[HiCache] Fix PP inconsistency with HiCache L3 (#22607) (#27010)

Co-authored-by: ybyang <ybyang7@iflytek.com>
Co-authored-by: hzh0425 <hzh0425@apache.org>
Co-authored-by: shangmingc <csmthu@gmail.com>
Co-authored-by: 晟海 <huangtingwei.htw@antgroup.com>
This commit is contained in:
Chao Shi
2026-08-25 20:49:57 +08:00
committed by GitHub
co-authored by ybyang hzh0425 shangmingc 晟海
parent c3947eeada
commit 829138a31e
10 changed files with 929 additions and 353 deletions
@@ -1,7 +1,12 @@
import os
import shutil
import tempfile
import unittest
from types import SimpleNamespace
from urllib.parse import urlparse
from test_unified_radix_cache_kl_nightly import AccuracyTwoPassMixin
from sglang.srt.utils import kill_process_tree
from sglang.test.ci.ci_register import register_cuda_ci
from sglang.test.kits.unified_radix_cache_kit import UnifiedRadixTreeTestMixin
@@ -102,5 +107,59 @@ class TestUnifiedQwen3HiCachePP(UnifiedRadixTreeTestMixin, CustomTestCase):
kill_process_tree(cls.process.pid)
class TestUnifiedQwen3HiCachePPL3(AccuracyTwoPassMixin, CustomTestCase):
"""Qwen3-32B + HiCache L3 (file backend) + PP + UnifiedRadixCache."""
gsm8k_threshold = 0.8
@classmethod
def setUpClass(cls):
cls.model = QWEN3_32B_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST
cls.hicache_dir = tempfile.mkdtemp(prefix="hicache_l3_pp_")
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=[
"--trust-remote-code",
"--tp-size",
"2",
"--pp-size",
"2",
"--mem-fraction-static",
"0.8",
"--cuda-graph-max-bs",
"32",
"--max-total-tokens",
"14000",
"--disable-piecewise-cuda-graph",
"--model-loader-extra-config",
'{"enable_multithread_load": true, "num_threads": 64}',
"--enable-hierarchical-cache",
"--hicache-write-policy",
"write_through",
"--hicache-storage-prefetch-policy",
"wait_complete",
"--hicache-io-backend",
"direct",
"--hicache-mem-layout",
"page_first_direct",
"--hicache-storage-backend",
"file",
],
env={
"SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",
"SGLANG_HICACHE_FILE_BACKEND_STORAGE_DIR": cls.hicache_dir,
},
)
@classmethod
def tearDownClass(cls):
kill_process_tree(cls.process.pid)
if os.path.isdir(cls.hicache_dir):
shutil.rmtree(cls.hicache_dir, ignore_errors=True)
if __name__ == "__main__":
unittest.main()