biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V1] Set structured output backend to auto by default (#15724)

Browse Source 
Signed-off-by: Russell Bryant <rbryant@redhat.com>
This commit is contained in:
Russell Bryant
2025-04-10 13:53:26 -04:00
committed by GitHub
parent 0c54fc7273
commit 9665313c39
4 changed files with 22 additions and 68 deletions
Download Patch File Download Diff File Expand all files Collapse all files

74
tests/entrypoints/openai/test_chat.py
View File

@@ -20,8 +20,6 @@ from .test_completion import zephyr_lora_files # noqa: F401
# any model with a chat template should work here
MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
GUIDED_DECODING_BACKENDS = ["outlines", "lm-format-enforcer", "xgrammar"]
@pytest.fixture(scope="module")
def monkeypatch_module():
@@ -487,20 +485,9 @@ async def test_chat_completion_stream_options(client: openai.AsyncOpenAI,
assert last_completion_tokens == 10
# NOTE: Not sure why, but when I place this after `test_guided_regex_chat`
# (i.e. using the same ordering as in the Completions API tests), the test
# will fail on the second `guided_decoding_backend` even when I swap their order
# (ref: https://github.com/vllm-project/vllm/pull/5526#issuecomment-2173772256)
@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
async def test_guided_choice_chat(client: openai.AsyncOpenAI,
is_v1_server: bool,
guided_decoding_backend: str,
sample_guided_choice):
if is_v1_server and guided_decoding_backend != 'xgrammar':
pytest.skip("Only xgrammar backend is supported with V1")
messages = [{
"role": "system",
"content": "you are a helpful assistant"
@@ -515,8 +502,7 @@ async def test_guided_choice_chat(client: openai.AsyncOpenAI,
messages=messages,
max_completion_tokens=10,
temperature=0.7,
extra_body=dict(guided_choice=sample_guided_choice,
guided_decoding_backend=guided_decoding_backend))
extra_body=dict(guided_choice=sample_guided_choice))
choice1 = chat_completion.choices[0].message.content
assert choice1 in sample_guided_choice
@@ -530,22 +516,16 @@ async def test_guided_choice_chat(client: openai.AsyncOpenAI,
messages=messages,
max_completion_tokens=10,
temperature=0.7,
extra_body=dict(guided_choice=sample_guided_choice,
guided_decoding_backend=guided_decoding_backend))
extra_body=dict(guided_choice=sample_guided_choice))
choice2 = chat_completion.choices[0].message.content
assert choice2 in sample_guided_choice
assert choice1 != choice2
@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
async def test_guided_json_chat(client: openai.AsyncOpenAI, is_v1_server: bool,
guided_decoding_backend: str,
async def test_guided_json_chat(client: openai.AsyncOpenAI,
sample_json_schema):
if is_v1_server:
pytest.skip("sample_json_schema has features unsupported in V1")
messages = [{
"role": "system",
"content": "you are a helpful assistant"
@@ -560,8 +540,7 @@ async def test_guided_json_chat(client: openai.AsyncOpenAI, is_v1_server: bool,
model=MODEL_NAME,
messages=messages,
max_completion_tokens=1000,
extra_body=dict(guided_json=sample_json_schema,
guided_decoding_backend=guided_decoding_backend))
extra_body=dict(guided_json=sample_json_schema))
message = chat_completion.choices[0].message
assert message.content is not None
json1 = json.loads(message.content)
@@ -578,8 +557,7 @@ async def test_guided_json_chat(client: openai.AsyncOpenAI, is_v1_server: bool,
model=MODEL_NAME,
messages=messages,
max_completion_tokens=1000,
extra_body=dict(guided_json=sample_json_schema,
guided_decoding_backend=guided_decoding_backend))
extra_body=dict(guided_json=sample_json_schema))
message = chat_completion.choices[0].message
assert message.content is not None
json2 = json.loads(message.content)
@@ -589,13 +567,7 @@ async def test_guided_json_chat(client: openai.AsyncOpenAI, is_v1_server: bool,
@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
async def test_guided_regex_chat(client: openai.AsyncOpenAI,
is_v1_server: bool,
guided_decoding_backend: str, sample_regex):
if is_v1_server and guided_decoding_backend != 'xgrammar':
pytest.skip("Only xgrammar backend is supported with V1")
async def test_guided_regex_chat(client: openai.AsyncOpenAI, sample_regex):
messages = [{
"role": "system",
@@ -610,8 +582,7 @@ async def test_guided_regex_chat(client: openai.AsyncOpenAI,
model=MODEL_NAME,
messages=messages,
max_completion_tokens=20,
extra_body=dict(guided_regex=sample_regex,
guided_decoding_backend=guided_decoding_backend))
extra_body=dict(guided_regex=sample_regex))
ip1 = chat_completion.choices[0].message.content
assert ip1 is not None
assert re.fullmatch(sample_regex, ip1) is not None
@@ -622,8 +593,7 @@ async def test_guided_regex_chat(client: openai.AsyncOpenAI,
model=MODEL_NAME,
messages=messages,
max_completion_tokens=20,
extra_body=dict(guided_regex=sample_regex,
guided_decoding_backend=guided_decoding_backend))
extra_body=dict(guided_regex=sample_regex))
ip2 = chat_completion.choices[0].message.content
assert ip2 is not None
assert re.fullmatch(sample_regex, ip2) is not None
@@ -652,15 +622,9 @@ async def test_guided_decoding_type_error(client: openai.AsyncOpenAI):
@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
async def test_guided_choice_chat_logprobs(client: openai.AsyncOpenAI,
is_v1_server: bool,
guided_decoding_backend: str,
sample_guided_choice):
if is_v1_server and guided_decoding_backend != 'xgrammar':
pytest.skip("Only xgrammar backend is supported with V1")
messages = [{
"role": "system",
"content": "you are a helpful assistant"
@@ -676,8 +640,7 @@ async def test_guided_choice_chat_logprobs(client: openai.AsyncOpenAI,
max_completion_tokens=10,
logprobs=True,
top_logprobs=5,
extra_body=dict(guided_choice=sample_guided_choice,
guided_decoding_backend=guided_decoding_backend))
extra_body=dict(guided_choice=sample_guided_choice))
assert chat_completion.choices[0].logprobs is not None
assert chat_completion.choices[0].logprobs.content is not None
@@ -689,14 +652,7 @@ async def test_guided_choice_chat_logprobs(client: openai.AsyncOpenAI,
@pytest.mark.asyncio
@pytest.mark.parametrize("guided_decoding_backend", GUIDED_DECODING_BACKENDS)
async def test_named_tool_use(client: openai.AsyncOpenAI, is_v1_server: bool,
guided_decoding_backend: str,
sample_json_schema):
if is_v1_server:
pytest.skip("sample_json_schema has features unsupported on V1")
async def test_named_tool_use(client: openai.AsyncOpenAI, sample_json_schema):
messages = [{
"role": "system",
"content": "you are a helpful assistant"
@@ -728,7 +684,7 @@ async def test_named_tool_use(client: openai.AsyncOpenAI, is_v1_server: bool,
"name": "dummy_function_name"
}
},
extra_body=dict(guided_decoding_backend=guided_decoding_backend))
)
message = chat_completion.choices[0].message
assert len(message.content) == 0
json_string = message.tool_calls[0].function.arguments
@@ -763,7 +719,6 @@ async def test_named_tool_use(client: openai.AsyncOpenAI, is_v1_server: bool,
"name": "dummy_function_name"
}
},
extra_body=dict(guided_decoding_backend=guided_decoding_backend),
stream=True)
output = []
@@ -888,7 +843,6 @@ async def test_required_tool_use(client: openai.AsyncOpenAI,
model=model_name,
tools=tools,
tool_choice="required",
extra_body=dict(guided_decoding_backend="outlines"),
)
assert chat_completion.choices[0].message.tool_calls is not None
@@ -900,7 +854,6 @@ async def test_required_tool_use(client: openai.AsyncOpenAI,
model=model_name,
tools=tools,
tool_choice="required",
extra_body=dict(guided_decoding_backend="outlines"),
stream=True,
)
@@ -914,12 +867,7 @@ async def test_required_tool_use(client: openai.AsyncOpenAI,
@pytest.mark.asyncio
async def test_inconsistent_tool_choice_and_tools(client: openai.AsyncOpenAI,
is_v1_server: bool,
sample_json_schema):
if is_v1_server:
pytest.skip("sample_json_schema has features unsupported on V1")
messages = [{
"role": "system",
"content": "you are a helpful assistant"