256 lines
7.6 KiB
Python
256 lines
7.6 KiB
Python
"""
|
|
Usage:
|
|
python3 -m unittest test_pp_single_node.TestPPAccuracy.test_gsm8k
|
|
python3 -m unittest test_pp_single_node.TestQwenPPAccuracy.test_pp_consistency
|
|
python3 -m unittest test_pp_single_node.TestFixedBugs.test_chunked_prefill_with_small_bs
|
|
"""
|
|
|
|
import time
|
|
import unittest
|
|
from types import SimpleNamespace
|
|
|
|
from sglang.bench_one_batch_server import BenchArgs as OneBatchBenchArgs
|
|
from sglang.srt.server_args import ServerArgs
|
|
from sglang.srt.utils import kill_process_tree
|
|
from sglang.test.few_shot_gsm8k import run_eval
|
|
from sglang.test.test_utils import (
|
|
DEFAULT_MODEL_NAME_FOR_TEST,
|
|
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
|
DEFAULT_URL_FOR_TEST,
|
|
is_in_ci,
|
|
popen_launch_server,
|
|
run_bench_one_batch_server,
|
|
)
|
|
|
|
|
|
class TestPPAccuracy(unittest.TestCase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
cls.base_url = "http://127.0.0.1:23333"
|
|
cls.process = popen_launch_server(
|
|
DEFAULT_MODEL_NAME_FOR_TEST,
|
|
cls.base_url,
|
|
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
|
other_args=[
|
|
"--tp-size",
|
|
2,
|
|
"--pp-size",
|
|
2,
|
|
"--chunked-prefill-size",
|
|
256,
|
|
],
|
|
)
|
|
|
|
@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.assertGreater(metrics["accuracy"], 0.74)
|
|
# Wait a little bit so that the memory check happens.
|
|
time.sleep(4)
|
|
|
|
|
|
class TestQwenPPAccuracy(unittest.TestCase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
cls.base_url = "http://127.0.0.1:23334" # different ports to avoid conflicts
|
|
cls.model_name = "Qwen/Qwen3-8B" # replace with your Qwen Model if needed
|
|
|
|
def run_gsm8k_test(self, pp_size):
|
|
process = popen_launch_server(
|
|
self.model_name,
|
|
self.base_url,
|
|
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
|
other_args=[
|
|
"--pp-size",
|
|
pp_size,
|
|
"--chunked-prefill-size",
|
|
256,
|
|
],
|
|
)
|
|
|
|
try:
|
|
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)
|
|
time.sleep(5)
|
|
return metrics
|
|
finally:
|
|
kill_process_tree(process.pid)
|
|
|
|
@unittest.skipIf(is_in_ci(), "To reduce the CI execution time.")
|
|
def test_pp_consistency(self):
|
|
baseline = self.run_gsm8k_test(pp_size=1)
|
|
pp_metrics = self.run_gsm8k_test(pp_size=2)
|
|
|
|
print(f"[Qwen PP Comparison] Baseline: {baseline} | PP: {pp_metrics}")
|
|
|
|
self.assertGreaterEqual(baseline["accuracy"], 0.74)
|
|
self.assertGreaterEqual(
|
|
pp_metrics["accuracy"],
|
|
baseline["accuracy"] - 0.02,
|
|
msg=(
|
|
f"PP accuracy dropped more than 1% compared to baseline. "
|
|
f"Baseline: {baseline['accuracy']:.2%}, PP: {pp_metrics['accuracy']:.2%}"
|
|
),
|
|
)
|
|
|
|
|
|
class TestQwenPPTieWeightsAccuracy(unittest.TestCase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
cls.base_url = "http://127.0.0.1:23335" # different ports to avoid conflicts
|
|
cls.model_name = (
|
|
"Qwen/Qwen3-0.6B" # qwen3 < 8B all have tie_word_embeddings = True
|
|
)
|
|
|
|
def run_gsm8k_test(self, pp_size):
|
|
process = popen_launch_server(
|
|
self.model_name,
|
|
self.base_url,
|
|
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
|
other_args=[
|
|
"--pp-size",
|
|
pp_size,
|
|
"--chunked-prefill-size",
|
|
256,
|
|
],
|
|
)
|
|
|
|
try:
|
|
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)
|
|
time.sleep(5)
|
|
return metrics
|
|
finally:
|
|
kill_process_tree(process.pid)
|
|
|
|
def test_pp_consistency(self):
|
|
baseline = self.run_gsm8k_test(pp_size=1)
|
|
pp_metrics = self.run_gsm8k_test(pp_size=2)
|
|
|
|
print(f"[Qwen PP Comparison] Baseline: {baseline} | PP: {pp_metrics}")
|
|
|
|
self.assertGreaterEqual(baseline["accuracy"], 0.38)
|
|
self.assertGreaterEqual(
|
|
pp_metrics["accuracy"],
|
|
baseline["accuracy"] - 0.02,
|
|
msg=(
|
|
f"PP accuracy dropped more than 1% compared to baseline. "
|
|
f"Baseline: {baseline['accuracy']:.2%}, PP: {pp_metrics['accuracy']:.2%}"
|
|
),
|
|
)
|
|
|
|
|
|
class TestQwenMoePPAccuracy(unittest.TestCase):
|
|
@classmethod
|
|
def setUpClass(cls):
|
|
cls.base_url = "http://127.0.0.1:23336" # different ports to avoid conflicts
|
|
cls.model_name = "Qwen/Qwen3-30B-A3B" # replace with your Qwen Model if needed
|
|
|
|
def run_gsm8k_test(self, pp_size):
|
|
process = popen_launch_server(
|
|
self.model_name,
|
|
self.base_url,
|
|
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
|
other_args=[
|
|
"--pp-size",
|
|
pp_size,
|
|
"--chunked-prefill-size",
|
|
256,
|
|
],
|
|
)
|
|
|
|
try:
|
|
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)
|
|
time.sleep(5)
|
|
return metrics
|
|
finally:
|
|
kill_process_tree(process.pid)
|
|
|
|
def test_pp_consistency(self):
|
|
baseline = self.run_gsm8k_test(pp_size=1)
|
|
pp_metrics = self.run_gsm8k_test(pp_size=2)
|
|
|
|
print(f"[Qwen PP Comparison] Baseline: {baseline} | PP: {pp_metrics}")
|
|
|
|
self.assertGreaterEqual(baseline["accuracy"], 0.74)
|
|
self.assertGreaterEqual(
|
|
pp_metrics["accuracy"],
|
|
baseline["accuracy"] - 0.02,
|
|
msg=(
|
|
f"PP accuracy dropped more than 1% compared to baseline. "
|
|
f"Baseline: {baseline['accuracy']:.2%}, PP: {pp_metrics['accuracy']:.2%}"
|
|
),
|
|
)
|
|
|
|
|
|
class TestFixedBugs(unittest.TestCase):
|
|
def test_chunked_prefill_with_small_bs(self):
|
|
model = DEFAULT_MODEL_NAME_FOR_TEST
|
|
server_args = ServerArgs(model_path=model)
|
|
bench_args = OneBatchBenchArgs(
|
|
batch_size=(1,),
|
|
input_len=(1,),
|
|
output_len=(1,),
|
|
base_url=DEFAULT_URL_FOR_TEST,
|
|
)
|
|
other_server_args = [
|
|
"--tp-size",
|
|
2,
|
|
"--pp-size",
|
|
2,
|
|
"--chunked-prefill",
|
|
256,
|
|
"--max-running-requests",
|
|
2,
|
|
]
|
|
run_bench_one_batch_server(
|
|
model,
|
|
DEFAULT_URL_FOR_TEST,
|
|
server_args,
|
|
bench_args,
|
|
other_server_args,
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|