feat(fireworks): native content-block streaming events

libs/partners/fireworks/langchain_fireworks/_stream_events.py

@@ -0,0 +1,160 @@
+"""Native content-block streaming-event converter for Fireworks.
+
+Builds text, reasoning, and tool-call blocks directly from Fireworks' raw
+OpenAI-shaped delta (Fireworks streams tool-call args incrementally), feeding
+the shared `BlockStreamTracker`. Unlike the compat bridge (which leaves
+Fireworks' `tool_calls`/`reasoning_content` in `additional_kwargs`), this
+surfaces them as first-class blocks.
+
+Reasoning delta key is ``reasoning_content`` (not ``reasoning``).
+"""
+
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Any
+
+from langchain_core.language_models.stream_events import (
+    BlockStreamTracker,
+    accumulate_usage,
+    build_message_finish,
+)
+
+from langchain_fireworks.chat_models import _usage_to_metadata
+
+if TYPE_CHECKING:
+    from collections.abc import AsyncIterator, Iterator
+
+    from langchain_protocol.protocol import (
+        MessageMetadata,
+        MessagesData,
+        MessageStartData,
+    )
+
+
+def _message_start(message_id: str | None, model: str | None) -> MessageStartData:
+    # Do not use a provider completion id here: on the v3 path core seeds the
+    # stream with the LangChain run id, and an empty id lets that stand
+    # (matching the compat bridge). Only an explicit `message_id` wins.
+    metadata: MessageMetadata = {"provider": "fireworks"}
+    if model:
+        metadata["model"] = model
+    return {
+        "event": "message-start",
+        "role": "ai",
+        "id": message_id or "",
+        "metadata": metadata,
+    }
+
+
+def _feed_delta(
+    tracker: BlockStreamTracker, delta: dict[str, Any]
+) -> Iterator[MessagesData]:
+    """Yield block events for one OpenAI-shaped Fireworks delta."""
+    if reasoning := delta.get("reasoning_content"):
+        yield from tracker.feed(
+            "reasoning", {"type": "reasoning", "reasoning": reasoning}
+        )
+    if content := delta.get("content"):
+        yield from tracker.feed("text", {"type": "text", "text": content})
+    # Deferred-as-scope-cut: function_call (legacy) and logprobs are not
+    # surfaced as blocks; the bridge already drops them.
+    for tc in delta.get("tool_calls") or []:
+        idx = tc.get("index", 0)
+        fn = tc.get("function") or {}
+        args = fn.get("arguments")
+        if args == "null":  # normalize JSON null for no-arg tools
+            args = "{}"
+        yield from tracker.feed(
+            f"tool:{idx}",
+            {
+                "type": "tool_call_chunk",
+                "id": tc.get("id"),
+                "name": fn.get("name"),
+                "args": args or "",
+                "index": idx,
+            },
+        )
+
+
+def convert_fireworks_stream(
+    raw: Iterator[Any], *, message_id: str | None = None
+) -> Iterator[MessagesData]:
+    """Convert a raw Fireworks chat stream to protocol events.
+
+    Args:
+        raw: Raw Fireworks chat-completion stream chunks (dicts or SDK objects).
+        message_id: Overrides the provider message id on `message-start`.
+
+    Yields:
+        Protocol `MessagesData` lifecycle events.
+    """
+    tracker = BlockStreamTracker()
+    started = False
+    usage: dict[str, Any] | None = None
+    response_metadata: dict[str, Any] = {"model_provider": "fireworks"}
+    model: str | None = None
+
+    for chunk in raw:
+        if not isinstance(chunk, dict):
+            chunk = chunk.model_dump()
+        if model is None:
+            model = chunk.get("model")
+        if service_tier := chunk.get("service_tier"):
+            response_metadata["service_tier"] = service_tier
+        # Usage arrives on the final no-choices chunk when
+        # `stream_options.include_usage=True`; also tolerate it on other chunks.
+        if fw_usage := chunk.get("usage"):
+            usage = accumulate_usage(usage, dict(_usage_to_metadata(fw_usage)))
+        choices = chunk.get("choices") or []
+        if len(choices) == 0:
+            continue
+        if not started:
+            started = True
+            yield _message_start(message_id, model)
+        choice = choices[0]
+        yield from _feed_delta(tracker, choice.get("delta") or {})
+        if finish_reason := choice.get("finish_reason"):
+            response_metadata["finish_reason"] = finish_reason
+
+    if not started:
+        return
+    yield from tracker.finish_all()
+    yield build_message_finish(usage=usage, response_metadata=response_metadata)
+
+
+async def aconvert_fireworks_stream(
+    raw: AsyncIterator[Any], *, message_id: str | None = None
+) -> AsyncIterator[MessagesData]:
+    """Async twin of `convert_fireworks_stream`."""
+    tracker = BlockStreamTracker()
+    started = False
+    usage: dict[str, Any] | None = None
+    response_metadata: dict[str, Any] = {"model_provider": "fireworks"}
+    model: str | None = None
+
+    async for chunk in raw:
+        if not isinstance(chunk, dict):
+            chunk = chunk.model_dump()
+        if model is None:
+            model = chunk.get("model")
+        if service_tier := chunk.get("service_tier"):
+            response_metadata["service_tier"] = service_tier
+        if fw_usage := chunk.get("usage"):
+            usage = accumulate_usage(usage, dict(_usage_to_metadata(fw_usage)))
+        choices = chunk.get("choices") or []
+        if len(choices) == 0:
+            continue
+        if not started:
+            started = True
+            yield _message_start(message_id, model)
+        choice = choices[0]
+        for ev in _feed_delta(tracker, choice.get("delta") or {}):
+            yield ev
+        if finish_reason := choice.get("finish_reason"):
+            response_metadata["finish_reason"] = finish_reason
+
+    if not started:
+        return
+    for ev in tracker.finish_all():
+        yield ev
+    yield build_message_finish(usage=usage, response_metadata=response_metadata)

