| 1 | #include "models.h" |
| 2 | |
| 3 | llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_graph_context_mamba(params) { |
| 4 | const int64_t n_embd_head = hparams.n_embd_head_v; |
| 5 | |
| 6 | ggml_tensor * cur; |
| 7 | ggml_tensor * inpL; |
| 8 | |
| 9 | // {n_embd, n_tokens} |
| 10 | inpL = build_inp_embd(model.tok_embd); |
| 11 | |
| 12 | auto * inp_hybrid = build_inp_mem_hybrid(); |
| 13 | |
| 14 | ggml_tensor * inp_out_ids = build_inp_out_ids(); |
| 15 | |
| 16 | for (int il = 0; il < n_layer; ++il) { |
| 17 | const int64_t n_head_kv = hparams.n_head_kv(il); |
| 18 | |
| 19 | cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il); |
| 20 | cb(cur, "attn_norm", il); |
| 21 | |
| 22 | if (n_head_kv == 0) { |
| 23 | cur = build_mamba_layer(inp_hybrid->get_recr(), cur, model, ubatch, il); |
| 24 | } else { |
| 25 | // Attention |
| 26 | |
| 27 | struct ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); |
| 28 | struct ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); |
| 29 | struct ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); |
| 30 | |
| 31 | cb(Qcur, "Qcur", il); |
| 32 | cb(Kcur, "Kcur", il); |
| 33 | cb(Vcur, "Vcur", il); |
| 34 | |
| 35 | Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); |
| 36 | Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); |
| 37 | Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); |
| 38 | |
| 39 | cb(Qcur, "Qcur", il); |
| 40 | cb(Kcur, "Kcur", il); |
| 41 | cb(Vcur, "Vcur", il); |
| 42 | |
| 43 | // No RoPE :) |
| 44 | cur = build_attn(inp_hybrid->get_attn(), |
| 45 | model.layers[il].wo, NULL, |
| 46 | Qcur, Kcur, Vcur, NULL, NULL, NULL, 1.0f/sqrtf(float(n_embd_head)), il); |
| 47 | } |
| 48 | if (il == n_layer - 1 && inp_out_ids) { |
| 49 | cur = ggml_get_rows(ctx0, cur, inp_out_ids); |
| 50 | inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); |
| 51 | } |
| 52 | // residual |
| 53 | struct ggml_tensor * ffn_inp = ggml_add(ctx0, inpL, cur); |
| 54 | cb(cur, "ffn_inp", il); |
| 55 | |
| 56 | cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il); |
| 57 | cb(cur, "ffn_norm", il); |
| 58 | |
| 59 | // feed-forward network |
| 60 | if (model.layers[il].ffn_gate_inp == nullptr) { |
nothing calls this directly
no test coverage detected