[dynamo][guards] Skip no tensor aliasing guards on parameters #138954
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
This is another unsound guard eval optimization. Its rare in practice to compile a function with two different parameters as inputs, and then later call the function with the same input as two different inputs (aliasing). This further reduces guard overhead from 280 us to 240 us for the model in #138386 [ghstack-poisoned]
This was referenced Oct 25, 2024
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/138954
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 1 Unrelated FailureAs of commit b399150 with merge base f9ae3fa ( NEW FAILURE - The following job has failed:
BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
…ers" This is another unsound guard eval optimization. Its rare in practice to compile a function with two different parameters as inputs, and then later call the function with the same input as two different inputs (aliasing). This further reduces guard overhead from 280 us to 240 us for the model in #138386 cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
…ers" This is another unsound guard eval optimization. Its rare in practice to compile a function with two different parameters as inputs, and then later call the function with the same input as two different inputs (aliasing). This further reduces guard overhead from 280 us to 240 us for the model in #138386 cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
anijain2305
added
ciflow/trunk
Contributor
|
I am sympathetic to this but it doesn't seem like it should be taking 40us to test this in the first place |
ezyang
reviewed
Oct 28, 2024
torch/_dynamo/guards.py
Outdated
| else: | ||
| self.tensor_check_examples.append(value) | ||
| self.tensor_check_names.append(tensor_name) | ||
| self.tensor_check_guards.append(guard) |
Contributor
There was a problem hiding this comment.
There's a refactor going on here that makes it hard to see the substantive change that happened here
Contributor
Author
There was a problem hiding this comment.
I can pull the refactor in a separate PR.
ezyang
reviewed
Oct 28, 2024
torch/_dynamo/guards.py
Outdated
| add_code_part(code, gcl.guard, True) | ||
| seen.add(code) | ||
|
|
||
| tensor_check_names = builder.tensor_check_names |
Contributor
There was a problem hiding this comment.
Does the code in torch/csrc/dynamo/guards.cpp referencing tensor_check_names have to be adjusted?
Also, since this happened before, can you explicitly make sure our verbose reporting (e.g., in logs and tlparse) did not regress with these changes? (The switch to cpp guard manager caused us to lose all tlparse guard output)
Contributor
Author
The quantized model in question has a really large number of parameters, and therefore we end up calling this guard for every parameter. There is not much going on in the guard itself. Its just the sheer quantity of them. |
…ers" This is another unsound guard eval optimization. Its rare in practice to compile a function with two different parameters as inputs, and then later call the function with one parameter input as two different inputs (aliasing). This further reduces guard overhead from 280 us to 240 us for the model in #138386 cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
…ers" This is another unsound guard eval optimization. Its rare in practice to compile a function with two different parameters as inputs, and then later call the function with one parameter input as two different inputs (aliasing). This further reduces guard overhead from 280 us to 240 us for the model in #138386 cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames rec [ghstack-poisoned]
jansel
approved these changes
Oct 29, 2024
pytorchmergebot
pushed a commit
that referenced
this pull request
Oct 29, 2024
This brings some unsoundness in guards. Earlier we were skipping empty nn module hooks dict guard only on inbuilt nn modules, but as seen in #138386, there could be still be significant guard overhead. With this PR, we reduce the guard eval latency from 420 us to 280 us (1.5x reduction). Pull Request resolved: #138942 Approved by: https://github.com/ezyang, https://github.com/jansel ghstack dependencies: #139040, #138954
rahulsingh-intel
pushed a commit
to rahulsingh-intel/pytorch
that referenced
this pull request
Oct 29, 2024
…h#138954) This is another unsound guard eval optimization. Its rare in practice to compile a function with two different parameters as inputs, and then later call the function with one parameter input as two different inputs (aliasing). This further reduces guard overhead from 280 us to 240 us for the model in pytorch#138386 Pull Request resolved: pytorch#138954 Approved by: https://github.com/jansel ghstack dependencies: pytorch#139040
rahulsingh-intel
pushed a commit
to rahulsingh-intel/pytorch
that referenced
this pull request
Oct 29, 2024
This brings some unsoundness in guards. Earlier we were skipping empty nn module hooks dict guard only on inbuilt nn modules, but as seen in pytorch#138386, there could be still be significant guard overhead. With this PR, we reduce the guard eval latency from 420 us to 280 us (1.5x reduction). Pull Request resolved: pytorch#138942 Approved by: https://github.com/ezyang, https://github.com/jansel ghstack dependencies: pytorch#139040, pytorch#138954
rahulsingh-intel
pushed a commit
to rahulsingh-intel/pytorch
that referenced
this pull request
Nov 5, 2024
…h#138954) This is another unsound guard eval optimization. Its rare in practice to compile a function with two different parameters as inputs, and then later call the function with one parameter input as two different inputs (aliasing). This further reduces guard overhead from 280 us to 240 us for the model in pytorch#138386 Pull Request resolved: pytorch#138954 Approved by: https://github.com/jansel ghstack dependencies: pytorch#139040
rahulsingh-intel
pushed a commit
to rahulsingh-intel/pytorch
that referenced
this pull request
Nov 5, 2024
This brings some unsoundness in guards. Earlier we were skipping empty nn module hooks dict guard only on inbuilt nn modules, but as seen in pytorch#138386, there could be still be significant guard overhead. With this PR, we reduce the guard eval latency from 420 us to 280 us (1.5x reduction). Pull Request resolved: pytorch#138942 Approved by: https://github.com/ezyang, https://github.com/jansel ghstack dependencies: pytorch#139040, pytorch#138954
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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):
This is another unsound guard eval optimization. Its rare in practice to
compile a function with two different parameters as inputs, and then
later call the function with one parameter input as two different inputs
(aliasing). This further reduces guard overhead from 280 us to 240 us
for the model in #138386
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames @rec