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