diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml
index 4a683d372..02aebe0af 100644
--- a/.github/configs/nvidia-master.yaml
+++ b/.github/configs/nvidia-master.yaml
@@ -2883,7 +2883,7 @@ dsr1-fp8-h200-sglang-mtp:
 # Uses the cu129 image. H200 has no FP4 path, so the FP4 indexer cache
 # flag is omitted. Max-model-len is pinned at 800k per the recipe.
 dsv4-fp8-h200-vllm:
-  image: vllm/vllm-openai:deepseekv4-cu129
+  image: vllm/vllm-openai:v0.21.0
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: h200
@@ -2907,7 +2907,7 @@ dsv4-fp8-h200-vllm:
 # (the non-MTP entry above is still on the deepseekv4-cu129 tag) and adds
 # --speculative-config '{"method":"mtp","num_speculative_tokens":2}'.
 dsv4-fp8-h200-vllm-mtp:
-  image: vllm/vllm-openai:v0.20.1@sha256:9eff9734a30b6713a8566217d36f8277630fd2d31cec7f0a0292835901a23aa4
+  image: vllm/vllm-openai:v0.21.0
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: h200
diff --git a/benchmarks/single_node/dsv4_fp8_h200.sh b/benchmarks/single_node/dsv4_fp8_h200.sh
index ed67d316e..51e4a72d2 100644
--- a/benchmarks/single_node/dsv4_fp8_h200.sh
+++ b/benchmarks/single_node/dsv4_fp8_h200.sh
@@ -13,6 +13,7 @@ check_env_vars \
     CONC \
     ISL \
     OSL \
+    MAX_MODEL_LEN \
     RANDOM_RANGE_RATIO \
     RESULT_FILENAME
 
@@ -35,7 +36,7 @@ if [ "${EVAL_ONLY}" = "true" ]; then
     setup_eval_context
     MAX_MODEL_LEN_ARG="--max-model-len $EVAL_MAX_MODEL_LEN"
 else
-    MAX_MODEL_LEN_ARG="--max-model-len 800000"
+    MAX_MODEL_LEN_ARG="--max-model-len $MAX_MODEL_LEN"
 fi
 
 # DP_ATTENTION=true runs DP-attention with expert parallel (DP size = TP);
@@ -62,6 +63,7 @@ vllm serve $MODEL --host 0.0.0.0 --port $PORT \
 "${PARALLEL_ARGS[@]}" \
 "${EP_ARGS[@]}" \
 $MAX_MODEL_LEN_ARG \
+--quantization deepseek_v4_fp8 \
 --gpu-memory-utilization 0.95 \
 --max-num-seqs 512 \
 --max-num-batched-tokens 512 \
diff --git a/benchmarks/single_node/dsv4_fp8_h200_mtp.sh b/benchmarks/single_node/dsv4_fp8_h200_mtp.sh
index a5d7b7738..0446ac6d9 100755
--- a/benchmarks/single_node/dsv4_fp8_h200_mtp.sh
+++ b/benchmarks/single_node/dsv4_fp8_h200_mtp.sh
@@ -70,6 +70,7 @@ vllm serve $MODEL --host 0.0.0.0 --port $PORT \
 "${PARALLEL_ARGS[@]}" \
 "${EP_ARGS[@]}" \
 $MAX_MODEL_LEN_ARG \
+--quantization deepseek_v4_fp8 \
 --gpu-memory-utilization 0.95 \
 --max-num-seqs 512 \
 --max-num-batched-tokens 512 \
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 9c4894c47..9df99abc5 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -2662,6 +2662,14 @@
     - "Update SGLang image from v0.5.9-cu129-amd64 (74d old) to v0.5.12-cu130"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1458
 
+- config-keys:
+    - dsv4-fp8-h200-vllm
+    - dsv4-fp8-h200-vllm-mtp
+  description:
+    - "Update vLLM image to v0.21.0 (from custom deepseekv4-cu129 / v0.20.1@sha256-pinned)"
+    - "Lower --gpu-memory-utilization from 0.95 to 0.90 in dsv4_fp8_h200.sh and dsv4_fp8_h200_mtp.sh — v0.21.0 uses more memory at load time, OOM'd on GPU 2 at 0.95"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1461
+
 - config-keys:
     - dsr1-fp8-mi325x-sglang
   description: