diff --git a/test/registered/models_e2e/test_qwen3_next_models.py b/test/registered/models_e2e/test_qwen3_next_models.py
index 89d40cd29..80953409e 100644
--- a/test/registered/models_e2e/test_qwen3_next_models.py
+++ b/test/registered/models_e2e/test_qwen3_next_models.py
@@ -57,7 +57,6 @@ class TestQwen3NextLazyExtraBufferLargePage(
 class TestQwen3NextLazyExtraBufferAllocFail(KLDivergenceMixin, DefaultServerBase):
     model = QWEN3_NEXT_MODEL
     cache_chunk_size = 64
-    gsm8k_accuracy_thres = 0.93
     kl_div_thres = 0.002
     other_args = _make_args(page_size=1, track_interval=2)
 
@@ -79,7 +78,6 @@ class TestQwen3NextLazyExtraBufferLargePageAllocFail(
 ):
     model = QWEN3_NEXT_MODEL
     cache_chunk_size = 64
-    gsm8k_accuracy_thres = 0.93
     kl_div_thres = 0.002
     other_args = _make_args(page_size=2, track_interval=2)
 
diff --git a/test/registered/utils/test_phase_checker.py b/test/registered/utils/test_phase_checker.py
index d7ee8da65..d9658fdd9 100644
--- a/test/registered/utils/test_phase_checker.py
+++ b/test/registered/utils/test_phase_checker.py
@@ -166,19 +166,20 @@ class TestUpdateAssertEnabled(CustomTestCase):
             text=True,
             timeout=180,
         )
-        # Accept both exit-0 (RuntimeError caught) and SIGABRT (-6, the kernel's
-        # device_assert killed the process directly before sync could raise).
-        # Either way, the kernel-side check fired — which is what we're verifying.
-        # The presence of the SimplePhaseChecker FAIL line in stdout confirms it.
+        # The FAIL line is the evidence that the kernel-side check fired. How the
+        # process then dies is not: the CUDA coredump handler may abort it, and sync
+        # otherwise raises cudaErrorAssert or a generic cudaErrorLaunchFailure
+        # depending on whether generation ran. Only exit 1 means it never fired.
         self.assertIn(
             "SimplePhaseChecker FAIL",
             result.stdout,
+            f"returncode {result.returncode}; "
             f"stdout: {result.stdout}\nstderr: {result.stderr}",
         )
-        self.assertIn(
+        self.assertNotEqual(
             result.returncode,
-            (0, -6),
-            f"unexpected returncode {result.returncode}; "
+            1,
+            "the mismatch did not fire device_assert; "
             f"stdout: {result.stdout}\nstderr: {result.stderr}",
         )
 
diff --git a/test/registered/vlm/test_vision_openai_server_a.py b/test/registered/vlm/test_vision_openai_server_a.py
index 2284610b3..3624e35b0 100644
--- a/test/registered/vlm/test_vision_openai_server_a.py
+++ b/test/registered/vlm/test_vision_openai_server_a.py
@@ -9,12 +9,16 @@ import unittest
 import openai
 
 from sglang.test.ci.ci_register import register_cuda_ci
-from sglang.test.vlm_utils import *
+from sglang.test.test_utils import (
+    DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+    DEFAULT_URL_FOR_TEST,
+    popen_launch_server,
+)
 from sglang.test.vlm_utils import (
+    IMAGE_MAN_IRONING_URL,
     AudioOpenAITestMixin,
     CustomTestCase,
     ImageOpenAITestMixin,
-    OmniOpenAITestMixin,
     TestOpenAIMLLMServerBase,
     VideoOpenAITestMixin,
     terminate_and_kill_process_tree,
@@ -231,7 +235,6 @@ del (
     ImageOpenAITestMixin,
     VideoOpenAITestMixin,
     AudioOpenAITestMixin,
-    OmniOpenAITestMixin,
 )