Reverb/open3dforge
0
1{2 "*": {3 "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm.sparse_submanifold_conv_bwd_input_implicit_gemm_kernel": {4 "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {5 "kwargs": {6 "B1": 64,7 "B2": 64,8 "BK": 649 },10 "num_warps": 2,11 "num_ctas": 1,12 "num_stages": 4,13 "num_buffers_warp_spec": 0,14 "num_consumer_groups": 0,15 "reg_dec_producer": 0,16 "reg_inc_consumer": 0,17 "maxnreg": null,18 "pre_hook": null19 },20 "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {21 "kwargs": {22 "B1": 64,23 "B2": 64,24 "BK": 6425 },26 "num_warps": 2,27 "num_ctas": 1,28 "num_stages": 4,29 "num_buffers_warp_spec": 0,30 "num_consumer_groups": 0,31 "reg_dec_producer": 0,32 "reg_inc_consumer": 0,33 "maxnreg": null,34 "pre_hook": null35 },36 "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {37 "kwargs": {38 "B1": 64,39 "B2": 256,40 "BK": 6441 },42 "num_warps": 4,43 "num_ctas": 1,44 "num_stages": 4,45 "num_buffers_warp_spec": 0,46 "num_consumer_groups": 0,47 "reg_dec_producer": 0,48 "reg_inc_consumer": 0,49 "maxnreg": null,50 "pre_hook": null51 },52 "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {53 "kwargs": {54 "B1": 128,55 "B2": 128,56 "BK": 6457 },58 "num_warps": 4,59 "num_ctas": 1,60 "num_stages": 4,61 "num_buffers_warp_spec": 0,62 "num_consumer_groups": 0,63 "reg_dec_producer": 0,64 "reg_inc_consumer": 0,65 "maxnreg": null,66 "pre_hook": null67 },68 "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {69 "kwargs": {70 "B1": 128,71 "B2": 128,72 "BK": 6473 },74 "num_warps": 4,75 "num_ctas": 1,76 "num_stages": 4,77 "num_buffers_warp_spec": 0,78 "num_consumer_groups": 0,79 "reg_dec_producer": 0,80 "reg_inc_consumer": 0,81 "maxnreg": null,82 "pre_hook": null83 },84 "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {85 "kwargs": {86 "B1": 128,87 "B2": 128,88 "BK": 6489 },90 "num_warps": 4,91 "num_ctas": 1,92 "num_stages": 4,93 "num_buffers_warp_spec": 0,94 "num_consumer_groups": 0,95 "reg_dec_producer": 0,96 "reg_inc_consumer": 0,97 "maxnreg": null,98 "pre_hook": null99 },100 "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {101 "kwargs": {102 "B1": 64,103 "B2": 64,104 "BK": 64105 },106 "num_warps": 2,107 "num_ctas": 1,108 "num_stages": 4,109 "num_buffers_warp_spec": 0,110 "num_consumer_groups": 0,111 "reg_dec_producer": 0,112 "reg_inc_consumer": 0,113 "maxnreg": null,114 "pre_hook": null115 },116 "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {117 "kwargs": {118 "B1": 128,119 "B2": 128,120 "BK": 32121 },122 "num_warps": 4,123 "num_ctas": 1,124 "num_stages": 4,125 "num_buffers_warp_spec": 0,126 "num_consumer_groups": 0,127 "reg_dec_producer": 0,128 "reg_inc_consumer": 0,129 "maxnreg": null,130 "pre_hook": null131 },132 "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {133 "kwargs": {134 "B1": 128,135 "B2": 128,136 "BK": 32137 },138 "num_warps": 4,139 "num_ctas": 1,140 "num_stages": 4,141 "num_buffers_warp_spec": 0,142 "num_consumer_groups": 0,143 "reg_dec_producer": 0,144 "reg_inc_consumer": 0,145 "maxnreg": null,146 "pre_hook": null147 },148 "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {149 "kwargs": {150 "B1": 128,151 "B2": 128,152 "BK": 32153 },154 "num_warps": 4,155 "num_ctas": 1,156 "num_stages": 4,157 "num_buffers_warp_spec": 0,158 "num_consumer_groups": 0,159 "reg_dec_producer": 0,160 "reg_inc_consumer": 0,161 "maxnreg": null,162 "pre_hook": null163 },164 "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {165 "kwargs": {166 "B1": 128,167 "B2": 128,168 "BK": 32169 },170 "num_warps": 4,171 "num_ctas": 1,172 "num_stages": 4,173 "num_buffers_warp_spec": 0,174 "num_consumer_groups": 0,175 "reg_dec_producer": 0,176 "reg_inc_consumer": 0,177 "maxnreg": null,178 "pre_hook": null179 },180 "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {181 "kwargs": {182 "B1": 128,183 "B2": 128,184 "BK": 32185 },186 "num_warps": 4,187 "num_ctas": 1,188 "num_stages": 4,189 "num_buffers_warp_spec": 0,190 "num_consumer_groups": 0,191 "reg_dec_producer": 0,192 "reg_inc_consumer": 0,193 "maxnreg": null,194 "pre_hook": null195 }196 },197 "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm.sparse_submanifold_conv_bwd_weight_implicit_gemm_kernel": {198 "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {199 "kwargs": {200 "B1": 128,201 "B2": 128,202 "BK": 32203 },204 "num_warps": 4,205 "num_ctas": 1,206 "num_stages": 4,207 "num_buffers_warp_spec": 0,208 "num_consumer_groups": 0,209 "reg_dec_producer": 0,210 "reg_inc_consumer": 0,211 "maxnreg": null,212 "pre_hook": null213 },214 "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {215 "kwargs": {216 "B1": 256,217 "B2": 64,218 "BK": 32219 },220 "num_warps": 4,221 "num_ctas": 1,222 "num_stages": 4,223 "num_buffers_warp_spec": 0,224 "num_consumer_groups": 0,225 "reg_dec_producer": 0,226 "reg_inc_consumer": 0,227 "maxnreg": null,228 "pre_hook": null229 },230 "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {231 "kwargs": {232 "B1": 256,233 "B2": 64,234 "BK": 32235 },236 "num_warps": 4,237 "num_ctas": 1,238 "num_stages": 4,239 "num_buffers_warp_spec": 0,240 "num_consumer_groups": 0,241 "reg_dec_producer": 0,242 "reg_inc_consumer": 0,243 "maxnreg": null,244 "pre_hook": null245 },246 "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {247 "kwargs": {248 "B1": 256,249 "B2": 128,250 "BK": 64251 },252 "num_warps": 8,253 "num_ctas": 1,254 "num_stages": 4,255 "num_buffers_warp_spec": 0,256 "num_consumer_groups": 0,257 "reg_dec_producer": 0,258 "reg_inc_consumer": 0,259 "maxnreg": null,260 "pre_hook": null261 },262 "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {263 "kwargs": {264 "B1": 256,265 "B2": 64,266 "BK": 32267 },268 "num_warps": 4,269 "num_ctas": 1,270 "num_stages": 4,271 "num_buffers_warp_spec": 0,272 "num_consumer_groups": 0,273 "reg_dec_producer": 0,274 "reg_inc_consumer": 0,275 "maxnreg": null,276 "pre_hook": null277 },278 "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {279 "kwargs": {280 "B1": 128,281 "B2": 64,282 "BK": 32283 },284 "num_warps": 4,285 "num_ctas": 1,286 "num_stages": 4,287 "num_buffers_warp_spec": 0,288 "num_consumer_groups": 0,289 "reg_dec_producer": 0,290 "reg_inc_consumer": 0,291 "maxnreg": null,292 "pre_hook": null293 },294 "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {295 "kwargs": {296 "B1": 128,297 "B2": 128,298 "BK": 32299 },300 "num_warps": 4,301 "num_ctas": 1,302 "num_stages": 4,303 "num_buffers_warp_spec": 0,304 "num_consumer_groups": 0,305 "reg_dec_producer": 0,306 "reg_inc_consumer": 0,307 "maxnreg": null,308 "pre_hook": null309 },310 "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {311 "kwargs": {312 "B1": 256,313 "B2": 128,314 "BK": 32315 },316 "num_warps": 8,317 "num_ctas": 1,318 "num_stages": 4,319 "num_buffers_warp_spec": 0,320 "num_consumer_groups": 0,321 "reg_dec_producer": 0,322 "reg_inc_consumer": 0,323 "maxnreg": null,324 "pre_hook": null325 },326 "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {327 "kwargs": {328 "B1": 256,329 "B2": 128,330 "BK": 32331 },332 "num_warps": 8,333 "num_ctas": 1,334 "num_stages": 4,335 "num_buffers_warp_spec": 0,336 "num_consumer_groups": 0,337 "reg_dec_producer": 0,338 "reg_inc_consumer": 0,339 "maxnreg": null,340 "pre_hook": null341 },342 "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {343 "kwargs": {344 "B1": 256,345 "B2": 128,346 "BK": 32347 },348 "num_warps": 8,349 "num_ctas": 1,350 "num_stages": 4,351 "num_buffers_warp_spec": 0,352 "num_consumer_groups": 0,353 "reg_dec_producer": 0,354 "reg_inc_consumer": 0,355 "maxnreg": null,356 "pre_hook": null357 },358 "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {359 "kwargs": {360 "B1": 256,361 "B2": 128,362 "BK": 32363 },364 "num_warps": 8,365 "num_ctas": 1,366 "num_stages": 4,367 "num_buffers_warp_spec": 0,368 "num_consumer_groups": 0,369 "reg_dec_producer": 0,370 "reg_inc_consumer": 0,371 "maxnreg": null,372 "pre_hook": null373 },374 "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {375 "kwargs": {376 "B1": 256,377 "B2": 64,378 "BK": 32379 },380 "num_warps": 4,381 "num_ctas": 1,382 "num_stages": 4,383 "num_buffers_warp_spec": 0,384 "num_consumer_groups": 0,385 "reg_dec_producer": 0,386 "reg_inc_consumer": 0,387 "maxnreg": null,388 "pre_hook": null389 }390 },391 "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_implicit_gemm_kernel": {392 "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {393 "kwargs": {394 "B1": 64,395 "B2": 64,396 "BK": 64397 },398 "num_warps": 2,399 "num_ctas": 1,400 "num_stages": 4,401 "num_buffers_warp_spec": 0,402 "num_consumer_groups": 0,403 "reg_dec_producer": 0,404 "reg_inc_consumer": 0,405 "maxnreg": null,406 "pre_hook": null407 },408 "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {409 "kwargs": {410 "B1": 64,411 "B2": 64,412 "BK": 64413 },414 "num_warps": 2,415 "num_ctas": 1,416 "num_stages": 4,417 "num_buffers_warp_spec": 0,418 "num_consumer_groups": 0,419 "reg_dec_producer": 0,420 "reg_inc_consumer": 0,421 "maxnreg": null,422 "pre_hook": null423 },424 "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {425 "kwargs": {426 "B1": 64,427 "B2": 256,428 "BK": 64429 },430 "num_warps": 4,431 "num_ctas": 1,432 "num_stages": 4,433 "num_buffers_warp_spec": 0,434 "num_consumer_groups": 0,435 "reg_dec_producer": 0,436 "reg_inc_consumer": 0,437 "maxnreg": null,438 "pre_hook": null439 },440 "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {441 "kwargs": {442 "B1": 128,443 "B2": 128,444 "BK": 64445 },446 "num_warps": 4,447 "num_ctas": 1,448 "num_stages": 4,449 "num_buffers_warp_spec": 0,450 "num_consumer_groups": 0,451 "reg_dec_producer": 0,452 "reg_inc_consumer": 0,453 "maxnreg": null,454 "pre_hook": null455 },456 "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {457 "kwargs": {458 "B1": 128,459 "B2": 128,460 "BK": 64461 },462 "num_warps": 4,463 "num_ctas": 1,464 "num_stages": 