chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 11 Packages Projects Releases Wiki Activity

Migrate 4-GPU/8-GPU workflow jobs to stage-c and add CI registry decorators (#17299)

Browse Source
This commit is contained in:
Alison Shao
2026-01-31 22:37:22 -08:00
committed by GitHub
parent 95180484e9
commit a0bae4c343
39 changed files with 189 additions and 284 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py
+164
View File
@@ -0,0 +1,164 @@
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.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_DEEPSEEK_NVFP4_MODEL_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
try_cached_model,
)
register_cuda_ci(est_time=1800, suite="stage-c-test-4-gpu-gb200")
class TestDeepseekR1Nvfp4CuteDSLDeepEP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = try_cached_model(DEFAULT_DEEPSEEK_NVFP4_MODEL_FOR_TEST)
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--disable-radix-cache",
"--mem-fraction-static",
"0.8",
"--max-prefill-tokens",
"16384",
"--max-running-requests",
"256",
"--chunked-prefill-size",
"1024",
"--tp",
"4",
"--dp",
"4",
"--ep",
"4",
"--moe-dense-tp-size",
"1",
"--enable-dp-attention",
"--quantization",
"modelopt_fp4",
"--attention-backend",
"trtllm_mla",
"--moe-a2a-backend",
"deepep",
"--moe-runner-backend",
"flashinfer_cutedsl",
"--deepep-mode",
"low_latency",
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
env={
**os.environ,
"SGLANG_DEEPEP_BF16_DISPATCH": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "256",
"SGLANG_MOE_NVFP4_DISPATCH": "0",
},
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=512,
parallel=512,
max_new_tokens=512,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
print(f"Eval accuracy of GSM8K: {metrics=}")
self.assertGreater(metrics["accuracy"], 0.92)
class TestDummyWithSBO(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = try_cached_model(DEFAULT_DEEPSEEK_NVFP4_MODEL_FOR_TEST)
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--disable-radix-cache",
"--mem-fraction-static",
"0.05",
"--max-prefill-tokens",
"16384",
"--max-running-requests",
"256",
"--chunked-prefill-size",
"1024",
"--cuda-graph-bs",
"64",
"--tp",
"4",
"--dp",
"4",
"--ep",
"4",
"--moe-dense-tp-size",
"1",
"--enable-dp-attention",
"--quantization",
"modelopt_fp4",
"--attention-backend",
"trtllm_mla",
"--moe-a2a-backend",
"deepep",
"--moe-runner-backend",
"flashinfer_cutedsl",
"--deepep-mode",
"low_latency",
"--json-model-override-args",
'{"num_hidden_layers": 1, "first_k_dense_replace": 0, "n_routed_experts": 24}',
"--enable-single-batch-overlap",
"--load-format",
"dummy",
]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
env={
**os.environ,
"SGLANG_DEEPEP_BF16_DISPATCH": "1",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "256",
"SGLANG_MOE_NVFP4_DISPATCH": "0",
},
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=0,
data_path=None,
num_questions=512,
parallel=512,
max_new_tokens=16,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval_few_shot_gsm8k(args)
print(f"Eval accuracy of GSM8K: {metrics=}")
if __name__ == "__main__":
unittest.main()
test/registered/4-gpu-models/test_gpt_oss_4gpu.py
+40
View File
@@ -0,0 +1,40 @@
import unittest
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.gpt_oss_common import BaseTestGptOss
register_cuda_ci(est_time=300, suite="stage-c-test-4-gpu-h100")
register_cuda_ci(est_time=300, suite="stage-c-test-4-gpu-b200")
class TestGptOss4Gpu(BaseTestGptOss):
def test_bf16_120b(self):
self.run_test(
model_variant="120b",
quantization="bf16",
expected_score_of_reasoning_effort={
"low": 0.60,
},
other_args=["--tp", "4", "--cuda-graph-max-bs", "200"],
)
def test_mxfp4_120b(self):
self.run_test(
model_variant="120b",
quantization="mxfp4",
expected_score_of_reasoning_effort={
"low": 0.60,
},
other_args=[
"--tp",
"4",
"--cuda-graph-max-bs",
"200",
"--mem-fraction-static",
"0.93",
],
)
if __name__ == "__main__":
unittest.main()
test/registered/4-gpu-models/test_qwen3_next_models.py
+130
View File
@@ -0,0 +1,130 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.kl_test_utils import (
test_input_output_logprobs_match_decode_cache_hit_helper,
test_input_output_logprobs_match_prefill_cache_hit_helper,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
register_cuda_ci(est_time=350, suite="stage-c-test-4-gpu-h100")
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
def send_request_helper(base_url: str, text: str):
response = requests.post(
base_url + "/generate",
json={
"text": text,
"sampling_params": {
"max_new_tokens": 1,
},
},
)
return response.json()
