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

Update GLM tests to 5.2 and delete redundant tests (#29686)

Browse Source
This commit is contained in:
Baizhou Zhang
2026-06-29 23:04:26 -07:00
committed by GitHub
parent b6907d9664
commit cc238270b1
16 changed files with 39 additions and 508 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/8-gpu-models/test_deepseek_v32.py
-200
View File
@@ -1,200 +0,0 @@
import unittest
from sglang.test.accuracy_test_runner import AccuracyTestParams
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.performance_test_runner import PerformanceTestParams
from sglang.test.run_combined_tests import run_combined_tests
from sglang.test.test_utils import ModelLaunchSettings, is_blackwell_system
from sglang.test.tool_call_test_runner import ToolCallTestParams
register_cuda_ci(est_time=5400, suite="nightly-8-gpu-common", nightly=True)
DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
BASE_ARGS = [
"--trust-remote-code",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
]
TOOL_CALL_ARGS = [
"--tool-call-parser=deepseekv32",
"--reasoning-parser=deepseek-v3",
]
DP_ARGS = [
"--tp=8",
"--dp=8",
"--enable-dp-attention",
]
# Accuracy thresholds
GSM8K_BASELINE = 0.935
GPQA_BASELINE = 0.83
class TestDeepseekV32(unittest.TestCase):
"""Unified test class for DeepSeek V3.2 performance and accuracy.
Tests multiple variants with both performance and accuracy tests:
- dp: Standard TP=8 + DP=8 with dp-attention
- dp+mtp: DP + EAGLE speculative decoding
- tp: Pure TP=8 only
- tp+mtp: Pure TP=8 + EAGLE speculative decoding
"""
def test_deepseek_v32_all_variants(self):
"""Run performance and accuracy for all DeepSeek V3.2 variants."""
TP_ARGS = [
"--tp=8",
]
MTP_ARGS = [
"--speculative-algorithm=EAGLE",
"--speculative-num-steps=3",
"--speculative-eagle-topk=1",
"--speculative-num-draft-tokens=4",
"--mem-frac=0.85",
]
variants = [
# Variant: "dp" - Standard TP=8 + DP=8 with dp-attention
ModelLaunchSettings(
DEEPSEEK_V32_MODEL_PATH,
tp_size=8,
extra_args=BASE_ARGS + DP_ARGS + TOOL_CALL_ARGS,
variant="DP8",
),
# Variant: "dp+mtp" - DP + EAGLE speculative decoding
ModelLaunchSettings(
DEEPSEEK_V32_MODEL_PATH,
tp_size=8,
extra_args=BASE_ARGS + DP_ARGS + TOOL_CALL_ARGS + MTP_ARGS,
variant="DP8+MTP",
),
# Variant: "tp" - Pure TP=8 only
ModelLaunchSettings(
DEEPSEEK_V32_MODEL_PATH,
tp_size=8,
extra_args=BASE_ARGS + TP_ARGS + TOOL_CALL_ARGS,
variant="TP8",
),
# Variant: "tp+mtp" - Pure TP=8 + EAGLE speculative decoding
ModelLaunchSettings(
DEEPSEEK_V32_MODEL_PATH,
tp_size=8,
extra_args=BASE_ARGS + TP_ARGS + TOOL_CALL_ARGS + MTP_ARGS,
variant="TP8+MTP",
),
]
run_combined_tests(
models=variants,
test_name="DeepSeek-V3.2",
accuracy_params=AccuracyTestParams(
dataset="gsm8k", baseline_accuracy=GSM8K_BASELINE
),
performance_params=PerformanceTestParams(
batch_sizes=[1, 8, 16, 64],
profile_dir="performance_profiles_deepseek_v32",
),
tool_call_params=ToolCallTestParams(
test_thinking=True, test_reasoning_usage=True
),
)
@unittest.skipIf(is_blackwell_system(), "Requires H200 system")
def test_deepseek_v32_dsa_backends(self):
"""Test DSA attention backend variants (H200 only).
