fix: improve Windows self-hosted runner workflow and update docs

- Add pwsh and vswhere to prerequisites list
- Increase timeout to 2160 minutes for long builds
- Improve CUDA cleanup using proper Windows uninstaller
- Update README platform table and manylinux compatibility note

.github/workflows/test-windows-self-hosted.yml

@@ -2,12 +2,14 @@
 # Test build wheels with self-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)
@@ -43,7 +45,7 @@ jobs:
   build_windows_wheels_self_hosted:
     name: Build wheels and Test (Windows x86_64, self-hosted runner)
     runs-on: ["self-hosted", "windows", "x64"]
-    timeout-minutes: 360
+    timeout-minutes: 2160
     env:
       MAX_JOBS: 2
       NVCC_THREADS: 2
@@ -81,6 +83,7 @@ jobs:
 
       - 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 }}"
           $wheelName = Get-ChildItem -Path "flash-attention\dist\*.whl" | Select-Object -First 1 | ForEach-Object { $_.Name }
@@ -138,16 +141,37 @@ jobs:
             Write-Host "[4/6] uv cache not found, skipping"
           }
 
-          # 5. Remove CUDA installation
+          # 5. Uninstall CUDA using proper Windows uninstaller
-          $cudaBaseDir = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA"
+          Write-Host "[5/6] Uninstalling CUDA packages..."
-          if (Test-Path $cudaBaseDir) {
+          $cudaPackages = Get-Package -Name "*CUDA*" -ErrorAction SilentlyContinue
-            Write-Host "[5/6] Removing CUDA installations: $cudaBaseDir"
+          if ($cudaPackages) {
-            Get-ChildItem -Path $cudaBaseDir -Directory -ErrorAction SilentlyContinue | ForEach-Object {
+            foreach ($package in $cudaPackages) {
-              Write-Host "  Removing: $($_.FullName)"
+              Write-Host "  Uninstalling: $($package.Name)"
-              Remove-Item -Path $_.FullName -Recurse -Force -ErrorAction SilentlyContinue
+              try {
+                $package | Uninstall-Package -Force -ErrorAction SilentlyContinue
+              } catch {
+                Write-Host "    Warning: Could not uninstall via Package Manager, trying alternative method..."
+              }
             }
           } else {
-            Write-Host "[5/6] CUDA directory not found, skipping"
+            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

README.md

@@ -45,6 +45,7 @@ pip install ./flash_attn-2.6.3+cu124torch2.5-cp312-cp312-linux_x86_64.whl
 
 > [!NOTE]
 > Since v0.7.0, wheels are built with manylinux2_28 platform.
+> These wheels compatible with old glibc versions (<=2.17).
 
 > [!NOTE]
 > Since v0.5.0, wheels are built with a local version label indicating the CUDA and PyTorch versions.  
@@ -227,10 +228,10 @@ docker compose up -d runner-arm
 
 This repository builds wheels across multiple platforms and environments:
 
-| Platform           | Runner Type                    | Container Image                            |
+| Platform           | Runner Type                        | Container Image                           |
-| ------------------ | ------------------------------ | ------------------------------------------ |
+| ------------------ | ---------------------------------- | ----------------------------------------- |
-| **Linux x86_64**   | GitHub-hosted (`ubuntu-22.04`) | -                                          |
+| **Linux x86_64**   | GitHub-hosted (`ubuntu-22.04`)     | -                                         |
-| **Linux x86_64**   | Self-hosted                    | `ubuntu:22.04` or `manylinux_2_28_x86_64`  |
+| **Linux x86_64**   | Self-hosted                        | `ubuntu:22.04` or `manylinux_2_28_x86_64` |
-| **Linux ARM64**    | Self-hosted (`linux/arm64`)    | `ubuntu:22.04` or `manylinux_2_28_aarch64` |
+| **Linux ARM64**    | GitHub-hosted (`ubuntu-22.04-arm`) | -                                         |
-| **Windows x86_64** | GitHub-hosted (`windows-2022`) | -                                          |
+| **Windows x86_64** | GitHub-hosted (`windows-2022`)     | -                                         |
-| **Windows x86_64** | AWS CodeBuild                  | -                                          |
+| **Windows x86_64** | AWS CodeBuild                      | -                                         |