1#include "models.h"
2
3
4llm_build_bitnet::llm_build_bitnet(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
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(tok_embd: model.tok_embd);
13
14 // inp_pos - contains the positions
15 ggml_tensor * inp_pos = build_inp_pos();
16
17 auto * inp_attn = build_attn_inp_kv();
18
19 ggml_tensor * inp_out_ids = build_inp_out_ids();
20
21 for (int il = 0; il < n_layer; ++il) {
22 ggml_tensor * inpSA = inpL;
23
24 cur = build_norm(cur: inpL,
25 mw: model.layers[il].attn_norm, NULL,
26 type: LLM_NORM_RMS, il);
27 cb(cur, name: "attn_norm", il);
28
29 // self-attention
30 {
31 // compute Q and K and RoPE them
32 ggml_tensor * Qcur = build_lora_mm(w: model.layers[il].wq, cur);
33 if (model.layers[il].wq_scale) {
34 Qcur = ggml_mul(ctx: ctx0, a: Qcur, b: model.layers[il].wq_scale);
35 }
36 cb(cur: Qcur, name: "Qcur", il);
37 if (model.layers[il].bq) {
38 Qcur = ggml_add(ctx: ctx0, a: Qcur, b: model.layers[il].bq);
39 cb(cur: Qcur, name: "Qcur", il);
40 }
41
42 // B1.K
43 ggml_tensor * Kcur = build_lora_mm(w: model.layers[il].wk, cur);
44 if (model.layers[il].wk_scale) {
45 Kcur = ggml_mul(ctx: ctx0, a: Kcur, b: model.layers[il].wk_scale);
46 }
47 cb(cur: Kcur, name: "Kcur", il);
48 if (model.layers[il].bk) {
49 Kcur = ggml_add(ctx: ctx0, a: Kcur, b: model.layers[il].bk);
50 cb(cur: Kcur, name: "Kcur", il);
51 }
52
53 // B1.V
54 ggml_tensor * Vcur = build_lora_mm(w: model.layers[il].wv, cur);
55 if (model.layers[il].wv_scale) {
56 Vcur = ggml_mul(ctx: ctx0, a: Vcur, b: model.layers[il].wv_scale);
57 }
58 cb(cur: Vcur, name: "Vcur", il);
59 if (model.layers[il].bv) {
60 Vcur = ggml_add(ctx: ctx0, a: Vcur, b: model.layers[il].bv);
61 cb(cur: Vcur, name: "Vcur", il);
62 }
63
64 Qcur = ggml_reshape_3d(ctx: ctx0, a: Qcur, ne0: n_embd_head, ne1: n_head, ne2: n_tokens);
65 Kcur = ggml_reshape_3d(ctx: ctx0, a: Kcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
66 Vcur = ggml_reshape_3d(ctx: ctx0, a: Vcur, ne0: n_embd_head, ne1: n_head_kv, ne2: n_tokens);
67
68 Qcur = ggml_rope_ext(
69 ctx: ctx0, a: Qcur, b: inp_pos, c: nullptr,
70 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
71 ext_factor, attn_factor, beta_fast, beta_slow
72 );
73
74 Kcur = ggml_rope_ext(
75 ctx: ctx0, a: Kcur, b: inp_pos, c: nullptr,
76 n_dims: n_rot, mode: rope_type, n_ctx_orig, freq_base, freq_scale,
77 ext_factor, attn_factor, beta_fast, beta_slow
78 );
79
80 cb(cur: Qcur, name: "Qcur", il);
81 cb(cur: Kcur, name: "Kcur", il);
82 cb(cur: Vcur, name: "Vcur", il);
83
84 cur = build_attn(inp: inp_attn,
85 NULL, NULL,
86 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);
87
88 cur = build_norm(cur,
89 mw: model.layers[il].attn_sub_norm, NULL,
90 type: LLM_NORM_RMS, il);
91 cb(cur, name: "attn_sub_norm", il);
92
93 cur = build_lora_mm(w: model.layers[il].wo, cur);
94 if (model.layers[il].wo_scale) {
95 cur = ggml_mul(ctx: ctx0, a: cur, b: model.layers[il].wo_scale);
96 }
97 if (model.layers[il].bo) {
98 cur = ggml_add(ctx: ctx0, a: cur, b: model.layers[il].bo);
99 }
100 cb(cur, name: "attn_out", il);
101 }
102
103 if (il == n_layer - 1 && inp_out_ids) {
104 cur = ggml_get_rows(ctx: ctx0, a: cur, b: inp_out_ids);
105 inpSA = ggml_get_rows(ctx: ctx0, a: inpSA, b: inp_out_ids);
106 }
107
108 ggml_tensor * ffn_inp = ggml_add(ctx: ctx0, a: cur, b: inpSA);
109 cb(cur: ffn_inp, name: "ffn_inp", il);
110
111 // feed-forward forward
112 cur = build_norm(cur: ffn_inp,
113 mw: model.layers[il].ffn_norm, NULL,
114 type: LLM_NORM_RMS, il);
115 cb(cur, name: "ffn_norm", il);
116
117 cur = build_ffn(cur,
118 up: model.layers[il].ffn_up, NULL, up_s: model.layers[il].ffn_up_scale,
119 gate: model.layers[il].ffn_gate, NULL, gate_s: model.layers[il].ffn_gate_scale,
120 NULL, NULL, NULL,
121 NULL,
122 type_op: LLM_FFN_SILU, type_gate: LLM_FFN_PAR, il);
123 cb(cur, name: "ffn_sub_out", il);
124
125 cur = build_norm(cur,
126 mw: model.layers[il].ffn_sub_norm, NULL,
127 type: LLM_NORM_RMS, il);
128 cb(cur, name: "ffn_sub_norm", il);
129
130 cur = build_lora_mm(w: model.layers[il].ffn_down, cur);
131 if (model.layers[il].ffn_down_scale) {
132 cur = ggml_mul(ctx: ctx0, a: cur, b: model.layers[il].ffn_down_scale);
133 }
134 cb(cur, name: "ffn_down", il);
135
136 cur = ggml_add(ctx: ctx0, a: cur, b: ffn_inp);
137 cb(cur, name: "l_out", il);
138
139 // input for next layer
140 inpL = cur;
141 }
142
143 cur = inpL;
144
145 cur = build_norm(cur,
146 mw: model.output_norm, NULL,
147 type: LLM_NORM_RMS, il: -1);
148
149 cb(cur, name: "result_norm", il: -1);
150 res->t_embd = cur;
151
152 // lm_head
153 // FIXME: do not use model.tok_embd directly, duplicate as model.output
154 cur = build_lora_mm(w: model.tok_embd, cur);
155
156 cb(cur, name: "result_output", il: -1);
157 res->t_logits = cur;
158
159 ggml_build_forward_expand(cgraph: gf, tensor: cur);
160}
161