transformers

9dff7ca5 - Fix EP: RouterParallel shape, tp_plan property, grouped_mm sentinels (#45473)

Fix EP: RouterParallel shape, tp_plan property, grouped_mm sentinels (#45473) * Fix EP: RouterParallel shape, tp_plan property, grouped_mm sentinels * Fix GroupedGemmParallel.shard_tensor: use self.rank for expert sharding * Fix expert parallel attribute mapping and update tests Add `attribute_map` to GptOssConfig to map `num_experts` to `num_local_experts`, and fix GroupedGemmParallel to use `empty_param.shape[0]` instead of `module.num_experts`. Remove RouterParallel tests and add new expert parallel forward/backward tests to TensorParallelTesterMixin. * Harden _can_set_{attn,experts}_implementation against sys.modules eviction Use sys.modules.get(cls.__module__) and treat a missing entry the same as the existing Jupyter/REPL case (no __file__) -> return False. Without this, PreTrainedModel.__init__ crashes with KeyError whenever another test (e.g. tests/utils/test_auto_docstring._clear) has evicted transformers.models.* entries from sys.modules while their class objects are still live.