scriptoriumadmin/litellm
1
0
Fork 0
mirror of https://github.com/BerriAI/litellm.git synced 2026-09-09 22:31:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 86

fix(fireworks_ai): fold instructions and developer items into one leading system message on the Responses path

Browse source
This commit is contained in:
mateo-berri 2026-09-08 11:08:03 -07:00
parent 82e6b84f5a
commit deaadc21d3
2 changed files with 100 additions and 14 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

44
litellm/llms/fireworks_ai/responses/transformation.py
View file

@ -5,6 +5,7 @@ from urllib.parse import unquote
import httpx
from openai.types.responses import EasyInputMessageParam, ResponseInputItemParam
from pydantic import TypeAdapter
from litellm.llms.fireworks_ai.common_utils import (
resolve_fireworks_api_key,
@ -31,16 +32,32 @@ def _session_params(litellm_params: GenericLiteLLMParams) -> Mapping[str, object
)
def _developer_item_as_system(item: ResponseInputItemParam) -> ResponseInputItemParam:
if "role" not in item or item["role"] != "developer":
return item
return EasyInputMessageParam(role="system", content=item["content"], type="message")
_instructions_adapter: Final = TypeAdapter[str | None](str | None)
def _developer_items_as_system(input: str | ResponseInputParam) -> str | ResponseInputParam:
if isinstance(input, str):
def _instruction_text(item: ResponseInputItemParam) -> str | None:
if "role" not in item or (item["role"] != "system" and item["role"] != "developer"):
return None
content: Final = item["content"]
if isinstance(content, str):
return content
return "\n\n".join(part["text"] for part in content if part["type"] == "input_text")
def _with_single_leading_system_item(
input: str | ResponseInputParam, instructions: str | None
) -> str | ResponseInputParam:
items: Final = () if isinstance(input, str) else tuple(input)
instruction_texts: Final = tuple(text for text in (instructions, *map(_instruction_text, items)) if text)
if not instruction_texts:
return input
return [_developer_item_as_system(item) for item in input]
leading: Final = EasyInputMessageParam(role="system", content="\n\n".join(instruction_texts), type="message")
rest: Final = (
(EasyInputMessageParam(role="user", content=input),)
if isinstance(input, str)
else tuple(item for item in items if _instruction_text(item) is None)
)
return [leading, *rest]
class FireworksAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
@ -68,9 +85,6 @@ class FireworksAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
base: Final = (api_base or get_secret_str("FIREWORKS_API_BASE") or FIREWORKS_AI_DEFAULT_API_BASE).rstrip("/")
return f"{base}/responses"
def _validate_input_param(self, input: str | ResponseInputParam) -> str | ResponseInputParam:
return _developer_items_as_system(super()._validate_input_param(input))
def transform_responses_api_request(
self,
model: str,
@ -79,10 +93,16 @@ class FireworksAIResponsesAPIConfig(OpenAIResponsesAPIConfig):
litellm_params: GenericLiteLLMParams,
headers: dict, # mutable-ok: overrides the base class signature
) -> dict: # mutable-ok: overrides the base class signature
instructions: Final = _instructions_adapter.validate_python(
response_api_optional_request_params.get("instructions")
)
folded_params: Final = { # mutable-ok: the base class takes the optional params as a dict
key: value for key, value in response_api_optional_request_params.items() if key != "instructions"
}
return super().transform_responses_api_request(
model=resolve_fireworks_resource_name(model),
input=input,
response_api_optional_request_params=response_api_optional_request_params,
input=_with_single_leading_system_item(self._validate_input_param(input), instructions),
response_api_optional_request_params=folded_params,
litellm_params=litellm_params,
headers=headers,
)

70
tests/test_litellm/llms/fireworks_ai/responses/test_fireworks_ai_responses_transformation.py
View file

@ -162,7 +162,7 @@ def test_responses_call_forwards_previous_response_id_and_store() -> None:
assert body["input"][0]["call_id"] == "call_abc123"
def test_responses_call_sends_developer_items_as_system_messages() -> None:
def test_responses_call_hoists_developer_items_into_one_leading_system_message() -> None:
client: Final = _mock_http_client(_fireworks_response("accounts/fireworks/models/kimi-k3"))
with patch(HTTPX_CLIENT_FACTORY, return_value=client):
litellm.responses(
@ -176,12 +176,78 @@ def test_responses_call_sends_developer_items_as_system_messages() -> None:
)
_, _, body = _sent_request(client)
assert tuple(body["input"]) == (
{"role": "user", "content": "Hi there"},
{"role": "system", "content": "Answer with exactly one word.", "type": "message"},
{"role": "user", "content": "Hi there"},
{"role": "user", "content": [{"type": "input_text", "text": "What is the capital of France?"}]},
)
def test_responses_call_folds_instructions_and_developer_item_into_one_leading_system_message() -> None:
client: Final = _mock_http_client(_fireworks_response("accounts/fireworks/models/qwen3p8-2p4t-a95b"))
with patch(HTTPX_CLIENT_FACTORY, return_value=client):
litellm.responses(
model="fireworks_ai/accounts/fireworks/models/qwen3p8-2p4t-a95b",
instructions="You are a coding agent running in the Codex CLI.",
input=[ # mutable-ok: the Responses API takes input as a JSON list
{
"role": "developer",
"content": [{"type": "input_text", "text": "<permissions instructions>read-only</permissions instructions>"}],
},
{"role": "user", "content": [{"type": "input_text", "text": "What is the capital of France?"}]},
{"id": "rs_1", "type": "reasoning", "summary": [{"type": "summary_text", "text": "A trivial question."}]},
{
"id": "msg_1",
"type": "message",
"role": "assistant",
"status": "completed",
"content": [{"type": "output_text", "text": "Paris is the capital of France.", "annotations": []}],
},
{"role": "user", "content": [{"type": "input_text", "text": "And of Spain?"}]},
],
store=False,
api_key="fw-test-key",
)
_, _, body = _sent_request(client)
assert "instructions" not in body
assert tuple(body["input"]) == (
{
"role": "system",
"content": (
"You are a coding agent running in the Codex CLI.\n\n"
"<permissions instructions>read-only</permissions instructions>"
),
"type": "message",
},
{"role": "user", "content": [{"type": "input_text", "text": "What is the capital of France?"}]},
{"id": "rs_1", "type": "reasoning", "summary": [{"type": "summary_text", "text": "A trivial question."}]},
{
"id": "msg_1",
"type": "message",
"role": "assistant",
"status": "completed",
"content": [{"type": "output_text", "text": "Paris is the capital of France.", "annotations": []}],
},
{"role": "user", "content": [{"type": "input_text", "text": "And of Spain?"}]},
)
def test_responses_call_turns_string_input_with_instructions_into_system_then_user_messages() -> None:
client: Final = _mock_http_client(_fireworks_response("accounts/fireworks/models/kimi-k3"))
with patch(HTTPX_CLIENT_FACTORY, return_value=client):
litellm.responses(
model="fireworks_ai/accounts/fireworks/models/kimi-k3",
instructions="Answer with exactly one word.",
input="What is the capital of France?",
api_key="fw-test-key",
)
_, _, body = _sent_request(client)
assert "instructions" not in body
assert tuple(body["input"]) == (
{"role": "system", "content": "Answer with exactly one word.", "type": "message"},
{"role": "user", "content": "What is the capital of France?"},
)
def test_responses_call_maps_pydantic_developer_items_and_replays_pydantic_output_items() -> None:
client: Final = _mock_http_client(_fireworks_response("accounts/fireworks/models/kimi-k3"))
pydantic_input: Final = cast(