[Feature Request] FP16-Fixed version.
#4
by
gelukuMLG
- opened
Currently the model does not work in FP16, and causes NaN/black imagen when forced into that precision. BF16 isn't really good for people with older gpus than the 30 series since it upcasts it to fp32 making generation several times slower.
On my RTX 2060 in m getting 14s/it with BF16 upcast to fp32, and 3s/it in FP16 forced which leads to the mentioned above black images.