NVIDIALaunchDevelopersSeptember 21, 2026

NVIDIA Dynamo-Triton 26.07 adds TensorRT multi-device inference

Read original source →developer.nvidia.com

TensorRT 11.0's multi-device inference lets one TensorRT network run across multiple GPUs via NCCL collectives; Dynamo-Triton 26.07 exposes it so one KIND_MODEL instance owns several GPUs. Demo serves Cosmos 3 Nano's 36-layer denoising transformer, splitting 44,160 video tokens across up to eight GPUs.

1 source

More stories today

Open the live feed