fix: asyncio-friendly nccl operations #52
Merged
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Outdated suggestions cannot be applied.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.
Description
NCCL operation in PyTorch's distributed package needs to set up NCCL communicator so that ranks can talk to one another. To set up the communicator, c10d key-value store needs to be consulted. This is a blocking call, which blocks asyncio's loop. This prevents the loop from scheduling different coroutines. The issue is mitigated by using run_in_executor().
Note that this doesn't seem to be a permanent fix. Depending on timing, blocking appears from time to time and leads to an exception whose example may looks like "torch.distributed.DistBackendError: [1] is setting up NCCL communicator and retrieving ncclUniqueId from [0] via c10d key-value store by key '0:1', but store->get('0:1') got error: Socket Timeout".
Type of Change
Checklist