| 1 | #include "llama-graph.h" |
| 2 | |
| 3 | #include "llama-impl.h" |
| 4 | #include "llama-batch.h" |
| 5 | #include "llama-cparams.h" |
| 6 | |
| 7 | #include "llama-kv-cache.h" |
| 8 | #include "llama-kv-cache-iswa.h" |
| 9 | #include "llama-memory-hybrid.h" |
| 10 | #include "llama-memory-recurrent.h" |
| 11 | |
| 12 | #include <cassert> |
| 13 | #include <cmath> |
| 14 | #include <cstring> |
| 15 | |
| 16 | void llm_graph_input_embd::set_input(const llama_ubatch * ubatch) { |
| 17 | if (ubatch->token) { |
| 18 | const int64_t n_tokens = ubatch->n_tokens; |
| 19 | |
| 20 | ggml_backend_tensor_set(tensor: tokens, data: ubatch->token, offset: 0, size: n_tokens*ggml_element_size(tensor: tokens)); |
| 21 | } |
| 22 | |
| 23 | if (ubatch->embd) { |
| 24 | const int64_t n_embd = embd->ne[0]; |
| 25 | const int64_t n_tokens = ubatch->n_tokens; |
| 26 | |
| 27 | ggml_backend_tensor_set(tensor: embd, data: ubatch->embd, offset: 0, size: n_tokens*n_embd*ggml_element_size(tensor: embd)); |
| 28 | } |
| 29 | } |
| 30 | |
| 31 | bool llm_graph_input_embd::can_reuse(const llm_graph_params & params) { |
| 32 | bool res = true; |
| 33 | |
| 34 | res &= (!tokens && !params.ubatch.token) || (tokens && tokens->ne[0] == params.ubatch.n_tokens); |
| 35 | res &= (!embd && !params.ubatch.embd) || (embd && embd->ne[0] == params.ubatch.n_tokens); |
| 36 | |
| 37 | return res; |
| 38 | } |
| 39 | |
| 40 | void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) { |
| 41 | if (ubatch->pos && pos) { |
| 42 | const int64_t n_tokens = ubatch->n_tokens; |
| 43 | |
| 44 | if (ubatch->token && n_pos_per_embd == 4) { |
| 45 | // in case we're using M-RoPE with text tokens, convert the 1D positions to 4D |
| 46 | // the 3 first dims are the same, and 4th dim is all 0 |
| 47 | std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd); |
| 48 | // copy the first dimension |
| 49 | for (int i = 0; i < n_tokens; ++i) { |
| 50 | pos_data[ i] = ubatch->pos[i]; |
| 51 | pos_data[ n_tokens + i] = ubatch->pos[i]; |
| 52 | pos_data[2 * n_tokens + i] = ubatch->pos[i]; |
| 53 | pos_data[3 * n_tokens + i] = 0; // 4th dim is 0 |
| 54 | } |
| 55 | ggml_backend_tensor_set(tensor: pos, data: pos_data.data(), offset: 0, size: pos_data.size()*ggml_element_size(tensor: pos)); |
| 56 | } else { |
| 57 | ggml_backend_tensor_set(tensor: pos, data: ubatch->pos, offset: 0, size: n_tokens*n_pos_per_embd*ggml_element_size(tensor: pos)); |
| 58 | } |
| 59 | } |
| 60 | } |
| 61 | |
| 62 | bool llm_graph_input_pos::can_reuse(const llm_graph_params & params) { |
| 63 | bool res = true; |
| 64 | |
| 65 | res &= pos->ne[0] == params.ubatch.n_tokens; |
| 66 | |
| 67 | return res; |
| 68 | } |
| 69 | |
| 70 | void llm_graph_input_attn_temp::set_input(const llama_ubatch * ubatch) { |
| 71 | if (ubatch->pos && attn_scale) { |
| 72 | const int64_t n_tokens = ubatch->n_tokens; |
| 73 | |
| 74 | std::vector<float> attn_scale_data(n_tokens, 0.0f); |
| 75 | for (int i = 0; i < n_tokens; ++i) { |
| 76 | const float pos = ubatch->pos[i]; |
| 77 | attn_scale_data[i] = std::log( |
| 78 | x: std::floor(x: (pos + 1.0f) / n_attn_temp_floor_scale) + 1.0 |
| 79 | ) * f_attn_temp_scale + 1.0; |
| 80 | } |
| 81 | |
| 82 | ggml_backend_tensor_set(tensor: attn_scale, data: attn_scale_data.data(), offset: 0, size: n_tokens*ggml_element_size(tensor: attn_scale)); |
| 83 | } |
| 84 | } |
| 85 | |
| 86 | void llm_graph_input_pos_bucket::set_input(const llama_ubatch * ubatch) { |
| 87 | if (pos_bucket) { |
| 88 | const int64_t n_tokens = ubatch->n_tokens; |
| 89 | |
| 90 | GGML_ASSERT(ggml_backend_buffer_is_host(pos_bucket->buffer)); |
| 91 | GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing |
| 92 | |
| 93 | int32_t * data = (int32_t *) pos_bucket->data; |
| 94 | |
| 95 | for (int h = 0; h < 1; ++h) { |
| 96 | for (int j = 0; j < n_tokens; ++j) { |
| 97 | for (int i = 0; i < n_tokens; ++i) { |
| 98 | data[h*(n_tokens*n_tokens) + j*n_tokens + i] = llama_relative_position_bucket(x: ubatch->pos[i], y: ubatch->pos[j], n_buckets: hparams.n_rel_attn_bkts, bidirectional: true); |
| 99 | } |
| 100 | } |
| 101 | } |
| 102 | } |
| 103 | } |
| 104 | |
| 105 | void llm_graph_input_pos_bucket_kv::set_input(const llama_ubatch * ubatch) { |
| 106 | if (pos_bucket) { |
| 107 | mctx->set_input_pos_bucket(dst: pos_bucket, ubatch); |
| 108 | } |
| 109 | } |
| 110 | |
| 111 | void llm_graph_input_out_ids::set_input(const llama_ubatch * ubatch) { |
| 112 | GGML_ASSERT(out_ids); |
| 113 | |
| 114 | const int64_t n_tokens = ubatch->n_tokens; |
| 115 | |
| 116 | GGML_ASSERT(ggml_backend_buffer_is_host(out_ids->buffer)); |
| 117 | int32_t * data = (int32_t *) out_ids->data; |
| 118 | |
| 119 | if (n_outputs == n_tokens) { |
| 120 | for (int i = 0; i < n_tokens; ++i) { |
| 121 | data[i] = i; |
| 122 | } |
| 123 | |
| 124 | return; |
| 125 | } |
| 126 | |
| 127 | GGML_ASSERT(ubatch->output); |
| 128 | |
| 129 | int n_outputs = 0; |
| 130 | |
| 131 | for (int i = 0; i < n_tokens; ++i) { |
| 132 | if (ubatch->output[i]) { |
| 133 | data[n_outputs++] = i; |
| 134 | } |
| 135 | } |
| 136 | } |
| 137 | |
| 138 | bool llm_graph_input_out_ids::can_reuse(const llm_graph_params & params) { |
| 139 | bool res = true; |
| 140 | |
| 141 | res &= n_outputs == params.n_outputs; |
| 142 | |
| 143 | return res; |
| 144 | } |
| 145 | |
| 146 | void llm_graph_input_mean::set_input(const llama_ubatch * ubatch) { |
| 147 | if (cparams.embeddings && cparams.pooling_type == LLAMA_POOLING_TYPE_MEAN) { |
| 148 | const int64_t n_tokens = ubatch->n_tokens; |
| 149 | const int64_t n_seq_tokens = ubatch->n_seq_tokens; |
| 150 | const int64_t n_seqs_unq = ubatch->n_seqs_unq; |
| 151 | |
| 152 | GGML_ASSERT(mean); |
| 153 | GGML_ASSERT(ggml_backend_buffer_is_host(mean->buffer)); |
| 154 | |
| 155 | float * data = (float *) mean->data; |
| 156 | memset(s: mean->data, c: 0, n: n_tokens*n_seqs_unq*ggml_element_size(tensor: mean)); |
| 157 | |
| 158 | std::vector<uint64_t> sums(n_seqs_unq, 0); |
| 159 | for (int i = 0; i < n_tokens; i += n_seq_tokens) { |
| 160 | for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { |
| 161 | const llama_seq_id seq_id = ubatch->seq_id[i][s]; |
| 162 | const int32_t seq_idx = ubatch->seq_idx[seq_id]; |
| 163 | |
| 164 | sums[seq_idx] += ubatch->n_seq_tokens; |
| 165 | } |
| 166 | } |
| 167 | |
| 168 | std::vector<float> div(n_seqs_unq, 0.0f); |
| 169 | for (int s = 0; s < n_seqs_unq; ++s) { |
| 170 | const uint64_t sum = sums[s]; |
| 171 | if (sum > 0) { |
| 172 | div[s] = 1.0f/float(sum); |
| 173 | } |
| 174 | } |
| 175 | |
| 176 | for (int i = 0; i < n_tokens; i += n_seq_tokens) { |
| 177 | for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { |
| 178 | const llama_seq_id seq_id = ubatch->seq_id[i][s]; |
| 179 | const int32_t seq_idx = ubatch->seq_idx[seq_id]; |
| 180 | |
| 181 | for (int j = 0; j < n_seq_tokens; ++j) { |
| 182 | data[seq_idx*n_tokens + i + j] = div[seq_idx]; |
| 183 | } |
| 184 | } |
| 185 | } |
| 186 | } |
| 187 | } |
| 188 | |
| 189 | void llm_graph_input_cls::set_input(const llama_ubatch * ubatch) { |
| 190 | const int64_t n_tokens = ubatch->n_tokens; |
| 191 | const int64_t n_seqs_unq = ubatch->n_seqs_unq; |
| 192 | |
| 193 | if (cparams.embeddings && ( |
| 194 | cparams.pooling_type == LLAMA_POOLING_TYPE_CLS || |
| 195 | cparams.pooling_type == LLAMA_POOLING_TYPE_RANK || |
| 196 | cparams.pooling_type == LLAMA_POOLING_TYPE_LAST |
| 197 | )) { |
| 198 | GGML_ASSERT(cls); |
| 199 | GGML_ASSERT(ggml_backend_buffer_is_host(cls->buffer)); |
| 200 | |
| 201 | uint32_t * data = (uint32_t *) cls->data; |
| 202 | memset(s: cls->data, c: 0, n: n_seqs_unq*ggml_element_size(tensor: cls)); |
| 203 | |
| 204 | std::vector<int> target_pos(n_seqs_unq, -1); |
| 205 | std::vector<int> target_row(n_seqs_unq, -1); |
| 206 | |
| 207 | const bool last = ( |
| 208 | cparams.pooling_type == LLAMA_POOLING_TYPE_LAST || |
| 209 | (cparams.pooling_type == LLAMA_POOLING_TYPE_RANK && arch == LLM_ARCH_QWEN3) // qwen3 reranking & embedding models use last token |
| 210 | ); |
| 211 | |
| 212 | for (int i = 0; i < n_tokens; ++i) { |
| 213 | const llama_pos pos = ubatch->pos[i]; |
| 214 | |
| 215 | for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { |
| 216 | const llama_seq_id seq_id = ubatch->seq_id[i][s]; |
| 217 | const int32_t seq_idx = ubatch->seq_idx[seq_id]; |
| 218 | |
| 219 | if ( |
| 220 | (target_pos[seq_idx] == -1) || |
| 221 | ( last && pos >= target_pos[seq_idx]) || |
| 222 | (!last && pos < target_pos[seq_idx]) |
| 223 | ) { |
| 224 | target_pos[seq_idx] = pos; |
| 225 | target_row[seq_idx] = i; |
| 226 | } |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | for (int s = 0; s < n_seqs_unq; ++s) { |
| 231 | if (target_row[s] >= 0) { |
| 232 | data[s] = target_row[s]; |
| 233 | } |
| 234 | } |
| 235 | } |
| 236 | } |
| 237 | |
| 238 | void llm_graph_input_rs::set_input(const llama_ubatch * ubatch) { |
| 239 | GGML_UNUSED(ubatch); |
| 240 | |
| 241 | const int64_t n_rs = mctx->get_n_rs(); |
| 242 | |
| 243 | if (s_copy) { |
| 244 | GGML_ASSERT(ggml_backend_buffer_is_host(s_copy->buffer)); |
| 245 | int32_t * data = (int32_t *) s_copy->data; |
| 246 | |
| 247 | // assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n |
| 248 | for (uint32_t i = 0; i < n_rs; ++i) { |
| 249 | data[i] = mctx->s_copy(i); |
| 250 | } |
| 251 | } |
| 252 | } |
| 253 | |
| 254 | void llm_graph_input_cross_embd::set_input(const llama_ubatch * ubatch) { |
| 255 | GGML_UNUSED(ubatch); |
| 256 | |
| 257 | if (cross_embd && !cross->v_embd.empty()) { |
| 258 | assert(cross_embd->type == GGML_TYPE_F32); |
| 259 | |
| 260 | ggml_backend_tensor_set(tensor: cross_embd, data: cross->v_embd.data(), offset: 0, size: ggml_nbytes(tensor: cross_embd)); |
| 261 | } |
| 262 | } |
| 263 | |
| 264 | static void print_mask(const float * data, int64_t n_tokens, int64_t n_kv, int64_t n_swa, llama_swa_type swa_type) { |
| 265 | LLAMA_LOG_DEBUG("%s: === Attention mask ===\n" , __func__); |
| 266 | const char * swa_type_str = "unknown" ; |
| 267 | |
| 268 | switch (swa_type) { |
| 269 | case LLAMA_SWA_TYPE_NONE: swa_type_str = "LLAMA_SWA_TYPE_NONE" ; break; |
| 270 | case LLAMA_SWA_TYPE_STANDARD: swa_type_str = "LLAMA_SWA_TYPE_STANDARD" ; break; |
| 271 | case LLAMA_SWA_TYPE_CHUNKED: swa_type_str = "LLAMA_SWA_TYPE_CHUNKED" ; break; |
| 272 | case LLAMA_SWA_TYPE_SYMMETRIC: swa_type_str = "LLAMA_SWA_TYPE_SYMMETRIC" ; break; |
