This was exactly the case for me on iMac Pro 2017 with Radeon Vega 64. With the default_size = 128, one epoch on GPU took 27 s, while on CPU 28 s. With batch_size = 1024, one epoch on GPU took 6 s, while on CPU 21 s.

Could you please update the installation instructions? It says now that only Python 3.8 and 3.9 are supported.