EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

[CI] Add more multi-gpu tests (#1280)

Browse Source
This commit is contained in:
Lianmin Zheng
2024-09-01 00:27:25 -07:00
committed by GitHub
parent d134c139a1
commit 1b5d56f7f8
11 changed files with 271 additions and 128 deletions
Download Patch File Download Diff File Expand all files Collapse all files

19
test/srt/test_serving_throughput.py
View File

@@ -33,7 +33,7 @@ class TestServingThroughput(unittest.TestCase):
)
# Run benchmark
num_prompts = 400
num_prompts = 500
args = SimpleNamespace(
backend="sglang",
base_url=base_url,
@@ -74,8 +74,7 @@ class TestServingThroughput(unittest.TestCase):
)
if os.getenv("SGLANG_IS_IN_CI", "false") == "true":
# A100 (PCIE): 1450, H100 (SMX): 2550
assert res["output_throughput"] > 2500
assert res["output_throughput"] > 2400
def test_default_without_radix_cache(self):
res = self.run_test(
@@ -85,7 +84,6 @@ class TestServingThroughput(unittest.TestCase):
)
if os.getenv("SGLANG_IS_IN_CI", "false") == "true":
# A100 (PCIE): 1500, H100 (SMX): 2850
assert res["output_throughput"] > 2800
def test_default_without_chunked_prefill(self):
@@ -96,18 +94,7 @@ class TestServingThroughput(unittest.TestCase):
)
if os.getenv("SGLANG_IS_IN_CI", "false") == "true":
# A100 (PCIE): 1450, H100 (SMX): 2550
assert res["output_throughput"] > 2500
def test_all_cases(self):
for disable_radix_cache in [False, True]:
for disable_flashinfer in [False, True]:
for chunked_prefill_size in [-1, 2048]:
self.run_test(
disable_radix_cache=False,
disable_flashinfer=False,
chunked_prefill_size=-1,
)
assert res["output_throughput"] > 2400
if __name__ == "__main__":