diff --git a/test/registered/models_e2e/test_llama8b_nvfp4_kv_cache_sm120.py b/test/registered/models_e2e/test_llama8b_nvfp4_kv_cache_sm120.py
deleted file mode 100644
index d516ef121..000000000
--- a/test/registered/models_e2e/test_llama8b_nvfp4_kv_cache_sm120.py
+++ /dev/null
@@ -1,64 +0,0 @@
-import unittest
-
-from sglang.srt.utils.common import is_sm120_supported
-from sglang.test.accuracy_test_runner import AccuracyTestParams
-from sglang.test.ci.ci_register import register_cuda_ci
-from sglang.test.run_combined_tests import run_combined_tests
-from sglang.test.test_utils import CustomTestCase, ModelLaunchSettings
-
-register_cuda_ci(est_time=300, stage="extra-a", runner_config="1-gpu-small")
-
-LLAMA8B_NVFP4_MODEL = "nvidia/Llama-3.1-8B-Instruct-NVFP4"
-TP_SIZE = 1
-
-
-@unittest.skipUnless(
-    is_sm120_supported(), "requires at least 1 SM120 GPU with CUDA 12.8+"
-)
-class TestLlama8BNVFP4KVCacheSM120(CustomTestCase):
-    """Llama-3.1-8B-Instruct-NVFP4 with NVFP4 KV cache on SM120."""
-
-    def test_gsm8k(self):
-        variants = [
-            ModelLaunchSettings(
-                LLAMA8B_NVFP4_MODEL,
-                tp_size=TP_SIZE,
-                extra_args=[
-                    "--quantization",
-                    "modelopt_fp4",
-                    "--fp4-gemm-backend",
-                    "auto",
-                    "--kv-cache-dtype",
-                    "nvfp4",
-                    "--prefill-attention-backend",
-                    "flashinfer",
-                    "--decode-attention-backend",
-                    "trtllm_mha",
-                    "--page-size",
-                    "64",
-                    "--cuda-graph-backend-prefill=disabled",
-                ],
-                variant="NVFP4-GEMM+NVFP4-KV+SM120-XQA",
-            )
-        ]
-
-        run_combined_tests(
-            models=variants,
-            test_name="Llama-3.1-8B-Instruct-NVFP4-KV-SM120",
-            accuracy_params=AccuracyTestParams(
-                dataset="gsm8k",
-                # Full GSM8K measured locally with 1319 requested / 1314 scored:
-                # - FP8 KV: 0.6461187214611872
-                # - NVFP4 KV: 0.632420091324201
-                # Keep the threshold 0.015 below the NVFP4 KV score.
-                baseline_accuracy=0.632420091324201 - 0.015,
-                num_examples=1319,
-                num_threads=200,
-                max_tokens=512,
-                api="completion",
-            ),
-        )
-
-
-if __name__ == "__main__":
-    unittest.main()
diff --git a/test/registered/quant/test_llama8b_nvfp4_kv_cache_sm120.py b/test/registered/quant/test_llama8b_nvfp4_kv_cache_sm120.py
new file mode 100644
index 000000000..cbbf6349e
--- /dev/null
+++ b/test/registered/quant/test_llama8b_nvfp4_kv_cache_sm120.py
@@ -0,0 +1,51 @@
+import unittest
+
+from sglang.srt.utils.common import is_sm120_supported
+from sglang.test.ci.ci_register import register_cuda_ci
+from sglang.test.kits.eval_accuracy_kit import GSM8KMixin
+from sglang.test.server_fixtures.default_fixture import DefaultServerBase
+
+register_cuda_ci(
+    est_time=300,
+    stage="extra-a",
+    runner_config="1-gpu-small",
+    disabled="Temporarily disabled due to failing accuracy",
+)
+
+
+@unittest.skipUnless(
+    is_sm120_supported(), "requires at least 1 SM120 GPU with CUDA 12.8+"
+)
+class TestLlama8BNVFP4KVCacheSM120(GSM8KMixin, DefaultServerBase):
+    """Llama-3.1-8B-Instruct-NVFP4 with NVFP4 KV cache on SM120."""
+
+    model = "nvidia/Llama-3.1-8B-Instruct-NVFP4"
+    # Full GSM8K measured locally with 1319 requested / 1314 scored:
+    # - FP8 KV: 0.6461187214611872
+    # - NVFP4 KV: 0.632420091324201
+    # Keep the threshold 0.015 below the NVFP4 KV score.
+    gsm8k_accuracy_thres = 0.632420091324201 - 0.015
+    gsm8k_num_questions = 1319
+    gsm8k_num_threads = 200
+
+    other_args = [
+        "--quantization",
+        "modelopt_fp4",
+        "--fp4-gemm-backend",
+        "auto",
+        "--kv-cache-dtype",
+        "nvfp4",
+        "--prefill-attention-backend",
+        "flashinfer",
+        "--decode-attention-backend",
+        "trtllm_mha",
+        "--page-size",
+        "64",
+        "--mem-fraction-static",
+        "0.87",
+        "--cuda-graph-backend-prefill=disabled",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()