humans& announced on Nov. 2 that it had open-sourced “weight teleportation,” a method for moving reinforcement-learning model updates from trainers to samplers. In reinforcement learning, trainers update a model’s weights while samplers use those weights to generate data for subsequent training steps. The company’s technical post describes keeping the two sets of weights closely matched as a challenge when updates involve large amounts of data.
The design starts by comparing old and new weights to produce a bitwise XOR mask of what changed. The trainer transfers that mask, which samplers apply to their existing weights to obtain the updated version. humans& describes massively parallel GPU compression and a double-compression approach with decompression split between CPUs and GPUs. It says the pipeline was designed around features of NVIDIA’s Blackwell chips and can move updates even when trainers and samplers are not in the same cluster.
In its Ampersand production RL configuration, humans& reports sampler pauses below 0.23 seconds and end-to-end weight synchronization below four seconds. Its announcement separately puts non-blocking compression and preparation at about three seconds. That preparation time is not sampler downtime. For its technical explanation, the company specifies an NVFP4 GLM-5.3 workload proxy with 445 GB of weight tensors, 0.085% of the checkpoint changed per update, a 64-B300-GPU trainer and an eight-B300-GPU rollout receiver. The reported timings are the company’s results for its configuration, not a general speed guarantee.
Comments