INNER CODE UNIT · Python
get_default_compile_config
pytorch/executorch · backends/apple/coreai/__init__.py:31
def get_default_compile_config() -> EdgeCompileConfig:
"""Default ``EdgeCompileConfig`` for Core AI lowering.
``_skip_dim_order=True`` keeps ExecuTorch on ``aten._to_copy`` instead of
emitting ``dim_order_ops._to_dim_order_copy``, which coreai-torch cannot
lower (its validator requires dim-order ops be decomposed).
``_check_ir_validity=False`` disables the edge-dialect verifier. Core AI's
quantized ops legitimately mix dtypes, which the verifier forbids: an FP8 or
FP4 weight reaches ``coreai::constexpr_blockwise_shift_scale`` alongside an
fp32 or ``float8_e8m0fnu`` scale, and the check rejects any op whose tensor
inputs disagree on dtype.
"""
return EdgeCompileConfig(_skip_dim_order=True, _check_ir_validity=False)
__all__ = [
"AOTCompileConfig",