libs/partners/fireworks/langchain_fireworks/chat_models.py

@@ -8,6 +8,7 @@ import logging
 from collections.abc import AsyncIterator, Callable, Iterator, Mapping, Sequence
 from operator import itemgetter
 from typing import (
+    TYPE_CHECKING,
     Any,
     Literal,
     NoReturn,
@@ -103,6 +104,9 @@ from langchain_fireworks._compat import _convert_from_v1_to_chat_completions
 from langchain_fireworks._version import __version__
 from langchain_fireworks.data._profiles import _PROFILES
 
+if TYPE_CHECKING:
+    from langchain_protocol.protocol import MessagesData
+
 logger = logging.getLogger(__name__)
 
 
@@ -1115,6 +1119,74 @@ class ChatFireworks(BaseChatModel):
                 )
             yield generation_chunk
 
+    def _stream_chat_model_events(
+        self,
+        messages: list[BaseMessage],
+        stop: list[str] | None = None,
+        run_manager: CallbackManagerForLLMRun | None = None,
+        *,
+        message_id: str | None = None,
+        **kwargs: Any,
+    ) -> Iterator[MessagesData]:
+        """Emit Fireworks-native content-block protocol events.
+
+        Detected by `langchain-core`'s `_iter_v2_events`; powers
+        `stream_events(version="v3")`. Falls through to the compat bridge only
+        if this method is absent. `message_id` is threaded from the stream so
+        `message-start` matches the bridge's LangChain run id.
+        """
+        from langchain_fireworks._stream_events import (
+            convert_fireworks_stream,
+        )
+
+        kwargs.pop("stream_usage", None)  # never leak into the Fireworks payload
+        message_dicts, params = self._create_message_dicts(messages, stop)
+        params = {**params, **kwargs, "stream": True}
+        if self.stream_usage and "stream_options" not in params:
+            params["stream_options"] = {"include_usage": True}
+        raw = _completion_with_retry(
+            self, run_manager=run_manager, messages=message_dicts, **params
+        )
+        for event in convert_fireworks_stream(raw, message_id=message_id):
+            if (
+                run_manager is not None
+                and event["event"] == "content-block-delta"
+                and event["delta"].get("type") == "text-delta"
+            ):
+                run_manager.on_llm_new_token(str(event["delta"].get("text", "")))
+            yield event
+
+    async def _astream_chat_model_events(
+        self,
+        messages: list[BaseMessage],
+        stop: list[str] | None = None,
+        run_manager: AsyncCallbackManagerForLLMRun | None = None,
+        *,
+        message_id: str | None = None,
+        **kwargs: Any,
+    ) -> AsyncIterator[MessagesData]:
+        """Async twin of `_stream_chat_model_events`."""
+        from langchain_fireworks._stream_events import (
+            aconvert_fireworks_stream,
+        )
+
+        kwargs.pop("stream_usage", None)  # never leak into the Fireworks payload
+        message_dicts, params = self._create_message_dicts(messages, stop)
+        params = {**params, **kwargs, "stream": True}
+        if self.stream_usage and "stream_options" not in params:
+            params["stream_options"] = {"include_usage": True}
+        raw = await _acompletion_with_retry(
+            self, run_manager=run_manager, messages=message_dicts, **params
+        )
+        async for event in aconvert_fireworks_stream(raw, message_id=message_id):
+            if (
+                run_manager is not None
+                and event["event"] == "content-block-delta"
+                and event["delta"].get("type") == "text-delta"
+            ):
+                await run_manager.on_llm_new_token(str(event["delta"].get("text", "")))
+            yield event
+
     async def _agenerate(
         self,
         messages: list[BaseMessage],

libs/partners/fireworks/tests/unit_tests/test_stream_events.py

@@ -0,0 +1,317 @@
+"""Unit tests for the Fireworks native stream-events converter."""
+
+import os
+from typing import Any, cast
+from unittest.mock import MagicMock, patch
+
+from langchain_tests.utils.stream_lifecycle import assert_valid_event_stream
+
+from langchain_fireworks import ChatFireworks
+from langchain_fireworks._stream_events import (
+    aconvert_fireworks_stream,
+    convert_fireworks_stream,
+)
+
+if "FIREWORKS_API_KEY" not in os.environ:
+    os.environ["FIREWORKS_API_KEY"] = "fake-key"
+
+
+def _reasoning_text_tool() -> list[dict]:
+    m = "accounts/fireworks/models/deepseek-r1"
+    return [
+        {
+            "model": m,
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {"role": "assistant", "reasoning_content": "Let me "},
+                    "finish_reason": None,
+                }
+            ],
+        },
+        {
+            "model": m,
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {"reasoning_content": "think."},
+                    "finish_reason": None,
+                }
+            ],
+        },
+        {
+            "model": m,
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {"content": "Hi "},
+                    "finish_reason": None,
+                }
+            ],
+        },
+        {
+            "model": m,
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {"content": "there"},
+                    "finish_reason": None,
+                }
+            ],
+        },
+        {
+            "model": m,
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {
+                        "tool_calls": [
+                            {
+                                "index": 0,
+                                "id": "t1",
+                                "function": {
+                                    "name": "get_weather",
+                                    "arguments": '{"city":',
+                                },
+                            }
+                        ]
+                    },
+                    "finish_reason": None,
+                }
+            ],
+        },
+        {
+            "model": m,
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {
+                        "tool_calls": [
+                            {"index": 0, "function": {"arguments": ' "Paris"}'}}
+                        ]
+                    },
+                    "finish_reason": "tool_calls",
+                }
+            ],
+        },
+        # Final no-choices chunk carrying usage (Fireworks, not under x_groq)
+        {
+            "model": m,
+            "choices": [],
+            "usage": {
+                "prompt_tokens": 10,
+                "completion_tokens": 5,
+                "total_tokens": 15,
+            },
+        },
+    ]
+
+
+def test_convert_fireworks_stream_lifecycle() -> None:
+    events: list[Any] = list(convert_fireworks_stream(iter(_reasoning_text_tool())))
+    assert_valid_event_stream(events)
+    assert events[0]["event"] == "message-start"
+    assert events[0]["id"] == ""
+    assert events[0]["metadata"]["provider"] == "fireworks"
+    assert events[0]["metadata"]["model"] == "accounts/fireworks/models/deepseek-r1"
+
+    finishes = [e for e in events if e["event"] == "content-block-finish"]
+    types = [f["content"]["type"] for f in finishes]
+    assert types == ["reasoning", "text", "tool_call"]
+    assert (
+        cast("dict[str, Any]", finishes[0]["content"])["reasoning"] == "Let me think."
+    )
+    assert cast("dict[str, Any]", finishes[1]["content"])["text"] == "Hi there"
