EnragedAntelope
162c4b7dbe
v0.7.0 - Standalone-only loading, remove base_model input
Breaking Changes:
- ModelOptUNetLoader now requires v0.6.0+ checkpoints with embedded config
- Removed base_model input - standalone loading only
- Removed overlay loading path for legacy checkpoints
Features:
- count_model_storage() for accurate buffer+parameter counting
- Conv2d MXFP8/NVFP4 flatten path (no more FP8 fallback)
- Strip ModelOpt quantizers BEFORE native quantization
- Integration test: quantize → save → load → inference roundtrip
Documentation:
- README.md updated for standalone-only workflow
- All references to base_model removed from docs