1#include "models.h"
2
3llm_build_qwen2vl::llm_build_qwen2vl(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 GGML_ASSERT(n_embd_head == hparams.n_rot);
8
9 ggml_tensor * cur;
10 ggml_tensor * inpL;
11
12 inpL = build_inp_embd(tok_embd: model.tok_embd);
13
14 // inp_pos - contains the positions
15 ggml_tensor * inp_pos = build_inp_pos();
16
17 auto * inp_attn = build_attn_inp_kv();
18
19 int sections[4];
20 std::copy(first: std::begin(cont: hparams.rope_sections), last: std::begin(cont: hparams.rope_sections) + 4, result: sections);
21
22 ggml_tensor * inp_out_ids = build_inp_out_ids();
23
24 for (int il = 0; il < n_layer; ++il) {
25 ggml_tensor * inpSA = inpL;
26
27 // norm
28 cur = build_norm(cur: inpL,
29 mw: model.layers[il].attn_norm, NULL,
30 type: LLM_NORM_RMS, il);
31 cb(cur, name: "attn_norm", il);
32
33 // self-attention
34 {
35 // compute Q and K and RoPE them
36 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
37 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
38 cb(cur: Qcur, name: "Qcur", il);
39
40 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
41 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
42 cb(cur: Kcur, name: "Kcur", il);
43
44 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
45 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
46 cb(cur: Vcur, name: "Vcur", il);
47
48 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
49 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
50 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
51
52 Qcur = ggml_rope_multi(
53 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
54 n_dims: n_rot, sections, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
55 ext_factor, attn_factor, beta_fast, beta_slow
56 );
57
58 Kcur = ggml_rope_multi(
59 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
60 n_dims: n_rot, sections, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
61 ext_factor, attn_factor, beta_fast, beta_slow
62 );
63
64 cb(cur: Qcur, name: "Qcur", il);
65 cb(cur: Kcur, name: "Kcur", il);
66 cb(cur: Vcur, name: "Vcur", il);
67
68 cur = build_attn(inp: inp_attn,
69 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
70 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);
71 }
72 if (il == n_layer - 1 && inp_out_ids) {
73 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
74 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
75 }
76 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
77 cb(cur: ffn_inp, name: "ffn_inp", il);
78
79 // feed-forward network
80 cur = build_norm(cur: ffn_inp,
81 mw: model.layers[il].ffn_norm, NULL,
82 type: LLM_NORM_RMS, il);
83 cb(cur, name: "ffn_norm", il);
84
85 cur = build_ffn(cur,
86 up: model.layers[il].ffn_up, NULL, NULL,
87 gate: model.layers[il].ffn_gate, NULL, NULL,
88 down: model.layers[il].ffn_down, NULL, NULL,
89 NULL,
90 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
91 cb(cur, name: "ffn_out", il);
92
93 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
94
95 cur = build_cvec(cur, il);
96 cb(cur, name: "l_out", il);
97
98 // input for next layer
99 inpL = cur;
100 }
101 cur = inpL;
102
103 cur = build_norm(cur,
104 mw: model.output_norm, NULL,
105 type: LLM_NORM_RMS, il: -1);
106
107 cb(cur, name: "result_norm", il: -1);
108 res->t_embd = cur;
109
110 // lm_head
111 cur = build_lora_mm(w: model.output, cur);
112
113 cb(cur, name: "result_output", il: -1);
114 res->t_logits = cur;
115
116 ggml_build_forward_expand(cgraph: gf, tensor: cur);
117}
118