fix: Add attention_mask to improve generation reliability

Warning: The attention mask is not set and cannot be inferred from input

🔧 Fix:
- Use tokenizer() instead of tokenizer.encode()
- Explicitly create and pass attention_mask to generate()
- Add padding=True to ensure proper mask generation
- Pass attention_mask parameter to model.generate()

✨ Benefits:
- More reliable text generation
- Eliminates attention mask warning
- Proper handling of padded sequences
- Better model behavior with variable length inputs

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <[email protected]>

app.py

@@ -106,18 +106,22 @@ def generate_response_impl(message, history):
 
         conversation += f"사용자: {message}\n어시스턴트:"
 
-        # Tokenize
-        inputs = current_tokenizer.encode(
+        # Tokenize with attention_mask
+        encoded = current_tokenizer(
             conversation,
             return_tensors="pt",
             truncation=True,
             max_length=512,
-        ).to(device)
+            padding=True,
+        )
+        inputs = encoded['input_ids'].to(device)
+        attention_mask = encoded['attention_mask'].to(device)
 
         # Generate response
         with torch.no_grad():
             outputs = current_model.generate(
                 inputs,
+                attention_mask=attention_mask,
                 max_new_tokens=MODEL_CONFIG["max_length"],
                 temperature=0.7,
                 top_p=0.9,