aitune.torch.libs.torch_compile | NVIDIA AITune Documentation

Shared torch.compile helpers.

Module Contents

Functions

Name	Description
`resolve_compile_dynamic`	Resolve effective dynamic compilation setting without mutating backend config.

Data

TorchCompileMode

API

aitune.torch.libs.torch_compile.resolve_compile_dynamic(
    config_dynamic: bool | None,
    graph_spec: aitune.torch.module.graph_spec.GraphSpec
) -> bool | None

Resolve effective dynamic compilation setting without mutating backend config.

Parameters:

config_dynamic

bool | None

Explicit dynamic setting from backend config.

graph_spec

GraphSpec

Graph specification used for dynamic-axis detection.

Returns: bool | None

Effective value to pass to torch.compile(dynamic=...).

aitune.torch.libs.torch_compile.TorchCompileMode = Literal['default', 'reduce-overhead', 'max-autotune', 'max-autotune-no-cudagraph...