What an interesting question! I'm always curious to see how different quantization approaches can impact the performance of QAT models. This seems like it could be a really insightful study. https://www.reddit.com/user/we_are_mammals