Environment:
Model: gemma-4-E4B-it-Q8_0.gguf (from unsloth/gemma-4-e4b-it-GGUF)
mmproj file downloaded and placed in the same folder as the model
llama-cpp-python: v0.3.35+
OS: Windows
Problem :
The model loads without errors and the family is correctly detected (Model family: gemma4), but captions are completely hallucinated, suggesting the vision encoder is not active. The mmproj file is present in the model directory but there is no dedicated field in the UI to specify its path explicitly, so I'm not sure if it's being picked up at all.
Expected behavior
The mmproj file gets passed to the chat handler so vision actually works.
Question:
Is the mmproj path inferred automatically from the model folder, or does it need to be specified somewhere? If automatic, is there a specific filename pattern it expects? A note in the docs about this would help since it's not obvious from the current GGUF setup instructions.
Environment:
Model: gemma-4-E4B-it-Q8_0.gguf (from unsloth/gemma-4-e4b-it-GGUF)
mmproj file downloaded and placed in the same folder as the model
llama-cpp-python: v0.3.35+
OS: Windows
Problem :
The model loads without errors and the family is correctly detected (Model family: gemma4), but captions are completely hallucinated, suggesting the vision encoder is not active. The mmproj file is present in the model directory but there is no dedicated field in the UI to specify its path explicitly, so I'm not sure if it's being picked up at all.
Expected behavior
The mmproj file gets passed to the chat handler so vision actually works.
Question:
Is the mmproj path inferred automatically from the model folder, or does it need to be specified somewhere? If automatic, is there a specific filename pattern it expects? A note in the docs about this would help since it's not obvious from the current GGUF setup instructions.