[XPU] Add registry mechanism for XPU CI tests (#25405)
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
This commit is contained in:
vikram singh shekhawat
co-authored by
Ma Mingfei
Ma Mingfei
parent
87c3171aaa
commit
737c6cd6d1
@@ -10,7 +10,7 @@ from sglang.srt.utils import get_device
|
||||
from sglang.test.ci.ci_register import register_cuda_ci, register_xpu_ci
|
||||
|
||||
register_cuda_ci(est_time=11, stage="base-b", runner_config="1-gpu-large")
|
||||
register_xpu_ci(est_time=30, suite="xpu")
|
||||
register_xpu_ci(est_time=900, suite="stage-b-test-1-gpu-xpu")
|
||||
|
||||
|
||||
@unittest.skipIf(
|
||||
|
||||
@@ -0,0 +1,113 @@
|
||||
"""
|
||||
python3 -m unittest test_deepseek_ocr.py
|
||||
"""
|
||||
|
||||
import json
|
||||
import os
|
||||
import unittest
|
||||
from pathlib import Path
|
||||
|
||||
import requests
|
||||
|
||||
from sglang.srt.utils import kill_process_tree
|
||||
from sglang.srt.utils.hf_transformers import get_tokenizer
|
||||
from sglang.test.ci.ci_register import register_xpu_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
CustomTestCase,
|
||||
popen_launch_server,
|
||||
)
|
||||
|
||||
register_xpu_ci(est_time=360, suite="stage-b-test-1-gpu-xpu")
|
||||
|
||||
|
||||
class TestDeepSeekOCR(CustomTestCase):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = "deepseek-ai/DeepSeek-OCR"
|
||||
cls.tokenizer = get_tokenizer(cls.model)
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.image_path = str(
|
||||
(Path(__file__).resolve().parents[3] / "examples/assets/example_image.png")
|
||||
)
|
||||
if not os.path.exists(cls.image_path):
|
||||
raise FileNotFoundError(f"Image not found: {cls.image_path}")
|
||||
cls.common_args = [
|
||||
"--device",
|
||||
"xpu",
|
||||
"--attention-backend",
|
||||
"intel_xpu",
|
||||
]
|
||||
os.environ["SGLANG_USE_SGL_XPU"] = "1"
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
other_args=[
|
||||
*cls.common_args,
|
||||
],
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def tearDownClass(cls):
|
||||
"""Fixture that is run once after all tests in the class."""
|
||||
if hasattr(cls, "process") and cls.process:
|
||||
cls.process.terminate()
|
||||
try:
|
||||
cls.process.wait(timeout=30)
|
||||
except Exception:
|
||||
# Force kill if it didn't exit cleanly in time
|
||||
kill_process_tree(cls.process.pid)
|
||||
|
||||
def get_request_json(self, max_new_tokens=32, n=1):
|
||||
response = requests.post(
|
||||
self.base_url + "/generate",
|
||||
json={
|
||||
"text": "<image>\n<|grounding|>Convert the document to pure text.",
|
||||
"image_data": self.image_path,
|
||||
"sampling_params": {
|
||||
"temperature": 0 if n == 1 else 0.5,
|
||||
"max_new_tokens": max_new_tokens,
|
||||
},
|
||||
},
|
||||
)
|
||||
return response.json()
|
||||
|
||||
def run_decode(
|
||||
self,
|
||||
max_new_tokens=128,
|
||||
n=1,
|
||||
):
|
||||
|
||||
ret = self.get_request_json(max_new_tokens=max_new_tokens, n=n)
|
||||
print(json.dumps(ret, indent=2))
|
||||
|
||||
def assert_one_item(item):
|
||||
if item["meta_info"]["finish_reason"]["type"] == "stop":
|
||||
self.assertEqual(
|
||||
item["meta_info"]["finish_reason"]["matched"],
|
||||
self.tokenizer.eos_token_id,
