Flash Attention regurgitates repeated tokens - seq2seq

Having some generation issues with NMT models trained with OpenNMT-py, which include OpenNMT-py versions before flash attention existed and one I'm currently training with the most recent which include flash attention. Models were converted using onmt_release_model with storage quantization set to int8.

Happens when turning flash_attention=True on when creating the ctranslate.Translator object. GPU is a RTX 3090.

Don't know if this is just an arch issue or something to do with the conversion process from opennmt-py.

Examples of some outputs given the Flores200 benchmark

sss of of of of of of of of of of of
sss                                                     in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in in          in in in in in in   patients patients patients in in in in in in in  in in in in in   in in in in in in in in in in in in in in in in in in in in in in patients in in in in in      in in in in in in in in patients patients patients patients patients patients patients patients patients patients patients patients patients in in patients patients patients patients patients in countries countries in in in in in in in in in             in in in in in in in in in in     in in in in in
ssss
ssmmmmmmmm
ss
__opt_src_en__opt_src_en__opt_src_en
sss
sss                       of of of of of of of of of                         of of
sss                                                tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax                  tax tax tax tax tax tax tax tax tax tax tax tax tax     tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax tax

sssmmmmmmmmmmmmmmmmmmm

Start by reproducing the Flores200 outputs with an OpenNMT-py seq2seq model converted by onmt_release_model using int8 storage quantization, then compare translation with flash_attention=True and False on an RTX 3090. The investigation should determine whether repeated tokens originate in flash attention, the GPU architecture, or model conversion; no source file or test is identified in the issue.