MLPerf Inference V3.1 on AGX Orin

[dino99]

Hi, I’m trying to reproduce MLPerf inference results v3.1 on AGX Orin system by following steps from https://github.com/mlcommons/inference_results_v3.1/closed/NVIDIA. System is flashed with Jetpack 6.0.
However, docker build fail with "/usr/local/cuda/include/cuda_bf16.hpp(4132): error: no suitable conversion function from “const __nv_bfloat16” to “unsigned short” exists
".
Does anyone know how to resolve it?
Detail error message is included inside errorMessage.txt.
errorMessage.txt (48.1 KB)

[kayccc]

The JetPack version with that MLPerf Inference 3.1 was originally tested with - JP 5.1.1.

inference_results_v3.1/closed/NVIDIA/systems/Orin_TRT.json at main · mlcommons/inference_results_v3.1 · GitHub

There is no reference for JetPack 6.0 yet, please roll back to use JetPack 5.1.1.

Thanks