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 const int64_t n_embd = hparams.n_embd;
6 const int64_t n_embd_head = hparams.n_embd_head_v;
7
8 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
9 GGML_ASSERT(n_embd_head == hparams.n_rot);
10
11 ggml_tensor * cur;
12 ggml_tensor * inpL;
13
14 inpL = build_inp_embd(tok_embd: model.tok_embd);
15
16 int sections[4];
17 std::copy(first: std::begin(cont: hparams.rope_sections), last: std::begin(cont: hparams.rope_sections) + 4, result: sections);
18
19 std::vector<ggml_tensor *> deepstack_features(n_deepstack_layers, nullptr);
20
21 if (ubatch.embd) {
22 // Image input: split main embd and deepstack embds
23 ggml_tensor * inpL_main = ggml_view_2d(ctx: ctx0, a: inpL, ne0: n_embd, ne1: n_tokens, nb1: inpL->nb[1], offset: 0);
24 for (size_t i = 0; i < n_deepstack_layers; i++) {
25 deepstack_features[i] = ggml_view_2d(ctx: ctx0, a: inpL, ne0: n_embd, ne1: n_tokens, nb1: inpL->nb[1], offset: (i + 1) * n_embd * sizeof(float));
26 }
27 inpL = inpL_main;
28 }
29
30 // inp_pos - contains the positions
31 ggml_tensor * inp_pos = build_inp_pos();
32
33 auto * inp_attn = build_attn_inp_kv();
34
35 ggml_tensor * inp_out_ids = build_inp_out_ids();
36
37 for (int il = 0; il < n_layer; ++il) {
38 ggml_tensor * inpSA = inpL;
39
40 // norm
41 cur = build_norm(cur: inpL,
42 mw: model.layers[il].attn_norm, NULL,
43 type: LLM_NORM_RMS, il);
44 cb(cur, name: "attn_norm", il);
45
46 // self_attention
47 {
48 // compute Q and K and RoPE them
49 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
50 cb(cur: Qcur, name: "Qcur", il);
51
52 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
53 cb(cur: Kcur, name: "Kcur", il);
54
55 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
56 cb(cur: Vcur, name: "Vcur", il);
57
58 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
59 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
60 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
61
62 Qcur = build_norm(cur: Qcur, mw: model.layers[il].attn_q_norm, NULL, type: LLM_NORM_RMS, il);
63 cb(cur: Qcur, name: "Qcur_normed", il);
64
65 Qcur = ggml_rope_multi(
66 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
67 n_dims: n_rot, sections, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
68 ext_factor, attn_factor, beta_fast, beta_slow
69 );
70
71 Kcur = build_norm(cur: Kcur, mw: model.layers[il].attn_k_norm, NULL, type: LLM_NORM_RMS, il);
72 cb(cur: Kcur, name: "Kcur_normed", il);
73
74 Kcur = ggml_rope_multi(
75 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
76 n_dims: n_rot, sections, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
77 ext_factor, attn_factor, beta_fast, beta_slow
78 );
79
80 cb(cur: Qcur, name: "Qcur", il);
81 cb(cur: Kcur, name: "Kcur", il);
82 cb(cur: Vcur, name: "Vcur", il);
83
84 cur = build_attn(inp: inp_attn,
85 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
86 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);
87 }
88
89 if (il == n_layer - 1 && inp_out_ids) {
90 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
91 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
92 }
93
94 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
95 cb(cur: ffn_inp, name: "ffn_inp", il);
96
97 // MoE branch
98 cur = build_norm(cur: ffn_inp,
99 mw: model.layers[il].ffn_norm, NULL,
100 type: LLM_NORM_RMS, il);
101 cb(cur, name: "ffn_norm", il);
102
103 ggml_tensor * moe_out =
104 build_moe_ffn(cur,
105 gate_inp: model.layers[il].ffn_gate_inp,
106 up_exps: model.layers[il].ffn_up_exps,
107 gate_exps: model.layers[il].ffn_gate_exps,
108 down_exps: model.layers[il].ffn_down_exps,
109 exp_probs_b: nullptr,
110 n_expert, n_expert_used,
111 type_op: LLM_FFN_SILU, norm_w: true,
112 scale_w: false, w_scale: 0.0,
113 gating_op: LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
114 il);
115 cb(cur: moe_out, name: "ffn_moe_out", il);
116 cur = moe_out;
117
118 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
119
120 cur = build_cvec(cur, il);
121 cb(cur, name: "l_out", il);
122
123 if (ubatch.embd && (size_t)il < n_deepstack_layers) {
124 cur = ggml_add(ctx: ctx0, a: cur, b: deepstack_features[il]);
125 cb(cur, name: "deepstack_out", il);
126 }
127
128 // input for next layer
129 inpL = cur;
130 }
131
132 cur = inpL;
133
134 cur = build_norm(cur,
135 mw: model.output_norm, NULL,
136 type: LLM_NORM_RMS, il: -1);
137
138 cb(cur, name: "result_norm", il: -1);
139 res->t_embd = cur;
140
141 // lm_head
142 cur = build_lora_mm(w: model.output, cur);
143
144 cb(cur, name: "result_output", il: -1);
145 res->t_logits = cur;
146
147 ggml_build_forward_expand(cgraph: gf, tensor: cur);
148}
149
150