fixing vllms

handler.py

@@ -59,6 +59,8 @@ class EndpointHandler:
                 pass
 
         # Auto-detect tensor parallel size from visible devices
+        # Some server environments require 'fork' for worker processes
+        os.environ.setdefault("VLLM_WORKER_MULTIPROC_METHOD", "fork")
         visible = os.environ.get("CUDA_VISIBLE_DEVICES")
         if visible and visible.strip():
             try:
@@ -80,6 +82,8 @@ class EndpointHandler:
             pipeline_parallel_size=1,
             gpu_memory_utilization=0.95,
             dtype="auto",
+            distributed_executor_backend="mp",
+            enforce_eager=True,
             trust_remote_code=True,
         )
         self.processor = AutoProcessor.from_pretrained(