Tests three DSA backend configurations:
- flashmla: flashmla_sparse prefill + flashmla_kv decode
- fa3: FA3 prefill + FA3 decode
- fp8kvcache: default backends with FP8 KV cache
"""
DSA_FLASHMLA_ARGS = [
"--attention-backend=dsa",
"--dsa-prefill-backend=flashmla_sparse",
"--dsa-decode-backend=flashmla_kv",
]
DSA_FA3_ARGS = [
"--attention-backend=dsa",
"--dsa-prefill-backend=fa3",
"--dsa-decode-backend=fa3",
]
DSA_FP8KV_ARGS = [
"--attention-backend=dsa",
"--kv-cache-dtype=fp8_e4m3",
]
dsa_variants = [
# flashmla backend
ModelLaunchSettings(
DEEPSEEK_V32_MODEL_PATH,
tp_size=8,
extra_args=BASE_ARGS + DP_ARGS + DSA_FLASHMLA_ARGS,
),
# fa3 backend
ModelLaunchSettings(
DEEPSEEK_V32_MODEL_PATH,
tp_size=8,
extra_args=BASE_ARGS + DP_ARGS + DSA_FA3_ARGS,
),
# fp8 kv cache
ModelLaunchSettings(
DEEPSEEK_V32_MODEL_PATH,
tp_size=8,
extra_args=BASE_ARGS + DP_ARGS + DSA_FP8KV_ARGS,
),
]
run_combined_tests(
models=dsa_variants,
test_name="DeepSeek-V3.2 DSA Backends",
accuracy_params=AccuracyTestParams(
dataset="gsm8k", baseline_accuracy=GSM8K_BASELINE
),
performance_params=None,
)
@unittest.skipIf(
not is_blackwell_system(),
"Hardware agnostic - just using B200 for efficiency reasons",
)
def test_deepseek_v32_b200(self):
"""Test DeepSeek V3.2 with GPQA evaluation using thinking mode (B200 only).
This test runs GPQA evaluation with the reasoning parser enabled.
"""
B200_REASONING_ARGS = [
"--tool-call-parser=deepseekv32",
"--reasoning-parser=deepseek-v3",
]
variants = [
ModelLaunchSettings(
DEEPSEEK_V32_MODEL_PATH,
tp_size=8,
extra_args=BASE_ARGS + DP_ARGS + B200_REASONING_ARGS,
),
]
run_combined_tests(
models=variants,
test_name="DeepSeek-V3.2 GPQA (B200)",
accuracy_params=AccuracyTestParams(
dataset="gpqa",
baseline_accuracy=GPQA_BASELINE,
num_examples=198,
num_threads=198,
max_tokens=120000,
thinking_mode="deepseek-v3",
temperature=0.1,
repeat=4,
),
performance_params=None, # Skip performance test for GPQA
)
if __name__ == "__main__":
unittest.main()
test/registered/8-gpu-models/test_glm_51_fp8.py → test/registered/8-gpu-models/test_glm52_fp8.py
+9 -9
View File
@@ -9,7 +9,7 @@ from sglang.test.test_utils import ModelLaunchSettings
# Runs on both H200 and B200 via nightly-8-gpu-common suite # Runs on both H200 and B200 via nightly-8-gpu-common suite
register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True) register_cuda_ci(est_time=1800, suite="nightly-8-gpu-common", nightly=True)
GLM_51_FP8_MODEL_PATH = "zai-org/GLM-5.1-FP8" GLM_52_FP8_MODEL_PATH = "zai-org/GLM-5.2-FP8"
COMMON_ARGS = [ COMMON_ARGS = [
"--trust-remote-code", "--trust-remote-code",
@@ -27,27 +27,27 @@ MTP_ARGS = [
] ]
class TestGlm51Fp8(unittest.TestCase): class TestGlm52Fp8(unittest.TestCase):
"""GLM-5.1 FP8 on H200/B200 (8-GPU, tp=8).""" """GLM-5.2 FP8 on H200/B200 (8-GPU, tp=8)."""
