| """Applied automatically in every Python process on PYTHONPATH. | |
| vLLM spawns worker subprocesses that re-import modules; without this file | |
| the monkey-patch would be missing in workers and weight loading would fail | |
| for the heterogeneous per-layer expert counts. | |
| """ | |
| import os | |
| if os.environ.get('VLLM_PRUNED_PATCH_DISABLE') != '1': | |
| try: | |
| import vllm_pruned_patch | |
| vllm_pruned_patch.apply() | |
| except Exception: | |
| pass | |