diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index fdf705be71..463b2e83ad 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2515,7 +2515,7 @@ dsv4-fp8-h200-vllm:
 # field, so dp-attn=true is used as the existing vLLM script switch for DP4
 # layouts on 4 allocated GPUs.
 dsv4-fp4-b300-vllm:
-  image: vllm/vllm-openai:deepseekv4-cu130
+  image: vllm/vllm-openai:v0.20.0-cu130
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: b300
diff --git a/benchmarks/single_node/dsv4_fp4_b300_vllm.sh b/benchmarks/single_node/dsv4_fp4_b300_vllm.sh
index 6bb5b90496..3dfa992d78 100755
--- a/benchmarks/single_node/dsv4_fp4_b300_vllm.sh
+++ b/benchmarks/single_node/dsv4_fp4_b300_vllm.sh
@@ -42,8 +42,15 @@ if [ "${EP_SIZE:-1}" -gt 1 ]; then
     EP_ARGS=(--enable-expert-parallel)
 fi
 
+# Mega-MoE backend and the lower GMU only kick in on the DP-attn path,
+# per the vLLM v0.20.0 DeepSeek-V4-Pro recipe. All configs share the
+# FULL_AND_PIECEWISE compilation config.
+GMU_ARGS=()
+MOE_ARGS=()
 if [ "${DP_ATTENTION}" = "true" ]; then
     MAX_NUM_BATCHED_TOKENS=2048
+    GMU_ARGS=(--gpu-memory-utilization 0.85)
+    MOE_ARGS=(--moe-backend deep_gemm_mega_moe)
 else
     MAX_NUM_BATCHED_TOKENS=$(( ISL * 2 ))
 fi
@@ -66,15 +73,16 @@ start_gpu_monitor
 
 set -x
 vllm serve "$MODEL" --host 0.0.0.0 --port "$PORT" \
-    "${PARALLEL_ARGS[@]}" \
-    --pipeline-parallel-size 1 \
-    --kv-cache-dtype fp8 \
     --trust-remote-code \
+    --kv-cache-dtype fp8 \
     --block-size 256 \
     --no-enable-prefix-caching \
+    "${PARALLEL_ARGS[@]}" \
     "${EP_ARGS[@]}" \
+    "${GMU_ARGS[@]}" \
+    "${MOE_ARGS[@]}" \
     --compilation-config '{"cudagraph_mode":"FULL_AND_PIECEWISE","custom_ops":["all"]}' \
-    --attention_config.use_fp4_indexer_cache True \
+    --attention_config.use_fp4_indexer_cache=True \
     --tokenizer-mode deepseek_v4 \
     --tool-call-parser deepseek_v4 \
     --enable-auto-tool-choice \
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 2be38aefe8..025b1c0c2a 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -1969,3 +1969,12 @@
     - "Keeps the three validated 8k/1k points: low-latency 1P/1D TP8 conc=1, mid-curve 1P/1D DEP8 conc=256, and max-tpt 3P/1D DEP8 conc=4096"
     - "All three recipes run NATS/etcd on a dedicated infra node and use compute-node local NVMe model weights via /mnt/numa1/models/deepseek-v4-pro/"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1163
+
+- config-keys:
+    - dsv4-fp4-b300-vllm
+  description:
+    - "Pin image to vllm/vllm-openai:v0.20.0-cu130 (was floating deepseekv4-cu130 tag); DeepGEMM is preinstalled in this image"
+    - "Use --attention_config.use_fp4_indexer_cache=True and --compilation-config {\"cudagraph_mode\": \"FULL_AND_PIECEWISE\", \"custom_ops\": [\"all\"]} for all configs"
+    - "Gate --moe-backend deep_gemm_mega_moe and --gpu-memory-utilization 0.85 on DP_ATTENTION=true per the v0.20.0 recipe"
+    - "Drop --pipeline-parallel-size 1; keep --no-enable-prefix-caching and --max-cudagraph-capture-size 2048"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1220