track number of cpp->python exceptions thrown in torch.compile benchmark suite #131481
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
…ark suite [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/131481
Note: Links to docs will display an error until the docs builds have been completed. ❌ 42 New Failures, 1 Cancelled Job, 14 Unrelated FailuresAs of commit 09b09fb with merge base 932ae13 ( NEW FAILURES - The following jobs have failed:
CANCELLED JOB - The following job was cancelled. Please retry:
BROKEN TRUNK - The following jobs 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. |
pytorch-bot
bot
added
ciflow/inductor
module: dynamo
release notes: jit
This was referenced Jul 23, 2024
github-actions
bot
requested review from
SherlockNoMad,
albanD,
antoniojkim,
ezyang and
miladm
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. [ghstack-poisoned]
bdhirsh
commented
Jul 23, 2024
| TORCH_API void runJITCPPTests(); | ||
| #endif | ||
|
|
||
| static thread_local uint64_t cpp_to_python_translated_exception_count{0}; |
Contributor
Author
There was a problem hiding this comment.
My thought process here was:
(1) I figured this can't be a plain uint64_t, since multiple threads may be raising exceptions and can increment the counter.
(2) I originally made it a std::atomic<uint64_t>, but then I didn't want to risk the increment code being very slow, and exacerbating slowness problems in cases where there are many cpp-to-python exceptions. Alternatively, we could make it slow but gate it behind a config / env var, but this seems like a nice metric to be able to track in an always-on way.
(3) thread_local will under-count if multiple threads are all raising exceptions. My hope is that we are very unlikely to hit a situation where the main thread raises no cpp-to-python exception, but a side thread raises a large number (although if this happens, then... we will not be able to track it in metrics).
Contributor
There was a problem hiding this comment.
Don't worry about (2), you're already in slow pain with exceptions, the atomic is a rounding error.
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. [ghstack-poisoned]
Contributor
Author
|
I gave up looking for a way to grab the output from CI, so I just ran a script that adds the extra column to every existing csv we check in (set to zero - if there are any exceptions, I'll expect CI to fail) |
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
…pile benchmark suite" cpp->python exceptions can be slow - **especially* if you are running pytorch with infra that tries to symbolize C++ stacktraces. This PR attempts to: (1) add a counter every time pybind catches a C++ exception that enters python (2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero. It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops. cc albanD voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames [ghstack-poisoned]
Contributor
|
Looks like this PR hasn't been updated in a while so we're going to go ahead and mark this as |
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.
cpp->python exceptions can be slow - *especially if you are running pytorch with infra that tries to symbolize C++ stacktraces.
This PR attempts to:
(1) add a counter every time pybind catches a C++ exception that enters python
(2) track the value of this counter in our benchmark suite, so we can hopefully drive it to zero.
It might be the case that we have very few (or even zero) of these cases showing up in torchbench. Emprically from internal, it also appears that this can show up when torch.compiling custom ops. So if few-or-zero exceptions are encountered in torchbench, we should also add some tests for avoiding C++ exceptions in the common path of torch.compile usage with custom ops.
Stack from ghstack (oldest at bottom):
cc @albanD @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames