I get polygons instances with a simple python list but it's really interesting how I could get these instances with Numpy as it should be faster.
poly_indices = [tuple(poly.vertices) for poly in bpy.context.active_object.data.polygons]
The main problem for Numpy Array is that a length of the polygon's tuple can be differen. For a tringle the tuple length is 3. For a quad the tuple length is 4. For an NGon the tuple length is from 5 to N.