Update app.py

app.py

@@ -56,7 +56,7 @@ random_seed = 42
 video_length = 201
 W = 1024
 H = W
-device = "cuda"
+device = "cuda" if torch.cuda.is_available() else "cpu"
 
 def get_pipe_image_and_video_predictor():
     vae = AutoencoderKLWan.from_pretrained("./model/vae", torch_dtype=torch.float16)
@@ -177,7 +177,7 @@ def preprocess_for_removal(images, masks):
         out_masks.append(msk_resized)
     arr_images = np.stack(out_images)
     arr_masks = np.stack(out_masks)
-    return torch.from_numpy(arr_images).half().to(device), torch.from_numpy(arr_masks).half().to(device)
+    return torch.from_numpy(arr_images).half(), torch.from_numpy(arr_masks).half()
 
 @spaces.GPU(duration=300)
 def inference_and_return_video(dilation_iterations, num_inference_steps, video_state=None):
@@ -189,7 +189,8 @@ def inference_and_return_video(dilation_iterations, num_inference_steps, video_s
     images = np.array(images)
     masks = np.array(masks)
     img_tensor, mask_tensor = preprocess_for_removal(images, masks)
-    mask_tensor = mask_tensor[:,:,:,:1]
+    img_tensor=img_tensor.to(device)
+    mask_tensor = mask_tensor[:,:,:,:1].to(device)
 
     if mask_tensor.shape[1] < mask_tensor.shape[2]:
         height = 480
@@ -218,7 +219,7 @@ def inference_and_return_video(dilation_iterations, num_inference_steps, video_s
     clip.write_videofile(video_file, codec='libx264', audio=False, verbose=False, logger=None)
     return video_file
 
-@spaces.GPU(duration=200)
+@spaces.GPU(duration=100)
 def track_video(n_frames, video_state):
 
     input_points = video_state["input_points"]