Unexpected results with CuDNNLSTM (instead of LSTM) layer

I have posted this question as an issue in Keras' Github but figured it might reach a broader audience here.

---

System information

• Have I written custom code (as opposed to using example directory): Minimal change to official Keras tutorial
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04.2 LTS
• TensorFlow backend (yes / no): yes
• TensorFlow version: 1.13.1
• Keras version: 2.2.4
• Python version: 3.6.5
• CUDA/cuDNN version: 10.1
• GPU model and memory: Tesla K80 11G

Describe the current behavior
I am executing the code from the Seq2Seq tutorial. The one and only change I made was to swap the LSTM layers for CuDNNLSTM. What happens is that the model predicts a fixed output for any input I give it. When I run the original code, I get sensible results.

Describe the expected behavior
See preceding section.

Code to reproduce the issue
Taken from here. Simply replace LSTM with CuDNNLSTM.

---

Any insights are greatly appreciated.

from Unexpected results with CuDNNLSTM (instead of LSTM) layer