diff --git a/.github/actions/check-maintenance/action.yml b/.github/actions/check-maintenance/action.yml
index 595283dcd..f064cad52 100644
--- a/.github/actions/check-maintenance/action.yml
+++ b/.github/actions/check-maintenance/action.yml
@@ -1,5 +1,5 @@
 name: Check Maintenance Mode
-description: Blocks CI when maintenance mode is active (issue #21065 is open), unless the PR has the bypass-maintenance label, or env SGLANG_PR_TEST_BYPASS_MAINTENANCE_ON_MAIN=true (PR Test workflow on main only). Merging non-CI-fix PRs is prohibited during maintenance mode; in severe cases, merge permissions may be revoked.
+description: Blocks CI when maintenance mode is active (issue #21065 is open), unless the PR has the bypass-maintenance label, or env PR_TEST_BYPASS_MAINTENANCE_ON_MAIN=true (PR Test workflow on main only). Merging non-CI-fix PRs is prohibited during maintenance mode; in severe cases, merge permissions may be revoked.
 
 inputs:
   github-token:
@@ -20,7 +20,7 @@ runs:
         PR_NUMBER="${{ github.event.pull_request.number }}"
 
         # PR Test workflow only: scheduled runs and runs on main (dispatch / workflow_call) set this env
-        if [[ "${SGLANG_PR_TEST_BYPASS_MAINTENANCE_ON_MAIN:-}" == "true" ]]; then
+        if [[ "${PR_TEST_BYPASS_MAINTENANCE_ON_MAIN:-}" == "true" ]]; then
           echo "✅ PR Test on main branch; bypassing maintenance gate."
           exit 0
         fi
diff --git a/.github/workflows/nightly-test-nvidia.yml b/.github/workflows/nightly-test-nvidia.yml
index d5ce7930b..0f4c754ab 100644
--- a/.github/workflows/nightly-test-nvidia.yml
+++ b/.github/workflows/nightly-test-nvidia.yml
@@ -91,7 +91,7 @@ jobs:
       # Match pr-test-jit-kernel workflow for consistent JIT warmup behavior
       SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
       # Allow maintenance bypass on default branch (same semantics as PR JIT workflow)
-      SGLANG_PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
+      PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
@@ -121,7 +121,7 @@ jobs:
     env:
       SGLANG_JIT_KERNEL_RUN_FULL_TESTS: "1"
       SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
-      SGLANG_PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
+      PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
     steps:
       - name: Checkout code
         uses: actions/checkout@v4
diff --git a/.github/workflows/pr-test-jit-kernel.yml b/.github/workflows/pr-test-jit-kernel.yml
index 772e69392..90b232b07 100644
--- a/.github/workflows/pr-test-jit-kernel.yml
+++ b/.github/workflows/pr-test-jit-kernel.yml
@@ -36,7 +36,7 @@ env:
   SGLANG_IS_IN_CI: true
   SGLANG_CUDA_COREDUMP: "1"
   SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
-  SGLANG_PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
+  PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
   SKIP_STAGE_HEALTH_CHECK: ${{ inputs.skip_stage_health_check == true && 'true' || 'false' }}
 
 jobs:
diff --git a/.github/workflows/pr-test-multimodal-gen.yml b/.github/workflows/pr-test-multimodal-gen.yml
index 742e43757..110c812c0 100644
--- a/.github/workflows/pr-test-multimodal-gen.yml
+++ b/.github/workflows/pr-test-multimodal-gen.yml
@@ -46,7 +46,7 @@ on:
 env:
   SGLANG_IS_IN_CI: true
   SGLANG_CUDA_COREDUMP: "1"
-  SGLANG_PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
+  PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
   SKIP_STAGE_HEALTH_CHECK: ${{ inputs.skip_stage_health_check == 'true' }}
 
 jobs:
diff --git a/.github/workflows/pr-test-sgl-kernel.yml b/.github/workflows/pr-test-sgl-kernel.yml
index 0cfaf43d4..43d490a31 100644
--- a/.github/workflows/pr-test-sgl-kernel.yml
+++ b/.github/workflows/pr-test-sgl-kernel.yml
@@ -27,7 +27,7 @@ on:
 env:
   SGLANG_IS_IN_CI: true
   SGLANG_CUDA_COREDUMP: "1"
-  SGLANG_PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
+  PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
   SKIP_STAGE_HEALTH_CHECK: ${{ inputs.skip_stage_health_check == true && 'true' || 'false' }}
 
 jobs:
diff --git a/.github/workflows/pr-test.yml b/.github/workflows/pr-test.yml
index 28f310926..93ff2b38e 100644
--- a/.github/workflows/pr-test.yml
+++ b/.github/workflows/pr-test.yml
@@ -69,7 +69,7 @@ env:
   SGLANG_JIT_DEEPGEMM_FAST_WARMUP: true
   SKIP_STAGE_HEALTH_CHECK: ${{ inputs.skip_stage_health_check == true && 'true' || 'false' }}
   # Schedule / main-branch dispatch / workflow_call from main use refs/heads/main; PR events use refs/pull/*/merge
-  SGLANG_PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
+  PR_TEST_BYPASS_MAINTENANCE_ON_MAIN: ${{ github.ref == 'refs/heads/main' && 'true' || 'false' }}
   USE_VENV: false
 
 permissions:
@@ -337,20 +337,20 @@ jobs:
           {
             echo "## Change Detection"
             echo ""
-            echo "| Component         | Changed |"
-            echo "|-------------------|---------|"
-            echo "| main_package      | ${{ steps.filter-api.outputs.main_package || steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }} |"
-            echo "| sgl_kernel (raw)  | ${{ steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel }} |"
-            echo "| sgl_kernel (used) | ${{ (!inputs.target_stage || inputs.include_wheel_build) && (steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel) }} |"
-            echo "| jit_kernel        | ${{ steps.filter-api.outputs.jit_kernel || steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }} |"
-            echo "| multimodal_gen    | ${{ steps.filter-api.outputs.multimodal_gen || steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }} |"
-            echo "| target_stage      | ${{ inputs.target_stage || '(none)' }} |"
-            echo "| detection_method  | ${{ inputs.target_stage && 'GitHub API' || 'dorny/paths-filter' }} |"
-            echo "| max_parallel      | ${{ steps.set-parallel.outputs.parallel_level }} (h100=${{ steps.set-parallel.outputs.max_parallel }}, 5090=${{ steps.set-parallel.outputs.max_parallel_small }}, 2gpu=${{ steps.set-parallel.outputs.max_parallel_2gpu }}) |"
-            echo "| b200_runner       | ${{ steps.set-runner.outputs.b200_runner }} |"
+            echo "| Component            | Changed |"
+            echo "|----------------------|---------|"
+            echo "| main_package         | ${{ steps.filter-api.outputs.main_package || steps.filter.outputs.main_package || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| sgl_kernel (raw)     | ${{ steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel }} |"
+            echo "| sgl_kernel (used)    | ${{ (!inputs.target_stage || inputs.include_wheel_build) && (steps.filter-api.outputs.sgl_kernel || steps.filter.outputs.sgl_kernel) }} |"
+            echo "| jit_kernel           | ${{ steps.filter-api.outputs.jit_kernel || steps.filter.outputs.jit_kernel || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| multimodal_gen       | ${{ steps.filter-api.outputs.multimodal_gen || steps.filter.outputs.multimodal_gen || steps.run-mode.outputs.run_all_tests }} |"
+            echo "| target_stage         | ${{ inputs.target_stage || '(none)' }} |"
+            echo "| detection_method     | ${{ inputs.target_stage && 'GitHub API' || 'dorny/paths-filter' }} |"
+            echo "| max_parallel         | ${{ steps.set-parallel.outputs.parallel_level }} (h100=${{ steps.set-parallel.outputs.max_parallel }}, 5090=${{ steps.set-parallel.outputs.max_parallel_small }}, 2gpu=${{ steps.set-parallel.outputs.max_parallel_2gpu }}) |"
+            echo "| b200_runner          | ${{ steps.set-runner.outputs.b200_runner }} |"
             echo "| b200_low_disk_runner | ${{ steps.set-runner.outputs.b200_low_disk_runner }} |"
-            echo "| enable_retry      | ${{ steps.set-retry.outputs.enable_retry }} |"
-            echo "| continue_on_error | ${{ steps.set-continue-on-error.outputs.continue_on_error }} |"
+            echo "| enable_retry         | ${{ steps.set-retry.outputs.enable_retry }} |"
+            echo "| continue_on_error    | ${{ steps.set-continue-on-error.outputs.continue_on_error }} |"
           } >> $GITHUB_STEP_SUMMARY
 
