xx

Dockerfile

@@ -16,8 +16,8 @@ COPY benchmarks/run_vllm_bench.py /opt/script/run_vllm_bench.py
 # Make scripts executable
 RUN chmod +x /opt/script/start-vllm
 
-# Create symlink for backward compatibility
-RUN ln -sf /opt/script/start-vllm /usr/local/bin/start-vllm
+# Replace the container's start-vllm with our improved version
+RUN cp /opt/script/start-vllm /usr/local/bin/start-vllm
 
 # Set working directory
 WORKDIR /opt

scripts/start_vllm.py

@@ -138,8 +138,11 @@ def nuke_vllm_cache():
             print(f" Failed: {e}")
 
 def configure_and_launch(model_idx, gpu_count):
+    print(f"DEBUG: configure_and_launch called with model_idx={model_idx}, gpu_count={gpu_count}")
     model_id = MODELS_TO_RUN[model_idx]
     config = MODEL_TABLE[model_id]
+    print(f"DEBUG: model_id={model_id}")
+    print(f"DEBUG: LOCAL_MODEL_DIR={LOCAL_MODEL_DIR}")
     
     # Determine whether we have a local copy to serve. Try multiple fallbacks:
     # 1) LOCAL_MODEL_DIR/<owner>/<repo>