1#include "models.h"
2
3
4
5llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
6 const int64_t n_embd_head = hparams.n_embd_head_v;
7 const int64_t n_embd_gqa = hparams.n_embd_v_gqa();
8
9 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
10
11 ggml_tensor * cur;
12 ggml_tensor * inpL;
13 ggml_tensor * inp_pos = nullptr;
14
15 if (model.arch != LLM_ARCH_JINA_BERT_V2) {
16 inp_pos = build_inp_pos();
17 }
18
19 // construct input embeddings (token, type, position)
20 inpL = build_inp_embd(model.tok_embd);
21
22 // token types are hardcoded to zero ("Sentence A")
23 if (model.type_embd) {
24 ggml_tensor * type_row0 = ggml_view_1d(ctx0, model.type_embd, n_embd, 0);
25 inpL = ggml_add(ctx0, inpL, type_row0);
26 }
27 if (model.arch == LLM_ARCH_BERT) {
28 inpL = ggml_add(ctx0, ggml_get_rows(ctx0, model.pos_embd, inp_pos), inpL);
29 }
30 cb(inpL, "inp_embd", -1);
31
32 // embed layer norm
33 inpL = build_norm(inpL, model.tok_norm, model.tok_norm_b, LLM_NORM, -1);
34 cb(inpL, "inp_norm", -1);
35
36 auto * inp_attn = build_attn_inp_no_cache();
37
38 ggml_tensor * inp_out_ids = build_inp_out_ids();
39
40 for (int il = 0; il < n_layer; ++il) {
41 ggml_tensor * cur = inpL;
42
43 {
44 ggml_tensor * Qcur;
45 ggml_tensor * Kcur;
46 ggml_tensor * Vcur;
47
48 // self-attention
49 if (model.layers[il].wqkv) {
50 cur = build_lora_mm(model.layers[il].wqkv, cur);
51 cb(cur, "wqkv", il);
52
53 if (model.layers[il].bqkv) {
54 cur = ggml_add(ctx0, cur, model.layers[il].bqkv);
55 cb(cur, "bqkv", il);
56 }
57
58 Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head * sizeof(float), cur->nb[1],
59 0 * sizeof(float) * (n_embd));
60 Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
61 cur->nb[1], 1 * sizeof(float) * (n_embd));
62 Vcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head * sizeof(float),
63 cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa));
64 } else {
65 Qcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wq, cur), model.layers[il].bq);
66 Kcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wk, cur), model.layers[il].bk);
67 Vcur = ggml_add(ctx0, build_lora_mm(model.layers[il].wv, cur), model.layers[il].bv);
68
69 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
70 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
71 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
72 }
73
74 if (model.layers[il].attn_q_norm) {
75 Qcur = ggml_reshape_2d(ctx0, Qcur, n_embd_head * n_head, n_tokens);
76
77 Qcur = build_norm(Qcur, model.layers[il].attn_q_norm, model.layers[il].attn_q_norm_b, LLM_NORM, il);
78
79 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens);
80 }
81
82 if (model.layers[il].attn_k_norm) {
83 Kcur = ggml_reshape_2d(ctx0, Kcur, n_embd_head * n_head_kv, n_tokens);
84
85 Kcur = build_norm(Kcur, model.layers[il].attn_k_norm, model.layers[il].attn_k_norm_b, LLM_NORM, il);
86
87 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
88 }
89
90 // RoPE
91 if (model.arch == LLM_ARCH_NOMIC_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
92 model.arch == LLM_ARCH_JINA_BERT_V3) {
93 Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
94 ext_factor, attn_factor, beta_fast, beta_slow);
95
96 Kcur = ggml_rope_ext(ctx0, Kcur, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
97 ext_factor, attn_factor, beta_fast, beta_slow);
98 }
99
100 cb(Qcur, "Qcur", il);
101 cb(Kcur, "Kcur", il);
102 cb(Vcur, "Vcur", il);
103
104 cur = build_attn(inp_attn,
105 model.layers[il].wo, model.layers[il].bo,
106 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, 1.0f / sqrtf(float(n_embd_head)), il);
107 cb(cur, "kqv_out", il);
108 }
109
110 if (il == n_layer - 1 && inp_out_ids) {
111 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
112 inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
113 }
114
115 // re-add the layer input
116 cur = ggml_add(ctx0, cur, inpL);
117
118 // attention layer norm
119 cur = build_norm(cur, model.layers[il].attn_out_norm, model.layers[il].attn_out_norm_b, LLM_NORM, il);
120
121 if (model.layers[il].attn_norm_2 != nullptr) {
122 cur = ggml_add(ctx0, cur, inpL); // re-add the layer input
123 cur = build_norm(cur, model.layers[il].attn_norm_2, model.layers[il].attn_norm_2_b, LLM_NORM, il);
124 }
125
126 ggml_tensor * ffn_inp = cur;
127 cb(ffn_inp, "ffn_inp", il);
128
129 // feed-forward network
130 if (hparams.moe_every_n_layers > 0 && il % hparams.moe_every_n_layers == 1) {
131 // MoE branch
132 cur = build_moe_ffn(cur, model.layers[il].ffn_gate_inp, model.layers[il].ffn_up_exps, nullptr,
133 model.layers[il].ffn_down_exps, nullptr, hparams.n_expert, hparams.n_expert_used,
134 LLM_FFN_GELU, false, false, 0.0f, LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX, il);
135 cb(cur, "ffn_moe_out", il);
136 } else if (model.arch == LLM_ARCH_BERT || model.arch == LLM_ARCH_NOMIC_BERT_MOE ||
137 model.arch == LLM_ARCH_JINA_BERT_V3) {
138 cur = build_ffn(cur,
139 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
140 NULL, NULL, NULL,
141 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
142 LLM_FFN_GELU, LLM_FFN_SEQ, il);
143 cb(cur, "ffn_out", il);
144 } else if (model.arch == LLM_ARCH_JINA_BERT_V2) {
145 const bool up_contains_gate = !model.layers[il].ffn_gate && model.layers[il].ffn_up->ne[1] != hparams.n_ff();
146 auto type_op = up_contains_gate ? LLM_FFN_GEGLU : LLM_FFN_GELU;
147 cur = build_ffn(cur,
148 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
149 model.layers[il].ffn_gate, NULL, NULL,
150 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL, NULL,
151 type_op, LLM_FFN_PAR, il);
152 cb(cur, "ffn_out", il);
153 } else {
154 cur = build_ffn(cur,
155 model.layers[il].ffn_up, NULL, NULL,
156 model.layers[il].ffn_gate, NULL, NULL,
157 model.layers[il].ffn_down, NULL, NULL,
158 NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
159 cb(cur, "ffn_out", il);
160 }
161
162 // attentions bypass the intermediate layer
163 cur = ggml_add(ctx0, cur, ffn_inp);
164
165 // output layer norm
166 cur = build_norm(cur, model.layers[il].layer_out_norm, model.layers[il].layer_out_norm_b, LLM_NORM, il);
167
168 // input for next layer
169 inpL = cur;
170 }
171
172 cur = inpL;
173
174 cb(cur, "result_embd", -1);
175 res->t_embd = cur;
176
177 ggml_build_forward_expand(gf, cur);
178}