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 ggml_tensor * cur;
9 ggml_tensor * inpL;
10
11 inpL = build_inp_embd(tok_embd: model.tok_embd);
12
13 // inp_pos - contains the positions
14 ggml_tensor * inp_pos = build_inp_pos();
15
16 auto * inp_attn = build_attn_inp_kv();
17
18 ggml_tensor * inp_out_ids = build_inp_out_ids();
19
20 // Only process up to last layer (skip final NextN layer)
21 // Final layer tensors are loaded but not processed in forward pass
22 const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
23 for (int il = 0; il < n_transformer_layers; ++il) {
24 ggml_tensor * inpSA = inpL;
25
26 // Pre-attention norm
27 cur = build_norm(cur: inpL, mw: model.layers[il].attn_norm, NULL, type: LLM_NORM_RMS, il);
28 cb(cur, name: "attn_norm", il);
29
30 // self-attention
31 {
32 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
33 if (model.layers[il].bq) {
34 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
35 }
36 cb(cur: Qcur, name: "Qcur", il);
37
38 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
39 if (model.layers[il].bk) {
40 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
41 }
42 cb(cur: Kcur, name: "Kcur", il);
43
44 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
45 if (model.layers[il].bv) {
46 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
47 }
48 cb(cur: Vcur, name: "Vcur", il);
49
50 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
51 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
52 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
53
54 // Apply Q/K norm if available (GLM-4.5 355B variant)
55 if (model.layers[il].attn_q_norm) {
56 Qcur = build_norm(cur: Qcur, mw: model.layers[il].attn_q_norm, NULL, type: LLM_NORM_RMS, il);
57 cb(cur: Qcur, name: "Qcur_normed", il);
58 }
59 if (model.layers[il].attn_k_norm) {
60 Kcur = build_norm(cur: Kcur, mw: model.layers[il].attn_k_norm, NULL, type: LLM_NORM_RMS, il);
61 cb(cur: Kcur, name: "Kcur_normed", il);
62 }
63 Qcur = ggml_rope_ext(
64 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
65 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
66 ext_factor, attn_factor, beta_fast, beta_slow
67 );
68
69 Kcur = ggml_rope_ext(
70 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
71 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
72 ext_factor, attn_factor, beta_fast, beta_slow
73 );
74
75 cb(cur: Qcur, name: "Qcur", il);
76 cb(cur: Kcur, name: "Kcur", il);
77 cb(cur: Vcur, name: "Vcur", il);
78
79 cur = build_attn(inp: inp_attn,
80 wo: model.layers[il].wo, NULL,
81 q_cur: Qcur, k_cur: Kcur, v_cur: Vcur, kq_b: nullptr, sinks: nullptr, v_mla: nullptr, kq_scale: 1.0f/sqrtf(x: float(n_embd_head)), il);
82 }
83 if (il == n_transformer_layers - 1 && inp_out_ids) {
84 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
85 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
86 }
87 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
88 cb(cur: ffn_inp, name: "ffn_inp", il);
89
90 // Post-attention norm
91 cur = build_norm(cur: ffn_inp, mw: model.layers[il].attn_post_norm, NULL, type: LLM_NORM_RMS, il);
92 cb(cur, name: "post_attn_norm", il);
93
94 // Check if this is a dense layer (n_layer_dense_lead=1, so layer 0 is dense)
95 if (static_cast<uint32_t>(il) < hparams.n_layer_dense_lead) {
96 // Dense FFN layer
97 cur = build_ffn(cur,
98 up: model.layers[il].ffn_up, NULL, NULL,
99 gate: model.layers[il].ffn_gate, NULL, NULL,
100 down: model.layers[il].ffn_down, NULL, NULL,
101 NULL,
102 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
103 cb(cur, name: "ffn_out", il);
104 } else {
105 // Process routed experts using existing MoE infrastructure
106 ggml_tensor * routed_out = build_moe_ffn(cur,
107 gate_inp: model.layers[il].ffn_gate_inp,
108 up_exps: model.layers[il].ffn_up_exps,
109 gate_exps: model.layers[il].ffn_gate_exps,
110 down_exps: model.layers[il].ffn_down_exps,
111 exp_probs_b: model.layers[il].ffn_exp_probs_b,
112 n_expert, n_expert_used,
113 type_op: LLM_FFN_SILU, norm_w: hparams.expert_weights_norm,
114 scale_w: true, w_scale: hparams.expert_weights_scale,
115 gating_op: (llama_expert_gating_func_type) hparams.expert_gating_func,
116 il);
117 cb(cur: routed_out, name: "ffn_moe_out", il);
118
119 // Process shared expert on original input
120 ggml_tensor * shared_out = build_ffn(cur,
121 up: model.layers[il].ffn_up_shexp, NULL, NULL,
122 gate: model.layers[il].ffn_gate_shexp, NULL, NULL,
123 down: model.layers[il].ffn_down_shexp, NULL, NULL,
124 NULL,
125 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
126 cb(cur: shared_out, name: "ffn_shexp_out", il);
127
128 // Final output: routed_output + shared_output
129 cur = ggml_add(ctx: ctx0, a: routed_out, b: shared_out);
130 cb(cur, name: "ffn_out", il);
131 }
132 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
133
134 cur = build_cvec(cur, il);
135 cb(cur, name: "l_out", il);
136
137 // input for next layer
138 inpL = cur;
139 }
140 cur = inpL;
141 cur = build_norm(cur, mw: model.output_norm, NULL, type: LLM_NORM_RMS, il: -1);
142
143 cb(cur, name: "result_norm", il: -1);
144 res->t_embd = cur;
145
146 // lm_head
147 cur = build_lora_mm(w: model.output, cur);
148
149 cb(cur, name: "result_output", il: -1);
150 res->t_logits = cur;
151
152 ggml_build_forward_expand(cgraph: gf, tensor: cur);
153}
154