   # =============================================== Wait Jobs for Sequential PR Execution ====================================================
@@ -757,7 +757,6 @@ jobs:
       - name: Install dependencies
         timeout-minutes: 20
         run: |
-          source /etc/profile.d/sglang-ci.sh
           CUSTOM_BUILD_SGL_KERNEL=${{needs.check-changes.outputs.sgl_kernel}} bash scripts/ci/cuda/ci_install_dependency.sh
 
       - name: Run test
@@ -765,7 +764,6 @@ jobs:
         env:
           CONTINUE_ON_ERROR_FLAG: ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
         run: |
-          source /etc/profile.d/sglang-ci.sh
           cd test/
           python3 run_suite.py --hw cuda --suite stage-b-test-1-gpu-small --auto-partition-id ${{ matrix.partition }} --auto-partition-size 8 $CONTINUE_ON_ERROR_FLAG
 
diff --git a/scripts/ci/cuda/ci_install_dependency.sh b/scripts/ci/cuda/ci_install_dependency.sh
index 1a8294aae..bd0f6b650 100755
--- a/scripts/ci/cuda/ci_install_dependency.sh
+++ b/scripts/ci/cuda/ci_install_dependency.sh
@@ -1,93 +1,16 @@
 #!/bin/bash
-# Install the dependency in CI.
+# Install dependencies for CUDA CI jobs.
 #
-# Structure (see section banners below):
-# - Configuration & timing
-# - Host / runner detection (arch, Blackwell, pip vs uv)
-# - Kill existing processes
-# - Install apt packages
-# - Python package site hygiene & install protoc
-# - Pip / uv toolchain & stale package cleanup
-# - Uninstall Flashinfer
-# - Install main package
-# - Install sglang-kernel
-# - Install sglang-router
-# - Download flashinfer artifacts
-# - Install extra dependency
-# - Fix other dependencies
-# - Prepare runner
-# - Verify imports
+# CU_VERSION (default: cu130) controls PyTorch index URL, FlashInfer JIT cache
+# index, and nvrtc variant selection.
 set -euxo pipefail
 
-# ------------------------------------------------------------------------------
-# Configuration & timing
-# ------------------------------------------------------------------------------
-# Set up environment variables
-#
-# CU_VERSION controls:
-#   - PyTorch index URL (pytorch.org/whl/${CU_VERSION})
-#   - FlashInfer JIT cache index (flashinfer.ai/whl/${CU_VERSION})
-#   - nvrtc variant selection (cu12 vs cu13)
-
-CU_VERSION="${CU_VERSION:-cu130}"
-CU_STRIP="${CU_VERSION#cu}"
-CU_MAJOR="${CU_STRIP:0:2}"
-
-# Nvidia package versions we override (torch pins older versions).
-# Used both as pip constraints during install and for post-install verification.
-NVIDIA_CUDNN_VERSION="9.16.0.29"
-NVIDIA_NVSHMEM_VERSION="3.4.5"
-OPTIONAL_DEPS="${1:-}"
-
-# Whether to create a uv venv. Default false; set USE_VENV=false to install
-# directly into system Python (useful for runners where uv venv misbehaves).
-USE_VENV="${USE_VENV:-0}"
-echo "USE_VENV=${USE_VENV}"
-
-# uv must be available on system Python (to create the venv, or to run
-# `uv pip install --system` when venv mode is disabled). Install if missing.
-python3 -m pip install --upgrade pip
-if ! command -v uv >/dev/null 2>&1; then
-    pip install uv
-fi
-
-SYS_PYTHON_VER=$(python3 -c "import sys; print(f'{sys.version_info.major}.{sys.version_info.minor}')")
-
-if [ "$USE_VENV" = "1" ]; then
-    # Per-job unique path. Include $$ (shell PID) so concurrent/back-to-back jobs
-    # on the same runner never target the same directory even if GITHUB_JOB
-    # doesn't differentiate matrix partitions.
-    UV_VENV="/tmp/sglang-ci-${GITHUB_RUN_ID:-norun}-${GITHUB_JOB:-nojob}-$$"
-    # --seed installs pip/setuptools into the venv so bare `pip` calls in
-    # cache_nvidia_wheels.sh and the human-eval setup resolve to the venv's
-    # pip (rather than silently falling back to system Python).
-    uv venv "$UV_VENV" --python "python${SYS_PYTHON_VER}" --seed
-    # shellcheck disable=SC1091
-    source "$UV_VENV/bin/activate"
-    # Assert activation actually took effect. A misconfigured activate script
-    # would otherwise leave us silently running against system Python.
-    [ "${VIRTUAL_ENV:-}" = "$UV_VENV" ] || { echo "FATAL: venv activation did not set VIRTUAL_ENV correctly"; exit 1; }
-    [ "$(command -v python3)" = "$UV_VENV/bin/python3" ] || { echo "FATAL: python3 still resolves outside venv (got $(command -v python3))"; exit 1; }
-
-    # Propagate to subsequent workflow steps. GITHUB_ENV/GITHUB_PATH only
-    # affect *later* steps, never the current one.
-    if [ -n "${GITHUB_ENV:-}" ]; then
-        echo "VIRTUAL_ENV=$UV_VENV" >> "$GITHUB_ENV"
-        echo "SGLANG_CI_VENV_PATH=$UV_VENV" >> "$GITHUB_ENV"
-        # Set BASH_ENV early so subsequent steps auto-source the venv's env script.
-        # LD_LIBRARY_PATH is written to this file later (after packages are installed)
-        # and gets picked up even if GITHUB_ENV becomes unavailable at that point.
-        echo "BASH_ENV=$UV_VENV/env.sh" >> "$GITHUB_ENV"
-        touch "$UV_VENV/env.sh"
-    fi
-    if [ -n "${GITHUB_PATH:-}" ]; then
-        echo "$UV_VENV/bin" >> "$GITHUB_PATH"
-    fi
-else
-    echo "USE_VENV=0: skipping uv venv creation, installing into system Python"
-    UV_VENV=""
-fi
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+REPO_ROOT="$(cd "${SCRIPT_DIR}/../../.." && pwd)"
 
+# ---------------------------------------------------------------------------
+# Timing helper
+# ---------------------------------------------------------------------------
 SECONDS=0
 _CI_MARK_PREV=${SECONDS}
 
@@ -100,323 +23,313 @@ mark_step_done() {
     _CI_MARK_PREV=${now}
 }
 
-mark_step_done "Configuration"
+# ---------------------------------------------------------------------------
+# Functions
+# ---------------------------------------------------------------------------
 
-# ------------------------------------------------------------------------------
-# Host / runner detection (CPU arch, Blackwell, USE_UV)
-# ------------------------------------------------------------------------------
-# Detect CPU architecture (x86_64 or aarch64)
-ARCH=$(uname -m)
-echo "Detected architecture: ${ARCH}"
+configure_environment() {
+    # CU_VERSION controls PyTorch index URL, FlashInfer JIT cache index, and
+    # nvrtc variant selection (cu12 vs cu13).
+    CU_VERSION="${CU_VERSION:-cu130}"
+    CU_STRIP="${CU_VERSION#cu}"
+    CU_MAJOR="${CU_STRIP:0:2}"
 
-# Detect GPU architecture (blackwell or not)
-if [ "${IS_BLACKWELL+set}" = set ]; then
-    case "$IS_BLACKWELL" in 1 | true | yes) IS_BLACKWELL=1 ;; *) IS_BLACKWELL=0 ;; esac
-    echo "IS_BLACKWELL=${IS_BLACKWELL} (manually set via environment)"
-else
-    IS_BLACKWELL=0
-    if command -v nvidia-smi >/dev/null 2>&1; then
-        while IFS= read -r cap; do
-            major="${cap%%.*}"
-            if [ "${major:-0}" -ge 10 ] 2>/dev/null; then
-                IS_BLACKWELL=1
-                break
-            fi
-        done <<< "$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)"
+    # Nvidia package versions we pin (torch ships older versions).
+    NVIDIA_CUDNN_VERSION="9.16.0.29"
+    NVIDIA_NVSHMEM_VERSION="3.4.5"
+    OPTIONAL_DEPS="${1:-}"
+
+    # Whether to create a uv venv (set USE_VENV=1). Default: 0.
+    USE_VENV="${USE_VENV:-0}"
+    echo "USE_VENV=${USE_VENV}"
+
+    python3 -m pip install --upgrade pip
+    if ! command -v uv >/dev/null 2>&1; then
+        pip install uv
     fi
-    echo "IS_BLACKWELL=${IS_BLACKWELL} (auto-detected via nvidia-smi)"
-fi
 
-# Whether to use pip or uv to install dependencies
-if [ "${USE_UV+set}" != set ]; then
-    if [ "$IS_BLACKWELL" = "1" ]; then
-        # Our current b200 runners have some issues with uv, so we default to pip
-        # It is a runner specific issue, not a GPU architecture issue.
-        USE_UV=false
-    else
-        USE_UV=true
-    fi
-fi
-case "$(printf '%s' "$USE_UV" | tr '[:upper:]' '[:lower:]')" in 1 | true | yes) USE_UV=1 ;; *) USE_UV=0 ;; esac
-echo "USE_UV=${USE_UV}"
+    SYS_PYTHON_VER=$(python3 -c "import sys; print(f'{sys.version_info.major}.{sys.version_info.minor}')")
 
-mark_step_done "Host / runner detection"
+    if [ "$USE_VENV" = "1" ]; then
+        UV_VENV="/tmp/sglang-ci-${GITHUB_RUN_ID:-norun}-${GITHUB_JOB:-nojob}-$$"
+        uv venv "$UV_VENV" --python "python${SYS_PYTHON_VER}" --seed
+        # shellcheck disable=SC1091
+        source "$UV_VENV/bin/activate"
+        [ "${VIRTUAL_ENV:-}" = "$UV_VENV" ] || { echo "FATAL: venv activation did not set VIRTUAL_ENV correctly"; exit 1; }
+        [ "$(command -v python3)" = "$UV_VENV/bin/python3" ] || { echo "FATAL: python3 still resolves outside venv (got $(command -v python3))"; exit 1; }
 
-# ------------------------------------------------------------------------------
-# Kill existing processes
-# ------------------------------------------------------------------------------
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-REPO_ROOT="$(cd "${SCRIPT_DIR}/../../.." && pwd)"
-python3 "${REPO_ROOT}/python/sglang/cli/killall.py"
-KILLALL_EXIT=$?
-echo "CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-}"
-
-if [ $KILLALL_EXIT -ne 0 ]; then
-    echo "ERROR: killall.py detected uncleanable GPU memory. Aborting CI."
-    exit 1
-fi
-
-mark_step_done "Kill existing processes"
-
-# ------------------------------------------------------------------------------
-# Install apt packages
-# ------------------------------------------------------------------------------
-# Install apt packages (including python3/pip which may be missing on some runners)
-# Use --no-install-recommends and ignore errors from unrelated broken packages on the runner
-# The NVIDIA driver packages may have broken dependencies that are unrelated to these packages
-# Run apt-get update first to refresh package index (stale index causes 404 on security.ubuntu.com)
-apt-get update || true
-CI_APT_PACKAGES=(
-    python3 python3-pip python3-venv python3-dev git libnuma-dev libssl-dev pkg-config
-    libibverbs-dev libibverbs1 ibverbs-providers ibverbs-utils
-    ffmpeg libavcodec-dev libavformat-dev libavutil-dev libswscale-dev
-)
-apt-get install -y --no-install-recommends "${CI_APT_PACKAGES[@]}" || {
-    echo "Warning: apt-get install failed, checking if required packages are available..."
-    for pkg in "${CI_APT_PACKAGES[@]}"; do
-        if ! dpkg -l "$pkg" 2>/dev/null | grep -q "^ii"; then
-            echo "ERROR: Required package $pkg is not installed and apt-get failed"
-            exit 1
+        if [ -n "${GITHUB_ENV:-}" ]; then
+            echo "VIRTUAL_ENV=$UV_VENV" >> "$GITHUB_ENV"
+            echo "SGLANG_CI_VENV_PATH=$UV_VENV" >> "$GITHUB_ENV"
+            echo "BASH_ENV=$UV_VENV/env.sh" >> "$GITHUB_ENV"
+            touch "$UV_VENV/env.sh"
         fi
-    done
-    echo "All required packages are already installed, continuing..."
+        if [ -n "${GITHUB_PATH:-}" ]; then
+            echo "$UV_VENV/bin" >> "$GITHUB_PATH"
+        fi
+    else
+        echo "USE_VENV=0: skipping uv venv creation, installing into system Python"
+        UV_VENV=""
+    fi
+
+    mark_step_done "${FUNCNAME[0]}"
 }
 
-mark_step_done "Install apt packages"
+detect_host() {
+    ARCH=$(uname -m)
+    echo "Detected architecture: ${ARCH}"
 
-# ------------------------------------------------------------------------------
-# Python package site hygiene & install protoc
-# ------------------------------------------------------------------------------
-# Clear torch compilation cache
-python3 -c 'import os, shutil, tempfile, getpass; cache_dir = os.environ.get("TORCHINDUCTOR_CACHE_DIR") or os.path.join(tempfile.gettempdir(), "torchinductor_" + getpass.getuser()); shutil.rmtree(cache_dir, ignore_errors=True)'
-
-# Remove broken dist-info directories (missing METADATA per PEP 376)
-SITE_PACKAGES=$(python3 -c "import site; print(site.getsitepackages()[0])")
-if [ -d "$SITE_PACKAGES" ]; then
-    { set +x; } 2>/dev/null
-    find "$SITE_PACKAGES" -maxdepth 1 -name "*.dist-info" -type d | while read -r d; do
-        if [ ! -f "$d/METADATA" ]; then
-            echo "Removing broken dist-info: $d"
-            rm -rf "$d"
+    if [ "${IS_BLACKWELL+set}" = set ]; then
+        case "$IS_BLACKWELL" in 1 | true | yes) IS_BLACKWELL=1 ;; *) IS_BLACKWELL=0 ;; esac
+        echo "IS_BLACKWELL=${IS_BLACKWELL} (manually set via environment)"
+    else
+        IS_BLACKWELL=0
+        if command -v nvidia-smi >/dev/null 2>&1; then
+            while IFS= read -r cap; do
+                major="${cap%%.*}"
+                if [ "${major:-0}" -ge 10 ] 2>/dev/null; then
+                    IS_BLACKWELL=1
+                    break
+                fi
+            done <<< "$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)"
         fi
-    done
-    set -x
-fi
+        echo "IS_BLACKWELL=${IS_BLACKWELL} (auto-detected via nvidia-smi)"
+    fi
 
-# Install protoc
-bash "${SCRIPT_DIR}/../utils/install_protoc.sh"
+    if [ "${USE_UV+set}" != set ]; then
+        if [ "$IS_BLACKWELL" = "1" ]; then
+            USE_UV=false
+        else
+            USE_UV=true
+        fi
+    fi
+    case "$(printf '%s' "$USE_UV" | tr '[:upper:]' '[:lower:]')" in 1 | true | yes) USE_UV=1 ;; *) USE_UV=0 ;; esac
+    echo "USE_UV=${USE_UV}"
 
