llama_guard inference fix

llama_toolchain/safety/shields/llama_guard.py

@@ -240,9 +240,39 @@ class LlamaGuardShield(ShieldBase):
             response = self.tokenizer.decode(
                 generated_tokens[0], skip_special_tokens=True
             )
-
+            cprint(f" Llama Guard response {response}", color="magenta")
             response = response.strip() 
             shield_response = self.get_shield_response(response)
-
             cprint(f"Final Llama Guard response {shield_response}", color="magenta")
             return shield_response
+
+
+
+
+
+        '''if  self.disable_input_check and messages[-1].role == "user":
+            return ShieldResponse(is_violation=False)
+        elif self.disable_output_check and messages[-1].role == "assistant":
+            return ShieldResponse(is_violation=False)
+        else:
+            prompt = self.build_prompt(messages)
+            llama_guard_input = {
+                "role": "user",
+                "content": prompt,
+            }
+            input_ids = self.tokenizer.apply_chat_template(
+                [llama_guard_input], return_tensors="pt", tokenize=True
+            ).to(self.device)
+            prompt_len = input_ids.shape[1]
+            output = self.model.generate(
+                input_ids=input_ids,
+                max_new_tokens=50,
+                output_scores=True,
+                return_dict_in_generate=True,
+                pad_token_id=0
+            )
+            generated_tokens = output.sequences[:, prompt_len:]
+            response = self.tokenizer.decode(generated_tokens[0], skip_special_tokens=True)
+            response = response.strip()
+            shield_response = self.get_shield_response(response)
+            return shield_response'''