ci(test): migrate OpenAI server tests to registered CI system (#16326)
Co-authored-by: Kangyan-Zhou <zky314343421@gmail.com>
This commit is contained in:
Alison Shao
co-authored by
Kangyan-Zhou
Kangyan-Zhou
parent
6e3fff1352
commit
0cbd8f3247
@@ -0,0 +1,107 @@
|
||||
"""
|
||||
python3 -m unittest openai_server.validation.test_large_max_new_tokens.TestLargeMaxNewTokens.test_chat_completion
|
||||
"""
|
||||
|
||||
import os
|
||||
import time
|
||||
import unittest
|
||||
from concurrent.futures import ThreadPoolExecutor
|
||||
|
||||
import openai
|
||||
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.srt.utils.hf_transformers_utils import get_tokenizer
|
||||
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
STDERR_FILENAME,
|
||||
STDOUT_FILENAME,
|
||||
CustomTestCase,
|
||||
popen_launch_server,
|
||||
)
|
||||
|
||||
register_cuda_ci(est_time=41, suite="stage-b-test-small-1-gpu")
|
||||
register_amd_ci(est_time=41, suite="stage-b-test-small-1-gpu")
|
||||
|
||||
|
||||
class TestLargeMaxNewTokens(CustomTestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = DEFAULT_SMALL_MODEL_NAME_FOR_TEST
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.api_key = "sk-123456"
|
||||
|
||||
cls.stdout = open(STDOUT_FILENAME, "w")
|
||||
cls.stderr = open(STDERR_FILENAME, "w")
|
||||
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
api_key=cls.api_key,
|
||||
other_args=(
|
||||
"--max-total-token",
|
||||
"1536",
|
||||
"--context-len",
|
||||
"8192",
|
||||
"--decode-log-interval",
|
||||
"2",
|
||||
),
|
||||
env={"SGLANG_CLIP_MAX_NEW_TOKENS_ESTIMATION": "256", **os.environ},
|
||||
return_stdout_stderr=(cls.stdout, cls.stderr),
|
||||
)
|
||||
cls.base_url += "/v1"
|
||||
cls.tokenizer = get_tokenizer(DEFAULT_SMALL_MODEL_NAME_FOR_TEST)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
kill_process_tree(cls.process.pid)
|
||||
cls.stdout.close()
|
||||
cls.stderr.close()
|
||||
os.remove(STDOUT_FILENAME)
|
||||
os.remove(STDERR_FILENAME)
|
||||
|
||||
def run_chat_completion(self):
|
||||
client = openai.Client(api_key=self.api_key, base_url=self.base_url)
|
||||
response = client.chat.completions.create(
|
||||
model=self.model,
|
||||
messages=[
|
||||
{"role": "system", "content": "You are a helpful AI assistant"},
|
||||
{
|
||||
"role": "user",
|
||||
"content": "Please repeat the world 'hello' for 10000 times.",
|
||||
},
|
||||
],
|
||||
temperature=0,
|
||||
)
|
||||
return response
|
||||
|
||||
def test_chat_completion(self):
|
||||
num_requests = 4
|
||||
|
||||
futures = []
|
||||
with ThreadPoolExecutor(num_requests) as executor:
|
||||
# Send multiple requests
|
||||
for i in range(num_requests):
|
||||
futures.append(executor.submit(self.run_chat_completion))
|
||||
|
||||
# Ensure that they are running concurrently
|
||||
pt = 0
|
||||
while pt >= 0:
|
||||
time.sleep(5)
|
||||
lines = open(STDERR_FILENAME).readlines()
|
||||
for line in lines[pt:]:
|
||||
print(line, end="", flush=True)
|
||||
if f"#running-req: {num_requests}" in line:
|
||||
all_requests_running = True
|
||||
pt = -1
|
||||
break
|
||||
pt += 1
|
||||
|
||||
assert all_requests_running
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,84 @@
|
||||
import unittest
|
||||
|
||||
from sglang.srt.sampling.sampling_params import MAX_LEN, get_max_seq_length
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
|
||||
from sglang.test.kits.matched_stop_kit import MatchedStopMixin
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_MODEL_NAME_FOR_TEST,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
CustomTestCase,
|
||||
popen_launch_server,
|
||||
)
|
||||
|
||||
register_cuda_ci(est_time=40, suite="stage-b-test-small-1-gpu")
|
||||
register_amd_ci(est_time=60, suite="stage-b-test-small-1-gpu")
|
||||
|
||||
|
||||
class TestMatchedStop(CustomTestCase, MatchedStopMixin):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = DEFAULT_MODEL_NAME_FOR_TEST
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=300,
|
||||
other_args=["--max-running-requests", "10"],
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
kill_process_tree(cls.process.pid)
|
||||
|
||||
|
||||
class TestRegexPatternMaxLength(unittest.TestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.regex_str_to_max_len = {
