Update src/ai_processor.py

src/ai_processor.py

@@ -159,11 +159,24 @@ def _medgemma_generate_gpu_with_pipeline(
     except Exception:
         BitsAndBytesConfig = None
 
+    # <<< START OF FIX >>>
+    # Force CUDA initialization to prevent IndexError in bitsandbytes/triton check.
+    # This ensures the CUDA context is ready before transformers and bnb probe the device.
+    use_cuda = torch.cuda.is_available()
+    if use_cuda:
+        try:
+            torch.tensor([1.0]).cuda()
+        except Exception as e:
+            # If even this fails, CUDA is truly not working.
+            print(f"WARNING: CUDA pre-initialization failed: {e}")
+            use_cuda = False
+    # <<< END OF FIX >>>
+
     hf_token = token or os.getenv("HF_TOKEN")
     mid = model_id or "unsloth/medgemma-4b-it-bnb-4bit"
 
     # device / dtype
-    use_cuda = torch.cuda.is_available()
+    # use_cuda is already defined above
     device   = 0 if use_cuda else -1
     dtype    = torch.bfloat16 if use_cuda else torch.float32
 
@@ -224,6 +237,8 @@ def _medgemma_generate_gpu_with_pipeline(
     return (text or "").strip() or "⚠️ Empty response"
 
 
+
+
 def generate_medgemma_report(
     patient_info: str,
     visual_results: Dict,