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

Fix GPU OOM (#6564)

Browse Source 
Co-authored-by: michael <michael.zhang@amd.com>
This commit is contained in:
kk
2025-05-25 07:38:39 +08:00
committed by GitHub
parent 24c035f2e3
commit 7a5e6ce1cb
8 changed files with 15 additions and 16 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
test/srt/test_bench_one_batch.py
View File

@@ -62,7 +62,10 @@ class TestBenchOneBatch(CustomTestCase):
f"### test_torch_compile_tp2_bs1 (Mixtral-8x7B)\n"
f"output_throughput: {output_throughput:.2f} token/s\n"
)
self.assertGreater(output_throughput, 220)
if os.getenv("SGLANG_AMD_CI") == "1":
self.assertGreater(output_throughput, 200)
else:
self.assertGreater(output_throughput, 220)
if __name__ == "__main__":

6
test/srt/test_bench_serving.py
View File

@@ -32,7 +32,7 @@ class TestBenchServing(CustomTestCase):
f'Output throughput: {res["output_throughput"]:.2f} token/s\n'
)
if os.getenv("SGLANG_AMD_CI") == "1":
self.assertGreater(res["output_throughput"], 3500)
self.assertGreater(res["output_throughput"], 3150)
else:
self.assertGreater(res["output_throughput"], 3800)
@@ -70,7 +70,7 @@ class TestBenchServing(CustomTestCase):
f'Output throughput: {res["output_throughput"]:.2f} token/s\n'
)
if os.getenv("SGLANG_AMD_CI") == "1":
self.assertGreater(res["output_throughput"], 3500)
self.assertGreater(res["output_throughput"], 3050)
else:
self.assertGreater(res["output_throughput"], 3800)
@@ -126,7 +126,7 @@ class TestBenchServing(CustomTestCase):
f'Output throughput: {res["output_throughput"]:.2f} token/s\n'
)
if os.getenv("SGLANG_AMD_CI") == "1":
self.assertGreater(res["output_throughput"], 4000)
self.assertGreater(res["output_throughput"], 3500)
else:
self.assertGreater(res["output_throughput"], 4300)

5
test/srt/test_eval_accuracy_large.py
View File

@@ -37,11 +37,6 @@ class TestEvalAccuracyLarge(CustomTestCase):
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def tearDown(self):
# Delay between tests to allow GPU memory cleanup
if os.getenv("SGLANG_AMD_CI") == "1":
time.sleep(180)
def test_mmlu(self):
args = SimpleNamespace(
base_url=self.base_url,

4
test/srt/test_full_deepseek_v3.py
View File

@@ -90,9 +90,9 @@ class TestDeepseekV3MTP(CustomTestCase):
"2",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.7",
]
if os.environ.get("SGLANG_AMD_CI") != "1":
other_args += ["--mem-frac", "0.7"]
cls.process = popen_launch_server(
cls.model,
cls.base_url,