EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Split test_intel_amx_attention_backend.py to pass CI of timeout (#11370)

Browse Source 
Co-authored-by: Ma Mingfei <mingfei.ma@intel.com>
This commit is contained in:
YanbingJiang
2025-10-16 10:22:32 +08:00
committed by GitHub
parent 476c67d7fc
commit cbac499750
7 changed files with 197 additions and 140 deletions
Download Patch File Download Diff File Expand all files Collapse all files

86
test/srt/cpu/test_cpu_graph.py Normal file
View File

@@ -0,0 +1,86 @@
"""
Usage:
python3 -m unittest test_cpu_graph.TestCPUGraph.test_mmlu_torch_compile_cpu
"""
import copy
import os
import unittest
from types import SimpleNamespace
from sglang.srt.utils import get_cpu_ids_by_node, 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,
intel_amx_benchmark,
is_in_ci,
popen_launch_server,
)
class TestCPUGraph(CustomTestCase):
@intel_amx_benchmark(
extra_args=[
"--batch-size",
"1",
"--mem-fraction-static",
"0.05",
"--enable-torch-compile",
"--torch-compile-max-bs",
"1",
],
min_throughput=10,
)
def test_latency_torch_compile_cpu(self):
return DEFAULT_MLA_MODEL_NAME_FOR_TEST
def test_mmlu_torch_compile_cpu(self):
model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
base_url = DEFAULT_URL_FOR_TEST
cpu_ids_by_node = get_cpu_ids_by_node()
n_numa_node = len(cpu_ids_by_node)
env = copy.deepcopy(os.environ)
env["SGLANG_CPU_OMP_THREADS_BIND"] = "all"
process = popen_launch_server(
model,
base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--attention-backend",
"intel_amx",
"--mem-fraction-static",
"0.05",
"--disable-radix",
"--trust-remote-code",
"--disable-overlap-schedule",
"--enable-torch-compile",
"--torch-compile-max-bs",
"1",
"--tp",
f"{n_numa_node}",
],
env=env,
)
try:
args = SimpleNamespace(
base_url=base_url,
model=model,
eval_name="mmlu",
num_examples=64,
num_threads=32,
)
metrics = run_eval(args)
if is_in_ci():
self.assertGreater(metrics["score"], 0.45)
finally:
kill_process_tree(process.pid)
if __name__ == "__main__":
unittest.main()

73
test/srt/cpu/test_intel_amx_attention_backend_a.py Normal file
View File

@@ -0,0 +1,73 @@
"""
Usage:
python3 -m unittest test_intel_amx_attention_backend.TestIntelAMXAttnBackend.test_latency_default_model
"""
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_MODEL_NAME_FOR_TEST,
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
intel_amx_benchmark,
is_in_ci,
popen_launch_server,
)
class TestIntelAMXAttnBackend(CustomTestCase):
@intel_amx_benchmark(
extra_args=["--batch-size", "4", "--mem-fraction-static", "0.3"],
min_throughput=10,
)
def test_latency_mla_model(self):
return DEFAULT_MLA_MODEL_NAME_FOR_TEST
@intel_amx_benchmark(
extra_args=["--batch-size", "4", "--mem-fraction-static", "0.1"],
min_throughput=40,
)
def test_latency_default_model(self):
return DEFAULT_MODEL_NAME_FOR_TEST
def test_mmlu(self):
model = DEFAULT_MLA_MODEL_NAME_FOR_TEST
base_url = DEFAULT_URL_FOR_TEST
process = popen_launch_server(
model,
base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--attention-backend",
"intel_amx",
"--mem-fraction-static",
"0.3",
"--disable-radix",
"--trust-remote-code",
"--disable-overlap-schedule",
],
)
try:
args = SimpleNamespace(
base_url=base_url,
model=model,
eval_name="mmlu",
num_examples=64,
num_threads=32,
)
metrics = run_eval(args)
if is_in_ci():
self.assertGreater(metrics["score"], 0.45)
finally:
kill_process_tree(process.pid)
if __name__ == "__main__":
unittest.main()

35
test/srt/cpu/test_intel_amx_attention_backend_b.py Normal file
View File

@@ -0,0 +1,35 @@
"""
For intel_amx attention backend FP8 tests
Usage:
python3 -m unittest test_intel_amx_attention_backend_1.TestIntelAMXAttnBackendQuant.test_latency_fp8_qwen
"""
import unittest
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_FP8_WITH_MOE,
DEFAULT_MODEL_NAME_FOR_TEST_QWEN_FP8,
CustomTestCase,
intel_amx_benchmark,
)
class TestIntelAMXAttnBackendQuant(CustomTestCase):
@intel_amx_benchmark(
extra_args=["--batch-size", "4", "--mem-fraction-static", "0.1"],
min_throughput=150,
)
def test_latency_fp8_qwen(self):
return DEFAULT_MODEL_NAME_FOR_TEST_QWEN_FP8
@intel_amx_benchmark(
extra_args=["--batch-size", "4", "--mem-fraction-static", "0.1"],
min_throughput=50,
)
def test_latency_fp8_moe_model(self):
return DEFAULT_MODEL_NAME_FOR_TEST_FP8_WITH_MOE
if __name__ == "__main__":
unittest.main()

53
test/srt/cpu/test_intel_amx_attention_backend_c.py Normal file
View File

@@ -0,0 +1,53 @@
"""
For intel_amx attention backend w8a8 tests
Usage:
python3 -m unittest test_intel_amx_attention_backend_2.TestIntelAMXAttnBackendQuant.test_latency_w8a8_default_model
"""
import unittest
from sglang.test.test_utils import (
DEFAULT_MODEL_NAME_FOR_TEST_W8A8,
DEFAULT_MODEL_NAME_FOR_TEST_W8A8_WITH_MOE,
CustomTestCase,
intel_amx_benchmark,
)
class TestIntelAMXAttnBackendQuant(CustomTestCase):
@intel_amx_benchmark(
extra_args=[
"--batch-size",
"4",
"--quantization",
"w8a8_int8",
"--mem-fraction-static",
"0.1",
],
min_throughput=100,
)
def test_latency_w8a8_default_model(self):
return DEFAULT_MODEL_NAME_FOR_TEST_W8A8
@intel_amx_benchmark(
extra_args=[
"--batch-size",
"4",
"--quantization",
"w8a8_int8",
"--mem-fraction-static",
"0.9",
"--max-total-tokens",
"65536",
"--tp",
"6",
],
min_throughput=100,
)
def test_latency_w8a8_moe_model(self):
return DEFAULT_MODEL_NAME_FOR_TEST_W8A8_WITH_MOE
if __name__ == "__main__":
unittest.main()