class TestQwen3Next(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
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=[
"--tp-size",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_prefix_cache_branching(self):
print("running test_prefix_cache_branching")
requests.get(self.base_url + "/flush_cache")
branching_pos = 257
text_prefix = "hi" * branching_pos
suffix_list = ["this" * 256, "here" * 256, "that" * 256]
cache_hit_list = [False, False, True]
# First request only prefill the entire sequence
# Second request won't have cache hit, but will cache the branching point
# Third request will have cache hit on the branching point
for i, (suffix, cache_hit) in enumerate(
zip(suffix_list, cache_hit_list, strict=True)
):
result = send_request_helper(self.base_url, text_prefix + suffix)
cached_tokens = result["meta_info"]["cached_tokens"]
if cache_hit:
expected_cached_tokens = branching_pos // 64 * 64
assert (
cached_tokens == expected_cached_tokens
), f"{i=}, {cache_hit=}, {cached_tokens=} is not equal to {expected_cached_tokens=}, {branching_pos=}"
else:
assert (
cached_tokens == 0
), f"{i=}, {cache_hit=}, {cached_tokens=} is not 0"
print("test_prefix_cache_branching passed")
if __name__ == "__main__":
unittest.main()
test/registered/4-gpu-models/test_qwen3_next_models_mtp.py
+215
View File
@@ -0,0 +1,215 @@
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.few_shot_gsm8k import run_eval
from sglang.test.kl_test_utils import (
test_input_output_logprobs_match_decode_cache_hit_helper,
test_input_output_logprobs_match_prefill_cache_hit_helper,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
register_cuda_ci(est_time=500, suite="stage-c-test-4-gpu-h100")
QWEN3_NEXT_MODEL = "Qwen/Qwen3-Next-80B-A3B-Instruct"
ACC_THRESHOLDS = {
QWEN3_NEXT_MODEL: {"kl_div": 0.0025, "gsm8k": 0.93},
}
# MTP has higher KL divergence threshold
ACC_THRESHOLDS_MTP = {
QWEN3_NEXT_MODEL: {"kl_div": 0.008, "gsm8k": 0.93},
}
def send_request_helper(base_url: str, text: str):
response = requests.post(
base_url + "/generate",
json={
"text": text,
"sampling_params": {
"max_new_tokens": 1,
},
},
)
return response.json()
class TestQwen3NextMTP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
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=[
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"no_buffer",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS[self.model]["gsm8k"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS,
self.model,
max_samples=32,
max_new_tokens=512,
)
class TestQwen3NextMTPTopk(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = QWEN3_NEXT_MODEL
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=[
"--trust-remote-code",
"--speculative-algorithm",
"NEXTN",
"--speculative-num-steps",
"5",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"8",
"--mem-fraction-static",
"0.8",
"--tp",
"4",
"--chunked-prefill-size",
"2048",
"--mamba-scheduler-strategy",
"extra_buffer",
"--mamba-track-interval",
"128",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
metrics = run_eval(args)
print(f"{metrics=}")
self.assertGreaterEqual(
metrics["accuracy"], ACC_THRESHOLDS_MTP[self.model]["gsm8k"]
)
def test_input_output_logprobs_match_prefill_cache_hit(self):
test_input_output_logprobs_match_prefill_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS_MTP,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_input_output_logprobs_match_decode_cache_hit(self):
test_input_output_logprobs_match_decode_cache_hit_helper(
self.base_url,
ACC_THRESHOLDS_MTP,
self.model,
max_samples=32,
max_new_tokens=512,
)
def test_prefix_cache_branching(self):
print("running test_prefix_cache_branching")
requests.get(self.base_url + "/flush_cache")
branching_pos = 257
text_prefix = "hi" * branching_pos
suffix_list = ["this" * 256, "here" * 256, "that" * 256]
cache_hit_list = [False, False, True]
# First request only prefill the entire sequence
# Second request won't have cache hit, but will cache the branching point
# Third request will have cache hit on the branching point
for i, (suffix, cache_hit) in enumerate(
zip(suffix_list, cache_hit_list, strict=True)
):
result = send_request_helper(self.base_url, text_prefix + suffix)
cached_tokens = result["meta_info"]["cached_tokens"]
if cache_hit:
expected_cached_tokens = branching_pos // 64 * 64
assert (
cached_tokens == expected_cached_tokens
), f"{i=}, {cache_hit=}, {cached_tokens=} is not equal to {expected_cached_tokens=}, {branching_pos=}"
else:
assert (
cached_tokens == 0
), f"{i=}, {cache_hit=}, {cached_tokens=} is not 0"
print("test_prefix_cache_branching passed")
if __name__ == "__main__":
unittest.main()