convert_hf_to_gguf.py: ValueError when converting Gemma3n models due to double tensor mapping

Description
I encountered a ValueError: Can not map tensor when trying to convert a Gemma 3n (Vision/Audio) model using convert_hf_to_gguf.py.

python .\convert_hf_to_gguf.py ..\gemma-3n-E4B-it\ --mmproj

And I print the name and new name varriable in ModelBase

def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str:
        new_name = self.tensor_map.get_name(key=name, try_suffixes=try_suffixes)
        print(f"name: {name}, new_name: {new_name}")   # Here
        if new_name is None:
            raise ValueError(f"Can not map tensor {name!r}")
        return new_name

Operating systems

Windows

GGML backends

CUDA

Hardware

12700KF + 5090

Models

Gemma3n

Problem description & steps to reproduce

Analysis In Gemma3nVisionAudioModel.modify_tensors:

• The script correctly maps the HF name to the GGUF name (e.g., v.blk.0.0.bn1.weight).

• However, it then calls yield from super().modify_tensors(data_torch, new_name, bid).

• The parent class (Model.modify_tensors) receives new_name (the GGUF name) as the input name.

• It calls self.map_tensor_name(name) on this already-converted name.

• Since v.blk.0.0.bn1.weight is not in the tensor_map keys (it is a value/result), the mapping fails with ValueError.

Proposed Fix We should yield the result directly instead of passing the already-mapped name back to the parent class for re-mapping.

modify_tensors function in Gemma3nVisionAudioModel class

        # yield from super().modify_tensors(data_torch, new_name, bid)
        yield (new_name, data_torch)

First Bad Commit

No response

Relevant log output