redo ci: cross python wheel (#824)

apply #662 again, since
we have #823 merged now.

---------

Signed-off-by: youkaichao <[email protected]>

Author: youkaichao

.github/workflows/release_wheel.yml

@@ -25,7 +25,6 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        python: ["3.8", "3.9", "3.10", "3.11", "3.12"]
         cuda: ["11.8", "12.1", "12.4"]
         torch: ["2.3", "2.4", "2.5"]
         exclude: # for cuda 12.4, we only support torch 2.4+
@@ -49,7 +48,6 @@ jobs:
               -v "$CI_RUNNER_CACHE_DIR":/ci-cache \
               -v "$GITHUB_WORKSPACE":/app \
               -e FLASHINFER_CI_CACHE=/ci-cache \
-              -e FLASHINFER_CI_PYTHON_VERSION=${{ matrix.python }} \
               -e FLASHINFER_CI_CUDA_VERSION=${{ matrix.cuda }} \
               -e FLASHINFER_CI_TORCH_VERSION=${{ matrix.torch }} \
               -e TORCH_CUDA_ARCH_LIST="$TORCH_CUDA_ARCH_LIST" \
@@ -62,7 +60,7 @@ jobs:
 
       - uses: actions/upload-artifact@v4
         with:
-          name: wheel-cuda${{ matrix.cuda }}-torch${{ matrix.torch }}-python${{ matrix.python }}
+          name: wheel-cuda${{ matrix.cuda }}-torch${{ matrix.torch }}
           path: dist/*
 
   release:
@@ -81,31 +79,7 @@ jobs:
         with:
           tag_name: ${{ inputs.tag_name }}
           files: |
-            dist/flashinfer*cp38*.whl
-
-      - uses: softprops/action-gh-release@v1
-        with:
-          tag_name: ${{ inputs.tag_name }}
-          files: |
-            dist/flashinfer*cp39*.whl
-
-      - uses: softprops/action-gh-release@v1
-        with:
-          tag_name: ${{ inputs.tag_name }}
-          files: |
-            dist/flashinfer*cp310*.whl
-
-      - uses: softprops/action-gh-release@v1
-        with:
-          tag_name: ${{ inputs.tag_name }}
-          files: |
-            dist/flashinfer*cp311*.whl
-
-      - uses: softprops/action-gh-release@v1
-        with:
-          tag_name: ${{ inputs.tag_name }}
-          files: |
-            dist/flashinfer*cp312*.whl
+            dist/flashinfer*.whl
 
       - uses: softprops/action-gh-release@v1
         with:

scripts/run-ci-build-wheel.sh

@@ -11,7 +11,6 @@ assert_env() {
 }
 
 assert_env FLASHINFER_CI_CACHE
-assert_env FLASHINFER_CI_PYTHON_VERSION
 assert_env FLASHINFER_CI_CUDA_VERSION
 assert_env FLASHINFER_CI_TORCH_VERSION
 assert_env TORCH_CUDA_ARCH_LIST
@@ -26,8 +25,6 @@ CUDA_MAJOR="${FLASHINFER_CI_CUDA_VERSION%.*}"
 CUDA_MINOR="${FLASHINFER_CI_CUDA_VERSION#*.}"
 TORCH_MAJOR="${FLASHINFER_CI_TORCH_VERSION%.*}"
 TORCH_MINOR="${FLASHINFER_CI_TORCH_VERSION#*.}"
-PYVER="${FLASHINFER_CI_PYTHON_VERSION//./}"
-export PATH="/opt/python/cp${PYVER}-cp${PYVER}/bin:$PATH"
 
 FLASHINFER_LOCAL_VERSION="cu${CUDA_MAJOR}${CUDA_MINOR}torch${FLASHINFER_CI_TORCH_VERSION}"
 if [ -n "${FLASHINFER_GIT_SHA}" ]; then

setup.py

@@ -212,6 +212,7 @@ def __init__(self, *args, **kwargs) -> None:
     cxx_flags = [
         "-O3",
         "-Wno-switch-bool",
+        "-DPy_LIMITED_API=0x03080000",
     ]
     nvcc_flags = [
         "-O3",
@@ -220,6 +221,7 @@ def __init__(self, *args, **kwargs) -> None:
         "-Xfatbin",
         "-compress-all",
         "-use_fast_math",
+        "-DPy_LIMITED_API=0x03080000",
     ]
     libraries = [
         "cublas",
@@ -264,6 +266,7 @@ def __init__(self, *args, **kwargs) -> None:
                 "cxx": cxx_flags,
                 "nvcc": nvcc_flags,
             },
+            py_limited_api=True,
         )
     ]
     if enable_sm90:
@@ -277,6 +280,7 @@ def __init__(self, *args, **kwargs) -> None:
                     "cxx": cxx_flags,
                     "nvcc": nvcc_flags + sm90a_flags,
                 },
+                py_limited_api=True,
             ),
         ]
 
@@ -285,4 +289,5 @@ def __init__(self, *args, **kwargs) -> None:
     ext_modules=ext_modules,
     cmdclass=cmdclass,
     install_requires=install_requires,
+    options={"bdist_wheel": {"py_limited_api": "cp38"}},
 )