docs: Adds Qwen3-235B-A22B-FP8 recipes for TRTLLM agg and disagg (#4179)

Signed-off-by: Elnifio <elnifio0519@gmail.com>

recipes/qwen3-235b-a22b-fp8/model-cache/model-cache.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: model-cache
+spec:
+  accessModes:
+    - ReadWriteMany
+  resources:
+    requests:
+      storage: 300Gi
+  storageClassName: "your-storage-class-name"

recipes/qwen3-235b-a22b-fp8/model-cache/model-download.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: model-download
+spec:
+  backoffLimit: 3
+  completions: 1
+  parallelism: 1
+  template:
+    metadata:
+      labels:
+        app: model-download
+    spec:
+      restartPolicy: Never
+      containers:
+        - name: model-download
+          image: python:3.10-slim
+          command: ["sh", "-c"]
+          envFrom:
+            - secretRef:
+                name: hf-token-secret
+          env:
+            - name: MODEL_NAME
+              value: Qwen/Qwen3-235B-A22B-FP8
+            - name: HF_HOME
+              value: /model-store
+            - name: HF_HUB_ENABLE_HF_TRANSFER
+              value: "1"
+            - name: MODEL_REVISION
+              value: 39eb2b067ea6b8e3e1dd97d3cd0c7ffeaf3e1a35
+          args:
+            - |
+              set -eux
+              pip install --no-cache-dir huggingface_hub hf_transfer
+              hf download $MODEL_NAME --revision $MODEL_REVISION
+          volumeMounts:
+            - name: model-cache
+              mountPath: /model-store
+      volumes:
+      - name: model-cache
+        persistentVolumeClaim:
+          claimName: model-cache

recipes/qwen3-235b-a22b-fp8/trtllm/agg/deploy.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: agg-config
+data:
+  agg.yaml: |
+    backend: pytorch
+    trust_remote_code: true
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    moe_tensor_parallel_size: 1
+    enable_attention_dp: false
+    enable_chunked_prefill: true
+    build_config:
+      max_batch_size: 128
+      max_num_tokens: 8192
+      max_seq_len: 8192
+    kv_cache_config:
+      enable_block_reuse: true
+      free_gpu_memory_fraction: 0.8
+      dtype: auto
+    cache_transceiver_config:
+      backend: DEFAULT
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 128
+    disable_overlap_scheduler: false
+    print_iter_log: false
+---
+apiVersion: nvidia.com/v1alpha1
+kind: DynamoGraphDeployment
+metadata:
+  name: qwen3-235b-a22b-agg
+spec:
+  backendFramework: trtllm
+  pvcs:
+    - name: model-cache
+      create: false
+  services:
+    Frontend:
+      componentType: frontend
+      dynamoNamespace: qwen3-235b-a22b-agg
+      replicas: 1
+      extraPodSpec:
+        affinity:
+          podAntiAffinity:
+            requiredDuringSchedulingIgnoredDuringExecution:
+            - labelSelector:
+                matchExpressions:
+                - key: nvidia.com/dynamo-graph-deployment-name
+                  operator: In
+                  values:
+                    - qwen3-235b-a22b-agg-frontend
+              topologyKey: kubernetes.io/hostname
+        mainContainer:
+          image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:my-tag
+          args:
+            - python3 -m dynamo.frontend --router-mode kv --http-port 8000
+          command:
+            - /bin/sh
+            - -c
+    TrtllmWorker:
+      componentType: main
+      dynamoNamespace: qwen3-235b-a22b-agg
+      envFromSecret: hf-token-secret
+      sharedMemory:
+        size: 256Gi
+      extraPodSpec:
+        affinity:
+          nodeAffinity:
+            requiredDuringSchedulingIgnoredDuringExecution:
+              nodeSelectorTerms:
+              - matchExpressions:
+                - key: nvidia.com/gpu.present
+                  operator: In
+                  values:
+                  - "true"
+        mainContainer:
+          env:
+            - name: MODEL_PATH
+              value: /mnt/model-cache/hub/models--Qwen--Qwen3-235B-A22B-FP8/snapshots/39eb2b067ea6b8e3e1dd97d3cd0c7ffeaf3e1a35
+            - name: ENGINE_ARGS
+              value: /engine_configs/agg.yaml
+          command:
+            - /bin/sh
+            - -c
+          args:
+            - |
+              python3 -m dynamo.trtllm \
+              --model-path "${MODEL_PATH}" \
+              --served-model-name "Qwen/Qwen3-235B-A22B-FP8" \
+              --extra-engine-args "${ENGINE_ARGS}"
+          image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:my-tag
+          workingDir: /workspace/components/backends/trtllm
+          volumeMounts:
+            - name: agg-config
+              mountPath: /engine_configs
+            - name: model-cache
+              mountPath: /mnt/model-cache
+        volumes:
+          - name: agg-config
+            configMap:
+              name: agg-config
+          - name: model-cache
+            persistentVolumeClaim:
+              claimName: model-cache
+      replicas: 4
+      resources:
+        limits:
+          gpu: "4"
+        requests:
+          gpu: "4"

recipes/qwen3-235b-a22b-fp8/trtllm/agg/perf.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: qwen3-235b-a22b-bench
+spec:
+  backoffLimit: 1
+  completions: 1
+  parallelism: 1
+  template:
+    metadata:
+      labels:
+        app: qwen3-235b-a22b-bench
+    spec:
+      affinity:
+        podAntiAffinity:
+          requiredDuringSchedulingIgnoredDuringExecution:
+            - labelSelector:
+                matchExpressions:
+                  - key: nvidia.com/dynamo-graph-deployment-name
+                    operator: In
+                    values:
+                      - qwen3-235b-a22b-agg
+              topologyKey: kubernetes.io/hostname
+      containers:
+      - command:
+        - /bin/sh
+        - -c
+        - |
+          apt-get update && apt-get install -y curl jq procps git && apt-get clean
+          pip install git+https://github.com/ai-dynamo/aiperf.git@70af59489df24a601dba57604a7341966150b366;
+          echo "aiperf installation completed";
+          sysctl -w net.ipv4.ip_local_port_range="1024 65000"
+          cat /proc/sys/net/ipv4/ip_local_port_range
+          export COLUMNS=200
+          EPOCH=$(date +%s)
+          ## utility functions -- can be moved to a bash script / configmap
+          wait_for_model_ready() {
+            echo "Waiting for model '$TARGET_MODEL' at $ENDPOINT/v1/models (checking every 5s)..."
+            while ! curl -s "http://$ENDPOINT/v1/models" | jq -e --arg model "$TARGET_MODEL" '.data[]? | select(.id == $model)' >/dev/null 2>&1; do
+                echo "[$(date '+%H:%M:%S')] Model not ready yet, sleeping 5s before checking again http://$ENDPOINT/v1/models"
+                sleep 5
+            done
+            echo "✅ Model '$TARGET_MODEL' is now available!"
+            echo "Model '$TARGET_MODEL' is now available!"
+            curl -s "http://$ENDPOINT/v1/models" | jq .
+          }
+          run_perf() {
+            local concurrency=$1
+            local isl=$2
+            local osl=$3
+            key=concurrency_${concurrency}
+            export ARTIFACT_DIR="${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}/${key}"
+            mkdir -p "$ARTIFACT_DIR"
+            echo "ARTIFACT_DIR: $ARTIFACT_DIR"
+            aiperf profile --artifact-dir $ARTIFACT_DIR \
+                --model $TARGET_MODEL \
+                --tokenizer $TARGET_MODEL  \
+                --endpoint-type chat  \
+                --endpoint /v1/chat/completions \
+                --streaming \
+                --url http://$ENDPOINT \
+                --synthetic-input-tokens-mean $isl \
+                --synthetic-input-tokens-stddev 0 \
+                --output-tokens-mean $osl \
+                --output-tokens-stddev 0 \
+                --extra-inputs "max_tokens:$osl" \
+                --extra-inputs "min_tokens:$osl" \
+                --extra-inputs "ignore_eos:true" \
+                --extra-inputs "{\"nvext\":{\"ignore_eos\":true}}" \
+                --extra-inputs "repetition_penalty:1.0" \
+                --extra-inputs "temperature: 0.0" \
+                --concurrency $concurrency \
+                --request-count $((10*concurrency)) \
+                --warmup-request-count $concurrency \
+                --conversation-num 12800 \
+                --random-seed 100 \
+                --workers-max 252 \
+                -H 'Authorization: Bearer NOT USED' \
+                -H 'Accept: text/event-stream'\
+                --record-processors 32 \
+                --ui simple
+            echo "ARTIFACT_DIR: $ARTIFACT_DIR"
+            ls -la $ARTIFACT_DIR
+          }
+          #### Actual execution ####
+          wait_for_model_ready
+          mkdir -p "${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}"
+          # Calculate total concurrency based on per-GPU concurrency and GPU count
+          TOTAL_CONCURRENCY=$((CONCURRENCY_PER_GPU * DEPLOYMENT_GPU_COUNT))
+          echo "Calculated total concurrency: $TOTAL_CONCURRENCY (${CONCURRENCY_PER_GPU} per GPU × ${DEPLOYMENT_GPU_COUNT} GPUs)"
+          # Write input_config.json
+          cat > "${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}/input_config.json" <<EOF
+          {
+            "gpu_count": $DEPLOYMENT_GPU_COUNT,
+            "concurrency_per_gpu": $CONCURRENCY_PER_GPU,
+            "total_concurrency": $TOTAL_CONCURRENCY,
+            "mode": "$DEPLOYMENT_MODE",
+            "isl": $ISL,
+            "osl": $OSL,
+            "endpoint": "$ENDPOINT",
+            "model endpoint": "$TARGET_MODEL"
+          }
+          EOF
+
+          # Run perf with calculated total concurrency
+          run_perf $TOTAL_CONCURRENCY $ISL $OSL
+          echo "done with concurrency $TOTAL_CONCURRENCY"
+        env:
+        - name: TARGET_MODEL
+          value: Qwen/Qwen3-235B-A22B-FP8
+        - name: ENDPOINT
+          value: qwen3-235b-a22b-agg-frontend:8000
+        - name: CONCURRENCY_PER_GPU
+          value: "2"
+        - name: DEPLOYMENT_GPU_COUNT
+          value: "16"
+        - name: ISL
+          value: "4000"
+        - name: OSL
+          value: "200"
+        - name: DEPLOYMENT_MODE
+          value: agg
+        - name: AIPERF_HTTP_CONNECTION_LIMIT
+          value: "200"
+        - name: JOB_NAME
+          valueFrom:
+            fieldRef:
+              apiVersion: v1
+              fieldPath: metadata.labels['job-name']
+        - name: ROOT_ARTIFACT_DIR
+          value: /model-cache/perf
+        - name: HF_HOME
+          value: /model-cache
+        - name: PYTHONUNBUFFERED
+          value: "1"
+        image: python:3.12-slim
+        imagePullPolicy: IfNotPresent
+        name: perf
+        securityContext:
+          privileged: true
+        volumeMounts:
+        - name: model-cache
+          mountPath: /model-cache
+        workingDir: /workspace
+      imagePullSecrets:
+      - name: nvcrimagepullsecret
+      restartPolicy: Never
+      volumes:
+      - name: model-cache
