From 2ad9a5b576d61d61ac8132e1c1440a0cfd85dd9a Mon Sep 17 00:00:00 2001 From: Vladislav Nosivskoy Date: Fri, 19 Jun 2026 19:18:01 +0300 Subject: [PATCH] [UnifiedTree] Use dense model for HiCache+CP KL tests (#28726) --- .../test_unified_radix_cache_kl_cp.py | 12 +++++------- 1 file changed, 5 insertions(+), 7 deletions(-) diff --git a/test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_cp.py b/test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_cp.py index e103e8d73..5798895a8 100644 --- a/test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_cp.py +++ b/test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_cp.py @@ -13,11 +13,11 @@ from sglang.test.test_utils import ( register_cuda_ci(est_time=950, stage="extra-b", runner_config="4-gpu-h100") -QWEN3_30B_MODEL = "Qwen/Qwen3-30B-A3B-FP8" +QWEN3_32B_MODEL = "Qwen/Qwen3-32B" class TestUnifiedQwen3HiCacheCP(UnifiedRadixTreeTestMixin, CustomTestCase): - """Qwen3-30B-A3B-FP8 + HiCache + CP + UnifiedRadixCache.""" + """Qwen3-32B + HiCache + CP + UnifiedRadixCache.""" hicache_io_backend = "kernel" hicache_mem_layout = "page_first" @@ -28,7 +28,7 @@ class TestUnifiedQwen3HiCacheCP(UnifiedRadixTreeTestMixin, CustomTestCase): @classmethod def setUpClass(cls): - cls.model = QWEN3_30B_MODEL + cls.model = QWEN3_32B_MODEL cls.base_url = DEFAULT_URL_FOR_TEST cls.process = popen_launch_server( cls.model, @@ -38,10 +38,6 @@ class TestUnifiedQwen3HiCacheCP(UnifiedRadixTreeTestMixin, CustomTestCase): "--trust-remote-code", "--tp-size", "4", - "--moe-dp-size", - "1", - "--ep-size", - "4", "--attn-cp-size", "2", "--enable-prefill-context-parallel", @@ -51,6 +47,8 @@ class TestUnifiedQwen3HiCacheCP(UnifiedRadixTreeTestMixin, CustomTestCase): "32", "--max-running-requests", str(cls.max_running_requests), + "--max-total-tokens", + "14000", # loadback trigger "--disable-piecewise-cuda-graph", "--model-loader-extra-config", '{"enable_multithread_load": true, "num_threads": 64}',