| 273 | }; |
| 274 | |
| 275 | LLAMA_LOG_DEBUG("%s: n_swa : %d, n_kv: %d, swq_type: %s\n" , __func__, (int)n_swa, (int)n_kv, swa_type_str); |
| 276 | LLAMA_LOG_DEBUG("%s: '0' = can attend, '∞' = masked\n" , __func__); |
| 277 | LLAMA_LOG_DEBUG("%s: Rows = query tokens, Columns = key/value tokens\n\n" , __func__); |
| 278 | |
| 279 | LLAMA_LOG_DEBUG(" " ); |
| 280 | for (int j = 0; j < std::min(a: (int64_t)20, b: n_kv); ++j) { |
| 281 | LLAMA_LOG_DEBUG("%2d" , j); |
| 282 | } |
| 283 | LLAMA_LOG_DEBUG("\n" ); |
| 284 | |
| 285 | for (int i = 0; i < std::min(a: (int64_t)20, b: n_tokens); ++i) { |
| 286 | LLAMA_LOG_DEBUG(" %2d " , i); |
| 287 | for (int j = 0; j < std::min(a: (int64_t)20, b: n_kv); ++j) { |
| 288 | float val = data[i * n_kv + j]; |
| 289 | if (val == -INFINITY) { |
| 290 | LLAMA_LOG_DEBUG(" ∞" ); |
| 291 | } else { |
| 292 | LLAMA_LOG_DEBUG(" 0" ); |
| 293 | } |
| 294 | } |
| 295 | LLAMA_LOG_DEBUG("\n" ); |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | void llm_graph_input_attn_no_cache::set_input(const llama_ubatch * ubatch) { |
| 300 | const int64_t n_kv = ubatch->n_tokens; |
| 301 | const int64_t n_tokens = ubatch->n_tokens; |
| 302 | |
| 303 | const auto fill_mask = [&](float * data, int n_swa, llama_swa_type swa_type) { |
| 304 | for (int h = 0; h < 1; ++h) { |
| 305 | for (int i1 = 0; i1 < n_tokens; ++i1) { |
| 306 | const llama_seq_id s1 = ubatch->seq_id[i1][0]; |
| 307 | const llama_pos p1 = ubatch->pos[i1]; |
| 308 | |
| 309 | const uint64_t idst = h*(n_kv*n_tokens) + i1*n_kv; |
| 310 | |
| 311 | for (int i0 = 0; i0 < n_tokens; ++i0) { |
| 312 | const llama_seq_id s0 = ubatch->seq_id[i0][0]; |
| 313 | const llama_pos p0 = ubatch->pos[i0]; |
| 314 | |
| 315 | // mask different sequences |
| 316 | if (s0 != s1) { |
| 317 | continue; |
| 318 | } |
| 319 | |
| 320 | // mask future tokens |
| 321 | if (cparams.causal_attn && p0 > p1) { |
| 322 | continue; |
| 323 | } |
| 324 | |
| 325 | // apply SWA if any |
| 326 | if (llama_hparams::is_masked_swa(n_swa, swa_type, p0, p1)) { |
| 327 | continue; |
| 328 | } |
| 329 | |
| 330 | data[idst + i0] = hparams.use_alibi ? -std::abs(x: p0 - p1) : 0.0f; |
| 331 | } |
| 332 | } |
| 333 | } |
| 334 | }; |
| 335 | |
| 336 | { |
| 337 | GGML_ASSERT(self_kq_mask); |
| 338 | GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask->buffer)); |
| 339 | |
| 340 | float * data = (float *) self_kq_mask->data; |
| 341 | |
| 342 | std::fill(first: data, last: data + ggml_nelements(tensor: self_kq_mask), value: -INFINITY); |
| 343 | |
| 344 | fill_mask(data, 0, LLAMA_SWA_TYPE_NONE); |
| 345 | |
| 346 | if (debug) { |
| 347 | print_mask(data, n_tokens, n_kv, n_swa: 0, swa_type: LLAMA_SWA_TYPE_NONE); |
| 348 | } |
| 349 | } |
| 350 | |
| 351 | if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) { |
| 352 | GGML_ASSERT(self_kq_mask_swa); |
| 353 | GGML_ASSERT(ggml_backend_buffer_is_host(self_kq_mask_swa->buffer)); |
| 354 | |
| 355 | float * data = (float *) self_kq_mask_swa->data; |
| 356 | |
| 357 | std::fill(first: data, last: data + ggml_nelements(tensor: self_kq_mask_swa), value: -INFINITY); |
| 358 | |
| 359 | fill_mask(data, hparams.n_swa, hparams.swa_type); |
| 360 | |
| 361 | if (debug) { |
| 362 | print_mask(data, n_tokens, n_kv, n_swa: hparams.n_swa, swa_type: hparams.swa_type); |
| 363 | } |
| 364 | } |
| 365 | } |
| 366 | |
| 367 | void llm_graph_input_attn_kv::set_input(const llama_ubatch * ubatch) { |
| 368 | mctx->set_input_k_idxs(dst: self_k_idxs, ubatch); |
| 369 | mctx->set_input_v_idxs(dst: self_v_idxs, ubatch); |
| 370 | |
| 371 | mctx->set_input_kq_mask(dst: self_kq_mask, ubatch, causal_attn: cparams.causal_attn); |
| 372 | } |
| 373 | |
| 374 | bool llm_graph_input_attn_kv::can_reuse(const llm_graph_params & params) { |
| 375 | const auto * mctx = static_cast<const llama_kv_cache_context *>(params.mctx); |
| 376 | |
| 377 | this->mctx = mctx; |
| 378 | |
| 379 | bool res = true; |
| 380 | |
| 381 | res &= self_k_idxs->ne[0] == params.ubatch.n_tokens; |
| 382 | //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there |
| 383 | |
| 384 | res &= self_kq_mask->ne[0] == mctx->get_n_kv(); |
| 385 | res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD); |
| 386 | |
| 387 | return res; |
| 388 | } |
| 389 | |
| 390 | void llm_graph_input_attn_kv_iswa::set_input(const llama_ubatch * ubatch) { |
| 391 | mctx->get_base()->set_input_k_idxs(dst: self_k_idxs, ubatch); |
| 392 | mctx->get_base()->set_input_v_idxs(dst: self_v_idxs, ubatch); |
| 393 | |
| 394 | mctx->get_base()->set_input_kq_mask(dst: self_kq_mask, ubatch, causal_attn: cparams.causal_attn); |
| 395 | |
| 396 | mctx->get_swa()->set_input_k_idxs(dst: self_k_idxs_swa, ubatch); |
| 397 | mctx->get_swa()->set_input_v_idxs(dst: self_v_idxs_swa, ubatch); |
| 398 | |
| 399 | mctx->get_swa()->set_input_kq_mask(dst: self_kq_mask_swa, ubatch, causal_attn: cparams.causal_attn); |
| 400 | } |
| 401 | |
| 402 | bool llm_graph_input_attn_kv_iswa::can_reuse(const llm_graph_params & params) { |
| 403 | const auto * mctx = static_cast<const llama_kv_cache_iswa_context *>(params.mctx); |
| 404 | |
| 405 | this->mctx = mctx; |
| 406 | |
| 407 | bool res = true; |
| 408 | |
| 409 | res &= self_k_idxs->ne[0] == params.ubatch.n_tokens; |
| 410 | //res &= self_v_idxs->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there |
| 411 | |
| 412 | res &= self_k_idxs_swa->ne[0] == params.ubatch.n_tokens; |
| 413 | //res &= self_v_idxs_swa->ne[0] == params.ubatch.n_tokens; // TODO: need to move this to the unified cache and check there |
| 414 | |
| 415 | res &= self_kq_mask->ne[0] == mctx->get_base()->get_n_kv(); |
| 416 | res &= self_kq_mask->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD); |
| 417 | |
| 418 | res &= self_kq_mask_swa->ne[0] == mctx->get_swa()->get_n_kv(); |
| 419 | res &= self_kq_mask_swa->ne[1] == GGML_PAD(params.ubatch.n_tokens, GGML_KQ_MASK_PAD); |
| 420 | |
| 421 | return res; |
| 422 | } |
| 423 | |
| 424 | void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) { |
| 425 | GGML_ASSERT(cross_kq_mask); |
| 426 | |
| 427 | const int64_t n_enc = cross_kq_mask->ne[0]; |
| 428 | const int64_t n_tokens = ubatch->n_tokens; |
| 429 | |
| 430 | GGML_ASSERT(ggml_backend_buffer_is_host(cross_kq_mask->buffer)); |
| 431 | GGML_ASSERT(!ubatch->equal_seqs()); // TODO: use ubatch->n_seqs instead of failing |
| 432 | |
| 433 | float * data = (float *) cross_kq_mask->data; |
| 434 | |
| 435 | for (int h = 0; h < 1; ++h) { |
| 436 | for (int i = 0; i < n_tokens; ++i) { |
| 437 | for (int j = 0; j < n_enc; ++j) { |
| 438 | float f = -INFINITY; |
| 439 | |
| 440 | for (int s = 0; s < ubatch->n_seq_id[i]; ++s) { |
| 441 | const llama_seq_id seq_id = ubatch->seq_id[i][s]; |
| 442 | |
| 443 | if (cross->seq_ids_enc[j].find(x: seq_id) != cross->seq_ids_enc[j].end()) { |
| 444 | f = 0.0f; |
| 445 | } |
| 446 | } |
| 447 | |
| 448 | data[h*(n_enc*n_tokens) + i*n_enc + j] = f; |
| 449 | } |
| 450 | } |
| 451 | |
| 452 | for (int i = n_tokens; i < GGML_PAD(n_tokens, GGML_KQ_MASK_PAD); ++i) { |
| 453 | for (int j = 0; j < n_enc; ++j) { |
| 454 | data[h*(n_enc*n_tokens) + i*n_enc + j] = -INFINITY; |
| 455 | } |
| 456 | } |
| 457 | } |
| 458 | } |
| 459 | |
| 460 | void llm_graph_input_mem_hybrid::set_input(const llama_ubatch * ubatch) { |
| 461 | inp_attn->set_input(ubatch); |
| 462 | inp_rs->set_input(ubatch); |
| 463 | } |
| 464 | |
| 465 | // |
| 466 | // llm_graph_result |
| 467 | // |
| 468 | |
| 469 | llm_graph_result::llm_graph_result(int64_t max_nodes) : max_nodes(max_nodes) { |
| 470 | reset(); |
| 471 | |
| 472 | const char * LLAMA_GRAPH_RESULT_DEBUG = getenv(name: "LLAMA_GRAPH_RESULT_DEBUG" ); |
| 473 | debug = LLAMA_GRAPH_RESULT_DEBUG ? atoi(nptr: LLAMA_GRAPH_RESULT_DEBUG) : 0; |
| 474 | } |
| 475 | |
| 476 | int64_t llm_graph_result::get_max_nodes() const { |
| 477 | return max_nodes; |
| 478 | } |
| 479 | |
| 480 | void llm_graph_result::reset() { |
| 481 | t_tokens = nullptr; |
| 482 | t_logits = nullptr; |
| 483 | t_embd = nullptr; |
| 484 | t_embd_pooled = nullptr; |
| 485 | |
| 486 | params = {}; |
| 487 | |
| 488 | inputs.clear(); |
| 489 | |
| 490 | buf_compute_meta.resize(new_size: ggml_tensor_overhead()*max_nodes + ggml_graph_overhead_custom(size: max_nodes, grads: false)); |
| 491 | |
| 492 | ggml_init_params params = { |
| 493 | /*.mem_size =*/ buf_compute_meta.size(), |
| 494 | /*.mem_buffer =*/ buf_compute_meta.data(), |
| 495 | /*.no_alloc =*/ true, |
| 496 | }; |
| 497 | |
| 498 | ctx_compute.reset(p: ggml_init(params)); |
| 499 | |
| 500 | gf = ggml_new_graph_custom(ctx: ctx_compute.get(), size: max_nodes, grads: false); |
| 501 | } |
| 502 | |
| 503 | void llm_graph_result::set_inputs(const llama_ubatch * ubatch) { |
| 504 | for (auto & input : inputs) { |
| 505 | input->set_input(ubatch); |
| 506 | } |
| 507 | } |
| 508 | |
| 509 | bool llm_graph_result::can_reuse(const llm_graph_params & params) { |
| 510 | if (!this->params.allow_reuse(other: params)) { |
| 511 | if (debug > 1) { |
| 512 | LLAMA_LOG_DEBUG("%s: cannot reuse graph due to incompatible graph parameters\n" , __func__); |
| 513 | } |
| 514 | |
| 515 | return false; |
| 516 | } |
| 517 | |
| 518 | if (debug > 1) { |
| 519 | LLAMA_LOG_DEBUG("%s: checking compatibility of %d inputs:\n" , __func__, (int) inputs.size()); |
| 520 | } |
| 521 | |
| 522 | bool res = true; |
| 523 | |
| 524 | for (auto & input : inputs) { |
| 525 | const bool cur = input->can_reuse(params); |
| 526 | |
| 527 | if (debug > 1) { |
| 528 | LLAMA_LOG_DEBUG("%s: can_reuse = %d\n" , "placeholder" , cur); |
| 529 | } |
| 530 | |
| 531 | res = res && cur; |
| 532 | } |
| 533 | |
| 534 | if (debug > 0) { |
| 535 | LLAMA_LOG_DEBUG("%s: can reuse graph = %d\n" , __func__, res); |
| 536 | } |
| 537 | |
| 538 | return res; |
| 539 | } |
| 540 | |
| 541 | llm_graph_input_i * llm_graph_result::add_input(llm_graph_input_ptr input) { |
| 542 | inputs.emplace_back(args: std::move(input)); |
| 543 | return inputs.back().get(); |
| 544 | } |
| 545 | |
| 546 | void llm_graph_result::set_params(const llm_graph_params & params) { |
| 547 | this->params = params; |
| 548 | } |
| 549 | |
| 550 | // |
| 551 | // llm_graph_context |
| 552 | // |
| 553 | |
| 554 | llm_graph_context::llm_graph_context(const llm_graph_params & params) : |
| 555 | arch (params.arch), |
| 556 | hparams (params.hparams), |
| 557 | cparams (params.cparams), |
