Open arjunsuresh opened 9 months ago
I am getting similar errors with scipy==1.9.1. In my case it seems that the root cause is numpy>=1.23 used.
@nvyihengz that's true. Considering numpy 1.22 is quite old we should fix this for inference 5.0
@anandhu-eng
@arjunsuresh so far NVIDIA has been using a separate virtual env for SDXL accuracy testing (https://github.com/mlcommons/inference_results_v4.0/blob/main/closed/NVIDIA/code/stable-diffusion-xl/tensorrt/accuracy_requirements.txt), which is not a very elegant approach. I agree this issue should be looked into in the next round.
Thank you @nvyihengz Interestingly there is no numpy
in the accuracy environment for SDXL whereas the original Nvidia requirements has numpy==1.22
. We had to install numpy==1.22
in the accuracy environment to get rid of this error while running Nvidia v4.0 code for SDXL. (We were reproducing this environment in a non-Nvidia docker container).
@arjunsuresh If I remember correctly, scipy will install numpy as a dependency. The venv is built from bare metal following the reference implementation (https://github.com/mlcommons/inference/blob/master/text_to_image/requirements.txt) so it does not inherit NVIDIA's container packages.
Please see below for the detailed output. The run is done on Nvidia RTX 4090 GPU.