| 1 | #include "models.h" |
| 2 | |
| 3 | llm_build_smollm3::llm_build_smollm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { |
| 4 | const int64_t n_embd_head = hparams.n_embd_head_v; |
| 5 | |
| 6 | GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); |
| 7 | GGML_ASSERT(n_embd_head == hparams.n_rot); |
| 8 | |
| 9 | ggml_tensor * cur; |
| 10 | ggml_tensor * inpL; |
| 11 | |
| 12 | inpL = build_inp_embd(model.tok_embd); |
| 13 | |
| 14 | // inp_pos - contains the positions |
| 15 | ggml_tensor * inp_pos = build_inp_pos(); |
| 16 | |
| 17 | auto * inp_attn = build_attn_inp_kv(); |
| 18 | |
| 19 | const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale; |
| 20 | |
| 21 | ggml_tensor * inp_out_ids = build_inp_out_ids(); |
| 22 | |
| 23 | for (int il = 0; il < n_layer; ++il) { |
| 24 | ggml_tensor * inpSA = inpL; |
| 25 | |
| 26 | const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0; |
| 27 | |
| 28 | // norm |
| 29 | cur = build_norm(inpL, |
| 30 | model.layers[il].attn_norm, NULL, |
| 31 | LLM_NORM_RMS, il); |
| 32 | cb(cur, "attn_norm", il); |
| 33 | |
| 34 | // self-attention |
| 35 | { |
| 36 | // compute Q and K and RoPE them |
| 37 | ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); |
| 38 | cb(Qcur, "Qcur", il); |
| 39 | if (model.layers[il].bq) { |
| 40 | Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); |
| 41 | cb(Qcur, "Qcur", il); |
| 42 | } |
| 43 | ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); |
| 44 | cb(Kcur, "Kcur", il); |
| 45 | if (model.layers[il].bk) { |
| 46 | Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); |
| 47 | cb(Kcur, "Kcur", il); |
| 48 | } |
| 49 | ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); |
| 50 | cb(Vcur, "Vcur", il); |
| 51 | if (model.layers[il].bv) { |
| 52 | Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); |
| 53 | cb(Vcur, "Vcur", il); |
| 54 | } |
| 55 | Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); |
| 56 | Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); |
| 57 | Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); |
| 58 | |
| 59 | if (use_rope) { |
| 60 | Qcur = ggml_rope_ext( |
nothing calls this directly
no test coverage detected