Running PYTORCH_TEST_WITH_ROCM=1 python test_nn.py TestNN.test_rnn_cpu_vs_cudnn_no_dropout does not run MIOpen path.
This line in test_nn.py sets data type of tensors for the test to double, which is not supported by MIOpen which is why during execution we fall back to HIP kernels rather than using MIOpen. The change came from PR #27628 where the default datatype of the test was changed from float to double. Can we get a little more clarity on this change or if there could be a way to run MIOpen RNN in the CI tests.

cc: @mruberry @ezyang @iotamudelta