Support B200 in CI (#8861)

.github/workflows/pr-test.yml

@@ -401,3 +401,25 @@ jobs:
           done
           echo "All jobs completed successfully"
           exit 0
+
+  unit-test-backend-8-gpu-b200:
+    needs: [check-changes, unit-test-frontend, unit-test-backend-2-gpu]
+    if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') &&
+      github.event.pull_request.draft == false &&
+      needs.check-changes.outputs.src == 'true'
+    runs-on: b200-runner
+    strategy:
+      fail-fast: false
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+
+      - name: Install dependencies
+        run: |
+          MODE_BLACKWELL=1 bash scripts/ci_install_dependency.sh
+
+      - name: Run test
+        timeout-minutes: 20
+        run: |
+          cd test/srt
+          python3 run_suite.py --suite per-commit-8-gpu-b200 --auto-partition-id 0 --auto-partition-size 1

scripts/ci_install_dependency.sh

@@ -2,44 +2,63 @@
 # Install the dependency in CI.
 set -euxo pipefail
 
+MODE_BLACKWELL=${MODE_BLACKWELL:-0}
+
+CU_VERSION="cu126"
+if [ "$MODE_BLACKWELL" = "1" ]; then
+    CU_VERSION="cu129"
+fi
+
 # Kill existing processes
 SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
 bash "${SCRIPT_DIR}/killall_sglang.sh"
 
+if ! command -v git >/dev/null 2>&1; then
+    apt update
+    apt install -y git
+fi
+
 # Update pip
-pip install --upgrade pip
+if [ "$MODE_BLACKWELL" != "1" ]; then
+    pip install --upgrade pip --break-system-packages
+fi
 
 # Clean up existing installations
-pip uninstall -y flashinfer flashinfer_python sgl-kernel sglang vllm || true
+pip uninstall -y flashinfer flashinfer_python sgl-kernel sglang vllm --break-system-packages || true
 pip cache purge || true
 rm -rf /root/.cache/flashinfer
+# TODO handle other python versions
 rm -rf /usr/local/lib/python3.10/dist-packages/flashinfer*
 rm -rf /usr/local/lib/python3.10/dist-packages/sgl_kernel*
 
 # Install the main package
-pip install -e "python[dev]" --extra-index-url https://download.pytorch.org/whl/test/cu126
+pip install -e "python[dev]" --extra-index-url https://download.pytorch.org/whl/test/${CU_VERSION} --break-system-packages
 
 # Show current packages
 pip list
 
 # Install additional dependencies
-pip install mooncake-transfer-engine==0.3.5 nvidia-cuda-nvrtc-cu12
+pip install mooncake-transfer-engine==0.3.5 nvidia-cuda-nvrtc-cu12 --break-system-packages
 
-# For lmms_evals evaluating MMMU
+if [ "$MODE_BLACKWELL" != "1" ]; then
-git clone --branch v0.3.3 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
+    # For lmms_evals evaluating MMMU
-pip install -e lmms-eval/
+    git clone --branch v0.3.3 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
+    pip install -e lmms-eval/ --break-system-packages
+fi
 
 # Install FlashMLA for attention backend tests
-# pip install git+https://github.com/deepseek-ai/FlashMLA.git
+# pip install git+https://github.com/deepseek-ai/FlashMLA.git --break-system-packages
 
 # Install hf_xet
-pip install huggingface_hub[hf_xet]
+pip install huggingface_hub[hf_xet] --break-system-packages
 
-# Install xformers
+if [ "$MODE_BLACKWELL" != "1" ]; then
-pip install -U xformers --index-url https://download.pytorch.org/whl/cu126 --no-deps --force-reinstall
+    # Install xformers
+    pip install -U xformers --index-url https://download.pytorch.org/whl/${CU_VERSION} --no-deps --force-reinstall --break-system-packages
+fi
 
 # To help dumping traces when timeout occurred
-pip install py-spy
+pip install py-spy --break-system-packages
 
 # Show current packages
 pip list

test/srt/run_suite.py

@@ -203,6 +203,9 @@ suites = {
     "per-commit-8-gpu-amd": [
         TestFile("test_full_deepseek_v3.py", 250),
     ],
+    "per-commit-8-gpu-b200": [
+        # add more here
+    ],
     "per-commit-cpu": [
         TestFile("cpu/test_activation.py"),
         TestFile("cpu/test_binding.py"),