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

[NPU] Add GitHub test summary and deduplicate test code. Part 1 (#23835)

Browse Source 
Co-authored-by: Elizaveta Martirosian <elizaveta.martirosian@gmail.com>
Co-authored-by: root <root@localhost.localdomain>
Co-authored-by: Elizaveta Martirosian <you@example.com>
Co-authored-by: ronnie_zheng <zl19940307@163.com>
This commit is contained in:
Elizaveta Martirosian
2026-05-02 14:18:18 +03:00
committed by GitHub
co-authored by Elizaveta Martirosian root Elizaveta Martirosian ronnie_zheng
parent 3259a2c789
commit ebbaab5597
8 changed files with 327 additions and 326 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/ascend/basic_function/dllm/test_npu_llada2_mini.py
+22 -58
View File
@@ -1,17 +1,13 @@
import os
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
from sglang.test.ascend.test_ascend_utils import LLaDA2_0_MINI_WEIGHTS_PATH
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.send_one import BenchArgs, send_one_prompt
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
is_in_ci,
popen_launch_server,
write_github_step_summary,
)
@@ -19,59 +15,27 @@ register_npu_ci(est_time=400, suite="stage-b-test-4-npu-a3", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
class TestLLaDA2Mini(CustomTestCase):
@classmethod
def setUpClass(cls):
cls._old_disable_acl = os.environ.get("SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT")
os.environ["SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT"] = "1"
class TestLLaDA2Mini(GSM8KAscendMixin, CustomTestCase):
model = LLaDA2_0_MINI_WEIGHTS_PATH
cls.model = "/root/.cache/modelscope/hub/models/inclusionAI/LLaDA2.0-mini"
cls.base_url = DEFAULT_URL_FOR_TEST
other_args = [
"--trust-remote-code",
"--disable-radix-cache",
"--mem-fraction-static",
"0.9",
"--max-running-requests",
"1",
"--attention-backend",
"ascend",
"--dllm-algorithm",
"LowConfidence", # TODO: Add dLLM configurations
]
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)
if cls._old_disable_acl is None:
os.environ.pop("SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT", None)
else:
os.environ["SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT"] = cls._old_disable_acl
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(f"{metrics=}")
self.assertGreater(metrics["accuracy"], 0.88)
self.assertGreater(metrics["output_throughput"], 70)
other_args = [
"--trust-remote-code",
"--disable-radix-cache",
"--mem-fraction-static",
"0.9",
"--max-running-requests",
"1",
"--attention-backend",
"ascend",
"--dllm-algorithm",
"LowConfidence", # TODO: Add dLLM configurations
]
env = {
**os.environ,
"SGLANG_NPU_DISABLE_ACL_FORMAT_WEIGHT": "1", # Need to avoid OOM issue
}
accuracy = 0.88
output_throughput = 70
def test_bs_1_speed(self):
args = BenchArgs(port=int(self.base_url.split(":")[-1]), max_new_tokens=2048)
test/registered/ascend/basic_function/optimization_debug/test_npu_piecewise_graph_prefill.py
+54 -70
View File
@@ -1,92 +1,76 @@
import subprocess
import unittest
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
from sglang.test.ascend.test_ascend_utils import (
QWEN2_5_7B_INSTRUCT_WEIGHTS_PATH,
write_results_to_github_step_summary,
)
from sglang.test.ci.ci_register import register_npu_ci
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,
SimpleNamespace,
popen_launch_server,
run_bench_one_batch,
)
register_npu_ci(est_time=400, suite="stage-b-test-1-npu-a2", nightly=False)
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
MODEL = "/root/.cache/modelscope/hub/models/Qwen/Qwen2.5-7B-Instruct"
GSM8K_EXP_ACCURACY = 0.84
EXP_PREFILL_LATENCY = 0.045
TOKENS_TO_CAPTURE = [i for i in range(128, 4096, 128)]
class TestPiecewiseGraphPrefillCorrectness(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.url = urlparse(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",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--cuda-graph-bs",
128,
"--enforce-piecewise-cuda-graph",
"--piecewise-cuda-graph-tokens",
*TOKENS_TO_CAPTURE,
],
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
def test_gsm8k(self):
print(f"##=== Testing accuracy: {self.model} ===##")
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=1319,
max_new_tokens=512,
parallel=128,
host=f"http://{self.url.hostname}",
