[Bugfix] use float32 precision in samplers/test_logprobs.py for comparing with HF

[tdoublep]

Fixes #6408

This PR changes the precision in tests/samplers/test_logprobs.py from half to float.

This is needed because the test is comparing the actual values of the logprobs against the equivalent outputs from HF. There is precedent established for doing this in other tests (see e.g., here or here).

This change ensures that the test does not fail on an H100 GPU.

[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 trigger fastcheck CI to run, which consists only a small and essential subset of tests to quickly catch errors with the flexibility to run extra individual tests on top (you can do this by unblocking test steps in the Buildkite run).

Full CI run is still required to merge this PR so once the PR is ready to go, please make sure to run it. If you need all test signals in between PR commits, you can trigger full CI as well.

To run full CI, you can do one of these:

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

🚀

[tdoublep]

/ready

[tdoublep]

CI failure looks real - I can reproduce locally on an L4 GPU

triton.runtime.autotuner.OutOfResources: out of resource: shared memory, Required: 132096, Hardware limit: 101376. Reducing block sizes or `num_stages` may help.

Haven't seen that before...

[tdoublep]

I was able to solve the error by halving the number of blocks in the prefix prefill kernel when torch.float32 is used. This should not affect normal runtime behaviour when using torch.float16, but allows us to run CI tests in torch.float32 on L4 GPUs.

[tdoublep]

@simon-mo auto-merge got disabled on this one, but now it is good.