Merge pull request #64 from amd/add-windows-wheels-release

erwei-xilinx · web-flow · commit 93c8cc64e30c · 2026-05-08T16:41:10.000-07:00
Add Windows wheel release job to nightly-wheels CI
diff --git a/.github/workflows/nightly-wheels.yml b/.github/workflows/nightly-wheels.yml
@@ -139,11 +139,192 @@ jobs:
           name: triton-wheel-py${{ matrix.python_version }}
           path: wheelhouse/*.whl
 
-      - name: Release wheels
-        if: |
-          github.event_name == 'workflow_dispatch' ||
-          github.event_name == 'schedule' ||
-          (github.event_name == 'push' && startsWith(github.ref, 'refs/tags/'))
+  build-wheels-windows:
+    name: Build triton wheel Windows (Python ${{ matrix.python_version }})
+    needs: [check-changes]
+    if: >-
+      always() &&
+      (needs.check-changes.result == 'skipped' ||
+       needs.check-changes.outputs.has_new_commits == 'true')
+
+    runs-on: windows-latest
+
+    permissions:
+      id-token: write
+      contents: write
+      packages: read
+
+    strategy:
+      fail-fast: false
+      matrix:
+        # Python matrix is narrower than Linux: Xilinx publishes mlir-air
+        # Windows wheels only for cp310/cp311/cp312 (no cp313/cp314).
+        python_version: ["3.10", "3.11", "3.12"]
+
+    env:
+      # Pinned XRT Windows SDK release. Provides headers, xrt_coreutil.lib,
+      # and xclbinutil/aiebu-asm. The build expects the SDK at
+      # C:\Program Files\AMD\xrt (see utils/env_setup.ps1).
+      XRT_WINDOWS_SDK_URL: "https://github.com/Xilinx/XRT/releases/download/2.21.75/xrt_windows_sdk.zip"
+
+    steps:
+      - name: Disable Git CRLF conversion
+        # Must run before checkout. windows-latest defaults to
+        # core.autocrlf=true, which rewrites text files to CRLF on
+        # checkout. The third_party/triton_shared.patch hunks were
+        # generated with LF context lines, so the rewrite makes
+        # `git apply --check` fail and apply_patches.py aborts —
+        # the build then dies with a C2397 narrowing-conversion
+        # error in PtrAnalysis.cpp that the patch was supposed to
+        # fix. Forcing LF avoids the conversion entirely.
+        shell: bash
+        run: |
+          git config --global core.autocrlf false
+          git config --global core.eol lf
+
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 2
+          submodules: recursive
+
+      - name: Set up Python ${{ matrix.python_version }}
+        uses: actions/setup-python@v5
+        with:
+          python-version: ${{ matrix.python_version }}
+
+      - name: Get commit info
+        id: commit-info
+        run: |
+          if [ -n "${{ inputs.TRITON_XDNA_COMMIT }}" ]; then
+            COMMIT="${{ inputs.TRITON_XDNA_COMMIT }}"
+          else
+            COMMIT=$(git rev-parse --short=7 HEAD)
+          fi
+          echo "commit=$COMMIT" >> $GITHUB_OUTPUT
+          echo "datetime=$(date +%Y%m%d%H)" >> $GITHUB_OUTPUT
+          echo "Building triton-xdna commit: $COMMIT"
+
+      - name: Install XRT Windows SDK
+        shell: pwsh
+        run: |
+          Write-Host "Downloading $env:XRT_WINDOWS_SDK_URL"
+          Invoke-WebRequest -Uri $env:XRT_WINDOWS_SDK_URL -OutFile xrt_windows_sdk.zip
+          # Zip layout is xrt_sdk/xrt/{include,lib,...}; extract to a temp
+          # location and move the inner xrt/ folder to where the build
+          # expects it (C:\Program Files\AMD\xrt).
+          $tempDir = Join-Path $env:RUNNER_TEMP "xrt_extract"
+          New-Item -ItemType Directory -Force -Path $tempDir | Out-Null
+          Expand-Archive -Path xrt_windows_sdk.zip -DestinationPath $tempDir -Force
+          $srcXrt = Join-Path $tempDir "xrt_sdk\xrt"
+          if (-not (Test-Path $srcXrt)) {
+            Write-Error "Unexpected SDK zip layout: $srcXrt not found"
+            exit 1
+          }
+          New-Item -ItemType Directory -Force -Path "C:\Program Files\AMD" | Out-Null
+          Move-Item -Path $srcXrt -Destination "C:\Program Files\AMD\xrt" -Force
+          $hdr = "C:\Program Files\AMD\xrt\include\xrt\xrt_bo.h"
+          if (-not (Test-Path $hdr)) {
+            Write-Error "XRT SDK extraction failed: $hdr not found"
+            exit 1
+          }
+          Write-Host "XRT SDK installed at C:\Program Files\AMD\xrt"
+          Remove-Item xrt_windows_sdk.zip
+          Remove-Item -Recurse -Force $tempDir
+
+      - name: Set up MSVC developer environment
+        # cibuildwheel doesn't activate vcvars on Windows; without this step
+        # cmake fails with "Could not find compiler set in environment
+        # variable CXX: cl.exe" because PATH/INCLUDE/LIB aren't populated.
+        uses: ilammy/msvc-dev-cmd@v1
+        with:
+          arch: x64
+
+      - name: Install cibuildwheel
+        # Use `python -m pip` instead of `pip` because Windows can't replace
+        # the running pip.exe wrapper while it holds the file open.
+        run: |
+          python -m pip install --upgrade pip
+          python -m pip install cibuildwheel
+
+      - name: Build wheels with cibuildwheel
+        env:
+          TRITON_XDNA_PROJECT_COMMIT: ${{ steps.commit-info.outputs.commit }}
+          DATETIME: ${{ steps.commit-info.outputs.datetime }}
+          XILINX_XRT: "C:\\Program Files\\AMD\\xrt"
+        run: |
+          # Convert python version (e.g., "3.11" -> "cp311")
+          PY_VERSION="${{ matrix.python_version }}"
+          CIBW_BUILD="cp${PY_VERSION//./}-win_amd64"
+          echo "Building for: $CIBW_BUILD"
+
+          CIBW_BUILD="$CIBW_BUILD" cibuildwheel --platform windows --output-dir wheelhouse
+
+      - name: List built wheels
+        run: |
+          ls -la wheelhouse/
+          echo "Built wheels:"
+          ls wheelhouse/*.whl
+
+      - name: Upload wheel artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: triton-wheel-windows-py${{ matrix.python_version }}
+          path: wheelhouse/*.whl
+
+  publish-release:
+    name: Publish wheels to GitHub release
+    needs: [build-wheels, build-wheels-windows]
+    # Run after both build matrices finish, regardless of partial failures —
+    # publish whatever wheels did build. Skipped on pull_request / merge_group
+    # to avoid mutating release tags from non-mainline events.
+    if: |
+      always() &&
+      (
+        github.event_name == 'workflow_dispatch' ||
+        github.event_name == 'schedule' ||
+        (github.event_name == 'push' && startsWith(github.ref, 'refs/tags/'))
+      ) &&
+      (
+        needs.build-wheels.result == 'success' ||
+        needs.build-wheels-windows.result == 'success'
+      )
+
+    runs-on: ubuntu-latest
+
+    permissions:
+      contents: write
+
+    steps:
+      - name: Checkout repository
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 1
+
+      - name: Get commit info
+        id: commit-info
+        run: |
+          if [ -n "${{ inputs.TRITON_XDNA_COMMIT }}" ]; then
+            COMMIT="${{ inputs.TRITON_XDNA_COMMIT }}"
+          else
+            COMMIT=$(git rev-parse --short=7 HEAD)
+          fi
+          echo "commit=$COMMIT" >> $GITHUB_OUTPUT
+          echo "datetime=$(date +%Y%m%d%H)" >> $GITHUB_OUTPUT
+
+      - name: Download all wheel artifacts
+        uses: actions/download-artifact@v4
+        with:
+          pattern: triton-wheel-*
+          merge-multiple: true
+          path: wheelhouse
+
+      - name: List collected wheels
+        run: |
+          ls -la wheelhouse/
+          echo "Wheel count: $(ls wheelhouse/*.whl 2>/dev/null | wc -l)"
+
+      - name: Publish release
         uses: ncipollo/release-action@v1.12.0
         with:
           artifacts: wheelhouse/*.whl
@@ -155,9 +336,11 @@ jobs:
 