| 558 | ubatch (params.ubatch), |
| 559 | n_embd (hparams.n_embd), |
| 560 | n_layer (hparams.n_layer), |
| 561 | n_rot (hparams.n_rot), |
| 562 | n_ctx (cparams.n_ctx), |
| 563 | n_head (hparams.n_head()), |
| 564 | n_head_kv (hparams.n_head_kv()), |
| 565 | n_embd_head_k (hparams.n_embd_head_k), |
| 566 | n_embd_k_gqa (hparams.n_embd_k_gqa()), |
| 567 | n_embd_head_v (hparams.n_embd_head_v), |
| 568 | n_embd_v_gqa (hparams.n_embd_v_gqa()), |
| 569 | n_expert (hparams.n_expert), |
| 570 | n_expert_used (cparams.warmup ? hparams.n_expert : hparams.n_expert_used), |
| 571 | freq_base (cparams.rope_freq_base), |
| 572 | freq_scale (cparams.rope_freq_scale), |
| 573 | ext_factor (cparams.yarn_ext_factor), |
| 574 | attn_factor (cparams.yarn_attn_factor), |
| 575 | beta_fast (cparams.yarn_beta_fast), |
| 576 | beta_slow (cparams.yarn_beta_slow), |
| 577 | norm_eps (hparams.f_norm_eps), |
| 578 | norm_rms_eps (hparams.f_norm_rms_eps), |
| 579 | n_tokens (ubatch.n_tokens), |
| 580 | n_outputs (params.n_outputs), |
| 581 | n_ctx_orig (cparams.n_ctx_orig_yarn), |
| 582 | pooling_type (cparams.pooling_type), |
| 583 | rope_type (hparams.rope_type), |
| 584 | sched (params.sched), |
| 585 | backend_cpu (params.backend_cpu), |
| 586 | cvec (params.cvec), |
| 587 | loras (params.loras), |
| 588 | mctx (params.mctx), |
| 589 | cross (params.cross), |
| 590 | cb_func (params.cb), |
| 591 | res (params.res), |
| 592 | ctx0 (res->get_ctx()), |
| 593 | gf (res->get_gf()) { |
| 594 | res->set_params(params); |
| 595 | } |
| 596 | |
| 597 | void llm_graph_context::cb(ggml_tensor * cur, const char * name, int il) const { |
| 598 | if (cb_func) { |
| 599 | cb_func(ubatch, cur, name, il); |
| 600 | } |
| 601 | } |
| 602 | |
| 603 | ggml_tensor * llm_graph_context::build_cvec( |
| 604 | ggml_tensor * cur, |
| 605 | int il) const { |
| 606 | return cvec->apply_to(ctx: ctx0, cur, il); |
| 607 | } |
| 608 | |
| 609 | ggml_tensor * llm_graph_context::build_lora_mm( |
| 610 | ggml_tensor * w, |
| 611 | ggml_tensor * cur) const { |
| 612 | ggml_tensor * res = ggml_mul_mat(ctx: ctx0, a: w, b: cur); |
| 613 | |
| 614 | for (const auto & lora : *loras) { |
| 615 | llama_adapter_lora_weight * lw = lora.first->get_weight(w); |
| 616 | if (lw == nullptr) { |
| 617 | continue; |
| 618 | } |
| 619 | |
| 620 | const float adapter_scale = lora.second; |
| 621 | const float scale = lw->get_scale(alpha: lora.first->alpha, adapter_scale); |
| 622 | |
| 623 | ggml_tensor * ab_cur = ggml_mul_mat( |
| 624 | ctx: ctx0, a: lw->b, |
| 625 | b: ggml_mul_mat(ctx: ctx0, a: lw->a, b: cur) |
| 626 | ); |
| 627 | |
| 628 | ab_cur = ggml_scale(ctx: ctx0, a: ab_cur, s: scale); |
| 629 | res = ggml_add(ctx: ctx0, a: res, b: ab_cur); |
| 630 | } |
| 631 | |
| 632 | return res; |
| 633 | } |
| 634 | |
| 635 | ggml_tensor * llm_graph_context::build_lora_mm_id( |
| 636 | ggml_tensor * w, // ggml_tensor * as |
| 637 | ggml_tensor * cur, // ggml_tensor * b |
| 638 | ggml_tensor * ids) const { |
| 639 | ggml_tensor * res = ggml_mul_mat_id(ctx: ctx0, as: w, b: cur, ids); |
| 640 | for (const auto & lora : *loras) { |
| 641 | llama_adapter_lora_weight * lw = lora.first->get_weight(w); |
| 642 | if (lw == nullptr) { |
| 643 | continue; |
| 644 | } |
| 645 | |
| 646 | const float alpha = lora.first->alpha; |
| 647 | const float rank = (float) lw->b->ne[0]; |
| 648 | const float scale = alpha ? lora.second * alpha / rank : lora.second; |
| 649 | |
| 650 | ggml_tensor * ab_cur = ggml_mul_mat_id( |
| 651 | ctx: ctx0, as: lw->b, |
| 652 | b: ggml_mul_mat_id(ctx: ctx0, as: lw->a, b: cur, ids), |
| 653 | ids |
| 654 | ); |
| 655 | |
| 656 | ab_cur = ggml_scale(ctx: ctx0, a: ab_cur, s: scale); |
| 657 | res = ggml_add(ctx: ctx0, a: res, b: ab_cur); |
| 658 | } |
| 659 | |
| 660 | return res; |
| 661 | } |
| 662 | |
| 663 | ggml_tensor * llm_graph_context::build_norm( |
| 664 | ggml_tensor * cur, |
| 665 | ggml_tensor * mw, |
| 666 | ggml_tensor * mb, |
| 667 | llm_norm_type type, |
| 668 | int il) const { |
| 669 | switch (type) { |
| 670 | case LLM_NORM: cur = ggml_norm (ctx: ctx0, a: cur, eps: hparams.f_norm_eps); break; |
| 671 | case LLM_NORM_RMS: cur = ggml_rms_norm(ctx: ctx0, a: cur, eps: hparams.f_norm_rms_eps); break; |
| 672 | case LLM_NORM_GROUP: |
| 673 | { |
| 674 | cur = ggml_reshape_3d(ctx: ctx0, a: cur, ne0: cur->ne[0], ne1: 1, ne2: cur->ne[1]); |
| 675 | cur = ggml_group_norm(ctx: ctx0, a: cur, n_groups: hparams.n_norm_groups, eps: hparams.f_norm_group_eps); |
| 676 | cur = ggml_reshape_2d(ctx: ctx0, a: cur, ne0: cur->ne[0], ne1: cur->ne[2]); |
| 677 | } break; |
| 678 | } |
| 679 | |
| 680 | if (mw || mb) { |
| 681 | cb(cur, name: "norm" , il); |
| 682 | } |
| 683 | |
| 684 | if (mw) { |
| 685 | cur = ggml_mul(ctx: ctx0, a: cur, b: mw); |
| 686 | if (mb) { |
| 687 | cb(cur, name: "norm_w" , il); |
| 688 | } |
| 689 | } |
| 690 | |
| 691 | if (mb) { |
| 692 | cur = ggml_add(ctx: ctx0, a: cur, b: mb); |
| 693 | } |
| 694 | |
| 695 | return cur; |
| 696 | } |
| 697 | |
| 698 | ggml_tensor * llm_graph_context::build_ffn( |
| 699 | ggml_tensor * cur, |
| 700 | ggml_tensor * up, |
| 701 | ggml_tensor * up_b, |
| 702 | ggml_tensor * up_s, |
| 703 | ggml_tensor * gate, |
| 704 | ggml_tensor * gate_b, |
| 705 | ggml_tensor * gate_s, |
| 706 | ggml_tensor * down, |
| 707 | ggml_tensor * down_b, |
| 708 | ggml_tensor * down_s, |
| 709 | ggml_tensor * act_scales, |
| 710 | llm_ffn_op_type type_op, |
| 711 | llm_ffn_gate_type type_gate, |
| 712 | int il) const { |
| 713 | ggml_tensor * tmp = up ? build_lora_mm(w: up, cur) : cur; |
| 714 | cb(cur: tmp, name: "ffn_up" , il); |
| 715 | |
| 716 | if (up_b) { |
| 717 | tmp = ggml_add(ctx: ctx0, a: tmp, b: up_b); |
| 718 | cb(cur: tmp, name: "ffn_up_b" , il); |
| 719 | } |
| 720 | |
| 721 | if (up_s) { |
| 722 | tmp = ggml_mul(ctx: ctx0, a: tmp, b: up_s); |
| 723 | cb(cur: tmp, name: "ffn_up_s" , il); |
| 724 | } |
| 725 | |
| 726 | if (gate) { |
| 727 | switch (type_gate) { |
| 728 | case LLM_FFN_SEQ: |
| 729 | { |
| 730 | cur = build_lora_mm(w: gate, cur: tmp); |
| 731 | cb(cur, name: "ffn_gate" , il); |
| 732 | } break; |
| 733 | case LLM_FFN_PAR: |
| 734 | { |
| 735 | cur = build_lora_mm(w: gate, cur); |
| 736 | cb(cur, name: "ffn_gate" , il); |
| 737 | } break; |
| 738 | } |
| 739 | |
| 740 | if (gate_b) { |
| 741 | cur = ggml_add(ctx: ctx0, a: cur, b: gate_b); |
| 742 | cb(cur, name: "ffn_gate_b" , il); |
| 743 | } |
| 744 | |
| 745 | if (gate_s) { |
| 746 | cur = ggml_mul(ctx: ctx0, a: cur, b: gate_s); |
| 747 | cb(cur, name: "ffn_gate_s" , il); |
| 748 | } |
| 749 | |
| 750 | } else { |
| 751 | cur = tmp; |
| 752 | } |
| 753 | |
| 754 | switch (type_op) { |
| 755 | case LLM_FFN_SILU: |
| 756 | if (gate && type_gate == LLM_FFN_PAR) { |
| 757 | cur = ggml_swiglu_split(ctx: ctx0, a: cur, b: tmp); |
| 758 | cb(cur, name: "ffn_swiglu" , il); |
| 759 | type_gate = LLM_FFN_SEQ; |
| 760 | } else { |
| 761 | cur = ggml_silu(ctx: ctx0, a: cur); |
| 762 | cb(cur, name: "ffn_silu" , il); |
| 763 | } break; |
| 764 | case LLM_FFN_GELU: |
| 765 | if (gate && type_gate == LLM_FFN_PAR) { |
| 766 | cur = ggml_geglu_split(ctx: ctx0, a: cur, b: tmp); |
| 767 | cb(cur, name: "ffn_geglu" , il); |
| 768 | type_gate = LLM_FFN_SEQ; |
| 769 | } else { |
| 770 | cur = ggml_gelu(ctx: ctx0, a: cur); |
| 771 | cb(cur, name: "ffn_gelu" , il); |
| 772 | if (act_scales != NULL) { |
| 773 | cur = ggml_div(ctx: ctx0, a: cur, b: act_scales); |
| 774 | cb(cur, name: "ffn_act" , il); |
| 775 | } |
| 776 | } break; |
| 777 | case LLM_FFN_RELU: |
| 778 | if (gate && type_gate == LLM_FFN_PAR) { |
| 779 | cur = ggml_reglu_split(ctx: ctx0, a: cur, b: tmp); |
| 780 | cb(cur, name: "ffn_reglu" , il); |
| 781 | type_gate = LLM_FFN_SEQ; |
| 782 | } else { |
| 783 | cur = ggml_relu(ctx: ctx0, a: cur); |
| 784 | cb(cur, name: "ffn_relu" , il); |
| 785 | } break; |
| 786 | case LLM_FFN_RELU_SQR: |
| 787 | { |
| 788 | cur = ggml_relu(ctx: ctx0, a: cur); |
| 789 | cb(cur, name: "ffn_relu" , il); |
| 790 | |
| 791 | cur = ggml_sqr(ctx: ctx0, a: cur); |
| 792 | cb(cur, name: "ffn_sqr(relu)" , il); |
| 793 | } break; |
| 794 | case LLM_FFN_SWIGLU: |
| 795 | { |
| 796 | cur = ggml_swiglu(ctx: ctx0, a: cur); |
| 797 | cb(cur, name: "ffn_swiglu" , il); |
| 798 | } break; |
| 799 | case LLM_FFN_GEGLU: |
| 800 | { |
| 801 | cur = ggml_geglu(ctx: ctx0, a: cur); |
| 802 | cb(cur, name: "ffn_geglu" , il); |
| 803 | } break; |
| 804 | case LLM_FFN_REGLU: |
| 805 | { |
| 806 | cur = ggml_reglu(ctx: ctx0, a: cur); |
| 807 | cb(cur, name: "ffn_reglu" , il); |
| 808 | } break; |
| 809 | default: |
| 810 | GGML_ABORT("fatal error" ); |
| 811 | } |
| 812 | |
| 813 | //expand here so that we can fuse ffn gate |
| 814 | ggml_build_forward_expand(cgraph: gf, tensor: cur); |
| 815 | |
| 816 | if (gate && type_gate == LLM_FFN_PAR) { |
| 817 | cur = ggml_mul(ctx: ctx0, a: cur, b: tmp); |
| 818 | cb(cur, name: "ffn_gate_par" , il); |
| 819 | } |
| 820 | |
| 821 | if (down) { |
| 822 | cur = build_lora_mm(w: down, cur); |
| 823 | if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) { |
| 824 | // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators |
| 825 | ggml_mul_mat_set_prec(a: cur, prec: GGML_PREC_F32); |
| 826 | } |
| 827 | } |
| 828 | |
| 829 | if (down_b) { |
| 830 | cb(cur, name: "ffn_down" , il); |
| 831 | } |
| 832 | |
| 833 | if (down_b) { |
| 834 | cur = ggml_add(ctx: ctx0, a: cur, b: down_b); |
| 835 | } |
| 836 | |
| 837 | if (down_s) { |
| 838 | cur = ggml_mul(ctx: ctx0, a: cur, b: down_s); |
| 839 | cb(cur, name: "ffn_down_s" , il); |
| 840 | } |
| 841 | |
| 842 | return cur; |
| 843 | } |
| 844 | |
| 845 | ggml_tensor * llm_graph_context::build_moe_ffn( |
| 846 | ggml_tensor * cur, |
| 847 | ggml_tensor * gate_inp, |
| 848 | ggml_tensor * up_exps, |
| 849 | ggml_tensor * gate_exps, |
| 850 | ggml_tensor * down_exps, |
| 851 | ggml_tensor * exp_probs_b, |
| 852 | int64_t n_expert, |
| 853 | int64_t n_expert_used, |
| 854 | llm_ffn_op_type type_op, |
| 855 | bool norm_w, |
| 856 | bool scale_w, |
| 857 | float w_scale, |
| 858 | llama_expert_gating_func_type gating_op, |
| 859 | int il, |
| 860 | ggml_tensor * probs_in) const { |
| 861 | return build_moe_ffn( |
| 862 | cur, |
| 863 | gate_inp, /* gate_inp_b */ nullptr, |
| 864 | up_exps, /* up_exps_b */ nullptr, |
| 865 | gate_exps, /* gate_exps_b */ nullptr, |
| 866 | down_exps, /* down_exps_b */ nullptr, |
| 867 | exp_probs_b, |
| 868 | n_expert, |
| 869 | n_expert_used, |
| 870 | type_op, |
| 871 | norm_w, |
| 872 | scale_w, |
| 873 | w_scale, |
| 874 | gating_op, |
| 875 | il, |
| 876 | probs_in |
| 877 | ); |
| 878 | } |
| 879 | |
| 880 | ggml_tensor * llm_graph_context::build_moe_ffn( |
| 881 | ggml_tensor * cur, |
| 882 | ggml_tensor * gate_inp, |
| 883 | ggml_tensor * gate_inp_b, |
| 884 | ggml_tensor * up_exps, |
| 885 | ggml_tensor * up_exps_b, |
| 886 | ggml_tensor * gate_exps, |
| 887 | ggml_tensor * gate_exps_b, |
| 888 | ggml_tensor * down_exps, |
| 889 | ggml_tensor * down_exps_b, |
| 890 | ggml_tensor * exp_probs_b, |
| 891 | int64_t n_expert, |
| 892 | int64_t n_expert_used, |
| 893 | llm_ffn_op_type type_op, |
| 894 | bool norm_w, |
| 895 | bool scale_w, |
| 896 | float w_scale, |
| 897 | llama_expert_gating_func_type gating_op, |
| 898 | int il, |
| 899 | ggml_tensor * probs_in) const { |
| 900 | const int64_t n_embd = cur->ne[0]; |
| 901 | const int64_t n_tokens = cur->ne[1]; |
| 902 | const bool weight_before_ffn = arch == LLM_ARCH_LLAMA4; // for llama4, we apply the sigmoid-ed weights before the FFN |
| 903 | |
| 904 | ggml_tensor * logits = nullptr; |
| 905 | |
| 906 | if (probs_in == nullptr) { |
| 907 | logits = build_lora_mm(w: gate_inp, cur); // [n_expert, n_tokens] |
| 908 | cb(cur: logits, name: "ffn_moe_logits" , il); |
| 909 | } else { |
| 910 | logits = probs_in; |
| 911 | } |
| 912 | |
| 913 | if (gate_inp_b) { |
| 914 | logits = ggml_add(ctx: ctx0, a: logits, b: gate_inp_b); |
| 915 | cb(cur: logits, name: "ffn_moe_logits_biased" , il); |
| 916 | } |
| 917 | |
| 918 | ggml_tensor * probs = nullptr; |
| 919 | switch (gating_op) { |
| 920 | case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX: |
| 921 | { |
| 922 | probs = ggml_soft_max(ctx: ctx0, a: logits); // [n_expert, n_tokens] |
| 923 | } break; |
| 924 | case LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID: |
| 925 | { |
| 926 | probs = ggml_sigmoid(ctx: ctx0, a: logits); // [n_expert, n_tokens] |
| 927 | } break; |
| 928 | case LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT: |
| 929 | { |
| 930 | probs = logits; // [n_expert, n_tokens] |
| 931 | } break; |
| 932 | default: |
| 933 | GGML_ABORT("fatal error" ); |
| 934 | } |
| 935 | cb(cur: probs, name: "ffn_moe_probs" , il); |
| 936 | |
| 937 | // add experts selection bias - introduced in DeepSeek V3 |
| 938 | // leave probs unbiased as it's later used to get expert weights |
| 939 | ggml_tensor * selection_probs = probs; |
| 940 | if (exp_probs_b != nullptr) { |
| 941 | selection_probs = ggml_add(ctx: ctx0, a: probs, b: exp_probs_b); |
| 942 | cb(cur: selection_probs, name: "ffn_moe_probs_biased" , il); |
| 943 | } |
| 944 | |
| 945 | // llama4 doesn't have exp_probs_b, and sigmoid is only used after top_k |
| 946 | // see: https://github.com/meta-llama/llama-models/blob/699a02993512fb36936b1b0741e13c06790bcf98/models/llama4/moe.py#L183-L198 |
| 947 | if (arch == LLM_ARCH_LLAMA4) { |
| 948 | selection_probs = logits; |
| 949 | } |
| 950 | |
| 951 | if (arch == LLM_ARCH_GROVEMOE) { |
| 952 | selection_probs = ggml_sigmoid(ctx: ctx0, a: logits); // [n_expert, n_tokens] |
| 953 | cb(cur: selection_probs, name: "ffn_moe_probs_biased" , il); |
| 954 | } |
| 955 | |
| 956 | // select top n_group_used expert groups |
| 957 | // https://huggingface.co/deepseek-ai/DeepSeek-V3/blob/e815299b0bcbac849fa540c768ef21845365c9eb/modeling_deepseek.py#L440-L457 |
| 958 | if (hparams.n_expert_groups > 1 && n_tokens > 0) { |
| 959 | const int64_t n_exp_per_group = n_expert / hparams.n_expert_groups; |
| 960 | |
| 961 | // organize experts into n_expert_groups |
| 962 | ggml_tensor * selection_groups = ggml_reshape_3d(ctx: ctx0, a: selection_probs, ne0: n_exp_per_group, ne1: hparams.n_expert_groups, ne2: n_tokens); // [n_exp_per_group, n_expert_groups, n_tokens] |
| 963 | |
| 964 | ggml_tensor * group_scores = ggml_top_k(ctx: ctx0, a: selection_groups, k: 2); // [2, n_expert_groups, n_tokens] |
| 965 | group_scores = ggml_get_rows(ctx: ctx0, a: ggml_reshape_4d(ctx: ctx0, a: selection_groups, ne0: 1, ne1: selection_groups->ne[0], ne2: selection_groups->ne[1], ne3: selection_groups->ne[2]), b: group_scores); // [1, 2, n_expert_groups, n_tokens] |
| 966 | |
| 967 | // get top n_group_used expert groups |
| 968 | group_scores = ggml_sum_rows(ctx: ctx0, a: ggml_reshape_3d(ctx: ctx0, a: group_scores, ne0: group_scores->ne[1], ne1: group_scores->ne[2], ne2: group_scores->ne[3])); // [1, n_expert_groups, n_tokens] |
| 969 | group_scores = ggml_reshape_2d(ctx: ctx0, a: group_scores, ne0: group_scores->ne[1], ne1: group_scores->ne[2]); // [n_expert_groups, n_tokens] |
| 970 | |
| 971 | ggml_tensor * expert_groups = ggml_top_k(ctx: ctx0, a: group_scores, k: hparams.n_group_used); // [n_group_used, n_tokens] |
| 972 | cb(cur: expert_groups, name: "ffn_moe_group_topk" , il); |
| 973 | |
| 974 | // mask out the other groups |
| 975 | selection_probs = ggml_get_rows(ctx: ctx0, a: selection_groups, b: expert_groups); // [n_exp_per_group, n_group_used, n_tokens] |
| 976 | selection_probs = ggml_set_rows(ctx: ctx0, a: ggml_scale_bias(ctx: ctx0, a: selection_groups, s: 0.0f, b: -INFINITY), b: selection_probs, c: expert_groups); // [n_exp_per_group, n_expert_groups, n_tokens] |
| 977 | selection_probs = ggml_reshape_2d(ctx: ctx0, a: selection_probs, ne0: n_expert, ne1: n_tokens); // [n_expert, n_tokens] |
| 978 | cb(cur: selection_probs, name: "ffn_moe_probs_masked" , il); |
| 979 | } |
| 980 | |
| 981 | // select experts |
| 982 | ggml_tensor * selected_experts = ggml_top_k(ctx: ctx0, a: selection_probs, k: n_expert_used); // [n_expert_used, n_tokens] |
| 983 | cb(cur: selected_experts->src[0], name: "ffn_moe_argsort" , il); |
| 984 | cb(cur: selected_experts, name: "ffn_moe_topk" , il); |
| 985 | |
| 986 | if (arch == LLM_ARCH_GROVEMOE && n_expert != hparams.n_expert) { |
| 987 | // TODO: Use scalar div instead when/if implemented |
| 988 | ggml_tensor * f_sel = ggml_cast(ctx: ctx0, a: selected_experts, type: GGML_TYPE_F32); |
| 989 | selected_experts = ggml_cast(ctx: ctx0, a: ggml_scale(ctx: ctx0, a: f_sel, s: 1.0f / float(hparams.n_group_experts)), type: GGML_TYPE_I32); |
| 990 | probs = ggml_reshape_3d(ctx: ctx0, a: probs, ne0: 1, ne1: hparams.n_expert, ne2: n_tokens); |
| 991 | } else { |
| 992 | probs = ggml_reshape_3d(ctx: ctx0, a: probs, ne0: 1, ne1: n_expert, ne2: n_tokens); |
| 993 | } |
| 994 | |
| 995 | ggml_tensor * weights = ggml_get_rows(ctx: ctx0, a: probs, b: selected_experts); // [1, n_expert_used, n_tokens] |
| 996 | cb(cur: weights, name: "ffn_moe_weights" , il); |
| 997 | |
| 998 | |
| 999 | if (gating_op == LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX_WEIGHT) { |
| 1000 | weights = ggml_reshape_2d(ctx: ctx0, a: weights, ne0: n_expert_used, ne1: n_tokens); |
| 1001 | weights = ggml_soft_max(ctx: ctx0, a: weights); // [n_expert_used, n_tokens] |
| 1002 | weights = ggml_reshape_3d(ctx: ctx0, a: weights, ne0: 1, ne1: n_expert_used, ne2: n_tokens); |
| 1003 | cb(cur: weights, name: "ffn_moe_weights_softmax" , il); |
| 1004 | } |
| 1005 | |
| 1006 | if (norm_w) { |
| 1007 | weights = ggml_reshape_2d(ctx: ctx0, a: weights, ne0: n_expert_used, ne1: n_tokens); |
| 1008 | |
| 1009 | ggml_tensor * weights_sum = ggml_sum_rows(ctx: ctx0, a: weights); // [1, n_tokens] |
| 1010 | cb(cur: weights_sum, name: "ffn_moe_weights_sum" , il); |
| 1011 | |
| 1012 | // Avoid division by zero, clamp to smallest number representable by F16 |
| 1013 | weights_sum = ggml_clamp(ctx: ctx0, a: weights_sum, min: 6.103515625e-5, INFINITY); |
| 1014 | cb(cur: weights_sum, name: "ffn_moe_weights_sum_clamped" , il); |
| 1015 | |
| 1016 | weights = ggml_div(ctx: ctx0, a: weights, b: weights_sum); // [n_expert_used, n_tokens] |
| 1017 | cb(cur: weights, name: "ffn_moe_weights_norm" , il); |
| 1018 | |
| 1019 | weights = ggml_reshape_3d(ctx: ctx0, a: weights, ne0: 1, ne1: n_expert_used, ne2: n_tokens); |
| 1020 | } |
| 1021 | if (scale_w) { |
| 1022 | weights = ggml_scale(ctx: ctx0, a: weights, s: w_scale); |
| 1023 | cb(cur: weights, name: "ffn_moe_weights_scaled" , il); |
| 1024 | } |
| 1025 | |
| 1026 | //call early so that topk-moe can be used |
| 1027 | ggml_build_forward_expand(cgraph: gf, tensor: weights); |
| 1028 | |
| 1029 | cur = ggml_reshape_3d(ctx: ctx0, a: cur, ne0: n_embd, ne1: 1, ne2: n_tokens); |
| 1030 | |
| 1031 | if (weight_before_ffn) { |
| 1032 | // repeat cur to [n_embd, n_expert_used, n_tokens] |
| 1033 | ggml_tensor * repeated = ggml_repeat_4d(ctx: ctx0, a: cur, ne0: n_embd, ne1: n_expert_used, ne2: n_tokens, ne3: 1); |
| 1034 | cur = ggml_mul(ctx: ctx0, a: repeated, b: weights); |
| 1035 | cb(cur, name: "ffn_moe_weighted" , il); |
| 1036 | } |
| 1037 | |
| 1038 | ggml_tensor * up = build_lora_mm_id(w: up_exps, cur, ids: selected_experts); // [n_ff, n_expert_used, n_tokens] |
| 1039 | cb(cur: up, name: "ffn_moe_up" , il); |
| 1040 | |
| 1041 | if (up_exps_b) { |
| 1042 | up = ggml_add_id(ctx: ctx0, a: up, b: up_exps_b, ids: selected_experts); |
| 1043 | cb(cur: up, name: "ffn_moe_up_biased" , il); |
| 1044 | } |
| 1045 | |
| 1046 | ggml_tensor * experts = nullptr; |
| 1047 | if (gate_exps) { |
| 1048 | cur = build_lora_mm_id(w: gate_exps, cur, ids: selected_experts); // [n_ff, n_expert_used, n_tokens] |
| 1049 | cb(cur, name: "ffn_moe_gate" , il); |
| 1050 | } else { |
| 1051 | cur = up; |
| 1052 | } |
| 1053 | |
| 1054 | if (gate_exps_b) { |
| 1055 | cur = ggml_add_id(ctx: ctx0, a: cur, b: gate_exps_b, ids: selected_experts); |
| 1056 | cb(cur, name: "ffn_moe_gate_biased" , il); |
| 1057 | } |
| 1058 | |
| 1059 | switch (type_op) { |
| 1060 | case LLM_FFN_SILU: |
| 1061 | if (gate_exps) { |
| 1062 | cur = ggml_swiglu_split(ctx: ctx0, a: cur, b: up); |
| 1063 | cb(cur, name: "ffn_moe_swiglu" , il); |
| 1064 | } else { |
| 1065 | cur = ggml_silu(ctx: ctx0, a: cur); |
| 1066 | cb(cur, name: "ffn_moe_silu" , il); |
| 1067 | } break; |
| 1068 | case LLM_FFN_GELU: |
| 1069 | if (gate_exps) { |
| 1070 | cur = ggml_geglu_split(ctx: ctx0, a: cur, b: up); |
| 1071 | cb(cur, name: "ffn_moe_geglu" , il); |
| 1072 | } else { |
| 1073 | cur = ggml_gelu(ctx: ctx0, a: cur); |
| 1074 | cb(cur, name: "ffn_moe_gelu" , il); |
| 1075 | } break; |
| 1076 | case LLM_FFN_SWIGLU_OAI_MOE: |
| 1077 | { |
| 1078 | // TODO: move to hparams? |
