1#include "models.h"
2
3
4llm_build_gptneox::llm_build_gptneox(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 const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
7
8 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
9
10 ggml_tensor * cur;
11 ggml_tensor * inpL;
12
13 inpL = build_inp_embd(model.tok_embd);
14
15 // inp_pos - contains the positions
16 ggml_tensor * inp_pos = build_inp_pos();
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 cur = build_norm(inpL,
24 model.layers[il].attn_norm,
25 model.layers[il].attn_norm_b,
26 LLM_NORM, il);
27 cb(cur, "attn_norm", il);
28
29 // self-attention
30 {
31 cur = build_lora_mm(model.layers[il].wqkv, cur);
32 cb(cur, "wqkv", il);
33
34 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
35 cb(cur, "bqkv", il);
36
37 ggml_tensor * Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0*sizeof(float)*(n_embd));
38 ggml_tensor * Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd));
39 ggml_tensor * Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1*sizeof(float)*(n_embd + n_embd_gqa));
40
41 Qcur = ggml_rope_ext(
42 ctx0, Qcur, inp_pos, nullptr,
43 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
44 ext_factor, attn_factor, beta_fast, beta_slow
45 );
46
47 Kcur = ggml_rope_ext(
48 ctx0, Kcur, inp_pos, nullptr,
49 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
50 ext_factor, attn_factor, beta_fast, beta_slow
51 );
52
53 cb(Qcur, "Qcur", il);
54 cb(Kcur, "Kcur", il);
55 cb(Vcur, "Vcur", il);
56
57 cur = build_attn(inp_attn,
58 model.layers[il].wo, model.layers[il].bo,
59 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f/sqrtf(float(n_embd_head)), il);
60 }
61
62 if (il == n_layer - 1 && inp_out_ids) {
63 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
64 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
65 }
66
67 // ffn
68 if (hparams.use_par_res) {
69 // attention and ffn are computed in parallel
70 // x = x + attn(ln1(x)) + ffn(ln2(x))
71
72 ggml_tensor * attn_out = cur;
73
74 cur = build_norm(inpL,
75 model.layers[il].ffn_norm,
76 model.layers[il].ffn_norm_b,
77 LLM_NORM, il);
78 cb(cur, "ffn_norm", il);
79
80 cur = build_ffn(cur,
81 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
82 NULL, NULL, NULL,
83 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
84 NULL,
85 LLM_FFN_GELU, LLM_FFN_SEQ, il);
86 cb(cur, "ffn_out", il);
87
88 cur = ggml_add(ctx0, cur, inpL);
89 cb(cur, "ffn_out", il);
90
91 cur = ggml_add(ctx0, cur, attn_out);
92
93 cur = build_cvec(cur, il);
94 cb(cur, "l_out", il);
95
96 // input for next layer
97 inpL = cur;
98 } else {
99 // attention and ffn are computed sequentially
100 // x = x + attn(ln1(x))
101 // x = x + ffn(ln2(x))
102
103 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpL);
104 cb(ffn_inp, "ffn_inp", il);
105
106 cur = build_norm(ffn_inp,
107 model.layers[il].ffn_norm,
108 model.layers[il].ffn_norm_b,
109 LLM_NORM, il);
110 cb(cur, "ffn_norm", il);
111
112 cur = build_ffn(cur,
113 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
114 NULL, NULL, NULL,
115 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
116 NULL,
117 LLM_FFN_GELU, LLM_FFN_SEQ, il);
118 cb(cur, "ffn_out", il);
119
120 cur = ggml_add(ctx0, cur, ffn_inp);
121
122 cur = build_cvec(cur, il);
123 cb(cur, "l_out", il);
124
125 // input for next layer
126 inpL = cur;
127 }
128 }
129
130 cur = build_norm(inpL,
131 model.output_norm,
132 model.output_norm_b,
133 LLM_NORM, -1);
134
135 cb(cur, "result_norm", -1);
136 res->t_embd = cur;
137
138 cur = build_lora_mm(model.output, cur);
139
140 cb(cur, "result_output", -1);
141 res->t_logits = cur;
142
143 ggml_build_forward_expand(gf, cur);
144}