gguf_convert_endian.py: skip MXFP4 data

[AlekseiNikiforovIBM]

MXFP4 data only uses 1-byte data, there's nothing to byteswap.

Also use gguf.constants.GGML_QUANT_SIZES to determine block sizes. Don't duplicate block size constants.

[CISC]

Hmmm, won't converting gpt-oss mess up for the same reason? I think GGUFWriter.add_tensor/write_tensor_data needs a check.

[AlekseiNikiforovIBM]

Hmmm, won't converting gpt-oss mess up for the same reason? I think GGUFWriter.add_tensor/write_tensor_data needs a check.

byteswap calls in GGUFWriter.add_tensor/write_tensor_data should be already a no-op if data elements are 1-byte-sized. But this conversion script requires explicit handling for each data structure.

[CISC]

@AlekseiNikiforovIBM Mind having a look at gguf_new_metadata.py and gguf_editor_gui.py? I have a feeling they might mess up when saving a non-native endian GGUF.

[AlekseiNikiforovIBM]

I can take a look at these scripts. Are there any examples on how to use gguf_new_metadata.py?

[CISC]

I can take a look at these scripts. Are there any examples on how to use gguf_new_metadata.py?

A very simple test would just be gguf-new-metadata --general-name foo input.gguf output.gguf