From 84d9ce30cb427759547bc7876ed80ab91787d175 Mon Sep 17 00:00:00 2001 From: brkirch Date: Tue, 24 Jan 2023 23:51:45 -0500 Subject: 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(). --- modules/shared.py | 1 + 1 file changed, 1 insertion(+) (limited to 'modules/shared.py') 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") -- cgit v1.2.1 From e3b53fd295aca784253dfc8668ec87b537a72f43 Mon Sep 17 00:00:00 2001 From: brkirch Date: Wed, 25 Jan 2023 00:23:10 -0500 Subject: Add UI setting for upcasting attention to float32 Adds "Upcast cross attention layer to float32" option in Stable Diffusion settings. This allows for generating images using SD 2.1 models without --no-half or xFormers. In order to make upcasting cross attention layer optimizations possible it is necessary to indent several sections of code in sd_hijack_optimizations.py so that a context manager can be used to disable autocast. Also, even though Stable Diffusion (and Diffusers) only upcast q and k, unfortunately my findings were that most of the cross attention layer optimizations could not function unless v is upcast also. --- modules/shared.py | 1 + 1 file changed, 1 insertion(+) (limited to 'modules/shared.py') diff --git a/modules/shared.py b/modules/shared.py index 4ce1209b..6a0b96cb 100644 --- a/modules/shared.py +++ b/modules/shared.py @@ -410,6 +410,7 @@ options_templates.update(options_section(('sd', "Stable Diffusion"), { "comma_padding_backtrack": OptionInfo(20, "Increase coherency by padding from the last comma within n tokens when using more than 75 tokens", gr.Slider, {"minimum": 0, "maximum": 74, "step": 1 }), "CLIP_stop_at_last_layers": OptionInfo(1, "Clip skip", gr.Slider, {"minimum": 1, "maximum": 12, "step": 1}), "extra_networks_default_multiplier": OptionInfo(1.0, "Multiplier for extra networks", gr.Slider, {"minimum": 0.0, "maximum": 1.0, "step": 0.01}), + "upcast_attn": OptionInfo(False, "Upcast cross attention layer to float32"), })) options_templates.update(options_section(('compatibility', "Compatibility"), { -- cgit v1.2.1