Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[Submodule] Change FlashInfer to import #156

Merged
merged 1 commit into from
Feb 7, 2024
Merged

[Submodule] Change FlashInfer to import #156

merged 1 commit into from
Feb 7, 2024

Conversation

comaniac
Copy link
Contributor

@comaniac comaniac commented Feb 7, 2024

This PR removes submodule flashinfer from this repo. As flashinfer is now officially released, we can directly install it via pip. However, I didn't add it to pyproject.toml but just updated the README due to two reasons:

  1. flashinfer is not available on PYPI yet.
  2. It supports limited CUDA version and requires to manually select the wheel path.

Meanwhile, the latest version of flashinfer has some interface changes. I modified them according to this comparison:
flashinfer-ai/flashinfer@88b9496...3fc62cb

I verified two things with this PR:

  1. My test prompt gets the same output tokens before and after this PR.
  2. benchmark/multi_chain_reasoning generates different results even temperature=0. There might be due to the misuse of the new interface or the randomness in the kernel. I'll do more investigation.

cc @Ying1123

@comaniac comaniac requested a review from merrymercy February 7, 2024 00:23
@comaniac
Copy link
Contributor Author

comaniac commented Feb 7, 2024

Per offline discussion, I instead used gsm8k with 20 questions. The outputs with and without this PR are 99% the same, so we should be good to go.

@Ying1123 Ying1123 merged commit 26c3494 into main Feb 7, 2024
@Ying1123 Ying1123 deleted the cody/flashinfer branch February 7, 2024 03:28
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants