[BUG] Unsloth booster fails on composite VLM LoRA: "No layers to finetune?" due to target_modules format mismatch

Reminder

• I have read the above rules and searched the existing issues.

System Info

Environment

• LLaMA-Factory: 0.9.6.dev0
• Unsloth: 2026.x
• Model: Qwen/Qwen3.5-4B
• Stage: SFT, LoRA
• Booster: unsloth
• GPU: NVIDIA A100 40GB
• OS: Linux

Reproduction

Describe the bug

When Booster = unsloth is enabled for LoRA SFT on a composite / vision-language model (e.g. Qwen/Qwen3.5-4B, model_type=qwen3_5), training crashes during LoRA adapter setup with:

The same training config works when Booster = auto (standard PEFT path).

To Reproduce

Steps to reproduce the behavior:

1. Install LLaMA-Factory with Unsloth support (use_unsloth / Booster unsloth).
2. Configure SFT with a composite VLM:
   • Model: Qwen/Qwen3.5-4B
   • Template: qwen3_5_nothink
   • Finetuning type: LoRA (lora_target=all)
   • Freeze vision tower: true
   • Booster: unsloth
3. Use any valid SFT dataset (text-only is sufficient).
4. Start training via CLI or WebUI.

Training fails in get_unsloth_peft_model() → FastLanguageModel.get_peft_model().

Expected behavior

Unsloth booster should attach LoRA adapters to the intended language-model linear layers on composite VLMs, consistent with the standard PEFT path (with vision/projector frozen when configured).

Actual behavior

LoRA setup fails because LlamaFactory passes full module paths as target_modules, while Unsloth expects leaf module names.

Root cause

There is a format mismatch between two LlamaFactory code paths:

1. adapter.py builds LoRA targets, then calls patch_target_modules() in model_utils/visual.py.
2. For models in COMPOSITE_MODELS (e.g. qwen3_5, qwen2_vl, llava, …), patch_target_modules() expands short targets like out_proj into full paths: model.language_model.layers.0.linear_attn.out_proj

This is correct for standard HuggingFace PEFT.

1. When use_unsloth=True, those expanded paths are passed to:

FastLanguageModel.get_peft_model(target_modules=...)

1. Unsloth’s get_peft_regex() treats target_modules as leaf names and builds regexes that match module paths ending in those names. Passing a full path as the “leaf” matches nothing → 0 trainable layers → crash.

Proposed fix

In the Unsloth integration path only, convert expanded VLM target_modules back to unique leaf names before calling FastLanguageModel.get_peft_model().

Suggested change in model_utils/unsloth.py:

def _coerce_unsloth_target_modules(model, target_modules: list[str]) -> list[str]:
    """Unsloth expects LoRA leaf names; VLMs get full paths from patch_target_modules."""
    model_type = getattr(model.config, "model_type", None)
    if model_type not in COMPOSITE_MODELS:
        return target_modules
    if not target_modules or all("." not in name for name in target_modules):
        return target_modules
    return sorted({name.rsplit(".", 1)[-1] for name in target_modules})

Others

No response