Missing package: medpython

[alonyado]

Package name

medpython

Package version

1.1.0

PyPI URL

https://pypi.org/project/medpython/

piwheels URL

https://www.piwheels.org/project/medpython/

Python version

• Python 3.9
• Python 3.11
• Python 3.13

This is for Raspberry Pi users only. If your issue does not relate to usage on Raspberry Pi using the official distro, please don't create an issue here.

yes

The project page includes links to existing issues about the package. Please check this issue doesn't already exist.

yes

Are you the project maintainer?

• I am the maintainer

More information

​Memory Issues during Build:

The package build currently fails at approximately 30% due to insufficient memory. If you can limit the build process to a single core to reduce peak memory consumption it might solve the issue.

Note: The build has been verified to succeed on Python 3.11.

[MichaIng]

CMAKE_BUILD_PARALLEL_LEVEL=1 or CMAKE_BUILD_PARALLEL_LEVEL=2 might solve it. See also here: #341

[alonyado]

Please see my suggested changes in medpython.
We will check if that works in the next version.

[alonyado]

I tested the suggestion to restrict the parallel build. The log timestamps confirm that the code is no longer compiling in parallel. However, it didn't resolve the issue.
We are still hitting the same virtual memory exhausted: Cannot allocate memory error.
The only difference now is that cmake ends the process immediately upon hitting the error, whereas before, the build would continue running other parallel processes.
Given that a single job is still exhausting the memory, I suspect we might actually need a 64-bit compiler for this library.

I guess the current build is only slightly above the 32bits memory limit, because in python 3.11 with Gcc 12.2 the memory was sufficient.

[MichaIng]

So this means a single job is hitting the limit all alone 🤔. Indeed it would only work on 64-bit then, but piwheels is explicitly 32-bit (PyPI provides aarch64 wheels for almost everything). It would hence require some cross-compiling, outside of scope, AFAIK.

So this means the project currently cannot build on 32-bit hosts at all. I guess there are some compiler and/or linker flags to reduce memory usage. I remember some Rust projects which had similar issues, with a single job hitting >3 GiB memory at LTO was the issue. A low-memory target with adjusted LTO flags, used by default on 32-bit systems, solved things. Maybe for medpython something similar is possible?

[alonyado]

Just to clarify, PyPI doesn't build aarch64 wheels (or any wheels) automatically.
The package maintainers have to build and upload them. I did already ship the aarch64 wheels for medpython.

Your diagnosis makes perfect sense. It looks like the build is just barely crossing that 3GB memory limit, which perfectly explains why it succeeded specifically under GCC 12.2 with Python 3.11.

I will try to add some flags to limit the memory usage when building on 32-bit and report. Thanks for the great suggestion.

[MichaIng]

Just to clarify, PyPI doesn't build aarch64 wheels (or any wheels) automatically.
The package maintainers have to build and upload them. I did already ship the aarch64 wheels for medpython.

Right, thanks for clarifying. Most project devs/maintainers (including yourself) upload aarch64 wheels to PyPI since those are sufficiently easy to build, with wide support across distros, trusted container base images, hosted runners (like the GitHub ARM runners) etc. For armv7l it is a lot more tedious to do builds in CI/CD, and for that rare armv6l/armv6hf that RPi 1/Zero models are, it is almost impossible, without even a single trusted (verified publisher) Raspbian Docker base image (after balenalib archived theirs), no manylinux images, annoying workarounds/mocks, and even QEMU emulation is at best armv7l with NEON. Hence piwheels has its value providing wheels for these architectures on native 32-bit RPi hosts and official Raspberry Pi OS (32-bit) base images. And consequently, it hits 32-bit only issues, as these are not commonly tested in upstream CI/CD pipelines anymore 😄.

I will try to add some flags to limit the memory usage when building on 32-bit and report.

Awesome! 👍