Cosmos Predict2 just doesn't work at all in fp8. Both e4m3fn and e5m2 are very noisy, melted, and fried, in different ways. Makes it annoying to test out the 14B, because while it can run in bf16 on a 4090 since ComfyUI does some kind of auto layer offloading, it's really slow. Maybe GGUF q8 fixes this problem?