[Frontend] Consolidate online serving utils. (#44479)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
wang.yuqi
2026-06-04 06:49:31 +00:00
committed by GitHub
parent b4b4aaa70e
commit d01d0b4646
81 changed files with 466 additions and 435 deletions
@@ -0,0 +1,120 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import pytest
from vllm.entrypoints.openai.engine.protocol import StreamOptions
from vllm.entrypoints.serve.utils.api_utils import (
get_max_tokens,
sanitize_message,
should_include_usage,
)
def test_sanitize_message():
assert (
sanitize_message("<_io.BytesIO object at 0x7a95e299e750>")
== "<_io.BytesIO object>"
)
@pytest.mark.parametrize(
("stream_options", "expected"),
[
(None, (True, True)),
(StreamOptions(include_usage=False), (True, True)),
(
StreamOptions(include_usage=False, continuous_usage_stats=False),
(True, True),
),
(
StreamOptions(include_usage=True, continuous_usage_stats=False),
(True, True),
),
],
)
def test_should_include_usage_force_enables_continuous_usage(stream_options, expected):
assert should_include_usage(stream_options, True) == expected
class TestGetMaxTokens:
"""Tests for get_max_tokens() to ensure generation_config's max_tokens
acts as a default when from model author, and as a ceiling when
explicitly set by the user."""
def test_default_sampling_params_used_when_no_request_max_tokens(self):
"""When user doesn't specify max_tokens, generation_config default
should apply."""
result = get_max_tokens(
max_model_len=24000,
max_tokens=None,
input_length=100,
default_sampling_params={"max_tokens": 2048},
)
assert result == 2048
def test_request_max_tokens_not_capped_by_default_sampling_params(self):
"""When user specifies max_tokens in request, model author's
generation_config max_tokens must NOT cap it (fixes #34005)."""
result = get_max_tokens(
max_model_len=24000,
max_tokens=5000,
input_length=100,
default_sampling_params={"max_tokens": 2048},
)
assert result == 5000
def test_override_max_tokens_caps_request(self):
"""When user explicitly sets max_tokens, it acts as a ceiling."""
result = get_max_tokens(
max_model_len=24000,
max_tokens=5000,
input_length=100,
default_sampling_params={"max_tokens": 2048},
override_max_tokens=2048,
)
assert result == 2048
def test_override_max_tokens_used_as_default(self):
"""When no request max_tokens, override still applies as default."""
result = get_max_tokens(
max_model_len=24000,
max_tokens=None,
input_length=100,
default_sampling_params={"max_tokens": 2048},
override_max_tokens=2048,
)
assert result == 2048
def test_max_model_len_still_caps_output(self):
"""max_model_len - input_length is always the hard ceiling."""
result = get_max_tokens(
max_model_len=3000,
max_tokens=5000,
input_length=100,
default_sampling_params={"max_tokens": 2048},
)
assert result == 2900 # 3000 - 100
def test_request_max_tokens_smaller_than_default(self):
"""When user explicitly requests fewer tokens than gen_config default,
that should be respected."""
result = get_max_tokens(
max_model_len=24000,
max_tokens=512,
input_length=100,
default_sampling_params={"max_tokens": 2048},
)
assert result == 512
def test_input_length_exceeds_max_model_len(self):
with pytest.raises(
ValueError,
match="Input length .* exceeds model's maximum context length .*",
):
get_max_tokens(
max_model_len=100,
max_tokens=50,
input_length=150,
default_sampling_params={"max_tokens": 2048},
)
@@ -0,0 +1,76 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Unit tests for ``system_fingerprint`` construction."""
from types import SimpleNamespace
import pytest
from vllm.entrypoints.serve.utils import fingerprint as fp
def _cfg(tp=1, pp=1, dp=1, ep=False, digest="a3b21f94deadbeef"):
c = SimpleNamespace(
parallel_config=SimpleNamespace(
tensor_parallel_size=tp,
pipeline_parallel_size=pp,
data_parallel_size=dp,
enable_expert_parallel=ep,
)
)
c.compute_hash = lambda: digest # type: ignore[attr-defined]
return c
@pytest.fixture(autouse=True)
def _reset():
fp.set_default_fingerprint_mode("full")
yield
fp.set_default_fingerprint_mode("full")
def test_four_modes_produce_expected_shapes():
from vllm import __version__ as v
cfg = _cfg(tp=8, ep=True)
assert fp.build_system_fingerprint(cfg, "full") == (f"vllm-{v}-tp8-ep-a3b21f94")
assert fp.build_system_fingerprint(cfg, "hash") == f"vllm-{v}-a3b21f94"
assert fp.build_system_fingerprint(cfg, "custom", "my-fp") == "my-fp"
assert fp.build_system_fingerprint(cfg, "none") is None
def test_full_mode_emits_only_non_trivial_parallelism():
from vllm import __version__ as v
# Single-GPU: nothing between version and hash.
assert fp.build_system_fingerprint(_cfg(), "full") == f"vllm-{v}-a3b21f94"
