[benchmarks][dynamo] Trying CI - Set train() for TIMM models accuracy tests (#89780)
Moving to train mode for TIMM models and also raising batch size for accuracy testing.
Raising batch size seems to remove a lot of noise/instability coming from batch_norm decomposition.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/89780
Approved by: https://github.com/ngimel
