more aggressive persistent reduction #161055
Closed
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/161055
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ✅ No FailuresAs of commit 9887644 with merge base 18b4fdd ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Gives 18% speedup on rms norm (2048, 32768). And we have seen other instances where inductor is not aggressive enough about codegening persistent reductions - e.g. 39% on [this kernel from torch ao](#159769 (comment)). Codegen-ing persistent reductions can be risky if you run out of registers. Here, I'm effectively making persistent reductions an option of looped reductions by setting RBLOCK == rnumel, so that we can still fallback to looped reductions as needed. As criteria: - there needs to be significant memory savings from doing a persistent reduction (by keeping memory in register and avoiding another iteration over input) - we should not be coalescing on x dimension, otherwise large rblock will inhibit coalescing - we should not be especially register or arithmetic intensive (this last part uses mem_ops_per_thread, but could be improved). Still need to do dashboard run, although I'm not sure we get a lot of large rblock in our benchmarks. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben [ghstack-poisoned]
Gives 18% speedup on rms norm (2048, 32768). And we have seen other instances where inductor is not aggressive enough about codegening persistent reductions - e.g. 39% on [this kernel from torch ao](#159769 (comment)). Codegen-ing persistent reductions can be risky if you run out of registers. Here, I'm effectively making persistent reductions an option of looped reductions by setting RBLOCK == rnumel, so that we can still fallback to looped reductions as needed. As criteria: - there needs to be significant memory savings from doing a persistent reduction (by keeping memory in register and avoiding another iteration over input) - we should not be coalescing on x dimension, otherwise large rblock will inhibit coalescing - we should not be especially register or arithmetic intensive (this last part uses mem_ops_per_thread, but could be improved). Still need to do dashboard run, although I'm not sure we get a lot of large rblock in our benchmarks. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben [ghstack-poisoned]
PaulZhang12
reviewed
Aug 22, 2025
Gives 18% speedup on rms norm (2048, 32768). And we have seen other instances where inductor is not aggressive enough about codegening persistent reductions - e.g. 39% on [this kernel from torch ao](#159769 (comment)). Codegen-ing persistent reductions can be risky if you run out of registers. Here, I'm effectively making persistent reductions an option of looped reductions by setting RBLOCK == rnumel, so that we can still fallback to looped reductions as needed. As criteria: - there needs to be significant memory savings from doing a persistent reduction (by keeping memory in register and avoiding another iteration over input) - we should not be coalescing on x dimension, otherwise large rblock will inhibit coalescing - we should not be especially register or arithmetic intensive (this last part uses mem_ops_per_thread, but could be improved). Still need to do dashboard run, although I'm not sure we get a lot of large rblock in our benchmarks. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben [ghstack-poisoned]
|
@pytorchbot merge |
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
Merge failedReason: Command Details for Dev Infra teamRaised by workflow job |
Gives 18% speedup on rms norm (2048, 32768). And we have seen other instances where inductor is not aggressive enough about codegening persistent reductions - e.g. 39% on [this kernel from torch ao](#159769 (comment)). Codegen-ing persistent reductions can be risky if you run out of registers. Here, I'm effectively making persistent reductions an option of looped reductions by setting RBLOCK == rnumel, so that we can still fallback to looped reductions as needed. As criteria: - there needs to be significant memory savings from doing a persistent reduction (by keeping memory in register and avoiding another iteration over input) - we should not be coalescing on x dimension, otherwise large rblock will inhibit coalescing - we should not be especially register or arithmetic intensive (this last part uses mem_ops_per_thread, but could be improved). Still need to do dashboard run, although I'm not sure we get a lot of large rblock in our benchmarks. cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben [ghstack-poisoned]
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
eellison
changed the title
markc-614
pushed a commit
to markc-614/pytorch
that referenced
this pull request
Sep 17, 2025
Gives 18% speedup on rms norm (2048, 32768). And we have seen other instances where inductor is not aggressive enough about codegening persistent reductions - e.g. 39% on [this kernel from torch ao](pytorch#159769 (comment)). Codegen-ing persistent reductions can be risky if you run out of registers. Here, I'm effectively making persistent reductions an option of looped reductions by setting RBLOCK == rnumel, so that we can still fallback to looped reductions as needed. As criteria: - there needs to be significant memory savings from doing a persistent reduction (by keeping memory in register and avoiding another iteration over input) - we should not be coalescing on x dimension, otherwise large rblock will inhibit coalescing - we should not be especially register or arithmetic intensive (this last part uses mem_ops_per_thread, but could be improved). Still need to do dashboard run, although I'm not sure we get a lot of large rblock in our benchmarks. Pull Request resolved: pytorch#161055 Approved by: https://github.com/jansel
mansiag05
pushed a commit
to mansiag05/pytorch
that referenced
this pull request
Sep 22, 2025
Gives 18% speedup on rms norm (2048, 32768). And we have seen other instances where inductor is not aggressive enough about codegening persistent reductions - e.g. 39% on [this kernel from torch ao](pytorch#159769 (comment)). Codegen-ing persistent reductions can be risky if you run out of registers. Here, I'm effectively making persistent reductions an option of looped reductions by setting RBLOCK == rnumel, so that we can still fallback to looped reductions as needed. As criteria: - there needs to be significant memory savings from doing a persistent reduction (by keeping memory in register and avoiding another iteration over input) - we should not be coalescing on x dimension, otherwise large rblock will inhibit coalescing - we should not be especially register or arithmetic intensive (this last part uses mem_ops_per_thread, but could be improved). Still need to do dashboard run, although I'm not sure we get a lot of large rblock in our benchmarks. Pull Request resolved: pytorch#161055 Approved by: https://github.com/jansel
shunting314
reviewed
Oct 17, 2025
|
|
||
| def __init__(self, *args, dynamic_scale_rblock=True, **kwargs): | ||
| super().__init__(*args, **kwargs) | ||
| self.dynamic_scale_rblock = dynamic_scale_rblock |
There was a problem hiding this comment.
is this used anywhere? @eellison
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Gives 18% speedup on rms norm (2048, 32768). And we have seen other instances where inductor is not aggressive enough about codegening persistent reductions - e.g. 39% on this kernel from torch ao.
Codegen-ing persistent reductions can be risky if you run out of registers. Here, I'm effectively making persistent reductions an option of looped reductions by setting RBLOCK == rnumel, so that we can still fallback to looped reductions as needed.
As criteria:
Still need to do dashboard run, although I'm not sure we get a lot of large rblock in our benchmarks.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @chenyang78 @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben