Make pypdfium2 an optional backend dependency

[mgiessing]

If I understand correctly docling-parse-v2 is the default backend used for pdf processing and pypdfium2 an alternative backend. Is it possible to make pypdfium2 an optional dependency during the installation of docling?

The reason I'm asking is, that the pdfium binary is difficult to compile on IBM Power (ppc64le) which is required for pypdfium2 and potentially also IBM z (s390x).
Most of the other dependencies/requirements are more or less available or easy to build but I'm struggling with pypdfium2.

Thank you very much!

[dolfim-ibm]

pypdfium2 is actually used also by the docling-parse-v2 to generate page images, so it cannot easily be disabled.

Are you by any change running on RHEL? Docling is available on RHEL AI with all its dependencies, and I think IBM Power is a supported architecture there.

[mgiessing]

Ah I wasn't aware pypdfium2 is used there underneath! In general I'm trying to run instructlab which depends on docling and ultimately provide manylinux compliant wheels (2014 is based on centos7 and 2_28 is based on Alma8) for ppc64le (for docling-parse).

I'm not sure RHEL AI is already supported on ppc64le, but if pdfium is part of that I'll get in contact with our porting team and ask about the status.

Thank you very much!

[tiran]

I feel your pain. PDFium and pypdfium2 are very hard to build from sources. It took me a long time to turn PDFium into an RPM. You can find a copy of my work at https://github.com/tiran/dist-git/tree/main/libpdfium .

I have packaged and built libpdfium for RHEL AI, which is based on RHEL 9.4. PDFium builds on x86_64, aarch64, ppc64le, and s390x. libpdfium is hard to build from sources. It starts with the fact that Google does not provide source tar balls for PDFium. One has to use the proprietary depot tool to retrieve the sources. The upstream project does not build a shared library, too. There are downstream patches to enable public headers and shared library support.