[RFC] Pass the original input to all PP stages #1130
Conversation
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
facebook-github-bot
added
the
CLA Signed
H-Huang
approved these changes
Apr 22, 2025
|
@H-Huang llama3 doesn't have this. And document masking + PP is one missing feature for llama4. |
We need the original tokens to generate the document masks/block causal masks. Since TorchTitan currently let all ranks perform data loading, there will be no performance regressions.
wconstab
approved these changes
Apr 24, 2025
tianyu-l
approved these changes
Apr 27, 2025
| If pipeline parallelism is enabled, this will be the input token indices | ||
| for the ranks on the first pipeline stage. This will be the activation of the | ||
| previous pipeline stage if the current rank is not on the first stage. | ||
| input_batch (torch.Tensor): The input batch read from the dataloader. |
There was a problem hiding this comment.
please add a comment that -- this field is needed for non-first PP stages to obtain proper document masks
| @@ -351,7 +355,7 @@ def train_step(self, input_dict: dict[str, torch.Tensor], labels: torch.Tensor): | |||
| # Non-PP forward / backward | |||
| with self.train_context(optional_context_parallel_ctx): | |||
| assert len(model_parts) == 1 | |||
| pred = model_parts[0](inputs) | |||
| pred = model_parts[0](inputs, input_batch=inputs) | |||
There was a problem hiding this comment.
non PP branch looks a bit strange -- I slightly prefer the alternative way of making input_batch optional, and let init_attention_mask use tokens if input_batch is None. The idea is that non PP users see less universal usage of input_batch
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.
We need the original tokens to generate the document masks/block causal masks. Since TorchTitan currently let all ranks perform data loading, there will be no performance regressions.
This is required to support document masking attention with PP.