test: a few test configuration updates to have chunked prefill tests pass on Spyre

[tjohnson31415]

Description

In our internal CI we run unit tests on Spyre devices. This PR fixes some of the new Chunked Prefill tests to be able to pass on Spyre.

The two issues causing tests to fail:

• chunk size greater than max model length (due to vLLM default of 2048 or granite TP4 detection setting it to 4096)
• some edge case I don't fully understand in test_single_cp_prefill causing failure during inference with DtException: No matching compiler iter found

Also found that having @pytest.mark.cpu on all tests in test_spyre_cp_scheduler_steps.py is incorrect (the mark is automatically applied from the backend parameterization).

[github-actions]

👋 Hi! Thank you for contributing to vLLM support on Spyre.
Just a reminder: Make sure that your code passes all the linting checks, otherwise your PR won't be able to be merged. To do so, first install the linting requirements, then run format.sh and commit the changes. This can be done with uv directly:

uv sync --frozen --group lint --active --inexact

Or this can be done with pip:

uv pip compile --group lint > requirements-lint.txt
pip install -r requirements-lint.txt
bash format.sh

Now you are good to go 🚀

[maxdebayser]

Nice simplification!

[maxdebayser]

Nice catch. Maybe in sendnn the max model len has to be a multiple of 64

[tjohnson31415]

I was wondering something like that as well, but values like 576, 640, 768 also didn't work in my testing.

[maxdebayser]

LGTM, thanks for the fixes!

[joerunde]

In that case wouldn't we also need to re-override the internal config for max_num_batched_tokens as well?

[tjohnson31415]

VLLM_DT_CHUNK_LEN currently takes precedence over user / vllm setting max_num_batch_tokens after the changes in #571

[joerunde]

very hot 🌶️