Robotics
TensorRT

FLUX 3 Action DROID on Jetson AGX Thor

Optimized runtime for the FLUX 3 Action DROID policy on NVIDIA Jetson AGX Thor, in two precisions: nvfp4-fp8 and bf16.

Requirements

  • NVIDIA Jetson AGX Thor (compute capability 11.0), JetPack 7.2.1, TensorRT 10.16.2.10, CUDA 13.2, PyTorch with CUDA.

Test

python run.py --precision nvfp4-fp8 --inputs example/nvfp4-fp8/synthetic-inputs.safetensors --reference example/nvfp4-fp8/synthetic-reference.safetensors
python run.py --precision bf16 --inputs example/bf16/synthetic-inputs.safetensors --reference example/bf16/synthetic-reference.safetensors
python verify.py

Prepare real inputs from Black Forest Labs' released FLUX 3 Action DROID policy and checkpoint (FLUX Kommunity License). A synthetic example is included so the runtime can be tested without that download.

Licence

FLUX Kommunity License, non-commercial. See LICENSE.md and NOTICE.

Downloads last month
-
Video Preview
loading