The great challenge of these brute-force implementations is probabilistically minimizing the chances of key "losses." In such large search spaces, this is very difficult. You can never be 100% sure that you have verified all the keys. There is much work to improve these algorithms.
I addressed this issue some long time ago. In a nutshell, there is no need to probabilistically minimize bad coding, it is enough to validate the results of some range, using a second ground-truth validation run, which is guaranteed 100% to be valid. For example, testing some random range both on the GPU, and on the CPU (using a stable implementation, like libsecp256k1), and comparing some hash of all the results.
If the hash matches, and all GPU models are tested properly, then everything went OK and the GPU code is thus reliable.
I would never, ever, ever, EVER have blind trust on some anonymous code which lacks basic testing scenarios, especially more so if that code is supposed to be run billions of times in a row. The only result is what happened here, which can be painful.