Convert formatting to use ruff instead of yapf + isort

[hmellor]

Closes #17657

This is a massive change and would be impossible to merge during the week. Any PR's merged while this was waiting for CI would cause merge conflicts which would be time consuming to solve in this PR.

Therefore, the plan to getting this PR merged quickly is as follows:

• First commit:
  • Make the changes to the pre-commit
  • Run pre-commit run -a and note all the things ruff couldn't automatically fix
  • Add these as pre-file-ignores to pyproject.toml
• Second commit:
  • Run pre-commit run -a
  • ruff and ruff format should pass immediately

Then, these ignores will be systematically removed in smaller more manageable PRs afterwards.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @hmellor.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[kfirwolfson]

Hi @hmellor, has anyone run into an issue with the new ruff definitions?
On one hand it straightens out if-else clauses to one liners, but on the other hand it can make these lines too long (>=80)?
This causes pre-commit to constantly fail.

Edit: seems this is a known issue. Appending " #noqa E501" seems to overcome it temporarily. Would love to know how to solve it going forward

[hmellor]

I think you may have a faulty configuration. This is not a known issue and I don't like using #noqa E501 as a workaorund.

In your Slack message you mentioned that the SIM rule was making lines to long (>80), but the new line length limit is 88.

[hmellor]

Having just found your PR I think what you have done is reasonable https://github.com/vllm-project/vllm/pull/24520/files#diff-9eeca590fd99f15621897e559dba39b3ec4e7c2c65ec3c3229711689e008b5f4R435

Sorry for the false alarm!

[kfirwolfson]

Thanks @hmellor. Yes, sorry, the row is actually 89 chars long. I also prefer not using #noqa E501. Is there another way around the problem?

[hmellor]

This would do it:

                    scheduler_config = self.vllm_config.scheduler_config
                    cache_hit_threshold = (
                        request.cache_hit_threshold
                        if request.cache_hit_threshold is not None
                        else scheduler_config.global_cache_hit_threshold
                    )

[kfirwolfson]

hmm, nice idea.

If we're adding another assignment, one can argue that we can might as well assign the global value and override if request has it.

                    cache_hit_threshold = \
                        self.vllm_config.scheduler_config.global_cache_hit_threshold
                    # Cache hit threshold in request overrides global setting
                    if request.cache_hit_threshold is not None:
                        cache_hit_threshold = request.cache_hit_threshold

I guess it's a matter of preference. The way you suggested looks more Pythonic to me, I'll go with that. Thanks

[ExtReMLapin]

Fucking hell, how to create conflicts in one step on opened PRs

[hmellor]

Some disruption was inevitable. However:

• We were going to do this eventually
• It's better to do it all at once rather than creating continued disruption over a longer period
• We did it at the weekend when vLLM development is less active
• We provided detailed instructions on how to resolve the conflicts here