Hello author, In the "Batch Size" section of the KTO paper, it mentions that "KTO needs a microbatch size ≥ 2 to estimate the reference point in a single step. The experiments in this paper all use an effective batch size of 32, and in general we recommend using a batch size between 8 and 128." In the code repository, I noticed that when using four GPUs with the KTO algorithm, the batch size is set to 32 (i.e., a microbatch size of 8 per GPU). Does the "effective batch size of 32" mentioned in the paper refer to the macrobatch size being 32?
Hello author, In the "Batch Size" section of the KTO paper, it mentions that "KTO needs a microbatch size ≥ 2 to estimate the reference point in a single step. The experiments in this paper all use an effective batch size of 32, and in general we recommend using a batch size between 8 and 128." In the code repository, I noticed that when using four GPUs with the KTO algorithm, the batch size is set to 32 (i.e., a microbatch size of 8 per GPU). Does the "effective batch size of 32" mentioned in the paper refer to the macrobatch size being 32?