• Simon Layton's avatar
    Fix fp16 masking in PoolerEndLogits · ec94f4e0
    Simon Layton authored
    Necessary to run xlnet (at least in squad) with `--fp16 --fp16_opt_level="O2"`, otherwise loss is immediately `NaN` and fine-tuning cannot proceed.
    ec94f4e0
modeling_utils.py 39.6 KB