diff --git a/benchmarks/llm-d/Dockerfile b/benchmarks/llm-d/Dockerfile
index 8494027604..cfc04557c5 100644
--- a/benchmarks/llm-d/Dockerfile
+++ b/benchmarks/llm-d/Dockerfile
@@ -22,7 +22,7 @@
 #     --build-arg EPP_FROM_IMAGE="$EPP_FROM_IMAGE" \
 #     --build-arg ROUTING_SIDECAR_IMAGE="$ROUTING_SIDECAR_IMAGE" \
 #     --build-arg ENVOY_FROM_IMAGE="$ENVOY_FROM_IMAGE" ...
-ARG VLLM_BASE=vllm/vllm-openai:v0.23.0
+ARG VLLM_BASE=vllm/vllm-openai:v0.26.0
 ARG EPP_FROM_IMAGE=ghcr.io/llm-d/llm-d-router-endpoint-picker:v0.9.0
 ARG ROUTING_SIDECAR_IMAGE=ghcr.io/llm-d/llm-d-router-disagg-sidecar:v0.9.0
 ARG ENVOY_FROM_IMAGE=envoyproxy/envoy:distroless-v1.33.2
diff --git a/benchmarks/multi_node/llm-d-recipes/dsv4-fp4-gb200-low-latency.yaml b/benchmarks/multi_node/llm-d-recipes/dsv4-fp4-gb200-low-latency.yaml
index f0cd876e4c..e29106062d 100644
--- a/benchmarks/multi_node/llm-d-recipes/dsv4-fp4-gb200-low-latency.yaml
+++ b/benchmarks/multi_node/llm-d-recipes/dsv4-fp4-gb200-low-latency.yaml
@@ -18,7 +18,7 @@
 #      decode ep:1/dp-attn:false in the master key is what selects the
 #      pure-TP path in server.sh. Decode knobs are latency-sized: max-num-seqs
 #      256, max-num-batched-tokens 256, max-cudagraph-capture-size 256,
-#      FULL_DECODE_ONLY graphs, stream-interval 50, gpu-memory-utilization 0.9.
+#      FULL_DECODE_ONLY graphs, gpu-memory-utilization 0.9.
 #   2. PREFILL is IDENTICAL to the mid-curve/high-tpt prefill (copied verbatim,
 #      DEP8 with enable-ep-weight-filter + moe-backend deep_gemm_mega_moe). An
 #      earlier srt-slurm-style low-latency prefill (weight-offload, no filter,
@@ -51,7 +51,7 @@ plugins:
   - type: prefix-cache-scorer
   - type: queue-scorer
   - type: active-request-scorer
-  - type: max-score-picker
+  - type: weighted-random-picker
 
 schedulingProfiles:
   - name: prefill
@@ -63,12 +63,13 @@ schedulingProfiles:
         weight: 2
       - pluginRef: active-request-scorer
         weight: 2
-      - pluginRef: max-score-picker
+      - pluginRef: weighted-random-picker
   - name: decode
     plugins:
       - pluginRef: decode-filter
       - pluginRef: active-request-scorer
-      - pluginRef: max-score-picker
+        weight: 2
+      - pluginRef: weighted-random-picker
 
