Open javafa opened 5 months ago
loss calculation is wrong.
Thank you @javafa for adding the quantization options! I will check it soon and merge it if it is good to go👍
calculation nagative values should be changed ... selected_index[selected_index < 0] += filtered_logits.size(1) ...
Added quantization options when use quantization, selected_indexs in compute_logps() could be negative values