# All parallelism axes.
assert (
fp.build_system_fingerprint(_cfg(tp=8, pp=2, dp=4, ep=True), "full")
== f"vllm-{v}-tp8-pp2-dp4-ep-a3b21f94"
)
def test_get_respects_set_default():
cfg = _cfg(tp=8)
full = fp.get_system_fingerprint(cfg)
assert full == fp.get_system_fingerprint(cfg)
fp.set_default_fingerprint_mode("hash")
hashed = fp.get_system_fingerprint(cfg)
assert hashed != full
assert "tp8" not in hashed
fp.set_default_fingerprint_mode("custom", "deploy-42")
assert fp.get_system_fingerprint(cfg) == "deploy-42"
fp.set_default_fingerprint_mode("none")
assert fp.get_system_fingerprint(cfg) is None
def test_compute_hash_failure_does_not_raise():
cfg = _cfg()
cfg.compute_hash = lambda: (_ for _ in ()).throw(RuntimeError("boom"))
assert fp.build_system_fingerprint(cfg, "full").endswith("-nohash")
assert fp.build_system_fingerprint(cfg, "hash").endswith("-nohash")
@@ -0,0 +1,248 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from unittest.mock import MagicMock, patch
from vllm.entrypoints.serve.utils.request_logger import RequestLogger
def test_request_logger_log_outputs():
"""Test the new log_outputs functionality."""
# Create a mock logger to capture log calls
mock_logger = MagicMock()
with patch("vllm.entrypoints.serve.utils.request_logger.logger", mock_logger):
request_logger = RequestLogger(max_log_len=None)
# Test basic output logging
request_logger.log_outputs(
request_id="test-123",
outputs="Hello, world!",
output_token_ids=[1, 2, 3, 4],
finish_reason="stop",
is_streaming=False,
delta=False,
)
mock_logger.info.assert_called_once()
call_args = mock_logger.info.call_args.args
assert "Generated response %s%s" in call_args[0]
assert call_args[1] == "test-123"
assert call_args[3] == "Hello, world!"
assert call_args[4] == [1, 2, 3, 4]
assert call_args[5] == "stop"
def test_request_logger_log_outputs_streaming_delta():
"""Test log_outputs with streaming delta mode."""
mock_logger = MagicMock()
with patch("vllm.entrypoints.serve.utils.request_logger.logger", mock_logger):
request_logger = RequestLogger(max_log_len=None)
# Test streaming delta logging
request_logger.log_outputs(
request_id="test-456",
outputs="Hello",
output_token_ids=[1],
finish_reason=None,
is_streaming=True,
delta=True,
)
mock_logger.info.assert_called_once()
call_args = mock_logger.info.call_args.args
assert "Generated response %s%s" in call_args[0]
assert call_args[1] == "test-456"
assert call_args[2] == " (streaming delta)"
assert call_args[3] == "Hello"
assert call_args[4] == [1]
assert call_args[5] is None
def test_request_logger_log_outputs_streaming_complete():
"""Test log_outputs with streaming complete mode."""
mock_logger = MagicMock()
with patch("vllm.entrypoints.serve.utils.request_logger.logger", mock_logger):
request_logger = RequestLogger(max_log_len=None)
# Test streaming complete logging
request_logger.log_outputs(
request_id="test-789",
outputs="Complete response",
output_token_ids=[1, 2, 3],
finish_reason="length",
is_streaming=True,
delta=False,
)
mock_logger.info.assert_called_once()
call_args = mock_logger.info.call_args.args
assert "Generated response %s%s" in call_args[0]
assert call_args[1] == "test-789"
assert call_args[2] == " (streaming complete)"
assert call_args[3] == "Complete response"
assert call_args[4] == [1, 2, 3]
assert call_args[5] == "length"
def test_request_logger_log_outputs_with_truncation():
"""Test log_outputs respects max_log_len setting."""
mock_logger = MagicMock()
with patch("vllm.entrypoints.serve.utils.request_logger.logger", mock_logger):
# Set max_log_len to 10
request_logger = RequestLogger(max_log_len=10)
# Test output truncation
long_output = "This is a very long output that should be truncated"
long_token_ids = list(range(20)) # 20 tokens
request_logger.log_outputs(
request_id="test-truncate",
outputs=long_output,
output_token_ids=long_token_ids,
finish_reason="stop",
is_streaming=False,
delta=False,
)
mock_logger.info.assert_called_once()
call_args = mock_logger.info.call_args
# Check that output was truncated to first 10 characters
logged_output = call_args[0][3]
assert logged_output == "This is a "
assert len(logged_output) == 10
# Check that token IDs were truncated to first 10 tokens
logged_token_ids = call_args[0][4]
assert logged_token_ids == list(range(10))
assert len(logged_token_ids) == 10
def test_request_logger_log_outputs_none_values():
"""Test log_outputs handles None values correctly."""
