Issue with building PyTorch 2.1.2 using EasyBuild 5

[lcniel]

There seems to be an issue with the new handling of use_pip and buildcmd. Basically, for some reason the code ends up here:

easybuild-easyblocks/easybuild/easyblocks/generic/pythonpackage.py

Line 836 in e0de69f

build_cmd = f"{self.python_cmd} setup.py {build_cmd}"

        if self.use_setup_py:

...
            if not build_cmd:
                build_cmd = 'build'  # Default value for setup.py
            build_cmd = f"{self.python_cmd} setup.py {build_cmd}"

But in the easyconfig, e.g. PyTorch-2.1.2-foss-2023a-CUDA-12.1.1.eb, 'buildcmd' is set: buildcmd = '%(python)s setup.py build' # Run the (long) build in the build step which leads to an incorrect command being parsed together and the build failing. From the log, it seems like use_pip should be correctly set in the PyTorch easyblock, but it does not appear to actually be used... Not sure what's going on here.

[lcniel]

ping @Micket

[boegel]

Possibly related relatively recent PR:

• inject custom '%(python)s' template value before getting value of 'buildcmd' custom easyconfig parameter in PythonPackage easyblock #3539

[Micket]

build_cmd is a dangerous name. In some easyblocks, it means the first, primary, executable to run for the build step. But in this easyblock it seems to mean the first argument, i.e. the "target" for python setup.py.

Now, i would very much like to figure out how and where this changed as to track down if we have a larger systematic error with EB5, or if it just affects PyTorch somehow.

[Flamefire]

One major issue is that the attempt to use pip only for the latest PyTorch versions introduced in #3079 does not work as expected:

• We set the default to None
• Call PythonPackage.__init__
• That calls determine_install_command
• Then we change the current value of use_pip
• in build_step use_setup_py is not updated and we run into the branch that appends the build_cmd to setup.py

As I agree with @Micket that buildcmd is wrongly named I opened a PR to use a better name: #3575

The issue here should be fixed with #3574 which is related and required independent of the other PR although that helps too.

[boegel]

This may be introduced via #3539, but not sure...

[Flamefire]

This may be introduced via #3539, but not sure...

No that didn't change the behavior. It would be the same without but possibly not resolving the template so fail for that additionally.

The issue is rather easybuilders/easybuild-easyconfigs#20004 that remove use_pip = True from the EC causing use_setup_py to be set by PythonPackage which is never reset