+    tool = cast("dict[str, Any]", finishes[2]["content"])
+    assert tool["name"] == "get_weather"
+    assert tool["args"] == {"city": "Paris"}
+
+    message_finish = events[-1]
+    assert message_finish["event"] == "message-finish"
+    assert message_finish["usage"] == {
+        "input_tokens": 10,
+        "output_tokens": 5,
+        "total_tokens": 15,
+    }
+
+
+def test_fireworks_stream_events_v3_lifecycle() -> None:
+    """Drive `stream_events(version="v3")` through the native sync hook.
+
+    Confirms the model-level path threads the LangChain run id onto
+    `message-start` (non-empty, unlike the converter's empty default) and
+    surfaces reasoning/text/tool blocks with `model_provider == "fireworks"`.
+    """
+    llm = ChatFireworks(model="accounts/fireworks/models/deepseek-r1")
+    mock_client = MagicMock()
+
+    def mock_create(*_a: Any, **_k: Any) -> Any:
+        return iter(_reasoning_text_tool())
+
+    mock_client.create = mock_create
+
+    with patch.object(llm, "client", mock_client):
+        events = list(llm.stream_events("Test query", version="v3"))
+
+    assert_valid_event_stream(events)
+
+    message_start = cast("dict[str, Any]", events[0])
+    assert message_start["event"] == "message-start"
+    assert message_start["id"]  # core seeds a non-empty LangChain run id
+    assert message_start["metadata"]["provider"] == "fireworks"
+
+    finishes = [e for e in events if e["event"] == "content-block-finish"]
+    assert [f["content"]["type"] for f in finishes] == [
+        "reasoning",
+        "text",
+        "tool_call",
+    ]
+    tool = cast("dict[str, Any]", finishes[2]["content"])
+    assert tool["name"] == "get_weather"
+    assert tool["args"] == {"city": "Paris"}
+
+    message_finish = cast("dict[str, Any]", events[-1])
+    assert message_finish["event"] == "message-finish"
+    assert message_finish["metadata"]["model_provider"] == "fireworks"
+
+
+async def test_fireworks_astream_events_v3_lifecycle() -> None:
+    """Async twin of `test_fireworks_stream_events_v3_lifecycle`."""
+    llm = ChatFireworks(model="accounts/fireworks/models/deepseek-r1")
+
+    async def _araw() -> Any:
+        for chunk in _reasoning_text_tool():
+            yield chunk
+
+    # `_acompletion_with_retry` does `await llm.async_client.create(**kwargs)`
+    # (1.x SDK: `create()` is a coroutine resolving to an `AsyncStream`), so
+    # the mock must be a coroutine returning an async iterator.
+    async def mock_create(*a: Any, **k: Any) -> Any:
+        return _araw()
+
+    mock_client = MagicMock()
+    mock_client.create = mock_create
+
+    with patch.object(llm, "async_client", mock_client):
+        stream = await llm.astream_events("Test query", version="v3")
+        events = [e async for e in stream]
+
+    assert_valid_event_stream(events)
+
+    message_start = cast("dict[str, Any]", events[0])
+    assert message_start["event"] == "message-start"
+    assert message_start["id"]
+    assert message_start["metadata"]["provider"] == "fireworks"
+
+    finishes = [e for e in events if e["event"] == "content-block-finish"]
+    assert [f["content"]["type"] for f in finishes] == [
+        "reasoning",
+        "text",
+        "tool_call",
+    ]
+    message_finish = cast("dict[str, Any]", events[-1])
+    assert message_finish["metadata"]["model_provider"] == "fireworks"
+
+
+async def test_aconvert_fireworks_stream_lifecycle() -> None:
+    async def _araw() -> Any:
+        for chunk in _reasoning_text_tool():
+            yield chunk
+
+    events: list[Any] = [e async for e in aconvert_fireworks_stream(_araw())]
