/src/llama.cpp/src/models/refact.cpp
Line | Count | Source |
1 | | #include "models.h" |
2 | | |
3 | 0 | llm_build_refact::llm_build_refact(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) { |
4 | 0 | const int64_t n_embd_head = hparams.n_embd_head_v; |
5 | |
|
6 | 0 | GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); |
7 | |
|
8 | 0 | ggml_tensor * cur; |
9 | 0 | ggml_tensor * inpL; |
10 | |
|
11 | 0 | inpL = build_inp_embd(model.tok_embd); |
12 | |
|
13 | 0 | auto * inp_attn = build_attn_inp_kv(); |
14 | |
|
15 | 0 | ggml_tensor * inp_out_ids = build_inp_out_ids(); |
16 | |
|
17 | 0 | for (int il = 0; il < n_layer; ++il) { |
18 | 0 | ggml_tensor * inpSA = inpL; |
19 | |
|
20 | 0 | cur = build_norm(inpL, |
21 | 0 | model.layers[il].attn_norm, NULL, |
22 | 0 | LLM_NORM_RMS, il); |
23 | 0 | cb(cur, "attn_norm", il); |
24 | | |
25 | | // self-attention |
26 | 0 | { |
27 | 0 | ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); |
28 | 0 | cb(Qcur, "Qcur", il); |
29 | |
|
30 | 0 | ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); |
31 | 0 | cb(Kcur, "Kcur", il); |
32 | |
|
33 | 0 | ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); |
34 | 0 | cb(Vcur, "Vcur", il); |
35 | |
|
36 | 0 | Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); |
37 | 0 | Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); |
38 | 0 | Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); |
39 | |
|
40 | 0 | cb(Qcur, "Qcur", il); |
41 | 0 | cb(Kcur, "Kcur", il); |
42 | 0 | cb(Vcur, "Vcur", il); |
43 | |
|
44 | 0 | cur = build_attn(inp_attn, |
45 | 0 | model.layers[il].wo, NULL, |
46 | 0 | Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); |
47 | 0 | } |
48 | 0 | if (il == n_layer - 1 && inp_out_ids) { |
49 | 0 | cur = ggml_get_rows(ctx0, cur, inp_out_ids); |
50 | 0 | inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); |
51 | 0 | } |
52 | 0 | ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); |
53 | 0 | cb(ffn_inp, "ffn_inp", il); |
54 | | |
55 | | // feed-forward network |
56 | 0 | { |
57 | 0 | cur = build_norm(ffn_inp, |
58 | 0 | model.layers[il].ffn_norm, NULL, |
59 | 0 | LLM_NORM_RMS, il); |
60 | 0 | cb(cur, "ffn_norm", il); |
61 | |
|
62 | 0 | cur = build_ffn(cur, |
63 | 0 | model.layers[il].ffn_up, NULL, NULL, |
64 | 0 | model.layers[il].ffn_gate, NULL, NULL, |
65 | 0 | model.layers[il].ffn_down, NULL, NULL, |
66 | 0 | NULL, |
67 | 0 | LLM_FFN_SILU, LLM_FFN_PAR, il); |
68 | 0 | cb(cur, "ffn_out", il); |
69 | 0 | } |
70 | 0 | cur = ggml_add(ctx0, cur, ffn_inp); |
71 | |
|
72 | 0 | cur = build_cvec(cur, il); |
73 | 0 | cb(cur, "l_out", il); |
74 | | |
75 | | // input for next layer |
76 | 0 | inpL = cur; |
77 | 0 | } |
78 | 0 | cur = inpL; |
79 | |
|
80 | 0 | cur = build_norm(cur, |
81 | 0 | model.output_norm, NULL, |
82 | 0 | LLM_NORM_RMS, -1); |
83 | |
|
84 | 0 | cb(cur, "result_norm", -1); |
85 | 0 | res->t_embd = cur; |
86 | | |
87 | | // lm_head |
88 | 0 | cur = build_lora_mm(model.output, cur); |
89 | |
|
90 | 0 | cb(cur, "result_output", -1); |
91 | 0 | res->t_logits = cur; |
92 | |
|
93 | 0 | ggml_build_forward_expand(gf, cur); |
94 | 0 | } |