coreai_opt.casting.cast_to_16_bit_precision¶

coreai_opt.casting.cast_to_16_bit_precision(exported_program, ignored_ops=None)[source]¶

Convert a torch exported program to 16-bit precision: FP32→FP16 and INT32/64→INT16.

Runs both cast passes sequentially: 1. cast_fp32_to_fp16: FP32→FP16 2. cast_int32_to_int16: INT32/INT64→INT16

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. Forwarded to cast_fp32_to_fp16().

Returns:

The modified exported program.

Return type:

ExportedProgram