1#include "models.h"
2
3llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
4 llm_graph_context(params) {
5 const bool is_mla = hparams.is_mla();
6
7 // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
8 const int64_t n_embd_head_k = hparams.n_embd_head_k_mla();
9 const int64_t n_embd_head_v = hparams.n_embd_head_v_mla();
10
11 const int64_t n_embd_head_qk_rope = hparams.n_rot;
12 const int64_t n_embd_head_qk_nope = n_embd_head_k - n_embd_head_qk_rope;
13
14 const uint32_t kv_lora_rank = hparams.n_lora_kv;
15
16 // We have to pre-scale kq_scale and attn_factor to make the YaRN RoPE work correctly.
17 // See https://github.com/ggml-org/llama.cpp/discussions/7416 for detailed explanation.
18 // And also: https://github.com/ggml-org/llama.cpp/pull/17945 [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
19
20 // first cancel the adjustment from llama_hparams::yarn_attn_factor_adjust to get the original attn_factor
21 GGML_ASSERT(ext_factor >= 0.0f);
22 const float attn_factor_org = attn_factor * (1.0f + 0.1f * logf(1.0f / freq_scale));
23
24 // use the original attn_factor to pre-scale the kq_scale
25 const float mscale = attn_factor_org * (1.0f + 0.1f * hparams.rope_yarn_log_mul * logf(1.0f / freq_scale));
26 const float kq_scale = 1.0f * mscale * mscale / sqrtf(float(n_embd_head_k));
27
28 ggml_tensor * cur;
29 ggml_tensor * inpL;
30
31 // {n_embd, n_tokens}
32 inpL = build_inp_embd(model.tok_embd);
33
34 // (optional) temperature tuning - used by mistral-large
35 ggml_tensor * inp_attn_scale = nullptr;
36 if (hparams.f_attn_temp_scale != 0.0f) {
37 inp_attn_scale = build_inp_attn_scale();
38 }
39
40 // inp_pos - contains the positions
41 ggml_tensor * inp_pos = build_inp_pos();
42
43 auto * inp_attn_kv = !is_mla ? build_attn_inp_kv() : nullptr;
44 auto * inp_attn_k = is_mla ? build_attn_inp_k() : nullptr;
45
46 ggml_tensor * inp_out_ids = build_inp_out_ids();
47
48 for (int il = 0; il < n_layer; ++il) {
49 ggml_tensor * inpSA = inpL;
50
51 // norm
52 cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
53 cb(cur, "attn_norm", il);
54
55 // self_attention
56 {
57 ggml_tensor * q = NULL;
58
59 const bool is_lite = model.layers[il].wq;
60
61 if (!is_lite) {
62 q = ggml_mul_mat(ctx0, model.layers[il].wq_a, cur);
63 cb(q, "q", il);
64
65 q = build_norm(q, model.layers[il].attn_q_a_norm, nullptr, LLM_NORM_RMS, il);
66 cb(q, "q", il);
67
68 q = ggml_mul_mat(ctx0, model.layers[il].wq_b, q);
69 cb(q, "q", il);
70 } else {
71 q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
72 cb(q, "q", il);
73 }
74 // split into {n_embd_head_qk_nope, n_head, n_tokens}
75 ggml_tensor * q_nope =
76 ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
77 ggml_row_size(q->type, n_embd_head_k) * n_head, 0);
78 cb(q_nope, "q_nope", il);
79
80 // and {n_embd_head_qk_rope, n_head, n_tokens}
81 ggml_tensor * q_pe = ggml_view_3d(
82 ctx0, q, n_embd_head_qk_rope, n_head, n_tokens, ggml_row_size(q->type, n_embd_head_k),
83 ggml_row_size(q->type, n_embd_head_k) * n_head, ggml_row_size(q->type, n_embd_head_qk_nope));
84 cb(q_pe, "q_pe", il);
85
