Add workers CLI argument to pass to uvicorn app invocation

Author: tgasser-nv

nemoguardrails/benchmark/Procfile

@@ -1,9 +1,9 @@
 # Procfile
 
 # NeMo Guardrails server
-# main: uvicorn nemoguardrails.server.api:app --workers 1 --port 9000
+# main: uvicorn nemoguardrails.server.api:app --workers 4 --port 9000
 gr: poetry run nemoguardrails server --config configs/guardrail_configs --default-config-id content_safety_colang1 --port 9000
 
 # Guardrails NIMs for inference
-app_llm: poetry run python mock_llm_server/run_server.py --port 8000 --config-file configs/mock_configs/meta-llama-3.3-70b-instruct.env
-cs_llm: poetry run python mock_llm_server/run_server.py --port 8001 --config-file configs/mock_configs/nvidia-llama-3.1-nemoguard-8b-content-safety.env
+app_llm: poetry run python mock_llm_server/run_server.py --workers 4 --port 8000 --config-file configs/mock_configs/meta-llama-3.3-70b-instruct.env
+cs_llm: poetry run python mock_llm_server/run_server.py --workers 4 --port 8001 --config-file configs/mock_configs/nvidia-llama-3.1-nemoguard-8b-content-safety.env

nemoguardrails/benchmark/mock_llm_server/run_server.py

@@ -71,7 +71,12 @@ def parse_arguments():
     parser.add_argument(
         "--config-file", help=".env file to configure model", required=True
     )
-
+    parser.add_argument(
+        "--workers",
+        type=int,
+        default=1,
+        help="Number of uvicorn worker processes (default: 1)",
+    )
     return parser.parse_args()
 
 
@@ -110,6 +115,7 @@ def main():  # pragma: no cover
             reload=args.reload,
             log_level=args.log_level,
             env_file=config_file,
+            workers=args.workers,
         )
     except KeyboardInterrupt:
         log.info("\nServer stopped by user")