Update (base update)

[ghstack-poisoned]

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -3,18 +3,8 @@ set -eux -o pipefail
 
 GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
-# Set CUDA architecture lists to match x86 build_cuda.sh
-if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
-elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
+if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
-elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
-fi
-
-# Compress the fatbin with -compress-mode=size for CUDA 13
-if [[ "$DESIRED_CUDA" == *"13"* ]]; then
-    export TORCH_NVCC_FLAGS="-compress-mode=size"
 fi
 
 SCRIPTPATH="$( cd -- "$(dirname "$0")" >/dev/null 2>&1 ; pwd -P )"
@@ -28,22 +18,14 @@ cd /
 # on the mounted pytorch repo
 git config --global --add safe.directory /pytorch
 pip install -r /pytorch/requirements.txt
-pip install auditwheel==6.2.0 wheel
+pip install auditwheel==6.2.0
 if [ "$DESIRED_CUDA" = "cpu" ]; then
     echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
+    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
+    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
 else
     echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
     export USE_SYSTEM_NCCL=1
-
-    # Check if we should use NVIDIA libs from PyPI (similar to x86 build_cuda.sh logic)
-    if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
-        echo "Bundling CUDA libraries with wheel for aarch64."
-    else
-        echo "Using nvidia libs from pypi for aarch64."
-        echo "Updated PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64: $PYTORCH_EXTRA_INSTALL_REQUIREMENTS"
-        export USE_NVIDIA_PYPI_LIBS=1
-    fi
-
-    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
+    #USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
+    USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -69,186 +69,62 @@ def replace_tag(filename) -> None:
         f.writelines(lines)
 
 
-def patch_library_rpath(
-    folder: str,
-    lib_name: str,
-    use_nvidia_pypi_libs: bool = False,
-    desired_cuda: str = "",
-) -> None:
-    """Apply patchelf to set RPATH for a library in torch/lib"""
-    lib_path = f"{folder}/tmp/torch/lib/{lib_name}"
-
-    if use_nvidia_pypi_libs:
-        # For PyPI NVIDIA libraries, construct CUDA RPATH
-        cuda_rpaths = [
-            "$ORIGIN/../../nvidia/cudnn/lib",
-            "$ORIGIN/../../nvidia/nvshmem/lib",
-            "$ORIGIN/../../nvidia/nccl/lib",
-            "$ORIGIN/../../nvidia/cusparselt/lib",
-        ]
-
-        if "130" in desired_cuda:
-            cuda_rpaths.append("$ORIGIN/../../nvidia/cu13/lib")
-        else:
-            cuda_rpaths.extend(
-                [
-                    "$ORIGIN/../../nvidia/cublas/lib",
-                    "$ORIGIN/../../nvidia/cuda_cupti/lib",
-                    "$ORIGIN/../../nvidia/cuda_nvrtc/lib",
-                    "$ORIGIN/../../nvidia/cuda_runtime/lib",
-                    "$ORIGIN/../../nvidia/cufft/lib",
-                    "$ORIGIN/../../nvidia/curand/lib",
-                    "$ORIGIN/../../nvidia/cusolver/lib",
-                    "$ORIGIN/../../nvidia/cusparse/lib",
-                    "$ORIGIN/../../nvidia/nvtx/lib",
-                    "$ORIGIN/../../nvidia/cufile/lib",
-                ]
-            )
-
-        # Add $ORIGIN for local torch libs
-        rpath = ":".join(cuda_rpaths) + ":$ORIGIN"
-    else:
-        # For bundled libraries, just use $ORIGIN
-        rpath = "$ORIGIN"
-
-    if os.path.exists(lib_path):
-        os.system(
-            f"cd {folder}/tmp/torch/lib/; "
-            f"patchelf --set-rpath '{rpath}' --force-rpath {lib_name}"
-        )
-
-
-def copy_and_patch_library(
-    src_path: str,
-    folder: str,
-    use_nvidia_pypi_libs: bool = False,
-    desired_cuda: str = "",
-) -> None:
-    """Copy a library to torch/lib and patch its RPATH"""
-    if os.path.exists(src_path):
-        lib_name = os.path.basename(src_path)
-        shutil.copy2(src_path, f"{folder}/tmp/torch/lib/{lib_name}")
-        patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
-
-
 def package_cuda_wheel(wheel_path, desired_cuda) -> None:
     """
     Package the cuda wheel libraries
     """
     folder = os.path.dirname(wheel_path)
+    wheelname = os.path.basename(wheel_path)
     os.mkdir(f"{folder}/tmp")
     os.system(f"unzip {wheel_path} -d {folder}/tmp")
-    # Delete original wheel since it will be repackaged
-    os.system(f"rm {wheel_path}")
+    libs_to_copy = [
+        "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
+        "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
+        "/usr/local/cuda/lib64/libcudnn.so.9",
+        "/usr/local/cuda/lib64/libcublas.so.12",
+        "/usr/local/cuda/lib64/libcublasLt.so.12",
+        "/usr/local/cuda/lib64/libcudart.so.12",
+        "/usr/local/cuda/lib64/libcufft.so.11",
+        "/usr/local/cuda/lib64/libcusparse.so.12",
+        "/usr/local/cuda/lib64/libcusparseLt.so.0",
+        "/usr/local/cuda/lib64/libcusolver.so.11",
+        "/usr/local/cuda/lib64/libcurand.so.10",
+        "/usr/local/cuda/lib64/libnccl.so.2",
+        "/usr/local/cuda/lib64/libnvJitLink.so.12",
+        "/usr/local/cuda/lib64/libnvrtc.so.12",
+        "/usr/local/cuda/lib64/libnvshmem_host.so.3",
+        "/usr/local/cuda/lib64/libcudnn_adv.so.9",
+        "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
+        "/usr/local/cuda/lib64/libcudnn_graph.so.9",
+        "/usr/local/cuda/lib64/libcudnn_ops.so.9",
+        "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9",
+        "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9",
+        "/usr/local/cuda/lib64/libcudnn_heuristic.so.9",
+        "/lib64/libgomp.so.1",
+        "/usr/lib64/libgfortran.so.5",
+        "/acl/build/libarm_compute.so",
+        "/acl/build/libarm_compute_graph.so",
+        "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
+        "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
+        "/usr/local/lib/libnvpl_lapack_core.so.0",
+        "/usr/local/lib/libnvpl_blas_core.so.0",
+    ]
 
-    # Check if we should use PyPI NVIDIA libraries or bundle system libraries
-    use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
-
-    if use_nvidia_pypi_libs:
-        print("Using nvidia libs from pypi - skipping CUDA library bundling")
-        # For PyPI approach, we don't bundle CUDA libraries - they come from PyPI packages
-        # We only need to bundle non-NVIDIA libraries
-        minimal_libs_to_copy = [
-            "/lib64/libgomp.so.1",
-            "/usr/lib64/libgfortran.so.5",
-            "/acl/build/libarm_compute.so",
-            "/acl/build/libarm_compute_graph.so",
-            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
-            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
-            "/usr/local/lib/libnvpl_lapack_core.so.0",
-            "/usr/local/lib/libnvpl_blas_core.so.0",
-        ]
-
-        # Copy minimal libraries to unzipped_folder/torch/lib
-        for lib_path in minimal_libs_to_copy:
-            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
-
-        # Patch torch libraries used for searching libraries
-        torch_libs_to_patch = [
-            "libtorch.so",
-            "libtorch_cpu.so",
-            "libtorch_cuda.so",
-            "libtorch_cuda_linalg.so",
-            "libtorch_global_deps.so",
-            "libtorch_python.so",
-            "libtorch_nvshmem.so",
-            "libc10.so",
-            "libc10_cuda.so",
-            "libcaffe2_nvrtc.so",
-            "libshm.so",
-        ]
-        for lib_name in torch_libs_to_patch:
-            patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
-    else:
-        print("Bundling CUDA libraries with wheel")
-        # Original logic for bundling system CUDA libraries
-        # Common libraries for all CUDA versions
-        common_libs = [
-            # Non-NVIDIA system libraries
-            "/lib64/libgomp.so.1",
-            "/usr/lib64/libgfortran.so.5",
-            "/acl/build/libarm_compute.so",
-            "/acl/build/libarm_compute_graph.so",
-            # Common CUDA libraries (same for all versions)
-            "/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
-            "/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
-            "/usr/local/lib/libnvpl_lapack_core.so.0",
-            "/usr/local/lib/libnvpl_blas_core.so.0",
-            "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
-            "/usr/local/cuda/lib64/libcudnn.so.9",
-            "/usr/local/cuda/lib64/libcusparseLt.so.0",
-            "/usr/local/cuda/lib64/libcurand.so.10",
-            "/usr/local/cuda/lib64/libnccl.so.2",
-            "/usr/local/cuda/lib64/libnvshmem_host.so.3",
-            "/usr/local/cuda/lib64/libcudnn_adv.so.9",
-            "/usr/local/cuda/lib64/libcudnn_cnn.so.9",
-            "/usr/local/cuda/lib64/libcudnn_graph.so.9",
-            "/usr/local/cuda/lib64/libcudnn_ops.so.9",
-            "/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9",
-            "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9",
-            "/usr/local/cuda/lib64/libcudnn_heuristic.so.9",
+    if "129" in desired_cuda:
+        libs_to_copy += [
+            "/usr/local/cuda/lib64/libnvrtc-builtins.so.12.9",
             "/usr/local/cuda/lib64/libcufile.so.0",
             "/usr/local/cuda/lib64/libcufile_rdma.so.1",
-            "/usr/local/cuda/lib64/libcusparse.so.12",
         ]
 
-        # CUDA version-specific libraries
-        if "13" in desired_cuda:
-            minor_version = desired_cuda[-1]
-            version_specific_libs = [
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
-                "/usr/local/cuda/lib64/libcublas.so.13",
-                "/usr/local/cuda/lib64/libcublasLt.so.13",
-                "/usr/local/cuda/lib64/libcudart.so.13",
-                "/usr/local/cuda/lib64/libcufft.so.12",
-                "/usr/local/cuda/lib64/libcusolver.so.12",
-                "/usr/local/cuda/lib64/libnvJitLink.so.13",
-                "/usr/local/cuda/lib64/libnvrtc.so.13",
-                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.13.{minor_version}",
-            ]
-        elif "12" in desired_cuda:
-            # Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
-            minor_version = desired_cuda[-1]
-            version_specific_libs = [
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
-                "/usr/local/cuda/lib64/libcublas.so.12",
-                "/usr/local/cuda/lib64/libcublasLt.so.12",
-                "/usr/local/cuda/lib64/libcudart.so.12",
-                "/usr/local/cuda/lib64/libcufft.so.11",
-                "/usr/local/cuda/lib64/libcusolver.so.11",
-                "/usr/local/cuda/lib64/libnvJitLink.so.12",
-                "/usr/local/cuda/lib64/libnvrtc.so.12",
-                f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
-            ]
-        else:
-            raise ValueError(f"Unsupported CUDA version: {desired_cuda}.")
-
-        # Combine all libraries
-        libs_to_copy = common_libs + version_specific_libs
-
-        # Copy libraries to unzipped_folder/torch/lib
-        for lib_path in libs_to_copy:
-            copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
+    # Copy libraries to unzipped_folder/a/lib
+    for lib_path in libs_to_copy:
+        lib_name = os.path.basename(lib_path)
+        shutil.copy2(lib_path, f"{folder}/tmp/torch/lib/{lib_name}")
+        os.system(
+            f"cd {folder}/tmp/torch/lib/; "
+            f"patchelf --set-rpath '$ORIGIN' --force-rpath {folder}/tmp/torch/lib/{lib_name}"
+        )
 
     # Make sure the wheel is tagged with manylinux_2_28
     for f in os.scandir(f"{folder}/tmp/"):
@@ -256,8 +132,14 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
             replace_tag(f"{f.path}/WHEEL")
             break
 
-    os.system(f"wheel pack {folder}/tmp/ -d {folder}")
-    os.system(f"rm -rf {folder}/tmp/")
+    os.mkdir(f"{folder}/cuda_wheel")
+    os.system(f"cd {folder}/tmp/; zip -r {folder}/cuda_wheel/{wheelname} *")
+    shutil.move(
+        f"{folder}/cuda_wheel/{wheelname}",
+        f"{folder}/{wheelname}",
+        copy_function=shutil.copy2,
+    )
+    os.system(f"rm -rf {folder}/tmp/ {folder}/cuda_wheel/")
 
 
 def complete_wheel(folder: str) -> str:
@@ -280,7 +162,14 @@ def complete_wheel(folder: str) -> str:
             f"/{folder}/dist/{repaired_wheel_name}",
         )
     else:
-        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
+        repaired_wheel_name = wheel_name.replace(
+            "linux_aarch64", "manylinux_2_28_aarch64"
+        )
+        print(f"Renaming {wheel_name} wheel to {repaired_wheel_name}")
+        os.rename(
+            f"/{folder}/dist/{wheel_name}",
+            f"/{folder}/dist/{repaired_wheel_name}",
+        )
 
     print(f"Copying {repaired_wheel_name} to artifacts")
     shutil.copy2(
@@ -317,21 +206,11 @@ if __name__ == "__main__":
     ).decode()
 
     print("Building PyTorch wheel")
-    build_vars = ""
+    build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
     # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
     if enable_cuda:
         build_vars += "MAX_JOBS=5 "
 
-        # Handle PyPI NVIDIA libraries vs bundled libraries
-        use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
-        if use_nvidia_pypi_libs:
-            print("Configuring build for PyPI NVIDIA libraries")
-            # Configure for dynamic linking (matching x86 logic)
-            build_vars += "ATEN_STATIC_CUDA=0 USE_CUDA_STATIC_LINK=0 USE_CUPTI_SO=1 "
-        else:
-            print("Configuring build for bundled NVIDIA libraries")
-            # Keep existing static linking approach - already configured above
-
     override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
     desired_cuda = os.getenv("DESIRED_CUDA")
     if override_package_version is not None:

.ci/aarch64_linux/build_aarch64_wheel.py

@@ -241,7 +241,7 @@ def wait_for_connection(addr, port, timeout=15, attempt_cnt=5):
         try:
             with socket.create_connection((addr, port), timeout=timeout):
                 return
-        except (ConnectionRefusedError, TimeoutError):  # noqa: PERF203
+        except (ConnectionRefusedError, socket.timeout):  # noqa: PERF203
             if i == attempt_cnt - 1:
                 raise
             time.sleep(timeout)

.ci/docker/README.md

@@ -120,8 +120,8 @@ If your new Docker image needs a library installed from a specific pinned commit
    If you're introducing a new argument to the Docker build, make sure to add it in the Docker build step in `.ci/docker/build.sh`:
    ```bash
    docker build \
-     ....
-     --build-arg "NEW_ARG_1=${NEW_ARG_1}"
+      ....
+      --build-arg "NEW_ARG_1=${NEW_ARG_1}"
    ```
 
 3. **Update Dockerfile logic**:

.ci/docker/build.sh

@@ -81,8 +81,8 @@ elif [[ "$image" == *riscv* ]]; then
   DOCKERFILE="ubuntu-cross-riscv/Dockerfile"
 fi
 
-_UCX_COMMIT=7836b165abdbe468a2f607e7254011c07d788152
-_UCC_COMMIT=430e241bf5d38cbc73fc7a6b89155397232e3f96
+_UCX_COMMIT=7bb2722ff2187a0cad557ae4a6afa090569f83fb
+_UCC_COMMIT=20eae37090a4ce1b32bcce6144ccad0b49943e0b
 if [[ "$image" == *rocm* ]]; then
   _UCX_COMMIT=cc312eaa4655c0cc5c2bcd796db938f90563bcf6
   _UCC_COMMIT=0c0fc21559835044ab107199e334f7157d6a0d3d
@@ -114,19 +114,31 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
     ;;
-  pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
-    CUDA_VERSION=13.0.0
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.8.1
     ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
+    GCC_VERSION=9
     VISION=yes
     KATEX=yes
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks)
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks)
     CUDA_VERSION=12.8.1
-    ANACONDA_PYTHON_VERSION=3.10
+    ANACONDA_PYTHON_VERSION=3.12
+    GCC_VERSION=9
+    VISION=yes
+    KATEX=yes
+    UCX_COMMIT=${_UCX_COMMIT}
+    UCC_COMMIT=${_UCC_COMMIT}
+    TRITON=yes
+    INDUCTOR_BENCHMARKS=yes
+    ;;
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks)
+    CUDA_VERSION=12.8.1
+    ANACONDA_PYTHON_VERSION=3.13
     GCC_VERSION=9
     VISION=yes
     KATEX=yes
@@ -161,8 +173,8 @@ case "$tag" in
     VISION=yes
     ONNX=yes
     ;;
-  pytorch-linux-jammy-py3.10-clang12)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-py3.9-clang12)
+    ANACONDA_PYTHON_VERSION=3.9
     CLANG_VERSION=12
     VISION=yes
     TRITON=yes
@@ -197,24 +209,24 @@ case "$tag" in
     UCC_COMMIT=${_UCC_COMMIT}
     PYTORCH_ROCM_ARCH="gfx90a;gfx942;gfx950"
     ;;
-  pytorch-linux-jammy-xpu-n-1-py3)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-xpu-2025.0-py3)
+    ANACONDA_PYTHON_VERSION=3.9
+    GCC_VERSION=11
+    VISION=yes
+    XPU_VERSION=2025.0
+    NINJA_VERSION=1.9.0
+    TRITON=yes
+    ;;
+  pytorch-linux-jammy-xpu-2025.1-py3)
+    ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
     XPU_VERSION=2025.1
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-xpu-n-py3)
-    ANACONDA_PYTHON_VERSION=3.10
-    GCC_VERSION=11
-    VISION=yes
-    XPU_VERSION=2025.2
-    NINJA_VERSION=1.9.0
-    TRITON=yes
-    ;;
-  pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks)
+    ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -222,8 +234,8 @@ case "$tag" in
     DOCS=yes
     INDUCTOR_BENCHMARKS=yes
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang12)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12)
+    ANACONDA_PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
     CLANG_VERSION=12
     VISION=yes
@@ -234,8 +246,8 @@ case "$tag" in
     CLANG_VERSION=18
     VISION=yes
     ;;
-  pytorch-linux-jammy-py3.10-gcc11)
-    ANACONDA_PYTHON_VERSION=3.10
+  pytorch-linux-jammy-py3.9-gcc11)
+    ANACONDA_PYTHON_VERSION=3.9
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -262,10 +274,13 @@ case "$tag" in
     TRITON_CPU=yes
     ;;
   pytorch-linux-jammy-linter)
-    PYTHON_VERSION=3.10
+    # TODO: Use 3.9 here because of this issue https://github.com/python/mypy/issues/13627.
+    # We will need to update mypy version eventually, but that's for another day. The task
+    # would be to upgrade mypy to 1.0.0 with Python 3.11
+    PYTHON_VERSION=3.9
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter)
-    PYTHON_VERSION=3.10
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter)
+    PYTHON_VERSION=3.9
     CUDA_VERSION=12.8.1
     ;;
   pytorch-linux-jammy-aarch64-py3.10-gcc11)

.ci/docker/centos-rocm/Dockerfile

@@ -56,13 +56,9 @@ ENV INSTALLED_VISION ${VISION}
 
 # Install rocm
 ARG ROCM_VERSION
-RUN mkdir ci_commit_pins
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh common_utils.sh
-RUN rm -r ci_commit_pins
+RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh

.ci/docker/ci_commit_pins/executorch.txt

@@ -1 +1 @@
-e0dda9059d082537cee36be6c5e4fe3b18c880c0
+56392aa978594cc155fa8af48cd949f5b5f1823a

.ci/docker/ci_commit_pins/huggingface-requirements.txt

@@ -1,2 +1,2 @@
-transformers==4.56.0
+transformers==4.54.0
 soxr==0.5.0

.ci/docker/ci_commit_pins/rocm-composable-kernel.txt

@@ -1 +0,0 @@
-7fe50dc3da2069d6645d9deb8c017a876472a977

.ci/docker/ci_commit_pins/torchbench.txt

@@ -1 +1 @@
-74a23feff57432129df84d8099e622773cf77925
+e03a63be43e33596f7f0a43b0f530353785e4a59

.ci/docker/ci_commit_pins/triton-xpu.txt

@@ -1 +1 @@
-1b0418a9a454b2b93ab8d71f40e59d2297157fae
+0958dc9b2bb815e428f721f9da599dab0dc1c5d7

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1 @@
-bbb06c0334a6772b92d24bde54956e675c8c6604
+f7888497a1eb9e98d4c07537f0d0bcfe180d1363

.ci/docker/common/install_cpython.sh

@@ -83,9 +83,9 @@ function build_cpython {
         py_suffix=${py_ver::-1}
         py_folder=$py_suffix
     fi
-    # Update to rc2 due to https://github.com/python/cpython/commit/c72699086fe4
+    # Only b3 is available now
     if [ "$py_suffix" == "3.14.0" ]; then
-        py_suffix="3.14.0rc2"
+        py_suffix="3.14.0b3"
     fi
     wget -q $PYTHON_DOWNLOAD_URL/$py_folder/Python-$py_suffix.tgz -O Python-$py_ver.tgz
     do_cpython_build $py_ver Python-$py_suffix

.ci/docker/common/install_cuda.sh

@@ -10,7 +10,7 @@ else
   arch_path='sbsa'
 fi
 
-NVSHMEM_VERSION=3.3.24
+NVSHMEM_VERSION=3.3.20
 
 function install_cuda {
   version=$1
@@ -65,7 +65,7 @@ function install_nvshmem {
   # This pattern is a lie as it is not consistent across versions, for 3.3.9 it was cuda_ver-arch-nvshhem-ver
   filename="libnvshmem-linux-${arch_path}-${nvshmem_version}_cuda${cuda_major_version}-archive"
   suffix=".tar.xz"
-  url="https://developer.download.nvidia.com/compute/nvshmem/redist/libnvshmem/linux-${arch_path}/${filename}${suffix}"
+  url="https://developer.download.nvidia.com/compute/redist/nvshmem/${nvshmem_version}/builds/cuda${cuda_major_version}/txz/agnostic/${dl_arch}/${filename}${suffix}"
 
   # download, unpack, install
   wget -q "${url}"
@@ -147,7 +147,8 @@ function install_128 {
 }
 
 function install_130 {
-  CUDNN_VERSION=9.13.0.50
+  CUDNN_VERSION=9.12.0.46
+  NVSHMEM_VERSION=3.3.20
   echo "Installing CUDA 13.0 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
   # install CUDA 13.0 in the same container
   install_cuda 13.0.0 cuda_13.0.0_580.65.06_linux

.ci/docker/common/install_executorch.sh

@@ -42,27 +42,22 @@ install_pip_dependencies() {
   # A workaround, ExecuTorch has moved to numpy 2.0 which is not compatible with the current
   # numba and scipy version used in PyTorch CI
   conda_run pip uninstall -y numba scipy
-  # Yaspin is needed for running CI test (get_benchmark_analysis_data.py)
-  pip_install yaspin==3.1.0
 
   popd
 }
 
 setup_executorch() {
+  pushd executorch
+
   export PYTHON_EXECUTABLE=python
-  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON -DEXECUTORCH_BUILD_TESTS=ON"
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
 
   as_jenkins .ci/scripts/setup-linux.sh --build-tool cmake || true
+  popd
 }
 
-if [ $# -eq 0 ]; then
-  clone_executorch
-  install_buck2
-  install_conda_dependencies
-  install_pip_dependencies
-  pushd executorch
-  setup_executorch
-  popd
-else
-  "$@"
-fi
+clone_executorch
+install_buck2
+install_conda_dependencies
+install_pip_dependencies
+setup_executorch

.ci/docker/common/install_rocm.sh

@@ -2,11 +2,6 @@
 
 set -ex
 
-# for pip_install function
-source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
-
-ROCM_COMPOSABLE_KERNEL_VERSION="$(cat $(dirname $0)/../ci_commit_pins/rocm-composable-kernel.txt)"
-
 ver() {
     printf "%3d%03d%03d%03d" $(echo "$1" | tr '.' ' ');
 }
@@ -118,8 +113,6 @@ EOF
         rm -rf HIP clr
     fi
 
-    pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
-
     # Cleanup
     apt-get autoclean && apt-get clean
     rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
@@ -183,8 +176,6 @@ install_centos() {
       sqlite3 $kdb "PRAGMA journal_mode=off; PRAGMA VACUUM;"
   done
 
-  pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
-
   # Cleanup
   yum clean all
   rm -rf /var/cache/yum

.ci/docker/common/install_triton.sh

@@ -57,7 +57,7 @@ if [ ! -f setup.py ]; then
   cd python
 fi
 
-pip_install pybind11==3.0.1
+pip_install pybind11==2.13.6
 
 # TODO: remove patch setup.py once we have a proper fix for https://github.com/triton-lang/triton/issues/4527
 as_jenkins sed -i -e 's/https:\/\/tritonlang.blob.core.windows.net\/llvm-builds/https:\/\/oaitriton.blob.core.windows.net\/public\/llvm-builds/g' setup.py

.ci/docker/common/install_ucc.sh

@@ -44,12 +44,8 @@ function install_ucc() {
 
   ./autogen.sh
 
-  if [[ -n "$CUDA_VERSION"  && $CUDA_VERSION == 13* ]]; then
-    NVCC_GENCODE="-gencode=arch=compute_86,code=compute_86"
-  else
-    # We only run distributed tests on Tesla M60 and A10G
-    NVCC_GENCODE="-gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_86,code=compute_86"
-  fi
+  # We only run distributed tests on Tesla M60 and A10G
+  NVCC_GENCODE="-gencode=arch=compute_52,code=sm_52 -gencode=arch=compute_86,code=compute_86"
 
   if [[ -n "$ROCM_VERSION" ]]; then
     if [[ -n "$PYTORCH_ROCM_ARCH" ]]; then

.ci/docker/common/install_xpu.sh

@@ -65,14 +65,10 @@ function install_ubuntu() {
 
 function install_rhel() {
     . /etc/os-release
-    if [[ "${ID}" == "rhel" ]]; then
-        if [[ ! " 8.8 8.9 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
-            echo "RHEL version ${VERSION_ID} not supported"
-            exit
-        fi
-    elif [[ "${ID}" == "almalinux" ]]; then
-        # Workaround for almalinux8 which used by quay.io/pypa/manylinux_2_28_x86_64
-        VERSION_ID="8.8"
+
+    if [[ ! " 8.8 8.10 9.0 9.2 9.3 " =~ " ${VERSION_ID} " ]]; then
+        echo "RHEL version ${VERSION_ID} not supported"
+        exit
     fi
 
     dnf install -y 'dnf-command(config-manager)'
@@ -150,11 +146,11 @@ if [[ "${XPU_DRIVER_TYPE,,}" == "lts" ]]; then
     XPU_DRIVER_VERSION="/lts/2350"
 fi
 
-# Default use Intel® oneAPI Deep Learning Essentials 2025.1
-if [[ "$XPU_VERSION" == "2025.2" ]]; then
-    XPU_PACKAGES="intel-deep-learning-essentials-2025.2"
-else
+# Default use Intel® oneAPI Deep Learning Essentials 2025.0
+if [[ "$XPU_VERSION" == "2025.1" ]]; then
     XPU_PACKAGES="intel-deep-learning-essentials-2025.1"
+else
+    XPU_PACKAGES="intel-deep-learning-essentials-2025.0"
 fi
 
 # The installation depends on the base OS

.ci/docker/libtorch/Dockerfile

@@ -74,14 +74,6 @@ RUN bash ./install_cuda.sh 13.0
 RUN bash ./install_magma.sh 13.0
 RUN ln -sf /usr/local/cuda-13.0 /usr/local/cuda
 
-# Install libibverbs for libtorch and copy to CUDA directory
-RUN apt-get update -y && \
-    apt-get install -y libibverbs-dev librdmacm-dev && \
-    cp /usr/lib/x86_64-linux-gnu/libmlx5.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/librdmacm.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/libibverbs.so* /usr/local/cuda/lib64/ && \
-    cp /usr/lib/x86_64-linux-gnu/libnl* /usr/local/cuda/lib64/
-
 FROM cpu as rocm
 ARG ROCM_VERSION
 ARG PYTORCH_ROCM_ARCH

.ci/docker/manywheel/Dockerfile_2_28

@@ -175,6 +175,6 @@ ENV XPU_DRIVER_TYPE ROLLING
 RUN python3 -m pip install --upgrade pip && \
     python3 -mpip install cmake==3.28.4
 ADD ./common/install_xpu.sh install_xpu.sh
-ENV XPU_VERSION 2025.2
+ENV XPU_VERSION 2025.1
 RUN bash ./install_xpu.sh && rm install_xpu.sh
 RUN pushd /opt/_internal && tar -xJf static-libs-for-embedding-only.tar.xz && popd

.ci/docker/requirements-ci.txt

@@ -93,9 +93,8 @@ librosa==0.10.2 ; python_version == "3.12" and platform_machine != "s390x"
 #Pinned versions:
 #test that import:
 
-mypy==1.16.0 ; platform_system != "Windows"
+mypy==1.16.0
 # Pin MyPy version because new errors are likely to appear with each release
-# Skip on Windows as lots of type annotations are POSIX specific
 #Description: linter
 #Pinned versions: 1.16.0
 #test that import: test_typing.py, test_type_hints.py
@@ -380,7 +379,7 @@ dataclasses_json==0.6.7
 cmake==4.0.0
 #Description: required for building
 
-tlparse==0.4.0
+tlparse==0.3.30
 #Description: required for log parsing
 
 cuda-bindings>=12.0,<13.0 ; platform_machine != "s390x"

.ci/docker/requirements-docs.txt

@@ -1,7 +1,7 @@
 sphinx==5.3.0
 #Description: This is used to generate PyTorch docs
 #Pinned versions: 5.3.0
--e git+https://github.com/pytorch/pytorch_sphinx_theme.git@d53b0ffb9b1cda68260693ea98f3483823c88d8e#egg=pytorch_sphinx_theme2
+-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@722b7e6f9ca512fcc526ad07d62b3d28c50bb6cd#egg=pytorch_sphinx_theme2
 
 # TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
 # but it doesn't seem to work and hangs around idly. The initial thought that it is probably

.ci/docker/triton_version.txt

@@ -1 +1 @@
-3.5.0
+3.4.0

.ci/docker/triton_xpu_version.txt

@@ -1 +1 @@
-3.5.0
+3.4.0

.ci/docker/ubuntu-rocm/Dockerfile

@@ -52,13 +52,9 @@ ENV INSTALLED_VISION ${VISION}
 
 # Install rocm
 ARG ROCM_VERSION
-RUN mkdir ci_commit_pins
-COPY ./common/common_utils.sh common_utils.sh
-COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
 COPY ./common/install_rocm.sh install_rocm.sh
 RUN bash ./install_rocm.sh
-RUN rm install_rocm.sh common_utils.sh
-RUN rm -r ci_commit_pins
+RUN rm install_rocm.sh
 COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
 RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
 RUN rm install_rocm_magma.sh

.ci/docker/ubuntu/Dockerfile

@@ -66,7 +66,6 @@ ENV NCCL_LIB_DIR="/usr/local/cuda/lib64/"
 # (optional) Install UCC
 ARG UCX_COMMIT
 ARG UCC_COMMIT
-ARG CUDA_VERSION
 ENV UCX_COMMIT $UCX_COMMIT
 ENV UCC_COMMIT $UCC_COMMIT
 ENV UCX_HOME /usr

.ci/libtorch/build.sh

@@ -7,4 +7,4 @@ set -ex
 
 SCRIPTPATH="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"
 
-USE_NVSHMEM=0 USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.10" ${SCRIPTPATH}/../manywheel/build.sh
+USE_NVSHMEM=0 USE_CUSPARSELT=0 BUILD_PYTHONLESS=1 DESIRED_PYTHON="3.9" ${SCRIPTPATH}/../manywheel/build.sh

.ci/lumen_cli/cli/lib/common/gh_summary.py

@@ -1,143 +0,0 @@
-from __future__ import annotations
-
-import logging
-import os
-import textwrap
-from pathlib import Path
-from typing import TYPE_CHECKING
-
-from cli.lib.common.utils import get_wheels
-from jinja2 import Template
-
-
-if TYPE_CHECKING:
-    from collections.abc import Iterable, Mapping
-
-
-logger = logging.getLogger(__name__)
-
-_TPL_CONTENT = Template(
-    textwrap.dedent("""\
-    ## {{ title }}
-
-    ```{{ lang }}
-    {{ content }}
-    ```
-""")
-)
-
-_TPL_LIST_ITEMS = Template(
-    textwrap.dedent("""\
-    ## {{ title }}
-    {% for it in items %}
-    - {{ it.pkg }}: {{ it.relpath }}
-    {% else %}
-    _(no item found)_
-    {% endfor %}
-    """)
-)
-
-_TPL_TABLE = Template(
-    textwrap.dedent("""\
-    {%- if rows %}
-    | {{ cols | join(' | ') }} |
-    |{%- for _ in cols %} --- |{%- endfor %}
-    {%- for r in rows %}
-    | {%- for c in cols %} {{ r.get(c, "") }} |{%- endfor %}
-    {%- endfor %}
-    {%- else %}
-    _(no data)_
-    {%- endif %}
-""")
-)
-
-
-def gh_summary_path() -> Path | None:
-    """Return the Path to the GitHub step summary file, or None if not set."""
-    p = os.environ.get("GITHUB_STEP_SUMMARY")
-    return Path(p) if p else None
-
-
-def write_gh_step_summary(md: str, *, append_content: bool = True) -> bool:
-    """
-    Write Markdown content to the GitHub Step Summary file if GITHUB_STEP_SUMMARY is set.
-    append_content: default true, if True, append to the end of the file, else overwrite the whole file
-
-    Returns:
-        True if written successfully (in GitHub Actions environment),
-        False if skipped (e.g., running locally where the variable is not set).
-    """
-    sp = gh_summary_path()
-    if not sp:
-        logger.info("[gh-summary] GITHUB_STEP_SUMMARY not set, skipping write.")
-        return False
-
-    md_clean = textwrap.dedent(md).strip() + "\n"
-
-    mode = "a" if append_content else "w"
-    with sp.open(mode, encoding="utf-8") as f:
-        f.write(md_clean)
-    return True
-
-
-def md_heading(text: str, level: int = 2) -> str:
-    """Generate a Markdown heading string with the given level (1-6)."""
-    return f"{'#' * max(1, min(level, 6))} {text}\n"
-
-
-def md_details(summary: str, content: str) -> str:
-    """Generate a collapsible <details> block with a summary and inner content."""
-    return f"<details>\n<summary>{summary}</summary>\n\n{content}\n\n</details>\n"
-
-
-def summarize_content_from_file(
-    output_dir: Path,
-    freeze_file: str,
-    title: str = "Content from file",
-    code_lang: str = "",  # e.g. "text" or "ini"
-) -> bool:
-    f = Path(output_dir) / freeze_file
-    if not f.exists():
-        return False
-    content = f.read_text(encoding="utf-8").strip()
-    md = render_content(content, title=title, lang=code_lang)
-    return write_gh_step_summary(md)
-
-
-def summarize_wheels(path: Path, title: str = "Wheels", max_depth: int = 3):
-    items = get_wheels(path, max_depth=max_depth)
-    if not items:
-        return False
-    md = render_list(items, title=title)
-    return write_gh_step_summary(md)
-
-
-def md_kv_table(rows: Iterable[Mapping[str, str | int | float]]) -> str:
-    """
-    Render a list of dicts as a Markdown table using Jinja template.
-    """
-    rows = list(rows)
-    cols = list({k for r in rows for k in r.keys()})
-    md = _TPL_TABLE.render(cols=cols, rows=rows).strip() + "\n"
-    return md
-
-
-def render_list(
-    items: Iterable[str],
-    *,
-    title: str = "List",
-) -> str:
-    tpl = _TPL_LIST_ITEMS
-    md = tpl.render(title=title, items=items)
-    return md
-
-
-def render_content(
-    content: str,
-    *,
-    title: str = "Content",
-    lang: str = "text",
-) -> str:
-    tpl = _TPL_CONTENT
-    md = tpl.render(title=title, content=content, lang=lang)
-    return md

.ci/lumen_cli/cli/lib/common/git_helper.py

@@ -45,7 +45,7 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
 
         # Checkout pinned commit
         commit = get_post_build_pinned_commit(target)
-        logger.info("Checking out pinned %s commit %s", target, commit)
+        logger.info("Checking out pinned commit %s", commit)
         r.git.checkout(commit)
 
         # Update submodules if requested
@@ -55,7 +55,7 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
                 sm.update(init=True, recursive=True, progress=PrintProgress())
 
         logger.info("Successfully cloned %s", target)
-        return r, commit
+        return r
 
     except GitCommandError as e:
         logger.error("Git operation failed: %s", e)

.ci/lumen_cli/cli/lib/common/pip_helper.py

@@ -4,7 +4,7 @@ import shlex
 import shutil
 import sys
 from collections.abc import Iterable
-from importlib.metadata import PackageNotFoundError, version  # noqa: UP035
+from importlib.metadata import PackageNotFoundError, version
 from typing import Optional, Union
 
 from cli.lib.common.utils import run_command

.ci/lumen_cli/cli/lib/common/utils.py

@@ -8,7 +8,6 @@ import shlex
 import subprocess
 import sys
 from contextlib import contextmanager
-from pathlib import Path
 from typing import Optional
 
 
@@ -116,24 +115,3 @@ def working_directory(path: str):
         yield
     finally:
         os.chdir(prev_cwd)
-
-
-def get_wheels(
-    output_dir: Path,
-    max_depth: Optional[int] = None,
-) -> list[str]:
-    """Return a list of wheels found in the given output directory."""
-    root = Path(output_dir)
-    if not root.exists():
-        return []
-    items = []
-    for dirpath, _, filenames in os.walk(root):
-        depth = Path(dirpath).relative_to(root).parts
-        if max_depth is not None and len(depth) > max_depth:
-            continue
-        for fname in sorted(filenames):
-            if fname.endswith(".whl"):
-                pkg = fname.split("-")[0]
-                relpath = str((Path(dirpath) / fname).relative_to(root))
-                items.append({"pkg": pkg, "relpath": relpath})
-    return items

.ci/lumen_cli/cli/lib/core/vllm/lib.py

@@ -1,27 +1,13 @@
 import logging
-import os
-import textwrap
 from typing import Any
 
-from cli.lib.common.gh_summary import write_gh_step_summary
 from cli.lib.common.git_helper import clone_external_repo
 from cli.lib.common.pip_helper import pip_install_packages
 from cli.lib.common.utils import run_command, temp_environ, working_directory
-from jinja2 import Template
 
 
 logger = logging.getLogger(__name__)
 
