A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI][gpt-oss] Enable python tool tests in CI (#24315)
Signed-off-by: wuhang <wuhang6@huawei.com>
wuhang committed
Oct 6, 2025 at 12:20 UTC
91ac7f764d04e7a9103e3c839244ce241a43b45e
3 files changed
+24
-26
requirements/common.txt
+1
@@ -49,3 +49,4 @@ pybase64 # fast base64 implementation
49
cbor2 # Required for cross-language serialization of hashable objects
50
setproctitle # Used to set process names for better debugging and monitoring
51
openai-harmony >= 0.0.3 # Required for gpt-oss
52
+gpt-oss >= 0.0.7
tests/entrypoints/openai/test_response_api_with_harmony.py
+16
-22
@@ -15,22 +15,15 @@ MODEL_NAME = "openai/gpt-oss-20b"
15
16
17
@pytest.fixture(scope="module")
18
-def monkeypatch_module():
19
- from _pytest.monkeypatch import MonkeyPatch
20
-
21
- mpatch = MonkeyPatch()
22
- yield mpatch
23
- mpatch.undo()
24
-
25
-
26
-@pytest.fixture(scope="module")
27
-def server(monkeypatch_module: pytest.MonkeyPatch):
18
+def server():
19
args = ["--enforce-eager", "--tool-server", "demo"]
20
+ env_dict = dict(
21
+ VLLM_ENABLE_RESPONSES_API_STORE="1",
22
+ PYTHON_EXECUTION_BACKEND="dangerously_use_uv",
23
+ )
24
30
- with monkeypatch_module.context() as m:
31
- m.setenv("VLLM_ENABLE_RESPONSES_API_STORE", "1")
32
- with RemoteOpenAIServer(MODEL_NAME, args) as remote_server:
33
- yield remote_server
25
+ with RemoteOpenAIServer(MODEL_NAME, args, env_dict=env_dict) as remote_server:
26
+ yield remote_server
27
28
29
@pytest_asyncio.fixture
@@ -316,7 +309,7 @@ async def test_streaming(client: OpenAI, model_name: str, background: bool):
309
# TODO: Add back when web search and code interpreter are available in CI
310
prompts = [
311
"tell me a story about a cat in 20 words",
319
- # "What is 13 * 24? Use python to calculate the result.",
312
+ "What is 13 * 24? Use python to calculate the result.",
313
# "When did Jensen found NVIDIA? Search it and answer the year only.",
314
]
315
@@ -329,12 +322,7 @@ async def test_streaming(client: OpenAI, model_name: str, background: bool):
322
# {
323
# "type": "web_search_preview"
324
# },
332
- # {
333
- # "type": "code_interpreter",
334
- # "container": {
335
- # "type": "auto"
336
- # }
337
- # },
325
+ {"type": "code_interpreter", "container": {"type": "auto"}},
326
],
327
stream=True,
328
background=background,
@@ -412,6 +400,7 @@ async def test_streaming(client: OpenAI, model_name: str, background: bool):
400
async for event in stream:
401
counter += 1
402
assert event == events[counter]
403
+ assert counter == len(events) - 1
404
405
406
@pytest.mark.asyncio
@@ -429,7 +418,6 @@ async def test_web_search(client: OpenAI, model_name: str):
418
419
@pytest.mark.asyncio
420
@pytest.mark.parametrize("model_name", [MODEL_NAME])
432
-@pytest.mark.skip(reason="Code interpreter tool is not available in CI yet.")
421
async def test_code_interpreter(client: OpenAI, model_name: str):
422
response = await client.responses.create(
423
model=model_name,
@@ -443,10 +431,16 @@ async def test_code_interpreter(client: OpenAI, model_name: str):
431
"and you must print to see the output."
432
),
433
tools=[{"type": "code_interpreter", "container": {"type": "auto"}}],
434
+ temperature=0.0, # More deterministic output in response
435
)
436
assert response is not None
437
assert response.status == "completed"
438
assert response.usage.output_tokens_details.tool_output_tokens > 0
439
+ for item in response.output:
440
+ if item.type == "message":
441
+ output_string = item.content[0].text
442
+ print("output_string: ", output_string, flush=True)
443
+ assert "5846" in output_string
444
445
446
def get_weather(latitude, longitude):
vllm/entrypoints/tool.py
+7
-4
@@ -14,10 +14,12 @@ if TYPE_CHECKING:
14
15
logger = init_logger(__name__)
16
17
+MIN_GPT_OSS_VERSION = "0.0.7"
18
+
19
20
def validate_gpt_oss_install():
21
"""
20
- Check if the gpt-oss is installed and its version is at least 0.0.3.
22
+ Check if the gpt-oss is installed and its version is at least 0.0.7.
23
If not, raise an ImportError.
24
"""
25
from importlib.metadata import PackageNotFoundError, version
@@ -25,16 +27,17 @@ def validate_gpt_oss_install():
27
from packaging.version import InvalidVersion, Version
28
29
try:
28
- pkg_version_str = version("gpt_oss") # e.g., "0.0.5"
30
+ pkg_version_str = version("gpt_oss")
31
pkg_version = Version(pkg_version_str)
32
except PackageNotFoundError:
33
raise ImportError("Package 'gpt_oss' is not installed.") from None
34
except InvalidVersion as e:
35
raise ImportError(f"Invalid version string for 'gpt_oss': {e}") from None
36
35
- if pkg_version < Version("0.0.3"):
37
+ if pkg_version < Version(MIN_GPT_OSS_VERSION):
38
raise ImportError(
37
- f"gpt_oss >= 0.0.3 is required, but {pkg_version} is installed."
39
+ f"gpt_oss >= {MIN_GPT_OSS_VERSION} is required, "
40
+ f"but {pkg_version} is installed."
41
) from None
42
43