Skip to content

The agent engineering platform.

fix(fireworks): avoid repeated `service_tier` [closes #39619] (#39620)

ccurme7fb045b744ca
diff --git a/libs/partners/fireworks/langchain_fireworks/chat_models.py b/libs/partners/fireworks/langchain_fireworks/chat_models.py
index fd5fd07957..42b8f15513 100644
--- a/libs/partners/fireworks/langchain_fireworks/chat_models.py
+++ b/libs/partners/fireworks/langchain_fireworks/chat_models.py
@@ -471,8 +471,6 @@ def _convert_chunk_to_message_chunk(
) -> BaseMessageChunk:
choices = chunk.get("choices") or []
response_metadata: dict[str, Any] = {"model_provider": "fireworks"}
- if service_tier := chunk.get("service_tier"):
- response_metadata["service_tier"] = service_tier
if not choices:
# Final chunk emitted when `stream_options.include_usage=True`:
# `choices` is empty and the chunk carries only `usage`.
@@ -1103,6 +1101,8 @@ class ChatFireworks(BaseChatModel):
if finish_reason := choice.get("finish_reason"):
generation_info["finish_reason"] = finish_reason
generation_info["model_name"] = self.model_name
+ if service_tier := chunk.get("service_tier"):
+ generation_info["service_tier"] = service_tier
logprobs = choice.get("logprobs")
if logprobs:
generation_info["logprobs"] = logprobs
@@ -1214,6 +1214,8 @@ class ChatFireworks(BaseChatModel):
if finish_reason := choice.get("finish_reason"):
generation_info["finish_reason"] = finish_reason
generation_info["model_name"] = self.model_name
+ if service_tier := chunk.get("service_tier"):
+ generation_info["service_tier"] = service_tier
logprobs = choice.get("logprobs")
if logprobs:
generation_info["logprobs"] = logprobs
diff --git a/libs/partners/fireworks/tests/unit_tests/test_chat_models.py b/libs/partners/fireworks/tests/unit_tests/test_chat_models.py
index 87088b5dcf..35263d1407 100644
--- a/libs/partners/fireworks/tests/unit_tests/test_chat_models.py
+++ b/libs/partners/fireworks/tests/unit_tests/test_chat_models.py
@@ -1668,7 +1668,7 @@ class TestServiceTier:
assert isinstance(result, AIMessage)
assert result.response_metadata["service_tier"] == "priority"
- def test_service_tier_echoed_in_stream_chunks(self) -> None:
+ def test_service_tier_echoed_once_in_stream_chunks(self) -> None:
model = _make_model(service_tier="priority")
model.client = MagicMock()
chunks: list[dict[str, Any]] = [
@@ -1676,6 +1676,10 @@ class TestServiceTier:
"choices": [{"delta": {"role": "assistant", "content": "hi"}}],
"service_tier": "priority",
},
+ {
+ "choices": [{"delta": {}, "finish_reason": "stop"}],
+ "service_tier": "priority",
+ },
{
"choices": [],
"usage": {
@@ -1687,10 +1691,42 @@ class TestServiceTier:
},
]
model.client.create.return_value = iter(chunks)
- out = list(model.stream("Hello"))
- tagged = [c for c in out if c.response_metadata.get("service_tier")]
- assert tagged
- assert all(c.response_metadata["service_tier"] == "priority" for c in tagged)
+ output = list(model.stream("Hello"))
+ tagged = [c for c in output if c.response_metadata.get("service_tier")]
+ assert len(tagged) == 1
+ assert tagged[0].response_metadata["service_tier"] == "priority"
+
+ combined = output[0]
+ for chunk in output[1:]:
+ combined += chunk
+ assert combined.response_metadata["service_tier"] == "priority"
+
+ async def test_service_tier_echoed_once_in_async_stream_chunks(self) -> None:
+ model = _make_model(service_tier="priority")
+ model.async_client = MagicMock()
+ chunks: list[dict[str, Any]] = [
+ {
+ "choices": [{"delta": {"role": "assistant", "content": "hi"}}],
+ "service_tier": "priority",
+ },
+ {
+ "choices": [{"delta": {}, "finish_reason": "stop"}],
+ "service_tier": "priority",
+ },
+ ]
+
+ async def _aiter() -> Any:
+ for chunk in chunks:
+ yield chunk
+
+ async def _create(**_kwargs: Any) -> Any:
+ return _aiter()
+
+ model.async_client.create = MagicMock(side_effect=_create)
+ output = [chunk async for chunk in model.astream("Hello")]
+ tagged = [c for c in output if c.response_metadata.get("service_tier")]
+ assert len(tagged) == 1
+ assert tagged[0].response_metadata["service_tier"] == "priority"
def test_service_tier_absent_when_not_in_response(self) -> None:
model = _make_model()