chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

[Spec V2] Support specV2 for mamba hybrid attention (#18808)

Browse Source 
Co-authored-by: Yi Zhong <207368749+vincentzed@users.noreply.github.com>
Co-authored-by: yizhang2077 <1109276519@qq.com>
Co-authored-by: Hanming Lu <hanming@x.ai>
This commit is contained in:
zhangheng
2026-02-27 00:36:01 +08:00
committed by GitHub
parent 78d6674c45
commit e4b708d3e9
8 changed files with 205 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

84
test/registered/4-gpu-models/test_qwen3_next_models_mtp.py
View File

@@ -3,6 +3,7 @@ from types import SimpleNamespace
import requests
from sglang.srt.environ import envs
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
@@ -211,5 +212,88 @@ class TestQwen3NextMTPTopk(CustomTestCase):
print("test_prefix_cache_branching passed")
class TestQwen3NextMTPV2(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
envs.SGLANG_ENABLE_SPEC_V2.set(True)
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",
"extra_buffer",
"--mamba-track-interval",
"128",
],
)
@classmethod
def tearDownClass(cls):
envs.SGLANG_ENABLE_SPEC_V2.set(False)
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"]
)
# TODO(hzh): After merging the PR that fixes specv2 to correctly return log probs, re-open the tests below. https://github.com/sgl-project/sglang/pull/18645
# def test_input_output_logprobs_match(self):
# test_input_output_logprobs_match_helper(
# self.base_url,
# ACC_THRESHOLDS,
# self.model,
# max_samples=32,
# max_new_tokens=512,
# )
# 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,
# )
if __name__ == "__main__":
unittest.main()