NVIDIA把大模型冷启动快了4倍，DeepSeek-V4 Pro启动只要2分钟，靠ModelExpress和RDMA，推理和训练都能提速。
NVIDIA AI使用ModelExpress(MX)将DeepSeek-V4 Pro的启动时间从8分钟缩短至不到2分钟，靠的是GPU-to-GPU RDMA快速传输权重。MX是NVIDIA Dynamo中的权重分发与缓存管理服务，它复用内核缓存，并通过NIXL让推理worker直接从其他GPU获取更新权重，避免集中广播。这一方法同样能加速推理和强化学习后训练。
We cut DeepSeek-V4 Pro startup from 8 minutes to under 2 minutes by moving weights over the fastest ...
We cut DeepSeek-V4 Pro startup from 8 minutes to under 2 minutes by moving weights over the fastest path to GPU memory with GPU-to-GPU RDMA. This was achieved using NVIDIA ModelExpress (MX), the weight distribution and cache management service in NVIDIA Dynamo, and this same approach speeds up both inference and RL post-training too. MX reuses kernel caches, while inference workers fetch updated weights directly from other GPUs over NIXL—avoiding centralized broadcasts and keeping weight movement off the critical path. 💬 2 🔄 1 ❤️ 13 👀 1462 📊 4 ⚡