-_TPL_VLLM_INFO = Template(
-    textwrap.dedent("""\
-    ##  Vllm against Pytorch CI Test Summary
-    **Vllm Commit**: [{{ vllm_commit }}](https://github.com/vllm-project/vllm/commit/{{ vllm_commit }})
-    {%- if torch_sha %}
-    **Pytorch Commit**: [{{ torch_sha }}](https://github.com/pytorch/pytorch/commit/{{ torch_sha }})
-    {%- endif %}
-""")
-)
-
 
 def sample_vllm_test_library():
     """
@@ -41,6 +27,7 @@ def sample_vllm_test_library():
                 "pytest -v -s basic_correctness/test_cumem.py",
                 "pytest -v -s basic_correctness/test_basic_correctness.py",
                 "pytest -v -s basic_correctness/test_cpu_offload.py",
+                "VLLM_TEST_ENABLE_ARTIFICIAL_PREEMPT=1 pytest -v -s basic_correctness/test_preemption.py",
             ],
         },
         "vllm_basic_models_test": {
@@ -67,12 +54,16 @@ def sample_vllm_test_library():
                         "-v",
                         "-s",
                         "entrypoints/llm",
+                        "--ignore=entrypoints/llm/test_lazy_outlines.py",
                         "--ignore=entrypoints/llm/test_generate.py",
+                        "--ignore=entrypoints/llm/test_generate_multiple_loras.py",
                         "--ignore=entrypoints/llm/test_collective_rpc.py",
                     ]
                 ),
-                "pytest -v -s entrypoints/llm/test_generate.py",
-                "pytest -v -s entrypoints/offline_mode",
+                "pytest -v -s entrypoints/llm/test_lazy_outlines.py",
+                "pytest -v -s entrypoints/llm/test_generate.py ",
+                "pytest -v -s entrypoints/llm/test_generate_multiple_loras.py",
+                "VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode",
             ],
         },
         "vllm_regression_test": {
@@ -92,24 +83,14 @@ def sample_vllm_test_library():
             "num_gpus": 4,
             "steps": [
                 "pytest -v -s -x lora/test_chatglm3_tp.py",
+                "echo $VLLM_WORKER_MULTIPROC_METHOD",
                 "pytest -v -s -x lora/test_llama_tp.py",
-                "pytest -v -s -x lora/test_llm_with_multi_loras.py",
+                "pytest -v -s -x lora/test_multi_loras_with_tp.py",
             ],
         },
-        "vllm_distributed_test_28_failure_test": {
-            "title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
-            "id": "vllm_distributed_test_28_failure_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "num_gpus": 4,
-            "steps": [
-                "pytest -v -s distributed/test_sequence_parallel.py",
-            ],
-        },
-        "vllm_lora_28_failure_test": {
-            "title": "LoRA pytorch 2.8 failure test",
-            "id": "vllm_lora_28_failure_test",
+        "vllm_lora_280_failure_test": {
+            "title": "LoRA 280 failure test",
+            "id": "vllm_lora_280_failure_test",
             "steps": ["pytest -v lora/test_quant_model.py"],
         },
         "vllm_multi_model_processor_test": {
@@ -120,15 +101,6 @@ def sample_vllm_test_library():
                 "pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py",
             ],
         },
-        "vllm_multi_model_test_28_failure_test": {
-            "title": "Multi-Model Test (Failed 2.8 release)",
-            "id": "vllm_multi_model_test_28_failure_test",
-            "package_install": ["git+https://github.com/TIGER-AI-Lab/Mantis.git"],
-            "steps": [
-                "pytest -v -s models/multimodal/generation/test_voxtral.py",
-                "pytest -v -s models/multimodal/pooling",
-            ],
-        },
         "vllm_pytorch_compilation_unit_tests": {
             "title": "PyTorch Compilation Unit Tests",
             "id": "vllm_pytorch_compilation_unit_tests",
@@ -143,28 +115,6 @@ def sample_vllm_test_library():
                 "pytest -v -s compile/test_decorator.py",
             ],
         },
-        "vllm_languagde_model_test_extended_generation_28_failure_test": {
-            "title": "Language Models Test (Extended Generation) 2.8 release failure",
-            "id": "vllm_languagde_model_test_extended_generation_28_failure_test",
-            "package_install": [
-                "--no-build-isolation",
-                "git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8",
-            ],
-            "steps": [
-                "pytest -v -s models/language/generation/test_mistral.py",
-            ],
-        },
-        "vllm_distributed_test_2_gpu_28_failure_test": {
-            "title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
-            "id": "vllm_distributed_test_2_gpu_28_failure_test",
-            "env_vars": {
-                "VLLM_WORKER_MULTIPROC_METHOD": "spawn",
-            },
-            "num_gpus": 4,
-            "steps": [
-                "pytest -v -s distributed/test_sequence_parallel.py",
-            ],
-        },
         # TODO(elainewy):need to add g6 with 4 gpus to run this test
         "vllm_lora_test": {
             "title": "LoRA Test %N",
@@ -264,13 +214,12 @@ def run_test_plan(
 
 
 def clone_vllm(dst: str = "vllm"):
-    _, commit = clone_external_repo(
+    clone_external_repo(
         target="vllm",
         repo="https://github.com/vllm-project/vllm.git",
         dst=dst,
         update_submodules=True,
     )
-    return commit
 
 
 def replace_buildkite_placeholders(step: str, shard_id: int, num_shards: int) -> str:
@@ -281,12 +230,3 @@ def replace_buildkite_placeholders(step: str, shard_id: int, num_shards: int) ->
     for k in sorted(mapping, key=len, reverse=True):
         step = step.replace(k, mapping[k])
     return step
-
-
-def summarize_build_info(vllm_commit: str) -> bool:
-    torch_sha = os.getenv("GITHUB_SHA")
-    md = (
-        _TPL_VLLM_INFO.render(vllm_commit=vllm_commit, torch_sha=torch_sha).strip()
-        + "\n"
-    )
-    return write_gh_step_summary(md)

.ci/lumen_cli/cli/lib/core/vllm/vllm_build.py

@@ -13,11 +13,6 @@ from cli.lib.common.envs_helper import (
     env_str_field,
     with_params_help,
 )
-from cli.lib.common.gh_summary import (
-    gh_summary_path,
-    summarize_content_from_file,
-    summarize_wheels,
-)
 from cli.lib.common.path_helper import (
     copy,
     ensure_dir_exists,
@@ -26,7 +21,7 @@ from cli.lib.common.path_helper import (
     is_path_exist,
 )
 from cli.lib.common.utils import run_command
-from cli.lib.core.vllm.lib import clone_vllm, summarize_build_info
+from cli.lib.core.vllm.lib import clone_vllm
 
 
 logger = logging.getLogger(__name__)
@@ -66,11 +61,6 @@ class VllmBuildParameters:
         "DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
     )
 
-    # the cleaning script to remove torch dependencies from pip
-    cleaning_script: Path = env_path_field(
-        "cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
-    )
-
     # OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
     output_dir: Path = env_path_field("OUTPUT_DIR", "external/vllm")
 
@@ -163,44 +153,18 @@ class VllmBuildRunner(BaseRunner):
         """
         inputs = VllmBuildParameters()
         logger.info("Running vllm build with inputs: %s", inputs)
-        vllm_commit = clone_vllm()
+        clone_vllm()
 
-        self.cp_torch_cleaning_script(inputs)
         self.cp_dockerfile_if_exist(inputs)
+
         # cp torch wheels from root direct to vllm workspace if exist
         self.cp_torch_whls_if_exist(inputs)
 
-        # make sure the output dir to store the build artifacts exist
-        ensure_dir_exists(Path(inputs.output_dir))
+        ensure_dir_exists(inputs.output_dir)
 
         cmd = self._generate_docker_build_cmd(inputs)
         logger.info("Running docker build: \n %s", cmd)
-
-        try:
-            run_command(cmd, cwd="vllm", env=os.environ.copy())
-        finally:
-            self.genearte_vllm_build_summary(vllm_commit, inputs)
-
-    def genearte_vllm_build_summary(
-        self, vllm_commit: str, inputs: VllmBuildParameters
-    ):
-        if not gh_summary_path():
-            return logger.info("Skipping, not detect GH Summary env var....")
-        logger.info("Generate GH Summary ...")
-        # summarize vllm build info
-        summarize_build_info(vllm_commit)
-
-        # summarize vllm build artifacts
-        vllm_artifact_dir = inputs.output_dir / "wheels"
-        summarize_content_from_file(
-            vllm_artifact_dir,
-            "build_summary.txt",
-            title="Vllm build env pip package summary",
-        )
-        summarize_wheels(
-            inputs.torch_whls_path, max_depth=3, title="Torch Wheels Artifacts"
-        )
-        summarize_wheels(vllm_artifact_dir, max_depth=3, title="Vllm Wheels Artifacts")
+        run_command(cmd, cwd="vllm", env=os.environ.copy())
 
     def cp_torch_whls_if_exist(self, inputs: VllmBuildParameters) -> str:
         if not inputs.use_torch_whl:
@@ -211,11 +175,6 @@ class VllmBuildRunner(BaseRunner):
         copy(inputs.torch_whls_path, tmp_dir)
         return tmp_dir
 
-    def cp_torch_cleaning_script(self, inputs: VllmBuildParameters):
-        script = get_path(inputs.cleaning_script, resolve=True)
-        vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
-        copy(script, vllm_script)
-
     def cp_dockerfile_if_exist(self, inputs: VllmBuildParameters):
         if not inputs.use_local_dockerfile:
             logger.info("using vllm default dockerfile.torch_nightly for build")

.ci/lumen_cli/cli/lib/core/vllm/vllm_test.py

@@ -11,7 +11,7 @@ from typing import Any
 
 from cli.lib.common.cli_helper import BaseRunner
 from cli.lib.common.envs_helper import env_path_field, env_str_field, get_env
-from cli.lib.common.path_helper import copy, get_path, remove_dir
+from cli.lib.common.path_helper import copy, remove_dir
 from cli.lib.common.pip_helper import (
     pip_install_first_match,
     pip_install_packages,
@@ -43,10 +43,6 @@ class VllmTestParameters:
 
     torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
 
-    cleaning_script: Path = env_path_field(
-        "cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
-    )
-
     def __post_init__(self):
         if not self.torch_whls_path.exists():
             raise ValueError("missing torch_whls_path")
@@ -96,13 +92,11 @@ class VllmTestRunner(BaseRunner):
         self._set_envs(params)
 
         clone_vllm(dst=self.work_directory)
-        self.cp_torch_cleaning_script(params)
         with working_directory(self.work_directory):
             remove_dir(Path("vllm"))
             self._install_wheels(params)
             self._install_dependencies()
         # verify the torches are not overridden by test dependencies
-
         check_versions()
 
     def run(self):
@@ -110,31 +104,20 @@ class VllmTestRunner(BaseRunner):
         main function to run vllm test
         """
         self.prepare()
-        try:
-            with working_directory(self.work_directory):
-                if self.test_type == TestInpuType.TEST_PLAN:
-                    if self.num_shards > 1:
-                        run_test_plan(
-                            self.test_plan,
-                            "vllm",
-                            sample_vllm_test_library(),
-                            self.shard_id,
-                            self.num_shards,
-                        )
-                    else:
-                        run_test_plan(
-                            self.test_plan, "vllm", sample_vllm_test_library()
-                        )
+        with working_directory(self.work_directory):
+            if self.test_type == TestInpuType.TEST_PLAN:
+                if self.num_shards > 1:
+                    run_test_plan(
+                        self.test_plan,
+                        "vllm",
+                        sample_vllm_test_library(),
+                        self.shard_id,
+                        self.num_shards,
+                    )
                 else:
-                    raise ValueError(f"Unknown test type {self.test_type}")
-        finally:
-            # double check the torches are not overridden by other packages
-            check_versions()
-
-    def cp_torch_cleaning_script(self, params: VllmTestParameters):
-        script = get_path(params.cleaning_script, resolve=True)
-        vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
-        copy(script, vllm_script)
+                    run_test_plan(self.test_plan, "vllm", sample_vllm_test_library())
+            else:
+                raise ValueError(f"Unknown test type {self.test_type}")
 
     def _install_wheels(self, params: VllmTestParameters):
         logger.info("Running vllm test with inputs: %s", params)
@@ -237,8 +220,6 @@ def preprocess_test_in(
     target_path = Path(target_file)
     lines = target_path.read_text().splitlines()
 
-    pkgs_to_add = []
-
     # Remove lines starting with the package names (==, @, >=) — case-insensitive
     pattern = re.compile(rf"^({'|'.join(pkgs_to_remove)})\s*(==|@|>=)", re.IGNORECASE)
     kept_lines = [line for line in lines if not pattern.match(line)]
@@ -255,11 +236,7 @@ def preprocess_test_in(
     ]
 
     # Write back: header_lines + blank + kept_lines
-    out_lines = header_lines + [""] + kept_lines
-    if pkgs_to_add:
-        out_lines += [""] + pkgs_to_add
-
-    out = "\n".join(out_lines) + "\n"
+    out = "\n".join(header_lines + [""] + kept_lines) + "\n"
     target_path.write_text(out)
     logger.info("[INFO] Updated %s", target_file)
 

.ci/manywheel/build_cuda.sh

@@ -124,7 +124,6 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
     fi
     if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
         echo "Bundling with cudnn and cublas."
-
         DEPS_LIST+=(
             "/usr/local/cuda/lib64/libcudnn_adv.so.9"
             "/usr/local/cuda/lib64/libcudnn_cnn.so.9"
@@ -134,11 +133,16 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             "/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9"
             "/usr/local/cuda/lib64/libcudnn_heuristic.so.9"
             "/usr/local/cuda/lib64/libcudnn.so.9"
+            "/usr/local/cuda/lib64/libcublas.so.12"
+            "/usr/local/cuda/lib64/libcublasLt.so.12"
             "/usr/local/cuda/lib64/libcusparseLt.so.0"
+            "/usr/local/cuda/lib64/libcudart.so.12"
+            "/usr/local/cuda/lib64/libnvrtc.so.12"
             "/usr/local/cuda/lib64/libnvrtc-builtins.so"
             "/usr/local/cuda/lib64/libcufile.so.0"
             "/usr/local/cuda/lib64/libcufile_rdma.so.1"
             "/usr/local/cuda/lib64/libnvshmem_host.so.3"
+            "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12"
             "/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so"
         )
         DEPS_SONAME+=(
@@ -150,56 +154,22 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             "libcudnn_engines_precompiled.so.9"
             "libcudnn_heuristic.so.9"
             "libcudnn.so.9"
+            "libcublas.so.12"
+            "libcublasLt.so.12"
             "libcusparseLt.so.0"
+            "libcudart.so.12"
+            "libnvrtc.so.12"
             "libnvrtc-builtins.so"
             "libnvshmem_host.so.3"
             "libcufile.so.0"
             "libcufile_rdma.so.1"
+            "libcupti.so.12"
             "libnvperf_host.so"
         )
         # Add libnvToolsExt only if CUDA version is not 12.9
-        if [[ $CUDA_VERSION == 13* ]]; then
-            DEPS_LIST+=(
-                "/usr/local/cuda/lib64/libcublas.so.13"
-                "/usr/local/cuda/lib64/libcublasLt.so.13"
-                "/usr/local/cuda/lib64/libcudart.so.13"
-                "/usr/local/cuda/lib64/libnvrtc.so.13"
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13"
-                "/usr/local/cuda/lib64/libibverbs.so.1"
-                "/usr/local/cuda/lib64/librdmacm.so.1"
-                "/usr/local/cuda/lib64/libmlx5.so.1"
-                "/usr/local/cuda/lib64/libnl-3.so.200"
-                "/usr/local/cuda/lib64/libnl-route-3.so.200")
-            DEPS_SONAME+=(
-                "libcublas.so.13"
-                "libcublasLt.so.13"
-                "libcudart.so.13"
-                "libnvrtc.so.13"
-                "libcupti.so.13"
-                "libibverbs.so.1"
-                "librdmacm.so.1"
-                "libmlx5.so.1"
-                "libnl-3.so.200"
-                "libnl-route-3.so.200")
-            export USE_CUPTI_SO=1
-            export ATEN_STATIC_CUDA=0
-            export USE_CUDA_STATIC_LINK=0
-            export USE_CUFILE=0
-        else
-            DEPS_LIST+=(
-                "/usr/local/cuda/lib64/libnvToolsExt.so.1"
-                "/usr/local/cuda/lib64/libcublas.so.12"
-                "/usr/local/cuda/lib64/libcublasLt.so.12"
-                "/usr/local/cuda/lib64/libcudart.so.12"
-                "/usr/local/cuda/lib64/libnvrtc.so.12"
-                "/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12")
-            DEPS_SONAME+=(
-                "libnvToolsExt.so.1"
-                "libcublas.so.12"
-                "libcublasLt.so.12"
-                "libcudart.so.12"
-                "libnvrtc.so.12"
-                "libcupti.so.12")
+        if [[ $CUDA_VERSION != 12.9* ]]; then
+            DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
+            DEPS_SONAME+=("libnvToolsExt.so.1")
         fi
     else
         echo "Using nvidia libs from pypi."

.ci/pytorch/check_binary.sh

@@ -300,3 +300,24 @@ except RuntimeError as e:
     exit 1
   fi
 fi
+
+###############################################################################
+# Check for C++ ABI compatibility to GCC-11 - GCC 13
+###############################################################################
+if [[ "$(uname)" == 'Linux' &&  "$PACKAGE_TYPE" == 'manywheel' ]]; then
+  pushd /tmp
+  # Per https://gcc.gnu.org/onlinedocs/gcc/C_002b_002b-Dialect-Options.html
+  # gcc-11 is ABI16, gcc-13 is ABI18, gcc-14 is ABI19
+  # gcc 11 - CUDA 11.8, xpu, rocm
+  # gcc 13 - CUDA 12.6, 12.8 and cpu
+  # Please see issue for reference: https://github.com/pytorch/pytorch/issues/152426
+  if [[ "$(uname -m)" == "s390x" ]]; then
+    cxx_abi="19"
+  elif [[ "$DESIRED_CUDA" != 'xpu' && "$DESIRED_CUDA" != 'rocm'* ]]; then
+    cxx_abi="18"
+  else
+    cxx_abi="16"
+  fi
+  python -c "import torch; exit(0 if torch._C._PYBIND11_BUILD_ABI == '_cxxabi10${cxx_abi}' else 1)"
+  popd
+fi

.ci/pytorch/common_utils.sh

@@ -149,22 +149,13 @@ function get_pinned_commit() {
   cat .github/ci_commit_pins/"${1}".txt
 }
 
-function detect_cuda_arch() {
-  if [[ "${BUILD_ENVIRONMENT}" == *cuda* ]]; then
-    if command -v nvidia-smi; then
-      TORCH_CUDA_ARCH_LIST=$(nvidia-smi --query-gpu=compute_cap --format=csv | tail -n 1)
-    elif [[ "${TEST_CONFIG}" == *nogpu* ]]; then
-      # There won't be nvidia-smi in nogpu tests, so just set TORCH_CUDA_ARCH_LIST to the default
-      # minimum supported value here
-      TORCH_CUDA_ARCH_LIST=8.0
-    fi
-    export TORCH_CUDA_ARCH_LIST
-  fi
-}
-
 function install_torchaudio() {
   local commit
   commit=$(get_pinned_commit audio)
+  if [[ "${BUILD_ENVIRONMENT}" == *cuda* ]] && command -v nvidia-smi; then
+    TORCH_CUDA_ARCH_LIST=$(nvidia-smi --query-gpu=compute_cap --format=csv | tail -n 1)
+    export TORCH_CUDA_ARCH_LIST
+  fi
   pip_build_and_install "git+https://github.com/pytorch/audio.git@${commit}" dist/audio
 }
 
@@ -258,19 +249,11 @@ function install_torchrec_and_fbgemm() {
       git clone --recursive https://github.com/pytorch/fbgemm
       pushd fbgemm/fbgemm_gpu
       git checkout "${fbgemm_commit}" --recurse-submodules
-      # until the fbgemm_commit includes the tbb patch
-      patch <<'EOF'
---- a/FbgemmGpu.cmake
-+++ b/FbgemmGpu.cmake
-@@ -184,5 +184,6 @@ gpu_cpp_library(
-     fbgemm_gpu_tbe_cache
-     fbgemm_gpu_tbe_optimizers
-     fbgemm_gpu_tbe_utils
-+    tbb
-   DESTINATION
-     fbgemm_gpu)
-EOF
-      python setup.py bdist_wheel --build-variant=rocm
+      python setup.py bdist_wheel \
+        --build-variant=rocm \
+        -DHIP_ROOT_DIR="${ROCM_PATH}" \
+        -DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
+        -DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
       popd
 
       # Save the wheel before cleaning up

.ci/pytorch/functorch_doc_push_script.sh

@@ -0,0 +1,40 @@
+#!/bin/bash
+
+# This is where the local pytorch install in the docker image is located
+pt_checkout="/var/lib/jenkins/workspace"
+source "$pt_checkout/.ci/pytorch/common_utils.sh"
+echo "functorch_doc_push_script.sh: Invoked with $*"
+
+set -ex -o pipefail
+
+version=${DOCS_VERSION:-nightly}
+echo "version: $version"
+
+# Build functorch docs
+pushd $pt_checkout/functorch/docs
+make html
+popd
+
+git clone https://github.com/pytorch/functorch -b gh-pages --depth 1 functorch_ghpages
+pushd functorch_ghpages
+
+if [ "$version" == "main" ]; then
+  version=nightly
+fi
+
+git rm -rf "$version" || true
+mv "$pt_checkout/functorch/docs/build/html" "$version"
+
+git add "$version" || true
+git status
+git config user.email "soumith+bot@pytorch.org"
+git config user.name "pytorchbot"
+# If there aren't changes, don't make a commit; push is no-op
+git commit -m "Generate Python docs from pytorch/pytorch@${GITHUB_SHA}" || true
+git status
+
+if [[ "${WITH_PUSH:-}" == true ]]; then
+  git push -u origin gh-pages
+fi
+
+popd

.ci/pytorch/macos-build.sh

@@ -35,10 +35,11 @@ fi
 
 print_cmake_info
 if [[ ${BUILD_ENVIRONMENT} == *"distributed"* ]]; then
-  USE_OPENMP=1 WERROR=1 python setup.py bdist_wheel
+  # Needed for inductor benchmarks, as lots of HF networks make `torch.distribtued` calls
+  USE_DISTRIBUTED=1 USE_OPENMP=1 WERROR=1 python setup.py bdist_wheel
 else
-  # NB: we always build with distributed; USE_DISTRIBUTED turns off all
-  # backends (specifically the gloo backend), so test that this case works too
+  # Explicitly set USE_DISTRIBUTED=0 to align with the default build config on mac. This also serves as the sole CI config that tests
+  # that building with USE_DISTRIBUTED=0 works at all. See https://github.com/pytorch/pytorch/issues/86448
   USE_DISTRIBUTED=0 USE_OPENMP=1 MACOSX_DEPLOYMENT_TARGET=11.0 WERROR=1 BUILD_TEST=OFF USE_PYTORCH_METAL=1 python setup.py bdist_wheel --plat-name macosx_11_0_arm64
 fi
 if which sccache > /dev/null; then

.ci/pytorch/macos-test.sh

@@ -13,13 +13,9 @@ if [[ ! $(python -c "import torch; print(int(torch.backends.openmp.is_available(
 fi
 popd
 
-python -mpip install -r requirements.txt
-
 # enable debug asserts in serialization
 export TORCH_SERIALIZATION_DEBUG=1
 
-python -mpip install --no-input -r requirements.txt
-
 setup_test_python() {
   # The CircleCI worker hostname doesn't resolve to an address.
   # This environment variable makes ProcessGroupGloo default to
@@ -199,7 +195,7 @@ torchbench_setup_macos() {
   git checkout "$(cat ../.github/ci_commit_pins/vision.txt)"
   git submodule update --init --recursive
   python setup.py clean
-  python -m pip install -e . -v --no-build-isolation
+  python setup.py develop
   popd
 
   pushd torchaudio
@@ -208,7 +204,7 @@ torchbench_setup_macos() {
   git submodule update --init --recursive
   python setup.py clean
   #TODO: Remove me, when figure out how to make TorchAudio find brew installed openmp
-  USE_OPENMP=0 python -m pip install -e . -v --no-build-isolation
+  USE_OPENMP=0 python setup.py develop
   popd
 
   checkout_install_torchbench
@@ -306,47 +302,6 @@ test_torchbench_smoketest() {
     fi
 
   done
-  echo "Pytorch benchmark on mps device completed"
-}
-
-test_aoti_torchbench_smoketest() {
-  print_cmake_info
-
-  echo "Launching AOTInductor torchbench setup"
-  pip_benchmark_deps
-  # shellcheck disable=SC2119,SC2120
-  torchbench_setup_macos
-
-  TEST_REPORTS_DIR=$(pwd)/test/test-reports
-  mkdir -p "$TEST_REPORTS_DIR"
-
-  local device=mps
-  local dtypes=(undefined float16 bfloat16 notset)
-  local dtype=${dtypes[$1]}
-  local models=(hf_T5 llama BERT_pytorch dcgan hf_GPT2 yolov3 resnet152 sam sam_fast pytorch_unet stable_diffusion_text_encoder speech_transformer Super_SloMo doctr_det_predictor doctr_reco_predictor timm_resnet timm_vovnet vgg16)
-
-  echo "Launching torchbench inference performance run for AOT Inductor and dtype ${dtype}"
-  local dtype_arg="--${dtype}"
-  if [ "$dtype" == notset ]; then
-      dtype_arg="--float32"
-  fi
-  touch "$TEST_REPORTS_DIR/aot_inductor_torchbench_${dtype}_inference_${device}_performance.csv"
-  for model in "${models[@]}"; do
-    PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-      --performance --only "$model" --export-aot-inductor --inference --devices "$device" "$dtype_arg" \
-      --output "$TEST_REPORTS_DIR/aot_inductor_torchbench_${dtype}_inference_${device}_performance.csv" || true
-    PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/torchbench.py \
-      --accuracy --only "$model" --export-aot-inductor --inference --devices "$device" "$dtype_arg" \
-      --output "$TEST_REPORTS_DIR/aot_inductor_torchbench_${dtype}_inference_${device}_accuracy.csv" || true
-  done
-
-  echo "Launching HuggingFace inference performance run for AOT Inductor and dtype ${dtype}"
-  PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
-    --performance --export-aot-inductor --inference --devices "$device" "$dtype_arg" \
-    --output "$TEST_REPORTS_DIR/aot_inductor_huggingface_${dtype}_inference_${device}_performance.csv" || true
-  PYTHONPATH="$(pwd)"/torchbench python benchmarks/dynamo/huggingface.py \
-    --accuracy --export-aot-inductor --inference --devices "$device" "$dtype_arg" \
-    --output "$TEST_REPORTS_DIR/aot_inductor_huggingface_${dtype}_inference_${device}_accuracy.csv" || true
 
   echo "Pytorch benchmark on mps device completed"
 }
@@ -395,8 +350,6 @@ elif [[ $TEST_CONFIG == *"perf_timm"* ]]; then
   test_timm_perf
 elif [[ $TEST_CONFIG == *"perf_smoketest"* ]]; then
   test_torchbench_smoketest "${SHARD_NUMBER}"
-elif [[ $TEST_CONFIG == *"aot_inductor_perf_smoketest"* ]]; then
-  test_aoti_torchbench_smoketest "${SHARD_NUMBER}"
 elif [[ $TEST_CONFIG == *"mps"* ]]; then
   test_python_mps
 elif [[ $NUM_TEST_SHARDS -gt 1 ]]; then

.ci/pytorch/multigpu-test.sh

@@ -45,7 +45,6 @@ if [[ "${SHARD_NUMBER:-2}" == "2" ]]; then
     # DTensor tests
     time python test/run_test.py --verbose -i distributed/tensor/test_random_ops
     time python test/run_test.py --verbose -i distributed/tensor/test_dtensor_compile
-    time python test/run_test.py --verbose -i distributed/tensor/test_utils.py
 
     # DeviceMesh test
     time python test/run_test.py --verbose -i distributed/test_device_mesh

.ci/pytorch/numba-cuda-13.patch

@@ -1,25 +0,0 @@
-From 6e08c9d08e9de59c7af28b720289debbbd384764 Mon Sep 17 00:00:00 2001
-From: Michael Wang <13521008+isVoid@users.noreply.github.com>
-Date: Tue, 1 Apr 2025 17:28:05 -0700
-Subject: [PATCH] Avoid bumping certain driver API to avoid future breakage
- (#185)
-
-Co-authored-by: isVoid <isVoid@users.noreply.github.com>
----
- numba_cuda/numba/cuda/cudadrv/driver.py | 3 +++
- 1 file changed, 3 insertions(+)
-
-diff --git a/numba_cuda/numba/cuda/cudadrv/driver.py b/numba_cuda/numba/cuda/cudadrv/driver.py
-index 1641bf77..233e9ed7 100644
---- a/numba_cuda/numba/cuda/cudadrv/driver.py
-+++ b/numba_cuda/numba/cuda/cudadrv/driver.py
-@@ -365,6 +365,9 @@ def _find_api(self, fname):
-         else:
-             variants = ('_v2', '')
- 
-+        if fname in ("cuCtxGetDevice", "cuCtxSynchronize"):
-+            return getattr(self.lib, fname)
-+
-         for variant in variants:
-             try:
-                 return getattr(self.lib, f'{fname}{variant}')

.ci/pytorch/smoke_test/smoke_test.py

@@ -386,8 +386,8 @@ def smoke_test_compile(device: str = "cpu") -> None:
 
 
 def smoke_test_nvshmem() -> None:
-    if not torch.cuda.is_available() or target_os == "windows":
-        print("Windows platform or CUDA is not available, skipping NVSHMEM test")
+    if not torch.cuda.is_available():
+        print("CUDA is not available, skipping NVSHMEM test")
         return
 
     # Check if NVSHMEM is compiled in current build
@@ -396,9 +396,7 @@ def smoke_test_nvshmem() -> None:
     except ImportError:
         # Not built with NVSHMEM support.
         # torch is not compiled with NVSHMEM prior to 2.9
-        from torch.torch_version import TorchVersion
-
-        if TorchVersion(torch.__version__) < (2, 9):
+        if torch.__version__ < "2.9":
             return
         else:
             # After 2.9: NVSHMEM is expected to be compiled in current build

.ci/pytorch/test.sh

@@ -32,16 +32,6 @@ if [[ "$BUILD_ENVIRONMENT" != *rocm* && "$BUILD_ENVIRONMENT" != *s390x* && -d /v
   git config --global --add safe.directory /var/lib/jenkins/workspace
 fi
 
-
-# Patch numba to avoid CUDA-13 crash, see https://github.com/pytorch/pytorch/issues/162878
-NUMBA_CUDA_DIR=$(python -c "import os;import numba.cuda; print(os.path.dirname(numba.cuda.__file__))" 2>/dev/null || true)
-if [ -n "$NUMBA_CUDA_DIR" ]; then
-  NUMBA_PATCH="$(dirname "$(realpath "${BASH_SOURCE[0]}")")/numba-cuda-13.patch"
-  pushd "$NUMBA_CUDA_DIR"
-  patch -p4 <"$NUMBA_PATCH"
-  popd
-fi
-
 echo "Environment variables:"
 env
 
@@ -101,7 +91,6 @@ if [[ "$BUILD_ENVIRONMENT" == *clang9* || "$BUILD_ENVIRONMENT" == *xpu* ]]; then
   export VALGRIND=OFF
 fi
 
-detect_cuda_arch
 
 if [[ "$BUILD_ENVIRONMENT" == *s390x* ]]; then
   # There are additional warnings on s390x, maybe due to newer gcc.
@@ -334,17 +323,11 @@ test_python() {
 }
 
 test_python_smoke() {
-  # Smoke tests for H100/B200
+  # Smoke tests for H100
   time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
   assert_git_not_dirty
 }
 
-test_python_smoke_b200() {
-  # Targeted smoke tests for B200 - staged approach to avoid too many failures
-  time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
-  assert_git_not_dirty
-}
-
 test_h100_distributed() {
   # Distributed tests at H100
   time python test/run_test.py --include distributed/_composable/test_composability/test_pp_composability.py  $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
@@ -512,14 +495,6 @@ test_inductor_cpp_wrapper_shard() {
     -k 'take' \
     --shard "$1" "$NUM_TEST_SHARDS" \
     --verbose
-
-  if [[ "${BUILD_ENVIRONMENT}" == *xpu* ]]; then
-    python test/run_test.py \
-      --include inductor/test_mkldnn_pattern_matcher \
-      -k 'xpu' \
-      --shard "$1" "$NUM_TEST_SHARDS" \
-      --verbose
-  fi
 }
 
 # "Global" flags for inductor benchmarking controlled by TEST_CONFIG
@@ -1556,10 +1531,14 @@ test_executorch() {
   install_torchvision
   install_torchaudio
 
-  INSTALL_SCRIPT="$(pwd)/.ci/docker/common/install_executorch.sh"
-
   pushd /executorch
-  "${INSTALL_SCRIPT}" setup_executorch
+
+  export PYTHON_EXECUTABLE=python
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON"
+
+  # NB: We need to rebuild ExecuTorch runner here because it depends on PyTorch
+  # from the PR
+  bash .ci/scripts/setup-linux.sh --build-tool cmake
 
   echo "Run ExecuTorch unit tests"
   pytest -v -n auto
@@ -1573,6 +1552,10 @@ test_executorch() {
 
   popd
 
+  # Test torchgen generated code for Executorch.
+  echo "Testing ExecuTorch op registration"
+  "$BUILD_BIN_DIR"/test_edge_op_registration
+
   assert_git_not_dirty
 }
 
@@ -1580,7 +1563,6 @@ test_linux_aarch64() {
   python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
         test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
         test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
-        distributed/elastic/timer/api_test distributed/elastic/timer/local_timer_example distributed/elastic/timer/local_timer_test \
         --shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
 
   # Dynamo tests
@@ -1648,7 +1630,11 @@ elif [[ "${TEST_CONFIG}" == *xla* ]]; then
   build_xla
   test_xla
 elif [[ "$TEST_CONFIG" == *vllm* ]]; then
-    echo "vLLM CI uses TORCH_CUDA_ARCH_LIST: $TORCH_CUDA_ARCH_LIST"
+    if [[ "${BUILD_ENVIRONMENT}" == *cuda* ]]; then
+      TORCH_CUDA_ARCH_LIST=$(nvidia-smi --query-gpu=compute_cap --format=csv | tail -n 1)
+      export TORCH_CUDA_ARCH_LIST
+    fi
+    echo "VLLM CI TORCH_CUDA_ARCH_LIST: $TORCH_CUDA_ARCH_LIST"
     (cd .ci/lumen_cli && python -m pip install -e .)
     python -m cli.run test external vllm --test-plan "$TEST_CONFIG" --shard-id "$SHARD_NUMBER" --num-shards "$NUM_TEST_SHARDS"
 elif [[ "${TEST_CONFIG}" == *executorch* ]]; then
@@ -1730,6 +1716,11 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision
   test_inductor_shard "${SHARD_NUMBER}"
+  if [[ "${SHARD_NUMBER}" == 1 ]]; then
+    if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.9-gcc11-build ]]; then
+      test_inductor_distributed
+    fi
+  fi
 elif [[ "${TEST_CONFIG}" == *einops* ]]; then
   test_einops
 elif [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then
@@ -1779,8 +1770,6 @@ elif [[ "${BUILD_ENVIRONMENT}" == *xpu* ]]; then
   test_xpu_bin
 elif [[ "${TEST_CONFIG}" == smoke ]]; then
   test_python_smoke
-elif [[ "${TEST_CONFIG}" == smoke_b200 ]]; then
-  test_python_smoke_b200
 elif [[ "${TEST_CONFIG}" == h100_distributed ]]; then
   test_h100_distributed
 elif [[ "${TEST_CONFIG}" == "h100-symm-mem" ]]; then

.ci/pytorch/win-test-helpers/build_pytorch.bat

@@ -137,7 +137,7 @@ sccache --show-stats
 python -c "import os, glob; os.system('python -mpip install --no-index --no-deps ' + glob.glob('dist/*.whl')[0])"
 (
   if "%BUILD_ENVIRONMENT%"=="" (
-    echo NOTE: To run `import torch`, please make sure to activate the conda environment by running `call %CONDA_ROOT_DIR%\Scripts\activate.bat %CONDA_ROOT_DIR%\envs\py_tmp` in Command Prompt before running Git Bash.
+    echo NOTE: To run `import torch`, please make sure to activate the conda environment by running `call %CONDA_PARENT_DIR%\Miniconda3\Scripts\activate.bat %CONDA_PARENT_DIR%\Miniconda3` in Command Prompt before running Git Bash.
   ) else (
     copy /Y "dist\*.whl" "%PYTORCH_FINAL_PACKAGE_DIR%"
 

.ci/pytorch/win-test-helpers/installation-helpers/activate_miniconda3.bat

@@ -3,12 +3,12 @@ if "%BUILD_ENVIRONMENT%"=="" (
 ) else (
   set CONDA_PARENT_DIR=C:\Jenkins
 )
-set CONDA_ROOT_DIR=%CONDA_PARENT_DIR%\Miniconda3
+
 
 :: Be conservative here when rolling out the new AMI with conda. This will try
 :: to install conda as before if it couldn't find the conda installation. This
 :: can be removed eventually after we gain enough confidence in the AMI
-if not exist %CONDA_ROOT_DIR% (
+if not exist %CONDA_PARENT_DIR%\Miniconda3 (
   set INSTALL_FRESH_CONDA=1
 )
 
@@ -17,14 +17,10 @@ if "%INSTALL_FRESH_CONDA%"=="1" (
   if errorlevel 1 exit /b
   if not errorlevel 0 exit /b
 
-  %TMP_DIR_WIN%\Miniconda3-latest-Windows-x86_64.exe /InstallationType=JustMe /RegisterPython=0 /S /AddToPath=0 /D=%CONDA_ROOT_DIR%
+  %TMP_DIR_WIN%\Miniconda3-latest-Windows-x86_64.exe /InstallationType=JustMe /RegisterPython=0 /S /AddToPath=0 /D=%CONDA_PARENT_DIR%\Miniconda3
   if errorlevel 1 exit /b
   if not errorlevel 0 exit /b
 )
 
 :: Activate conda so that we can use its commands, i.e. conda, python, pip
-call %CONDA_ROOT_DIR%\Scripts\activate.bat %CONDA_ROOT_DIR%
-:: Activate conda so that we can use its commands, i.e. conda, python, pip
-call conda activate py_tmp
-
-call pip install -r .ci/docker/requirements-ci.txt
+call %CONDA_PARENT_DIR%\Miniconda3\Scripts\activate.bat %CONDA_PARENT_DIR%\Miniconda3

