Export¶

Quantize and export PyTorch models for DeepGate edge deployment.

Functions:

Name	Description
`trace`	Trace `model` into the graph that `dg.export` lowers.
`export`	Lower a quantized model to a `schema.json` dict for deployment.

dg.trace ¶

trace(
    model: Module,
    example_inputs: Tensor | tuple[Tensor, ...],
    *,
    dynamic_batch: bool = True
) -> torch.fx.GraphModule

Trace model into the graph that dg.export lowers.

This is the whole export path for a logic model, which has no quantization:

schema = dg.export(dg.trace(model, (example,)))

A quantized model is traced by enable_quantization instead, so you do not call this yourself.

Parameters:

Name	Type	Description	Default
`model`	`Module`	The model to trace. Tracing runs on the model's own device, and the traced graph comes back on that same device.	required
`example_inputs`	`Tensor \| tuple[Tensor, ...]`	Example inputs to trace with. Pass a batch of at least 2 samples when `dynamic_batch` is set.	required
`dynamic_batch`	`bool`	Let the traced graph accept any batch size. A batch of 1 would otherwise be baked in as a fixed size.	`True`

Returns:

Type	Description
`GraphModule`	The traced graph, ready for `dg.export`.

dg.export ¶

export(
    converted: GraphModule,
    *,
    input_boundary: str = DType.FLOAT32,
    output_boundary: str = DType.FLOAT32,
    preprocess: Preprocessor | None = None
) -> dict[str, Any]

Lower a quantized model to a schema.json dict for deployment.

Write the returned dict out with json.dump and upload it to the DeepGate platform to compile or benchmark. Any operation that cannot be lowered raises an error rather than being silently dropped.

Parameters:

Name	Type	Description	Default
`converted`	`GraphModule`	The quantized graph from `freeze_quantization` or `post_training_quantize`, or a traced logic model from `trace`.	required
`input_boundary`	`str`	Dtype the deployed model accepts, `"float32"` (the default) or `"int8"`. Use `"int8"` to hand the model already-quantized input, which moves the input quantize step onto the host.	`FLOAT32`
`output_boundary`	`str`	Dtype the deployed model returns, `"float32"` (the default) or `"int8"`. Use `"int8"` to dequantize on the host.	`FLOAT32`
`preprocess`	`Preprocessor \| None`	An audio or image pipeline to compile into the model, so deployed inference runs the same preprocessing you trained on. Requires the default float32 `input_boundary`.	`None`

Returns:

Type	Description
`dict[str, Any]`	The schema dict.