diff --git a/.github/workflows/create_release.yml b/.github/workflows/create_release.yml
index af63c8931233..076e8e114b73 100644
--- a/.github/workflows/create_release.yml
+++ b/.github/workflows/create_release.yml
@@ -43,30 +43,31 @@ jobs:
           tag_or_branch="${tag_or_branch#refs/heads/}"
           # replace directory separators with _ in branch name
           tag_or_branch="${tag_or_branch//\//_}"
+          if [[ ${tag_or_branch} == v* ]]; then
+            # strip trailing v from tag name
+            tag_or_branch="${tag_or_branch#v}"
+            # important: version must be fixed in setup.py
+            sed -i -e "s:^TRITON_VERSION = .*:TRITON_VERSION = '${tag_or_branch}':" setup.py || exit 1
+          fi
           echo "RELEASE_NAME=triton-$tag_or_branch" >> "$GITHUB_ENV"
-          echo "RELEASE_FILE=triton-$tag_or_branch.tar.gz" >> "$GITHUB_ENV"
       - name: Create source distribution
         run: |
-            # Create new folder with specified name so extracting the archive yields that
-            rm -rf "/tmp/$RELEASE_NAME"
-            cp -r "$PWD" "/tmp/$RELEASE_NAME"
-            mv "/tmp/$RELEASE_NAME" .
-            # Cleanup
-            find "$RELEASE_NAME" -name '.git*' -exec rm -rv {} \; || true
-            # Create archive
-            tar -czf "$RELEASE_FILE" "$RELEASE_NAME"
-            echo "Created source archive $RELEASE_FILE with content: $(ls -a "$RELEASE_NAME")"
+          pip install build || exit 1
+          python -m build -s || exit 1
+          cd dist || exit 1
+          release_file=( *.tar.gz )
+          echo "RELEASE_FILE=${release_file}" >> "$GITHUB_ENV"
       - name: Upload source distribution for release
         if: ${{ github.event_name == 'release' }}
         uses: softprops/action-gh-release@v2
         with:
-          files: ${{env.RELEASE_FILE}}
+          files: dist/${{env.RELEASE_FILE}}
       - name: Upload source distribution to GHA artifacts for release tags
         if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
         uses: actions/upload-artifact@v4.4.0
         with:
           name: ${{ env.RELEASE_FILE }}
-          path: ${{ env.RELEASE_FILE }}
+          path: dist/${{ env.RELEASE_FILE }}
       - name: Set output
         id: release_name
         run: echo "name=release_name::${{ env.RELEASE_NAME }}.tar.gz" >> "${GITHUB_OUTPUT}"
diff --git a/.github/workflows/wheels.yml b/.github/workflows/wheels.yml
index f516d6a3a85f..361effbfe39c 100644
--- a/.github/workflows/wheels.yml
+++ b/.github/workflows/wheels.yml
@@ -82,7 +82,12 @@ jobs:
           export CIBW_BUILD="cp3{9,10,11,12,13,13t}-manylinux_${{ matrix.config.arch }}"
           export CIBW_SKIP="cp{35,36,37,38}-*"
           export CIBW_FREE_THREADED_SUPPORT=1
-          python3 -m cibuildwheel python --output-dir wheelhouse
+          python3 -m cibuildwheel . --output-dir wheelhouse
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: cibw-wheels-manylinux_2_28_${{ matrix.config.arch }}-wheels-upload
+          path: ./wheelhouse/*.whl
 
       - name: Install Azure CLI
         if: ${{ steps.check-version.outputs.new_commit == 'true' }}
diff --git a/MANIFEST.in b/MANIFEST.in
new file mode 100644
index 000000000000..aadefd90e4e5
--- /dev/null
+++ b/MANIFEST.in
@@ -0,0 +1,18 @@
+graft bin
+graft cmake
+graft docs
+graft include
+graft lib
+graft python/src
+graft python/test
+graft python/triton/backends/amd
+graft python/triton/backends/nvidia
+graft python/triton/tools/extra/cuda
+graft test
+graft third_party
+graft unittest
+include CMakeLists.txt
+include Makefile
+include python/build_helpers.py
+include python/requirements.txt
+include python/test-requirements.txt
diff --git a/Makefile b/Makefile
index f984decf3194..ce3b44c9e649 100644
--- a/Makefile
+++ b/Makefile
@@ -79,7 +79,7 @@ dev-install-torch:
 