+        persistentVolumeClaim:
+          claimName: model-cache

recipes/qwen3-235b-a22b-fp8/trtllm/disagg/deploy.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: prefill-config
+data:
+  prefill.yaml: |
+    backend: pytorch
+    trust_remote_code: true
+    tensor_parallel_size: 2
+    moe_tensor_parallel_size: 2
+    moe_expert_parallel_size: 1
+    enable_attention_dp: false
+    enable_chunked_prefill: false
+    build_config:
+      max_batch_size: 2
+      max_num_tokens: 8192
+      max_seq_len: 8192
+    kv_cache_config:
+      enable_block_reuse: true
+      free_gpu_memory_fraction: 0.7
+      dtype: fp8
+    cache_transceiver_config:
+      backend: DEFAULT
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 2
+    disable_overlap_scheduler: true
+    print_iter_log: false
+---
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: decode-config
+data:
+  decode.yaml: |
+    backend: pytorch
+    trust_remote_code: true
+    tensor_parallel_size: 4
+    moe_expert_parallel_size: 4
+    moe_tensor_parallel_size: 1
+    enable_attention_dp: false
+    enable_chunked_prefill: false
+    build_config:
+      max_batch_size: 512
+      max_num_tokens: 1024
+      max_seq_len: 8192
+    kv_cache_config:
+      enable_block_reuse: false
+      free_gpu_memory_fraction: 0.95
+      dtype: fp8
+    cache_transceiver_config:
+      backend: DEFAULT
+    cuda_graph_config:
+      enable_padding: true
+      max_batch_size: 512
+    disable_overlap_scheduler: false
+    print_iter_log: false
+---
+apiVersion: nvidia.com/v1alpha1
+kind: DynamoGraphDeployment
+metadata:
+  name: qwen3-235b-a22b-disagg
+spec:
+  backendFramework: trtllm
+  pvcs:
+    - name: model-cache
+      create: false
+  services:
+    Frontend:
+      componentType: frontend
+      dynamoNamespace: qwen3-235b-a22b-disagg
+      replicas: 1
+      extraPodSpec:
+        affinity:
+          podAntiAffinity:
+            requiredDuringSchedulingIgnoredDuringExecution:
+            - labelSelector:
+                matchExpressions:
+                - key: nvidia.com/dynamo-graph-deployment-name
+                  operator: In
+                  values:
+                    - qwen3-235b-a22b-disagg-frontend
+              topologyKey: kubernetes.io/hostname
+        mainContainer:
+          image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:my-tag
+          args:
+            - python3 -m dynamo.frontend --router-mode kv --http-port 8000
+          command:
+            - /bin/sh
+            - -c
+    TRTLLMPrefillWorker:
+      componentType: worker
+      subComponentType: prefill
+      dynamoNamespace: qwen3-235b-a22b-disagg
+      envFromSecret: hf-token-secret
+      replicas: 6
+      resources:
+        limits:
+          gpu: "2"
+      sharedMemory:
+        size: 256Gi
+      extraPodSpec:
+        affinity:
+          nodeAffinity:
+            requiredDuringSchedulingIgnoredDuringExecution:
+              nodeSelectorTerms:
+              - matchExpressions:
+                - key: nvidia.com/gpu.present
+                  operator: In
+                  values:
+                  - "true"
+        mainContainer:
+          env:
+            - name: MODEL_PATH
+              value: /mnt/model-cache/hub/models--Qwen--Qwen3-235B-A22B-FP8/snapshots/39eb2b067ea6b8e3e1dd97d3cd0c7ffeaf3e1a35
+            - name: ENGINE_ARGS
+              value: /engine_configs/prefill.yaml
+          image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:my-tag
+          workingDir: /workspace/components/backends/trtllm
+          command:
+            - /bin/sh
+            - -c
+          args:
+            - |
+              python3 -m dynamo.trtllm \
+                --model-path "${MODEL_PATH}" \
+                --served-model-name "Qwen/Qwen3-235B-A22B-FP8" \
+                --extra-engine-args "${ENGINE_ARGS}" \
+                --disaggregation-mode prefill \
+                --disaggregation-strategy prefill_first
+          volumeMounts:
+            - name: prefill-config
+              mountPath: /engine_configs