|
||||
"((ab|cd(e|f){2}){3,5}g|hij)*k": MAX_LEN,
|
||||
# - '*' → infinite tokens need to be stored
|
||||
"abc*?k": MAX_LEN,
|
||||
# - '*?' → infinite tokens still need to be stored even if lazy matching used
|
||||
"^spec(foo|at)$": 7,
|
||||
# - '^' and '$' don't add any characters to the max length
|
||||
# "spec" → 4
|
||||
# "(foo|at)" → max(3, 2) = 3
|
||||
# Whole regex = 7
|
||||
"(a(bca|de(fg|hi){2,3})j){2}kl": 22,
|
||||
# - Innermost alt: "fg" vs "hi" → 2
|
||||
# - Repeat {2,3}: max = 3 * 2 = 6
|
||||
# - Inner group "de(...)": 2 (for "de") + 6 = 8.
|
||||
# - "bca" or "de(...)" → max(3, 8) = 8
|
||||
# - Whole group: "a" (1) + group (8) + "j"(1) = 10
|
||||
# - Repeat {2} → 20
|
||||
# - Add "kl"(2) → 22
|
||||
"(foo(bar|baz(qux){1,2}))|(x(yz){5,10})": 21,
|
||||
# Branch 1:
|
||||
# "foo"(3) + max("bar"(3), "baz"(3)+"qux"{2} = 3 + 6 = 9) = 3 + 9 = 12
|
||||
# Branch 2:
|
||||
# "x"(1) + "yz"{10} = 1 + 20 =21
|
||||
# Whole regex = max(12, 21) = 21
|
||||
"(((a|bc){1,3}(d(e|f){2}|gh){2,4})|(ijk|lmp(no|p){3})){5}": 90,
|
||||
# Branch A:
|
||||
# (a|bc){1,3} → max = 3 * 2 = 6
|
||||
# Inside: d(e|f){2} = 1 + 2 * 1 = 3 vs gh = 2 → max = 3
|
||||
# Repeat {2,4} → 4 * 3 = 12
|
||||
# Branch A total = 18
|
||||
# Branch B:
|
||||
# "ijk"(3) vs "lmp(no|p){3}" = 3 + 3 * max(2, 1) = 3 + 6 = 9 → max = 9
|
||||
# Branch B total = 9
|
||||
# Whole outer alt = max(18, 9) = 18
|
||||
# Repeat {5} → 90
|
||||
}
|
||||
|
||||
def test_get_max_length(self):
|
||||
for regex_str, max_len in self.regex_str_to_max_len.items():
|
||||
if max_len == MAX_LEN:
|
||||
self.assertGreaterEqual(get_max_seq_length(regex_str), MAX_LEN)
|
||||
else:
|
||||
self.assertEqual(get_max_seq_length(regex_str), max_len)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,94 @@
|
||||
import openai
|
||||
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.srt.utils.hf_transformers_utils import get_tokenizer
|
||||
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
CustomTestCase,
|
||||
popen_launch_server,
|
||||
)
|
||||
|
||||
register_cuda_ci(est_time=6, suite="stage-b-test-small-1-gpu")
|
||||
register_amd_ci(est_time=47, suite="stage-b-test-small-1-gpu")
|
||||
|
||||
|
||||
class TestOpenAIServerIgnoreEOS(CustomTestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = DEFAULT_SMALL_MODEL_NAME_FOR_TEST
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.api_key = "sk-123456"
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
api_key=cls.api_key,
|
||||
)
|
||||
cls.base_url += "/v1"
|
||||
cls.tokenizer = get_tokenizer(DEFAULT_SMALL_MODEL_NAME_FOR_TEST)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
kill_process_tree(cls.process.pid)
|
||||
|
||||
def test_ignore_eos(self):
|
||||
"""
|
||||
Test that ignore_eos=True allows generation to continue beyond EOS token
|
||||
and reach the max_tokens limit.