 .PHONY: dev-install-triton
 dev-install-triton:
-	$(PYTHON) -m pip install -e python --no-build-isolation -v
+	$(PYTHON) -m pip install -e . --no-build-isolation -v
 
 .PHONY: dev-install
 .NOPARALLEL: dev-install
diff --git a/README.md b/README.md
index 979fe434a758..a4fa113078c9 100644
--- a/README.md
+++ b/README.md
@@ -71,7 +71,7 @@ git clone https://github.com/triton-lang/triton.git
 cd triton
 
 pip install -r python/requirements.txt # build-time dependencies
-pip install -e python
+pip install -e .
 ```
 
 Or with a virtualenv:
@@ -84,7 +84,7 @@ python -m venv .venv --prompt triton
 source .venv/bin/activate
 
 pip install -r python/requirements.txt # build-time dependencies
-pip install -e python
+pip install -e .
 ```
 
 # Building with a custom LLVM
@@ -124,7 +124,7 @@ arbitrary LLVM version.
        $ LLVM_INCLUDE_DIRS=$LLVM_BUILD_DIR/include \
          LLVM_LIBRARY_DIR=$LLVM_BUILD_DIR/lib \
          LLVM_SYSPATH=$LLVM_BUILD_DIR \
-         pip install -e python
+         pip install -e .
 
 # Tips for building
 
@@ -139,7 +139,7 @@ arbitrary LLVM version.
   can be changed anytime.
 
 - If you're running out of memory when building Triton, specify the `MAX_JOBS`
-  environment variable (to the `pip install -e python` command) to limit the
+  environment variable (to the `pip install -e .` command) to limit the
   number of jobs.
 
 - Pass `--no-build-isolation` to `pip install` to make nop builds faster.
@@ -150,7 +150,7 @@ arbitrary LLVM version.
   (probably because, in our build, users don't invoke cmake directly, but
   instead use setup.py).  Teach vscode how to compile Triton as follows.
 
-    - Do a local build. Run command `pip install -e python`
+    - Do a local build. Run command `pip install -e .`
     - Get the full path to the `compile_commands.json` file produced by the build:
       `find python/build -name 'compile_commands.json' | xargs readlink -f`.
       You might get a full path similar to `/Users/{username}/triton/python/build/cmake.macosx-11.1-arm64-cpython-3.12/compile_commands.json`
diff --git a/pyproject.toml b/pyproject.toml
index 525f303efb6f..09dda337f1ce 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,5 +1,6 @@
 [build-system]
-requires = ["setuptools>=40.8.0", "wheel", "cmake>=3.18", "ninja>=1.11.1"]
+requires = ["setuptools>=40.8.0", "cmake>=3.18", "ninja>=1.11.1", "pybind11>=2.13.1"]
+build-backend = "setuptools.build_meta"
 
 [tool.yapf]
 based_on_style = "pep8"
@@ -9,6 +10,12 @@ each_dict_entry_on_separate_line=false
 split_before_named_assigns = false
 split_complex_comprehension = true
 
+# We're incrementally switching from autopep8 to ruff.
+[tool.autopep8]
+aggressive = 1
+ignore = "E501,E701,E731,W690,W503"
+max_line_length = 88
+
 [tool.ruff]
 line-length = 120
 
