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

ci: migrate MLA tests to test/registered/mla/ (#15798)

Browse Source
This commit is contained in:
Alison Shao
2025-12-24 23:29:30 -08:00
committed by GitHub
parent ccf2330b28
commit 72a980c6d5
8 changed files with 37 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/srt/run_suite.py
-9
View File
@@ -58,7 +58,6 @@ suites = {
TestFile("test_constrained_decoding.py", 111),
TestFile("test_eval_fp8_accuracy.py", 250),
TestFile("test_external_models.py", 30),
TestFile("test_flashmla.py", 230),
TestFile("test_fp8_utils.py", 9),
TestFile("rotary_embedding/test_mrope.py", 10),
TestFile("test_fused_moe.py", 80),
@@ -71,11 +70,6 @@ suites = {
TestFile("test_mamba_unittest.py", 9),
TestFile("test_metrics.py", 32),
TestFile("test_metrics_utils.py", 1),
TestFile("test_mla.py", 194),
TestFile("test_mla_deepseek_v3.py", 442),
TestFile("test_mla_flashinfer.py", 302),
TestFile("test_mla_fp8.py", 77),
TestFile("test_mla_int8_deepseek_v3.py", 300),
TestFile("test_model_hooks.py", 6),
TestFile("test_modelopt_loader.py", 11),
TestFile("test_multi_tokenizer.py", 230),
@@ -247,8 +241,6 @@ suite_amd = {
TestFile("test_jinja_template_utils.py", 1),
TestFile("test_metrics.py", 32),
TestFile("test_metrics_utils.py", 1),
# TestFile("test_mla.py", 242), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107
# TestFile("test_mla_deepseek_v3.py", 221), # Temporarily disabled, see https://github.com/sgl-project/sglang/issues/12574
TestFile("test_no_chunked_prefill.py", 108),
TestFile("test_page_size.py", 60),
TestFile("test_penalty.py", 180),
@@ -275,7 +267,6 @@ suite_amd = {
],
"per-commit-amd-mi35x": [
TestFile("test_gpt_oss_1gpu.py", 750),
TestFile("test_mla.py", 242),
],
"per-commit-2-gpu-amd": [
# TestFile("lora/test_lora_tp.py", 116), # Disabled temporarily, see https://github.com/sgl-project/sglang/issues/13107. Moved to test/registered/lora/
test/srt/test_flashmla.py
-134
View File
@@ -1,134 +0,0 @@
"""
Usage:
python3 test/srt/test_flashmla.py
"""
import unittest
from types import SimpleNamespace
import requests
import torch
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_MLA,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
class TestFlashMLAAttnBackend(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MODEL_NAME_FOR_TEST_MLA
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"2",
"--attention-backend",
"flashmla",
]
)
# Use longer timeout for DeepGEMM JIT compilation which can take 10-20 minutes
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 2,
other_args=other_args,
)
@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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
class TestFlashMLAMTP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"4",
"--disable-radix",
"--enable-torch-compile",
"--torch-compile-max-bs",
"1",
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
"lmsys/sglang-ci-dsv3-test-NextN",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"3",
"--attention-backend",
"flashmla",
]
)
# Use longer timeout for DeepGEMM JIT compilation which can take 10-20 minutes
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 2,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
requests.get(self.base_url + "/flush_cache")
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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
server_info = requests.get(self.base_url + "/get_server_info")
print(f"{server_info=}")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 2.4)
if __name__ == "__main__":
unittest.main()
test/srt/test_mla.py
-52
View File
@@ -1,52 +0,0 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MLA_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
class TestMLA(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
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",
"--enable-torch-compile",
"--torch-compile-max-bs",
"4",
"--chunked-prefill-size",
"256",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_mgsm_en(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mgsm_en",
num_examples=None,
num_threads=1024,
)
metrics = run_eval(args)
self.assertGreater(metrics["score"], 0.8)
if __name__ == "__main__":
unittest.main()
test/srt/test_mla_deepseek_v3.py
-193
View File
@@ -1,193 +0,0 @@
import os
import unittest
from types import SimpleNamespace
import requests
from sglang.srt.utils import is_cuda, is_hip, kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
)
class TestMLADeepseekV3(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code", "--chunked-prefill-size", "256"]
if is_cuda():
other_args.extend(["--enable-torch-compile", "--cuda-graph-max-bs", "2"])
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.62)
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
class TestMLADeepseekV3DisableFusedFunc(CustomTestCase):
@classmethod
def setUpClass(cls):
os.environ["SGLANG_CI_DISABLE_MOE_FUSED_FUNC"] = "1"
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code", "--chunked-prefill-size", "256"]
if is_cuda():
other_args.extend(["--cuda-graph-max-bs", "2"])
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.62)
@unittest.skipIf(is_hip(), "FA is not available.")
