1#include "models.h"
2
3llm_build_mistral3::llm_build_mistral3(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 // (optional) temperature tuning
18 ggml_tensor * inp_attn_scale = nullptr;
19 if (hparams.f_attn_temp_scale != 0.0f) {
20 inp_attn_scale = build_inp_attn_scale();
21 }
22
23 auto * inp_attn = build_attn_inp_kv();
24
25 const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
26
27 ggml_tensor * inp_out_ids = build_inp_out_ids();
28
29 for (int il = 0; il < n_layer; ++il) {
30 ggml_tensor * inpSA = inpL;
31
32 // norm
33 cur = build_norm(inpL,
34 model.layers[il].attn_norm, NULL,
35 LLM_NORM_RMS, il);
36 cb(cur, "attn_norm", il);
37
38 // self-attention
39 {
40 // rope freq factors for llama3; may return nullptr for llama2 and other models
41 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
42
43 // compute Q and K and RoPE them
44 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
45 cb(Qcur, "Qcur", il);
46 if (model.layers[il].bq) {
47 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
48 cb(Qcur, "Qcur", il);
49 }
50 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
51 cb(Kcur, "Kcur", il);
52 if (model.layers[il].bk) {
53 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
54 cb(Kcur, "Kcur", il);
55 }
56 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
57 cb(Vcur, "Vcur", il);
58 if (model.layers[il].bv) {
59 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
60 cb(Vcur, "Vcur", il);
61 }
62 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
63 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
64 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
65
66 Qcur = ggml_rope_ext(
67 ctx0, Qcur, inp_pos, rope_factors,
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, rope_factors,
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 if (inp_attn_scale) {
83 // apply llama 4 temperature scaling
84 Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
85 cb(Qcur, "Qcur_attn_temp_scaled", il);
86 }
87
88 cur = build_attn(inp_attn,
89 model.layers[il].wo, model.layers[il].bo,
90 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
91 cb(cur, "attn_out", il);
92 }
93 if (il == n_layer - 1 && inp_out_ids) {
94 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
95 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
96 }
97 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
98 cb(ffn_inp, "ffn_inp", il);
99
100 // feed-forward network (non-MoE)
101 if (model.layers[il].ffn_gate_inp == nullptr) {
102
103 cur = build_norm(ffn_inp,
104 model.layers[il].ffn_norm, NULL,
105 LLM_NORM_RMS, il);
106 cb(cur, "ffn_norm", il);
107
108 cur = build_ffn(cur,
109 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
110 model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
111 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
112 NULL,
113 LLM_FFN_SILU, LLM_FFN_PAR, il);
114 cb(cur, "ffn_out", il);
115 } else {
116 // MoE branch
117 cur = build_norm(ffn_inp,
118 model.layers[il].ffn_norm, NULL,
119 LLM_NORM_RMS, il);
120 cb(cur, "ffn_norm", il);
121
122 cur = build_moe_ffn(cur,
123 model.layers[il].ffn_gate_inp,
124 model.layers[il].ffn_up_exps,
125 model.layers[il].ffn_gate_exps,
126 model.layers[il].ffn_down_exps,
127 nullptr,
128 n_expert, n_expert_used,
129 LLM_FFN_SILU, true,
130 false, 0.0,
131 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
132 il);
133 cb(cur, "ffn_moe_out", il);
134 }
135 cur = ggml_add(ctx0, cur, ffn_inp);
136 cb(cur, "ffn_out", il);
137
138 cur = build_cvec(cur, il);
139 cb(cur, "l_out", il);
140
141 // input for next layer
142 inpL = cur;
143 }
144 cur = inpL;
145
146 cur = build_norm(cur,
147 model.output_norm, NULL,
148 LLM_NORM_RMS, -1);
149
150 cb(cur, "result_norm", -1);
151 res->t_embd = cur;
152
153 // lm_head
154 cur = build_lora_mm(model.output, cur);
155
156 cb(cur, "result_output", -1);
157 res->t_logits = cur;
158
159 ggml_build_forward_expand(gf, cur);
160}