Enable compilation #6
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
when direct assignment is used in inputs, apparently the compiled model can give wrong results, as explained in this issue: pytorch/xla#6796 The workaround seems to be to use the `index_put_` method.
This improves inference time a lot.
mfuntowicz
approved these changes
Mar 26, 2024
| @@ -493,8 +492,8 @@ def decode(self, batches: List[CachedBatch]) -> Tuple[List[Generation], CachedBa | |||
| dtype=torch.int64, | |||
| device=self.model.device, | |||
| ) | |||
| attention_mask[i, :] = slot.attention_mask | |||
| position_ids[i, 0] = slot.cur_position | |||
| attention_mask.index_put_([torch.tensor([i])], slot.attention_mask) | |||
There was a problem hiding this comment.
Can't we just put 1 here? I.e. the new token in the attention_mask we want to attend to
There was a problem hiding this comment.
arf, nope, it was correct, and I just broke it: I need to put the slot attention mask in the i-th line (corresponding the batch i). I am going to fix it.
It's useless to put another value, and the index was put there by mistake.
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Outdated suggestions cannot be applied.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.
What does this PR do?
This model implements a workaround to a problem appearing when compilation is used in torch_xla + TPU and direct assignment is used on inputs.
After fixing that, it is possible to enable compilation by default on models that support static cache, such as gemma, leading to an improved performance in decoding after the first token is decoded.