fix: close stream on cancellation for SambaNova and Google OpenAI services
This commit is contained in:
Luke Payyapilli
@@ -91,52 +91,55 @@ class GoogleLLMOpenAIBetaService(OpenAILLMService):
|
||||
ChatCompletionChunk
|
||||
] = await self._stream_chat_completions_specific_context(context)
|
||||
|
||||
async for chunk in chunk_stream:
|
||||
if chunk.usage:
|
||||
tokens = LLMTokenUsage(
|
||||
prompt_tokens=chunk.usage.prompt_tokens or 0,
|
||||
completion_tokens=chunk.usage.completion_tokens or 0,
|
||||
total_tokens=chunk.usage.total_tokens or 0,
|
||||
)
|
||||
await self.start_llm_usage_metrics(tokens)
|
||||
# Use context manager to ensure stream is closed on cancellation/exception.
|
||||
# Without this, CancelledError during iteration leaves the underlying socket open.
|
||||
async with chunk_stream:
|
||||
async for chunk in chunk_stream:
|
||||
if chunk.usage:
|
||||
tokens = LLMTokenUsage(
|
||||
prompt_tokens=chunk.usage.prompt_tokens or 0,
|
||||
completion_tokens=chunk.usage.completion_tokens or 0,
|
||||
total_tokens=chunk.usage.total_tokens or 0,
|
||||
)
|
||||
await self.start_llm_usage_metrics(tokens)
|
||||
|
||||
if chunk.choices is None or len(chunk.choices) == 0:
|
||||
continue
|
||||
if chunk.choices is None or len(chunk.choices) == 0:
|
||||
continue
|
||||
|
||||
await self.stop_ttfb_metrics()
|
||||
await self.stop_ttfb_metrics()
|
||||
|
||||
if not chunk.choices[0].delta:
|
||||
continue
|
||||
if not chunk.choices[0].delta:
|
||||
continue
|
||||
|
||||
if chunk.choices[0].delta.tool_calls:
|
||||
# We're streaming the LLM response to enable the fastest response times.
|
||||
# For text, we just yield each chunk as we receive it and count on consumers
|
||||
# to do whatever coalescing they need (eg. to pass full sentences to TTS)
|
||||
#
|
||||
# If the LLM is a function call, we'll do some coalescing here.
|
||||
# If the response contains a function name, we'll yield a frame to tell consumers
|
||||
# that they can start preparing to call the function with that name.
|
||||
# We accumulate all the arguments for the rest of the streamed response, then when
|
||||
# the response is done, we package up all the arguments and the function name and
|
||||
# yield a frame containing the function name and the arguments.
|
||||
logger.debug(f"Tool call: {chunk.choices[0].delta.tool_calls}")
|
||||
tool_call = chunk.choices[0].delta.tool_calls[0]
|
||||
if tool_call.index != func_idx:
|
||||
functions_list.append(function_name)
|
||||
arguments_list.append(arguments)
|
||||
tool_id_list.append(tool_call_id)
|
||||
function_name = ""
|
||||
arguments = ""
|
||||
tool_call_id = ""
|
||||
func_idx += 1
|
||||
if tool_call.function and tool_call.function.name:
|
||||
function_name += tool_call.function.name
|
||||
tool_call_id = tool_call.id
|
||||
if tool_call.function and tool_call.function.arguments:
|
||||
# Keep iterating through the response to collect all the argument fragments
|
||||
arguments += tool_call.function.arguments
|
||||
elif chunk.choices[0].delta.content:
|
||||
await self.push_frame(LLMTextFrame(chunk.choices[0].delta.content))
|
||||
if chunk.choices[0].delta.tool_calls:
|
||||
# We're streaming the LLM response to enable the fastest response times.
|
||||
# For text, we just yield each chunk as we receive it and count on consumers
|
||||
# to do whatever coalescing they need (eg. to pass full sentences to TTS)
|
||||
#
|
||||
# If the LLM is a function call, we'll do some coalescing here.
|
||||
# If the response contains a function name, we'll yield a frame to tell consumers
|
||||
# that they can start preparing to call the function with that name.
|
||||
# We accumulate all the arguments for the rest of the streamed response, then when
|
||||
# the response is done, we package up all the arguments and the function name and
|
||||
# yield a frame containing the function name and the arguments.
|
||||
logger.debug(f"Tool call: {chunk.choices[0].delta.tool_calls}")
|
||||
tool_call = chunk.choices[0].delta.tool_calls[0]
|
||||
if tool_call.index != func_idx:
|
||||
functions_list.append(function_name)
|
||||
arguments_list.append(arguments)
|
||||
tool_id_list.append(tool_call_id)
|
||||
function_name = ""
|
||||
arguments = ""
|
||||
tool_call_id = ""
|
||||
func_idx += 1
|
||||
if tool_call.function and tool_call.function.name:
|
||||
function_name += tool_call.function.name
|
||||
tool_call_id = tool_call.id
|
||||
if tool_call.function and tool_call.function.arguments:
|
||||
# Keep iterating through the response to collect all the argument fragments
|
||||
arguments += tool_call.function.arguments
|
||||
elif chunk.choices[0].delta.content:
|
||||
await self.push_frame(LLMTextFrame(chunk.choices[0].delta.content))
|
||||
|
||||
# if we got a function name and arguments, check to see if it's a function with
|
||||
# a registered handler. If so, run the registered callback, save the result to
|
||||
|
||||
@@ -131,59 +131,62 @@ class SambaNovaLLMService(OpenAILLMService): # type: ignore
|
||||
else self._stream_chat_completions_universal_context(context)
|
||||
)
|
||||
|
||||
async for chunk in chunk_stream:
|
||||
if chunk.usage:
|
||||
tokens = LLMTokenUsage(
|
||||
prompt_tokens=chunk.usage.prompt_tokens,
|
||||
completion_tokens=chunk.usage.completion_tokens,
|
||||
total_tokens=chunk.usage.total_tokens,
|
||||
)
|
||||
await self.start_llm_usage_metrics(tokens)
|
||||
# Use context manager to ensure stream is closed on cancellation/exception.
|
||||
# Without this, CancelledError during iteration leaves the underlying socket open.
|
||||
async with chunk_stream:
|
||||
async for chunk in chunk_stream:
|
||||
if chunk.usage:
|
||||
tokens = LLMTokenUsage(
|
||||
prompt_tokens=chunk.usage.prompt_tokens,
|
||||
completion_tokens=chunk.usage.completion_tokens,
|
||||
total_tokens=chunk.usage.total_tokens,
|
||||
)
|
||||
await self.start_llm_usage_metrics(tokens)
|
||||
|
||||
if chunk.choices is None or len(chunk.choices) == 0:
|
||||
continue
|
||||
if chunk.choices is None or len(chunk.choices) == 0:
|
||||
continue
|
||||
|
||||
await self.stop_ttfb_metrics()
|
||||
await self.stop_ttfb_metrics()
|
||||
|
||||
if not chunk.choices[0].delta:
|
||||
continue
|
||||
if not chunk.choices[0].delta:
|
||||
continue
|
||||
|
||||
if chunk.choices[0].delta.tool_calls:
|
||||
# We're streaming the LLM response to enable the fastest response times.
|
||||
# For text, we just yield each chunk as we receive it and count on consumers
|
||||
# to do whatever coalescing they need (eg. to pass full sentences to TTS)
|
||||
#
|
||||
# If the LLM is a function call, we'll do some coalescing here.
|
||||
# If the response contains a function name, we'll yield a frame to tell consumers
|
||||
# that they can start preparing to call the function with that name.
|
||||
# We accumulate all the arguments for the rest of the streamed response, then when
|
||||
# the response is done, we package up all the arguments and the function name and
|
||||
# yield a frame containing the function name and the arguments.
|
||||
if chunk.choices[0].delta.tool_calls:
|
||||
# We're streaming the LLM response to enable the fastest response times.
|
||||
# For text, we just yield each chunk as we receive it and count on consumers
|
||||
# to do whatever coalescing they need (eg. to pass full sentences to TTS)
|
||||
#
|
||||
# If the LLM is a function call, we'll do some coalescing here.
|
||||
# If the response contains a function name, we'll yield a frame to tell consumers
|
||||
# that they can start preparing to call the function with that name.
|
||||
# We accumulate all the arguments for the rest of the streamed response, then when
|
||||
# the response is done, we package up all the arguments and the function name and
|
||||
# yield a frame containing the function name and the arguments.
|
||||
|
||||
tool_call = chunk.choices[0].delta.tool_calls[0]
|
||||
if tool_call.index != func_idx:
|
||||
functions_list.append(function_name)
|
||||
arguments_list.append(arguments)
|
||||
tool_id_list.append(tool_call_id)
|
||||
function_name = ""
|
||||
arguments = ""
|
||||
tool_call_id = ""
|
||||
func_idx += 1
|
||||
if tool_call.function and tool_call.function.name:
|
||||
function_name += tool_call.function.name
|
||||
tool_call_id = tool_call.id # type: ignore
|
||||
if tool_call.function and tool_call.function.arguments:
|
||||
# Keep iterating through the response to collect all the argument fragments
|
||||
arguments += tool_call.function.arguments
|
||||
elif chunk.choices[0].delta.content:
|
||||
await self.push_frame(LLMTextFrame(chunk.choices[0].delta.content))
|
||||
tool_call = chunk.choices[0].delta.tool_calls[0]
|
||||
if tool_call.index != func_idx:
|
||||
functions_list.append(function_name)
|
||||
arguments_list.append(arguments)
|
||||
tool_id_list.append(tool_call_id)
|
||||
function_name = ""
|
||||
arguments = ""
|
||||
tool_call_id = ""
|
||||
func_idx += 1
|
||||
if tool_call.function and tool_call.function.name:
|
||||
function_name += tool_call.function.name
|
||||
tool_call_id = tool_call.id # type: ignore
|
||||
if tool_call.function and tool_call.function.arguments:
|
||||
# Keep iterating through the response to collect all the argument fragments
|
||||
arguments += tool_call.function.arguments
|
||||
elif chunk.choices[0].delta.content:
|
||||
await self.push_frame(LLMTextFrame(chunk.choices[0].delta.content))
|
||||
|
||||
# When gpt-4o-audio / gpt-4o-mini-audio is used for llm or stt+llm
|
||||
# we need to get LLMTextFrame for the transcript
|
||||
elif hasattr(chunk.choices[0].delta, "audio") and chunk.choices[0].delta.audio.get(
|
||||
"transcript"
|
||||
):
|
||||
await self.push_frame(LLMTextFrame(chunk.choices[0].delta.audio["transcript"]))
|
||||
# When gpt-4o-audio / gpt-4o-mini-audio is used for llm or stt+llm
|
||||
# we need to get LLMTextFrame for the transcript
|
||||
elif hasattr(chunk.choices[0].delta, "audio") and chunk.choices[0].delta.audio.get(
|
||||
"transcript"
|
||||
):
|
||||
await self.push_frame(LLMTextFrame(chunk.choices[0].delta.audio["transcript"]))
|
||||
|
||||
# if we got a function name and arguments, check to see if it's a function with
|
||||
# a registered handler. If so, run the registered callback, save the result to
|
||||
|
||||
81
tests/test_google_llm_openai.py
Normal file
81
tests/test_google_llm_openai.py
Normal file
@@ -0,0 +1,81 @@
|
||||
#
|
||||
# Copyright (c) 2024-2026, Daily
|
||||
#
|
||||
# SPDX-License-Identifier: BSD 2-Clause License
|
||||
#
|
||||
|
||||
"""Unit tests for Google LLM OpenAI Beta service."""
|
||||
|
||||
import asyncio
|
||||
import warnings
|
||||
from unittest.mock import AsyncMock, patch
|
||||
|
||||
import pytest
|
||||
|
||||
from pipecat.processors.aggregators.openai_llm_context import OpenAILLMContext
|
||||
|
||||
try:
|
||||
from pipecat.services.google.llm_openai import GoogleLLMOpenAIBetaService
|
||||
|
||||
google_available = True
|
||||
except Exception:
|
||||
google_available = False
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
@pytest.mark.skipif(not google_available, reason="Google dependencies not installed")
|
||||
async def test_google_llm_openai_stream_closed_on_cancellation():
|
||||
"""Test that the stream is closed when CancelledError occurs during iteration.
|
||||
|
||||
This prevents socket leaks when the pipeline is interrupted (e.g., user interruption).
|
||||
See issue #3639.
|
||||
"""
|
||||
with patch.object(GoogleLLMOpenAIBetaService, "create_client"):
|
||||
with warnings.catch_warnings():
|
||||
warnings.simplefilter("ignore", DeprecationWarning)
|
||||
service = GoogleLLMOpenAIBetaService(api_key="test-key", model="test-model")
|
||||
service._client = AsyncMock()
|
||||
|
||||
stream_closed = False
|
||||
|
||||
class MockAsyncStream:
|
||||
"""Mock AsyncStream that tracks close() calls and raises CancelledError."""
