From 771e38633216f709e754252a5c52b5b58843c2d3 Mon Sep 17 00:00:00 2001 From: Alison Shao <54658187+alisonshao@users.noreply.github.com> Date: Tue, 14 Jul 2026 16:21:10 -0700 Subject: [PATCH] Disable flaky DSV4-Flash FP4 BCG determinism test (nondeterminism from #30898 idle-rank dummy extend) (#31125) --- .../models_e2e/test_deepseek_v4_flash_fp4_b200.py | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/test/registered/models_e2e/test_deepseek_v4_flash_fp4_b200.py b/test/registered/models_e2e/test_deepseek_v4_flash_fp4_b200.py index 6f5530e63..af7f02902 100644 --- a/test/registered/models_e2e/test_deepseek_v4_flash_fp4_b200.py +++ b/test/registered/models_e2e/test_deepseek_v4_flash_fp4_b200.py @@ -170,6 +170,17 @@ class TestDSV4FlashFP4BreakableCudaGraphB200( gsm8k_accuracy_thres = 0.93 + @unittest.skip( + "Flaky: temp-0 outputs are nondeterministic under this recipe " + "(sparse-DP prefill replays the breakable CUDA graph with a " + "fabricated idle-rank dummy extend; its hidden states vary run to " + "run and perturb real tokens' logits through the shared EP grouped " + "GEMMs at capture buckets 4/16). Introduced by #30898; disabled " + "pending a proper fix that keeps BCG enabled. See #31125." + ) + def test_determinism_temp_zero(self): + pass + @classmethod def setUpClass(cls): cls.model = try_cached_model(MODEL)