Upgrade CI default CUDA version from 12.9 to 13.0 (#21441)

scripts/ci/cuda/ci_download_flashinfer_jit_cache.sh

@@ -5,7 +5,7 @@
 # Required environment (caller must export or set):
 #   UNINSTALL_JIT_CACHE          — literal true/false (skip download when false)
 #   FLASHINFER_PYTHON_REQUIRED   — e.g. from python/pyproject.toml (flashinfer_python)
-#   CU_VERSION                   — e.g. cu129
+#   CU_VERSION                   — e.g. cu130
 #   PIP_CMD                      — e.g. "pip" or "uv pip"
 #   PIP_INSTALL_SUFFIX           — extra pip args for this runner
 set -euxo pipefail

scripts/ci/cuda/ci_install_deepep.sh

@@ -100,13 +100,8 @@ 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
-        # 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'
+        # CUDA > 12.8 supports sm_103 (Blackwell)
+        CHOSEN_TORCH_CUDA_ARCH_LIST='10.0;10.3'
     else
         echo "Unsupported CUDA version for Grace Blackwell: $CUDA_VERSION" && exit 1
     fi && \

scripts/ci/cuda/ci_install_dependency.sh

@@ -23,7 +23,7 @@ set -euxo pipefail
 # Configuration & timing
 # ------------------------------------------------------------------------------
 # Set up environment variables
-CU_VERSION="cu129"
+CU_VERSION="cu130"
 
 # Nvidia package versions we override (torch pins older versions).
 # Used both as pip constraints during install and for post-install verification.
@@ -244,7 +244,15 @@ if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ] && [ -d "sgl-kernel/dist" ]; then
     else
         WHEEL_ARCH="x86_64"
     fi
-    $PIP_CMD install sgl-kernel/dist/sglang_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl --force-reinstall $PIP_INSTALL_SUFFIX
+    # Wheel may have +cuXYZ suffix (e.g. sglang_kernel-0.4.0+cu130-...) depending on CUDA version
+    KERNEL_WHL=$(ls sgl-kernel/dist/sglang_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}*-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl 2>/dev/null | head -1)
+    if [ -z "$KERNEL_WHL" ]; then
+        echo "ERROR: No matching sgl-kernel wheel found in sgl-kernel/dist/ for version ${SGL_KERNEL_VERSION_FROM_KERNEL} arch ${WHEEL_ARCH}"
+        ls -alh sgl-kernel/dist/
+        exit 1
+    fi
+    echo "Installing sgl-kernel wheel: $KERNEL_WHL"
+    $PIP_CMD install "$KERNEL_WHL" --force-reinstall $PIP_INSTALL_SUFFIX
 elif [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ] && [ ! -d "sgl-kernel/dist" ]; then
     # CUSTOM_BUILD_SGL_KERNEL was set but artifacts not available (e.g., stage rerun without wheel build)
     # Fail instead of falling back to PyPI - we need to test the built kernel, not PyPI version
@@ -318,10 +326,10 @@ mark_step_done "Install extra dependency"
 # Fix other dependencies
 # ------------------------------------------------------------------------------
 # Fix CUDA version mismatch between torch and torchaudio.
-# PyPI's torch 2.9.1 bundles cu128 but torchaudio from pytorch.org/cu129 uses cu129.
+# PyPI's torch bundles a specific CUDA version but torchaudio from pytorch.org/cu130 may use a different one.
 # This mismatch causes torchaudio's C extension to fail loading, producing:
 #   "partially initialized module 'torchaudio' has no attribute 'lib'"
-# We cannot replace torch with cu129 (breaks sgl_kernel ABI), so instead we reinstall
+# We cannot replace torch with a different CUDA version (breaks sgl_kernel ABI), so instead we reinstall
 # torchaudio/torchvision from an index matching torch's CUDA version.
 TORCH_CUDA_VER=$(python3 -c "import torch; v=torch.version.cuda; parts=v.split('.'); print(f'cu{parts[0]}{parts[1]}')")
 echo "Detected torch CUDA version: ${TORCH_CUDA_VER}"