Basically stuff a 32 bit value into an 8 bit value (and lose precision). Apparently it doesn't affect the results significantly. More info: https://github.co

Basically stuff a 32 bit value into an 8 bit value (and lose precision). Apparently it doesn't affect the results significantly. More info: https://github.com/huggingface/transformers/pull/17901 https://github.com/huggingface/transformers/pull/17901