feat(cuda): add CUDA 13.3 Support

.github/workflows/cuda13-windows.yml

@@ -0,0 +1,104 @@
+name: cuda13-windows
+
+# Builds + packaging-smokes the transcribe-cpp-native-cu13 wheel for
+# win_amd64. Same design as the Windows CUDA 12 wheel:
+#   - the CUDA runtime is NEVER vendored: cudart/cublas come from the
+#     nvidia-*-cu13 runtime wheels (win_amd64 wheels exist — the PyTorch
+#     layout), preloaded by the package's prepare() hook, which knows the
+#     Windows DLL names (cudart64_13.dll, cublasLt64_13.dll,
+#     cublas64_13.dll under nvidia/<pkg>/bin/); nvcuda.dll is the driver.
+#   - the ggml-vulkan module is bundled alongside CUDA (superset-of-default
+#     rationale in bindings/python-native-cu13/pyproject.toml: a CUDA
+#     install on an AMD/Intel-GPU Windows box must not silently lose Vulkan
+#     acceleration), built against the LunarG SDK like wheel-windows.
+#   - delvewheel repair EXCLUDES the CUDA 13 DLLs + the driver + vulkan-1.dll,
+#     mirroring auditwheel's excludes on Linux.
+#   - GPU-less smoke: cu13 provider selected, declared backends
+#     {cuda, vulkan, cpu}, the cuda/vulkan modules quietly absent without a
+#     driver, CPU transcribes, full pytest suite.
+#
+# Trimmed cadence: NOT per PR push. Releases (tags) and manual dispatch only
+# — nvcc over ggml-cuda × 5 arches is the heaviest build in the project.
+#
+# CUDA 13.3 toolkit.
+
+on:
+  # Called by publish.yml on release tags.
+  workflow_call:
+  workflow_dispatch:
+
+env:
+  PYTHONUTF8: "1"  # hf CLI prints ✓; Windows cp1252 console chokes
+
+jobs:
+  cuda13-wheel-windows:
+    # 16vcpu: nvcc over ggml-cuda x 5 arches under MSVC.
+    runs-on: blacksmith-16vcpu-windows-2025
+    timeout-minutes: 180
+    env:
+      HF_TOKEN: ${{ secrets.HF_TOKEN }}
+      CMAKE_GENERATOR: Ninja
+      # LunarG prunes old SDK downloads — when bumping, verify the URL exists.
+      # Keep in lockstep with wheel-windows (python-wheels.yml) and
+      # rust-windows-deep-path (rust-ci.yml).
+      VULKAN_VERSION: "1.4.350.0"
+    steps:
+      - uses: actions/checkout@v6
+      # vcvars for the whole job: nvcc needs cl.exe, and the Ninja generator
+      # needs the MSVC environment.
+      - uses: ilammy/msvc-dev-cmd@v1
+      - name: Install Vulkan SDK ${{ env.VULKAN_VERSION }} (glslc for the ggml-vulkan module)
+        run: |
+          curl.exe -o "$env:RUNNER_TEMP\vulkan_sdk.exe" -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkan_sdk.exe"
+          & "$env:RUNNER_TEMP\vulkan_sdk.exe" --accept-licenses --default-answer --confirm-command install
+          Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
+          Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
+      - name: Install CUDA toolkit 13.3 (network installer, lean subset)
+        uses: Jimver/cuda-toolkit@v0.2.35
+        with:
+          cuda: "13.3.0"
+          method: network
+          use-github-cache: false
+          use-local-cache: false
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
+      - uses: astral-sh/setup-uv@v8.2.0
+      - uses: ./.github/actions/fetch-canary
+        with:
+          hf-token: ${{ secrets.HF_TOKEN }}
+      - name: Build the raw cu13 wheel (scikit-build-core; lane posture in
+              bindings/python-native-cu13/pyproject.toml)
+        run: uv build --wheel bindings/python-native-cu13 --out-dir dist-cu13-raw
+      - name: Repair (delvewheel; CUDA runtime + driver + Vulkan loader stay OUT)
+        run: |
+          uvx delvewheel repair `
+            --exclude cudart64_13.dll `
+            --exclude cublas64_13.dll `
+            --exclude cublasLt64_13.dll `
+            --exclude nvcuda.dll `
+            --exclude vulkan-1.dll `
+            -w wheelhouse-cu13 (Get-Item dist-cu13-raw/*.whl).FullName
+          Get-ChildItem wheelhouse-cu13
+      - name: GPU-less packaging smoke (mirrors Modal's packaging_check)
+        run: |
+          uv venv --seed --python 3.12 smoke-venv
+          smoke-venv/Scripts/pip install -q --find-links wheelhouse-cu13 `
+            transcribe-cpp-native-cu13 "pytest>=7" numpy huggingface_hub
+          $env:TRANSCRIBE_SMOKE_PROVIDER = "transcribe-cpp-native-cu13"
+          # The default provider is deliberately absent (its pin can't
+          # resolve pre-release); wheel_smoke installs the API package
+          # dep-free. Declared-backends gate asserts {cuda, vulkan, cpu}.
+          $env:TRANSCRIBE_SMOKE_PIP_NO_DEPS = "1"
+          $env:CI = "1"
+          smoke-venv/Scripts/python scripts/ci/wheel_smoke.py .
+      - uses: actions/upload-artifact@v7
+        with:
+          name: cuda-dist-windows-x86_64-cu13
+          path: wheelhouse-cu13/*.whl
+      - name: Extract the native bundle
+        run: |
+          python scripts/ci/extract_native_bundle.py --wheel-dir wheelhouse-cu13 `
+            --tuple windows-x86_64-cuda --out bundles
+      - uses: actions/upload-artifact@v7
+        with:
+          name: native-windows-x86_64-cuda-cu13
+          path: bundles/*.tar.gz

