1#include "models.h"
2
3llm_build_hunyuan_moe::llm_build_hunyuan_moe(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(tok_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 = 1.0f / sqrtf(x: float(n_embd_head));
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 // norm
27 cur = build_norm(cur: inpL,
28 mw: model.layers[il].attn_norm, NULL,
29 type: LLM_NORM_RMS, il);
30 cb(cur, name: "attn_norm", il);
31
32 // self-attention
33 {
34 // rope freq factors for llama3; may return nullptr for llama2 and other models
35 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
36
37 // compute Q and K and RoPE them
38 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
39 cb(cur: Qcur, name: "Qcur", il);
40 if (model.layers[il].bq) {
41 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
42 cb(cur: Qcur, name: "Qcur", il);
43 }
44 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
45 cb(cur: Kcur, name: "Kcur", il);
46 if (model.layers[il].bk) {
47 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
48 cb(cur: Kcur, name: "Kcur", il);
49 }
50 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
51 cb(cur: Vcur, name: "Vcur", il);
52 if (model.layers[il].bv) {
53 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
54 cb(cur: Vcur, name: "Vcur", il);
55 }
56 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
57 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
58 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
59
60 Qcur = ggml_rope_ext(
61 ctx: ctx0, a: Qcur, b: inp_pos, c: rope_factors,
62 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
63 ext_factor, attn_factor, beta_fast, beta_slow
64 );
65
66 cb(cur: Qcur, name: "Qcur", il);
67 cb(cur: Kcur, name: "Kcur", il);
68 cb(cur: Vcur, name: "Vcur", il);
69
70 Kcur = ggml_rope_ext(
71 ctx: ctx0, a: Kcur, b: inp_pos, c: rope_factors,
72 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
73 ext_factor, attn_factor, beta_fast, beta_slow
74 );
75
76 Kcur = build_norm(cur: Kcur,
77 mw: model.layers[il].attn_k_norm, mb: nullptr,
78 type: LLM_NORM_RMS, il);
79 cb(cur: Kcur, name: "Kcur_norm", il);
80
81 Qcur = build_norm(cur: Qcur,
82 mw: model.layers[il].attn_q_norm, mb: nullptr,
83 type: LLM_NORM_RMS, il);
84 cb(cur: Qcur, name: "Qcur_norm", il);
85
86 cur = build_attn(inp: inp_attn,
87 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
88 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale, il);
89 cb(cur, name: "attn_out", il);
90 }
91 if (il == n_layer - 1 && inp_out_ids) {
92 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
93 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
94 }
95 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
96 cb(cur: ffn_inp, name: "ffn_inp", il);
97
98 cur = build_norm(cur: ffn_inp,
99 mw: model.layers[il].ffn_norm, NULL,
100 type: LLM_NORM_RMS, il);
101 cb(cur, name: "ffn_norm", il);
102
103 // feed-forward network (non-MoE)
104 ggml_tensor * cur_mlp = build_ffn(cur,
105 up: model.layers[il].ffn_up_shexp, NULL, NULL,
106 gate: model.layers[il].ffn_gate_shexp, NULL, NULL,
107 down: model.layers[il].ffn_down_shexp, NULL, NULL,
108 NULL,
109 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
110 cb(cur: cur_mlp, name: "ffn_mlp", il);
111
112 // MoE branch
113 ggml_tensor * cur_moe = build_moe_ffn(cur,
114 gate_inp: model.layers[il].ffn_gate_inp,
115 up_exps: model.layers[il].ffn_up_exps,
116 gate_exps: model.layers[il].ffn_gate_exps,
117 down_exps: model.layers[il].ffn_down_exps,
118 exp_probs_b: nullptr,
119 n_expert, n_expert_used,
120 type_op: LLM_FFN_SILU,
121 norm_w: true, // norm_topk_prob
122 scale_w: false,
123 w_scale: 0.0,
124 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
125 il);
126 cb(cur: cur_moe, name: "ffn_moe_out", il);
127
128 ggml_tensor * ffn_out = ggml_add(ctx: ctx0, a: cur_moe, b: cur_mlp);
129 cb(cur: ffn_out, name: "ffn_out", il);
130
131 cur = ggml_add(ctx: ctx0, a: ffn_out, b: ffn_inp);
132
133 cur = build_cvec(cur, il);
134 cb(cur, name: "l_out", il);
135
136 // input for next layer
137 inpL = cur;
138 }
139 cur = inpL;
140
141 cur = build_norm(cur,
142 mw: model.output_norm, NULL,
143 type: LLM_NORM_RMS, il: -1);
144
145 cb(cur, name: "result_norm", il: -1);
146 res->t_embd = cur;
147
148 // lm_head
149 cur = build_lora_mm(w: model.output, cur);
150 cb(cur, name: "result_output", il: -1);
151 res->t_logits = cur;
152
153 ggml_build_forward_expand(cgraph: gf, tensor: cur);
154}
155