Fix UnboundLocalError and Triton cache permission error

app.py

@@ -3,6 +3,14 @@ import os
 if 'PYTORCH_ALLOC_CONF' not in os.environ:
     os.environ['PYTORCH_ALLOC_CONF'] = 'expandable_segments:True,max_split_size_mb=128'
 
+if 'TRITON_CACHE_DIR' not in os.environ:
+    triton_cache = '/tmp/.triton'
+    os.environ['TRITON_CACHE_DIR'] = triton_cache
+    try:
+        os.makedirs(triton_cache, exist_ok=True)
+    except:
+        pass
+
 import tempfile
 import base64
 import torch

inference.py

@@ -148,6 +148,8 @@ def predict_volume(nifti_file, modality, slice_idx=None):
         print("Running inference...")
         inference_start = time.time()
         
+        y1, y2, y3, y4 = None, None, None, None
+        
         gpu_handle = None
         if model_loader.DEVICE.type == 'cuda':
             allocated = torch.cuda.memory_allocated(0) / (1024**3)
@@ -370,8 +372,13 @@ def predict_volume(nifti_file, modality, slice_idx=None):
         finally:
             timeout_timer.cancel()
             
-            print(f"  → Raw model output shapes: y1={y1.shape if hasattr(y1, 'shape') else 'N/A'}, y2={y2.shape if hasattr(y2, 'shape') else 'N/A'}")
-            sys.stdout.flush()
+            if y1 is not None:
+                print(f"  → Raw model output shapes: y1={y1.shape if hasattr(y1, 'shape') else 'N/A'}, y2={y2.shape if hasattr(y2, 'shape') else 'N/A'}")
+                sys.stdout.flush()
+            else:
+                print("  → Inference failed before model output was generated")
+                sys.stdout.flush()
+                raise RuntimeError("Inference failed: model did not produce output")
             
             pred = y1
             if isinstance(pred, torch.Tensor):