phi_4_gpu_context_dy.json

{
    "input_model": { "type": "HfModel", "model_path": "microsoft/Phi-4" },
    "systems": {
        "local_system": {
            "type": "LocalSystem",
            "accelerators": [ { "execution_providers": [ "OpenVINOExecutionProvider" ] } ]
        }
    },
    "passes": {
        "optimum_convert": {
            "type": "OpenVINOOptimumConversion",
            "extra_args": { "device": "gpu", "task": "text-generation-with-past" },
            "ov_quant_config": { "weight_format": "int4", "group_size": 128, "ratio": 0.8 }
        },
        "io_update": { "type": "OpenVINOIoUpdate", "static": false, "reuse_cache": true },
        "encapsulation": {
            "type": "OpenVINOEncapsulation",
            "target_device": "gpu",
            "keep_ov_dynamic_dims": true,
            "ov_version": "2025.2",
            "reuse_cache": true
        }
    },
    "search_strategy": false,
    "host": "local_system",
    "cache_dir": "cache",
    "clean_cache": true,
    "evaluate_input_model": false,
    "output_dir": "models/Phi_4_gpu_context_dy"
}