qlora-rs 1.3.0

[1.3.0] - 2026-08-20

Added

• merge_lora_into_dequantized — F32 NF4 dequant (dequantize_nf4, not
  compute_dtype / cached BF16) + LoRA residual using [LoraLayer::scaling]
  (rsLoRA α/√r when set). Portable merge for HF model.safetensors.
• .cz.toml (commitizen conventional commits, semver 1.x; no major_version_zero).

Changed

• ExportFormat::Gguf is documented as custom NF4 (not ggml Q4_0). A second
  public variant was not added (1.x exhaustive-match break). merge_and_export_gguf
  is labeled internal / non-portable.
• CLAUDE.md: GGUF export is no longer instructed as llama.cpp-compatible.
  Custom NF4 GGUF is internal; portable path is axolotl dense HF +
  convert_hf_to_gguf.py.

Notes

• Erratum: 1.0-era wording that GGUF export was llama.cpp-compatible (including
  0.1.0-alpha "GGUF (llama.cpp compatible)" and 1.0.0 merge_and_export_gguf as a
  complete GGUF writer) was incorrect. Custom GGUF_TYPE_QLORA_NF4 (0x4E4634)
  is internal; portable path is axolotl dense HF + convert_hf_to_gguf.py.
  Historical version sections are left unchanged.