.ci/pytorch/win-test-helpers/setup_pytorch_env.bat

@@ -14,7 +14,7 @@ if not errorlevel 0 exit /b
 :: build\torch. Rather than changing all these references, making a copy of torch folder
 :: from conda to the current workspace is easier. The workspace will be cleaned up after
 :: the job anyway
-xcopy /s %CONDA_ROOT_DIR%\envs\py_tmp\Lib\site-packages\torch %TMP_DIR_WIN%\build\torch\
+xcopy /s %CONDA_PARENT_DIR%\Miniconda3\Lib\site-packages\torch %TMP_DIR_WIN%\build\torch\
 
 pushd .
 if "%VC_VERSION%" == "" (

.ci/pytorch/win-test.sh

@@ -38,20 +38,13 @@ if [[ "$BUILD_ENVIRONMENT" == *cuda* ]]; then
 fi
 
 # TODO: Move both of them to Windows AMI
-python -m pip install tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
-
-# Copied from https://github.com/pytorch/test-infra/blob/be01a40157c36cd5a48391fdf44a7bc3ebd4c7e3/aws/ami/windows/scripts/Installers/Install-Pip-Dependencies.ps1#L16 with some adjustments
-# pytest-rerunfailures==10.3 as 10.2 fails with INTERNALERROR> pluggy._manager.PluginValidationError: unknown hook 'pytest_configure_node'
-# scipy from 1.6.3 to 1.10
-# expecttest from 0.1.3 to 0.3.0
-# xdoctest from 1.0.2 to 1.3.0
-python -m pip install "future==0.18.2" "hypothesis==5.35.1" "expecttest==0.3.0" "librosa>=0.6.2" "scipy==1.10.1" "psutil==5.9.1" "pynvml==11.4.1" "pillow==9.2.0" "unittest-xml-reporting<=3.2.0,>=2.0.0" "pytest==7.1.3" "pytest-xdist==2.5.0" "pytest-flakefinder==1.1.0" "pytest-rerunfailures==10.3" "pytest-shard==0.1.2" "sympy==1.11.1" "xdoctest==1.3.0" "pygments==2.12.0" "opt-einsum>=3.3" "networkx==2.8.8" "mpmath==1.2.1" "pytest-cpp==2.3.0" "boto3==1.35.42"
+python -m pip install pytest-rerunfailures==10.3 pytest-cpp==2.3.0 tensorboard==2.13.0 protobuf==5.29.4 pytest-subtests==0.13.1
 
 # Install Z3 optional dependency for Windows builds.
 python -m pip install z3-solver==4.15.1.0
 
 # Install tlparse for test\dynamo\test_structured_trace.py UTs.
-python -m pip install tlparse==0.4.0
+python -m pip install tlparse==0.3.30
 
 # Install parameterized
 python -m pip install parameterized==0.8.1
@@ -59,6 +52,12 @@ python -m pip install parameterized==0.8.1
 # Install pulp for testing ilps under torch\distributed\_tools
 python -m pip install pulp==2.9.0
 
+# Install expecttest to merge https://github.com/pytorch/pytorch/pull/155308
+python -m pip install expecttest==0.3.0
+
+# Install intel-openmp
+python -m pip install intel-openmp==2025.1.1
+
 run_tests() {
     # Run nvidia-smi if available
     for path in '/c/Program Files/NVIDIA Corporation/NVSMI/nvidia-smi.exe' /c/Windows/System32/nvidia-smi.exe; do

.ci/pytorch/windows/internal/copy.bat

@@ -1,20 +1,12 @@
-
-if %CUDA_VERSION% geq 130 (
-    set "dll_path=bin\x64"
-) else (
-    set "dll_path=bin"
-)
-
-copy "%CUDA_PATH%\%dll_path%\cusparse*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cublas*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cudart*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\curand*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cufft*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\cusolver*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\nvrtc*64_*.dll*" pytorch\torch\lib
-copy "%CUDA_PATH%\%dll_path%\nvJitLink_*.dll*"  pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cusparse*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cublas*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cudart*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\curand*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cufft*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\cusolver*64_*.dll*" pytorch\torch\lib
 
 copy "%CUDA_PATH%\bin\cudnn*64_*.dll*" pytorch\torch\lib
+copy "%CUDA_PATH%\bin\nvrtc*64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\extras\CUPTI\lib64\cupti64_*.dll*" pytorch\torch\lib
 copy "%CUDA_PATH%\extras\CUPTI\lib64\nvperf_host*.dll*" pytorch\torch\lib
 
@@ -28,3 +20,8 @@ copy "%libuv_ROOT%\bin\uv.dll" pytorch\torch\lib
 if exist "C:\Windows\System32\zlibwapi.dll" (
     copy "C:\Windows\System32\zlibwapi.dll"  pytorch\torch\lib
 )
+
+::copy nvJitLink dll is requires for cuda 12+
+if exist "%CUDA_PATH%\bin\nvJitLink_*.dll*" (
+    copy "%CUDA_PATH%\bin\nvJitLink_*.dll*"  pytorch\torch\lib
+)

.ci/pytorch/windows/internal/driver_update.bat

@@ -1,9 +1,9 @@
-set WIN_DRIVER_VN=580.88
-set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe" & REM @lint-ignore
-curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe
+set WIN_DRIVER_VN=528.89
+set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe" & REM @lint-ignore
+curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe
 if errorlevel 1 exit /b 1
 
-start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe -s -noreboot
+start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe -s -noreboot
 if errorlevel 1 exit /b 1
 
-del %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe || ver > NUL
+del %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe || ver > NUL

.ci/pytorch/windows/internal/xpu_install.bat

@@ -13,9 +13,9 @@ if not exist "%SRC_DIR%\temp_build" mkdir "%SRC_DIR%\temp_build"
 :xpu_bundle_install_start
 
 set XPU_BUNDLE_PARENT_DIR=C:\Program Files (x86)\Intel\oneAPI
-set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
+set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/9d6d6c17-ca2d-4735-9331-99447e4a1280/intel-deep-learning-essentials-2025.0.1.28_offline.exe
 set XPU_BUNDLE_PRODUCT_NAME=intel.oneapi.win.deep-learning-essentials.product
-set XPU_BUNDLE_VERSION=2025.1.3+5
+set XPU_BUNDLE_VERSION=2025.0.1+20
 set XPU_BUNDLE_INSTALLED=0
 set XPU_BUNDLE_UNINSTALL=0
 set XPU_EXTRA_URL=NULL
@@ -24,9 +24,9 @@ set XPU_EXTRA_VERSION=2025.0.1+1226
 set XPU_EXTRA_INSTALLED=0
 set XPU_EXTRA_UNINSTALL=0
 
-if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.2] (
-    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/24751ead-ddc5-4479-b9e6-f9fe2ff8b9f2/intel-deep-learning-essentials-2025.2.1.25_offline.exe
-    set XPU_BUNDLE_VERSION=2025.2.1+20
+if not [%XPU_VERSION%]==[] if [%XPU_VERSION%]==[2025.1] (
+    set XPU_BUNDLE_URL=https://registrationcenter-download.intel.com/akdlm/IRC_NAS/75d4eb97-914a-4a95-852c-7b9733d80f74/intel-deep-learning-essentials-2025.1.3.8_offline.exe
+    set XPU_BUNDLE_VERSION=2025.1.3+5
 )
 
 :: Check if XPU bundle is target version or already installed
@@ -90,3 +90,14 @@ if errorlevel 1 exit /b 1
 del xpu_extra.exe
 
 :xpu_install_end
+
+if not "%XPU_ENABLE_KINETO%"=="1" goto install_end
+:: Install Level Zero SDK
+set XPU_EXTRA_LZ_URL=https://github.com/oneapi-src/level-zero/releases/download/v1.14.0/level-zero-sdk_1.14.0.zip
+curl -k -L %XPU_EXTRA_LZ_URL% --output "%SRC_DIR%\temp_build\level_zero_sdk.zip"
+echo "Installing level zero SDK..."
+7z x "%SRC_DIR%\temp_build\level_zero_sdk.zip" -o"%SRC_DIR%\temp_build\level_zero"
+set "INCLUDE=%SRC_DIR%\temp_build\level_zero\include;%INCLUDE%"
+del "%SRC_DIR%\temp_build\level_zero_sdk.zip"
+
+:install_end

.ci/wheel/build_wheel.sh

@@ -85,7 +85,7 @@ mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
 # Create an isolated directory to store this builds pytorch checkout and conda
 # installation
 if [[ -z "$MAC_PACKAGE_WORK_DIR" ]]; then
-    MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_${DESIRED_PYTHON}_$(date +%H%M%S)"
+    MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_conda_${DESIRED_PYTHON}_$(date +%H%M%S)"
 fi
 mkdir -p "$MAC_PACKAGE_WORK_DIR" || true
 if [[ -n ${GITHUB_ACTIONS} ]]; then
@@ -96,11 +96,11 @@ fi
 whl_tmp_dir="${MAC_PACKAGE_WORK_DIR}/dist"
 mkdir -p "$whl_tmp_dir"
 
-mac_version='macosx-11_0-arm64'
+mac_version='macosx_11_0_arm64'
 libtorch_arch='arm64'
 
 # Create a consistent wheel package name to rename the wheel to
-wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version//[-,]/_}.whl"
+wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version}.whl"
 
 ###########################################################
 
@@ -124,61 +124,91 @@ popd
 
 export TH_BINARY_BUILD=1
 export INSTALL_TEST=0 # dont install test binaries into site-packages
-export MACOSX_DEPLOYMENT_TARGET=11.0
+export MACOSX_DEPLOYMENT_TARGET=10.15
+export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 
+SETUPTOOLS_PINNED_VERSION="==70.1.0"
+PYYAML_PINNED_VERSION="=5.3"
 EXTRA_CONDA_INSTALL_FLAGS=""
 CONDA_ENV_CREATE_FLAGS=""
 RENAME_WHEEL=true
 case $desired_python in
     3.14t)
         echo "Using 3.14 deps"
-        mac_version='macosx-11.0-arm64'
-        NUMPY_PINNED_VERSION="==2.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
+        NUMPY_PINNED_VERSION="=2.1.0"
+        CONDA_ENV_CREATE_FLAGS="python-freethreading"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+        desired_python="3.14.0rc1"
         RENAME_WHEEL=false
         ;;
     3.14)
         echo "Using 3.14t deps"
-        mac_version='macosx-11.0-arm64'
-        NUMPY_PINNED_VERSION="==2.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
+        NUMPY_PINNED_VERSION="=2.1.0"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+        desired_python="3.14.0rc1"
         RENAME_WHEEL=false
         ;;
     3.13t)
         echo "Using 3.13 deps"
-        NUMPY_PINNED_VERSION="==2.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
+        NUMPY_PINNED_VERSION="=2.1.0"
+        CONDA_ENV_CREATE_FLAGS="python-freethreading"
+        EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+        desired_python="3.13"
         RENAME_WHEEL=false
         ;;
     3.13)
         echo "Using 3.13 deps"
-        NUMPY_PINNED_VERSION="==2.1.0"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
+        NUMPY_PINNED_VERSION="=2.1.0"
         ;;
     3.12)
         echo "Using 3.12 deps"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=6.0.1"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.11)
         echo "Using 3.11 deps"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=5.3"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     3.10)
         echo "Using 3.10 deps"
-        NUMPY_PINNED_VERSION="==2.0.2"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=5.3"
+        NUMPY_PINNED_VERSION="=2.0.2"
+        ;;
+    3.9)
+        echo "Using 3.9 deps"
+        SETUPTOOLS_PINNED_VERSION=">=70.1.0"
+        PYYAML_PINNED_VERSION=">=5.3"
+        NUMPY_PINNED_VERSION="=2.0.2"
         ;;
     *)
-        echo "Unsupported version $desired_python"
-        exit 1
+        echo "Using default deps"
+        NUMPY_PINNED_VERSION="=1.11.3"
         ;;
 esac
 
-PINNED_PACKAGES=(
-    "numpy${NUMPY_PINNED_VERSION}"
-)
-python -mvenv ~/${desired_python}-build
-source ~/${desired_python}-build/bin/activate
-retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements.txt"
+# Install into a fresh env
+tmp_env_name="wheel_py$python_nodot"
+conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
+source activate "$tmp_env_name"
+
+retry pip install -r "${pytorch_rootdir}/requirements-build.txt"
+pip install "numpy=${NUMPY_PINNED_VERSION}"  "pyyaml${PYYAML_PINNED_VERSION}" requests ninja "setuptools${SETUPTOOLS_PINNED_VERSION}" typing-extensions
+retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
 retry brew install libomp
 
-# For USE_DISTRIBUTED=1 on macOS, this enables gloo, which needs libuv, which
-# is build as part of tensorpipe submodule
+# For USE_DISTRIBUTED=1 on macOS, need libuv, which is build as part of tensorpipe submodule
 export USE_DISTRIBUTED=1
 
 export USE_MKLDNN=OFF
@@ -188,7 +218,7 @@ export BUILD_TEST=OFF
 pushd "$pytorch_rootdir"
 echo "Calling setup.py bdist_wheel at $(date)"
 
-_PYTHON_HOST_PLATFORM=${mac_version} ARCHFLAGS="-arch arm64" python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name "${mac_version//[-.]/_}"
+python setup.py bdist_wheel -d "$whl_tmp_dir"
 
 echo "Finished setup.py bdist_wheel at $(date)"
 

.circleci/scripts/binary_populate_env.sh

@@ -75,8 +75,8 @@ TRITON_VERSION=$(cat $PYTORCH_ROOT/.ci/docker/triton_version.txt)
 # Here PYTORCH_EXTRA_INSTALL_REQUIREMENTS is already set for the all the wheel builds hence append TRITON_CONSTRAINT
 TRITON_CONSTRAINT="platform_system == 'Linux' and platform_machine == 'x86_64'"
 
-# CUDA 12.9/13.0 builds have triton for Linux and Linux aarch64 binaries.
-if [[ "$DESIRED_CUDA" == "cu129" ]] || [[ "$DESIRED_CUDA" == "cu130" ]]; then
+# CUDA 12.9 builds have triton for Linux and Linux aarch64 binaries.
+if [[ "$DESIRED_CUDA" == "cu129" ]]; then
   TRITON_CONSTRAINT="platform_system == 'Linux'"
 fi
 

.circleci/scripts/binary_windows_build.sh

@@ -15,7 +15,8 @@ fi
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
     export VC_YEAR=2022
     export USE_SCCACHE=0
-    export XPU_VERSION=2025.2
+    export XPU_VERSION=2025.1
+    export XPU_ENABLE_KINETO=1
 fi
 
 echo "Free space on filesystem before build:"

.circleci/scripts/binary_windows_test.sh

@@ -8,7 +8,7 @@ export VC_YEAR=2022
 
 if [[ "$DESIRED_CUDA" == 'xpu' ]]; then
     export VC_YEAR=2022
-    export XPU_VERSION=2025.2
+    export XPU_VERSION=2025.1
 fi
 
 pushd "$PYTORCH_ROOT/.ci/pytorch/"

.flake8

@@ -73,7 +73,7 @@ exclude =
     ./docs/src,
     ./functorch/docs,
     ./functorch/examples,
-    ./functorch/docs/source/tutorials,
+    ./functorch/notebooks,
     ./scripts,
     ./test/generated_type_hints_smoketest.py,
     ./third_party,

.github/actionlint.yaml

@@ -12,16 +12,13 @@ self-hosted-runner:
     - linux.9xlarge.ephemeral
     - am2.linux.9xlarge.ephemeral
     - linux.12xlarge
-    - linux.12xlarge.memory
     - linux.24xlarge
-    - linux.24xlarge.memory
     - linux.24xlarge.ephemeral
     - linux.24xlarge.amd
     - linux.arm64.2xlarge
     - linux.arm64.2xlarge.ephemeral
     - linux.arm64.m7g.4xlarge
     - linux.arm64.m7g.4xlarge.ephemeral
-    - linux.arm64.r7g.12xlarge.memory
     - linux.4xlarge.nvidia.gpu
     - linux.8xlarge.nvidia.gpu
     - linux.16xlarge.nvidia.gpu

.github/actions/build-external-packages/action.yml

@@ -4,11 +4,6 @@ name: Build External packages
 description: build external packages for PyTorch
 
 inputs:
-  cuda-version:
-    description: CUDA version to use
-    type: string
-    required: true
-    default: '12.8.1'
   cuda-arch-list:
     description: TORCH_CUDA_ARCH_LIST (e.g., "8.0;8.9;9.0")
     type: string
@@ -49,12 +44,10 @@ runs:
       env:
         SCCACHE_BUCKET: ossci-compiler-cache-circleci-v2
         SCCACHE_REGION: us-east-1
-        CUDA_VERSION: ${{ inputs.cuda-version }}
         TORCH_CUDA_ARCH_LIST: ${{ inputs.cuda-arch-list }}
         BASE_IMAGE: ${{ inputs.docker-image }}
         BUILD_TARGETS: ${{ inputs.build-targets }}
-        PARENT_OUTPUT_DIR: ${{ inputs.output-dir }}
-        TORCH_WHEELS_PATH: ${{ inputs.torch-wheel-dir }}
+        PARENT_OUTPUT_DIR: ${{ inputs.output-dir}}
       shell: bash
       run: |
         set -euo pipefail
@@ -75,6 +68,7 @@ runs:
           export OUTPUT_DIR
           echo "Building external package: $target in directory $OUTPUT_DIR"
           python3 -m cli.run build external "$target"
+
         done
 
         END_TIME=$(date +%s)

.github/actions/checkout-pytorch/action.yml

@@ -57,21 +57,6 @@ runs:
         submodules: ${{ inputs.submodules }}
         show-progress: false
 
-    - name: Clean submodules post checkout
-      id: clean-submodules
-      if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' }}
-      shell: bash
-      env:
-        NO_SUDO: ${{ inputs.no-sudo }}
-      run: |
-        cd "${GITHUB_WORKSPACE}"
-        # Clean stale submodule dirs
-        if [ -z "${NO_SUDO}" ]; then
-          sudo git submodule foreach --recursive git clean -ffdx
-        else
-          git submodule foreach --recursive git clean -ffdx
-        fi
-
     - name: Clean workspace (try again)
       if: ${{ steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' &&
         (steps.first-clean.outcome != 'success' || steps.first-checkout-attempt.outcome != 'success') }}

.github/actions/reuse-old-whl/reuse_old_whl.py

@@ -264,7 +264,7 @@ def unzip_artifact_and_replace_files() -> None:
         change_content_to_new_version(f"artifacts/dist/{old_stem}/torch/version.py")
 
         for file in Path(f"artifacts/dist/{old_stem}").glob(
-            "*.dist-info/*",
+            "*.dist-info/**",
         ):
             change_content_to_new_version(file)
 

.github/actions/setup-win/action.yml

@@ -6,12 +6,6 @@ inputs:
   cuda-version:
     description: which cuda version to install, 'cpu' for none
     required: true
-  python-version:
-    required: false
-    type: string
-    default: "3.10"
-    description: |
-      The python version to be used. Will be 3.10 by default
 
 runs:
   using: composite
@@ -44,24 +38,18 @@ runs:
         CONDA="C:\Jenkins\Miniconda3\condabin\conda.bat"
 
         {
-          echo "CONDA=${CONDA}";
           echo "CONDA_RUN=${CONDA} run --no-capture-output";
           echo "CONDA_BUILD=${CONDA} run conda-build";
           echo "CONDA_INSTALL=${CONDA} install";
         } >> "${GITHUB_ENV}"
 
     - name: Setup Python3
-      env:
-          PYTHON_VERSION: ${{ inputs.python-version }}
       shell: bash
       run: |
         set +e
         set -x
 
-        # Create new py_tmp env with python-version
-        ${CONDA} create -y -n py_tmp python=${PYTHON_VERSION} intel-openmp
-
-        PYTHON3=$(${CONDA_RUN} -n py_tmp which python3)
+        PYTHON3=$(${CONDA_RUN} which python3)
         EXIT_CODE=$?
 
         if [[ "${EXIT_CODE}" == "0" ]]; then
@@ -74,7 +62,7 @@ runs:
           # installation, which is Python 3 based. Its Python is default to Python 3. Further, there
           # is also the Miniconda installation that is Python 2 based, and both can be installed if
           # needed. In both cases, Python binary is just called python
-          PYTHON=$(${CONDA_RUN} -n py_tmp which python)
+          PYTHON=$(${CONDA_RUN} which python)
           EXIT_CODE=$?
 
           if [[ "${EXIT_CODE}" == "0" ]]; then

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-87ff22e49ed0e92576c4935ccb8c143daac4a3cd
+10a5002c6195bd95e34df8fe28ff8a2d55a2a922

.github/ci_commit_pins/fbgemm_rocm.txt

@@ -1 +1 @@
-08ae0af1395c8d8471f4025deb6af9aef90b342f
+7f1de94a4c2d14f59ad4ca84538c36084ea6b2c8

.github/ci_commit_pins/vllm.txt

@@ -1 +1 @@
-090197034faf3b193c4467cedeb9281e3078892d
+add1adfec742dfb13e614dab3372b5aafd1ff046

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-c77852e117bdf056c8e9a087e51d6f65cf6ba53d
+a1c6ee92c85e8b0955c20892ed68f032a6015c09

.github/ci_configs/vllm/Dockerfile.tmp_vllm

@@ -12,46 +12,54 @@ ARG BUILD_BASE_IMAGE=torch-nightly-base
 # by default, it uses devel-ubuntu22.04 official image.
 ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
 
-# The logic is copied from https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile
-ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
 
-
-#################### TORCH NIGHTLY BASE IMAGE ####################
+#################### TORCH NIGHTLY  BASE IMAGE ####################
 # A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
-FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
+From nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
+ARG CUDA_VERSION=12.8.1
+ARG PYTHON_VERSION=3.12
+ARG TARGETPLATFORM
+ENV DEBIAN_FRONTEND=noninteractive
 
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-ARG GET_PIP_URL
+RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
+    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
 
-# Install Python and other dependencies
-RUN apt-get update -y \
-    && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-    && add-apt-repository -y ppa:deadsnakes/ppa \
-    && apt-get update -y \
-    && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-    && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-    && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-    && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-    && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
-    && python3 --version && python3 -m pip --version
+# Install Python and other dependencies if it does not existed
+RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
+      echo "Installing Python ${PYTHON_VERSION}..." && \
+      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
+      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
+      apt-get update -y && \
+      apt-get install -y ccache software-properties-common git curl sudo && \
+      for i in 1 2 3; do \
+        add-apt-repository -y ppa:deadsnakes/ppa && break || \
+        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+      done && \
+      apt-get update -y && \
+      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
+      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
+      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
+      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
+      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
+   else \
+      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
+   fi \
+   && python3 --version && python3 -m pip --version
 
 # Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
 # as it was causing spam when compiling the CUTLASS kernels
 # Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
 RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
-    if command -v apt-get >/dev/null; then \
-        if [ "$current_gcc_version" -lt 10 ]; then \
-            echo "GCC version is $current_gcc_version, installing gcc-10..."; \
-            apt-get update \
-            && apt-get install -y gcc-10 g++-10 \
-            && update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
-            && update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
-        else \
-            echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
-        fi \
-    fi \
-    && gcc --version && g++ --version
+    if [ "$current_gcc_version" -lt 10 ]; then \
+      echo "GCC version is $current_gcc_version, installing gcc-10..."; \
+      apt-get update && \
+      apt-get install -y gcc-10 g++-10 && \
+      update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 && \
+      update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
+    else \
+      echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
+    fi && \
+    gcc --version && g++ --version
 
 # install uv for faster pip installs
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -71,20 +79,11 @@ ENV UV_LINK_MODE=copy
 FROM ${BUILD_BASE_IMAGE} AS base
 USER root
 
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
-# Install some system dependencies and double check python version
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim; \
-    else \
-        dnf install -y git curl wget sudo; \
-    fi \
-    && python3 --version && python3 -m pip --version
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
 
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
@@ -119,15 +118,17 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
         echo "[INFO] Installing torch wheels to build vllm"; \
         torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist -name 'torchaudio*.whl' | head -n1 | xargs); \
-        uv pip install --system "${torch_whl}[opt-einsum]" "${vision_whl}" "${audio_whl}" /dist/*.whl; \
+        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
+        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
+        uv pip install --system "${torch_whl}[opt-einsum]"; \
+        uv pip install --system "${vision_whl}"; \
+        uv pip install --system "${audio_whl}"; \
     elif [ -n "$PINNED_TORCH_VERSION" ]; then \
         echo "[INFO] Installing pinned torch nightly version to build vllm: $PINNED_TORCH_VERSION"; \
-        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        uv pip install --system "$PINNED_TORCH_VERSION" --index-url https://download.pytorch.org/whl/nightly/cu128; \
     else \
         echo "[INFO] Installing torch nightly with latest one to build vllm"; \
-        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        uv pip install --system torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128; \
     fi
 
 # Install numba 0.61.2 for cuda environment
@@ -136,11 +137,12 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # Install common dependencies from vllm common.txt
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system -r requirements/common.txt
+uv pip install --system -r requirements/common.txt
+
 
 # Must put before installing xformers, so it can install the correct version of xfomrers.
-ARG xformers_cuda_arch_list='7.5;8.0+PTX;9.0a'
-ENV TORCH_CUDA_ARCH_LIST=${xformers_cuda_arch_list}
+ARG exformer_cuda_arch_list='7.5;8.0+PTX;9.0a'
+ENV TORCH_CUDA_ARCH_LIST=${exformer_cuda_arch_list}
 
 ARG max_jobs=16
 ENV MAX_JOBS=${max_jobs}
@@ -151,8 +153,8 @@ RUN pip freeze | grep -E 'ninja'
 
 # Build xformers with cuda and torch nightly/wheel
 # following official xformers guidance: https://github.com/facebookresearch/xformers#build
-# sha for https://github.com/facebookresearch/xformers/tree/v0.0.32.post2
-ARG XFORMERS_COMMIT=5d4b92a5e5a9c6c6d4878283f47d82e17995b468
+# sha for https://github.com/facebookresearch/xformers/tree/v0.0.31
+ARG XFORMERS_COMMIT=eb0946a363464da96ea40afd1a7f72a907c25497
 ENV CCACHE_DIR=/root/.cache/ccache
 
 RUN --mount=type=cache,target=/root/.cache/ccache \
@@ -174,7 +176,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 # Build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
 # track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
 RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
-
 RUN cat torch_build_versions.txt
 RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
 
@@ -186,6 +187,11 @@ RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
 FROM base AS build
 ARG TARGETPLATFORM
 
+ENV UV_HTTP_TIMEOUT=500
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
+# Use copy mode to avoid hardlink failures with Docker cache mounts
+ENV UV_LINK_MODE=copy
+
 COPY . .
 
 RUN python3 use_existing_torch.py
@@ -214,16 +220,11 @@ ARG SCCACHE_S3_NO_CREDENTIALS=0
 RUN --mount=type=cache,target=/root/.cache/uv \
     --mount=type=bind,source=.git,target=.git \
     if [ "$USE_SCCACHE" = "1" ]; then \
-        echo "Installing sccache..."; \
-        if [ "$TARGETPLATFORM" = "linux/arm64" ]; then \
-            SCCACHE_ARCHIVE="sccache-v0.8.1-aarch64-unknown-linux-musl"; \
-        else \
-            SCCACHE_ARCHIVE="sccache-v0.8.1-x86_64-unknown-linux-musl"; \
-        fi; \
-        curl -L -o sccache.tar.gz "https://github.com/mozilla/sccache/releases/download/v0.8.1/${SCCACHE_ARCHIVE}.tar.gz" \
+        echo "Installing sccache..." \
+        && curl -L -o sccache.tar.gz https://github.com/mozilla/sccache/releases/download/v0.8.1/sccache-v0.8.1-x86_64-unknown-linux-musl.tar.gz \
         && tar -xzf sccache.tar.gz \
-        && sudo mv "${SCCACHE_ARCHIVE}"/sccache /usr/bin/sccache \
-        && rm -rf sccache.tar.gz "${SCCACHE_ARCHIVE}" \
+        && sudo mv sccache-v0.8.1-x86_64-unknown-linux-musl/sccache /usr/bin/sccache \
+        && rm -rf sccache.tar.gz sccache-v0.8.1-x86_64-unknown-linux-musl \
         && export SCCACHE_BUCKET=${SCCACHE_BUCKET_NAME} \
         && export SCCACHE_REGION=${SCCACHE_REGION_NAME} \
         && export SCCACHE_S3_NO_CREDENTIALS=${SCCACHE_S3_NO_CREDENTIALS} \
@@ -249,9 +250,9 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
         python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
     fi
 
-RUN echo "[INFO] Listing current directory:" && \
+RUN echo "[DEBUG] Listing  current directory:" && \
     ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
+    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
 #################### WHEEL BUILD IMAGE ####################
@@ -261,42 +262,51 @@ RUN echo "[INFO] Listing current directory:" && \
 # Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
 FROM ${FINAL_BASE_IMAGE} AS vllm-base
 USER root
-
-ARG CUDA_VERSION
-ARG PYTHON_VERSION
-ARG GET_PIP_URL
-
-# TODO (huydhn): Only work with PyTorch manylinux builder
-ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
-
 # prepare for environment starts
 WORKDIR /workspace
 
-# Install Python and other dependencies
-RUN if command -v apt-get >/dev/null; then \
-        apt-get update -y \
-        && apt-get install -y ccache software-properties-common git curl wget sudo vim \
-        && add-apt-repository -y ppa:deadsnakes/ppa \
-        && apt-get update -y \
-        && apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
-        && update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
-        && update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
-        && ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
-        && curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
-    else \
-        dnf install -y git curl wget sudo; \
-    fi \
-    && python3 --version && python3 -m pip --version
+RUN PYTHON_VERSION_STR=$(echo ${PYTHON_VERSION} | sed 's/\.//g') && \
+    echo "export PYTHON_VERSION_STR=${PYTHON_VERSION_STR}" >> /etc/environment
+
+# Install Python and other dependencies if it does not existed
+RUN if ! command -v python3 >/dev/null || ! python3 --version | grep -q "${PYTHON_VERSION}"; then \
+      echo "Installing Python ${PYTHON_VERSION}..." && \
+      echo 'tzdata tzdata/Areas select America' | debconf-set-selections && \
+      echo 'tzdata tzdata/Zones/America select Los_Angeles' | debconf-set-selections && \
+      apt-get update -y && \
+      apt-get install -y ccache software-properties-common git curl sudo && \
+      for i in 1 2 3; do \
+        add-apt-repository -y ppa:deadsnakes/ppa && break || \
+        { echo "Attempt $i failed, retrying in 5s..."; sleep 5; }; \
+      done && \
+      apt-get update -y && \
+      apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv && \
+      update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 && \
+      update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} && \
+      ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config && \
+      curl -sS https://bootstrap.pypa.io/get-pip.py | python${PYTHON_VERSION}; \
+   else \
+      echo "Python ${PYTHON_VERSION} already present, skipping setup."; \
+   fi \
+   && python3 --version && python3 -m pip --version
+
 
 # Get the torch versions, and whls used in previous stagtes for consistency
 COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
-RUN echo "[INFO] Listing current directory before torch install step:" && \
+RUN echo "[DEBUG] Listing current directory before torch install step:" && \
     ls -al && \
-    echo "[INFO] Showing torch_build_versions.txt content:" && \
+    echo "[DEBUG] Showing torch_build_versions.txt content:" && \
     cat torch_build_versions.txt
 
+# Workaround for https://github.com/openai/triton/issues/2507 and
+# https://github.com/pytorch/pytorch/issues/107960 -- hopefully
+# this won't be needed for future versions of this docker image
+# or future versions of triton.
+RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
+
+
 # Install uv for faster pip installs if not existed
 RUN --mount=type=cache,target=/root/.cache/uv \
     if ! python3 -m uv --version > /dev/null 2>&1; then \
@@ -316,13 +326,15 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
     --mount=type=cache,target=/root/.cache/uv \
     if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
         torch_whl=$(find /dist -maxdepth 1 -name 'torch-*.whl' -print -quit); \
-        vision_whl=$(find /dist -name 'torchvision*.whl' | head -n1 | xargs); \
-        audio_whl=$(find /dist -name 'torchaudio*.whl' | head -n1 | xargs); \
+        vision_whl=$(find /dist/vision -name 'torchvision*.whl' | head -n1 | xargs); \
+        audio_whl=$(find /dist/audio -name 'torchaudio*.whl' | head -n1 | xargs); \
         echo "[INFO] Use wheels to build : '${torch_whl}' '${audio_whl}' '${vision_whl}'"; \
-        uv pip install --system "${torch_whl}[opt-einsum]" "${vision_whl}" "${audio_whl}" /dist/*.whl; \
+        uv pip install --system "${torch_whl}[opt-einsum]"; \
+        uv pip install --system "${vision_whl}"; \
+        uv pip install --system "${audio_whl}"; \
     else \
         echo "[INFO] Installing torch versions from torch_build_versions.txt"; \
-        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
+        uv pip install --system $(cat torch_build_versions.txt | xargs) --index-url https://download.pytorch.org/whl/nightly/cu128; \
     fi
 
 # Install the vllm wheel from previous stage
@@ -333,8 +345,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/xformers/*.whl --verbose
 
+
 # Build flashinfer from source.
-ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
+ARG torch_cuda_arch_list='8.0;8.9;9.0a'
 # install package for build flashinfer
 # see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
 
@@ -345,7 +358,7 @@ ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 # Build flashinfer for torch nightly from source around 10 mins
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
 # Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
-ARG FLASHINFER_GIT_REF="v0.2.14.post1"
+ARG FLASHINFER_GIT_REF="v0.2.9rc2"
 RUN --mount=type=cache,target=/root/.cache/uv \
     git clone --depth 1 --recursive --shallow-submodules \
         --branch ${FLASHINFER_GIT_REF} \
@@ -363,7 +376,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 
 # Logging to confirm the torch versions
 RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
-RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm\|^flashinfer' > build_summary.txt
 ################### VLLM INSTALLED IMAGE ####################
 
 
@@ -402,6 +414,11 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system -r requirements/nightly_torch_test.txt
 
+# Workaround for #17068
+# pinned commit for v2.2.4
+RUN --mount=type=cache,target=/root/.cache/uv \
+    uv pip install --system --no-build-isolation "git+https://github.com/state-spaces/mamba@95d8aba8a8c75aedcaa6143713b11e745e7cd0d9#egg=mamba-ssm"
+
 # Logging to confirm the torch versions
 RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
 
@@ -416,5 +433,4 @@ FROM scratch as export-wheels
 # Just copy the wheels we prepared in previous stages
 COPY --from=base /workspace/xformers-dist /wheels/xformers
 COPY --from=build /workspace/vllm-dist /wheels/vllm
-COPY --from=vllm-base /workspace/build_summary.txt /wheels/build_summary.txt
 COPY --from=vllm-base /workspace/wheels/flashinfer /wheels/flashinfer-python

.github/ci_configs/vllm/use_existing_torch.py

@@ -1,17 +0,0 @@
-import glob
-
-
-requires_files = glob.glob("requirements/*.txt")
-requires_files += ["pyproject.toml"]
-for file in requires_files:
-    print(f">>> cleaning {file}")
-    with open(file) as f:
-        lines = f.readlines()
-    if "torch" in "".join(lines).lower():
-        print("removed:")
-        with open(file, "w") as f:
-            for line in lines:
-                if "torch" not in line.lower():
-                    f.write(line)
-    print(f"<<< done cleaning {file}")
-    print()

.github/labeler.yml

@@ -130,6 +130,3 @@
 - torch/csrc/inductor/aoti_include/**
 - torchgen/aoti/**
 - torchgen/gen_aoti_c_shim.py
-
-"ciflow/vllm":
-- .github/ci_commit_pins/vllm.txt

.github/pytorch-probot.yml

@@ -36,7 +36,6 @@ ciflow_push_tags:
 - ciflow/win-arm64
 - ciflow/h100-symm-mem
 - ciflow/h100-cutlass-backend
-- ciflow/b200
 retryable_workflows:
 - pull
 - trunk

.github/requirements/pip-requirements-macOS.txt

@@ -15,7 +15,7 @@ optree==0.13.0
 packaging==23.1
 parameterized==0.8.1
 pillow==10.3.0
-protobuf==5.29.5
+protobuf==5.29.4
 psutil==5.9.8
 pygments==2.15.0
 pytest-cpp==2.3.0
@@ -26,9 +26,9 @@ pytest-xdist==3.3.1
 pytest==7.3.2
 pyyaml==6.0.2
 scipy==1.12.0
-setuptools==78.1.1
+setuptools==72.1.0
 sympy==1.13.3
-tlparse==0.4.0
+tlparse==0.3.30
 tensorboard==2.13.0
 typing-extensions==4.12.2
 unittest-xml-reporting<=3.2.0,>=2.0.0

.github/scripts/build_triton_wheel.py

@@ -84,7 +84,6 @@ def build_triton(
                 ["git", "checkout", f"release/{ver}.{rev}.x"], cwd=triton_basedir
             )
         else:
-            check_call(["git", "fetch", "origin", commit_hash], cwd=triton_basedir)
             check_call(["git", "checkout", commit_hash], cwd=triton_basedir)
 
         # change built wheel name and version

.github/scripts/docathon-label-sync.py

@@ -39,9 +39,7 @@ def main() -> None:
     pull_request_label_names = [label.name for label in pull_request_labels]
     issue_label_names = [label.name for label in issue_labels]
     labels_to_add = [
-        label
-        for label in issue_label_names
-        if label not in pull_request_label_names and label != "actionable"
+        label for label in issue_label_names if label not in pull_request_label_names
     ]
     if not labels_to_add:
         print("The pull request already has the same labels.")