86 ggml_tensor * kv_cmpr_pe = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
87 cb(kv_cmpr_pe, "kv_cmpr_pe", il);
88
89 // split into {kv_lora_rank, n_tokens}
90 ggml_tensor * kv_cmpr =
91 ggml_view_2d(ctx0, kv_cmpr_pe, kv_lora_rank, n_tokens,
92 ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope), 0);
93 cb(kv_cmpr, "kv_cmpr", il);
94
95 // and {n_embd_head_qk_rope, 1, n_tokens}
96 ggml_tensor * k_pe = ggml_view_3d(ctx0, kv_cmpr_pe, n_embd_head_qk_rope, 1, n_tokens,
97 ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
98 ggml_row_size(kv_cmpr_pe->type, kv_lora_rank + n_embd_head_qk_rope),
99 ggml_row_size(kv_cmpr_pe->type, kv_lora_rank));
100 cb(k_pe, "k_pe", il);
101
102 q_pe = ggml_rope_ext(ctx0, q_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
103 ext_factor, attn_factor, beta_fast, beta_slow);
104 cb(q_pe, "q_pe", il);
105
106 k_pe = ggml_rope_ext(ctx0, k_pe, inp_pos, nullptr, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
107 ext_factor, attn_factor, beta_fast, beta_slow);
108 cb(k_pe, "k_pe", il);
109
110 kv_cmpr = build_norm(kv_cmpr, model.layers[il].attn_kv_a_norm, nullptr, LLM_NORM_RMS, il);
111 cb(kv_cmpr, "kv_cmpr", il);
112
113 if (is_mla) {
114 // {n_embd_head_qk_nope, n_tokens, n_head}
115 q_nope = ggml_permute(ctx0, q_nope, 0, 2, 1, 3);
116 cb(q_nope, "q_nope_perm", il);
117
118 // {n_embd_head_qk_nope, kv_lora_rank, n_head} x {n_embd_head_qk_nope, n_tokens, n_head}
119 ggml_tensor * q_nope_absorbed = ggml_mul_mat(ctx0, model.layers[il].wk_b, q_nope);
120 cb(q_nope_absorbed, "q_nope_absorbed", il);
121
122 // {kv_lora_rank, n_head, n_tokens}
123 q_nope_absorbed = ggml_permute(ctx0, q_nope_absorbed, 0, 2, 1, 3);
124 cb(q_nope_absorbed, "q_nope_absorbed_perm", il);
125
126 // {n_embd_head_qk_rope + kv_lora_rank, n_head, n_tokens}
127 // note: rope must go first for in-place context shifting in build_rope_shift()
128 ggml_tensor * Qcur = ggml_concat(ctx0, q_nope_absorbed, q_pe, 0);
129 cb(Qcur, "Qcur", il);
130
131 kv_cmpr = ggml_reshape_3d(ctx0, kv_cmpr, kv_lora_rank, 1, n_tokens);
132 cb(kv_cmpr, "kv_cmpr_reshape", il);
133
134 // {n_embd_head_qk_rope + kv_lora_rank, 1, n_tokens}
135 ggml_tensor * Kcur = ggml_concat(ctx0, kv_cmpr, k_pe, 0);
136 cb(Kcur, "Kcur", il);
137
138 // {kv_lora_rank, 1, n_tokens}
139 ggml_tensor * Vcur = kv_cmpr;
140 cb(Vcur, "Vcur", il);
141
142 if (inp_attn_scale) {
143 // apply llama 4 temperature scaling
144 Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
145 cb(Qcur, "Qcur_attn_temp_scaled", il);
146 }
147
148 // note: MLA with the absorption optimzation converts into MQA (ie: GQA with 1 group)
149 cur = build_attn(inp_attn_k,
150 model.layers[il].wo, NULL,
151 Qcur, Kcur, Vcur, nullptr, nullptr, model.layers[il].wv_b, kq_scale, il);
152 } else {
153 ggml_tensor * kv = ggml_mul_mat(ctx0, model.layers[il].wkv_b, kv_cmpr);
154 cb(kv, "kv", il);
155
156 // split into {n_embd_head_qk_nope, n_head, n_tokens}
157 ggml_tensor * k_nope =
158 ggml_view_3d(ctx0, kv, n_embd_head_qk_nope, n_head, n_tokens,
159 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
160 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head, 0);
