LLM_ARCH_GEMMA,
LLM_ARCH_STARCODER2,
LLM_ARCH_MAMBA,
+ LLM_ARCH_COMMAND_R,
LLM_ARCH_UNKNOWN,
};
{ LLM_ARCH_GEMMA, "gemma" },
{ LLM_ARCH_STARCODER2, "starcoder2" },
{ LLM_ARCH_MAMBA, "mamba" },
+ { LLM_ARCH_COMMAND_R, "command-r" },
{ LLM_ARCH_UNKNOWN, "(unknown)" },
};
LLM_KV_EXPERT_COUNT,
LLM_KV_EXPERT_USED_COUNT,
LLM_KV_POOLING_TYPE,
+ LLM_KV_LOGIT_SCALE,
LLM_KV_ATTENTION_HEAD_COUNT,
LLM_KV_ATTENTION_HEAD_COUNT_KV,
{ LLM_KV_EXPERT_COUNT, "%s.expert_count" },
{ LLM_KV_EXPERT_USED_COUNT, "%s.expert_used_count" },
{ LLM_KV_POOLING_TYPE , "%s.pooling_type" },
+ { LLM_KV_LOGIT_SCALE, "%s.logit_scale" },
{ LLM_KV_ATTENTION_HEAD_COUNT, "%s.attention.head_count" },
{ LLM_KV_ATTENTION_HEAD_COUNT_KV, "%s.attention.head_count_kv" },
{ LLM_TENSOR_SSM_OUT, "blk.%d.ssm_out" },
},
},
+ {
+ LLM_ARCH_COMMAND_R,
+ {
+ { LLM_TENSOR_TOKEN_EMBD, "token_embd" },
+ { LLM_TENSOR_OUTPUT_NORM, "output_norm" },
+ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" },
+ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" },
+ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" },
+ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" },
+ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" },
+ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" },
+ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" },
+ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" },
+ },
+ },
{
LLM_ARCH_UNKNOWN,
{
MODEL_20B,
MODEL_30B,
MODEL_34B,
+ MODEL_35B,
MODEL_40B,
MODEL_65B,
MODEL_70B,
float f_clamp_kqv = 0.0f;
float f_max_alibi_bias = 0.0f;
+ float f_logit_scale = 0.0f;
bool causal_attn = true;
bool need_kq_pos = false;
case MODEL_20B: return "20B";
case MODEL_30B: return "30B";
case MODEL_34B: return "34B";
+ case MODEL_35B: return "35B";
case MODEL_40B: return "40B";
case MODEL_65B: return "65B";
case MODEL_70B: return "70B";
default: model.type = e_model::MODEL_UNKNOWN;
}
} break;
+ case LLM_ARCH_COMMAND_R:
+ {
+ ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale);
+ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+ switch (hparams.n_layer) {
+ case 40: model.type = e_model::MODEL_35B; break;
+ default: model.type = e_model::MODEL_UNKNOWN;
+ }
+ } break;
default: (void)0;
}
LLAMA_LOG_INFO("%s: f_norm_rms_eps = %.1e\n", __func__, hparams.f_norm_rms_eps);
LLAMA_LOG_INFO("%s: f_clamp_kqv = %.1e\n", __func__, hparams.f_clamp_kqv);
LLAMA_LOG_INFO("%s: f_max_alibi_bias = %.1e\n", __func__, hparams.f_max_alibi_bias);
+ LLAMA_LOG_INFO("%s: f_logit_scale = %.1e\n", __func__, hparams.f_logit_scale);
LLAMA_LOG_INFO("%s: n_ff = %u\n", __func__, hparams.n_ff);
LLAMA_LOG_INFO("%s: n_expert = %u\n", __func__, hparams.n_expert);
LLAMA_LOG_INFO("%s: n_expert_used = %u\n", __func__, hparams.n_expert_used);
layer.ssm_out = ml.create_tensor(ctx_split, tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd});
}
} break;
+ case LLM_ARCH_COMMAND_R:
+ {
+ model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+
+ // output
+ {
+ model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd});
+ // init output from the input tok embed
+ model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab});
+ ml.n_created--; // artificial tensor
+ ml.size_data += ggml_nbytes(model.output);
+ }
+
+ for (int i = 0; i < n_layer; ++i) {
+ ggml_context * ctx_layer = ctx_for_layer(i);
+ ggml_context * ctx_split = ctx_for_layer_split(i);
+
+ auto & layer = model.layers[i];
+
+ layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd});
+
+ layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd});
+ layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_gqa});
+ layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_gqa});
+ layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd});