.github/workflows/wheel-index.yml

@@ -1,12 +1,15 @@
 name: wheel-index
 
 # PEP 503 "simple repository" index on GitHub Pages for wheels that can't ship
-# on PyPI: the cu12 CUDA provider (~197 MB, over PyPI's 100 MB cap) and future
-# Windows-CUDA / cu13 / ROCm providers. GitHub hosts the wheels (as release
-# assets); this workflow generates the few KB of HTML that links to them and
-# publishes it to Pages. Consumed as:
+# on PyPI: the cu12/cu13 CUDA providers (~197 MB, over PyPI's 100 MB cap) and
+# future ROCm providers. GitHub hosts the wheels (as release assets); this
+# workflow generates the few KB of HTML that links to them and publishes it to
+# Pages. Consumed as:
 #   pip install "transcribe-cpp[cu12]" \
 #     --extra-index-url https://<owner>.github.io/<repo>/whl/cu12
+#   # or for CUDA 13:
+#   pip install "transcribe-cpp[cu13]" \
+#     --extra-index-url https://<owner>.github.io/<repo>/whl/cu13
 #
 # ONE-TIME PREREQ: repo Settings -> Pages -> Source = "GitHub Actions".
 #

bindings/python-native-cu12/transcribe_cpp_native_cu12/__init__.py

@@ -27,12 +27,18 @@
 #: DLLs under <pkg>/bin/ (the layout PyTorch relies on).
 if sys.platform == "win32":
     _NVIDIA_LIBS = (
+        "cuda_runtime/bin/cudart64_13.dll",
+        "cublas/bin/cublasLt64_13.dll",
+        "cublas/bin/cublas64_13.dll",
         "cuda_runtime/bin/cudart64_12.dll",
         "cublas/bin/cublasLt64_12.dll",
         "cublas/bin/cublas64_12.dll",
     )
 else:
     _NVIDIA_LIBS = (
+        "cuda_runtime/lib/libcudart.so.13",
+        "cublas/lib/libcublasLt.so.13",
+        "cublas/lib/libcublas.so.13",
         "cuda_runtime/lib/libcudart.so.12",
         "cublas/lib/libcublasLt.so.12",
         "cublas/lib/libcublas.so.12",

