1#include "models.h"
2
3
4llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
5 ggml_tensor * cur;
6 ggml_tensor * inpL;
7
8 inpL = build_inp_embd(model.tok_embd);
9
10 // inp_pos - contains the positions
11 ggml_tensor * inp_pos = build_inp_pos();
12
13 auto * inp_attn = build_attn_inp_kv();
14
15 ggml_tensor * inp_out_ids = build_inp_out_ids();
16
17 for (int il = 0; il < n_layer; ++il) {
18 ggml_tensor * inpSA = inpL;
19
20 // norm
21 cur = build_norm(inpL,
22 model.layers[il].attn_norm, NULL,
23 LLM_NORM_RMS, il);
24 cb(cur, "attn_norm", il);
25
26 // self-attention
27 {
28 // rope freq factors for llama3; may return nullptr for llama2 and other models
29 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
30
31 // compute Q and K and RoPE them
32 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
33 cb(Qcur, "Qcur", il);
34 if (model.layers[il].bq) {
35 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
36 cb(Qcur, "Qcur", il);
37 }
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
46 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
47 cb(Vcur, "Vcur", il);
48 if (model.layers[il].bv) {
49 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
50 cb(Vcur, "Vcur", il);
51 }
52
53 Qcur = ggml_reshape_3d(ctx0, Qcur, n_rot, n_head, n_tokens);
54 Kcur = ggml_reshape_3d(ctx0, Kcur, n_rot, n_head_kv, n_tokens);
55 Vcur = ggml_reshape_3d(ctx0, Vcur, n_rot, n_head_kv, n_tokens);
56
57 Qcur = ggml_rope_ext(
58 ctx0, Qcur, inp_pos, rope_factors,
59 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
60 ext_factor, attn_factor, beta_fast, beta_slow
61 );
62
63 Kcur = ggml_rope_ext(
64 ctx0, Kcur, inp_pos, rope_factors,
65 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
66 ext_factor, attn_factor, beta_fast, beta_slow
67 );
68
69 cb(Qcur, "Qcur", il);
70 cb(Kcur, "Kcur", il);
71 cb(Vcur, "Vcur", il);
72
73 cur = build_attn(inp_attn,
74 model.layers[il].wo, model.layers[il].bo,
75 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_rot)), il);
76 }
77
78 if (il == n_layer - 1 && inp_out_ids) {
79 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
80 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
81 }
82
83 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
84 cb(ffn_inp, "ffn_inp", il);
85
86 cur = build_norm(ffn_inp,
87 model.layers[il].ffn_norm, NULL,
88 LLM_NORM_RMS, il);
89 cb(cur, "ffn_norm", il);
90
91 ggml_tensor * moe_out =
92 build_moe_ffn(cur,
93 model.layers[il].ffn_gate_inp,
94 model.layers[il].ffn_up_exps,
95 model.layers[il].ffn_gate_exps,
96 model.layers[il].ffn_down_exps,
97 nullptr,
98 n_expert, n_expert_used,
99 LLM_FFN_SILU, hparams.expert_weights_norm,
100 false, hparams.expert_weights_scale,
101 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
102 il);
103 cb(moe_out, "ffn_moe_out", il);
104
105 // FFN shared expert
106 {
107 ggml_tensor * ffn_shexp = build_ffn(cur,
108 model.layers[il].ffn_up_shexp, NULL, NULL,
109 model.layers[il].ffn_gate_shexp, NULL, NULL,
110 model.layers[il].ffn_down_shexp, NULL, NULL,
111 NULL,
112 LLM_FFN_SILU, LLM_FFN_PAR, il);
113 cb(ffn_shexp, "ffn_shexp", il);
114
115 cur = ggml_add(ctx0, moe_out, ffn_shexp);
116 cb(cur, "ffn_out", il);
117 }
118
119 cur = ggml_add(ctx0, cur, ffn_inp);
120
121 cur = build_cvec(cur, il);
122 cb(cur, "l_out", il);
123
124 // input for next layer
125 inpL = cur;
126 }
127
128 cur = inpL;
129
130 cur = build_norm(cur,
131 model.output_norm, NULL,
132 LLM_NORM_RMS, -1);
133
134 cb(cur, "result_norm", -1);
135 res->t_embd = cur;
136
137 // lm_head
138 cur = build_lora_mm(model.output, cur);
139
140 cb(cur, "result_output", -1);
141 res->t_logits = cur;
142
143 ggml_build_forward_expand(gf, cur);
144}