def test_glm51_fp8(self): def test_glm52_fp8(self):
dp_args = ["--dp=8", "--enable-dp-attention"] dp_args = ["--dp=8", "--enable-dp-attention"]
variants = [ variants = [
ModelLaunchSettings( ModelLaunchSettings(
GLM_51_FP8_MODEL_PATH, GLM_52_FP8_MODEL_PATH,
tp_size=8, tp_size=8,
extra_args=COMMON_ARGS, extra_args=COMMON_ARGS,
variant="TP8", variant="TP8",
), ),
ModelLaunchSettings( ModelLaunchSettings(
GLM_51_FP8_MODEL_PATH, GLM_52_FP8_MODEL_PATH,
tp_size=8, tp_size=8,
extra_args=COMMON_ARGS + dp_args, extra_args=COMMON_ARGS + dp_args,
variant="TP8+DP8", variant="TP8+DP8",
), ),
ModelLaunchSettings( ModelLaunchSettings(
GLM_51_FP8_MODEL_PATH, GLM_52_FP8_MODEL_PATH,
tp_size=8, tp_size=8,
extra_args=COMMON_ARGS + dp_args + MTP_ARGS, extra_args=COMMON_ARGS + dp_args + MTP_ARGS,
variant="TP8+DP8+MTP", variant="TP8+DP8+MTP",
@@ -56,10 +56,10 @@ class TestGlm51Fp8(unittest.TestCase):
run_combined_tests( run_combined_tests(
models=variants, models=variants,
test_name="GLM-5.1-FP8", test_name="GLM-5.2-FP8",
accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92), accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92),
performance_params=PerformanceTestParams( performance_params=PerformanceTestParams(
profile_dir="performance_profiles_glm_51_fp8", profile_dir="performance_profiles_glm_52_fp8",
), ),
) )
test/registered/cuda_graph/piecewise/test_pcg_glm5_fp4.py → test/registered/cuda_graph/piecewise/test_pcg_glm52_fp4.py
+5 -5
View File
@@ -13,20 +13,20 @@ from sglang.test.test_utils import (
register_cuda_ci(est_time=900, stage="base-c", runner_config="4-gpu-b200") register_cuda_ci(est_time=900, stage="base-c", runner_config="4-gpu-b200")
GLM5_FP4_MODEL = "nvidia/GLM-5-NVFP4" GLM52_FP4_MODEL = "nvidia/GLM-5.2-NVFP4"
class TestPCGGlm5Fp4(CustomTestCase): class TestPCGGlm52Fp4(CustomTestCase):
"""PCG prefill on GLM-5-NVFP4 (DSA model, TP=4, B200). """PCG prefill on GLM-5.2-NVFP4 (DSA model, TP=4, B200).
GLM-5 uses GlmMoeDsaForCausalLM (DSA attention). This test verifies that GLM-5.2 uses GlmMoeDsaForCausalLM (DSA attention). This test verifies that
piecewise CUDA graph works correctly after the DSA indexer was updated to piecewise CUDA graph works correctly after the DSA indexer was updated to
cache k_fp8/k_scale for PCG-compatible prefill. cache k_fp8/k_scale for PCG-compatible prefill.
""" """
@classmethod @classmethod
def setUpClass(cls): def setUpClass(cls):
cls.model = GLM5_FP4_MODEL cls.model = GLM52_FP4_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server( cls.process = popen_launch_server(
cls.model, cls.model,
test/registered/cuda_graph/piecewise/test_pcg_glm5_fp8_tp8.py → test/registered/cuda_graph/piecewise/test_pcg_glm52_fp8_tp8.py
+4 -4
View File
@@ -13,15 +13,15 @@ from sglang.test.test_utils import (
register_cuda_ci(est_time=900, stage="base-c", runner_config="8-gpu-h200") register_cuda_ci(est_time=900, stage="base-c", runner_config="8-gpu-h200")
GLM5_FP8_MODEL = "zai-org/GLM-5-FP8" GLM52_FP8_MODEL = "zai-org/GLM-5.2-FP8"
class TestBCGGlm5Fp8TP8(CustomTestCase): class TestBCGGlm52Fp8TP8(CustomTestCase):
"""Breakable CUDA graph prefill on GLM-5-FP8 (DSA model, TP=8, H200).""" """Breakable CUDA graph prefill on GLM-5.2-FP8 (DSA model, TP=8, H200)."""
@classmethod @classmethod
def setUpClass(cls): def setUpClass(cls):
cls.model = GLM5_FP8_MODEL cls.model = GLM52_FP8_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server( cls.process = popen_launch_server(
cls.model, cls.model,
test/registered/debug_utils/test_nightly_precision_regression.py
+2 -2
View File
@@ -2,7 +2,7 @@
compare day-over-day against a rolling baseline. compare day-over-day against a rolling baseline.
Env knobs: Env knobs:
SGLANG_PRECISION_MODELS comma-separated model ids (default GLM-5.1-FP8) SGLANG_PRECISION_MODELS comma-separated model ids (default GLM-5.2-FP8)
SGLANG_PRECISION_BASELINE_DIR local baseline dir SGLANG_PRECISION_BASELINE_DIR local baseline dir
SGLANG_PRECISION_DIFF_THRESHOLD per-tensor rel_diff cutoff (default 1e-3) SGLANG_PRECISION_DIFF_THRESHOLD per-tensor rel_diff cutoff (default 1e-3)
SGLANG_PRECISION_FORCE_UPDATE=1 skip comparison, refresh baseline SGLANG_PRECISION_FORCE_UPDATE=1 skip comparison, refresh baseline
@@ -49,7 +49,7 @@ except Exception: # pragma: no cover
register_cuda_ci(est_time=3600, suite="nightly-precision-8-gpu-h200", nightly=True) register_cuda_ci(est_time=3600, suite="nightly-precision-8-gpu-h200", nightly=True)
DEFAULT_MODELS_FOR_NIGHTLY_PRECISION = "zai-org/GLM-5.1-FP8" DEFAULT_MODELS_FOR_NIGHTLY_PRECISION = "zai-org/GLM-5.2-FP8"
DEFAULT_DIFF_THRESHOLD = 1e-3 DEFAULT_DIFF_THRESHOLD = 1e-3
# Fallback when the layer count can't be resolved: never silently shrink coverage. # Fallback when the layer count can't be resolved: never silently shrink coverage.
DUMPER_FILTER_ALL_LAYERS = ( DUMPER_FILTER_ALL_LAYERS = (
test/registered/gb300/test_glm5_nvfp4.py → test/registered/gb300/test_glm52_nvfp4.py
+5 -5
View File
@@ -8,7 +8,7 @@ from sglang.test.test_utils import ModelLaunchSettings
register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300-glm5-nvfp4", nightly=True) register_cuda_ci(est_time=7200, suite="nightly-4-gpu-gb300-glm5-nvfp4", nightly=True)
MODEL_PATH = "nvidia/GLM-5-NVFP4" MODEL_PATH = "nvidia/GLM-5.2-NVFP4"
COMMON_ARGS = [ COMMON_ARGS = [
"--trust-remote-code", "--trust-remote-code",
@@ -35,10 +35,10 @@ DP_MTP_ARGS = [
] ]
class TestGlm5Nvfp4(unittest.TestCase): class TestGlm52Nvfp4(unittest.TestCase):
"""GLM-5 NVFP4 on GB300 (4x GB300 NVL4, tp=4).""" """GLM-5.2 NVFP4 on GB300 (4x GB300 NVL4, tp=4)."""
def test_glm5_nvfp4(self): def test_glm52_nvfp4(self):
variants = [ variants = [
ModelLaunchSettings( ModelLaunchSettings(
MODEL_PATH, MODEL_PATH,
@@ -58,7 +58,7 @@ class TestGlm5Nvfp4(unittest.TestCase):
run_combined_tests( run_combined_tests(
models=variants, models=variants,
test_name="GLM-5-NVFP4", test_name="GLM-5.2-NVFP4",
accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92), accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92),
performance_params=PerformanceTestParams( performance_params=PerformanceTestParams(
profile_dir="performance_profiles_gb300", profile_dir="performance_profiles_gb300",
test/registered/gb300/test_glm5_fp8.py
-72
View File
@@ -1,72 +0,0 @@
import unittest
from sglang.test.accuracy_test_runner import AccuracyTestParams
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.performance_test_runner import PerformanceTestParams
from sglang.test.run_combined_tests import run_combined_tests
from sglang.test.test_utils import ModelLaunchSettings
register_cuda_ci(
est_time=7200,
suite="nightly-4-gpu-gb300-glm5-fp8",
nightly=True,
disabled="not needed",
)
MODEL_PATH = "zai-org/GLM-5.1-FP8"
COMMON_ARGS = [
"--trust-remote-code",
"--reasoning-parser=glm45",
"--tool-call-parser=glm47",
"--mem-fraction-static=0.9",
"--enable-metrics",
]
MTP_ARGS = [
"--speculative-algorithm=EAGLE",
"--speculative-num-steps=3",
"--speculative-eagle-topk=1",
"--speculative-num-draft-tokens=4",
]
class TestGlm5Fp8(unittest.TestCase):
"""GLM-5.1 FP8 on GB300 (4x GB300 NVL4, tp=4)."""
def test_glm5_fp8(self):
variants = [
ModelLaunchSettings(
MODEL_PATH,
tp_size=4,
extra_args=COMMON_ARGS,
variant="TP4",
),
ModelLaunchSettings(
MODEL_PATH,
tp_size=4,
extra_args=COMMON_ARGS + ["--dp-size=4", "--enable-dp-attention"],
variant="TP4+DP4+DPA",
),
ModelLaunchSettings(
MODEL_PATH,
tp_size=4,
extra_args=COMMON_ARGS
+ ["--dp-size=4", "--enable-dp-attention"]
+ MTP_ARGS,
variant="TP4+DP4+DPA+MTP",
),
]
run_combined_tests(
models=variants,
test_name="GLM-5.1-FP8",
accuracy_params=AccuracyTestParams(dataset="gsm8k", baseline_accuracy=0.92),
performance_params=PerformanceTestParams(
profile_dir="performance_profiles_gb300",
),
)
if __name__ == "__main__":
unittest.main()
test/registered/models_e2e/test_deepseek_v32_fp4_mtp_dp.py
-68
View File
@@ -1,68 +0,0 @@
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
from sglang.test.test_utils import is_in_ci, write_github_step_summary
register_cuda_ci(est_time=400, stage="base-c", runner_config="4-gpu-b200")
DSV32_FP4_MODEL = "nvidia/DeepSeek-V3.2-NVFP4"
class TestDeepseekV32FP4DPSpec(GSM8KMixin, DefaultServerBase):
model = DSV32_FP4_MODEL
timeout = 1200
other_args = [
"--tp",
"4",
"--dp",
"4",
"--enable-dp-attention",
"--attention-backend",
"dsa",
"--moe-runner-backend",
"flashinfer_trtllm",
"--quantization",
"modelopt_fp4",
"--tool-call-parser",
"deepseekv32",
"--reasoning-parser",
"deepseek-v3",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
]
gsm8k_accuracy_thres = 0.93
gsm8k_num_questions = 500
gsm8k_num_threads = 500
gsm8k_num_shots = 20
gsm8k_accept_length_thres = 2.7
def test_z_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
acc_length, speed = send_one_prompt(args)
print(f"{acc_length=:.2f} {speed=:.2f}")
if is_in_ci():
write_github_step_summary(
f"### test_bs_1_speed (deepseek-v32 mtp dp)\n"
f"{acc_length=:.2f}\n"
f"{speed=:.2f} token/s\n"
)
self.assertGreater(acc_length, 2.7)
self.assertGreater(speed, 90)
if __name__ == "__main__":
unittest.main()
test/registered/models_e2e/test_deepseek_v32_fp4_mtp_tp.py
-73
View File
@@ -1,73 +0,0 @@
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
from sglang.test.test_utils import is_in_ci, write_github_step_summary
register_cuda_ci(est_time=400, stage="base-c", runner_config="4-gpu-b200")
DSV32_FP4_MODEL = "nvidia/DeepSeek-V3.2-NVFP4"
class TestDeepseekV32FP4TPSpec(GSM8KMixin, DefaultServerBase):
model = DSV32_FP4_MODEL
timeout = 1200
other_args = [
"--tp",
"4",
"--attention-backend",
"dsa",
"--moe-runner-backend",
"flashinfer_trtllm",
"--quantization",
"modelopt_fp4",
"--tool-call-parser",
"deepseekv32",
"--reasoning-parser",
"deepseek-v3",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
]
gsm8k_accuracy_thres = 0.93
gsm8k_num_questions = 500
gsm8k_num_threads = 500
gsm8k_num_shots = 20
gsm8k_accept_length_thres = 2.7
def test_z_bs_1_speed(self):
args = BenchArgs(
port=int(self.base_url.split(":")[-1]),
max_new_tokens=2048,
prompt=(
"Human: Think carefully before answering. Build a fully functional FastAPI todo server. "
"Start with a short design plan, then output the complete Python code, then show how to run it "
"and test three endpoints.\n\nAssistant:"
),
)
acc_length, speed = send_one_prompt(args)