-# Install Rust toolchain (needed by crates built via setuptools-rust, e.g. the
-# native gRPC extension bundled into the sglang wheel).
-bash "${SCRIPT_DIR}/../utils/install_rustup.sh"
-export PATH="${CARGO_HOME:-$HOME/.cargo}/bin:${PATH}"
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-mark_step_done "Python package site hygiene & install protoc + rust"
+kill_existing_processes() {
+    python3 "${REPO_ROOT}/python/sglang/cli/killall.py"
+    KILLALL_EXIT=$?
+    echo "CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-}"
 
-# ------------------------------------------------------------------------------
-# Pip / uv toolchain & stale package cleanup
-# ------------------------------------------------------------------------------
-# Install pip and uv (use python3 -m pip for robustness since some runners only have pip3).
-# In venv mode this upgrades the venv's pip (the bootstrap block near the top
-# already upgraded system pip before `uv venv`).
-python3 -m pip install --upgrade pip
+    if [ $KILLALL_EXIT -ne 0 ]; then
+        echo "ERROR: killall.py detected uncleanable GPU memory. Aborting CI."
+        exit 1
+    fi
 
-# uv is already installed on system Python (above).
-# - Venv mode: the venv is active and `uv pip` targets it automatically.
-# - Non-venv mode: UV_SYSTEM_PYTHON=1 makes `uv pip` operate on system Python
-#   (otherwise uv refuses to run outside a venv).
-if [ "$USE_VENV" != "1" ]; then
-    export UV_SYSTEM_PYTHON=1
-fi
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-export UV_LINK_MODE=copy
-PIP_CMD="uv pip"
-PIP_INSTALL_SUFFIX="--index-strategy unsafe-best-match --prerelease allow"
-PIP_UNINSTALL_CMD="uv pip uninstall"
-PIP_UNINSTALL_SUFFIX=""
+install_apt_packages() {
+    apt-get update || true
+    CI_APT_PACKAGES=(
+        python3 python3-pip python3-venv python3-dev git libnuma-dev libssl-dev pkg-config
+        libibverbs-dev libibverbs1 ibverbs-providers ibverbs-utils
+        ffmpeg libavcodec-dev libavformat-dev libavutil-dev libswscale-dev
+    )
+    apt-get install -y --no-install-recommends "${CI_APT_PACKAGES[@]}" || {
+        echo "Warning: apt-get install failed, checking if required packages are available..."
+        for pkg in "${CI_APT_PACKAGES[@]}"; do
+            if ! dpkg -l "$pkg" 2>/dev/null | grep -q "^ii"; then
+                echo "ERROR: Required package $pkg is not installed and apt-get failed"
+                exit 1
+            fi
+        done
+        echo "All required packages are already installed, continuing..."
+    }
 
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-# Clean up existing installations
-$PIP_UNINSTALL_CMD sgl-kernel sglang-kernel sglang sgl-fa4 flash-attn-4 $PIP_UNINSTALL_SUFFIX || true
+clean_site_packages() {
+    # Clear torch compilation cache
+    python3 -c 'import os, shutil, tempfile, getpass; cache_dir = os.environ.get("TORCHINDUCTOR_CACHE_DIR") or os.path.join(tempfile.gettempdir(), "torchinductor_" + getpass.getuser()); shutil.rmtree(cache_dir, ignore_errors=True)'
 
-mark_step_done "Pip / uv toolchain & stale package cleanup"
+    # Remove broken dist-info directories (missing METADATA per PEP 376)
+    SITE_PACKAGES=$(python3 -c "import site; print(site.getsitepackages()[0])")
+    if [ -d "$SITE_PACKAGES" ]; then
+        { set +x; } 2>/dev/null
+        find "$SITE_PACKAGES" -maxdepth 1 -name "*.dist-info" -type d | while read -r d; do
+            if [ ! -f "$d/METADATA" ]; then
+                echo "Removing broken dist-info: $d"
+                rm -rf "$d"
+            fi
+        done
+        set -x
+    fi
 
-# ------------------------------------------------------------------------------
-# Uninstall Flashinfer
-# ------------------------------------------------------------------------------
-# Keep flashinfer packages installed if version matches to avoid re-downloading:
-# - flashinfer-cubin: 150+ MB
-# - flashinfer-jit-cache: 1.2+ GB, by far the largest download in CI
-FLASHINFER_PYTHON_REQUIRED=$(grep -Po -m1 '(?<=flashinfer_python==)[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
-FLASHINFER_CUBIN_REQUIRED=$(grep -Po -m1 '(?<=flashinfer_cubin==)[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
-FLASHINFER_CUBIN_INSTALLED=$(pip show flashinfer-cubin 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
-FLASHINFER_JIT_INSTALLED=$(pip show flashinfer-jit-cache 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//' || echo "")
-FLASHINFER_JIT_CU_VERSION=$(pip show flashinfer-jit-cache 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed -n 's/.*+//p' || echo "")
+    # Install protoc
+    bash "${SCRIPT_DIR}/../utils/install_protoc.sh"
 
-UNINSTALL_CUBIN=true
-UNINSTALL_JIT_CACHE=true
+    # Install Rust toolchain (needed by setuptools-rust, e.g. the native gRPC extension)
+    bash "${SCRIPT_DIR}/../utils/install_rustup.sh"
+    export PATH="${CARGO_HOME:-$HOME/.cargo}/bin:${PATH}"
 
-if [ "$FLASHINFER_CUBIN_INSTALLED" = "$FLASHINFER_CUBIN_REQUIRED" ] && [ -n "$FLASHINFER_CUBIN_REQUIRED" ]; then
-    echo "flashinfer-cubin==${FLASHINFER_CUBIN_REQUIRED} already installed, keeping it"
-    UNINSTALL_CUBIN=false
-else
-    echo "flashinfer-cubin version mismatch (installed: ${FLASHINFER_CUBIN_INSTALLED:-none}, required: ${FLASHINFER_CUBIN_REQUIRED}), reinstalling"
-fi
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-if [ "$FLASHINFER_JIT_INSTALLED" = "$FLASHINFER_PYTHON_REQUIRED" ] && [ -n "$FLASHINFER_PYTHON_REQUIRED" ]; then
-    echo "flashinfer-jit-cache==${FLASHINFER_PYTHON_REQUIRED} already installed, keeping it"
-    UNINSTALL_JIT_CACHE=false
-else
-    echo "flashinfer-jit-cache version mismatch (installed: ${FLASHINFER_JIT_INSTALLED:-none}, required: ${FLASHINFER_PYTHON_REQUIRED}), will reinstall"
-fi
+setup_pip_toolchain() {
+    python3 -m pip install --upgrade pip
 
-if [ "$UNINSTALL_JIT_CACHE" = false ] && [ "$FLASHINFER_JIT_CU_VERSION" != "$CU_VERSION" ]; then
-    echo "flashinfer-jit-cache CUDA version mismatch (installed: ${FLASHINFER_JIT_CU_VERSION:-none}, required: ${CU_VERSION}), will reinstall"
+    if [ "$USE_VENV" != "1" ]; then
+        export UV_SYSTEM_PYTHON=1
+    fi
+
+    export UV_LINK_MODE=copy
+    PIP_CMD="uv pip"
+    PIP_INSTALL_SUFFIX="--index-strategy unsafe-best-match --prerelease allow"
+    PIP_UNINSTALL_CMD="uv pip uninstall"
+    PIP_UNINSTALL_SUFFIX=""
+
+    $PIP_UNINSTALL_CMD sgl-kernel sglang-kernel sglang sgl-fa4 flash-attn-4 $PIP_UNINSTALL_SUFFIX || true
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
+uninstall_stale_flashinfer() {
+    # Keep flashinfer packages if version matches to avoid re-downloading:
+    # - flashinfer-cubin: 150+ MB
+    # - flashinfer-jit-cache: 1.2+ GB
+    FLASHINFER_PYTHON_REQUIRED=$(grep -Po -m1 '(?<=flashinfer_python==)[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
+    FLASHINFER_CUBIN_REQUIRED=$(grep -Po -m1 '(?<=flashinfer_cubin==)[0-9A-Za-z\.\-]+' python/pyproject.toml || echo "")
+    FLASHINFER_CUBIN_INSTALLED=$(pip show flashinfer-cubin 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
+    FLASHINFER_JIT_INSTALLED=$(pip show flashinfer-jit-cache 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//' || echo "")
+    FLASHINFER_JIT_CU_VERSION=$(pip show flashinfer-jit-cache 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed -n 's/.*+//p' || echo "")
+
+    UNINSTALL_CUBIN=true
     UNINSTALL_JIT_CACHE=true
-fi
 
