1#include "models.h"
2
3
4
5llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
6 llm_graph_context(params) {
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(tok_embd: model.tok_embd);
16
17 // inp_pos - contains the positions
18 ggml_tensor * inp_pos = build_inp_pos();
19
20 auto * inp_attn = build_attn_inp_kv();
21
22 for (int il = 0; il < n_layer; ++il) {
23 ggml_tensor * inpSA = inpL;
24
25 // norm
26 {
27 cur = build_norm(cur: inpL, mw: model.layers[il].attn_norm, NULL, 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 if (model.layers[il].bq) {
35 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
36 cb(cur: Qcur, name: "Qcur", il);
37 }
38 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
39 cb(cur: Kcur, name: "Kcur", il);
40 if (model.layers[il].bk) {
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 cb(cur: Vcur, name: "Vcur", il);
46 if (model.layers[il].bv) {
47 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
48 cb(cur: Vcur, name: "Vcur", il);
49 }
50 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
51 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
52 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
53
54 Qcur = ggml_rope_ext(ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr, 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 Kcur = ggml_rope_ext(ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr, 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 cb(cur: Qcur, name: "Qcur", il);
61 cb(cur: Kcur, name: "Kcur", il);
62 cb(cur: Vcur, name: "Vcur", il);
63
64 cur = build_attn(inp: inp_attn,
65 wo: model.layers[il].wo, NULL,
66 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);
67 }
68 if (il == n_layer - 1) {
69 // skip computing output for unused tokens
70 ggml_tensor * inp_out_ids = build_inp_out_ids();
71 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
72 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
73 }
74 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
75 cb(cur: ffn_inp, name: "ffn_inp", il);
76
77 // feed-forward network
78 {
79 cur = build_norm(cur: ffn_inp, mw: model.layers[il].ffn_norm, NULL, 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, type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
87 cb(cur, name: "ffn_out", il);
88 }
89 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
90
91 cur = build_cvec(cur, il);
92 cb(cur, name: "l_out", il);
93
94 // input for next layer
95 inpL = cur;
96 }
97 cur = inpL;
98
99 cur = build_norm(cur, mw: model.output_norm, NULL, type: LLM_NORM_RMS, il: -1);
100
101 cb(cur, name: "result_norm", il: -1);
102 res->t_embd = cur;
103
104 // lm_head
105 cur = build_lora_mm(w: model.output, cur);
106
107 cb(cur, name: "result_output", il: -1);
108 res->t_logits = cur;
109
110 ggml_build_forward_expand(cgraph: gf, tensor: cur);
111}
112