+            - name: model-cache
+              mountPath: /mnt/model-cache
+        volumes:
+          - name: prefill-config
+            configMap:
+              name: prefill-config
+          - name: model-cache
+            persistentVolumeClaim:
+              claimName: model-cache
+    TRTLLMDecodeWorker:
+      componentType: worker
+      subComponentType: decode
+      dynamoNamespace: qwen3-235b-a22b-disagg
+      envFromSecret: hf-token-secret
+      replicas: 1
+      resources:
+        limits:
+          gpu: "4"
+      sharedMemory:
+        size: 256Gi
+      extraPodSpec:
+        affinity:
+          nodeAffinity:
+            requiredDuringSchedulingIgnoredDuringExecution:
+              nodeSelectorTerms:
+              - matchExpressions:
+                - key: nvidia.com/gpu.present
+                  operator: In
+                  values:
+                  - "true"
+        mainContainer:
+          env:
+            - name: MODEL_PATH
+              value: /mnt/model-cache/hub/models--Qwen--Qwen3-235B-A22B-FP8/snapshots/39eb2b067ea6b8e3e1dd97d3cd0c7ffeaf3e1a35
+            - name: ENGINE_ARGS
+              value: /engine_configs/decode.yaml
+          image: nvcr.io/nvidia/ai-dynamo/tensorrtllm-runtime:my-tag
+          workingDir: /workspace/components/backends/trtllm
+          command:
+            - /bin/sh
+            - -c
+          args:
+            - |
+              python3 -m dynamo.trtllm \
+                --model-path "${MODEL_PATH}" \
+                --served-model-name "Qwen/Qwen3-235B-A22B-FP8" \
+                --extra-engine-args "${ENGINE_ARGS}" \
+                --disaggregation-mode decode \
+                --disaggregation-strategy prefill_first
+          volumeMounts:
+            - name: decode-config
+              mountPath: /engine_configs
+            - name: model-cache
+              mountPath: /mnt/model-cache
+        volumes:
+          - name: decode-config
+            configMap:
+              name: decode-config
+          - name: model-cache
+            persistentVolumeClaim:
+              claimName: model-cache

recipes/qwen3-235b-a22b-fp8/trtllm/disagg/perf.yaml

+# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: batch/v1
+kind: Job
+metadata:
+  name: qwen3-235b-a22b-disagg-bench
+spec:
+  backoffLimit: 1
+  completions: 1
+  parallelism: 1
+  template:
+    metadata:
+      labels:
+        app: qwen3-235b-a22b-disagg-bench
+    spec:
+      affinity:
+        podAntiAffinity:
+          requiredDuringSchedulingIgnoredDuringExecution:
+            - labelSelector:
+                matchExpressions:
+                  - key: nvidia.com/dynamo-graph-deployment-name
+                    operator: In
+                    values:
+                      - qwen3-235b-a22b-disagg
+              topologyKey: kubernetes.io/hostname
+      containers:
+      - command:
+        - /bin/sh
+        - -c
+        - |
+          apt-get update && apt-get install -y curl jq procps git && apt-get clean
+          pip install git+https://github.com/ai-dynamo/aiperf.git@70af59489df24a601dba57604a7341966150b366;
+          echo "aiperf installation completed";
+          sysctl -w net.ipv4.ip_local_port_range="1024 65000"
+          cat /proc/sys/net/ipv4/ip_local_port_range
+          export COLUMNS=200
+          EPOCH=$(date +%s)
+          ## utility functions -- can be moved to a bash script / configmap
+          wait_for_model_ready() {
+            echo "Waiting for model '$TARGET_MODEL' at $ENDPOINT/v1/models (checking every 5s)..."
+            while ! curl -s "http://$ENDPOINT/v1/models" | jq -e --arg model "$TARGET_MODEL" '.data[]? | select(.id == $model)' >/dev/null 2>&1; do
+                echo "[$(date '+%H:%M:%S')] Model not ready yet, sleeping 5s before checking again http://$ENDPOINT/v1/models"
+                sleep 5
+            done
+            echo "✅ Model '$TARGET_MODEL' is now available!"
+            echo "Model '$TARGET_MODEL' is now available!"
+            curl -s "http://$ENDPOINT/v1/models" | jq .