|
||||
"""
|
||||
client = openai.Client(api_key=self.api_key, base_url=self.base_url)
|
||||
|
||||
max_tokens = 200
|
||||
|
||||
response_default = client.chat.completions.create(
|
||||
model=self.model,
|
||||
messages=[
|
||||
{"role": "system", "content": "You are a helpful assistant."},
|
||||
{"role": "user", "content": "Count from 1 to 20."},
|
||||
],
|
||||
temperature=0,
|
||||
max_tokens=max_tokens,
|
||||
extra_body={"ignore_eos": False},
|
||||
)
|
||||
|
||||
response_ignore_eos = client.chat.completions.create(
|
||||
model=self.model,
|
||||
messages=[
|
||||
{"role": "system", "content": "You are a helpful assistant."},
|
||||
{"role": "user", "content": "Count from 1 to 20."},
|
||||
],
|
||||
temperature=0,
|
||||
max_tokens=max_tokens,
|
||||
extra_body={"ignore_eos": True},
|
||||
)
|
||||
|
||||
default_tokens = len(
|
||||
self.tokenizer.encode(response_default.choices[0].message.content)
|
||||
)
|
||||
ignore_eos_tokens = len(
|
||||
self.tokenizer.encode(response_ignore_eos.choices[0].message.content)
|
||||
)
|
||||
|
||||
# Check if ignore_eos resulted in more tokens or exactly max_tokens
|
||||
# The ignore_eos response should either:
|
||||
# 1. Have more tokens than the default response (if default stopped at EOS before max_tokens)
|
||||
# 2. Have exactly max_tokens (if it reached the max_tokens limit)
|
||||
self.assertTrue(
|
||||
ignore_eos_tokens > default_tokens or ignore_eos_tokens >= max_tokens,
|
||||
f"ignore_eos did not generate more tokens: {ignore_eos_tokens} vs {default_tokens}",
|
||||
)
|
||||
|
||||
self.assertEqual(
|
||||
response_ignore_eos.choices[0].finish_reason,
|
||||
"length",
|
||||
f"Expected finish_reason='length' for ignore_eos=True, got {response_ignore_eos.choices[0].finish_reason}",
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
import unittest
|
||||
|
||||
unittest.main()
|
||||
@@ -0,0 +1,92 @@
|
||||
import unittest
|
||||
|
||||
import openai
|
||||
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.test.ci.ci_register import register_amd_ci, register_cuda_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
CustomTestCase,
|
||||
popen_launch_server,
|
||||
)
|
||||
|
||||
register_cuda_ci(est_time=38, suite="stage-b-test-small-1-gpu")
|
||||
register_amd_ci(est_time=31, suite="stage-b-test-small-1-gpu")
|
||||
|
||||
|
||||
class TestRequestLengthValidation(CustomTestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.api_key = "sk-123456"
|
||||
|
||||
# Start server with auto truncate disabled
|
||||
cls.process = popen_launch_server(
|
||||
DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
|
||||
cls.base_url,
|
||||
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
api_key=cls.api_key,
|
||||
other_args=("--max-total-tokens", "1000", "--context-length", "1000"),
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
kill_process_tree(cls.process.pid)
|
||||
|
||||
def test_input_length_longer_than_context_length(self):
|
||||
client = openai.Client(api_key=self.api_key, base_url=f"{self.base_url}/v1")
|
||||
|
||||
long_text = "hello " * 1200 # Will tokenize to more than context length
|
||||
|
||||
with self.assertRaises(openai.BadRequestError) as cm:
|
||||
client.chat.completions.create(
|
||||
model=DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
|
||||
messages=[
|
||||
{"role": "user", "content": long_text},
|
||||
],
|
||||
temperature=0,
|
||||
)
|
||||
|
||||
self.assertIn("is longer than the model's context length", str(cm.exception))
|
||||
|
||||
def test_input_length_longer_than_maximum_allowed_length(self):
|
||||
client = openai.Client(api_key=self.api_key, base_url=f"{self.base_url}/v1")
|
||||
|
||||
long_text = "hello " * 999 # the maximum allowed length is 994 tokens
|
||||
|
||||
with self.assertRaises(openai.BadRequestError) as cm:
|
||||
client.chat.completions.create(
|
||||
model=DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
|
||||
messages=[
|
||||
{"role": "user", "content": long_text},
|
||||
],
|
||||
temperature=0,
|
||||
)
|
||||
|
||||
self.assertIn("is longer than the model's context length", str(cm.exception))
|
||||
|
||||
def test_max_tokens_validation(self):
|
||||
client = openai.Client(api_key=self.api_key, base_url=f"{self.base_url}/v1")
|
||||
|
||||
long_text = "hello "
|
||||
|
||||
with self.assertRaises(openai.BadRequestError) as cm:
|
||||
client.chat.completions.create(
|
||||
model=DEFAULT_SMALL_MODEL_NAME_FOR_TEST,
|
||||
messages=[
|
||||
{"role": "user", "content": long_text},
|
||||
],
|
||||
temperature=0,
|
||||
max_tokens=1200,
|
||||
)
|
||||
|
||||
self.assertIn(
|
||||
"max_completion_tokens is too large",
|
||||
str(cm.exception),
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user