nccl

nccl_test/simple-test/simple_nccl.py

@@ -0,0 +1,17 @@
+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()

nccl_test/simple-test/template.yaml

@@ -365,8 +365,8 @@ spec:
   template:
     metadata:
       labels: { app: nccl }
-      annotations:
-        k8s.v1.cni.cncf.io/networks: hostdevice-net  # HostDeviceNetwork (RDMA)
+#      annotations:
+#        k8s.v1.cni.cncf.io/networks: hostdevice-net  # HostDeviceNetwork (RDMA)
     spec:
 #      hostNetwork: true
 #      dnsPolicy: ClusterFirstWithHostNet
@@ -392,15 +392,17 @@ spec:
           resources:
             limits:
               nvidia.com/gpu: "8"
-              nvidia.com/hostdev: "8"   # SR-IOV Device Plugin이 광고한 RDMA NIC 리소스
+#              nvidia.com/hostdev: "8"   # SR-IOV Device Plugin이 광고한 RDMA NIC 리소스
             requests:
               nvidia.com/gpu: "8"
-              nvidia.com/hostdev: "8"
+#              nvidia.com/hostdev: "8"
           volumeMounts:
             - name: scripts
               mountPath: /config_scripts
             - name: shared-memory
               mountPath: /dev/shm
+            - name: infiniband
+              mountPath: /dev/infiniband
       volumes:
         - name: scripts
           configMap:
@@ -410,5 +412,9 @@ spec:
           emptyDir:
             medium: Memory
             sizeLimit: 118541097369600m
+        - name: infiniband
+          hostPath:
+            path: /dev/infiniband
+            type: Directory