--Fnonhomo_split

Applicability

Description

Sets non-uniform subgraph partition tuning. If the batch input to a model is an odd number or the number of AI Cores is not in a factor or multiple relationship with the batch, you are advised to use this option to obtain better performance.

Example: Assume that the batch is 7. Under uniform partition, you can only choose not to partition or partition it into seven copies. Under non-uniform partition, you can partition it to more non-factor copies, for example, 4+3.

For the Atlas training products, this option is not supported.

For the Atlas A2 training products/Atlas A2 inference products, this option is not supported.

For the Atlas A3 training products/Atlas A3 inference products, this option is not supported.

See Also

To use this option, set job_type to 1.

Arguments

None

Suggestions and Benefits

None

Examples

The following is an example:

--Fnonhomo_split --job_type=1

Restrictions

None