diff --git a/.github/workflows/_build_windows_self_host.yml b/.github/workflows/_build_windows_self_host.yml
new file mode 100644
index 0000000..738a259
--- /dev/null
+++ b/.github/workflows/_build_windows_self_host.yml
@@ -0,0 +1,209 @@
+# #########################################################
+# Build wheels with GitHub hosted runner on Windows x86_64
+# #########################################################
+# Prerequisites (must be pre-installed on the runner):
+#   - pwsh
+#   - Git
+#   - Chocolatey
+#   - Visual Studio BuildTools 2022 with:
+#     - Microsoft.VisualStudio.Component.VC.Tools.x86.x64
+#     - Microsoft.VisualStudio.Component.VC.CMake.Project
+#     - Microsoft.VisualStudio.Component.Windows11SDK.22621
+#   - vswhere
+#   - CMake
+#   - Ninja
+#   - Make (optional)
+# #########################################################
+
+name: "[Windows x86_64, self-hosted] Build wheels and upload to GitHub Releases"
+
+on:
+  workflow_call:
+    inputs:
+      flash-attn-version:
+        description: "Flash-Attention version"
+        required: true
+        type: string
+      python-version:
+        description: "Python version"
+        required: true
+        type: string
+      torch-version:
+        description: "PyTorch version"
+        required: true
+        type: string
+      cuda-version:
+        description: "CUDA version"
+        required: true
+        type: string
+      runner:
+        description: "Runner type"
+        required: false
+        type: string
+        default: '["self-hosted", "windows", "x64"]'
+      is-upload:
+        description: "Whether to upload the release asset"
+        required: false
+        type: boolean
+        default: true
+
+jobs:
+  build_windows_wheels_self_hosted:
+    name: Build wheels and Upload (Windows x86_64, self-hosted runner)
+    runs-on: ${{ fromjson(inputs.runner) }}
+    timeout-minutes: 2160
+    env:
+      MAX_JOBS: 2
+      NVCC_THREADS: 2
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Enable Git long paths
+        shell: pwsh
+        run: git config --system core.longpaths true
+
+      # Install Python using uv because setup-python needs newer version of glibc
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+
+      - name: Install Python
+        shell: pwsh
+        run: |
+          uv venv -p ${{ inputs.python-version }}
+          uv pip install -U pip setuptools==75.8.0 wheel packaging psutil numpy ninja
+          $current_dir = (Get-Location).Path
+          echo "$current_dir\.venv\Scripts" >> $env:GITHUB_PATH
+
+      - uses: mjun0812/setup-cuda@v1
+        with:
+          version: ${{ inputs.cuda-version }}
+
+      # Visual Studio BuildTools is pre-installed on the runner
+      - name: Setup MSVC Developer Command Prompt
+        uses: TheMrMilchmann/setup-msvc-dev@v3
+        with:
+          arch: x64
+
+      - name: Add msbuild to PATH
+        uses: microsoft/setup-msbuild@v2
+
+      - name: Build wheels
+        shell: pwsh
+        timeout-minutes: 2160
+        run: |
+          .\build_windows.ps1 -FlashAttnVersion "${{ inputs.flash-attn-version }}" -PythonVersion "${{ inputs.python-version }}" -TorchVersion "${{ inputs.torch-version }}" -CudaVersion "${{ inputs.cuda-version }}"
+          # Script changes to flash-attention directory, wheel is in dist/
+          $wheelName = Get-ChildItem -Path "dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
+          echo "wheel_name=$wheelName" >> $env:GITHUB_ENV
+
+      - name: Install Test
+        shell: pwsh
+        working-directory: flash-attention
+        run: |
+          pip install --no-cache-dir dist/$env:wheel_name
+          python -c "import flash_attn; print(flash_attn.__version__)"
+
+      - name: Upload Release Asset
+        if: ${{ inputs.is-upload }}
+        shell: pwsh
+        working-directory: flash-attention
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+        run: |
+          $tag_name = "${env:GITHUB_REF}".Replace("refs/tags/", "")
+          $wheel_path = "dist/$env:wheel_name"
+          gh release upload "$tag_name" "$wheel_path" --clobber
+
+      # Cleanup step - always runs even if previous steps fail
+      # Only cleans up Python and CUDA installations (VS BuildTools is pre-installed)
+      - name: Cleanup (always run)
+        if: always()
+        shell: pwsh
+        run: |
+          Write-Host "=========================================="
+          Write-Host "Starting cleanup for self-hosted runner..."
