Revert "Upgrade CI default CUDA version from 12.9 to 13.0" (#22727)

2026-04-13 14:39:24 -07:00
parent ba7bcca6b3
commit b441317aa4
5 changed files with 33 additions and 33 deletions
@@ -100,8 +100,13 @@ if [ "$GRACE_BLACKWELL" = "1" ]; then
    if [ "$CUDA_VERSION" = "12.8" ]; then
        CHOSEN_TORCH_CUDA_ARCH_LIST='10.0'
    elif awk -v ver="$CUDA_VERSION" 'BEGIN {exit !(ver > 12.8)}'; then
-        # CUDA > 12.8 supports sm_103 (Blackwell)
-        CHOSEN_TORCH_CUDA_ARCH_LIST='10.0;10.3'
+        # With cuda > 12.8, the compiler supports 10.3, so we should use
+        # CHOSEN_TORCH_CUDA_ARCH_LIST='10.0;10.3'
+        #
+        # However, our CI machine has a weird setup and nvidia-smi reports wrong CUDA version in the container.
+        # The container is actually cuda 12.8, but nvidia-smi reports 13.0, leading to compilation errors. so we
+        # drop 10.3.
+        CHOSEN_TORCH_CUDA_ARCH_LIST='10.0'
    else
        echo "Unsupported CUDA version for Grace Blackwell: $CUDA_VERSION" && exit 1
    fi && \