1#include "models.h"
2
3
4llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
5 const int64_t n_embd_head = hparams.n_embd_head_v;
6 const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
7
8 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
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();
19
20 ggml_tensor * inp_out_ids = build_inp_out_ids();
21
22 for (int il = 0; il < n_layer; ++il) {
23 ggml_tensor * inpSA = inpL;
24
25 cur = build_norm(inpL,
26 model.layers[il].attn_norm,
27 NULL,
28 LLM_NORM_RMS, il);
29 cb(cur, "attn_norm", il);
30
31 // self-attention
32 {
33 ggml_tensor * Qcur = nullptr;
34 ggml_tensor * Kcur = nullptr;
35 ggml_tensor * Vcur = nullptr;
36
37 if (model.layers[il].wqkv == nullptr) {
38 Qcur = build_lora_mm(model.layers[il].wq, cur);
39 if (model.layers[il].bq) {
40 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
41 }
42 Kcur = build_lora_mm(model.layers[il].wk, cur);
43 if (model.layers[il].bk) {
44 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
45 }
46 Vcur = build_lora_mm(model.layers[il].wv, cur);
47 if (model.layers[il].bv) {
48 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
49 }
50 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
51 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
52 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
53 } else {
54 cur = build_lora_mm(model.layers[il].wqkv, cur);
55 cb(cur, "wqkv", il);
56 if (model.layers[il].bqkv) {
57 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
58 cb(cur, "bqkv", il);
59 }
60 Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
61 Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
62 Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
63 }
64
65 //printf("freq_base: %f freq_scale: %f ext_factor: %f attn_factor: %f\n", freq_base, freq_scale, ext_factor, attn_factor);
66 Qcur = ggml_rope_ext(
67 ctx0, Qcur, inp_pos, nullptr,
68 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
69 ext_factor, attn_factor, beta_fast, beta_slow
70 );
71
72 Kcur = ggml_rope_ext(
73 ctx0, Kcur, inp_pos, nullptr,
74 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
75 ext_factor, attn_factor, beta_fast, beta_slow
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, NULL,
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 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
90 }
91
92 // Add the input
93 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
94 cb(ffn_inp, "ffn_inp", il);
95
96 // FF
97 {
98 cur = build_norm(ffn_inp,
99 model.layers[il].ffn_norm,
100 NULL,
101 LLM_NORM_RMS, il);
102 cb(cur, "ffn_norm", il);
103
104 cur = build_ffn(cur,
105 model.layers[il].ffn_up, NULL, NULL,
106 NULL, NULL, NULL,
107 model.layers[il].ffn_down, NULL, NULL,
108 NULL,
109 LLM_FFN_SWIGLU, LLM_FFN_SEQ, il);
110 cb(cur, "ffn_out", il);
111
112 }
113
114 inpL = ggml_add(ctx0, cur, ffn_inp);
115 cb(inpL, "l_out", il);
116 }
117
118 cur = build_norm(inpL,
119 model.output_norm,
120 NULL,
121 LLM_NORM_RMS, -1);
122
123 cb(cur, "result_norm", -1);
124 res->t_embd = cur;
125
126 cur = build_lora_mm(model.output, cur);
127
128 cb(cur, "result_output", -1);
129 res->t_logits = cur;
130
131 ggml_build_forward_expand(gf, cur);
132}