/src/llama.cpp/src/models/cohere2-iswa.cpp
Line | Count | Source |
1 | | #include "models.h" |
2 | | |
3 | 0 | llm_build_cohere2_iswa::llm_build_cohere2_iswa(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 | const float f_logit_scale = hparams.f_logit_scale; |
9 | |
|
10 | 0 | ggml_tensor * cur; |
11 | 0 | ggml_tensor * inpL; |
12 | |
|
13 | 0 | inpL = build_inp_embd(model.tok_embd); |
14 | | |
15 | | // inp_pos - contains the positions |
16 | 0 | ggml_tensor * inp_pos = build_inp_pos(); |
17 | |
|
18 | 0 | auto * inp_attn = build_attn_inp_kv_iswa(); |
19 | |
|
20 | 0 | ggml_tensor * inp_out_ids = build_inp_out_ids(); |
21 | |
|
22 | 0 | for (int il = 0; il < n_layer; ++il) { |
23 | 0 | const bool is_swa = hparams.is_swa(il); |
24 | | |
25 | | // norm |
26 | 0 | cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il); |
27 | 0 | cb(cur, "attn_norm", il); |
28 | 0 | ggml_tensor * ffn_inp = cur; |
29 | | |
30 | | // self-attention |
31 | 0 | { |
32 | | // rope freq factors for 128k context |
33 | 0 | ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); |
34 | | |
35 | | // compute Q and K and RoPE them |
36 | 0 | ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur); |
37 | 0 | cb(Qcur, "Qcur", il); |
38 | 0 | if (model.layers[il].bq) { |
39 | 0 | Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq); |
40 | 0 | cb(Qcur, "Qcur", il); |
41 | 0 | } |
42 | |
|
43 | 0 | ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur); |
44 | 0 | cb(Kcur, "Kcur", il); |
45 | 0 | if (model.layers[il].bk) { |
46 | 0 | Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk); |
47 | 0 | cb(Kcur, "Kcur", il); |
48 | 0 | } |
49 | |
|
50 | 0 | ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur); |
51 | 0 | cb(Vcur, "Vcur", il); |
52 | 0 | if (model.layers[il].bv) { |
53 | 0 | Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv); |
54 | 0 | cb(Vcur, "Vcur", il); |
55 | 0 | } |
56 | |
|
57 | 0 | Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); |
58 | 0 | Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); |
59 | 0 | Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens); |
60 | |
|
61 | 0 | if (is_swa) { |
62 | 0 | Qcur = ggml_rope_ext( |
63 | 0 | ctx0, Qcur, inp_pos, rope_factors, |
64 | 0 | n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, |
65 | 0 | ext_factor, attn_factor, beta_fast, beta_slow |
66 | 0 | ); |
67 | |
|
68 | 0 | Kcur = ggml_rope_ext( |
69 | 0 | ctx0, Kcur, inp_pos, rope_factors, |
70 | 0 | n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, |
71 | 0 | ext_factor, attn_factor, beta_fast, beta_slow |
72 | 0 | ); |
73 | 0 | } |
74 | |
|
75 | 0 | cb(Qcur, "Qcur", il); |
76 | 0 | cb(Kcur, "Kcur", il); |
77 | 0 | cb(Vcur, "Vcur", il); |
78 | |
|
79 | 0 | cur = build_attn(inp_attn, |
80 | 0 | model.layers[il].wo, model.layers[il].bo, |
81 | 0 | Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il); |
82 | 0 | } |
83 | |
|
84 | 0 | if (il == n_layer - 1 && inp_out_ids) { |
85 | 0 | cur = ggml_get_rows(ctx0, cur, inp_out_ids); |
86 | 0 | inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); |
87 | 0 | ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids); |
88 | 0 | } |
89 | |
|
90 | 0 | ggml_tensor * attn_out = cur; |
91 | | |
92 | | // feed-forward network |
93 | 0 | { |
94 | 0 | cur = build_ffn(ffn_inp, |
95 | 0 | model.layers[il].ffn_up, NULL, NULL, |
96 | 0 | model.layers[il].ffn_gate, NULL, NULL, |
97 | 0 | model.layers[il].ffn_down, NULL, NULL, |
98 | 0 | NULL, LLM_FFN_SILU, LLM_FFN_PAR, il); |
99 | 0 | cb(cur, "ffn_out", il); |
100 | 0 | } |
101 | | |
102 | | // add together residual + FFN + self-attention |
103 | 0 | cur = ggml_add(ctx0, cur, inpL); |
104 | 0 | cur = ggml_add(ctx0, cur, attn_out); |
105 | |
|
106 | 0 | cur = build_cvec(cur, il); |
107 | 0 | cb(cur, "l_out", il); |
108 | | |
109 | | // input for next layer |
110 | 0 | inpL = cur; |
111 | 0 | } |
112 | |
|
113 | 0 | cur = inpL; |
114 | |
|
115 | 0 | cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1); |
116 | |
|
117 | 0 | cb(cur, "result_norm", -1); |
118 | 0 | res->t_embd = cur; |
119 | | |
120 | | // lm_head |
121 | 0 | cur = build_lora_mm(model.output, cur); |
122 | |
|
123 | 0 | if (f_logit_scale) { |
124 | 0 | cur = ggml_scale(ctx0, cur, f_logit_scale); |
125 | 0 | } |
126 | |
|
127 | 0 | cb(cur, "result_output", -1); |
128 | 0 | res->t_logits = cur; |
129 | |
|
130 | 0 | ggml_build_forward_expand(gf, cur); |
131 | 0 | } |