NEWS · CODING · #3235
NVIDIA's NCCL M2N Reshard implements Reshard collective to speed RL weight refit up to 5.7x
The Reshard collective reformulates RL "refit" as a DTensor-to-DTensor transfer that moves only the theoretical minimum data between arbitrary source and destination layouts. NVIDIA's open-source NCCL M2N Reshard reference implementation (nvidia/nccl-extensions) reports 2.7×–5.7× refit speedups in NeMo RL benchmarks on DGX H100 and DGX GB200 (tested on models including Qwen 3 235B and DeepSeek v3 671B), and uses a hybrid mode that applies Reshard to dominant feed-forward projection weights while retaining gather-and-broadcast for smaller tensors.
KEY POINTS
- The Reshard collective reformulates RL "refit" as a DTensor-to-DTensor transfer that moves only the theoretical minimum data between arbitrary source and destination layouts.
- NVIDIA's open-source NCCL M2N Reshard reference implementation (nvidia/nccl-extensions) reports 2.7×–5.7× refit speedups in NeMo RL benchmarks on DGX H100 and DGX GB200 (tested on models including Qwen 3 235B and DeepSeek v3 671B), and uses a hybrid mode that applies Reshard to dominant feed-forward projection weights while retaining gather-and-broadcast for smaller tensors.
- By cutting redundant all-gather/broadcast transfers—which can reach ~10 TB per refit for current largest open models—Reshard reduces communication overhead and improves scalability of RL post-training in disaggregated GPU setups.
WHY IT MATTERS
By cutting redundant all-gather/broadcast transfers—which can reach ~10 TB per refit for current largest open models—Reshard reduces communication overhead and improves scalability of RL post-training in disaggregated GPU setups.