-# Build uninstall list based on what needs updating
-FLASHINFER_UNINSTALL="flashinfer-python"
-[ "$UNINSTALL_CUBIN" = true ] && FLASHINFER_UNINSTALL="$FLASHINFER_UNINSTALL flashinfer-cubin"
-[ "$UNINSTALL_JIT_CACHE" = true ] && FLASHINFER_UNINSTALL="$FLASHINFER_UNINSTALL flashinfer-jit-cache"
-$PIP_UNINSTALL_CMD $FLASHINFER_UNINSTALL $PIP_UNINSTALL_SUFFIX || true
-$PIP_UNINSTALL_CMD opencv-python opencv-python-headless $PIP_UNINSTALL_SUFFIX || true
-
-mark_step_done "Uninstall Flashinfer"
-
-# ------------------------------------------------------------------------------
-# Install main package
-# ------------------------------------------------------------------------------
-# Install the main package
-EXTRAS="dev,runai,tracing"
-if [ -n "$OPTIONAL_DEPS" ]; then
-    EXTRAS="dev,runai,tracing,${OPTIONAL_DEPS}"
-fi
-echo "Installing python extras: [${EXTRAS}]"
-# source "${SCRIPT_DIR}/cache_nvidia_wheels.sh"
-$PIP_CMD install -e "python[${EXTRAS}]" $PIP_INSTALL_SUFFIX
-
-mark_step_done "Install main package"
-
-# ------------------------------------------------------------------------------
-# Install torch/sglang-kernel
-# ------------------------------------------------------------------------------
-SGL_KERNEL_VERSION_FROM_KERNEL=$(grep -Po '(?<=^version = ")[^"]*' sgl-kernel/pyproject.toml)
-SGL_KERNEL_VERSION_FROM_SRT=$(grep -Po -m1 '(?<=sglang-kernel==)[0-9A-Za-z\.\-]+' python/pyproject.toml)
-echo "SGL_KERNEL_VERSION_FROM_KERNEL=${SGL_KERNEL_VERSION_FROM_KERNEL} SGL_KERNEL_VERSION_FROM_SRT=${SGL_KERNEL_VERSION_FROM_SRT}"
-
-
-if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ] && [ -d "sgl-kernel/dist" ]; then
-    ls -alh sgl-kernel/dist
-    # Determine wheel architecture
-    if [ "$ARCH" = "aarch64" ] || [ "$ARCH" = "arm64" ]; then
-        WHEEL_ARCH="aarch64"
+    if [ "$FLASHINFER_CUBIN_INSTALLED" = "$FLASHINFER_CUBIN_REQUIRED" ] && [ -n "$FLASHINFER_CUBIN_REQUIRED" ]; then
+        echo "flashinfer-cubin==${FLASHINFER_CUBIN_REQUIRED} already installed, keeping it"
+        UNINSTALL_CUBIN=false
     else
-        WHEEL_ARCH="x86_64"
+        echo "flashinfer-cubin version mismatch (installed: ${FLASHINFER_CUBIN_INSTALLED:-none}, required: ${FLASHINFER_CUBIN_REQUIRED}), reinstalling"
     fi
-    # Wheel filenames carry a +cuXYZ local version tag (e.g. sglang_kernel-0.4.0+cu130-...).
-    # Pick the one matching the test runner's $CU_VERSION; any other tag (or an untagged
-    # wheel of unknown CUDA origin) would at best hit the PyPI-reinstall branch below and
-    # silently replace the PR-built wheel with the public main-branch wheel, and at worst
-    # ABI-crash at cuInit. Better to fail loudly — the error message + dist/ listing below
-    # makes the diagnosis obvious.
-    # `|| true` swallows `ls`'s exit-2-on-no-match so `set -o pipefail` doesn't abort the
-    # script before we reach the explicit error check.
-    KERNEL_WHL=$(ls sgl-kernel/dist/sglang_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}+${CU_VERSION}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl 2>/dev/null | head -1 || true)
-    if [ -z "$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} cuda ${CU_VERSION}"
-        ls -alh sgl-kernel/dist/
-        exit 1
-    fi
-    echo "Installing sgl-kernel wheel: $KERNEL_WHL"
-    $PIP_CMD install "$KERNEL_WHL" --force-reinstall $PIP_INSTALL_SUFFIX
-else
-    if [ "${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
-        echo "ERROR: CUSTOM_BUILD_SGL_KERNEL=true but sgl-kernel/dist not found."
-        echo "This usually happens when rerunning a stage without the sgl-kernel-build-wheels job."
-        echo "Please re-run the full workflow using /tag-and-rerun-ci to rebuild the kernel."
-        exit 1
-    fi
-fi
 
-# Now we are running torch with cuda13 in CI environment, so the torch packages will be reinstalled if they are still at CU129 version
-# TODO: Remove this part after torch has been upgraded to 2.11, where cu13 is enabled by default
-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: ${TORCH_CUDA_VER}"
-if [ "${TORCH_CUDA_VER}" != "${CU_VERSION}" ]; then
-    TORCH_VER=$(pip show torch 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//')
-    TORCHAUDIO_VER=$(pip show torchaudio 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//')
-    TORCHVISION_VER=$(pip show torchvision 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//')
-    echo "Reinstalling torch==${TORCH_VER} torchaudio==${TORCHAUDIO_VER} torchvision==${TORCHVISION_VER} from ${CU_VERSION} index to match torch..."
-    $PIP_CMD install "torch==${TORCH_VER}" "torchaudio==${TORCHAUDIO_VER}" "torchvision==${TORCHVISION_VER}" --index-url "https://download.pytorch.org/whl/${CU_VERSION}" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
-fi
-
-# sglang-kernel wheels carry a +cuXYZ local version tag (e.g. 0.4.1+cu130).
-# If it doesn't match CU_VERSION, reinstall from the matching index.
-SGL_KERNEL_FULL_VER=$(pip show sglang-kernel 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
-SGL_KERNEL_CUDA_VER=$(printf '%s' "$SGL_KERNEL_FULL_VER" | sed -n 's/.*+//p')
-echo "Detected sglang-kernel version: ${SGL_KERNEL_FULL_VER} (CUDA tag: ${SGL_KERNEL_CUDA_VER:-none})"
-if [ -n "$SGL_KERNEL_CUDA_VER" ] && [ "$SGL_KERNEL_CUDA_VER" != "$CU_VERSION" ]; then
-    SGL_KERNEL_VER="${SGL_KERNEL_FULL_VER%+*}"
-    echo "Reinstalling sglang-kernel==${SGL_KERNEL_VER} from ${CU_VERSION} index to match torch..."
-    if [ "$CU_MAJOR" = "13" ]; then
-        $PIP_CMD install "sglang-kernel==${SGL_KERNEL_VER}" --index-url "https://docs.sglang.ai/whl/${CU_VERSION}/" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
+    if [ "$FLASHINFER_JIT_INSTALLED" = "$FLASHINFER_PYTHON_REQUIRED" ] && [ -n "$FLASHINFER_PYTHON_REQUIRED" ]; then
+        echo "flashinfer-jit-cache==${FLASHINFER_PYTHON_REQUIRED} already installed, keeping it"
+        UNINSTALL_JIT_CACHE=false
     else
-        $PIP_CMD install "sglang-kernel==${SGL_KERNEL_VER}" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
+        echo "flashinfer-jit-cache version mismatch (installed: ${FLASHINFER_JIT_INSTALLED:-none}, required: ${FLASHINFER_PYTHON_REQUIRED}), will reinstall"
     fi
-fi
 
