You.com、NVIDIA 与 CoreWeave 直播演示 RL 权重热加载部署方案
做 RL 训练部署的工程师值得看这场直播：You.com、NVIDIA 和 CoreWeave 会现场演示用 Dynamo 在不中断服务的情况下热加载新权重，还公开了 HGX B200 上的实测数据和成本。想搭 on-policy RL 生产管线的团队可以直接抄方案。
You.com、NVIDIA 和 CoreWeave 将在 11 月 19 日（下周四）进行一场实操直播，解决 RL 训练中的工程瓶颈：每隔几小时更新策略后，如何在不中断推理服务的情况下把新 checkpoint 加载到推理 worker 上。
演示基于 NVIDIA Nemotron 3.5 Lightning 模型，使用 RL Rollouts 和 NVIDIA Dynamo 在 CoreWeave 基础设施上完成 checkpoint 热加载与 on-policy 权重切换。在 NVIDIA HGX B200 上的实测结果是准确率提升 22.94%，工具调用次数减少 30.24%。
对做 RL 训练部署的工程团队来说，这类 on-policy 权重同步是生产环境的核心难点，直播会给出具体蓝图、计算成本和前后对比 trace。
Updating RL policies every few hours creates an immediate bottleneck: getting new checkpoints onto inference workers without dropping serving availability.
Next Thursday, Nov 19th, engineers from @youdotcom, @NVIDIA and CoreWeave break down how to solve this in production today.
We’re walking through a live deployment of NVIDIA Nemotron 3.5 Lightning, using RL Rollouts and NVIDIA Dynamo on CoreWeave infrastructure to handle checkpoint hot-loading and on-policy weight transitions.
The direct outcomes on NVIDIA HGX B200: ✅ 22.94% accuracy gain ✅ 30.24% fewer tool calls
Get the exact blueprint, compute costs, and before-and-after traces. Register here: https://t.co/bQcQcevvH2
This report is categorized under AITOP's "Products" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include NVIDIA, Nemotron, CoreWeave, RL Rollouts.
Source attribution: curated from CoreWeave (@CoreWeave).
