1#include "models.h"
2
3llm_build_grok::llm_build_grok(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 ggml_tensor * inp_out_ids = build_inp_out_ids();
20
21 for (int il = 0; il < n_layer; ++il) {
22 ggml_tensor * inpSA = inpL;
23
24 // norm
25 cur = build_norm(inpL,
26 model.layers[il].attn_norm, NULL,
27 LLM_NORM_RMS, il);
28 cb(cur, "attn_norm", il);
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 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
40 cb(Kcur, "Kcur", il);
41 if (model.layers[il].bk) {
42 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
43 cb(Kcur, "Kcur", il);
44 }
45 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
46 cb(Vcur, "Vcur", il);
47 if (model.layers[il].bv) {
48 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
49 cb(Vcur, "Vcur", il);
50 }
51 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
52 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
53 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
54
55 Qcur = ggml_rope_ext(
56 ctx0, Qcur, inp_pos, nullptr,
57 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
58 ext_factor, attn_factor, beta_fast, beta_slow
59 );
60
61 Kcur = ggml_rope_ext(
62 ctx0, Kcur, inp_pos, nullptr,
63 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
64 ext_factor, attn_factor, beta_fast, beta_slow
65 );
66
67 cb(Qcur, "Qcur", il);
68 cb(Kcur, "Kcur", il);
69 cb(Vcur, "Vcur", il);
70
71 cur = build_attn(inp_attn,
72 model.layers[il].wo, model.layers[il].bo,
73 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
74 }
75 if (il == n_layer - 1 && inp_out_ids) {
76 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
77 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
78 }
79 cur = build_norm(cur,
80 model.layers[il].attn_out_norm, NULL,
81 LLM_NORM_RMS, il);
82 cb(cur, "attn_out_norm", il);
83
84 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
85 cb(ffn_inp, "ffn_inp", il);
86
87 // feed-forward network
88 cur = build_norm(ffn_inp,
89 model.layers[il].ffn_norm, NULL,
90 LLM_NORM_RMS, il);
91 cb(cur, "ffn_norm", il);
92
93 // MoE branch
94 ggml_tensor * moe_out = build_moe_ffn(cur,
95 model.layers[il].ffn_gate_inp,
96 model.layers[il].ffn_up_exps,
97 model.layers[il].ffn_gate_exps,
98 model.layers[il].ffn_down_exps,
99 nullptr,
100 n_expert, n_expert_used,
101 LLM_FFN_GELU, true,
102 false, 0.0,
103 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
104 il);
105 cb(moe_out, "ffn_moe_out", il);
106
107 if (model.layers[il].ffn_up) {
108 ggml_tensor * ffn_out = build_ffn(cur,
109 model.layers[il].ffn_up, NULL, NULL,
110 model.layers[il].ffn_gate, NULL, NULL,
111 model.layers[il].ffn_down, NULL, NULL,
112 NULL,
113 LLM_FFN_GELU, LLM_FFN_PAR, il);
114 cb(ffn_out, "ffn_out", il);
115
116 cur = ggml_scale(ctx0, ggml_add(ctx0, ffn_out, moe_out), std::sqrt(2) / 2);
117 cb(cur, "ffn_out", il);
118 } else {
119 cur = moe_out;
120 }
121 cur = build_norm(cur,
122 model.layers[il].ffn_post_norm, NULL,
123 LLM_NORM_RMS, il);
124 cb(cur, "ffn_post_norm", il);
125
126 cur = ggml_add(ctx0, cur, ffn_inp);
127 cb(cur, "ffn_out", il);
128
129 cur = build_cvec(cur, il);
130 cb(cur, "l_out", il);
131
132 // input for next layer
133 inpL = cur;
134 }
135 cur = inpL;
136
137 cur = build_norm(cur,
138 model.output_norm, NULL,
139 LLM_NORM_RMS, -1);
140
141 cb(cur, "result_norm", -1);
142 res->t_embd = cur;
143
144 // lm_head
145 cur = build_lora_mm(model.output, cur);
146
147 cur = ggml_scale(ctx0, cur, hparams.f_logit_scale);
148
149 // final logit soft-capping
150 if (hparams.f_final_logit_softcapping) {
151 cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_final_logit_softcapping);
152 cur = ggml_tanh(ctx0, cur);
153 cur = ggml_scale(ctx0, cur, hparams.f_final_logit_softcapping);
154 }
155 cb(cur, "result_output", -1);
156 res->t_logits = cur;
157
158 ggml_build_forward_expand(gf, cur);
159}