diff --git a/test/registered/models_e2e/test_dsa_glm52_cache_layer_split.py b/test/registered/models_e2e/test_dsa_glm52_pd_mtp_cp_layersplit.py
similarity index 91%
rename from test/registered/models_e2e/test_dsa_glm52_cache_layer_split.py
rename to test/registered/models_e2e/test_dsa_glm52_pd_mtp_cp_layersplit.py
index 8fbe21aa6..5bf33d972 100644
--- a/test/registered/models_e2e/test_dsa_glm52_cache_layer_split.py
+++ b/test/registered/models_e2e/test_dsa_glm52_pd_mtp_cp_layersplit.py
@@ -9,7 +9,7 @@ PD-disaggregated GLM-5.2 deployment: a layer-split prefill worker running
 interleave prefill-CP + layer split, and an ordinary decode worker that receives
 full cache shards via PD transfer.
 
-Runs nightly on an 8-GPU B200 runner (prefill TP=4 + decode TP=4).
+Runs on an 8-GPU B300 runner (prefill TP=4 + decode TP=4).
 """
 
 import unittest
@@ -20,11 +20,11 @@ from sglang.test.server_fixtures.disaggregation_fixture import (
     PDDisaggregationServerBase,
 )
 
-register_cuda_ci(est_time=750, stage="nightly", runner_config="8-gpu-b200")
+register_cuda_ci(est_time=750, stage="base-c", runner_config="8-gpu-b300")
 
 
 class TestGLM52DSACacheLayerSplit(PDDisaggregationServerBase, GSM8KMixin):
-    model = "nvidia/GLM-5.2-NVFP4"
+    model = "/data/radixark/model-cache/hub/models--nvidia--GLM-5.2-NVFP4/snapshots/aec724e8c7b8ee9db3b48c01c320f63f9cdaf8aa"
 
     # Full GSM8K test set (1319 questions) with a tight accuracy floor.
     gsm8k_accuracy_thres = 0.935