[Frontend] Publish Prometheus metrics in run_batch API (#7641)

vllm/entrypoints/openai/run_batch.py

@@ -3,6 +3,7 @@ from io import StringIO
 from typing import Awaitable, Callable, List
 
 import aiohttp
+from prometheus_client import start_http_server
 
 from vllm.engine.arg_utils import AsyncEngineArgs, nullable_str
 from vllm.engine.async_llm_engine import AsyncLLMEngine
@@ -56,6 +57,24 @@ def parse_args():
                         'ID numbers being printed in log.'
                         '\n\nDefault: Unlimited')
 
+    parser.add_argument("--enable-metrics",
+                        action="store_true",
+                        help="Enable Prometheus metrics")
+    parser.add_argument(
+        "--url",
+        type=str,
+        default="0.0.0.0",
+        help="URL to the Prometheus metrics server "
+        "(only needed if enable-metrics is set).",
+    )
+    parser.add_argument(
+        "--port",
+        type=int,
+        default=8000,
+        help="Port number for the Prometheus metrics server "
+        "(only needed if enable-metrics is set).",
+    )
+
     return parser.parse_args()
 
 
@@ -184,4 +203,12 @@ if __name__ == "__main__":
     logger.info("vLLM batch processing API version %s", VLLM_VERSION)
     logger.info("args: %s", args)
 
+    # Start the Prometheus metrics server. LLMEngine uses the Prometheus client
+    # to publish metrics at the /metrics endpoint.
+    if args.enable_metrics:
+        logger.info("Prometheus metrics enabled")
+        start_http_server(port=args.port, addr=args.url)
+    else:
+        logger.info("Prometheus metrics disabled")
+
     asyncio.run(main(args))