Update app.py

app.py

@@ -1,4 +1,3 @@
-# float16 +32 
 import os
 import random
 import numpy as np
@@ -23,11 +22,35 @@ STORAGE_PSWD = os.getenv('STORAGE_PSWD', '').strip()  # SFTP Passwort
 STORAGE_PORT = int(os.getenv('STORAGE_PORT', '22').strip())  # SFTP Port
 STORAGE_SECRET = os.getenv('STORAGE_SECRET', '').strip()  # Secret Token
 
-# Modell-Konfiguration
-available_models = {
-    "sd3-medium": "stabilityai/stable-diffusion-3-medium-diffusers",
-    "sd2-base": "stabilityai/stable-diffusion-2-1-base"
-}
+# Modell-Konfiguration und Device-Setup
+device = "cuda" if torch.cuda.is_available() else "cpu"
+print(f"Using device: {device}")
+
+# Stelle fest, ob auf CPU oder GPU-System
+is_gpu_available = torch.cuda.is_available()
+
+# Modell laden - passend zur Hardware
+repo = "stabilityai/stable-diffusion-3-medium-diffusers"
+
+# Die Standard-Präzision basiert auf verfügbarer Hardware
+DEFAULT_PRECISION = "float16" if is_gpu_available else "float32"
+print(f"Default precision: {DEFAULT_PRECISION}")
+
+# Modell beim Start laden
+try:
+    # Wähle Präzision basierend auf Hardware
+    if DEFAULT_PRECISION == "float16":
+        pipe = StableDiffusionPipeline.from_pretrained(repo, torch_dtype=torch.float16).to(device)
+    else:  # float32 für CPU
+        pipe = StableDiffusionPipeline.from_pretrained(repo, torch_dtype=torch.float32).to(device)
+    
+    print("Model loaded successfully")
+except Exception as e:
+    raise RuntimeError(f"Failed to load the model. Ensure the token has access to the repo. Error: {e}")
+
+# Maximalwerte
+MAX_SEED = np.iinfo(np.int32).max
+MAX_IMAGE_SIZE = 1344
 
 # SFTP-Funktion
 def upload_to_sftp(local_file, remote_path):
@@ -44,46 +67,8 @@ def upload_to_sftp(local_file, remote_path):
         print(f"Error during SFTP upload: {e}")
         return False
 
-# Modell laden Funktion
-def load_model(model_name, precision):
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-    repo = available_models.get(model_name, available_models["sd3-medium"])
-    
-    try:
-        # Wähle Präzision basierend auf Auswahl
-        if precision == "float16":
-            torch_dtype = torch.float16
-        else:  # float32
-            torch_dtype = torch.float32
-            
-        pipe = StableDiffusionPipeline.from_pretrained(
-            repo, 
-            torch_dtype=torch_dtype
-        ).to(device)
-        
-        # Wenn auf CPU und Speicheroptimierung gewünscht
-        if device == "cpu":
-            pipe.enable_sequential_cpu_offload()
-            
-        return pipe
-    except Exception as e:
-        raise RuntimeError(f"Failed to load the model. Ensure the token has access to the repo. Error: {e}")
-
-# Maximalwerte
-MAX_SEED = np.iinfo(np.int32).max
-MAX_IMAGE_SIZE = 1344
-
-# Globale Pipe-Variable
-pipe = None
-
 # Inferenz-Funktion
-def infer(prompt, width, height, guidance_scale, num_inference_steps, seed, randomize_seed, model_name, precision):
-    global pipe
-    
-    # Prüfe, ob Modell neu geladen werden muss
-    if pipe is None:
-        pipe = load_model(model_name, precision)
-    
+def infer(prompt, width, height, guidance_scale, num_inference_steps, seed, randomize_seed):
     if randomize_seed:
         seed = random.randint(0, MAX_SEED)
     
@@ -109,41 +94,9 @@ def infer(prompt, width, height, guidance_scale, num_inference_steps, seed, rand
     else:
         return "Failed to upload image", seed
 
-# Modell neu laden
-def reload_model(model_name, precision):
-    global pipe
-    pipe = load_model(model_name, precision)
-    return f"Model loaded: {model_name} with {precision} precision"
-
 # Gradio-App
 with gr.Blocks() as demo:
-    gr.Markdown("### Stable Diffusion - Test App")
-    
-    with gr.Row():
-        with gr.Column():
-            # Modell Auswahl
-            model_name = gr.Radio(
-                choices=list(available_models.keys()),
-                value="sd3-medium",
-                label="Model"
-            )
-            
-            # Präzision Auswahl
-            precision = gr.Radio(
-                choices=["float16", "float32"],
-                value="float16",
-                label="Precision"
-            )
-            
-            reload_button = gr.Button("Load/Reload Model")
-            model_status = gr.Textbox(label="Model Status")
-            
-            # Modell laden Button
-            reload_button.click(
-                reload_model,
-                inputs=[model_name, precision],
-                outputs=[model_status]
-            )
+    gr.Markdown(f"### Stable Diffusion 3 - Test App (Running on {device.upper()} with {DEFAULT_PRECISION})")
     
     with gr.Row():
         with gr.Column():
@@ -161,8 +114,7 @@ with gr.Blocks() as demo:
         infer,
         inputs=[
             prompt, width, height, guidance_scale, 
-            num_inference_steps, seed, randomize_seed,
-            model_name, precision
+            num_inference_steps, seed, randomize_seed
         ],
         outputs=[output, seed]
     )