 dataLayer:
   discovery:
@@ -89,7 +90,10 @@ prefill:
   extra-args: >-
     --kv-cache-dtype fp8
     --enforce-eager
-    --gpu-memory-utilization 0.9
+    --gpu-memory-utilization 0.95
+    --max-model-len 9280
+    --max-num-seqs 16
+    --max-num-batched-tokens 32768
     --enable-cumem-allocator
     --no-enable-prefix-caching
     --no-async-scheduling
@@ -98,6 +102,7 @@ prefill:
     --moe-backend deep_gemm_mega_moe
     --enable-ep-weight-filter
     --no-disable-hybrid-kv-cache-manager
+    --no-enable-flashinfer-autotune
     --numa-bind
   env:
     NCCL_CUMEM_ENABLE: "1"
@@ -123,14 +128,16 @@ prefill:
     NVSHMEM_REMOTE_TRANSPORT: "none"
     NVSHMEM_ENABLE_NIC_PE_MAPPING: "1"
     NVSHMEM_HCA_PE_MAPPING: "mlx5_0:1:1,mlx5_1:1:1,mlx5_3:1:1,mlx5_4:1:1"
-    VLLM_USE_NCCL_SYMM_MEM: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "0"
     VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
     VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
     TILELANG_CLEANUP_TEMP_FILES: "1"
-    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_CUMEM_HANDLE_TYPE: "FABRIC"
+    NVSHMEM_DISABLE_CUDA_VMM: "0"
     VLLM_SKIP_P2P_CHECK: "1"
-    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "0"
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
     VLLM_USE_DEEP_GEMM: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
     NVIDIA_GDRCOPY: "enabled"
     VLLM_HTTP_TIMEOUT_KEEP_ALIVE: "120"
     # Harden distributed coordination against transient gloo/NCCL resets under
@@ -156,7 +163,6 @@ decode:
     --no-enable-flashinfer-autotune
     --block-size 256
     --compilation-config {"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}
-    --stream-interval 50
     --tokenizer-mode deepseek_v4
     --no-disable-hybrid-kv-cache-manager
     --enable-sleep-mode
@@ -171,9 +177,13 @@ decode:
     # rc added so KV falls back to InfiniBand, not TCP (see prefill note).
     UCX_TLS: "cuda_copy,cuda_ipc,rc,tcp"
     UCX_CUDA_IPC_ENABLE_MNNVL: "y"
-    VLLM_USE_NCCL_SYMM_MEM: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "0"
     TILELANG_CLEANUP_TEMP_FILES: "1"
+    NVSHMEM_CUMEM_HANDLE_TYPE: "FABRIC"
+    NVSHMEM_DISABLE_CUDA_VMM: "0"
     VLLM_SKIP_P2P_CHECK: "1"
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
     NVIDIA_GDRCOPY: "enabled"
     TORCH_DISTRIBUTED_DEFAULT_TIMEOUT: "1800"
 
diff --git a/benchmarks/multi_node/llm-d-recipes/dsv4-fp4-gb200-mid-curve-megamoe.yaml b/benchmarks/multi_node/llm-d-recipes/dsv4-fp4-gb200-mid-curve-megamoe.yaml
index 2103ecf5c7..f82a79f28c 100644
--- a/benchmarks/multi_node/llm-d-recipes/dsv4-fp4-gb200-mid-curve-megamoe.yaml
+++ b/benchmarks/multi_node/llm-d-recipes/dsv4-fp4-gb200-mid-curve-megamoe.yaml
@@ -45,7 +45,7 @@ plugins:
   - type: prefix-cache-scorer
   - type: queue-scorer
   - type: active-request-scorer
-  - type: max-score-picker
+  - type: weighted-random-picker
 
 schedulingProfiles:
   - name: prefill
@@ -57,12 +57,13 @@ schedulingProfiles:
         weight: 2
       - pluginRef: active-request-scorer
         weight: 2
-      - pluginRef: max-score-picker
+      - pluginRef: weighted-random-picker
   - name: decode
     plugins:
       - pluginRef: decode-filter
       - pluginRef: active-request-scorer
-      - pluginRef: max-score-picker
+        weight: 2
+      - pluginRef: weighted-random-picker
 
 dataLayer:
   discovery:
@@ -87,7 +88,10 @@ prefill:
   extra-args: >-
     --kv-cache-dtype fp8
     --enforce-eager
-    --gpu-memory-utilization 0.9
+    --gpu-memory-utilization 0.95
+    --max-model-len 9280
+    --max-num-seqs 16
+    --max-num-batched-tokens 32768
     --enable-cumem-allocator
     --no-enable-prefix-caching
     --no-async-scheduling
@@ -96,6 +100,7 @@ prefill:
     --moe-backend deep_gemm_mega_moe
     --enable-ep-weight-filter
     --no-disable-hybrid-kv-cache-manager
+    --no-enable-flashinfer-autotune
     --numa-bind
   env:
     NCCL_CUMEM_ENABLE: "1"
@@ -121,14 +126,16 @@ prefill:
     NVSHMEM_REMOTE_TRANSPORT: "none"
     NVSHMEM_ENABLE_NIC_PE_MAPPING: "1"
     NVSHMEM_HCA_PE_MAPPING: "mlx5_0:1:1,mlx5_1:1:1,mlx5_3:1:1,mlx5_4:1:1"
-    VLLM_USE_NCCL_SYMM_MEM: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "0"
     VLLM_SPARSE_INDEXER_MAX_LOGITS_MB: "1024"
     VLLM_MAX_TOKENS_PER_EXPERT_FP4_MOE: "2048"
     TILELANG_CLEANUP_TEMP_FILES: "1"
-    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_CUMEM_HANDLE_TYPE: "FABRIC"
+    NVSHMEM_DISABLE_CUDA_VMM: "0"
     VLLM_SKIP_P2P_CHECK: "1"
-    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "0"
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
     VLLM_USE_DEEP_GEMM: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
     NVIDIA_GDRCOPY: "enabled"
     VLLM_HTTP_TIMEOUT_KEEP_ALIVE: "120"
     # Harden distributed coordination against transient gloo/NCCL connection
@@ -146,12 +153,13 @@ decode:
     --max-num-seqs 512
     --max-num-batched-tokens 512
     --max-cudagraph-capture-size 512
-    --gpu-memory-utilization 0.85
+    --gpu-memory-utilization 0.9
+    --max-model-len 9280
     --enable-cumem-allocator
     --no-enable-prefix-caching
+    --no-enable-flashinfer-autotune
     --block-size 256
     --compilation-config {"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}
-    --stream-interval 50
     --tokenizer-mode deepseek_v4
     --moe-backend deep_gemm_mega_moe
     --enable-ep-weight-filter
@@ -170,12 +178,14 @@ decode:
     NVSHMEM_REMOTE_TRANSPORT: "none"
     NVSHMEM_ENABLE_NIC_PE_MAPPING: "1"
     NVSHMEM_HCA_PE_MAPPING: "mlx5_0:1:1,mlx5_1:1:1,mlx5_3:1:1,mlx5_4:1:1"
-    VLLM_USE_NCCL_SYMM_MEM: "1"
+    VLLM_USE_NCCL_SYMM_MEM: "0"
     TILELANG_CLEANUP_TEMP_FILES: "1"
-    TORCH_SYMMMEM: "NVSHMEM"
+    NVSHMEM_CUMEM_HANDLE_TYPE: "FABRIC"
+    NVSHMEM_DISABLE_CUDA_VMM: "0"
     VLLM_SKIP_P2P_CHECK: "1"
-    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "0"
+    VLLM_RANDOMIZE_DP_DUMMY_INPUTS: "1"
     VLLM_USE_DEEP_GEMM: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
     NVIDIA_GDRCOPY: "enabled"
     # See prefill note: harden the decode's cross-node DP8/EP8 coordination
     # against transient gloo resets under the conc4096 warmup burst.
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 73a83024c9..7853626445 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -5107,7 +5107,7 @@ dsv4-fp4-gb200-dynamo-vllm:
 # TODO: change image to official llmd image.
 # Build source: benchmarks/llm-d/Dockerfile.
 dsv4-fp4-gb200-llmd-vllm:
-  image: ghcr.io/ezrasilvera/llm-d-nokube-vllm:vllm0.23
+  image: quay.io/rh-ee-imarkov/llm-d-nokube-vllm:vllm0.26@sha256:a9095d4c835935c4070be2040de0a5ef3b44098f603092ab66d743b0e731b7b4
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: gb200
@@ -5143,28 +5143,6 @@ dsv4-fp4-gb200-llmd-vllm:
           - "DECODE_NODES=2"
           - "GPUS_PER_NODE=4"
 
-      # Low-middle: 1 prefill DEP8 + 4 decode TP8.
-      - spec-decoding: "none"
-        conc-list: [256, 512]
-        prefill:
-          num-worker: 1
-          tp: 1
-          ep: 8
-          dp-attn: true
-          additional-settings:
-          - "PREFILL_NODES=2"
-          - "GPUS_PER_NODE=4"
-          - "CONFIG_FILE=dsv4-fp4-gb200-low-latency.yaml"
-        decode:
-          num-worker: 4
-          tp: 8
-          ep: 1
-          dp-attn: false
-          additional-settings:
-          - "DECODE_NODES=8"
-          - "DECODE_WORKERS=4"
-          - "GPUS_PER_NODE=4"
-
       # Mid curve: 1 prefill DEP8 + 1 decode DEP8.
       - spec-decoding: "none"
         conc-list: [256, 512, 1024]
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 4dd819f46d..abae8b7b42 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5460,3 +5460,13 @@
     - "Use vllm/vllm-openai:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 with prefix caching enabled, default KV-cache dtype, FlashInfer TRT-LLM attention with FP8 indexer KV, and the EAGLE3 drafter on FLASH_ATTN."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2328
   
+
+- config-keys:
+    - dsv4-fp4-gb200-llmd-vllm
+  description:
+    - "Decode: gpu-memory-utilization 0.85->0.9, add max-model-len 9280, disable NCCL symmetric memory, add no-enable-flashinfer-autotune, drop stream-interval (unsupported by rust frontend)"
+    - "Prefill: gpu-memory-utilization 0.9->0.95, add max-model-len 9280, add max-num-seqs 16 + max-num-batched-tokens 32768, enable rust frontend, enable randomize-dp-dummy-inputs"
+    - "EPP: switch from max-score-picker to weighted-random-picker for better load distribution under high concurrency"
+    - "Low-latency recipe: same prefill optimizations, decode NCCL symmetric memory disabled + rust frontend enabled, drop stream-interval"
+    - "Sweep: drop low-middle point (c256/512, 1P×4D) from llmd-vllm key; low-latency recipe runs at c1 only"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2498
diff --git a/runners/launch_gb200-nv.sh b/runners/launch_gb200-nv.sh
index 672ccd0414..a54f568d32 100755
--- a/runners/launch_gb200-nv.sh
+++ b/runners/launch_gb200-nv.sh
@@ -16,6 +16,89 @@ SQUASH_DIR="/mnt/lustre01/users-public/sa-shared"
 POWER_SRT_SLURM_URL="https://github.com/edwingao28/srt-slurm.git"
 POWER_SRT_SLURM_PIN="6fc1bed01a0b82dae0088a105c03ce0cfb353443"
 
+# Enroot 3.x does not parse Docker's tag@digest syntax. For digest-pinned
+# images, use its explicit registry syntax and pass the digest as the
+# manifest reference so the import remains immutable.
+enroot_uri_for_image() {
+    local image="$1"
+    local image_without_digest="$image"
+    local digest=""
+    local first_component registry repository repository_dir repository_name
+
+    if [[ "$image" == *@sha256:* ]]; then
+        image_without_digest="${image%@*}"
+        digest="${image##*@}"
+    fi
+
+    first_component="${image_without_digest%%/*}"
+    if [[ "$image_without_digest" == */* && ( "$first_component" == *.* || "$first_component" == *:* || "$first_component" == "localhost" ) ]]; then
+        registry="$first_component"
+        repository="${image_without_digest#*/}"
+    else
+        registry="registry-1.docker.io"
+        repository="$image_without_digest"
+    fi
+
+    if [[ -z "$digest" ]]; then
+        if [[ "$registry" == "registry-1.docker.io" ]]; then
+            printf 'docker://%s\n' "$image"
+        else
+            printf 'docker://%s#%s\n' "$registry" "$repository"
+        fi
+        return
+    fi
+
+    repository_dir="${repository%/*}"
+    repository_name="${repository##*/}"
+    repository_name="${repository_name%%:*}"
+    if [[ "$repository" == */* ]]; then
+        repository="${repository_dir}/${repository_name}"
+    else
+        repository="$repository_name"
+    fi
+    if [[ "$registry" == "registry-1.docker.io" && "$repository" != */* ]]; then
+        repository="library/$repository"
+    fi
+
+    printf 'docker://%s#%s:%s\n' "$registry" "$repository" "$digest"
+}
+
+# Concurrent matrix jobs import to the same shared-FS squash path.
+# Serialize imports and atomically replace invalid images so readers never
+# observe a partially written squash file.
+import_squash() {
+    local squash="$1" image="$2"
+    local lock="${squash}.lock"
+    local tmp="${squash}.tmp.$$"
+    local enroot_uri
+    enroot_uri=$(enroot_uri_for_image "$image") || exit 1
+    (
+        exec 9>"$lock"
+        flock -w 1800 9 || { echo "Failed to acquire lock for $squash" >&2; exit 1; }
+        if unsquashfs -l "$squash" > /dev/null 2>&1; then
+            echo "Squash file already exists and is valid, skipping import: $squash"
+        else
+            local enroot_runtime
+            enroot_runtime=$(mktemp -d "${TMPDIR:-/tmp}/enroot-import.XXXXXX") || exit 1
+            trap 'rm -rf -- "$enroot_runtime"' EXIT
+            export ENROOT_RUNTIME_PATH="$enroot_runtime"
+
+            rm -f "$squash" "$squash".tmp.*
+            if ! enroot import -o "$tmp" "$enroot_uri"; then
+                rm -f "$tmp"
+                echo "Error: enroot import failed for $enroot_uri" >&2
+                exit 1
+            fi
+            if ! unsquashfs -l "$tmp" > /dev/null 2>&1; then
+                rm -f "$tmp"
+                echo "Error: enroot import produced an invalid squash file: $tmp" >&2
+                exit 1
+            fi
+            mv -f "$tmp" "$squash" || exit 1
+        fi
+    ) || exit 1
+}
+
 if [[ "$FRAMEWORK" == "llmd-vllm" ]]; then
     if [[ "$MODEL_PREFIX" == "dsv4" && "$PRECISION" == "fp4" ]]; then
         export MODEL_PATH="/mnt/numa1/models/DeepSeek-V4-Pro"
@@ -26,31 +109,7 @@ if [[ "$FRAMEWORK" == "llmd-vllm" ]]; then
     fi
 
     SQUASH_FILE="${SQUASH_DIR}/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
-
-    # Enroot uses '#' between the registry and repository, e.g. docker://ghcr.io#org/image:tag.
-    case "$IMAGE" in
-        */*)
-            _registry="${IMAGE%%/*}"
-            _rest="${IMAGE#*/}"
-            if [[ "$_registry" == *.* || "$_registry" == *:* ]]; then
-                ENROOT_URL="docker://${_registry}#${_rest}"
-            else
-                ENROOT_URL="docker://${IMAGE}"  # bare hub repo
-            fi
-            ;;
-        *)  ENROOT_URL="docker://${IMAGE}" ;;
-    esac
-    echo "ENROOT_URL=$ENROOT_URL"
-
-    if [[ ! -s "$SQUASH_FILE" ]]; then
-        echo "enroot import -> $SQUASH_FILE"
-        enroot import -o "$SQUASH_FILE" "$ENROOT_URL" || {
-            echo "Error: enroot import failed for $ENROOT_URL" >&2
-            exit 1
-        }
-    else
-        echo "Reusing existing squash: $SQUASH_FILE"
-    fi
+    import_squash "$SQUASH_FILE" "$IMAGE"
 