+          Write-Host "=========================================="
+
+          # 1. Remove flash-attention directory (source and build artifacts)
+          $flashAttnDir = Join-Path (Get-Location) "flash-attention"
+          if (Test-Path $flashAttnDir) {
+            Write-Host "[1/6] Removing flash-attention directory: $flashAttnDir"
+            Remove-Item -Path $flashAttnDir -Recurse -Force -ErrorAction SilentlyContinue
+          } else {
+            Write-Host "[1/6] flash-attention directory not found, skipping"
+          }
+
+          # 2. Remove Python virtual environment (.venv)
+          $venvDir = Join-Path (Get-Location) ".venv"
+          if (Test-Path $venvDir) {
+            Write-Host "[2/6] Removing Python virtual environment: $venvDir"
+            Remove-Item -Path $venvDir -Recurse -Force -ErrorAction SilentlyContinue
+          } else {
+            Write-Host "[2/6] .venv directory not found, skipping"
+          }
+
+          # 3. Remove pip cache
+          $pipCacheDir = Join-Path $env:LOCALAPPDATA "pip\cache"
+          if (Test-Path $pipCacheDir) {
+            Write-Host "[3/6] Removing pip cache: $pipCacheDir"
+            Remove-Item -Path $pipCacheDir -Recurse -Force -ErrorAction SilentlyContinue
+          } else {
+            Write-Host "[3/6] pip cache not found, skipping"
+          }
+
+          # 4. Remove uv cache
+          $uvCacheDir = Join-Path $env:LOCALAPPDATA "uv"
+          if (Test-Path $uvCacheDir) {
+            Write-Host "[4/6] Removing uv cache: $uvCacheDir"
+            Remove-Item -Path $uvCacheDir -Recurse -Force -ErrorAction SilentlyContinue
+          } else {
+            Write-Host "[4/6] uv cache not found, skipping"
+          }
+
+          # 5. Uninstall CUDA using proper Windows uninstaller
+          Write-Host "[5/6] Uninstalling CUDA packages..."
+          $cudaPackages = Get-Package -Name "*CUDA*" -ErrorAction SilentlyContinue
+          if ($cudaPackages) {
+            foreach ($package in $cudaPackages) {
+              Write-Host "  Uninstalling: $($package.Name)"
+              try {
+                $package | Uninstall-Package -Force -ErrorAction SilentlyContinue
+              } catch {
+                Write-Host "    Warning: Could not uninstall via Package Manager, trying alternative method..."
+              }
+            }
+          } else {
+            Write-Host "  No CUDA packages found in Package Manager"
+          }
+
+          # Also try NVIDIA uninstaller if exists
+          $cudaBaseDir = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA"
+          if (Test-Path $cudaBaseDir) {
+            Get-ChildItem -Path $cudaBaseDir -Directory -ErrorAction SilentlyContinue | ForEach-Object {
+              $uninstaller = Join-Path $_.FullName "Uninstall.exe"
+              if (Test-Path $uninstaller) {
+                Write-Host "  Running CUDA uninstaller: $uninstaller"
+                Start-Process -FilePath $uninstaller -ArgumentList "/S" -Wait -ErrorAction SilentlyContinue
+              }
+            }
+            # Remove remaining directory if exists
+            if (Test-Path $cudaBaseDir) {
+              Write-Host "  Removing remaining CUDA directory: $cudaBaseDir"
+              Remove-Item -Path $cudaBaseDir -Recurse -Force -ErrorAction SilentlyContinue
+            }
+          }
+
+          # 6. Remove temp files
+          Write-Host "[6/6] Removing temporary files"
+          $tempPatterns = @("pip-*", "torch*", "cuda*", "flash*", "uv-*")
+          foreach ($pattern in $tempPatterns) {
+            $tempPath = Join-Path $env:TEMP $pattern
+            Get-ChildItem -Path $tempPath -ErrorAction SilentlyContinue | ForEach-Object {
+              Write-Host "  Removing temp: $($_.FullName)"
+              Remove-Item -Path $_.FullName -Recurse -Force -ErrorAction SilentlyContinue
+            }
+          }
+
+          Write-Host "=========================================="
+          Write-Host "Cleanup completed."
