mistralai

/

Mistral-Small-3.1-24B-Instruct-2503

like 1.19k

Follow

Mistral AI_ 8.03k

Hi team,

I'm encountering an issue when running the mistralai/Mistral-Small-3.1-24B-Instruct-2503 model on vLLM (latest version). Even though my usage is text-only, the engine is attempting to invoke PixtralProcessor with image inputs, which leads to the following error:
AttributeError: 'MistralTokenizer' object has no attribute 'init_kwargs' ... RuntimeError: Failed to apply PixtralProcessor on data={'text': '[IMG]', 'images': [...]}

🧠 Context

• Model: mistralai/Mistral-Small-3.1-24B-Instruct-2503
• Setup: vLLM inference with LLM(model=model_name, tokenizer_mode="mistral")
• Task: Simple chat completion, no multimodal inputs
• Goal: Just use the model for basic conversational text

💡 Observations

1. When tokenizer_mode="mistral" is set, vLLM seems to treat it like the Pixtral (multimodal) variant.
2. The issue disappears if we use tokenizer_mode="auto"
3. If this is the expected behaviour maybe the documentation should be be updated to make this more clear.