diff --git a/.github/workflows/pr-test.yml b/.github/workflows/pr-test.yml
index 458ffc783..2161c6443 100644
--- a/.github/workflows/pr-test.yml
+++ b/.github/workflows/pr-test.yml
@@ -915,10 +915,16 @@ jobs:
         with:
           python-version: '3.10'
 
+      - name: Install uv
+        uses: astral-sh/setup-uv@v5
+
+      # uv pip targets a venv by default; setup-python has no venv — install into that interpreter (see UV_SYSTEM_PYTHON in https://docs.astral.sh/uv/guides/integration/github/)
       - name: Install dependencies
         timeout-minutes: 20
+        env:
+          UV_SYSTEM_PYTHON: "1"
         run: |
-          pip install -e "python/[dev]"
+          uv pip install -e "python[dev]" --index-strategy unsafe-best-match --prerelease allow
 
       - name: Run test
         timeout-minutes: 10
diff --git a/scripts/ci/cuda/ci_download_flashinfer_jit_cache.sh b/scripts/ci/cuda/ci_download_flashinfer_jit_cache.sh
new file mode 100755
index 000000000..bca94477c
--- /dev/null
+++ b/scripts/ci/cuda/ci_download_flashinfer_jit_cache.sh
@@ -0,0 +1,71 @@
+#!/bin/bash
+# Install flashinfer-jit-cache with caching and retry logic (flashinfer.ai can have transient DNS issues).
+# The jit-cache wheel is 1.2+ GB, so we skip the download entirely if already installed.
+#
+# 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
+#   PIP_CMD                      — e.g. "pip" or "uv pip"
+#   PIP_INSTALL_SUFFIX           — extra pip args for this runner
+set -euxo pipefail
+
+: "${UNINSTALL_JIT_CACHE:?must be set}"
+: "${FLASHINFER_PYTHON_REQUIRED:?must be set}"
+: "${CU_VERSION:?must be set}"
+: "${PIP_CMD:?must be set}"
+
+FLASHINFER_JIT_CACHE_INSTALLED=false
+if [ "$UNINSTALL_JIT_CACHE" = false ]; then
+    FLASHINFER_JIT_CACHE_INSTALLED=true
+    echo "flashinfer-jit-cache already at correct version, skipping download"
+fi
+
+if [ "$FLASHINFER_JIT_CACHE_INSTALLED" = false ]; then
+    FLASHINFER_CACHE_DIR="${HOME}/.cache/flashinfer-wheels"
+    mkdir -p "${FLASHINFER_CACHE_DIR}"
+
+    find "${FLASHINFER_CACHE_DIR}" -name "flashinfer_jit_cache-*.whl" ! -name "flashinfer_jit_cache-${FLASHINFER_PYTHON_REQUIRED}*" -type f -delete 2>/dev/null || true
+
+    FLASHINFER_WHEEL_PATTERN="flashinfer_jit_cache-${FLASHINFER_PYTHON_REQUIRED}*.whl"
+    CACHED_WHEEL=$(find "${FLASHINFER_CACHE_DIR}" -name "${FLASHINFER_WHEEL_PATTERN}" -type f 2>/dev/null | head -n 1)
+
+    if [ -n "$CACHED_WHEEL" ] && [ -f "$CACHED_WHEEL" ]; then
+        echo "Found cached flashinfer wheel: $CACHED_WHEEL"
+        if $PIP_CMD install "$CACHED_WHEEL" $PIP_INSTALL_SUFFIX; then
+            FLASHINFER_JIT_CACHE_INSTALLED=true
+            echo "Successfully installed flashinfer-jit-cache from cache"
+        else
+            echo "Failed to install from cache, will try downloading..."
+            rm -f "$CACHED_WHEEL"
+        fi
+    fi
+
+    if [ "$FLASHINFER_JIT_CACHE_INSTALLED" = false ]; then
+        for i in {1..5}; do
+            # Download wheel to cache directory (use pip directly as uv pip doesn't support download)
+            if timeout 600 pip download "flashinfer-jit-cache==${FLASHINFER_PYTHON_REQUIRED}" \
+                --index-url "https://flashinfer.ai/whl/${CU_VERSION}" \
+                -d "${FLASHINFER_CACHE_DIR}"; then
+
+                CACHED_WHEEL=$(find "${FLASHINFER_CACHE_DIR}" -name "${FLASHINFER_WHEEL_PATTERN}" -type f 2>/dev/null | head -n 1)
+                if [ -n "$CACHED_WHEEL" ] && [ -f "$CACHED_WHEEL" ]; then
+                    if $PIP_CMD install "$CACHED_WHEEL" $PIP_INSTALL_SUFFIX; then
+                        FLASHINFER_JIT_CACHE_INSTALLED=true
+                        echo "Successfully downloaded and installed flashinfer-jit-cache"
+                        break
+                    fi
+                else
+                    echo "Warning: Download succeeded but wheel file not found"
+                fi
+            fi
+            echo "Attempt $i to download flashinfer-jit-cache failed, retrying in 10 seconds..."
+            sleep 10
+        done
+    fi
+fi
+
+if [ "$FLASHINFER_JIT_CACHE_INSTALLED" = false ]; then
+    echo "ERROR: Failed to install flashinfer-jit-cache after 5 attempts"
+    exit 1
+fi
diff --git a/scripts/ci/cuda/ci_install_dependency.sh b/scripts/ci/cuda/ci_install_dependency.sh
index 86e4fb4e9..53cfc5e90 100755
--- a/scripts/ci/cuda/ci_install_dependency.sh
+++ b/scripts/ci/cuda/ci_install_dependency.sh
@@ -1,37 +1,110 @@
 #!/bin/bash
 # Install the dependency in CI.
+#
+# 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
 set -euxo pipefail
 
