optimum-intel-internal-testing/tiny-random-gemma4-31B

from transformers import (
    AutoTokenizer,
    Gemma4ForConditionalGeneration,
)


def generate_vlm_model(output_dir="./tiny-random-gemma4-31B"):
    from transformers import AutoConfig, AutoProcessor, AutoTokenizer, Gemma4ForConditionalGeneration
    model_id = "google/gemma-4-31B-it"
    config = AutoConfig.from_pretrained(model_id)

    # Text config
    config.text_config.global_head_dim = 4
    config.text_config.head_dim = 4
    config.text_config.hidden_size = 32
    config.text_config.hidden_size_per_layer_input = 0
    config.text_config.num_hidden_layers = 2
    config.text_config.layer_types = ["sliding_attention", "full_attention"]
    config.text_config.num_kv_shared_layers = 0
    config.text_config.intermediate_size = 64
    config.text_config.dtype = "float32"

    # Vision config
    config.vision_config.head_dim = 4
    config.vision_config.hidden_size = 8
    config.vision_config.intermediate_size = 32
    config.vision_config.num_hidden_layers = 1
    config.vision_config.num_key_value_heads = 2

    model = Gemma4ForConditionalGeneration(config)
    model.eval()
    model.save_pretrained(str(output_dir))

    tokenizer = AutoTokenizer.from_pretrained(model_id)
    tokenizer.save_pretrained(str(output_dir))

    processor = AutoProcessor.from_pretrained(model_id)
    processor.save_pretrained(str(output_dir))

    return model


if __name__ == "__main__":
    generate_vlm_model()