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

ci: prune per-commit CUDA tests — move 25 files + 13 testcases to test/manual/ (#24721)

Browse Source
This commit is contained in:
Alison Shao
2026-05-08 15:53:23 -07:00
committed by GitHub
parent aefd8e257f
commit 5fbec0e445
45 changed files with 818 additions and 525 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/manual/models/test_nvidia_nemotron_3_nano_archived.py
+47
View File
@@ -0,0 +1,47 @@
"""Archived test classes split out of test/registered/models/test_nvidia_nemotron_3_nano.py.
Originally registered with `register_cuda_ci(...)`. Moved here as part of
the per-commit pruning effort to keep the code reachable manually.
Run with `python3 test/manual/models/test_nvidia_nemotron_3_nano_archived.py`.
"""
import unittest
from sglang.test.kits.lm_eval_kit import LMEvalMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
NEMOTRON_3_NANO_THINKING_ARGS = [
"--trust-remote-code",
"--tool-call-parser",
"qwen3_coder",
"--reasoning-parser",
"deepseek-r1",
]
class TestNvidiaNemotron3Nano30BBF16(LMEvalMixin, DefaultServerBase):
"""Test Nemotron-3-Nano-30B BF16 model with lm-eval GSM8K evaluation."""
model = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
model_config_name = "lm_eval_configs/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.yaml"
other_args = [
"--tp-size",
"2",
] + NEMOTRON_3_NANO_THINKING_ARGS
class TestNvidiaNemotron3Nano30BBF16FlashInfer(LMEvalMixin, DefaultServerBase):
"""Test Nemotron-3-Nano-30B BF16 model with lm-eval GSM8K evaluation using flashinfer mamba backend."""
model = "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16"
model_config_name = "lm_eval_configs/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16.yaml"
other_args = [
"--tp-size",
"2",
"--mamba-backend",
"flashinfer",
] + NEMOTRON_3_NANO_THINKING_ARGS
if __name__ == "__main__":
unittest.main()
test/manual/models/test_nvidia_nemotron_nano_v2.py
+99
View File
@@ -0,0 +1,99 @@
import unittest
from sglang.srt.utils import is_blackwell
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
register_cuda_ci(est_time=249, suite="stage-b-test-2-gpu-large")
class TestNvidiaNemotronNanoV2BF16(GSM8KMixin, DefaultServerBase):
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
gsm8k_accuracy_thres = 0.87
other_args = ["--max-mamba-cache-size", "256"]
class TestNvidiaNemotronNanoV2BF16PP(GSM8KMixin, DefaultServerBase):
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
gsm8k_accuracy_thres = 0.87
other_args = ["--max-mamba-cache-size", "256", "--pp-size", "2"]
class TestNvidiaNemotronNanoV2FP8(GSM8KMixin, DefaultServerBase):
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-FP8"
other_args = ["--max-mamba-cache-size", "256"]
@unittest.skipIf(not is_blackwell(), "NVFP4 only supported on blackwell")
class TestNvidiaNemotronNanoV2NVFP4(GSM8KMixin, DefaultServerBase):
gsm8k_accuracy_thres = 0.855
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2-NVFP4"
other_args = ["--max-mamba-cache-size", "256"]
@unittest.skip(
"STANDALONE speculative decoding does not yet support target and draft models "
"with different hidden sizes (Nemotron-9B: 4480, Llama-3.2-1B: 2048)"
)
class TestNvidiaNemotronNanoV2SpeculativeDecoding(GSM8KMixin, DefaultServerBase):
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
other_args = [
"--speculative-algorithm",
"STANDALONE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"3",
"--speculative-num-draft-tokens",
"5",
"--speculative-draft-model-path",
"meta-llama/Llama-3.2-1B",
"--speculative-draft-load-format",
"dummy",
"--max-running-requests",
"8",
"--max-total-tokens",
"2048",
"--json-model-override-args",
'{"vocab_size": 131072}',
]
@unittest.skip(
"STANDALONE speculative decoding does not yet support target and draft models "
"with different hidden sizes (Nemotron-9B: 4480, Llama-3.2-1B: 2048)"
)
class TestNvidiaNemotronNanoV2SpeculativeDecodingBF16Cache(
GSM8KMixin, DefaultServerBase
):
gsm8k_accuracy_thres = 0.87
model = "nvidia/NVIDIA-Nemotron-Nano-9B-v2"
other_args = [
"--speculative-algorithm",
"STANDALONE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"3",
"--speculative-num-draft-tokens",
"5",
"--speculative-draft-model-path",
"meta-llama/Llama-3.2-1B",
"--speculative-draft-load-format",
"dummy",
"--max-running-requests",
"8",
"--max-total-tokens",
"2048",
"--json-model-override-args",
'{"vocab_size": 131072}',
"--mamba-ssm-dtype",
"bfloat16",
]
if __name__ == "__main__":
unittest.main()
test/manual/models/test_nvidia_nemotron_nano_v2_vl.py
+33
View File
@@ -0,0 +1,33 @@
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.mmmu_vlm_kit import MMMUMixin
from sglang.test.server_fixtures.default_fixture import DefaultServerBase
from sglang.test.server_fixtures.mmmu_fixture import MMMUServerBase
# NVIDIA Nemotron Nano V2 VL model tests (CUDA only)
# GSM8k + MMMU evaluation
register_cuda_ci(est_time=256, suite="stage-b-test-1-gpu-large")
MODEL = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"
class TestNvidiaNemotronNanoV2VLTextOnly(GSM8KMixin, DefaultServerBase):
gsm8k_accuracy_thres = 0.85
model = MODEL
other_args = ["--max-mamba-cache-size", "256", "--trust-remote-code"]
class TestNvidiaNemotronNanoV2VLMMMU(MMMUMixin, MMMUServerBase):
accuracy = 0.444
model = MODEL
other_args = ["--max-mamba-cache-size", "128", "--trust-remote-code"]
mmmu_args = ["--limit=0.1"]
"""`--limit=0.1`: 10 percent of each task - this is fine for testing since the nominal result isn't interesting - this run is just to prevent relative regressions."""
if __name__ == "__main__":
unittest.main()
test/manual/models/test_qwen_models.py
+83
View File
@@ -0,0 +1,83 @@
# Qwen model tests
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
register_cuda_ci(est_time=108, suite="stage-b-test-1-gpu-small")
register_amd_ci(est_time=130, suite="stage-b-test-1-gpu-small-amd")
class TestQwen2(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "Qwen/Qwen2-7B-Instruct"
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreater(metrics["score"], 0.78)
class TestQwen2FP8(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "neuralmagic/Qwen2-7B-Instruct-FP8"
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=200,
num_threads=128,
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreater(metrics["score"], 0.78)
if __name__ == "__main__":
unittest.main()