Test in the Playground

Chat with your endpoint in the portal to see how the model answers, before you connect an application. The Playground is a chat window for an endpoint that serves chat models.

Before you begin

• An endpoint on vLLM or another framework of the vLLM family, in the Running, Idle, or Partially Available state.
• Permission to read model endpoints in the project. A Member needs a policy that grants read permission for the Inference service.
• An instruct or chat model. To test any other model, use your own client or the sample requests; see Call an endpoint with a token.

Open the Playground

In the Model Endpoints list, select Playground on the endpoint’s row. You can also select Playground on the endpoint’s details page. The chat opens with that endpoint selected. Playground is also its own entry in the side menu, under Inference.

Set the generation options

Five sliders shape the answer. Set them before you send a prompt; the defaults suit a first try.

Option	Range	Starts at	What it changes
Output Length	1 to 1024	512	The most tokens in the answer. A higher value allows a longer answer, which takes longer to write.
Temperature	0 to 2	0.5	How random the answer is. Around 0.1 to 0.3 gives steady, repeatable answers; around 0.7 to 1.0 gives more varied ones.
Top-P	0 to 1	1	Limits each word to the most likely choices whose chances add up to this share, such as 0.9. A lower value gives more focused text.
Top-K	1 to 100	1	Limits each word to this many of the most likely choices. A small value keeps the text focused, and a larger one, such as 50, allows more variety.
Repetition Penalty	0 to 2	0	Discourages words that already appeared. A higher value pushes the model harder to avoid repeating itself.

Check the answer

Type a prompt in the chat, and the model’s reply appears in the window. Try the same prompt at two temperatures to see how much the answers change. Playground becomes available once the endpoint is serving, so wait until it is Running.

When the answers look right, connect your application and send the same settings as request fields. See Use the OpenAI client and stream responses.