move models

app.py

@@ -10,13 +10,14 @@ import torchvision.transforms.functional as TF
 from diffusion import make_sigmas
 from huggingface_hub import hf_hub_download
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
-pipe = AutoPipelineForInpainting.from_pretrained("diffusers/stable-diffusion-xl-1.0-inpainting-0.1", torch_dtype=torch.float16, variant="fp16").to(device)
+pipe = AutoPipelineForInpainting.from_pretrained("diffusers/stable-diffusion-xl-1.0-inpainting-0.1", torch_dtype=torch.float16, variant="fp16")
+pipe.text_encoder.to("cuda")
+pipe.text_encoder_2.to("cuda")
 
-comparing_unet = SDXLUNet.load(hf_hub_download("stabilityai/stable-diffusion-xl-base-1.0", "unet/diffusion_pytorch_model.fp16.safetensors"), device=device)
-comparing_vae = SDXLVae.load(hf_hub_download("madebyollin/sdxl-vae-fp16-fix", "diffusion_pytorch_model.safetensors"), device=device)
+comparing_unet = SDXLUNet.load(hf_hub_download("stabilityai/stable-diffusion-xl-base-1.0", "unet/diffusion_pytorch_model.fp16.safetensors"))
+comparing_vae = SDXLVae.load(hf_hub_download("madebyollin/sdxl-vae-fp16-fix", "diffusion_pytorch_model.safetensors"))
 comparing_vae.to(torch.float16)
-comparing_controlnet = SDXLControlNetPreEncodedControlnetCond.load(hf_hub_download("williamberman/sdxl_controlnet_inpainting", "sdxl_controlnet_inpaint_pre_encoded_controlnet_cond_checkpoint_200000.safetensors"), device=device)
+comparing_controlnet = SDXLControlNetPreEncodedControlnetCond.load(hf_hub_download("williamberman/sdxl_controlnet_inpainting", "sdxl_controlnet_inpaint_pre_encoded_controlnet_cond_checkpoint_200000.safetensors"))
 comparing_controlnet.to(torch.float16)
 
 def read_content(file_path: str) -> str:
@@ -44,7 +45,17 @@ def predict(dict, prompt="", negative_prompt="", guidance_scale=7.5, steps=20, s
     init_image = dict["image"].convert("RGB").resize((1024, 1024))
     mask = dict["mask"].convert("RGB").resize((1024, 1024))
     
+    pipe.vae.to('cuda')
+    pipe.unet.to('cuda')
+    pipe.controlnet.to('cuda')
     output = pipe(prompt = prompt, negative_prompt=negative_prompt, image=init_image, mask_image=mask, guidance_scale=guidance_scale, num_inference_steps=int(steps), strength=strength)
+    pipe.vae.to('cpu')
+    pipe.unet.to('cpu')
+    pipe.controlnet.to('cpu')
+
+    comparing_unet.to('cuda')
+    comparing_vae.to('cuda')
+    comparing_controlnet.to('cuda')
 
     image = TF.to_tensor(dict["image"].convert("RGB").resize((1024, 1024)))
     mask = TF.to_tensor(dict["mask"].convert("L").resize((1024, 1024)))
@@ -62,6 +73,10 @@ def predict(dict, prompt="", negative_prompt="", guidance_scale=7.5, steps=20, s
         text_encoder_one=pipe.text_encoder, text_encoder_two=pipe.text_encoder_2, unet=comparing_unet, controlnet=comparing_controlnet
     )
     out = comparing_vae.output_tensor_to_pil(comparing_vae.decode(out))
+
+    comparing_unet.to('cpu')
+    comparing_vae.to('cpu')
+    comparing_controlnet.to('cpu')
     
     return output.images[0], out[0], gr.update(visible=True)