发布

  • 2.0.3 19b6bf7590

    fix: handle LoRA rank mismatch during inference in lora_ft_webui

    frostbyte_neo 发布于 2026-04-28 02:52:57 +00:00 | 33 次提交 在此版本后已推送到 main

    Pass the selected LoRA checkpoint to load_model() on first load so the
    model initializes with the correct rank from lora_config.json instead of
    always defaulting to r=32.

    On subsequent LoRA hot-swaps, detect rank incompatibility and
    automatically reload the model with the new checkpoint's config,
    preventing tensor shape mismatch errors (fixes #283).

    Made-with: Cursor

    下载附件