161 cb(k_nope, "k_nope_view", il);
162
163 // and {n_embd_head_v, n_head, n_tokens}
164 ggml_tensor * Vcur = ggml_view_3d(ctx0, kv, n_embd_head_v, n_head, n_tokens,
165 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v),
166 ggml_row_size(kv->type, n_embd_head_qk_nope + n_embd_head_v) * n_head,
167 ggml_row_size(kv->type, n_embd_head_qk_nope));
168 cb(Vcur, "Vcur_view", il);
169
170 Vcur = ggml_cont(ctx0, Vcur);
171 cb(Vcur, "Vcur_cont", il);
172
173 ggml_tensor * Qcur = ggml_concat(ctx0, q_nope, q_pe, 0);
174 cb(Qcur, "Qcur", il);
175
176 ggml_tensor * Kcur = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
177 cb(Kcur, "Kcur", il);
178
179 if (inp_attn_scale) {
180 // apply llama 4 temperature scaling
181 Qcur = ggml_mul(ctx0, Qcur, inp_attn_scale);
182 cb(Qcur, "Qcur_attn_temp_scaled", il);
183 }
184
185 // note: MLA without the absorption optimization converts into MHA (ie: GQA with full n_head groups)
186 cur = build_attn(inp_attn_kv,
187 model.layers[il].wo, NULL,
188 Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
189 }
190 }
191 if (il == n_layer - 1 && inp_out_ids) {
192 cur = ggml_get_rows(ctx0, cur, inp_out_ids);
193 inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
194 }
195 ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
196 cb(ffn_inp, "ffn_inp", il);
197
198 cur = build_norm(ffn_inp, model.layers[il].ffn_norm, NULL, LLM_NORM_RMS, il);
199 cb(cur, "ffn_norm", il);
200
201 if ((uint32_t) il < hparams.n_layer_dense_lead) {
202 cur = build_ffn(cur,
203 model.layers[il].ffn_up, NULL, NULL,
204 model.layers[il].ffn_gate, NULL, NULL,
205 model.layers[il].ffn_down, NULL, NULL,
206 NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
207 cb(cur, "ffn_out", il);
208 } else {
209 // MoE branch
210 ggml_tensor * moe_out = build_moe_ffn(cur,
211 model.layers[il].ffn_gate_inp,
212 model.layers[il].ffn_up_exps,
213 model.layers[il].ffn_gate_exps,
214 model.layers[il].ffn_down_exps,
215 model.layers[il].ffn_exp_probs_b,
216 n_expert, n_expert_used,
217 LLM_FFN_SILU, hparams.expert_weights_norm,
218 hparams.expert_weights_scale, hparams.expert_weights_scale,
219 (llama_expert_gating_func_type) hparams.expert_gating_func,
220 il);
221 cb(moe_out, "ffn_moe_out", il);
222
223 // FFN shared expert
224 {
225 ggml_tensor * ffn_shexp =
226 build_ffn(cur,
227 model.layers[il].ffn_up_shexp, NULL, NULL,
228 model.layers[il].ffn_gate_shexp, NULL, NULL,
229 model.layers[il].ffn_down_shexp, NULL, NULL,
230 NULL, LLM_FFN_SILU, LLM_FFN_PAR, il);
231 cb(ffn_shexp, "ffn_shexp", il);
232
233 cur = ggml_add(ctx0, moe_out, ffn_shexp);
234 cb(cur, "ffn_out", il);
235 }
236 }
237 cur = ggml_add(ctx0, cur, ffn_inp);
238
239 cur = build_cvec(cur, il);
240 cb(cur, "l_out", il);
241
242 // input for next layer
243 inpL = cur;
244 }
245 cur = inpL;
246
247 cur = build_norm(cur, model.output_norm, NULL, LLM_NORM_RMS, -1);
248
249 cb(cur, "result_norm", -1);
250 res->t_embd = cur;
251
252 // lm_head
253 cur = ggml_mul_mat(ctx0, model.output, cur);
254
255 cb(cur, "result_output", -1);
256 res->t_logits = cur;
257
258 ggml_build_forward_expand(gf, cur);
259}