Store lazy-compile and model-adapter cached outputs as weakref.ref and add resolve/prune helpers so caches don't strongly retain base models. Eviction and lookup now resolve weakrefs, prune dead entries, and only dispose non-collected patchers. Added tests to assert caches don't retain base models and to ensure pre-quantized 'as_needed' inputs still trigger prior-cache cleanup. Bumped package version to 2.2.1 and updated CHANGELOG.md to document the fix.
Introduce experimental asymmetric 4-bit W4A8 format (asym_w4a8_int8) across the toolkit: native quantization, state-dict round-trip, LoRA handling, save/load, lazy-compile handling, and runtime dispatch. Add a temporary torch.compile compatibility shim (w4a8_compile_compat.py) that provides a toolkit custom-op / FakeTensor boundary while upstream compiler-safe operator support is missing. Update int8_quant, model adapter, LoRA patching, lazy compile, model save, loaders, and tests; add docs/benchmarks.md, README/CHANGELOG updates, and unit tests for W4A8 compile compat. Bump package version to 2.2.0. Requires ComfyUI 0.31.0 / compatible comfy-kitchen; marked experimental.
Rebrand the project as "ComfyUI Quantization Toolkit" and introduce native INT4 (ConvRot W4A4) support, mixed INT4/INT8 modes, and expanded quantization policies. Add an INT4 torch.compile compatibility shim, INT4-aware lazy-compile changes, richer LoRA stack nodes with autogrow, and a Quantized LoRA entry/apply pair. Replace ad-hoc prints with structured logging, bump package metadata to 2.0.0, and add documentation (advanced usage, checkpoints, policies) plus many unit tests for compile, lazy-compile, LoRA, adapter caching, saving, and modes. Misc: kernel config, model save/load, and numerous runtime/adapter refactors and robustness fixes.
ThereforeGames