Humans& introduced a GPU-accelerated method for lossless model weight updates in reinforcement learning. It compresses and streams delta weight masks between trainers and samplers. The approach uses Blackwell-native decompression and parallel GPU compression to reduce payload size. Sampler pause time drops below 0.23 seconds in production. End-to-end weight sync now takes under 4 seconds. The method avoids disk round trips by applying updates directly in GPU memory. It works with low-precision weights without loss. The design overlaps decompression with ordinary GPU compute. The pipeline saturates the Blackwell Decompression Engine. Intra-cluster RDMA is now slower for most workloads. The implementation is open source.
What changed
Weight updates were transferred as full model checkpoints or uncompressed delta masks, causing high latency and blocking samplers.
What it unlocks
Near-instant weight synchronization between trainers and samplers in reinforcement learning without blocking or high-speed intra-cluster networking.
- 445 GB model weights
- 0.085% weights updated per step
- 0.23 s sampler pause time
- 4 s end-to-end weight sync time
What you need to act on it
- Blackwell GPUs with Decompression Engine (DE) support
- NVFP4 GLM-5.3 model or similar architecture
- Distributed GPU cluster for training and rollout
Sources