+
+ layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff});
+ layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd});
+ layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff});
+ }
+ } break;
default:
throw std::runtime_error("unknown architecture");
}
return gf;
}
+
+ struct ggml_cgraph * build_command_r() {
+
+ struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
+
+ const int64_t n_embd_head = hparams.n_embd_head_v;
+ GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+ const float f_logit_scale = hparams.f_logit_scale;
+
+ struct ggml_tensor * cur;
+ struct ggml_tensor * inpL;
+
+ inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb);
+
+ // inp_pos - contains the positions
+ struct ggml_tensor * inp_pos = build_inp_pos();
+
+ // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
+ struct ggml_tensor * KQ_mask = build_inp_KQ_mask();
+
+ for (int il = 0; il < n_layer; ++il) {
+
+ // norm
+ cur = llm_build_norm(ctx0, inpL, hparams,
+ model.layers[il].attn_norm, NULL,
+ LLM_NORM, cb, il);
+ cb(cur, "attn_norm", il);
+ struct ggml_tensor * ffn_inp = cur;
+
+ // self-attention
+ {
+ // compute Q and K and RoPE them
+ struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+ cb(Qcur, "Qcur", il);
+ if (model.layers[il].bq) {
+ Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
+ cb(Qcur, "Qcur", il);
+ }
+
+ struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+ cb(Kcur, "Kcur", il);
+ if (model.layers[il].bk) {
+ Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
+ cb(Kcur, "Kcur", il);
+ }
+
+ struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+ cb(Vcur, "Vcur", il);
+ if (model.layers[il].bv) {
+ Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
+ cb(Vcur, "Vcur", il);
+ }
+
+ Qcur = ggml_rope_custom(
+ ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
+ n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow
+ );
+ cb(Qcur, "Qcur", il);
+
+ Kcur = ggml_rope_custom(
+ ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+ n_rot, rope_type, 0, n_orig_ctx, freq_base, freq_scale,
+ ext_factor, attn_factor, beta_fast, beta_slow
+ );
+ cb(Kcur, "Kcur", il);
+
+ cur = llm_build_kv(ctx0, model, hparams, kv_self, gf,
+ model.layers[il].wo, model.layers[il].bo,
+ Kcur, Vcur, Qcur, KQ_mask, nullptr, n_ctx, n_tokens, kv_head, n_kv, 1.0f/sqrtf(float(n_embd_head)), cb, il);
+ }
+
+ struct ggml_tensor * attn_out = cur;
+
+ // feed-forward network
+ {
+ cur = llm_build_ffn(ctx0, ffn_inp,
+ model.layers[il].ffn_up, NULL,
+ model.layers[il].ffn_gate, NULL,
+ model.layers[il].ffn_down, NULL,
+ NULL,
+ LLM_FFN_SILU, LLM_FFN_PAR, cb, il);
+ cb(cur, "ffn_out", il);
+ }
+
+ // add together residual + FFN + self-attention
+ cur = ggml_add(ctx0, cur, inpL);
+ cur = ggml_add(ctx0, cur, attn_out);
+ cb(cur, "l_out", il);
+
+ // input for next layer
+ inpL = cur;
+ }
+
+ cur = inpL;
+
+ cur = llm_build_norm(ctx0, cur, hparams,
+ model.output_norm, NULL,
+ LLM_NORM, cb, -1);
+ cb(cur, "result_norm", -1);
+
+ // lm_head
+ cur = ggml_mul_mat(ctx0, model.output, cur);
+
+ if (f_logit_scale) {
+ cur = ggml_scale(ctx0, cur, f_logit_scale);
+ }
+
+ cb(cur, "result_output", -1);
+
+ ggml_build_forward_expand(gf, cur);
+
+ return gf;
+
+ }
};
static struct ggml_cgraph * llama_build_graph_defrag(llama_context & lctx, const std::vector<uint32_t> & ids) {
{
result = llm.build_mamba();
} break;
+ case LLM_ARCH_COMMAND_R:
+ {
+ result = llm.build_command_r();
+ } break;
default:
GGML_ASSERT(false);
}
case LLM_ARCH_ORION:
case LLM_ARCH_INTERNLM2:
case LLM_ARCH_MINICPM:
+ case LLM_ARCH_COMMAND_R:
return LLAMA_ROPE_TYPE_NORM;
// the pairs of head values are offset by n_rot/2
overview / pkg / ggml / sources / llama.cpp / commitdiff