Found 1 document(s)
Page 1 of 1
-
Low-precision arithmetic can make reinforcement learning (RL) for large language models cheaper, but RL requires the sampler and the learner to agree. Several frameworks now run FP8 on both sides. We implement end-to-end...