update EFA version in cu130 base image (#5468)

* update EFA version in base image

* rebuild after efa version update

* renamed the NCCL OFI plugin library

* fix NCCL OFI plugin library issue

* temporarily disable NCCL check

* Enable NCCL check and build

* Enable efa version based NCCL check

* fix typo

* test build for cuda129

* revert back toml file

---------

Co-authored-by: DevakiBolleneni <[email protected]>

Author: DevakiBolleneni

base/x86_64/gpu/cu130/ubuntu22.04/Dockerfile

@@ -3,7 +3,7 @@ ARG PYTHON_VERSION="3.12.10"
 ARG PYTHON_SHORT_VERSION="3.12"
 ARG CUDA_MAJOR="13"
 ARG CUDA_MINOR="0"
-ARG EFA_VERSION="1.43.3"
+ARG EFA_VERSION="1.44.0"
 ARG OS_VERSION="ubuntu22.04"
 FROM nvidia/cuda:13.0.0-base-${OS_VERSION} AS base-builder
 

dlc_developer_config.toml

@@ -36,7 +36,7 @@ deep_canary_mode = false
 
 [build]
 # Add in frameworks you would like to build. By default, builds are disabled unless you specify building an image.
-# available frameworks - ["base", "vllm", "sglang", "autogluon", "huggingface_tensorflow", "huggingface_pytorch", "huggingface_tensorflow_trcomp", "huggingface_pytorch_trcomp", "pytorch_trcomp", "tensorflow", "pytorch", "stabilityai_pytorch"]
+# available frameworks - ["base", "vllm", "autogluon", "huggingface_tensorflow", "huggingface_pytorch", "huggingface_tensorflow_trcomp", "huggingface_pytorch_trcomp", "pytorch_trcomp", "tensorflow", "pytorch", "stabilityai_pytorch"]
 build_frameworks = []
 
 

scripts/install_efa.sh

@@ -17,13 +17,26 @@ case $ARCH in
 esac
 
 function check_libnccl_net_so {
-    OFI_LIB_DIR="/opt/amazon/ofi-nccl/lib/${ARCH_DIR}"
-    NCCL_NET_SO="$OFI_LIB_DIR/libnccl-net.so"
+    
+    if [[ "$EFA_VERSION" > "1.44.0" ]] || [[ "$EFA_VERSION" == "1.44.0" ]]; then  # version threshold
+        # Newer EFA version - no ARCH_DIR, different filename
+        OFI_LIB_DIR="/opt/amazon/ofi-nccl/lib/"
+        NCCL_NET_SO="$OFI_LIB_DIR/libnccl-net-ofi.so"
+        echo "Using newer EFA path structure"
+    else
+        # Older EFA version - uses ARCH_DIR
+        OFI_LIB_DIR="/opt/amazon/ofi-nccl/lib/${ARCH_DIR}"
+        NCCL_NET_SO="$OFI_LIB_DIR/libnccl-net.so"
+        echo "Using older EFA path structure with ARCH_DIR: $ARCH_DIR"
+    fi
 
     # Check if file exists
     if [ ! -f "$NCCL_NET_SO" ]; then
         echo "ERROR: $NCCL_NET_SO does not exist"
         return 1
+    else
+        echo "NCCL OFI plugin found at: $NCCL_NET_SO"
+        return 0
     fi
 }