1#include "models.h"
2
3llm_build_qwen3vlmoe::llm_build_qwen3vlmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
4 const size_t n_deepstack_layers = hparams.n_deepstack_layers;
5
6 const int64_t n_embd = hparams.n_embd;
7 const int64_t n_embd_head = hparams.n_embd_head_v;
8
9 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
10 GGML_ASSERT(n_embd_head == hparams.n_rot);
11
12 ggml_tensor * cur;
13 ggml_tensor * inpL;
14
15 inpL = build_inp_embd(model.tok_embd);
16
17 int sections[4];
18 std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
19
20 // inp_pos - contains the positions
21 ggml_tensor * inp_pos = build_inp_pos();
22
23 auto * inp_attn = build_attn_inp_kv();
24
25 ggml_tensor * inp_out_ids = build_inp_out_ids();
26
27 for (int il = 0; il < n_layer; ++il) {
28 ggml_tensor * inpSA = inpL;
29
30 // norm
31 cur = build_norm(inpL,
32 model.layers[il].attn_norm, NULL,
33 LLM_NORM_RMS, il);
34 cb(cur, "attn_norm", il);
35
36 // self_attention
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 cb(Qcur, "Qcur_normed", il);
54
55 Qcur = ggml_rope_multi(
56 ctx0, Qcur, inp_pos, nullptr,
57 n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
58 ext_factor, attn_factor, beta_fast, beta_slow
59 );
60
61 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, NULL, LLM_NORM_RMS, il);
62 cb(Kcur, "Kcur_normed", il);
63
64 Kcur = ggml_rope_multi(
65 ctx0, Kcur, inp_pos, nullptr,
66 n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
67 ext_factor, attn_factor, beta_fast, beta_slow
68 );
69
70 cb(Qcur, "Qcur", il);
71 cb(Kcur, "Kcur", il);
72 cb(Vcur, "Vcur", il);
73
74 cur = build_attn(inp_attn,
75 model.layers[il].wo, model.layers[il].bo,
76 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
77 }
78
79 if (il == n_layer - 1 && inp_out_ids) {
80 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
81 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
82 }
83
84 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
85 cb(ffn_inp, "ffn_inp", il);
86
87 // MoE branch
88 cur = build_norm(ffn_inp,
89 model.layers[il].ffn_norm, NULL,
90 LLM_NORM_RMS, il);
91 cb(cur, "ffn_norm", il);
92
93 ggml_tensor * moe_out =
94 build_moe_ffn(cur,
95 model.layers[il].ffn_gate_inp,
96 model.layers[il].ffn_up_exps,
97 model.layers[il].ffn_gate_exps,
98 model.layers[il].ffn_down_exps,
99 nullptr,
100 n_expert, n_expert_used,
101 LLM_FFN_SILU, true,
102 false, 0.0,
103 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
104 il);
105 cb(moe_out, "ffn_moe_out", il);
106 cur = moe_out;
107
108 cur = ggml_add(ctx0, cur, ffn_inp);
109
110 cur = build_cvec(cur, il);
111 cb(cur, "l_out", il);
112
113 if (il < (int) n_deepstack_layers) {
114 ggml_tensor * ds = ggml_view_2d(ctx0, res->t_inp_embd, n_embd, n_tokens, res->t_inp_embd->nb[1], (il + 1) * n_embd * sizeof(float));
115 cur = ggml_add(ctx0, cur, ds);
116 cb(cur, "deepstack_out", il);
117 }
118
119 // input for next layer
120 inpL = cur;
121 }
122
123 cur = inpL;
124
125 cur = build_norm(cur,
126 model.output_norm, NULL,
127 LLM_NORM_RMS, -1);
128
129 cb(cur, "result_norm", -1);
130 res->t_embd = cur;
131
132 // lm_head
133 cur = build_lora_mm(model.output, cur);
134
135 cb(cur, "result_output", -1);
136 res->t_logits = cur;
137
138 ggml_build_forward_expand(gf, cur);
139}
140