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

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

Browse Source
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
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/ascend/test_ascend_autoround_dense.py
-76
View File
@@ -1,76 +0,0 @@
import logging
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,
popen_launch_server,
)
logger = logging.getLogger(__name__)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/Intel/Qwen3-8B-int4-AutoRound": {
"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()
test/srt/ascend/test_ascend_autoround_moe.py
-76
View File
@@ -1,76 +0,0 @@
import logging
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,
popen_launch_server,
)
logger = logging.getLogger(__name__)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/Intel/Qwen3-30B-A3B-Instruct-2507-int4-AutoRound": {
"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()
test/srt/ascend/test_ascend_compile_graph_tp1_bf16.py
-80
View File
@@ -1,80 +0,0 @@
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,
popen_launch_server,
)
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()
test/srt/ascend/test_ascend_deepep.py
-96
View File
@@ -1,96 +0,0 @@
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_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-R1-0528-W8A8": {
"accuracy": 0.95,
"latency": 1000,
"output_throughput": 6,
},
}
class TestAscendDeepEP(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",
"--attention-backend",
"ascend",
"--mem-fraction-static",
0.8,
"--disable-radix-cache",
"--chunked-prefill-size",
32768,
"--tp-size",
16,
"--dp-size",
1,
"--ep-size",
16,
"--max-running-requests",
24,
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"auto",
]
cls.extra_envs = {
"HCCL_BUFFSIZE": "1000",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "32",
"SGLANG_NPU_USE_MLAPO": "1",
}
os.environ.update(cls.extra_envs)
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=2400,
other_args=[
*self.common_args,
],
)
try:
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=500,
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()
test/srt/ascend/test_ascend_gptq_moe.py
-76
View File
@@ -1,76 +0,0 @@
import logging
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,
popen_launch_server,
)
logger = logging.getLogger(__name__)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/Qwen/Qwen3-30B-A3B-GPTQ-Int4": {
"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()
test/srt/ascend/test_ascend_graph_tp1_bf16.py
-73
View File
@@ -1,73 +0,0 @@
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,
popen_launch_server,
)
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()
test/srt/ascend/test_ascend_graph_tp2_bf16.py
-75
View File
@@ -1,75 +0,0 @@
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,
popen_launch_server,
)
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()
test/srt/ascend/test_ascend_hicache_mha.py
-76
View File
@@ -1,76 +0,0 @@
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,
popen_launch_server,
)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct": {
"accuracy": 0.85,
"latency": 150,
"output_throughput": 30,
},
}
class TestAscendMhaHicache(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",
"--enable-hierarchical-cache",
"--hicache-ratio",
1.2,
]
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()
test/srt/ascend/test_ascend_hicache_mla.py
-78
View File
@@ -1,78 +0,0 @@
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,
popen_launch_server,
)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/vllm-ascend/DeepSeek-V2-Lite-W8A8": {
"accuracy": 0.34,
"latency": 1000,
"output_throughput": 6,
},
}
class TestAscendMlaHicache(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",
4,
"--enable-hierarchical-cache",
"--hicache-ratio",
1.2,
]
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()
test/srt/ascend/test_ascend_mla_fia_w8a8int8.py
-79
View File
@@ -1,79 +0,0 @@
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,
popen_launch_server,
)
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",
"--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)
if __name__ == "__main__":
unittest.main()
test/srt/ascend/test_ascend_mla_w8a8int8.py
-77
View File
@@ -1,77 +0,0 @@
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,
popen_launch_server,
)
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",
"--tp-size",
4,
"--disable-radix-cache",
]
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()
test/srt/ascend/test_ascend_piecewise_graph_prefill.py
-89
View File
@@ -1,89 +0,0 @@
import unittest
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,
SimpleNamespace,
popen_launch_server,
run_bench_one_batch,
)
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()
test/srt/ascend/test_ascend_sampling_backend.py
-96
View File
@@ -1,96 +0,0 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
class TestAscendSamplingBackend(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct"
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=[
"--sampling-backend",
"ascend",
"--disable-radix-cache",
"--disable-cuda-graph",
"--mem-fraction-static",
0.85,
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_mmlu(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mmlu",
num_examples=64,
num_threads=32,
temperature=0.1,
)
metrics = run_eval(args)
self.assertGreaterEqual(metrics["score"], 0.65)
def test_greedy(self):
first_text = None
# ensure the answer is identical across single response
for _ in range(5):
response_single = requests.post(
self.base_url + "/generate",
json={
"text": "The capital of Germany is",
"sampling_params": {
"temperature": 0,
"max_new_tokens": 32,
},
},
).json()
text = response_single["text"]
if first_text is None:
first_text = text
self.assertEqual(text, first_text)
first_text = None
response_batch = requests.post(
self.base_url + "/generate",
json={
"text": ["The capital of Germany is"] * 10,
"sampling_params": {
"temperature": 0,
"max_new_tokens": 32,
},
},
).json()
# ensure the answer is identical among the batch
for i in range(10):
text = response_batch[i]["text"]
if first_text is None:
first_text = text
self.assertEqual(text, first_text)
if __name__ == "__main__":
unittest.main()
test/srt/ascend/test_ascend_tp1_bf16.py
-74
View File
@@ -1,74 +0,0 @@
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,
popen_launch_server,
)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct": {
"accuracy": 0.84,
"latency": 150,
"output_throughput": 30,
},
}
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",
"--disable-cuda-graph",
"--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()
test/srt/ascend/test_ascend_tp2_bf16.py
-76
View File
@@ -1,76 +0,0 @@
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,
popen_launch_server,
)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/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,
]
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()
test/srt/ascend/test_ascend_tp2_fia_bf16.py
-79
View File
@@ -1,79 +0,0 @@
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,
popen_launch_server,
)
TEST_MODEL_MATRIX = {
"/root/.cache/modelscope/hub/models/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)
if __name__ == "__main__":
unittest.main()
test/srt/ascend/test_ascend_tp4_bf16.py
-79
View File
@@ -1,79 +0,0 @@
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_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
TEST_MODEL_MATRIX = {
"Qwen/Qwen3-30B-A3B-Instruct-2507": {
"accuracy": 0.90,
"latency": 180,
"output_throughput": 20,
},
}
class TestAscendTp4Bf16(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",
0.7,
"--max-running-requests",
32,
"--attention-backend",
"ascend",
"--disable-radix-cache",
"--cuda-graph-max-bs",
32,
"--tp-size",
4,
]
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=1800,
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()
test/srt/ascend/test_ascend_w4a4_quantization.py
-108
View File
@@ -1,108 +0,0 @@
"""
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.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,
)
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()
test/srt/ascend/test_ascend_w8a8_quantization.py
-103
View File
@@ -1,103 +0,0 @@
"""
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.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,
)
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()
test/srt/ascend/test_llada2_mini_ascend.py
-87
View File
@@ -1,87 +0,0 @@
import os
import unittest
from types import SimpleNamespace
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.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
class TestLLaDA2Mini(CustomTestCase):
@classmethod
def setUpClass(cls):
cls._old_disable_acl = os.environ.get("SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT")
os.environ["SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT"] = "1"
cls.model = "/root/.cache/modelscope/hub/models/inclusionAI/LLaDA2.0-mini"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--disable-radix-cache",
"--mem-fraction-static",
"0.9",
"--max-running-requests",
"1",
"--attention-backend",
"ascend",
"--dllm-algorithm",
"LowConfidence", # TODO: Add dLLM configurations
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
if cls._old_disable_acl is None:
os.environ.pop("SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT", None)
else:
os.environ["SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT"] = cls._old_disable_acl
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
print(f"{metrics=}")
self.assertGreater(metrics["accuracy"], 0.88)
self.assertGreater(metrics["output_throughput"], 70)
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
acc_length, speed = send_one_prompt(args)
print(f"{speed=:.2f}")
if is_in_ci():
write_github_step_summary(
f"### test_bs_1_speed (llada2-mini) with tp1\n"
f"{speed=:.2f} token/s\n"
)
self.assertGreater(speed, 130)
if __name__ == "__main__":
unittest.main()