Starting a RunPod Serverless Endpoint

Starting a RunPod Serverless Endpoint | Generated by AI

Home 2026.07

Question: How do I actually start a RunPod Serverless vLLM endpoint when it shows 0 running workers, 0 jobs in progress, and 0 jobs waiting in queue?

Answer:

That status is normal. RunPod Serverless is scale-to-zero.

0 running workers
0 jobs

does not mean something is broken. It means:

No GPU is currently allocated.
No request has arrived.
You’re not paying for idle GPUs.

The lifecycle is:

Create Endpoint
        │
        ▼
0 workers (idle)
        │
        ▼
Send first request
        │
        ▼
RunPod allocates GPU
        │
        ▼
Downloads model
        │
        ▼
Starts vLLM
        │
        ▼
Processes request
        │
        ▼
Returns result
        │
        ▼
After idle timeout
        │
        ▼
Worker shuts down again

Option 1 — Use the RunPod UI (easiest)

Open your endpoint:

Requests

Use something like:

{
  "input": {
    "prompt": "Hello"
  }
}

Click Run.

The queue will become

1 waiting

then

1 running

then you’ll see

1 worker

during execution. The first request may take a few minutes because RunPod must provision a GPU and download the model. (Runpod Documentation)

Option 2 — Use curl

curl -X POST \
  https://api.runpod.ai/v2/<ENDPOINT_ID>/runsync \
  -H "Authorization: Bearer <RUNPOD_API_KEY>" \
  -H "Content-Type: application/json" \
  -d '{
        "input": {
          "prompt": "Hello"
        }
      }'

That single request automatically starts a worker. (Runpod Documentation)

Option 3 — OpenAI-compatible API

The vLLM template also exposes an OpenAI-compatible API after the worker is running, so you can point the OpenAI SDK or LiteLLM/OpenWebUI at it. (Runpod Documentation)

If it stays at 0 workers forever

Then something is wrong. Check:

Endpoint status is Active
A GPU type is selected
MIN_WORKERS isn’t unintentionally set to 0 if you expected an always-warm worker
The model name is correct (for example Qwen/Qwen3-8B)
Hugging Face token is configured if the model is gated
Endpoint logs for model download or OOM errors (Runpod Documentation)

Since you’ve been experimenting with SGLang and vLLM on RunPod, I’d recommend sending a simple runsync request first. If you paste:

the endpoint configuration,
the selected GPU,
the model name, and
the endpoint logs,

I can usually identify the startup issue within a minute.

References:

Back Donate