actually invoke vllm this time

vllm_shim_module.py

@@ -267,9 +267,12 @@ def strip_shim_from_pythonpath():
 def main():
     args = sys.argv[1:]
 
+    # Determine which vllm module was actually invoked so we exec the real one
+    # (could be vllm.entrypoints.cli.main, vllm.entrypoints.openai.api_server, etc.)
+    invoked_module = __name__  # e.g. "vllm.entrypoints.cli.main" or "vllm.entrypoints.openai.api_server"
     log("=" * 50)
     log("vLLM Custom Weights Shim")
-    log(f"  Invoked as: python -m {__name__} {' '.join(args)}")
+    log(f"  Invoked as: python -m {invoked_module} {' '.join(args)}")
     log("=" * 50)
 
     # Intercept --model / positional model if it's a URL
@@ -278,8 +281,8 @@ def main():
     # Strip our shim from PYTHONPATH so the real vLLM resolves correctly
     strip_shim_from_pythonpath()
 
-    # Build the real vLLM command
-    vllm_cmd = [sys.executable, "-m", "vllm.entrypoints.openai.api_server"] + modified_args
+    # Build the real vLLM command using the same module that was invoked
+    vllm_cmd = [sys.executable, "-m", invoked_module] + modified_args
 
     log(f"Launching vLLM: {' '.join(vllm_cmd)}")