tree 4050e545c37dc329193c01105d8a3d98605dfbbb
parent e826feabce35024cd00ce27d76db6196f4628970
author Haoliang Zhang <haoliang@google.com> 1646784471 -0800
committer TensorFlower Gardener <gardener@tensorflow.org> 1646785243 -0800

Add a new converter flag `_experimental_guarantee_all_funcs_one_use` to allow cloning function calls so that each function have exactly one use.

By default the flag is False, since duplicating FuncOps will significantly bloat the model size in certain scenarios. With this flag set to True, user can enable the `mlir::TF::CreateGuaranteeAllFuncsOneUsePass()` pass so that function inlining can succeed when multiple `StatefulPartitionedCall` ops call into the same function.

PiperOrigin-RevId: 433335089