Check if completion_tokens is none in error handler.

abetlen · abetlen · commit 99ff17556284 · 2023-12-22T13:41:06.000-05:00
diff --git a/llama_cpp/server/errors.py b/llama_cpp/server/errors.py
@@ -72,7 +72,7 @@ def context_length_exceeded(
         return 400, ErrorResponse(
             message=message.format(
                 context_window,
-                completion_tokens + prompt_tokens,
+                (completion_tokens or 0) + prompt_tokens,
                 prompt_tokens,
                 completion_tokens,
             ), # type: ignore
