[inductor] Reduce block sizes when using Triton CPU backend #136612
Conversation
This greatly reduces compile time. [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/136612
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 1 Unrelated FailureAs of commit c418224 with merge base 6966811 ( NEW FAILURE - The following job has failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This greatly reduces compile time. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
This greatly reduces compile time. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
This greatly reduces compile time. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
This greatly reduces compile time. ghstack-source-id: 0144cdd Pull Request resolved: pytorch/pytorch#136612
This greatly reduces compile time. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
This greatly reduces compile time. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
This greatly reduces compile time; TorchBench models that were previously 50-100x slower (vs the cpp backend) are now ~20x slower. More work needs to be done on the Triton side, but smaller block sizes will still be helpful. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy yf225 chenyang78 kadeng muchulee8 ColinPeppler amjames desertfire chauhang [ghstack-poisoned]
|
@pytorchbot merge -f "test failure looks unrelated" |
Merge startedYour change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes). Please use Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
|
Are you primarily running with max-autotune? |
|
Yeah, the compile time measurements are using max-autotune. |
|
Perhaps it's better to just turn off max-autotune? The biggest impact it has is in codegening triton matmuls (and autotuning those). If you just want to search better block sizes perhaps it'd be better to try coordinate descent tuning? pytorch/torch/_inductor/config.py Line 346 in 7dc1788 |
|
Oh interesting. Switching from max-autotune to coordinate descent tuning makes 60% of the models faster at runtime. Compile times improve for just above 50% of cases, but there are also major regressions for some models... |
4 participants
Stack from ghstack (oldest at bottom):
This greatly reduces compile time; TorchBench models that were previously 50-100x slower (vs the cpp backend) are now ~20x slower. More work needs to be done on the Triton side, but smaller block sizes will still be helpful.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang