Remove the logic for bundling tileiras into wheel

Signed-off-by: Jay Gu <jagu@nvidia.com>

Author: haijieg

setup.py

@@ -2,11 +2,8 @@
 #
 # SPDX-License-Identifier: Apache-2.0
 from distutils import file_util
-import shutil
-
 from setuptools import setup
 from setuptools.command.build_ext import build_ext
-from setuptools.command.bdist_wheel import bdist_wheel
 from setuptools.extension import Extension
 import os
 import sys
@@ -76,51 +73,6 @@ def run(self):
                                 dry_run=self.dry_run)
 
 
-class BdistWheelWithDeps(bdist_wheel):
-    user_options = bdist_wheel.user_options + [
-        ('include-tileir', None, 'Include tileir dependency in the package'),
-        ('tileir-deps=', None, 'Path to extra tileir dependency to package'),
-    ]
-
-    def initialize_options(self):
-        super().initialize_options()
-        self.include_tileir = False
-        self.tileir_deps = None
-
-    def finalize_options(self):
-        super().finalize_options()
-
-    def _copy_tileir_deps(self, src, dst):
-        bin_dir = os.path.join(dst, 'bin')
-        lib_dir = os.path.join(dst, 'lib')
-        file_util.copy_file(os.path.join(src, 'tileiras'), bin_dir,
-                            dry_run=self.dry_run)
-        file_util.copy_file(os.path.join(src, 'ptxas'), bin_dir,
-                            dry_run=self.dry_run)
-        file_util.copy_file(os.path.join(src, 'libnvvm.so'), lib_dir,
-                            dry_run=self.dry_run)
-
-    def run(self):
-        build_py = self.get_finalized_command('build_py')
-        build_lib = build_py.build_lib
-        dst_dir = os.path.join(build_lib, 'cuda', 'tile', '_deps')
-        dst_bin_dir = os.path.join(dst_dir, 'bin')
-        dst_lib_dir = os.path.join(dst_dir, 'lib')
-        if self.include_tileir:
-            os.makedirs(dst_bin_dir, exist_ok=True)
-            os.makedirs(dst_lib_dir, exist_ok=True)
-            if self.tileir_deps:
-                src_dir = os.path.abspath(self.tileir_deps)
-                self._copy_tileir_deps(src_dir, dst_dir)
-            else:
-                compiler_src = shutil.which('tileiras')
-                if compiler_src is None:
-                    raise FileNotFoundError("Cannot find `tileiras`. Make sure it is in PATH.")
-                file_util.copy_file(compiler_src, dst_bin_dir, update=True, dry_run=self.dry_run)
-
-        super().run()
-
-
 def _get_csrc_dir(ext_name: str):
     prefix = "cuda.tile._"
     assert ext_name.startswith(prefix)
@@ -141,6 +93,5 @@ def _get_build_lib_filename(ext_name: str):
     ],
     cmdclass=dict(
         build_ext=BuildExtWithCmake,
-        bdist_wheel=BdistWheelWithDeps
     )
 )

src/cuda/tile/_compile.py

@@ -319,12 +319,6 @@ def _get_cuda_home() -> Optional[str]:
     return os.environ.get("CUDA_HOME")
 
 
-def _local_deps_dir():
-    import cuda.tile
-    package_dir = os.path.dirname(os.path.abspath(cuda.tile.__file__))
-    return os.path.join(package_dir, '_deps')
-
-
 @dataclass
 class _CompilerBinary:
     path: str
@@ -410,19 +404,9 @@ def _find_pip_tileiras() -> Optional[str]:
 
 @cache
 def _find_compiler_bin() -> _CompilerBinary:
-    # search under cuda/tile/_deps
     bin_path = os.environ.get('PATH', '')
     ld_path = os.environ.get('LD_LIBRARY_PATH', "") if not is_windows() else ""
 
-    deps_bin_dir = os.path.join(_local_deps_dir(), 'bin')
-    deps_lib_dir = os.path.join(_local_deps_dir(), 'lib')
-    if os.path.exists(deps_bin_dir):
-        logger.debug(f"Searching tileiras: {deps_bin_dir}")
-        if (res := shutil.which("tileiras", path=deps_bin_dir)):
-            bin_path = deps_bin_dir + ":" + bin_path
-            ld_path = deps_lib_dir + ":" + ld_path
-            return _CompilerBinary(res, bin_path, ld_path, pass_cuda_home_var=False)
-
     # search from nvidia-cuda-tileiras pip package
     logger.debug("Searching tileiras from nvidia pip package")
     res = _find_pip_tileiras()

test/conftest.py

@@ -12,7 +12,8 @@
 from functools import cache
 
 from cuda.tile._bytecode.version import BytecodeVersion
-from cuda.tile._compile import _get_max_supported_bytecode_version
+from cuda.tile._compile import _get_max_supported_bytecode_version, _SUPPORTED_VERSIONS
+from cuda.tile._cext import dev_features_enabled
 
 
 def pytest_addoption(parser):
@@ -45,11 +46,19 @@ def get_tileiras_version():
 
 def requires_tileiras(version: BytecodeVersion):
     """Skip test if tileiras version is lower than required."""
+
+    def vstr(v):
+        return f"{v.major()}.{v.minor()}"
+
+    if version not in _SUPPORTED_VERSIONS and not dev_features_enabled():
+        return pytest.mark.skip(
+            reason=f"Requires dev features enabled for version {vstr(version)}"
+        )
+
     current = get_tileiras_version()
     return pytest.mark.skipif(
         current < version,
-        reason=f"Requires tileiras {version.major()}.{version.minor()}, "
-               f"found {current.major()}.{current.minor()}"
+        reason=f"Requires tileiras {vstr(version)}, found {vstr(current)}"
     )