RL rollouts
Keep your trainer. Run rollout inference on Grid. Reserved capacity, quoted.
RL rollouts are sampling for teams that already have a trainer. You keep the algorithm. Grid serves the completions.
This is distinct from Scoped Agent RL. In an Agent RL pilot, Forward scopes the trajectory dataset, deterministic verifiers, and policy optimization loop with your team. Rollouts are the inference sampling layer only for teams that already run their own trainer.
When to use it#
Use rollouts when the trainer is yours and the bottleneck is high-throughput, sticky, multi-turn sampling.
Use Scoped Agent RL when you want Forward to co-design the verifier harness and run policy updates with your team.
What Grid provides#
- OpenAI-compatible chat and completions for rollout traffic.
- Session stickiness so turns of one trajectory can reuse prefix cache on the same replica.
- Capacity that can sit next to production inference, quoted as a reservation.
Weight-update mechanics, regions, and concurrency are sized in the engagement. We do not publish swap times or compression ratios we have not measured on our fleet.
Availability#
Reserved capacity. Talk to us from RL rollouts. There is no anonymous Serverless path for this product.