Update qwentts.cpp build pin

.github/workflows/publish-hf-wheels.yml

@@ -6,7 +6,7 @@ name: Publish Hugging Face Wheels
       qwentts_ref:
         description: qwentts.cpp commit, tag, or branch to build
         required: true
-        default: 9dbe7ea26a01b30fccb117ae5e86807c1dc23d42
+        default: 7df559a8ca25f66fee02970514ebe5f01dee9055
       hf_repo_id:
         description: Hugging Face dataset repo for wheel hosting
         required: true
@@ -17,7 +17,7 @@ permissions:
 
 env:
   HF_WHEEL_REPO_ID: ${{ inputs.hf_repo_id || vars.HF_WHEEL_REPO_ID || 'andito/qwentts-cpp-python-wheels' }}
-  QWENTTS_REF: ${{ inputs.qwentts_ref || '9dbe7ea26a01b30fccb117ae5e86807c1dc23d42' }}
+  QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
 
 jobs:
   build-wheels:

.github/workflows/publish.yml

@@ -9,7 +9,7 @@ name: Publish
       qwentts_ref:
         description: qwentts.cpp commit, tag, or branch to build
         required: true
-        default: 9dbe7ea26a01b30fccb117ae5e86807c1dc23d42
+        default: 7df559a8ca25f66fee02970514ebe5f01dee9055
 
 permissions:
   contents: read
@@ -40,7 +40,7 @@ jobs:
       QWENTTS_CPP_BACKEND: cuda
       QWENTTS_CPP_BUILD_JOBS: ${{ matrix.build_jobs }}
       QWENTTS_CPP_WHEEL_BUILD_TAG: 1cu128
-      QWENTTS_REF: ${{ inputs.qwentts_ref || '9dbe7ea26a01b30fccb117ae5e86807c1dc23d42' }}
+      QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
       CUDA_ARCHITECTURES: ${{ matrix.cuda_architectures }}
     steps:
       - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10 # v6.0.3

.github/workflows/wheels.yml

@@ -1,12 +1,22 @@
 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: 9dbe7ea26a01b30fccb117ae5e86807c1dc23d42
+        default: 7df559a8ca25f66fee02970514ebe5f01dee9055
 
 jobs:
   linux-cuda:
@@ -86,7 +96,7 @@ jobs:
       QWENTTS_CPP_BACKEND: cuda
       QWENTTS_CPP_BUILD_JOBS: ${{ matrix.build_jobs }}
       QWENTTS_CPP_WHEEL_BUILD_TAG: ${{ matrix.wheel_build_tag }}
-      QWENTTS_REF: ${{ inputs.qwentts_ref || '9dbe7ea26a01b30fccb117ae5e86807c1dc23d42' }}
+      QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
       CUDA_ARCHITECTURES: ${{ matrix.cuda_architectures }}
     steps:
       - uses: actions/checkout@v4
@@ -168,7 +178,7 @@ jobs:
       QWENTTS_CPP_BACKEND: cpu
       QWENTTS_CPP_BUILD_JOBS: "2"
       QWENTTS_CPP_WHEEL_BUILD_TAG: 1cpu
-      QWENTTS_REF: ${{ inputs.qwentts_ref || '9dbe7ea26a01b30fccb117ae5e86807c1dc23d42' }}
+      QWENTTS_REF: ${{ inputs.qwentts_ref || '7df559a8ca25f66fee02970514ebe5f01dee9055' }}
     steps:
       - uses: actions/checkout@v4
 

AGENTS.md

@@ -29,7 +29,25 @@ the `HF_TOKEN` repository secret. Do not upload those local-version variants to
 PyPI.
 
 The full Linux wheel matrix is intentionally not run for every pull request.
-Use the manual wheel workflows when validating or publishing wheel artifacts.
+Relevant pull requests run the CPU and CUDA validation matrix automatically.
+Use the publishing workflow for the additional older-Linux variants.
+
+## Updating the qwentts.cpp Pin
+
+To rebuild against a newer qwentts.cpp revision:
+
+1. Resolve the latest upstream `master` commit to its full SHA.
+2. Update every default and event fallback for `QWENTTS_REF` in
+   `.github/workflows/wheels.yml`, `.github/workflows/publish.yml`, and
+   `.github/workflows/publish-hf-wheels.yml`.
+3. Update the pinned revision and its summary in `README.md`.
+4. Open a focused pull request containing the pin and documentation changes.
+   The pull request workflow builds the validation wheel matrix automatically;
+   do not dispatch it manually.
+5. Review the runner results and downloadable wheel artifacts before merging.
+
+Do not include locally generated libraries, wheels, sdists, `build/`, or
+`dist/` contents in the pull request.
 
 CUDA release wheels keep native cubins for sm_86, sm_90, and sm_120, plus PTX
 fallbacks for sm_75 and the newest supported CUDA architecture. This keeps the

README.md

@@ -69,14 +69,13 @@ the same backend flavor. For example, the `cu128` page can host both
 `manylinux_2_35` wheels for Ubuntu 22.04+ and `manylinux_2_39` wheels for
 Ubuntu 24.04+. Pip selects the newest compatible wheel for the current machine.
 
-The full wheel matrix is published manually through the `Publish Hugging Face
-Wheels` workflow. Pull requests do not build every CUDA/Linux variant by
-default.
+Relevant pull requests automatically build the CPU and CUDA validation matrix.
+The additional older-Linux variants are built when publishing the full matrix
+through the `Publish Hugging Face Wheels` workflow.
 
 The CI wheel build defaults to qwentts.cpp
-`9dbe7ea26a01b30fccb117ae5e86807c1dc23d42`, which includes the scheduler
-resets and ABI v2 cached voice-reference fields. CPU and CUDA both stay on the
-backend prompt-projection path.
+`7df559a8ca25f66fee02970514ebe5f01dee9055`, which retains ABI v2 and includes
+the latest static-graph, streaming-decode, and widened voice-route changes.
 
 `QWENTTS_CPP_WHEEL_BUILD_TAG` is useful for local wheelhouses. For public
 indexes, publish one backend flavor per package/version/platform compatibility