|
||||
|
||||
def __init__(self):
|
||||
self.iteration_count = 0
|
||||
|
||||
async def __aenter__(self):
|
||||
return self
|
||||
|
||||
async def __aexit__(self, exc_type, exc_val, exc_tb):
|
||||
nonlocal stream_closed
|
||||
stream_closed = True
|
||||
return False
|
||||
|
||||
def __aiter__(self):
|
||||
return self
|
||||
|
||||
async def __anext__(self):
|
||||
self.iteration_count += 1
|
||||
if self.iteration_count > 1:
|
||||
raise asyncio.CancelledError()
|
||||
mock_chunk = AsyncMock()
|
||||
mock_chunk.usage = None
|
||||
mock_chunk.choices = []
|
||||
return mock_chunk
|
||||
|
||||
mock_stream = MockAsyncStream()
|
||||
|
||||
service._stream_chat_completions_specific_context = AsyncMock(return_value=mock_stream)
|
||||
service.start_ttfb_metrics = AsyncMock()
|
||||
service.stop_ttfb_metrics = AsyncMock()
|
||||
service.start_llm_usage_metrics = AsyncMock()
|
||||
|
||||
context = OpenAILLMContext(
|
||||
messages=[{"role": "user", "content": "Hello"}],
|
||||
)
|
||||
|
||||
with pytest.raises(asyncio.CancelledError):
|
||||
await service._process_context(context)
|
||||
|
||||
assert stream_closed, "Stream should be closed even when CancelledError occurs"
|
||||
72
tests/test_sambanova_llm.py
Normal file
72
tests/test_sambanova_llm.py
Normal file
@@ -0,0 +1,72 @@
|
||||
#
|
||||
# Copyright (c) 2024-2026, Daily
|
||||
#
|
||||
# SPDX-License-Identifier: BSD 2-Clause License
|
||||
#
|
||||
|
||||
"""Unit tests for SambaNova LLM service."""
|
||||
|
||||
import asyncio
|
||||
from unittest.mock import AsyncMock, patch
|
||||
|
||||
import pytest
|
||||
|
||||
from pipecat.processors.aggregators.llm_context import LLMContext
|
||||
from pipecat.services.sambanova.llm import SambaNovaLLMService
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_sambanova_llm_stream_closed_on_cancellation():
|
||||
"""Test that the stream is closed when CancelledError occurs during iteration.
|
||||
|
||||
This prevents socket leaks when the pipeline is interrupted (e.g., user interruption).
|
||||
See issue #3639.
|
||||
"""
|
||||
with patch.object(SambaNovaLLMService, "create_client"):
|
||||
service = SambaNovaLLMService(api_key="test-key", model="test-model")
|
||||
service._client = AsyncMock()
|
||||
|
||||
stream_closed = False
|
||||
|
||||
class MockAsyncStream:
|
||||
"""Mock AsyncStream that tracks close() calls and raises CancelledError."""
|
||||
|
||||
def __init__(self):
|
||||
self.iteration_count = 0
|
||||
|
||||
async def __aenter__(self):
|
||||
return self
|
||||
|
||||
async def __aexit__(self, exc_type, exc_val, exc_tb):
|
||||
nonlocal stream_closed
|
||||
stream_closed = True
|
||||
return False
|
||||
|
||||
def __aiter__(self):
|
||||
return self
|
||||
|
||||
async def __anext__(self):
|
||||
self.iteration_count += 1
|
||||
if self.iteration_count > 1:
|
||||
raise asyncio.CancelledError()
|
||||
mock_chunk = AsyncMock()
|
||||
mock_chunk.usage = None
|
||||
mock_chunk.choices = []
|
||||
return mock_chunk
|
||||
|
||||
mock_stream = MockAsyncStream()
|
||||
|
||||
service._stream_chat_completions_specific_context = AsyncMock(return_value=mock_stream)
|
||||
service._stream_chat_completions_universal_context = AsyncMock(return_value=mock_stream)
|
||||
service.start_ttfb_metrics = AsyncMock()
|
||||
service.stop_ttfb_metrics = AsyncMock()
|
||||
service.start_llm_usage_metrics = AsyncMock()
|
||||
|
||||
context = LLMContext(
|
||||
messages=[{"role": "user", "content": "Hello"}],
|
||||
)
|
||||
|
||||
with pytest.raises(asyncio.CancelledError):
|
||||
await service._process_context(context)
|
||||
|
||||
assert stream_closed, "Stream should be closed even when CancelledError occurs"
|
||||
Reference in New Issue
Block a user