print(f"{acc_length=:.2f} {speed=:.2f}")
if is_in_ci():
write_github_step_summary(
f"### test_bs_1_speed (deepseek-v32 mtp tp)\n"
f"{acc_length=:.2f}\n"
f"{speed=:.2f} token/s\n"
)
self.assertGreater(acc_length, 2.7)
self.assertGreater(speed, 150)
if __name__ == "__main__":
unittest.main()
test/registered/models_e2e/test_dsa_dsv32_dp_mtp.py
-28
View File
@@ -1,28 +0,0 @@
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.kits.spec_decoding_kit import SpecDecodingMixin
from sglang.test.server_fixtures.dsa_mtp_fixture import (
DsaMtpEvalConfigDefaults,
DsaMtpServerBase,
)
register_cuda_ci(
est_time=600,
stage="extra-b",
runner_config="8-gpu-h200",
)
class TestDeepseekV32DPMTP(
DsaMtpServerBase, DsaMtpEvalConfigDefaults, GSM8KMixin, SpecDecodingMixin
):
model = "deepseek-ai/DeepSeek-V3.2"
mem_fraction_static = 0.85
enable_dp_attention = True
bs_1_speed_thres = 90
if __name__ == "__main__":
unittest.main()
test/registered/models_e2e/test_dsa_dsv32_tp_mtp.py
-27
View File
@@ -1,27 +0,0 @@
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.kits.spec_decoding_kit import SpecDecodingMixin
from sglang.test.server_fixtures.dsa_mtp_fixture import (
DsaMtpEvalConfigDefaults,
DsaMtpServerBase,
)
register_cuda_ci(
est_time=400,
stage="extra-b",
runner_config="8-gpu-h200",
)
class TestDeepseekV32TPMTP(
DsaMtpServerBase, DsaMtpEvalConfigDefaults, GSM8KMixin, SpecDecodingMixin
):
model = "deepseek-ai/DeepSeek-V3.2"
mem_fraction_static = 0.85
bs_1_speed_thres = 180
if __name__ == "__main__":
unittest.main()
test/registered/models_e2e/test_dsa_glm5_dp_mtp.py → test/registered/models_e2e/test_dsa_glm52_dp_mtp.py
+2 -2
View File
@@ -15,10 +15,10 @@ register_cuda_ci(
) )
class TestGLM5DPMTP( class TestGLM52DPMTP(
DsaMtpServerBase, DsaMtpEvalConfigDefaults, GSM8KMixin, SpecDecodingMixin DsaMtpServerBase, DsaMtpEvalConfigDefaults, GSM8KMixin, SpecDecodingMixin
): ):
model = "zai-org/GLM-5-FP8" model = "zai-org/GLM-5.2-FP8"
mem_fraction_static = 0.88 mem_fraction_static = 0.88
enable_dp_attention = True enable_dp_attention = True
bs_1_speed_thres = 70 bs_1_speed_thres = 70
test/registered/models_e2e/test_dsa_glm5_hisparse.py → test/registered/models_e2e/test_dsa_glm52_hisparse.py
+5 -5
View File
@@ -9,18 +9,18 @@ from sglang.test.server_fixtures.default_fixture import DefaultServerBase
register_cuda_ci(est_time=720, stage="extra-b", runner_config="8-gpu-h200") register_cuda_ci(est_time=720, stage="extra-b", runner_config="8-gpu-h200")
GLM5_FP8_MODEL_PATH = "zai-org/GLM-5-FP8" GLM52_FP8_MODEL_PATH = "zai-org/GLM-5.2-FP8"
class TestGLM5HiSparse(DefaultServerBase, GSM8KMixin): class TestGLM52HiSparse(DefaultServerBase, GSM8KMixin):
"""GLM-5 FP8 with HiSparse (host-to-device sparse KV offload) on DSA decode. """GLM-5.2 FP8 with HiSparse (host-to-device sparse KV offload) on DSA decode.
HiSparse targets the high-concurrency regime and is not used together with HiSparse targets the high-concurrency regime and is not used together with
EAGLE MTP, so this variant runs without speculative decoding (unlike the EAGLE MTP, so this variant runs without speculative decoding (unlike the
DSA-MTP variants in test_dsa_glm5_{dp,tp}_mtp.py). DSA-MTP variants in test_dsa_glm52_{dp,tp}_mtp.py).
