From 1ac75c346355ac11e47d53e85af1c86a1a4bacce Mon Sep 17 00:00:00 2001 From: Liangsheng Yin Date: Thu, 11 Jun 2026 14:17:03 -0700 Subject: [PATCH] [CI] Fix GB300 TestDummyWithSBO crash: disable NaN assert and coredump for dummy weights (#27955) --- .../4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py | 10 ++++++++++ 1 file changed, 10 insertions(+) diff --git a/test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py b/test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py index b5ce3d47a..a4411f1a1 100644 --- a/test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py +++ b/test/registered/4-gpu-models/test_deepseek_v3_cutedsl_4gpu.py @@ -141,6 +141,16 @@ class TestDummyWithSBO(CustomTestCase): **os.environ, "SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK": "256", "SGLANG_MOE_NVFP4_DISPATCH": "0", + # Dummy random weights legitimately produce NaN logits; turn + # off the CI crash machinery (async assert, coredump on GPU + # exception, crash-time coredump) so NaN is sanitized with a + # warning instead of killing the scheduler. + "SGLANG_ENABLE_ASYNC_ASSERT": "0", + "SGLANG_CUDA_COREDUMP": "0", + # Already injected into os.environ by the test process when + # SGLANG_CUDA_COREDUMP=1, so it must be overridden explicitly. + "CUDA_ENABLE_COREDUMP_ON_EXCEPTION": "0", + "SGLANG_CUDA_COREDUMP_BEFORE_CRASH": "0", }, )