refactor(kda): remove native CUDA extension

.coderabbit.yaml

@@ -16,16 +16,9 @@ reviews:
       mode: "off"
 
   path_filters:
-    - "!csrc/cutlass/**"
     - "!third_party/**"
 
   path_instructions:
-    - path: "csrc/**"
-      instructions: |
-        Review CUDA changes for synchronization and memory-ordering correctness,
-        shared-memory and barrier lifetimes, out-of-bounds access, divergence,
-        architecture guards, numerical precision, and concrete performance regressions.
-
     - path: "cula/**"
       instructions: |
         Review CuTe DSL kernels and Python APIs for architecture dispatch,

.github/workflows/build-release.yml

@@ -1,4 +1,4 @@
-name: Build & Release Wheels
+name: Build & Release Wheel
 
 on:
   push:
@@ -12,127 +12,50 @@ concurrency:
 
 jobs:
   build-wheel:
-    name: "wheel / ${{ matrix.cuda }} / cp312 / ${{ matrix.arch }}"
-    runs-on: ${{ matrix.arch == 'aarch64' && 'ubuntu-24.04-arm' || 'ubuntu-latest' }}
-    defaults:
-      run:
-        shell: bash
-    strategy:
-      fail-fast: false
-      matrix:
-        cuda:
-          - cu129
-          - cu130
-        arch:
-          - x86_64
-          - aarch64
-    container:
-      # UBI 8 provides the glibc 2.28 baseline required by manylinux_2_28.
-      image: "nvidia/cuda:${{ matrix.cuda == 'cu129' && '12.9.0' || '13.0.0' }}-devel-ubi8"
-
+    name: wheel / py3-none-any
+    runs-on: ubuntu-latest
     steps:
-      - name: Free disk space
-        run: |
-          rm -rf /opt/hostedtoolcache /usr/local/lib/android /usr/share/dotnet \
-                 /usr/local/share/boost /opt/ghc 2>/dev/null || true
-          dnf clean all 2>/dev/null || true
-          df -h / || true
-
-      - name: Install system dependencies
-        run: |
-          dnf install -y \
-            git \
-            gcc-toolset-13-gcc \
-            gcc-toolset-13-gcc-c++ \
-            python3.12 \
-            python3.12-devel \
-            python3.12-pip
-          dnf clean all
-
       - name: Checkout
         uses: actions/checkout@v5
         with:
           fetch-depth: 0
-          submodules: recursive
-
-      - name: Configure git safe directory
-        run: git config --global --add safe.directory "$GITHUB_WORKSPACE"
 
-      - name: Install Python dependencies
-        run: |
-          python3.12 -m pip install --no-cache-dir --upgrade pip
-          python3.12 -m pip install --no-cache-dir torch==2.9.1 --index-url ${{ matrix.cuda == 'cu129' && 'https://download.pytorch.org/whl/cu129' || 'https://download.pytorch.org/whl/cu130' }}
-          python3.12 -m pip install --no-cache-dir setuptools wheel "setuptools_scm>=6.0" build ninja auditwheel patchelf
+      - name: Set up Python
+        uses: actions/setup-python@v6
+        with:
+          python-version: "3.12"
 
-      - name: Expose PyTorch CUDA libraries
-        # The nvidia/cuda devel image ships the CUDA toolkit but not cuDNN.
-        # torch's pip wheels bundle cuDNN inside site-packages/nvidia/*/lib/,
-        # which is not on the linker search path — add it so torch._C can load.
-        run: |
-          LIB_DIRS="$(python3.12 <<'EOF'
-          import glob, os, site
-          dirs = []
-          for sp in site.getsitepackages():
-              for p in [os.path.join(sp, "torch", "lib")] + glob.glob(os.path.join(sp, "nvidia", "*", "lib")):
-                  if os.path.isdir(p):
-                      dirs.append(p)
-          print(":".join(dirs))
-          EOF
-          )"
-          echo "LD_LIBRARY_PATH=${LIB_DIRS}${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}" >> "$GITHUB_ENV"
+      - name: Install build tools
+        run: python -m pip install --upgrade build setuptools wheel setuptools_scm
 
       - name: Compute version
         id: version
+        shell: bash
         run: |
           if [[ "$GITHUB_REF" == refs/tags/v* ]]; then
             BASE="${GITHUB_REF#refs/tags/v}"
           else
-            # Strip any local segment (+gXXX) so we get a clean base
-            BASE=$(python3.12 -c "from setuptools_scm import get_version; print(get_version().split('+')[0])")
+            BASE=$(python -c "from setuptools_scm import get_version; print(get_version().split('+')[0])")
           fi
-          echo "version=${BASE}+${{ matrix.cuda }}" >> "$GITHUB_OUTPUT"
+          echo "version=${BASE}" >> "$GITHUB_OUTPUT"
 
