I’d expect that to depend on the sort of workload you’re running. All those frameworks are typically used in Python because they optimize for the use-case where you’re doing most of your work on the GPU, where it’s actually some compiled kernel running mostly matrix multiplications, and the Python part of your code is slow but doesn’t take up a significant part of the runtime. This will not be the case if your runtime is, say, dominated by loading data in Python, or such.

Or if you aren’t running on a GPU.