Drop the mmlu case from the unified radix cache kit (#34667)

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_hicache_pp_kl.py

@@ -36,7 +36,6 @@ class TestUnifiedQwen3HiCachePP(UnifiedRadixTreeTestMixin, CustomTestCase):
     kl_threshold = 0.005
     gsm8k_threshold = 0.7
     num_gsm8k_questions = 50
-    mmlu_threshold = 0.7
     decode_cache_assert = staticmethod(_assert_pp_decode_cached_tokens)
 
     def test_gsm8k(self):

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_cp.py

@@ -24,7 +24,6 @@ class TestUnifiedQwen3HiCacheCP(UnifiedRadixTreeTestMixin, CustomTestCase):
     max_running_requests = 32
     kl_threshold = 0.005
     gsm8k_threshold = 0.7
-    mmlu_threshold = 0.7
 
     @classmethod
     def setUpClass(cls):

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dcp.py

@@ -42,7 +42,6 @@ class TestUnifiedKimiLinearDcpHiCache(UnifiedRadixTreeTestMixin, CustomTestCase)
 
     kl_threshold = 0.01
     gsm8k_threshold = 0.85
-    mmlu_threshold = 0.4
     prefill_cache_assert = staticmethod(
         make_mamba_prefill_assert(chunk_size=WIDENED_PAGE)
     )

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_hybrid_bitexact.py

@@ -25,9 +25,9 @@ Measured avg_kl_div:
 extra coverage rather than as the guard for one fix. CI runs `1-gpu-large`, which
 is SM90.
 
-These classes do not use UnifiedRadixTreeTestMixin: it bundles gsm8k and mmlu,
-which an undertrained checkpoint cannot gate on, and each class here runs the
-harness its regression was actually reproduced with.
+These classes do not use UnifiedRadixTreeTestMixin: it bundles a gsm8k case an
+undertrained checkpoint cannot gate on, and each class here runs the harness its
+regression was actually reproduced with.
 
 The imported `test_`-prefixed helpers are aliased so pytest does not collect them
 as tests.

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_swa.py

@@ -8,7 +8,6 @@ from sglang.test.test_utils import (
     DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
     DEFAULT_URL_FOR_TEST,
     CustomTestCase,
-    is_in_ci,
     popen_launch_server,
 )
 
@@ -22,11 +21,6 @@ class TestUnifiedSWARadixCache(UnifiedRadixTreeTestMixin, CustomTestCase):
 
     kl_threshold = 0.03
     gsm8k_threshold = 0.7
-    mmlu_threshold = 0.7
-
-    @unittest.skipIf(is_in_ci(), "SWA model mmlu eval not stable enough")
-    def test_mmlu(self):
-        super().test_mmlu()
 
     @classmethod
     def setUpClass(cls):