Set gloo process group for FSDP with CPU offload #2108
Merged
Conversation
This file contains 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
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/torchtune/2108
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit e4f00c4 with merge base 32e265d ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
facebook-github-bot
added
the
CLA Signed
felipemello1
reviewed
Dec 3, 2024
recipes/full_finetune_distributed.py
Outdated
| if cfg.get("fsdp_cpu_offload", False): | ||
| # Utilize all available CPU cores for intra-op parallelism. This provides ~2x | ||
| # speed up when benchmarking fused AdamW on CPU | ||
| training.set_torch_num_threads() | ||
| process_group = "cuda:nccl,cpu:gloo" |
There was a problem hiding this comment.
n00b question: why not just make this the default every time, instead of "gloo" if cfg.device == "cpu" else "nccl"
There was a problem hiding this comment.
I actually wasn't sure myself why we do this. Well now I know. So yes, I think we can take your suggestion here
8 tasks
felipemello1
approved these changes
Dec 4, 2024
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Outdated suggestions cannot be applied.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.
Addresses #1977.
As discussed in the issue (see this comment), FSDP's implementation of gradient clipping uses _NormPartial, which requires comms primitives (specifically
all_reduce). This means that when running with CPU offloading we need to initialize the gloo process group to calculate the grad norm for DTensors on CPU. For simplicity this PR enables it whenever CPU offloading is enabled regardless of gradient clipping.Test plan:
Added a test case for gradient clipping + CPU offload to
test_full_finetune_distributed.py.