transformers

0201b877 - Guard torch.distributed.tensor.device_mesh import in continuous_batching

Guard torch.distributed.tensor.device_mesh import in continuous_batching The unconditional `from torch.distributed.tensor.device_mesh import DeviceMesh` in `generation/continuous_batching/distributed.py` crashes at module load time on torch versions that predate `torch.distributed.tensor.device_mesh`. Since this triggers from `from transformers import AutoModelForCausalLM`, it breaks downstream test collection on older-torch CI matrices (e.g. accelerate's minimum-torch job). `DeviceMesh` is only used as a type annotation, so move the import under `TYPE_CHECKING` and stringify the annotation. Runtime behavior unchanged.