Identical responses returned for similar-but-different input images

[BelieveDiffusion]

I'm using vllm-mlx on macOS (MBP M5 Max, 128GB, Tahoe 26.3.1 (a)) to host mlx-community/Qwen3-VL-30B-A3B-Instruct-bf16. Overall it works great, but I'm running into one annoying issue.

If I make sequential calls via the OpenAI API to describe a series of images one at a time, and several of the images are similar (but not the same), I find I get back identical text descriptions for multiple images in a row. The images are sufficiently different that the repeated returned description is wrong in at least some regard for the second and subsequent similar images, but they are word-for-word the same description as the first similar one in the sequence.

Is there some caching going on for similar-but-not-the-same images, which is causing a previously-generated similar-image input to be re-used for later queries (or something)? If so, is there a way to opt out of it?

For more context, I'm sending requests to http://localhost:8000/v1/chat/completions, with a query that includes a text message and a single image message containing a base64-encoded JPEG (max 2048px in each dimension).

[BelieveDiffusion]

I figured out the cause of the issue. It's this line in mllm.py:

    image_hash = hashlib.md5(base64_string[:1000].encode()).hexdigest()

Truncating the base64 string of a JPG to the first 1000 characters before calculating the hash seems to return the same hash for images that are overall visually similar.

[BelieveDiffusion]

I posted a one-line PR to remove the [:1000] truncation. This fixes the incorrectly-reused cached base64 images for me locally.