prefill.yaml

apiVersion: leaderworkerset.x-k8s.io/v1
kind: LeaderWorkerSet
metadata:
  name: wide-ep-llm-d-prefill
  labels:
    llm-d.ai/inferenceServing: "true"
    llm-d.ai/model: Qwen3-30B-A3B
    llm-d.ai/role: prefill
spec:
  replicas: 1
  leaderWorkerTemplate:
    size: 1
    workerTemplate:
      metadata:
        labels:
          llm-d.ai/inferenceServing: "true"
          llm-d.ai/model: Qwen3-30B-A3B
          llm-d.ai/role: prefill
      spec:
        serviceAccountName: deepseek-r1
        imagePullSecrets:
        - name: rh-ee-ecrncevi-ecrncevi-vllm-pull-pull-secret
        volumes:
          - name: dshm
            emptyDir:
              medium: Memory
              sizeLimit: 2Gi # 2Gi for NCCL 2.27 to be safe
          - name: shared-cuda
            emptyDir: {}
        resourceClaims:
          - name: compute-domain-channel
            resourceClaimTemplateName: llm-d-dev-claim
        containers:
          - name: vllm-worker
            image: quay.io/rh-ee-ecrncevi/llm-d-dev@sha256:032980d0a779f6e32f9606becf66841a0e46501aaf155b5678ff73c17c381992
            securityContext:
              privileged: true
              runAsGroup: 0
              runAsUser: 0
            imagePullPolicy: Always
            command:
              - /bin/bash
              - -c
            args:
              - |-
                #################
                # RUN vLLM prefill worker
                #################
                START_RANK=$(( ${LWS_WORKER_INDEX:-0} * DP_SIZE_LOCAL ))
                source /opt/vllm/bin/activate
                exec vllm serve \
                  Qwen/Qwen3-30B-A3B \
                  --port 8000 \
                  --disable-uvicorn-access-log \
                  --enable-expert-parallel \
                  --data-parallel-hybrid-lb \
                  --data-parallel-size $((LWS_GROUP_SIZE * DP_SIZE_LOCAL)) \
                  --data-parallel-size-local $DP_SIZE_LOCAL \
                  --data-parallel-address ${LWS_LEADER_ADDRESS} \
                  --data-parallel-rpc-port 5555 \
                  --data-parallel-start-rank $START_RANK \
                  --trust-remote-code \
                  --kv_transfer_config '{"kv_connector":"NixlConnector","kv_role":"kv_both"}'
            env:
              - name: DP_SIZE_LOCAL
                value: "4"
              - name: TRITON_LIBCUDA_PATH
                value: /usr/lib64
              - name: VLLM_DEEPEP_HIGH_THROUGHPUT_FORCE_INTRA_NODE
                value: "1"
              - name: VLLM_RANDOMIZE_DP_DUMMY_INPUTS
                value: "1"
              - name: VLLM_USE_DEEP_GEMM
                value: "1"
              - name: VLLM_ALL2ALL_BACKEND
                value: deepep_high_throughput
              - name: NVIDIA_GDRCOPY
                value: enabled
              - name: NVSHMEM_DEBUG
                value: INFO
              - name: VLLM_LOGGING_LEVEL
                value: INFO
              - name: NCCL_DEBUG
                value: INFO
              - name: VLLM_NIXL_SIDE_CHANNEL_HOST
                valueFrom:
                  fieldRef:
                    fieldPath: status.podIP
            ports:
              - containerPort: 8000
                name: metrics
                protocol: TCP
            readinessProbe:
              httpGet:
                path: /health
                port: 8000
            resources:
              limits:
                ephemeral-storage: 75Gi
                memory: 512Gi
                nvidia.com/gpu: "4"
              requests:
                cpu: 32
                ephemeral-storage: 75Gi
                memory: 512Gi
                nvidia.com/gpu: "4"
            volumeMounts:
              - mountPath: /dev/shm
                name: dshm
              - name: shared-cuda
                mountPath: /usr/local/cuda
            workingDir: /code