humans& proposes a GPU-accelerated weight transfer solution, reducing the RL weight synchronization pause to 0.23 seconds

Ziang Li and others at humans& have released a GPU-accelerated weight transfer solution that, under the Ampersand production RL configuration, reduces sampler pause time to under 0.23 seconds and end-to-end weight synchronization time to under 4 seconds. The solution is based on a bitwise XOR mask of incremental weights, combined with GPU-parallel compression, two-layer Snappy plus Zstd compression, the Blackwell decompression engine, and Triton kernels for in-place weight updates, and it overlaps decompression with SM computation. The related implementation has been open-sourced in Miles and SGLang.