Enable nightly CUDA 11.8 builds (#1206)

* enable nightly builds for CUDA 11.8

* add CUDA 11.8 version to manywheel, remove 11.3 and 11.5

Author: ptrblck

conda/pytorch-nightly/build.sh

@@ -69,6 +69,13 @@ if [[ -n "$build_with_cuda" ]]; then
         export USE_STATIC_CUDNN=0
         #for cuda 11.7 include all dynamic loading libraries
         DEPS_LIST=(/usr/local/cuda/lib64/libcudnn*.so.8 /usr/local/cuda-11.7/extras/CUPTI/lib64/libcupti.so.11.7)
+    elif [[ $CUDA_VERSION == 11.8* ]]; then
+	export TORCH_CUDA_ARCH_LIST="$TORCH_CUDA_ARCH_LIST;6.0;6.1;7.0;7.5;8.0;8.6"
+	#for cuda 11.8 we use cudnn 8.5
+	#which does not have single static libcudnn_static.a deliverable to link with
+	export USE_STATIC_CUDNN=0
+	#for cuda 11.8 include all dynamic loading libraries
+	DEPS_LIST=(/usr/local/cuda/lib64/libcudnn*.so.8 /usr/local/cuda-11.8/extras/CUPTI/lib64/libcupti.so.11.8)
     fi
     export NCCL_ROOT_DIR=/usr/local/cuda
     export USE_STATIC_NCCL=1  # links nccl statically (driven by tools/setup_helpers/nccl.py, some of the NCCL cmake files such as FindNCCL.cmake and gloo/FindNCCL.cmake)

manywheel/build_cuda.sh

@@ -58,7 +58,7 @@ cuda_version_nodot=$(echo $CUDA_VERSION | tr -d '.')
 
 TORCH_CUDA_ARCH_LIST="3.7;5.0;6.0;7.0"
 case ${CUDA_VERSION} in
-    11.[3567])
+    11.[678])
         TORCH_CUDA_ARCH_LIST="${TORCH_CUDA_ARCH_LIST};7.5;8.0;8.6"
         EXTRA_CAFFE2_CMAKE_FLAGS+=("-DATEN_NO_TEST=ON")
         ;;
@@ -196,6 +196,60 @@ elif [[ $CUDA_VERSION == "11.7" ]]; then
         export LIB_SO_RPATH=$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 "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
+        echo "Bundling with cudnn and cublas."
+        DEPS_LIST+=(
+            "/usr/local/cuda/lib64/libcudnn_adv_infer.so.8"
+            "/usr/local/cuda/lib64/libcudnn_adv_train.so.8"
+            "/usr/local/cuda/lib64/libcudnn_cnn_infer.so.8"
+            "/usr/local/cuda/lib64/libcudnn_cnn_train.so.8"
+            "/usr/local/cuda/lib64/libcudnn_ops_infer.so.8"
+            "/usr/local/cuda/lib64/libcudnn_ops_train.so.8"
+            "/usr/local/cuda/lib64/libcudnn.so.8"
+	    "/usr/local/cuda/lib64/libcublas.so.11"
+            "/usr/local/cuda/lib64/libcublasLt.so.11"
+        )
+	DEPS_SONAME+=(
+            "libcudnn_adv_infer.so.8"
+            "libcudnn_adv_train.so.8"
+            "libcudnn_cnn_infer.so.8"
+            "libcudnn_cnn_train.so.8"
+            "libcudnn_ops_infer.so.8"
+            "libcudnn_ops_train.so.8"
+            "libcudnn.so.8"
+            "libcublas.so.11"
+            "libcublasLt.so.11"
+        )
+    else
+        echo "Using cudnn and cublas from pypi."
+        CUDA_RPATHS=(
+            '$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
 else
     echo "Unknown cuda version $CUDA_VERSION"
     exit 1