v2.3.0rc2 — CLI-only

CLI/library release candidate from commit b4278e71ed61283ae5de1684859d770ca062a6b2. Studio is deferred and excluded from this version.

• Modular backend, typed API, portable CLI commands and explicit validation gates.
• New native-layout policies for Iris-3B, Kandinsky 6 video/audio and Kandinsky 6 SR, with conservative precision boundaries and Iris alignment protection.
• Safer telemetry/output paths, API option validation and bounded self-test scratch cleanup.

Download the CLI wheel, source archive and SHA256SUMS below. The artifacts were built and verified by the passing release checks at this exact commit.

sha256sum -c SHA256SUMS
python -m pip install ./comfyui_native_quantizer-2.3.0rc2-py3-none-any.whl
comfyquant --no-config --version

Requires Python 3.12–3.13. Use an environment supplying the intended PyTorch build for conversion; optional native-layout dependencies are described in the README.

Local validation includes 57 self-tests, 102 core tests, tiny native Iris loading/forward and 18 bounded CUDA cases. Hosted Python 3.12/3.13 and distribution checks passed. These establish converter/package and scoped consumer contracts; full-model quality, complete Kandinsky extension forwards, offload and LoRA acceptance remain unverified. Unevaluated combinations remain experimental. See the validation record and support matrix.

Licensed under Apache-2.0. This GitHub prerelease does not publish to PyPI.