v0.19.0

0.19.0 (2026-07-31)

Features

• cli: add -v/-q verbosity and report setup phases during training (220bc8c)

Bug Fixes

• cli: close countable phases, route progress to stderr, pin the -v ladder (22e842d)
• trainer: end the run on a negative MSE — the visible signature of a swallowed OOM (7e58981)
• trainer: type check_step_loss's step as u64, matching the loop counter (4faffb3)

v0.18.0

0.18.0 (2026-07-27)

Features

• core: time real training steps — the #110 bench adapter and driver (#161) (eb546ea)

v0.17.0

0.17.0 (2026-07-26)

Features

• core: embed LoRA metadata (trigger words, training record) in exports (bc02d60)

Bug Fixes

• core: f32 metadata precision, checkpoint semantics, consumer contract (0b2f16c)

v0.16.0

0.16.0 (2026-07-24)

Features

• api: expose GPU backend feature passthrough (wgpu/candle/cuda/tch) (#151) (0b1ccae)

v0.15.0

0.15.0 (2026-07-23)

Features

• trainer: optional Krea-2-Turbo training-adapter merge-at-load (#83) (1600ba3)

Bug Fixes

• trainer: address #83 review — widen merge scope, real-adapter alpha, frozen base (958f762)

Performance Improvements

• trainer: read training-adapter alpha metadata header-only (#83 review) (66f27b7)

v0.14.0

0.14.0 (2026-07-22)

Features

• brand: add loractl logo and GitHub asset set (7110bce)

v0.13.0

0.13.0 (2026-07-19)

Features

• core: block-level gradient checkpointing for the MMDiT trunk (#134) (#135) (a1964af)
• core: retention-ledger attribution wiring (#132) (b7be0b8)
• quant: chunk large frozen-base dequants into row chunks (#128) (#130) (5bc9902)

Bug Fixes

• core: retention-ledger review fixes — deny allowlist, per-step report (d112876)

v0.12.0

0.12.0 (2026-07-18)

Features

• core: ADR-0005 step-VRAM probe + memory-lever configs (#126) (f5bf4aa)

v0.11.0

0.11.0 (2026-07-18)

Features

• bench: scaffold loractl-bench measurement harness crate (#115) (d5fa97b)
• cli: per-component model path-override flags (#108) (42d219b)
• core: QuantMatmulT-backward GPU diagnostic (quant_grad_compare) (#114) (c3f0c76)
• flow: resolution-based timestep shift with Krea 2 endpoints (#107) (7ff5ef5)
• quant: int4 (Q4S) frozen-base quantization to fit a 24 GB step (#119) (a2149db)

Bug Fixes

• core: restore example loads with skip_enum_variants (#117) (4bc4757)

v0.10.0

0.10.0 (2026-07-16)

Features

• config: per-component model paths — load scattered ComfyUI files (#101) (20a3d7a)
• encode: load a ComfyUI native-keyed Qwen-Image VAE (#104) (ad87df4)
• encode: load a ComfyUI-fp8 Qwen3-VL text encoder (#105) (7166c64)
• hf: fetch the Qwen3-VL tokenizer on demand for ComfyUI layouts (#103) (248d2e3)