[Bugfix] Fix ray instance detect issue

[yma11]

Fix ray instance detect so that will first try connecting to latest launched instance and if not, create a new one with num_gpus=parallel_config.world_size.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[yma11]

@youkaichao can you help review this change? Thanks.

[comaniac]

Overall LGTM. One question is should we also change the other branch?

[comaniac]

Use double quotes for consistency.

[yma11]

Overall LGTM. One question is should we also change the other branch?

Agree. So I unify the init logic which should make sense for all platforms, please help take a look again. Thanks.

[comaniac]

The update code seems not equivalent to the original one? Currently for non-hip and non-xpu cases we don't init Ray with all GPUs.

[yma11]

The update code seems not equivalent to the original one? Currently for non-hip and non-xpu cases we don't init Ray with all GPUs.

For non-hip and non-xpu cases, it will finally create a local instance with detected gpus if fails to connect existing cluster based on explanation.
Actually I intended to fix an error "When connecting to an existing cluster, num_cpus and num_gpus must not be provided." in xpu case. It happens when a valid ray_address and num_gpus are both given. I want to respect both of these values but seems the confliction can't be resolved. Maybe it's more reasonable to do ray.init(address=ray_address, ignore_reinit_error=True) for all platforms. num_gpus=parallel_config.world_size is expected to take affect only when new local instance created but it's not so meaningful in that case. What do you think?

[comaniac]

Sounds reasonable to me, but cc @rkooo567 @richardliaw to double check.

[youkaichao]

@yma11 please resolve the conflict

[yma11]

@youkaichao Thanks for reminder. @comaniac I switched the fix back to only change hip and xpu code path since there is an possible issue on these platforms. When there is no ray cluster existing and trying to launch a new instance, Ray may can't detect correct GPU numbers thus will cause no GPU resources available for ray worker allocation. So we need give num_gpus as the argument in this case. That's why this specific code path exists here. FYI and thanks for your review.

[youkaichao]

@DarkLight1337 please help check is the error related or it occurs in the main branch previously?

[DarkLight1337]

It is a failure from main branch that has since been fixed. You can force merge this.