Bump to include user-defined autotuner fix and load-with-other codegen workaround

It is now always added by Triton's own setup.py (see
https://github.com/triton-lang/triton/pull/4812/)

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-cf34004b8a67d290a962da166f5aa2fc66751326
+35c6c7c6284582b3f41c71c150e11b517acf074a

.ci/docker/triton_version.txt

@@ -1 +1 @@
-3.1.0
+3.2.0

.github/scripts/build_triton_wheel.py

@@ -48,25 +48,6 @@ def patch_init_py(
         f.write(orig)
 
 
-# TODO: remove patch_setup_py() once we have a proper fix for https://github.com/triton-lang/triton/issues/4527
-def patch_setup_py(path: Path) -> None:
-    with open(path) as f:
-        orig = f.read()
-    try:
-        orig = check_and_replace(
-            orig,
-            "https://tritonlang.blob.core.windows.net/llvm-builds/",
-            "https://oaitriton.blob.core.windows.net/public/llvm-builds/",
-        )
-        with open(path, "w") as f:
-            f.write(orig)
-    except RuntimeError as e:
-        print(
-            f"Applying patch_setup_py() for llvm-build package failed: {e}.",
-            "If you are trying to build a newer version of Triton, you can ignore this.",
-        )
-
-
 def build_triton(
     *,
     version: str,
@@ -81,13 +62,6 @@ def build_triton(
         max_jobs = os.cpu_count() or 1
         env["MAX_JOBS"] = str(max_jobs)
 
-    version_suffix = ""
-    if not release:
-        # Nightly binaries include the triton commit hash, i.e. 2.1.0+e6216047b8
-        # while release build should only include the version, i.e. 2.1.0
-        version_suffix = f"+{commit_hash[:10]}"
-        version += version_suffix
-
     with TemporaryDirectory() as tmpdir:
         triton_basedir = Path(tmpdir) / "triton"
         triton_pythondir = triton_basedir / "python"
@@ -108,9 +82,6 @@ def build_triton(
         else:
             check_call(["git", "checkout", commit_hash], cwd=triton_basedir)
 
-        # TODO: remove this and patch_setup_py() once we have a proper fix for https://github.com/triton-lang/triton/issues/4527
-        patch_setup_py(triton_pythondir / "setup.py")
-
         if build_conda:
             with open(triton_basedir / "meta.yaml", "w") as meta:
                 print(
@@ -125,8 +96,8 @@ def build_triton(
                     file=meta,
                 )
                 print(
-                    "requirements:\n  host:\n    - python\n    - setuptools\n  run:\n    - python\n"
-                    "    - filelock\n    - pytorch\n",
+                    "requirements:\n  host:\n    - python\n    - setuptools\n    - pybind11\n"
+                    "  run:\n    - python\n    - filelock\n    - pytorch\n",
                     file=meta,
                 )
                 print(
@@ -162,7 +133,8 @@ def build_triton(
 
         # change built wheel name and version
         env["TRITON_WHEEL_NAME"] = triton_pkg_name
-        env["TRITON_WHEEL_VERSION_SUFFIX"] = version_suffix
+        if device == "cuda":
+            env["TRITON_BUILD_WITH_CLANG_LLD"] = "1"
         patch_init_py(
             triton_pythondir / "triton" / "__init__.py",
             version=f"{version}",

.github/workflows/build-triton-wheel.yml

@@ -53,7 +53,7 @@ jobs:
             rocm_version: ""
     timeout-minutes: 40
     env:
-      DOCKER_IMAGE: ${{ matrix.device == 'rocm' && format('pytorch/manylinux-builder:rocm{0}', matrix.rocm_version) || 'pytorch/manylinux-builder:cpu' }}
+      DOCKER_IMAGE: ${{ matrix.device == 'rocm' && format('pytorch/manylinux-builder:rocm{0}', matrix.rocm_version) || 'pytorch/manylinux2_28-builder:cpu' }}
       PY_VERS: ${{ matrix.py_vers }}
       BUILD_DEVICE: ${{ matrix.device }}
     steps:
@@ -119,14 +119,12 @@ jobs:
           fi
 
           docker exec -t "${container_name}" yum install -y zlib-devel zip
-          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}"  -m pip install -U setuptools==67.4.0
-          # Triton xpu build use GCC11
-          if [[ "${BUILD_DEVICE}" == xpu ]]; then
-            docker exec -t "${container_name}" yum install -y devtoolset-11-gcc-c++
-            docker exec -t "${container_name}" bash -c "source /opt/rh/devtoolset-11/enable && ${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE"
-          else
-            docker exec -t "${container_name}" bash -c "${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE"
+          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}"  -m pip install -U setuptools==67.4.0 pybind11==2.13.1
+          if [[ "${{ matrix.device }}" == "cuda" ]]; then
+            # With this install, it gets clang 16.0.6.
+            docker exec -t "${container_name}" dnf install clang lld -y
           fi
+          docker exec -t "${container_name}" bash -c "${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE"
           docker exec -t "${container_name}" chown -R 1000.1000 /artifacts
 
       - uses: actions/upload-artifact@v4.4.0