| 1079 | constexpr float alpha = 1.702f; |
| 1080 | constexpr float limit = 7.0f; |
| 1081 | cur = ggml_swiglu_oai(ctx: ctx0, a: cur, b: up, alpha, limit); |
| 1082 | cb(cur, name: "ffn_moe_swiglu_oai" , il); |
| 1083 | } break; |
| 1084 | case LLM_FFN_RELU: |
| 1085 | if (gate_exps) { |
| 1086 | cur = ggml_reglu_split(ctx: ctx0, a: cur, b: up); |
| 1087 | cb(cur, name: "ffn_moe_reglu" , il); |
| 1088 | } else { |
| 1089 | cur = ggml_relu(ctx: ctx0, a: cur); |
| 1090 | cb(cur, name: "ffn_moe_relu" , il); |
| 1091 | } break; |
| 1092 | default: |
| 1093 | GGML_ABORT("fatal error" ); |
| 1094 | } |
| 1095 | |
| 1096 | //expand here so that we can fuse ffn gate |
| 1097 | ggml_build_forward_expand(cgraph: gf, tensor: cur); |
| 1098 | |
| 1099 | experts = build_lora_mm_id(w: down_exps, cur, ids: selected_experts); // [n_embd, n_expert_used, n_tokens] |
| 1100 | cb(cur: experts, name: "ffn_moe_down" , il); |
| 1101 | |
| 1102 | if (down_exps_b) { |
| 1103 | experts = ggml_add_id(ctx: ctx0, a: experts, b: down_exps_b, ids: selected_experts); |
| 1104 | cb(cur: experts, name: "ffn_moe_down_biased" , il); |
| 1105 | } |
| 1106 | |
| 1107 | if (!weight_before_ffn) { |
| 1108 | experts = ggml_mul(ctx: ctx0, a: experts, b: weights); |
| 1109 | cb(cur, name: "ffn_moe_weighted" , il); |
| 1110 | } |
| 1111 | |
| 1112 | ggml_tensor * cur_experts[LLAMA_MAX_EXPERTS] = { nullptr }; |
| 1113 | |
| 1114 | assert(n_expert_used > 0); |
| 1115 | |
| 1116 | // order the views before the adds |
| 1117 | for (uint32_t i = 0; i < hparams.n_expert_used; ++i) { |
| 1118 | cur_experts[i] = ggml_view_2d(ctx: ctx0, a: experts, ne0: n_embd, ne1: n_tokens, nb1: experts->nb[2], offset: i*experts->nb[1]); |
| 1119 | |
| 1120 | ggml_build_forward_expand(cgraph: gf, tensor: cur_experts[i]); |
| 1121 | } |
| 1122 | |
| 1123 | // aggregate experts |
| 1124 | // note: here we explicitly use hparams.n_expert_used instead of n_expert_used |
| 1125 | // to avoid potentially a large number of add nodes during warmup |
| 1126 | // ref: https://github.com/ggml-org/llama.cpp/pull/14753 |
| 1127 | ggml_tensor * moe_out = cur_experts[0]; |
| 1128 | |
| 1129 | for (uint32_t i = 1; i < hparams.n_expert_used; ++i) { |
| 1130 | moe_out = ggml_add(ctx: ctx0, a: moe_out, b: cur_experts[i]); |
| 1131 | } |
| 1132 | |
| 1133 | if (hparams.n_expert_used == 1) { |
| 1134 | // avoid returning a non-contiguous tensor |
| 1135 | moe_out = ggml_cont(ctx: ctx0, a: moe_out); |
| 1136 | } |
| 1137 | |
| 1138 | cb(cur: moe_out, name: "ffn_moe_out" , il); |
| 1139 | |
| 1140 | return moe_out; |
| 1141 | } |
| 1142 | |
| 1143 | // input embeddings with optional lora |
| 1144 | ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const { |
| 1145 | const int64_t n_embd = hparams.n_embd_inp(); |
| 1146 | |
| 1147 | auto inp = std::make_unique<llm_graph_input_embd>(); |
| 1148 | |
| 1149 | ggml_tensor * cur = nullptr; |
| 1150 | |
| 1151 | if (ubatch.token) { |
| 1152 | inp->tokens = ggml_new_tensor_1d(ctx: ctx0, type: GGML_TYPE_I32, ne0: ubatch.n_tokens); |
| 1153 | //cb(inp->tokens, "inp_tokens", -1); |
| 1154 | ggml_set_input(tensor: inp->tokens); |
| 1155 | res->t_tokens = inp->tokens; |
| 1156 | |
| 1157 | cur = ggml_get_rows(ctx: ctx0, a: tok_embd, b: inp->tokens); |
| 1158 | |
| 1159 | // apply lora for embedding tokens if needed |
| 1160 | for (const auto & lora : *loras) { |
| 1161 | llama_adapter_lora_weight * lw = lora.first->get_weight(w: tok_embd); |
| 1162 | if (lw == nullptr) { |
| 1163 | continue; |
| 1164 | } |
| 1165 | |
| 1166 | const float adapter_scale = lora.second; |
| 1167 | const float scale = lw->get_scale(alpha: lora.first->alpha, adapter_scale); |
| 1168 | |
| 1169 | ggml_tensor * inpL_delta = ggml_scale(ctx: ctx0, a: ggml_mul_mat( |
| 1170 | ctx: ctx0, a: lw->b, // non-transposed lora_b |
| 1171 | b: ggml_get_rows(ctx: ctx0, a: lw->a, b: inp->tokens) |
| 1172 | ), s: scale); |
| 1173 | |
| 1174 | cur = ggml_add(ctx: ctx0, a: cur, b: inpL_delta); |
| 1175 | } |
| 1176 | } else { |
| 1177 | inp->embd = ggml_new_tensor_2d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_embd, ne1: ubatch.n_tokens); |
| 1178 | ggml_set_input(tensor: inp->embd); |
| 1179 | |
| 1180 | cur = inp->embd; |
| 1181 | } |
| 1182 | |
| 1183 | // For Granite architecture |
| 1184 | if (hparams.f_embedding_scale != 0.0f) { |
| 1185 | cur = ggml_scale(ctx: ctx0, a: cur, s: hparams.f_embedding_scale); |
| 1186 | } |
| 1187 | |
| 1188 | cb(cur, name: "inp_embd" , il: -1); |
| 1189 | |
| 1190 | res->add_input(input: std::move(inp)); |
| 1191 | |
| 1192 | return cur; |
| 1193 | } |
| 1194 | |
| 1195 | ggml_tensor * llm_graph_context::build_inp_pos() const { |
| 1196 | auto inp = std::make_unique<llm_graph_input_pos>(args: hparams.n_pos_per_embd()); |
| 1197 | |
| 1198 | auto & cur = inp->pos; |
| 1199 | |
| 1200 | cur = ggml_new_tensor_1d(ctx: ctx0, type: GGML_TYPE_I32, ne0: (int64_t)n_tokens*hparams.n_pos_per_embd()); |
| 1201 | ggml_set_input(tensor: cur); |
| 1202 | |
| 1203 | res->add_input(input: std::move(inp)); |
| 1204 | |
| 1205 | return cur; |
| 1206 | } |
| 1207 | |
| 1208 | ggml_tensor * llm_graph_context::build_inp_attn_scale() const { |
| 1209 | auto inp = std::make_unique<llm_graph_input_attn_temp>(args: hparams.n_attn_temp_floor_scale, args: hparams.f_attn_temp_scale); |
| 1210 | |
| 1211 | auto & cur = inp->attn_scale; |
| 1212 | |
| 1213 | // this need to be 1x1xN for broadcasting |
| 1214 | cur = ggml_new_tensor_3d(ctx: ctx0, type: GGML_TYPE_F32, ne0: 1, ne1: 1, ne2: n_tokens); |
| 1215 | ggml_set_input(tensor: cur); |
| 1216 | |
| 1217 | res->add_input(input: std::move(inp)); |
| 1218 | |
| 1219 | return cur; |
| 1220 | } |
| 1221 | |
| 1222 | ggml_tensor * llm_graph_context::build_inp_out_ids() const { |
| 1223 | // note: when all tokens are output, we could skip this optimization to spare the ggml_get_rows() calls, |
| 1224 | // but this would make the graph topology depend on the number of output tokens, which can interere with |
| 1225 | // features that require constant topology such as pipline parallelism |
| 1226 | // ref: https://github.com/ggml-org/llama.cpp/pull/14275#issuecomment-2987424471 |
| 1227 | //if (n_outputs < n_tokens) { |
| 1228 | // return nullptr; |
| 1229 | //} |
| 1230 | |
| 1231 | auto inp = std::make_unique<llm_graph_input_out_ids>(args: hparams, args: cparams, args: n_outputs); |
| 1232 | |
| 1233 | auto & cur = inp->out_ids; |
| 1234 | |
| 1235 | cur = ggml_new_tensor_1d(ctx: ctx0, type: GGML_TYPE_I32, ne0: n_outputs); |
| 1236 | ggml_set_input(tensor: cur); |
| 1237 | |
| 1238 | res->add_input(input: std::move(inp)); |
| 1239 | |
| 1240 | return cur; |
| 1241 | } |
| 1242 | |
| 1243 | ggml_tensor * llm_graph_context::build_inp_mean() const { |
| 1244 | auto inp = std::make_unique<llm_graph_input_mean>(args: cparams); |
| 1245 | |
| 1246 | auto & cur = inp->mean; |
| 1247 | |
| 1248 | cur = ggml_new_tensor_2d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_tokens, ne1: ubatch.n_seqs_unq); |
| 1249 | ggml_set_input(tensor: cur); |
| 1250 | |
| 1251 | res->add_input(input: std::move(inp)); |
| 1252 | |
| 1253 | return cur; |
| 1254 | } |
| 1255 | |
| 1256 | ggml_tensor * llm_graph_context::build_inp_cls() const { |
| 1257 | auto inp = std::make_unique<llm_graph_input_cls>(args: cparams, args: arch); |
| 1258 | |
| 1259 | auto & cur = inp->cls; |
| 1260 | |
| 1261 | cur = ggml_new_tensor_1d(ctx: ctx0, type: GGML_TYPE_I32, ne0: ubatch.n_seqs_unq); |
| 1262 | ggml_set_input(tensor: cur); |
| 1263 | |
| 1264 | res->add_input(input: std::move(inp)); |
| 1265 | |
| 1266 | return cur; |
| 1267 | } |
| 1268 | |
| 1269 | ggml_tensor * llm_graph_context::build_inp_cross_embd() const { |
| 1270 | auto inp = std::make_unique<llm_graph_input_cross_embd>(args: cross); |
| 1271 | |
| 1272 | auto & cur = inp->cross_embd; |
| 1273 | |
| 1274 | // if we have the output embeddings from the encoder, use them directly |
| 1275 | // TODO: needs more work to be correct, for now just use the tensor shape |
| 1276 | //if (cross->t_embd) { |
| 1277 | // cur = ggml_view_tensor(ctx0, cross->t_embd); |
| 1278 | |
| 1279 | // return cur; |
| 1280 | //} |
| 1281 | |
| 1282 | const auto n_embd = !cross->v_embd.empty() ? cross->n_embd : hparams.n_embd_inp(); |
| 1283 | const auto n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train; |
| 1284 | |
| 1285 | cur = ggml_new_tensor_2d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_embd, ne1: n_enc); |
| 1286 | ggml_set_input(tensor: cur); |
| 1287 | |
| 1288 | res->add_input(input: std::move(inp)); |
| 1289 | |
| 1290 | return cur; |
| 1291 | } |
| 1292 | |
| 1293 | ggml_tensor * llm_graph_context::build_inp_pos_bucket_enc() const { |
| 1294 | auto inp = std::make_unique<llm_graph_input_pos_bucket>(args: hparams); |
| 1295 | |
| 1296 | auto & cur = inp->pos_bucket; |
| 1297 | |
| 1298 | cur = ggml_new_tensor_2d(ctx: ctx0, type: GGML_TYPE_I32, ne0: n_tokens, ne1: n_tokens); |
| 1299 | ggml_set_input(tensor: cur); |
| 1300 | |
| 1301 | res->add_input(input: std::move(inp)); |
| 1302 | |
| 1303 | return cur; |
| 1304 | } |
| 1305 | |
| 1306 | ggml_tensor * llm_graph_context::build_inp_pos_bucket_dec() const { |
| 1307 | const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx); |
| 1308 | |
| 1309 | auto inp = std::make_unique<llm_graph_input_pos_bucket_kv>(args: hparams, args&: mctx_cur); |
| 1310 | |
| 1311 | const auto n_kv = mctx_cur->get_n_kv(); |
| 1312 | |
| 1313 | auto & cur = inp->pos_bucket; |
| 1314 | |
| 1315 | cur = ggml_new_tensor_2d(ctx: ctx0, type: GGML_TYPE_I32, ne0: n_kv, ne1: n_tokens); |
| 1316 | ggml_set_input(tensor: cur); |
| 1317 | |
| 1318 | res->add_input(input: std::move(inp)); |
| 1319 | |
| 1320 | return cur; |
| 1321 | } |
| 1322 | |
| 1323 | ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const { |
| 1324 | ggml_tensor * pos_bucket_1d = ggml_reshape_1d(ctx: ctx0, a: pos_bucket, ne0: pos_bucket->ne[0] * pos_bucket->ne[1]); |
| 1325 | cb(cur: pos_bucket_1d, name: "pos_bucket_1d" , il: -1); |
| 1326 | |
| 1327 | ggml_tensor * pos_bias = ggml_get_rows(ctx: ctx0, a: attn_rel_b, b: pos_bucket_1d); |
| 1328 | |
| 1329 | pos_bias = ggml_reshape_3d(ctx: ctx0, a: pos_bias, ne0: pos_bias->ne[0], ne1: pos_bucket->ne[0], ne2: pos_bucket->ne[1]); |
| 1330 | pos_bias = ggml_permute (ctx: ctx0, a: pos_bias, axis0: 2, axis1: 0, axis2: 1, axis3: 3); |
| 1331 | pos_bias = ggml_cont (ctx: ctx0, a: pos_bias); |
