1#include "models.h"
2
3llm_build_afmoe::llm_build_afmoe(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 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
6
7 ggml_tensor * cur;
8 ggml_tensor * inpL;
9
10 inpL = build_inp_embd(model.tok_embd);
11
12 // MuP scaling: embeddings * sqrt(hidden_size)
13 // mup_enabled = true, hidden_size = 1024, scale = 32.0
14 inpL = ggml_scale(ctx0, inpL, sqrtf(float(n_embd)));
15 cb(inpL, "inp_embd_scaled", -1);
16
17 // inp_pos - contains the positions
18 ggml_tensor * inp_pos = build_inp_pos();
19 auto * inp_attn = build_attn_inp_kv_iswa();
20 ggml_tensor * inp_out_ids = build_inp_out_ids();
21
22 const float kq_scale = 1.0f/sqrtf(float(n_embd_head));
23
24 for (int il = 0; il < n_layer; ++il) {
25 const float freq_base_l = model.get_rope_freq_base (cparams, il);
26 const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
27
28 ggml_tensor * inpSA = inpL;
29
30 // This overlaps with SWA layers in current models, so get_rope_freq_base/scale may be superfluous
31 const bool use_rope = hparams.n_no_rope_layer_step > 0 &&
32 (il + 1) % hparams.n_no_rope_layer_step != 0;
33
34 // dual attention normalization (pre)
35 cur = build_norm(inpL,
36 model.layers[il].attn_norm, NULL,
37 LLM_NORM_RMS, il);
38 cb(cur, "attn_norm", il);
39
40 // self-attention
41 {
42 ggml_tensor * attn_inp = cur; // save input for gate computation
43
44 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
45 cb(Qcur, "Qcur", il);
46
47 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
48 cb(Kcur, "Kcur", il);
49
50 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
51 cb(Vcur, "Vcur", il);
52
53 // compute gate from input
54 ggml_tensor * gate = build_lora_mm(model.layers[il].wqkv_gate, attn_inp);
55 cb(gate, "attn_gate_proj", il);
56
57 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
58 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
59
60 // Q/K normalization
61 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
62 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
63 cb(Qcur, "Qcur_normed", il);
64 cb(Kcur, "Kcur_normed", il);
65
66 if (use_rope) {
67 Qcur = ggml_rope_ext(
68 ctx0, Qcur, inp_pos, nullptr,
69 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
70 ext_factor, attn_factor, beta_fast, beta_slow);
71 cb(Qcur, "Qcur_rope", il);
72
73 Kcur = ggml_rope_ext(
74 ctx0, Kcur, inp_pos, nullptr,
75 n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
76 ext_factor, attn_factor, beta_fast, beta_slow);
77 cb(Kcur, "Kcur_rope", il);
78 }
79
80 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
81
82 cur = build_attn(inp_attn,
83 NULL, NULL, // wo will be applied after gating
84 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
85 cb(cur, "attn_out", il);
86
87 // attention gating: attn_out * sigmoid(gate) BEFORE o_proj
88 gate = ggml_sigmoid(ctx0, gate);
89 cb(gate, "attn_gate_sig", il);
90 cur = ggml_mul(ctx0, cur, gate);
91 cb(cur, "attn_gated", il);
92
93 // now apply output projection
94 cur = build_lora_mm(model.layers[il].wo, cur);
95 cb(cur, "attn_o_proj", il);
96 }
97
98 // dual attention normalization (post)
99 cur = build_norm(cur,
100 model.layers[il].attn_post_norm, NULL,
101 LLM_NORM_RMS, il);
102 cb(cur, "attn_post_norm", il);
103
104 if (il == n_layer - 1 && inp_out_ids) {
105 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
106 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
107 }
108
109 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
110 cb(ffn_inp, "ffn_inp", il);
111
112 // dual ffn normalization (pre)
113 cur = build_norm(ffn_inp,
114 model.layers[il].ffn_norm, NULL,
115 LLM_NORM_RMS, il);
116 cb(cur, "ffn_norm", il);
117
118 // MoE or dense FFN
119 if ((uint32_t)il >= hparams.n_layer_dense_lead) {
120 // MoE layer with sigmoid routing, normalization, and scaling
121 ggml_tensor * moe_out = build_moe_ffn(cur,
122 model.layers[il].ffn_gate_inp,
123 model.layers[il].ffn_up_exps,
124 model.layers[il].ffn_gate_exps,
125 model.layers[il].ffn_down_exps,
126 model.layers[il].ffn_exp_probs_b,
127 n_expert, n_expert_used,
128 LLM_FFN_SILU,
129 hparams.expert_weights_norm, // norm_w (route_norm=True)
130 hparams.expert_weights_scale, // scale_w
131 hparams.expert_weights_scale, // w_scale (route_scale=2.826)
132 (llama_expert_gating_func_type) hparams.expert_gating_func,
133 il);
134 cb(moe_out, "ffn_moe_out", il);
135
136 // shared expert
137 if (hparams.n_expert_shared > 0) {
138 ggml_tensor * ffn_shexp = build_ffn(cur,
139 model.layers[il].ffn_up_shexp, NULL, NULL,
140 model.layers[il].ffn_gate_shexp, NULL, NULL,
141 model.layers[il].ffn_down_shexp, NULL, NULL,
142 NULL,
143 LLM_FFN_SILU, LLM_FFN_PAR, il);
144 cb(ffn_shexp, "ffn_shexp", il);
145
146 cur = ggml_add(ctx0, moe_out, ffn_shexp);
147 cb(cur, "ffn_out", il);
148 } else {
149 cur = moe_out;
150 }
151 } else {
152 // dense layer
153 cur = build_ffn(cur,
154 model.layers[il].ffn_up, NULL, NULL,
155 model.layers[il].ffn_gate, NULL, NULL,
156 model.layers[il].ffn_down, NULL, NULL,
157 NULL,
158 LLM_FFN_SILU, LLM_FFN_PAR, il);
159 cb(cur, "ffn_out", il);
160 }
161
162 // dual ffn normalization (post)
163 cur = build_norm(cur,
164 model.layers[il].ffn_post_norm, NULL,
165 LLM_NORM_RMS, il);
166 cb(cur, "ffn_post_norm", il);
167
168 cur = ggml_add(ctx0, cur, ffn_inp);
169 cur = build_cvec(cur, il);
170 cb(cur, "l_out", il);
171
172 // input for next layer
173 inpL = cur;
174 }
175
176 cur = inpL;
177
178 cur = build_norm(cur,
179 model.output_norm, NULL,
180 LLM_NORM_RMS, -1);
181 cb(cur, "result_norm", -1);
182
183 res->t_embd = cur;
184
185 // lm_head
186 cur = build_lora_mm(model.output, cur);
187 cb(cur, "result_output", -1);
188 res->t_logits = cur;
189
190 ggml_build_forward_expand(gf, cur);
191}