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

Migrate 4-GPU/8-GPU workflow jobs to stage-c and add CI registry decorators (#17299)

Browse Source
This commit is contained in:
Alison Shao
2026-01-31 22:37:22 -08:00
committed by GitHub
parent 95180484e9
commit a0bae4c343
39 changed files with 189 additions and 284 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/models/test_dummy_grok_models.py
-34
View File
@@ -1,34 +0,0 @@
import unittest
from sglang.test.test_utils import CustomTestCase, is_in_ci, run_bench_one_batch
class TestDummyGrok1(CustomTestCase):
def test_dummy_grok_1(self):
_, output_throughput, _ = run_bench_one_batch(
None,
[
"--model",
"/dummy-grok",
"--tokenizer-path",
"Xenova/grok-1-tokenizer",
"--batch-size",
"2",
"--tp",
"2",
"--quantization",
"fp8",
"--load-format",
"dummy",
"--json-model-override-args",
'{"num_hidden_layers": 2}',
],
)
if is_in_ci():
self.assertGreater(output_throughput, 0)
if __name__ == "__main__":
unittest.main()
test/srt/models/test_kimi_k2_models.py
-70
View File
@@ -1,70 +0,0 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
class TestKimiK2Thinking(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "moonshotai/Kimi-K2-Thinking"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--tp",
"8",
"--trust-remote-code",
"--tool-call-parser",
"kimi_k2",
"--reasoning-parser",
"kimi_k2",
"--model-loader-extra-config",
'{"enable_multithread_load": true, "num_threads": 64}',
]
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_a_gsm8k(
self,
):
requests.get(self.base_url + "/flush_cache")
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_few_shot_gsm8k(args)
print(f"{metrics=}")
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (Kimi-K2-Thinking)\n" f'{metrics["accuracy"]=:.3f}\n'
)
self.assertGreater(metrics["accuracy"], 0.95)
if __name__ == "__main__":
unittest.main()
test/srt/models/test_mimo_models.py
-47
View File
@@ -1,47 +0,0 @@
import unittest
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.kits.spec_decoding_kit import SpecDecodingMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
class TestMiMoV2Flash(GSM8KMixin, SpecDecodingMixin, DefaultServerBase):
gsm8k_accuracy_thres = 0.75
gsm8k_num_questions = 1319
gsm8k_parallel = 1319
model = "XiaomiMiMo/MiMo-V2-Flash"
other_args = [
"--tp",
"4",
"--dp",
"2",
"--enable-dp-attention",
"--trust-remote-code",
"--attention-backend",
"fa3",
"--max-running-requests",
"128",
"--cuda-graph-max-bs",
"64",
"--mem-fraction-static",
"0.75",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--enable-multi-layer-eagle",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
]
bs_1_speed_thres = 170
accept_length_thres = 3.2
if __name__ == "__main__":
unittest.main()
test/srt/models/test_ministral3_models.py
-26
View File
@@ -1,26 +0,0 @@
import unittest
from sglang.test.kits.gsm8k_accuracy_kit import GSM8KMixin
from sglang.test.kits.mmmu_vlm_kit import MMMUMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
from sglang.test.server_fixtures.mmmu_fixture import MMMUServerBase
MODEL = "mistralai/Ministral-3-3B-Instruct-2512"
class TestMinistral3TextOnly(GSM8KMixin, DefaultServerBase):
gsm8k_accuracy_thres = 0.6
model = MODEL
other_args = ["--trust-remote-code"]
class TestMinistral3MMMU(MMMUMixin, MMMUServerBase):
accuracy = 0.3
model = MODEL
other_args = ["--trust-remote-code"]
mmmu_args = ["--limit=0.1"]
"""`--limit=0.1`: 10 percent of each task - this is fine for testing since the nominal result isn't interesting - this run is just to prevent relative regressions."""
