Update app.py

app.py

@@ -10,9 +10,9 @@ from diffusers import EulerDiscreteScheduler
 import gradio as gr
 
 # Download the model files
-#ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors")
+ckpt_dir = snapshot_download(repo_id="Kwai-Kolors/Kolors")
 
-# Function to load models on demand
+# Function to load models
 def load_models():
     text_encoder = ChatGLMModel.from_pretrained(
         os.path.join(ckpt_dir, 'text_encoder'),
@@ -29,21 +29,20 @@ def load_models():
         unet=unet,
         scheduler=scheduler,
         force_zeros_for_empty_prompt=False
-    ).to("cuda")
+    )
 
 # Create a global variable to hold the pipeline
 pipe = load_models()
 
-@spaces.GPU(duration=200)
 def generate_image(prompt, negative_prompt, height, width, num_inference_steps, guidance_scale, num_images_per_prompt, use_random_seed, seed, progress=gr.Progress(track_tqdm=True)):
     if use_random_seed:
         seed = random.randint(0, 2**32 - 1)
     else:
         seed = int(seed)  # Ensure seed is an integer
     
-    # Move the model to the GPU for inference and clear unnecessary variables
+    # Move the model to the CPU for inference and clear unnecessary variables
     with torch.no_grad():
-        generator = torch.Generator(pipe.device).manual_seed(seed)
+        generator = torch.Generator().manual_seed(seed)
         result = pipe(
             prompt=prompt,
             negative_prompt=negative_prompt,
@@ -58,8 +57,6 @@ def generate_image(prompt, negative_prompt, height, width, num_inference_steps,
     
     return image, seed
 
-
-
 # Gradio interface
 iface = gr.Interface(
     fn=generate_image,