minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

migrate NPU PR/nightly test cases to a3-560T (#33346)

Browse Source 
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: sglang-npu-bot <sglangnpu@163.com>
This commit is contained in:
Sugar920
2026-08-05 01:17:58 +08:00
committed by GitHub
co-authored by Claude sglang-npu-bot
parent 95d0e57e83
commit e76d0acdc9
33 changed files with 502 additions and 487 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/npu/basic_function/HiCache/test_npu_hicache_mha.py
+21 -22
View File
@@ -3,6 +3,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -12,7 +13,7 @@ from sglang.test.test_utils import (
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="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
@@ -45,33 +46,31 @@ class TestAscendMhaHicache(CustomTestCase):
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,
],
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"],
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {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"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
finally:
kill_process_tree(process.pid)
test/registered/npu/basic_function/HiCache/test_npu_hicache_mla.py
+20 -16
View File
@@ -3,6 +3,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -47,8 +48,6 @@ class TestAscendMlaHicache(CustomTestCase):
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,
@@ -59,21 +58,26 @@ class TestAscendMlaHicache(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/backends/test_npu_sampling_backend.py
+5 -1
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
@@ -13,7 +14,7 @@ from sglang.test.test_utils import (
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="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
@@ -41,6 +42,9 @@ class TestAscendSamplingBackend(CustomTestCase):
kill_process_tree(cls.process.pid)
def test_mmlu(self):
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
return
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
test/registered/npu/basic_function/dp_attn/test_npu_dp_attention.py
+3 -3
View File
@@ -7,6 +7,7 @@ from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import NPUGSM8KMixin
from sglang.test.ascend.test_ascend_utils import (
DEEPSEEK_V2_LITE_W8A8_WEIGHTS_PATH,
IMAGES_EXAMPLE_PATH,
KIMI_VL_A3B_INSTRUCT_WEIGHTS_PATH,
QWEN3_32B_WEIGHTS_PATH,
)
@@ -16,7 +17,6 @@ from sglang.test.kits.json_constrained_kit import JSONConstrainedMixin
from sglang.test.kits.radix_cache_server_kit import run_radix_attention_test
from sglang.test.kits.regex_constrained_kit import RegexConstrainedMixin
from sglang.test.test_utils import (
DEFAULT_IMAGE_URL,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -74,7 +74,7 @@ class TestDPAttentionMixedChunk(
CustomTestCase,
NPUGSM8KMixin,
):
gsm8k_accuracy_thres = 0.35
gsm8k_accuracy_thres = 0.34
@classmethod
def setUpClass(cls):
@@ -151,7 +151,7 @@ class TestDPAttentionDP2TP2VLM(CustomTestCase):
def setUpClass(cls):
cls.model = KIMI_VL_A3B_INSTRUCT_WEIGHTS_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
cls.image_url = DEFAULT_IMAGE_URL
cls.image_url = IMAGES_EXAMPLE_PATH
cls.process = popen_launch_server(
cls.model,
cls.base_url,
test/registered/npu/basic_function/optimization_debug/test_npu_compile_graph_tp1_bf16.py
+21 -17
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -13,7 +14,7 @@ from sglang.test.test_utils import (
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="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
@@ -49,8 +50,6 @@ class TestAscendTp1Bf16(CustomTestCase):
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,
@@ -61,21 +60,26 @@ class TestAscendTp1Bf16(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/optimization_debug/test_npu_piecewise_graph_prefill.py
+1 -1
View File
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
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="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
test/registered/npu/basic_function/parallel_strategy/data_parallelism/test_npu_load_balance_method.py
+4
View File
@@ -6,6 +6,7 @@ from urllib.parse import urlparse
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import DEEPSEEK_R1_0528_W8A8_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
@@ -72,6 +73,9 @@ class TestDPAttentionRoundBinLoadBalance(CustomTestCase):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
return
args = SimpleNamespace(
num_shots=5,
data_path=None,
test/registered/npu/basic_function/parallel_strategy/data_parallelism/test_npu_load_balance_method_pd_disaggregation.py
+4
View File
@@ -11,6 +11,7 @@ from urllib.parse import urlparse
import requests
from sglang.test.ascend.disaggregation_utils import TestDisaggregationBase
