Skip to content

torch_inductor

torch_inductor

Torch Inductor deployment backend for .pt2 artifacts, to deploy policies on x86 CPUs with optimized C++/Triton kernels, ref. https://docs.pytorch.org/ao/stable/pt2e_quantization/pt2e_quant_x86_inductor.html.

TorchInductorBackend

Bases: DeploymentBackend

Backend that saves .pt2 artifacts for torch.compile lowering.

export 

export(model, example_inputs)

Return a .pt2 deployment artifact descriptor.

Source code in src/versatil/post_training_compression/deployment_backends/torch_inductor.py 
24
25
26
27
28
29
30
31
32
33
34
35
36
def export(
    self,
    model: nn.Module,
    example_inputs: tuple[torch.Tensor, ...],
) -> DeploymentArtifact:
    """Return a .pt2 deployment artifact descriptor."""
    return DeploymentArtifact(
        converted_model=model,
        example_inputs=example_inputs,
        model_filename=self.model_filename,
        artifact_format=self.artifact_format,
        backend_name=self.name,
    )