opencv cuda frame

[aafaqin]

In my pipeline i wanted to use few custom matrix opertaion which i created in jax which are accelerated using GPU.
something like this

  # Convert the frames to JAX arrays and stack along a new dimension.
  stacked_frames = jnp.stack([jnp.asarray(frame) for frame in frames_list], axis=0)

  # Calculate and return the median along the newly created dimension.
  median_frame = jnp.median(stacked_frames, axis=0)

But i wanted to use median_frame in some opencv cuda operations.
but i wanted that frame when used should not to device to host and then again host to device computation.
How do i do this?