+# ------------------------------------------------------------------------------
+# Configuration & timing
+# ------------------------------------------------------------------------------
 # Set up environment variables
-IS_BLACKWELL=${IS_BLACKWELL:-0}
 CU_VERSION="cu129"
-FLASHINFER_VERSION=0.6.6
 OPTIONAL_DEPS="${1:-}"
 
-# Detect system architecture
+SECONDS=0
+_CI_MARK_PREV=${SECONDS}
+
+mark_step_done() {
+    local label=$1
+    local now=${SECONDS}
+    local step=$((now - _CI_MARK_PREV))
+    printf '\n[STEP DONE] %s,  step: %ss,  total: %ss,  date: %s\n' \
+        "${label}" "${step}" "${now}" "$(date -u '+%Y-%m-%dT%H:%M:%SZ')"
+    _CI_MARK_PREV=${now}
+}
+
+mark_step_done "Configuration"
+
+# ------------------------------------------------------------------------------
+# Host / runner detection (CPU arch, Blackwell, USE_UV)
+# ------------------------------------------------------------------------------
+# Detect CPU architecture (x86_64 or aarch64)
 ARCH=$(uname -m)
 echo "Detected architecture: ${ARCH}"
 
-if [ "$CU_VERSION" = "cu130" ]; then
-    NVRTC_SPEC="nvidia-cuda-nvrtc"
+# Detect GPU architecture (blackwell or not)
+if [ "${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
-    NVRTC_SPEC="nvidia-cuda-nvrtc-cu12"
+    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 <<< "$(nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)"
+    fi
+    echo "IS_BLACKWELL=${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=${USE_UV}"
+
+mark_step_done "Host / runner detection"
+
+# ------------------------------------------------------------------------------
 # Kill existing processes
-SCRIPT_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" && pwd )"
+# ------------------------------------------------------------------------------
+SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
 bash "${SCRIPT_DIR}/../../killall_sglang.sh"
 echo "CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES:-}"
 
+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
-apt-get install -y --no-install-recommends 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 || {
+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 "${CI_APT_PACKAGES[@]}" || {
     echo "Warning: apt-get install failed, checking if required packages are available..."
-    # Verify the packages we need are actually installed
-    for pkg in python3 python3-pip python3-venv python3-dev git libnuma-dev libssl-dev pkg-config libibverbs-dev libibverbs1 ibverbs-providers ibverbs-utils ffmpeg; do
+    for pkg in "${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
@@ -40,66 +113,19 @@ apt-get install -y --no-install-recommends python3 python3-pip python3-venv pyth
     echo "All required packages are already installed, continuing..."
 }
 
