Olive export of meta-llama/Llama-3.2-11B-Vision

llama.json

{
    "input_model": {
        "type": "PyTorchModel",
        "model_script": "model_loader.py",
        "model_loader": "load_model",
        "model_path": "meta-llama/Llama-3.2-11B-Vision",
        "io_config": {
            "input_names": [
                "input_ids",
                "attention_mask",
                "pixel_values",
                "aspect_ratio_ids",
                "aspect_ratio_mask",
                "cross_attention_mask"
            ],
            "output_names": [
                "logits"
            ],
            "input_shapes": [
                [
                    1,
                    14
                ],
                [
                    1,
                    14
                ],
                [
                    1,
                    1,
                    4,
                    3,
                    448,
                    448
                ],
                [
                    1,
                    1
                ],
                [
                    1,
                    1,
                    4
                ],
                [
                    1,
                    14,
                    1,
                    4
                ]
            ],
            "input_types": [
                "int32",
                "int32",
                "float32",
                "int32",
                "int32",
                "int32"
            ]
        }
    },
    "passes": {
        "onnx_conversion": {
            "type": "OnnxConversion",
            "use_dynamo_exporter": true,
            "target_opset": 18
        }
    },
    "engine": {
        "log_severity_level": 0
    }
}

model_loader.py

from transformers import MllamaForConditionalGeneration


def load_model(model_path):
    return MllamaForConditionalGeneration.from_pretrained(model_path)

olive.sh

olive run --config llama.json