Skip to content

sentencepiece riscv64 support #164

Open

Description

@gounthar

Tracking riscv64 wheel support for sentencepiece (no riscv64 wheels on PyPI for 0.2.2).

The native build trips abseil's subword-atomics static_assert under GCC 13, but RISE's cibuildwheel in manylinux_2_39_riscv64 (which ships GCC 14) builds it cleanly.

Build workflow: #165 (validated on native riscv64 hardware, 60 tests passing).

Activity

  1. gounthar commented on Jul 15, 2026

    @gounthar
    CollaboratorAuthor

    Licensing check (per the development guide's Licensing section), for the record before merge:

    I pulled the official sentencepiece-0.2.2-cp310-cp310-manylinux_2_27_x86_64...whl from PyPI and inspected its .dist-info. It carries License-Expression: Apache-2.0 in the metadata, but ships no LICENSE file, and no notices for the abseil (Apache-2.0) and protobuf (BSD-3) sources that are statically linked into _sentencepiece.so.

    Our riscv64 wheels build from the same setup.py/pyproject.toml, so they will match upstream's packaging exactly, including this gap. It is an upstream omission rather than anything the riscv64 build introduces.

    Flagging it rather than silently patching: the guide treats this case by case. Options are to accept parity with the upstream wheel, add a build patch that bundles the license texts into our wheels, or raise it upstream so the fix lands for every platform. Happy to do whichever the maintainers prefer.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions