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

ci: prune per-commit CUDA tests — move 25 files + 13 testcases to test/manual/ (#24721)

Browse Source
This commit is contained in:
Alison Shao
2026-05-08 15:53:23 -07:00
committed by GitHub
parent aefd8e257f
commit 5fbec0e445
45 changed files with 818 additions and 525 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/mla/test_flashmla.py
+1 -45
View File
@@ -14,7 +14,6 @@ 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_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -22,50 +21,7 @@ from sglang.test.test_utils import (
)
# FlashMLA attention backend tests with MTP speculative decoding
register_cuda_ci(est_time=314, suite="stage-b-test-1-gpu-large")
class TestFlashMLAAttnBackend(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"2",
"--attention-backend",
"flashmla",
]
)
# Use longer timeout for DeepGEMM JIT compilation which can take 10-20 minutes
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 2,
other_args=other_args,
)
@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",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
register_cuda_ci(est_time=160, suite="stage-b-test-1-gpu-large")
class TestFlashMLAMTP(CustomTestCase):
test/registered/mla/test_mla.py
-46
View File
@@ -1,46 +0,0 @@
import unittest
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import MGSMEnMixin
from sglang.test.test_utils import (
DEFAULT_MLA_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
# MLA attention test with MGSM evaluation
register_cuda_ci(est_time=181, suite="stage-b-test-1-gpu-large")
register_amd_ci(est_time=1100, suite="stage-b-test-1-gpu-small-amd")
class TestMLA(CustomTestCase, MGSMEnMixin):
mgsm_en_score_threshold = 0.8
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
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=[
"--trust-remote-code",
"--enable-torch-compile",
"--torch-compile-max-bs",
"4",
"--chunked-prefill-size",
"256",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
if __name__ == "__main__":
unittest.main()
test/registered/mla/test_mla_deepseek_v3.py
-211
View File
@@ -1,211 +0,0 @@
import os
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import is_cuda, is_hip, kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci, 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,
is_in_ci,
popen_launch_server,
)
# DeepSeek-V3 MLA tests with torch compile, FA3, and MTP speculative decoding
register_cuda_ci(est_time=543, suite="stage-b-test-1-gpu-large")
register_amd_ci(
est_time=221,
suite="stage-b-test-1-gpu-small-amd",
disabled="see https://github.com/sgl-project/sglang/issues/12574",
)
class TestMLADeepseekV3(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code", "--chunked-prefill-size", "256"]
if is_cuda():
other_args.extend(["--enable-torch-compile", "--cuda-graph-max-bs", "2"])
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
class TestMLADeepseekV3DisableFusedFunc(CustomTestCase):
@classmethod
def setUpClass(cls):
os.environ["SGLANG_CI_DISABLE_MOE_FUSED_FUNC"] = "1"
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code", "--chunked-prefill-size", "256"]
if is_cuda():
other_args.extend(["--cuda-graph-max-bs", "2"])
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.62)
@unittest.skipIf(is_hip(), "FA is not available.")
class TestMLADeepseekV3Fa3Fp8Kvcache(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--chunked-prefill-size",
"256",
"--kv-cache-dtype",
"fp8_e4m3",
]
if is_cuda():
other_args.extend(
[
"--attention-backend",
"fa3",
"--mem-fraction-static",
"0.8",
"--cuda-graph-max-bs",
"2",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
class TestDeepseekV3MTP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--cuda-graph-max-bs",
"2",
"--disable-radix",
"--enable-torch-compile",
"--torch-compile-max-bs",
"1",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"4",
]
# This test runs first (alphabetically) and needs longer timeout for
# DeepGEMM JIT compilation which is required for DeepSeek-V3's FP8 MoE layers
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 2,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
requests.get(self.base_url + "/flush_cache")
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.60)
server_info = requests.get(self.base_url + "/server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 2.5)
if __name__ == "__main__":
unittest.main()
test/registered/mla/test_mla_flashinfer.py
+1 -44
View File
@@ -15,50 +15,7 @@ from sglang.test.test_utils import (
)
# FlashInfer MLA backend tests with MTP speculative decoding
register_cuda_ci(est_time=260, suite="stage-b-test-1-gpu-large")
class TestFlashinferMLA(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--enable-torch-compile",
"--cuda-graph-max-bs",
"4",
"--attention-backend",
"flashinfer",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.615)
register_cuda_ci(est_time=130, suite="stage-b-test-1-gpu-large")
class TestFlashinferMLAMTP(CustomTestCase):
test/registered/mla/test_mla_int8_deepseek_v3.py
+1 -88
View File
@@ -16,53 +16,9 @@ from sglang.test.test_utils import (
)
# DeepSeek-V3 INT8 quantization tests (channel and block INT8)
register_cuda_ci(est_time=313, suite="stage-b-test-1-gpu-large")
register_cuda_ci(est_time=160, suite="stage-b-test-1-gpu-large")
class TestMLADeepseekV3ChannelInt8(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-channel-int8-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"16",
"--enable-torch-compile",
"--torch-compile-max-bs",
"2",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreaterEqual(metrics["score"], 0.61)
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
class TestDeepseekV3MTPChannelInt8(CustomTestCase):
@classmethod
def setUpClass(cls):
@@ -126,49 +82,6 @@ class TestDeepseekV3MTPChannelInt8(CustomTestCase):
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
class TestMLADeepseekV3BlockInt8(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-block-int8-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"16",
"--enable-torch-compile",
"--torch-compile-max-bs",
"2",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(metrics)
self.assertGreater(metrics["score"], 0.62)
class TestDeepseekV3MTPBlockInt8(CustomTestCase):
@classmethod
def setUpClass(cls):