Calling a model deployment

OpenAI-compatible inference against a promoted deployment revision.

curl https://starforge.your-company.com/inference/dep-7c1e/v1/chat/completions \
  -H "Authorization: Bearer $DEPLOYMENT_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "default",
    "messages": [{"role": "user", "content": "Summarise this contract clause."}]
  }'

The path is stable, the body is whatever the serving engine accepts, and the response is whatever it returns. StarForge authenticates the call, routes it to the deployment’s currently promoted revision, and records it — it does not reshape the payload.

Base URL

https://<your deployment domain>/inference/{deployment_id}/v1

Point any OpenAI-compatible client at that, and it works unchanged:

Python

from openai import OpenAI

client = OpenAI(
    base_url="https://starforge.your-company.com/inference/dep-7c1e/v1",
    api_key=os.environ["DEPLOYMENT_TOKEN"],
)

response = client.chat.completions.create(
    model="default",
    messages=[{"role": "user", "content": "Summarise this contract clause."}],
)

Everything under /v1 is proxied through: chat/completions, completions, embeddings, models, and anything else the serving engine exposes. Streaming works.

Authentication

Only a deployment token opens this path. Not your login token, not an ingest token.

export DEPLOYMENT_TOKEN=sfd_...

Property	Behaviour
Scope	One deployment. A token for dep-7c1e is rejected by dep-9a44
Expiry	None. It works until somebody revokes it
Identity	The token, not a person. It keeps working after its creator leaves the company
Revocation	Immediate, from the console’s deployment page

Create and revoke tokens on the deployment’s page in the console. Each token is identified separately in the traffic record, so you can tell one caller’s usage from another’s inside a single deployment.

A deployment token is a bearer credential with no expiry. Put it in a secret store, not in a repository, and issue one per calling service so you can revoke one without breaking the others.

Which model answers

The model field in the body is passed to the serving engine. Use default unless the engine was configured to serve several names.

What actually answers is the deployment’s promoted revision — the one snapshot of model source and serving configuration that is currently live. Promoting a new revision changes what this URL returns without the URL changing. Rolling back does the same in reverse. Callers are not affected and do not need to know.

Errors

Code	Meaning
401	Missing, malformed, or revoked deployment token
404	No such deployment
503	The deployment is suspended, or has no ready revision to route to

Anything else comes from the serving engine and carries its wording.

Confirm it worked

curl -s https://starforge.your-company.com/inference/dep-7c1e/v1/models \
  -H "Authorization: Bearer $DEPLOYMENT_TOKEN" | jq

A model list means the token is good and a revision is serving. From there, the console’s deployment page shows request volume, latency, and — if Reflow is enabled — the buffer filling with traffic you can mine into the next version’s training data.

Playground sessions are not this

A Playground session is a short-lived vLLM instance for a human to chat with, started from a run’s artifacts and expiring on its own. It has no stable address and no deployment token. If you are wiring an application to a model, you want a deployment; if you are deciding whether a checkpoint is any good, you want the Playground.