Quantization-aware adapters and safe device movement
First-class quantization metadata, safer device movement across quantized models, auto-routing based on checkpoint info, and major test coverage expansion.
Release: InvarLock 0.3.0 - Capabilities metadata, auto-routing, and broader coverage
Highlights
- Quantization metadata and detection (methods/config/capabilities).
- Safe device movement to avoid
.to()foot-guns on quantized models. - Smarter auto-adapter routing + broader tests.
0.3.0 is foundational if you’re working with quantized or pre-quantized checkpoints. Instead of treating quantization like a weird edge case, the framework can represent it, detect it, and route adapters appropriately—so you spend less time fighting “why did this load differently?” issues.
The safe device movement work is especially practical: it helps avoid a common class of runtime errors where models manage device placement internally and don’t tolerate generic .to() calls.
If you’ve been building guard/eval workflows across different quantization setups, this release is essentially the “make this feel normal” moment.
For the immutable release record, read the tagged CHANGELOG.md for v0.3.0.
More in Release
Continue through nearby posts in the same reading thread.
Release
Large-model reload stability and B200 controls
Fixes a GPU memory leak during reload fallback, hardens B200 scripts, and adds practical controls for acceptance ranges and overhead measurement.
Release
Public evaluate pipeline and report schema v1
The initial public release on GitHub and PyPI: core evaluate pipeline, guard chain, schema v1, and the first docs/CLI surface.
Release
Calibration, determinism, and regression protection
`invarlock calibrate` arrives, determinism utilities mature, and regression harness + golden tracking help prevent silent policy drift.