NVIDIA releases open-source Reshard collective to speed RL weight refit

NVIDIA published an open-source reference implementation of Reshard, a new collective communication operation for copying updated model weights from training GPUs to inference GPUs during reinforcement learning post-training. The code, called NCCL M2N Reshard, is available in the NVIDIA/nccl-extensions GitHub repository. Rather than gathering each parameter in full and broadcasting it, Reshard frames refit as a DTensor-to-DTensor transfer, sending only the tensor regions each destination shard actually needs while handling differing parallelism layouts and replicated weights. On DGX H100 and DGX GB200 systems running NeMo RL, NVIDIA reports refit speedups of 2.7x to 5.7x over gather-and-broadcast across models including Qwen 3 235B and DeepSeek v3 671B. NeMo RL uses a hybrid scheme that applies Reshard to feed-forward projection weights and keeps gather-and-broadcast for smaller ones.