diff --git a/litellm/llms/custom_httpx/sse_lines.py b/litellm/llms/custom_httpx/sse_lines.py new file mode 100644 index 00000000000..2a31f49e162 --- /dev/null +++ b/litellm/llms/custom_httpx/sse_lines.py @@ -0,0 +1,56 @@ +""" +Split a decoded Server-Sent Events body into lines the way the SSE spec does. + +httpx's `iter_lines()` / `aiter_lines()` follow `str.splitlines()`, which also breaks on U+0085, U+2028, +U+2029, \\x0b, \\x0c and \\x1c-\\x1e. JSON leaves U+0085, U+2028 and U+2029 unescaped inside strings, so a +`data:` line carrying one of them gets cut in half. SSE lines end only at CRLF, LF or CR +(https://html.spec.whatwg.org/multipage/server-sent-events.html#parsing-an-event-stream) +""" + +import io +import re +from collections.abc import AsyncIterable, AsyncIterator, Iterable, Iterator +from typing import Final + +_SSE_LINE_BREAK: Final = re.compile(r"\r\n|\r|\n") + + +class _SSELineSplitter: + def __init__(self) -> None: + self._partial_line: io.StringIO = io.StringIO() + self._after_cr: bool = False + + def feed(self, text: str) -> tuple[str, ...]: + if not text: + return () + fresh: Final = text[1:] if self._after_cr and text.startswith("\n") else text + self._after_cr = fresh.endswith("\r") + *ended, tail = _SSE_LINE_BREAK.split(fresh) + if not ended: + _ = self._partial_line.write(tail) + return () + first: Final = self._partial_line.getvalue() + ended[0] + self._partial_line = io.StringIO() + _ = self._partial_line.write(tail) + return (first, *ended[1:]) + + def flush(self) -> tuple[str, ...]: + rest: Final = self._partial_line.getvalue() + self._partial_line = io.StringIO() + return (rest,) if rest else () + + +def iter_sse_lines(text_chunks: Iterable[str]) -> Iterator[str]: + splitter: Final = _SSELineSplitter() + for text in text_chunks: + yield from splitter.feed(text) + yield from splitter.flush() + + +async def aiter_sse_lines(text_chunks: AsyncIterable[str]) -> AsyncIterator[str]: + splitter: Final = _SSELineSplitter() + async for text in text_chunks: + for line in splitter.feed(text): + yield line + for line in splitter.flush(): + yield line diff --git a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py index 9acbc9c459c..a06a380658d 100644 --- a/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py +++ b/litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py @@ -41,6 +41,7 @@ from litellm.llms.custom_httpx.http_handler import ( get_async_httpx_client, get_httpx_client, ) +from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines from litellm.types.llms.anthropic import AnthropicThinkingParam from litellm.types.llms.base import CachedTokensDetails from litellm.types.llms.gemini import BidiGenerateContentServerMessage @@ -3241,7 +3242,7 @@ async def make_call( raise completion_stream: Final = ModelResponseIterator( - streaming_response=response.aiter_lines(), + streaming_response=aiter_sse_lines(response.aiter_text()), sync_stream=False, logging_obj=logging_obj, response_headers=response.headers, @@ -3312,7 +3313,7 @@ def make_sync_call( raise completion_stream: Final = ModelResponseIterator( - streaming_response=response.iter_lines(), + streaming_response=iter_sse_lines(response.iter_text()), sync_stream=True, logging_obj=logging_obj, response_headers=response.headers, @@ -4190,7 +4191,7 @@ class ModelResponseIterator: log_failed_hop(e) return self._truncated_stream_end(token) self._start_hop(hop_response, token) - self.streaming_response = hop_response.aiter_lines() + self.streaming_response = aiter_sse_lines(hop_response.aiter_text()) self.async_response_iterator = self.streaming_response return None @@ -4204,7 +4205,7 @@ class ModelResponseIterator: log_failed_hop(e) return self._truncated_stream_end(token) self._start_hop(hop_response, token) - self.streaming_response = hop_response.iter_lines() + self.streaming_response = iter_sse_lines(hop_response.iter_text()) self.response_iterator = self.streaming_response return None diff --git a/tests/unit/litellm_core_utils/test_streaming_handler.py b/tests/unit/litellm_core_utils/test_streaming_handler.py index cc1fa1b8803..200ebb0a63c 100644 --- a/tests/unit/litellm_core_utils/test_streaming_handler.py +++ b/tests/unit/litellm_core_utils/test_streaming_handler.py @@ -7704,7 +7704,7 @@ def gemini_mock_post_streaming(url, **kwargs): mock_response = MagicMock() mock_response.status_code = 200 mock_response.headers = {"Content-Type": "text/event-stream"} - mock_response.iter_lines = MagicMock(return_value=stream_response()) + mock_response.iter_text = MagicMock(return_value=stream_response()) return mock_response diff --git a/tests/unit/llms/custom_httpx/test_sse_lines.py b/tests/unit/llms/custom_httpx/test_sse_lines.py new file mode 100644 index 00000000000..10a9f9a251f --- /dev/null +++ b/tests/unit/llms/custom_httpx/test_sse_lines.py @@ -0,0 +1,38 @@ +from collections.abc import AsyncIterator, Sequence +from typing import Final + +import pytest + +from litellm.llms.custom_httpx.sse_lines import aiter_sse_lines, iter_sse_lines + +CASES: Final = ( + pytest.param(("data: a
b\r\n\r\n",), ("data: a
b", ""), id="u2028-stays-inside-the-line"), + pytest.param( + ("data: \u0085
\x0b\x0c\x1c\x1d\x1e\n",), + ("data: \u0085
\x0b\x0c\x1c\x1d\x1e",), + id="no-splitlines-only-separator-breaks-a-line", + ), + pytest.param(("a\r\nb\nc\rd\n",), ("a", "b", "c", "d"), id="crlf-lf-and-cr-each-end-a-line"), + pytest.param(("a\r", "\nb\n"), ("a", "b"), id="crlf-split-across-chunks-is-one-break"), + pytest.param(("a\r", "\n\r\n"), ("a", ""), id="crlf-split-then-blank-line"), + pytest.param(("a\r", "", "\nb\n"), ("a", "b"), id="empty-chunk-between-cr-and-lf"), + pytest.param(("a\r", "\r"), ("a", ""), id="cr-then-cr-is-two-breaks"), + pytest.param(("da", "ta: x", "y\n"), ("data: xy",), id="line-spanning-chunks-is-joined"), + pytest.param(("data: x",), ("data: x",), id="unterminated-last-line-is-kept"), + pytest.param(("",), (), id="empty-body"), +) + + +async def _async_chunks(chunks: Sequence[str]) -> AsyncIterator[str]: + for chunk in chunks: + yield chunk + + +@pytest.mark.parametrize(("chunks", "expected"), CASES) +def test_iter_sse_lines_splits_only_on_cr_lf_and_crlf(chunks: Sequence[str], expected: Sequence[str]) -> None: + assert tuple(iter_sse_lines(chunks)) == tuple(expected) + + +@pytest.mark.parametrize(("chunks", "expected"), CASES) +async def test_aiter_sse_lines_splits_only_on_cr_lf_and_crlf(chunks: Sequence[str], expected: Sequence[str]) -> None: + assert tuple([line async for line in aiter_sse_lines(_async_chunks(chunks))]) == tuple(expected) diff --git a/tests/unit/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py b/tests/unit/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py index 3bc864cc047..f8671ddbbd0 100644 --- a/tests/unit/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py +++ b/tests/unit/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py @@ -3,8 +3,9 @@ import json import logging import re import threading -from collections.abc import Awaitable, Callable, Iterator, Mapping +from collections.abc import AsyncIterator, Awaitable, Callable, Iterator, Mapping from copy import deepcopy +from itertools import pairwise from typing import Final, List, Literal, cast, get_args from unittest.mock import MagicMock, Mock, patch @@ -10091,7 +10092,7 @@ def test_vertex_ai_streaming_response_id(): mock_response = MagicMock() mock_response.status_code = 200 - mock_response.iter_lines = MagicMock(return_value=stream_response()) + mock_response.iter_text = MagicMock(return_value=stream_response()) return mock_response logging_obj = MagicMock() @@ -10654,3 +10655,60 @@ async def test_gemini_sends_a_boolean_for_a_gateway_default_written_as_text_or_n sent: Final = await _sent_thinking_config("gemini-3.8-flash", {"reasoning_effort": "high"}) assert sent["includeThoughts"] is expected + + +class _BodyInChunks(httpx.SyncByteStream, httpx.AsyncByteStream): + def __init__(self, chunks: tuple[bytes, ...]) -> None: + self._chunks: Final = chunks + + def __iter__(self) -> Iterator[bytes]: + yield from self._chunks + + async def __aiter__(self) -> AsyncIterator[bytes]: + for chunk in self._chunks: + yield chunk + + +def _split_at(body: bytes, cuts: tuple[int, ...]) -> tuple[bytes, ...]: + return tuple(body[start:end] for start, end in pairwise((0, *sorted(cuts), len(body)))) + + +async def _gemini_stream(is_async: bool, chunks: tuple[bytes, ...]) -> list: + def handler(request: httpx.Request) -> httpx.Response: + return httpx.Response(200, stream=_BodyInChunks(chunks), headers={"content-type": "text/event-stream"}) + + kwargs: Final = { + "model": "gemini/gemini-2.5-flash", + "messages": [{"role": "user", "content": "hi"}], + "stream": True, + "stream_options": {"include_usage": True}, + "api_key": "fake-gemini-key", + } + if is_async: + stream: Final = await litellm.acompletion( + **kwargs, client=AsyncHTTPHandler(transport=httpx.MockTransport(handler)) + ) + return [chunk async for chunk in stream] + return list( + litellm.completion(**kwargs, client=HTTPHandler(client=httpx.Client(transport=httpx.MockTransport(handler)))) + ) + + +@pytest.mark.parametrize("is_async", [True, False]) +async def test_gemini_stream_keeps_text_and_usage_when_a_part_contains_a_unicode_line_separator(is_async: bool): + events: Final = ( + {"candidates": [{"content": {"role": "model", "parts": [{"text": "line one\u2028line two"}]}}]}, + { + "candidates": [{"content": {"role": "model", "parts": [{"text": " and three"}]}, "finishReason": "STOP"}], + "usageMetadata": {"promptTokenCount": 10, "candidatesTokenCount": 8, "totalTokenCount": 18}, + }, + ) + body: Final = "".join(f"data: {json.dumps(event, ensure_ascii=False)}\r\n\r\n" for event in events).encode() + cut_inside_separator_and_crlf: Final = (body.index(b"\xe2\x80\xa8") + 1, body.index(b"\r") + 1) + + received: Final = await _gemini_stream(is_async, _split_at(body, cut_inside_separator_and_crlf)) + + usage: Final = next(chunk.usage for chunk in reversed(received) if getattr(chunk, "usage", None)) + text: Final = "".join(c.choices[0].delta.content or "" for c in received if c.choices) + assert text == "line one\u2028line two and three" + assert (usage.prompt_tokens, usage.completion_tokens) == (10, 8) diff --git a/tests/unit/llms/vertex_ai/test_http_status_201.py b/tests/unit/llms/vertex_ai/test_http_status_201.py index 34502d5cebe..0ed469acfa7 100644 --- a/tests/unit/llms/vertex_ai/test_http_status_201.py +++ b/tests/unit/llms/vertex_ai/test_http_status_201.py @@ -39,8 +39,8 @@ class TestVertexAIHTTPStatus201(unittest.TestCase): # Create a mock response with status code 201 mock_response = MagicMock() mock_response.status_code = 201 - mock_response.aiter_lines = MagicMock() - mock_response.aiter_lines.return_value = ["test response"] + mock_response.aiter_text = MagicMock() + mock_response.aiter_text.return_value = ["test response"] # Setup mock client mock_client = MagicMock() @@ -76,8 +76,8 @@ class TestVertexAIHTTPStatus201(unittest.TestCase): # Create a mock response with status code 201 mock_response = MagicMock() mock_response.status_code = 201 - mock_response.iter_lines = MagicMock() - mock_response.iter_lines.return_value = ["test response"] + mock_response.iter_text = MagicMock() + mock_response.iter_text.return_value = ["test response"] mock_post.return_value = mock_response # Call the make_sync_call function