What happened
NVIDIA published “Accelerating Reinforcement Learning Weight Refit with the Reshard Collective” on November 6, 2026. Refit is the transfer of updated policy weights from training GPUs to inference GPUs in disaggregated RL post-training. NVIDIA estimates gather-and-broadcast logical payloads near 10 TB per step for the largest open models (about 11.2 TB for Kimi K3 BF16 and 9.6 TB for Qwen 3.8 2.4T). Reshard reformulates the transfer as a DTensor-to-DTensor move of only the theoretical minimum shards. NCCL M2N Reshard is open source in nvidia/nccl-extensions. In NeMo RL benchmarks, Reshard refit was 1.9 s vs 6.4 s for Qwen 3 235B on DGX H100 (about 3.4×), 2.6 s vs 14.9 s for DeepSeek v3 671B on H100 (about 5.7×), and 2.3 s vs 6.1 s for the same DeepSeek layout on GB200. A hybrid path reshards feed-forward projections (~95% of parameters) and keeps gather-and-broadcast for smaller weights.
Why it matters
Teams running large MoE RL loops can spend seconds every step just moving weights. A named collective and a published latency table let you ask whether the bottleneck is still the optimizer or the refit path.
Enable in NeMo RL with policy.generation.refit_transport: nccl_reshard when not colocated.
A refit-only microbenchmark for Qwen3.8 2.4T BF16 took 4.3 seconds on NVIDIA’s account.
NVLink-domain grouping avoids sending the same shard twice to data-parallel replicas.