port=int(self.url.port),
)
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
GSM8K_EXP_ACCURACY,
)
class TestPiecewiseGraphPrefillCorrectness(GSM8KAscendMixin, CustomTestCase):
model = QWEN2_5_7B_INSTRUCT_WEIGHTS_PATH
other_args = [
"--trust-remote-code",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--cuda-graph-bs",
128,
"--enforce-piecewise-cuda-graph",
"--piecewise-cuda-graph-tokens",
*TOKENS_TO_CAPTURE,
]
accuracy = 0.84
num_questions = 1319
class TestPiecewiseGraphPrefillBenchmark(CustomTestCase):
model = QWEN2_5_7B_INSTRUCT_WEIGHTS_PATH
other_args = [
"--trust-remote-code",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--enforce-piecewise-cuda-graph",
"--piecewise-cuda-graph-tokens",
] + TOKENS_TO_CAPTURE
latency = 0.045
def test_latency(self):
print(f"##=== Testing prefill latency: {MODEL} ===##")
prefill_latency, _, _ = run_bench_one_batch(
MODEL,
other_args=[
"--trust-remote-code",
"--mem-fraction-static",
0.8,
"--attention-backend",
"ascend",
"--enforce-piecewise-cuda-graph",
"--piecewise-cuda-graph-tokens",
]
+ TOKENS_TO_CAPTURE,
)
self.assertLess(prefill_latency, EXP_PREFILL_LATENCY)
print(f"##=== Testing prefill latency: {self.model} ===##")
model_metrics = {
"server": subprocess.list2cmdline(map(str, self.other_args)),
"client": "bench_one_batch",
"latency_threshold": self.latency,
}
try:
prefill_latency, _, _ = run_bench_one_batch(
self.model,
other_args=self.other_args,
)
model_metrics["latency"] = float(prefill_latency)
self.assertLess(prefill_latency, self.latency)
except Exception as e:
model_metrics["error"] = e
print(f"Error testing {self.model}: {e}")
self.fail(f"Test failed for {self.model}: {e}")
finally:
write_results_to_github_step_summary({self.model: model_metrics})
if __name__ == "__main__":
test/registered/ascend/basic_function/parallel_strategy/expert_parallelism/test_npu_deepep_auto_deepseek_v3_2_w8a8.py
+40 -85
View File
@@ -1,22 +1,16 @@
import os
import unittest
from types import SimpleNamespace
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
from sglang.test.ascend.test_ascend_utils import DEEPSEEK_V3_2_W8A8_WEIGHTS_PATH
from sglang.test.ascend.test_mmlu import TestMMLU
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.few_shot_gsm8k import run_eval as run_gsm8k
from sglang.test.run_eval import run_eval
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.test_utils import CustomTestCase
register_npu_ci(est_time=400, suite="nightly-16-npu-a3", nightly=True)
class TestDeepEpDeepseekV32(CustomTestCase):
class TestDeepEpDeepseekV32(GSM8KAscendMixin, TestMMLU, CustomTestCase):
"""Testcase: Verify that for the DeepSeek V3.2 model in the single-machine colocation scenario,
its inference accuracy on the MMLU and GSM8K dataset meets the preset standard when the parameter --deepep-mode auto is configured.
@@ -24,84 +18,45 @@ class TestDeepEpDeepseekV32(CustomTestCase):
[Test Target] --moe-a2a-backend deepep;--deepep-mode
"""
@classmethod
def setUpClass(cls):
cls.model = DEEPSEEK_V3_2_W8A8_WEIGHTS_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=6000,
other_args=[
"--trust-remote-code",
"--tp-size",
"16",
"--quantization",
"modelslim",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"auto",
"--mem-fraction-static",
0.82,
"--disable-cuda-graph",
"--disable-radix-cache",
"--context-length",
40960,
"--max-prefill-tokens",
40960,
"--max-total-tokens",
40960,
],
env={
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "16",
"HCCL_BUFFSIZE": "1600",
"HCCL_OP_EXPANSION_MODE": "AIV",
"SGLANG_NPU_USE_MLAPO": "0",
"SGLANG_NPU_USE_MULTI_STREAM": "1",
"TASK_QUEUE_ENABLE": "0",
**os.environ,
},
)
model = DEEPSEEK_V3_2_W8A8_WEIGHTS_PATH
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
timeout_for_server_launch = 60000
other_args = [
"--trust-remote-code",
"--tp-size",
"16",
"--quantization",
"modelslim",
"--moe-a2a-backend",
"deepep",
"--deepep-mode",
"auto",
"--mem-fraction-static",
0.82,
"--disable-cuda-graph",
"--disable-radix-cache",
"--context-length",
40960,
"--max-prefill-tokens",
40960,
"--max-total-tokens",
40960,
]
def test_mmlu(self):
expect_score = 0.85
args = SimpleNamespace(
base_url=self.base_url,
model=self.model,
eval_name="mmlu",
num_examples=128,
num_threads=32,
)
print("Starting mmlu test...")
