This model is broken with asian characters
#3
by nyxia - opened
This model seems to be broken with asian (Chinese/Japanese characters) outputting mainly U+FFFD � (vllm 0.29.1.rc1)
Agreed. It gets lost in garbled output a lot
I used redhat’s nvfp4 quant and seeing other comments, looks like they fucked up the quant.
Typically it works and suddenly in the middle of a response it starts outputting paragraphs of completely random words. Usually English/Russian/Chinese
Is it possible to update it?
I used redhat’s nvfp4 quant and seeing other comments, looks like they fucked up the quant.
lol try higher bit or another provider ig