diff --git a/python/build_helpers.py b/python/build_helpers.py
index da2efd30cac3..98f7facc5772 100644
--- a/python/build_helpers.py
+++ b/python/build_helpers.py
@@ -12,6 +12,6 @@ def get_cmake_dir():
     plat_name = sysconfig.get_platform()
     python_version = sysconfig.get_python_version()
     dir_name = f"cmake.{plat_name}-{sys.implementation.name}-{python_version}"
-    cmake_dir = Path(get_base_dir()) / "python" / "build" / dir_name
+    cmake_dir = Path(get_base_dir()) / "build" / dir_name
     cmake_dir.mkdir(parents=True, exist_ok=True)
     return cmake_dir
diff --git a/python/pyproject.toml b/python/pyproject.toml
deleted file mode 100644
index d96af50a543e..000000000000
--- a/python/pyproject.toml
+++ /dev/null
@@ -1,15 +0,0 @@
-
-[build-system]
-requires = ["setuptools>=40.8.0", "wheel", "cmake>=3.18", "ninja>=1.11.1", "pybind11>=2.13.1"]
-
-# We're incrementally switching from autopep8 to ruff.
-[tool.autopep8]
-aggressive = 1
-ignore = "E501,E701,E731,W690,W503"
-max_line_length = 88
-
-[tool.ruff]
-line-length = 120
-
-[tool.ruff.lint]
-ignore = ["E501", "E701", "E731", "E741"]
diff --git a/python/triton/_C/include b/python/triton/_C/include
deleted file mode 120000
index b85a409837d1..000000000000
--- a/python/triton/_C/include
+++ /dev/null
@@ -1 +0,0 @@
-../../../include/
\ No newline at end of file
diff --git a/python/triton/backends/__init__.py b/python/triton/backends/__init__.py
index f2c46ea9e0fe..5725a6d03214 100644
--- a/python/triton/backends/__init__.py
+++ b/python/triton/backends/__init__.py
@@ -1,10 +1,15 @@
 import importlib
 import inspect
-import os
+import sys
 from dataclasses import dataclass
 from .driver import DriverBase
 from .compiler import BaseBackend
 
+if sys.version_info >= (3, 10):
+    from importlib.metadata import entry_points
+else:
+    from importlib_metadata import entry_points
+
 
 def _find_concrete_subclasses(module, base_class):
     ret = []
@@ -27,16 +32,11 @@ class Backend:
 
 def _discover_backends():
     backends = dict()
-    root = os.path.dirname(__file__)
-    for name in os.listdir(root):
-        if not os.path.isdir(os.path.join(root, name)):
-            continue
-        if name.startswith('__'):
-            continue
-        compiler = importlib.import_module(f"triton.backends.{name}.compiler")
-        driver = importlib.import_module(f"triton.backends.{name}.driver")
-        backends[name] = Backend(_find_concrete_subclasses(compiler, BaseBackend),
-                                 _find_concrete_subclasses(driver, DriverBase))
+    for ep in entry_points().select(group="triton.backends"):
+        compiler = importlib.import_module(f"{ep.value}.compiler")
+        driver = importlib.import_module(f"{ep.value}.driver")
+        backends[ep.name] = Backend(_find_concrete_subclasses(compiler, BaseBackend),
+                                    _find_concrete_subclasses(driver, DriverBase))
     return backends
 
 
diff --git a/python/setup.py b/setup.py
similarity index 74%
rename from python/setup.py
rename to setup.py
index 333dd58eabea..965b4c2d839d 100644
--- a/python/setup.py
+++ b/setup.py
@@ -14,29 +14,32 @@
 from io import BytesIO
 from distutils.command.clean import clean
 from pathlib import Path
-from typing import List, Optional
+from typing import Optional
 
-from setuptools import Extension, setup
+from setuptools import Extension, find_packages, setup
 from setuptools.command.build_ext import build_ext
 from setuptools.command.build_py import build_py
-from dataclasses import dataclass
-
-from distutils.command.install import install
 from setuptools.command.develop import develop
-from setuptools.command.egg_info import egg_info
-from wheel.bdist_wheel import bdist_wheel
+from dataclasses import dataclass
 
 import pybind11
 
-from build_helpers import get_base_dir, get_cmake_dir
+try:
+    from setuptools.command.editable_wheel import editable_wheel
+except ImportError:
+    # create a dummy class, since there is no command to override
+    class editable_wheel:
+        pass
+
+
+sys.path.insert(0, os.path.dirname(__file__))
+
+from python.build_helpers import get_base_dir, get_cmake_dir
 
 
 @dataclass
 class Backend:
     name: str
-    package_data: List[str]
-    language_package_data: List[str]
-    tools_package_data: List[str]
     src_dir: str
     backend_dir: str
     language_dir: Optional[str]
@@ -51,7 +54,7 @@ class BackendInstaller:
     def prepare(backend_name: str, backend_src_dir: str = None, is_external: bool = False):
         # Initialize submodule if there is one for in-tree backends.
         if not is_external:
-            root_dir = os.path.join(os.pardir, "third_party")
+            root_dir = "third_party"
             assert backend_name in os.listdir(
                 root_dir), f"{backend_name} is requested for install but not present in {root_dir}"
 
@@ -65,34 +68,24 @@ def prepare(backend_name: str, backend_src_dir: str = None, is_external: bool =
 
             backend_src_dir = os.path.join(root_dir, backend_name)
 
-        backend_path = os.path.abspath(os.path.join(backend_src_dir, "backend"))
+        backend_path = os.path.join(backend_src_dir, "backend")
         assert os.path.exists(backend_path), f"{backend_path} does not exist!"
 
-        language_dir = os.path.abspath(os.path.join(backend_src_dir, "language"))
+        language_dir = os.path.join(backend_src_dir, "language")
         if not os.path.exists(language_dir):
             language_dir = None
 
-        tools_dir = os.path.abspath(os.path.join(backend_src_dir, "tools"))
+        tools_dir = os.path.join(backend_src_dir, "tools")
         if not os.path.exists(tools_dir):
             tools_dir = None
 
         for file in ["compiler.py", "driver.py"]:
             assert os.path.exists(os.path.join(backend_path, file)), f"${file} does not exist in ${backend_path}"
 
-        install_dir = os.path.join(os.path.dirname(__file__), "triton", "backends", backend_name)
-        package_data = [f"{os.path.relpath(p, backend_path)}/*" for p, _, _, in os.walk(backend_path)]
-
-        language_package_data = []
-        if language_dir is not None:
-            language_package_data = [f"{os.path.relpath(p, language_dir)}/*" for p, _, _, in os.walk(language_dir)]
+        install_dir = os.path.join(os.path.dirname(__file__), "python", "triton", "backends", backend_name)
 
-        tools_package_data = []
-        if tools_dir is not None:
-            tools_package_data = [f"{os.path.relpath(p, tools_dir)}/*" for p, _, _, in os.walk(tools_dir)]
-
-        return Backend(name=backend_name, package_data=package_data, language_package_data=language_package_data,
-                       tools_package_data=tools_package_data, src_dir=backend_src_dir, backend_dir=backend_path,
-                       language_dir=language_dir, tools_dir=tools_dir, install_dir=install_dir, is_external=is_external)
+        return Backend(name=backend_name, src_dir=backend_src_dir, backend_dir=backend_path, language_dir=language_dir,
+                       tools_dir=tools_dir, install_dir=install_dir, is_external=is_external)
 
     # Copy all in-tree backends under triton/third_party.
     @staticmethod
@@ -270,7 +263,7 @@ def update_symlink(link_path, source_path):
 
     print(f"creating symlink: {link_path} -> {source_path}", file=sys.stderr)
     link_path.absolute().parent.mkdir(parents=True, exist_ok=True)  # Ensure link's parent directory exists
-    link_path.symlink_to(source_path, target_is_directory=True)
+    link_path.symlink_to(source_path.absolute(), target_is_directory=True)
 
 
 def get_thirdparty_packages(packages: list):
@@ -331,7 +324,7 @@ def download_and_copy(name, src_func, dst_path, variable, version, url_func):
     url = url_func(supported[system], arch, version)
     src_path = src_func(supported[system], arch, version)
     tmp_path = os.path.join(triton_cache_path, "nvidia", name)  # path to cache the download
-    dst_path = os.path.join(base_dir, os.pardir, "third_party", "nvidia", "backend", dst_path)  # final binary path
+    dst_path = os.path.join(base_dir, "third_party", "nvidia", "backend", dst_path)  # final binary path
     src_path = os.path.join(tmp_path, src_path)
     download = not os.path.exists(src_path)
     if os.path.exists(dst_path) and system == "Linux" and shutil.which(dst_path) is not None:
@@ -389,6 +382,8 @@ def finalize_options(self):
         build_ext.finalize_options(self)
 
     def run(self):
+        download_and_copy_dependencies()
+
         try:
             out = subprocess.check_output(["cmake", "--version"])
         except OSError:
@@ -511,80 +506,105 @@ def build_extension(self, ext):
         subprocess.check_call(["cmake", "--build", ".", "--target", "mlir-doc"], cwd=cmake_dir)
 
