Update clip_model.py

clip_model.py

@@ -56,8 +56,14 @@ class ClipModel:
         batch_size = 16  # Process 16 vocab at a time
         similarities = []
 
-        # Release unused memory
-        torch.cuda.empty_cache()
+        # Check memory usage before calling empty_cache
+        reserved_memory = torch.cuda.memory_reserved()
+        allocated_memory = torch.cuda.memory_allocated()
+        
+        # Only call empty_cache if reserved memory exceeds a threshold
+        if reserved_memory > 0.8 * torch.cuda.get_device_properties(0).total_memory:
+            # Release unused memory
+            torch.cuda.empty_cache()
 
         with torch.no_grad():
             for i in range(0, len(self.vocab), batch_size):