Nvidia releases Reshard implementation to speed reinforcement learning weight transfers

Nvidia has made an open source implementation of Reshard available to transfer updated model weights between training and inference GPUs. The company reports speedups of 2.7x to 5.7x over conventional gather-and-broadcast transfers in NeMo RL benchmarks. Reshard sends the portions of model weights each destination needs, then uses NVLink exchanges to supply replicas without duplicating network traffic. This addresses differing distributions of weights across training and inference GPUs during reinforcement learning. The NeMo RL implementation takes a hybrid approach: dominant feed-forward projection weights use Reshard, while other weights retain conventional transfers. The reported improvements measure weight refit latency rather than overall training speed.