NVIDIA abre el colectivo Reshard para acelerar el refit de pesos RL hasta 5,7×
![[object Object]](https://developer-blogs.nvidia.com/wp-content/uploads/2026/11/green-cube.webp)
NCCL M2N Reshard mueve solo shards que se solapan; NeMo RL muestra 2,7–5,7× frente a gather-and-broadcast.
Por qué importa
NVIDIA describió Reshard, un colectivo que trata el refit RL como transferencia DTensor-a-DTensor.
Datos clave
- 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