NVIDIA · RL Weight Refit

Reshard collective (NCCL M2N Reshard)

NVIDIA introduces Reshard, a collective communication method that treats RL weight refit as a DTensor-to-DTensor transfer, moving only the data each inference GPU needs rather than gathering full tensors and broadcasting them. An open source reference implementation, NCCL M2N Reshard, is available in the NVIDIA/nccl-extensions repository. In NeMo RL benchmarks on DGX H100 and DGX GB200 systems, the company reports 2.7x to 5.7x faster refit than gather-and-broadcast across models including Qwen 3 235B and DeepSeek v3 671B. A hybrid mode applies Reshard to the large feed-forward projection weights and keeps gather-and-broadcast for smaller tensors.