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

Fix Deepseek nightly tests (#12906)

Browse Source
This commit is contained in:
Kangyan-Zhou
2025-11-09 00:26:06 -08:00
committed by GitHub
parent b5e0417392
commit 93cf60fc64
5 changed files with 250 additions and 204 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/nightly/test_deepseek_v31_perf.py
+59 -68
View File
@@ -8,7 +8,7 @@ DEEPSEEK_V31_MODEL_PATH = "deepseek-ai/DeepSeek-V3.1"
PROFILE_DIR = "performance_profiles_deepseek_v31"
class TestNightlyDeepseekV31Basic(unittest.TestCase):
class TestNightlyDeepseekV31Performance(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEEPSEEK_V31_MODEL_PATH
@@ -16,81 +16,72 @@ class TestNightlyDeepseekV31Basic(unittest.TestCase):
cls.batch_sizes = [1, 1, 8, 16, 64]
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
cls.other_args = [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
# Define variant configurations
cls.variants = [
{
"name": "basic",
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
],
},
{
"name": "mtp",
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-frac",
"0.7",
],
},
]
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
cls.runner.setup_profile_directory()
def test_bench_one_batch(self):
results, success = self.runner.run_benchmark_for_model(
model_path=self.model,
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=self.other_args,
variant="basic",
)
failed_variants = []
self.runner.add_report(results)
self.runner.write_final_report()
try:
for variant_config in self.variants:
with self.subTest(variant=variant_config["name"]):
results, success = self.runner.run_benchmark_for_model(
model_path=self.model,
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=variant_config["other_args"],
variant=variant_config["name"],
)
if not success:
if not success:
failed_variants.append(variant_config["name"])
self.runner.add_report(results)
finally:
self.runner.write_final_report()
if failed_variants:
raise AssertionError(
f"Benchmark failed for {self.model} with basic configuration"
)
class TestNightlyDeepseekV31MTP(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEEPSEEK_V31_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
cls.batch_sizes = [1, 1, 8, 16, 64]
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
cls.other_args = [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-frac",
"0.7",
]
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
cls.runner.setup_profile_directory()
def test_bench_one_batch(self):
results, success = self.runner.run_benchmark_for_model(
model_path=self.model,
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=self.other_args,
variant="mtp",
)
self.runner.add_report(results)
self.runner.write_final_report()
if not success:
raise AssertionError(
f"Benchmark failed for {self.model} with MTP configuration"
f"Benchmark failed for {self.model} with the following variants: "
f"{', '.join(failed_variants)}"
)
test/srt/nightly/test_deepseek_v32_perf.py
+76 -112
View File
@@ -8,7 +8,7 @@ DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2-Exp"
PROFILE_DIR = "performance_profiles_deepseek_v32"
class TestNightlyDeepseekV32Basic(unittest.TestCase):
class TestNightlyDeepseekV32Performance(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEEPSEEK_V32_MODEL_PATH
@@ -16,125 +16,89 @@ class TestNightlyDeepseekV32Basic(unittest.TestCase):
cls.batch_sizes = [1, 1, 8, 16, 64]
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
cls.other_args = [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
# Define variant configurations
cls.variants = [
{
"name": "basic",
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
],
},
{
"name": "mtp",
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-frac",
"0.7",
],
},
{
"name": "nsa",
"other_args": [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--attention-backend",
"nsa",
"--nsa-prefill-backend",
"flashmla_sparse",
"--nsa-decode-backend",
"flashmla_kv",
],
},
]
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
cls.runner.setup_profile_directory()
def test_bench_one_batch(self):
results, success = self.runner.run_benchmark_for_model(
model_path=self.model,
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=self.other_args,
variant="basic",
)
failed_variants = []
self.runner.add_report(results)
self.runner.write_final_report()
try:
for variant_config in self.variants:
with self.subTest(variant=variant_config["name"]):
results, success = self.runner.run_benchmark_for_model(
model_path=self.model,
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=variant_config["other_args"],
variant=variant_config["name"],
)
if not success:
if not success:
failed_variants.append(variant_config["name"])
self.runner.add_report(results)
finally:
self.runner.write_final_report()
if failed_variants:
raise AssertionError(
f"Benchmark failed for {self.model} with basic configuration"
)
class TestNightlyDeepseekV32MTP(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEEPSEEK_V32_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
cls.batch_sizes = [1, 1, 8, 16, 64]
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
cls.other_args = [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-frac",
"0.7",
]
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
cls.runner.setup_profile_directory()
def test_bench_one_batch(self):
results, success = self.runner.run_benchmark_for_model(
model_path=self.model,
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=self.other_args,
variant="mtp",
)
self.runner.add_report(results)
self.runner.write_final_report()
if not success:
raise AssertionError(
f"Benchmark failed for {self.model} with MTP configuration"
)
class TestNightlyDeepseekV32NSA(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEEPSEEK_V32_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
cls.batch_sizes = [1, 1, 8, 16, 64]
cls.input_lens = tuple(_parse_int_list_env("NIGHTLY_INPUT_LENS", "4096"))
cls.output_lens = tuple(_parse_int_list_env("NIGHTLY_OUTPUT_LENS", "512"))
cls.other_args = [
"--trust-remote-code",
"--tp",
"8",
"--dp",
"8",
"--enable-dp-attention",
"--attention-backend",
"nsa",
"--nsa-prefill-backend",
"flashmla_sparse",
"--nsa-decode-backend",
"flashmla_kv",
]
cls.runner = NightlyBenchmarkRunner(PROFILE_DIR, cls.__name__, cls.base_url)
cls.runner.setup_profile_directory()
def test_bench_one_batch(self):
results, success = self.runner.run_benchmark_for_model(
model_path=self.model,
batch_sizes=self.batch_sizes,
input_lens=self.input_lens,
output_lens=self.output_lens,
other_args=self.other_args,
variant="nsa",
)
self.runner.add_report(results)
self.runner.write_final_report()
if not success:
raise AssertionError(
f"Benchmark failed for {self.model} with NSA configuration"
f"Benchmark failed for {self.model} with the following variants: "
f"{', '.join(failed_variants)}"
)
test/srt/run_suite.py
+1 -1
View File
@@ -220,7 +220,7 @@ suites = {
],
"nightly-4-gpu-b200": [
TestFile("test_fp4_moe.py", 300),
TestFile("nightly/test_nightly_gpt_oss_4gpu_perf.py", 600),
TestFile("nightly/test_gpt_oss_4gpu_perf.py", 600),
],
"nightly-8-gpu-b200": [],
"nightly-4-gpu": [],