1#include "models.h"
2
3llm_build_glm4_moe::llm_build_glm4_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
8 int sections[4];
9 std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
10
11 ggml_tensor * cur;
12 ggml_tensor * inpL;
13
14 inpL = build_inp_embd(model.tok_embd);
15
16 bool use_mrope = hparams.use_mrope();
17 if (ubatch.embd && !use_mrope) {
18 // unfortunately, we need to forcefully stop here, to avoid users complaining about wrong results
19 GGML_ABORT("This GGUF does not support multimodal. Please reconvert it.");
20 }
21
22 // inp_pos - contains the positions
23 ggml_tensor * inp_pos = build_inp_pos();
24
25 auto * inp_attn = build_attn_inp_kv();
26
27 ggml_tensor * inp_out_ids = build_inp_out_ids();
28
29 // Only process up to last layer (skip final NextN layer)
30 // Final layer tensors are loaded but not processed in forward pass
31 const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
32 for (int il = 0; il < n_transformer_layers; ++il) {
33 ggml_tensor * inpSA = inpL;
34
35 // Pre-attention norm
36 cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
37 cb(cur, "attn_norm", il);
38
39 // self-attention
40 {
41 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
42 if (model.layers[il].bq) {
43 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
44 }
45 cb(Qcur, "Qcur", il);
46
47 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
48 if (model.layers[il].bk) {
49 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
50 }
51 cb(Kcur, "Kcur", il);
52
53 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
54 if (model.layers[il].bv) {
55 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
56 }
57 cb(Vcur, "Vcur", il);
58
59 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
60 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
61 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
62
63 // Apply Q/K norm if available (GLM-4.5 355B variant)
64 if (model.layers[il].attn_q_norm) {
65 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, NULL, LLM_NORM_RMS, il);
66 cb(Qcur, "Qcur_normed", il);
67 }
68 if (model.layers[il].attn_k_norm) {
69 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
70 cb(Kcur, "Kcur_normed", il);
71 }
72
73 if (use_mrope) {
74 Qcur = ggml_rope_multi(ctx0, Qcur, inp_pos, nullptr,
75 n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
76 ext_factor, attn_factor, beta_fast, beta_slow);
77
78 Kcur = ggml_rope_multi(ctx0, Kcur, inp_pos, nullptr,
79 n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
80 ext_factor, attn_factor, beta_fast, beta_slow);
81 } else {
82 // Normal RoPE
83 Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot,
84 rope_type, n_ctx_orig, freq_base, freq_scale,
85 ext_factor, attn_factor, beta_fast, beta_slow);
86
87 Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot,
88 rope_type, n_ctx_orig, freq_base, freq_scale,
89 ext_factor, attn_factor, beta_fast, beta_slow);
90 }
91
92 cb(Qcur, "Qcur", il);
93 cb(Kcur, "Kcur", il);
94 cb(Vcur, "Vcur", il);
95
96 cur = build_attn(inp_attn,
97 model.layers[il].wo, NULL,
98 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
99 }
100 if (il == n_transformer_layers - 1 && inp_out_ids) {
101 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
102 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
103 }
104 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
105 cb(ffn_inp, "ffn_inp", il);
106
107 // Post-attention norm
108 cur = build_norm(ffn_inp, model.layers[il].attn_post_norm, NULL, LLM_NORM_RMS, il);
109 cb(cur, "post_attn_norm", il);
110
111 // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
112 if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
113 // Dense FFN layer
114 cur = build_ffn(cur,
115 model.layers[il].ffn_up, NULL, NULL,
116 model.layers[il].ffn_gate, NULL, NULL,
117 model.layers[il].ffn_down, NULL, NULL,
118 NULL,
119 LLM_FFN_SILU, LLM_FFN_PAR, il);
120 cb(cur, "ffn_out", il);
121 } else {
122 // Process routed experts using existing MoE infrastructure
123 ggml_tensor * routed_out = build_moe_ffn(cur,
124 model.layers[il].ffn_gate_inp,
125 model.layers[il].ffn_up_exps,
126 model.layers[il].ffn_gate_exps,
127 model.layers[il].ffn_down_exps,
128 model.layers[il].ffn_exp_probs_b,
129 n_expert, n_expert_used,
130 LLM_FFN_SILU, hparams.expert_weights_norm,
131 true, hparams.expert_weights_scale,
132 (llama_expert_gating_func_type) hparams.expert_gating_func,
133 il);
134 cb(routed_out, "ffn_moe_out", il);
135
136 // Process shared expert on original input
137 ggml_tensor * shared_out = build_ffn(cur,
138 model.layers[il].ffn_up_shexp, NULL, NULL,
139 model.layers[il].ffn_gate_shexp, NULL, NULL,
140 model.layers[il].ffn_down_shexp, NULL, NULL,
141 NULL,
142 LLM_FFN_SILU, LLM_FFN_PAR, il);
143 cb(shared_out, "ffn_shexp_out", il);
144
145 // Final output: routed_output + shared_output
146 cur = ggml_add(ctx0, routed_out, shared_out);
147 cb(cur, "ffn_out", il);
148 }
149 cur = ggml_add(ctx0, cur, ffn_inp);
150
151 cur = build_cvec(cur, il);
152 cb(cur, "l_out", il);
153
154 // input for next layer
155 inpL = cur;
156 }
157 cur = inpL;
158 cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
159
160 cb(cur, "result_norm", -1);
161 res->t_embd = cur;
162
163 // lm_head
164 cur = build_lora_mm(model.output, cur);
165
166 cb(cur, "result_output", -1);
167 res->t_logits = cur;
168
169 ggml_build_forward_expand(gf, cur);
170}