metrics = run_eval(args)
self.assertGreater(metrics["score"], expect_score)
env = {
**os.environ,
"PYTORCH_NPU_ALLOC_CONF": "expandable_segments:True",
"STREAMS_PER_DEVICE": "32",
"SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "16",
"HCCL_BUFFSIZE": "1600",
"HCCL_OP_EXPANSION_MODE": "AIV",
"SGLANG_NPU_USE_MLAPO": "0",
"SGLANG_NPU_USE_MULTI_STREAM": "1",
"TASK_QUEUE_ENABLE": "0",
}
def test_gsm8k(self):
expect_accuracy = 0.95
args = SimpleNamespace(
num_shots=8,
data_path=None,
timeout=60000,
num_questions=200,
max_new_tokens=512,
parallel=128,
host="http://127.0.0.1",
port=int(self.base_url.split(":")[-1]),
)
print("Starting gsm8k test...")
metrics = run_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
expect_accuracy,
f'Accuracy of {self.model} is {str(metrics["accuracy"])}, is lower than {expect_accuracy}',
)
accuracy = 0.95 # Test GSM8K accuracy ≥0.95
accuracy_mmlu = 0.85 # Test MMLU accuracy ≥0.85
if __name__ == "__main__":
test/registered/ascend/basic_function/speculative_inference/test_npu_eagle3.py
+38 -76
View File
@@ -1,99 +1,61 @@
import os
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
from sglang.srt.utils import kill_process_tree
from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
from sglang.test.ascend.test_ascend_utils import (
QWEN3_8B_EAGLE3_WEIGHTS_PATH,
QWEN3_8B_WEIGHTS_PATH,
)
from sglang.test.ci.ci_register import register_npu_ci
from sglang.test.few_shot_gsm8k import run_eval as run_eval_few_shot_gsm8k
from sglang.test.test_utils import (
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
from sglang.test.test_utils import CustomTestCase
register_npu_ci(est_time=400, suite="nightly-1-npu-a3", nightly=True)
class TestNpuEagle3(CustomTestCase):
class TestNpuEagle3(GSM8KAscendMixin, CustomTestCase):
"""Testcase: Verify GSM8K inference accuracy ≥0.81 for model with specified EAGLE3 speculative inference parameters.
[Test Category] Speculative Decoding
[Test Target] --speculative-draft-model-quantization; --speculative-algorithm; --speculative-draft-model-path; --speculative-num-steps; --speculative-eagle-topk; --speculative-num-draft-tokens; --speculative-attention-mode
"""
@classmethod
def setUpClass(cls):
cls.model = QWEN3_8B_WEIGHTS_PATH
cls.accuracy = 0.81
cls.base_url = DEFAULT_URL_FOR_TEST
cls.url = urlparse(DEFAULT_URL_FOR_TEST)
model = QWEN3_8B_WEIGHTS_PATH
timeout_for_server_launch = 1500
other_args = [
"--trust-remote-code",
"--attention-backend",
"ascend",
"--disable-radix-cache",
"--speculative-draft-model-quantization",
"unquant",
"--speculative-algorithm",
"EAGLE3",
"--speculative-draft-model-path",
QWEN3_8B_EAGLE3_WEIGHTS_PATH,
"--speculative-num-steps",
"4",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"5",
"--speculative-attention-mode",
"decode",
"--tp-size",
"1",
"--mem-fraction-static",
"0.7",
"--disable-cuda-graph",
"--dtype",
"bfloat16",
]
cls.common_args = [
"--trust-remote-code",
"--attention-backend",
"ascend",
"--disable-radix-cache",
"--speculative-draft-model-quantization",
"unquant",
"--speculative-algorithm",
"EAGLE3",
"--speculative-draft-model-path",
QWEN3_8B_EAGLE3_WEIGHTS_PATH,
"--speculative-num-steps",
"4",
"--speculative-eagle-topk",
"1",
"--speculative-num-draft-tokens",
"5",
"--speculative-attention-mode",
"decode",
"--tp-size",
"1",
"--mem-fraction-static",
"0.7",
"--disable-cuda-graph",
"--dtype",
"bfloat16",
]
env = {
**os.environ,
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
}
cls.extra_envs = {
"SGLANG_ENABLE_OVERLAP_PLAN_STREAM": "1",
}
os.environ.update(cls.extra_envs)
def test_gsm8k(self):
process = popen_launch_server(
self.model,
self.base_url,
timeout=1500,
other_args=[
*self.common_args,
],
)
try:
args = SimpleNamespace(
num_shots=5,
data_path=None,
num_questions=1319,
max_new_tokens=512,
parallel=128,
host=f"http://{self.url.hostname}",
port=int(self.url.port),
)
metrics = run_eval_few_shot_gsm8k(args)
self.assertGreaterEqual(
metrics["accuracy"],
self.accuracy,
)
finally:
kill_process_tree(process.pid)
accuracy = 0.81
num_questions = 1319
if __name__ == "__main__":