| 2 | |
| 3 | |
| 4 | llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { |
| 5 | const int64_t n_embd_head = hparams.n_embd_head_v; |
| 6 | const int64_t n_embd_gqa = hparams.n_embd_v_gqa(); |
| 7 | |
| 8 | GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); |
| 9 | GGML_ASSERT(n_embd_head == hparams.n_rot); |
| 10 | |
| 11 | ggml_tensor * cur; |
| 12 | ggml_tensor * inpL; |
| 13 | |
| 14 | inpL = build_inp_embd(model.tok_embd); |
| 15 | |
| 16 | // inp_pos - contains the positions |
| 17 | ggml_tensor * inp_pos = build_inp_pos(); |
| 18 | |
| 19 | auto * inp_attn = build_attn_inp_kv(); |
| 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 | // norm |
| 27 | cur = build_norm(inpL, |
| 28 | model.layers[il].attn_norm, NULL, |
| 29 | LLM_NORM, il); |
| 30 | cb(cur, "attn_norm", il); |
| 31 | |
| 32 | // self-attention |
| 33 | { |
| 34 | ggml_tensor * Qcur = nullptr; |
| 35 | ggml_tensor * Kcur = nullptr; |
| 36 | ggml_tensor * Vcur = nullptr; |
| 37 | |
| 38 | cur = build_lora_mm(model.layers[il].wqkv, cur); |
| 39 | cb(cur, "wqkv", il); |
| 40 | |
| 41 | cur = ggml_clamp(ctx0, cur, -hparams.f_clamp_kqv, hparams.f_clamp_kqv); |
| 42 | cb(cur, "wqkv_clamped", il); |
| 43 | |
| 44 | Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd)); |
| 45 | Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd)); |
| 46 | Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa)); |
| 47 | |
| 48 | Qcur = ggml_rope_ext( |
| 49 | ctx0, Qcur, inp_pos, nullptr, |
| 50 | n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, |
| 51 | ext_factor, attn_factor, beta_fast, beta_slow |
| 52 | ); |
| 53 | |
| 54 | Kcur = ggml_rope_ext( |
| 55 | ctx0, Kcur, inp_pos, nullptr, |
| 56 | n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, |
| 57 | ext_factor, attn_factor, beta_fast, beta_slow |
| 58 | ); |
| 59 | |
| 60 | cb(Qcur, "Qcur", il); |
| 61 | cb(Kcur, "Kcur", il); |
nothing calls this directly
no test coverage detected