 
-nvidia_version_path = os.path.join(get_base_dir(), "cmake", "nvidia-toolchain-version.json")
-with open(nvidia_version_path, "r") as nvidia_version_file:
-    # parse this json file to get the version of the nvidia toolchain
-    NVIDIA_TOOLCHAIN_VERSION = json.load(nvidia_version_file)
-
-exe_extension = sysconfig.get_config_var("EXE")
-download_and_copy(
-    name="nvcc",
-    src_func=lambda system, arch, version: f"cuda_nvcc-{system}-{arch}-{version}-archive/bin/ptxas{exe_extension}",
-    dst_path="bin/ptxas",
-    variable="TRITON_PTXAS_PATH",
-    version=NVIDIA_TOOLCHAIN_VERSION["ptxas"],
-    url_func=lambda system, arch, version:
-    f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/{system}-{arch}/cuda_nvcc-{system}-{arch}-{version}-archive.tar.xz",
-)
-download_and_copy(
-    name="cuobjdump",
-    src_func=lambda system, arch, version:
-    f"cuda_cuobjdump-{system}-{arch}-{version}-archive/bin/cuobjdump{exe_extension}",
-    dst_path="bin/cuobjdump",
-    variable="TRITON_CUOBJDUMP_PATH",
-    version=NVIDIA_TOOLCHAIN_VERSION["cuobjdump"],
-    url_func=lambda system, arch, version:
-    f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_cuobjdump/{system}-{arch}/cuda_cuobjdump-{system}-{arch}-{version}-archive.tar.xz",
-)
-download_and_copy(
-    name="nvdisasm",
-    src_func=lambda system, arch, version:
-    f"cuda_nvdisasm-{system}-{arch}-{version}-archive/bin/nvdisasm{exe_extension}",
-    dst_path="bin/nvdisasm",
-    variable="TRITON_NVDISASM_PATH",
-    version=NVIDIA_TOOLCHAIN_VERSION["nvdisasm"],
-    url_func=lambda system, arch, version:
-    f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvdisasm/{system}-{arch}/cuda_nvdisasm-{system}-{arch}-{version}-archive.tar.xz",
-)
-download_and_copy(
-    name="nvcc",
-    src_func=lambda system, arch, version: f"cuda_nvcc-{system}-{arch}-{version}-archive/include",
-    dst_path="include",
-    variable="TRITON_CUDACRT_PATH",
-    version=NVIDIA_TOOLCHAIN_VERSION["cudacrt"],
-    url_func=lambda system, arch, version:
-    f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/{system}-{arch}/cuda_nvcc-{system}-{arch}-{version}-archive.tar.xz",
-)
-download_and_copy(
-    name="cudart",
-    src_func=lambda system, arch, version: f"cuda_cudart-{system}-{arch}-{version}-archive/include",
-    dst_path="include",
-    variable="TRITON_CUDART_PATH",
-    version=NVIDIA_TOOLCHAIN_VERSION["cudart"],
-    url_func=lambda system, arch, version:
-    f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/{system}-{arch}/cuda_cudart-{system}-{arch}-{version}-archive.tar.xz",
-)
-download_and_copy(
-    name="cupti",
-    src_func=lambda system, arch, version: f"cuda_cupti-{system}-{arch}-{version}-archive/include",
-    dst_path="include",
-    variable="TRITON_CUPTI_INCLUDE_PATH",
-    version=NVIDIA_TOOLCHAIN_VERSION["cupti"],
-    url_func=lambda system, arch, version:
-    f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_cupti/{system}-{arch}/cuda_cupti-{system}-{arch}-{version}-archive.tar.xz",
-)
-download_and_copy(
-    name="cupti",
-    src_func=lambda system, arch, version: f"cuda_cupti-{system}-{arch}-{version}-archive/lib",
-    dst_path="lib/cupti",
-    variable="TRITON_CUPTI_LIB_PATH",
-    version=NVIDIA_TOOLCHAIN_VERSION["cupti"],
-    url_func=lambda system, arch, version:
-    f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_cupti/{system}-{arch}/cuda_cupti-{system}-{arch}-{version}-archive.tar.xz",
-)
+def download_and_copy_dependencies():
+    nvidia_version_path = os.path.join(get_base_dir(), "cmake", "nvidia-toolchain-version.json")
+    with open(nvidia_version_path, "r") as nvidia_version_file:
+        # parse this json file to get the version of the nvidia toolchain
+        NVIDIA_TOOLCHAIN_VERSION = json.load(nvidia_version_file)
+
+    exe_extension = sysconfig.get_config_var("EXE")
+    download_and_copy(
+        name="nvcc",
+        src_func=lambda system, arch, version: f"cuda_nvcc-{system}-{arch}-{version}-archive/bin/ptxas{exe_extension}",
+        dst_path="bin/ptxas",
+        variable="TRITON_PTXAS_PATH",
+        version=NVIDIA_TOOLCHAIN_VERSION["ptxas"],
+        url_func=lambda system, arch, version:
+        f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/{system}-{arch}/cuda_nvcc-{system}-{arch}-{version}-archive.tar.xz",
+    )
+    download_and_copy(
+        name="cuobjdump",
+        src_func=lambda system, arch, version:
+        f"cuda_cuobjdump-{system}-{arch}-{version}-archive/bin/cuobjdump{exe_extension}",
+        dst_path="bin/cuobjdump",
+        variable="TRITON_CUOBJDUMP_PATH",
+        version=NVIDIA_TOOLCHAIN_VERSION["cuobjdump"],
+        url_func=lambda system, arch, version:
+        f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_cuobjdump/{system}-{arch}/cuda_cuobjdump-{system}-{arch}-{version}-archive.tar.xz",
+    )
+    download_and_copy(
+        name="nvdisasm",
+        src_func=lambda system, arch, version:
+        f"cuda_nvdisasm-{system}-{arch}-{version}-archive/bin/nvdisasm{exe_extension}",
+        dst_path="bin/nvdisasm",
+        variable="TRITON_NVDISASM_PATH",
+        version=NVIDIA_TOOLCHAIN_VERSION["nvdisasm"],
+        url_func=lambda system, arch, version:
+        f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvdisasm/{system}-{arch}/cuda_nvdisasm-{system}-{arch}-{version}-archive.tar.xz",
+    )
+    download_and_copy(
+        name="nvcc",
+        src_func=lambda system, arch, version: f"cuda_nvcc-{system}-{arch}-{version}-archive/include",
+        dst_path="include",
+        variable="TRITON_CUDACRT_PATH",
+        version=NVIDIA_TOOLCHAIN_VERSION["cudacrt"],
+        url_func=lambda system, arch, version:
+        f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_nvcc/{system}-{arch}/cuda_nvcc-{system}-{arch}-{version}-archive.tar.xz",
+    )
+    download_and_copy(
+        name="cudart",
+        src_func=lambda system, arch, version: f"cuda_cudart-{system}-{arch}-{version}-archive/include",
+        dst_path="include",
+        variable="TRITON_CUDART_PATH",
+        version=NVIDIA_TOOLCHAIN_VERSION["cudart"],
+        url_func=lambda system, arch, version:
+        f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_cudart/{system}-{arch}/cuda_cudart-{system}-{arch}-{version}-archive.tar.xz",
+    )
+    download_and_copy(
+        name="cupti",
+        src_func=lambda system, arch, version: f"cuda_cupti-{system}-{arch}-{version}-archive/include",
+        dst_path="include",
+        variable="TRITON_CUPTI_INCLUDE_PATH",
+        version=NVIDIA_TOOLCHAIN_VERSION["cupti"],
+        url_func=lambda system, arch, version:
+        f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_cupti/{system}-{arch}/cuda_cupti-{system}-{arch}-{version}-archive.tar.xz",
+    )
+    download_and_copy(
+        name="cupti",
+        src_func=lambda system, arch, version: f"cuda_cupti-{system}-{arch}-{version}-archive/lib",
+        dst_path="lib/cupti",
+        variable="TRITON_CUPTI_LIB_PATH",
+        version=NVIDIA_TOOLCHAIN_VERSION["cupti"],
+        url_func=lambda system, arch, version:
+        f"https://developer.download.nvidia.com/compute/cuda/redist/cuda_cupti/{system}-{arch}/cuda_cupti-{system}-{arch}-{version}-archive.tar.xz",
+    )
+
+
 backends = [*BackendInstaller.copy(["nvidia", "amd"]), *BackendInstaller.copy_externals()]
 
