[NVIDIA] Add CI workloads for GB200 (#12242)

scripts/ci/ci_install_dependency.sh

@@ -5,6 +5,10 @@ set -euxo pipefail
 IS_BLACKWELL=${IS_BLACKWELL:-0}
 CU_VERSION="cu129"
 
+# Detect system architecture
+ARCH=$(uname -m)
+echo "Detected architecture: ${ARCH}"
+
 if [ "$CU_VERSION" = "cu130" ]; then
     NVRTC_SPEC="nvidia-cuda-nvrtc"
 else
@@ -23,8 +27,20 @@ rm -rf /root/.cache/flashinfer
 # Install apt packages
 apt install -y git libnuma-dev libssl-dev pkg-config
 
+# Check if protoc of correct architecture is already installed
+if command -v protoc >/dev/null 2>&1; then
+    if protoc --version >/dev/null 2>&1; then
+        echo "protoc already installed: $(protoc --version)"
+    else
+        echo "protoc found but not runnable, reinstalling..."
+        INSTALL_PROTOC=1
+    fi
+else
+    INSTALL_PROTOC=1
+fi
+
 # Install protoc for router build (gRPC protobuf compilation)
-if ! command -v protoc &> /dev/null; then
+if [ "${INSTALL_PROTOC:-0}" = "1" ]; then
     echo "Installing protoc..."
     if command -v apt-get &> /dev/null; then
         # Ubuntu/Debian
@@ -37,9 +53,16 @@ if ! command -v protoc &> /dev/null; then
     fi
 
     cd /tmp
-    wget https://github.com/protocolbuffers/protobuf/releases/download/v32.0/protoc-32.0-linux-x86_64.zip
-    unzip protoc-32.0-linux-x86_64.zip -d /usr/local
-    rm protoc-32.0-linux-x86_64.zip
+    # Determine protoc architecture
+    if [ "$ARCH" = "aarch64" ] || [ "$ARCH" = "arm64" ]; then
+        PROTOC_ARCH="aarch_64"
+    else
+        PROTOC_ARCH="x86_64"
+    fi
+    PROTOC_ZIP="protoc-32.0-linux-${PROTOC_ARCH}.zip"
+    wget https://github.com/protocolbuffers/protobuf/releases/download/v32.0/${PROTOC_ZIP}
+    unzip -o ${PROTOC_ZIP} -d /usr/local
+    rm ${PROTOC_ZIP}
     protoc --version
     cd -
 else
@@ -52,6 +75,7 @@ if [ "$IS_BLACKWELL" = "1" ]; then
     # so we can only use pip with `--break-system-packages`
     PIP_CMD="pip"
     PIP_INSTALL_SUFFIX="--break-system-packages"
+    $PIP_CMD install --upgrade pip
 
     # Clean up existing installations
     $PIP_CMD uninstall -y sgl-kernel sglang $PIP_INSTALL_SUFFIX || true
@@ -81,7 +105,13 @@ echo "SGL_KERNEL_VERSION_FROM_KERNEL=${SGL_KERNEL_VERSION_FROM_KERNEL} SGL_KERNE
 
 if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ]; then
     ls -alh sgl-kernel/dist
-    $PIP_CMD install sgl-kernel/dist/sgl_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}-cp310-abi3-manylinux2014_x86_64.whl --force-reinstall $PIP_INSTALL_SUFFIX
+    # Determine wheel architecture
+    if [ "$ARCH" = "aarch64" ] || [ "$ARCH" = "arm64" ]; then
+        WHEEL_ARCH="aarch64"
+    else
+        WHEEL_ARCH="x86_64"
+    fi
+    $PIP_CMD install sgl-kernel/dist/sgl_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl --force-reinstall $PIP_INSTALL_SUFFIX
 else
     $PIP_CMD install sgl-kernel==${SGL_KERNEL_VERSION_FROM_SRT} --force-reinstall $PIP_INSTALL_SUFFIX
 fi

scripts/ci_monitor/ci_analyzer.py

@@ -112,6 +112,9 @@ class SGLangCIAnalyzer:
             "b200": [
                 "unit-test-backend-4-gpu-b200",
             ],
+            "gb200": [
+                "unit-test-backend-4-gpu-gb200",
+            ],
         }
 
         stats = {
@@ -180,6 +183,7 @@ class SGLangCIAnalyzer:
                     "unit-test-deepep-8-gpu",
                     "unit-test-backend-8-gpu-deepseek-v32",
                     "unit-test-backend-4-gpu-b200",
+                    "unit-test-backend-4-gpu-gb200",
                     "vllm-dependency-test",
                     "nightly-test-eval-text-models",
                     "nightly-test-perf-text-models",

scripts/ci_monitor/ci_analyzer_balance.py

@@ -181,6 +181,7 @@ class SGLangTestBalanceAnalyzer:
             "unit-test-backend-8-gpu-h200",
             "unit-test-backend-8-gpu-h20",
             "unit-test-backend-4-gpu-b200",
+            "unit-test-backend-4-gpu-gb200",
             "unit-test-deepep-4-gpu",
             "unit-test-deepep-8-gpu",
             "unit-test-backend-8-gpu-deepseek-v32",