Does BOFT-training switch to float32?

Based on the training speed, sample generation speed, and final file size compared to Diag-OFT, it appears that BOFT-training has switched to float32. Is there a possibility to implement mixed precision?

Start by locating the BOFT-training and Diag-OFT training entry points, then compare how each path selects numeric precision. Check the existing training configuration and precision-related tests or examples, if present, to establish the expected mixed-precision behavior. Done means BOFT training supports the agreed precision mode without regressing generation speed or output compatibility.

Written by the indexing model from the issue text.