|
||||
)
|
||||
elif item["meta_info"]["finish_reason"]["type"] == "length":
|
||||
self.assertEqual(
|
||||
len(item["output_ids"]), item["meta_info"]["completion_tokens"]
|
||||
)
|
||||
self.assertEqual(len(item["output_ids"]), max_new_tokens)
|
||||
|
||||
# Determine whether to assert a single item or multiple items based on n
|
||||
if n == 1:
|
||||
assert_one_item(ret)
|
||||
else:
|
||||
self.assertEqual(len(ret), n)
|
||||
for i in range(n):
|
||||
assert_one_item(ret[i])
|
||||
|
||||
print("=" * 100)
|
||||
|
||||
def test_moe(self):
|
||||
self.run_decode()
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,60 @@
|
||||
"""
|
||||
python3 -m unittest test_deepseek_ocr_triton.py
|
||||
"""
|
||||
|
||||
import os
|
||||
import unittest
|
||||
from pathlib import Path
|
||||
|
||||
from test_deepseek_ocr import TestDeepSeekOCR
|
||||
|
||||
from sglang.srt.utils.hf_transformers import get_tokenizer
|
||||
from sglang.test.ci.ci_register import register_xpu_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
DEFAULT_URL_FOR_TEST,
|
||||
popen_launch_server,
|
||||
)
|
||||
|
||||
register_xpu_ci(
|
||||
est_time=360,
|
||||
suite="stage-b-test-1-gpu-xpu",
|
||||
disabled="Temporarily disabled until Triton-XPU upgrade",
|
||||
)
|
||||
|
||||
|
||||
# TODO: Temporarily disable this test and re-enable it after Triton-XPU is upgraded.
|
||||
@unittest.skip("Temporarily disabled until Triton-XPU upgrade")
|
||||
class TestDeepSeekOCRTriton(TestDeepSeekOCR):
|
||||
@classmethod
|
||||
def setUpClass(cls):
|
||||
cls.model = "deepseek-ai/DeepSeek-OCR"
|
||||
cls.tokenizer = get_tokenizer(cls.model)
|
||||
cls.base_url = DEFAULT_URL_FOR_TEST
|
||||
cls.image_path = str(
|
||||
(Path(__file__).resolve().parents[3] / "examples/assets/example_image.png")
|
||||
)
|
||||
if not os.path.exists(cls.image_path):
|
||||
raise FileNotFoundError(f"Image not found: {cls.image_path}")
|
||||
cls.common_args = [
|
||||
"--device",
|
||||
"xpu",
|
||||
"--attention-backend",
|
||||
"intel_xpu",
|
||||
]
|
||||
os.environ["SGLANG_USE_SGL_XPU"] = "0"
|
||||
cls.process = popen_launch_server(
|
||||
cls.model,
|
||||
cls.base_url,
|
||||
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
|
||||
other_args=[
|
||||
*cls.common_args,
|
||||
],
|
||||
)
|
||||
|
||||
|
||||
# Prevent pytest from collecting the imported base test class here.
|
||||
del TestDeepSeekOCR
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,86 @@
|
||||
"""
|
||||
Usage:
|
||||
python3 -m unittest test_intel_xpu_backend.TestIntelXPUBackend.test_latency_qwen_model
|
||||
"""
|
||||
|
||||
import unittest
|
||||
from functools import wraps
|
||||
|
||||
from sglang.test.ci.ci_register import register_xpu_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_MODEL_NAME_FOR_TEST_FP8_WITH_MOE,
|
||||
DEFAULT_SMALL_MODEL_NAME_FOR_TEST_BASE,
|
||||
DEFAULT_SMALL_MODEL_NAME_FOR_TEST_QWEN,
|
||||
CustomTestCase,
|
||||
is_in_ci,
|
||||
run_bench_one_batch,
|
||||
)
|
||||
|
||||
register_xpu_ci(est_time=600, suite="stage-b-test-1-gpu-xpu")
|
||||
|
||||
|
||||
def intel_xpu_benchmark(
|
||||
extra_args=None, min_throughput=None, mem_fraction_static="0.4"
|
||||
):
|
||||
