Fix: Force CPU mode for LaMa + robust JSON parsing

app.py

@@ -39,24 +39,39 @@ executor = ThreadPoolExecutor(max_workers=min(NUM_CPUS, 4))
 # MODEL PRELOADING (Load once at startup, not per-request)
 # ============================================================
 
+# Force CPU mode (no GPU available)
+os.environ["CUDA_VISIBLE_DEVICES"] = ""
+import torch
+torch.set_default_device('cpu')
+
 print("🚀 BubbleScribe starting up...")
 print(f"   CPU Threads: {NUM_CPUS}")
+print(f"   Device: CPU (forced)")
 
 # Global LaMa model (loaded once)
 _lama_model = None
 _lama_lock = threading.Lock()
 
 def get_lama_model():
-    """Get or initialize LaMa model (singleton pattern)."""
+    """Get or initialize LaMa model (singleton pattern) - CPU only."""
     global _lama_model
     if _lama_model is None:
         with _lama_lock:
             if _lama_model is None:
-                print("📦 Loading LaMa inpainting model...")
+                print("📦 Loading LaMa inpainting model (CPU mode)...")
                 try:
                     from simple_lama_inpainting import SimpleLama
-                    _lama_model = SimpleLama()
-                    print("✅ LaMa model loaded!")
+                    _lama_model = SimpleLama(device=torch.device('cpu'))
+                    print("✅ LaMa model loaded on CPU!")
+                except TypeError:
+                    # Older version without device parameter
+                    try:
+                        from simple_lama_inpainting import SimpleLama
+                        _lama_model = SimpleLama()
+                        print("✅ LaMa model loaded!")
+                    except Exception as e:
+                        print(f"⚠️ LaMa failed to load: {e}")
+                        _lama_model = "fallback"
                 except Exception as e:
                     print(f"⚠️ LaMa failed to load: {e}")
                     _lama_model = "fallback"
@@ -152,6 +167,90 @@ def scale_bbox(bbox: list, original_size: tuple, processed_size: tuple) -> list:
         int(bbox[3] * scale_y)
     ]
 
+# ============================================================
+# JSON REPAIR (Handle malformed model responses)
+# ============================================================
+
+def repair_json(text: str) -> str:
+    """Attempt to repair common JSON issues from LLM responses."""
+    # Remove any markdown code blocks
+    text = re.sub(r'```json\s*', '', text)
+    text = re.sub(r'```\s*', '', text)
+    
+    # Fix unescaped newlines in strings
+    text = re.sub(r'(?<!\\)\n(?=[^"]*"[^"]*(?:"[^"]*"[^"]*)*$)', '\\n', text)
+    
+    # Fix trailing commas before ] or }
+    text = re.sub(r',\s*([}\]])', r'\1', text)
+    
+    # Fix missing commas between objects
+    text = re.sub(r'\}\s*\{', '},{', text)
+    
+    # Fix unescaped quotes inside strings (rough heuristic)
+    # Replace Japanese quotes with escaped ones
+    text = text.replace('「', '\\"').replace('」', '\\"')
+    text = text.replace('『', '\\"').replace('』', '\\"')
+    
+    return text
+
+def safe_parse_json(text: str) -> list:
+    """Safely parse JSON with multiple fallback strategies."""
+    # Strategy 1: Direct parse
+    try:
+        json_match = re.search(r'\[[\s\S]*\]', text)
+        if json_match:
+            return json.loads(json_match.group())
+    except json.JSONDecodeError:
+        pass
+    
+    # Strategy 2: Repair and parse
+    try:
+        repaired = repair_json(text)
+        json_match = re.search(r'\[[\s\S]*\]', repaired)
+        if json_match:
+            return json.loads(json_match.group())
+    except json.JSONDecodeError:
+        pass
+    
+    # Strategy 3: Extract individual objects
+    try:
+        objects = re.findall(r'\{[^{}]*\}', text)
+        results = []
+        for obj in objects:
+            try:
+                parsed = json.loads(repair_json(obj))
+                if 'bbox' in parsed:
+                    results.append(parsed)
+            except:
+                continue
+        if results:
+            return results
+    except:
+        pass
+    
+    # Strategy 4: Manual extraction with regex
+    try:
+        results = []
+        # Find bbox patterns
+        bbox_matches = re.findall(r'"bbox"\s*:\s*\[\s*(\d+)\s*,\s*(\d+)\s*,\s*(\d+)\s*,\s*(\d+)\s*\]', text)
+        original_matches = re.findall(r'"original"\s*:\s*"([^"]*)"', text)
+        translated_matches = re.findall(r'"translated"\s*:\s*"([^"]*)"', text)
+        
+        for i, bbox in enumerate(bbox_matches):
+            result = {
+                "bbox": [int(bbox[0]), int(bbox[1]), int(bbox[2]), int(bbox[3])],
+                "original": original_matches[i] if i < len(original_matches) else "",
+                "translated": translated_matches[i] if i < len(translated_matches) else ""
+            }
+            results.append(result)
+        
+        if results:
+            return results
+    except:
+        pass
+    
+    return []
+
 # ============================================================
 # DETECTION & TRANSLATION
 # ============================================================
@@ -222,11 +321,10 @@ Important:
         elif hasattr(msg, 'reasoning_content') and msg.reasoning_content:
             result_text = msg.reasoning_content
         
-        # Parse JSON from response
-        json_match = re.search(r'\[[\s\S]*\]', result_text)
-        if json_match:
-            detections = json.loads(json_match.group())
-            
+        # Parse JSON from response with robust error handling
+        detections = safe_parse_json(result_text)
+        
+        if detections:
             # Scale bboxes back to original size if needed
             if original_size != processed_size:
                 for det in detections: