minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 14 Packages Projects Releases Wiki Activity

[CI] Merge tokenizer worker tests and drop redundant triton attention e2e (#33641)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-08-05 11:55:11 -07:00
committed by GitHub
parent 5f79cf3511
commit 1a045669e4
8 changed files with 133 additions and 413 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/mla/test_flashmla.py
-93
View File
@@ -1,93 +0,0 @@
"""
Usage:
python3 test/registered/mla/test_flashmla.py
"""
import unittest
from types import SimpleNamespace
import requests
import torch
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,
)
# FlashMLA attention backend tests with MTP speculative decoding
register_cuda_ci(est_time=160, stage="base-b", runner_config="1-gpu-large")
class TestFlashMLAMTP(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(
[
"--cuda-graph-max-bs-decode",
"4",
"--disable-radix",
"--enable-torch-compile",
"--torch-compile-max-bs",
"1",
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
"lmsys/sglang-ci-dsv3-test-NextN",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"3",
"--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):
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").json()
avg_spec_accept_length = server_info["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 2.4)
if __name__ == "__main__":
unittest.main()
test/registered/mla/test_mla_flashinfer.py
-84
View File
@@ -1,84 +0,0 @@
import unittest
from types import SimpleNamespace
import requests
import torch
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,
)
# FlashInfer MLA backend tests with MTP speculative decoding
register_cuda_ci(est_time=130, stage="base-b", runner_config="1-gpu-large")
class TestFlashinferMLAMTP(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(
[
"--cuda-graph-max-bs-decode",
"4",
"--enable-torch-compile",
"--torch-compile-max-bs",
"1",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"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):
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").json()
avg_spec_accept_length = server_info["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_int8_deepseek_v3.py
+2 -63
View File
@@ -11,11 +11,11 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
)
# DeepSeek-V3 INT8 quantization tests (channel and block INT8)
# DeepSeek-V3 channel-INT8 + MTP smoke; int8 GEMM numerics live in
# unit/layers/quantization/test_int8_linear_methods.py
register_cuda_ci(est_time=160, stage="base-b", runner_config="1-gpu-large")
@@ -81,66 +81,5 @@ class TestDeepseekV3MTPChannelInt8(CustomTestCase):
self.assertGreater(avg_spec_accept_length, 2.5)
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
class TestDeepseekV3MTPBlockInt8(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-decode",
"16",
"--enable-torch-compile",
"--torch-compile-max-bs",
"2",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"4",
]
)
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):
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()