1#include "models.h"
2
3
4llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
5 const int64_t n_embd_head = hparams.n_embd_head_v;
6
7 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
8 GGML_ASSERT(n_embd_head == hparams.n_rot);
9
10 ggml_tensor * cur;
11 ggml_tensor * inpL;
12
13 inpL = build_inp_embd(tok_embd: model.tok_embd);
14
15 // inp_pos - contains the positions
16 ggml_tensor * inp_pos = model.type == LLM_TYPE_7B ? build_inp_pos() : nullptr;
17
18 auto * inp_attn = build_attn_inp_kv();
19
20 ggml_tensor * inp_out_ids = build_inp_out_ids();
21
22 for (int il = 0; il < n_layer; ++il) {
23 ggml_tensor * inpSA = inpL;
24
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 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
33 cb(cur: Qcur, name: "Qcur", il);
34
35 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
36 cb(cur: Kcur, name: "Kcur", il);
37
38 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
39 cb(cur: Vcur, name: "Vcur", il);
40
41 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
42 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
43 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
44
45 switch (model.type) {
46 case LLM_TYPE_7B:
47 Qcur = ggml_rope_ext(
48 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
49 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
50 ext_factor, attn_factor, beta_fast, beta_slow
51 );
52 Kcur = ggml_rope_ext(
53 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
54 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
55 ext_factor, attn_factor, beta_fast, beta_slow
56 );
57 break;
58 case LLM_TYPE_13B:
59 break;
60 default:
61 GGML_ABORT("fatal error");
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, NULL,
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
73 if (il == n_layer - 1 && inp_out_ids) {
74 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
75 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
76 }
77
78 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
79 cb(cur: ffn_inp, name: "ffn_inp", il);
80
81 // feed-forward network
82 {
83 cur = build_norm(cur: ffn_inp,
84 mw: model.layers[il].ffn_norm, NULL,
85 type: LLM_NORM_RMS, il);
86 cb(cur, name: "ffn_norm", il);
87
88 cur = build_ffn(cur,
89 up: model.layers[il].ffn_up, NULL, NULL,
90 gate: model.layers[il].ffn_gate, NULL, NULL,
91 down: model.layers[il].ffn_down, NULL, NULL,
92 NULL,
93 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
94 cb(cur, name: "ffn_out", il);
95 }
96
97 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
98
99 cur = build_cvec(cur, il);
100 cb(cur, name: "l_out", il);
101
102 // input for next layer
103 inpL = cur;
104 }
105
106 cur = inpL;
107
108 cur = build_norm(cur,
109 mw: model.output_norm, NULL,
110 type: LLM_NORM_RMS, il: -1);
111
112 cb(cur, name: "result_norm", il: -1);
113 res->t_embd = cur;
114
115 // lm_head
116 cur = build_lora_mm(w: model.output, cur);
117
118 cb(cur, name: "result_output", il: -1);
119 res->t_logits = cur;
120
121 ggml_build_forward_expand(cgraph: gf, tensor: cur);
122}
123