1#include "models.h"
2
3llm_build_hunyuan_dense::llm_build_hunyuan_dense(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 GGML_ASSERT(n_embd_head == hparams.n_rot);
8
9 ggml_tensor * cur;
10 ggml_tensor * inpL;
11
12 inpL = build_inp_embd(model.tok_embd);
13
14 // inp_pos - contains the positions
15 ggml_tensor * inp_pos = build_inp_pos();
16
17 auto * inp_attn = build_attn_inp_kv();
18
19 const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
20
21 ggml_tensor * inp_out_ids = build_inp_out_ids();
22
23 for (int il = 0; il < n_layer; ++il) {
24 ggml_tensor * inpSA = inpL;
25
26 // norm
27 cur = build_norm(inpL,
28 model.layers[il].attn_norm, NULL,
29 LLM_NORM_RMS, il);
30 cb(cur, "attn_norm", il);
31 // self-attention
32 {
33 // rope freq factors for llama3; may return nullptr for llama2 and other models
34 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
35
36 // compute Q and K and RoPE them
37 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
38 cb(Qcur, "Qcur", il);
39 if (model.layers[il].bq) {
40 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
41 cb(Qcur, "Qcur", il);
42 }
43 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
44 cb(Kcur, "Kcur", il);
45 if (model.layers[il].bk) {
46 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
47 cb(Kcur, "Kcur", il);
48 }
49 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
50 cb(Vcur, "Vcur", il);
51 if (model.layers[il].bv) {
52 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
53 cb(Vcur, "Vcur", il);
54 }
55 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
56 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
57 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
58
59 Qcur = ggml_rope_ext(
60 ctx0, Qcur, inp_pos, rope_factors,
61 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
62 ext_factor, attn_factor, beta_fast, beta_slow
63 );
64
65 cb(Qcur, "Qcur", il);
66 cb(Kcur, "Kcur", il);
67 cb(Vcur, "Vcur", il);
68
69 Kcur = ggml_rope_ext(
70 ctx0, Kcur, inp_pos, rope_factors,
71 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
72 ext_factor, attn_factor, beta_fast, beta_slow
73 );
74
75 Kcur = build_norm(Kcur,
76 model.layers[il].attn_k_norm, nullptr,
77 LLM_NORM_RMS, il);
78 cb(Kcur, "Kcur_norm", il);
79
80 Qcur = build_norm(Qcur,
81 model.layers[il].attn_q_norm, nullptr,
82 LLM_NORM_RMS, il);
83 cb(Qcur, "Qcur_norm", il);
84
85 cur = build_attn(inp_attn,
86 model.layers[il].wo, model.layers[il].bo,
87 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
88 cb(cur, "attn_out", il);
89 }
90 if (il == n_layer - 1 && inp_out_ids) {
91 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
92 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
93 }
94 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
95 cb(ffn_inp, "ffn_inp", il);
96
97 cur = build_norm(ffn_inp,
98 model.layers[il].ffn_norm, NULL,
99 LLM_NORM_RMS, il);
100 cb(cur, "ffn_norm", il);
101 // feed-forward network (non-MoE)
102 ggml_tensor * cur_mlp = build_ffn(cur,
103 model.layers[il].ffn_up, NULL, NULL,
104 model.layers[il].ffn_gate, NULL, NULL,
105 model.layers[il].ffn_down, NULL, NULL,
106 NULL,
107 LLM_FFN_SILU, LLM_FFN_PAR, il);
108 cb(cur_mlp, "ffn_out", il);
109
110 cur = ggml_add(ctx0, cur_mlp, ffn_inp);
111
112 cur = build_cvec(cur, il);
113 cb(cur, "l_out", il);
114
115 // input for next layer
116 inpL = cur;
117 }
118 cur = inpL;
119
120 cur = build_norm(cur,
121 model.output_norm, NULL,
122 LLM_NORM_RMS, -1);
123
124 cb(cur, "result_norm", -1);
125 res->t_embd = cur;
126 // lm_head
127 cur = build_lora_mm(model.output, cur);
128 cb(cur, "result_output", -1);
129 res->t_logits = cur;
130
131 ggml_build_forward_expand(gf, cur);
132}