[sharktank] Add common model config, export and compile

[sogartar]

We don't have a standard way to configure, export and compile sharktank models.

Here is introduced such mechanism and for demonstration the CLIP text model is refactored to utilize this new approach.

config.json:

{
    "model_type": "my_package.MyModel",
    "mlir_path": "model.mlir",
    "parameters_path": "model.irpa",
    "iree_module_path": "model.vmfb",
    "compile_args": ["--iree-hal-target-device=local"],
    "export_functions": [
        {
            "function": "forward",
            "batch_sizes": [1, 2, 3]
        }
    ]
    "my_model_arg": 5,
}

usage

model = create_model("config.json")
model.export()
model.compile()

[ScottTodd]

I'll need to set up some time aside to review this in detail. Can you tag related issues on this PR and connect this back to the overall goals? Bonus points if you could also comment on the issues and help coordinate their full resolution. Cross-cutting work like this goes more smoothly if the plans are communicated ahead of time, before jumping straight to code.

These issues appear related:

• [tracking] Shortfin usability issues #245
• Find More General and Easier to use Alternative For Compiling Models for Shortfin LLM Server #402
• Streamline LLM import/compile/serve user experience #691

[sogartar]

Yes the are related issues.

[sogartar]

This PR requires some modification to make it more interoperable with IREE build pipeline tooling.

[sogartar]

I added an export function to be used in an IREE build pipeline.

[renxida]

I have some py functions / classes for managing LLMs in app_tests/integration_tests/llm/model_management.py.

I like the class / function structure currently in there, but they use a very hacky bash-script-in-python-script kind of solution to call into sharktank and IREE to shard / export / compile models.

Would be very cool if I could switch to using this instead.