Describe the bug
Models loaded from .ckpt / .safetensors file are using Float32 automatically, with no obvious way to change this behavior.
Reproduction
- Import
load_pipeline_from_original_stable_diffusion_ckpt from diffusers.pipelines.stable_diffusion.convert_from_ckpt
- Load model from .ckpt file
- Observe extreme VRAM usage compared to Float16
Logs
No response
System Info
Diffusers: 0.14.0
OS: Linux (Kubuntu 22.04)
Python 3.10.6