1#include "models.h"
2
3
4llm_build_granite::llm_build_granite(
5 const llama_model & model,
6 const llm_graph_params & params)
7 : llm_graph_context(params) {
8
9 const int64_t n_embd_head = hparams.n_embd_head_v;
10
11 GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
12 GGML_ASSERT(n_embd_head == hparams.n_rot);
13
14 ggml_tensor * cur;
15 ggml_tensor * inpL;
16
17 inpL = build_inp_embd(model.tok_embd);
18
19 // inp_pos - built only if rope enabled
20 ggml_tensor * inp_pos = nullptr;
21 if (hparams.rope_finetuned) {
22 inp_pos = build_inp_pos();
23 }
24 auto * inp_attn = build_attn_inp_kv();
25
26 ggml_tensor * inp_out_ids = build_inp_out_ids();
27
28 for (int il = 0; il < n_layer; ++il) {
29 ggml_tensor * inpSA = inpL;
30
31 // norm
32 cur = build_norm(inpL,
33 model.layers[il].attn_norm, NULL,
34 LLM_NORM_RMS, il);
35 cb(cur, "attn_norm", il);
36
37 // self-attention
38 cur = build_attention_layer(
39 cur, inp_pos, inp_attn,
40 model, n_embd_head, il);
41
42 if (il == n_layer - 1 && inp_out_ids) {
43 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
44 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
45 }
46 // ffn
47 cur = build_layer_ffn(cur, inpSA, model, il);
48
49 // input for next layer
50 inpL = cur;
51 }
52 cur = inpL;
53
54 cur = build_norm(cur,
55 model.output_norm, NULL,
56 LLM_NORM_RMS, -1);
57
58 cb(cur, "result_norm", -1);
59 res->t_embd = cur;
60
61 // lm_head
62 cur = build_lora_mm(model.output, cur);
63
64 // For Granite architectures - scale logits
65 cur = ggml_scale(ctx0, cur, 1.0f / hparams.f_logit_scale);
66 cb(cur, "result_output", -1);
67 res->t_logits = cur;
68
69 ggml_build_forward_expand(gf, cur);
70}
71
72ggml_tensor * llm_build_granite::build_attention_layer(
73 ggml_tensor * cur,
74 ggml_tensor * inp_pos,
75 llm_graph_input_attn_kv * inp_attn,
76 const llama_model & model,
77 const int64_t n_embd_head,
78 const int il) {
79
80 // compute Q and K and (optionally) RoPE them
81 ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
82 cb(Qcur, "Qcur", il);
83 if (model.layers[il].bq) {
84 Qcur = ggml_add(ctx0, Qcur, model.layers[il].bq);
85 cb(Qcur, "Qcur", il);
86 }
87
88 ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
89 cb(Kcur, "Kcur", il);
90 if (model.layers[il].bk) {
91 Kcur = ggml_add(ctx0, Kcur, model.layers[il].bk);
92 cb(Kcur, "Kcur", il);
93 }
94
95 ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
96 cb(Vcur, "Vcur", il);
97 if (model.layers[il].bv) {
98 Vcur = ggml_add(ctx0, Vcur, model.layers[il].bv);
99 cb(Vcur, "Vcur", il);
100 }
101
102 Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, hparams.n_head(il), n_tokens);
103 Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
104 Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, hparams.n_head_kv(il), n_tokens);
105
106 const bool use_rope = hparams.rope_finetuned;
107 if (use_rope) {
108 ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
109 Qcur = ggml_rope_ext(
110 ctx0, Qcur, inp_pos, rope_factors,
111 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
112 ext_factor, attn_factor, beta_fast, beta_slow
113 );
114
115 Kcur = ggml_rope_ext(
116 ctx0, Kcur, inp_pos, rope_factors,
117 n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
118 ext_factor, attn_factor, beta_fast, beta_slow
119 );
120 }
121
122 cb(Qcur, "Qcur", il);
123 cb(Kcur, "Kcur", il);
124 cb(Vcur, "Vcur", il);
125
126 const float kq_scale = hparams.f_attention_scale == 0.0f ? 1.0f/sqrtf(float(n_embd_head)) : hparams.f_attention_scale;
127 cur = build_attn(inp_attn,
128 model.layers[il].wo, model.layers[il].bo,
129 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
130 cb(cur, "attn_out", il);
131 return cur;
132}
133
134ggml_tensor * llm_build_granite::build_layer_ffn(
135 ggml_tensor * cur,
136 ggml_tensor * inpSA,
137 const llama_model & model,
138 const int il) {
139
140 // For Granite architectures - scale residual
141 if (hparams.f_residual_scale) {
142 cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
143 }
144 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
145 cb(ffn_inp, "ffn_inp", il);
146
147 // feed-forward network (non-MoE)
148 if (model.layers[il].ffn_gate_inp == nullptr) {
149
150 cur = build_norm(ffn_inp,
151 model.layers[il].ffn_norm, NULL,
152 LLM_NORM_RMS, il);
153 cb(cur, "ffn_norm", il);
154
155 cur = build_ffn(cur,
156 model.layers[il].ffn_up, model.layers[il].ffn_up_b, NULL,
157 model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
158 model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
159 NULL,
160 LLM_FFN_SILU, LLM_FFN_PAR, il);
161 cb(cur, "ffn_out", il);
162
163 } else {
164 // MoE branch
165 cur = build_norm(ffn_inp,
166 model.layers[il].ffn_norm, NULL,
167 LLM_NORM_RMS, il);
168 cb(cur, "ffn_norm", il);
169
170 ggml_tensor * moe_out = build_moe_ffn(cur,
171 model.layers[il].ffn_gate_inp,
172 model.layers[il].ffn_up_exps,
173 model.layers[il].ffn_gate_exps,
174 model.layers[il].ffn_down_exps,
175 nullptr,
176 n_expert, n_expert_used,
177 LLM_FFN_SILU, true,
178 false, 0.0,
179 LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
180 il);
181 cb(moe_out, "ffn_moe_out", il);
182
183 // For Granite MoE Shared
184 if (hparams.n_ff_shexp > 0) {
185 ggml_tensor * ffn_shexp = build_ffn(cur,
186 model.layers[il].ffn_up_shexp, NULL, NULL,
187 model.layers[il].ffn_gate_shexp, NULL, NULL,
188 model.layers[il].ffn_down_shexp, NULL, NULL,
189 NULL,
190 LLM_FFN_SILU, LLM_FFN_PAR, il);
191 cb(ffn_shexp, "ffn_shexp", il);
192
193 cur = ggml_add(ctx0, moe_out, ffn_shexp);
194 cb(cur, "ffn_out", il);
195 } else {
196 cur = moe_out;
197 }
198 }
199
200 // For Granite architectures - scale residual
201 if (hparams.f_residual_scale) {
202 cur = ggml_scale(ctx0, cur, hparams.f_residual_scale);
203 }
204 cur = ggml_add(ctx0, cur, ffn_inp);
205 cb(cur, "ffn_out", il);
206
207 cur = build_cvec(cur, il);
208 cb(cur, "l_out", il);
209
210 return cur;
211}