--is_weight_clip

Applicable Products

All processors

Description

Specifies whether to clip floating-point weight data during conversion from a high-precision data type to a low-precision data type.

When the weight of the original model is of a high-precision data type (such as Float32), inserting a Cast operator during conversion to a low-precision data type (for example, Float16) may cause data overflow due to value-range constraints. By configuring the --is_weight_clip parameter, high-precision data can be clipped before the Cast operator to effectively prevent overflow and guarantee data integrity.

See Also

None

Arguments

• 0: disabled.
• 1 (default): enabled.

Suggestions and Benefits

None

Example

--is_weight_clip=1

Dependencies and Restrictions

None