Add option for float32 sampling with float16 UNet

This also handles type casting so that ROCm and MPS torch devices work correctly without --no-half. One cast is required for deepbooru in deepbooru_model.py, some explicit casting is required for img2img and inpainting. depth_model can't be converted to float16 or it won't work correctly on some systems (it's known to have issues on MPS) so in sd_models.py model.depth_model is removed for model.half().

1 files changed, 1 insertions, 0 deletions

diff --git a/modules/shared.py b/modules/shared.py
index 5f713bee..4ce1209b 100644
--- a/modules/shared.py
+++ b/modules/shared.py
@@ -45,6 +45,7 @@ parser.add_argument("--lowram", action='store_true', help="load stable diffusion
 parser.add_argument("--always-batch-cond-uncond", action='store_true', help="disables cond/uncond batching that is enabled to save memory with --medvram or --lowvram")

 parser.add_argument("--unload-gfpgan", action='store_true', help="does not do anything.")

 parser.add_argument("--precision", type=str, help="evaluate at this precision", choices=["full", "autocast"], default="autocast")

+parser.add_argument("--upcast-sampling", action='store_true', help="upcast sampling. No effect with --no-half. Usually produces similar results to --no-half with better performance while using less memory.")

 parser.add_argument("--share", action='store_true', help="use share=True for gradio and make the UI accessible through their site")

 parser.add_argument("--ngrok", type=str, help="ngrok authtoken, alternative to gradio --share", default=None)

 parser.add_argument("--ngrok-region", type=str, help="The region in which ngrok should start.", default="us")