fix: prevent panic in memory_cleanup on CUDA after linear forward #3932
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
After a linear layer forward call in autodiff mode, executing AutodiffBackend::memory_cleanup on CUDA would panic with 'The size should match'. This occurred because memory_cleanup was called without synchronizing pending asynchronous operations, causing size mismatches in CubeCL's memory management. The fix adds a sync() call before memory_cleanup in CubeBackend to ensure all operations are completed before cleanup. This resolves the issue for CUDA backends while having no impact on synchronous backends like ndarray. Added a test in burn-cubecl to verify memory_cleanup works after linear operations. Files changed: - crates/burn-cubecl/src/backend.rs: Added sync before memory_cleanup - crates/burn-cubecl/src/tests/memory_cleanup.rs: New test file - crates/burn-cubecl/src/tests/mod.rs: Added test module and macro
|
Hi, I've tried running with your changes and it still panics for me. The part that panics is block Z, which has a linear layer. Your test refers to block X (although block X and Y are pretty much the same). Also, I think you accidentally removed |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 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Outdated suggestions cannot be applied.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.
After a linear layer forward call in autodiff mode, executing AutodiffBackend::memory_cleanup on CUDA would panic with 'The size should match'. This occurred because memory_cleanup was called without synchronizing pending asynchronous operations, causing size mismatches in CubeCL's memory management.
The fix adds a sync() call before memory_cleanup in CubeBackend to ensure all operations are completed before cleanup. This resolves the issue for CUDA backends while having no impact on synchronous backends like ndarray.
Added a test in burn-cubecl to verify memory_cleanup works after linear operations.
Files changed:
Pull Request Template
Checklist
cargo run-checkscommand has been executed.Related Issues/PRs
Provide links to relevant issues and dependent PRs.
Changes
Summarize the problem being addressed and your solution.
Testing
Describe how these changes have been tested.
issue fixed: #3927