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

[NPU]Optimize the PR pipeline to reduce E2E runtime (#18767)

Browse Source 
Co-authored-by: Sugar920 <121632458+Sugar920@users.noreply.github.com>
This commit is contained in:
Cherry_ming
2026-02-28 17:06:03 +08:00
committed by GitHub
co-authored by Sugar920
parent 366574b2b8
commit 1eb40d8d45
12 changed files with 17 additions and 238 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/pr-test-npu.yml
+14 -14
View File
@@ -79,6 +79,10 @@ jobs:
needs: [check-changes, pr-gate] needs: [check-changes, pr-gate]
if: needs.check-changes.outputs.main_package == 'true' if: needs.check-changes.outputs.main_package == 'true'
runs-on: linux-aarch64-a2-1 runs-on: linux-aarch64-a2-1
strategy:
fail-fast: false
matrix:
part: [ 0, 1 ]
container: container:
image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11 image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11
steps: steps:
@@ -113,7 +117,7 @@ jobs:
STREAMS_PER_DEVICE: 32 STREAMS_PER_DEVICE: 32
run: | run: |
cd test/srt cd test/srt
python3 run_suite.py --suite per-commit-1-npu-a2 python3 run_suite.py --suite per-commit-1-npu-a2 --auto-partition-id ${{ matrix.part }} --auto-partition-size 2
per-commit-2-npu-a2: per-commit-2-npu-a2:
needs: [check-changes, pr-gate] needs: [check-changes, pr-gate]
@@ -122,7 +126,7 @@ jobs:
strategy: strategy:
fail-fast: true fail-fast: true
matrix: matrix:
part: [0, 1, 2] part: [0, 1]
container: container:
image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11 image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11
steps: steps:
@@ -157,14 +161,14 @@ jobs:
STREAMS_PER_DEVICE: 32 STREAMS_PER_DEVICE: 32
run: | run: |
cd test/srt cd test/srt
python3 run_suite.py --suite per-commit-2-npu-a2 --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 python3 run_suite.py --suite per-commit-2-npu-a2 --auto-partition-id ${{ matrix.part }} --auto-partition-size 2
per-commit-4-npu-a2: per-commit-4-npu-a3:
needs: [check-changes, pr-gate] needs: [check-changes, pr-gate]
if: needs.check-changes.outputs.main_package == 'true' if: needs.check-changes.outputs.main_package == 'true'
runs-on: linux-aarch64-a2-4 runs-on: linux-aarch64-a3-4
container: container:
image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-910b-ubuntu22.04-py3.11 image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-a3-ubuntu22.04-py3.11
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -180,7 +184,7 @@ jobs:
pip config set global.extra-index-url "https://pypi.tuna.tsinghua.edu.cn/simple" pip config set global.extra-index-url "https://pypi.tuna.tsinghua.edu.cn/simple"
pip config set global.trusted-host "${CACHING_URL} pypi.tuna.tsinghua.edu.cn" pip config set global.trusted-host "${CACHING_URL} pypi.tuna.tsinghua.edu.cn"
bash scripts/ci/npu/npu_ci_install_dependency.sh 910b bash scripts/ci/npu/npu_ci_install_dependency.sh a3
# copy required file from our daily cache # copy required file from our daily cache
cp ~/.cache/modelscope/hub/datasets/otavia/ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json /tmp cp ~/.cache/modelscope/hub/datasets/otavia/ShareGPT_Vicuna_unfiltered/ShareGPT_V3_unfiltered_cleaned_split.json /tmp
# copy download through proxy # copy download through proxy
@@ -197,16 +201,12 @@ jobs:
STREAMS_PER_DEVICE: 32 STREAMS_PER_DEVICE: 32
run: | run: |
cd test/srt cd test/srt
python3 run_suite.py --suite per-commit-4-npu-a2 --timeout-per-file 3600 python3 run_suite.py --suite per-commit-4-npu-a3 --timeout-per-file 3600
per-commit-16-npu-a3: per-commit-16-npu-a3:
needs: [check-changes, pr-gate] needs: [check-changes, pr-gate]
if: needs.check-changes.outputs.main_package == 'true' if: needs.check-changes.outputs.main_package == 'true'
runs-on: linux-aarch64-a3-16 runs-on: linux-aarch64-a3-16
strategy:
fail-fast: true
matrix:
part: [0, 1]
container: container:
image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-a3-ubuntu22.04-py3.11 image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.5.0-a3-ubuntu22.04-py3.11
steps: steps:
@@ -241,12 +241,12 @@ jobs:
STREAMS_PER_DEVICE: 32 STREAMS_PER_DEVICE: 32
run: | run: |
cd test/srt cd test/srt
python3 run_suite.py --suite per-commit-16-npu-a3 --timeout-per-file 3600 --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 python3 run_suite.py --suite per-commit-16-npu-a3 --timeout-per-file 3600