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import (
QWEN3_30B_A3B_INSTRUCT_2507_WEIGHTS_PATH,
)
@@ -134,6 +135,9 @@ class BaseTestNPULoadBalanceMethodDPDisaggregation(TestDisaggregationBase):
)
def test_gsm8k(self):
if _is_pr_pipeline:
run_npu_pr_smoke(self.lb_url)
return
args = SimpleNamespace(
num_shots=5,
data_path=None,
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep.py
+20 -16
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -67,8 +68,6 @@ class TestAscendDeepEP(CustomTestCase):
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,
@@ -79,21 +78,26 @@ class TestAscendDeepEP(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_eplb_min_rebalancing_utilization_threshold.py
+4
View File
@@ -3,6 +3,7 @@ import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import QWEN3_30B_A3B_W8A8_WEIGHTS_PATH
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.run_eval import run_eval
@@ -94,6 +95,9 @@ class TestEplbMinRebalancingUtilizationThresholdBase(CustomTestCase):
os.remove("./rebalance_err_log.txt")
def test_gsm8k(self):
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
return
args = SimpleNamespace(
max_tokens=512,
base_url=DEFAULT_URL_FOR_TEST,
test/registered/npu/basic_function/parallel_strategy/expert_parallelism/test_npu_moe_dense_tp_size.py
+4
View File
@@ -2,6 +2,7 @@ import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import DEEPSEEK_CODER_V2_LITE_WEIGHTS_PATH
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.run_eval import run_eval
@@ -61,6 +62,9 @@ class TestAscendMoeDenseTPSize(CustomTestCase):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
if _is_pr_pipeline:
run_npu_pr_smoke(DEFAULT_URL_FOR_TEST)
return
args = SimpleNamespace(
max_tokens=512,
base_url=DEFAULT_URL_FOR_TEST,
test/registered/npu/basic_function/quant/test_npu_autoround_dense.py
+21 -17
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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
@@ -14,7 +15,7 @@ from sglang.test.test_utils import (
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="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
@@ -46,8 +47,6 @@ class TestAscendAutoRoundDense(CustomTestCase):
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,
@@ -58,21 +57,26 @@ class TestAscendAutoRoundDense(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
logger.info(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/quant/test_npu_autoround_moe.py
+21 -17
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import (
QWEN3_30B_A3B_INSTRUCT_2507_INT4_AUTOROUND_WEIGHTS_PATH,
)
@@ -15,7 +16,7 @@ from sglang.test.test_utils import (
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="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
@@ -47,8 +48,6 @@ class TestAscendAutoRoundMoE(CustomTestCase):
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,
@@ -59,21 +58,26 @@ class TestAscendAutoRoundMoE(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
logger.info(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/quant/test_npu_gptq_moe.py
+21 -17
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import (
QWEN3_30B_A3B_GPTQ_2507_INT4_WEIGHTS_PATH,
)
@@ -16,7 +17,7 @@ from sglang.test.test_utils import (
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="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
logger = logging.getLogger(__name__)
@@ -48,8 +49,6 @@ class TestAscendGPTQMoEInt4(CustomTestCase):
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,
@@ -60,21 +59,26 @@ class TestAscendGPTQMoEInt4(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
logger.info(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/runtime_opts/test_npu_mla_fia_w8a8int8.py
+21 -17
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -13,7 +14,7 @@ from sglang.test.test_utils import (
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="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
@@ -48,8 +49,6 @@ class TestAscendMlaW8A8Int8(CustomTestCase):
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,
@@ -60,21 +59,26 @@ class TestAscendMlaW8A8Int8(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/runtime_opts/test_npu_mla_w8a8int8.py
+20 -16
View File
@@ -3,6 +3,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -46,8 +47,6 @@ class TestAscendMlaW8A8Int8(CustomTestCase):
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,
@@ -58,21 +57,26 @@ class TestAscendMlaW8A8Int8(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/runtime_opts/test_npu_tp1_bf16.py
+21 -17
View File
@@ -3,6 +3,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -12,7 +13,7 @@ from sglang.test.test_utils import (
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="stage-b-test-1-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
@@ -43,8 +44,6 @@ class TestAscendTp1Bf16(CustomTestCase):
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,
@@ -55,21 +54,26 @@ class TestAscendTp1Bf16(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/runtime_opts/test_npu_tp2_bf16.py
+21 -17
View File
@@ -3,6 +3,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -12,7 +13,7 @@ from sglang.test.test_utils import (
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="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
@@ -45,8 +46,6 @@ class TestAscendTp2Bf16(CustomTestCase):
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,
@@ -57,21 +56,26 @@ class TestAscendTp2Bf16(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/runtime_opts/test_npu_tp2_fia_bf16.py
+21 -17
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -13,7 +14,7 @@ from sglang.test.test_utils import (
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="stage-b-test-2-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
TEST_MODEL_MATRIX = {
@@ -48,8 +49,6 @@ class TestAscendTp2Bf16(CustomTestCase):
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,
@@ -60,21 +59,26 @@ class TestAscendTp2Bf16(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/runtime_opts/test_npu_tp4_bf16.py
+20 -16
View File
@@ -3,6 +3,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
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 (
@@ -48,8 +49,6 @@ class TestAscendTp4Bf16(CustomTestCase):
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,
@@ -60,21 +59,26 @@ class TestAscendTp4Bf16(CustomTestCase):
)
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),
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {model} ===##")
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
TEST_MODEL_MATRIX[model]["accuracy"],
)
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)
test/registered/npu/basic_function/speculative_inference/test_npu_basic_sanity_eagle3.py
+2 -2
View File
@@ -1,6 +1,7 @@
import unittest
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import NPUGSM8KMixin
from sglang.test.ascend.test_ascend_utils import (
QWEN3_8B_EAGLE3_WEIGHTS_PATH,
QWEN3_8B_WEIGHTS_PATH,
@@ -9,7 +10,6 @@ from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.kits.basic_api_contract_kit import BasicAPIContractMixin
from sglang.test.kits.basic_decode_correctness_kit import BasicDecodeCorrectnessMixin
from sglang.test.kits.basic_scheduler_stress_kit import BasicSchedulerStressMixin
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
@@ -25,7 +25,7 @@ class TestBasicSanityEagle3(
BasicAPIContractMixin,
BasicDecodeCorrectnessMixin,
BasicSchedulerStressMixin,
GSM8KMixin,
NPUGSM8KMixin,
CustomTestCase,
):
served_model_name = QWEN3_8B_WEIGHTS_PATH
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_attention_mode.py
+19 -12
View File
@@ -3,6 +3,7 @@ import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import (
QWEN3_32B_EAGLE3_WEIGHTS_PATH,
QWEN3_32B_W8A8_MINDIE_WEIGHTS_PATH,
@@ -103,12 +104,15 @@ class TestNpuSpeculativeAttentionMode(CustomTestCase):
)
try:
metrics = self._run_gsm8k_eval()
self.assertGreaterEqual(
metrics["score"],
0.83,
f"GSM8K score {metrics['score']} below threshold 0.83",
)
if _is_pr_pipeline:
run_npu_pr_smoke(DEFAULT_URL_FOR_TEST)
else:
metrics = self._run_gsm8k_eval()
self.assertGreaterEqual(
metrics["score"],
0.83,
f"GSM8K score {metrics['score']} below threshold 0.83",
)
finally:
kill_process_tree(process.pid)
@@ -164,12 +168,15 @@ class TestNpuSpeculativeAttentionMode(CustomTestCase):
)
try:
metrics = self._run_gsm8k_eval()
self.assertGreaterEqual(
metrics["score"],
0.83,
f"GSM8K score {metrics['score']} below threshold 0.83",
)
if _is_pr_pipeline:
run_npu_pr_smoke(DEFAULT_URL_FOR_TEST)
else:
metrics = self._run_gsm8k_eval()
self.assertGreaterEqual(
metrics["score"],
0.83,
f"GSM8K score {metrics['score']} below threshold 0.83",
)
finally:
kill_process_tree(process.pid)
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_draft_attention_backend.py
+4
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import (
DEEPSEEK_R1_0528_W4A8_PER_CHANNEL_WEIGHTS_PATH,
)
@@ -87,6 +88,9 @@ class TestAscendSpeculativeDraftAttentionAndMoeRunner(CustomTestCase):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(self):
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
return
args = SimpleNamespace(
base_url=self.base_url,
eval_name="gsm8k",
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_moe_a2a_backend.py
+22 -17
View File
@@ -4,6 +4,7 @@ from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import (
DEEPSEEK_R1_0528_W8A8_WEIGHTS_PATH,
)
@@ -64,7 +65,6 @@ class TestAscendDistTimeout(CustomTestCase):
]
def test_a_gsm8k(self):