     export LLMD_CONTAINER_ENGINE=pyxis
     export LLMD_SQUASH_FILE="$SQUASH_FILE"
@@ -207,84 +266,6 @@ uses_watchtower_shared_fs() {
 SQUASH_FILE="${SQUASH_DIR}/$(echo "$IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
 NGINX_SQUASH_FILE="${SQUASH_DIR}/$(echo "$NGINX_IMAGE" | sed 's/[\/:@#]/_/g').sqsh"
 
-# Enroot 3.x does not parse Docker's tag@digest syntax. For digest-pinned
-# images, use its explicit registry syntax and pass the digest as the
-# manifest reference so the import remains immutable.
-enroot_uri_for_image() {
-    local image="$1"
-    local image_without_digest="$image"
-    local digest=""
-    local first_component registry repository repository_dir repository_name
-
-    if [[ "$image" == *@sha256:* ]]; then
-        image_without_digest="${image%@*}"
-        digest="${image##*@}"
-    fi
-
-    first_component="${image_without_digest%%/*}"
-    if [[ "$image_without_digest" == */* && ( "$first_component" == *.* || "$first_component" == *:* || "$first_component" == "localhost" ) ]]; then
-        registry="$first_component"
-        repository="${image_without_digest#*/}"
-    else
-        registry="registry-1.docker.io"
-        repository="$image_without_digest"
-    fi
-
-    if [[ -z "$digest" ]]; then
-        if [[ "$registry" == "registry-1.docker.io" ]]; then
-            printf 'docker://%s\n' "$image"
-        else
-            printf 'docker://%s#%s\n' "$registry" "$repository"
-        fi
-        return
-    fi
-
-    repository_dir="${repository%/*}"
-    repository_name="${repository##*/}"
-    repository_name="${repository_name%%:*}"
-    if [[ "$repository" == */* ]]; then
-        repository="${repository_dir}/${repository_name}"
-    else
-        repository="$repository_name"
-    fi
-    if [[ "$registry" == "registry-1.docker.io" && "$repository" != */* ]]; then
-        repository="library/$repository"
-    fi
-
-    printf 'docker://%s#%s:%s\n' "$registry" "$repository" "$digest"
-}
-
-# Concurrent matrix jobs import to the same shared-FS squash path.
-# Serialize imports and atomically replace invalid images so readers never
-# observe a partially written squash file.
-import_squash() {
-    local squash="$1" image="$2"
-    local lock="${squash}.lock"
-    local tmp="${squash}.tmp.$$"
-    local enroot_uri
-    enroot_uri=$(enroot_uri_for_image "$image") || exit 1
-    (
-        exec 9>"$lock"
-        flock -w 1800 9 || { echo "Failed to acquire lock for $squash" >&2; exit 1; }
-        if unsquashfs -l "$squash" > /dev/null 2>&1; then
-            echo "Squash file already exists and is valid, skipping import: $squash"
-        else
-            rm -f "$squash" "$squash".tmp.*
-            if ! enroot import -o "$tmp" "$enroot_uri"; then
-                rm -f "$tmp"
-                echo "Error: enroot import failed for $enroot_uri" >&2
-                exit 1
-            fi
-            if ! unsquashfs -l "$tmp" > /dev/null 2>&1; then
-                rm -f "$tmp"
-                echo "Error: enroot import produced an invalid squash file: $tmp" >&2
-                exit 1
-            fi
-            mv -f "$tmp" "$squash" || exit 1
-        fi
-    ) || exit 1
-}
-
 import_squash "$SQUASH_FILE" "$IMAGE"
 import_squash "$NGINX_SQUASH_FILE" "$NGINX_IMAGE"