fix: patch PEFT for Gemma4ClippableLinear in loader checkpoint path (fixes export)

[rolandtannous]

Problem

Exporting or loading a Gemma-4 LoRA checkpoint fails with:

ValueError: Target module Gemma4ClippableLinear(
  (linear): Linear(in_features=768, out_features=768, bias=False)
) is not supported. Currently, only the following modules are supported:
`torch.nn.Linear`, `torch.nn.Embedding`, `torch.nn.Conv1d`, ...

PEFT doesn't support Gemma4ClippableLinear yet — it wraps nn.Linear but doesn't subclass it, so PEFT's LoRA injection doesn't recognize it as a valid target.

Tracked upstream:

• Add LoRA support for Gemma 4 (Gemma4ClippableLinear) huggingface/peft#3130
• Add support for Gemma4ClippableLinear (Gemma 4 QLoRA fails) huggingface/peft#3129

Why training works but export/inference doesn't

Training works because unsloth/models/vision.py (lines 1301-1353) already has a monkey-patch that intercepts LoraModel._create_and_replace and redirects Gemma4ClippableLinear to its inner .linear child:

# vision.py — training path (already exists)
_clippable_linear_cls = None
try:
    from transformers.models.gemma4.modeling_gemma4 import (
        Gemma4ClippableLinear as _clippable_linear_cls,
    )
except ImportError:
    pass
if _clippable_linear_cls is not None:
    from peft.tuners.lora.model import LoraModel as _LoraModel
    _original_car = _LoraModel._create_and_replace

    def _patched_car(self, peft_config, adapter_name, target,
                     target_name, parent, current_key=None, **kwargs):
        if isinstance(target, _clippable_linear_cls):
            return _original_car(
                self, peft_config, adapter_name,
                target.linear, "linear", target,  # redirect to inner nn.Linear
                current_key=current_key, **kwargs,
            )
        return _original_car(
            self, peft_config, adapter_name,
            target, target_name, parent,
            current_key=current_key, **kwargs,
        )

    _LoraModel._create_and_replace = _patched_car

model = _get_peft_model(model, lora_config)

# Restore original PEFT method
if _clippable_linear_cls is not None:
    _LoraModel._create_and_replace = _original_car

But when loading an existing checkpoint (export, inference), the code goes through loader.py → PeftModel.from_pretrained() which is a completely separate code path. The monkey-patch from vision.py is not active there.

Fix

Apply the same monkey-patch pattern from vision.py around the PeftModel.from_pretrained() call in loader.py's vision model loading path. The patch:

1. Saves the original _create_and_replace method
2. Installs a wrapper that checks if the target is Gemma4ClippableLinear — if so, redirects to its inner .linear attribute
3. Calls PeftModel.from_pretrained() with the patch active
4. Restores the original method immediately after

This is a temporary workaround until PEFT adds native support for Gemma4ClippableLinear.

Test plan

• Export a Gemma-4 LoRA checkpoint (previously crashed with ValueError)
• Load a Gemma-4 LoRA checkpoint for inference
• Verify Gemma-4 training still works (unchanged path in vision.py)
• Verify non-Gemma-4 models are unaffected (patch is guarded by ImportError + isinstance check)

export_gemma4.mp4

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 5847fbe2b5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[rolandtannous]

resolved reviewer comments and performed additional end to end tests:

1- finetuned text model requiring transformers 4.57.6
2- finetuned gemma4
3-exported gemma4
4-exported gemma3
5- chat with gguf
6- chat with gemma4
7-chat with Phi model

[BenjaminBossan]

Shouldn't it be possible to target the linear submodule of Gemma4ClippableLinear by setting a proper target_modules? It might require a bit of regex acrobatics to achieve that, but it should work, e.g. to target q_proj and v_proj of the audio tower:

target_modules=r".*\.audio_tower.*\.(q_proj|v_proj)\.linear"