-      - name: Build fat-binary wheel
+      - name: Build wheel
         env:
-          CC: /opt/rh/gcc-toolset-13/root/usr/bin/gcc
-          CXX: /opt/rh/gcc-toolset-13/root/usr/bin/g++
-          CUDAHOSTCXX: /opt/rh/gcc-toolset-13/root/usr/bin/g++
-          CULA_BUILD_ALL_ARCHS: "1"
           SETUPTOOLS_SCM_PRETEND_VERSION: "${{ steps.version.outputs.version }}"
-          NVCC_THREADS: "4"
-          MAX_JOBS: "4"
-        run: |
-          "$CC" --version
-          "$CXX" --version
-          python3.12 -m build --wheel --no-isolation --outdir dist-raw
-
-      - name: Repair wheel for manylinux_2_28
-        run: |
-          # These libraries are supplied by the NVIDIA driver, PyTorch, or
-          # PyTorch's CUDA runtime dependency and must remain external.
-          python3.12 -m auditwheel repair \
-            --plat manylinux_2_28_${{ matrix.arch }} \
-            --exclude libcuda.so.1 \
-            --exclude 'libcudart.so.*' \
-            --exclude 'libc10*.so' \
-            --exclude 'libtorch*.so' \
-            --wheel-dir dist \
-            dist-raw/*.whl
+        run: python -m build --wheel --outdir dist
 
       - name: Verify wheel
+        shell: bash
         run: |
-          echo "Built wheel:"
-          ls -lh dist/*.whl
-          ls dist/*.whl | grep -q "+${{ matrix.cuda }}" \
-            || { echo "ERROR: wheel name missing +${{ matrix.cuda }} suffix"; exit 1; }
-          ls dist/*.whl | grep -q "manylinux_2_28_${{ matrix.arch }}" \
-            || { echo "ERROR: wheel is not tagged manylinux_2_28_${{ matrix.arch }}"; exit 1; }
-          python3.12 -m auditwheel show dist/*.whl
+          python -m pip install --no-deps dist/*.whl
+          python -c "import cula; print(cula.__version__)"
+          ls dist/*.whl | grep -q -- '-py3-none-any.whl$' \
+            || { echo "ERROR: wheel is not platform-independent"; exit 1; }
 
       - name: Upload wheel artifact
         uses: actions/upload-artifact@v6
         with:
-          name: wheel-${{ matrix.cuda }}-${{ matrix.arch }}
+          name: wheel-py3-none-any
           path: dist/*.whl
 
   release:
@@ -143,16 +66,15 @@ jobs:
     permissions:
       contents: write
     steps:
-      - name: Download all artifacts
+      - name: Download wheel
         uses: actions/download-artifact@v6
         with:
           path: artifacts/
 
       - name: Create release
         uses: softprops/action-gh-release@v3
         with:
-          files: |
-            artifacts/wheel-*/*.whl
+          files: artifacts/wheel-py3-none-any/*.whl
           generate_release_notes: true
           draft: true
           prerelease: ${{ contains(github.ref, 'rc') || contains(github.ref, 'beta') || contains(github.ref, 'alpha') }}

.github/workflows/lint.yml

@@ -13,17 +13,6 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  clang-format:
-    name: clang-format
-    runs-on: ubuntu-latest
-    steps:
-      - uses: actions/checkout@v6
-      - uses: j178/prek-action@v2
-        with:
-          install-only: true
-      - name: Check clang-format
-        run: prek run clang-format --all-files --show-diff-on-failure --color=always
-
   ruff:
     name: Ruff
     runs-on: ubuntu-latest

.gitmodules

@@ -1,6 +1,3 @@
-[submodule "csrc/cutlass"]
-	path = csrc/cutlass
-	url = https://github.com/NVIDIA/cutlass.git
 [submodule "third_party/flash-linear-attention"]
 	path = third_party/flash-linear-attention
 	url = https://github.com/fla-org/flash-linear-attention.git

.pre-commit-config.yaml

@@ -1,12 +1,4 @@
 repos:
-  - repo: https://github.com/pre-commit/mirrors-clang-format
-    rev: v15.0.7
-    hooks:
-      - id: clang-format
-        files: ^csrc/
-        types_or: [c, c++, cuda]
-        exclude: ^csrc/(cutlass|kerutils)/
-
   - repo: https://github.com/astral-sh/ruff-pre-commit
     rev: v0.15.0
     hooks:

MANIFEST.in

@@ -3,6 +3,4 @@ include LICENSE
 include pyproject.toml
 include setup.py
 
-recursive-include csrc *.h *.hpp *.cuh *.cu *.cpp *.cc
-recursive-include csrc/cutlass *
 recursive-include cula *.py

README.md

@@ -4,7 +4,7 @@
 
 # cuLA — CUDA Linear Attention
 
-**High-performance CUDA kernels for linear attention variants, written in [CuTe DSL](https://github.com/NVIDIA/cutlass/tree/main/python/CuTeDSL) and CUTLASS C++.**
+**High-performance CUDA kernels for linear attention variants, written entirely in [CuTe DSL](https://github.com/NVIDIA/cutlass/tree/main/python/CuTeDSL).**
 
 </div>
 
@@ -20,17 +20,21 @@ Linear attention mechanisms reformulate standard attention to use linear-time st
 
 cuLA supports both **Hopper (SM90)** and **Blackwell (SM10X)** GPUs.
 
-> **Requirements (Hopper & Blackwell):** Python 3.12+, CUDA Toolkit 12.9+, PyTorch 2.9.1+. Building the optional SM90 native extension also requires NVCC.
+> **Requirements (Hopper & Blackwell):** Python 3.12+, CUDA Toolkit 12.9+, PyTorch 2.8.0+, and a matching `nvidia-cutlass-dsl` installation. cuLA does not build a C++/CUDA extension.
 
 > **Note:** The PyTorch CUDA version must match your system CUDA Toolkit version. Check with `nvcc --version` and `python -c "import torch; print(torch.version.cuda)"`.
 
 ### Pre-built Wheels
 
-Pre-built wheels with the SM90 native extension and CuTeDSL JIT kernels for SM100/SM103 are available on [GitHub Releases](https://github.com/inclusionAI/cuLA/releases). Linux wheels target `manylinux_2_28` and require glibc 2.28 or newer:
+Platform-independent wheels containing the CuTeDSL sources are available on [GitHub Releases](https://github.com/inclusionAI/cuLA/releases):
 
-    pip install "cuda-linear-attention==<VERSION>+<CUDA_TAG>" -f https://github.com/inclusionAI/cuLA/releases/expanded_assets/<TAG>
+    pip install "cuda-linear-attention==<VERSION>" -f https://github.com/inclusionAI/cuLA/releases/expanded_assets/<TAG>
 
-Replace `<TAG>` with the release tag (e.g., `v0.2.0`), `<VERSION>` with the base version (e.g., `0.2.0`), and `<CUDA_TAG>` with your PyTorch CUDA build tag (e.g., `cu129` or `cu130`). Or download the `.whl` file directly from the [Releases page](https://github.com/inclusionAI/cuLA/releases) and install it with `pip install <filename>.whl`.
+Replace `<TAG>` with the release tag (e.g., `v0.2.0`) and `<VERSION>` with the version (e.g., `0.2.0`). Or download the `.whl` file directly from the [Releases page](https://github.com/inclusionAI/cuLA/releases) and install it with `pip install <filename>.whl`.
+
+The cuLA wheel does not install FLA. Install `flash-linear-attention` before importing `chunk_kda`:
+
+    pip install flash-linear-attention
 
 ### Build from Source
 
@@ -41,10 +45,20 @@ git clone https://github.com/inclusionAI/cuLA.git
 git submodule update --init --recursive
 ```
 
-**Install PyTorch:**
+**Install PyTorch and CuTeDSL:**
+
+For CUDA 12.9:
+
+```bash
+pip install torch==2.8.0 --index-url https://download.pytorch.org/whl/cu129
+pip install 'nvidia-cutlass-dsl>=4.6.2,<4.9'
+```
+
+For CUDA 13:
 
 ```bash
-pip install torch==2.9.1 --index-url https://download.pytorch.org/whl/cu129
+pip install torch==2.9.1 --index-url https://download.pytorch.org/whl/cu130
+pip install 'nvidia-cutlass-dsl[cu13]>=4.6.2,<4.9'
 ```
 
 **Install cuLA & dependencies:**
@@ -57,10 +71,10 @@ pip install -e third_party/flash-linear-attention
 pip install -e . --no-build-isolation
 ```
 
-**Build a wheel with the SM90 extension (SM100/SM103 use CuTeDSL JIT):**
+**Build the platform-independent wheel:**
 
 ```bash
-CULA_BUILD_ALL_ARCHS=1 python -m build --wheel --no-isolation
+python -m build --wheel
 ```
 
 ## Quick Start