             **Commit:** ${{ steps.commit-info.outputs.commit }}
             **Build Date:** ${{ steps.commit-info.outputs.datetime }}
-            **Python Version:** ${{ matrix.python_version }}
 
-            ### Installation
+            Linux wheels: Python 3.10–3.14 (manylinux_x86_64)
+            Windows wheels: Python 3.10–3.12 (win_amd64)
+
+            ### Installation (Linux)
 
             ```bash
             pip install triton-xdna \
@@ -166,6 +349,16 @@ jobs:
               --find-links https://github.com/Xilinx/llvm-aie/releases/expanded_assets/nightly \
               --find-links https://github.com/Xilinx/mlir-air/releases/expanded_assets/latest-air-wheels-no-rtti
             ```
+
+            ### Installation (Windows)
+
+            ```powershell
+            pip install triton-xdna `
+              --find-links https://github.com/${{ github.repository }}/releases/expanded_assets/latest-wheels `
+              --find-links https://github.com/Xilinx/mlir-aie/releases/expanded_assets/latest-wheels-no-rtti `
+              --find-links https://github.com/Xilinx/llvm-aie/releases/expanded_assets/nightly `
+              --find-links https://github.com/Xilinx/mlir-air/releases/expanded_assets/latest-air-wheels-no-rtti
+            ```
           allowUpdates: true
           replacesArtifacts: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/') }}
           makeLatest: ${{ github.event_name == 'push' }}
diff --git a/README.md b/README.md
@@ -129,7 +129,7 @@ compilation pipeline (Triton → MLIR → xclbin → XRT dispatch) runs natively
 
 - **Windows 10/11** (x64)
 - **Visual Studio 2022** with "Desktop development with C++" workload
-- **Python 3.12+**
+- **Python 3.10, 3.11, or 3.12** (Xilinx Windows wheels do not yet support 3.13+)
 - **CMake 3.20+** and **Ninja** (via pip or standalone)
 - **AMD NPU driver** (installs `xrt_coreutil.dll` runtime)
 
@@ -147,52 +147,45 @@ pip install --upgrade pip setuptools wheel
 
 Prepare XRT development files (headers, import library, xclbinutil). Download
 `xrt_windows_sdk.zip` from [Xilinx/XRT releases](https://github.com/Xilinx/XRT/releases)
-and extract the `xrt/` directory to `C:\Program Files\AMD\xrt`:
+and extract the inner `xrt_sdk/xrt/` directory (note the zip's top-level
+folder is `xrt_sdk/`) to `C:\Program Files\AMD\xrt`:
 
 ```powershell
-# The xrt/ folder inside the zip should end up at:
+# The contents of xrt_sdk/xrt/ inside the zip should end up at:
 #   C:\Program Files\AMD\xrt\include\xrt\xrt_bo.h
 #   C:\Program Files\AMD\xrt\lib\xrt_coreutil.lib
 ```
 
-Run the automated build:
+Run the automated environment setup (must be dot-sourced so PATH/env vars
+persist in the current shell):
 
 ```powershell
-.\utils\build_windows.ps1
+. .\utils\env_setup.ps1
 ```
 
-This installs pre-built wheels (triton-windows, mlir-aie, llvm-aie), builds mlir-air
-from source, and installs the Triton-XDNA backend. Takes approximately 30–60 minutes.
+This installs the pre-built wheels (`triton-windows`, `mlir-air[aie]` which
+transitively pulls `mlir-aie` and `llvm-aie`) and the Triton-XDNA backend.
 
 ### Windows Manual Build
 
+Install build tools, PyTorch, and the MLIR-AIE/AIR/LLVM-AIE stack. The
+`mlir_air[aie]` extra transitively pins matching `mlir-aie` and pulls
+`llvm-aie`, so a single resolver pass installs the whole stack from the
+Xilinx release pages:
+
 ```powershell
 pip install cmake ninja lit numpy PyYAML nanobind scipy
 pip install torch --index-url https://download.pytorch.org/whl/cpu
 pip install triton-windows
-pip install mlir-aie -f https://github.com/Xilinx/mlir-aie/releases/expanded_assets/latest-wheels-no-rtti
-pip install llvm-aie -f https://github.com/Xilinx/llvm-aie/releases/expanded_assets/nightly
+pip install "mlir_air[aie]" `
+  -f https://github.com/Xilinx/mlir-air/releases/expanded_assets/latest-air-wheels-no-rtti `
+  -f https://github.com/Xilinx/mlir-aie/releases/expanded_assets/latest-wheels-no-rtti `
+  -f https://github.com/Xilinx/llvm-aie/releases/expanded_assets/nightly
 ```
 
-mlir-air must be built from source (no Windows wheels yet):
-
-```powershell
-git clone https://github.com/Xilinx/mlir-air.git
-cd mlir-air
-git checkout <commit-from-utils/mlir-air-hash.txt>
-git submodule update --init --recursive
-
-cmake -G Ninja -DCMAKE_BUILD_TYPE=Release `
-  -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=cl `
-  -DMLIR_DIR=<mlir-distro>/lib/cmake/mlir `
-  -DLLVM_DIR=<mlir-distro>/lib/cmake/llvm `
-  -DAIE_DIR=<mlir-aie-python-pkg>/lib/cmake/aie `
-  -DLLVM_ENABLE_RTTI=OFF -DBUILD_SHARED_LIBS=OFF `
-  -DAIR_RUNTIME_TARGETS="" -DAIR_ENABLE_GPU=OFF `
-  -B build -S .
-ninja -C build -j $env:NUMBER_OF_PROCESSORS
-ninja -C build install
-```
+To pin a specific mlir-air version, use the values from
+`utils/mlir-air-hash.txt`:
+`mlir_air[aie]==<Version>.<Timestamp>+<short-commit>.no.rtti`.
 