print(f"##=== Testing accuracy: {self.model} ===##")
process = popen_launch_server(
self.model,
self.base_url,
@@ -74,23 +74,28 @@ class TestAscendDistTimeout(CustomTestCase):
)
try:
args = SimpleNamespace(
base_url=self.base_url,
eval_name="gsm8k",
api="completion",
num_examples=1319,
num_threads=128,
max_tokens=512,
num_shots=5,
temperature=0.0,
)
if _is_pr_pipeline:
run_npu_pr_smoke(self.base_url)
else:
print(f"##=== Testing accuracy: {self.model} ===##")
metrics = run_eval(args)
self.assertGreaterEqual(
metrics["score"],
self.accuracy,
f"GSM8K score {metrics['score']} below threshold {self.accuracy}",
)
args = SimpleNamespace(
base_url=self.base_url,
eval_name="gsm8k",
api="completion",
num_examples=1319,
num_threads=128,
max_tokens=512,
num_shots=5,
temperature=0.0,
)
metrics = run_eval(args)
self.assertGreaterEqual(
metrics["score"],
self.accuracy,
f"GSM8K score {metrics['score']} below threshold {self.accuracy}",
)
finally:
kill_process_tree(process.pid)
test/registered/npu/basic_function/speculative_inference/test_npu_speculative_token_map.py
+33 -27
View File
@@ -3,6 +3,7 @@ import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.npu_eval_accuracy_kit import _is_pr_pipeline, run_npu_pr_smoke
from sglang.test.ascend.test_ascend_utils import (
FR_SPEC_TOKEN_MAP_PATH,
LLAMA_3_8B_EAGLE_WEIGHTS_PATH,
@@ -77,24 +78,26 @@ class TestNpuSpeculativeTokenMap(CustomTestCase):
env=env,
)
try:
eval_args = SimpleNamespace(
base_url=DEFAULT_URL_FOR_TEST,
eval_name="gsm8k",
api="completion",
num_examples=1319,
num_threads=128,
max_tokens=512,
num_shots=5,
temperature=0.0,
)
metrics = run_eval(eval_args)
self.assertGreaterEqual(metrics["score"], 0.83)
if _is_pr_pipeline:
run_npu_pr_smoke(DEFAULT_URL_FOR_TEST)
else:
eval_args = SimpleNamespace(
base_url=DEFAULT_URL_FOR_TEST,
eval_name="gsm8k",
api="completion",
num_examples=1319,
num_threads=128,
max_tokens=512,
num_shots=5,
temperature=0.0,
)
metrics = run_eval(eval_args)
self.assertGreaterEqual(metrics["score"], 0.83)
finally:
kill_process_tree(process.pid)
def test_eagle_with_valid_token_map_gsm8k(self):
"""EAGLE (EAGLE-2) with valid token map; GSM8K accuracy should meet threshold."""
args = [
"--trust-remote-code",
"--attention-backend",
@@ -135,20 +138,23 @@ class TestNpuSpeculativeTokenMap(CustomTestCase):
env=env,
)
try:
eval_args = SimpleNamespace(
base_url=DEFAULT_URL_FOR_TEST,
eval_name="gsm8k",
api="completion",
num_examples=1319,
num_threads=128,
max_tokens=512,
num_shots=5,
temperature=0.0,
)
metrics = run_eval(eval_args)
self.assertGreaterEqual(
metrics["score"], 0.75
) # adjust threshold as needed
if _is_pr_pipeline:
run_npu_pr_smoke(DEFAULT_URL_FOR_TEST)
else:
eval_args = SimpleNamespace(
base_url=DEFAULT_URL_FOR_TEST,
eval_name="gsm8k",
api="completion",
num_examples=1319,
num_threads=128,
max_tokens=512,
num_shots=5,
temperature=0.0,
)
metrics = run_eval(eval_args)
self.assertGreaterEqual(
metrics["score"], 0.75
) # adjust threshold as needed
finally:
kill_process_tree(process.pid)
test/registered/npu/llm_models/test_npu_llama_2_7b_communications_compression.py
+1 -1
View File
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import LLAMA_2_7B_WEIGHTS_PATH
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.test_utils import CustomTestCase
register_npu_ci(est_time=400, suite="nightly-2-npu-a3")
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
class TestLlama(GSM8KAscendMixin, CustomTestCase):
test/registered/npu/llm_models/test_npu_qwen3_8b_communications_quantization.py
+1 -1
View File
@@ -5,7 +5,7 @@ from sglang.test.ascend.test_ascend_utils import QWEN3_8B_WEIGHTS_PATH
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.test_utils import CustomTestCase
register_npu_ci(est_time=400, suite="nightly-2-npu-a3")
register_npu_ci(est_time=400, suite="nightly-2-npu-a3", nightly=True)
class TestQwen38BCommQuantization(GSM8KAscendMixin, CustomTestCase):
test/registered/unit/layers/attention/test_vision_backend_selection.py
+2 -1
View File
@@ -11,7 +11,8 @@ from sglang.srt.layers.attention import vision
from sglang.test.ci.ci_register import register_cpu_ci, register_npu_ci
register_cpu_ci(est_time=2, suite="base-a-test-cpu")
register_npu_ci(est_time=2, suite="stage-b-test-1-npu-a2")
register_npu_ci(est_time=2, suite="stage-b-test-1-npu-a3")
register_npu_ci(est_time=2, suite="nightly-1-npu-a3", nightly=True)
@pytest.fixture
test/run_suite.py
+3 -2
View File
@@ -100,9 +100,10 @@ PER_COMMIT_SUITES = {
HWBackend.NPU: [
"base-a-test-1-gpu-small",
"stage-a-unit-test-npu",
"stage-b-test-1-npu-a2",
"stage-b-test-2-npu-a2",
"stage-b-test-1-npu-a3",
"stage-b-test-2-npu-a3",
"stage-b-test-4-npu-a3",
"stage-b-test-8-npu-a3",
"stage-b-test-16-npu-a3",
],
HWBackend.XPU: [