.github/scripts/generate_binary_build_matrix.py

@@ -16,16 +16,18 @@ from typing import Optional
 
 
 # NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
-CUDA_ARCHES = ["12.6", "12.8", "13.0"]
+CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
+    "12.9": "12.9.1",
     "13.0": "13.0.0",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
     "12.8": "9",
+    "12.9": "9",
     "13.0": "9",
 }
 
@@ -38,82 +40,99 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
 
 CPU_S390X_ARCH = ["cpu-s390x"]
 
-CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
+CUDA_AARCH64_ARCHES = ["12.9-aarch64"]
 
 
 PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
     "12.6": (
-        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "12.8": (
-        "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'"
+    ),
+    "12.9": (
+        "nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "13.0": (
-        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | "
-        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' | "
-        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' | "
-        "nvidia-curand==10.4.0.35; platform_system == 'Linux' | "
-        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
-        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
-        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
-        "nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
-        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
-        "nvidia-cufile==1.15.0.42; platform_system == 'Linux'"
+        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cudnn-cu13==9.12.0.46; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvshmem-cu13==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'"
     ),
     "xpu": (
-        "intel-cmplr-lib-rt==2025.2.1 | "
-        "intel-cmplr-lib-ur==2025.2.1 | "
-        "intel-cmplr-lic-rt==2025.2.1 | "
-        "intel-sycl-rt==2025.2.1 | "
-        "oneccl-devel==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "oneccl==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "impi-rt==2021.16.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
-        "onemkl-sycl-blas==2025.2.0 | "
-        "onemkl-sycl-dft==2025.2.0 | "
-        "onemkl-sycl-lapack==2025.2.0 | "
-        "onemkl-sycl-rng==2025.2.0 | "
-        "onemkl-sycl-sparse==2025.2.0 | "
-        "dpcpp-cpp-rt==2025.2.1 | "
-        "intel-opencl-rt==2025.2.1 | "
-        "mkl==2025.2.0 | "
-        "intel-openmp==2025.2.1 | "
-        "tbb==2022.2.0 | "
-        "tcmlib==1.4.0 | "
-        "umf==0.11.0 | "
-        "intel-pti==0.13.1"
+        "intel-cmplr-lib-rt==2025.1.1 | "
+        "intel-cmplr-lib-ur==2025.1.1 | "
+        "intel-cmplr-lic-rt==2025.1.1 | "
+        "intel-sycl-rt==2025.1.1 | "
+        "oneccl-devel==2021.15.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "oneccl==2021.15.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "impi-rt==2021.15.0; platform_system == 'Linux' and platform_machine == 'x86_64' | "
+        "onemkl-sycl-blas==2025.1.0 | "
+        "onemkl-sycl-dft==2025.1.0 | "
+        "onemkl-sycl-lapack==2025.1.0 | "
+        "onemkl-sycl-rng==2025.1.0 | "
+        "onemkl-sycl-sparse==2025.1.0 | "
+        "dpcpp-cpp-rt==2025.1.1 | "
+        "intel-opencl-rt==2025.1.1 | "
+        "mkl==2025.1.0 | "
+        "intel-openmp==2025.1.1 | "
+        "tbb==2022.1.0 | "
+        "tcmlib==1.3.0 | "
+        "umf==0.10.0 | "
+        "intel-pti==0.12.3"
     ),
 }
 
@@ -191,7 +210,7 @@ LIBTORCH_CONTAINER_IMAGES: dict[str, str] = {
     "cpu": "libtorch-cxx11-builder:cpu",
 }
 
-FULL_PYTHON_VERSIONS = ["3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
+FULL_PYTHON_VERSIONS = ["3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t"]
 
 
 def translate_desired_cuda(gpu_arch_type: str, gpu_arch_version: str) -> str:
@@ -221,8 +240,12 @@ def generate_libtorch_matrix(
         if os == "linux":
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
+            if "13.0" in arches:
+                arches.remove("13.0")
         elif os == "windows":
             arches += CUDA_ARCHES
+            if "13.0" in arches:
+                arches.remove("13.0")
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -287,6 +310,8 @@ def generate_wheels_matrix(
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
             arches += CUDA_ARCHES + XPU_ARCHES
+            if "13.0" in arches:
+                arches.remove("13.0")
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts
@@ -309,20 +334,19 @@ def generate_wheels_matrix(
                 else arch_version
             )
 
+            # TODO: Enable python 3.13t on cpu-s390x
+            if gpu_arch_type == "cpu-s390x" and python_version == "3.13t":
+                continue
             # TODO: Enable python 3.14 for rest
-            if os not in [
-                "linux",
-                "linux-aarch64",
-                "linux-s390x",
-                "macos-arm64",
-                "windows",
-            ] and (python_version == "3.14" or python_version == "3.14t"):
+            if os not in ["linux", "linux-aarch64", "macos-arm64", "windows"] and (
+                python_version == "3.14" or python_version == "3.14t"
+            ):
                 continue
 
             # cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
 
             if (
-                arch_version in ["13.0", "12.8", "12.6"]
+                arch_version in ["13.0", "12.9", "12.8", "12.6"]
                 and os == "linux"
                 or arch_version in CUDA_AARCH64_ARCHES
             ):
@@ -386,5 +410,6 @@ def generate_wheels_matrix(
 
 
 validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
 validate_nccl_dep_consistency("12.8")
 validate_nccl_dep_consistency("12.6")

.github/scripts/generate_ci_workflows.py

@@ -135,7 +135,7 @@ ROCM_SMOKE_WORKFLOWS = [
         build_configs=generate_binary_build_matrix.generate_wheels_matrix(
             OperatingSystem.LINUX,
             arches=["6.4"],
-            python_versions=["3.10"],
+            python_versions=["3.9"],
         ),
         ciflow_config=CIFlowConfig(
             labels={

.github/scripts/prepare_vllm_wheels.sh

@@ -1,94 +0,0 @@
-#!/usr/bin/env bash
-
-set -eux
-
-torch_version=$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
-nightly=$(echo ${torch_version} | cut -d'.' -f4)
-
-# Copied from .ci/manywheel/build_common.sh
-make_wheel_record() {
-  fpath=$1
-  if echo $fpath | grep RECORD >/dev/null 2>&1; then
-    echo "$fpath,,"
-  else
-    fhash=$(openssl dgst -sha256 -binary $fpath | openssl base64 | sed -e 's/+/-/g' | sed -e 's/\//_/g' | sed -e 's/=//g')
-    fsize=$(ls -nl $fpath | awk '{print $5}')
-    echo "$fpath,sha256=$fhash,$fsize"
-  fi
-}
-
-change_wheel_version() {
-  local package=$1
-  local wheel=$2
-  local f_version=$3
-  local t_version=$4
-
-  # Extract the wheel
-  ${PYTHON_EXECUTABLE} -mwheel unpack $wheel
-
-  mv "${package}-${f_version}" "${package}-${t_version}"
-  # Change the version from f_version to t_version in the dist-info dir
-  pushd "${package}-${t_version}"
-  mv "${package}-${f_version}.dist-info" "${package}-${t_version}.dist-info"
-
-  pushd "${package}-${t_version}.dist-info"
-  sed -i "s/${package}-${f_version}.dist-info/${package}-${t_version}.dist-info/g" RECORD
-
-  # Update the version in METADATA and its SHA256 hash
-  sed -i "s/Version: ${f_version}/Version: ${t_version}/g" METADATA
-  # then add PyTorch nightly dependency of vLLM
-  if [[ "${package}" == vllm ]] || [[ "${package}" == xformers ]]; then
-    sed -i "/License-File/a\Requires-Dist: torch==${torch_version}" METADATA
-  fi
-  sed -i '/METADATA,sha256/d' RECORD
-  popd
-
-  make_wheel_record "${package}-${t_version}.dist-info/METADATA" >> "${package}-${t_version}.dist-info/RECORD"
-  popd
-
-  # Repack the wheel
-  ${PYTHON_EXECUTABLE} -mwheel pack "${package}-${t_version}"
-
-  # Clean up
-  rm -rf "${package}-${t_version}"
-}
-
-repackage_wheel() {
-  local package=$1
-  pushd $package
-
-  local orig_wheel=$(find . -name *${package//-/_}*)
-  local orig_version=$(unzip -p $orig_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
-
-  local version=""
-  if [[ "${package}" == vllm ]]; then
-    # Copied from vllm/.buildkite/scripts/upload-wheels.sh
-    version=1.0.0
-  else
-    version=$(echo $orig_version | tr '.+' '.' | cut -d'.' -f1-3)
-  fi
-  local nightly_version=$version.$nightly
-
-  # Use nightly version
-  change_wheel_version ${package//-/_} $orig_wheel $orig_version $nightly_version
-  # Clean up
-  rm "${orig_wheel}"
-
-  auditwheel repair --plat $PLATFORM *.whl \
-    --exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
-  local repair_wheel=$(find wheelhouse -name *${PLATFORM}*)
-  local repair_wheel=$(basename ${repair_wheel})
-  popd
-
-  cp ${package}/wheelhouse/${repair_wheel} .
-  rm -rf $package
-}
-
-# Require to re-package the wheel
-${PYTHON_EXECUTABLE} -mpip install wheel==0.45.1
-
-pushd externals/vllm/wheels
-for package in xformers flashinfer-python vllm; do
-  repackage_wheel $package
-done
-popd

.github/scripts/test_trymerge.py

@@ -27,7 +27,6 @@ from trymerge import (
     get_drci_classifications,
     gh_get_team_members,
     GitHubPR,
-    iter_issue_timeline_until_comment,
     JobCheckState,
     main as trymerge_main,
     MandatoryChecksMissingError,
@@ -35,8 +34,6 @@ from trymerge import (
     RE_GHSTACK_DESC,
     read_merge_rules,
     remove_job_name_suffix,
-    sha_from_committed_event,
-    sha_from_force_push_after,
     validate_revert,
 )
 
@@ -127,7 +124,7 @@ def mock_parse_args(revert: bool = False, force: bool = False) -> Any:
             self.force = force
             self.pr_num = 76123
             self.dry_run = True
-            self.comment_id = 12345  # Set to non-zero value
+            self.comment_id = 0
             self.reason = "this is for testing"
             self.ignore_current = False
             self.check_mergeability = False
@@ -155,9 +152,9 @@ def mock_revert(
 def mock_merge(
     pr: GitHubPR,
     repo: GitRepo,
-    comment_id: int,
     dry_run: bool = False,
     skip_mandatory_checks: bool = False,
+    comment_id: Optional[int] = None,
     timeout_minutes: int = 400,
     stale_pr_days: int = 3,
     ignore_current: bool = False,
@@ -473,9 +470,9 @@ class TestTryMerge(TestCase):
         mock_merge.assert_called_once_with(
             mock.ANY,
             mock.ANY,
-            comment_id=mock.ANY,
             dry_run=mock.ANY,
             skip_mandatory_checks=True,
+            comment_id=mock.ANY,
             ignore_current=False,
         )
 
@@ -488,9 +485,9 @@ class TestTryMerge(TestCase):
         mock_merge.assert_called_once_with(
             mock.ANY,
             mock.ANY,
-            comment_id=mock.ANY,
             dry_run=mock.ANY,
             skip_mandatory_checks=False,
+            comment_id=mock.ANY,
             ignore_current=False,
         )
 
@@ -1141,176 +1138,5 @@ Pull Request resolved: https://github.com/pytorch/pytorch/pull/154394"""
         )
 
 
-@mock.patch("trymerge.gh_graphql", side_effect=mocked_gh_graphql)
-@mock.patch("trymerge.gh_fetch_merge_base", return_value="")
-@mock.patch(
-    "trymerge.get_drci_classifications", side_effect=mocked_drci_classifications
-)
-class TestTimelineFunctions(TestCase):
-    """Tests for the new timeline-related functions"""
-
-    def test_sha_from_committed_event(self, *args: Any) -> None:
-        """Test extracting SHA from committed event"""
-        # Based on actual GitHub API format - committed events have "sha" at top level
-        event = {
-            "event": "committed",
-            "sha": "fb21ce932ded6670c918804a0d9151b773770a7c",
-        }
-        self.assertEqual(
-            sha_from_committed_event(event), "fb21ce932ded6670c918804a0d9151b773770a7c"
-        )
-
-        # Test with missing SHA
-        event_no_sha = {"event": "committed"}
-        self.assertIsNone(sha_from_committed_event(event_no_sha))
-
-    def test_sha_from_force_push_after(self, *args: Any) -> None:
-        """Test extracting SHA from force push event"""
-        # NOTE: The current function doesn't handle the actual GitHub API format
-        # Real force push events have "commit_id" at top level, but this function
-        # looks for "after", "after_commit", "after_sha", or "head_sha" fields
-
-        # Test with the legacy format the current function handles
-        event_legacy = {
-            "event": "head_ref_force_pushed",
-            "after": {"sha": "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e"},
-        }
-        self.assertEqual(
-            sha_from_force_push_after(event_legacy),
-            "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
-        )
-
-        # Test with current GitHub API format (should return None with current implementation)
-        event_real_api = {
-            "event": "head_ref_force_pushed",
-            "commit_id": "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
-        }
-        self.assertEqual(
-            sha_from_force_push_after(event_real_api),
-            "ef22bcbc54bb0f787e1e4ffd3d83df18fc407f5e",
-        )  # Current function doesn't handle commit_id
-
-        # Test with missing SHA
-        event_no_sha = {"event": "head_ref_force_pushed"}
-        self.assertIsNone(sha_from_force_push_after(event_no_sha))
-
-    @mock.patch("trymerge.gh_fetch_json_list")
-    def test_iter_issue_timeline_until_comment(
-        self, mock_gh_fetch_json_list: Any, *args: Any
-    ) -> None:
-        """Test timeline iteration until target comment"""
-        # Mock timeline data based on actual GitHub API format
-        timeline_data = [
-            {"event": "commented", "id": 100, "body": "first comment"},
-            {"event": "committed", "sha": "fb21ce932ded6670c918804a0d9151b773770a7c"},
-            {"event": "commented", "id": 200, "body": "target comment"},
-            {"event": "commented", "id": 300, "body": "after target"},
-        ]
-        mock_gh_fetch_json_list.return_value = timeline_data
-
-        # Test iteration stops at target comment
-        events = list(iter_issue_timeline_until_comment("pytorch", "pytorch", 123, 200))
-        self.assertEqual(len(events), 3)  # Should stop at target comment
-        self.assertEqual(events[0]["event"], "commented")
-        self.assertEqual(events[0]["id"], 100)
-        self.assertEqual(events[1]["event"], "committed")
-        self.assertEqual(events[1]["sha"], "fb21ce932ded6670c918804a0d9151b773770a7c")
-        self.assertEqual(events[2]["event"], "commented")
-        self.assertEqual(events[2]["id"], 200)
-
-    @mock.patch("trymerge.gh_fetch_json_list")
-    def test_iter_issue_timeline_until_comment_not_found(
-        self, mock_gh_fetch_json_list: Any, *args: Any
-    ) -> None:
-        """Test timeline iteration when target comment is not found"""
-        # Mock empty timeline
-        mock_gh_fetch_json_list.return_value = []
-
-        events = list(iter_issue_timeline_until_comment("pytorch", "pytorch", 123, 999))
-        self.assertEqual(len(events), 0)
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_commit_after_comment(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        """Test get_commit_sha_at_comment returns correct SHA after comment"""
-        mock_iter_timeline.return_value = [
-            {"event": "committed", "sha": "commit1"},
-            {"event": "committed", "sha": "commit2"},
-            {"event": "commented", "id": 100},
-            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertEqual(sha, "commit2")
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_force_push_before_comment(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.return_value = [
-            {"event": "committed", "sha": "commit1"},
-            {"event": "committed", "sha": "commit2"},
-            {"event": "head_ref_force_pushed", "commit_id": "commit3"},
-            {"event": "commented", "id": 100},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertEqual(sha, "commit3")
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_force_push_before_comment_legacy_mode(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.return_value = [
-            {"event": "committed", "sha": "commit1"},
-            {"event": "committed", "sha": "commit2"},
-            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
-            {"event": "commented", "id": 100},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertEqual(sha, "commit3")
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_multiple_comments(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.return_value = [
-            {"event": "committed", "sha": "commit1"},
-            {"event": "commented", "id": 100},
-            {"event": "committed", "sha": "commit2"},
-            {"event": "commented", "id": 200},
-            {"event": "head_ref_force_pushed", "after": {"sha": "commit3"}},
-            {"event": "commented", "id": 300},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(200)
-        self.assertEqual(sha, "commit2")
-        sha = pr.get_commit_sha_at_comment(300)
-        self.assertEqual(sha, "commit3")
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_no_events(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.return_value = [
-            {"event": "commented", "id": 100},
-            {"event": "labeled", "label": {"name": "test"}},
-        ]
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertIsNone(sha)
-
-    @mock.patch("trymerge.iter_issue_timeline_until_comment")
-    def test_get_commit_sha_at_comment_exception(
-        self, mock_iter_timeline: Any, *args: Any
-    ) -> None:
-        mock_iter_timeline.side_effect = Exception("API error")
-        pr = GitHubPR("pytorch", "pytorch", 77700)
-        sha = pr.get_commit_sha_at_comment(100)
-        self.assertIsNone(sha)
-
-
 if __name__ == "__main__":
     main()

.github/scripts/trymerge.py

@@ -450,63 +450,6 @@ HAS_NO_CONNECTED_DIFF_TITLE = (
 IGNORABLE_FAILED_CHECKS_THESHOLD = 10
 
 
-def iter_issue_timeline_until_comment(
-    org: str, repo: str, issue_number: int, target_comment_id: int, max_pages: int = 200
-) -> Any:
-    """
-    Yield timeline entries in order until (and including) the entry whose id == target_comment_id
-    for a 'commented' event. Stops once the target comment is encountered.
-    """
-    page = 1
-
-    while page <= max_pages:
-        url = (
-            f"https://api.github.com/repos/{org}/{repo}/issues/{issue_number}/timeline"
-        )
-        params = {"per_page": 100, "page": page}
-
-        batch = gh_fetch_json_list(url, params)
-
-        if not batch:
-            return
-        for ev in batch:
-            # The target is the issue comment row with event == "commented" and id == issue_comment_id
-            if ev.get("event") == "commented" and ev.get("id") == target_comment_id:
-                yield ev  # nothing in the timeline after this matters, so stop early
-                return
-            yield ev
-        if len(batch) < 100:
-            return
-        page += 1
-
-    # If we got here without finding the comment, then we either hit a bug or some github PR
-    # has a _really_ long timeline.
-    # The max # of pages found on any pytorch/pytorch PR at the time of this change was 41
-    raise RuntimeError(
-        f"Could not find a merge commit in the first {max_pages} pages of the timeline at url {url}."
-        f"This is most likely a bug, please report it to the @pytorch/pytorch-dev-infra team."
-    )
-
-
-def sha_from_committed_event(ev: dict[str, Any]) -> Optional[str]:
-    """Extract SHA from committed event in timeline"""
-    return ev.get("sha")
-
-
-def sha_from_force_push_after(ev: dict[str, Any]) -> Optional[str]:
-    """Extract SHA from force push event in timeline"""
-    # The current GitHub API format
-    commit_id = ev.get("commit_id")
-    if commit_id:
-        return str(commit_id)
-
-    # Legacy format
-    after = ev.get("after") or ev.get("after_commit") or {}
-    if isinstance(after, dict):
-        return after.get("sha") or after.get("oid")
-    return ev.get("after_sha") or ev.get("head_sha")
-
-
 def gh_get_pr_info(org: str, proj: str, pr_no: int) -> Any:
     rc = gh_graphql(GH_GET_PR_INFO_QUERY, name=proj, owner=org, number=pr_no)
     return rc["data"]["repository"]["pullRequest"]
@@ -794,24 +737,16 @@ class GitHubPR:
     def last_commit(self) -> Any:
         return self.info["commits"]["nodes"][-1]["commit"]
 
-    def last_commit_sha(self, default: Optional[str] = None) -> str:
-        # for commits, the oid is the sha
-
-        if default is None:
-            return str(self.last_commit()["oid"])
-
-        return str(self.last_commit().get("oid", default))
-
     def get_merge_base(self) -> str:
         if self.merge_base:
             return self.merge_base
 
-        last_commit_sha = self.last_commit_sha()
+        last_commit_oid = self.last_commit()["oid"]
         # NB: We could use self.base_ref() here for regular PR, however, that doesn't
         # work for ghstack where the base is the custom branch, i.e. gh/USER/ID/base,
         # so let's just use main instead
         self.merge_base = gh_fetch_merge_base(
-            self.org, self.project, last_commit_sha, self.default_branch()
+            self.org, self.project, last_commit_oid, self.default_branch()
         )
 
