Repository navigation
Avoid mixing allocators in mGPU fused losses - #840
Merged
Merged
Conversation
Avoid mixing the CUDA caching allocator with the PrivateUse1 allocator for temporary fused loss buffers. Use cudaMallocAsync and stream-ordered cleanup on mGPU, while keeping CUDA scratch in torch::empty tensors. Move allocation and reduction into the backend entry points and pass caller-owned buffers to the shared tile helpers. Signed-off-by: Matthew Cong <mcong@nvidia.com>
swahtz
reviewed
Oct 30, 2026
swahtz
reviewed
Oct 30, 2026
swahtz
approved these changes
Oct 30, 2026
swahtz
left a comment
swahtz
left a comment
Contributor
There was a problem hiding this comment.
Looks good to me. Two cleanup comments, feel free to address if you think they're useful, otherwise I don't see any issues
Signed-off-by: Matthew Cong <mcong@nvidia.com>
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fused L1 and L1 + SSIM currently allocate CUDA tensors for scratch in the mGPU path. The
CUDACachingAllocatorretains unused blocks, causing GPU memory contention with the kPrivateUse1 unified-memory pool. This PR switches mGPU scratch tocudaMallocAsyncwith stream-ordered cleanup while retainingtorch::emptyfor kCUDA. Allocation and reduction now live in the backend entry points, and the shared tile helpers take caller-owned buffers.This also removes the prematurely added mGPU test variants from #820. mGPU tests do not yet exist elsewhere and will be added in the future. CUDA test coverage is kept.