Merge 0.2.0 into main: propagate_iterative non-convergence fix + 0.2.0 release

.github/workflows/wheels.yml

@@ -49,7 +49,6 @@ jobs:
       - name: Build the universal CPU wheel
         env:
           SFA_BUILD_CUDA: "0"
-          SFA_PACKAGE_NAME: "sfa"
         run: |
           python -m pip install --upgrade pip build
           python -m build --wheel
@@ -72,7 +71,6 @@ jobs:
       - name: Build sdist
         env:
           SFA_BUILD_CUDA: "0"
-          SFA_PACKAGE_NAME: "sfa"
         run: |
           python -m pip install --upgrade pip build
           python -m build --sdist
@@ -221,6 +219,22 @@ jobs:
           echo "--- $CUDA_HOME/lib64 (cublas only) ---"
           ls -la "$CUDA_HOME/lib64/" | grep -E 'cublas|nvrtc' || true
 
+      - name: Rename wheel project to ${{ matrix.cuda.pkg }}
+        # PEP 621 forbids `name` from being declared dynamic in
+        # pyproject.toml, so the same source tree's static
+        # `name = "sfa"` must be sed-replaced in place per CUDA cell
+        # to produce the per-CUDA wheel names (sfa-cu128 / sfa-cu132).
+        # The CPU and sdist jobs leave this line untouched and ship
+        # under the default `sfa` name. Runs on both Linux and
+        # Windows runners via Git Bash; cibuildwheel afterwards reads
+        # the modified pyproject.toml inside its build sandbox.
+        shell: bash
+        run: |
+          set -euo pipefail
+          sed -i 's/^name = "sfa"$/name = "${{ matrix.cuda.pkg }}"/' pyproject.toml
+          echo "--- name line after rename ---"
+          grep '^name' pyproject.toml
+
       - name: Build wheels
         uses: pypa/cibuildwheel@v2.21
         env:
@@ -238,9 +252,10 @@ jobs:
           CIBW_CONTAINER_ENGINE: >-
             docker; create_args: -v /usr/local/cuda:/usr/local/cuda:ro
           # Linux-only env: bind-mounted CUDA path + PATH update.
+          # The wheel project name is set by the sed step above
+          # (pyproject.toml's `[project] name`), not by an env var.
           CIBW_ENVIRONMENT_LINUX: >-
             SFA_BUILD_CUDA=1
-            SFA_PACKAGE_NAME=${{ matrix.cuda.pkg }}
             SFA_CUDA_ARCH="${{ matrix.cuda.archs }}"
             SFA_CUDA_RUNTIME_REQUIRES="${{ matrix.cuda.runtime_requires }}"
             CUDA_PATH=/usr/local/cuda
@@ -250,7 +265,6 @@ jobs:
           # subprocess automatically.
           CIBW_ENVIRONMENT_WINDOWS: >-
             SFA_BUILD_CUDA=1
-            SFA_PACKAGE_NAME=${{ matrix.cuda.pkg }}
             SFA_CUDA_ARCH="${{ matrix.cuda.archs }}"
             SFA_CUDA_RUNTIME_REQUIRES="${{ matrix.cuda.runtime_requires }}"
           # GitHub runners have no NVIDIA GPU, so the CUDA-gated pytest
@@ -280,17 +294,51 @@ jobs:
           name: cuda-wheels-${{ matrix.cuda.pkg }}-${{ matrix.os }}
           path: ./wheelhouse/*.whl
 
+  # ---------------------------------------------------------------------------
+  # Publish all build artifacts (CPU universal wheel, sdist, and every
+  # CUDA wheel produced by the matrix) to the GitHub Release that the
+  # `v*` tag triggers. End users install from these URLs directly:
+  #
+  #   pip install https://github.com/dwgoon/sfa/releases/download/v0.2.0/<wheel-name>
+  #
+  # This is the primary distribution channel while the PyPI publish job
+  # below is disabled.
+  # ---------------------------------------------------------------------------
+  publish_github_release:
+    name: publish-github-release
+    needs: [build_cpu_wheel, build_sdist, build_cuda_wheels]
+    runs-on: ubuntu-latest
+    if: startsWith(github.ref, 'refs/tags/v')
+    permissions:
+      contents: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          path: dist
+          merge-multiple: true
+      - name: List collected artifacts
+        run: ls -la dist/
+      - uses: softprops/action-gh-release@v2
+        with:
+          files: dist/*
+          generate_release_notes: true
+          prerelease: false
+          fail_on_unmatched_files: true
+
   # ---------------------------------------------------------------------------
   # Optional: publish to PyPI on tag pushes. Requires the OIDC trusted-
   # publisher relationship to be configured at https://pypi.org for each
-  # of `sfa`, `sfa-cu128`, `sfa-cu132`, `sfa-cu133`. Disabled by default;
-  # change the `if:` guard to enable.
+  # of `sfa`, `sfa-cu128`, `sfa-cu132`. Currently disabled because the
+  # maintainer's PyPI account is unavailable; re-enable by flipping the
+  # `if:` guard back to `startsWith(github.ref, 'refs/tags/v')` once
+  # trusted publishers are in place. Existing GitHub Release artifacts
+  # are unaffected by this gate.
   # ---------------------------------------------------------------------------
-  publish:
+  publish_pypi:
     name: publish-to-pypi
     needs: [build_cpu_wheel, build_sdist, build_cuda_wheels]
     runs-on: ubuntu-latest
-    if: startsWith(github.ref, 'refs/tags/v')
+    if: false
     permissions:
       id-token: write
     steps:

INSTALL.md

@@ -20,6 +20,15 @@ runtime they link against.
 This file is a quick reference. See
 [doc/install.md](doc/install.md) for the full guide.
 
