Model Binders

Binders provide the mapping between architecture-specific parameter names and universal semantic roles.

Architecture Detection

UAL automatically detects model architecture from:

Model class name
Configuration attributes
Parameter name patterns

Binder Structure

Each binder defines:

{
    "attention_query": {
        "pattern": "model.layers.{layer}.self_attn.q_proj",
        "in_features": 4096,
        "out_features": 4096
    },
    ...
}

Pattern Matching

Patterns use {layer} placeholder for layer indexing:

# Matches: model.layers.0.self_attn.q_proj
#          model.layers.1.self_attn.q_proj
#          ...
pattern = "model.layers.{layer}.self_attn.q_proj"

Supported Architectures

GPT-2 Family

Parameter prefix: transformer.h.{layer}
Fused attention: c_attn for Q/K/V
MLP: c_fc (up), c_proj (down)

LLaMA Family

Parameter prefix: model.layers.{layer}
Separate Q/K/V projections
Gated MLP with gate_proj

Qwen Family

Similar to GPT-2 structure
Fused attention with c_attn
Custom MLP naming

Custom Binders

To add a new architecture:

from ual_adapter.binders import BaseBinder

class MyModelBinder(BaseBinder):
    def get_attention_query_pattern(self, layer: int) -> str:
        return f"model.layer.{layer}.attention.query"

    def get_dimensions(self, param_name: str) -> tuple:
        # Return (in_features, out_features)
        return (768, 768)