[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 +1,84 @@
import os
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
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 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)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct": {
"accuracy": 0.84,
"latency": 150,
"output_throughput": 30,
},
}
os.environ["ASCEND_USE_FIA"] = "true"
class TestAscendTp1Bf16(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",
0.6,
"--attention-backend",
"ascend",
"--disable-radix-cache",
"--enable-torch-compile",
"--watchdog-timeout",
30000,
]
def test_a_gsm8k(self):
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=32,
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 +1,77 @@
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
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 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)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct": {
"accuracy": 0.85,
"latency": 150,
"output_throughput": 30,
},
}
class TestAscendGraphTp1Bf16(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",
]
def test_a_gsm8k(self):
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)
if __name__ == "__main__":
unittest.main()
@@ -0,0 +1,79 @@
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
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 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-2-npu-a2", nightly=False)
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct": {
"accuracy": 0.85,
"latency": 180,
"output_throughput": 20,
},
}
class TestAscendGraphTp2Bf16(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",
"--tp-size",
2,
]
def test_a_gsm8k(self):
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)
if __name__ == "__main__":
unittest.main()
@@ -0,0 +1,93 @@
import unittest
from urllib.parse import urlparse
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 as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
SimpleNamespace,
popen_launch_server,
run_bench_one_batch,
)
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)
MODEL = "/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct"
GSM8K_EXP_ACCURACY = 0.84
EXP_PREFILL_LATENCY = 0.045
TOKENS_TO_CAPTURE = [i for i in range(128, 4096, 128)]
class TestPiecewiseGraphPrefillCorrectness(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.url = urlparse(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",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--cuda-graph-bs",
128,
"--enforce-piecewise-cuda-graph",
"--piecewise-cuda-graph-tokens",
*TOKENS_TO_CAPTURE,
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
print(f"##=== Testing accuracy: {self.model} ===##")
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"],
GSM8K_EXP_ACCURACY,
)
class TestPiecewiseGraphPrefillBenchmark(CustomTestCase):
def test_latency(self):
print(f"##=== Testing prefill latency: {MODEL} ===##")
prefill_latency, _, _ = run_bench_one_batch(
MODEL,
other_args=[
"--trust-remote-code",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--enforce-piecewise-cuda-graph",
"--piecewise-cuda-graph-tokens",
]
+ TOKENS_TO_CAPTURE,
)
self.assertLess(prefill_latency, EXP_PREFILL_LATENCY)
if __name__ == "__main__":
unittest.main()