From 1ebd6fab6c6cf8910bd3f8a13c69d5c6720a6703 Mon Sep 17 00:00:00 2001 From: Liangsheng Yin Date: Sat, 8 Aug 2026 20:28:42 -0700 Subject: [PATCH] [CI] Gate Kimi-K3 acceptance length on the GSM8K average (#34145) --- test/registered/models_e2e/test_kimi_k3_b300.py | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/test/registered/models_e2e/test_kimi_k3_b300.py b/test/registered/models_e2e/test_kimi_k3_b300.py index b38004670..33fec0b49 100644 --- a/test/registered/models_e2e/test_kimi_k3_b300.py +++ b/test/registered/models_e2e/test_kimi_k3_b300.py @@ -40,8 +40,15 @@ class TestKimiK3B300LowLatency(GSM8KMixin, SpecDecodingMixin, CustomTestCase): gsm8k_score_threshold = 0.95 gsm8k_num_examples = 200 - accept_length_thres = 6.6 - bs_1_speed_thres = 440 + # Gated on GSM8K rather than on test_bs_1_speed below: a 200-question + # average holds steady when a numerics change moves where the single + # greedy prompt hits EOS. + gsm8k_accept_length_thres = 4.5 + # Both scale with how far that one greedy prompt runs, and speed is + # end-to-end, so launch and TTFT are amortized over the output -- it sits + # well below the steady decode rate the server logs. Coarse guards only. + accept_length_thres = 4.0 + bs_1_speed_thres = 300 @classmethod def setUpClass(cls):