+> [!NOTE]
+> For the **0.2.0** line, the primary distribution channel is the
+> [GitHub Releases page](https://github.com/dwgoon/sfa/releases),
+> not PyPI. The `pip install sfa-cuXYZ` shortcuts shown below assume
+> the wheels are on PyPI; until they land there, use the explicit
+> release URLs in the "Install from GitHub Releases" section. The
+> 0.1.0 `sfa` CPU package remains installable from PyPI through the
+> `pip install sfa` snippet.
+
 ## CPU (any OS)
 
 ```bash
@@ -53,6 +62,56 @@ NVIDIA driver development in 2019.
 > Install **only one** `sfa-cuXYZ` per environment; mixing them causes
 > import conflicts.
 
+## Install from GitHub Releases
+
+When a `v*` tag is pushed, `wheels.yml` attaches every artifact (one
+universal CPU wheel, the sdist, and a CUDA wheel for every supported
+Python + OS + CUDA combination) to the corresponding GitHub Release.
+`pip` can install those wheels directly:
+
+```bash
+# CPU (universal; any OS, any Python from 3.10 to 3.13).
+pip install https://github.com/dwgoon/sfa/releases/download/v0.2.0/sfa-0.2.0-py3-none-any.whl
+```
+
+For CUDA wheels, the filename encodes Python version, OS, and CUDA
+major.minor. The pattern is
+
+```
+sfa_cuXYZ-VER-cpPP-cpPP-PLAT.whl
+```
+
+with
+
+| Token  | Meaning                                              | Possible values                                  |
+|--------|------------------------------------------------------|--------------------------------------------------|
+| `XYZ`  | CUDA major.minor (no dot)                            | `128` (CUDA 12.8) or `132` (CUDA 13.2)           |
+| `VER`  | sfa version (matches the tag without the `v` prefix) | e.g. `0.2.0`                                     |
+| `PP`   | CPython major+minor (no dot)                         | `310`, `311`, `312`, `313`                       |
+| `PLAT` | Wheel platform tag                                   | `manylinux_2_28_x86_64` (Linux) / `win_amd64` (Windows) |
+
+Examples:
+
+```bash
+# Linux + Python 3.12 + CUDA 13.2
+pip install https://github.com/dwgoon/sfa/releases/download/v0.2.0/sfa_cu132-0.2.0-cp312-cp312-manylinux_2_28_x86_64.whl
+
+# Windows + Python 3.10 + CUDA 12.8
+pip install https://github.com/dwgoon/sfa/releases/download/v0.2.0/sfa_cu128-0.2.0-cp310-cp310-win_amd64.whl
+```
+
+If none of the prebuilt wheels match your environment, install the
+sdist or a tagged git source instead:
+
+```bash
+# sdist (compiles a pure-Python install; no CUDA extension).
+pip install https://github.com/dwgoon/sfa/releases/download/v0.2.0/sfa-0.2.0.tar.gz
+
+# git source at the v0.2.0 tag (set SFA_BUILD_CUDA=1 + have nvcc to
+# build the CUDA extension; otherwise a pure-Python install is made).
+pip install git+https://github.com/dwgoon/sfa.git@v0.2.0
+```
+
 ## Optional extras
 
 Matplotlib-based plotting helpers in `sfa.plot`:

README.md

@@ -74,6 +74,25 @@ pip install sfa-cu132
 > `sfa-cuXYZ` share the `sfa` Python namespace and will conflict if
 > stacked.
 
+### Install from GitHub Releases (current 0.2.0 primary channel)
+
+The 0.2.0 line is distributed through the project's
+[GitHub Releases page](https://github.com/dwgoon/sfa/releases) until
+the new CUDA wheels land on PyPI. Each `v*` tag attaches one universal
+CPU wheel, the sdist, and a per-Python / per-OS / per-CUDA wheel for
+each `sfa-cuXYZ` variant. Examples:
+
+```bash
+# CPU (universal, any OS / Python 3.10 - 3.13)
+pip install https://github.com/dwgoon/sfa/releases/download/v0.2.0/sfa-0.2.0-py3-none-any.whl
+
+# CUDA 13.2, Linux, Python 3.12
+pip install https://github.com/dwgoon/sfa/releases/download/v0.2.0/sfa_cu132-0.2.0-cp312-cp312-manylinux_2_28_x86_64.whl
+```
+
+See [INSTALL.md](INSTALL.md#install-from-github-releases) for the
+full wheel-filename pattern and Windows / older-Python URLs.
+
 ### Build from source
 
 For a new CUDA major version, a custom GPU architecture, or

pyproject.toml

@@ -2,9 +2,17 @@
 requires = ["setuptools>=68", "wheel", "pybind11>=2.13", "numpy>=2.0"]
 build-backend = "setuptools.build_meta"
 