| 1332 | |
| 1333 | cb(cur: pos_bias, name: "pos_bias" , il: -1); |
| 1334 | |
| 1335 | return pos_bias; |
| 1336 | } |
| 1337 | |
| 1338 | ggml_tensor * llm_graph_context::build_attn_mha( |
| 1339 | ggml_tensor * q, |
| 1340 | ggml_tensor * k, |
| 1341 | ggml_tensor * v, |
| 1342 | ggml_tensor * kq_b, |
| 1343 | ggml_tensor * kq_mask, |
| 1344 | ggml_tensor * sinks, |
| 1345 | ggml_tensor * v_mla, |
| 1346 | float kq_scale, |
| 1347 | int il) const { |
| 1348 | const bool v_trans = v->nb[1] > v->nb[2]; |
| 1349 | |
| 1350 | // split the batch into streams if needed |
| 1351 | const auto n_stream = k->ne[3]; |
| 1352 | |
| 1353 | q = ggml_view_4d(ctx: ctx0, a: q, ne0: q->ne[0], ne1: q->ne[1], ne2: q->ne[2]/n_stream, ne3: n_stream, nb1: q->nb[1], nb2: q->nb[2], nb3: q->nb[3]/n_stream, offset: 0); |
| 1354 | |
| 1355 | q = ggml_permute(ctx: ctx0, a: q, axis0: 0, axis1: 2, axis2: 1, axis3: 3); |
| 1356 | k = ggml_permute(ctx: ctx0, a: k, axis0: 0, axis1: 2, axis2: 1, axis3: 3); |
| 1357 | v = ggml_permute(ctx: ctx0, a: v, axis0: 0, axis1: 2, axis2: 1, axis3: 3); |
| 1358 | |
| 1359 | ggml_tensor * cur; |
| 1360 | |
| 1361 | if (cparams.flash_attn && kq_b == nullptr) { |
| 1362 | GGML_ASSERT(kq_b == nullptr && "Flash attention does not support KQ bias yet" ); |
| 1363 | |
| 1364 | if (v_trans) { |
| 1365 | v = ggml_transpose(ctx: ctx0, a: v); |
| 1366 | } |
| 1367 | |
| 1368 | // this can happen when KV cache is not used (e.g. an embedding model with non-causal attn) |
| 1369 | if (k->type == GGML_TYPE_F32) { |
| 1370 | k = ggml_cast(ctx: ctx0, a: k, type: GGML_TYPE_F16); |
| 1371 | } |
| 1372 | |
| 1373 | if (v->type == GGML_TYPE_F32) { |
| 1374 | v = ggml_cast(ctx: ctx0, a: v, type: GGML_TYPE_F16); |
| 1375 | } |
| 1376 | |
| 1377 | cur = ggml_flash_attn_ext(ctx: ctx0, q, k, v, mask: kq_mask, scale: kq_scale, max_bias: hparams.f_max_alibi_bias, |
| 1378 | logit_softcap: hparams.attn_soft_cap ? hparams.f_attn_logit_softcapping : 0.0f); |
| 1379 | cb(cur, LLAMA_TENSOR_NAME_FATTN, il); |
| 1380 | |
| 1381 | ggml_flash_attn_ext_add_sinks(a: cur, sinks); |
| 1382 | ggml_flash_attn_ext_set_prec (a: cur, prec: GGML_PREC_F32); |
| 1383 | |
| 1384 | if (v_mla) { |
| 1385 | #if 0 |
| 1386 | // v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens. |
| 1387 | // However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient. |
| 1388 | cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens); |
| 1389 | cur = ggml_mul_mat(ctx0, v_mla, cur); |
| 1390 | #else |
| 1391 | // It's preferable to do the calculation as a matrix-matrix multiplication with n_tokens in dimension 1. |
| 1392 | // The permutations are noops and only change how the tensor data is interpreted. |
| 1393 | cur = ggml_permute(ctx: ctx0, a: cur, axis0: 0, axis1: 2, axis2: 1, axis3: 3); |
| 1394 | cur = ggml_mul_mat(ctx: ctx0, a: v_mla, b: cur); |
| 1395 | cb(cur, name: "fattn_mla" , il); |
| 1396 | cur = ggml_permute(ctx: ctx0, a: cur, axis0: 0, axis1: 2, axis2: 1, axis3: 3); |
| 1397 | cur = ggml_cont(ctx: ctx0, a: cur); // Needed because ggml_reshape_2d expects contiguous inputs. |
| 1398 | #endif |
| 1399 | } |
| 1400 | |
| 1401 | cur = ggml_reshape_2d(ctx: ctx0, a: cur, ne0: cur->ne[0]*cur->ne[1], ne1: cur->ne[2]*cur->ne[3]); |
| 1402 | } else { |
| 1403 | ggml_tensor * kq = ggml_mul_mat(ctx: ctx0, a: k, b: q); |
| 1404 | cb(cur: kq, name: "kq" , il); |
| 1405 | |
| 1406 | // note: this op tends to require high floating point range |
| 1407 | // while for some models F16 is enough, for others it is not, so we default to F32 here |
| 1408 | ggml_mul_mat_set_prec(a: kq, prec: GGML_PREC_F32); |
| 1409 | |
| 1410 | if (arch == LLM_ARCH_GROK) { |
| 1411 | // need to do the following: |
| 1412 | // multiply by attn_output_multiplier |
| 1413 | // and then : |
| 1414 | // kq = 30 * tanh(kq / 30) |
| 1415 | // before the softmax below |
| 1416 | |
| 1417 | kq = ggml_tanh(ctx: ctx0, a: ggml_scale(ctx: ctx0, a: kq, s: hparams.f_attn_out_scale / hparams.f_attn_logit_softcapping)); |
| 1418 | cb(cur: kq, name: "kq_tanh" , il); |
| 1419 | kq = ggml_scale(ctx: ctx0, a: kq, s: hparams.f_attn_logit_softcapping); |
| 1420 | cb(cur: kq, name: "kq_scaled" , il); |
| 1421 | } |
| 1422 | |
| 1423 | if (hparams.attn_soft_cap) { |
| 1424 | kq = ggml_scale(ctx: ctx0, a: kq, s: 1.0f / hparams.f_attn_logit_softcapping); |
| 1425 | cb(cur: kq, name: "kq_scaled_1" , il); |
| 1426 | kq = ggml_tanh (ctx: ctx0, a: kq); |
| 1427 | cb(cur: kq, name: "kq_tanh" , il); |
| 1428 | kq = ggml_scale(ctx: ctx0, a: kq, s: hparams.f_attn_logit_softcapping); |
| 1429 | cb(cur: kq, name: "kq_scaled_2" , il); |
| 1430 | } |
| 1431 | |
| 1432 | if (kq_b) { |
| 1433 | kq = ggml_add(ctx: ctx0, a: kq, b: kq_b); |
| 1434 | cb(cur: kq, name: "kq_plus_kq_b" , il); |
| 1435 | } |
| 1436 | |
| 1437 | kq = ggml_soft_max_ext(ctx: ctx0, a: kq, mask: kq_mask, scale: kq_scale, max_bias: hparams.f_max_alibi_bias); |
| 1438 | ggml_soft_max_add_sinks(a: kq, sinks); |
| 1439 | cb(cur: kq, name: "kq_soft_max" , il); |
| 1440 | |
| 1441 | if (!v_trans) { |
| 1442 | // note: avoid this branch |
| 1443 | v = ggml_cont(ctx: ctx0, a: ggml_transpose(ctx: ctx0, a: v)); |
| 1444 | cb(cur: v, name: "v_cont" , il); |
| 1445 | } |
| 1446 | |
| 1447 | ggml_tensor * kqv = ggml_mul_mat(ctx: ctx0, a: v, b: kq); |
| 1448 | cb(cur: kqv, name: "kqv" , il); |
| 1449 | |
| 1450 | // for MLA with the absorption optimization, we need to "decompress" from MQA back to MHA |
| 1451 | if (v_mla) { |
| 1452 | kqv = ggml_mul_mat(ctx: ctx0, a: v_mla, b: kqv); |
| 1453 | cb(cur: kqv, name: "kqv_mla" , il); |
| 1454 | } |
| 1455 | |
| 1456 | cur = ggml_permute(ctx: ctx0, a: kqv, axis0: 0, axis1: 2, axis2: 1, axis3: 3); |
| 1457 | |
| 1458 | // recombine streams |
| 1459 | cur = ggml_cont_2d(ctx: ctx0, a: cur, ne0: cur->ne[0]*cur->ne[1], ne1: cur->ne[2]*cur->ne[3]); |
| 1460 | |
| 1461 | if (!cparams.offload_kqv) { |
| 1462 | // all nodes between the KV store and the attention output are run on the CPU |
| 1463 | ggml_backend_sched_set_tensor_backend(sched, node: cur, backend: backend_cpu); |
| 1464 | } |
| 1465 | } |
| 1466 | |
| 1467 | ggml_build_forward_expand(cgraph: gf, tensor: cur); |
| 1468 | |
| 1469 | return cur; |
| 1470 | } |
| 1471 | |
| 1472 | llm_graph_input_attn_no_cache * llm_graph_context::build_attn_inp_no_cache() const { |
| 1473 | auto inp = std::make_unique<llm_graph_input_attn_no_cache>(args: hparams, args: cparams); |
| 1474 | |
| 1475 | // note: there is no KV cache, so the number of KV values is equal to the number of tokens in the batch |
| 1476 | inp->self_kq_mask = ggml_new_tensor_4d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), ne2: 1, ne3: 1); |
| 1477 | ggml_set_input(tensor: inp->self_kq_mask); |
| 1478 | |
| 1479 | inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx: ctx0, a: inp->self_kq_mask, type: GGML_TYPE_F16) : inp->self_kq_mask; |
| 1480 | |
| 1481 | if (hparams.swa_type != LLAMA_SWA_TYPE_NONE) { |
| 1482 | inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), ne2: 1, ne3: 1); |
| 1483 | ggml_set_input(tensor: inp->self_kq_mask_swa); |
| 1484 | |
| 1485 | inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx: ctx0, a: inp->self_kq_mask_swa, type: GGML_TYPE_F16) : inp->self_kq_mask_swa; |
| 1486 | } else { |
| 1487 | inp->self_kq_mask_swa = nullptr; |
| 1488 | inp->self_kq_mask_swa_cnv = nullptr; |
| 1489 | } |
| 1490 | |
| 1491 | return (llm_graph_input_attn_no_cache *) res->add_input(input: std::move(inp)); |
| 1492 | } |
| 1493 | |
| 1494 | ggml_tensor * llm_graph_context::build_attn( |
| 1495 | llm_graph_input_attn_no_cache * inp, |
| 1496 | ggml_tensor * wo, |
| 1497 | ggml_tensor * wo_b, |
| 1498 | ggml_tensor * q_cur, |
| 1499 | ggml_tensor * k_cur, |
| 1500 | ggml_tensor * v_cur, |
| 1501 | ggml_tensor * kq_b, |
| 1502 | ggml_tensor * sinks, |
| 1503 | ggml_tensor * v_mla, |
| 1504 | float kq_scale, |
| 1505 | int il) const { |
| 1506 | GGML_UNUSED(n_tokens); |
| 1507 | |
| 1508 | // these nodes are added to the graph together so that they are not reordered |
| 1509 | // by doing so, the number of splits in the graph is reduced |
| 1510 | ggml_build_forward_expand(cgraph: gf, tensor: q_cur); |
| 1511 | ggml_build_forward_expand(cgraph: gf, tensor: k_cur); |
| 1512 | ggml_build_forward_expand(cgraph: gf, tensor: v_cur); |
| 1513 | |
| 1514 | const bool is_swa = hparams.is_swa(il); |
| 1515 | |
| 1516 | const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask(); |
| 1517 | |
| 1518 | // [TAG_NO_CACHE_PAD] |
| 1519 | // TODO: if ubatch.equal_seqs() == true, we can split the three tensors below into ubatch.n_seqs_unq streams |
| 1520 | // but it might not be worth it: https://github.com/ggml-org/llama.cpp/pull/15636 |
| 1521 | //assert(!ubatch.equal_seqs() || (k_cur->ne[3] == 1 && k_cur->ne[3] == ubatch.n_seqs_unq)); |
| 1522 | |
| 1523 | ggml_tensor * q = q_cur; |
| 1524 | ggml_tensor * k = k_cur; |
| 1525 | ggml_tensor * v = v_cur; |
| 1526 | |
| 1527 | ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il); |
| 1528 | cb(cur, name: "kqv_out" , il); |
| 1529 | |
| 1530 | if (wo) { |
| 1531 | cur = build_lora_mm(w: wo, cur); |
| 1532 | } |
| 1533 | |
| 1534 | if (wo_b) { |
| 1535 | //cb(cur, "kqv_wo", il); |
| 1536 | } |
| 1537 | |
| 1538 | if (wo_b) { |
| 1539 | cur = ggml_add(ctx: ctx0, a: cur, b: wo_b); |
| 1540 | } |
| 1541 | |
| 1542 | return cur; |
| 1543 | } |
| 1544 | |
| 1545 | static std::unique_ptr<llm_graph_input_attn_kv> build_attn_inp_kv_impl( |
| 1546 | ggml_context * ctx0, |
| 1547 | const llama_ubatch & ubatch, |
| 1548 | const llama_hparams & hparams, |
| 1549 | const llama_cparams & cparams, |
| 1550 | const llama_kv_cache_context * mctx_cur) { |
| 1551 | |
| 1552 | auto inp = std::make_unique<llm_graph_input_attn_kv>(args: hparams, args: cparams, args&: mctx_cur); |
| 1553 | |
| 1554 | { |
| 1555 | GGML_ASSERT(hparams.swa_type == LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache_iswa for SWA" ); |
| 1556 | |
| 1557 | const auto n_kv = mctx_cur->get_n_kv(); |
| 1558 | const auto n_tokens = ubatch.n_tokens; |
| 1559 | const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq; |
| 1560 | |
| 1561 | inp->self_k_idxs = mctx_cur->build_input_k_idxs(ctx: ctx0, ubatch); |
| 1562 | inp->self_v_idxs = mctx_cur->build_input_v_idxs(ctx: ctx0, ubatch); |
| 1563 | |
