Repository navigation
Fix 'Adaptive pool MPS' crash in post-optimization search - #63
Merged
Merged
Conversation
F.interpolate(mode="area") uses adaptive_avg_pool2d, which MPS only supports for evenly divisible sizes. Fixes hvoss-tech#62.
Owner
|
Hey Floris, Thank you for the quick fix! I'll merge it. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Fixes #62
Problem
On Apple Silicon (MPS), runs crash when the post-optimization search starts:
For large images, the search first shrinks the target with
F.interpolate(mode="area"). That usesadaptive_avg_pool2dinternally, and MPS only supports it when the old size divides evenly by the new one (pytorch#96056). The new size comes from a scale factor, so in practice it almost never does.Fix
On MPS only, this one resize now runs on the CPU, and the result moves straight back to MPS. Everything else in the search, including scoring, stays on the GPU.
CUDA, ROCm and CPU runs are unchanged.
Testing
mps:0.mainwithout this change.Risk
Low. The change only affects MPS runs with images large enough to be downsampled for the search, and it's easy to revert.