""" """
model = GLM5_FP8_MODEL_PATH model = GLM52_FP8_MODEL_PATH
other_args = [ other_args = [
"--trust-remote-code", "--trust-remote-code",
"--tp", "--tp",
test/registered/models_e2e/test_dsa_glm5_tp_mtp.py → test/registered/models_e2e/test_dsa_glm52_tp_mtp.py
+2 -2
View File
@@ -15,10 +15,10 @@ register_cuda_ci(
) )
class TestGLM5TPMTP( class TestGLM52TPMTP(
DsaMtpServerBase, DsaMtpEvalConfigDefaults, GSM8KMixin, SpecDecodingMixin DsaMtpServerBase, DsaMtpEvalConfigDefaults, GSM8KMixin, SpecDecodingMixin
): ):
model = "zai-org/GLM-5-FP8" model = "zai-org/GLM-5.2-FP8"
mem_fraction_static = 0.8 mem_fraction_static = 0.8
bs_1_speed_thres = 150 bs_1_speed_thres = 150
test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_nightly.py
+5 -5
View File
@@ -23,7 +23,7 @@ from sglang.test.test_utils import (
popen_launch_server, popen_launch_server,
) )
GLM5_MODEL = "zai-org/GLM-5.1-FP8" GLM5_MODEL = "zai-org/GLM-5.2-FP8"
GLM5_LAUNCH_TIMEOUT = 3600 GLM5_LAUNCH_TIMEOUT = 3600
register_cuda_ci(est_time=900, suite="nightly-8-gpu-h200", nightly=True) register_cuda_ci(est_time=900, suite="nightly-8-gpu-h200", nightly=True)
@@ -141,7 +141,7 @@ class AccuracyTwoPassMixin:
class TestGLM5HiRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase): class TestGLM5HiRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase):
"""GLM-5.1-FP8 + HiCache L3 (file backend), with HiRadixTree.""" """GLM-5.2-FP8 + HiCache L3 (file backend), with HiRadixTree."""
@classmethod @classmethod
def setUpClass(cls): def setUpClass(cls):
@@ -159,7 +159,7 @@ class TestGLM5HiRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase):
"--page-size", "--page-size",
"64", "64",
"--mem-fraction-static", "--mem-fraction-static",
"0.85", "0.8",
"--model-loader-extra-config", "--model-loader-extra-config",
'{"enable_multithread_load": true, "num_threads": 64}', '{"enable_multithread_load": true, "num_threads": 64}',
"--enable-hierarchical-cache", "--enable-hierarchical-cache",
@@ -189,7 +189,7 @@ class TestGLM5HiRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase):
class TestGLM5UnifiedRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase): class TestGLM5UnifiedRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase):
"""GLM-5.1-FP8 + HiCache L3 (file backend), with UnifiedRadixTree.""" """GLM-5.2-FP8 + HiCache L3 (file backend), with UnifiedRadixTree."""
@classmethod @classmethod
def setUpClass(cls): def setUpClass(cls):
@@ -207,7 +207,7 @@ class TestGLM5UnifiedRadixCacheL3Accuracy(AccuracyTwoPassMixin, CustomTestCase):
"--page-size", "--page-size",
"64", "64",
"--mem-fraction-static", "--mem-fraction-static",
"0.85", "0.8",
"--model-loader-extra-config", "--model-loader-extra-config",
'{"enable_multithread_load": true, "num_threads": 64}', '{"enable_multithread_load": true, "num_threads": 64}',
"--enable-hierarchical-cache", "--enable-hierarchical-cache",
test/run_suite.py
-1
View File
@@ -131,7 +131,6 @@ NIGHTLY_SUITES = {
# GB300 (4x GB300 NVL4) nightly suites # GB300 (4x GB300 NVL4) nightly suites
"nightly-4-gpu-gb300", "nightly-4-gpu-gb300",
"nightly-4-gpu-gb300-deepseek-v4-pro-fp4", "nightly-4-gpu-gb300-deepseek-v4-pro-fp4",
"nightly-4-gpu-gb300-glm5-fp8",
"nightly-4-gpu-gb300-glm5-nvfp4", "nightly-4-gpu-gb300-glm5-nvfp4",
"nightly-4-gpu-gb300-kimi-k25", "nightly-4-gpu-gb300-kimi-k25",
"nightly-4-gpu-gb300-kimi-k25-nvfp4", "nightly-4-gpu-gb300-kimi-k25-nvfp4",