cuda-bindings (https://pypi.org/project/cuda-bindings/) needs to be installed. If it is not, the code will silently transfer the gpu data back to cpu.
Be sure to install the version of cuda-bindings with the same major version as your CUDA toolkit version.

NOTE: cuda-bindings 12.9.3 and 13.0.2 do not work for this PR. (reported in NVIDIA/cuda-python#1122)

import cupy as xp
import pyqtgraph as pg

pg.setConfigOption('imageAxisOrder', 'row-major')
pg.setConfigOption('useCupy', True)

X = xp.linspace(-1, 1, 1024, dtype=xp.float32)
Y = X[:, None]
Z = xp.sin(20 * 2 * xp.pi * (Y**2 + X**2))

pg.mkQApp()
win = pg.RawImageGLWidget()
win.setImage(Z, levels=[-1, 1])
win.show()
pg.exec()
print(win.gfx_resource)    # None if it didn't go through GPU
win.cleanup()

My apologies @pijyoi for keeping this PR in the queue for so long. Getting this functionality baked into the library has been on the wish list for quite some time. Thank you for making it happen!