NVIDIA's internal tests found its Reshard method shortened DeepSeek-V3's weight refit on DGX H100 to 2.6 seconds from 14.9, trimming a recurring transfer step in reinforcement-learning workflows.
The test used a 671-billion-parameter model. On DGX GB200, the same transfer took 2.3 seconds, compared with 6.1 seconds under gather-and-broadcast. Across both DGX systems, reported speedups ranged from 2.7 to 5.7 times.
Training GPUs update a model during reinforcement-learning workflows, while separate GPUs run the revised version. The latter produce output sequences, called trajectories, that feed into later rounds. Those model outputs also serve as feedback for later updates. Because the process repeats, each weight transfer adds time to the workflow.
Reshard handles the exchange by matching tensor layouts across sending and receiving GPUs and routing the needed shards without making a full temporary tensor. By contrast, gather-and-broadcast rebuilds a complete tensor before distribution, although inference GPUs may use only selected portions. This can send data the receiving hardware does not need. NVIDIA says the approach is meant to reduce repeated traffic and spread transfers among GPUs. Collective operations let participating processors coordinate those exchanges.
In a separate NVIDIA test, a 256-GPU NVL72 cluster using NDR InfiniBand synchronized weights in 2.77 seconds, down from 5.78 seconds; the full reinforcement-learning step was 12.7% shorter. In a feed-forward mixture-of-experts layer test, the transfer took 9.8 milliseconds; flat direct sends required 77.3 milliseconds.
NVIDIA's NCCL extensions repository hosts an open-source implementation.