bindings/typescript/src/cuda.ts

@@ -30,11 +30,17 @@ import koffi from "koffi";
 const RUNTIME_LIBS: ReadonlyArray<readonly [string, string, string]> =
   process.platform === "win32"
     ? [
+        ["cuda_runtime", "bin", "cudart64_13.dll"],
+        ["cublas", "bin", "cublasLt64_13.dll"],
+        ["cublas", "bin", "cublas64_13.dll"],
         ["cuda_runtime", "bin", "cudart64_12.dll"],
         ["cublas", "bin", "cublasLt64_12.dll"],
         ["cublas", "bin", "cublas64_12.dll"],
       ]
     : [
+        ["cuda_runtime", "lib", "libcudart.so.13"],
+        ["cublas", "lib", "libcublasLt.so.13"],
+        ["cublas", "lib", "libcublas.so.13"],
         ["cuda_runtime", "lib", "libcudart.so.12"],
         ["cublas", "lib", "libcublasLt.so.12"],
         ["cublas", "lib", "libcublas.so.12"],

scripts/ci/build_wheel_index.py

@@ -4,20 +4,23 @@
 """Generate a PEP 503 "simple repository" index from this repo's GitHub
 release assets, for wheels too large or too specialized for PyPI.
 
-The cu12 CUDA provider is ~197 MB — over PyPI's 100 MB cap — so it cannot live
-on PyPI; future Windows-CUDA / cu13 / ROCm providers ride the same index.
+The cu12 and cu13 CUDA providers are ~197 MB — over PyPI's 100 MB cap — so they
+cannot live on PyPI; future ROCm providers ride the same index.
 GitHub hosts the wheels (as release assets); GitHub Pages hosts the few KB of
 HTML this produces. pip/uv consume it as:
 
     pip install "transcribe-cpp[cu12]" \
         --extra-index-url https://<owner>.github.io/<repo>/whl/cu12
+    # or for CUDA 13:
+    pip install "transcribe-cpp[cu13]" \
+        --extra-index-url https://<owner>.github.io/<repo>/whl/cu13
 
 Modeled on llama-cpp-python's releases-to-pep-503.sh, with one simplification:
-our flavors are distinguished by PACKAGE NAME (transcribe-cpp-native-cu12), not
-by a build-tag suffix on the release, so a single release can carry every
-flavor's wheels and we just filter assets by distribution name. No sha256 in
-the URLs (would mean downloading every 197 MB wheel each run — the same
-trade-off llama-cpp-python makes; assets are served over HTTPS from GitHub).
+our flavors are distinguished by PACKAGE NAME (e.g. transcribe-cpp-native-cu12,
+transcribe-cpp-native-cu13), not by a build-tag suffix on the release, so a single
+release can carry every flavor's wheels and we just filter assets by distribution
+name. No sha256 in the URLs (would mean downloading every 197 MB wheel each run —
+the same trade-off llama-cpp-python makes; assets are served over HTTPS from GitHub).
 
 Usage (CI):
     GITHUB_REPOSITORY=owner/repo GITHUB_TOKEN=*** \
@@ -35,10 +38,11 @@
 
 #: flavor (index subdir) -> provider packages whose wheels live under it.
 #: A wheel is matched to a flavor purely by its distribution name, so the same
-#: flavor holds every platform tag of that package (linux + windows cu12).
+#: flavor holds every platform tag of that package (linux + windows cu12/cu13).
 FLAVORS: dict[str, list[str]] = {
     "cu12": ["transcribe-cpp-native-cu12"],
-    # future: "cu13": ["transcribe-cpp-native-cu13"], "rocm": [...]
+    "cu13": ["transcribe-cpp-native-cu13"],
+    # future: "rocm": [...]
 }
 
 _API = "https://api.github.com"