AI BriefingNVIDIA Reshard Cuts RL Weight Transfer Time by Up to 5.7x
·2026.11.10 15:30
Key point
The Reshard operation announced by NVIDIA reduces weight transfer time in disaggregated RL environments by up to 5.7x.
Details
- NVIDIA announced the collective communication operation Reshard via its technical blog on November 6, 2026
- Achieved up to a 5.7x performance improvement by reducing unnecessary data during weight transfers between training and inference GPUs
- Reduced time from 14.9 seconds to 2.6 seconds on DGX H100 for the DeepSeek v3 671B model
- Calculates and transfers only overlapping shard regions by leveraging Device Mesh and Placement information from PyTorch DTensor
- Requires the Megatron training backend when applying NeMo RL, and supports vLLM or Megatron generation backends
- Supports BF16 and specific FP8/MXFP8 combinations; quantized weights are converted to BF16 before transfer
- Since the official documentation is for the nightly version, it is necessary to re-verify NeMo RL version support conditions when actually adopting it
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
NVIDIA Releases GR00T 2 Foundation World-Action Models for Cross-Embodiment Robotics
Nvidia Used Circular Financing to Keep GPU Prices High, WSJ Report ClaimsNVIDIA Tutorial: Serving GLM-5.3 NVFP4 on Dual DGX Stations with vLLM