quantity with auto_gptq avg loss: nan
hgcdanniel opened this issue · comments
hi,when I try to use https://github.com/PanQiWei/AutoGPTQ to quantity the model finetune with startcoder, the Quantizing attn.c_attn in layer 1/40...2/40..... are all avg loss: nan, no mather 8/4/bit ,could you share your quantity script??
thanks