-mark_step_done "Install sglang-kernel"
+    if [ "$UNINSTALL_JIT_CACHE" = false ] && [ "$FLASHINFER_JIT_CU_VERSION" != "$CU_VERSION" ]; then
+        echo "flashinfer-jit-cache CUDA version mismatch (installed: ${FLASHINFER_JIT_CU_VERSION:-none}, required: ${CU_VERSION}), will reinstall"
+        UNINSTALL_JIT_CACHE=true
+    fi
 
-# ------------------------------------------------------------------------------
-# Install sglang-router
-# ------------------------------------------------------------------------------
-# Install router for pd-disagg test
-$PIP_CMD install sglang-router $PIP_INSTALL_SUFFIX
+    FLASHINFER_UNINSTALL="flashinfer-python"
+    [ "$UNINSTALL_CUBIN" = true ] && FLASHINFER_UNINSTALL="$FLASHINFER_UNINSTALL flashinfer-cubin"
+    [ "$UNINSTALL_JIT_CACHE" = true ] && FLASHINFER_UNINSTALL="$FLASHINFER_UNINSTALL flashinfer-jit-cache"
+    $PIP_UNINSTALL_CMD $FLASHINFER_UNINSTALL $PIP_UNINSTALL_SUFFIX || true
+    $PIP_UNINSTALL_CMD opencv-python opencv-python-headless $PIP_UNINSTALL_SUFFIX || true
 
-# Show current packages
-$PIP_CMD list
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-mark_step_done "Install sglang-router"
+install_sglang() {
+    EXTRAS="dev,runai,tracing"
+    if [ -n "$OPTIONAL_DEPS" ]; then
+        EXTRAS="dev,runai,tracing,${OPTIONAL_DEPS}"
+    fi
+    echo "Installing python extras: [${EXTRAS}]"
+    $PIP_CMD install -e "python[${EXTRAS}]" $PIP_INSTALL_SUFFIX
 
-# ------------------------------------------------------------------------------
-# Download flashinfer artifacts
-# ------------------------------------------------------------------------------
-# Download flashinfer jit cache
-UNINSTALL_JIT_CACHE="$UNINSTALL_JIT_CACHE" \
-    FLASHINFER_PYTHON_REQUIRED="$FLASHINFER_PYTHON_REQUIRED" \
-    CU_VERSION="$CU_VERSION" \
-    PIP_CMD="$PIP_CMD" \
-    PIP_INSTALL_SUFFIX="$PIP_INSTALL_SUFFIX" \
-    bash "${SCRIPT_DIR}/ci_download_flashinfer_jit_cache.sh"
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-mark_step_done "Download flashinfer artifacts"
+install_sglang_kernel() {
+    SGL_KERNEL_VERSION_FROM_KERNEL=$(grep -Po '(?<=^version = ")[^"]*' sgl-kernel/pyproject.toml)
+    SGL_KERNEL_VERSION_FROM_SRT=$(grep -Po -m1 '(?<=sglang-kernel==)[0-9A-Za-z\.\-]+' python/pyproject.toml)
+    echo "SGL_KERNEL_VERSION_FROM_KERNEL=${SGL_KERNEL_VERSION_FROM_KERNEL} SGL_KERNEL_VERSION_FROM_SRT=${SGL_KERNEL_VERSION_FROM_SRT}"
+
+    if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ] && [ -d "sgl-kernel/dist" ]; then
+        ls -alh sgl-kernel/dist
+        if [ "$ARCH" = "aarch64" ] || [ "$ARCH" = "arm64" ]; then
+            WHEEL_ARCH="aarch64"
+        else
+            WHEEL_ARCH="x86_64"
+        fi
+        KERNEL_WHL=$(ls sgl-kernel/dist/sglang_kernel-${SGL_KERNEL_VERSION_FROM_KERNEL}+${CU_VERSION}-cp310-abi3-manylinux2014_${WHEEL_ARCH}.whl 2>/dev/null | head -1 || true)
+        if [ -z "$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} cuda ${CU_VERSION}"
+            ls -alh sgl-kernel/dist/
+            exit 1
+        fi
+        echo "Installing sgl-kernel wheel: $KERNEL_WHL"
+        $PIP_CMD install "$KERNEL_WHL" --force-reinstall $PIP_INSTALL_SUFFIX
+    else
+        if [ "${CUSTOM_BUILD_SGL_KERNEL:-}" = "true" ] && [ ! -d "sgl-kernel/dist" ]; then
+            echo "ERROR: CUSTOM_BUILD_SGL_KERNEL=true but sgl-kernel/dist not found."
+            echo "This usually happens when rerunning a stage without the sgl-kernel-build-wheels job."
+            echo "Please re-run the full workflow using /tag-and-rerun-ci to rebuild the kernel."
+            exit 1
+        fi
+    fi
+
+    # Reinstall torch with matching CUDA version if needed
+    # TODO: Remove after torch 2.11 where cu13 is enabled by default
+    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: ${TORCH_CUDA_VER}"
+    if [ "${TORCH_CUDA_VER}" != "${CU_VERSION}" ]; then
+        TORCH_VER=$(pip show torch 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//')
+        TORCHAUDIO_VER=$(pip show torchaudio 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//')
+        TORCHVISION_VER=$(pip show torchvision 2>/dev/null | grep "^Version:" | awk '{print $2}' | sed 's/+.*//')
+        echo "Reinstalling torch==${TORCH_VER} torchaudio==${TORCHAUDIO_VER} torchvision==${TORCHVISION_VER} from ${CU_VERSION} index to match torch..."
+        $PIP_CMD install "torch==${TORCH_VER}" "torchaudio==${TORCHAUDIO_VER}" "torchvision==${TORCHVISION_VER}" --index-url "https://download.pytorch.org/whl/${CU_VERSION}" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
+    fi
+
+    # Reinstall sglang-kernel with matching CUDA version if needed
+    SGL_KERNEL_FULL_VER=$(pip show sglang-kernel 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
+    SGL_KERNEL_CUDA_VER=$(printf '%s' "$SGL_KERNEL_FULL_VER" | sed -n 's/.*+//p')
+    echo "Detected sglang-kernel version: ${SGL_KERNEL_FULL_VER} (CUDA tag: ${SGL_KERNEL_CUDA_VER:-none})"
+    if [ -n "$SGL_KERNEL_CUDA_VER" ] && [ "$SGL_KERNEL_CUDA_VER" != "$CU_VERSION" ]; then
+        SGL_KERNEL_VER="${SGL_KERNEL_FULL_VER%+*}"
+        echo "Reinstalling sglang-kernel==${SGL_KERNEL_VER} from ${CU_VERSION} index to match torch..."
+        if [ "$CU_MAJOR" = "13" ]; then
+            $PIP_CMD install "sglang-kernel==${SGL_KERNEL_VER}" --index-url "https://docs.sglang.ai/whl/${CU_VERSION}/" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
+        else
+            $PIP_CMD install "sglang-kernel==${SGL_KERNEL_VER}" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
+        fi
+    fi
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
+install_sglang_router() {
+    $PIP_CMD install sglang-router $PIP_INSTALL_SUFFIX
+    $PIP_CMD list
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
+download_flashinfer_cache() {
+    UNINSTALL_JIT_CACHE="$UNINSTALL_JIT_CACHE" \
+        FLASHINFER_PYTHON_REQUIRED="$FLASHINFER_PYTHON_REQUIRED" \
+        CU_VERSION="$CU_VERSION" \
+        PIP_CMD="$PIP_CMD" \
+        PIP_INSTALL_SUFFIX="$PIP_INSTALL_SUFFIX" \
+        bash "${SCRIPT_DIR}/ci_download_flashinfer_jit_cache.sh"
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
+stabilize_flashinfer_jit_paths() {
+    # In venv mode, FlashInfer JIT writes build.ninja with hardcoded -isystem
+    # paths. Per-job venvs get unique paths, but the JIT cache is shared on the
+    # host mount. Fix by symlinking venv copies to a stable host-mounted path.
+    if [ "$USE_VENV" != "1" ]; then
+        return
+    fi
 