if __name__ == "__main__":
unittest.main()
test/srt/models/test_qwen3_next_models.py
-127
View File
@@ -1,127 +0,0 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.kl_test_utils import (
test_input_output_logprobs_match_decode_cache_hit_helper,
test_input_output_logprobs_match_prefill_cache_hit_helper,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
def send_request_helper(base_url: str, text: str):
response = requests.post(
base_url + "/generate",
json={
"text": text,
"sampling_params": {
"max_new_tokens": 1,
},
},
)
return response.json()
class TestQwen3Next(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-size",
"4",
"--chunked-prefill-size",
"2048",
"--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"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_prefix_cache_branching(self):
print("running test_prefix_cache_branching")
requests.get(self.base_url + "/flush_cache")
branching_pos = 257
text_prefix = "hi" * branching_pos
suffix_list = ["this" * 256, "here" * 256, "that" * 256]
cache_hit_list = [False, False, True]
# First request only prefill the entire sequence
# Second request won't have cache hit, but will cache the branching point
# Third request will have cache hit on the branching point
for i, (suffix, cache_hit) in enumerate(
zip(suffix_list, cache_hit_list, strict=True)
):
result = send_request_helper(self.base_url, text_prefix + suffix)
cached_tokens = result["meta_info"]["cached_tokens"]
if cache_hit:
expected_cached_tokens = branching_pos // 64 * 64
assert (
cached_tokens == expected_cached_tokens
), f"{i=}, {cache_hit=}, {cached_tokens=} is not equal to {expected_cached_tokens=}, {branching_pos=}"
else:
assert (
cached_tokens == 0
), f"{i=}, {cache_hit=}, {cached_tokens=} is not 0"
print("test_prefix_cache_branching passed")
if __name__ == "__main__":
unittest.main()
test/srt/models/test_qwen3_next_models_mtp.py
-212
View File
@@ -1,212 +0,0 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.kl_test_utils import (
test_input_output_logprobs_match_decode_cache_hit_helper,
test_input_output_logprobs_match_prefill_cache_hit_helper,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
# MTP has higher KL divergence threshold
ACC_THRESHOLDS_MTP = {
QWEN3_NEXT_MODEL: {"kl_div": 0.008, "gsm8k": 0.93},
}
def send_request_helper(base_url: str, text: str):
response = requests.post(
base_url + "/generate",
json={
"text": text,
"sampling_params": {
"max_new_tokens": 1,
},
},
)
return response.json()
class TestQwen3NextMTP(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=[
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"no_buffer",
],
)
@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"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
class TestQwen3NextMTPTopk(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=[
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"8",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--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_MTP[self.model]["gsm8k"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS_MTP,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS_MTP,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_prefix_cache_branching(self):
print("running test_prefix_cache_branching")
requests.get(self.base_url + "/flush_cache")
branching_pos = 257
text_prefix = "hi" * branching_pos
suffix_list = ["this" * 256, "here" * 256, "that" * 256]
cache_hit_list = [False, False, True]
# First request only prefill the entire sequence
# Second request won't have cache hit, but will cache the branching point
# Third request will have cache hit on the branching point
for i, (suffix, cache_hit) in enumerate(
zip(suffix_list, cache_hit_list, strict=True)
):
result = send_request_helper(self.base_url, text_prefix + suffix)
cached_tokens = result["meta_info"]["cached_tokens"]
if cache_hit:
expected_cached_tokens = branching_pos // 64 * 64
assert (
cached_tokens == expected_cached_tokens
), f"{i=}, {cache_hit=}, {cached_tokens=} is not equal to {expected_cached_tokens=}, {branching_pos=}"
else:
assert (
cached_tokens == 0
), f"{i=}, {cache_hit=}, {cached_tokens=} is not 0"
print("test_prefix_cache_branching passed")
if __name__ == "__main__":
unittest.main()
test/srt/models/test_qwen3_next_models_pcg.py
-70
View File
@@ -1,70 +0,0 @@
"""
Qwen3 Next piecewise CUDA graph tests.
DISABLED: See https://github.com/sgl-project/sglang/issues/17039
PCG tests for Qwen3 Next have intermittent failures (5-10% probability).
Investigation ongoing by @YuweiAn.
"""
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
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,
)
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
@unittest.skip("Disabled: intermittent failures, see #17039")
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",
"--enable-piecewise-cuda-graph",
"--piecewise-cuda-graph-compiler",
"eager",
],
)
@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"]
)
if __name__ == "__main__":
unittest.main()