1#include "models.h"
2
3llm_build_smollm3::llm_build_smollm3(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 = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
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 const bool use_rope = (il + 1) % hparams.n_no_rope_layer_step != 0;
27
28 // norm
29 cur = build_norm(inpL,
30 model.layers[il].attn_norm, NULL,
31 LLM_NORM_RMS, il);
32 cb(cur, "attn_norm", il);
33
34 // self-attention
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 if (use_rope) {
60 Qcur = ggml_rope_ext(
61 ctx0, Qcur, inp_pos, nullptr,
62 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
63 ext_factor, attn_factor, beta_fast, beta_slow
64 );
65
66 Kcur = ggml_rope_ext(
67 ctx0, Kcur, 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 cb(Qcur, "Qcur", il);
73 cb(Kcur, "Kcur", il);
74 cb(Vcur, "Vcur", il);
75
76 cur = build_attn(inp_attn,
77 model.layers[il].wo, model.layers[il].bo,
78 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
79 cb(cur, "attn_out", il);
80 }
81 if (il == n_layer - 1 && inp_out_ids) {
82 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
83 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
84 }
85 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
86 cb(ffn_inp, "ffn_inp", il);
87
88 // feed-forward network
89 {
90 cur = build_norm(ffn_inp,
91 model.layers[il].ffn_norm, NULL,
92 LLM_NORM_RMS, il);
93 cb(cur, "ffn_norm", il);
94
95 cur = build_ffn(cur,
96 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
97 model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
98 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
99 NULL,
100 LLM_FFN_SILU, LLM_FFN_PAR, il);
101 cb(cur, "ffn_out", il);
102 }
103 cur = ggml_add(ctx0, cur, ffn_inp);
104 cb(cur, "ffn_out", il);
105
106 cur = build_cvec(cur, il);
107 cb(cur, "l_out", il);
108
109 // input for next layer
110 inpL = cur;
111 }
112 cur = inpL;
113
114 cur = build_norm(cur,
115 model.output_norm, NULL,
116 LLM_NORM_RMS, -1);
117
118 cb(cur, "result_norm", -1);
119 res->t_embd = cur;
120
121 // lm_head
122 cur = build_lora_mm(model.output, cur);
123
124 cb(cur, "result_output", -1);
125 res->t_logits = cur;
126
127 ggml_build_forward_expand(gf, cur);
128}