| 1 | # Sparse MLA benchmark: forward_mha vs forward_mqa |
| 2 | # |
| 3 | # Usage: |
| 4 | # python benchmark.py --config configs/mla_sparse_mha_vs_mqa.yaml |
| 5 | # |
| 6 | # Heatmap grid: |
| 7 | # - batch_size: 1, 2, 4, 8, 16, 32 |
| 8 | # - seq_len: 32, 64, 128, 256, 512, 1024, 2048 |
| 9 | # - q_len: powers of two through seq_len |
| 10 | # |
| 11 | # Specs with q_len < seq_len include context; the q_len == seq_len diagonal |
| 12 | # covers pure prefill. |
| 13 | # The model shape below is the DP case. For the TP8 run, manually change |
| 14 | # model.num_q_heads from 128 to 16 before rerunning this benchmark. |
| 15 | |
| 16 | mode: mha_vs_mqa |
| 17 | |
| 18 | model: |
| 19 | name: "deepseek-v3" |
| 20 | num_layers: 60 |
| 21 | num_q_heads: 128 |
| 22 | num_kv_heads: 1 |
| 23 | head_dim: 576 |
| 24 | kv_lora_rank: 512 |
| 25 | qk_nope_head_dim: 128 |
| 26 | qk_rope_head_dim: 64 |
| 27 | v_head_dim: 128 |
| 28 | block_size: 128 |
| 29 | max_model_len: 2048 |
| 30 | |
| 31 | batch_specs: |
| 32 | # Batch size 1 |
| 33 | # seq_len = 32 |
| 34 | - "1q1s32" |
| 35 | - "1q2s32" |
| 36 | - "1q4s32" |
| 37 | - "1q8s32" |
| 38 | - "1q16s32" |
| 39 | - "1q32" |
| 40 | # seq_len = 64 |
| 41 | - "1q1s64" |
| 42 | - "1q2s64" |
| 43 | - "1q4s64" |
| 44 | - "1q8s64" |
| 45 | - "1q16s64" |
| 46 | - "1q32s64" |
| 47 | - "1q64" |
| 48 | # seq_len = 128 |
| 49 | - "1q1s128" |
| 50 | - "1q2s128" |
| 51 | - "1q4s128" |
| 52 | - "1q8s128" |
| 53 | - "1q16s128" |
| 54 | - "1q32s128" |
| 55 | - "1q64s128" |
| 56 | - "1q128" |
| 57 | # seq_len = 256 |
| 58 | - "1q1s256" |
| 59 | - "1q2s256" |
| 60 | - "1q4s256" |
| 61 | - "1q8s256" |
| 62 | - "1q16s256" |
| 63 | - "1q32s256" |
| 64 | - "1q64s256" |
| 65 | - "1q128s256" |
| 66 | - "1q256" |
| 67 | # seq_len = 512 |
| 68 | - "1q1s512" |
| 69 | - "1q2s512" |
| 70 | - "1q4s512" |
| 71 | - "1q8s512" |
| 72 | - "1q16s512" |
| 73 | - "1q32s512" |
| 74 | - "1q64s512" |
| 75 | - "1q128s512" |
| 76 | - "1q256s512" |
| 77 | - "1q512" |
| 78 | # seq_len = 1024 |
| 79 | - "1q1s1024" |
| 80 | - "1q2s1024" |
| 81 | - "1q4s1024" |
| 82 | - "1q8s1024" |
| 83 | - "1q16s1024" |
| 84 | - "1q32s1024" |
| 85 | - "1q64s1024" |
| 86 | - "1q128s1024" |
| 87 | - "1q256s1024" |
| 88 | - "1q512s1024" |
| 89 | - "1q1024" |
| 90 | # seq_len = 2048 |
| 91 | - "1q1s2048" |
| 92 | - "1q2s2048" |
| 93 | - "1q4s2048" |
| 94 | - "1q8s2048" |
| 95 | - "1q16s2048" |
| 96 | - "1q32s2048" |
| 97 | - "1q64s2048" |
| 98 | - "1q128s2048" |
| 99 | - "1q256s2048" |
| 100 | - "1q512s2048" |
| 101 | - "1q1024s2048" |
| 102 | - "1q2048" |
| 103 | |
| 104 | # Batch size 2 |
| 105 | # seq_len = 32 |
| 106 | - "2q1s32" |
| 107 | - "2q2s32" |
| 108 | - "2q4s32" |
| 109 | - "2q8s32" |
| 110 | - "2q16s32" |
| 111 | - "2q32" |
| 112 | # seq_len = 64 |
| 113 | - "2q1s64" |
| 114 | - "2q2s64" |
| 115 | - "2q4s64" |
| 116 | - "2q8s64" |
| 117 | - "2q16s64" |
| 118 | - "2q32s64" |
| 119 | - "2q64" |
| 120 | # seq_len = 128 |
| 121 | - "2q1s128" |
| 122 | - "2q2s128" |
| 123 | - "2q4s128" |
| 124 | - "2q8s128" |
| 125 | - "2q16s128" |
| 126 | - "2q32s128" |
| 127 | - "2q64s128" |
| 128 | - "2q128" |
| 129 | # seq_len = 256 |
| 130 | - "2q1s256" |
| 131 | - "2q2s256" |
| 132 | - "2q4s256" |
| 133 | - "2q8s256" |
| 134 | - "2q16s256" |
| 135 | - "2q32s256" |
| 136 | - "2q64s256" |
| 137 | - "2q128s256" |
| 138 | - "2q256" |
| 139 | # seq_len = 512 |
| 140 | - "2q1s512" |
| 141 | - "2q2s512" |
| 142 | - "2q4s512" |
| 143 | - "2q8s512" |
| 144 | - "2q16s512" |
| 145 | - "2q32s512" |
| 146 | - "2q64s512" |
| 147 | - "2q128s512" |
| 148 | - "2q256s512" |
| 149 | - "2q512" |
| 150 | # seq_len = 1024 |
| 151 | - "2q1s1024" |
| 152 | - "2q2s1024" |
| 153 | - "2q4s1024" |
| 154 | - "2q8s1024" |
| 155 | - "2q16s1024" |
| 156 | - "2q32s1024" |
| 157 | - "2q64s1024" |
| 158 | - "2q128s1024" |
| 159 | - "2q256s1024" |
| 160 | - "2q512s1024" |
| 161 | - "2q1024" |
| 162 | # seq_len = 2048 |
| 163 | - "2q1s2048" |
| 164 | - "2q2s2048" |
| 165 | - "2q4s2048" |
| 166 | - "2q8s2048" |
| 167 | - "2q16s2048" |
| 168 | - "2q32s2048" |
| 169 | - "2q64s2048" |
| 170 | - "2q128s2048" |
| 171 | - "2q256s2048" |
| 172 | - "2q512s2048" |
| 173 | - "2q1024s2048" |
| 174 | - "2q2048" |
| 175 | |
| 176 | # Batch size 4 |
| 177 | # seq_len = 32 |
| 178 | - "4q1s32" |
| 179 | - "4q2s32" |
| 180 | - "4q4s32" |
| 181 | - "4q8s32" |
| 182 | - "4q16s32" |
| 183 | - "4q32" |
| 184 | # seq_len = 64 |
| 185 | - "4q1s64" |
| 186 | - "4q2s64" |
| 187 | - "4q4s64" |
| 188 | - "4q8s64" |
| 189 | - "4q16s64" |
| 190 | - "4q32s64" |
| 191 | - "4q64" |
| 192 | # seq_len = 128 |
| 193 | - "4q1s128" |
| 194 | - "4q2s128" |
| 195 | - "4q4s128" |
| 196 | - "4q8s128" |
| 197 | - "4q16s128" |
| 198 | - "4q32s128" |
| 199 | - "4q64s128" |
| 200 | - "4q128" |
| 201 | # seq_len = 256 |
| 202 | - "4q1s256" |
| 203 | - "4q2s256" |
| 204 | - "4q4s256" |
| 205 | - "4q8s256" |
| 206 | - "4q16s256" |
| 207 | - "4q32s256" |
| 208 | - "4q64s256" |
| 209 | - "4q128s256" |
| 210 | - "4q256" |
| 211 | # seq_len = 512 |
| 212 | - "4q1s512" |
| 213 | - "4q2s512" |
| 214 | - "4q4s512" |
| 215 | - "4q8s512" |
| 216 | - "4q16s512" |
| 217 | - "4q32s512" |
| 218 | - "4q64s512" |
| 219 | - "4q128s512" |
| 220 | - "4q256s512" |
| 221 | - "4q512" |
| 222 | # seq_len = 1024 |
| 223 | - "4q1s1024" |
| 224 | - "4q2s1024" |
| 225 | - "4q4s1024" |
| 226 | - "4q8s1024" |
| 227 | - "4q16s1024" |
| 228 | - "4q32s1024" |
| 229 | - "4q64s1024" |
| 230 | - "4q128s1024" |
| 231 | - "4q256s1024" |
| 232 | - "4q512s1024" |
| 233 | - "4q1024" |
| 234 | # seq_len = 2048 |
| 235 | - "4q1s2048" |
| 236 | - "4q2s2048" |
| 237 | - "4q4s2048" |
| 238 | - "4q8s2048" |
| 239 | - "4q16s2048" |
| 240 | - "4q32s2048" |
| 241 | - "4q64s2048" |
| 242 | - "4q128s2048" |
| 243 | - "4q256s2048" |
| 244 | - "4q512s2048" |
| 245 | - "4q1024s2048" |
| 246 | - "4q2048" |
| 247 | |
| 248 | # Batch size 8 |
| 249 | # seq_len = 32 |
| 250 | - "8q1s32" |
| 251 | - "8q2s32" |
| 252 | - "8q4s32" |
| 253 | - "8q8s32" |
| 254 | - "8q16s32" |
| 255 | - "8q32" |
| 256 | # seq_len = 64 |
| 257 | - "8q1s64" |
| 258 | - "8q2s64" |
| 259 | - "8q4s64" |
| 260 | - "8q8s64" |
| 261 | - "8q16s64" |
| 262 | - "8q32s64" |
| 263 | - "8q64" |
| 264 | # seq_len = 128 |
| 265 | - "8q1s128" |
| 266 | - "8q2s128" |
| 267 | - "8q4s128" |
| 268 | - "8q8s128" |
| 269 | - "8q16s128" |
| 270 | - "8q32s128" |
| 271 | - "8q64s128" |
| 272 | - "8q128" |
| 273 | # seq_len = 256 |
| 274 | - "8q1s256" |
| 275 | - "8q2s256" |
| 276 | - "8q4s256" |
| 277 | - "8q8s256" |
| 278 | - "8q16s256" |
| 279 | - "8q32s256" |
| 280 | - "8q64s256" |
| 281 | - "8q128s256" |
| 282 | - "8q256" |
| 283 | # seq_len = 512 |
| 284 | - "8q1s512" |
| 285 | - "8q2s512" |
| 286 | - "8q4s512" |
| 287 | - "8q8s512" |
| 288 | - "8q16s512" |
| 289 | - "8q32s512" |
| 290 | - "8q64s512" |
| 291 | - "8q128s512" |
| 292 | - "8q256s512" |
| 293 | - "8q512" |
| 294 | # seq_len = 1024 |
| 295 | - "8q1s1024" |
| 296 | - "8q2s1024" |
| 297 | - "8q4s1024" |
| 298 | - "8q8s1024" |
| 299 | - "8q16s1024" |
| 300 | - "8q32s1024" |
| 301 | - "8q64s1024" |
| 302 | - "8q128s1024" |
| 303 | - "8q256s1024" |
| 304 | - "8q512s1024" |
| 305 | - "8q1024" |
| 306 | # seq_len = 2048 |
| 307 | - "8q1s2048" |
| 308 | - "8q2s2048" |
| 309 | - "8q4s2048" |
| 310 | - "8q8s2048" |
| 311 | - "8q16s2048" |
| 312 | - "8q32s2048" |
| 313 | - "8q64s2048" |
| 314 | - "8q128s2048" |
| 315 | - "8q256s2048" |
| 316 | - "8q512s2048" |
| 317 | - "8q1024s2048" |
| 318 | - "8q2048" |
| 319 | |
| 320 | # Batch size 16 |
| 321 | # seq_len = 32 |
| 322 | - "16q1s32" |
| 323 | - "16q2s32" |
| 324 | - "16q4s32" |
| 325 | - "16q8s32" |
| 326 | - "16q16s32" |
| 327 | - "16q32" |
| 328 | # seq_len = 64 |
| 329 | - "16q1s64" |
| 330 | - "16q2s64" |
| 331 | - "16q4s64" |
| 332 | - "16q8s64" |
| 333 | - "16q16s64" |
| 334 | - "16q32s64" |
| 335 | - "16q64" |
| 336 | # seq_len = 128 |
| 337 | - "16q1s128" |
| 338 | - "16q2s128" |
| 339 | - "16q4s128" |
| 340 | - "16q8s128" |
| 341 | - "16q16s128" |
| 342 | - "16q32s128" |
| 343 | - "16q64s128" |
| 344 | - "16q128" |
| 345 | # seq_len = 256 |
| 346 | - "16q1s256" |
| 347 | - "16q2s256" |
| 348 | - "16q4s256" |
| 349 | - "16q8s256" |
| 350 | - "16q16s256" |
| 351 | - "16q32s256" |
| 352 | - "16q64s256" |
| 353 | - "16q128s256" |
| 354 | - "16q256" |
| 355 | # seq_len = 512 |
| 356 | - "16q1s512" |
| 357 | - "16q2s512" |
| 358 | - "16q4s512" |
| 359 | - "16q8s512" |
| 360 | - "16q16s512" |
| 361 | - "16q32s512" |
| 362 | - "16q64s512" |
| 363 | - "16q128s512" |
| 364 | - "16q256s512" |
| 365 | - "16q512" |
| 366 | # seq_len = 1024 |
| 367 | - "16q1s1024" |
| 368 | - "16q2s1024" |
| 369 | - "16q4s1024" |
| 370 | - "16q8s1024" |
| 371 | - "16q16s1024" |
| 372 | - "16q32s1024" |
| 373 | - "16q64s1024" |
| 374 | - "16q128s1024" |
| 375 | - "16q256s1024" |
| 376 | - "16q512s1024" |
| 377 | - "16q1024" |
| 378 | # seq_len = 2048 |
| 379 | - "16q1s2048" |
| 380 | - "16q2s2048" |
| 381 | - "16q4s2048" |
| 382 | - "16q8s2048" |
| 383 | - "16q16s2048" |
| 384 | - "16q32s2048" |
| 385 | - "16q64s2048" |
| 386 | - "16q128s2048" |
| 387 | - "16q256s2048" |
| 388 | - "16q512s2048" |
| 389 | - "16q1024s2048" |
| 390 | - "16q2048" |
| 391 | |
| 392 | # Batch size 32 |
| 393 | # seq_len = 32 |
| 394 | - "32q1s32" |
| 395 | - "32q2s32" |
| 396 | - "32q4s32" |
| 397 | - "32q8s32" |
| 398 | - "32q16s32" |
| 399 | - "32q32" |
| 400 | # seq_len = 64 |
| 401 | - "32q1s64" |
| 402 | - "32q2s64" |
| 403 | - "32q4s64" |
| 404 | - "32q8s64" |
| 405 | - "32q16s64" |
| 406 | - "32q32s64" |
| 407 | - "32q64" |
| 408 | # seq_len = 128 |
| 409 | - "32q1s128" |
| 410 | - "32q2s128" |
| 411 | - "32q4s128" |
| 412 | - "32q8s128" |
| 413 | - "32q16s128" |
| 414 | - "32q32s128" |
| 415 | - "32q64s128" |
| 416 | - "32q128" |
| 417 | # seq_len = 256 |
| 418 | - "32q1s256" |
| 419 | - "32q2s256" |
| 420 | - "32q4s256" |
| 421 | - "32q8s256" |
| 422 | - "32q16s256" |
| 423 | - "32q32s256" |
| 424 | - "32q64s256" |
| 425 | - "32q128s256" |
| 426 | - "32q256" |
| 427 | # seq_len = 512 |
| 428 | - "32q1s512" |
| 429 | - "32q2s512" |
| 430 | - "32q4s512" |
| 431 | - "32q8s512" |
| 432 | - "32q16s512" |
| 433 | - "32q32s512" |
| 434 | - "32q64s512" |
| 435 | - "32q128s512" |
| 436 | - "32q256s512" |
| 437 | - "32q512" |
| 438 | # seq_len = 1024 |
| 439 | - "32q1s1024" |
| 440 | - "32q2s1024" |
| 441 | - "32q4s1024" |
| 442 | - "32q8s1024" |
| 443 | - "32q16s1024" |
| 444 | - "32q32s1024" |
| 445 | - "32q64s1024" |
| 446 | - "32q128s1024" |
| 447 | - "32q256s1024" |
| 448 | - "32q512s1024" |
| 449 | - "32q1024" |
| 450 | # seq_len = 2048 |
| 451 | - "32q1s2048" |
| 452 | - "32q2s2048" |
| 453 | - "32q4s2048" |
| 454 | - "32q8s2048" |
| 455 | - "32q16s2048" |
| 456 | - "32q32s2048" |
| 457 | - "32q64s2048" |
| 458 | - "32q128s2048" |
| 459 | - "32q256s2048" |
| 460 | - "32q512s2048" |
| 461 | - "32q1024s2048" |
| 462 | - "32q2048" |
| 463 | |
| 464 | backends: |
| 465 | - FLASHMLA_SPARSE |
| 466 | |
| 467 | device: "cuda:0" |
| 468 | profile_memory: false |
| 469 | sparse_mla_dense_mha_max_seq_len: 2048 |
| 470 | sparse_mla_topk_pattern: "random" |
| 471 | |
| 472 | output: |
| 473 | csv: "benchmark_output/mla_sparse_mha_vs_mqa.csv" |
| 474 | json: "benchmark_output/mla_sparse_mha_vs_mqa.json" |