minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 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/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__":