EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[new feat] ascend backend support fia fusion kernel (#8328)

Browse Source 
Co-authored-by: Even Zhou <even.y.zhou@outlook.com>
This commit is contained in:
ZhengdQin
2025-08-26 14:13:08 +08:00
committed by GitHub
parent e2e378caba
commit f92b729d52
9 changed files with 509 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

103
test/srt/ascend/test_ascend_mla_fia_w8a8int8.py Normal file
View File

@@ -0,0 +1,103 @@
import os
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
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,
run_bench_offline_throughput,
)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V2-Lite-W8A8": {
"accuracy": 0.34,
"latency": 1000,
"output_throughput": 6,
},
}
class TestAscendMlaW8A8Int8(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.models = TEST_MODEL_MATRIX.keys()
cls.base_url = DEFAULT_URL_FOR_TEST
cls.url = urlparse(DEFAULT_URL_FOR_TEST)
cls.common_args = [
"--trust-remote-code",
"--disable-cuda-graph",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--quantization",
"w8a8_int8",
"--tp-size",
2,
"--disable-radix-cache",
]
def test_a_gsm8k(self):
os.environ["ASCEND_USE_FIA"] = "true"
for model in self.models:
with self.subTest(model=model):
print(f"##=== Testing accuracy: {model} ===##")
process = popen_launch_server(
model,
self.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
*self.common_args,
],
)
try:
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=1319,
max_new_tokens=512,
parallel=128,
host=f"http://{self.url.hostname}",
port=int(self.url.port),
)
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
finally:
kill_process_tree(process.pid)
def test_b_throughput(self):
for model in self.models:
with self.subTest(model=model):
print(f"##=== Testing throughput: {model} ===##")
output_throughput = run_bench_offline_throughput(
model,
[
*self.common_args,
],
)
print(f"##=== {model} throughput: {output_throughput} ===##")
if is_in_ci():
self.assertGreater(
output_throughput,
TEST_MODEL_MATRIX[model]["output_throughput"],
)
if __name__ == "__main__":
unittest.main()

1
test/srt/ascend/test_ascend_mla_w8a8int8.py
View File

@@ -40,6 +40,7 @@ class TestAscendMlaW8A8Int8(CustomTestCase):
"w8a8_int8",
"--tp-size",
4,
"--disable-radix-cache",
]
def test_a_gsm8k(self):

101
test/srt/ascend/test_ascend_tp2_fia_bf16.py Normal file
View File

@@ -0,0 +1,101 @@
import os
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
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,
run_bench_offline_throughput,
)
TEST_MODEL_MATRIX = {
"Qwen/Qwen2.5-7B-Instruct": {
"accuracy": 0.85,
"latency": 180,
"output_throughput": 20,
},
}
class TestAscendTp2Bf16(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.models = TEST_MODEL_MATRIX.keys()
cls.base_url = DEFAULT_URL_FOR_TEST
cls.url = urlparse(DEFAULT_URL_FOR_TEST)
cls.common_args = [
"--trust-remote-code",
"--disable-cuda-graph",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--tp-size",
2,
"--disable-radix-cache",
]
def test_a_gsm8k(self):
os.environ["ASCEND_USE_FIA"] = "true"
for model in self.models:
with self.subTest(model=model):
print(f"##=== Testing accuracy: {model} ===##")
process = popen_launch_server(
model,
self.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
*self.common_args,
],
)
try:
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=1319,
max_new_tokens=512,
parallel=128,
host=f"http://{self.url.hostname}",
port=int(self.url.port),
)
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
finally:
kill_process_tree(process.pid)
def test_b_throughput(self):
for model in self.models:
with self.subTest(model=model):
print(f"##=== Testing throughput: {model} ===##")
output_throughput = run_bench_offline_throughput(
model,
[
*self.common_args,
],
)
print(f"##=== {model} throughput: {output_throughput} ===##")
if is_in_ci():
self.assertGreater(
output_throughput,
TEST_MODEL_MATRIX[model]["output_throughput"],
)
if __name__ == "__main__":
unittest.main()

2
test/srt/run_suite.py
View File

@@ -275,6 +275,8 @@ suite_ascend = {
"per-commit-2-ascend-npu": [
TestFile("ascend/test_ascend_tp2_bf16.py", 400),
TestFile("ascend/test_ascend_graph_tp2_bf16.py", 400),
TestFile("ascend/test_ascend_tp2_fia_bf16.py", 400),
TestFile("ascend/test_ascend_mla_fia_w8a8int8.py", 400),
],
"per-commit-4-ascend-npu": [
TestFile("ascend/test_ascend_mla_w8a8int8.py", 400),