Update app.py

app.py

@@ -70,6 +70,25 @@ class BasicAgent:
                     device_map="auto",
                 )
 
+                if hasattr(self.model, 'tokenizer') and self.model.tokenizer is not None:
+                    # Set left padding for better batching with causal models
+                    self.model.tokenizer.padding_side = "left"
+                    # Ensure pad token is set
+                    if self.model.tokenizer.pad_token is None:
+                        self.model.tokenizer.pad_token = self.model.tokenizer.eos_token
+                    # Set default padding behavior for FlashAttention
+                    self.model.tokenizer.pad_to_multiple_of = 8
+                    print("✅ Applied tokenizer padding fix for FlashAttention alignment")
+                
+                # If the model has a processor with tokenizer, fix that too
+                if hasattr(self.model, 'processor') and hasattr(self.model.processor, 'tokenizer'):
+                    self.model.processor.tokenizer.padding_side = "left"
+                    if self.model.processor.tokenizer.pad_token is None:
+                        self.model.processor.tokenizer.pad_token = self.model.processor.tokenizer.eos_token
+                    self.model.processor.tokenizer.pad_to_multiple_of = 8
+                    print("✅ Applied processor tokenizer padding fix")
+
+
                 # Verify where model actually loaded
                 if hasattr(self.model, 'device'):
                     print(f"✅ Model loaded on device: {self.model.device}")