Use VLLM_WORKER_MULTIPROC_METHOD=spawn instead of --forked for tests

[tjohnson31415]

This PR lets us remove the requirement of --forked from our pytest tests.

The hang that is observed without --forked is due to the known issue with libgomp and threading (see this gcc bug report that is a "won't fix"). It is a common problem in Python dueo to the usage of native libraries behind the scenes. If a process is forked after an OpenMP thread pool has been created, then the child will not have a threadpool and the code hangs the next time code enters a parallel context.

Where this comes up in our tests is actually because we use transformers to compare the generation results. vLLM and PyTorch delay initializing the thread pool until it is needed. When just using vLLM in V1, this does not happen in the frontend process, so it is ok to use fork(), but using transformer's model.generate in the main process during the tests initializes the thread pool and causes the next attempt to create a vllm.LLM to hang in the forked worker process.

With spawn, the new process is created from scratch and creates a new OpenMP thread pool. But there are trade-offs here too, eg. using spawn in offline mode requires particular handling in a script. vLLM docs have a good summary of the trade-offs of this setting in REF. Because of that, I just set in in the test environment. Using the vllm cli to run the code actually defaults to spawn anyways (REF).

FIX #146

[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 🚀

[prashantgupta24]

bot:test
MARKERS="spyre"

[prashantgupta24]

bot:test
MARKERS="spyre and not cb"

[prashantgupta24]

bot:test
MARKERS="spyre and not cb"

[prashantgupta24]

bot:test
MARKERS="spyre and not quantized and not multi and not embedding and not cb"

[joerunde]

lpgtm