Broken python environment in Docker images

[napaalm]

🐛 Describe the bug

On image rocm/pytorch:rocm6.1_ubuntu22.04_py3.10_pytorch_2.4 (and probably all the others) pip check returns dependencies errors. This leads to pip reinstalling torch without ROCm support when any pip package which lists it as a dependency is installed. In particular, python-triton-rocm is not installed, and the wrong version of numpy is installed. See also pypa/pip#12902.

pip check output:

$ docker run --rm -it rocm/pytorch:rocm6.1_ubuntu22.04_py3.10_pytorch_2.4 bash
root@2857e8f82486:/var/lib/jenkins# pip check
torch 2.4.0.dev20240520+rocm6.0 requires pytorch-triton-rocm, which is not installed.
numba 0.55.2 has requirement numpy<1.23,>=1.18, but you have numpy 1.26.4.
scipy 1.8.1 has requirement numpy<1.25.0,>=1.17.3, but you have numpy 1.26.4.

Versions

No response

[ppanchad-amd]

Hi @napaalm. Internal ticket has been created to investigate your issue. Thanks!

[napaalm]

@ppanchad-amd thank you. Please check all published images for these issues, and not just the one I reported.

[zichguan-amd]

Hi @napaalm, thanks for reporting. I've checked the recent images for ROCm 6.2.2 and 6.2.3 and they don't have this issue anymore.