centralize VLLM_WORKER_MULTIPROC_METHOD (#728)

Signed-off-by: Kyle Sayers <[email protected]>

Author: kylesayrs

examples/models/vllm_qwen2vl.sh

@@ -4,7 +4,6 @@
 
 # cd ~/prod/lmms-eval-public
 # pip3 install -e .
-export VLLM_WORKER_MULTIPROC_METHOD=spawn
 export NCCL_BLOCKING_WAIT=1
 export NCCL_TIMEOUT=18000000
 export NCCL_DEBUG=DEBUG

lmms_eval/models/vllm.py

@@ -60,6 +60,7 @@ def __init__(
                     eval_logger.warning(f"Failed to parse JSON-like string for argument '{key}': {value}")
 
         # Set up vllm client
+        os.environ["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn"
         self.client = LLM(
             model=self.model_version,
             tensor_parallel_size=tensor_parallel_size,