 Install Triton-XDNA:
 
@@ -228,6 +221,7 @@ python vec-add.py
 
 ### Windows Known Limitations
 
-- mlir-air must be built from source (no Windows wheels published)
+- Python 3.10, 3.11, and 3.12 only — Xilinx does not publish `mlir-air` /
+  `mlir-aie` Windows wheels for 3.13+ yet
 - xclbinutil and aiebu-asm must be on PATH (from XRT Windows SDK)
 - NPU driver must be installed
diff --git a/scripts/apply_patches.py b/scripts/apply_patches.py
@@ -77,8 +77,11 @@ def check_patch_applicable(patch_file: Path, target_dir: Path) -> tuple[bool, st
     Returns:
         (can_apply, reason) tuple
     """
+    # --ignore-whitespace tolerates CRLF/LF differences in context lines,
+    # which matters on Windows where git checkout converts text files to
+    # CRLF by default but our patches were generated with LF.
     result = run_git(
-        ["apply", "--check", str(patch_file)],
+        ["apply", "--check", "--ignore-whitespace", str(patch_file)],
         cwd=target_dir,
         check=False,
     )
@@ -88,7 +91,7 @@ def check_patch_applicable(patch_file: Path, target_dir: Path) -> tuple[bool, st
 
     # Check if patch is already applied by trying reverse
     result_reverse = run_git(
-        ["apply", "--check", "--reverse", str(patch_file)],
+        ["apply", "--check", "--reverse", "--ignore-whitespace", str(patch_file)],
         cwd=target_dir,
         check=False,
     )
@@ -102,7 +105,7 @@ def check_patch_applicable(patch_file: Path, target_dir: Path) -> tuple[bool, st
 def apply_patch(patch_file: Path, target_dir: Path) -> bool:
     """Apply a patch file to the target directory."""
     try:
-        run_git(["apply", str(patch_file)], cwd=target_dir)
+        run_git(["apply", "--ignore-whitespace", str(patch_file)], cwd=target_dir)
         return True
     except subprocess.CalledProcessError as e:
         print(f"  ✗ Failed to apply patch: {e.stderr}", file=sys.stderr)
diff --git a/setup.py b/setup.py
@@ -380,7 +380,14 @@ def download_llvm_for_triton_windows(triton_dir: Path) -> Path:
 
         print("  Extracting...")
         with tarfile.open(download_path, "r:gz") as tar:
-            tar.extractall(triton_dir.parent, filter="data")
+            # filter="data" requires Python 3.12+ (PEP 706) or a backport
+            # patch release (3.10.12+, 3.11.4+). cibuildwheel's bundled
+            # nuget-cpython for 3.10/3.11 isn't always a backported version,
+            # so guard the kwarg.
+            if sys.version_info >= (3, 12):
+                tar.extractall(triton_dir.parent, filter="data")
+            else:
+                tar.extractall(triton_dir.parent)
 
         if not llvm_dir.exists():
             raise RuntimeError(f"Extracted LLVM directory not found: {llvm_dir}")
@@ -701,12 +708,21 @@ def _rename_package(self, unpack_dir: Path):
                     new_lines.append("Name: triton-xdna\n")
                 elif line.startswith("Version: "):
                     new_lines.append(f"Version: {our_version}\n")
+                elif line.startswith("Author: "):
+                    new_lines.append("Author: Erwei Wang\n")
+                elif line.startswith("Author-email: "):
+                    new_lines.append("Author-email: erwei.wang@amd.com\n")
+                elif line.startswith("Home-page: "):
+                    new_lines.append("Home-page: https://github.com/amd/Triton-XDNA\n")
                 else:
                     new_lines.append(line)
 
             with open(metadata_path, "w") as f:
                 f.writelines(new_lines)
-            print("  Updated Name and Version in METADATA", file=sys.stderr)
+            print(
+                "  Updated Name/Version/Author/Author-email/Home-page in METADATA",
+                file=sys.stderr,
+            )
 
         # Update WHEEL file if needed (usually doesn't need changes)
         wheel_path = new_dist_info / "WHEEL"
diff --git a/utils/env_setup.ps1 b/utils/env_setup.ps1