+mark_step_done "Install apt packages"
+
+# ------------------------------------------------------------------------------
+# 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)'
 
-# 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: $(protoc --version)"
-    else
-        echo "protoc found but not runnable, reinstalling..."
-        INSTALL_PROTOC=1
-    fi
-else
-    INSTALL_PROTOC=1
-fi
-
-# Install protoc for router build (gRPC protobuf compilation)
-if [ "${INSTALL_PROTOC:-0}" = "1" ]; then
-    # TODO: move this to a separate script
-    echo "Installing protoc..."
-    if command -v apt-get &> /dev/null; then
-        # Ubuntu/Debian
-        apt-get update || true  # May fail due to unrelated broken packages
-        apt-get install -y --no-install-recommends wget unzip gcc g++ perl make || {
-            echo "Warning: apt-get install failed, checking if required packages are available..."
-            for pkg in wget unzip gcc g++ perl make; do
-                if ! dpkg -l "$pkg" 2>/dev/null | grep -q "^ii"; then
-                    echo "ERROR: Required package $pkg is not installed and apt-get failed"
-                    exit 1
-                fi
-            done
-            echo "All required packages are already installed, continuing..."
-        }
-    elif command -v yum &> /dev/null; then
-        # RHEL/CentOS
-        yum update -y
-        yum install -y wget unzip gcc gcc-c++ perl-core make
-    fi
-
-    cd /tmp
-    # Determine protoc architecture
-    if [ "$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
-    echo "protoc already installed: $(protoc --version)"
-fi
-
 # Remove broken dist-info directories (missing METADATA per PEP 376)
 SITE_PACKAGES=$(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 d; do
+    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"
@@ -108,18 +134,23 @@ if [ -d "$SITE_PACKAGES" ]; then
     set -x
 fi
 
-# Install uv (use python3 -m pip for robustness since some runners only have pip3)
+# Install protoc
+bash "${SCRIPT_DIR}/../utils/install_protoc.sh"
+
+mark_step_done "Python package site hygiene & install protoc"
+
+# ------------------------------------------------------------------------------
+# Pip / uv toolchain & stale package cleanup
+# ------------------------------------------------------------------------------
+# Install pip and uv (use python3 -m pip for robustness since some runners only have pip3)
 python3 -m pip install --upgrade pip
 
-if [ "$IS_BLACKWELL" = "1" ]; then
-    # The blackwell CI runner has some issues with pip and uv,
-    # so we can only use pip with `--break-system-packages`
+if [ "$USE_UV" = "0" ]; then
     PIP_CMD="pip"
     PIP_INSTALL_SUFFIX="--break-system-packages"
     PIP_UNINSTALL_CMD="pip uninstall -y"
     PIP_UNINSTALL_SUFFIX="--break-system-packages"
 else
-    # In normal cases, we use uv, which is much faster than pip.
     pip install uv
     export UV_SYSTEM_PYTHON=true
 
@@ -132,9 +163,15 @@ fi
 # Clean up existing installations
 $PIP_UNINSTALL_CMD sgl-kernel sglang-kernel sglang sgl-fa4 flash-attn-4 $PIP_UNINSTALL_SUFFIX || true
 
+mark_step_done "Pip / uv toolchain & stale package cleanup"
+
+# ------------------------------------------------------------------------------
+# Uninstall Flashinfer
+# ------------------------------------------------------------------------------
 # Keep flashinfer packages installed if version matches to avoid re-downloading:
 # - flashinfer-cubin: 150+ MB, plus extra cubins from ci_download_flashinfer_cubin.sh
 # - 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 "")
@@ -149,11 +186,11 @@ else
     echo "flashinfer-cubin version mismatch (installed: ${FLASHINFER_CUBIN_INSTALLED:-none}, required: ${FLASHINFER_CUBIN_REQUIRED}), reinstalling"
 fi
 
