Python: Troubleshooting

Common problems installing and building Python packages on Isambard systems. For everyday Conda, uv, and pip/venv usage, see the basics guide. For background on building packages from source on aarch64, see the advanced guide.

Installing a package fails

ERROR: Could not find a version that satisfies the requirement <package> / ERROR: No matching distribution found for <package>

No wheel is published for aarch64 (or for your Python version), and pip could not build the package from source either. Check whether the package is available through Conda instead — see finding aarch64-compatible packages. If not, you will need to build it from source; see the advanced guide.

Package fails to build with a compiler error (see CC/CXX)

The default system compiler is often too old to build modern research software. Set CC and CXX to a newer compiler, as described in compilers in the advanced guide, and retry the install.

Build succeeds, but fails to detect a pre-installed dependency (e.g. torch) or ARM-specific compiler flags

Pip's isolated build environment can hide pre-installed packages and system compiler flags from the build process. Try again with --no-build-isolation — see disabling build isolation in the advanced guide.

mpi4py installs but fails at runtime, or brings in its own MPI implementation

The Conda package for mpi4py bundles its own MPI implementation, which does not support the system's high-speed interconnect. Force a build from source against the system MPI instead — see MPI libraries in the advanced guide.

Environment problems

conda: command not found after starting a new session or an interactive job on a compute node

Conda is not activated in this shell. Activation does not persist between the login node and compute nodes, so run source ~/miniforge3/bin/activate again in the new session.

A package is installed, but import <package> still fails

This usually means the package was installed into the wrong environment, or pip/python3 is not resolving to the environment you think it is. Check the active environment with which python3, and use python3 -m pip rather than a bare pip to be sure which installation is being used — see pip vs python3 -m pip in the basics guide.