Question about auto_sharding.py

[kevinsouthxu]

I implement a new intra-operator-strategy searching algorithm to replace the _call_solver_serialized_args in auto_sharding.py
but I got errors when executing:
(MeshHostWorker pid=77816) 2023-06-15 02:37:13.289767: E external/org_tensorflow/tensorflow/compiler/xla/pjrt/[pjrt_stream_executor_client.cc:2156](http://pjrt_stream_executor_client.cc:2156/)] Execution of replica 0 failed: INVALID_ARGUMENT: Executable expected shape f32[8192,2048]{1,0} for argument 1 but got incompatible shape f32[4096,2048]{1,0} (MeshHostWorker pid=77816) 2023-06-15 02:37:13.289837: E external/org_tensorflow/tensorflow/compiler/xla/pjrt/[pjrt_stream_executor_client.cc:2156](http://pjrt_stream_executor_client.cc:2156/)] Execution of replica 0 failed: INVALID_ARGUMENT: Executable expected shape f32[8192,2048]{1,0} for argument 1 but got incompatible shape f32[4096,2048]{1,0}

I wonder where there exists some constraint when selecting parallelism strategy?
But I think I consider all the constraints in the _call_solver_serialized_args implemented with ILP solver.