[CodeGen] Aligned the output format and fixed acc benchmark issues. (#1981)

Author: Zhenzhong1

CodeGen/benchmark/accuracy/README.md

@@ -34,6 +34,7 @@ We utilize the [bigcode-evaluation-harness](https://github.com/bigcode-project/b
 
 2.  **Install Dependencies:**
     ```shell
+    # pip install -r requirements-hpu.txt, if you deploy microservices on the Gaudi.
     pip install -r requirements.txt
     pip install -e .
     ```

CodeGen/codegen.py

@@ -280,7 +280,15 @@ async def handle_request(self, request: Request):
 
         # Get the response from the last node in the runtime graph
         last_node = runtime_graph.all_leaves()[-1]
-        response = result_dict[last_node]["text"]
+
+        try:
+            response = result_dict[last_node]["choices"][0]["text"]
+        except (KeyError, IndexError, TypeError):
+            try:
+                response = result_dict[last_node]["text"]
+            except (KeyError, TypeError):
+                response = "Response Error"
+
         choices = []
         usage = UsageInfo()
         choices.append(