switch to native offloading

Author: cquil11

.github/configs/nvidia-master.yaml

@@ -1781,7 +1781,7 @@ dsv4-fp4-b200-vllm-agentic:
       # Experimental LMCache MP offload. LMCacheMPConnector currently requires
       # vLLM's hybrid KV manager to be disabled, so this is not an HMA/CSA/HCA
       # parity run against the no-offload path.
-      - { tp: 8, ep: 8, dp-attn: true, offloading: lmcache-mp, conc-list: [12, 16, 24, 32, 48, 64] }
+      - { tp: 8, ep: 8, dp-attn: true, offloading: cpu, conc-list: [12, 16, 24, 32, 48, 64] }
 
 dsv4-fp4-b200-trt:
   image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-9aa3715