Adding Custom Architectures

Guide for adding support for new model architectures.

Creating a Binder

from ual_adapter.binders import BaseBinder

class MyModelBinder(BaseBinder):
    \"\"\"Binder for MyCustomModel architecture.\"\"\"

    architecture_name = "my_model"

    def get_parameter_mapping(self) -> dict:
        return {
            "attention_query": {
                "pattern": "encoder.layer.{layer}.attention.self.query",
                "in_features": 768,
                "out_features": 768
            },
            "attention_key": {
                "pattern": "encoder.layer.{layer}.attention.self.key",
                "in_features": 768,
                "out_features": 768
            },
            # ... more mappings
        }

Registering the Binder

from ual_adapter.binders import BinderRegistry

registry = BinderRegistry()
registry.register("my_model", MyModelBinder)

Testing the Binder

# Test detection
from transformers import AutoModel
model = AutoModel.from_pretrained("path/to/my_model")

from ual_adapter.utils import detect_architecture
arch = detect_architecture(model)
assert arch == "my_model"

# Test parameter mapping
ual = UniversalAdapter(model, tokenizer)
ual.export_adapter("test", "test.air")

# Verify AIR format
import json
with open("test.air") as f:
    air = json.load(f)
assert air["metadata"]["source_architecture"] == "my_model"

Contributing

To contribute your binder to UAL Adapter:

Fork the repository
Add binder to ual_adapter/binders/architectures.py
Add tests to tests/test_binders.py
Submit pull request