Upload processor_config.json

#2
by BrokeC - opened

Fix: Add missing spatial_merge_size to processor config

Problem

Missing spatial_merge_size parameter causes vision inference to crash:
ValueError: Number of image token positions (5476) does not match number of image features (1369)

Solution

Add "spatial_merge_size": 2 to processor_config.json:

{
  "image_break_token": "[IMG_BREAK]",
  "image_end_token": "[IMG_END]",
  "image_token": "[IMG]",
  "patch_size": 14,
  "processor_class": "PixtralProcessor",
  "spatial_merge_size": 2
}

This matches the original https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503/blob/main/processor_config.json and config.json.

βœ… Tested on M2 Max 64GB - vision inference works after fix.

Same fixes needed for: -4bit, -6bit, -8bit, -bf16 (all 0.1.19 conversions)
Ready to merge
This branch is ready to get merged automatically.

Sign up or log in to comment