[inductor] Skip cudagraph capture when CUDA caching allocator is bypassed#183780
Open
jeffdaily wants to merge 1 commit into
Open
[inductor] Skip cudagraph capture when CUDA caching allocator is bypassed#183780jeffdaily wants to merge 1 commit into
jeffdaily wants to merge 1 commit into
Conversation
…ssed `PYTORCH_NO_CUDA_MEMORY_CACHING=1` (and the equivalent `PYTORCH_NO_HIP_MEMORY_CACHING`) forces every allocation through `cudaMalloc` / `hipMalloc`, bypassing the caching allocator. Cudagraph capture pools allocations through that allocator; without it, capture appears to complete normally but the pool tracking diverges, and the first replay fires a cryptic `"storage data ptrs are not allocated in pool ..."` error. Detect the bypass at the same layer as the existing device-topology check (`cudagraph_utils.check_lowering_disable_cudagraph`) and skip cudagraph capture with a clear reason. Tests and other code that opportunistically enable cudagraphs (`test_padding`, `test_callback`, inductor's `compile_fx` pipeline, etc.) still run — they just exercise the eager triton kernels instead of the cudagraph wrapper. Two existing tests in `test/inductor/test_cuda_repro.py` (`test_unused_cpu_input_cudagraphs` and `test_cpu_index`) explicitly assert `graph.disable_cudagraphs_reason is None`. Add `skipTest()` guards in those tests so they bow out cleanly when the env bypass is set, keeping the new inductor invariant strict. This is a debugging-experience fix for anyone using `PYTORCH_NO_CUDA_MEMORY_CACHING` (memory-debug builds, sanitizer builds, etc.), not a ROCm-specific change. Authored with Claude.
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/183780
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: ❌ 1 New Failure, 2 Unrelated FailuresAs of commit c5270cc with merge base a1cc64b ( NEW FAILURE - The following job has failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
UNSTABLE - The following job is marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Contributor
|
@claude explain those CI failures, check if they are related, and review the changes in this PR |
|
Claude Code is working… I'll analyze this and get back to you. |
ezyang
reviewed
May 15, 2026
| Truthy match follows c10::utils::check_env: only the literal "1" counts.""" | ||
| if ( | ||
| os.environ.get("PYTORCH_NO_CUDA_MEMORY_CACHING") == "1" | ||
| or os.environ.get("PYTORCH_NO_HIP_MEMORY_CACHING") == "1" |
Contributor
There was a problem hiding this comment.
It would be better to query the memory allocator state directly to get this info rather than re-inferring it from env
ezyang
reviewed
May 15, 2026
| os.environ.get("PYTORCH_NO_CUDA_MEMORY_CACHING") == "1" | ||
| or os.environ.get("PYTORCH_NO_HIP_MEMORY_CACHING") == "1" | ||
| ): | ||
| self.skipTest("cudagraphs disabled when caching allocator is bypassed") |
Contributor
There was a problem hiding this comment.
Surely this is worth reifying into a decorator
Contributor
|
This seems fine, although I'm not sure if there's a more structural fix we should do for the direct mode / other consequences when the caching allocator is disabled. |
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.
PYTORCH_NO_CUDA_MEMORY_CACHING=1(and the equivalentPYTORCH_NO_HIP_MEMORY_CACHING) forces every allocation throughcudaMalloc/hipMalloc, bypassing the caching allocator. Cudagraph capture pools allocations through that allocator; without it, capture appears to complete normally but the pool tracking diverges, and the first replay fires a cryptic"storage data ptrs are not allocated in pool ..."error.Detect the bypass at the same layer as the existing device-topology check (
cudagraph_utils.check_lowering_disable_cudagraph) and skip cudagraph capture with a clear reason. Tests and other code that opportunistically enable cudagraphs (test_padding,test_callback, inductor'scompile_fxpipeline, etc.) still run — they just exercise the eager triton kernels instead of the cudagraph wrapper.Two existing tests in
test/inductor/test_cuda_repro.py(test_unused_cpu_input_cudagraphsandtest_cpu_index) explicitly assertgraph.disable_cudagraphs_reason is None. AddskipTest()guards in those tests so they bow out cleanly when the env bypass is set, keeping the new inductor invariant strict.This is a debugging-experience fix for anyone using
PYTORCH_NO_CUDA_MEMORY_CACHING(memory-debug builds, sanitizer builds, etc.), not a ROCm-specific change.Authored with Claude.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @mlazos