4,465 "num_buffers_warp_spec": 0,466 "num_consumer_groups": 0,467 "reg_dec_producer": 0,468 "reg_inc_consumer": 0,469 "maxnreg": null,470 "pre_hook": null471 },472 "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {473 "kwargs": {474 "B1": 128,475 "B2": 128,476 "BK": 64477 },478 "num_warps": 4,479 "num_ctas": 1,480 "num_stages": 4,481 "num_buffers_warp_spec": 0,482 "num_consumer_groups": 0,483 "reg_dec_producer": 0,484 "reg_inc_consumer": 0,485 "maxnreg": null,486 "pre_hook": null487 },488 "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {489 "kwargs": {490 "B1": 64,491 "B2": 64,492 "BK": 64493 },494 "num_warps": 2,495 "num_ctas": 1,496 "num_stages": 4,497 "num_buffers_warp_spec": 0,498 "num_consumer_groups": 0,499 "reg_dec_producer": 0,500 "reg_inc_consumer": 0,501 "maxnreg": null,502 "pre_hook": null503 },504 "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {505 "kwargs": {506 "B1": 128,507 "B2": 128,508 "BK": 32509 },510 "num_warps": 4,511 "num_ctas": 1,512 "num_stages": 4,513 "num_buffers_warp_spec": 0,514 "num_consumer_groups": 0,515 "reg_dec_producer": 0,516 "reg_inc_consumer": 0,517 "maxnreg": null,518 "pre_hook": null519 },520 "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {521 "kwargs": {522 "B1": 128,523 "B2": 128,524 "BK": 32525 },526 "num_warps": 4,527 "num_ctas": 1,528 "num_stages": 4,529 "num_buffers_warp_spec": 0,530 "num_consumer_groups": 0,531 "reg_dec_producer": 0,532 "reg_inc_consumer": 0,533 "maxnreg": null,534 "pre_hook": null535 },536 "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {537 "kwargs": {538 "B1": 128,539 "B2": 128,540 "BK": 32541 },542 "num_warps": 4,543 "num_ctas": 1,544 "num_stages": 4,545 "num_buffers_warp_spec": 0,546 "num_consumer_groups": 0,547 "reg_dec_producer": 0,548 "reg_inc_consumer": 0,549 "maxnreg": null,550 "pre_hook": null551 },552 "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {553 "kwargs": {554 "B1": 128,555 "B2": 128,556 "BK": 32557 },558 "num_warps": 4,559 "num_ctas": 1,560 "num_stages": 4,561 "num_buffers_warp_spec": 0,562 "num_consumer_groups": 0,563 "reg_dec_producer": 0,564 "reg_inc_consumer": 0,565 "maxnreg": null,566 "pre_hook": null567 },568 "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {569 "kwargs": {570 "B1": 128,571 "B2": 128,572 "BK": 32573 },574 "num_warps": 4,575 "num_ctas": 1,576 "num_stages": 4,577 "num_buffers_warp_spec": 0,578 "num_consumer_groups": 0,579 "reg_dec_producer": 0,580 "reg_inc_consumer": 0,581 "maxnreg": null,582 "pre_hook": null583 }584 },585 "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_implicit_gemm_kernel": {586 "(7, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {587 "kwargs": {588 "B1": 128,589 "B2": 128,590 "BK": 32591 },592 "num_warps": 4,593 "num_ctas": 1,594 "num_stages": 4,595 "num_buffers_warp_spec": 0,596 "num_consumer_groups": 0,597 "reg_dec_producer": 0,598 "reg_inc_consumer": 0,599 "maxnreg": null,600 "pre_hook": null601 },602 "(9, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {603 "kwargs": {604 "B1": 256,605 "B2": 64,606 "BK": 32607 },608 "num_warps": 4,609 "num_ctas": 1,610 "num_stages": 4,611 "num_buffers_warp_spec": 0,612 "num_consumer_groups": 0,613 "reg_dec_producer": 0,614 "reg_inc_consumer": 0,615 "maxnreg": null,616 "pre_hook": null617 },618 "(11, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {619 "kwargs": {620 "B1": 256,621 "B2": 64,622 "BK": 32623 },624 "num_warps": 4,625 "num_ctas": 1,626 "num_stages": 4,627 "num_buffers_warp_spec": 0,628 "num_consumer_groups": 0,629 "reg_dec_producer": 0,630 "reg_inc_consumer": 0,631 "maxnreg": null,632 "pre_hook": null633 },634 "(13, 1024, 1024, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {635 "kwargs": {636 "B1": 256,637 "B2": 128,638 "BK": 64639 },640 "num_warps": 8,641 "num_ctas": 1,642 "num_stages": 4,643 "num_buffers_warp_spec": 0,644 "num_consumer_groups": 0,645 "reg_dec_producer": 0,646 "reg_inc_consumer": 0,647 "maxnreg": null,648 "pre_hook": null649 },650 "(15, 512, 512, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {651 "kwargs": {652 "B1": 256,653 "B2": 64,654 "BK": 32655 },656 "num_warps": 4,657 "num_ctas": 1,658 "num_stages": 4,659 "num_buffers_warp_spec": 0,660 "num_consumer_groups": 0,661 "reg_dec_producer": 0,662 "reg_inc_consumer": 0,663 "maxnreg": null,664 "pre_hook": null665 },666 "(17, 256, 256, 27, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float16')": {667 "kwargs": {668 "B1": 128,669 "B2": 64,670 "BK": 32671 },672 "num_warps": 4,673 "num_ctas": 1,674 "num_stages": 4,675 "num_buffers_warp_spec": 0,676 "num_consumer_groups": 0,677 "reg_dec_producer": 0,678 "reg_inc_consumer": 0,679 "maxnreg": null,680 "pre_hook": null681 },682 "(7, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {683 "kwargs": {684 "B1": 128,685 "B2": 128,686 "BK": 32687 },688 "num_warps": 4,689 "num_ctas": 1,690 "num_stages": 4,691 "num_buffers_warp_spec": 0,692 "num_consumer_groups": 0,693 "reg_dec_producer": 0,694 "reg_inc_consumer": 0,695 "maxnreg": null,696 "pre_hook": null697 },698 "(9, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {699 "kwargs": {700 "B1": 256,701 "B2": 128,702 "BK": 32703 },704 "num_warps": 8,705 "num_ctas": 1,706 "num_stages": 4,707 "num_buffers_warp_spec": 0,708 "num_consumer_groups": 0,709 "reg_dec_producer": 0,710 "reg_inc_consumer": 0,711 "maxnreg": null,712 "pre_hook": null713 },714 "(11, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {715 "kwargs": {716 "B1": 256,717 "B2": 128,718 "BK": 32719 },720 "num_warps": 8,721 "num_ctas": 1,722 "num_stages": 4,723 "num_buffers_warp_spec": 0,724 "num_consumer_groups": 0,725 "reg_dec_producer": 0,726 "reg_inc_consumer": 0,727 "maxnreg": null,728 "pre_hook": null729 },730 "(13, 1024, 1024, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {731 "kwargs": {732 "B1": 256,733 "B2": 128,734 "BK": 32735 },736 "num_warps": 8,737 "num_ctas": 1,738 "num_stages": 4,739 "num_buffers_warp_spec": 0,740 "num_consumer_groups": 0,741 "reg_dec_producer": 0,742 "reg_inc_consumer": 0,743 "maxnreg": null,744 "pre_hook": null745 },746 "(15, 512, 512, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {747 "kwargs": {748 "B1": 256,749 "B2": 128,750 "BK": 32751 },752 "num_warps": 8,753 "num_ctas": 1,754 "num_stages": 4,755 "num_buffers_warp_spec": 0,756 "num_consumer_groups": 0,757 "reg_dec_producer": 0,758 "reg_inc_consumer": 0,759 "maxnreg": null,760 "pre_hook": null761 },762 "(17, 256, 256, 27, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {763 "kwargs": {764 "B1": 256,765 "B2": 64,766 "BK": 32767 },768 "num_warps": 4,769 "num_ctas": 1,770 "num_stages": 4,771 "num_buffers_warp_spec": 0,772 "num_consumer_groups": 0,773 "reg_dec_producer": 0,774 "reg_inc_consumer": 0,775 "maxnreg": null,776 "pre_hook": null777 }778 },779 "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_input_implicit_gemm_splitk_kernel": {780 "(7, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {781 "kwargs": {782 "B1": 64,783 "B2": 64,784 "BK": 64785 },786 "num_warps": 2,787 "num_ctas": 1,788 "num_stages": 4,789 "num_buffers_warp_spec": 0,790 "num_consumer_groups": 0,791 "reg_dec_producer": 0,792 "reg_inc_consumer": 0,793 "maxnreg": null,794 "pre_hook": null795 },796 "(7, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {797 "kwargs": {798 "B1": 64,799 "B2": 64,800 "BK": 64801 },802 "num_warps": 2,803 "num_ctas": 1,804 "num_stages": 4,805 "num_buffers_warp_spec": 0,806 "num_consumer_groups": 0,807 "reg_dec_producer": 0,808 "reg_inc_consumer": 0,809 "maxnreg": null,810 "pre_hook": null811 },812 "(7, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {813 "kwargs": {814 "B1": 64,815 "B2": 256,816 "BK": 64817 },818 "num_warps": 4,819 "num_ctas": 1,820 "num_stages": 4,821 "num_buffers_warp_spec": 0,822 "num_consumer_groups": 0,823 "reg_dec_producer": 0,824 "reg_inc_consumer": 0,825 "maxnreg": null,826 "pre_hook": null827 },828 "(7, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {829 "kwargs": {830 "B1": 64,831 "B2": 256,832 "BK": 64833 },834 "num_warps": 4,835 "num_ctas": 1,836 "num_stages": 4,837 "num_buffers_warp_spec": 0,838 "num_consumer_groups": 0,839 "reg_dec_producer": 0,840 "reg_inc_consumer": 0,841 "maxnreg": null,842 "pre_hook": null843 },844 "(7, 1024, 1024, 27, 64, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {845 "kwargs": {846 "B1": 64,847 "B2": 256,848 "BK": 64849 },850 "num_warps": 4,851 "num_ctas": 1,852 "num_stages": 4,853 "num_buffers_warp_spec": 0,854 "num_consumer_groups": 0,855 "reg_dec_producer": 0,856 "reg_inc_consumer": 0,857 "maxnreg": null,858 "pre_hook": null859 },860 "(7, 1024, 1024, 27, 128, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {861 "kwargs": {862 "B1": 64,863 "B2": 256,864 "BK": 64865 },866 "num_warps": 4,867 "num_ctas": 1,868 "num_stages": 4,869 "num_buffers_warp_spec": 0,870 "num_consumer_groups": 0,871 "reg_dec_producer": 0,872 "reg_inc_consumer": 0,873 "maxnreg": null,874 "pre_hook": null875 },876 "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {877 "kwargs": {878 "B1": 64,879 "B2": 128,880 "BK": 32881 },882 "num_warps": 4,883 "num_ctas": 1,884 "num_stages": 4,885 "num_buffers_warp_spec": 0,886 "num_consumer_groups": 0,887 "reg_dec_producer": 0,888 "reg_inc_consumer": 0,889 "maxnreg": null,890 "pre_hook": null891 },892 "(9, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {893 "kwargs": {894 "B1": 128,895 "B2": 128,896 "BK": 64897 },898 "num_warps": 4,899 "num_ctas": 1,900 "num_stages": 4,901 "num_buffers_warp_spec": 0,902 "num_consumer_groups": 0,903 "reg_dec_producer": 0,904 "reg_inc_consumer": 0,905 "maxnreg": null,906 "pre_hook": null907 },908 "(9, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {909 "kwargs": {910 "B1": 128,911 "B2": 128,912 "BK": 64913 },914 "num_warps": 4,915 "num_ctas": 1,916 "num_stages": 4,917 "num_buffers_warp_spec": 0,918 "num_consumer_groups": 0,919 "reg_dec_producer": 0,920 "reg_inc_consumer": 0,921 "maxnreg": null,922 "pre_hook": null923 },924 "(9, 1024, 1024, 27, 16, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {925 "kwargs": {926 "B1": 128,927 "B2": 128,928 "BK": 64929 },930 "num_warps": 4,931 "num_ctas": 1,932 "num_stages": 4,933 "num_buffers_warp_spec": 0,934 "num_consumer_groups": 0,935 "reg_dec_producer": 0,936 "reg_inc_consumer": 0,937 "maxnreg": null,938 "pre_hook": null939 },940 "(9, 1024, 1024, 27, 32, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {941 "kwargs": {942 "B1": 128,943 "B2": 128,944 "BK": 64945 },946 "num_warps": 4,947 "num_ctas": 1,948 "num_stages": 4,949 "num_buffers_warp_spec": 0,950 "num_consumer_groups": 0,951 "reg_dec_producer": 0,952 "reg_inc_consumer": 0,953 "maxnreg": null,954 "pre_hook": null955 },956 "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {957 "kwargs": {958 "B1": 128,959 "B2": 128,960 "BK": 64961 },962 "num_warps": 4,963 "num_ctas": 1,964 "num_stages": 4,965 "num_buffers_warp_spec": 0,966 "num_consumer_groups": 0,967 "reg_dec_producer": 0,968 "reg_inc_consumer": 0,969 "maxnreg": null,970 "pre_hook": null971 },972 "(11, 1024, 1024, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {973 "kwargs": {974 "B1": 128,975 "B2": 128,976 "BK": 64977 },978 "num_warps": 4,979 "num_ctas": 1,980 "num_stages": 4,981 "num_buffers_warp_spec": 0,982 "num_consumer_groups": 0,983 "reg_dec_producer": 0,984 "reg_inc_consumer": 0,985 "maxnreg": null,986 "pre_hook": null987 },988 "(11, 1024, 1024, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {989 "kwargs": {990 "B1": 128,991 "B2": 128,992 "BK": 64993 },994 "num_warps": 4,995 "num_ctas": 1,996 "num_stages": 4,997 "num_buffers_warp_spec": 0,998 "num_consumer_groups": 0,999 "reg_dec_producer": 0,1000 "reg_inc_consumer": 0,1001 "maxnreg": null,1002 "pre_hook": null1003 },1004 "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1005 "kwargs": {1006 "B1": 128,1007 "B2": 256,1008 "BK": 641009 },1010 "num_warps": 8,1011 "num_ctas": 1,1012 "num_stages": 4,1013 "num_buffers_warp_spec": 0,1014 "num_consumer_groups": 0,1015 "reg_dec_producer": 0,1016 "reg_inc_consumer": 0,1017 "maxnreg": null,1018 "pre_hook": null1019 },1020 "(7, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {1021 "kwargs": {1022 "B1": 64,1023 "B2": 128,1024 "BK": 321025 },1026 "num_warps": 4,1027 "num_ctas": 1,1028 "num_stages": 4,1029 "num_buffers_warp_spec": 0,1030 "num_consumer_groups": 0,1031 "reg_dec_producer": 0,1032 "reg_inc_consumer": 0,1033 "maxnreg": null,1034 "pre_hook": null1035 },1036 "(9, 1024, 1024, 27, 8, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {1037 "kwargs": {1038 "B1": 128,1039 "B2": 128,1040 "BK": 321041 },1042 "num_warps": 4,1043 "num_ctas": 1,1044 "num_stages": 4,1045 "num_buffers_warp_spec": 0,1046 "num_consumer_groups": 0,1047 "reg_dec_producer": 0,1048 "reg_inc_consumer": 0,1049 "maxnreg": null,1050 "pre_hook": null1051 },1052 "(11, 1024, 1024, 27, 4, True, 'torch.float32', 'torch.float32', 'torch.uint32', 'torch.float32')": {1053 "kwargs": {1054 "B1": 64,1055 "B2": 64,1056 "BK": 641057 },1058 "num_warps": 2,1059 "num_ctas": 1,1060 "num_stages": 4,1061 "num_buffers_warp_spec": 0,1062 "num_consumer_groups": 0,1063 "reg_dec_producer": 0,1064 "reg_inc_consumer": 0,1065 "maxnreg": null,1066 "pre_hook": null1067 }1068 },1069 "flex_gemm.kernels.triton.spconv.sparse_submanifold_conv_bwd_implicit_gemm_splitk.sparse_submanifold_conv_bwd_weight_implicit_gemm_splitk_kernel": {1070 "(7, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1071 "kwargs": {1072 "B1": 128,1073 "B2": 64,1074 "BK": 321075 },1076 "num_warps": 2,1077 "num_ctas": 1,1078 "num_stages": 4,1079 "num_buffers_warp_spec": 0,1080 "num_consumer_groups": 0,1081 "reg_dec_producer": 0,1082 "reg_inc_consumer": 0,1083 "maxnreg": null,1084 "pre_hook": null1085 },1086 "(9, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1087 "kwargs": {1088 "B1": 256,1089 "B2": 64,1090 "BK": 321091 },1092 "num_warps": 4,1093 "num_ctas": 1,1094 "num_stages": 4,1095 "num_buffers_warp_spec": 0,1096 "num_consumer_groups": 0,1097 "reg_dec_producer": 0,1098 "reg_inc_consumer": 0,1099 "maxnreg": null,1100 "pre_hook": null1101 },1102 "(11, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1103 "kwargs": {1104 "B1": 256,1105 "B2": 64,1106 "BK": 321107 },1108 "num_warps": 4,1109 "num_ctas": 1,1110 "num_stages": 4,1111 "num_buffers_warp_spec": 0,1112 "num_consumer_groups": 0,1113 "reg_dec_producer": 0,1114 "reg_inc_consumer": 0,1115 "maxnreg": null,1116 "pre_hook": null1117 },1118 "(13, 1024, 1024, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1119 "kwargs": {1120 "B1": 64,1121 "B2": 64,1122 "BK": 321123 },1124 "num_warps": 2,1125 "num_ctas": 1,1126 "num_stages": 4,1127 "num_buffers_warp_spec": 0,1128 "num_consumer_groups": 0,1129 "reg_dec_producer": 0,1130 "reg_inc_consumer": 0,1131 "maxnreg": null,1132 "pre_hook": null1133 },1134 "(15, 512, 512, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1135 "kwargs": {1136 "B1": 256,1137 "B2": 64,1138 "BK": 321139 },1140 "num_warps": 4,1141 "num_ctas": 1,1142 "num_stages": 4,1143 "num_buffers_warp_spec": 0,1144 "num_consumer_groups": 0,1145 "reg_dec_producer": 0,1146 "reg_inc_consumer": 0,1147 "maxnreg": null,1148 "pre_hook": null1149 },1150 "(15, 512, 512, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1151 "kwargs": {1152 "B1": 256,1153 "B2": 64,1154 "BK": 321155 },1156 "num_warps": 4,1157 "num_ctas": 1,1158 "num_stages": 4,1159 "num_buffers_warp_spec": 0,1160 "num_consumer_groups": 0,1161 "reg_dec_producer": 0,1162 "reg_inc_consumer": 0,1163 "maxnreg": null,1164 "pre_hook": null1165 },1166 "(15, 512, 512, 27, 8, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1167 "kwargs": {1168 "B1": 256,1169 "B2": 128,1170 "BK": 641171 },1172 "num_warps": 8,1173 "num_ctas": 1,1174 "num_stages": 4,1175 "num_buffers_warp_spec": 0,1176 "num_consumer_groups": 0,1177 "reg_dec_producer": 0,1178 "reg_inc_consumer": 0,1179 "maxnreg": null,1180 "pre_hook": null1181 },1182 "(17, 256, 256, 27, 2, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1183 "kwargs": {1184 "B1": 128,1185 "B2": 128,1186 "BK": 321187 },1188 "num_warps": 4,1189 "num_ctas": 1,1190 "num_stages": 4,1191 "num_buffers_warp_spec": 0,1192 "num_consumer_groups": 0,1193 "reg_dec_producer": 0,1194 "reg_inc_consumer": 0,1195 "maxnreg": null,1196 "pre_hook": null1197 },1198 "(17, 256, 256, 27, 4, True, 'torch.float16', 'torch.float16', 'torch.uint32', 'torch.float32')": {1199 "kwargs": {1200 "B1": 128,