/playground) is for short-lived model trials. Sessions have a TTL and are reclaimed automatically. Use Model Deployments when you need a stable endpoint, tokens, recovery, or revisions.
Playground: start a vLLM service from a training export or an HF id.
Starting a service
- Choose a model: the training artifacts dropdown (runs with an hf_export), or directly enter an HF id / shared-storage path;
- Choose GPU count (= tensor parallelism) and TTL;
- Launch. The status card updates in real time: starting → ready → remaining TTL.
Chat
- Streaming output, adjustable system prompt / temperature / max tokens;
- Conversation history stays on the page (cleared on refresh, never stored).
A/B comparison
Pick two ready services side by side: the same input is sent to both simultaneously and the outputs render in parallel — alignment differences between SFT vs GRPO, or v1 vs v2, become directly visible.Management
The page lists all of my services: endpoint address, GPUs occupied, remaining TTL, renew / stop buttons. Services are automatically reclaimed when their TTL expires (guarding against “started it and forgot to shut it down”).Playground is a configurable feature that admins can open up by role; the slurm backend does not support it yet (training nodes usually have no inbound network). See the Playground guide for full semantics.