coreai_opt.casting.cast_fp32_to_fp16¶
- coreai_opt.casting.cast_fp32_to_fp16(exported_program, ignored_ops=None, calibration_data=None)[source]¶
Convert a torch exported program from FP32 to FP16 where applicable.
Converts parameters, user inputs, and compute ops to FP16, inserting casts only where values would overflow FP16 range.
- Parameters:
exported_program (ExportedProgram) – Exported program to convert.
ignored_ops (Collection[OpOverload | OpOverloadPacket] | OpOverload | OpOverloadPacket | Callable[[Node], bool] | None) – Optional operations or predicate to exclude from FP16 casting. Can be a collection/set of
OpOverloadorOpOverloadPacketinstances (e.g.{torch.ops.aten.exp, torch.ops.aten.exp.default}), a single op instance, or a predicate callable taking atorch.fx.Nodeand returning a boolean (e.g.lambda node: node.name == "exp_1"). Ignored ops are kept in FP32 with boundary casts inserted.calibration_data (Iterable[Tensor | tuple[Any, ...] | list[Any] | dict[str, Any]] | None) – Optional iterable yielding sample inputs (each sample a Tensor, tuple/list of Tensors, or dict of kwargs) for dynamic activation range observation. If provided, operations whose intermediate activations overflow the FP16 representable range (> 65504) are automatically identified and excluded from FP16 casting, remaining in FP32 precision with boundary casts.
- Returns:
The modified exported program.
- Return type:
ExportedProgram