 
+def get_package_dirs():
+    yield ("", "python")
+
+    for backend in backends:
+        yield (f"triton.backends.{backend.name}", backend.backend_dir)
+
+        if backend.language_dir:
+            # Install the contents of each backend's `language` directory into
+            # `triton.language.extra`.
+            for x in os.listdir(backend.language_dir):
+                yield (f"triton.language.extra.{x}", os.path.join(backend.language_dir, x))
+
+        if backend.tools_dir:
+            # Install the contents of each backend's `tools` directory into
+            # `triton.tools.extra`.
+            for x in os.listdir(backend.tools_dir):
+                yield (f"triton.tools.extra.{x}", os.path.join(backend.tools_dir, x))
+
+    if check_env_flag("TRITON_BUILD_PROTON", "ON"):  # Default ON
+        yield ("triton.profiler", "third_party/proton/proton")
+
+
 def add_link_to_backends():
     for backend in backends:
         update_symlink(backend.install_dir, backend.backend_dir)
@@ -592,7 +612,8 @@ def add_link_to_backends():
         if backend.language_dir:
             # Link the contents of each backend's `language` directory into
             # `triton.language.extra`.
-            extra_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "triton", "language", "extra"))
+            extra_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "python", "triton", "language",
+                                                     "extra"))
             for x in os.listdir(backend.language_dir):
                 src_dir = os.path.join(backend.language_dir, x)
                 install_dir = os.path.join(extra_dir, x)
@@ -601,7 +622,7 @@ def add_link_to_backends():
         if backend.tools_dir:
             # Link the contents of each backend's `tools` directory into
             # `triton.tools.extra`.
-            extra_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "triton", "tools", "extra"))
+            extra_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "python", "triton", "tools", "extra"))
             for x in os.listdir(backend.tools_dir):
                 src_dir = os.path.join(backend.tools_dir, x)
                 install_dir = os.path.join(extra_dir, x)
@@ -609,8 +630,8 @@ def add_link_to_backends():
 
