Configure release CI - #12

* Migrate _C_gguf to PyTorch stable C++ ABI + Py_LIMITED_API

Builds the CUDA extension against torch::stable::* + STABLE_TORCH_LIBRARY
so a single abi3 wheel is binary-compatible across libtorch >= 2.9 and
CPython >= 3.10 without per-version rebuilds.

- torch_bindings.cpp: TORCH_LIBRARY -> STABLE_TORCH_LIBRARY + TORCH_BOX(...).
  PYBIND11_MODULE replaced by a minimal PyInit__C_gguf so the module loads
  under Py_LIMITED_API (no pybind11).
- gguf/gguf_kernel.cu: wrappers take torch::stable::Tensor.
  at::cuda::OptionalCUDAGuard -> torch::stable::accelerator::DeviceGuard,
  at::cuda::getCurrentCUDAStream() -> aoti_torch_get_current_cuda_stream,
  torch::empty/zeros(opts) -> torch::stable::new_empty/new_zeros(self, sizes, dtype).
  Inner CUDA kernels (mmq.cuh, mmvq.cuh, moe.cuh, ...) are unchanged.
- dispatch_utils.h: AT_DISPATCH_* (depends on ATen internals) -> THO_DISPATCH_V2
  from torch/headeronly (header-only, stable-ABI safe).
- setup.py: py_limited_api=True, bdist_wheel.py_limited_api="cp310",
  nvcc -DUSE_CUDA so the AOTI CUDA-stream shim decl is visible.
- pyproject.toml: torch>=2.9 (torch::stable::Tensor minimum).

Verified by building the wheel against torch 2.12.0+cu130 and loading it
under torch 2.11.0+cu130: ggml_dequantize (CUDA) and
ggml_moe_get_block_size (composite) both run end-to-end.

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude

* Add release workflow with sdist + manylinux wheel

Extends #12's sdist publish flow with a build-wheel job so PyPI gets a
prebuilt binary alongside the source distribution. The stable-ABI move
in the previous commit is what makes a single wheel viable across
libtorch / CPython versions.

- scripts/build_release_wheel.sh: derives TORCH_CUDA_ARCH_LIST from
  `nvcc --version`, mirroring vllm-project/vllm CMakeLists.txt
  CUDA_SUPPORTED_ARCHS. 10.1 is dropped from the 12.8 list because
  PyTorch's torch.utils.cpp_extension._get_cuda_arch_flags validator
  does not list it.
- .github/workflows/release.yml:
  * build-sdist: replicates #12 verbatim.
  * build-wheel: nvidia/cuda:12.8.1-devel-ubuntu22.04 container
    (glibc 2.35 floor). Installs torch+cu128, runs the build script,
    retags the wheel `linux_x86_64` -> `manylinux_2_35_x86_64` via
    `python -m wheel tags`, twine-checks it.
  * publish-pypi / publish-testpypi: download both artifacts and
    publish them together.
- MANIFEST.in: pulls vllm_gguf_plugin/csrc/**/*.{h,cpp,cu,cuh} into the
  sdist so source builds work from PyPI tarballs.
- setup.py: gains the `_should_build_extension()` guard from #12 so
  sdist / egg_info / dist_info commands don't import torch.

Local sanity check on CUDA 13.0:
  `pip install -e .` (no env var): 23s, 4.8 MB .so, local arch only.
  `./scripts/build_release_wheel.sh`: 101s, 6.0 MB wheel, sm_75..sm_120.
  `python -m build --sdist`: <2s, 77 KB tar.gz.
`twine check` passes on both the sdist and the retagged wheel.

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude

---------

Signed-off-by: mgoin <mgoin64@gmail.com>

…oject#14)

* Migrate _C_gguf to PyTorch stable C++ ABI + Py_LIMITED_API

Builds the CUDA extension against torch::stable::* + STABLE_TORCH_LIBRARY
so a single abi3 wheel is binary-compatible across libtorch >= 2.9 and
CPython >= 3.10 without per-version rebuilds.

- torch_bindings.cpp: TORCH_LIBRARY -> STABLE_TORCH_LIBRARY + TORCH_BOX(...).
  PYBIND11_MODULE replaced by a minimal PyInit__C_gguf so the module loads
  under Py_LIMITED_API (no pybind11).
- gguf/gguf_kernel.cu: wrappers take torch::stable::Tensor.
  at::cuda::OptionalCUDAGuard -> torch::stable::accelerator::DeviceGuard,
  at::cuda::getCurrentCUDAStream() -> aoti_torch_get_current_cuda_stream,
  torch::empty/zeros(opts) -> torch::stable::new_empty/new_zeros(self, sizes, dtype).
  Inner CUDA kernels (mmq.cuh, mmvq.cuh, moe.cuh, ...) are unchanged.
- dispatch_utils.h: AT_DISPATCH_* (depends on ATen internals) -> THO_DISPATCH_V2
  from torch/headeronly (header-only, stable-ABI safe).
- setup.py: py_limited_api=True, bdist_wheel.py_limited_api="cp310",
  nvcc -DUSE_CUDA so the AOTI CUDA-stream shim decl is visible.
- pyproject.toml: torch>=2.9 (torch::stable::Tensor minimum).

Verified by building the wheel against torch 2.12.0+cu130 and loading it
under torch 2.11.0+cu130: ggml_dequantize (CUDA) and
ggml_moe_get_block_size (composite) both run end-to-end.

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude

* Add release workflow with sdist + manylinux wheel

Extends vllm-project#12's sdist publish flow with a build-wheel job so PyPI gets a
prebuilt binary alongside the source distribution. The stable-ABI move
in the previous commit is what makes a single wheel viable across
libtorch / CPython versions.

- scripts/build_release_wheel.sh: derives TORCH_CUDA_ARCH_LIST from
  `nvcc --version`, mirroring vllm-project/vllm CMakeLists.txt
  CUDA_SUPPORTED_ARCHS. 10.1 is dropped from the 12.8 list because
  PyTorch's torch.utils.cpp_extension._get_cuda_arch_flags validator
  does not list it.
- .github/workflows/release.yml:
  * build-sdist: replicates vllm-project#12 verbatim.
  * build-wheel: nvidia/cuda:12.8.1-devel-ubuntu22.04 container
    (glibc 2.35 floor). Installs torch+cu128, runs the build script,
    retags the wheel `linux_x86_64` -> `manylinux_2_35_x86_64` via
    `python -m wheel tags`, twine-checks it.
  * publish-pypi / publish-testpypi: download both artifacts and
    publish them together.
- MANIFEST.in: pulls vllm_gguf_plugin/csrc/**/*.{h,cpp,cu,cuh} into the
  sdist so source builds work from PyPI tarballs.
- setup.py: gains the `_should_build_extension()` guard from vllm-project#12 so
  sdist / egg_info / dist_info commands don't import torch.

Local sanity check on CUDA 13.0:
  `pip install -e .` (no env var): 23s, 4.8 MB .so, local arch only.
  `./scripts/build_release_wheel.sh`: 101s, 6.0 MB wheel, sm_75..sm_120.
  `python -m build --sdist`: <2s, 77 KB tar.gz.
`twine check` passes on both the sdist and the retagged wheel.

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Claude

---------

Signed-off-by: mgoin <mgoin64@gmail.com>