Comment by llm_trw

None of those papers actually use quantized training, they are all about quantized inference.

Which is rather unfortunate as it means that the difference between what you can train locally and what you can run locally is growing ever larger.