NVIDIA open-sourceia o coletivo Reshard para acelerar refit de pesos RL até 5,7×
![[object Object]](https://developer-blogs.nvidia.com/wp-content/uploads/2026/11/green-cube.webp)
NCCL M2N Reshard move só shards que se intersectam; NeMo RL mostra 2,7–5,7× vs gather-and-broadcast.
Por que importa
A NVIDIA descreveu o Reshard, coletivo que trata o refit RL como transferência DTensor-a-DTensor.
Fatos principais
- New Reshard collective for DTensor-to-DTensor refit
- Open NCCL M2N Reshard reference implementation
- 2.7×–5.7× faster than gather-and-broadcast in NeMo RL
- Benchmarked on DGX H100 and GB200
- Hybrid path: FFN projections via Reshard; small weights legacy
- Aimed at trillion-parameter RL post-training payloads