[Bugfix][Frontend] Preserve Cohere and Anthropic streaming errors (#61947)

Signed-off-by: Kevin <jia__kuan@163.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Jovan
2026-10-17 06:10:16 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 90d6d871d6
commit 5be4266f5c
6 changed files with 531 additions and 375 deletions
@@ -28,6 +28,7 @@ from vllm.entrypoints.generate.base.serving import build_per_request_timing_metr
from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionRequest,
ChatCompletionResponse,
ChatCompletionStreamResponse,
)
from vllm.entrypoints.openai.chat_completion.serving import (
OpenAIServingChat,
@@ -43,7 +44,7 @@ from vllm.entrypoints.openai.models.serving import (
from vllm.entrypoints.openai.parser.harmony_utils import get_encoding
from vllm.entrypoints.serve.engine.protocol import ErrorResponse
from vllm.entrypoints.serve.utils.tool_calls_utils import resolve_finish_reason
from vllm.exceptions import QueueOverflowError, VLLMValidationError
from vllm.exceptions import GenerationError, QueueOverflowError, VLLMValidationError
from vllm.inputs import TokensPrompt
from vllm.logprobs import Logprob
from vllm.multimodal.inputs import PlaceholderRange
@@ -716,6 +717,149 @@ async def _collect_metrics_stream_chunks(
return chunks
@pytest.mark.asyncio
@pytest.mark.parametrize("raw", [False, True])
async def test_streaming_chunks_keep_usage_and_metadata(raw):
serving = _build_minimal_metrics_serving_chat(False)
request = ChatCompletionRequest(
model="test-model",
messages=[{"role": "user", "content": "Test"}],
stream=True,
stream_options={"include_usage": True},
)
metadata = RequestResponseMetadata(request_id="chatcmpl-test")
method = (
serving.chat_completion_stream_generator_raw
if raw
else serving.chat_completion_stream_generator
)
chunks = [
chunk
async for chunk in method(
request,
_single_request_output(_make_metrics_request_output()),
"chatcmpl-test",
"test-model",
conversation=[],
tokenizer=MagicMock(),
request_metadata=metadata,
)
]
if raw:
assert all(isinstance(chunk, ChatCompletionStreamResponse) for chunk in chunks)
else:
assert chunks[-1] == "data: [DONE]\n\n"
first = json.loads(chunks[0].removeprefix("data: "))
assert first["choices"][0]["logprobs"] is None
usage = json.loads(chunks[-2].removeprefix("data: "))
assert "system_fingerprint" not in usage
assert "prompt_tokens_details" not in usage["usage"]
chunks = [
ChatCompletionStreamResponse.model_validate_json(
chunk.removeprefix("data: ")
)
for chunk in chunks[:-1]
]
assert len(chunks) == 3
assert chunks[0].choices[0].delta.role == "assistant"
assert chunks[1].choices[0].delta.content == "Hello"
assert chunks[1].choices[0].finish_reason == "stop"
assert chunks[2].choices == []
assert chunks[2].usage.prompt_tokens == 3
assert chunks[2].usage.completion_tokens == 2
assert metadata.final_usage_info.total_tokens == 5
@pytest.mark.asyncio
async def test_streaming_serialization_failure_emits_error_and_done():
serving = _build_minimal_metrics_serving_chat(False)
serving.parser_cls = HarmonyParser
parser = MagicMock()
parser.parse_delta.return_value = DeltaMessage(
content="Hello", extra_object=object()
)
serving._make_parser = MagicMock(return_value=parser)
request = ChatCompletionRequest(
model="test-model", messages=[{"role": "user", "content": "Test"}], stream=True
)
frames = [
frame
async for frame in serving.chat_completion_stream_generator(
request,
_single_request_output(_make_metrics_request_output()),
"chatcmpl-test",
"test-model",
conversation=[],
tokenizer=MagicMock(),
request_metadata=RequestResponseMetadata(request_id="chatcmpl-test"),
)
]
assert frames[-1] == "data: [DONE]\n\n"
error = ErrorResponse.model_validate_json(frames[-2].removeprefix("data: "))
assert "serialize" in error.error.message
@pytest.mark.asyncio
@pytest.mark.parametrize("raw", [False, True])
@pytest.mark.parametrize(
"failure", ["parser", "before_first", "after_content", "generation"]
)
async def test_streaming_errors_preserve_envelope(raw, failure):
serving = _build_minimal_metrics_serving_chat(False)
error = (
GenerationError("backend failed")
if failure == "generation"
else ValueError("backend failed")
)
if failure == "parser":
serving.parser_cls = HarmonyParser
serving._make_parser = MagicMock(side_effect=error)
async def results():
if failure == "after_content":
output = _make_metrics_request_output()
output.outputs[0].finish_reason = None
output.finished = False
yield output
raise error
request = ChatCompletionRequest(
model="test-model", messages=[{"role": "user", "content": "Test"}], stream=True
)
method = (
serving.chat_completion_stream_generator_raw
if raw
else serving.chat_completion_stream_generator
)
chunks = [
chunk
async for chunk in method(
request,
results(),
"chatcmpl-test",
"test-model",
conversation=[],
tokenizer=MagicMock(),
request_metadata=RequestResponseMetadata(request_id="chatcmpl-test"),
)
]
if raw:
assert all(
isinstance(chunk, (ChatCompletionStreamResponse, ErrorResponse))
for chunk in chunks
)
assert isinstance(chunks[-1], ErrorResponse)
error_response = chunks[-1]
else:
assert chunks[-1] == "data: [DONE]\n\n"
assert chunks.count("data: [DONE]\n\n") == 1
error_response = ErrorResponse.model_validate_json(
chunks[-2].removeprefix("data: ")
)
assert error_response.error.message == "backend failed"
assert error_response.error.code == (500 if failure == "generation" else 400)
def test_build_per_request_timing_metrics_valid_timestamps():
metrics = build_per_request_timing_metrics(
_PER_REQUEST_STATS, num_generation_tokens=10