attempt to make space run on cpu

removed GPU decorator from predict_onnx()

dont try to load model every time, use global onnx session var

explicitly setting queue params demo.queue(concurrency_count=1, max_size=None, status_update_rate=1)

app.py

@@ -286,9 +286,13 @@ g_idx_to_tag = None
 g_tag_to_category = None
 g_current_model = None
 
+# --- Global ONNX session ---
+g_session = None
+
 # --- Initialization Function ---
 def initialize_onnx_paths(model_choice=DEFAULT_MODEL):
     global g_onnx_model_path, g_tag_mapping_path, g_labels_data, g_idx_to_tag, g_tag_to_category, g_current_model
+    global g_session
     
     if not model_choice in MODEL_OPTIONS:
         print(f"Invalid model choice: {model_choice}, falling back to default: {DEFAULT_MODEL}")
@@ -301,19 +305,40 @@ def initialize_onnx_paths(model_choice=DEFAULT_MODEL):
     
     print(f"Initializing ONNX paths and labels for model: {model_choice}...")
     hf_token = os.environ.get("HF_TOKEN")
+    
     try:
         print(f"Attempting to download ONNX model: {onnx_filename}")
-        g_onnx_model_path = hf_hub_download(repo_id=REPO_ID, filename=onnx_filename, cache_dir=CACHE_DIR, token=hf_token, force_download=False)
+        g_onnx_model_path = hf_hub_download(
+            repo_id=REPO_ID,
+            filename=onnx_filename,
+            cache_dir=CACHE_DIR,
+            token=hf_token,
+            force_download=False
+        )
         print(f"ONNX model path: {g_onnx_model_path}")
 
         print(f"Attempting to download Tag mapping: {tag_mapping_filename}")
-        g_tag_mapping_path = hf_hub_download(repo_id=REPO_ID, filename=tag_mapping_filename, cache_dir=CACHE_DIR, token=hf_token, force_download=False)
+        g_tag_mapping_path = hf_hub_download(
+            repo_id=REPO_ID,
+            filename=tag_mapping_filename,
+            cache_dir=CACHE_DIR,
+            token=hf_token,
+            force_download=False
+        )
         print(f"Tag mapping path: {g_tag_mapping_path}")
 
         print("Loading labels from mapping...")
         g_labels_data, g_idx_to_tag, g_tag_to_category = load_tag_mapping(g_tag_mapping_path)
         print(f"Labels loaded. Count: {len(g_labels_data.names)}")
-        
+
+        # Load ONNX session ONCE here
+        print("Creating ONNX Runtime session (CPUExecutionProvider)...")
+        g_session = ort.InferenceSession(
+            g_onnx_model_path,
+            providers=["CPUExecutionProvider"]
+        )
+        print("ONNX Runtime session ready.")
+
         return True
 
     except Exception as e:
@@ -341,7 +366,6 @@ def change_model(model_choice):
         return f"Error changing model: {str(e)}"
         
 # --- Main Prediction Function (ONNX) ---
-@spaces.GPU()
 def predict_onnx(image_input, model_choice, gen_threshold, char_threshold, output_mode):
     print(f"--- predict_onnx function started (GPU worker) with model {model_choice} ---")
     
@@ -371,7 +395,7 @@ def predict_onnx(image_input, model_choice, gen_threshold, char_threshold, outpu
             providers.append('CUDAExecutionProvider')
         providers.append('CPUExecutionProvider')
         print(f"Attempting to load session with providers: {providers}")
-        session = ort.InferenceSession(g_onnx_model_path, providers=providers)
+        session = g_session
         print(f"ONNX session loaded using: {session.get_providers()[0]}")
     except Exception as e:
         message = f"Error loading ONNX session in worker: {e}"
@@ -544,5 +568,6 @@ if __name__ == "__main__":
     if not os.environ.get("HF_TOKEN"): print("Warning: HF_TOKEN environment variable not set.")
     # Initialize paths and labels at startup (with default model)
     initialize_onnx_paths(DEFAULT_MODEL)
+    demo.queue(concurrency_count=1, max_size=None, status_update_rate=1)
     # Launch Gradio app
     demo.launch()