EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Add CI for gpt-oss model on hopper (#8851)

Browse Source
This commit is contained in:
fzyzcjy
2025-08-09 15:34:23 +08:00
committed by GitHub
parent de8b8b6e5c
commit 442534aa44
7 changed files with 187 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
test/srt/run_suite.py
View File

@@ -63,6 +63,7 @@ suites = {
TestFile("test_fp8_kernel.py", 8),
TestFile("test_function_call_parser.py", 10),
TestFile("test_fused_moe.py", 30),
TestFile("test_gpt_oss_1gpu.py", 600),
TestFile("test_hicache.py", 116),
TestFile("test_hicache_mla.py", 127),
TestFile("test_hicache_storage.py", 127),
@@ -104,7 +105,7 @@ suites = {
TestFile("test_utils_update_weights.py", 48),
TestFile("test_vision_chunked_prefill.py", 175),
TestFile("test_vlm_input_format.py", 300),
TestFile("test_vision_openai_server_a.py", 584),
TestFile("test_vision_openai_server_a.py", 989),
TestFile("test_vision_openai_server_b.py", 620),
TestFile("test_w8a8_quantization.py", 46),
TestFile("test_reasoning_parser.py", 5),
@@ -176,6 +177,7 @@ suites = {
TestFile("test_update_weights_from_distributed.py", 103),
],
"per-commit-4-gpu": [
TestFile("test_gpt_oss_4gpu.py", 600),
TestFile("test_local_attn.py", 250),
TestFile("test_pp_single_node.py", 372),
TestFile("test_multi_instance_release_memory_occupation.py", 64),

31
test/srt/test_gpt_oss_1gpu.py Normal file
View File

@@ -0,0 +1,31 @@
import unittest
from test_gpt_oss_common import BaseTestGptOss
class TestGptOss1Gpu(BaseTestGptOss):
def test_mxfp4_20b(self):
self.run_test(
model_variant="20b",
quantization="mxfp4",
expected_score_of_reasoning_effort={
"low": 0.38,
"medium": 0.38,
"high": 0.29, # TODO investigate
},
)
def test_bf16_20b(self):
self.run_test(
model_variant="20b",
quantization="bf16",
expected_score_of_reasoning_effort={
"low": 0.38,
"medium": 0.38,
"high": 0.29, # TODO investigate
},
)
if __name__ == "__main__":
unittest.main()

42
test/srt/test_gpt_oss_4gpu.py Normal file
View File

@@ -0,0 +1,42 @@
import unittest
from test_gpt_oss_common import BaseTestGptOss
class TestGptOss4Gpu(BaseTestGptOss):
def test_bf16_120b(self):
self.run_test(
model_variant="120b",
quantization="bf16",
expected_score_of_reasoning_effort={
"low": 0.61,
# remove to speed up
# "medium": 0.61,
# "high": 0.61,
},
other_args=["--tp", "4", "--cuda-graph-max-bs", "200"],
)
def test_mxfp4_120b(self):
self.run_test(
model_variant="120b",
quantization="mxfp4",
expected_score_of_reasoning_effort={
"low": 0.61,
# remove to speed up
# "medium": 0.61,
# "high": 0.61,
},
other_args=[
"--tp",
"4",
"--cuda-graph-max-bs",
"200",
"--mem-fraction-static",
"0.93",
],
)
if __name__ == "__main__":
unittest.main()

99
test/srt/test_gpt_oss_common.py Normal file
View File

@@ -0,0 +1,99 @@
from concurrent.futures import ThreadPoolExecutor
from types import SimpleNamespace
from typing import Dict, List, Literal, Optional
from sglang.srt.utils import kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
_base_url = DEFAULT_URL_FOR_TEST
class BaseTestGptOss(CustomTestCase):
def run_test(
self,
model_variant: Literal["20b", "120b"],
quantization: Literal["mxfp4", "bf16"],
expected_score_of_reasoning_effort: Dict[str, float],
other_args: Optional[List[str]] = None,
):
if other_args is None:
other_args = []
model = {
("20b", "bf16"): "lmsys/gpt-oss-20b-bf16",
("120b", "bf16"): "lmsys/gpt-oss-120b-bf16",
("20b", "mxfp4"): "openai/gpt-oss-20b",
("120b", "mxfp4"): "openai/gpt-oss-120b",
}[(model_variant, quantization)]
if model_variant == "20b":
other_args += ["--cuda-graph-max-bs", "600"]
self._run_test_raw(
model=model,
expected_score_of_reasoning_effort=expected_score_of_reasoning_effort,
other_args=other_args,
)
def _run_test_raw(
self,
model: str,
expected_score_of_reasoning_effort: Dict[str, float],
other_args: List[str],
):
process = popen_launch_server(
model,
_base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
try:
# run multiple tests in parallel since we are mostly bound by the longest generate sequence
# instead of the number of questions
with ThreadPoolExecutor(max_workers=4) as executor:
list(
executor.map(
lambda d: self._run_one_eval(**d),
[
dict(
model=model,
reasoning_effort=reasoning_effort,
expected_score=expected_score,
)
for reasoning_effort, expected_score in expected_score_of_reasoning_effort.items()
],
)
)
finally:
kill_process_tree(process.pid)
def _run_one_eval(self, model, reasoning_effort, expected_score):
args = SimpleNamespace(
base_url=_base_url,
model=model,
eval_name="gpqa",
num_examples=198,
# use enough threads to allow parallelism
num_threads=198,
# TODO 4k is still not enough, we need e.g. 64k token, but that is super slow
# otherwise a lot of questions are not answered
max_tokens=4096,
# simple-evals by default use 0.5 and is better than 0.0 temperature
# but here for reproducibility, we use 0.1
temperature=0.1,
reasoning_effort=reasoning_effort,
)
print(f"Evaluation start: {model=} {reasoning_effort=} {expected_score=}")
metrics = run_eval(args)
print(
f"Evaluation end: {model=} {reasoning_effort=} {expected_score=} {metrics=}"
)
self.assertGreaterEqual(metrics["score"], expected_score)