Small optimizations #5
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
This will only enable compilation for decoding. Note that there is not a big speedup for now, probably due to slot increasing buffer size over time, triggering recompilation.
Logits post-processing is not very heavyweight, and doing it on CPU actually accelerates decoding, because compilation is not re-triggered.
mfuntowicz
approved these changes
Mar 22, 2024
| @@ -512,8 +523,11 @@ def _generate_token( | |||
| # Save KV cache | |||
| self.past_key_values = outputs.past_key_values | |||
| # Barrier for XLA model | |||
| xm.mark_step(wait=False) | |||
| xm.mark_step() | |||
There was a problem hiding this comment.
For my knowledge: We were not waiting before, why this has changed here?
There was a problem hiding this comment.
the default is wait=False, and I did not want to give the false impression I am changing the default behaviour, so I just removed the default parameter.
| @@ -44,13 +45,19 @@ def create_request( | |||
| seed: int = 0, | |||
| repetition_penalty: float = 1.0, | |||
| ): | |||
| # For these tests we can safely set typical_p to 1.0 (default) | |||
| typical_p = 1.0 | |||
| if do_sample == False: | |||
| @@ -35,13 +36,19 @@ def create_request( | |||
| seed: int = 0, | |||
| repetition_penalty: float = 1.0, | |||
| ): | |||
| # For these tests we can safely set typical_p to 1.0 (default) | |||
| typical_p = 1.0 | |||
| if do_sample == False: | |||
tengomucho
force-pushed
the
small-optimizations
branch
from
4dda7e9
to
401eea6
Compare
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Outdated suggestions cannot be applied.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 PR adds few optimizations that have been added in preparation of model compilation for decoding. Note that compilation is still not enabled by default due to a bug I am currently investigating.
On the other hand, I spent some time profiling the code with the xla profiling API and I was able to understand that adding few
xm.mark_stepimproved the performance, and that the token processing code eventually runs faster if executed in the CPU, because it will avoid recompilation.