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

[NVIDIA] Integrate FlashInfer decode kernel (Blackwell) for Qwen3.5 (#19150)

Browse Source 
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Kaixi Hou
2026-03-17 22:11:18 -07:00
committed by GitHub
parent da24478913
commit 4cc19862ef
5 changed files with 161 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

108
test/registered/4-gpu-models/test_qwen35_models.py
View File

@@ -7,15 +7,18 @@ import requests
from sglang.srt.environ import envs
from sglang.srt.utils import kill_process_tree
from sglang.test.accuracy_test_runner import AccuracyTestParams
from sglang.test.ci.ci_register import register_cuda_ci
# This eval harness applies the chat_template, which is critical for qwen3.5
# to get good accuracy on gsm8k
from sglang.test.run_combined_tests import run_combined_tests
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
ModelLaunchSettings,
popen_launch_server,
)
@@ -26,62 +29,59 @@ QWEN35_27B_MODEL = "Qwen/Qwen3.5-27B"
ACC_THRESHOLDS = {QWEN35_FP4_MODEL: {"gsm8k": 0.95}, QWEN35_27B_MODEL: {"gsm8k": 0.8}}
class TestQwen35FP4(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN35_FP4_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",
"--mamba-ssm-dtype",
"bfloat16",
"--max-running-requests",
"128",
"--reasoning-parser",
"qwen3",
"--attention-backend",
"trtllm_mha",
"--quantization",
"modelopt_fp4",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
class TestQwen35FP4(unittest.TestCase):
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]),
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",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
]
variants = [
ModelLaunchSettings(
QWEN35_FP4_MODEL,
extra_args=base_args,
variant="Triton",
),
ModelLaunchSettings(
QWEN35_FP4_MODEL,
extra_args=base_args + ["--linear-attn-decode-backend", "flashinfer"],
variant="FlashInfer",
),
]
run_combined_tests(
models=variants,
test_name="Qwen3.5-397B-A17B-NVFP4",
accuracy_params=AccuracyTestParams(
dataset="gsm8k",
baseline_accuracy=ACC_THRESHOLDS[QWEN35_FP4_MODEL]["gsm8k"],
num_examples=200,
num_threads=128,
max_tokens=16000,
thinking_mode="qwen3",
temperature=0.6,
top_p=0.95,
top_k=20,
),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(metrics["score"], ACC_THRESHOLDS[self.model]["gsm8k"])
class TestQwen35FP4MTP(CustomTestCase):