1#include "models.h"
2
3llm_build_qwen2::llm_build_qwen2(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 ggml_tensor * inp_out_ids = build_inp_out_ids();
20
21 for (int il = 0; il < n_layer; ++il) {
22 ggml_tensor * inpSA = inpL;
23
24 // norm
25 cur = build_norm(cur: inpL,
26 mw: model.layers[il].attn_norm, NULL,
27 type: LLM_NORM_RMS, il);
28 cb(cur, name: "attn_norm", il);
29
30 // self-attention
31 {
32 // compute Q and K and RoPE them
33 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
34 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
35 cb(cur: Qcur, name: "Qcur", il);
36
37 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
38 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
39 cb(cur: Kcur, name: "Kcur", il);
40
41 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
42 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
43 cb(cur: Vcur, name: "Vcur", il);
44
45 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
46 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
47 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
48
49 Qcur = ggml_rope_ext(
50 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
51 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
52 ext_factor, attn_factor, beta_fast, beta_slow
53 );
54
55 Kcur = ggml_rope_ext(
56 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
57 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
58 ext_factor, attn_factor, beta_fast, beta_slow
59 );
60
61 cb(cur: Qcur, name: "Qcur", il);
62 cb(cur: Kcur, name: "Kcur", il);
63 cb(cur: Vcur, name: "Vcur", il);
64
65 cur = build_attn(inp: inp_attn,
66 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
67 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);
68 }
69 if (il == n_layer - 1 && inp_out_ids) {
70 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
71 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
72 }
73 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
74 cb(cur: ffn_inp, name: "ffn_inp", il);
75
76 // feed-forward network
77 cur = build_norm(cur: ffn_inp,
78 mw: model.layers[il].ffn_norm, NULL,
79 type: LLM_NORM_RMS, il);
80 cb(cur, name: "ffn_norm", il);
81
82 cur = build_ffn(cur,
83 up: model.layers[il].ffn_up, NULL, NULL,
84 gate: model.layers[il].ffn_gate, NULL, NULL,
85 down: model.layers[il].ffn_down, NULL, NULL,
86 NULL,
87 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
88 cb(cur, name: "ffn_out", il);
89
90 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
91
92 cur = build_cvec(cur, il);
93 cb(cur, name: "l_out", il);
94
95 // input for next layer
96 inpL = cur;
97 }
98 cur = inpL;
99
100 cur = build_norm(cur,
101 mw: model.output_norm, NULL,
102 type: LLM_NORM_RMS, il: -1);
103
104 cb(cur, name: "result_norm", il: -1);
105 res->t_embd = cur;
106
107 // lm_head
108 cur = build_lora_mm(w: model.output, cur);
109
110 if (model.output_b != nullptr) {
111 cur = ggml_add(ctx: ctx0, a: cur, b: model.output_b);
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