Since the version 1.17, onnxruntime only support GLIBC >= 2.28, but some old OS release (e.g. CentOS7, RHEL7) could only embed GLIBC < 2.28.
In this case, installing an environment with conda and onnxruntime >= 1.17, as it is documented in Install with conda, will fail with the following pip error :
ERROR: Could not find a version that satisfies the requirement onnxruntime==1.18.0 (from versions: 1.12.0, 1.12.1, 1.13.1, 1.14.0, 1.14.1, 1.15.0, 1.15.1, 1.16.0, 1.16.1, 1.16.2, 1.16.3)
ERROR: No matching distribution found for onnxruntime==1.18.0
Pip detect the conflict between the the GLIBC and onnxruntime, but is not able to resolve it.
The first alternative is to use the an containerizer environment (docker/podman/singularity). But, in case no one of these solution can be used, conda can resolve the issue.
Conda doen't use the pypi onnxruntime wheels, while it publishes its own packages. The conda onnxruntime libs are statically linked (in contrary to the pypi ones) and that seems to solve the issue (maybe they compile in their own for a glibc 2.27, and/or they embed many stuff that solves the issue ...).
It is then recommended to manually edit the requirements.txt file to move the onnxruntime dependency to the env.yaml file, such that conda install it instead of pip:
name: py4cast
channels:
- pytorch
- nvidia
dependencies:
- python=3.11.9 # The reference python version for the project
- pytorch-cuda=12.1 # The reference cuda version for the project
- pytorch==2.4.1 # The reference pytorch version for the project
- onnxruntime-gpu==1.19.2
- pip
- pip:
- -r requirements.txt
- -e . # install py4cast in editable mode