[NPU]Add a full test pipeline on NPU, resolve issues in the NPU test architecture (#20751)

This commit is contained in:
Cherry_ming
2026-04-01 19:56:31 +08:00
committed by GitHub
parent ac039bd04e
commit e67b95d66b
43 changed files with 673 additions and 106 deletions
@@ -0,0 +1,81 @@
import logging
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.test_ascend_utils import QWEN3_8B_INT4_AUTOROUND_WEIGHTS_PATH
from sglang.test.ci.ci_register import register_npu_ci
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,
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a2", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
TEST_MODEL_MATRIX = {
QWEN3_8B_INT4_AUTOROUND_WEIGHTS_PATH: {
"accuracy": 0.85,
},
}
class TestAscendAutoRoundDense(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",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--quantization",
"auto-round",
]
def test_a_gsm8k(self):
for model in self.models:
with self.subTest(model=model):
logger.info(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)
if __name__ == "__main__":
unittest.main()
@@ -0,0 +1,83 @@
import logging
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.test_ascend_utils import (
QWEN3_30B_A3B_INSTRUCT_2507_INT4_AUTOROUND_WEIGHTS_PATH,
)
from sglang.test.ci.ci_register import register_npu_ci
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,
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a2", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
TEST_MODEL_MATRIX = {
QWEN3_30B_A3B_INSTRUCT_2507_INT4_AUTOROUND_WEIGHTS_PATH: {
"accuracy": 0.85,
},
}
class TestAscendAutoRoundMoE(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",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--quantization",
"auto-round",
]
def test_a_gsm8k(self):
for model in self.models:
with self.subTest(model=model):
logger.info(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)
if __name__ == "__main__":
unittest.main()
@@ -0,0 +1,83 @@
import logging
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.test_ascend_utils import (
QWEN3_30B_A3B_GPTQ_2507_INT4_WEIGHTS_PATH,
)
from sglang.test.ci.ci_register import register_npu_ci
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,
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a2", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
TEST_MODEL_MATRIX = {
QWEN3_30B_A3B_GPTQ_2507_INT4_WEIGHTS_PATH: {
"accuracy": 0.85,
},
}
class TestAscendGPTQMoEInt4(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",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--quantization",
"gptq",
]
def test_a_gsm8k(self):
for model in self.models:
with self.subTest(model=model):
logger.info(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)
if __name__ == "__main__":
unittest.main()
@@ -0,0 +1,112 @@
"""
Usage:
python3 -m unittest test_ascend_w4a4_quantization.TestAscendW4A4.test_gsm8k
"""
import os
import time
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_npu_ci
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,
is_in_ci,
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-4-npu-a3", nightly=True)
if "ASCEND_RT_VISIBLE_DEVICES" not in os.environ:
os.environ["ASCEND_RT_VISIBLE_DEVICES"] = "0,1,2,3"
DEFAULT_PORT_FOR_SRT_TEST_RUNNER = (
7000 + int(os.environ.get("ASCEND_RT_VISIBLE_DEVICES", "0")[0]) * 100
)
DEFAULT_URL_FOR_TEST = f"http://127.0.0.1:{DEFAULT_PORT_FOR_SRT_TEST_RUNNER + 1000}"
class TestAscendW4A4(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "/root/.cache/modelscope/hub/models/Eco-Tech/Qwen3-32B-w4a4-LAOS"
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",
"--device",
"npu",
"--attention-backend",
"ascend",
"--tp-size",
"4",
"--mem-fraction-static",
"0.8",
"--cuda-graph-bs",
"64",
"--disable-radix-cache",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
base_url = DEFAULT_URL_FOR_TEST
url = urlparse(base_url)
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=1319,
max_new_tokens=512,
parallel=64,
host=f"http://{url.hostname}",
port=int(url.port),
)
metrics = run_eval(args)
print(metrics)
self.assertGreaterEqual(metrics["accuracy"], 0.80)
self.assertGreaterEqual(metrics["output_throughput"], 1000)
def run_decode(self, max_new_tokens):
response = requests.post(
self.base_url + "/generate",
json={
"text": "The capital of France is",
"sampling_params": {
"temperature": 0,
"max_new_tokens": max_new_tokens,
},
"ignore_eos": True,
},
)
return response.json()
def test_throughput(self):
max_tokens = 256
tic = time.perf_counter()
res = self.run_decode(max_tokens)
tok = time.perf_counter()
print(res["text"])
throughput = max_tokens / (tok - tic)
print(f"Throughput: {throughput} tokens/s")
if is_in_ci():
self.assertGreaterEqual(throughput, 35)
if __name__ == "__main__":
unittest.main()
@@ -0,0 +1,107 @@
"""
Usage:
python3 -m unittest test_ascend_w8a8_quantization.TestAscendW8A8.test_gsm8k
"""
import os
import time
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_npu_ci
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,
is_in_ci,
popen_launch_server,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a2", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
if "ASCEND_RT_VISIBLE_DEVICES" not in os.environ:
os.environ["ASCEND_RT_VISIBLE_DEVICES"] = "0,1"
DEFAULT_PORT_FOR_SRT_TEST_RUNNER = (
7000 + int(os.environ.get("ASCEND_RT_VISIBLE_DEVICES", "0")[0]) * 100
)
DEFAULT_URL_FOR_TEST = f"http://127.0.0.1:{DEFAULT_PORT_FOR_SRT_TEST_RUNNER + 1000}"
class TestAscendW8A8CompressedTensors(CustomTestCase):
@classmethod
def setUpClass(cls):
# TODO: Move model to CI or Modelscope
cls.model = "RedHatAI/Qwen2.5-0.5B-Instruct-quantized.w8a8"
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",
"--disable-cuda-graph",
"--device",
"npu",
"--attention-backend",
"ascend",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
base_url = DEFAULT_URL_FOR_TEST
url = urlparse(base_url)
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host=f"http://{url.hostname}",
port=int(url.port),
)
metrics = run_eval(args)
print(metrics)
self.assertGreaterEqual(metrics["accuracy"], 0.3)
self.assertGreaterEqual(metrics["output_throughput"], 700)
def run_decode(self, max_new_tokens):
response = requests.post(
self.base_url + "/generate",
json={
"text": "The capital of France is",
"sampling_params": {
"temperature": 0,
"max_new_tokens": max_new_tokens,
},
"ignore_eos": True,
},
)
return response.json()
def test_throughput(self):
max_tokens = 256
tic = time.perf_counter()
res = self.run_decode(max_tokens)
tok = time.perf_counter()
print(res["text"])
throughput = max_tokens / (tok - tic)
print(f"Throughput: {throughput} tokens/s")
if is_in_ci():
self.assertGreaterEqual(throughput, 25)
if __name__ == "__main__":
unittest.main()