mock_logger = MagicMock()
with patch("vllm.entrypoints.serve.utils.request_logger.logger", mock_logger):
request_logger = RequestLogger(max_log_len=None)
# Test with None output_token_ids
request_logger.log_outputs(
request_id="test-none",
outputs="Test output",
output_token_ids=None,
finish_reason="stop",
is_streaming=False,
delta=False,
)
mock_logger.info.assert_called_once()
call_args = mock_logger.info.call_args.args
assert "Generated response %s%s" in call_args[0]
assert call_args[1] == "test-none"
assert call_args[3] == "Test output"
assert call_args[4] is None
assert call_args[5] == "stop"
def test_request_logger_log_outputs_empty_output():
"""Test log_outputs handles empty output correctly."""
mock_logger = MagicMock()
with patch("vllm.entrypoints.serve.utils.request_logger.logger", mock_logger):
request_logger = RequestLogger(max_log_len=5)
# Test with empty output
request_logger.log_outputs(
request_id="test-empty",
outputs="",
output_token_ids=[],
finish_reason="stop",
is_streaming=False,
delta=False,
)
mock_logger.info.assert_called_once()
call_args = mock_logger.info.call_args.args
assert "Generated response %s%s" in call_args[0]
assert call_args[1] == "test-empty"
assert call_args[3] == ""
assert call_args[4] == []
assert call_args[5] == "stop"
def test_request_logger_log_outputs_integration():
"""Test that log_outputs can be called alongside log_inputs."""
mock_logger = MagicMock()
with patch("vllm.entrypoints.serve.utils.request_logger.logger", mock_logger):
request_logger = RequestLogger(max_log_len=None)
# Test that both methods can be called without interference
request_logger.log_inputs(
request_id="test-integration",
prompt="Test prompt",
prompt_token_ids=[1, 2, 3],
prompt_embeds=None,
params=None,
lora_request=None,
)
request_logger.log_outputs(
request_id="test-integration",
outputs="Test output",
output_token_ids=[4, 5, 6],
finish_reason="stop",
is_streaming=False,
delta=False,
)
# Should have been called twice - once for inputs, once for outputs
assert mock_logger.info.call_count == 2
# Check that the calls were made with correct patterns
input_call = mock_logger.info.call_args_list[0][0]
output_call = mock_logger.info.call_args_list[1][0]
assert "Received request %s" in input_call[0]
assert input_call[1] == "test-integration"
assert "Generated response %s%s" in output_call[0]
assert output_call[1] == "test-integration"
def test_streaming_complete_logs_full_text_content():
"""Test that streaming complete logging includes
full accumulated text, not just token count."""
mock_logger = MagicMock()
with patch("vllm.entrypoints.serve.utils.request_logger.logger", mock_logger):
request_logger = RequestLogger(max_log_len=None)
# Test with actual content instead of token count format
full_response = "This is a complete response from streaming"
request_logger.log_outputs(
request_id="test-streaming-full-text",
outputs=full_response,
output_token_ids=None,
finish_reason="streaming_complete",
is_streaming=True,
delta=False,
)
mock_logger.info.assert_called_once()
call_args = mock_logger.info.call_args.args
# Verify the logged output is the full text, not a token count format
logged_output = call_args[3]
assert logged_output == full_response
assert "tokens>" not in logged_output
assert "streaming_complete" not in logged_output
# Verify other parameters
assert call_args[1] == "test-streaming-full-text"
assert call_args[2] == " (streaming complete)"
assert call_args[5] == "streaming_complete"
@@ -0,0 +1,72 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import asyncio
import tempfile
from pathlib import Path
from ssl import SSLContext
import pytest
from vllm.entrypoints.serve.utils.ssl import SSLCertRefresher
class MockSSLContext(SSLContext):
def __init__(self):
self.load_cert_chain_count = 0
self.load_ca_count = 0
def load_cert_chain(
self,
certfile,
keyfile=None,
password=None,
):
self.load_cert_chain_count += 1
def load_verify_locations(
self,
cafile=None,
capath=None,
cadata=None,
):
self.load_ca_count += 1
def create_file() -> str:
with tempfile.NamedTemporaryFile(dir="/tmp", delete=False) as f:
return f.name
def touch_file(path: str) -> None:
Path(path).touch()
@pytest.mark.asyncio
async def test_ssl_refresher():
ssl_context = MockSSLContext()
key_path = create_file()
cert_path = create_file()
ca_path = create_file()
ssl_refresher = SSLCertRefresher(ssl_context, key_path, cert_path, ca_path)
await asyncio.sleep(1)
assert ssl_context.load_cert_chain_count == 0
assert ssl_context.load_ca_count == 0
touch_file(key_path)
await asyncio.sleep(1)
assert ssl_context.load_cert_chain_count == 1
assert ssl_context.load_ca_count == 0
touch_file(cert_path)
touch_file(ca_path)
await asyncio.sleep(1)
assert ssl_context.load_cert_chain_count == 2
assert ssl_context.load_ca_count == 1
ssl_refresher.stop()
touch_file(cert_path)
touch_file(ca_path)
await asyncio.sleep(1)
assert ssl_context.load_cert_chain_count == 2
assert ssl_context.load_ca_count == 1