feat(utils.py): support passing openai response headers to client, if enabled

Allows openai/openai-compatible provider response headers to be sent to client, if 'return_response_headers' is enabled

litellm/proxy/_new_secret_config.yaml

@@ -7,3 +7,4 @@ model_list:
 litellm_settings:
   callbacks: ["logfire"]
   redact_user_api_key_info: true
+  return_response_headers: true

litellm/tests/test_completion_cost.py

@@ -896,6 +896,9 @@ async def test_completion_cost_hidden_params(sync_mode):
 
     assert "response_cost" in response._hidden_params
     assert isinstance(response._hidden_params["response_cost"], float)
+    assert isinstance(
+        response._hidden_params["llm_provider-x-ratelimit-remaining-requests"], float
+    )
 
 
 def test_vertex_ai_gemini_predict_cost():

litellm/utils.py

@@ -5678,6 +5678,14 @@ def convert_to_model_response_object(
     _response_headers: Optional[dict] = None,
 ):
     received_args = locals()
+    if _response_headers is not None:
+        if hidden_params is not None:
+            hidden_params["additional_headers"] = {
+                "{}-{}".format("llm_provider", k): v
+                for k, v in _response_headers.items()
+            }
+        else:
+            hidden_params = {"additional_headers": _response_headers}
     ### CHECK IF ERROR IN RESPONSE ### - openrouter returns these in the dictionary
     if (
         response_object is not None