PyTorch and TensorFloat32

hardware-backends NVIDIA CUDA
5

I just installed 1.12 (pip; Windows; CUDA 11.6) and I don’t see torch.backends.matmul.allow_tf32:

>>> import torch
>>> torch.__version__
'1.12.0+cu116'
>>> torch.backends.matmul.allow_tf32
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: module 'torch.backends' has no attribute 'matmul'
>>> torch.backends.cudnn.allow_tf32
True
>>> dir(torch.backends)
['ContextProp', 'PropModule', '__allow_nonbracketed_mutation', '__allow_nonbracketed_mutation_flag', '__builtins__', '__cached__', '__doc__', '__file__', '__loader__', '__name__', '__package__', '__path__', '__spec__', 'contextmanager', 'cuda', 'cudnn', 'disable_global_flags', 'flags_frozen', 'mkl', 'mkldnn', 'mps', 'openmp', 'quantized', 'types']