1.1.1

Fixes the normal pip installation failure reported in #27. Version 1.1.0 had conflicting Transformers requirements, which could cause an unpinned install to fall back to 1.0.5.

Fixes

• Support Transformers 5 and remove the uv-only dependency override, allowing ordinary pip dependency resolution.
• Load the PLaMo tokenizer without importing the checkpoint's unused PyTorch model configuration. MLX translation does not require PyTorch.
• Retain the compatible mlx-lm 0.31.x range and MCP <2 constraint.
• Add clean distribution-wheel installation and tests on macOS/Linux with Python 3.10–3.14. The main-branch real-model integration job now tests the installed wheel as well.

Implementation: #28.

Install or upgrade

python -m pip install --upgrade plamo-translate==1.1.1
python -m pip check
plamo-translate --version

Available on PyPI. The release artifact remains a macOS Apple Silicon wheel. Linux wheel checks exercise package installation and mock tests; this release does not claim a new Linux/Intel inference distribution.

Stop an existing plamo-translate server before testing the updated installation, so the client uses the new runtime. No dependency override, --no-deps, or manual PyTorch installation is needed.

Thanks

Many thanks to @omonomo for the detailed report, exact reproduction steps, correct dependency diagnosis, and workaround in #27. This contribution identified a gap in our release validation and helped us add checks that protect future releases.

Post-publication validation

A fresh Python 3.14.8 / pip 26.2.1 venv on Apple M1 Max installed the published package from PyPI with unpinned ordinary pip and selected 1.1.1. pip check passed, PyTorch was absent, and the downloaded wheel hash matched the release artifact. All 62 unit/mock/numerical tests and 2 real-model CLI integration tests passed against that installed package.

The 4-bit model at revision 900f34ccc16d4e8592b1baa38889e2f53f6717e8 completed the long translation fixture at EOS, preserved eight paragraphs and the checked names, and produced matching full text through direct inference and both CLI output modes. Japanese-to-English translation also passed. These local checks cover specific regressions, not general translation quality.

The release PR CI passed all 17 applicable jobs, including ten clean wheel-install combinations. Hosted real-model CI runs separately on main; the real-model results above are from the published PyPI package tested locally.