/src/llama.cpp/src/models/t5-enc.cpp
Line | Count | Source |
1 | | #include "models.h" |
2 | | |
3 | 0 | llm_build_t5_enc::llm_build_t5_enc(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 | ggml_tensor * pos_bucket_enc = build_inp_pos_bucket_enc(); |
14 | |
|
15 | 0 | auto * inp_attn = build_attn_inp_no_cache(); |
16 | |
|
17 | 0 | ggml_tensor * inp_out_ids = build_inp_out_ids(); |
18 | |
|
19 | 0 | for (int il = 0; il < n_layer; ++il) { |
20 | 0 | ggml_tensor * inpSA = inpL; |
21 | | |
22 | | // norm |
23 | 0 | cur = build_norm(inpL, |
24 | 0 | model.layers[il].attn_norm_enc, NULL, |
25 | 0 | LLM_NORM_RMS, il); |
26 | 0 | cb(cur, "attn_norm", il); |
27 | | |
28 | | // self-attention |
29 | 0 | { |
30 | 0 | ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_enc, cur); |
31 | 0 | cb(Qcur, "Qcur", il); |
32 | |
|
33 | 0 | ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_enc, cur); |
34 | 0 | cb(Kcur, "Kcur", il); |
35 | |
|
36 | 0 | ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_enc, cur); |
37 | 0 | cb(Vcur, "Vcur", il); |
38 | |
|
39 | 0 | Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); |
40 | 0 | Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); |
41 | 0 | Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); |
42 | |
|
43 | 0 | ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b_enc ? model.layers[il].attn_rel_b_enc : model.layers[0].attn_rel_b_enc; |
44 | 0 | ggml_tensor * kq_b = build_pos_bias(pos_bucket_enc, attn_rel_b); |
45 | |
|
46 | 0 | cur = build_attn(inp_attn, |
47 | 0 | model.layers[il].wo_enc, nullptr, |
48 | 0 | Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il); |
49 | 0 | cb(cur, "kqv_out", il); |
50 | 0 | } |
51 | 0 | if (il == n_layer - 1 && inp_out_ids) { |
52 | 0 | cur = ggml_get_rows(ctx0, cur, inp_out_ids); |
53 | 0 | inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids); |
54 | 0 | } |
55 | 0 | ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA); |
56 | 0 | cb(ffn_inp, "ffn_inp", il); |
57 | | |
58 | | // feed-forward network |
59 | 0 | { |
60 | 0 | cur = build_norm(ffn_inp, |
61 | 0 | model.layers[il].ffn_norm_enc, NULL, |
62 | 0 | LLM_NORM_RMS, il); |
63 | 0 | cb(cur, "ffn_norm", il); |
64 | | |
65 | | // T5 uses relu, flan-T5 uses gelu-gated |
66 | 0 | cur = build_ffn(cur, |
67 | 0 | model.layers[il].ffn_up_enc, NULL, NULL, |
68 | 0 | model.layers[il].ffn_gate_enc, NULL, NULL, |
69 | 0 | model.layers[il].ffn_down_enc, NULL, NULL, |
70 | 0 | NULL, |
71 | 0 | model.layers[il].ffn_gate_enc ? LLM_FFN_GELU : LLM_FFN_RELU, |
72 | 0 | model.layers[il].ffn_gate_enc ? LLM_FFN_PAR : LLM_FFN_SEQ, |
73 | 0 | il); |
74 | 0 | cb(cur, "ffn_out", il); |
75 | 0 | } |
76 | 0 | cur = ggml_add(ctx0, cur, ffn_inp); |
77 | 0 | cb(cur, "ffn_out", il); |
78 | |
|
79 | 0 | cur = build_cvec(cur, il); |
80 | 0 | cb(cur, "l_out", il); |
81 | | |
82 | | // input for next layer |
83 | 0 | inpL = cur; |
84 | 0 | } |
85 | 0 | cur = inpL; |
86 | 0 | cb(cur, "result_embd", -1); |
87 | |
|
88 | 0 | cur = build_norm(cur, |
89 | 0 | model.output_norm_enc, NULL, |
90 | 0 | LLM_NORM_RMS, -1); |
91 | |
|
92 | 0 | cb(cur, "result_norm", -1); |
93 | 0 | res->t_embd = cur; |
94 | |
|
95 | 0 | ggml_build_forward_expand(gf, cur); |
96 | 0 | } |