Run wheel builds on pull requests

.github/workflows/wheels.yml

@@ -1,24 +1,26 @@
 name: qwentts-cpp-python wheels
 
 "on":
+  pull_request:
+    branches:
+      - main
+    paths:
+      - ".github/workflows/wheels.yml"
+      - "pyproject.toml"
+      - "setup.py"
+      - "scripts/**"
+      - "src/**"
+      - "tests/**"
   workflow_dispatch:
     inputs:
       qwentts_ref:
         description: qwentts.cpp commit, tag, or branch to build
         required: true
-        default: eda8b59092b8d7b7142177c599b08909b87f63b4
-      cuda_architectures:
-        description: CMAKE_CUDA_ARCHITECTURES for portable CUDA binaries
-        required: true
-        default: 75-virtual;80-virtual;86-real;89-real;120a-real;121a-real
-      wheel_build_tag:
-        description: Optional wheel build tag
-        required: true
-        default: 1cu130
+        default: f79b23a56702584d51e3a66509b881cd99a95bba
 
 jobs:
   linux-cuda:
-    name: Linux ${{ matrix.arch }} CUDA wheel
+    name: Linux ${{ matrix.arch }} CUDA ${{ matrix.cuda_version }} wheel
     runs-on: ${{ matrix.runner }}
     strategy:
       fail-fast: false
@@ -27,13 +29,41 @@ jobs:
           - arch: x86_64
             runner: ubuntu-24.04
             cuda_repo_arch: x86_64
+            cuda_version: "12.8"
+            cuda_package_suffix: "12-8"
+            cuda_home: /usr/local/cuda-12.8
+            wheel_build_tag: 1cu128
+            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual
           - arch: aarch64
             runner: ubuntu-24.04-arm
             cuda_repo_arch: sbsa
+            cuda_version: "12.8"
+            cuda_package_suffix: "12-8"
+            cuda_home: /usr/local/cuda-12.8
+            wheel_build_tag: 1cu128
+            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual
+          - arch: x86_64
+            runner: ubuntu-24.04
+            cuda_repo_arch: x86_64
+            cuda_version: "13.0"
+            cuda_package_suffix: "13-0"
+            cuda_home: /usr/local/cuda-13.0
+            wheel_build_tag: 1cu130
+            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual;121-real;121-virtual
+          - arch: aarch64
+            runner: ubuntu-24.04-arm
+            cuda_repo_arch: sbsa
+            cuda_version: "13.0"
+            cuda_package_suffix: "13-0"
+            cuda_home: /usr/local/cuda-13.0
+            wheel_build_tag: 1cu130
+            cuda_architectures: 75-virtual;80-virtual;86-real;89-real;90-real;100-real;120-real;120-virtual;121-real;121-virtual
     env:
       QWENTTS_CPP_BACKEND: cuda
       QWENTTS_CPP_BUILD_JOBS: "2"
-      QWENTTS_CPP_WHEEL_BUILD_TAG: ${{ inputs.wheel_build_tag }}
+      QWENTTS_CPP_WHEEL_BUILD_TAG: ${{ matrix.wheel_build_tag }}
+      QWENTTS_REF: ${{ inputs.qwentts_ref || 'f79b23a56702584d51e3a66509b881cd99a95bba' }}
+      CUDA_ARCHITECTURES: ${{ matrix.cuda_architectures }}
     steps:
       - uses: actions/checkout@v4
 
@@ -58,17 +88,17 @@ jobs:
           sudo dpkg -i cuda-keyring_1.1-1_all.deb
           sudo apt-get update
           sudo apt-get install -y --no-install-recommends \
-            cuda-nvcc-13-0 \
-            cuda-cudart-dev-13-0 \
-            libcublas-dev-13-0
-          echo "/usr/local/cuda-13.0/bin" >> "$GITHUB_PATH"
-          echo "CUDA_HOME=/usr/local/cuda-13.0" >> "$GITHUB_ENV"
-          echo "CMAKE_CUDA_COMPILER=/usr/local/cuda-13.0/bin/nvcc" >> "$GITHUB_ENV"
+            cuda-nvcc-${{ matrix.cuda_package_suffix }} \
+            cuda-cudart-dev-${{ matrix.cuda_package_suffix }} \
+            libcublas-dev-${{ matrix.cuda_package_suffix }}
+          echo "${{ matrix.cuda_home }}/bin" >> "$GITHUB_PATH"
+          echo "CUDA_HOME=${{ matrix.cuda_home }}" >> "$GITHUB_ENV"
+          echo "CMAKE_CUDA_COMPILER=${{ matrix.cuda_home }}/bin/nvcc" >> "$GITHUB_ENV"
 
       - name: Fetch qwentts.cpp
         run: |
           git clone --recursive https://github.com/ServeurpersoCom/qwentts.cpp third_party/qwentts.cpp
-          git -C third_party/qwentts.cpp checkout "${{ inputs.qwentts_ref }}"
+          git -C third_party/qwentts.cpp checkout "$QWENTTS_REF"
           git -C third_party/qwentts.cpp submodule update --init --recursive
 