| 1564 | inp->self_kq_mask = ggml_new_tensor_4d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), ne2: 1, ne3: n_stream); |
| 1565 | ggml_set_input(tensor: inp->self_kq_mask); |
| 1566 | |
| 1567 | inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx: ctx0, a: inp->self_kq_mask, type: GGML_TYPE_F16) : inp->self_kq_mask; |
| 1568 | } |
| 1569 | |
| 1570 | return inp; |
| 1571 | } |
| 1572 | |
| 1573 | llm_graph_input_attn_kv * llm_graph_context::build_attn_inp_kv() const { |
| 1574 | const auto * mctx_cur = static_cast<const llama_kv_cache_context *>(mctx); |
| 1575 | |
| 1576 | auto inp = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur); |
| 1577 | |
| 1578 | return (llm_graph_input_attn_kv *) res->add_input(input: std::move(inp)); |
| 1579 | } |
| 1580 | |
| 1581 | ggml_tensor * llm_graph_context::build_attn( |
| 1582 | llm_graph_input_attn_kv * inp, |
| 1583 | ggml_tensor * wo, |
| 1584 | ggml_tensor * wo_b, |
| 1585 | ggml_tensor * q_cur, |
| 1586 | ggml_tensor * k_cur, |
| 1587 | ggml_tensor * v_cur, |
| 1588 | ggml_tensor * kq_b, |
| 1589 | ggml_tensor * sinks, |
| 1590 | ggml_tensor * v_mla, |
| 1591 | float kq_scale, |
| 1592 | int il) const { |
| 1593 | // these nodes are added to the graph together so that they are not reordered |
| 1594 | // by doing so, the number of splits in the graph is reduced |
| 1595 | ggml_build_forward_expand(cgraph: gf, tensor: q_cur); |
| 1596 | ggml_build_forward_expand(cgraph: gf, tensor: k_cur); |
| 1597 | ggml_build_forward_expand(cgraph: gf, tensor: v_cur); |
| 1598 | |
| 1599 | const auto * mctx_cur = inp->mctx; |
| 1600 | |
| 1601 | // store to KV cache |
| 1602 | { |
| 1603 | const auto & k_idxs = inp->get_k_idxs(); |
| 1604 | const auto & v_idxs = inp->get_v_idxs(); |
| 1605 | |
| 1606 | ggml_build_forward_expand(cgraph: gf, tensor: mctx_cur->cpy_k(ctx: ctx0, k_cur, k_idxs, il)); |
| 1607 | ggml_build_forward_expand(cgraph: gf, tensor: mctx_cur->cpy_v(ctx: ctx0, v_cur, v_idxs, il)); |
| 1608 | } |
| 1609 | |
| 1610 | const auto & kq_mask = inp->get_kq_mask(); |
| 1611 | |
| 1612 | ggml_tensor * q = q_cur; |
| 1613 | ggml_tensor * k = mctx_cur->get_k(ctx: ctx0, il); |
| 1614 | ggml_tensor * v = mctx_cur->get_v(ctx: ctx0, il); |
| 1615 | |
| 1616 | ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il); |
| 1617 | cb(cur, name: "kqv_out" , il); |
| 1618 | |
| 1619 | if (wo) { |
| 1620 | cur = build_lora_mm(w: wo, cur); |
| 1621 | if (arch == LLM_ARCH_GLM4 || arch == LLM_ARCH_GLM4_MOE) { |
| 1622 | // GLM4 and GLM4_MOE seem to have numerical issues with half-precision accumulators |
| 1623 | ggml_mul_mat_set_prec(a: cur, prec: GGML_PREC_F32); |
| 1624 | } |
| 1625 | } |
| 1626 | |
| 1627 | if (wo_b) { |
| 1628 | cur = ggml_add(ctx: ctx0, a: cur, b: wo_b); |
| 1629 | } |
| 1630 | |
| 1631 | return cur; |
| 1632 | } |
| 1633 | |
| 1634 | ggml_tensor * llm_graph_context::build_attn( |
| 1635 | llm_graph_input_attn_kv_iswa * inp, |
| 1636 | ggml_tensor * wo, |
| 1637 | ggml_tensor * wo_b, |
| 1638 | ggml_tensor * q_cur, |
| 1639 | ggml_tensor * k_cur, |
| 1640 | ggml_tensor * v_cur, |
| 1641 | ggml_tensor * kq_b, |
| 1642 | ggml_tensor * sinks, |
| 1643 | ggml_tensor * v_mla, |
| 1644 | float kq_scale, |
| 1645 | int il) const { |
| 1646 | // these nodes are added to the graph together so that they are not reordered |
| 1647 | // by doing so, the number of splits in the graph is reduced |
| 1648 | ggml_build_forward_expand(cgraph: gf, tensor: q_cur); |
| 1649 | |
| 1650 | if (k_cur) { |
| 1651 | ggml_build_forward_expand(cgraph: gf, tensor: k_cur); |
| 1652 | } |
| 1653 | |
| 1654 | if (v_cur) { |
| 1655 | ggml_build_forward_expand(cgraph: gf, tensor: v_cur); |
| 1656 | } |
| 1657 | |
| 1658 | const auto * mctx_iswa = inp->mctx; |
| 1659 | |
| 1660 | const bool is_swa = hparams.is_swa(il); |
| 1661 | |
| 1662 | const auto * mctx_cur = is_swa ? mctx_iswa->get_swa() : mctx_iswa->get_base(); |
| 1663 | |
| 1664 | // optionally store to KV cache |
| 1665 | if (k_cur) { |
| 1666 | const auto & k_idxs = is_swa ? inp->get_k_idxs_swa() : inp->get_k_idxs(); |
| 1667 | |
| 1668 | ggml_build_forward_expand(cgraph: gf, tensor: mctx_cur->cpy_k(ctx: ctx0, k_cur, k_idxs, il)); |
| 1669 | } |
| 1670 | |
| 1671 | if (v_cur) { |
| 1672 | const auto & v_idxs = is_swa ? inp->get_v_idxs_swa() : inp->get_v_idxs(); |
| 1673 | |
| 1674 | ggml_build_forward_expand(cgraph: gf, tensor: mctx_cur->cpy_v(ctx: ctx0, v_cur, v_idxs, il)); |
| 1675 | } |
| 1676 | |
| 1677 | const auto & kq_mask = is_swa ? inp->get_kq_mask_swa() : inp->get_kq_mask(); |
| 1678 | |
| 1679 | ggml_tensor * q = q_cur; |
| 1680 | ggml_tensor * k = mctx_cur->get_k(ctx: ctx0, il); |
| 1681 | ggml_tensor * v = mctx_cur->get_v(ctx: ctx0, il); |
| 1682 | |
| 1683 | ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il); |
| 1684 | cb(cur, name: "kqv_out" , il); |
| 1685 | |
| 1686 | if (wo) { |
| 1687 | cur = build_lora_mm(w: wo, cur); |
| 1688 | } |
| 1689 | |
| 1690 | if (wo_b) { |
| 1691 | //cb(cur, "kqv_wo", il); |
| 1692 | } |
| 1693 | |
| 1694 | if (wo_b) { |
| 1695 | cur = ggml_add(ctx: ctx0, a: cur, b: wo_b); |
| 1696 | } |
| 1697 | |
| 1698 | return cur; |
| 1699 | } |
| 1700 | |
| 1701 | llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const { |
| 1702 | auto inp = std::make_unique<llm_graph_input_attn_cross>(args: cross); |
| 1703 | |
| 1704 | const int32_t n_enc = !cross->v_embd.empty() ? cross->n_enc : hparams.n_ctx_train; |
| 1705 | |
| 1706 | inp->cross_kq_mask = ggml_new_tensor_4d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_enc, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD), ne2: 1, ne3: 1); |
| 1707 | ggml_set_input(tensor: inp->cross_kq_mask); |
| 1708 | |
| 1709 | inp->cross_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx: ctx0, a: inp->cross_kq_mask, type: GGML_TYPE_F16) : inp->cross_kq_mask; |
| 1710 | |
| 1711 | return (llm_graph_input_attn_cross *) res->add_input(input: std::move(inp)); |
| 1712 | } |
| 1713 | |
| 1714 | ggml_tensor * llm_graph_context::build_attn( |
| 1715 | llm_graph_input_attn_cross * inp, |
| 1716 | ggml_tensor * wo, |
| 1717 | ggml_tensor * wo_b, |
| 1718 | ggml_tensor * q_cur, |
| 1719 | ggml_tensor * k_cur, |
| 1720 | ggml_tensor * v_cur, |
| 1721 | ggml_tensor * kq_b, |
| 1722 | ggml_tensor * sinks, |
| 1723 | ggml_tensor * v_mla, |
| 1724 | float kq_scale, |
| 1725 | int il) const { |
| 1726 | // these nodes are added to the graph together so that they are not reordered |
| 1727 | // by doing so, the number of splits in the graph is reduced |
| 1728 | ggml_build_forward_expand(cgraph: gf, tensor: q_cur); |
| 1729 | ggml_build_forward_expand(cgraph: gf, tensor: k_cur); |
| 1730 | ggml_build_forward_expand(cgraph: gf, tensor: v_cur); |
| 1731 | |
| 1732 | const auto & kq_mask = inp->get_kq_mask_cross(); |
| 1733 | |
| 1734 | ggml_tensor * q = q_cur; |
| 1735 | ggml_tensor * k = k_cur; |
| 1736 | ggml_tensor * v = v_cur; |
| 1737 | |
| 1738 | ggml_tensor * cur = build_attn_mha(q, k, v, kq_b, kq_mask, sinks, v_mla, kq_scale, il); |
| 1739 | cb(cur, name: "kqv_out" , il); |
| 1740 | |
| 1741 | if (wo) { |
| 1742 | cur = build_lora_mm(w: wo, cur); |
| 1743 | } |
| 1744 | |
| 1745 | if (wo_b) { |
| 1746 | //cb(cur, "kqv_wo", il); |
| 1747 | } |
| 1748 | |
| 1749 | if (wo_b) { |
| 1750 | cur = ggml_add(ctx: ctx0, a: cur, b: wo_b); |
| 1751 | } |
| 1752 | |
| 1753 | return cur; |
| 1754 | } |
| 1755 | |
| 1756 | // TODO: maybe separate the inner implementation into a separate function |
| 1757 | // like with the non-sliding window equivalent |
| 1758 | // once sliding-window hybrid caches are a thing. |
| 1759 | llm_graph_input_attn_kv_iswa * llm_graph_context::build_attn_inp_kv_iswa() const { |
| 1760 | const auto * mctx_cur = static_cast<const llama_kv_cache_iswa_context *>(mctx); |
| 1761 | |
| 1762 | auto inp = std::make_unique<llm_graph_input_attn_kv_iswa>(args: hparams, args: cparams, args&: mctx_cur); |
| 1763 | |
| 1764 | const auto n_stream = cparams.kv_unified ? 1 : ubatch.n_seqs_unq; |
| 1765 | |
| 1766 | { |
| 1767 | const auto n_kv = mctx_cur->get_base()->get_n_kv(); |
| 1768 | |
| 1769 | inp->self_k_idxs = mctx_cur->get_base()->build_input_k_idxs(ctx: ctx0, ubatch); |
| 1770 | inp->self_v_idxs = mctx_cur->get_base()->build_input_v_idxs(ctx: ctx0, ubatch); |
| 1771 | |
| 1772 | inp->self_kq_mask = ggml_new_tensor_4d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), ne2: 1, ne3: n_stream); |
| 1773 | ggml_set_input(tensor: inp->self_kq_mask); |
| 1774 | |
| 1775 | inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx: ctx0, a: inp->self_kq_mask, type: GGML_TYPE_F16) : inp->self_kq_mask; |
| 1776 | } |
| 1777 | |
| 1778 | { |
| 1779 | GGML_ASSERT(hparams.swa_type != LLAMA_SWA_TYPE_NONE && "Use llama_kv_cache for non-SWA" ); |
| 1780 | |
| 1781 | const auto n_kv = mctx_cur->get_swa()->get_n_kv(); |
| 1782 | |
| 1783 | inp->self_k_idxs_swa = mctx_cur->get_swa()->build_input_k_idxs(ctx: ctx0, ubatch); |
| 1784 | inp->self_v_idxs_swa = mctx_cur->get_swa()->build_input_v_idxs(ctx: ctx0, ubatch); |
| 1785 | |
| 1786 | inp->self_kq_mask_swa = ggml_new_tensor_4d(ctx: ctx0, type: GGML_TYPE_F32, ne0: n_kv, GGML_PAD(n_tokens/n_stream, GGML_KQ_MASK_PAD), ne2: 1, ne3: n_stream); |
| 1787 | ggml_set_input(tensor: inp->self_kq_mask_swa); |
| 1788 | |
| 1789 | inp->self_kq_mask_swa_cnv = cparams.flash_attn ? ggml_cast(ctx: ctx0, a: inp->self_kq_mask_swa, type: GGML_TYPE_F16) : inp->self_kq_mask_swa; |
| 1790 | } |
| 1791 | |
| 1792 | return (llm_graph_input_attn_kv_iswa *) res->add_input(input: std::move(inp)); |
| 1793 | } |
| 1794 | |
| 1795 | ggml_tensor * llm_graph_context::build_rs( |
| 1796 | ggml_tensor * s, |
| 1797 | ggml_tensor * state_copy_main, |
| 1798 | ggml_tensor * , |
| 1799 | int32_t state_size, |
| 1800 | int32_t n_seqs, |
| 1801 | uint32_t n_rs, |
| 1802 | uint32_t rs_head, |
| 1803 | uint32_t rs_size, |
| 1804 | int32_t rs_zero, |
| 1805 | const llm_graph_get_rows_fn & get_state_rows) const { |
| 1806 | |
| 1807 | ggml_tensor * states = ggml_reshape_2d(ctx: ctx0, a: s, ne0: state_size, ne1: rs_size); |
| 1808 | |
| 1809 | // Clear a single state which will then be copied to the other cleared states. |
| 1810 | // Note that this is a no-op when the view is zero-sized. |
| 1811 | ggml_tensor * state_zero = ggml_view_1d(ctx: ctx0, a: states, ne0: state_size*(rs_zero >= 0), offset: rs_zero*states->nb[1]*(rs_zero >= 0)); |
