1#include "models.h"
2
3llm_build_t5_dec::llm_build_t5_dec(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 //const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
6
7 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
8
9 ggml_tensor * cur;
10 ggml_tensor * inpL;
11
12 inpL = build_inp_embd(model.tok_embd);
13
14 ggml_tensor * embd_enc = build_inp_cross_embd();
15 ggml_tensor * pos_bucket_dec = build_inp_pos_bucket_dec();
16
17 const int64_t n_outputs_enc = embd_enc->ne[1];
18
19 auto * inp_attn_self = build_attn_inp_kv();
20 auto * inp_attn_cross = build_attn_inp_cross();
21
22 ggml_tensor * inp_out_ids = build_inp_out_ids();
23
24 const int64_t dec_n_layer = hparams.dec_n_layer;
25
26 for (int il = 0; il < dec_n_layer; ++il) {
27 ggml_tensor * inpSA = inpL;
28
29 // norm
30 cur = build_norm(inpL,
31 model.layers[il].attn_norm, NULL,
32 LLM_NORM_RMS, il);
33 cb(cur, "attn_norm", il);
34
35 // self-attention
36 {
37 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
38 cb(Qcur, "Qcur", il);
39
40 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
41 cb(Kcur, "Kcur", il);
42
43 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
44 cb(Vcur, "Vcur", il);
45
46 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
47 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
48 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
49
50 ggml_tensor * attn_rel_b = model.layers[il].attn_rel_b ? model.layers[il].attn_rel_b : model.layers[0].attn_rel_b;
51 ggml_tensor * kq_b = build_pos_bias(pos_bucket_dec, attn_rel_b);
52
53 cur = build_attn(inp_attn_self,
54 model.layers[il].wo, model.layers[il].bo,
55 Qcur, Kcur, Vcur, kq_b, nullptr, nullptr, 1.0f, il);
56 cb(cur, "kqv_out", il);
57 }
58 cur = ggml_add(ctx0, cur, inpSA);
59 cb(cur, "cross_inp", il);
60
61 ggml_tensor * inpCA = cur;
62
63 // norm
64 cur = build_norm(cur,
65 model.layers[il].attn_norm_cross, NULL,
66 LLM_NORM_RMS, il);
67 cb(cur, "attn_norm_cross", il);
68
69 // cross-attention
70 {
71 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq_cross, cur);
72 cb(Qcur, "Qcur", il);
73
74 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk_cross, embd_enc);
75 cb(Kcur, "Kcur", il);
76
77 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv_cross, embd_enc);
78 cb(Vcur, "Vcur", il);
79
80 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
81 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_outputs_enc);
82 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_outputs_enc);
83
84 cur = build_attn(inp_attn_cross,
85 model.layers[il].wo_cross, nullptr,
86 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f, il);
87 cb(cur, "kqv_out", il);
88
89 //ggml_tensor * q = ggml_permute(ctx0, Qcur, 0, 2, 1, 3);
90 //ggml_tensor * k = ggml_cont(ctx0, ggml_permute(ctx0, Kcur, 0, 2, 1, 3));
91
92 //ggml_tensor * kq = ggml_mul_mat(ctx0, k, q);
93 //cb(kq, "kq", il);
94
95 //kq = ggml_soft_max_ext(ctx0, kq, KQ_mask_cross, 1.0f, hparams.f_max_alibi_bias);
96 //cb(kq, "kq_soft_max_ext", il);
97
98 //ggml_tensor * v = ggml_cont(ctx0, ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd_gqa, n_outputs_enc)));
99 //cb(v, "v", il);
100
101 //ggml_tensor * kqv = ggml_mul_mat(ctx0, ggml_reshape_3d(ctx0, v, n_outputs_enc, n_embd_head, n_head_kv), kq);
102 //cb(kqv, "kqv", il);
103
104 //ggml_tensor * kqv_merged = ggml_permute(ctx0, kqv, 0, 2, 1, 3);
105 //cb(kqv_merged, "kqv_merged", il);
106
107 //cur = ggml_cont_2d(ctx0, kqv_merged, n_embd_gqa, n_tokens);
108 //cb(cur, "kqv_merged_cont", il);
109
110 //ggml_build_forward_expand(gf, cur);
111
112 //cur = build_lora_mm(model.layers[il].wo_cross, cur);
113 //cb(cur, "kqv_out", il);
114 }
115 if (il == dec_n_layer - 1 && inp_out_ids) {
116 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
117 inpCA = ggml_get_rows(ctx0, inpCA, inp_out_ids);
118 }
119 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpCA);
120 cb(ffn_inp, "ffn_inp", il);
121
122 // feed-forward network
123 {
124 cur = build_norm(ffn_inp,
125 model.layers[il].ffn_norm, NULL,
126 LLM_NORM_RMS, il);
127 cb(cur, "ffn_norm", il);
128
129 // T5 uses relu, flan-T5 uses gelu-gated
130 cur = build_ffn(cur,
131 model.layers[il].ffn_up, NULL, NULL,
132 model.layers[il].ffn_gate, NULL, NULL,
133 model.layers[il].ffn_down, NULL, NULL,
134 NULL,
135 model.layers[il].ffn_gate ? LLM_FFN_GELU : LLM_FFN_RELU,
136 model.layers[il].ffn_gate ? LLM_FFN_PAR : LLM_FFN_SEQ,
137 il);
138 cb(cur, "ffn_out", il);
139 }
140 cur = ggml_add(ctx0, cur, ffn_inp);
141 cb(cur, "ffn_out", il);
142
143 cur = build_cvec(cur, il);
144 cb(cur, "l_out", il);
145
146 // input for next layer
147 inpL = cur;
148 }
149 cur = inpL;
150 cb(cur, "result_embd", -1);
151
152 cur = build_norm(cur,
153 model.output_norm, NULL,
154 LLM_NORM_RMS, -1);
155
156 cb(cur, "result_norm", -1);
157 res->t_embd = cur;
158
159 // lm_head
160 cur = build_lora_mm(model.output, cur);
161
162 cb(cur, "result_output", -1);
163 res->t_logits = cur;
164
165 ggml_build_forward_expand(gf, cur);
166}