I've found that when calling "grid_sample" with a float16 "input" and float32 "grid" parameter combo, pytorch errors out with: RuntimeError: expected scalar type Half but found Float Now this is easy ...