zai-org

/

GLM-4.1V-9B-Thinking

like 755

Follow

Z.ai 5.5k

Fixed the model_type error in vision_config. This PR allows the visual model to be loaded correctly instead of loading the entire GLM-4V model.

A simple reproduction script is as follows:

import torch
from transformers import Glm46VForConditionalGeneration


def main():
    model = Glm46VForConditionalGeneration.from_pretrained(
        "THUDM/GLM-4.1V-9B-Thinking",
        dtype="auto",
        device_map="auto",
    )

    visual = model.model.visual
    language = model.model.language_model

    print(f"type(model.model.visual): {type(visual)}")
    print(f"type(model.model.language_model): {type(language)}")

    assert visual.__class__.__name__ == "Glm4vVisionModel", (
        "vision_config mistakenly sets model_type='glm4v', so AutoModel.from_config "
        "instantiates the full Glm4vModel (visual + language) instead of the pure "
        "vision backbone Glm4vVisionModel."
    )


if __name__ == "__main__":
    main()