Shard input_mask for Llama #905
Draft
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
sogartar
reviewed
Feb 5, 2025
There was a problem hiding this comment.
One thing that seems unaddressed is the scaled_dot_product_attention. It also applies the mask. How would that be reconciled? If is_causal=False and attention_mask=None does it not apply any mask?
| @@ -129,20 +129,25 @@ def prefill( | |||
| # [bs, batch_seq_len] | |||
| tokens: Union[torch.Tensor, ReplicatedTensor], | |||
| *, | |||
| # [bs, batch_seq_len] | |||
| input_mask: Union[torch.Tensor, ReplicatedTensor], | |||
There was a problem hiding this comment.
This input needs a check such that it should be mutually exclusive with the attention_mask arg.
| @@ -166,6 +171,8 @@ def decode( | |||
| # [bs, 1] | |||
| tokens: Union[torch.Tensor, ReplicatedTensor], | |||
| *, | |||
| # [bs, 1] | |||
| input_mask: Union[torch.Tensor, ReplicatedTensor], | |||
There was a problem hiding this comment.
This input needs a check such that it should be mutually exclusive with the attention_mask arg.
sogartar
reviewed
Feb 5, 2025
| self._assert_device(attention_mask, dtype=self.activation_dtype) | ||
| self._assert_device(seq_block_ids) | ||
| self._assert_device(*cache_state, dtype=self.activation_dtype) | ||
|
|
||
| h = self.token_embedding(tokens) | ||
| self.trace_tensor("llama.token_embedding", h) | ||
|
|
||
| h *= input_mask.unsqueeze(-1) |
There was a problem hiding this comment.
I don't think that if we assume this should somehow substitute the attention mask the math checks out.
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Outdated suggestions cannot be applied.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.
Currently, 405b OOMs when using long input prompts (issue here).
This PR implements a suggested fix for this, which is to shard the
input_mask. This makes sense, since the OOM error is dependent on the length of the input.I'm new to sharktank and still seeing the issue with current implementation, so wanted to have it double-checked to make sure it's properly implemented.
MLIR here