class TestMLADeepseekV3Fa3Fp8Kvcache(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--chunked-prefill-size",
"256",
"--kv-cache-dtype",
"fp8_e4m3",
]
if is_cuda():
other_args.extend(["--attention-backend", "fa3"])
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.62)
class TestDeepseekV3MTP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--cuda-graph-max-bs",
"2",
"--disable-radix",
"--enable-torch-compile",
"--torch-compile-max-bs",
"1",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"4",
]
# This test runs first (alphabetically) and needs longer timeout for
# DeepGEMM JIT compilation which is required for DeepSeek-V3's FP8 MoE layers
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH * 2,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
requests.get(self.base_url + "/flush_cache")
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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
server_info = requests.get(self.base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 2.5)
if __name__ == "__main__":
unittest.main()
test/srt/test_mla_flashinfer.py
-124
View File
@@ -1,124 +0,0 @@
import unittest
from types import SimpleNamespace
import requests
import torch
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
class TestFlashinferMLA(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--enable-torch-compile",
"--cuda-graph-max-bs",
"4",
"--attention-backend",
"flashinfer",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.615)
class TestFlashinferMLAMTP(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "lmsys/sglang-ci-dsv3-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"4",
"--enable-torch-compile",
"--torch-compile-max-bs",
"1",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"3",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"4",
"--attention-backend",
"flashinfer",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
requests.get(self.base_url + "/flush_cache")
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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
server_info = requests.get(self.base_url + "/get_server_info")
print(f"{server_info=}")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 2.5)
if __name__ == "__main__":
unittest.main()
test/srt/test_mla_fp8.py
-49
View File
@@ -1,49 +0,0 @@
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_MLA_FP8_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
class TestMLA(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = DEFAULT_MLA_FP8_MODEL_NAME_FOR_TEST
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",
"--kv-cache-dtype",
"fp8_e5m2",
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_mgsm_en(self):
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mgsm_en",
num_examples=None,
num_threads=1024,
)
metrics = run_eval(args)
assert metrics["score"] >= 0.8
if __name__ == "__main__":
unittest.main()
test/srt/test_mla_int8_deepseek_v3.py
-229
View File
@@ -1,229 +0,0 @@
import unittest
from types import SimpleNamespace
import requests
import torch
from sglang.srt.utils import kill_process_tree
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
)
class TestMLADeepseekV3ChannelInt8(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "sgl-project/sglang-ci-dsv3-channel-int8-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"16",
"--enable-torch-compile",
"--torch-compile-max-bs",
"2",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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_few_shot_gsm8k(args)
print(metrics)
self.assertGreaterEqual(metrics["accuracy"], 0.61)
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
class TestDeepseekV3MTPChannelInt8(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "sgl-project/sglang-ci-dsv3-channel-int8-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"16",
"--enable-torch-compile",
"--torch-compile-max-bs",
"2",
"--speculative-algorithm",
"EAGLE",
"--speculative-draft-model-path",
"sgl-project/sglang-ci-dsv3-channel-int8-test-NextN",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"4",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
requests.get(self.base_url + "/flush_cache")
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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
server_info = requests.get(self.base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 2.5)
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
class TestMLADeepseekV3BlockInt8(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "sgl-project/sglang-ci-dsv3-block-int8-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"16",
"--enable-torch-compile",
"--torch-compile-max-bs",
"2",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.62)
class TestDeepseekV3MTPBlockInt8(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = "sgl-project/sglang-ci-dsv3-block-int8-test"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = ["--trust-remote-code"]
if torch.cuda.is_available() and torch.version.cuda:
other_args.extend(
[
"--cuda-graph-max-bs",
"16",
"--enable-torch-compile",
"--torch-compile-max-bs",
"2",
"--speculative-algorithm",
"EAGLE",
"--speculative-num-steps",
"2",
"--speculative-eagle-topk",
"4",
"--speculative-num-draft-tokens",
"4",
]
)
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
requests.get(self.base_url + "/flush_cache")
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_few_shot_gsm8k(args)
print(metrics)
self.assertGreater(metrics["accuracy"], 0.60)
server_info = requests.get(self.base_url + "/get_server_info")
avg_spec_accept_length = server_info.json()["internal_states"][0][
"avg_spec_accept_length"
]
print(f"{avg_spec_accept_length=}")
self.assertGreater(avg_spec_accept_length, 2.5)
if __name__ == "__main__":
unittest.main()