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

[NVIDIA] [GDN] Enable FlashInfer MTP verify on SM100+ (Blackwell) (#23273)

Browse Source 
Co-authored-by: Yangmin Li <yangminl@nvidia.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Shu Wang
2026-06-01 18:56:42 -07:00
committed by GitHub
co-authored by Yangmin Li Claude Opus 4.7
parent 54143264bf
commit 0574d2b8a5
4 changed files with 163 additions and 86 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/models_e2e/test_qwen35_fp4_mtp.py
+95 -57
View File
@@ -15,11 +15,74 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=340, stage="base-c", runner_config="4-gpu-b200")
register_cuda_ci(est_time=740, stage="base-c", runner_config="4-gpu-b200")
QWEN35_FP4_MODEL = "nvidia/Qwen3.5-397B-A17B-NVFP4"
ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}}
MTP_BASE_ARGS = [
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
"--mamba-ssm-dtype",
"bfloat16",
"--max-running-requests",
"128",
"--reasoning-parser",
"qwen3",
"--attention-backend",
"trtllm_mha",
"--quantization",
"modelopt_fp4",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
]
def _run_mtp_gsm8k(test_case):
args = SimpleNamespace(
model=test_case.model,
eval_name="gsm8k",
num_shots=5,
num_examples=200,
max_tokens=16000,
num_threads=128,
repeat=1,
temperature=0.6,
top_p=0.95,
top_k=20,
base_url=test_case.base_url,
host="http://127.0.0.1",
port=int(test_case.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
test_case.assertGreaterEqual(
metrics["score"], ACC_THRESHOLDS[test_case.model]["gsm8k"]
)
server_info = requests.get(test_case.base_url + "/server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
test_case.assertGreater(avg_spec_accept_length, 3.3)
class TestQwen35FP4MTP(ReasoningTokenUsageMixin, CustomTestCase):
reasoning_parser_name = "qwen3"
@@ -34,37 +97,36 @@ class TestQwen35FP4MTP(ReasoningTokenUsageMixin, CustomTestCase):
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",
"--mamba-ssm-dtype",
"bfloat16",
"--max-running-requests",
"128",
"--reasoning-parser",
"qwen3",
"--attention-backend",
"trtllm_mha",
"--quantization",
"modelopt_fp4",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
other_args=MTP_BASE_ARGS,
)
@classmethod
def tearDownClass(cls):
envs.SGLANG_ENABLE_SPEC_V2.set(False)
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
_run_mtp_gsm8k(self)
class TestQwen35FP4MTPFlashInfer(ReasoningTokenUsageMixin, CustomTestCase):
reasoning_parser_name = "qwen3"
@classmethod
def setUpClass(cls):
cls.model = QWEN35_FP4_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.init_reasoning_token_verifier()
envs.SGLANG_ENABLE_SPEC_V2.set(True)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=MTP_BASE_ARGS
+ [
"--linear-attn-decode-backend",
"flashinfer",
"--enforce-disable-flashinfer-allreduce-fusion",
],
)
@@ -74,31 +136,7 @@ class TestQwen35FP4MTP(ReasoningTokenUsageMixin, CustomTestCase):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
model=self.model,
eval_name="gsm8k",
num_shots=5,
num_examples=200,
max_tokens=16000,
num_threads=128,
repeat=1,
temperature=0.6,
top_p=0.95,
top_k=20,
base_url=self.base_url,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(metrics["score"], ACC_THRESHOLDS[self.model]["gsm8k"])
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, 3.3)
_run_mtp_gsm8k(self)
if __name__ == "__main__":