1#include "models.h"
2
3
4llm_build_exaone_moe::llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params) :
5 llm_graph_context(params) {
6 const int64_t n_embd_head = hparams.n_embd_head_k;
7
8 GGML_ASSERT(n_embd_head == hparams.n_embd_head_v);
9 GGML_ASSERT(n_embd_head == hparams.n_rot);
10
11 ggml_tensor * cur;
12 ggml_tensor * inpL;
13
14 inpL = build_inp_embd(model.tok_embd);
15
16 // inp_pos - contains the positions
17 ggml_tensor * inp_pos = build_inp_pos();
18
19 auto * inp_attn_iswa = build_attn_inp_kv_iswa();
20
21 ggml_tensor * inp_out_ids = build_inp_out_ids();
22
23 const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
24 for (int il = 0; il < n_transformer_layers; ++il) {
25 ggml_tensor * inpSA = inpL;
26
27 // use RoPE for SWA layers
28 const bool is_local_layer = hparams.is_swa(il);
29
30 // norm
31 cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
32 cb(cur, "attn_norm", il);
33
34 // self-attention
35 {
36 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
37
38 // compute Q and K and RoPE them
39 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
40 cb(Qcur, "Qcur", il);
41
42 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
43 cb(Kcur, "Kcur", il);
44
45 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
46 cb(Vcur, "Vcur", il);
47
48 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
49 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
50 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
51
52 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
53 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
54 cb(Qcur, "Qcur_normed", il);
55 cb(Kcur, "Kcur_normed", il);
56
57 if (is_local_layer) {
58 Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
59 freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
60
61 Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base,
62 freq_scale, ext_factor, attn_factor, beta_fast, beta_slow);
63 }
64 cb(Qcur, "Qcur", il);
65 cb(Kcur, "Kcur", il);
66 cb(Vcur, "Vcur", il);
67
68 cur = build_attn(inp_attn_iswa,
69 model.layers[il].wo, NULL,
70 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
71 cb(cur, "attn_out", il);
72 }
73 if (il == n_transformer_layers - 1 && inp_out_ids) {
74 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
75 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
76 }
77 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
78 cb(ffn_inp, "ffn_inp", il);
79
80 // norm
81 cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
82 cb(cur, "ffn_norm", il);
83
84 // feed-forward network
85 if (model.layers[il].ffn_gate_inp == nullptr) {
86 // dense branch
87 cur = build_ffn(cur,
88 model.layers[il].ffn_up, NULL, NULL,
89 model.layers[il].ffn_gate, NULL, NULL,
90 model.layers[il].ffn_down, NULL, NULL, NULL,
91 LLM_FFN_SILU, LLM_FFN_PAR, il);
92 cb(cur, "ffn_out", il);
93 } else {
94 // MoE branch
95 ggml_tensor * moe_out = build_moe_ffn(cur,
96 model.layers[il].ffn_gate_inp,
97 model.layers[il].ffn_up_exps,
98 model.layers[il].ffn_gate_exps,
99 model.layers[il].ffn_down_exps,
100 model.layers[il].ffn_exp_probs_b,
101 n_expert, n_expert_used,
102 LLM_FFN_SILU, hparams.expert_weights_norm,
103 true, hparams.expert_weights_scale,
104 (llama_expert_gating_func_type) hparams.expert_gating_func,
105 il);
106 cb(moe_out, "ffn_moe_out", il);
107
108 // FFN shared expert
109 {
110 ggml_tensor * ffn_shexp =
111 build_ffn(cur,
112 model.layers[il].ffn_up_shexp, NULL, NULL,
113 model.layers[il].ffn_gate_shexp, NULL, NULL,
114 model.layers[il].ffn_down_shexp, NULL, NULL,
115 NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
116 cb(ffn_shexp, "ffn_shexp", il);
117
118 cur = ggml_add(ctx0, moe_out, ffn_shexp);
119 cb(cur, "ffn_out", il);
120 }
121 }
122
123 cur = ggml_add(ctx0, cur, ffn_inp);
124
125 cur = build_cvec(cur, il);
126 cb(cur, "l_out", il);
127
128 // input for next layer
129 inpL = cur;
130 }
131 cur = inpL;
132
133 // final norm
134 cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
135
136 cb(cur, "result_norm", -1);
137 res->t_embd = cur;
138
139 // lm_head
140 cur = build_lora_mm(model.output, cur);
141
142 cb(cur, "result_output", -1);
143 res->t_logits = cur;
144
145 ggml_build_forward_expand(gf, cur);
146}