Add support for context parallelism #174
Open
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
2015aroras
approved these changes
Feb 21, 2025
| for b in attn_buffers.values(): | ||
| mark_dynamic(b, 0) | ||
|
|
||
| with self._model_forward_context(): |
There was a problem hiding this comment.
It looks like PP doesn't support CP yet. Maybe cause a config error in that case.
| @@ -560,9 +592,14 @@ def train_batch(self, batch: Dict[str, Any], dry_run: bool = False): | |||
|
|
|||
| # Train one micro-batch at a time. | |||
| for micro_batch_idx, micro_batch in enumerate(micro_batches): | |||
| with self._train_microbatch_context(micro_batch_idx, num_micro_batches): | |||
| attn_buffers = self.model.get_attn_buffers( | |||
There was a problem hiding this comment.
Is there a need to pass the attention buffers back through forward in this manner? I thought the context parallelism would modify them in-place. I would prefer not doing this if it's not needed.
There was a problem hiding this comment.
This API needs some improvement but there's a couple reasons I chose to pass these buffers through the forward method:
- This allows the
TransformerTrainModuleto mark their sizes as dynamic fortorch.compile(). This is irrelevant to context parallelism but it's a good thing to do in general. Without this I think we were getting some unnecessary re-compilations during in-loop evals. - It guarantees we don't create duplicate versions of these buffers, which could happen if the
BufferCacheowned by eachAttentionmodule was a different instance. In our usual constructors we make sure the cache is shared, but if someone instantiated their model differently, the cache might not be shared.
There was a problem hiding this comment.
Alright refactored this so that the buffers are managed completely by the model.
Co-authored-by: Shane A <[email protected]>
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.
To enable, run with
--train_module.cp_config='{degree: 2}'(adjust the degree as needed). Probably don't try this with TP at the same time for now.