Adding nccl config and newer version

0e9e2d53 · Nicolas Richart · 72b357ab · 0e9e2d53 · 0e9e2d53 · 0e9e2d53
Commit 0e9e2d53 authored 8 months ago by Nicolas Richart
--- a/stacks/pinot-noir-gcc/definitions_cuda.yaml
+++ b/stacks/pinot-noir-gcc/definitions_cuda.yaml
@@ -6,4 +6,5 @@ definitions:

  - cuda_serial_codes:
      - nccl
+      - nccl@2.20.5-1
      - nvshmem
--- a/stacks/pinot-noir-gcc/modules_kuma_h100.yaml
+++ b/stacks/pinot-noir-gcc/modules_kuma_h100.yaml
@@ -7,3 +7,8 @@ modules:
            OMPI_MCA_pml: 'ucx'
            OMPI_MCA_osc: 'ucx'
            UCX_NET_DEVICES: 'mlx5_2:1,mlx5_3:1'
+    nccl:
+      environment:
+        set:
+          NCCL_IB_HCA: 'mlx5_2:1,mlx5_3:1'
+          NCCL_SOCKET_IFNAME: 'bond0'
--- a/stacks/pinot-noir/modules_kuma_h100.yaml
+++ b/stacks/pinot-noir/modules_kuma_h100.yaml
@@ -7,3 +7,8 @@ modules:
            OMPI_MCA_pml: 'ucx'
            OMPI_MCA_osc: 'ucx'
            UCX_NET_DEVICES: 'mlx5_2:1,mlx5_3:1'
+    nccl:
+      environment:
+        set:
+          NCCL_IB_HCA: 'mlx5_2:1,mlx5_3:1'
+          NCCL_SOCKET_IFNAME: 'bond0'