Is it possible to store half precision for training? #820
Description
For very large and detail scene, fp32 is slow and bloat, quickly eat up the memory. Is it possible to downscale it to fp16 or even fp8 in rerpresentation? Or at least for opacity, features_dc/rest, quats, scales (they are in good limited range)?