+    assert_valid_event_stream(events)
+    assert events[0]["event"] == "message-start"
+    assert events[0]["id"] == ""
+    assert events[0]["metadata"]["provider"] == "fireworks"
+
+    finishes = [e for e in events if e["event"] == "content-block-finish"]
+    types = [f["content"]["type"] for f in finishes]
+    assert types == ["reasoning", "text", "tool_call"]
+    assert (
+        cast("dict[str, Any]", finishes[0]["content"])["reasoning"] == "Let me think."
+    )
+    assert cast("dict[str, Any]", finishes[1]["content"])["text"] == "Hi there"
+    tool = cast("dict[str, Any]", finishes[2]["content"])
+    assert tool["name"] == "get_weather"
+    assert tool["args"] == {"city": "Paris"}
+
+    message_finish = events[-1]
+    assert message_finish["event"] == "message-finish"
+    assert message_finish["usage"] == {
+        "input_tokens": 10,
+        "output_tokens": 5,
+        "total_tokens": 15,
+    }
+
+
+def _parallel_tool_calls() -> list[dict]:
+    """Two tool calls interleaved at index 0 and 1 across chunks."""
+    m = "accounts/fireworks/models/deepseek-r1"
+    return [
+        {
+            "model": m,
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {
+                        "role": "assistant",
+                        "tool_calls": [
+                            {
+                                "index": 0,
+                                "id": "call_a",
+                                "function": {
+                                    "name": "get_weather",
+                                    "arguments": '{"city":',
+                                },
+                            },
+                            {
+                                "index": 1,
+                                "id": "call_b",
+                                "function": {
+                                    "name": "get_time",
+                                    "arguments": '{"zone":',
+                                },
+                            },
+                        ],
+                    },
+                    "finish_reason": None,
+                }
+            ],
+        },
+        {
+            "model": m,
+            "choices": [
+                {
+                    "index": 0,
+                    "delta": {
+                        "tool_calls": [
+                            {"index": 0, "function": {"arguments": ' "Paris"}'}},
+                            {"index": 1, "function": {"arguments": ' "UTC"}'}},
+                        ]
+                    },
+                    "finish_reason": "tool_calls",
+                }
+            ],
+        },
+    ]
+
+
+def test_convert_fireworks_stream_parallel_tool_calls() -> None:
+    """Args at distinct `tool:{idx}` keys finalize as separate tool calls."""
+    events: list[Any] = list(convert_fireworks_stream(iter(_parallel_tool_calls())))
+    assert_valid_event_stream(events)
+
+    finishes = [e for e in events if e["event"] == "content-block-finish"]
+    types = [f["content"]["type"] for f in finishes]
+    assert types == ["tool_call", "tool_call"]
+
+    first = cast("dict[str, Any]", finishes[0]["content"])
+    second = cast("dict[str, Any]", finishes[1]["content"])
+    assert first["name"] == "get_weather"
+    assert first["args"] == {"city": "Paris"}
+    assert second["name"] == "get_time"
+    assert second["args"] == {"zone": "UTC"}
+
+    # No usage chunk in this fixture: message-finish should carry no usage.
+    assert events[-1]["event"] == "message-finish"
+    assert events[-1].get("usage") is None