Handles 11.8

Author: syed-ahmed

common/install_cuda.sh

@@ -73,6 +73,16 @@ function install_118 {
     cd ..
     rm -rf tmp_cudnn
     ldconfig
+
+    # NCCL license: https://docs.nvidia.com/deeplearning/nccl/#licenses
+    mkdir tmp_nccl && cd tmp_nccl
+    wget -q https://developer.download.nvidia.com/compute/redist/nccl/v2.15.5/nccl_2.15.5-1+cuda11.8_x86_64.txz
+    tar xf nccl_2.15.5-1+cuda11.8_x86_64.txz
+    cp -a nccl_2.15.5-1+cuda11.8_x86_64/include/* /usr/local/cuda/include/
+    cp -a nccl_2.15.5-1+cuda11.8_x86_64/lib/* /usr/local/cuda/lib64/
+    cd ..
+    rm -rf tmp_nccl
+    ldconfig
 }
 
 function prune_116 {

manywheel/build_cuda.sh

@@ -142,7 +142,7 @@ DEPS_SONAME=(
     "libcublasLt.so.11"
     "libgomp.so.1"
 )
-elif [[ $CUDA_VERSION == "11.7" ]]; then
+elif [[ $CUDA_VERSION == "11.7" || $CUDA_VERSION == "11.8" ]]; then
     export USE_STATIC_CUDNN=0
     # Try parallelizing nvcc as well
     export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
@@ -181,64 +181,6 @@ elif [[ $CUDA_VERSION == "11.7" ]]; then
             "libcublas.so.11"
             "libcublasLt.so.11"
         )
-    else
-        echo "Using cudnn and cublas from pypi."
-        CUDA_RPATHS=(
-            '$ORIGIN/../../nvidia/cublas/lib'
-            '$ORIGIN/../../nvidia/cudnn/lib'
-        )
-        CUDA_RPATHS=$(IFS=: ; echo "${CUDA_RPATHS[*]}")
-        export C_SO_RPATH=$CUDA_RPATHS':$ORIGIN:$ORIGIN/lib'
-        export LIB_SO_RPATH=$CUDA_RPATHS':$ORIGIN'
-        export FORCE_RPATH="--force-rpath"
-    fi
-elif [[ $CUDA_VERSION == "11.8" ]]; then
-    export USE_STATIC_CUDNN=0
-    # Try parallelizing nvcc as well
-    export TORCH_NVCC_FLAGS="-Xfatbin -compress-all --threads 2"
-    DEPS_LIST=(
-        "/usr/local/cuda/lib64/libcudart.so.11.0"
-        "/usr/local/cuda/lib64/libnvToolsExt.so.1"
-        "/usr/local/cuda/lib64/libnvrtc.so.11.2"    # this is not a mistake for 11.8, it links to 11.8.89
-        "/usr/local/cuda/lib64/libnvrtc-builtins.so.11.8"
-        "$LIBGOMP_PATH"
-    )
-    DEPS_SONAME=(
-        "libcudart.so.11.0"
-        "libnvToolsExt.so.1"
-        "libnvrtc.so.11.2"
-        "libnvrtc-builtins.so.11.8"
-        "libgomp.so.1"
-    )
-
-    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
-        echo "Bundling with cudnn and cublas."
-        DEPS_LIST+=(
-            "/usr/local/cuda/lib64/libcudnn_adv_infer.so.8"
-            "/usr/local/cuda/lib64/libcudnn_adv_train.so.8"
-            "/usr/local/cuda/lib64/libcudnn_cnn_infer.so.8"
-            "/usr/local/cuda/lib64/libcudnn_cnn_train.so.8"
-            "/usr/local/cuda/lib64/libcudnn_ops_infer.so.8"
-            "/usr/local/cuda/lib64/libcudnn_ops_train.so.8"
-            "/usr/local/cuda/lib64/libcudnn.so.8"
-            "/usr/local/cuda/lib64/libcublas.so.11"
-            "/usr/local/cuda/lib64/libcublasLt.so.11"
-        )
-        DEPS_SONAME+=(
-            "libcudnn_adv_infer.so.8"
-            "libcudnn_adv_train.so.8"
-            "libcudnn_cnn_infer.so.8"
-            "libcudnn_cnn_train.so.8"
-            "libcudnn_ops_infer.so.8"
-            "libcudnn_ops_train.so.8"
-            "libcudnn.so.8"
-            "libcublas.so.11"
-            "libcublasLt.so.11"
-            "libnvrtc.so.11.2"
-            "libnvrtc-builtins.so.11.7"
-            "libcudart.so.11.0"
-            "libnvToolsExt.so.1"
-        )
     else
         echo "Using nvidia libs from pypi."
         CUDA_RPATHS=(