Base images with torch 2.6 and 2.7 pip broken

[bzikst]

Hi there. I'm testing new compatibility matrix in v0.15.1 and it seems like a problem with pip in new images.

At least with
r8.im/cog-base:cuda12.6-python3.12-torch2.6.0
r8.im/cog-base:cuda12.6-python3.12-torch2.7.0

Error

 > [stage-1  4/34] RUN --mount=type=cache,target=/root/.cache/pip pip install --no-cache-dir /tmp/cog-0.15.1-py3-none-any.whl 'pydantic>=1.9,<3':
0.423 Traceback (most recent call last):
0.423   File "/root/.pyenv/versions/3.12.10/bin/pip", line 5, in <module>
0.423     from pip._internal.cli.main import main
0.423 ModuleNotFoundError: No module named 'pip'

[bzikst]

Upd: r8.im/cog-base:cuda12.4-python3.12-torch2.6.0 works just fine, maybe it's on cuda12.6 then

[8W9aG]

Is this still the case? I believe we have updated these base images recently

[bzikst]

Is this still the case? I believe we have updated these base images recently

Just checked, same error still there. The error occurs during Docker build but works fine in interactive container shell. It seems like there's an issue with pyenv setup in the base image where pip isn't properly installed/configured.