LFM2-VL-1.6B-bf16-preprocessor-fix

LFM2-VL-1.6B-bf16 — Fixed preprocessor_config.json

This repo contains a corrected preprocessor_config.json for mlx-community/LFM2-VL-1.6B-bf16.

Problem

The mlx-community conversion of LFM2-VL-1.6B has two bugs in preprocessor_config.json:

1. image_processor_type is set to "Siglip2ImageProcessor" instead of "Lfm2VlImageProcessorFast"
2. input_data_format is set to "channels_last" instead of null

These cause inference to crash with:

RuntimeError: The size of tensor a (H) must match the size of tensor b (3) at non-singleton dimension 1

The error occurs because input_data_format: "channels_last" tells the image processor the input is already in (H, W, C) format, but torchvision's normalize expects (C, H, W). When set to null, the processor auto-detects the format correctly.

Fix

Compare with the original LiquidAI/LFM2-VL-1.6B preprocessor_config.json:

Usage

To use this fix, either:

1. Download this file and place it in your local mlx-community model cache, or
2. Apply at runtime after loading:

   from mlx_vlm import load
   model, processor = load("mlx-community/LFM2-VL-1.6B-bf16")
   processor.image_processor.input_data_format = None
   

Upstream

This should be fixed in the mlx-community conversion. Filed as a reference for anyone hitting this error.