def decorator(test_func):
|
||||
@wraps(test_func)
|
||||
def wrapper(self):
|
||||
common_args = [
|
||||
"--disable-radix-cache",
|
||||
"--trust-remote-code",
|
||||
"--mem-fraction-static",
|
||||
str(mem_fraction_static),
|
||||
"--batch-size",
|
||||
"1",
|
||||
"--device",
|
||||
"xpu",
|
||||
]
|
||||
ci_args = ["--input", "64", "--output", "4"] if is_in_ci() else []
|
||||
full_args = common_args + ci_args + (extra_args or [])
|
||||
|
||||
model = test_func(self)
|
||||
prefill_latency, decode_throughput, decode_latency = run_bench_one_batch(
|
||||
model, full_args
|
||||
)
|
||||
|
||||
print(f"{model=}")
|
||||
print(f"{prefill_latency=}")
|
||||
print(f"{decode_throughput=}")
|
||||
print(f"{decode_latency=}")
|
||||
|
||||
if is_in_ci() and min_throughput is not None:
|
||||
self.assertGreater(decode_throughput, min_throughput)
|
||||
|
||||
return wrapper
|
||||
|
||||
return decorator
|
||||
|
||||
|
||||
class TestIntelXPUBackend(CustomTestCase):
|
||||
|
||||
@intel_xpu_benchmark(min_throughput=10, mem_fraction_static="0.3")
|
||||
def test_latency_qwen_model(self):
|
||||
return DEFAULT_SMALL_MODEL_NAME_FOR_TEST_QWEN
|
||||
|
||||
@intel_xpu_benchmark(
|
||||
["--attention-backend", "intel_xpu", "--page-size", "128"],
|
||||
mem_fraction_static="0.5",
|
||||
)
|
||||
def test_attention_backend(self):
|
||||
return DEFAULT_SMALL_MODEL_NAME_FOR_TEST_BASE
|
||||
|
||||
@intel_xpu_benchmark(
|
||||
[
|
||||
"--json-model-override-args",
|
||||
'{"num_hidden_layers": 4}',
|
||||
"--decode-attention-backend",
|
||||
"intel_xpu",
|
||||
],
|
||||
min_throughput=32,
|
||||
)
|
||||
def test_mla_decode_attention_backend(self):
|
||||
return DEFAULT_MODEL_NAME_FOR_TEST_FP8_WITH_MOE
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
@@ -0,0 +1,47 @@
|
||||
"""
|
||||
Basic XPU test: verifies the server starts and produces a non-empty
|
||||
response on Intel XPU with the default attention backend.
|
||||
|
||||
Assigned to stage-a so it gates stage-b before the heavier tests run.
|
||||
|
||||
Usage:
|
||||
python3 -m unittest test_xpu_basic.TestXPUBasic.test_basic_generation
|
||||
"""
|
||||
|
||||
import unittest
|
||||
|
||||
from sglang.test.ci.ci_register import register_xpu_ci
|
||||
from sglang.test.test_utils import (
|
||||
DEFAULT_SMALL_MODEL_NAME_FOR_TEST_QWEN,
|
||||
CustomTestCase,
|
||||
is_in_ci,
|
||||
run_bench_one_batch,
|
||||
)
|
||||
|
||||
register_xpu_ci(est_time=300, suite="stage-a-test-1-gpu-xpu")
|
||||
|
||||
|
||||
class TestXPUBasic(CustomTestCase):
|
||||
|
||||
def test_basic_generation(self):
|
||||
"""Server starts on XPU and completes at least one decode step."""
|
||||
args = [
|
||||
"--device",
|
||||
"xpu",
|
||||
"--disable-radix-cache",
|
||||
"--mem-fraction-static",
|
||||
"0.6",
|
||||
"--batch-size",
|
||||
"1",
|
||||
]
|
||||
if is_in_ci():
|
||||
args += ["--input", "64", "--output", "4"]
|
||||
|
||||
_, decode_throughput, _ = run_bench_one_batch(
|
||||
DEFAULT_SMALL_MODEL_NAME_FOR_TEST_QWEN, args
|
||||
)
|
||||
self.assertGreater(decode_throughput, 0, "XPU decode throughput must be > 0")
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
unittest.main()
|
||||
Reference in New Issue
Block a user