Torch backend: LSTM fails to use cuDNN when criteria are satisfied

[DLumi]

If you create a LSTM layer that satisfies requirements stated in documentation, you cannot do a forward pass on the layer created with use_cudnn=True.
Error:

ValueError: Exception encountered when calling LSTM.call().

use_cudnn=True was specified, but cuDNN is not supported for this layer configuration with this backend. Pass use_cudnn='auto' to fallback to a non-cuDNN implementation.

Arguments received by LSTM.call():
  • sequences=torch.Tensor(shape=torch.Size([1, 256, 128]), dtype=float32)
  • initial_state=None
  • mask=None
  • training=None

A reference to the cuDNN requirements in code (pretty much the same):
https://github.com/keras-team/keras/blob/0ddf96200f781f95dea0289aa87d85eb5eb3e733/keras/src/backend/torch/rnn.py#L548C1-L564C6

Colab to reproduce:
https://colab.research.google.com/drive/1OFLXgZpiUXDXTO1eebQYhXnq2HevZpoe

P.S. I explicitly defined the requirements even though on paper you should be good to go with the default values
P.P.S this problem is observed starting with keras==3.12. On 3.11.3 this is not an issue. Or at least, the layer is created, I have no idea if it actually uses cudnn (likely not as the training is VERY slow)