1#include "models.h"
2
3llm_build_stablelm::llm_build_stablelm(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 ggml_tensor * cur;
9 ggml_tensor * inpL;
10
11 inpL = build_inp_embd(model.tok_embd);
12
13 // inp_pos - contains the positions
14 ggml_tensor * inp_pos = build_inp_pos();
15
16 auto * inp_attn = build_attn_inp_kv();
17
18 ggml_tensor * inp_out_ids = build_inp_out_ids();
19
20 for (int il = 0; il < n_layer; ++il) {
21 // norm
22 cur = build_norm(inpL,
23 model.layers[il].attn_norm,
24 model.layers[il].attn_norm_b,
25 LLM_NORM, il);
26 cb(cur, "attn_norm", il);
27
28 ggml_tensor * inpSA = cur;
29
30 // self-attention
31 {
32 // compute Q and K and RoPE them
33 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
34 cb(Qcur, "Qcur", il);
35 if (model.layers[il].bq) {
36 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
37 cb(Qcur, "Qcur", il);
38 }
39
40 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
41 cb(Kcur, "Kcur", il);
42 if (model.layers[il].bk) {
43 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
44 cb(Kcur, "Kcur", il);
45 }
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
54 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
55 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
56 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
57
58 if (model.layers[il].attn_q_norm) {
59 Qcur = build_norm(Qcur,
60 model.layers[il].attn_q_norm,
61 NULL,
62 LLM_NORM, il);
63 cb(Qcur, "Qcur", il);
64 }
65 if (model.layers[il].attn_k_norm) {
66 Kcur = build_norm(Kcur,
67 model.layers[il].attn_k_norm,
68 NULL,
69 LLM_NORM, il);
70 cb(Kcur, "Kcur", il);
71 }
72
73 Qcur = ggml_rope_ext(
74 ctx0, Qcur, inp_pos, nullptr,
75 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
76 ext_factor, attn_factor, beta_fast, beta_slow
77 );
78
79 Kcur = ggml_rope_ext(
80 ctx0, Kcur, inp_pos, nullptr,
81 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
82 ext_factor, attn_factor, beta_fast, beta_slow
83 );
84
85 cb(Qcur, "Qcur", il);
86 cb(Kcur, "Kcur", il);
87 cb(Vcur, "Vcur", il);
88
89 cur = build_attn(inp_attn,
90 model.layers[il].wo, NULL,
91 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
92 }
93 if (il == n_layer - 1 && inp_out_ids) {
94 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
95 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
96 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
97 }
98 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
99 cb(ffn_inp, "ffn_inp", il);
100
101 // feed-forward network
102 {
103 if (model.layers[il].ffn_norm) {
104 cur = build_norm(ffn_inp,
105 model.layers[il].ffn_norm,
106 model.layers[il].ffn_norm_b,
107 LLM_NORM, il);
108 cb(cur, "ffn_norm", il);
109 } else {
110 // parallel residual
111 cur = inpSA;
112 }
113 cur = build_ffn(cur,
114 model.layers[il].ffn_up, NULL, NULL,
115 model.layers[il].ffn_gate, NULL, NULL,
116 model.layers[il].ffn_down, NULL, NULL,
117 NULL,
118 LLM_FFN_SILU, LLM_FFN_PAR, il);
119 cb(cur, "ffn_out", il);
120 }
121 cur = ggml_add(ctx0, cur, ffn_inp);
122
123 cur = build_cvec(cur, il);
124 cb(cur, "l_out", il);
125
126 // input for next layer
127 inpL = cur;
128 }
129 cur = inpL;
130
131 cur = build_norm(cur,
132 model.output_norm,
133 model.output_norm_b,
134 LLM_NORM, -1);
135
136 cb(cur, "result_norm", -1);
137 res->t_embd = cur;
138
139 // lm_head
140 cur = build_lora_mm(model.output, cur);
141
142 cb(cur, "result_output", -1);
143 res->t_logits = cur;
144
145 ggml_build_forward_expand(gf, cur);
146}