apiVersion: v1
kind: ConfigMap
metadata:
  name: nccl-scripts
data:
  run.sh: |
    
    #!/usr/bin/env bash
    set -ex

    # 현재 실행 시작 시 타임스탬프 저장
    START_TIME=$(date +%s)

    # ----- RDMA/IB 강제 -----
    export NCCL_DEBUG=INFO
    export NCCL_DEBUG_SUBSYS=INIT,NET,IB
    #export NCCL_NET=IB
    #export NCCL_IB_DISABLE=0
    export NCCL_SOCKET_IFNAME="${NCCL_SOCKET_IFNAME:-net1}"   # HostDeviceNetwork로 붙인 NIC 이름
    export NCCL_IB_HCA="mlx5_0,mlx5_1,mlx5_2,mlx5_8,mlx5_9,mlx5_5,mlx5_6,mlx5_7"
    export NCCL_NET_GDR_LEVEL=2
    # (RoCE 환경에 따라 필요 시) export NCCL_IB_GID_INDEX=3

    # 프로세스별 로그 파일 경로 (노드/프로세스마다 별도 파일)
    export NCCL_DEBUG_FILE="/tmp/nccl-%h-%p.log"

    # ----- 학습 실행 (각 파드가 1개 프로세스씩 구동) -----
    echo "[RUN] RANK=$RANK on $(hostname)"
    torchrun \
      --nnodes=2 \
      --nproc_per_node=8 \
      --rdzv_backend=c10d \
      --node_rank=0 \
      --rdzv_endpoint=nccl-0.nccl:23456 \
      /workspace/train_nccl.py 2>&1 | tee /tmp/nccl-aggregate.log

    # ----- RDMA 사용 여부 집계 -----

    # 새로 생성된 로그 목록 가져오기: START_TIME 이후 생성된 파일만
    new_logs=$(find /tmp -name "nccl-nccl-*.log" -type f -newermt "@${START_TIME}")

    total=0; ok=0
    for f in $new_logs; do
      [[ -f "$f" ]] || continue
      total=$((total+1))
      if rdma_line=$(grep -E "Channel .*GDRDMA" "$f"); then
        echo "[OK] RDMA: $rdma_line"
        ok=$((ok+1))
      fi
    done
    echo "[RDMA-CHECK] used IB on $ok / $total local ranks (processes) in $(hostname)"

    # 모든 랭크가 IB를 썼는지(파드 단위) 판단: 로컬 프로세스 수와 비교
    if [[ "$ok" -eq 8 ]]; then
      echo "[RDMA-CHECK] ✅ RDMA path OK for all local GPUs"
      exit 0
    else
      echo "[RDMA-CHECK] ❌ RDMA path NOT used by all local GPUs"
      echo "---- NCCL NET/IB related lines ----"
      grep -h "NET/" /tmp/nccl-*.log || true
      exit 1
    fi
  train_nccl.py: |
    import os
    import torch
    import torch.distributed as dist

    def main():
        dist.init_process_group("nccl")
        local_rank = int(os.environ["LOCAL_RANK"])
        torch.cuda.set_device(local_rank)

        print(f"Rank {dist.get_rank()} initialized on GPU {local_rank}")

        x = torch.ones(10).cuda()
        dist.all_reduce(x)
        print(f"Rank {dist.get_rank()} result: {x[0].item()}")

    if __name__ == "__main__":
        main()
---
apiVersion: v1
kind: Service
metadata:
  name: nccl
  labels: { app: nccl }
spec:
  clusterIP: None  # Headless
  selector: { app: nccl }
  ports:
    - name: rdv
      port: 23456
---
apiVersion: apps/v1
kind: StatefulSet
metadata:
  name: nccl
spec:
  serviceName: nccl
  replicas: 2                     # >=2 노드에서 스케줄되도록 노드 리소스 준비 필요
  selector:
    matchLabels: { app: nccl }
  template:
    metadata:
      labels: { app: nccl }
      annotations:
        k8s.v1.cni.cncf.io/networks: hostdevice-net  # HostDeviceNetwork (RDMA)
    spec:
#      hostNetwork: true
#      dnsPolicy: ClusterFirstWithHostNet
      # RDMA 장치나 /dev/infiniband 노출이 안 보이면 네트워크 오퍼레이터의 RDMA 플러그인 구성을 확인하세요.
      # (환경에 따라 RDMA Shared Device Plugin이 필요할 수 있음) 
      containers:
        - name: worker
          image: nvcr.io/nvidia/pytorch:24.10-py3
          imagePullPolicy: IfNotPresent
          securityContext:
            privileged: true
            capabilities:
              add:
                - IPC_LOCK
                - NET_ADMIN
#            capabilities: { add: ["IPC_LOCK"] }
          command: [ "bash", "-lc", "cp /config_scripts/* /workspace/ && chmod +x /workspace/run.sh && sleep infinity" ]
          env:
            # --- (선택) 특정 GPU만 사용하려면 아래 값을 물리 인덱스로 설정 (예: "2") ---
            # - name: CUDA_VISIBLE_DEVICES_OVERRIDE
            #   value: "2"
            # RDMA 관련 기본값은 run.sh에서 설정
          resources:
            limits:
              nvidia.com/gpu: "8"
              nvidia.com/hostdev: "8"   # SR-IOV Device Plugin이 광고한 RDMA NIC 리소스
            requests:
              nvidia.com/gpu: "8"
              nvidia.com/hostdev: "8"
          volumeMounts:
            - name: scripts
              mountPath: /config_scripts
            - name: shared-memory
              mountPath: /dev/shm
      volumes:
        - name: scripts
          configMap:
            name: nccl-scripts
            defaultMode: 0755
        - name: shared-memory
          emptyDir:
            medium: Memory
            sizeLimit: 118541097369600m