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

[UnifiedTree] Use Qwen3-32B in unified radix pp kl tests and set KL threshold to 0.005 (#27931)

Browse Source
This commit is contained in:
Chao Shi
2026-06-14 13:15:35 +08:00
committed by GitHub
parent 91c63aeb4d
commit 5fb4e2d02e
1 changed files with 4 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_hicache_pp_kl.py
+4 -4
View File
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
register_cuda_ci(est_time=400, stage="base-c", runner_config="4-gpu-h100") register_cuda_ci(est_time=400, stage="base-c", runner_config="4-gpu-h100")
QWEN3_30B_MODEL = "Qwen/Qwen3-30B-A3B-FP8" QWEN3_32B_MODEL = "Qwen/Qwen3-32B"
def _assert_pp_decode_cached_tokens(result, history_len, output_len, label): def _assert_pp_decode_cached_tokens(result, history_len, output_len, label):
@@ -28,12 +28,12 @@ def _assert_pp_decode_cached_tokens(result, history_len, output_len, label):
class TestUnifiedQwen3HiCachePP(UnifiedRadixTreeTestMixin, CustomTestCase): class TestUnifiedQwen3HiCachePP(UnifiedRadixTreeTestMixin, CustomTestCase):
"""Qwen3-30B-A3B-FP8 + HiCache + PP + UnifiedRadixCache.""" """Qwen3-32B + HiCache + PP + UnifiedRadixCache."""
hicache_io_backend = "direct" hicache_io_backend = "direct"
hicache_mem_layout = "page_first_direct" hicache_mem_layout = "page_first_direct"
max_running_requests = 2 max_running_requests = 2
kl_threshold = 0.012 kl_threshold = 0.005
gsm8k_threshold = 0.7 gsm8k_threshold = 0.7
num_gsm8k_questions = 50 num_gsm8k_questions = 50
mmlu_threshold = 0.7 mmlu_threshold = 0.7
@@ -61,7 +61,7 @@ class TestUnifiedQwen3HiCachePP(UnifiedRadixTreeTestMixin, CustomTestCase):
@classmethod @classmethod
def setUpClass(cls): def setUpClass(cls):
cls.model = QWEN3_30B_MODEL cls.model = QWEN3_32B_MODEL
cls.base_url = DEFAULT_URL_FOR_TEST cls.base_url = DEFAULT_URL_FOR_TEST
cls.process = popen_launch_server( cls.process = popen_launch_server(
cls.model, cls.model,