ci(cu131+linux): build CU131 wheels with GGML dynamic backends for linux

- Replace the old CPU/AVX release tag matrix with a single CU131 backend
  wheel layout.
- Enable `GGML_BACKEND_DL` and `GGML_CPU_ALL_VARIANTS` so Linux wheels ship
  runtime-loadable GGML backend DLLs and CPU variant backends.
- Disable non-wheel targets such as examples, tests, tools, server, embedded UI, and curl.
- Remove the `.basic` style local version suffix and publish wheels
  as `+cu131`.
- Update CUDA architectures to CUDA 13.1 and simplify CMake argument handling.

Signed-off-by: JamePeng <jame_peng@sina.com>

Author: JamePeng

.github/workflows/build-wheels-cu130-linux.yml

@@ -0,0 +1,156 @@
+name: Build Wheels (CU131) for Linux
+
+on:
+  workflow_dispatch:
+
+permissions:
+  contents: write
+
+jobs:
+  build_wheels:
+    name: Build Wheel ${{ matrix.os }} py${{ matrix.pyver }} cu131
+    runs-on: ubuntu-22.04
+    container: nvidia/cuda:13.1.2-cudnn-devel-ubuntu22.04
+
+    strategy:
+      fail-fast: false
+      matrix:
+        os: ["ubuntu-22.04"]
+        pyver: ["3.10", "3.11", "3.12", "3.13", "3.14"] # Python versions
+        cuda: ["13.1.2"]
+        cudaarch: ["75-real;80-real;86-real;87-real;89-real;90-real;100-real;120-real;121-real"]
+
+    defaults:
+      run:
+        shell: bash
+
+    env:
+      CUDAVER: ${{ matrix.cuda }}
+      CUDAARCHVER: ${{ matrix.cudaarch }}
+      MAX_JOBS: 12
+
+    steps:
+      - name: Install dependencies
+        run: |
+          apt update
+          apt install -y \
+            build-essential \
+            ccache \
+            cmake \
+            curl \
+            git \
+            libgomp1 \
+            libjpeg-dev \
+            libssl-dev \
+            ninja-build
+
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          submodules: recursive
+
+      - name: Install uv and Python ${{ matrix.pyver }}
+        uses: astral-sh/setup-uv@v7
+        with:
+          python-version: ${{ matrix.pyver }}
+          activate-environment: true
+          enable-cache: true
+
+      - name: Show CUDA version
+        run: nvcc -V
+
+      - name: Build wheel
+        env:
+          LD_LIBRARY_PATH: "/usr/local/cuda/lib64:/usr/local/cuda/compat:/usr/lib/x86_64-linux-gnu:${LD_LIBRARY_PATH}"
+          VERBOSE: "1"
+          CUDA_HOME: "/usr/local/cuda"
+          CUDA_PATH: "/usr/local/cuda"
+          CUDA_TOOLKIT_ROOT_DIR: "/usr/local/cuda"
+        run: |
+          set -euo pipefail
+
+          echo "LD_LIBRARY_PATH=${LD_LIBRARY_PATH}"
+          find /usr/ -name 'libcuda.so.*' || true
+          find /usr/ -name 'libcudart.so.*' || true
+
+          cuda_ver_short=$(echo "${CUDAVER}" | cut -d'.' -f 1,2 | sed 's/\.//g')
+
+          # Build one CUDA wheel with dynamic GGML backends:
+          # - GGML_BACKEND_DL enables runtime-loadable backend shared libraries.
+          # - GGML_CPU_ALL_VARIANTS builds CPU variant backends when supported.
+          # - GGML_NATIVE=OFF avoids binding the wheel to the CI runner CPU.
+          CMAKE_ARGS_ARRAY=(
+            "-G Ninja"
+
+            # Disable non-wheel targets.
+            "-DLLAMA_BUILD_EXAMPLES=OFF"
+            "-DLLAMA_BUILD_TESTS=OFF"
+            "-DLLAMA_BUILD_TOOLS=OFF"
+            "-DLLAMA_BUILD_SERVER=OFF"
+            "-DLLAMA_BUILD_UI=OFF"
+            "-DLLAMA_USE_PREBUILT_UI=OFF"
+            "-DLLAMA_CURL=OFF"
+            "-DLLAMA_OPENSSL=ON"
+
+            # GGML dynamic backend layout.
+            "-DGGML_CPU=ON"
+            "-DGGML_CUDA=ON"
+            "-DGGML_NATIVE=OFF"
+            "-DGGML_BACKEND_DL=ON"
+            "-DGGML_CPU_ALL_VARIANTS=ON"
+            "-DGGML_OPENMP=ON"
+
+            # CUDA backend.
+            "-DCMAKE_CUDA_ARCHITECTURES=${CUDAARCHVER}"
+            "-DGGML_CUDA_FORCE_MMQ=ON"
+            "-DCUDA_SEPARABLE_COMPILATION=ON"
+            "-DCMAKE_CUDA_FLAGS=--diag-suppress=177,221,550"
+
+            # Build behavior.
+            "-DCMAKE_BUILD_PARALLEL_LEVEL=${MAX_JOBS}"
+            "-DGGML_CCACHE=ON"
+            "-DENABLE_CCACHE=ON"
+          )
+
+          CMAKE_ARGS="${CMAKE_ARGS_ARRAY[*]}"
+          echo "CMAKE_ARGS=${CMAKE_ARGS}"
+
+          uv pip install --upgrade build setuptools wheel packaging
+          CMAKE_ARGS="${CMAKE_ARGS}" uv build --wheel
+
+          if ! ls dist/*.whl >/dev/null 2>&1; then
+            echo "No wheel built in dist/ directory"
+            exit 1
+          fi
+
+          wheel_path=$(ls dist/*.whl | head -n 1)
+          filename=$(basename "$wheel_path")
+
+          # Wheel filename format:
+          #   name-version-python_tag-abi_tag-platform_tag.whl
+          IFS='-' read -r dist_name version py_tag abi_tag plat_tag <<< "$filename"
+
+          # CPU all-variants is now an internal runtime layout detail.
+          new_version="${version}+cu${cuda_ver_short}"
+          new_filename="${dist_name}-${new_version}-${py_tag}-${abi_tag}-${plat_tag}"
+
+          mv "$wheel_path" "dist/$new_filename"
+          echo "Renamed wheel to: $new_filename"
+
+          echo "CUDA_VERSION=$cuda_ver_short" >> "$GITHUB_ENV"
+          echo "TAG_VERSION=$version" >> "$GITHUB_ENV"
+
+      - name: Get current date
+        id: get-date
+        run: |
+          currentDate=$(date +%Y%m%d)
+          echo "BUILD_DATE=$currentDate" >> "$GITHUB_ENV"
+
+      - name: Create release
+        if: always() && env.TAG_VERSION != ''
+        uses: softprops/action-gh-release@v3
+        with:
+          files: dist/*
+          tag_name: v${{ env.TAG_VERSION }}-cu${{ env.CUDA_VERSION }}-linux-${{ env.BUILD_DATE }}
+        env:
+          GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}