[dynamo] Rehaul the autograd.Function support#166788
Closed
anijain2305 wants to merge 51 commits intogh/anijain2305/940/basefrom
Closed
[dynamo] Rehaul the autograd.Function support#166788anijain2305 wants to merge 51 commits intogh/anijain2305/940/basefrom
anijain2305 wants to merge 51 commits intogh/anijain2305/940/basefrom
Conversation
[ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
anijain2305
added
the
keep-going
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela [ghstack-poisoned]
umechand-amd
pushed a commit
to ROCm/pytorch
that referenced
this pull request
Dec 8, 2025
We make a rehaul because (1) we want to support non-proxyable outputs in the fwd method (2) we saw general softness in the support. I have put lot of comments in the code. Follow up * Graph break on backward stride dependent computation. This is BC breaking, so needs care. * Use DynamoAutogradFunctionTraceHelper for backward tracer. * Add test cases for module input, pytree input/outputs, pg groups * Consider unifying `automatic` and `automatic_with_forced_placeholders` * Better error messages - especially nonstrict trace for stride dependent backward computation. Pull Request resolved: pytorch#166788 Approved by: https://github.com/zou3519
umechand-amd
pushed a commit
to ROCm/pytorch
that referenced
this pull request
Dec 8, 2025
…d graph tracing (pytorch#169399)" This reverts commit f874542. Reverted pytorch#169399 on behalf of https://github.com/huydhn due to Sorry for reverting your change but I need to revert this to revert pytorch#166788 ([comment](pytorch#169399 (comment)))
umechand-amd
pushed a commit
to ROCm/pytorch
that referenced
this pull request
Dec 8, 2025
This reverts commit a84798e. Reverted pytorch#166788 on behalf of https://github.com/huydhn due to Sorry for reverting your change but it seems to cause some numerical error in trunk ([comment](pytorch#166788 (comment)))
JacobSzwejbka
pushed a commit
that referenced
this pull request
Dec 8, 2025
We make a rehaul because (1) we want to support non-proxyable outputs in the fwd method (2) we saw general softness in the support. I have put lot of comments in the code. Follow up * Graph break on backward stride dependent computation. This is BC breaking, so needs care. * Use DynamoAutogradFunctionTraceHelper for backward tracer. * Add test cases for module input, pytree input/outputs, pg groups * Consider unifying `automatic` and `automatic_with_forced_placeholders` * Better error messages - especially nonstrict trace for stride dependent backward computation. Pull Request resolved: #166788 Approved by: https://github.com/zou3519
JacobSzwejbka
pushed a commit
that referenced
this pull request
Dec 8, 2025
…d graph tracing (#169399)" This reverts commit f874542. Reverted #169399 on behalf of https://github.com/huydhn due to Sorry for reverting your change but I need to revert this to revert #166788 ([comment](#169399 (comment)))
JacobSzwejbka
pushed a commit
that referenced
this pull request
Dec 8, 2025
This reverts commit a84798e. Reverted #166788 on behalf of https://github.com/huydhn due to Sorry for reverting your change but it seems to cause some numerical error in trunk ([comment](#166788 (comment)))
tiendatngcs
pushed a commit
to tiendatngcs/pytorch-Dec25
that referenced
this pull request
Dec 10, 2025
ghstack-source-id: fa8ce98 Pull Request resolved: pytorch/pytorch#166788
tiendatngcs
pushed a commit
to tiendatngcs/pytorch-Dec25
that referenced
this pull request
Dec 10, 2025
ghstack-source-id: 62b4540 Pull Request resolved: pytorch/pytorch#166788
Contributor
Author
|
@zou3519 I am debugging why Observations
On
So my recommendation is to skip the test for now for this PR, because this is an existing issue. And handle that as a separate issue, opened here - #170160 |
Contributor
|
@anijain2305 can you add a test case that defends against the source bug? I've seen this issue actually but wasn't able to repro it. It can just be test_partitioner_recomputes_factory_ones_like_sin_op with a fullgraph=True and cleaned up. I'm curious about what the tlparse looks like for the original PR - like, was the autograd.Function captured in whole? If it was captured in whole, did it run into the problem? Was the source bug introduced later? I'm not completely sure I'd classify this as "a preexisting problem" without that analysis. This PR does increase the memory usage of that autograd.Function because inductor fails to reinplace the custom op in the backward, it trades-off the graph break for additional memory. That being said, it's pretty rare for someone to use ones_like to allocate a tensor for a kernel, so I'm fine if you want to merge this. It's also a good time to yolo merge this (it's after the branch cut) |
Collaborator
|
Starting merge as part of PR stack under #169399 |
1 similar comment
Collaborator
|
Starting merge as part of PR stack under #169399 |
pytorchmergebot
pushed a commit
that referenced
this pull request
Dec 12, 2025
…tracing (#169399) Pull Request resolved: #169399 Approved by: https://github.com/tugsbayasgalan, https://github.com/zou3519 ghstack dependencies: #166788
vishalgoyal316
pushed a commit
to vishalgoyal316/pytorch
that referenced
this pull request
Dec 17, 2025
We make a rehaul because (1) we want to support non-proxyable outputs in the fwd method (2) we saw general softness in the support. I have put lot of comments in the code. Follow up * Graph break on backward stride dependent computation. This is BC breaking, so needs care. * Use DynamoAutogradFunctionTraceHelper for backward tracer. * Add test cases for module input, pytree input/outputs, pg groups * Consider unifying `automatic` and `automatic_with_forced_placeholders` * Better error messages - especially nonstrict trace for stride dependent backward computation. Pull Request resolved: pytorch#166788 Approved by: https://github.com/zou3519
vishalgoyal316
pushed a commit
to vishalgoyal316/pytorch
that referenced
this pull request
Dec 17, 2025
…tracing (pytorch#169399) Pull Request resolved: pytorch#169399 Approved by: https://github.com/tugsbayasgalan, https://github.com/zou3519 ghstack dependencies: pytorch#166788
krastogi-in
pushed a commit
to krastogi-in/pytorch
that referenced
this pull request
Jan 9, 2026
We make a rehaul because (1) we want to support non-proxyable outputs in the fwd method (2) we saw general softness in the support. I have put lot of comments in the code. Follow up * Graph break on backward stride dependent computation. This is BC breaking, so needs care. * Use DynamoAutogradFunctionTraceHelper for backward tracer. * Add test cases for module input, pytree input/outputs, pg groups * Consider unifying `automatic` and `automatic_with_forced_placeholders` * Better error messages - especially nonstrict trace for stride dependent backward computation. Pull Request resolved: pytorch#166788 Approved by: https://github.com/zou3519
krastogi-in
pushed a commit
to krastogi-in/pytorch
that referenced
this pull request
Jan 9, 2026
…tracing (pytorch#169399) Pull Request resolved: pytorch#169399 Approved by: https://github.com/tugsbayasgalan, https://github.com/zou3519 ghstack dependencies: pytorch#166788
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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
4 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):
We make a rehaul because
(1) we want to support non-proxyable outputs in the fwd method
(2) we saw general softness in the support.
I have put lot of comments in the code.
Follow up
automaticandautomatic_with_forced_placeholderscc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @Lucaskabela @chenyang78