Implementation:Axolotl ai cloud Axolotl Load Processor

Knowledge Sources	Axolotl HuggingFace AutoProcessor
Domains	Multimodal, Vision_Language
Last Updated	2026-02-06 23:00 GMT

Overview

Concrete tool for loading multimodal processors for vision-language model training provided by the Axolotl framework.

Description

The load_processor function loads the appropriate processor for a vision-language model. It handles processor class selection (AutoProcessor or model-specific), Mistral-specific tokenizer patching, image size auto-detection from the processor, and trust_remote_code configuration. For Mistral models, it patches the tokenizer backend and uses a custom Mistral3Processor.

Usage

Called conditionally during dataset loading when cfg.processor_type is set or the model is detected as multimodal. The processor is passed through to data preparation and training.

Code Reference

Source Location

• Repository: axolotl
• File: src/axolotl/loaders/processor.py
• Lines: L17-82

Signature

def load_processor(
    cfg: DictDefault,
    tokenizer: PreTrainedTokenizerBase,
) -> ProcessorMixin:
    """Load a multimodal processor for vision-language models.

    Args:
        cfg: Config with processor_type, processor_config, trust_remote_code,
             tokenizer_use_mistral_common, image_size.
        tokenizer: Pre-loaded tokenizer instance.

    Returns:
        ProcessorMixin: Configured processor instance with image_size
        potentially populated from processor metadata.
    """

Import

from axolotl.loaders.processor import load_processor

I/O Contract

Inputs

Name	Type	Required	Description
cfg	DictDefault	Yes	Config with processor_type (explicit class name), processor_config (HF model ID), trust_remote_code, tokenizer_use_mistral_common, image_size
tokenizer	PreTrainedTokenizerBase	Yes	Pre-loaded tokenizer passed to processor

Outputs

Name	Type	Description
return	ProcessorMixin	Configured processor (AutoProcessor, Mistral3Processor, or VoxtralProcessor)

Usage Examples

Loading AutoProcessor

from axolotl.loaders.processor import load_processor
from axolotl.loaders.tokenizer import load_tokenizer

tokenizer = load_tokenizer(cfg)

# cfg.processor_type = "AutoProcessor"
# cfg.processor_config = "meta-llama/Llama-3.2-11B-Vision"
processor = load_processor(cfg, tokenizer)

# Image size auto-detected from processor
print(cfg.image_size)  # (560, 560)

Related Pages

Implements Principle

• Principle:Axolotl_ai_cloud_Axolotl_Multimodal_Processor_Loading

Requires Environment

• Environment:Axolotl_ai_cloud_Axolotl_Python_Runtime