Closed
Conversation
…aking into trace PR pytorch#1362 (D100731798) fixed a bug in activityBuffers() where readyGpuTraceBuffers_ was ignored when allocatedGpuTraceBuffers_ was empty. A side effect is that CUPTI buffers flushed during toggleCollectionDynamic(false) are now correctly returned during trace processing. This surfaced a latent issue: synchronizeGpuDevice() calls cudaDeviceSynchronize() while CUPTI is still enabled, so CUPTI captures that call as a "cudaDeviceSynchronize" runtime activity. The event contains "cuda" in its name and fails the PyTorch test_dynamic_toggle assertion that no GPU events appear after the toggle. Fix by disabling CUPTI activity collection before the device sync when toggling off. The sync still serves its purpose — draining in-flight GPU work and flushing pre-existing CUPTI buffers — but the sync call itself is no longer instrumented.
|
@scotts has imported this pull request. If you are a Meta employee, you can view this in D101110738. |
scotts
added a commit
to scotts/pytorch
that referenced
this pull request
Apr 16, 2026
Includes the following commits: - Fix stream wait events referencing future correlation IDs (pytorch/kineto#1339) 23b5bb5 - Remove kineto tb_plugin directory entirely (pytorch/kineto#1368) 9497960 - Move Stream Sync events to a new row in JSON trace export (pytorch/kineto#1356) 041e7ce - Expose isGpuCollectionStopped() through Kineto's public API (pytorch/kineto#1367) 17708f5 - Fix toggle test (pytorch/kineto#1369) ee2103c - Link to correct fmt repo (pytorch/kineto#1345) 3447834 - Fix data race on CuptiActivityApi::externalCorrelationEnabled_ (pytorch/kineto#1365) 0e86499 - Stop allocating CUPTI buffers after exceeding max buffer count (pytorch/kineto#1362) 666f62c - Add XPU workflow (pytorch/kineto#1302) 11cc1e0 - Remove RocprofActivity.h/RoctracerActivity.h from RocmActivityProfiler.h (pytorch/kineto#1357) 896068d - Split ActivityProfilerController into Sync and Async Handlers (pytorch/kineto#1269) 6d7f045 - Add priority field to kernel metadata (pytorch/kineto#1361) f2a7423 - Add kineto-release skill (pytorch/kineto#1360) 675b6cd Authored with Claude.
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
1 participant
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.
Fix
test_dynamic_togglefailure caused by toggle's own CUDA events leaking into trace.PR #1362 (D100731798) fixed a bug in
activityBuffers()wherereadyGpuTraceBuffers_was ignored whenallocatedGpuTraceBuffers_was empty. A side effect is that CUPTI buffers flushed duringtoggleCollectionDynamic(false)are now correctly returned during trace processing. This surfaced a latent issue:synchronizeGpuDevice()callscudaDeviceSynchronize()while CUPTI is still enabled, so CUPTI captures that call as a "cudaDeviceSynchronize" runtime activity. The event contains "cuda" in its name and fails the PyTorchtest_dynamic_toggleassertion that no GPU events appear after the toggle.Fix by disabling CUPTI activity collection before the device sync when toggling off. The sync still serves its purpose — draining in-flight GPU work and flushing pre-existing CUPTI buffers — but the sync call itself is no longer instrumented.