_token); + sampling.candidates = {(llama_token *) (base + offset), (size_t)(n_vocab*n_outputs_max)}; + offset += sampling.candidates.size * sizeof(llama_token); // The count vectors keep track of the actual number of logits/probs/candidates // copied from the backend for each output row. @@ -1939,7 +1917,16 @@ uint32_t llama_context::output_reserve(int32_t n_outputs) { std::fill(sampling.probs_count.begin(), sampling.probs_count.end(), 0); std::fill(sampling.candidates_count.begin(), sampling.candidates_count.end(), 0); - std::fill_n(sampling.sampled, sampling.sampled_size, LLAMA_TOKEN_NULL); + std::fill_n(sampling.sampled.data, sampling.sampled.size, LLAMA_TOKEN_NULL); + } else { + sampling.logits = {nullptr, 0}; + sampling.probs = {nullptr, 0}; + sampling.sampled = {nullptr, 0}; + sampling.candidates = {nullptr, 0}; + + sampling.logits_count.clear(); + sampling.probs_count.clear(); + sampling.candidates_count.clear(); } // set all ids as invalid (negative) @@ -1958,49 +1945,42 @@ void llama_context::output_reorder() { const uint64_t i0 = output_swaps[s].i0; const uint64_t i1 = output_swaps[s].i1; - if (logits_size > 0) { + if (logits.size > 0) { for (uint64_t k = 0; k < n_vocab; k++) { - std::swap(logits[i0*n_vocab + k], logits[i1*n_vocab + k]); + std::swap(logits.data[i0*n_vocab + k], logits.data[i1*n_vocab + k]); } } - if (embd_size > 0) { + if (embd.size > 0) { for (uint64_t k = 0; k < n_embd; k++) { - std::swap(embd[i0*n_embd + k], embd[i1*n_embd + k]); + std::swap(embd.data[i0*n_embd + k], embd.data[i1*n_embd + k]); } } - if (sampling.logits && sampling.logits_size > 0) { + if (!sampling.samplers.empty()) { + assert(sampling.logits.size > 0); + assert(sampling.probs.size > 0); + assert(sampling.candidates.size > 0); + assert(sampling.sampled.size > 0); + assert(sampling.logits_count.size() > 0); + assert(sampling.probs_count.size() > 0); + assert(sampling.candidates_count.size() > 0); + for (uint64_t k = 0; k < n_vocab; ++k) { - std::swap(sampling.logits[i0*n_vocab + k], sampling.logits[i1*n_vocab + k]); + std::swap(sampling.logits.data[i0*n_vocab + k], sampling.logits.data[i1*n_vocab + k]); } - } - if (sampling.probs && sampling.probs_size > 0) { for (uint64_t k = 0; k < n_vocab; ++k) { - std::swap(sampling.probs[i0*n_vocab + k], sampling.probs[i1*n_vocab + k]); + std::swap(sampling.probs.data[i0*n_vocab + k], sampling.probs.data[i1*n_vocab + k]); } - } - if (sampling.candidates && sampling.candidates_size > 0) { for (uint64_t k = 0; k < n_vocab; ++k) { - std::swap(sampling.candidates[i0*n_vocab + k], sampling.candidates[i1*n_vocab + k]); + std::swap(sampling.candidates.data[i0*n_vocab + k], sampling.candidates.data[i1*n_vocab + k]); } - } - - if (sampling.sampled && sampling.sampled_size > 0) { - std::swap(sampling.sampled[i0], sampling.sampled[i1]); - } - - if (!sampling.logits_count.empty()) { - std::swap(sampling.logits_count[i0], sampling.logits_count[i1]); - } - if (!sampling.probs_count.empty()) { - std::swap(sampling.probs_count[i0], sampling.probs_count[i1]); - } - - if (!sampling.candidates_count.empty()) { + std::swap(sampling.sampled.data[i0], sampling.sampled.data[i1]); + std::swap(sampling.logits_count[i0], sampling.logits_count[i1]); + std::swap(sampling.probs_count[i0], sampling.probs_count[i1]); std::swap(sampling.candidates_count[i0], sampling.candidates_count[i1]); } } @@ -2013,7 +1993,7 @@ void llama_context::output_reorder() { // uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const { - if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI_LINEAR) { + if (model.arch == LLM_ARCH_QWEN3NEXT || model.arch == LLM_ARCH_KIMI_LINEAR || model.arch == LLM_ARCH_QWEN35 || model.arch == LLM_ARCH_QWEN35MOE) { return std::max(n_tokens * 40, 32u * model.n_tensors()); } uint32_t res = std::max(1024u, 8u*model.n_tensors()); @@ -2533,12 +2513,12 @@ size_t llama_context::state_write_data(llama_io_write_i & io) { { LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__); - const uint64_t logits_size = std::min((uint64_t) this->logits_size, (uint64_t) n_outputs * model.vocab.n_tokens()); + const uint64_t logits_size = std::min((uint64_t) this->logits.size, (uint64_t) n_outputs * model.vocab.n_tokens()); io.write(&logits_size, sizeof(logits_size)); if (logits_size) { - io.write(logits, logits_size * sizeof(float)); + io.write(logits.data, logits_size * sizeof(float)); } } @@ -2546,12 +2526,12 @@ size_t llama_context::state_write_data(llama_io_write_i & io) { { LLAMA_LOG_DEBUG("%s: - writing embeddings\n", __func__); - const uint64_t embd_size = std::min((uint64_t) this->embd_size, (uint64_t) n_outputs * model.hparams.n_embd); + const uint64_t embd_size = std::min((uint64_t) this->embd.size, (uint64_t) n_outputs * model.hparams.n_embd); io.write(&embd_size, sizeof(embd_size)); if (embd_size) { - io.write(embd, embd_size * sizeof(float)); + io.write(embd.data, embd_size * sizeof(float)); } } @@ -2619,12 +2599,12 @@ size_t llama_context::state_read_data(llama_io_read_i & io) { uint64_t logits_size; io.read_to(&logits_size, sizeof(logits_size)); - if (this->logits_size < logits_size) { + if (this->logits.size < logits_size) { throw std::runtime_error("logits buffer too small"); } if (logits_size) { - io.read_to(this->logits, logits_size * sizeof(float)); + io.read_to(this->logits.data, logits_size * sizeof(float)); } } @@ -2635,12 +2615,12 @@ size_t llama_context::state_read_data(llama_io_read_i & io) { uint64_t embd_size; io.read_to(&embd_size, sizeof(embd_size)); - if (this->embd_size < embd_size) { + if (this->embd.size < embd_size) { throw std::runtime_error("embeddings buffer too small"); } if (embd_size) { - io.read_to(this->embd, embd_size * sizeof(float)); + io.read_to(this->embd.data, embd_size * sizeof(float)); } } @@ -3218,35 +3198,28 @@ uint32_t llama_get_sampled_probs_count_ith(llama_context * ctx, int32_t i) { // llama adapter API -int32_t llama_set_adapter_lora( +int32_t llama_set_adapters_lora( llama_context * ctx, - llama_adapter_lora * adapter, - float scale) { - ctx->set_adapter_lora(adapter, scale); - - return 0; -} - -int32_t llama_rm_adapter_lora( - llama_context * ctx, - llama_adapter_lora * adapter) { - bool res = ctx->rm_adapter_lora(adapter); + llama_adapter_lora ** adapters, + size_t n_adapters, + float * scales) { + if (adapters == nullptr || scales == nullptr) { + GGML_ASSERT(n_adapters == 0 && "invalid llama_set_adapters_lora call"); + } - return res ? 0 : -1; -} + ctx->set_adapters_lora(adapters, n_adapters, scales); -void llama_clear_adapter_lora(llama_context * ctx) { - ctx->clear_adapter_lora(); + return 0; } -int32_t llama_apply_adapter_cvec( +int32_t llama_set_adapter_cvec( llama_context * ctx, - const float * data, - size_t len, - int32_t n_embd, - int32_t il_start, - int32_t il_end) { - bool res = ctx->apply_adapter_cvec(data, len, n_embd, il_start, il_end); + const float * data, + size_t len, + int32_t n_embd, + int32_t il_start, + int32_t il_end) { + bool res = ctx->set_adapter_cvec(data, len, n_embd, il_start, il_end); return res ? 0 : -1; } diff --git a/examples/talk-llama/llama-context.h b/examples/talk-llama/llama-context.h index 8e71cdd1..a8e53f33 100644 --- a/examples/talk-llama/llama-context.h +++ b/examples/talk-llama/llama-context.h @@ -4,6 +4,7 @@ #include "llama-cparams.h" #include "llama-graph.h" #include "llama-adapter.h" +#include "llama-impl.h" #include "ggml-cpp.h" #include "ggml-opt.h" @@ -104,16 +105,11 @@ struct llama_context { void set_causal_attn(bool value); void set_warmup(bool value); - void set_adapter_lora( - llama_adapter_lora * adapter, - float scale); + void set_adapters_lora(llama_adapter_lora ** adapters, size_t n_adapters, float * scales); - bool rm_adapter_lora( - llama_adapter_lora * adapter); + bool adapters_lora_are_same(llama_adapter_lora ** adapters, size_t n_adapters, float * scales); - void clear_adapter_lora(); - - bool apply_adapter_cvec( + bool set_adapter_cvec( const float * data, size_t len, int32_t n_embd, @@ -269,34 +265,26 @@ private: std::unique_ptr memory; // decode output (2-dimensional array: [n_outputs][n_vocab]) - size_t logits_size = 0; // capacity (of floats) for logits - float * logits = nullptr; + buffer_view logits = {nullptr, 0}; // embeddings output (2-dimensional array: [n_outputs][n_embd]) // populated only when pooling_type == LLAMA_POOLING_TYPE_NONE - size_t embd_size = 0; // capacity (of floats) for embeddings - float * embd = nullptr; + buffer_view embd = {nullptr, 0}; - // TODO: simplify struct sampling_info { + // !samplers.empty() to check if any samplers are active std::map samplers; - float * logits = nullptr; - size_t logits_size = 0; - - llama_token * sampled = nullptr; - size_t sampled_size = 0; - - float * probs = nullptr; - size_t probs_size = 0; - - llama_token * candidates = nullptr; - size_t candidates_size = 0; + buffer_view logits = {nullptr, 0}; + buffer_view sampled = {nullptr, 0}; + buffer_view probs = {nullptr, 0}; + buffer_view candidates = {nullptr, 0}; std::vector logits_count; std::vector probs_count; std::vector candidates_count; + // optimization std::vector token_ids_full_vocab; }; diff --git a/examples/talk-llama/llama-hparams.h b/examples/talk-llama/llama-hparams.h index 6c695bdb..c4b2a99d 100644 --- a/examples/talk-llama/llama-hparams.h +++ b/examples/talk-llama/llama-hparams.h @@ -42,7 +42,6 @@ struct llama_hparams { uint32_t n_ctx_train; // context size the model was trained on uint32_t n_embd; - uint32_t n_embd_features = 0; uint32_t n_layer; int32_t n_layer_kv_from_start = -1; // if non-negative, the first n_layer_kv_from_start layers have KV cache uint32_t n_rot; @@ -194,6 +193,11 @@ struct llama_hparams { std::array xielu_beta; std::array xielu_eps; + // DSA (deepseek sparse attention) + uint32_t indexer_n_head = 0; + uint32_t indexer_head_size = 0; + uint32_t indexer_top_k = 0; + // qwen3vl deepstack uint32_t n_deepstack_layers = 0; diff --git a/examples/talk-llama/llama-impl.h b/examples/talk-llama/llama-impl.h index c3391e79..dfd9fee9 100644 --- a/examples/talk-llama/llama-impl.h +++ b/examples/talk-llama/llama-impl.h @@ -49,6 +49,16 @@ struct time_meas { int64_t & t_acc; }; +template +struct buffer_view { + T * data; + size_t size = 0; + + bool has_data() const { + return data && size > 0; + } +}; + void replace_all(std::string & s, const std::string & search, const std::string & replace); // TODO: rename to llama_format ? diff --git a/examples/talk-llama/llama-mmap.cpp b/examples/talk-llama/llama-mmap.cpp index 0261e4c7..c03228e9 100644 --- a/examples/talk-llama/llama-mmap.cpp +++ b/examples/talk-llama/llama-mmap.cpp @@ -504,6 +504,8 @@ struct llama_mmap::impl { } } #elif defined(_WIN32) + HANDLE hMapping = nullptr; + impl(struct llama_file * file, size_t prefetch, bool numa) { GGML_UNUSED(numa); @@ -511,7 +513,7 @@ struct llama_mmap::impl { HANDLE hFile = (HANDLE) _get_osfhandle(file->file_id()); - HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL); + hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL); if (hMapping == NULL) { DWORD error = GetLastError(); @@ -520,9 +522,9 @@ struct llama_mmap::impl { addr = MapViewOfFile(hMapping, FILE_MAP_READ, 0, 0, 0); DWORD error = GetLastError(); - CloseHandle(hMapping); if (addr == NULL) { + CloseHandle(hMapping); throw std::runtime_error(format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str())); } @@ -554,9 +556,17 @@ struct llama_mmap::impl { } ~impl() { - if (!UnmapViewOfFile(addr)) { - LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n", - llama_format_win_err(GetLastError()).c_str()); + if (hMapping) { + if (addr) { + if (!UnmapViewOfFile(addr)) { + LLAMA_LOG_WARN("warning: UnmapViewOfFile failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } + } + if (!CloseHandle(hMapping)) { + LLAMA_LOG_WARN("warning: CloseHandle failed: %s\n", + llama_format_win_err(GetLastError()).c_str()); + } } } #else diff --git a/examples/talk-llama/llama-model.cpp b/examples/talk-llama/llama-model.cpp index 674d06c8..c26584aa 100644 --- a/examples/talk-llama/llama-model.cpp +++ b/examples/talk-llama/llama-model.cpp @@ -125,6 +125,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_21B_A3B: return "21B.A3B"; case LLM_TYPE_30B_A3B: return "30B.A3B"; case LLM_TYPE_31B_A3_5B: return "31B.A3.5B"; + case LLM_TYPE_35B_A3B: return "35B.A3B"; case LLM_TYPE_48B_A3B: return "48B.A3B"; case LLM_TYPE_80B_A3B: return "80B.A3B"; case LLM_TYPE_100B_A6B: return "100B.A6B"; @@ -136,6 +137,7 @@ const char * llm_type_name(llm_type type) { case LLM_TYPE_300B_A47B: return "300B.A47B"; case LLM_TYPE_310B_A15B: return "310B.A15B"; case LLM_TYPE_355B_A32B: return "355B.A32B"; + case LLM_TYPE_744B_A40B: return "744B.A40B"; case LLM_TYPE_E2B: return "E2B"; case LLM_TYPE_E4B: return "E4B"; default: return "?B"; @@ -522,7 +524,8 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_EXPERT_GROUP_USED_COUNT, hparams.n_group_used, false); if (arch == LLM_ARCH_WAVTOKENIZER_DEC) { - ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd_features); + ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd); + ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd_out_impl); ml.get_key(LLM_KV_POSNET_EMBEDDING_LENGTH, hparams.posnet.n_embd); ml.get_key(LLM_KV_POSNET_BLOCK_COUNT, hparams.posnet.n_layer); @@ -1820,6 +1823,50 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_GLM_DSA: + { + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false); + + // MoE parameters + ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert); + ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used); + ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); + ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT, hparams.n_layer_dense_lead, false); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE, hparams.expert_weights_scale); + ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM, hparams.expert_weights_norm, false); + + // deepseek MLA parameters + ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q); + ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv); + ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA, hparams.n_embd_head_k_mla_impl, false); + ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false); + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp); + ml.get_key(LLM_KV_EXPERT_SHARED_COUNT, hparams.n_expert_shared); + + // DSA parameters + ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head); + ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size); + ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K, hparams.indexer_top_k); + + // Expert gating function (GLM-4.5 uses sigmoid) + ml.get_key(LLM_KV_EXPERT_GATING_FUNC, hparams.expert_gating_func, false); + if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) { + hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID; + } + + // NextN/MTP parameters + ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false); + + // TODO: when MTP is implemented, this should probably be updated if needed + hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers; + + switch (hparams.n_layer) { + case 79: type = LLM_TYPE_744B_A40B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_BITNET: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -2403,8 +2450,12 @@ void llama_model::load_hparams(llama_model_loader & ml) { ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); // Mark recurrent layers (linear attention layers) - for (uint32_t i = 0; i < hparams.n_layer; ++i) { - hparams.recurrent_layer_arr[i] = ((i + 1) % 4 != 0); // TODO: extract the magic 4 from "full_attention_interval" + { + uint32_t full_attn_interval = 4; + ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false); + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0); + } } switch (hparams.n_layer) { @@ -2412,6 +2463,62 @@ void llama_model::load_hparams(llama_model_loader & ml) { default: type = LLM_TYPE_UNKNOWN; } } break; + case LLM_ARCH_QWEN35: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true); + + // Load linear attention (gated delta net) parameters + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); + + // Mark recurrent layers (linear attention layers) + { + uint32_t full_attn_interval = 4; + ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false); + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0); + } + } + + switch (hparams.n_layer) { + case 24: type = LLM_TYPE_2B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; + case LLM_ARCH_QWEN35MOE: + { + ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false); + ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false); + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + + ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, true); + + // Load linear attention (gated delta net) parameters + ml.get_key(LLM_KV_SSM_CONV_KERNEL, hparams.ssm_d_conv); + ml.get_key(LLM_KV_SSM_INNER_SIZE, hparams.ssm_d_inner); + ml.get_key(LLM_KV_SSM_STATE_SIZE, hparams.ssm_d_state); + ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank); + ml.get_key(LLM_KV_SSM_GROUP_COUNT, hparams.ssm_n_group); + + // Mark recurrent layers (linear attention layers) + { + uint32_t full_attn_interval = 4; + ml.get_key(LLM_KV_FULL_ATTENTION_INTERVAL, full_attn_interval, false); + for (uint32_t i = 0; i < hparams.n_layer; ++i) { + hparams.recurrent_layer_arr[i] = ((i + 1) % full_attn_interval != 0); + } + } + + switch (hparams.n_layer) { + case 28: type = LLM_TYPE_35B_A3B; break; + case 48: type = LLM_TYPE_80B_A3B; break; + default: type = LLM_TYPE_UNKNOWN; + } + } break; case LLM_ARCH_MISTRAL3: { ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); @@ -5430,6 +5537,108 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } } break; + case LLM_ARCH_GLM_DSA: + { + const bool is_mla = hparams.is_mla(); + if (!is_mla) { + throw std::runtime_error("GLM_DSA architecture requires MLA"); + } + + // 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_mla = hparams.n_embd_head_k_mla(); + const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla(); + + const int64_t n_embd_head_qk_rope = hparams.n_rot; + const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope; + + const int64_t q_lora_rank = hparams.n_lora_q; + const int64_t kv_lora_rank = hparams.n_lora_kv; + + const int64_t n_ff_exp = hparams.n_ff_exp; + const int64_t n_expert_shared = hparams.n_expert_shared; + + 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); + // try to load output.weight, if not found, use token_embd (tied embeddings) + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED); + if (!output) { + output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); + } + + for (int i = 0; i < n_layer; ++i) { + int flags = 0; + if (hparams.nextn_predict_layers > 0 && static_cast(i) >= n_layer - hparams.nextn_predict_layers) { + // skip all tensors in the NextN layers + // TODO @ngxson : TENSOR_NOT_REQUIRED was a hack, need to remove it later + flags |= TENSOR_SKIP | TENSOR_NOT_REQUIRED; + } + + auto & layer = layers[i]; + + layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags); + layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, flags); + layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags); + + layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags); + layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, flags); + + layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags); + + // note: only old legacy GGUF files will have the unsplit wkv_b tensor in + layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, flags); + layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, flags); + + layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, flags); + + layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags); + + // DSA indexer + layer.indexer_k_norm = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "weight", i), {hparams.indexer_head_size}, flags); + layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM, "bias", i), {hparams.indexer_head_size}, flags); + layer.indexer_proj = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ, "weight", i), {n_embd, hparams.indexer_n_head}, flags); + layer.indexer_attn_k = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K, "weight", i), {n_embd, hparams.indexer_head_size}, flags); + layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags); + if (i < (int) hparams.n_layer_dense_lead) { + layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags); + layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}, flags); + layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, flags); + } else { + layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags); + layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED); + + if (n_expert == 0) { + throw std::runtime_error("n_expert must be > 0"); + } + if (n_expert_used == 0) { + throw std::runtime_error("n_expert_used must be > 0"); + } + + // MoE branch + layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, flags); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, flags); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert}, flags); + + // Shared expert branch + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_exp * n_expert_shared, n_embd}, flags); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags); + } + + // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers + if (hparams.nextn_predict_layers > 0 && static_cast(i) >= n_layer - hparams.nextn_predict_layers) { + layer.nextn.eh_proj = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags); + layer.nextn.enorm = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags); + layer.nextn.hnorm = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags); + + // Optional tensors + layer.nextn.embed_tokens = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED); + layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED); + } + } + } break; case LLM_ARCH_NEMOTRON: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -5985,9 +6194,9 @@ bool llama_model::load_tensors(llama_model_loader & ml) { } break; case LLM_ARCH_WAVTOKENIZER_DEC: { - tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {hparams.n_embd_features, n_vocab}, 0); + tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {hparams.n_embd, n_vocab}, 0); - conv1d = create_tensor(tn(LLM_TENSOR_CONV1D, "weight"), {7, hparams.n_embd_features, hparams.posnet.n_embd}, 0); + conv1d = create_tensor(tn(LLM_TENSOR_CONV1D, "weight"), {7, hparams.n_embd, hparams.posnet.n_embd}, 0); conv1d_b = create_tensor(tn(LLM_TENSOR_CONV1D, "bias"), {1, hparams.posnet.n_embd}, 0); // posnet @@ -6083,8 +6292,8 @@ bool llama_model::load_tensors(llama_model_loader & ml) { output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0); } - output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {hparams.convnext.n_embd, n_embd}, 0); - output_b = create_tensor(tn(LLM_TENSOR_OUTPUT, "bias"), {n_embd}, 0); + output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {hparams.convnext.n_embd, hparams.n_embd_out()}, 0); + output_b = create_tensor(tn(LLM_TENSOR_OUTPUT, "bias"), {hparams.n_embd_out()}, 0); } break; case LLM_ARCH_BAILINGMOE: { @@ -7101,6 +7310,131 @@ bool llama_model::load_tensors(llama_model_loader & ml) { layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { hparams.n_ff_shexp, n_embd }, 0); } } break; + case LLM_ARCH_QWEN35MOE: + { + 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); + } + + const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used; + + // Calculate dimensions from hyperparameters + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t head_v_dim = hparams.ssm_d_state; + const int64_t n_k_heads = hparams.ssm_n_group; + const int64_t n_v_heads = hparams.ssm_dt_rank; + const int64_t key_dim = head_k_dim * n_k_heads; + const int64_t value_dim = head_v_dim * n_v_heads; + const int64_t conv_dim = key_dim * 2 + value_dim; + + 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.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0); + + if (!hparams.is_recurrent(i)) { + // Attention layers + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 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); + + // Q/K normalization for attention layers + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0); + } else { + // Linear attention (gated delta net) specific tensors + // Create tensors with calculated dimensions + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED); + layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED); + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0); + layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0); + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), { hparams.ssm_dt_rank }, 0); + layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0); + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, 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_exp, n_expert }, 0); + layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, 0); + layer.ffn_up_exps = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, 0); + + // Shared experts + const int64_t n_ff_shexp = hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff; + + layer.ffn_gate_inp_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP_SHEXP, "weight", i), { n_embd }, 0); + layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0); + layer.ffn_up_shexp = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, 0); + layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, 0); + } + } break; + case LLM_ARCH_QWEN35: + { + 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); + } + + // Calculate dimensions from hyperparameters + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t head_v_dim = hparams.ssm_d_state; + const int64_t n_k_heads = hparams.ssm_n_group; + const int64_t n_v_heads = hparams.ssm_dt_rank; + const int64_t key_dim = head_k_dim * n_k_heads; + const int64_t value_dim = head_v_dim * n_v_heads; + const int64_t conv_dim = key_dim * 2 + value_dim; + + 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.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, 0); + + if (!hparams.is_recurrent(i)) { + // Attention layers + layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head * 2 }, 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); + + // Q/K normalization for attention layers + layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0); + layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0); + } else { + // Linear attention (gated delta net) specific tensors + // Create tensors with calculated dimensions + layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), { n_embd, key_dim * 2 + value_dim }, TENSOR_NOT_REQUIRED); + layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), { n_embd, value_dim }, TENSOR_NOT_REQUIRED); + layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), { hparams.ssm_d_conv, conv_dim }, 0); + layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), { hparams.ssm_dt_rank }, 0); + layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A_NOSCAN, i), { hparams.ssm_dt_rank }, 0); + layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_alpha = create_tensor(tn(LLM_TENSOR_SSM_ALPHA, "weight", i), { n_embd, n_v_heads }, 0); + layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), { head_v_dim }, 0); + layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), { value_dim, 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); + } + } break; case LLM_ARCH_MIMO2: { tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0); @@ -7545,6 +7879,8 @@ void llama_model::print_info() const { arch == LLM_ARCH_PLAMO2 || arch == LLM_ARCH_GRANITE_HYBRID || arch == LLM_ARCH_QWEN3NEXT || + arch == LLM_ARCH_QWEN35 || + arch == LLM_ARCH_QWEN35MOE || arch == LLM_ARCH_NEMOTRON_H || arch == LLM_ARCH_NEMOTRON_H_MOE) { LLAMA_LOG_INFO("%s: ssm_d_conv = %u\n", __func__, hparams.ssm_d_conv); @@ -7576,7 +7912,7 @@ void llama_model::print_info() const { LLAMA_LOG_INFO("%s: expert_weights_scale = %.1f\n", __func__, hparams.expert_weights_scale); } - if (arch == LLM_ARCH_DEEPSEEK2) { + if (arch == LLM_ARCH_DEEPSEEK2 || arch == LLM_ARCH_GLM_DSA) { LLAMA_LOG_INFO("%s: n_layer_dense_lead = %d\n", __func__, hparams.n_layer_dense_lead); LLAMA_LOG_INFO("%s: n_lora_q = %d\n", __func__, hparams.n_lora_q); LLAMA_LOG_INFO("%s: n_lora_kv = %d\n", __func__, hparams.n_lora_kv); @@ -7776,7 +8112,6 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, 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; @@ -7801,7 +8136,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, /* attn_type_v */ params.type_v, /* attn_v_trans */ !cparams.flash_attn, /* attn_swa_full */ params.swa_full, - /* attn_kv_size */ cparams.n_ctx, + /* attn_kv_size */ cparams.n_ctx_seq, /* attn_n_ubatch */ cparams.n_ubatch, /* attn_n_pad */ 1, /* recurrent_type_r */ GGML_TYPE_F32, @@ -7818,7 +8153,7 @@ llama_memory_i * llama_model::create_memory(const llama_memory_params & params, /* 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_kv_size */ cparams.n_ctx_seq, /* attn_n_pad */ 1, /* attn_n_swa */ hparams.n_swa, /* attn_swa_type */ hparams.swa_type, @@ -8149,6 +8484,7 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { llm = std::make_unique(*this, params); } break; case LLM_ARCH_DEEPSEEK2: + case LLM_ARCH_GLM_DSA: { llm = std::make_unique(*this, params); } break; @@ -8343,6 +8679,14 @@ ggml_cgraph * llama_model::build_graph(const llm_graph_params & params) const { { llm = std::make_unique(*this, params); } break; + case LLM_ARCH_QWEN35: + { + llm = std::make_unique(*this, params); + } break; + case LLM_ARCH_QWEN35MOE: + { + llm = std::make_unique(*this, params); + } break; case LLM_ARCH_MISTRAL3: { llm = std::make_unique(*this, params); @@ -8542,6 +8886,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { case LLM_ARCH_MISTRAL3: case LLM_ARCH_LLAMA_EMBED: case LLM_ARCH_MAINCODER: + case LLM_ARCH_GLM_DSA: return LLAMA_ROPE_TYPE_NORM; // the pairs of head values are offset by n_rot/2 @@ -8611,6 +8956,8 @@ llama_rope_type llama_model_rope_type(const llama_model * model) { return LLAMA_ROPE_TYPE_MROPE; case LLM_ARCH_QWEN3VL: case LLM_ARCH_QWEN3VLMOE: + case LLM_ARCH_QWEN35: + case LLM_ARCH_QWEN35MOE: return LLAMA_ROPE_TYPE_IMROPE; case LLM_ARCH_GLM4: diff --git a/examples/talk-llama/llama-model.h b/examples/talk-llama/llama-model.h index 7b580043..b3505914 100644 --- a/examples/talk-llama/llama-model.h +++ b/examples/talk-llama/llama-model.h @@ -118,6 +118,7 @@ enum llm_type { LLM_TYPE_21B_A3B, // Ernie MoE small LLM_TYPE_30B_A3B, LLM_TYPE_31B_A3_5B, + LLM_TYPE_35B_A3B, // Qwen3.5 LLM_TYPE_48B_A3B, // Kimi Linear LLM_TYPE_80B_A3B, // Qwen3 Next LLM_TYPE_100B_A6B, @@ -129,6 +130,7 @@ enum llm_type { LLM_TYPE_300B_A47B, // Ernie MoE big LLM_TYPE_310B_A15B, // /MiMo-V2-Flash LLM_TYPE_355B_A32B, // GLM-4.5 + LLM_TYPE_744B_A40B, // GLM-5 LLM_TYPE_E2B, LLM_TYPE_E4B, }; @@ -322,6 +324,9 @@ struct llama_layer { // qwen3next struct ggml_tensor * ssm_beta_alpha = nullptr; + // qwen3.5 + struct ggml_tensor * ssm_alpha = nullptr; + // rwkv struct ggml_tensor * time_mix_w1 = nullptr; struct ggml_tensor * time_mix_w2 = nullptr; @@ -425,6 +430,13 @@ struct llama_layer { struct ggml_tensor * ssm_g_b = nullptr; struct ggml_tensor * ssm_o_norm = nullptr; + // DSA (deepseek sparse attention) + struct ggml_tensor * indexer_k_norm = nullptr; + struct ggml_tensor * indexer_k_norm_b = nullptr; + struct ggml_tensor * indexer_proj = nullptr; + struct ggml_tensor * indexer_attn_k = nullptr; + struct ggml_tensor * indexer_attn_q_b = nullptr; // note: for lora a/b, not bias + struct llama_layer_posnet posnet; struct llama_layer_convnext convnext; diff --git a/examples/talk-llama/llama-vocab.cpp b/examples/talk-llama/llama-vocab.cpp index 6d6bdfa0..62e137fb 100644 --- a/examples/talk-llama/llama-vocab.cpp +++ b/examples/talk-llama/llama-vocab.cpp @@ -368,6 +368,13 @@ struct llm_tokenizer_bpe : llm_tokenizer { "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}| ?[^\\s\\p{L}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", }; break; + case LLAMA_VOCAB_PRE_TYPE_QWEN35: + regex_exprs = { + // original regex from tokenizer.json + // "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+" + "(?:'[sS]|'[tT]|'[rR][eE]|'[vV][eE]|'[mM]|'[lL][lL]|'[dD])|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+", + }; + break; case LLAMA_VOCAB_PRE_TYPE_PORO: case LLAMA_VOCAB_PRE_TYPE_BLOOM: case LLAMA_VOCAB_PRE_TYPE_GPT3_FINNISH: @@ -1926,6 +1933,10 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) { tokenizer_pre == "kormo") { pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN2; clean_spaces = false; + } else if ( + tokenizer_pre == "qwen35") { + pre_type = LLAMA_VOCAB_PRE_TYPE_QWEN35; + clean_spaces = false; } else if ( tokenizer_pre == "stablelm2") { pre_type = LLAMA_VOCAB_PRE_TYPE_STABLELM2; diff --git a/examples/talk-llama/llama-vocab.h b/examples/talk-llama/llama-vocab.h index 28c3a82b..718238fb 100644 --- a/examples/talk-llama/llama-vocab.h +++ b/examples/talk-llama/llama-vocab.h @@ -54,6 +54,7 @@ enum llama_vocab_pre_type { LLAMA_VOCAB_PRE_TYPE_SOLAR_OPEN = 43, LLAMA_VOCAB_PRE_TYPE_YOUTU = 44, LLAMA_VOCAB_PRE_TYPE_EXAONE_MOE = 45, + LLAMA_VOCAB_PRE_TYPE_QWEN35 = 46, }; struct LLM_KV; diff --git a/examples/talk-llama/llama.h b/examples/talk-llama/llama.h index bf4e28a8..d2d7f59e 100644 --- a/examples/talk-llama/llama.h +++ b/examples/talk-llama/llama.h @@ -482,7 +482,7 @@ extern "C" { enum llama_params_fit_status { LLAMA_PARAMS_FIT_STATUS_SUCCESS = 0, // found allocations that are projected to fit LLAMA_PARAMS_FIT_STATUS_FAILURE = 1, // could not find allocations that are projected to fit - LLAMA_PARAMS_FIT_STATUS_ERROR = 2, // a hard error occured, e.g. because no model could be found at the specified path + LLAMA_PARAMS_FIT_STATUS_ERROR = 2, // a hard error occurred, e.g. because no model could be found at the specified path }; // fits mparams and cparams to free device memory (assumes system memory is unlimited) @@ -656,21 +656,12 @@ extern "C" { // The following functions operate on a llama_context, hence the naming: llama_verb_... - // Add a loaded LoRA adapter to given context - // This will not modify model's weight - LLAMA_API int32_t llama_set_adapter_lora( + // Set LoRa adapters on the context. Will only modify if the adapters currently in context are different. + LLAMA_API int32_t llama_set_adapters_lora( struct llama_context * ctx, - struct llama_adapter_lora * adapter, - float scale); - - // Remove a specific LoRA adapter from given context - // Return -1 if the adapter is not present in the context - LLAMA_API int32_t llama_rm_adapter_lora( - struct llama_context * ctx, - struct llama_adapter_lora * adapter); - - // Remove all LoRA adapters from given context - LLAMA_API void llama_clear_adapter_lora(struct llama_context * ctx); + struct llama_adapter_lora ** adapters, + size_t n_adapters, + float * scales); // Apply a loaded control vector to a llama_context, or if data is NULL, clear // the currently loaded vector. @@ -678,7 +669,7 @@ extern "C" { // to an n_embd x n_layers buffer starting from layer 1. // il_start and il_end are the layer range the vector should apply to (both inclusive) // See llama_control_vector_load in common to load a control vector. - LLAMA_API int32_t llama_apply_adapter_cvec( + LLAMA_API int32_t llama_set_adapter_cvec( struct llama_context * ctx, const float * data, size_t len, @@ -1150,9 +1141,9 @@ extern "C" { // /// Apply chat template. Inspired by hf apply_chat_template() on python. - /// Both "model" and "custom_template" are optional, but at least one is required. "custom_template" has higher precedence than "model" + /// /// NOTE: This function does not use a jinja parser. It only support a pre-defined list of template. See more: https://github.com/ggml-org/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template - /// @param tmpl A Jinja template to use for this chat. If this is nullptr, the model’s default chat template will be used instead. + /// @param tmpl A Jinja template to use for this chat. /// @param chat Pointer to a list of multiple llama_chat_message /// @param n_msg Number of llama_chat_message in this chat /// @param add_ass Whether to end the prompt with the token(s) that indicate the start of an assistant message. diff --git a/examples/talk-llama/models/deepseek2.cpp b/examples/talk-llama/models/deepseek2.cpp index 987f4499..b2c1f160 100644 --- a/examples/talk-llama/models/deepseek2.cpp +++ b/examples/talk-llama/models/deepseek2.cpp @@ -45,7 +45,8 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr ggml_tensor * inp_out_ids = build_inp_out_ids(); - for (int il = 0; il < n_layer; ++il) { + int effective_n_layers = hparams.n_layer - hparams.nextn_predict_layers; + for (int il = 0; il < effective_n_layers; ++il) { ggml_tensor * inpSA = inpL; // norm @@ -188,7 +189,7 @@ llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_gr Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); } } - if (il == n_layer - 1 && inp_out_ids) { + if (il == effective_n_layers - 1 && inp_out_ids) { cur = ggml_get_rows(ctx0, cur, inp_out_ids); inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); } diff --git a/examples/talk-llama/models/kimi-linear.cpp b/examples/talk-llama/models/kimi-linear.cpp index 0f037d1a..942844d0 100644 --- a/examples/talk-llama/models/kimi-linear.cpp +++ b/examples/talk-llama/models/kimi-linear.cpp @@ -41,8 +41,11 @@ static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_t 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_x, - ggml_view_1d(ctx0, conv_states_all, conv_state_size * n_seqs, - (kv_head * n_embd_r_total + qkv * conv_state_size) * ggml_element_size(conv_states_all)))); + ggml_view_3d(ctx0, conv_states_all, + d_conv - 1, d_inner, n_seqs, + (d_conv - 1) * ggml_element_size(conv_states_all), // nb1: contiguous within one channel's conv taps + n_embd_r_total * ggml_element_size(conv_states_all), // nb2: stride between sequences (skip over K,V states) + (kv_head * n_embd_r_total + qkv * conv_state_size) * ggml_element_size(conv_states_all)))); // offset to first seq's Q/K/V state // Reshape conv weight: GGUF [d_conv, 1, d_inner, 1] -> ggml_ssm_conv expects [d_conv, d_inner] // GGUF stores as [d_conv, 1, d_inner, 1] with memory layout w[conv_step + channel * d_conv] // vLLM stores as [d_inner, d_conv] with memory layout w[channel * d_conv + conv_step] diff --git a/examples/talk-llama/models/models.h b/examples/talk-llama/models/models.h index cfcbb9aa..ec6f80e5 100644 --- a/examples/talk-llama/models/models.h +++ b/examples/talk-llama/models/models.h @@ -476,6 +476,7 @@ struct llm_build_qwen3vl : public llm_graph_context { struct llm_build_qwen3vlmoe : public llm_graph_context { llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params); }; + struct llm_build_qwen3next : public llm_graph_context_mamba { llm_build_qwen3next(const llama_model & model, const llm_graph_params & params); private: @@ -485,6 +486,118 @@ private: ggml_tensor * inp_pos, int il); + ggml_tensor * build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + int il); + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + int il); + + // returns pair of output and new state + std::pair build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il); + + // returns pair of output and new state + std::pair build_delta_net_autoregressive( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il); + + ggml_tensor * build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer); + + // returns pair of qkv, z + std::pair build_qkvz( + ggml_tensor * input, + int il); + + const llama_model & model; +}; + +struct llm_build_qwen35 : public llm_graph_context_mamba { + llm_build_qwen35(const llama_model & model, const llm_graph_params & params); +private: + ggml_tensor * build_layer_attn( + llm_graph_input_attn_kv * inp_attn, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int * sections, + int il); + + ggml_tensor * build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il); + + ggml_tensor * build_layer_ffn( + ggml_tensor * cur, + int il); + + // returns pair of output and new state + std::pair build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il); + + // returns pair of output and new state + std::pair build_delta_net_autoregressive( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il); + + ggml_tensor * build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer); + + // returns pair of qkv, z + std::pair build_qkvz( + ggml_tensor * input, + int il); + + const llama_model & model; +}; + +struct llm_build_qwen35moe : public llm_graph_context_mamba { + llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params); +private: + ggml_tensor * build_layer_attn( + llm_graph_input_attn_kv * inp_attn, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int * sections, + int il); + ggml_tensor * build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, diff --git a/examples/talk-llama/models/qwen35.cpp b/examples/talk-llama/models/qwen35.cpp new file mode 100644 index 00000000..592c1704 --- /dev/null +++ b/examples/talk-llama/models/qwen35.cpp @@ -0,0 +1,740 @@ +#include "ggml.h" +#include "models.h" + +#define CHUNK_SIZE 64 + +llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params), model(model) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + cb(inpL, "model.input_embed", -1); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_pos = build_inp_pos(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + ggml_tensor * causal_mask = + ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f), + GGML_TRI_TYPE_LOWER); + + ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f)); + ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity); + + ggml_build_forward_expand(gf, causal_mask); + ggml_build_forward_expand(gf, identity); + ggml_build_forward_expand(gf, diag_mask); + + 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); + + // Determine layer type and build appropriate attention mechanism + if (hparams.is_recurrent(il)) { + // Linear attention layer (gated delta net) + cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il); + } else { + // Full attention layer + cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, 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); + } + + // Residual connection + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "attn_residual", il); + + // Save the tensor before post-attention norm for residual connection + ggml_tensor * ffn_residual = cur; + + // Post-attention norm + ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il); + cb(attn_post_norm, "attn_post_norm", il); + + // Dense FFN layer - without residual connection + cur = build_layer_ffn(attn_post_norm, il); + cb(cur, "ffn_out", il); + + // Residual connection for FFN - add to the tensor from before post_attention_layernorm + cur = ggml_add(ctx0, cur, ffn_residual); + cb(cur, "post_ffn", il); + + // Input for next layer + inpL = cur; + } + cur = inpL; + + // Final norm + 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); +} + +// utility to get one slice from the third dimension +// input dim: [x, y, c, b] +// output dim: [x, y, 1, b] +static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) { + return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3], + t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c); +} + +std::pair llm_build_qwen35::build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il) { + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + + beta = ggml_sigmoid(ctx0, beta); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs); + + beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + cb(q, "q_perm", il); + cb(k, "k_perm", il); + cb(v, "v_perm", il); + cb(beta, "beta_perm", il); + cb(g, "g_perm", il); + cb(state, "state_in", il); + + GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs); + GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs); + + // Do padding + const int64_t chunk_size = CHUNK_SIZE; + + const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size; + const int64_t n_chunks = (n_tokens + pad) / chunk_size; + + q = ggml_pad(ctx0, q, 0, pad, 0, 0); + k = ggml_pad(ctx0, k, 0, pad, 0, 0); + v = ggml_pad(ctx0, v, 0, pad, 0, 0); + g = ggml_pad(ctx0, g, pad, 0, 0, 0); + beta = ggml_pad(ctx0, beta, 0, pad, 0, 0); + + cb(q, "q_pad", il); + cb(k, "k_pad", il); + cb(v, "v_pad", il); + cb(beta, "beta_pad", il); + cb(g, "g_pad", il); + + ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); + ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); + + cb(v_beta, "v_beta", il); + cb(k_beta, "k_beta", il); + + q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs); + k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs); + k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs); + v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs); + v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs); + + g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs); + beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs); + + ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); + cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs); + ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * gcs_j_broadcast = + ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); + cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + decay_mask = ggml_exp(ctx0, decay_mask); + decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + + ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); + + ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); + ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask)); + cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask); + ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower); + + ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); + attn = ggml_mul(ctx0, lin_solve, causal_mask); + attn = ggml_add(ctx0, attn, identity); + cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn); + + ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); + ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); + + ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); + cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * k_cumdecay = + ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); + cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q); + attn_kq = ggml_mul(ctx0, attn_kq, decay_mask); + attn_kq = ggml_mul(ctx0, attn_kq, diag_mask); + cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + + // vectorized calculation of key_gdiff + // improved from the chunked version: + // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1) + // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp() + // key_gdiff = key * g_diff.unsqueeze(-1) + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + + // get last element in g_cumsum along chunk_size dimension (ne0) + // example: [[x, y, z, ..., last], ...] -> [[last], ...] + ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3], + g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3], + (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum)); + g_last = ggml_cont(ctx0, g_last); + cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last); + cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last)); + cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); + ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp, + 1, chunk_size, n_chunks, g_diff_exp->ne[3]); + + ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t); + cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)); + cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs) + + // state to be updated per chunk + ggml_tensor * new_state = state; // ggml_dup(ctx0, state); + cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs) + + // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * core_attn_out = nullptr; + + for (int64_t chunk = 0; chunk < n_chunks; chunk++) { + // shape: (S_k, chunk_size, 1, H_k * n_seqs) + ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul + + // shape: (S_v, chunk_size, 1, H_v * n_seqs) + ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat + + // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul + + // shape: (chunk_size, 1, H_v * n_seqs) + ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat + + // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) + // replaced by precomputed attn_kq + ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk); + cb(attn_chunk, "attn_chunk", il); + + ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs); + + // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state + ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk); + cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs) + + // v_new = v_i - v_prime + ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime); + ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new)); + cb(v_new, "v_new_chunk", il); + + // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state + ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk); + ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp); + cb(attn_inter, "attn_inter_chunk", il); + + // core_attn_out[:, :, i] = attn_inter + attn @ v_new + ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk); + cb(v_attn, "v_attn_chunk", il); + + ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn); + cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs) + + core_attn_out = core_attn_out == nullptr + ? core_attn_out_chunk + : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2); + + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk); + //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why? + ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t); + + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk)); + new_state = ggml_add(ctx0, + ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)), + ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs)); + } + + // truncate padded tokens + ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, + S_v, n_tokens, H_v, n_seqs, + ggml_row_size(core_attn_out->type, S_v), + ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks), + ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0); + output_tokens = ggml_cont(ctx0, output_tokens); + cb(output_tokens, "output_tokens", il); + + // permute back to (S_v, H_v, n_tokens, n_seqs) + output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3); + output_tokens = ggml_cont(ctx0, output_tokens); + + return {output_tokens, new_state}; +} + +std::pair llm_build_qwen35::build_delta_net_autoregressive( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il) { + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + beta = ggml_sigmoid(ctx0, beta); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs); + ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs); + + // Apply exponential to g_t + g_t = ggml_exp(ctx0, g_t); + + // Apply the gated delta rule for the single timestep + // last_recurrent_state = last_recurrent_state * g_t + state = ggml_mul(ctx0, state, g_t); + + // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2) + ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs); + ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed); + // we need to sum over dim=-2, so we transpose, sum, then transpose again + kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem)))); + + // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v) + ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs); + // delta = (v_t - kv_mem) * beta_t + ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs] + ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t); + + // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta + ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta); + state = ggml_add(ctx0, state, k_t_delta); + + // Compute the attention output + // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2) + ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t + ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed); + // again, since it's over dim = -2, transpose, sum, transpose back + ggml_tensor * core_attn_out = + ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q)))); + + // core_attn_out should be [S_v, 1, H_v, n_seqs] after this + cb(core_attn_out, "output_tokens", il); + cb(state, "new_state", il); + + return {core_attn_out, state}; +} + +std::pair llm_build_qwen35::build_qkvz( + ggml_tensor * input, + int il) { + const int64_t n_seqs = ubatch.n_seqs; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input); + qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs); + cb(qkv_mixed, "linear_attn_qkv_mixed", il); + + ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input); + cb(z, "z", il); + + return { qkv_mixed, z }; +} + +ggml_tensor * llm_build_qwen35::build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer) { + ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer); + ggml_tensor * gated_silu = ggml_silu(ctx0, gate); + + return ggml_mul(ctx0, normalized, gated_silu); +} + +ggml_tensor * llm_build_qwen35::build_layer_attn( + llm_graph_input_attn_kv * inp, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int * sections, + int il) { + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention + + // Qwen3Next uses a single Q projection that outputs query + gate + ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ] + cb(Qcur_full, "Qcur_full", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, 0); + cb(Qcur, "Qcur_reshaped", il); + + // Apply Q normalization + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", 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); + + // Apply K normalization + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, + ggml_element_size(Qcur_full) * n_embd_head); + gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); + cb(gate, "gate_reshaped", il); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // Apply MRoPE + 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); + + // Attention computation + 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, + nullptr, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_pregate", il); + + ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); + cb(gate_sigmoid, "gate_sigmoid", il); + + cur = ggml_mul(ctx0, cur, gate_sigmoid); + cb(cur, "attn_gated", il); + + cur = build_lora_mm(model.layers[il].wo, cur); + cb(cur, "attn_output", il); + + return cur; +} + +ggml_tensor * llm_build_qwen35::build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il) { + const auto * mctx_cur = inp->mctx; + + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t n_seqs = ubatch.n_seqs; + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t num_k_heads = hparams.ssm_n_group; + const int64_t num_v_heads = hparams.ssm_dt_rank; + const int64_t head_v_dim = d_inner / num_v_heads; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + const auto kv_head = mctx_cur->get_head(); + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + // Input projections + auto qkvz = build_qkvz(cur, il); + ggml_tensor * qkv_mixed = qkvz.first; + ggml_tensor * z = qkvz.second; + + ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur); + beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs); + cb(beta, "beta", il); + ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur); + alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs); + cb(alpha, "alpha", il); + + ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); + ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); + cb(alpha_softplus, "a_softplus", il); + ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus + cb(gate, "gate", il); + + // Get convolution states from cache + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state(); + + // Build the convolution states tensor + ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); + cb(conv_states, "conv_states", il); + + // Calculate convolution kernel size + ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; + const int64_t conv_kernel_size = conv_kernel->ne[0]; + const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state; + conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); + cb(conv_states, "conv_states_reshaped", il); + + qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3); + cb(qkv_mixed, "qkv_mixed_permuted", il); + + ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0); + cb(conv_input, "conv_input", il); + + // Update convolution state cache + // Extract the last (conv_kernel_size - 1) states from conv_input + ggml_tensor * last_conv_states = + ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1], + conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input)); + cb(last_conv_states, "last_conv_states", il); + + ggml_tensor * state_update_target = + ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs, + kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all)); + cb(state_update_target, "state_update_target", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target)); + cb(conv_states_all, "conv_states_updated", il); + + // Apply SSM convolution + ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel); + cb(conv_output_proper, "conv_output_raw", il); + + ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper); + cb(conv_output_silu, "conv_output_silu", il); + + ggml_tensor * conv_qkv_mix = conv_output_silu; + + // Calculate the total conv dimension + int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads; + int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim); + + // Extract the convolved Q, K, V from conv_output + ggml_tensor * q_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0); + cb(q_conv, "q_conv", il); + ggml_tensor * k_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, + head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(k_conv, "k_conv", il); + ggml_tensor * v_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv, + 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(v_conv, "v_conv", il); + + // Unsqueeze them + q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); + + ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); + state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs); + cb(state, "state_predelta", il); + + // if head keys and value keys are different, repeat Q/K to match V's head count + // V heads are in tiled order (from conversion), so simple tiled repeat works + if (num_k_heads != num_v_heads) { + GGML_ASSERT(num_v_heads % num_k_heads == 0); + q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); + k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); + } + + cb(q_conv, "q_conv_predelta", il); + cb(k_conv, "k_conv_predelta", il); + cb(v_conv, "v_conv_predelta", il); + + // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens + std::pair attn_out; // pair of (output, new_state) + if (n_seq_tokens == 1) { + attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il); + } else { + attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il); + } + ggml_tensor * output = attn_out.first; + ggml_tensor * new_state = attn_out.second; + cb(output, "attn_output", il); + cb(new_state, "new_state", il); + + // Update the recurrent states + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, new_state, + ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, + kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); + + // Reshape both attn_out_final and z to 2D tensors for normalization + // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // Apply gated normalization: self.norm(core_attn_out, z) + ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il); + + // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim] + ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs); + cb(final_output, "final_output", il); + + // Output projection + cur = build_lora_mm(model.layers[il].ssm_out, final_output); + cb(cur, "linear_attn_out", il); + + // Reshape back to original dimensions + cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + return cur; +} + +ggml_tensor * llm_build_qwen35::build_layer_ffn(ggml_tensor * cur, const int il) { + // Qwen3.5 does not use MoE FFN + GGML_ASSERT(model.layers[il].ffn_gate_inp == nullptr); + + 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); + + return cur; +} diff --git a/examples/talk-llama/models/qwen35moe.cpp b/examples/talk-llama/models/qwen35moe.cpp new file mode 100644 index 00000000..0db8f825 --- /dev/null +++ b/examples/talk-llama/models/qwen35moe.cpp @@ -0,0 +1,774 @@ +#include "ggml.h" +#include "models.h" + +#define CHUNK_SIZE 64 + +llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_graph_params & params) : + llm_graph_context_mamba(params), model(model) { + const int64_t n_embd_head = hparams.n_embd_head_v; + + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + int sections[4]; + std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections); + + ggml_tensor * cur; + ggml_tensor * inpL; + + inpL = build_inp_embd(model.tok_embd); + + cb(inpL, "model.input_embed", -1); + + auto * inp = build_inp_mem_hybrid(); + + ggml_tensor * inp_pos = build_inp_pos(); + ggml_tensor * inp_out_ids = build_inp_out_ids(); + + ggml_tensor * causal_mask = + ggml_tri(ctx0, ggml_fill(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f), + GGML_TRI_TYPE_LOWER); + + ggml_tensor * identity = ggml_diag(ctx0, ggml_fill(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f)); + ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity); + + ggml_build_forward_expand(gf, causal_mask); + ggml_build_forward_expand(gf, identity); + ggml_build_forward_expand(gf, diag_mask); + + 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); + + // Determine layer type and build appropriate attention mechanism + if (hparams.is_recurrent(il)) { + // Linear attention layer (gated delta net) + cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il); + } else { + // Full attention layer + cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, 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); + } + + // Residual connection + cur = ggml_add(ctx0, cur, inpSA); + cb(cur, "attn_residual", il); + + // Save the tensor before post-attention norm for residual connection + ggml_tensor * ffn_residual = cur; + + // Post-attention norm + ggml_tensor * attn_post_norm = build_norm(cur, model.layers[il].attn_post_norm, nullptr, LLM_NORM_RMS, il); + cb(attn_post_norm, "attn_post_norm", il); + + // MOE FFN layer + cur = build_layer_ffn(attn_post_norm, il); + cb(cur, "ffn_out", il); + + // Residual connection for FFN - add to the tensor from before post_attention_layernorm + cur = ggml_add(ctx0, cur, ffn_residual); + cb(cur, "post_moe", il); + + // Input for next layer + inpL = cur; + } + cur = inpL; + + // Final norm + 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); +} + +// utility to get one slice from the third dimension +// input dim: [x, y, c, b] +// output dim: [x, y, 1, b] +static ggml_tensor * get_slice_2d(ggml_context * ctx0, ggml_tensor * t, int64_t c) { + return ggml_view_4d(ctx0, t, t->ne[0], t->ne[1], 1, t->ne[3], + t->nb[1], t->nb[2], t->nb[3], t->nb[2] * c); +} + +std::pair llm_build_qwen35moe::build_delta_net_chunking( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il) { + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + + beta = ggml_sigmoid(ctx0, beta); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); + g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs); + + beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + cb(q, "q_perm", il); + cb(k, "k_perm", il); + cb(v, "v_perm", il); + cb(beta, "beta_perm", il); + cb(g, "g_perm", il); + cb(state, "state_in", il); + + GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs); + GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs); + + // Do padding + const int64_t chunk_size = CHUNK_SIZE; + + const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size; + const int64_t n_chunks = (n_tokens + pad) / chunk_size; + + q = ggml_pad(ctx0, q, 0, pad, 0, 0); + k = ggml_pad(ctx0, k, 0, pad, 0, 0); + v = ggml_pad(ctx0, v, 0, pad, 0, 0); + g = ggml_pad(ctx0, g, pad, 0, 0, 0); + beta = ggml_pad(ctx0, beta, 0, pad, 0, 0); + + cb(q, "q_pad", il); + cb(k, "k_pad", il); + cb(v, "v_pad", il); + cb(beta, "beta_pad", il); + cb(g, "g_pad", il); + + ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); + ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); + + cb(v_beta, "v_beta", il); + cb(k_beta, "k_beta", il); + + q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs); + k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs); + k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs); + v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs); + v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs); + + g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs); + beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs); + + ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); + cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs); + ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * gcs_j_broadcast = + ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs); + + ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); + cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + decay_mask = ggml_exp(ctx0, decay_mask); + decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + + ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); + + ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); + ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask)); + cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask); + ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower); + + ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); + attn = ggml_mul(ctx0, lin_solve, causal_mask); + attn = ggml_add(ctx0, attn, identity); + cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn); + + ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); + ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); + + ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); + cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * k_cumdecay = + ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); + cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q); + attn_kq = ggml_mul(ctx0, attn_kq, decay_mask); + attn_kq = ggml_mul(ctx0, attn_kq, diag_mask); + cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + + + // vectorized calculation of key_gdiff + // improved from the chunked version: + // g_last = torch.clamp(g_cum[:, :, -1], max=50.0).exp().unsqueeze(-1).unsqueeze(-1) + // g_diff = torch.clamp(g_cum[:, :, -1:] - g_cum, max=50.0).exp() + // key_gdiff = key * g_diff.unsqueeze(-1) + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + + // get last element in g_cumsum along chunk_size dimension (ne0) + // example: [[x, y, z, ..., last], ...] -> [[last], ...] + ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3], + g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3], + (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum)); + g_last = ggml_cont(ctx0, g_last); + cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last); + cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last)); + cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + + ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); + ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp, + 1, chunk_size, n_chunks, g_diff_exp->ne[3]); + + ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t); + cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + + ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)); + cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs) + + + // state to be updated per chunk + ggml_tensor * new_state = state; // ggml_dup(ctx0, state); + cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs) + + // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * core_attn_out = nullptr; + + for (int64_t chunk = 0; chunk < n_chunks; chunk++) { + // shape: (S_k, chunk_size, 1, H_k * n_seqs) + ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul + + // shape: (S_v, chunk_size, 1, H_v * n_seqs) + ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat + + // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul + + // shape: (chunk_size, 1, H_v * n_seqs) + ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat + + // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) + // replaced by precomputed attn_kq + ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk); + cb(attn_chunk, "attn_chunk", il); + + ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs); + + // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state + ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk); + cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs) + + // v_new = v_i - v_prime + ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime); + ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new)); + cb(v_new, "v_new_chunk", il); + + // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state + ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk); + ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp); + cb(attn_inter, "attn_inter_chunk", il); + + // core_attn_out[:, :, i] = attn_inter + attn @ v_new + ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk); + cb(v_attn, "v_attn_chunk", il); + + ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn); + cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs) + + core_attn_out = core_attn_out == nullptr + ? core_attn_out_chunk + : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2); + + // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new + ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk); + //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why? + ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t); + + // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew + ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk)); + new_state = ggml_add(ctx0, + ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)), + ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs)); + } + + // truncate padded tokens + ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, + S_v, n_tokens, H_v, n_seqs, + ggml_row_size(core_attn_out->type, S_v), + ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks), + ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0); + output_tokens = ggml_cont(ctx0, output_tokens); + cb(output_tokens, "output_tokens", il); + + // permute back to (S_v, H_v, n_tokens, n_seqs) + output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3); + output_tokens = ggml_cont(ctx0, output_tokens); + + return {output_tokens, new_state}; +} + +std::pair llm_build_qwen35moe::build_delta_net_autoregressive( + ggml_tensor * q, + ggml_tensor * k, + ggml_tensor * v, + ggml_tensor * g, + ggml_tensor * beta, + ggml_tensor * state, + int il) { + const int64_t S_k = q->ne[0]; + const int64_t H_k = q->ne[1]; + const int64_t n_tokens = q->ne[2]; + const int64_t n_seqs = q->ne[3]; + + const int64_t S_v = v->ne[0]; + const int64_t H_v = v->ne[1]; + + GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing + GGML_ASSERT(v->ne[2] == n_tokens); + GGML_ASSERT(k->ne[2] == n_tokens); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); + GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + + GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); + GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + + GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case + + const float eps_norm = hparams.f_norm_rms_eps; + + q = ggml_l2_norm(ctx0, q, eps_norm); + k = ggml_l2_norm(ctx0, k, eps_norm); + + const float scale = 1.0f / sqrtf(S_v); + + q = ggml_scale(ctx0, q, scale); + beta = ggml_sigmoid(ctx0, beta); + + cb(q, "q_in", il); + cb(k, "k_in", il); + cb(v, "v_in", il); + cb(beta, "beta_in", il); + cb(g, "g_in", il); + + state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + + ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs); + ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs); + + // Apply exponential to g_t + g_t = ggml_exp(ctx0, g_t); + + // Apply the gated delta rule for the single timestep + // last_recurrent_state = last_recurrent_state * g_t + state = ggml_mul(ctx0, state, g_t); + + // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2) + ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs); + ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed); + // we need to sum over dim=-2, so we transpose, sum, then transpose again + kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem)))); + + // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v) + ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs); + // delta = (v_t - kv_mem) * beta_t + ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs] + ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t); + + // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta + ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta); + state = ggml_add(ctx0, state, k_t_delta); + + // Compute the attention output + // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2) + ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t + ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed); + // again, since it's over dim = -2, transpose, sum, transpose back + ggml_tensor * core_attn_out = + ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q)))); + + // core_attn_out should be [S_v, 1, H_v, n_seqs] after this + cb(core_attn_out, "output_tokens", il); + cb(state, "new_state", il); + + return {core_attn_out, state}; +} + +std::pair llm_build_qwen35moe::build_qkvz( + ggml_tensor * input, + int il) { + const int64_t n_seqs = ubatch.n_seqs; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + ggml_tensor * qkv_mixed = build_lora_mm(model.layers[il].wqkv, input); + qkv_mixed = ggml_reshape_3d(ctx0, qkv_mixed, qkv_mixed->ne[0], n_seq_tokens, n_seqs); + cb(qkv_mixed, "linear_attn_qkv_mixed", il); + + ggml_tensor * z = build_lora_mm(model.layers[il].wqkv_gate, input); + cb(z, "z", il); + + return { qkv_mixed, z }; +} + +ggml_tensor * llm_build_qwen35moe::build_norm_gated( + ggml_tensor * input, + ggml_tensor * weights, + ggml_tensor * gate, + int layer) { + ggml_tensor * normalized = build_norm(input, weights, nullptr, LLM_NORM_RMS, layer); + ggml_tensor * gated_silu = ggml_silu(ctx0, gate); + + return ggml_mul(ctx0, normalized, gated_silu); +} + +ggml_tensor * llm_build_qwen35moe ::build_layer_attn( + llm_graph_input_attn_kv * inp, + ggml_tensor * cur, + ggml_tensor * inp_pos, + int * sections, + int il) { + const int64_t n_embd_head = hparams.n_embd_head_v; + GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); + + // Order: joint QG projection, QG split, Q norm, KV projection, K norm, RoPE, attention + + // Qwen3Next uses a single Q projection that outputs query + gate + ggml_tensor * Qcur_full = build_lora_mm(model.layers[il].wq, cur); // [ (n_embd_head * 2) * n_head, n_tokens ] + cb(Qcur_full, "Qcur_full", il); + + ggml_tensor * Qcur = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, 0); + cb(Qcur, "Qcur_reshaped", il); + + // Apply Q normalization + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", 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); + + // Apply K normalization + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); + + ggml_tensor * gate = ggml_view_3d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, + ggml_element_size(Qcur_full) * n_embd_head * 2, + ggml_element_size(Qcur_full) * n_embd_head * 2 * n_head, + ggml_element_size(Qcur_full) * n_embd_head); + gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); + cb(gate, "gate_reshaped", il); + + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + + // Apply IMRoPE + 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); + + // Attention computation + 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, + nullptr, nullptr, + Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); + cb(cur, "attn_pregate", il); + + ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); + cb(gate_sigmoid, "gate_sigmoid", il); + + cur = ggml_mul(ctx0, cur, gate_sigmoid); + cb(cur, "attn_gated", il); + + cur = build_lora_mm(model.layers[il].wo, cur); + cb(cur, "attn_output", il); + + return cur; +} + +ggml_tensor * llm_build_qwen35moe ::build_layer_attn_linear( + llm_graph_input_rs * inp, + ggml_tensor * cur, + ggml_tensor * causal_mask, + ggml_tensor * identity, + ggml_tensor * diag_mask, + int il) { + const auto * mctx_cur = inp->mctx; + + const int64_t d_inner = hparams.ssm_d_inner; + const int64_t n_seqs = ubatch.n_seqs; + const int64_t head_k_dim = hparams.ssm_d_state; + const int64_t num_k_heads = hparams.ssm_n_group; + const int64_t num_v_heads = hparams.ssm_dt_rank; + const int64_t head_v_dim = d_inner / num_v_heads; + const int64_t n_seq_tokens = ubatch.n_seq_tokens; + + const auto kv_head = mctx_cur->get_head(); + + GGML_ASSERT(n_seqs != 0); + GGML_ASSERT(ubatch.equal_seqs()); + GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs); + + // Input projections + auto qkvz = build_qkvz(cur, il); + ggml_tensor * qkv_mixed = qkvz.first; + ggml_tensor * z = qkvz.second; + + ggml_tensor * beta = build_lora_mm(model.layers[il].ssm_beta, cur); + beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs); + cb(beta, "beta", il); + ggml_tensor * alpha = build_lora_mm(model.layers[il].ssm_alpha, cur); + alpha = ggml_cont_3d(ctx0, alpha, num_v_heads, n_seq_tokens, n_seqs); + cb(alpha, "alpha", il); + + ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); + ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); + cb(alpha_softplus, "a_softplus", il); + ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus + cb(gate, "gate", il); + + // Get convolution states from cache + ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); + ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); + + // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state(); + + // Build the convolution states tensor + ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); + cb(conv_states, "conv_states", il); + + // Calculate convolution kernel size + ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; + const int64_t conv_kernel_size = conv_kernel->ne[0]; + const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state; + conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); + cb(conv_states, "conv_states_reshaped", il); + + qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3); + cb(qkv_mixed, "qkv_mixed_permuted", il); + + ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0); + cb(conv_input, "conv_input", il); + + // Update convolution state cache + // Extract the last (conv_kernel_size - 1) states from conv_input + ggml_tensor * last_conv_states = + ggml_view_3d(ctx0, conv_input, conv_kernel_size - 1, conv_channels, n_seqs, conv_input->nb[1], + conv_input->nb[2], (conv_input->ne[0] - conv_states->ne[0]) * ggml_element_size(conv_input)); + cb(last_conv_states, "last_conv_states", il); + + ggml_tensor * state_update_target = + ggml_view_1d(ctx0, conv_states_all, (conv_kernel_size - 1) * conv_channels * n_seqs, + kv_head * (conv_kernel_size - 1) * conv_channels * ggml_element_size(conv_states_all)); + cb(state_update_target, "state_update_target", il); + + ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target)); + cb(conv_states_all, "conv_states_updated", il); + + // Apply SSM convolution + ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel); + cb(conv_output_proper, "conv_output_raw", il); + + ggml_tensor * conv_output_silu = ggml_silu(ctx0, conv_output_proper); + cb(conv_output_silu, "conv_output_silu", il); + + ggml_tensor * conv_qkv_mix = conv_output_silu; + + // Calculate the total conv dimension + int64_t qkv_dim = head_k_dim * num_k_heads * 2 + head_v_dim * num_v_heads; + int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim); + + // Extract the convolved Q, K, V from conv_output + ggml_tensor * q_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0); + cb(q_conv, "q_conv", il); + ggml_tensor * k_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, + head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(k_conv, "k_conv", il); + ggml_tensor * v_conv = + ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv, + 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + cb(v_conv, "v_conv", il); + + // Unsqueeze them + q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); + + ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); + state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs); + cb(state, "state_predelta", il); + + // if head keys and value keys are different, repeat Q/K to match V's head count + // V heads are in tiled order (from conversion), so simple tiled repeat works + if (num_k_heads != num_v_heads) { + GGML_ASSERT(num_v_heads % num_k_heads == 0); + q_conv = ggml_repeat_4d(ctx0, q_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); + k_conv = ggml_repeat_4d(ctx0, k_conv, head_k_dim, num_v_heads, n_seq_tokens, n_seqs); + } + + cb(q_conv, "q_conv_predelta", il); + cb(k_conv, "k_conv_predelta", il); + cb(v_conv, "v_conv_predelta", il); + + // Choose between build_delta_net_chunking, build_delta_net_recurrent, and build_delta_net_autoregressive based on n_tokens + std::pair attn_out; // pair of (output, new_state) + if (n_seq_tokens == 1) { + attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il); + } else { + attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il); + } + ggml_tensor * output = attn_out.first; + ggml_tensor * new_state = attn_out.second; + cb(output, "attn_output", il); + cb(new_state, "new_state", il); + + // Update the recurrent states + ggml_build_forward_expand(gf, + ggml_cpy(ctx0, new_state, + ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, + kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); + + // Reshape both attn_out_final and z to 2D tensors for normalization + // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] + ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + + // Apply gated normalization: self.norm(core_attn_out, z) + ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il); + + // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim] + ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs); + cb(final_output, "final_output", il); + + // Output projection + cur = build_lora_mm(model.layers[il].ssm_out, final_output); + cb(cur, "linear_attn_out", il); + + // Reshape back to original dimensions + cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + return cur; +} + +ggml_tensor * llm_build_qwen35moe ::build_layer_ffn(ggml_tensor * cur, const int il) { + // Check if this is an MoE layer + GGML_ASSERT(model.layers[il].ffn_gate_inp != nullptr); + + 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); + + // Add shared experts if present - following Qwen3Next reference implementation + if (model.layers[il].ffn_up_shexp != nullptr) { + 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); + + // Apply shared expert gating as in the reference implementation + // The shared expert has its own gate that is sigmoided + // Note: ffn_gate_inp_shexp is the shared expert gate (outputs 1 value per token) + ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); + cb(shared_gate, "shared_expert_gate", il); + + // Apply sigmoid to the gate + shared_gate = ggml_sigmoid(ctx0, shared_gate); + cb(shared_gate, "shared_expert_gate_sigmoid", il); + + + // Apply the gate to the shared expert output + ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate); + cb(ffn_shexp, "ffn_shexp_gated", il); + + cur = ggml_add(ctx0, moe_out, ffn_shexp); + cb(cur, "ffn_out", il); + } else { + cur = moe_out; + } + + return cur; +} diff --git a/examples/talk-llama/models/qwen3next.cpp b/examples/talk-llama/models/qwen3next.cpp index 99b1a76a..aea8b295 100644 --- a/examples/talk-llama/models/qwen3next.cpp +++ b/examples/talk-llama/models/qwen3next.cpp @@ -16,17 +16,6 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr ggml_tensor * inp_pos = build_inp_pos(); ggml_tensor * inp_out_ids = build_inp_out_ids(); - ggml_tensor * causal_mask = - ggml_tri(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, CHUNK_SIZE, CHUNK_SIZE), 1.0f), - GGML_TRI_TYPE_LOWER); - - ggml_tensor * identity = ggml_diag(ctx0, ggml_fill_inplace(ctx0, ggml_new_tensor_1d(ctx0, GGML_TYPE_F32, CHUNK_SIZE), 1.0f)); - ggml_tensor * diag_mask = ggml_add(ctx0, causal_mask, identity); - - ggml_build_forward_expand(gf, causal_mask); - ggml_build_forward_expand(gf, identity); - ggml_build_forward_expand(gf, diag_mask); - for (int il = 0; il < n_layer; ++il) { ggml_tensor * inpSA = inpL; @@ -36,7 +25,7 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr // Determine layer type and build appropriate attention mechanism if (hparams.is_recurrent(il)) { // Linear attention layer (gated delta net) - cur = build_layer_attn_linear(inp->get_recr(), cur, causal_mask, identity, diag_mask, il); + cur = build_layer_attn_linear(inp->get_recr(), cur, il); } else { // Full attention layer cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il); @@ -99,11 +88,8 @@ std::pair llm_build_qwen3next::build_delta_net_chu ggml_tensor * k, ggml_tensor * v, ggml_tensor * g, - ggml_tensor * beta, - ggml_tensor * state, - ggml_tensor * causal_mask, - ggml_tensor * identity, - ggml_tensor * diag_mask, + ggml_tensor * b, + ggml_tensor * s, int il) { const int64_t S_k = q->ne[0]; const int64_t H_k = q->ne[1]; @@ -113,134 +99,123 @@ std::pair llm_build_qwen3next::build_delta_net_chu const int64_t S_v = v->ne[0]; const int64_t H_v = v->ne[1]; - GGML_ASSERT(v->ne[2] == n_tokens); - GGML_ASSERT(k->ne[2] == n_tokens); - GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); - GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); - GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + GGML_ASSERT(S_k == S_v); + GGML_ASSERT(H_v % H_k == 0); GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs); - GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case - - const float eps_norm = hparams.f_norm_rms_eps; - - q = ggml_l2_norm(ctx0, q, eps_norm); - k = ggml_l2_norm(ctx0, k, eps_norm); + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs); + GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs); - const float scale = 1.0f / sqrtf(S_v); + const float scale = 1.0f / sqrtf(S_k); q = ggml_scale(ctx0, q, scale); - beta = ggml_sigmoid(ctx0, beta); - cb(q, "q_in", il); cb(k, "k_in", il); cb(v, "v_in", il); - cb(beta, "beta_in", il); + cb(b, "b_in", il); cb(g, "g_in", il); - q = ggml_cont_4d(ctx0, ggml_permute(ctx0, q, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); - k = ggml_cont_4d(ctx0, ggml_permute(ctx0, k, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); - v = ggml_cont_4d(ctx0, ggml_permute(ctx0, v, 0, 2, 1, 3), S_v, n_tokens, H_v, n_seqs); - g = ggml_cont_4d(ctx0, ggml_permute(ctx0, g, 2, 0, 3, 1), n_tokens, 1, H_k, n_seqs); - - beta = ggml_cont(ctx0, ggml_permute(ctx0, beta, 2, 0, 1, 3)); - state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); - - cb(q, "q_perm", il); - cb(k, "k_perm", il); - cb(v, "v_perm", il); - cb(beta, "beta_perm", il); - cb(g, "g_perm", il); - cb(state, "state_in", il); - - GGML_ASSERT(q->ne[1] == n_tokens && q->ne[0] == S_k && q->ne[2] == H_k && q->ne[3] == n_seqs); - GGML_ASSERT(k->ne[1] == n_tokens && k->ne[0] == S_k && k->ne[2] == H_k && k->ne[3] == n_seqs); - GGML_ASSERT(v->ne[1] == n_tokens && v->ne[0] == S_v && v->ne[2] == H_k && v->ne[3] == n_seqs); - GGML_ASSERT(beta->ne[1] == n_tokens && beta->ne[2] == H_k && beta->ne[0] == 1 && beta->ne[3] == n_seqs); + q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs] + k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs] + v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs] + g = ggml_permute(ctx0, g, 2, 1, 3, 0); // [ 1, n_tokens, H_v, n_seqs] + b = ggml_permute(ctx0, b, 2, 0, 1, 3); // [ 1, n_tokens, H_v, n_seqs] - // Do padding - const int64_t chunk_size = CHUNK_SIZE; + const int CS = CHUNK_SIZE; - const int64_t pad = (chunk_size - n_tokens % chunk_size) % chunk_size; - const int64_t n_chunks = (n_tokens + pad) / chunk_size; + const int pad = (CS - n_tokens % CS) % CS; + const int n_chunks = (n_tokens + pad) / CS; q = ggml_pad(ctx0, q, 0, pad, 0, 0); k = ggml_pad(ctx0, k, 0, pad, 0, 0); v = ggml_pad(ctx0, v, 0, pad, 0, 0); - g = ggml_pad(ctx0, g, pad, 0, 0, 0); - beta = ggml_pad(ctx0, beta, 0, pad, 0, 0); + g = ggml_pad(ctx0, g, 0, pad, 0, 0); + b = ggml_pad(ctx0, b, 0, pad, 0, 0); - cb(q, "q_pad", il); - cb(k, "k_pad", il); - cb(v, "v_pad", il); - cb(beta, "beta_pad", il); - cb(g, "g_pad", il); + ggml_tensor * v_b = ggml_mul(ctx0, v, b); + ggml_tensor * k_b = ggml_mul(ctx0, k, b); - ggml_tensor * v_beta = ggml_mul(ctx0, v, beta); - ggml_tensor * k_beta = ggml_mul(ctx0, k, beta); + cb(v_b, "v_b", il); + cb(k_b, "k_b", il); - cb(v_beta, "v_beta", il); - cb(k_beta, "k_beta", il); + q = ggml_reshape_4d(ctx0, q, S_k, CS, n_chunks, H_k * n_seqs); + k = ggml_reshape_4d(ctx0, k, S_k, CS, n_chunks, H_k * n_seqs); + k_b = ggml_reshape_4d(ctx0, k_b, S_k, CS, n_chunks, H_v * n_seqs); + v = ggml_reshape_4d(ctx0, v, S_v, CS, n_chunks, H_v * n_seqs); + v_b = ggml_reshape_4d(ctx0, v_b, S_v, CS, n_chunks, H_v * n_seqs); - q = ggml_reshape_4d(ctx0, q, S_k, chunk_size, n_chunks, H_k * n_seqs); - k = ggml_reshape_4d(ctx0, k, S_k, chunk_size, n_chunks, H_k * n_seqs); - k_beta = ggml_reshape_4d(ctx0, k_beta, S_k, chunk_size, n_chunks, H_k * n_seqs); - v = ggml_reshape_4d(ctx0, v, S_v, chunk_size, n_chunks, H_v * n_seqs); - v_beta = ggml_reshape_4d(ctx0, v_beta, S_v, chunk_size, n_chunks, H_v * n_seqs); + g = ggml_reshape_4d(ctx0, g, CS, 1, n_chunks, H_v * n_seqs); + b = ggml_reshape_4d(ctx0, b, 1, CS, n_chunks, H_v * n_seqs); - g = ggml_reshape_4d(ctx0, g, chunk_size, 1, n_chunks, H_k * n_seqs); - beta = ggml_reshape_4d(ctx0, beta, 1, chunk_size, n_chunks, H_k * n_seqs); + // [CS, 1, n_chunks, H_v * n_seqs] + ggml_tensor * g_cs = ggml_cumsum(ctx0, g); + cb(g_cs, "g_cs", il); - ggml_tensor * g_cumsum = ggml_cumsum(ctx0, g); - cb(g_cumsum, "g_cumsum", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + ggml_tensor * g_cs_i = g_cs; + ggml_tensor * g_cs_j = ggml_reshape_4d(ctx0, g_cs, 1, CS, n_chunks, H_v * n_seqs); - ggml_tensor * gcs_i = g_cumsum; // ggml_reshape_4d(ctx0, g_cumsum, chunk_size, 1, n_chunks, H_v * n_seqs); - ggml_tensor * gcs_j = ggml_reshape_4d(ctx0, g_cumsum, 1, chunk_size, n_chunks, H_v * n_seqs); + g_cs_j = ggml_repeat_4d(ctx0, g_cs_j, CS, CS, n_chunks, H_v * n_seqs); - ggml_tensor * gcs_j_broadcast = - ggml_repeat_4d(ctx0, gcs_j, chunk_size, chunk_size, n_chunks, H_v * n_seqs); + // [CS, CS, n_chunks, H_v * n_seqs] + ggml_tensor * decay_mask; + decay_mask = ggml_sub(ctx0, g_cs_j, g_cs_i); + decay_mask = ggml_tri(ctx0, decay_mask, GGML_TRI_TYPE_LOWER_DIAG); + decay_mask = ggml_exp(ctx0, decay_mask); + cb(decay_mask, "decay_mask", il); - ggml_tensor * decay_mask = ggml_sub(ctx0, gcs_j_broadcast, gcs_i); - cb(decay_mask, "decay_mask", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + // [CS, CS, n_chunks, H_k * n_seqs] + ggml_tensor * kb; + kb = ggml_mul_mat(ctx0, k, k_b); + kb = ggml_mul (ctx0, kb, decay_mask); - decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); - decay_mask = ggml_exp(ctx0, decay_mask); - decay_mask = ggml_mul(ctx0, decay_mask, diag_mask); + // [CS, CS, n_chunks, H_k * n_seqs] + ggml_tensor * attn; + attn = ggml_tri(ctx0, kb, GGML_TRI_TYPE_LOWER); - ggml_tensor * kmulkbeta = ggml_mul_mat(ctx0, k, k_beta); + ggml_tensor * identity; + identity = ggml_view_1d(ctx0, attn, CS, 0); + identity = ggml_fill (ctx0, identity, 1.0f); + identity = ggml_diag (ctx0, identity); - ggml_tensor * k_decay = ggml_mul(ctx0, kmulkbeta, decay_mask); - ggml_tensor * attn = ggml_neg(ctx0, ggml_mul(ctx0, k_decay, causal_mask)); - cb(attn, "attn_pre_solve", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * lhs = ggml_add(ctx0, attn, identity); + cb(lhs, "dnet_add_ch_lhs", il); - ggml_tensor * attn_lower = ggml_mul(ctx0, attn, causal_mask); - ggml_tensor * lhs = ggml_sub(ctx0, ggml_repeat(ctx0, identity, attn_lower), attn_lower); + attn = ggml_neg(ctx0, attn); - ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); - attn = ggml_mul(ctx0, lin_solve, causal_mask); - attn = ggml_add(ctx0, attn, identity); - cb(attn, "attn_solved", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * lin_solve = ggml_solve_tri(ctx0, lhs, attn, true, true, false); + attn = ggml_add(ctx0, lin_solve, identity); + cb(attn, "dnet_add_ch_attn_solved", il); // [CS, CS, n_chunks, H_k * n_seqs] - v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_beta)), attn); + // [S_v, CS, n_chunks, H_v * n_seqs] + v = ggml_mul_mat(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, v_b)), attn); - ggml_tensor * g_cumsum_t = ggml_cont(ctx0, ggml_transpose(ctx0, g_cumsum)); - ggml_tensor * gexp = ggml_exp(ctx0, g_cumsum_t); + // [CS, 1, n_chunks, H_v * n_seqs] + ggml_tensor * g_exp = ggml_exp(ctx0, g_cs); - ggml_tensor * kbeta_gexp = ggml_mul(ctx0, k_beta, gexp); - cb(kbeta_gexp, "kbeta_gexp", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + k_b = ggml_cont(ctx0, ggml_transpose(ctx0, k_b)); - ggml_tensor * k_cumdecay = - ggml_cont(ctx0, ggml_transpose(ctx0, ggml_mul_mat(ctx0, attn, ggml_cont(ctx0, ggml_transpose(ctx0, kbeta_gexp))))); - cb(k_cumdecay, "k_cumdecay", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + // [CS, S_k, n_chunks, H_k * n_seqs] + ggml_tensor * kbg = ggml_mul(ctx0, k_b, g_exp); + cb(kbg, "k_beta_g_exp", il); - ggml_tensor * attn_kq = ggml_mul_mat(ctx0, k, q); - attn_kq = ggml_mul(ctx0, attn_kq, decay_mask); - attn_kq = ggml_mul(ctx0, attn_kq, diag_mask); - cb(attn_kq, "attn_kq", il); // shape: (chunk_size, chunk_size, n_chunks, H_v * n_seqs) + // [S_k, CS, n_chunks, H_k * n_seqs] + ggml_tensor * k_cd = ggml_mul_mat(ctx0, kbg, attn); + cb(k_cd, "k_cumdecay", il); + // [S_k, CS, n_chunks, H_k * n_seqs] + ggml_tensor * g_exp_t = ggml_transpose(ctx0, g_exp); + ggml_tensor * q_g_exp = ggml_mul(ctx0, q, g_exp_t); + + // [CS, CS, n_chunks, H_k * n_seqs] + ggml_tensor * kq = ggml_mul_mat(ctx0, k, q); + kq = ggml_mul(ctx0, kq, decay_mask); + kq = ggml_tri(ctx0, kq, GGML_TRI_TYPE_LOWER_DIAG); + cb(kq, "kq", il); // vectorized calculation of key_gdiff // improved from the chunked version: @@ -250,109 +225,98 @@ std::pair llm_build_qwen3next::build_delta_net_chu // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew - // get last element in g_cumsum along chunk_size dimension (ne0) + // get last element in g_cumsum along CS dimension (ne0) // example: [[x, y, z, ..., last], ...] -> [[last], ...] - ggml_tensor * g_last = ggml_view_4d(ctx0, g_cumsum, 1, 1, g_cumsum->ne[2], g_cumsum->ne[3], - g_cumsum->nb[1], g_cumsum->nb[2], g_cumsum->nb[3], - (g_cumsum->ne[0] - 1) * ggml_element_size(g_cumsum)); + // [1, 1, n_chunks, H_v * n_seqs] + ggml_tensor * g_last = ggml_view_4d(ctx0, g_cs, 1, 1, g_cs->ne[2], g_cs->ne[3], + g_cs->nb[1], + g_cs->nb[2], + g_cs->nb[3], + ggml_row_size(g_cs->type, g_cs->ne[0] - 1)); + cb(g_last, "g_last", il); + + // TODO: remove this cont when CUDA supports non-cont unary ops g_last = ggml_cont(ctx0, g_last); - cb(g_last, "g_last", il); // shape: (1, 1, n_chunks, H_v * n_seqs) + // [1, 1, n_chunks, H_v * n_seqs] ggml_tensor * g_last_exp = ggml_exp(ctx0, g_last); - cb(g_last_exp, "g_last_exp", il); // shape: (1, 1, n_chunks, H_v * n_seqs) - - ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cumsum, g_last)); - cb(g_diff, "g_diff", il); // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) + cb(g_last_exp, "g_last_exp", il); - ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); - ggml_tensor * g_diff_exp_t = ggml_reshape_4d(ctx0, g_diff_exp, - 1, chunk_size, n_chunks, g_diff_exp->ne[3]); + // [CS, 1, n_chunks, H_v * n_seqs] + ggml_tensor * g_diff = ggml_neg(ctx0, ggml_sub(ctx0, g_cs, g_last)); + cb(g_diff, "g_diff", il); - ggml_tensor * key_gdiff = ggml_mul(ctx0, k, g_diff_exp_t); - cb(key_gdiff, "key_gdiff", il); // shape: (S_k, chunk_size, n_chunks, H_v * n_seqs) + ggml_tensor * g_diff_exp = ggml_exp(ctx0, g_diff); + ggml_tensor * g_diff_exp_t = ggml_transpose(ctx0, g_diff_exp); - ggml_tensor * key_gdiff_t = ggml_cont(ctx0, ggml_transpose(ctx0, key_gdiff)); - cb(key_gdiff_t, "key_gdiff_t", il); // shape: (chunk_size, S_k, n_chunks, H_v * n_seqs) + // [S_k, CS, n_chunks, H_v * n_seqs] + ggml_tensor * kg = ggml_mul(ctx0, k, g_diff_exp_t); + cb(kg, "key_gdiff", il); + // [CS, S_k, n_chunks, H_v * n_seqs] + ggml_tensor * kg_t = ggml_cont(ctx0, ggml_transpose(ctx0, kg)); + cb(kg_t, "key_gdiff_t", il); - // state to be updated per chunk - ggml_tensor * new_state = state; // ggml_dup(ctx0, state); - cb(new_state, "new_state", il); // shape: (S_v, S_v, H_v, n_seqs) + ggml_tensor * s_t = ggml_transpose(ctx0, s); + s_t = ggml_cont_4d(ctx0, s_t, S_v, S_v, 1, H_v * n_seqs); + cb(s_t, "dnet_add_ch_state", il); - // shape after loop of chunks: (S_v, chunk_size, n_chunks, H_v * n_seqs) - ggml_tensor * core_attn_out = nullptr; + // [CS, S_v, n_chunks, H_v * n_seqs] + ggml_tensor * v_t = ggml_cont(ctx0, ggml_transpose(ctx0, v)); for (int64_t chunk = 0; chunk < n_chunks; chunk++) { - // shape: (S_k, chunk_size, 1, H_k * n_seqs) - ggml_tensor * q_chunk = get_slice_2d(ctx0, q, chunk); // (no cont), next op: ggml_mul + ggml_tensor * ch_k_cd = get_slice_2d(ctx0, k_cd, chunk); // [S_k, CS, 1, H_k * n_seqs] + ggml_tensor * ch_v_t = get_slice_2d(ctx0, v_t, chunk); // [ CS, S_v, 1, H_v * n_seqs] + ggml_tensor * ch_kq = get_slice_2d(ctx0, kq, chunk); // [ CS, CS, 1, H_k * n_seqs] + ggml_tensor * ch_q_g_exp = get_slice_2d(ctx0, q_g_exp, chunk); // [S_k, CS, 1, H_k * n_seqs] + ggml_tensor * ch_kg_t = get_slice_2d(ctx0, kg_t, chunk); // [ CS, S_k, 1, H_v * n_seqs] - // shape: (S_v, chunk_size, 1, H_v * n_seqs) - ggml_tensor * v_chunk = get_slice_2d(ctx0, v, chunk); // (no cont), next op: ggml_repeat + // [CS, S_v, 1, H_v * n_seqs] + ggml_tensor * v_t_p = ggml_mul_mat(ctx0, ch_k_cd, s_t); + cb(v_t_p, "v_prime", il); - // shape: (chunk_size, 1, n_chunks, H_v * n_seqs) - ggml_tensor * gexp_chunk = get_slice_2d(ctx0, gexp, chunk); // (no cont), next op: ggml_mul + // [CS, S_v, 1, H_v * n_seqs] + ggml_tensor * v_t_new = ggml_sub(ctx0, ch_v_t, v_t_p); + cb(v_t_new, "v_t_new", il); - // shape: (chunk_size, 1, H_v * n_seqs) - ggml_tensor * k_cumdecay_chunk = get_slice_2d(ctx0, k_cumdecay, chunk); // (no cont), next op: ggml_mul_mat + // [S_v, CS, 1, H_v * n_seqs] + ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_t_new, ch_kq); + cb(v_attn, "v_attn", il); - // attn = (q_i @ k_i.transpose(-1, -2) * decay_mask[:, :, i]).masked_fill_(mask, 0) - // replaced by precomputed attn_kq - ggml_tensor * attn_chunk = get_slice_2d(ctx0, attn_kq, chunk); - cb(attn_chunk, "attn_chunk", il); + // [S_v, CS, 1, H_v * n_seqs] + ggml_tensor * attn_inter = ggml_mul_mat(ctx0, s_t, ch_q_g_exp); + cb(attn_inter, "attn_inter", il); - ggml_tensor * state_t = ggml_cont_4d(ctx0, ggml_permute(ctx0, new_state, 1, 0, 2, 3), S_v, S_v, 1, H_v * n_seqs); + // [S_v, CS, 1, H_v * n_seqs] + ggml_tensor * o_ch = ggml_add(ctx0, attn_inter, v_attn); + cb(o_ch, "dnet_add_ch_attn_out", il); - // v_prime = (k_cumdecay[:, :, i]) @ last_recurrent_state - ggml_tensor * v_prime = ggml_mul_mat(ctx0, state_t, k_cumdecay_chunk); - cb(v_prime, "v_prime_chunk", il); // shape: (S_v, 1, H_v * n_seqs) - - // v_new = v_i - v_prime - ggml_tensor * v_new = ggml_sub(ctx0, ggml_repeat(ctx0, v_chunk, v_prime), v_prime); - ggml_tensor * v_new_t = ggml_cont(ctx0, ggml_transpose(ctx0, v_new)); - cb(v_new, "v_new_chunk", il); - - // attn_inter = (q_i * g[:, :, i, :, None].exp()) @ last_recurrent_state - ggml_tensor * q_g_exp = ggml_mul(ctx0, q_chunk, gexp_chunk); - ggml_tensor * attn_inter = ggml_mul_mat(ctx0, state_t, q_g_exp); - cb(attn_inter, "attn_inter_chunk", il); - - // core_attn_out[:, :, i] = attn_inter + attn @ v_new - ggml_tensor * v_attn = ggml_mul_mat(ctx0, v_new_t, attn_chunk); - cb(v_attn, "v_attn_chunk", il); - - ggml_tensor * core_attn_out_chunk = ggml_add(ctx0, attn_inter, v_attn); - cb(core_attn_out_chunk, "core_attn_out_chunk", il); // shape: (S_v, chunk_size, 1, H_v * n_seqs) - - core_attn_out = core_attn_out == nullptr - ? core_attn_out_chunk - : ggml_concat(ctx0, core_attn_out, core_attn_out_chunk, 2); + v = ggml_set_inplace(ctx0, v, o_ch, v->nb[1], v->nb[2], v->nb[3], chunk * v->nb[2]); // kgdmulvnew = (key_gdiff).transpose(-1, -2) @ v_new - ggml_tensor * k_gdiff_t = get_slice_2d(ctx0, key_gdiff_t, chunk); - //ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, k_gdiff, v_new); // this is slower on metal, why? - ggml_tensor * kgdmulvnew = ggml_mul_mat(ctx0, v_new_t, k_gdiff_t); + // TODO: head broadcast might not work here - probably will need a transpose + ggml_tensor * kgv = ggml_mul_mat(ctx0, ch_kg_t, v_t_new); // [S_k, S_v, 1, H_k * n_seqs] // last_recurrent_state = last_recurrent_state * g_last + kgdmulvnew - ggml_tensor * gexp_last_chunk = ggml_cont(ctx0, get_slice_2d(ctx0, g_last_exp, chunk)); - new_state = ggml_add(ctx0, - ggml_mul(ctx0, new_state, ggml_reshape_4d(ctx0, gexp_last_chunk, gexp_last_chunk->ne[0], gexp_last_chunk->ne[1], H_v, n_seqs)), - ggml_reshape_4d(ctx0, kgdmulvnew, kgdmulvnew->ne[0], kgdmulvnew->ne[1], H_v, n_seqs)); + ggml_tensor * ch_g_last_exp = get_slice_2d(ctx0, g_last_exp, chunk); + s_t = ggml_mul(ctx0, s_t, ch_g_last_exp); + s_t = ggml_add(ctx0, s_t, kgv); + cb(s_t, "dnet_add_ch_state", il); } + s_t = ggml_reshape_4d(ctx0, s_t, S_v, S_v, H_v, n_seqs); + // truncate padded tokens - ggml_tensor * output_tokens = ggml_view_4d(ctx0, core_attn_out, + ggml_tensor * o = ggml_view_4d(ctx0, v, S_v, n_tokens, H_v, n_seqs, - ggml_row_size(core_attn_out->type, S_v), - ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks), - ggml_row_size(core_attn_out->type, S_v * chunk_size * n_chunks * H_v), 0); - output_tokens = ggml_cont(ctx0, output_tokens); - cb(output_tokens, "output_tokens", il); + ggml_row_size(v->type, S_v), + ggml_row_size(v->type, S_v * CS * n_chunks), + ggml_row_size(v->type, S_v * CS * n_chunks * H_v), 0); - // permute back to (S_v, H_v, n_tokens, n_seqs) - output_tokens = ggml_permute(ctx0, output_tokens, 0, 2, 1, 3); - output_tokens = ggml_cont(ctx0, output_tokens); + o = ggml_permute (ctx0, o, 0, 2, 1, 3); // [S_v, H_v, n_tokens, n_seqs] + s = ggml_transpose(ctx0, s_t); // [S_v, S_v, H_v, n_seqs] - return {output_tokens, new_state}; + return {o, s}; } std::pair llm_build_qwen3next::build_delta_net_autoregressive( @@ -360,8 +324,8 @@ std::pair llm_build_qwen3next::build_delta_net_aut ggml_tensor * k, ggml_tensor * v, ggml_tensor * g, - ggml_tensor * beta, - ggml_tensor * state, + ggml_tensor * b, // beta + ggml_tensor * s, // state int il) { const int64_t S_k = q->ne[0]; const int64_t H_k = q->ne[1]; @@ -371,75 +335,72 @@ std::pair llm_build_qwen3next::build_delta_net_aut const int64_t S_v = v->ne[0]; const int64_t H_v = v->ne[1]; - GGML_ASSERT(n_tokens == 1); // This function is optimized for single token processing - GGML_ASSERT(v->ne[2] == n_tokens); - GGML_ASSERT(k->ne[2] == n_tokens); - GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); - GGML_ASSERT(beta->ne[0] == H_v && beta->ne[2] == n_tokens && beta->ne[3] == n_seqs); - GGML_ASSERT(state->ne[0] == S_v && state->ne[1] == S_v * H_v && state->ne[2] == 1 && state->ne[3] == n_seqs); + GGML_ASSERT(n_tokens == 1); + + GGML_ASSERT(S_k == S_v); + GGML_ASSERT(H_v % H_k == 0); GGML_ASSERT(q->ne[0] == S_k && q->ne[1] == H_k && q->ne[2] == n_tokens && q->ne[3] == n_seqs); GGML_ASSERT(k->ne[0] == S_k && k->ne[1] == H_k && k->ne[2] == n_tokens && k->ne[3] == n_seqs); + GGML_ASSERT(v->ne[0] == S_v && v->ne[1] == H_v && v->ne[2] == n_tokens && v->ne[3] == n_seqs); - GGML_ASSERT(H_k == H_v); // we did a repeat to make sure this is the case - - const float eps_norm = hparams.f_norm_rms_eps; + GGML_ASSERT(g->ne[0] == H_v && g->ne[1] == n_tokens && g->ne[2] == n_seqs); + GGML_ASSERT(b->ne[0] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs); + GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs); - q = ggml_l2_norm(ctx0, q, eps_norm); - k = ggml_l2_norm(ctx0, k, eps_norm); + const float scale = 1.0f / sqrtf(S_k); - const float scale = 1.0f / sqrtf(S_v); + q = ggml_scale(ctx0, q, scale); - q = ggml_scale(ctx0, q, scale); - beta = ggml_sigmoid(ctx0, beta); + q = ggml_permute(ctx0, q, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs] + k = ggml_permute(ctx0, k, 0, 2, 1, 3); // [S_k, n_tokens, H_k, n_seqs] + v = ggml_permute(ctx0, v, 0, 2, 1, 3); // [S_v, n_tokens, H_v, n_seqs] cb(q, "q_in", il); cb(k, "k_in", il); cb(v, "v_in", il); - cb(beta, "beta_in", il); + cb(b, "b_in", il); cb(g, "g_in", il); - state = ggml_reshape_4d(ctx0, state, S_v, S_v, H_v, n_seqs); + g = ggml_reshape_4d(ctx0, g, 1, 1, H_v, n_seqs); + b = ggml_reshape_4d(ctx0, b, 1, 1, H_v, n_seqs); - ggml_tensor * g_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, g), 1, 1, H_k, n_seqs); - ggml_tensor * beta_t = ggml_reshape_4d(ctx0, ggml_transpose(ctx0, beta), 1, 1, H_k, n_seqs); + // [S_v, S_v, H_v, n_seqs] + g = ggml_exp(ctx0, g); + s = ggml_mul(ctx0, s, g); - // Apply exponential to g_t - g_t = ggml_exp(ctx0, g_t); + ggml_tensor * s_t = ggml_cont(ctx0, ggml_transpose(ctx0, s)); - // Apply the gated delta rule for the single timestep - // last_recurrent_state = last_recurrent_state * g_t - state = ggml_mul(ctx0, state, g_t); + // [1, S_v, H_v, n_seqs] + ggml_tensor * sk; + sk = ggml_mul (ctx0, s_t, k); + sk = ggml_sum_rows(ctx0, sk); - // kv_mem = (last_recurrent_state * k_t.unsqueeze(-1)).sum(dim=-2) - ggml_tensor * k_t_unsqueezed = ggml_reshape_4d(ctx0, k, 1, S_v, H_v, n_seqs); - ggml_tensor * kv_mem = ggml_mul(ctx0, state, k_t_unsqueezed); - // we need to sum over dim=-2, so we transpose, sum, then transpose again - kv_mem = ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, kv_mem)))); + // [S_v, 1, H_v, n_seqs] + ggml_tensor * d; + d = ggml_sub(ctx0, v, ggml_transpose(ctx0, sk)); + d = ggml_mul(ctx0, d, b); - // v_t = v.unsqueeze(2) (we insert the singleton dimension after n_seqs and H_v) - ggml_tensor * v_t = ggml_reshape_4d(ctx0, v, S_v, 1, H_v, n_seqs); - // delta = (v_t - kv_mem) * beta_t - ggml_tensor * v_diff = ggml_sub(ctx0, v_t, kv_mem); // both should be [S_v, 1, H_v, n_seqs] - ggml_tensor * delta = ggml_mul(ctx0, v_diff, beta_t); + // [1, S_v, H_v, n_seqs] + ggml_tensor * d_t; + d_t = ggml_transpose(ctx0, d); - // last_recurrent_state = last_recurrent_state + k_t.unsqueeze(-1) * delta - ggml_tensor * k_t_delta = ggml_mul(ctx0, ggml_repeat_4d(ctx0, k_t_unsqueezed, S_v, S_v, H_v, n_seqs), delta); - state = ggml_add(ctx0, state, k_t_delta); + // [S_v, S_v, H_v, n_seqs] + ggml_tensor * kd; + k = ggml_repeat(ctx0, k, s); + kd = ggml_mul (ctx0, k, d_t); - // Compute the attention output - // core_attn_out = (last_recurrent_state * q_t.unsqueeze(-1)).sum(dim=-2) - ggml_tensor * q_t_unsqueezed = ggml_reshape_4d(ctx0, q, 1, S_v, H_v, n_seqs); // unsqueeze q_t - ggml_tensor * state_q = ggml_mul(ctx0, state, q_t_unsqueezed); - // again, since it's over dim = -2, transpose, sum, transpose back - ggml_tensor * core_attn_out = - ggml_transpose(ctx0, ggml_sum_rows(ctx0, ggml_cont(ctx0, ggml_transpose(ctx0, state_q)))); + s_t = ggml_add(ctx0, s_t, kd); - // core_attn_out should be [S_v, 1, H_v, n_seqs] after this - cb(core_attn_out, "output_tokens", il); - cb(state, "new_state", il); + cb(s_t, "dnet_add_ar_state", il); - return {core_attn_out, state}; + ggml_tensor * s_q = ggml_mul (ctx0, s_t, q); + ggml_tensor * o = ggml_sum_rows(ctx0, s_q); + + o = ggml_permute (ctx0, o, 2, 0, 1, 3); // [S_v, H_v, n_tokens, n_seqs] + s = ggml_transpose(ctx0, s_t); // [S_v, S_v, H_v, n_seqs] + + return {o, s}; } ggml_tensor * llm_build_qwen3next::build_norm_gated( @@ -472,39 +433,29 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn( // Split Q projection into query and gate // The split should be along dimension 0 (the feature dimension) ggml_tensor * Qcur = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1, - Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0); + Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], 0); + cb(Qcur, "Qcur_view", il); + ggml_tensor * gate = ggml_view_4d(ctx0, Qcur_full, n_embd_head, n_head, n_tokens, 1, Qcur_full->nb[1], Qcur_full->nb[2], Qcur_full->nb[3], n_embd_head * ggml_element_size(Qcur_full)); - cb(Qcur, "Qcur", il); cb(gate, "gate", il); - // Now reshape Qcur to [n_embd_head, n_head, n_tokens] for multi-head attention - Qcur = ggml_cont_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - cb(Qcur, "Qcur_reshaped", il); - - // Apply Q normalization - Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); - cb(Qcur, "Qcur_normed", 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); - // Apply K normalization Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); - cb(Kcur, "Kcur_normed", il); + Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); - // Reshape gate to [n_embd, n_tokens] for the sigmoid gating (flatten the heads) - gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); - cb(gate, "gate_reshaped", il); + Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, nullptr, LLM_NORM_RMS, il); + cb(Qcur, "Qcur_normed", il); - Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); + Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, nullptr, LLM_NORM_RMS, il); + cb(Kcur, "Kcur_normed", il); - // Apply RoPE Qcur = ggml_rope_ext( ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, @@ -519,7 +470,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn( cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); - // Attention computation 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, @@ -527,10 +477,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn( Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il); cb(cur, "attn_pregate", il); - ggml_tensor * gate_sigmoid = ggml_sigmoid(ctx0, gate); - cb(gate_sigmoid, "gate_sigmoid", il); + // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont + gate = ggml_cont_2d(ctx0, gate, n_embd_head * n_head, n_tokens); + + gate = ggml_sigmoid(ctx0, gate); + cb(gate, "gate_sigmoid", il); + + gate = ggml_reshape_2d(ctx0, gate, n_embd_head * n_head, n_tokens); - cur = ggml_mul(ctx0, cur, gate_sigmoid); + cur = ggml_mul(ctx0, cur, gate); cb(cur, "attn_gated", il); cur = build_lora_mm(model.layers[il].wo, cur); @@ -560,7 +515,6 @@ std::pair llm_build_qwen3next::build_qkvz( cb(z, "z", il); return { qkv_mixed, z }; - } else { // legacy (slower) path ggml_tensor * mixed_qkvz = build_lora_mm(model.layers[il].ssm_in, input); @@ -624,9 +578,6 @@ std::pair llm_build_qwen3next::build_qkvz( ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( llm_graph_input_rs * inp, ggml_tensor * cur, - ggml_tensor * causal_mask, - ggml_tensor * identity, - ggml_tensor * diag_mask, int il) { const auto * mctx_cur = inp->mctx; @@ -671,7 +622,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( split_sizes_ba[0] * ggml_element_size(mixed_ba_reshaped)); cb(a, "a", il); - ggml_tensor * beta = ggml_cont_4d(ctx0, b, num_v_heads, 1, n_seq_tokens, n_seqs); + // TODO: CUDA is missing non-contiguous unary ops. when implemented: remove this cont + b = ggml_cont(ctx0, b); + + ggml_tensor * beta = ggml_sigmoid(ctx0, b); + + beta = ggml_reshape_4d(ctx0, beta, num_v_heads, 1, n_seq_tokens, n_seqs); // Reshape a to merge head dimensions: [batch, seq_len, num_k_heads, num_v_heads/num_k_heads] -> [batch, seq_len, num_v_heads] ggml_tensor * alpha = ggml_cont_3d(ctx0, a, num_v_heads, n_seq_tokens, n_seqs); @@ -679,6 +635,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_tensor * alpha_biased = ggml_add(ctx0, alpha, model.layers[il].ssm_dt); ggml_tensor * alpha_softplus = ggml_softplus(ctx0, alpha_biased); cb(alpha_softplus, "a_softplus", il); + ggml_tensor * gate = ggml_mul(ctx0, alpha_softplus, model.layers[il].ssm_a); // -A_log.exp() * softplus cb(gate, "gate", il); @@ -686,8 +643,6 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_tensor * conv_states_all = mctx_cur->get_r_l(il); ggml_tensor * ssm_states_all = mctx_cur->get_s_l(il); - // bool use_precomputed_states = n_seq_tokens == 1 && mctx_cur->has_previous_state(); - // Build the convolution states tensor ggml_tensor * conv_states = build_rs(inp, conv_states_all, hparams.n_embd_r(), n_seqs); cb(conv_states, "conv_states", il); @@ -696,11 +651,12 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_tensor * conv_kernel = model.layers[il].ssm_conv1d; const int64_t conv_kernel_size = conv_kernel->ne[0]; const int64_t conv_channels = d_inner + 2 * hparams.ssm_n_group * hparams.ssm_d_state; - conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); + + conv_states = ggml_reshape_3d(ctx0, conv_states, conv_kernel_size - 1, conv_channels, n_seqs); cb(conv_states, "conv_states_reshaped", il); - qkv_mixed = ggml_permute(ctx0, qkv_mixed, 1, 0, 2, 3); - cb(qkv_mixed, "qkv_mixed_permuted", il); + qkv_mixed = ggml_transpose(ctx0, qkv_mixed); + cb(qkv_mixed, "qkv_mixed_transposed", il); ggml_tensor * conv_input = ggml_concat(ctx0, conv_states, qkv_mixed, 0); cb(conv_input, "conv_input", il); @@ -720,7 +676,10 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( ggml_build_forward_expand(gf, ggml_cpy(ctx0, last_conv_states, state_update_target)); cb(conv_states_all, "conv_states_updated", il); - // Apply SSM convolution + ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); + state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim, num_v_heads, n_seqs); + cb(state, "state_predelta", il); + ggml_tensor * conv_output_proper = ggml_ssm_conv(ctx0, conv_input, conv_kernel); cb(conv_output_proper, "conv_output_raw", il); @@ -734,26 +693,36 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( int64_t nb1_qkv = ggml_row_size(conv_qkv_mix->type, qkv_dim); // Extract the convolved Q, K, V from conv_output - ggml_tensor * q_conv = - ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, 0); + ggml_tensor * q_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs, + ggml_row_size(conv_qkv_mix->type, head_k_dim), + nb1_qkv, + nb1_qkv * n_seq_tokens, + 0); + + ggml_tensor * k_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_k_dim, num_k_heads, n_seq_tokens, n_seqs, + ggml_row_size(conv_qkv_mix->type, head_k_dim), + nb1_qkv, + nb1_qkv * n_seq_tokens, + head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); + + ggml_tensor * v_conv = ggml_view_4d(ctx0, conv_qkv_mix, head_v_dim, num_v_heads, n_seq_tokens, n_seqs, + ggml_row_size(conv_qkv_mix->type, head_v_dim), + nb1_qkv, + nb1_qkv * n_seq_tokens, + ggml_row_size(conv_qkv_mix->type, 2 * head_k_dim * num_k_heads)); + cb(q_conv, "q_conv", il); - ggml_tensor * k_conv = - ggml_view_2d(ctx0, conv_qkv_mix, head_k_dim * num_k_heads, n_seq_tokens * n_seqs, nb1_qkv, - head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); cb(k_conv, "k_conv", il); - ggml_tensor * v_conv = - ggml_view_2d(ctx0, conv_qkv_mix, head_v_dim * num_v_heads, n_seq_tokens * n_seqs, nb1_qkv, - 2 * head_k_dim * num_k_heads * ggml_element_size(conv_qkv_mix)); cb(v_conv, "v_conv", il); - // Unsqueeze them - q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); - k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); - v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); + const float eps_norm = hparams.f_norm_rms_eps; - ggml_tensor * state = build_rs(inp, ssm_states_all, hparams.n_embd_s(), n_seqs); - state = ggml_reshape_4d(ctx0, state, head_v_dim, head_v_dim * num_v_heads, 1, n_seqs); - cb(state, "state_predelta", il); + q_conv = ggml_l2_norm(ctx0, q_conv, eps_norm); + k_conv = ggml_l2_norm(ctx0, k_conv, eps_norm); + + //q_conv = ggml_cont_4d(ctx0, q_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + //k_conv = ggml_cont_4d(ctx0, k_conv, head_k_dim, num_k_heads, n_seq_tokens, n_seqs); + //v_conv = ggml_cont_4d(ctx0, v_conv, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); // if head keys and value keys are different, repeat to force tensors into matching shapes if (num_k_heads != num_v_heads) { @@ -786,7 +755,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( if (n_seq_tokens == 1) { attn_out = build_delta_net_autoregressive(q_conv, k_conv, v_conv, gate, beta, state, il); } else { - attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, causal_mask, identity, diag_mask, il); + attn_out = build_delta_net_chunking(q_conv, k_conv, v_conv, gate, beta, state, il); } ggml_tensor * output = attn_out.first; ggml_tensor * new_state = attn_out.second; @@ -795,19 +764,15 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( // Update the recurrent states ggml_build_forward_expand(gf, - ggml_cpy(ctx0, new_state, - ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, - kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); - - // Reshape both attn_out_final and z to 2D tensors for normalization - // attn_out_final: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] - ggml_tensor * attn_out_2d_final = ggml_reshape_2d(ctx0, output, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + ggml_cpy(ctx0, new_state, + ggml_view_1d(ctx0, ssm_states_all, hparams.n_embd_s() * n_seqs, + kv_head * hparams.n_embd_s() * ggml_element_size(ssm_states_all)))); // z: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * n_tokens * n_seqs, head_dim] - ggml_tensor * z_2d = ggml_reshape_2d(ctx0, z, head_v_dim, num_v_heads * n_seq_tokens * n_seqs); + ggml_tensor * z_2d = ggml_reshape_4d(ctx0, z, head_v_dim, num_v_heads, n_seq_tokens, n_seqs); // Apply gated normalization: self.norm(core_attn_out, z) - ggml_tensor * attn_out_norm = build_norm_gated(attn_out_2d_final, model.layers[il].ssm_norm, z_2d, il); + ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il); // Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim] ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs); @@ -818,7 +783,8 @@ ggml_tensor * llm_build_qwen3next::build_layer_attn_linear( cb(cur, "linear_attn_out", il); // Reshape back to original dimensions - cur = ggml_cont_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs); + return cur; } @@ -839,7 +805,7 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int if (model.layers[il].ffn_up_shexp != nullptr) { ggml_tensor * ffn_shexp = build_ffn(cur, - model.layers[il].ffn_up_shexp, NULL, NULL, + 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, @@ -852,11 +818,9 @@ ggml_tensor * llm_build_qwen3next::build_layer_ffn(ggml_tensor * cur, const int ggml_tensor * shared_gate = build_lora_mm(model.layers[il].ffn_gate_inp_shexp, cur); cb(shared_gate, "shared_expert_gate", il); - // Apply sigmoid to the gate shared_gate = ggml_sigmoid(ctx0, shared_gate); cb(shared_gate, "shared_expert_gate_sigmoid", il); - // Apply the gate to the shared expert output ffn_shexp = ggml_mul(ctx0, ffn_shexp, shared_gate); cb(ffn_shexp, "ffn_shexp_gated", il); diff --git a/examples/talk-llama/unicode.cpp b/examples/talk-llama/unicode.cpp index adfc489d..b88d953b 100644 --- a/examples/talk-llama/unicode.cpp +++ b/examples/talk-llama/unicode.cpp @@ -1,16 +1,10 @@ -#if defined(_MSC_VER) -#define _SILENCE_CXX17_CODECVT_HEADER_DEPRECATION_WARNING -#endif - #include "unicode.h" #include "unicode-data.h" #include #include -#include #include #include -#include #include #include #include @@ -199,27 +193,6 @@ static std::unordered_map unicode_utf8_to_byte_map() { return map; } -static inline std::wstring unicode_wstring_from_utf8(const std::string & s) { -#if defined(__clang__) - // disable C++17 deprecation warning for std::codecvt_utf8 -# pragma clang diagnostic push -# pragma clang diagnostic ignored "-Wdeprecated-declarations" -#elif defined(__GNUC__) -# pragma GCC diagnostic push -# pragma GCC diagnostic ignored "-Wdeprecated-declarations" -#endif - - std::wstring_convert> conv; - -#if defined(__clang__) -# pragma clang diagnostic pop -#elif defined(__GNUC__) -# pragma GCC diagnostic pop -#endif - - return conv.from_bytes(s); -} - static std::vector unicode_byte_encoding_process(const std::vector & bpe_words) { std::vector bpe_encoded_words; for (const auto & word : bpe_words) { @@ -1028,10 +1001,10 @@ std::vector unicode_regex_split(const std::string & text, const std break; } } + const auto cpts_regex = unicode_cpts_from_utf8(regex_expr); if (use_collapsed) { // sanity-check that the original regex does not contain any non-ASCII characters - const auto cpts_regex = unicode_cpts_from_utf8(regex_expr); for (size_t i = 0; i < cpts_regex.size(); ++i) { if (cpts_regex[i] >= 128) { throw std::runtime_error("Regex includes both unicode categories and non-ASCII characters - not supported"); @@ -1087,7 +1060,7 @@ std::vector unicode_regex_split(const std::string & text, const std bpe_offsets = unicode_regex_split_stl(text_collapsed, regex_expr_collapsed, bpe_offsets); } else { // no unicode category used, we can use std::wregex directly - const std::wstring wregex_expr = unicode_wstring_from_utf8(regex_expr); + std::wstring wregex_expr(cpts_regex.begin(), cpts_regex.end()); // std::wregex \s does not mach non-ASCII whitespaces, using 0x0B as fallback std::wstring wtext(cpts.begin(), cpts.end());