[CI] Gate Kimi-K3 acceptance length on the GSM8K average (#34145)

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):