+          }
+          run_perf() {
+            local concurrency=$1
+            local isl=$2
+            local osl=$3
+            key=concurrency_${concurrency}
+            export ARTIFACT_DIR="${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}/${key}"
+            mkdir -p "$ARTIFACT_DIR"
+            echo "ARTIFACT_DIR: $ARTIFACT_DIR"
+            aiperf profile --artifact-dir $ARTIFACT_DIR \
+                --model $TARGET_MODEL \
+                --tokenizer $TARGET_MODEL  \
+                --endpoint-type chat  \
+                --endpoint /v1/chat/completions \
+                --streaming \
+                --url http://$ENDPOINT \
+                --synthetic-input-tokens-mean $isl \
+                --synthetic-input-tokens-stddev 0 \
+                --output-tokens-mean $osl \
+                --output-tokens-stddev 0 \
+                --extra-inputs "max_tokens:$osl" \
+                --extra-inputs "min_tokens:$osl" \
+                --extra-inputs "ignore_eos:true" \
+                --extra-inputs "{\"nvext\":{\"ignore_eos\":true}}" \
+                --extra-inputs "repetition_penalty:1.0" \
+                --extra-inputs "temperature: 0.0" \
+                --concurrency $concurrency \
+                --request-count $((10*concurrency)) \
+                --warmup-request-count $concurrency \
+                --conversation-num 12800 \
+                --random-seed 100 \
+                --workers-max 252 \
+                -H 'Authorization: Bearer NOT USED' \
+                -H 'Accept: text/event-stream'\
+                --record-processors 32 \
+                --ui simple
+            echo "ARTIFACT_DIR: $ARTIFACT_DIR"
+            ls -la $ARTIFACT_DIR
+          }
+          #### Actual execution ####
+          wait_for_model_ready
+          mkdir -p "${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}"
+          # Calculate total concurrency based on per-GPU concurrency and GPU count
+          TOTAL_CONCURRENCY=$((CONCURRENCY_PER_GPU * DEPLOYMENT_GPU_COUNT))
+          echo "Calculated total concurrency: $TOTAL_CONCURRENCY (${CONCURRENCY_PER_GPU} per GPU × ${DEPLOYMENT_GPU_COUNT} GPUs)"
+          # Write input_config.json
+          cat > "${ROOT_ARTIFACT_DIR}/${EPOCH}_${JOB_NAME}/input_config.json" <<EOF
+          {
+            "gpu_count": $DEPLOYMENT_GPU_COUNT,
+            "concurrency_per_gpu": $CONCURRENCY_PER_GPU,
+            "total_concurrency": $TOTAL_CONCURRENCY,
+            "mode": "$DEPLOYMENT_MODE",
+            "isl": $ISL,
+            "osl": $OSL,
+            "endpoint": "$ENDPOINT",
+            "model endpoint": "$TARGET_MODEL"
+          }
+          EOF
+
+          # Run perf with calculated total concurrency
+          run_perf $TOTAL_CONCURRENCY $ISL $OSL
+          echo "done with concurrency $TOTAL_CONCURRENCY"
+        env:
+        - name: TARGET_MODEL
+          value: Qwen/Qwen3-235B-A22B-FP8
+        - name: ENDPOINT
+          value: qwen3-235b-a22b-disagg-frontend:8000
+        - name: CONCURRENCY_PER_GPU
+          value: "2"
+        - name: DEPLOYMENT_GPU_COUNT
+          value: "16"
+        - name: ISL
+          value: "4000"
+        - name: OSL
+          value: "200"
+        - name: DEPLOYMENT_MODE
+          value: disagg
+        - name: AIPERF_HTTP_CONNECTION_LIMIT
+          value: "200"
+        - name: JOB_NAME
+          valueFrom:
+            fieldRef:
+              apiVersion: v1
+              fieldPath: metadata.labels['job-name']
+        - name: ROOT_ARTIFACT_DIR
+          value: /model-cache/perf
+        - name: HF_HOME
+          value: /model-cache
+        - name: PYTHONUNBUFFERED
+          value: "1"
+        image: python:3.12-slim
+        imagePullPolicy: IfNotPresent
+        name: perf
+        securityContext:
+          privileged: true
+        volumeMounts:
+        - name: model-cache
+          mountPath: /model-cache
+        workingDir: /workspace
+      imagePullSecrets:
+      - name: nvcrimagepullsecret
+      restartPolicy: Never
+      volumes:
+      - name: model-cache
+        persistentVolumeClaim:
+          claimName: model-cache