transformers

Fix tf32 issue: set `torch.backends.cudnn.conv.fp32_precision` explicitly.

#45248

Merged