minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[Test] Refactor KL divergence and prefix cache branching to kits (#19715)

Browse Source
This commit is contained in:
roikoren755
2026-03-12 16:11:59 +08:00
committed by GitHub
parent 46b558445d
commit 067353f67b
8 changed files with 244 additions and 552 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/models/test_qwen3_next_models_fp4.py
+18 -55
View File
@@ -1,70 +1,33 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import get_device_sm, kill_process_tree
from sglang.srt.utils import get_device_sm
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
register_cuda_ci(est_time=500, suite="nightly-4-gpu-b200", nightly=True)
QWEN3_NEXT_MODEL_FP4 = "nvidia/Qwen3-Next-80B-A3B-Instruct-NVFP4"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL_FP4: {"kl_div": 0.0025, "gsm8k": 0.93},
}
@unittest.skipIf(
get_device_sm() < 100, "Test requires CUDA SM 100 or higher (Blackwell)"
)
class TestQwen3NextFp4(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL_FP4
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--quantization",
"modelopt_fp4",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
class TestQwen3NextFp4(GSM8KMixin, DefaultServerBase):
model = QWEN3_NEXT_MODEL_FP4
gsm8k_accuracy_thres = 0.93
other_args = [
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--quantization",
"modelopt_fp4",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
]
if __name__ == "__main__":
test/registered/models/test_qwen3_next_models_pcg.py
+9 -48
View File
@@ -3,17 +3,10 @@ Qwen3 Next piecewise CUDA graph tests.
"""
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
register_cuda_ci(
est_time=400,
@@ -22,46 +15,14 @@ register_cuda_ci(
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
class TestQwen3NextPiecewiseCudaGraph(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--tp",
"4",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
class TestQwen3NextPiecewiseCudaGraph(GSM8KMixin, DefaultServerBase):
model = QWEN3_NEXT_MODEL
gsm8k_accuracy_thres = 0.93
other_args = [
"--tp",
"4",
]
if __name__ == "__main__":