GPU-accelerated Weight Teleportation

Miles v0.1 introduced a GPU-accelerated weight teleportation system that cuts end-to-end weight synchronization to under 4 seconds and sampler pause time below 0.23 seconds in its Ampersand production reinforcement learning configuration, using a Blackwell-native compression and decompression pipeline built on delta weight sync. The design combines massively parallel GPU compression, a curated double-compression algorithm with heterogeneous CPU-GPU decompression, and hardware-native decompression algorithms that fully overlap with ordinary GPU compute, while keeping memory buffer overhead tiny across trainers and samplers. In a proxy workload using an NVFP4 GLM-5.3 model with 445 GB of canonical weight tensors, only 0.085% of the checkpoint is mutated per steady-state update, with trainers on 8 nodes and 64 B300 GPUs and each rollout receiver running two TP4 engines on a single 8-GPU B300 node. The result matters because large-scale RL requires training and sampler weights to match closely, and terabyte-scale weight updates have typically forced a tradeoff between high-speed intra-cluster networking and meaningful synchronization delays. What to watch next is whether the approach holds up beyond the Ampersand configuration and sparse-update proxy, and how broadly the Blackwell Decompression Engine pipeline is adopted across Miles' production RL workloads.