-if [ "$FLASHINFER_JIT_INSTALLED" = "$FLASHINFER_VERSION" ] && [ -n "$FLASHINFER_VERSION" ]; then
-    echo "flashinfer-jit-cache==${FLASHINFER_VERSION} already installed, keeping it"
+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_VERSION}), will reinstall"
+    echo "flashinfer-jit-cache version mismatch (installed: ${FLASHINFER_JIT_INSTALLED:-none}, required: ${FLASHINFER_PYTHON_REQUIRED}), will reinstall"
 fi
 
 # Build uninstall list based on what needs updating
@@ -163,34 +200,24 @@ FLASHINFER_UNINSTALL="flashinfer-python"
 $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"
 if [ -n "$OPTIONAL_DEPS" ]; then
     EXTRAS="dev,${OPTIONAL_DEPS}"
 fi
 echo "Installing python extras: [${EXTRAS}]"
-
 $PIP_CMD install -e "python[${EXTRAS}]" --extra-index-url https://download.pytorch.org/whl/${CU_VERSION} $PIP_INSTALL_SUFFIX
 
-# Fix CUDA version mismatch between torch and torchaudio.
-# PyPI's torch 2.9.1 bundles cu128 but torchaudio from pytorch.org/cu129 uses cu129.
-# 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
-# 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: ${TORCH_CUDA_VER}"
-if [ "${TORCH_CUDA_VER}" != "${CU_VERSION}" ]; then
-    # Pin versions to match what was installed by pyproject.toml (strip +cuXYZ suffix)
-    TORCHAUDIO_VER=$(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 $2}' | sed 's/+.*//')
-    echo "Reinstalling torchaudio==${TORCHAUDIO_VER} torchvision==${TORCHVISION_VER} from ${TORCH_CUDA_VER} index to match torch..."
-    $PIP_CMD install "torchaudio==${TORCHAUDIO_VER}" "torchvision==${TORCHVISION_VER}" --index-url "https://download.pytorch.org/whl/${TORCH_CUDA_VER}" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
-fi
-
-# Install router for pd-disagg test
-$PIP_CMD install sglang-router $PIP_INSTALL_SUFFIX
+mark_step_done "Install main package"
 
+# ------------------------------------------------------------------------------
+# Install sglang-kernel
+# ------------------------------------------------------------------------------
 # Install sgl-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)
@@ -227,117 +254,106 @@ else
     fi
 fi
 
+mark_step_done "Install sglang-kernel"
+
+# ------------------------------------------------------------------------------
+# Install sglang-router
+# ------------------------------------------------------------------------------
+# Install router for pd-disagg test
+$PIP_CMD install sglang-router $PIP_INSTALL_SUFFIX
+
 # Show current packages
 $PIP_CMD list
 
+mark_step_done "Install sglang-router"
+
+# ------------------------------------------------------------------------------
+# 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"
+# Download flashinfer cubins
+bash "${SCRIPT_DIR}/ci_download_flashinfer_cubin.sh"
+
+mark_step_done "Download flashinfer artifacts"
+
+# ------------------------------------------------------------------------------
+# Install extra dependency
+# ------------------------------------------------------------------------------
 # Install other python dependencies
+if [ "$CU_VERSION" = "cu130" ]; then
+    NVRTC_SPEC="nvidia-cuda-nvrtc"
+else
+    NVRTC_SPEC="nvidia-cuda-nvrtc-cu12"
+fi
 $PIP_CMD install mooncake-transfer-engine==0.3.9 "${NVRTC_SPEC}" py-spy scipy huggingface_hub[hf_xet] pytest $PIP_INSTALL_SUFFIX
 
