fix(nccl-tests): upgrade to cuda-nvcc-12-8, add sm_100 (Blackwell B100/B200)

iso/builder/build-nccl-tests.sh

@@ -34,9 +34,9 @@ if [ -f "${CACHE_DIR}/bin/all_reduce_perf" ]; then
     exit 0
 fi
 
-# Resolve nvcc path (cuda-nvcc-12-6 installs to /usr/local/cuda-12.6/bin/nvcc)
+# Resolve nvcc path (cuda-nvcc-12-8 installs to /usr/local/cuda-12.8/bin/nvcc)
 NVCC=""
-for candidate in nvcc /usr/local/cuda-12.6/bin/nvcc /usr/local/cuda-12/bin/nvcc /usr/local/cuda/bin/nvcc; do
+for candidate in nvcc /usr/local/cuda-12.8/bin/nvcc /usr/local/cuda-12/bin/nvcc /usr/local/cuda/bin/nvcc; do
     if command -v "$candidate" >/dev/null 2>&1 || [ -x "$candidate" ]; then
         NVCC="$candidate"
         break
@@ -96,11 +96,12 @@ SRC_DIR=$(ls -d nccl-tests-* 2>/dev/null | head -1)
 cd "$SRC_DIR"
 
 echo "=== building all_reduce_perf ==="
-# CUDA 12.6 supports Volta through Hopper (sm_70..sm_90).
+# CUDA 12.8 supports Volta through Blackwell (sm_70..sm_100).
 GENCODE="-gencode=arch=compute_70,code=sm_70 \
          -gencode=arch=compute_80,code=sm_80 \
          -gencode=arch=compute_86,code=sm_86 \
-         -gencode=arch=compute_90,code=sm_90"
+         -gencode=arch=compute_90,code=sm_90 \
+         -gencode=arch=compute_100,code=sm_100"
 make MPI=0 \
     NVCC="$NVCC" \
     CUDA_HOME="$CUDA_HOME" \