       - name: Build wheel
@@ -79,12 +109,64 @@ jobs:
             --clean \
             --cmake-arg=-G \
             --cmake-arg=Ninja \
-            --cmake-arg="-DCMAKE_CUDA_ARCHITECTURES=${{ inputs.cuda_architectures }}"
+            --cmake-arg="-DCMAKE_CUDA_ARCHITECTURES=$CUDA_ARCHITECTURES"
+          python -m build --wheel
+          mkdir -p wheelhouse
+          cp dist/*.whl wheelhouse/
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: qwentts-cpp-python-linux-${{ matrix.arch }}-cuda-${{ matrix.wheel_build_tag }}
+          path: wheelhouse/*.whl
+
+  linux-cpu:
+    name: Linux ${{ matrix.arch }} CPU wheel
+    runs-on: ${{ matrix.runner }}
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - arch: x86_64
+            runner: ubuntu-24.04
+          - arch: aarch64
+            runner: ubuntu-24.04-arm
+    env:
+      QWENTTS_CPP_BACKEND: cpu
+      QWENTTS_CPP_BUILD_JOBS: "2"
+      QWENTTS_CPP_WHEEL_BUILD_TAG: 1cpu
+      QWENTTS_REF: ${{ inputs.qwentts_ref || 'f79b23a56702584d51e3a66509b881cd99a95bba' }}
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install build tools
+        run: |
+          python -m pip install --upgrade pip build wheel
+          sudo apt-get update
+          sudo apt-get install -y --no-install-recommends cmake ninja-build patchelf ca-certificates
+
+      - name: Fetch qwentts.cpp
+        run: |
+          git clone --recursive https://github.com/ServeurpersoCom/qwentts.cpp third_party/qwentts.cpp
+          git -C third_party/qwentts.cpp checkout "$QWENTTS_REF"
+          git -C third_party/qwentts.cpp submodule update --init --recursive
+
+      - name: Build wheel
+        run: |
+          python scripts/build_native.py \
+            --source third_party/qwentts.cpp \
+            --backend cpu \
+            --clean \
+            --cmake-arg=-G \
+            --cmake-arg=Ninja
           python -m build --wheel
           mkdir -p wheelhouse
           cp dist/*.whl wheelhouse/
 
       - uses: actions/upload-artifact@v4
         with:
-          name: qwentts-cpp-python-linux-${{ matrix.arch }}-cuda
+          name: qwentts-cpp-python-linux-${{ matrix.arch }}-cpu
           path: wheelhouse/*.whl

README.md

@@ -16,13 +16,28 @@ python scripts/build_native.py \
   --source /path/to/qwentts.cpp \
   --backend cuda \
   --clean
-QWENTTS_CPP_WHEEL_BUILD_TAG=1cu130 python -m build --wheel
+QWENTTS_CPP_WHEEL_BUILD_TAG=1cu128 python -m build --wheel
+```
+
+CPU development build:
+
+```bash
+python scripts/build_native.py \
+  --source /path/to/qwentts.cpp \
+  --backend cpu \
+  --clean
+QWENTTS_CPP_WHEEL_BUILD_TAG=1cpu python -m build --wheel
 ```
 
 `--backend cuda` is the default because `faster-qwen3-tts` is a CUDA-first
 package. CPU builds are still useful for development and smoke tests, but they
 are not the primary release target.
 
+The CI wheel build defaults to qwentts.cpp
+`f79b23a56702584d51e3a66509b881cd99a95bba`, which includes the scheduler reset
+before talker graph allocation. CPU and CUDA both stay on the backend
+prompt-projection path.
+
 `QWENTTS_CPP_WHEEL_BUILD_TAG` is useful for local wheelhouses. For public
 indexes, publish one backend flavor per package/version/platform compatibility
 tag; otherwise pip has no way to choose between CPU and CUDA binaries.

samples/pr-wheel-audio/README.md

@@ -0,0 +1,40 @@
+# PR wheel audio samples
+
+These samples were generated on the local GB10 machine, using cached GGUF
+files and no model downloads. All files were generated from native ref
+`eda8b59092b8d7b7142177c599b08909b87f63b4` plus the scheduler reset that is now
+merged upstream in qwentts.cpp `f79b23a56702584d51e3a66509b881cd99a95bba`.
+
+Common settings:
+
+- Model: `Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice`
+- Quantization: `BF16`
+- Speaker: `aiden`
+- Language: `english`
+- Seed: `42`
+- Sampling settings: `temperature=0.9`, `top_k=50`, `top_p=1.0`,
+  `repetition_penalty=1.05`
+- Streaming chunk: `8 / 12.5` seconds, with `2.0` seconds left context
+
+Files:
+
+| File | Mode | Prompt | Duration |
+| --- | --- | --- | --- |
+| `cpu_greedy_20f_customvoice_bf16.wav` | Greedy, 20 frames | `This is a short deterministic backend comparison.` | 1.60s |
+| `cu128_greedy_20f_customvoice_bf16.wav` | Greedy, 20 frames | `This is a short deterministic backend comparison.` | 1.60s |
+| `cu130_greedy_20f_customvoice_bf16.wav` | Greedy, 20 frames | `This is a short deterministic backend comparison.` | 1.60s |
+| `cpu_sampled_seed42_customvoice_bf16.wav` | Sampled, seed 42 | `Hello from the qwentts cpp Python wheel. This sample compares the same Qwen three custom voice settings across backends.` | 6.40s |
+| `cu128_sampled_seed42_customvoice_bf16.wav` | Sampled, seed 42 | `Hello from the qwentts cpp Python wheel. This sample compares the same Qwen three custom voice settings across backends.` | 6.40s |
+| `cu130_sampled_seed42_customvoice_bf16.wav` | Sampled, seed 42 | `Hello from the qwentts cpp Python wheel. This sample compares the same Qwen three custom voice settings across backends.` | 6.40s |
+
+Validation:
+
+- STT validator: `nano-parakeet` with `nvidia/parakeet-tdt-0.6b-v3`.
+- All six final WAVs produced non-empty, prompt-matching transcripts.
+- Before this patch, the same current-ref CPU wheel transcribed as only
+  `D` / `What uh` on this machine.
+- The broader all-options validation artifacts are linked from the PR comment
+  and include CPU reset-before local, CUDA 13 reset-before local, CUDA 12.8
+  wheel, and CUDA 13.0 wheel runs.
+
+See `manifest.json` for the measured generation metadata and STT transcripts.