1#include "models.h"
2
3
4
5llm_build_exaone::llm_build_exaone(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 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, mw: model.layers[il].attn_norm, NULL, type: LLM_NORM_RMS, il);
29 cb(cur, name: "attn_norm", il);
30
31 // self-attention
32 {
33 // rope freq factors for llama3; may return nullptr for llama2 and other models
34 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
35
36 // compute Q and K and RoPE them
37 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
38 cb(cur: Qcur, name: "Qcur", il);
39 if (model.layers[il].bq) {
40 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
41 cb(cur: Qcur, name: "Qcur", il);
42 }
43 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
44 cb(cur: Kcur, name: "Kcur", il);
45 if (model.layers[il].bk) {
46 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
47 cb(cur: Kcur, name: "Kcur", il);
48 }
49 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
50 cb(cur: Vcur, name: "Vcur", il);
51 if (model.layers[il].bv) {
52 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
53 cb(cur: Vcur, name: "Vcur", il);
54 }
55 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
56 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
57 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
58
59 Qcur = ggml_rope_ext(ctx: ctx0, a: Qcur, b: inp_pos, c: rope_factors, n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
60 ext_factor, attn_factor, beta_fast, beta_slow);
61
62 Kcur = ggml_rope_ext(ctx: ctx0, a: Kcur, b: inp_pos, c: rope_factors, n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
63 ext_factor, attn_factor, beta_fast, beta_slow);
64
65 cb(cur: Qcur, name: "Qcur", il);
66 cb(cur: Kcur, name: "Kcur", il);
67 cb(cur: Vcur, name: "Vcur", il);
68
69 cur = build_attn(inp: inp_attn,
70 wo: model.layers[il].wo, wo_b: model.layers[il].bo,
71 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);
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 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
78 cb(cur: ffn_inp, name: "ffn_inp", il);
79
80 // feed-forward network
81 cur = build_norm(cur: ffn_inp, mw: model.layers[il].ffn_norm, NULL, type: LLM_NORM_RMS, il);
82 cb(cur, name: "ffn_norm", il);
83
84 cur = build_ffn(cur,
85 up: model.layers[il].ffn_up, NULL, NULL,
86 gate: model.layers[il].ffn_gate, NULL, NULL,
87 down: model.layers[il].ffn_down, NULL, NULL,
88 NULL, type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
89 cb(cur, name: "ffn_out", il);
90
91 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
92 cb(cur, name: "ffn_out", il);
93
94 cur = build_cvec(cur, il);
95 cb(cur, name: "l_out", il);
96
97 // input for next layer
98 inpL = cur;
99 }
100 cur = inpL;
101
102 cur = build_norm(cur, mw: model.output_norm, NULL, 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 cb(cur, name: "result_output", il: -1);
111 res->t_logits = cur;
112
113 ggml_build_forward_expand(cgraph: gf, tensor: cur);
114}
115