update app

app.py

@@ -3,7 +3,6 @@ import os
 import shutil
 import time
 import uuid
-import spaces #[zeroGPU Spaces]
 import unicodedata
 from io import BytesIO
 from threading import Timer
@@ -12,6 +11,7 @@ from datetime import datetime
 
 import gradio as gr
 import torch
+import spaces  # <--- Added Spaces support
 from dotenv import load_dotenv
 from e2b_desktop import Sandbox
 from gradio_modal import Modal
@@ -39,7 +39,7 @@ load_dotenv(override=True)
 # -----------------------------------------------------------------------------
 
 E2B_API_KEY = os.getenv("E2B_API")
-HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_API")
+HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACE_API_KEY")
 if HF_TOKEN:
     login(token=HF_TOKEN)
 
@@ -58,7 +58,7 @@ if not os.path.exists(TMP_DIR):
 
 print("Loading Fara Model... This may take a moment.")
 DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
-MODEL_ID_F = "microsoft/Fara-7B"  # Ensure this ID is accessible or point to local path
+MODEL_ID_F = "microsoft/Fara-7B"  # Ensure this repository exists and you have access
 
 try:
     processor_f = AutoProcessor.from_pretrained(MODEL_ID_F, trust_remote_code=True)
@@ -66,19 +66,27 @@ try:
         MODEL_ID_F,
         trust_remote_code=True,
         torch_dtype=torch.bfloat16 if DEVICE == "cuda" else torch.float32,
-        device_map="auto" if DEVICE == "cuda" else None,
+        device_map="auto",
     )
-    if DEVICE == "cpu":
-        model_f.to(DEVICE)
-    
-    model_f.eval()
     print(f"Fara Model loaded successfully on {DEVICE}")
 except Exception as e:
     print(f"Error loading Fara Model: {e}")
-    print("Please ensure you have access to the model and enough GPU memory.")
-    # Fallback to prevent crash during import, though app won't work without model
-    model_f = None
-    processor_f = None
+    print("Falling back to Qwen/Qwen2.5-VL-7B-Instruct for demonstration if Fara is unavailable...")
+    try:
+        # Fallback to base Qwen-VL if Fara repo isn't public/accessible
+        MODEL_ID_F = "Qwen/Qwen2.5-VL-7B-Instruct" 
+        processor_f = AutoProcessor.from_pretrained(MODEL_ID_F, trust_remote_code=True)
+        model_f = Qwen2_5_VLForConditionalGeneration.from_pretrained(
+            MODEL_ID_F,
+            trust_remote_code=True,
+            torch_dtype=torch.bfloat16 if DEVICE == "cuda" else torch.float32,
+            device_map="auto",
+        )
+        print(f"Fallback Model ({MODEL_ID_F}) loaded successfully.")
+    except Exception as inner_e:
+        print(f"Critical error loading model: {inner_e}")
+        model_f = None
+        processor_f = None
 
 class FaraLocalModel(Model):
     """
@@ -98,12 +106,9 @@ class FaraLocalModel(Model):
         if self.model is None:
             raise ValueError("Fara Model is not loaded.")
 
-        # Convert SmolAgents messages to Qwen/Transformers format
-        # SmolAgents uses a specific dict structure for content. 
-        # We need to normalize it for process_vision_info / apply_chat_template
-        
         formatted_messages = []
         
+        # Convert SmolAgents messages to Qwen/Transformers format
         for msg in messages:
             role = msg["role"]
             content = msg["content"]
@@ -150,7 +155,7 @@ class FaraLocalModel(Model):
                 **inputs,
                 max_new_tokens=kwargs.get("max_tokens", 1024),
                 stop_strings=stop_sequences,
-                tokenizer=self.processor.tokenizer, # Specific for stop_strings in modern transformers
+                tokenizer=self.processor.tokenizer, 
             )
 
         # Decode
@@ -185,7 +190,7 @@ Action:
 click(254, 308)
 ```<end_code>
 
-Akways format your action ('Action:' part) as Python code blocks as shown above.
+Always format your action ('Action:' part) as Python code blocks as shown above.
 </action_process>
 
 <tools>
@@ -220,7 +225,7 @@ In browser, ignore any sign-in popups while they don't interfere with the elemen
 </general_guidelines>
 """.replace("<<current_date>>", datetime.now().strftime("%A, %d-%B-%Y"))
 
-@spaces.GPU
+
 def draw_marker_on_image(image_copy, click_coordinates):
     x, y = click_coordinates
     draw = ImageDraw.Draw(image_copy)
@@ -750,6 +755,7 @@ def initialize_session(interactive_mode, browser_uuid):
         return update_html(interactive_mode, browser_uuid), browser_uuid
 
 class EnrichedGradioUI(GradioUI):
+    @spaces.GPU(duration=180) # Allocate GPU for 3 minutes per interaction cycle
     def interact_with_agent(
         self,
         task_input,
@@ -836,7 +842,7 @@ with gr.Blocks(theme=theme, css=custom_css, js=custom_js) as demo:
         with gr.Sidebar(position="left"):
             with Modal(visible=True) as modal:
                 gr.Markdown("""### Welcome to Fara CUA Demo 🖥️
-This agent uses **microsoft/Fara-7B** (running locally) and **smolagents** to control a remote computer.
+This agent uses **microsoft/Fara-7B** (running locally via ZeroGPU) and **smolagents** to control a remote computer.
 
 👉 Type a task, click 'Let's go!', and watch the agent work.
 """)