1#include "models.h"
2
3llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4 const int64_t n_embd_head = hparams.n_embd_head_v;
5
6 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
7
8 const float f_logit_scale = hparams.f_logit_scale;
9
10 ggml_tensor * cur;
11 ggml_tensor * inpL;
12
13 inpL = build_inp_embd(model.tok_embd);
14
15 // inp_pos - contains the positions
16 ggml_tensor * inp_pos = build_inp_pos();
17
18 auto * inp_attn = build_attn_inp_kv_iswa();
19
20 ggml_tensor * inp_out_ids = build_inp_out_ids();
21
22 for (int il = 0; il < n_layer; ++il) {
23 const bool is_swa = hparams.is_swa(il);
24 // UNUSED:
25 // const float freq_base_l = model.get_rope_freq_base (cparams, il);
26 // const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
27
28 // norm
29 cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM, il);
30 cb(cur, "attn_norm", il);
31 ggml_tensor * ffn_inp = cur;
32
33 // self-attention
34 {
35 // rope freq factors for 128k context
36 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
37
38 // compute Q and K and RoPE them
39 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
40 cb(Qcur, "Qcur", il);
41 if (model.layers[il].bq) {
42 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
43 cb(Qcur, "Qcur", il);
44 }
45
46 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
47 cb(Kcur, "Kcur", il);
48 if (model.layers[il].bk) {
49 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
50 cb(Kcur, "Kcur", il);
51 }
52
53 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
54 cb(Vcur, "Vcur", il);
55 if (model.layers[il].bv) {
56 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
57 cb(Vcur, "Vcur", il);
58 }
59
60 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
61 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
62 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
63
64 if (is_swa) {
65 Qcur = ggml_rope_ext(
66 ctx0, Qcur, inp_pos, rope_factors,
67 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
68 ext_factor, attn_factor, beta_fast, beta_slow
69 );
70
71 Kcur = ggml_rope_ext(
72 ctx0, Kcur, inp_pos, rope_factors,
73 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
74 ext_factor, attn_factor, beta_fast, beta_slow
75 );
76 }
77
78 cb(Qcur, "Qcur", il);
79 cb(Kcur, "Kcur", il);
80 cb(Vcur, "Vcur", il);
81
82 cur = build_attn(inp_attn,
83 model.layers[il].wo, model.layers[il].bo,
84 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
85 }
86
87 if (il == n_layer - 1 && inp_out_ids) {
88 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
89 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
90 ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
91 }
92
93 ggml_tensor * attn_out = cur;
94
95 // feed-forward network
96 {
97 cur = build_ffn(ffn_inp,
98 model.layers[il].ffn_up, NULL, NULL,
99 model.layers[il].ffn_gate, NULL, NULL,
100 model.layers[il].ffn_down, NULL, NULL,
101 NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
102 cb(cur, "ffn_out", il);
103 }
104
105 // add together residual + FFN + self-attention
106 cur = ggml_add(ctx0, cur, inpL);
107 cur = ggml_add(ctx0, cur, attn_out);
108
109 cur = build_cvec(cur, il);
110 cb(cur, "l_out", il);
111
112 // input for next layer
113 inpL = cur;
114 }
115
116 cur = inpL;
117
118 cur = build_norm(cur, model.output_norm, NULL, LLM_NORM, -1);
119
120 cb(cur, "result_norm", -1);
121 res->t_embd = cur;
122
123 // lm_head
124 cur = build_lora_mm(model.output, cur);
125
126 if (f_logit_scale) {
127 cur = ggml_scale(ctx0, cur, f_logit_scale);
128 }
129
130 cb(cur, "result_output", -1);
131 res->t_logits = cur;
132
133 ggml_build_forward_expand(gf, cur);
134}