Conversion Plugin

alkaid always lowers computations through its NumPy-style symbolic tracing API. Conversion plugins bridge external model objects into that tracing API, so framework models can be replayed on FVArray inputs and then converted to ALIR.

Top-Level Tracer Plugins

Top-level plugins are registered in the alir_tracer.plugins entry-point group. Each entry point maps a framework key to an ALIRTracerPluginBase subclass:

get_input_shapes() returns input shapes when they can be inferred from the model, or None when callers must provide symbolic inputs.
apply_model(verbose, inputs) replays the model on FVArray inputs and returns a trace dictionary plus output names.

When alkaid.converter.trace_model() is called without an explicit framework, the framework key is inferred from type(model).__module__.split('.', 1)[0]. For example, a Keras model resolves to keras, so the built-in Keras tracer is selected.

The current package declares these top-level entries in pyproject.toml:

Entry key	Tracer
`alkaid`	`alkaid.converter.example:ExampleALIRTracer`
`keras`	`alkaid.converter.builtin.keras.main:KerasALIRTracer`
`torch`	`alkaid.converter.builtin.torch.main:TorchALIRTracer`

Second-Level Operation Plugins

The built-in Keras and Torch tracers can be extended without replacing the top-level tracer. Third-party packages can register a zero-argument callable in one of these groups:

Entry-point group	Purpose
`alkaid_keras`	Import/register Keras layer or operation replay handlers.
`alkaid_torch`	Import/register Torch module, function, or method replay handlers.

The callable should import code that registers handlers in the corresponding built-in registry. alkaid loads such plugins on demand based on the base module of the layer, operation, or module being replayed.

Known Integrations

Integration	Mechanism	Description
Keras 3	Built-in `keras` top-level tracer	Replays supported Keras graph operations.
Torch	Built-in `torch` top-level tracer	Replays supported `torch.fx` graph nodes; callers must provide symbolic inputs.
HGQ2	`alkaid_keras` second-level handlers	Adds HGQ-specific Keras layer support when HGQ2 is installed.
Example model	Built-in `alkaid` top-level tracer	Minimal tracer used as an implementation reference and in tests.

Additional frontends can still register their own top-level key under alir_tracer.plugins when they are not naturally expressed as Keras or Torch operation handlers.

Example Plugin

An example plugin class is provided in alkaid.converter.example. The package metadata registers it as:

entry-points."alir_tracer.plugins".alkaid = "alkaid.converter.example:ExampleALIRTracer"

The plugin is used in tests/test_plugin.py, which is the smallest complete example of the top-level tracer interface.