-# The project `name` is set dynamically from setup.py so the same source
-# tree can produce both "sfa" (CPU-only) and "sfa-cu130" (CUDA-enabled)
-# wheels under the SFA_PACKAGE_NAME env var.
+# PEP 621 requires `name` to be a static string here (it explicitly
+# forbids name from being dynamic). The CPU and sdist builds use this
+# default `sfa`; CUDA build cells in `wheels.yml` overwrite this line
+# in-place with `sed` so the same source tree produces sfa-cu128 /
+# sfa-cu132 wheels under their own PyPI names.
+#
+# `dependencies` is left dynamic so setup.py can append the per-CUDA
+# nvidia-cublas-cuXX / nvidia-cuda-runtime-cuXX runtime packages via
+# the SFA_CUDA_RUNTIME_REQUIRES env var. With no entry under
+# [tool.setuptools.dynamic], setuptools reads the resolved value from
+# setup.py's `install_requires=` call.
 [project]
 name = "sfa"
 version = "0.2.0"
@@ -13,13 +21,7 @@ readme = "README.md"
 license = { text = "MIT" }
 authors = [{ name = "Daewon Lee", email = "daewon4you@gmail.com" }]
 requires-python = ">=3.10"
-dependencies = [
-    "numpy",
-    "scipy",
-    "pandas",
-    "networkx",
-    "threadpoolctl",
-]
+dynamic = ["dependencies"]
 
 [project.optional-dependencies]
 cuda = []  # toolchain provided by conda env (environment-cuda.yml)
@@ -40,8 +42,9 @@ include = ["sfa*"]
 # Builds CPU-only wheels by default. To produce CUDA wheels, set
 # environment variables in the GitHub Actions workflow:
 #   SFA_BUILD_CUDA=1
-#   SFA_PACKAGE_NAME=sfa-cu130
-#   SFA_CUDA_ARCH=sm_70;sm_75;sm_80;sm_86;sm_89;sm_90  (AOT fat binary)
+#   SFA_CUDA_ARCH=sm_75;sm_80;sm_86;sm_89;sm_90;sm_100;sm_120
+# and rewrite the static `[project] name` line above to the target
+# wheel name (sfa-cu128 / sfa-cu132) via a sed step before running.
 # ---------------------------------------------------------------------------
 [tool.cibuildwheel]
 build = "cp310-* cp311-* cp312-* cp313-*"

setup.py

@@ -253,46 +253,29 @@ def _cuda_extension() -> Optional[Extension]:
     ext_modules.append(cuda_ext)
 
 
-# PyPI package name. Default is the cross-platform CPU-only package "sfa".
-# CUDA-enabled wheels are published under per-CUDA-version names
-# (sfa-cu128, sfa-cu132, sfa-cu133) so users can pick the variant that
-# matches their NVIDIA driver. Override at build time:
-#   SFA_PACKAGE_NAME=sfa-cu132 SFA_BUILD_CUDA=1 python -m build --wheel
-_pkg_name = os.environ.get("SFA_PACKAGE_NAME", "sfa")
-
-
 # Per-build extra install_requires fed in by the CI matrix. Used to pin
 # the NVIDIA runtime PyPI packages (nvidia-cublas-cuXX,
 # nvidia-cuda-runtime-cuXX) that match the CUDA major version we are
 # building against. The value is a whitespace-separated list of pip
 # requirement specifiers, e.g.
 #   SFA_CUDA_RUNTIME_REQUIRES="nvidia-cublas-cu13>=13.2,<13.3 \
 #                              nvidia-cuda-runtime-cu13>=13.2,<13.3"
+#
+# The wheel's static metadata (name, version, description, author, etc.)
+# lives in pyproject.toml's [project] table; setup.py only supplies what
+# isn't there: the CUDA build extension, the dynamic install_requires
+# composed from the base list plus _extra_runtime, and the
+# build-extension command class.
 _extra_runtime = [
     spec for spec in os.environ.get("SFA_CUDA_RUNTIME_REQUIRES", "").split()
     if spec.strip()
 ]
 
 
 setup(
-    name=_pkg_name,
-    version="0.2.0.dev0",
-    description="Signal flow analysis",
-    url="http://github.com/dwgoon/sfa",
-    author="Daewon Lee",
-    author_email="daewon4you@gmail.com",
-    license="MIT",
-    packages=find_packages(),
-    package_data={"": ["*.tsv", "*.sif", "*.json"]},
-    python_requires=">=3.10",
     install_requires=(["numpy", "scipy", "pandas", "networkx",
                        "threadpoolctl"]
                       + _extra_runtime),
-    extras_require={
-        "plot": ["matplotlib", "seaborn"],
-        "cuda": [],  # toolchain provided by conda env (environment-cuda.yml)
-        "test": ["pytest"],
-    },
     ext_modules=ext_modules,
     cmdclass={"build_ext": CudaBuildExt} if ext_modules else {},
     zip_safe=False,