+# 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/ $PIP_INSTALL_SUFFIX
 fi
-
-# DeepEP depends on nvshmem 3.4.5
-# On Blackwell machines, skip reinstall if correct version already installed to avoid race conditions
-if [ "$IS_BLACKWELL" = "1" ]; then
-    INSTALLED_NVSHMEM=$(pip show nvidia-nvshmem-cu12 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
-    if [ "$INSTALLED_NVSHMEM" = "3.4.5" ]; then
-        echo "nvidia-nvshmem-cu12==3.4.5 already installed, skipping reinstall"
-    else
-        $PIP_CMD install nvidia-nvshmem-cu12==3.4.5 $PIP_INSTALL_SUFFIX
-    fi
-else
-    $PIP_CMD install nvidia-nvshmem-cu12==3.4.5 --force-reinstall $PIP_INSTALL_SUFFIX
-fi
-
-# Cudnn with version less than 9.16.0.29 will cause performance regression on Conv3D kernel
-# On Blackwell machines, skip reinstall if correct version already installed to avoid race conditions
-if [ "$IS_BLACKWELL" = "1" ]; then
-    INSTALLED_CUDNN=$(pip show nvidia-cudnn-cu12 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
-    if [ "$INSTALLED_CUDNN" = "9.16.0.29" ]; then
-        echo "nvidia-cudnn-cu12==9.16.0.29 already installed, skipping reinstall"
-    else
-        $PIP_CMD install nvidia-cudnn-cu12==9.16.0.29 $PIP_INSTALL_SUFFIX
-    fi
-else
-    $PIP_CMD install nvidia-cudnn-cu12==9.16.0.29 --force-reinstall $PIP_INSTALL_SUFFIX
-fi
 $PIP_CMD uninstall xformers || true
 
-# Install flashinfer-jit-cache with caching and retry logic (flashinfer.ai can have transient DNS issues)
-# The jit-cache wheel is 1.2+ GB, so we skip the download entirely if already installed.
-FLASHINFER_INSTALLED=false
-if [ "$UNINSTALL_JIT_CACHE" = false ]; then
-    FLASHINFER_INSTALLED=true
-    echo "flashinfer-jit-cache already at correct version, skipping download"
+mark_step_done "Install extra dependency"
+
+# ------------------------------------------------------------------------------
+# Fix other dependencies
+# ------------------------------------------------------------------------------
+# Fix CUDA version mismatch between torch and torchaudio.
+# PyPI's torch 2.9.1 bundles cu128 but torchaudio from pytorch.org/cu129 uses cu129.
+# This mismatch causes torchaudio's C extension to fail loading, producing:
+#   "partially initialized module 'torchaudio' has no attribute 'lib'"
+# We cannot replace torch with cu129 (breaks sgl_kernel ABI), so instead we reinstall
+# torchaudio/torchvision from an index matching torch's CUDA version.
+TORCH_CUDA_VER=$(python3 -c "import torch; v=torch.version.cuda; parts=v.split('.'); print(f'cu{parts[0]}{parts[1]}')")
+echo "Detected torch CUDA version: ${TORCH_CUDA_VER}"
+if [ "${TORCH_CUDA_VER}" != "${CU_VERSION}" ]; then
+    # Pin versions to match what was installed by pyproject.toml (strip +cuXYZ suffix)
+    TORCHAUDIO_VER=$(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 $2}' | sed 's/+.*//')
+    echo "Reinstalling torchaudio==${TORCHAUDIO_VER} torchvision==${TORCHVISION_VER} from ${TORCH_CUDA_VER} index to match torch..."
+    $PIP_CMD install "torchaudio==${TORCHAUDIO_VER}" "torchvision==${TORCHVISION_VER}" --index-url "https://download.pytorch.org/whl/${TORCH_CUDA_VER}" --force-reinstall --no-deps $PIP_INSTALL_SUFFIX
 fi
 
-if [ "$FLASHINFER_INSTALLED" = false ]; then
-    # Cache directory for flashinfer wheels (persists across CI runs on self-hosted runners)
-    FLASHINFER_CACHE_DIR="${HOME}/.cache/flashinfer-wheels"
-    mkdir -p "${FLASHINFER_CACHE_DIR}"
-
-    # Clean up old versions to avoid cache bloat
-    find "${FLASHINFER_CACHE_DIR}" -name "flashinfer_jit_cache-*.whl" ! -name "flashinfer_jit_cache-${FLASHINFER_VERSION}*" -type f -delete 2>/dev/null || true
-
-    FLASHINFER_WHEEL_PATTERN="flashinfer_jit_cache-${FLASHINFER_VERSION}*.whl"
-    CACHED_WHEEL=$(find "${FLASHINFER_CACHE_DIR}" -name "${FLASHINFER_WHEEL_PATTERN}" -type f 2>/dev/null | head -n 1)
-
-    # Try to install from cache first
-    if [ -n "$CACHED_WHEEL" ] && [ -f "$CACHED_WHEEL" ]; then
-        echo "Found cached flashinfer wheel: $CACHED_WHEEL"
-        if $PIP_CMD install "$CACHED_WHEEL" $PIP_INSTALL_SUFFIX; then
-            FLASHINFER_INSTALLED=true
-            echo "Successfully installed flashinfer-jit-cache from cache"
-        else
-            echo "Failed to install from cache, will try downloading..."
-            rm -f "$CACHED_WHEEL"
-        fi
-    fi
-
-    # If not installed from cache, download with retry logic
-    if [ "$FLASHINFER_INSTALLED" = false ]; then
-        for i in {1..5}; do
-            # Download wheel to cache directory (use pip directly as uv pip doesn't support download)
-            # Timeout after 10 minutes — the wheel is ~1.2 GB
-            if timeout 600 pip download flashinfer-jit-cache==${FLASHINFER_VERSION} \
-                --index-url https://flashinfer.ai/whl/${CU_VERSION} \
-                -d "${FLASHINFER_CACHE_DIR}"; then
-
-                CACHED_WHEEL=$(find "${FLASHINFER_CACHE_DIR}" -name "${FLASHINFER_WHEEL_PATTERN}" -type f 2>/dev/null | head -n 1)
-                if [ -n "$CACHED_WHEEL" ] && [ -f "$CACHED_WHEEL" ]; then
-                    if $PIP_CMD install "$CACHED_WHEEL" $PIP_INSTALL_SUFFIX; then
-                        FLASHINFER_INSTALLED=true
-                        echo "Successfully downloaded and installed flashinfer-jit-cache"
-                        break
-                    fi
-                else
-                    echo "Warning: Download succeeded but wheel file not found"
-                fi
-            fi
-            echo "Attempt $i to download flashinfer-jit-cache failed, retrying in 10 seconds..."
-            sleep 10
-        done
-    fi
+# Fix dependencies: DeepEP depends on nvshmem 3.4.5 — skip reinstall when already correct (avoids pip races / wasted work)
+INSTALLED_NVSHMEM=$(pip show nvidia-nvshmem-cu12 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
+if [ "$INSTALLED_NVSHMEM" = "3.4.5" ]; then
+    echo "nvidia-nvshmem-cu12==3.4.5 already installed, skipping reinstall"
+else
+    $PIP_CMD install nvidia-nvshmem-cu12==3.4.5 $PIP_INSTALL_SUFFIX
 fi
 
-if [ "$FLASHINFER_INSTALLED" = false ]; then
-    echo "ERROR: Failed to install flashinfer-jit-cache after 5 attempts"
-    exit 1
+# Fix dependencies: Cudnn with version less than 9.16.0.29 will cause performance regression on Conv3D kernel
+INSTALLED_CUDNN=$(pip show nvidia-cudnn-cu12 2>/dev/null | grep "^Version:" | awk '{print $2}' || echo "")
+if [ "$INSTALLED_CUDNN" = "9.16.0.29" ]; then
+    echo "nvidia-cudnn-cu12==9.16.0.29 already installed, skipping reinstall"
+else
+    $PIP_CMD install nvidia-cudnn-cu12==9.16.0.29 $PIP_INSTALL_SUFFIX
 fi
 
-# Download flashinfer cubins if the local set is incomplete
-bash "${SCRIPT_DIR}/ci_download_flashinfer_cubin.sh"
+mark_step_done "Fix other dependencies"
 
-# Force reinstall nvidia-cutlass-dsl to avoid potential version mismatch issues
-$PIP_CMD install "nvidia-cutlass-dsl>=4.4.1" "nvidia-cutlass-dsl-libs-base>=4.4.1" --no-deps --force-reinstall $PIP_INSTALL_SUFFIX || true
+# ------------------------------------------------------------------------------
+# Prepare runner
+# ------------------------------------------------------------------------------
+# Prepare the CI runner (cleanup HuggingFace cache, etc.)
+bash "${SCRIPT_DIR}/prepare_runner.sh"
 