-# ------------------------------------------------------------------------------
-# Stabilize FlashInfer JIT cache paths
-# ------------------------------------------------------------------------------
-# FlashInfer JIT writes build.ninja with hardcoded -isystem paths pointing to the
-# venv's flashinfer/data/ and tvm_ffi/include/. With per-job venvs each job gets
-# a unique /tmp/sglang-ci-<run>-<job>-<pid>/ path, but the JIT cache is shared
-# on the host mount. When the next job's venv has a different path and the old one
-# is cleaned up, ninja fails because source files no longer exist at the cached path.
-#
-# Fix (two parts):
-# 1. Clear only STALE cached_ops (build.ninja referencing non-existent venv paths).
-#    Do NOT clear all cached_ops — they contain compiled .so files that take 10-20 min
-#    to recompile. Only remove entries where the source paths no longer exist.
-# 2. Copy source files to a stable host-mounted path and symlink each venv's
-#    copy there. build.ninja then references the stable path across all jobs.
-#
-# Part 1: Clear stale cached_ops (keep valid compiled kernels)
-if [ "$USE_VENV" = "1" ]; then
     STABLE_FI_DIR="${HOME}/.cache/flashinfer/_stable_src"
+
+    # Clear stale cached_ops (keep valid compiled kernels)
     if [ -d "${HOME}/.cache/flashinfer" ]; then
         STALE_COUNT=0
         while IFS= read -r ninja_file; do
-            # Check for stale venv paths (/tmp/sglang-ci-*) or old stable path (flashinfer-src)
             STALE_PATH=$(grep -o '/tmp/sglang-ci-[^ ]*\|flashinfer-src' "$ninja_file" 2>/dev/null | head -1 || true)
             if [ -n "$STALE_PATH" ]; then
                 if echo "$STALE_PATH" | grep -q "flashinfer-src" || [ ! -d "$STALE_PATH" ]; then
@@ -428,7 +341,7 @@ if [ "$USE_VENV" = "1" ]; then
         echo "Cleaned $STALE_COUNT stale FlashInfer cached_ops (kept valid ones)"
     fi
 
-    # Part 2: Stabilize paths (STABLE_FI_DIR set above in Part 1)
+    # Copy source files to stable path and symlink venv copies there
     FI_DATA=$(python3 -c "import flashinfer, os; print(os.path.join(os.path.dirname(flashinfer.__file__), 'data'))")
     TVM_INC=$(python3 -c "import tvm_ffi, os; print(os.path.join(os.path.dirname(tvm_ffi.__file__), 'include'))")
 
@@ -451,127 +364,132 @@ if [ "$USE_VENV" = "1" ]; then
     ln -s "$STABLE_FI_DIR/tvm-ffi-include" "$TVM_INC_PARENT/include"
     echo "Symlinked venv flashinfer/tvm_ffi -> $STABLE_FI_DIR"
 
-    mark_step_done "Stabilize FlashInfer JIT cache paths"
-fi
+    mark_step_done "${FUNCNAME[0]}"
+}
 
+install_extra_deps() {
+    if [ "$CU_MAJOR" = "13" ]; then
+        MOONCAKE_PKG="mooncake-transfer-engine-cuda13==0.3.10.post2"
+        EXTRA_NVIDIA_SPECS="nvidia-cuda-nvrtc"
+    else
+        MOONCAKE_PKG="mooncake-transfer-engine==0.3.10.post2"
+        EXTRA_NVIDIA_SPECS="nvidia-cuda-nvrtc-cu12"
+    fi
+    $PIP_CMD install ${MOONCAKE_PKG} ${EXTRA_NVIDIA_SPECS} py-spy scipy huggingface_hub[hf_xet] pytest $PIP_INSTALL_SUFFIX
 
-# ------------------------------------------------------------------------------
-# Install extra dependency
-# ------------------------------------------------------------------------------
-# Install other python dependencies.
-# Match on CUDA major version so future minor bumps (cu131, etc.) don't fall
-# through to the wrong branch. Prefer NVCC_VER (set in the venv path); otherwise
-# parse the first two digits of CU_VERSION (pytorch convention is cu{major}{minor}
-# with a single-digit minor, e.g. cu126, cu129, cu130).
-if [ "$CU_MAJOR" = "13" ]; then
-    MOONCAKE_PKG="mooncake-transfer-engine-cuda13==0.3.10.post2"
-    EXTRA_NVIDIA_SPECS="nvidia-cuda-nvrtc"
-else
-    MOONCAKE_PKG="mooncake-transfer-engine==0.3.10.post2"
-    EXTRA_NVIDIA_SPECS="nvidia-cuda-nvrtc-cu12"
-fi
-$PIP_CMD install ${MOONCAKE_PKG} ${EXTRA_NVIDIA_SPECS} py-spy scipy huggingface_hub[hf_xet] pytest $PIP_INSTALL_SUFFIX
+    if [ "$IS_BLACKWELL" != "1" ]; then
+        git clone --branch v0.5 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
+        $PIP_CMD install -e lmms-eval/ $PIP_INSTALL_SUFFIX
+    fi
+    $PIP_CMD uninstall xformers || true
 
-# Install other test dependencies
-if [ "$IS_BLACKWELL" != "1" ]; then
-    # For lmms_evals evaluating MMMU
-    git clone --branch v0.5 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
-    $PIP_CMD install -e lmms-eval/ $PIP_INSTALL_SUFFIX
-fi
-$PIP_CMD uninstall xformers || true
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-mark_step_done "Install extra dependency"
+fix_nvidia_deps() {
+    if [ "$CU_MAJOR" = "13" ]; then
+        NVSHMEM_PKG="nvidia-nvshmem-cu13"
+        CUDNN_PKG="nvidia-cudnn-cu13"
+    else
+        NVSHMEM_PKG="nvidia-nvshmem-cu12"
+        CUDNN_PKG="nvidia-cudnn-cu12"
+    fi
 
-# ------------------------------------------------------------------------------
-# Fix other dependencies
-# ------------------------------------------------------------------------------
+    # DeepEP depends on nvshmem 3.4.5
+    INSTALLED_NVSHMEM=$(pip show ${NVSHMEM_PKG} 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
+    if [ "$INSTALLED_NVSHMEM" = "$NVIDIA_NVSHMEM_VERSION" ]; then
+        echo "${NVSHMEM_PKG}==${NVIDIA_NVSHMEM_VERSION} already installed, skipping reinstall"
+    else
+        $PIP_CMD install ${NVSHMEM_PKG}==${NVIDIA_NVSHMEM_VERSION} $PIP_INSTALL_SUFFIX
+    fi
 
-# Pick cu12 vs cu13 variants of nvshmem / cudnn based on CU_VERSION
-if [ "$CU_MAJOR" = "13" ]; then
-    NVSHMEM_PKG="nvidia-nvshmem-cu13"
-    CUDNN_PKG="nvidia-cudnn-cu13"
-else
-    NVSHMEM_PKG="nvidia-nvshmem-cu12"
-    CUDNN_PKG="nvidia-cudnn-cu12"
-fi
+    # cudnn < 9.16.0.29 causes Conv3D performance regression
+    INSTALLED_CUDNN=$(pip show ${CUDNN_PKG} 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
+    if [ "$INSTALLED_CUDNN" = "$NVIDIA_CUDNN_VERSION" ]; then
+        echo "${CUDNN_PKG}==${NVIDIA_CUDNN_VERSION} already installed, skipping reinstall"
+    else
+        $PIP_CMD install ${CUDNN_PKG}==${NVIDIA_CUDNN_VERSION} $PIP_INSTALL_SUFFIX
+    fi
 
-# Fix dependencies: DeepEP depends on nvshmem 3.4.5 — skip reinstall when already correct (avoids pip races / wasted work)
-INSTALLED_NVSHMEM=$(pip show ${NVSHMEM_PKG} 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
-if [ "$INSTALLED_NVSHMEM" = "$NVIDIA_NVSHMEM_VERSION" ]; then
-    echo "${NVSHMEM_PKG}==${NVIDIA_NVSHMEM_VERSION} already installed, skipping reinstall"
-else
-    $PIP_CMD install ${NVSHMEM_PKG}==${NVIDIA_NVSHMEM_VERSION} $PIP_INSTALL_SUFFIX
-fi
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-# Fix dependencies: Cudnn with version less than 9.16.0.29 will cause performance regression on Conv3D kernel
-INSTALLED_CUDNN=$(pip show ${CUDNN_PKG} 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
-if [ "$INSTALLED_CUDNN" = "$NVIDIA_CUDNN_VERSION" ]; then
-    echo "${CUDNN_PKG}==${NVIDIA_CUDNN_VERSION} already installed, skipping reinstall"
-else
-    $PIP_CMD install ${CUDNN_PKG}==${NVIDIA_CUDNN_VERSION} $PIP_INSTALL_SUFFIX
-fi
+install_test_tools() {
+    # Download kernels from kernels community
+    kernels download python || true
+    kernels lock python || true
+    [ -e "${HOME}/.cache/sglang" ] && [ ! -d "${HOME}/.cache/sglang" ] && rm -f "${HOME}/.cache/sglang"
+    mkdir -p "${HOME}/.cache/sglang/"
+    mv python/kernels.lock "${HOME}/.cache/sglang/" || true
 
-mark_step_done "Fix other dependencies"
+    # Install human-eval (subshell keeps cd local)
+    $PIP_CMD install "setuptools==70.0.0" $PIP_INSTALL_SUFFIX
+    [ -d human-eval ] || git clone https://github.com/merrymercy/human-eval.git
+    (
+        cd human-eval
+        $PIP_CMD install -e . --no-build-isolation $PIP_INSTALL_SUFFIX
+    )
 
-# Download kernels from kernels community
-kernels download python || true
-kernels lock python || true
-# Ensure target is a directory — on fresh containers or after a previous buggy
-# `mv` that created a FILE at this path, mkdir -p would fail silently.
-[ -e "${HOME}/.cache/sglang" ] && [ ! -d "${HOME}/.cache/sglang" ] && rm -f "${HOME}/.cache/sglang"
-mkdir -p "${HOME}/.cache/sglang/"
-mv python/kernels.lock "${HOME}/.cache/sglang/" || true
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-# Install human-eval. This script is sourced from ci_install_deepep.sh, so a
-# bare `cd human-eval` would leave the caller stuck in that directory for the
-# rest of its execution. The subshell keeps the cd local to the pip install.
-$PIP_CMD install "setuptools==70.0.0" $PIP_INSTALL_SUFFIX
-[ -d human-eval ] || git clone https://github.com/merrymercy/human-eval.git
-(
-    cd human-eval
-    $PIP_CMD install -e . --no-build-isolation $PIP_INSTALL_SUFFIX)
+prepare_runner() {
+    bash "${SCRIPT_DIR}/prepare_runner.sh"
 
-# ------------------------------------------------------------------------------
-# Prepare runner
-# ------------------------------------------------------------------------------
-# Prepare the CI runner (cleanup HuggingFace cache, etc.)
-bash "${SCRIPT_DIR}/prepare_runner.sh"
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-mark_step_done "Prepare runner"
+setup_ld_library_path() {
+    # NVIDIA pip packages and torch ship .so files under site-packages that are
+    # not on the default LD_LIBRARY_PATH.
+    SITE_PACKAGES=$(python3 -c "import site, sys; print(site.getsitepackages()[0])")
+    NVIDIA_LIBS=$(find "$SITE_PACKAGES" -path "*/nvidia/*/lib" -type d 2>/dev/null | tr '\n' ':')
+    TORCH_LIB="$SITE_PACKAGES/torch/lib"
+    VENV_LD="${NVIDIA_LIBS}${TORCH_LIB}"
+    export LD_LIBRARY_PATH="${VENV_LD}${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
 
-# ------------------------------------------------------------------------------
-# LD_LIBRARY_PATH discovery
-# ------------------------------------------------------------------------------
-# NVIDIA pip packages (cublas, cudnn, nccl, nvrtc, ...) and torch ship .so files
-# under site-packages. In venv mode these are NOT on the default LD_LIBRARY_PATH,
-# so dlopen('libcublas.so.12') from torch would fail. Prepend them here.
-# In non-venv mode, system site-packages may also need this if the runner's
-# default ld config doesn't cover the NVIDIA pip layout.
-SITE_PACKAGES=$(python3 -c "import site, sys; print(site.getsitepackages()[0])")
-# Glob matches NVIDIA pip-package layout:
-# site-packages/nvidia/<component>/lib/lib*.so. If NVIDIA restructures
-# packaging, this may need updating.
-NVIDIA_LIBS=$(find "$SITE_PACKAGES" -path "*/nvidia/*/lib" -type d 2>/dev/null | tr '\n' ':')
-TORCH_LIB="$SITE_PACKAGES/torch/lib"
-VENV_LD="${NVIDIA_LIBS}${TORCH_LIB}"
-export LD_LIBRARY_PATH="${VENV_LD}${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
-# Write LD_LIBRARY_PATH to the venv's env.sh (always succeeds — local file)
-# so subsequent steps auto-source it via BASH_ENV. In non-venv mode, skip the
-# env.sh write and rely on GITHUB_ENV propagation.
-if [ "$USE_VENV" = "1" ] && [ -n "$UV_VENV" ]; then
-    echo "export LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH\"" >> "$UV_VENV/env.sh"
-fi
-# Also try GITHUB_ENV (may fail if runner temp file was cleaned up during long installs).
-if [ -n "${GITHUB_ENV:-}" ]; then
-    echo "LD_LIBRARY_PATH=$LD_LIBRARY_PATH" >> "$GITHUB_ENV" || echo "WARNING: GITHUB_ENV write failed; LD_LIBRARY_PATH will be set via BASH_ENV instead"
-fi
-echo "LD_LIBRARY_PATH=$LD_LIBRARY_PATH"
+    if [ "$USE_VENV" = "1" ] && [ -n "$UV_VENV" ]; then
+        echo "export LD_LIBRARY_PATH=\"$LD_LIBRARY_PATH\"" >> "$UV_VENV/env.sh"
+    fi
+    if [ -n "${GITHUB_ENV:-}" ]; then
+        echo "LD_LIBRARY_PATH=$LD_LIBRARY_PATH" >> "$GITHUB_ENV" || echo "WARNING: GITHUB_ENV write failed; LD_LIBRARY_PATH will be set via BASH_ENV instead"
+    fi
+    echo "LD_LIBRARY_PATH=$LD_LIBRARY_PATH"
 
+    mark_step_done "${FUNCNAME[0]}"
+}
 
-# ------------------------------------------------------------------------------
-# Verify imports
-# ------------------------------------------------------------------------------
-# Show current packages
-$PIP_CMD list
-python3 -c "import torch; print(torch.version.cuda)"
-python3 -c "import cutlass; import cutlass.cute;"
+verify_imports() {
+    $PIP_CMD list
+    python3 -c "import torch; print(torch.version.cuda)"
+    python3 -c "import cutlass; import cutlass.cute;"
+
+    mark_step_done "${FUNCNAME[0]}"
+}
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+
+main() {
+    configure_environment "$@"
+    detect_host
+    kill_existing_processes
+    install_apt_packages
+    clean_site_packages
+    setup_pip_toolchain
+    uninstall_stale_flashinfer
+    install_sglang
+    install_sglang_kernel
+    install_sglang_router
+    download_flashinfer_cache
+    stabilize_flashinfer_jit_paths
+    install_extra_deps
+    fix_nvidia_deps
+    install_test_tools
+    prepare_runner
+    setup_ld_library_path
+    verify_imports
+}
+
+main "$@"