multimodal-gen-test-1-npu-a3: multimodal-gen-test-1-npu-a3:
needs: [check-changes, pr-gate] needs: [check-changes, pr-gate]
if: needs.check-changes.outputs.multimodal_gen == 'true' if: needs.check-changes.outputs.multimodal_gen == 'true'
runs-on: linux-aarch64-a3-16 runs-on: linux-aarch64-a3-2
container: container:
image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.3.rc2-a3-ubuntu22.04-py3.11 image: swr.cn-southwest-2.myhuaweicloud.com/base_image/ascend-ci/cann:8.3.rc2-a3-ubuntu22.04-py3.11
steps: steps:
test/srt/ascend/test_ascend_compile_graph_tp1_bf16.py
-22
View File
@@ -9,9 +9,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -77,26 +75,6 @@ class TestAscendTp1Bf16(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_deepep.py
+1 -23
View File
@@ -8,9 +8,7 @@ from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import ( from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -78,7 +76,7 @@ class TestAscendDeepEP(CustomTestCase):
args = SimpleNamespace( args = SimpleNamespace(
num_shots=5, num_shots=5,
data_path=None, data_path=None,
num_questions=1319, num_questions=500,
max_new_tokens=512, max_new_tokens=512,
parallel=128, parallel=128,
host=f"http://{self.url.hostname}", host=f"http://{self.url.hostname}",
@@ -93,26 +91,6 @@ class TestAscendDeepEP(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_graph_tp1_bf16.py
-22
View File
@@ -8,9 +8,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -70,26 +68,6 @@ class TestAscendGraphTp1Bf16(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_graph_tp2_bf16.py
-22
View File
@@ -8,9 +8,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -72,26 +70,6 @@ class TestAscendGraphTp2Bf16(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_hicache_mha.py
-22
View File
@@ -8,9 +8,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -73,26 +71,6 @@ class TestAscendMhaHicache(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_hicache_mla.py
-22
View File
@@ -8,9 +8,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -75,26 +73,6 @@ class TestAscendMlaHicache(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_mla_fia_w8a8int8.py
-22
View File
@@ -9,9 +9,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -76,26 +74,6 @@ class TestAscendMlaW8A8Int8(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_mla_w8a8int8.py
-22
View File
@@ -8,9 +8,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -74,26 +72,6 @@ class TestAscendMlaW8A8Int8(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_tp2_bf16.py
-22
View File
@@ -8,9 +8,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -73,26 +71,6 @@ class TestAscendTp2Bf16(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/ascend/test_ascend_tp4_bf16.py
-22
View File
@@ -7,9 +7,7 @@ from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import ( from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_ci,
popen_launch_server, popen_launch_server,
run_bench_offline_throughput,
) )
TEST_MODEL_MATRIX = { TEST_MODEL_MATRIX = {
@@ -76,26 +74,6 @@ class TestAscendTp4Bf16(CustomTestCase):
finally: finally:
kill_process_tree(process.pid) 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__": if __name__ == "__main__":
unittest.main() unittest.main()
test/srt/run_suite.py
+2 -3
View File
@@ -99,15 +99,14 @@ suite_ascend = {
TestFile("ascend/test_ascend_tp2_bf16.py", 400), TestFile("ascend/test_ascend_tp2_bf16.py", 400),
TestFile("ascend/test_ascend_tp2_fia_bf16.py", 400), TestFile("ascend/test_ascend_tp2_fia_bf16.py", 400),
], ],
"per-commit-4-npu-a2": [ "per-commit-4-npu-a3": [
TestFile("ascend/test_ascend_mla_w8a8int8.py", 400), TestFile("ascend/test_ascend_mla_w8a8int8.py", 400),
TestFile("ascend/test_ascend_hicache_mla.py", 400), TestFile("ascend/test_ascend_hicache_mla.py", 400),
TestFile("ascend/test_ascend_tp4_bf16.py", 400), TestFile("ascend/test_ascend_tp4_bf16.py", 400),
TestFile("ascend/test_ascend_w4a4_quantization.py", 600),
], ],
"per-commit-16-npu-a3": [ "per-commit-16-npu-a3": [
TestFile("ascend/test_ascend_deepep.py", 3600), TestFile("ascend/test_ascend_deepep.py", 3600),
# TestFile("ascend/test_ascend_deepseek_mtp.py", 2800),
TestFile("ascend/test_ascend_w4a4_quantization.py", 600),
], ],
} }