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

ci: tag-gated nightly migration — foundation + 40 whole-file moves (#24725)

Browse Source 
Co-authored-by: hnyls2002 <lsyincs@gmail.com>
Co-authored-by: Liangsheng Yin <hnyls2002@gmail.com>
This commit is contained in:
Alison Shao
2026-05-14 16:11:15 -07:00
committed by GitHub
co-authored by hnyls2002 Liangsheng Yin
parent 67096f48bf
commit ba214ef3d3
78 changed files with 2254 additions and 2131 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/quant/test_deepseek_v3_fp4_4gpu.py
+3 -127
View File
@@ -1,11 +1,9 @@
import os
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
@@ -14,137 +12,15 @@ from sglang.test.test_utils import (
write_github_step_summary,
)
# Per-commit: SymmetricMemory variant only.
# - TestDeepseekV3FP4 (TRTLLM) archived to test/manual/quant/test_deepseek_v3_fp4_4gpu_trtllm.py
# - TestDeepseekV3FP4CutlassMoE moved to test_deepseek_v3_fp4_4gpu_extra.py
register_cuda_ci(est_time=960, stage="stage-c", runner_config="4-gpu-b200")
FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
SERVER_LAUNCH_TIMEOUT = 1200
class TestDeepseekV3FP4(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = FULL_DEEPSEEK_V3_FP4_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--tp",
"4",
"--attention-backend",
"trtllm_mla",
"--moe-runner-backend",
"flashinfer_trtllm",
"--quantization",
"modelopt_fp4",
"--kv-cache-dtype",
"fp8_e4m3",
"--model-loader-extra-config",
'{"enable_multithread_load": true,"num_threads": 64}',
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=SERVER_LAUNCH_TIMEOUT,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(
self,
): # Append an "a" to make this test run first (alphabetically) to warm up the server
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=1319,
num_threads=1319,
num_shots=8,
)
metrics = run_eval(args)
print(f"{metrics=}")
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v3-fp4)\n" f'{metrics["score"]=:.3f}\n'
)
self.assertGreater(metrics["score"], 0.93)
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
_, speed = send_one_prompt(args)
print(f"{speed=:.2f}")
if is_in_ci():
write_github_step_summary(
f"### test_bs_1_speed (deepseek-v3-fp4)\n" f"{speed=:.2f} token/s\n"
)
self.assertGreater(speed, 120)
class TestDeepseekV3FP4CutlassMoE(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = FULL_DEEPSEEK_V3_FP4_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--tp",
"4",
"--ep",
"4",
"--attention-backend",
"trtllm_mla",
"--moe-runner-backend",
"flashinfer_cutlass",
"--quantization",
"modelopt_fp4",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=SERVER_LAUNCH_TIMEOUT,
other_args=other_args,
env={
**os.environ,
"SGLANG_MOE_NVFP4_DISPATCH": "1", # Enable nvfp4 all gather
},
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(
self,
): # Append an "a" to make this test run first (alphabetically) to warm up the server
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=1319,
num_threads=1319,
num_shots=8,
)
metrics = run_eval(args)
print(f"{metrics=}")
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v3-fp4-cutlass-moe)\n"
f'{metrics["score"]=:.3f}\n'
)
self.assertGreater(metrics["score"], 0.93)
class TestDeepseekV3FP4SymmetricMemory(CustomTestCase):
@classmethod
def setUpClass(cls):
test/registered/quant/test_deepseek_v3_fp4_4gpu_extra.py
+87
View File
@@ -0,0 +1,87 @@
"""Extra: DeepSeek-V3 FP4 with FlashInfer Cutlass MoE backend.
Sibling per-commit file (test_deepseek_v3_fp4_4gpu.py) keeps the
SymmetricMemory variant.
"""
import os
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
register_cuda_ci(est_time=960, stage="extra-b", runner_config="4-gpu-b200")
FULL_DEEPSEEK_V3_FP4_MODEL_PATH = "nvidia/DeepSeek-V3-0324-FP4"
SERVER_LAUNCH_TIMEOUT = 1200
class TestDeepseekV3FP4CutlassMoE(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = FULL_DEEPSEEK_V3_FP4_MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--tp",
"4",
"--ep",
"4",
"--attention-backend",
"trtllm_mla",
"--moe-runner-backend",
"flashinfer_cutlass",
"--quantization",
"modelopt_fp4",
"--model-loader-extra-config",
'{"enable_multithread_load": true}',
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=SERVER_LAUNCH_TIMEOUT,
other_args=other_args,
env={
**os.environ,
"SGLANG_MOE_NVFP4_DISPATCH": "1", # Enable nvfp4 all gather
},
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_a_gsm8k(
self,
): # Append an "a" to make this test run first (alphabetically) to warm up the server
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="gsm8k",
api="completion",
max_tokens=512,
num_examples=1319,
num_threads=1319,
num_shots=8,
)
metrics = run_eval(args)
print(f"{metrics=}")
if is_in_ci():
write_github_step_summary(
f"### test_gsm8k (deepseek-v3-fp4-cutlass-moe)\n"
f'{metrics["score"]=:.3f}\n'
)
self.assertGreater(metrics["score"], 0.93)
if __name__ == "__main__":
unittest.main()
test/registered/quant/test_fp8_blockwise_gemm.py
+1 -1
View File
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
try_cached_model,
)
register_cuda_ci(est_time=430, suite="nightly-4-gpu-b200", nightly=True)
register_cuda_ci(est_time=430, stage="extra-b", runner_config="4-gpu-b200")
MODEL_PATH = "Qwen/Qwen3-4B-Instruct-2507-FP8"
MXFP8_MODEL_PATH = "zianglih/Qwen3-4B-Instruct-2507-MXFP8"
test/registered/quant/test_fp8_gemm_sm120.py
+1 -1
View File
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
try_cached_model,
)
register_cuda_ci(est_time=146, stage="stage-b", runner_config="1-gpu-small")
register_cuda_ci(est_time=146, stage="extra-a", runner_config="1-gpu-small")
PERTENSOR_MODEL_PATH = "nvidia/Llama-3.1-8B-Instruct-FP8"
BLOCKWISE_MODEL_PATH = "Qwen/Qwen3-4B-Instruct-2507-FP8"
test/registered/quant/test_fp8kv_triton.py
+1 -1
View File
@@ -12,7 +12,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=73, stage="stage-b", runner_config="1-gpu-large")
register_cuda_ci(est_time=73, stage="extra-a", runner_config="1-gpu-large")
class TestFP8KVCacheTritonBackend(CustomTestCase):
test/registered/quant/test_gptqmodel_dynamic.py
+1 -1
View File
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=100, stage="stage-b", runner_config="1-gpu-large")
register_cuda_ci(est_time=100, stage="extra-a", runner_config="1-gpu-large")
def check_quant_method(model_path: str, use_marlin_kernel: bool):
test/registered/quant/test_w8a8_quantization.py
+1 -1
View File
@@ -14,7 +14,7 @@ from sglang.test.test_utils import (
popen_launch_server,
)
register_cuda_ci(est_time=232, stage="stage-b", runner_config="1-gpu-large")
register_cuda_ci(est_time=232, stage="extra-a", runner_config="1-gpu-large")
class BaseW8A8Test(CustomTestCase):