+mark_step_done "Prepare runner"
+
+# ------------------------------------------------------------------------------
+# Verify imports
+# ------------------------------------------------------------------------------
 # Show current packages
 $PIP_CMD list
 python3 -c "import torch; print(torch.version.cuda)"
 python3 -c "import cutlass; import cutlass.cute;"
 
-# Prepare the CI runner (cleanup HuggingFace cache, etc.)
-bash "${SCRIPT_DIR}/prepare_runner.sh"
+mark_step_done "Verify imports"
diff --git a/scripts/ci/utils/install_protoc.sh b/scripts/ci/utils/install_protoc.sh
new file mode 100755
index 000000000..5d4bd2891
--- /dev/null
+++ b/scripts/ci/utils/install_protoc.sh
@@ -0,0 +1,53 @@
+#!/bin/bash
+# Ensure protoc is installed for router build (gRPC protobuf compilation).
+set -euxo pipefail
+
+if command -v protoc >/dev/null 2>&1 && protoc --version >/dev/null 2>&1; then
+    echo "protoc already installed: $(protoc --version)"
+    exit 0
+fi
+
+if command -v protoc >/dev/null 2>&1; then
+    echo "protoc found but not runnable, reinstalling..."
+else
+    echo "protoc not found, installing..."
+fi
+
+ARCH=$(uname -m)
+
+if command -v apt-get &> /dev/null; then
+    # Ubuntu/Debian
+    apt-get update || true # May fail due to unrelated broken packages
+    PROTOC_APT_PACKAGES=(wget unzip gcc g++ perl make)
+    apt-get install -y --no-install-recommends "${PROTOC_APT_PACKAGES[@]}" || {
+        echo "Warning: apt-get install failed, checking if required packages are available..."
+        for pkg in "${PROTOC_APT_PACKAGES[@]}"; do
+            if ! dpkg -l "$pkg" 2>/dev/null | grep -q "^ii"; then
+                echo "ERROR: Required package $pkg is not installed and apt-get failed"
+                exit 1
+            fi
+        done
+        echo "All required packages are already installed, continuing..."
+    }
+elif command -v yum &> /dev/null; then
+    # RHEL/CentOS
+    yum update -y
+    yum install -y wget unzip gcc gcc-c++ perl-core make
+else
+    echo "ERROR: Neither apt-get nor yum found; cannot install protoc build deps"
+    exit 1
+fi
+
+if [ "$ARCH" = "aarch64" ] || [ "$ARCH" = "arm64" ]; then
+    PROTOC_ARCH="aarch_64"
+else
+    PROTOC_ARCH="x86_64"
+fi
+PROTOC_ZIP="protoc-32.0-linux-${PROTOC_ARCH}.zip"
+(
+    cd /tmp
+    wget "https://github.com/protocolbuffers/protobuf/releases/download/v32.0/${PROTOC_ZIP}"
+    unzip -o "${PROTOC_ZIP}" -d /usr/local
+    rm -f "${PROTOC_ZIP}"
+)
+protoc --version
diff --git a/scripts/release/bump_flashinfer_version.py b/scripts/release/bump_flashinfer_version.py
index c2ecb33fd..6d873c16b 100755
--- a/scripts/release/bump_flashinfer_version.py
+++ b/scripts/release/bump_flashinfer_version.py
@@ -10,7 +10,6 @@ from utils import compare_versions, get_repo_root, normalize_version, validate_v
 FILES_TO_UPDATE = [
     Path("python/pyproject.toml"),
     Path("docker/Dockerfile"),
-    Path("scripts/ci/cuda/ci_install_dependency.sh"),
     Path("python/sglang/srt/entrypoints/engine.py"),
     Path("python/sglang/srt/utils/common.py"),
 ]
@@ -52,12 +51,6 @@ def replace_flashinfer_version(
             rf"\g<1>{new_version}",
             new_content,
         )
-    elif name == "ci_install_dependency.sh":
-        new_content = re.sub(
-            rf"(FLASHINFER_VERSION=){re.escape(old_version)}",
-            rf"\g<1>{new_version}",
-            new_content,
-        )
     elif name == "engine.py":
         new_content = re.sub(
             r'(assert_pkg_version\(\s*"flashinfer_python",\s*)"'