[BugFix] Increase timeout for startup failure test

[njhill]

This test is now failing on main most of the time due to the timeout not being large enough.

I'm still confused why this started failing recently. The problem is that just creating VllmConfig takes 10 seconds before we even create the LLM, which is due to the model class being loaded in a separate process to be able to read some config from it:

vllm/vllm/model_executor/models/registry.py

Lines 324 to 330 in 2858830

	def inspect_model_cls(self) -> _ModelInfo:
	return _run_in_subprocess(
	lambda: _ModelInfo.from_model_cls(self.load_model_cls()))
	def load_model_cls(self) -> Type[nn.Module]:
	mod = importlib.import_module(self.module_name)
	return getattr(mod, self.class_name)

It would be good if we could find a way to retrieve the info in question (like is_multimodal_model) without this overhead - I assume it's adding to startup time in all cases, cumulatively contributing to CI times quite a bit, etc.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

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 either: Add ready label to the PR or enable auto-merge.

🚀

[DarkLight1337]

@hmellor maybe this is related to the recent config refactor? Another possibility is that the overall import time of vLLM has increased significantly for whatever reason.

The code related to retrieving model info has been here for many months already, so that by itself couldn't be the cause. This code exists to get around CUDA re-initialization error because importing the model files may initialize CUDA before vLLM Engine is started.

[houseroad]

Good for short mitigation. Maybe we can create an issue for tracking this? Sounds like a good ramp up task?

[hmellor]

From this profile we can see that the actual retrieval of the model only takes about 1s (bottom thread) but the overhead of creating the new process for is_multimodal_model is unusually large (10s, as @njhill originally said)

[DarkLight1337]

Let's unblock CI first. We can solve the underlying problem later