fix: raise default VLLM_STARTUP_TIMEOUT to 1800s and expose it in the hub env schema

.runpod/hub.json

@@ -686,6 +686,16 @@
           "advanced": true
         }
       },
+      {
+        "key": "VLLM_STARTUP_TIMEOUT",
+        "input": {
+          "name": "vLLM Startup Timeout",
+          "type": "number",
+          "description": "Seconds to wait for the internal vLLM /health endpoint before failing the worker. Raise this for very large models (400GB+) where weight loading can exceed the default on contended hosts.",
+          "default": 1800,
+          "advanced": true
+        }
+      },
       {
         "key": "ENABLE_EXPERT_PARALLEL",
         "input": {

docs/configuration.md

@@ -72,7 +72,7 @@ These are consumed by the wrapper itself, not passed to vLLM:
 | `BASE_PATH`           | `/runpod-volume` | Build arg, not read at runtime: the image derives `HF_HOME` and `VLLM_CACHE_ROOT` (`$BASE_PATH/vllm-cache`) from it at build time, so both caches persist on a network volume. Setting it on an endpoint moves neither; set `HF_HUB_CACHE` (model downloads) and `VLLM_CACHE_ROOT` (compile cache) directly. |
 | `MAX_CONCURRENCY`     | `30`    | Max concurrent jobs per worker (RunPod concurrency modifier). vLLM queues internally beyond this. |
 | `VLLM_PORT`           | `8000`  | Loopback port the internal `vllm serve` binds to.                       |
-| `VLLM_STARTUP_TIMEOUT`| `1200`  | Seconds to wait for vLLM `/health` before failing the worker.           |
+| `VLLM_STARTUP_TIMEOUT`| `1800`  | Seconds to wait for vLLM `/health` before failing the worker.           |
 | `REQUEST_TIMEOUT`     | `3600`  | Per-request timeout to the vLLM server, in seconds.                     |
 | `VLLM_EXTRA_ARGS`     | —       | Raw extra CLI args (see above).                                         |
 | `VLLM_CONFIG_FILE`    | —       | Path to a `vllm serve` YAML config file (`--config`).                   |

src/main.py

@@ -47,7 +47,7 @@
 
 VLLM_HOST = "127.0.0.1"
 VLLM_PORT = os.getenv("VLLM_PORT", "8000")
-STARTUP_TIMEOUT = int(os.getenv("VLLM_STARTUP_TIMEOUT", "1200"))  # seconds
+STARTUP_TIMEOUT = int(os.getenv("VLLM_STARTUP_TIMEOUT", "1800"))  # seconds
 HEALTH_POLL_INTERVAL = 2  # seconds
 # How long to wait for vLLM to exit after SIGTERM before SIGKILL.
 SHUTDOWN_GRACE = 30  # seconds