Misc. bug: convert_hf_to_gguf.py crashes with KeyError: 'image_mean' when preprocessor_config.json is incomplete

[rnett]

Name and Version

version: 8757 (a29e4c0)
built with Clang 19.1.5 for Windows x86_64

Operating systems

Windows, Linux, Mac

Which llama.cpp modules do you know to be affected?

Python/Bash scripts

Command line

python convert_hf_to_gguf.py path/to/model --outfile mmproj.gguf --mmproj --verbose

Problem description & steps to reproduce

The MmprojModel class in convert_hf_to_gguf.py crashes with a KeyError: 'image_mean' when exporting a multimodal projector if the Hugging Face preprocessor_config.json is missing the image_mean or image_std keys. This occurs with several newer Gemma 4 and custom vision checkpoints that don't include these standard preprocessor keys.

The current implementation assumes these keys always exist:

image_mean = _MISTRAL_COMMON_DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]

Relevant log output

Logs

  File "convert_hf_to_gguf.py", line 2233, in set_gguf_parameters
    image_mean = _MISTRAL_COMMON_DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"]
KeyError: 'image_mean'

[mollahmasud977-sys]

Thanks for providing the detailed reproduction steps and logs. I see what’s happening: the MmprojModel class assumes that preprocessor_config.json contains certain standard keys like image_mean, which causes a KeyError when these are missing. في السبت، 11 أبريل 2026 22:17 Ryan Nett ***@***.***> كتب:

…

*rnett* created an issue (ggml-org/llama.cpp#21775) <#21775> Name and Version $./llama-cli --version version: 8757 (a29e4c0 <a29e4c0> ) built with Clang 19.1.5 for Windows x86_64 Operating systems Windows Which llama.cpp modules do you know to be affected? Python/Bash scripts Command line uv run python llama/quant/src/quantize_nvfp4/convert_hf_to_gguf.py llama/quant/models/gemma-4-E2B-nvfp4 --outfile llama/quant/models/gemma4-nvfp4-mmproj.gguf --mmproj --verbose Problem description & steps to reproduce The \MmprojModel\ class in \convert_hf_to_gguf.py\ crashes with a \KeyError: 'image_mean'\ when exporting a multimodal projector if the Hugging Face \preprocessor_config.json\ is missing standard keys. This is a follow-up to closed issue #18514 <#18514>, which suggested a manual workaround rather than fixing the script's robustness. Relevant log output Logs \\console File ".../convert_hf_to_gguf.py", line 2233, in set_gguf_parameters image_mean = _MISTRAL_COMMON_DATASET_MEAN if self.is_mistral_format else self.preprocessor_config["image_mean"] ~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^ KeyError: 'image_mean' \\ — Reply to this email directly, view it on GitHub <#21775>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/CBTQ2IU3AHBD5PI4DZK55HT4VKZADAVCNFSM6AAAAACXVGGWRCVHI2DSMVQWIX3LMV43ASLTON2WKOZUGI2DKOBQGUZDSNQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>