+          Write-Host "=========================================="
diff --git a/.github/workflows/build.yml b/.github/workflows/build.yml
index eb3bdf8..f38fc4b 100644
--- a/.github/workflows/build.yml
+++ b/.github/workflows/build.yml
@@ -126,6 +126,8 @@ jobs:
   # #########################################################
   # Windows
   # #########################################################
+
+  # ################ GitHub-hosted runner ################
   build_wheels_windows:
     name: Build Windows
     needs: [create_releases, create_matrix]
@@ -166,6 +168,30 @@ jobs:
       cuda-version: ${{ matrix.cuda-version }}
     secrets: inherit
 
+  # ################ Self-hosted runner ################
+  build_wheels_windows_self_hosted:
+    name: Build Windows (self-hosted)
+    needs: [create_releases, create_matrix]
+    if: ${{ fromjson(needs.create_matrix.outputs.matrix).windows_self_hosted }}
+    strategy:
+      fail-fast: false
+      matrix:
+        flash-attn-version: ${{ fromjson(needs.create_matrix.outputs.matrix).windows_self_hosted.flash-attn-version }}
+        python-version: ${{ fromjson(needs.create_matrix.outputs.matrix).windows_self_hosted.python-version }}
+        torch-version: ${{ fromjson(needs.create_matrix.outputs.matrix).windows_self_hosted.torch-version }}
+        cuda-version: ${{ fromjson(needs.create_matrix.outputs.matrix).windows_self_hosted.cuda-version }}
+        exclude: ${{ fromjson(needs.create_matrix.outputs.matrix).exclude }}
+    uses: ./.github/workflows/_build_windows_self_hosted.yml
+    with:
+      flash-attn-version: ${{ matrix.flash-attn-version }}
+      python-version: ${{ matrix.python-version }}
+      torch-version: ${{ matrix.torch-version }}
+      cuda-version: ${{ matrix.cuda-version }}
+    secrets: inherit
+
+  # #########################################################
+  # Post-processing: update release notes and docs
+  # #########################################################
   update_release_notes:
     name: Update Release Notes
     needs:
@@ -175,6 +201,7 @@ jobs:
       - build_wheels_linux_arm64_self_hosted
       - build_wheels_windows
       - build_wheels_windows_code_build
+      - build_wheels_windows_self_hosted
     if: always()
     runs-on: ubuntu-latest
     steps:
@@ -204,6 +231,7 @@ jobs:
       - build_wheels_linux_arm64_self_hosted
       - build_wheels_windows
       - build_wheels_windows_code_build
+      - build_wheels_windows_self_hosted
     permissions:
       contents: write
     if: always()
diff --git a/create_matrix.py b/create_matrix.py
index 08c1653..c802e3c 100644
--- a/create_matrix.py
+++ b/create_matrix.py
@@ -195,6 +195,34 @@ WINDOWS_CODEBUILD_MATRIX = {
     ],
 }
 
+WINDOWS_SELF_HOSTED_MATRIX = {
+    "flash-attn-version": [
+        # "2.6.3",
+        # "2.7.4",
+        "2.8.3",
+    ],
+    "python-version": [
+        # "3.10",
+        # "3.11",
+        "3.12",
+        # "3.13",
+    ],
+    "torch-version": [
+        # "2.5.1",
+        # "2.6.0",
+        # "2.7.1",
+        # "2.8.0",
+        "2.9.1",
+    ],
+    "cuda-version": [
+        # "12.4",
+        # "12.6",
+        "12.8",
+        # "12.9",
+        # "13.0",
+    ],
+}
+
 
 def main():
     print(
@@ -206,8 +234,8 @@ def main():
                 "linux_arm64": False,
                 # "linux_arm64": LINUX_ARM64_MATRIX,
                 #
-                # "linux_self_hosted": False,
-                "linux_self_hosted": LINUX_SELF_HOSTED_MATRIX,
+                "linux_self_hosted": False,
+                # "linux_self_hosted": LINUX_SELF_HOSTED_MATRIX,
                 #
                 "linux_arm64_self_hosted": False,
                 # "linux_arm64_self_hosted": LINUX_ARM64_SELF_HOSTED_MATRIX,
@@ -215,6 +243,9 @@ def main():
                 "windows": False,
                 # "windows": WINDOWS_MATRIX,
                 #
+                # "windows_self_hosted": False,
+                "windows_self_hosted": WINDOWS_SELF_HOSTED_MATRIX,
+                #
                 "windows_code_build": False,
                 # "windows_code_build": WINDOWS_CODEBUILD_MATRIX,
                 #