 
 def add_link_to_proton():
-    proton_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), os.pardir, "third_party", "proton", "proton"))
-    proton_install_dir = os.path.join(os.path.dirname(__file__), "triton", "profiler")
+    proton_dir = os.path.abspath(os.path.join(os.path.dirname(__file__), "third_party", "proton", "proton"))
+    proton_install_dir = os.path.join(os.path.dirname(__file__), "python", "triton", "profiler")
     update_symlink(proton_install_dir, proton_dir)
 
 
@@ -620,84 +641,18 @@ def add_links():
         add_link_to_proton()
 
 
-class plugin_install(install):
-
-    def run(self):
-        add_links()
-        install.run(self)
-
-
 class plugin_develop(develop):
 
     def run(self):
         add_links()
-        develop.run(self)
-
-
-class plugin_bdist_wheel(bdist_wheel):
-
-    def run(self):
-        add_links()
-        bdist_wheel.run(self)
+        super().run()
 
 
-class plugin_egginfo(egg_info):
+class plugin_editable_wheel(editable_wheel):
 
     def run(self):
         add_links()
-        egg_info.run(self)
-
-
-package_data = {
-    "triton/tools/extra": sum((b.tools_package_data for b in backends), []),
-    **{f"triton/backends/{b.name}": b.package_data
-       for b in backends}, "triton/language/extra": sum((b.language_package_data for b in backends), [])
-}
-
-
-def get_extra_packages(extra_name):
-    packages = []
-    extra_file_extensions = {"language": (".py"), "tools": (".c", ".h", ".cpp")}
-    assert extra_name in extra_file_extensions, f"{extra_name} extra is not valid"
-
-    for backend in backends:
-        backend_extra_dir = getattr(backend, f"{extra_name}_dir", None)
-        if backend_extra_dir is None:
-            continue
-
-        # Walk the specified directory of each backend to enumerate
-        # any subpackages, which will be added to extra_package.
-        for dir, dirs, files in os.walk(backend_extra_dir, followlinks=True):
-            if not any(f for f in files if f.endswith(extra_file_extensions[extra_name])) or dir == backend_extra_dir:
-                # Ignore directories with no relevant files
-                # or the root directory
-                continue
-            subpackage = os.path.relpath(dir, backend_extra_dir)
-            package = os.path.join(f"triton/{extra_name}/extra", subpackage)
-            packages.append(package)
-
-    return list(packages)
-
-
-def get_packages():
-    packages = [
-        "triton",
-        "triton/_C",
-        "triton/compiler",
-        "triton/language",
-        "triton/language/extra",
-        "triton/runtime",
-        "triton/backends",
-        "triton/tools",
-        "triton/tools/extra",
-    ]
-    packages += [f'triton/backends/{backend.name}' for backend in backends]
-    packages += get_extra_packages("language")
-    packages += get_extra_packages("tools")
-    if check_env_flag("TRITON_BUILD_PROTON", "ON"):  # Default ON
-        packages += ["triton/profiler"]
-
-    return packages
+        super().run()
 
 
 def get_entry_points():
@@ -707,6 +662,7 @@ def get_entry_points():
             "proton-viewer = triton.profiler.viewer:main",
             "proton = triton.profiler.proton:main",
         ]
+    entry_points["triton.backends"] = [f"{b.name} = triton.backends.{b.name}" for b in backends]
     return entry_points
 
 
@@ -734,27 +690,34 @@ def get_git_version_suffix():
         return get_git_commit_hash()
 
 
+# keep it separate for easy substitution
+TRITON_VERSION = "3.3.0" + get_git_version_suffix() + os.environ.get("TRITON_WHEEL_VERSION_SUFFIX", "")
+
+package_dirs = dict(get_package_dirs())
+extra_packages = [x for x in package_dirs if x != ""]
+
 setup(
     name=os.environ.get("TRITON_WHEEL_NAME", "triton"),
-    version="3.3.0" + get_git_version_suffix() + os.environ.get("TRITON_WHEEL_VERSION_SUFFIX", ""),
+    version=TRITON_VERSION,
     author="Philippe Tillet",
     author_email="phil@openai.com",
     description="A language and compiler for custom Deep Learning operations",
     long_description="",
-    install_requires=["setuptools>=40.8.0"],
-    packages=get_packages(),
+    install_requires=[
+        "setuptools>=40.8.0",
+        "importlib-metadata; python_version < '3.10'",
+    ],
+    packages=find_packages(where="python") + extra_packages,
+    package_dir=package_dirs,
     entry_points=get_entry_points(),
-    package_data=package_data,
     include_package_data=True,
     ext_modules=[CMakeExtension("triton", "triton/_C/")],
     cmdclass={
         "build_ext": CMakeBuild,
         "build_py": CMakeBuildPy,
         "clean": CMakeClean,
-        "install": plugin_install,
         "develop": plugin_develop,
-        "bdist_wheel": plugin_bdist_wheel,
-        "egg_info": plugin_egginfo,
+        "editable_wheel": plugin_editable_wheel,
     },
     zip_safe=False,
     # for PyPI
diff --git a/python/MANIFEST.in b/third_party/amd/backend/__init__.py
similarity index 100%
rename from python/MANIFEST.in
rename to third_party/amd/backend/__init__.py
diff --git a/third_party/proton/proton/_C/include b/third_party/proton/proton/_C/include
deleted file mode 120000
index fe4f4a1aa9bd..000000000000
--- a/third_party/proton/proton/_C/include
+++ /dev/null
@@ -1 +0,0 @@
-../../csrc/include/
\ No newline at end of file