From 1b14c26b7b774ac3e96d53533ca5474240180c8a Mon Sep 17 00:00:00 2001
From: Andres Marafioti <andimarafioti@gmail.com>
Date: Fri, 25 Sep 2026 09:39:08 +0200
Subject: [PATCH] Publish macOS Metal wheels alongside Linux wheels on PyPI

---
 .github/workflows/metal-wheel.yml | 10 ++++++++--
 .github/workflows/publish.yml     |  8 +++++++-
 AGENTS.md                         | 11 +++++++----
 3 files changed, 22 insertions(+), 7 deletions(-)

diff --git a/.github/workflows/metal-wheel.yml b/.github/workflows/metal-wheel.yml
index f8bde95..e0971a5 100644
--- a/.github/workflows/metal-wheel.yml
+++ b/.github/workflows/metal-wheel.yml
@@ -10,6 +10,10 @@ name: Apple Silicon Metal wheel
         default: 7df559a8ca25f66fee02970514ebe5f01dee9055
   workflow_call:
     inputs:
+      pypi:
+        description: Build a public version for PyPI instead of a Hugging Face local version
+        type: boolean
+        default: false
       qwentts_ref:
         type: string
         default: 7df559a8ca25f66fee02970514ebe5f01dee9055
@@ -39,10 +43,12 @@ jobs:
           git -C third_party/qwentts.cpp fetch --depth 1 origin "$QWENTTS_REF"
           git -C third_party/qwentts.cpp checkout --detach FETCH_HEAD
           git -C third_party/qwentts.cpp submodule update --init --recursive --depth 1
+      - name: Set Hugging Face local version
+        if: ${{ !inputs.pypi }}
+        run: python scripts/set_local_version.py metal
       - name: Build and repair Metal wheel
         run: |
           test "$(uname -m)" = arm64
-          python scripts/set_local_version.py metal
           python scripts/build_native.py --backend metal --clean
           # setup-python may provide universal2 Python; the bundled dylibs are arm64 only.
           python -m build --wheel --config-setting=--build-option=--plat-name=macosx_14_0_arm64
@@ -67,5 +73,5 @@ jobs:
             "from qwentts_cpp import QwenLibrary; lib = QwenLibrary(); print(lib.path); print(lib.version())"
       - uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
         with:
-          name: hf-wheel-metal-macosx-arm64
+          name: ${{ inputs.pypi && 'wheel-metal-macosx-arm64' || 'hf-wheel-metal-macosx-arm64' }}
           path: wheelhouse/*.whl
diff --git a/.github/workflows/publish.yml b/.github/workflows/publish.yml
index 7a70b04..e637e2c 100644
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -15,6 +15,12 @@ permissions:
   contents: read
 
 jobs:
+  build-metal-wheel:
+    uses: ./.github/workflows/metal-wheel.yml
+    with:
+      pypi: true
+      qwentts_ref: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
+
   build-wheels:
     name: Linux ${{ matrix.arch }} CUDA 12.8 wheel
     runs-on: ${{ matrix.runner }}
@@ -110,7 +116,7 @@ jobs:
 
   check:
     name: Check distributions
-    needs: build-wheels
+    needs: [build-wheels, build-metal-wheel]
     runs-on: ubuntu-latest
     steps:
       - uses: astral-sh/setup-uv@fac544c07dec837d0ccb6301d7b5580bf5edae39 # v8.2.0
diff --git a/AGENTS.md b/AGENTS.md
index 271d8be..f8288fe 100644
--- a/AGENTS.md
+++ b/AGENTS.md
@@ -8,11 +8,13 @@
 
 PyPI publishing is handled by GitHub Actions in `.github/workflows/publish.yml`.
 The workflow runs on pushed tags that match `v*`, builds the public Linux CUDA
-12.8 wheels, checks the artifacts with `twine check --strict`, and publishes
+12.8 and macOS Metal wheels, checks the artifacts with `twine check --strict`, and publishes
 through the configured `pypi` environment.
 
-The public `qwentts-cpp-python` package currently publishes CUDA 12.8 wheels
-only. Publishing multiple backend flavors under the same package name, version,
+The public `qwentts-cpp-python` package publishes CUDA 12.8 wheels for Linux
+and Metal wheels for macOS 14+ arm64, using plain public versions for both.
+Their platform tags let pip select the appropriate wheel automatically.
+Publishing multiple backend flavors under the same package name, version,
 and platform tag would leave pip with no reliable way to choose the intended
 runtime.
 
@@ -31,7 +33,8 @@ PyPI.
 The `+metal` flavor is a macOS 14+ arm64 wheel with embedded Metal shaders and
 bundled native dylibs. The reusable `.github/workflows/metal-wheel.yml` builds,
 repairs, validates ABI layouts, and tests a clean wheel install. It runs for PRs
-and is also called by validation and Hugging Face publishing. GPU synthesis is
+and is also called by validation, PyPI publishing (with `pypi: true` to omit
+the local version suffix), and Hugging Face publishing. GPU synthesis is
 tested locally with `scripts/smoke_stream.py --require-metal` and local weights.
 
 Pull requests do not run the Linux wheel matrix. Do not run the validation