Runtime configuration helper for using vLLM safely via SnakeBridge/Snakepit.

This module composes SnakeBridge.ConfigHelper and adds a vLLM-specific safeguard around vLLM v1 multiprocessing (which can spawn child processes).

v1 multiprocessing (vLLM 0.14+)

vLLM can spawn subprocesses for the engine core. Under Snakepit, those subprocesses must be cleaned up deterministically on shutdown to avoid orphaned GPU processes holding memory.

Configure with:

config :vllm, v1_multiprocessing: :auto | :on | :off

• :off - always forces VLLM_ENABLE_V1_MULTIPROCESSING=0
• :on - forces VLLM_ENABLE_V1_MULTIPROCESSING=1 and fails fast unless Snakepit process-group cleanup is available
• :auto - enables only when safe; otherwise forces off with a warning

When v1 multiprocessing is enabled, this helper also sets VLLM_WORKER_MULTIPROC_METHOD=spawn (unless already provided) to avoid forking a multi-threaded Python gRPC server.