| 1812 | ggml_build_forward_expand(cgraph: gf, tensor: ggml_scale_inplace(ctx: ctx0, a: state_zero, s: 0)); |
| 1813 | |
| 1814 | // copy states |
| 1815 | // NOTE: assuming the copy destinations are ALL contained between rs_head and rs_head + n_rs |
| 1816 | // {state_size, rs_size} -> {state_size, n_seqs} |
| 1817 | ggml_tensor * output_states = get_state_rows(ctx0, states, state_copy_main); |
| 1818 | ggml_build_forward_expand(cgraph: gf, tensor: output_states); |
| 1819 | |
| 1820 | // copy extra states which won't be changed further (between n_seqs and n_rs) |
| 1821 | ggml_tensor * = ggml_get_rows(ctx: ctx0, a: states, b: state_copy_extra); |
| 1822 | ggml_build_forward_expand(cgraph: gf, |
| 1823 | tensor: ggml_cpy(ctx: ctx0, |
| 1824 | a: states_extra, |
| 1825 | b: ggml_view_1d(ctx: ctx0, a: s, ne0: state_size*(n_rs - n_seqs), offset: (rs_head + n_seqs)*state_size*ggml_element_size(tensor: s)))); |
| 1826 | |
| 1827 | return output_states; |
| 1828 | } |
| 1829 | |
| 1830 | static std::unique_ptr<llm_graph_input_rs> build_rs_inp_impl( |
| 1831 | ggml_context * ctx0, |
| 1832 | const llama_ubatch & ubatch, |
| 1833 | const llama_memory_recurrent_context * mctx_cur) { |
| 1834 | |
| 1835 | auto inp = std::make_unique<llm_graph_input_rs>(args&: mctx_cur); |
| 1836 | |
| 1837 | const int64_t n_rs = mctx_cur->get_n_rs(); |
| 1838 | const int64_t n_seqs = ubatch.n_seqs; |
| 1839 | |
| 1840 | inp->s_copy = ggml_new_tensor_1d(ctx: ctx0, type: GGML_TYPE_I32, ne0: n_rs); |
| 1841 | ggml_set_input(tensor: inp->s_copy); |
| 1842 | |
| 1843 | inp->s_copy_main = ggml_view_1d(ctx: ctx0, a: inp->s_copy, ne0: n_seqs, offset: 0); |
| 1844 | inp->s_copy_extra = ggml_view_1d(ctx: ctx0, a: inp->s_copy, ne0: n_rs - n_seqs, offset: n_seqs * inp->s_copy->nb[0]); |
| 1845 | |
| 1846 | return inp; |
| 1847 | } |
| 1848 | |
| 1849 | llm_graph_input_rs * llm_graph_context::build_rs_inp() const { |
| 1850 | const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx); |
| 1851 | |
| 1852 | auto inp = build_rs_inp_impl(ctx0, ubatch, mctx_cur); |
| 1853 | |
| 1854 | return (llm_graph_input_rs *) res->add_input(input: std::move(inp)); |
| 1855 | } |
| 1856 | |
| 1857 | ggml_tensor * llm_graph_context::build_rs( |
| 1858 | llm_graph_input_rs * inp, |
| 1859 | ggml_tensor * s, |
| 1860 | int32_t state_size, |
| 1861 | int32_t n_seqs, |
| 1862 | const llm_graph_get_rows_fn & get_state_rows) const { |
| 1863 | const auto * kv_state = inp->mctx; |
| 1864 | |
| 1865 | return build_rs(s, state_copy_main: inp->s_copy_main, state_copy_extra: inp->s_copy_extra, state_size, n_seqs, |
| 1866 | n_rs: kv_state->get_n_rs(), rs_head: kv_state->get_head(), rs_size: kv_state->get_size(), rs_zero: kv_state->get_rs_z(), |
| 1867 | get_state_rows); |
| 1868 | } |
| 1869 | |
| 1870 | ggml_tensor * llm_graph_context::build_rwkv_token_shift_load( |
| 1871 | llm_graph_input_rs * inp, |
| 1872 | const llama_ubatch & ubatch, |
| 1873 | int il) const { |
| 1874 | const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx); |
| 1875 | |
| 1876 | const auto token_shift_count = hparams.token_shift_count; |
| 1877 | |
| 1878 | const int64_t n_seqs = ubatch.n_seqs; |
| 1879 | |
| 1880 | ggml_tensor * token_shift_all = mctx_cur->get_r_l(il); |
| 1881 | |
| 1882 | ggml_tensor * token_shift = build_rs( |
| 1883 | inp, s: token_shift_all, |
| 1884 | state_size: hparams.n_embd_r(), n_seqs); |
| 1885 | |
| 1886 | token_shift = ggml_reshape_3d(ctx: ctx0, a: token_shift, ne0: hparams.n_embd, ne1: token_shift_count, ne2: n_seqs); |
| 1887 | |
| 1888 | return token_shift; |
| 1889 | } |
| 1890 | |
| 1891 | ggml_tensor * llm_graph_context::build_rwkv_token_shift_store( |
| 1892 | ggml_tensor * token_shift, |
| 1893 | const llama_ubatch & ubatch, |
| 1894 | int il) const { |
| 1895 | const auto * mctx_cur = static_cast<const llama_memory_recurrent_context *>(mctx); |
| 1896 | |
| 1897 | const auto token_shift_count = hparams.token_shift_count; |
| 1898 | const auto n_embd = hparams.n_embd; |
| 1899 | |
| 1900 | const int64_t n_seqs = ubatch.n_seqs; |
| 1901 | |
| 1902 | const auto kv_head = mctx_cur->get_head(); |
| 1903 | |
| 1904 | return ggml_cpy( |
| 1905 | ctx: ctx0, |
| 1906 | a: ggml_view_1d(ctx: ctx0, a: token_shift, ne0: n_embd * n_seqs * token_shift_count, offset: 0), |
| 1907 | b: ggml_view_1d(ctx: ctx0, a: mctx_cur->get_r_l(il), ne0: hparams.n_embd_r()*n_seqs, offset: hparams.n_embd_r()*kv_head*ggml_element_size(tensor: mctx_cur->get_r_l(il))) |
| 1908 | ); |
| 1909 | } |
| 1910 | |
| 1911 | llm_graph_input_mem_hybrid * llm_graph_context::build_inp_mem_hybrid() const { |
| 1912 | const auto * mctx_cur = static_cast<const llama_memory_hybrid_context *>(mctx); |
| 1913 | |
| 1914 | auto inp_rs = build_rs_inp_impl(ctx0, ubatch, mctx_cur: mctx_cur->get_recr()); |
| 1915 | auto inp_attn = build_attn_inp_kv_impl(ctx0, ubatch, hparams, cparams, mctx_cur: mctx_cur->get_attn()); |
| 1916 | |
| 1917 | auto inp = std::make_unique<llm_graph_input_mem_hybrid>(args: std::move(inp_attn), args: std::move(inp_rs), args&: mctx_cur); |
| 1918 | |
| 1919 | return (llm_graph_input_mem_hybrid *) res->add_input(input: std::move(inp)); |
| 1920 | } |
| 1921 | |
| 1922 | void llm_graph_context::build_dense_out( |
| 1923 | ggml_tensor * dense_2, |
| 1924 | ggml_tensor * dense_3) const { |
| 1925 | if (!cparams.embeddings || dense_2 == nullptr || dense_3 == nullptr) { |
| 1926 | return; |
| 1927 | } |
| 1928 | ggml_tensor * cur = res->t_embd_pooled != nullptr ? res->t_embd_pooled : res->t_embd; |
| 1929 | GGML_ASSERT(cur != nullptr && "missing t_embd_pooled/t_embd" ); |
| 1930 | |
| 1931 | cur = ggml_mul_mat(ctx: ctx0, a: dense_2, b: cur); |
| 1932 | cur = ggml_mul_mat(ctx: ctx0, a: dense_3, b: cur); |
| 1933 | cb(cur, name: "result_embd_pooled" , il: -1); |
| 1934 | res->t_embd_pooled = cur; |
| 1935 | ggml_build_forward_expand(cgraph: gf, tensor: cur); |
| 1936 | } |
| 1937 | |
| 1938 | |
| 1939 | void llm_graph_context::build_pooling( |
| 1940 | ggml_tensor * cls, |
| 1941 | ggml_tensor * cls_b, |
| 1942 | ggml_tensor * cls_out, |
| 1943 | ggml_tensor * cls_out_b) const { |
| 1944 | if (!cparams.embeddings) { |
| 1945 | return; |
| 1946 | } |
| 1947 | |
| 1948 | ggml_tensor * inp = res->t_embd; |
| 1949 | |
| 1950 | //// find result_norm tensor for input |
| 1951 | //for (int i = ggml_graph_n_nodes(gf) - 1; i >= 0; --i) { |
| 1952 | // inp = ggml_graph_node(gf, i); |
| 1953 | // if (strcmp(inp->name, "result_norm") == 0 || strcmp(inp->name, "result_embd") == 0) { |
| 1954 | // break; |
| 1955 | // } |
| 1956 | |
| 1957 | // inp = nullptr; |
| 1958 | //} |
| 1959 | |
| 1960 | GGML_ASSERT(inp != nullptr && "missing result_norm/result_embd tensor" ); |
| 1961 | |
| 1962 | ggml_tensor * cur; |
| 1963 | |
| 1964 | switch (pooling_type) { |
| 1965 | case LLAMA_POOLING_TYPE_NONE: |
| 1966 | { |
| 1967 | cur = inp; |
| 1968 | } break; |
| 1969 | case LLAMA_POOLING_TYPE_MEAN: |
| 1970 | { |
| 1971 | ggml_tensor * inp_mean = build_inp_mean(); |
| 1972 | cur = ggml_mul_mat(ctx: ctx0, a: ggml_cont(ctx: ctx0, a: ggml_transpose(ctx: ctx0, a: inp)), b: inp_mean); |
| 1973 | } break; |
| 1974 | case LLAMA_POOLING_TYPE_CLS: |
| 1975 | case LLAMA_POOLING_TYPE_LAST: |
| 1976 | { |
| 1977 | ggml_tensor * inp_cls = build_inp_cls(); |
| 1978 | cur = ggml_get_rows(ctx: ctx0, a: inp, b: inp_cls); |
| 1979 | } break; |
| 1980 | case LLAMA_POOLING_TYPE_RANK: |
| 1981 | { |
| 1982 | ggml_tensor * inp_cls = build_inp_cls(); |
| 1983 | cur = ggml_get_rows(ctx: ctx0, a: inp, b: inp_cls); |
| 1984 | |
| 1985 | // classification head |
| 1986 | // https://github.com/huggingface/transformers/blob/5af7d41e49bbfc8319f462eb45253dcb3863dfb7/src/transformers/models/roberta/modeling_roberta.py#L1566 |
| 1987 | if (cls) { |
| 1988 | cur = ggml_mul_mat(ctx: ctx0, a: cls, b: cur); |
| 1989 | if (cls_b) { |
| 1990 | cur = ggml_add(ctx: ctx0, a: cur, b: cls_b); |
| 1991 | } |
| 1992 | cur = ggml_tanh(ctx: ctx0, a: cur); |
| 1993 | } |
| 1994 | |
| 1995 | // some models don't have `cls_out`, for example: https://huggingface.co/jinaai/jina-reranker-v1-tiny-en |
| 1996 | // https://huggingface.co/jinaai/jina-reranker-v1-tiny-en/blob/cb5347e43979c3084a890e3f99491952603ae1b7/modeling_bert.py#L884-L896 |
| 1997 | // Single layer classification head (direct projection) |
| 1998 | // https://github.com/huggingface/transformers/blob/f4fc42216cd56ab6b68270bf80d811614d8d59e4/src/transformers/models/bert/modeling_bert.py#L1476 |
| 1999 | if (cls_out) { |
| 2000 | cur = ggml_mul_mat(ctx: ctx0, a: cls_out, b: cur); |
| 2001 | if (cls_out_b) { |
| 2002 | cur = ggml_add(ctx: ctx0, a: cur, b: cls_out_b); |
| 2003 | } |
| 2004 | } |
| 2005 | |
| 2006 | // softmax for qwen3 reranker |
| 2007 | if (arch == LLM_ARCH_QWEN3) { |
| 2008 | cur = ggml_soft_max(ctx: ctx0, a: cur); |
| 2009 | } |
| 2010 | } break; |
| 2011 | default: |
| 2012 | { |
| 2013 | GGML_ABORT("unknown pooling type" ); |
| 2014 | } |
| 2015 | } |
| 2016 | |
| 2017 | cb(cur, name: "result_embd_pooled" , il: -1); |
| 2018 | res->t_embd_pooled = cur; |
| 2019 | |
| 2020 | ggml_build_forward_expand(cgraph: gf, tensor: cur); |
| 2021 | } |
| 2022 | |
| 2023 | int32_t llama_relative_position_bucket(llama_pos x, llama_pos y, uint64_t n_buckets, bool bidirectional) { |
| 2024 | // TODO move to hparams if a T5 variant appears that uses a different value |
| 2025 | const int64_t max_distance = 128; |
| 2026 | |
| 2027 | if (bidirectional) { |
| 2028 | n_buckets >>= 1; |
| 2029 | } |
| 2030 | |
| 2031 | const int64_t max_exact = n_buckets >> 1; |
| 2032 | |
| 2033 | int32_t relative_position = x - y; |
| 2034 | int32_t relative_bucket = 0; |
| 2035 | |
| 2036 | if (bidirectional) { |
| 2037 | relative_bucket += (relative_position > 0) * n_buckets; |
| 2038 | relative_position = std::abs(x: relative_position); |
| 2039 | } else { |
| 2040 | relative_position = -std::min<int32_t>(a: relative_position, b: 0); |
| 2041 | } |
| 2042 | |
| 2043 | int32_t relative_position_if_large = floorf(x: max_exact + logf(x: 1.0 * relative_position / max_exact) * (n_buckets - max_exact) / log(x: 1.0 * max_distance / max_exact)); |
| 2044 | relative_position_if_large = std::min<int32_t>(a: relative_position_if_large, b: n_buckets - 1); |
| 2045 | relative_bucket += (relative_position < max_exact ? relative_position : relative_position_if_large); |
| 2046 | |
| 2047 | return relative_bucket; |
| 2048 | } |
| 2049 | |