Remove CPU sync before Sampler #414
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
kdamaszk
requested review from
michalkuligowski,
kzawora-intel,
madamczykhabana and
mswiniarsk
michalkuligowski
approved these changes
Oct 22, 2024
michalkuligowski
deleted the
dev/kdamaszke/remove-cpu-sync-before-sampler
branch
xuechendi
pushed a commit
to xuechendi/vllm-fork
that referenced
this pull request
Oct 23, 2024
Currently before each Sampler call we have a CPU sync, which causes a host gap: <img width="226" alt="image" src="https://github.com/user-attachments/assets/4509e69b-0f16-4ac9-812e-a2a9bc43a6ad"> This PR is removing that sync, so the host gap is no longer visible: <img width="133" alt="image" src="https://github.com/user-attachments/assets/66c19e4b-d832-4955-848d-8ae4acd8d264"> NOTE: class `ApplyToppTopkScalar` still has some CPU syncs inside. It means that the biggest gain will be observed in the scenario without `top_p` or `top_k` parameters. I think it is worth to investigate if we can remove the syncs from this function too.
tzielinski-habana
pushed a commit
that referenced
this pull request
Oct 24, 2024
Currently before each Sampler call we have a CPU sync, which causes a host gap: <img width="226" alt="image" src="https://github.com/user-attachments/assets/4509e69b-0f16-4ac9-812e-a2a9bc43a6ad"> This PR is removing that sync, so the host gap is no longer visible: <img width="133" alt="image" src="https://github.com/user-attachments/assets/66c19e4b-d832-4955-848d-8ae4acd8d264"> NOTE: class `ApplyToppTopkScalar` still has some CPU syncs inside. It means that the biggest gain will be observed in the scenario without `top_p` or `top_k` parameters. I think it is worth to investigate if we can remove the syncs from this function too.
michalkuligowski
pushed a commit
that referenced
this pull request
Oct 30, 2024
This change is fixing the performance issue I have introduced in the PR #414 -- due to the usage of `torch.where` both functions have been called. Now we will run only the selected one.
imangohari1
pushed a commit
to imangohari1/vllm-fork
that referenced
this pull request
Oct 31, 2024
This change is fixing the performance issue I have introduced in the PR HabanaAI#414 -- due to the usage of `torch.where` both functions have been called. Now we will run only the selected one.
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 before each Sampler call we have a CPU sync, which causes a host gap:
This PR is removing that sync, so the host gap is no longer visible:
NOTE: class
ApplyToppTopkScalarstill has some CPU syncs inside. It means that the biggest gain will be observed in the scenario withouttop_portop_kparameters. I think it is worth to investigate if we can remove the syncs from this function too.