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

[CI] Trim redundant nightly test registrations (#34070)

Browse Source 
Co-authored-by: Baizhou Zhang <sobereddiezhang@gmail.com>
This commit is contained in:
Liangsheng Yin
2026-08-08 01:42:46 -07:00
committed by GitHub
co-authored by Baizhou Zhang
parent dd5d82bead
commit f6a6f5bf1e
24 changed files with 2 additions and 3571 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/cuda_graph/piecewise/test_pcg_glm52_fp4.py
-71
View File
@@ -1,71 +0,0 @@
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.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
register_cuda_ci(est_time=900, suite="nightly-4-gpu-b200", nightly=True)
GLM52_FP4_MODEL = "nvidia/GLM-5.2-NVFP4"
class TestPCGGlm52Fp4(CustomTestCase):
"""PCG prefill on GLM-5.2-NVFP4 (DSA model, TP=4, B200).
GLM-5.2 uses GlmMoeDsaForCausalLM (DSA attention). This test verifies that
piecewise CUDA graph works correctly after the DSA indexer was updated to
cache k_fp8/k_scale for PCG-compatible prefill.
"""
@classmethod
def setUpClass(cls):
cls.model = GLM52_FP4_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-size",
"4",
"--trust-remote-code",
"--reasoning-parser",
"glm45",
"--tool-call-parser",
"glm47",
"--quantization",
"modelopt_fp4",
"--disable-flashinfer-autotune",
"--cuda-graph-backend-prefill=tc_piecewise",
"--model-loader-extra-config",
'{"enable_multithread_load": true, "num_threads": 64}',
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
num_examples=200,
num_threads=200,
max_tokens=4096,
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreater(metrics["score"], 0.92)
if __name__ == "__main__":
unittest.main()
test/registered/cuda_graph/piecewise/test_pcg_glm52_fp8_tp8.py
-75
View File
@@ -1,75 +0,0 @@
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.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
register_cuda_ci(est_time=900, suite="nightly-8-gpu-h200", nightly=True)
GLM52_FP8_MODEL = "zai-org/GLM-5.2-FP8"
class TestBCGGlm52Fp8TP8(CustomTestCase):
"""Breakable CUDA graph prefill on GLM-5.2-FP8 (DSA model, TP=8, H200)."""
@classmethod
def setUpClass(cls):
cls.model = GLM52_FP8_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-size",
"8",
"--trust-remote-code",
"--reasoning-parser",
"glm45",
"--tool-call-parser",
"glm47",
"--mem-fraction-static",
"0.8",
"--disable-flashinfer-autotune",
"--cuda-graph-backend-prefill=breakable",
# Small chunks => many prefill iterations, each <= the 2048
# capture max, so every prefill batch replays the BCG graph and
# exercises the DSA split-op / dual-stream / MLA-fusion paths.
"--chunked-prefill-size",
"512",
"--model-loader-extra-config",
'{"enable_multithread_load": true, "num_threads": 64}',
],
env={
"SGLANG_ENABLE_PCG_DSV2_DUAL_STREAM": "1",
},
)
@classmethod
def tearDownClass(cls):
if hasattr(cls, "process") and cls.process:
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
num_examples=200,
num_threads=200,
max_tokens=4096,
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreater(metrics["score"], 0.92)
if __name__ == "__main__":
unittest.main()