Reuse Existing Cuda Streams each Epoch #308
Open
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Outdated suggestions cannot be applied.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.
This PR modifies the FFCV
Loader
andEpochIterator
to create one set of Cuda streams and reuse them each epoch instead of creating new Cuda streams every epoch.The current method of creating new Cuda streams every epoch can cause increasing memory allocation if using a GPU transform, as each epoch the GPU transform will allocate new memory in the new Cuda stream. This doesn't cause any errors, as the prior epochs' allocation can be reused. But it does make keeping track of GPU memory usage more difficult and can hide real memory overflow errors.
I don't think this will cause any issues with distributed training, but am unable to test.
If wanted, I can create a flag like
recompile
to recreate Cuda streams every epoch to match current behavior.