| |
|
| |
Overview
Join CoreWeave & NVIDIA Dynamo for an evening on the infrastructure behind production inference & reinforcement learning post-training.
In RL post-training, training & inference start to blend together. The policy generates rollouts, training produces new weights, & those weights need to be serving again before the next round can start. How fast that loop turns depends on how quickly your inference stack can pick up a new checkpoint.
We'll start with the recent CoreWeave Forge launch from Fully Connected 2026. One new piece is RL Rollouts on CoreWeave Dedicated Inference, built on NVIDIA Dynamo. It hot-loads updated checkpoints into a live deployment with minimal downtime, so each round of rollouts uses the latest policy. CoreWeave, NVIDIA & You.com used it to post-train NVIDIA Nemotron 3.5 Lightning with RL in NeMo Gym, using You.com's web search API. Model reload latency improved 15x over baseline.
This meetup is for developers & researchers building & scaling inference & post-training systems. Topics include high-throughput serving, disaggregated inference, efficient rollouts, weight hot-swapping, SFT & RL pipelines, & closing the loop from production signal to a better model. Expect technical talks, real lessons from teams running these systems, & plenty of time to connect.
Agenda:
6:00 pm: Doors open
6:30 pm: CoreWeave talk: CoreWeave Forge: from production inference to post-training
6:45 pm: NVIDIA Dynamo talk: Inference at scale & fast weight updates for RL rollouts
7:00-9:00 pm: Mingle with fellow developers!
Resources & Legal
NVIDIA Privacy Policy: https://www.nvidia.com/en-us/about-nvidia/privacy-policy/
|
|
|
|
|
|
|
|
