LSTM layer with dropout does not use fast CuDNN implementation in Keras 3

[giulatona]

I have a model that uses the keras LSTM layer with dropout not equal to zero. I am using TF backend.

I noticed a performance drop after updating to Keras 3 and after checking I found out that the requirements for using the fast CuDNN LSTM implementation.

With Keras 2.15, the requirements to use the cuDNN implementation are:

1. activation == tanh
2. recurrent_activation == sigmoid
3. recurrent_dropout == 0
4. unroll is False
5. use_bias is True
6. Inputs, if use masking, are strictly right-padded.
7. Eager execution is enabled in the outermost context.

With Keras 3, the requirements to use the cuDNN implementation are:

1. activation == tanh
2. recurrent_activation == sigmoid
3. dropout == 0 and recurrent_dropout == 0
4. unroll is False
5. use_bias is True
6. Inputs, if use masking, are strictly right-padded.
7. Eager execution is enabled in the outermost context.

This change has a significant impact on performance when using dropout. Can we expect a change or some alternative way to use dropout without performance degradation?

[fchollet]

Thanks for the report -- this is now resolved at HEAD.