         # Fallback to baseRefOid if the API call fails, i.e. rate limit. Note that baseRefOid
@@ -900,44 +835,6 @@ class GitHubPR:
     def get_commit_count(self) -> int:
         return int(self.info["commits_with_authors"]["totalCount"])
 
-    def get_commit_sha_at_comment(self, comment_id: int) -> Optional[str]:
-        """
-        Get the PR head commit SHA that was present when a specific comment was posted.
-        This ensures we only merge the state of the PR at the time the merge command was issued,
-        not any subsequent commits that may have been pushed after.
-
-        Returns None if no head-changing events found before the comment or if the comment was not found.
-        """
-        head = None
-
-        try:
-            for event in iter_issue_timeline_until_comment(
-                self.org, self.project, self.pr_num, comment_id
-            ):
-                etype = event.get("event")
-                if etype == "committed":
-                    sha = sha_from_committed_event(event)
-                    if sha:
-                        head = sha
-                        print(f"Timeline: Found commit event for SHA {sha}")
-                elif etype == "head_ref_force_pushed":
-                    sha = sha_from_force_push_after(event)
-                    if sha:
-                        head = sha
-                        print(f"Timeline: Found force push event for SHA {sha}")
-                elif etype == "commented":
-                    if event.get("id") == comment_id:
-                        print(f"Timeline: Found final comment with sha {sha}")
-                        return head
-        except Exception as e:
-            print(
-                f"Warning: Failed to reconstruct timeline for comment {comment_id}: {e}"
-            )
-            return None
-
-        print(f"Did not find comment with id {comment_id} in the PR timeline")
-        return None
-
     def get_pr_creator_login(self) -> str:
         return cast(str, self.info["author"]["login"])
 
@@ -1254,7 +1151,7 @@ class GitHubPR:
         *,
         skip_mandatory_checks: bool = False,
         dry_run: bool = False,
-        comment_id: int,
+        comment_id: Optional[int] = None,
         ignore_current_checks: Optional[list[str]] = None,
     ) -> None:
         # Raises exception if matching rule is not found
@@ -1270,7 +1167,7 @@ class GitHubPR:
             skip_internal_checks=can_skip_internal_checks(self, comment_id),
             ignore_current_checks=ignore_current_checks,
         )
-        additional_merged_prs = self.merge_changes_locally(
+        additional_merged_prs = self.merge_changes(
             repo, skip_mandatory_checks, comment_id
         )
 
@@ -1299,7 +1196,7 @@ class GitHubPR:
                 broken_trunk_checks=ignorable_checks.get("BROKEN_TRUNK", []),
                 flaky_checks=ignorable_checks.get("FLAKY", []),
                 unstable_checks=ignorable_checks.get("UNSTABLE", []),
-                last_commit_sha=self.last_commit_sha(default=""),
+                last_commit_sha=self.last_commit().get("oid", ""),
                 merge_base_sha=self.get_merge_base(),
                 merge_commit_sha=merge_commit_sha,
                 is_failed=False,
@@ -1320,7 +1217,7 @@ class GitHubPR:
             dry_run=dry_run,
         )
 
-    def merge_changes_locally(
+    def merge_changes(
         self,
         repo: GitRepo,
         skip_mandatory_checks: bool = False,
@@ -1329,15 +1226,27 @@ class GitHubPR:
         skip_all_rule_checks: bool = False,
     ) -> list["GitHubPR"]:
         """
-        :param skip_all_rule_checks: If true, skips all rule checks on ghstack PRs, useful for dry-running merge locally
+        :param skip_all_rule_checks: If true, skips all rule checks, useful for dry-running merge locally
         """
         branch_to_merge_into = self.default_branch() if branch is None else branch
         if repo.current_branch() != branch_to_merge_into:
             repo.checkout(branch_to_merge_into)
+        if not self.is_ghstack_pr():
+            msg = self.gen_commit_message()
+            pr_branch_name = f"__pull-request-{self.pr_num}__init__"
+            repo.fetch(self.last_commit()["oid"], pr_branch_name)
+            repo._run_git("merge", "--squash", pr_branch_name)
+            repo._run_git("commit", f'--author="{self.get_author()}"', "-m", msg)
 
-        # It's okay to skip the commit SHA check for ghstack PRs since
-        # authoring requires write access to the repo.
-        if self.is_ghstack_pr():
+            # Did the PR change since we started the merge?
+            pulled_sha = repo.show_ref(pr_branch_name)
+            latest_pr_status = GitHubPR(self.org, self.project, self.pr_num)
+            if pulled_sha != latest_pr_status.last_commit()["oid"]:
+                raise RuntimeError(
+                    "PR has been updated since CI checks last passed. Please rerun the merge command."
+                )
+            return []
+        else:
             return self.merge_ghstack_into(
                 repo,
                 skip_mandatory_checks,
@@ -1345,48 +1254,6 @@ class GitHubPR:
                 skip_all_rule_checks=skip_all_rule_checks,
             )
 
-        msg = self.gen_commit_message()
-        pr_branch_name = f"__pull-request-{self.pr_num}__init__"
-
-        # Determine which commit SHA to merge
-        commit_to_merge = None
-        if not comment_id:
-            raise ValueError("Must provide --comment-id when merging regular PRs")
-
-        # Get the commit SHA that was present when the comment was made
-        commit_to_merge = self.get_commit_sha_at_comment(comment_id)
-        if not commit_to_merge:
-            raise RuntimeError(
-                f"Could not find commit that was pushed before comment {comment_id}"
-            )
-
-        # Validate that this commit is the latest commit on the PR
-        latest_commit = self.last_commit_sha()
-        if commit_to_merge != latest_commit:
-            raise RuntimeError(
-                f"Commit {commit_to_merge} was HEAD when comment {comment_id} was posted "
-                f"but now the latest commit on the PR is {latest_commit}. "
-                f"Please re-issue the merge command to merge the latest commit."
-            )
-
-        print(f"Merging commit {commit_to_merge} locally")
-
-        repo.fetch(commit_to_merge, pr_branch_name)
-        repo._run_git("merge", "--squash", pr_branch_name)
-        repo._run_git("commit", f'--author="{self.get_author()}"', "-m", msg)
-
-        # Did the PR change since we started the merge?
-        pulled_sha = repo.show_ref(pr_branch_name)
-        latest_pr_status = GitHubPR(self.org, self.project, self.pr_num)
-        if (
-            pulled_sha != latest_pr_status.last_commit_sha()
-            or pulled_sha != commit_to_merge
-        ):
-            raise RuntimeError(
-                "PR has been updated since CI checks last passed. Please rerun the merge command."
-            )
-        return []
-
 
 class MergeRuleFailedError(RuntimeError):
     def __init__(self, message: str, rule: Optional["MergeRule"] = None) -> None:
@@ -1591,7 +1458,7 @@ def find_matching_merge_rule(
             pending_checks = []
             failed_checks = []
 
-        hud_link = f"https://hud.pytorch.org/{pr.org}/{pr.project}/commit/{pr.last_commit_sha()}"
+        hud_link = f"https://hud.pytorch.org/{pr.org}/{pr.project}/commit/{pr.last_commit()['oid']}"
         if len(failed_checks) > 0:
             if reject_reason_score < 30000:
                 reject_reason_score = 30000
@@ -2289,14 +2156,14 @@ def categorize_checks(
 def merge(
     pr: GitHubPR,
     repo: GitRepo,
-    comment_id: int,
     dry_run: bool = False,
     skip_mandatory_checks: bool = False,
+    comment_id: Optional[int] = None,
     timeout_minutes: int = 400,
     stale_pr_days: int = 3,
     ignore_current: bool = False,
 ) -> None:
-    initial_commit_sha = pr.last_commit_sha()
+    initial_commit_sha = pr.last_commit()["oid"]
     pr_link = f"https://github.com/{pr.org}/{pr.project}/pull/{pr.pr_num}"
     print(f"Attempting merge of {initial_commit_sha} ({pr_link})")
 
@@ -2367,7 +2234,7 @@ def merge(
             f"Attempting merge of https://github.com/{pr.org}/{pr.project}/pull/{pr.pr_num} ({elapsed_time / 60} minutes elapsed)"
         )
         pr = GitHubPR(pr.org, pr.project, pr.pr_num)
-        if initial_commit_sha != pr.last_commit_sha():
+        if initial_commit_sha != pr.last_commit()["oid"]:
             raise RuntimeError(
                 "New commits were pushed while merging. Please rerun the merge command."
             )
@@ -2534,7 +2401,7 @@ def main() -> None:
     if args.check_mergeability:
         if pr.is_ghstack_pr():
             get_ghstack_prs(repo, pr)  # raises error if out of sync
-        pr.merge_changes_locally(
+        pr.merge_changes(
             repo,
             skip_mandatory_checks=True,
             skip_all_rule_checks=True,
@@ -2549,18 +2416,12 @@ def main() -> None:
         gh_post_pr_comment(org, project, args.pr_num, message, dry_run=args.dry_run)
         return
     try:
-        # Ensure comment id is set, else fail
-        if not args.comment_id:
-            raise ValueError(
-                "Comment ID is required for merging PRs, please provide it using --comment-id"
-            )
-
         merge(
             pr,
             repo,
-            comment_id=args.comment_id,
             dry_run=args.dry_run,
             skip_mandatory_checks=args.force,
+            comment_id=args.comment_id,
             ignore_current=args.ignore_current,
         )
     except Exception as e:
@@ -2582,7 +2443,7 @@ def main() -> None:
                 broken_trunk_checks=[],
                 flaky_checks=[],
                 unstable_checks=[],
-                last_commit_sha=pr.last_commit_sha(default=""),
+                last_commit_sha=pr.last_commit().get("oid", ""),
                 merge_base_sha=pr.get_merge_base(),
                 is_failed=True,
                 skip_mandatory_checks=args.force,

.github/templates/common.yml.j2

@@ -4,7 +4,7 @@
 {%- set download_artifact_action = "actions/download-artifact@v4.1.7" -%}
 
 {%- set timeout_minutes = 240 -%}
-{%- set timeout_minutes_windows_binary = 360 -%}
+{%- set timeout_minutes_windows_binary = 300 -%}
 
 {%- macro concurrency(build_environment) -%}
 concurrency:

.github/templates/linux_binary_build_workflow.yml.j2

@@ -135,7 +135,7 @@ jobs:
       contents: read
     steps:
       - name: Setup XPU
-        uses: pytorch/pytorch/.github/actions/setup-xpu@main
+        uses: ./.github/actions/setup-xpu
       - name: configure aws credentials
         id: aws_creds
         uses: aws-actions/configure-aws-credentials@v4

.github/templates/macos_binary_build_workflow.yml.j2

@@ -22,16 +22,6 @@ name: !{{ build_environment }}
           echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
 {%- endmacro %}
 
-{%- macro setup_python(py_ver) -%}
-      - name: Setup Python
-        uses: actions/setup-python@v6
-        with:
-          # TODO: Removeme once 3.14 is out
-          # .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
-          python-version: "!{{ (py_ver.strip('t') + '.4') if '3.14' not in py_ver else '3.14.0-rc.2' }}"
-          freethreaded: !{{ "true" if py_ver.endswith('t') else "false" }}
-{%- endmacro %}
-
 on:
 # TODO: Migrate to new ciflow trigger, reference https://github.com/pytorch/pytorch/pull/70321
   push:
@@ -71,13 +61,28 @@ jobs:
     {%- endif %}
     steps:
       !{{ set_runner_specific_vars() }}
-      !{{ setup_python(config.get("python_version", "3.10")) }}
+      - name: Install conda and dependencies
+        run: |
+          # Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
+          curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
+          chmod +x "${RUNNER_TEMP}/conda.sh"
+          /bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
+          echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
+          if [ -d "/Applications/Xcode_14.3.1.app" ]; then
+            echo "DEVELOPER_DIR=/Applications/Xcode_14.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
+          elif [ -d "/Applications/Xcode_13.3.1.app" ]; then
+            echo "DEVELOPER_DIR=/Applications/Xcode_13.3.1.app/Contents/Developer" >> "${GITHUB_ENV}"
+          fi
       !{{ common.checkout(deep_clone=False, directory="pytorch") }}
       - name: Populate binary env
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
       - name: Build PyTorch binary
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -94,6 +99,8 @@ jobs:
 {%- if config["package_type"] == "wheel" %}
       - name: Test PyTorch wheel
         run: |
+          # shellcheck disable=SC1091
+          source "${RUNNER_TEMP}/anaconda/bin/activate"
           set -eux -o pipefail
           # shellcheck disable=SC1090
           source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@@ -104,9 +111,33 @@ jobs:
 
           SMOKE_TEST_PARAMS=""
 
+          EXTRA_CONDA_INSTALL_FLAGS=""
+          CONDA_ENV_CREATE_FLAGS=""
+          # shellcheck disable=SC2153
+          case $DESIRED_PYTHON in
+            3.14t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.14)
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
+              desired_python="3.14.0rc1"
+              ;;
+            3.13t)
+              CONDA_ENV_CREATE_FLAGS="python-freethreading"
+              EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
+              desired_python="3.13"
+              ;;
+            *)
+              # shellcheck disable=SC2153
+              desired_python=${DESIRED_PYTHON}
+              ;;
+          esac
+
           # shellcheck disable=SC2086
-          python -mvenv test_venv
-          source test_venv/bin/activate
+          conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
+          conda activate test_conda_env
           pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
 
           # shellcheck disable=SC2086

.github/templates/upload.yml.j2

@@ -33,7 +33,7 @@
   {%- if is_windows %}
       # This is a dummy value for libtorch to work correctly with our batch scripts
       # without this value pip does not get installed for some reason
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
   {%- endif %}
 
 {%- else %}

.github/workflows/_binary-test-linux.yml

@@ -187,6 +187,8 @@ jobs:
 
       - name: Install nvidia driver, nvidia-docker runtime, set GPU_FLAG
         uses: pytorch/test-infra/.github/actions/setup-nvidia@main
+        with:
+          driver-version: ${{ startsWith(inputs.GPU_ARCH_VERSION, '13') && '580.65.06' || '570.133.07' }}
         if: ${{ inputs.GPU_ARCH_TYPE == 'cuda' && steps.filter.outputs.is-test-matrix-empty == 'False' }}
 
       - name: configure aws credentials

.github/workflows/_docs.yml

@@ -75,6 +75,10 @@ jobs:
             runner: ${{ inputs.runner_prefix }}linux.2xlarge
             # It takes less than 30m to finish python docs unless there are issues
             timeout-minutes: 30
+          - docs_type: functorch
+            runner: ${{ inputs.runner_prefix }}linux.2xlarge
+            # It takes less than 15m to finish functorch docs unless there are issues
+            timeout-minutes: 15
     # Set a fixed name for this job instead of using the current matrix-generated name, i.e. build-docs (cpp, linux.12xlarge, 180)
     # The current name requires updating the database last docs push query from test-infra every time the matrix is updated
     name: build-docs-${{ matrix.docs_type }}-${{ inputs.push }}
@@ -207,6 +211,16 @@ jobs:
           path: cppdocs/
           s3-prefix: pytorch/pytorch/${{ github.event.pull_request.number }}/cppdocs
 
+      - name: Upload functorch Docs Preview
+        uses: seemethere/upload-artifact-s3@baba72d0712b404f646cebe0730933554ebce96a # v5.1.0
+        if: ${{ github.event_name == 'pull_request' && matrix.docs_type == 'functorch' && steps.build-docs.outcome == 'success' }}
+        with:
+          retention-days: 14
+          s3-bucket: doc-previews
+          if-no-files-found: error
+          path: functorch_ghpages/nightly/
+          s3-prefix: pytorch/pytorch/${{ github.event.pull_request.number }}/functorchdocs
+
       - name: Teardown Linux
         uses: pytorch/test-infra/.github/actions/teardown-linux@main
         if: always()

.github/workflows/_get-changed-files.yml

@@ -2,12 +2,6 @@ name: Get Changed Files
 
 on:
   workflow_call:
-    inputs:
-      all_files:
-        description: "Whether to return all files instead of just changed files"
-        required: false
-        type: boolean
-        default: false
     outputs:
       changed-files:
         description: "List of changed files (space-separated) or '*' if not in a PR"
@@ -32,23 +26,17 @@ jobs:
             # Get the PR number from the github context
             PR_NUMBER="${{ github.event.number }}"
 
-            # Check if all_files is requested
-            if [ "${{ inputs.all_files }}" = "true" ]; then
-              echo "all_files input is true, returning all files"
-              echo "changed-files=*" >> "$GITHUB_OUTPUT"
-            else
-              # Use gh CLI to get changed files in the PR with explicit repo
-              CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
+            # Use gh CLI to get changed files in the PR with explicit repo
+            CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
 
-              if [ -z "$CHANGED_FILES" ]; then
-                echo "No changed files found, setting to '*'"
-                CHANGED_FILES="*"
-              fi
-
-              echo "Changed files: $CHANGED_FILES"
-              echo "changed-files=$CHANGED_FILES" >> "$GITHUB_OUTPUT"
+            if [ -z "$CHANGED_FILES" ]; then
+              echo "No changed files found, setting to '*'"
+              CHANGED_FILES="*"
             fi
 
+            echo "Changed files: $CHANGED_FILES"
+            echo "changed-files=$CHANGED_FILES" >> "$GITHUB_OUTPUT"
+
           else
             echo "Not in PR context, setting changed files to '*'"
             echo "changed-files=*" >> "$GITHUB_OUTPUT"

.github/workflows/_link_check.yml

@@ -13,7 +13,6 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || !contains(github.event.pull_request.labels.*.name, 'skip-url-lint') }}
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
-      job-name: lint-urls
       timeout: 120
       runner: ${{ inputs.runner }}linux.2xlarge
       docker-image: ci-image:pytorch-linux-jammy-linter
@@ -39,7 +38,6 @@ jobs:
     if: ${{ github.event_name != 'pull_request' || !contains(github.event.pull_request.labels.*.name, 'skip-xref-lint') }}
     uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
     with:
-      job-name: lint-xrefs
       timeout: 60
       runner: ${{ inputs.runner }}linux.2xlarge
       docker-image: ci-image:pytorch-linux-jammy-linter

.github/workflows/_linux-test.yml

@@ -169,7 +169,7 @@ jobs:
         id: install-nvidia-driver
         uses: pytorch/test-infra/.github/actions/setup-nvidia@main
         with:
-          driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '580.82.07' }}
+          driver-version: ${{ matrix.config == 'legacy_nvidia_driver' && '525.105.17' || '570.133.07' }}
         if: ${{ contains(inputs.build-environment, 'cuda') && !contains(matrix.config, 'nogpu') && steps.check_container_runner.outputs.IN_CONTAINER_RUNNER == 'false' && !contains(matrix.runner, 'b200') }}
 
       - name: Setup GPU_FLAG for docker run
@@ -409,7 +409,7 @@ jobs:
           job_identifier: ${{ github.workflow }}_${{ inputs.build-environment }}
 
       - name: Authenticate with AWS
-        if: ${{ always() && contains(matrix.runner, 'b200') }}
+        if: ${{ contains(matrix.runner, 'b200') }}
         uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
         with:
           role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_upload-benchmark-results

.github/workflows/_rocm-test.yml

@@ -62,11 +62,6 @@ on:
         required: false
         type: number
         default: 1
-    secrets:
-      HUGGING_FACE_HUB_TOKEN:
-        required: false
-        description: |
-          HF Auth token to avoid rate limits when downloading models or datasets from hub
 env:
   GIT_DEFAULT_BRANCH: ${{ github.event.repository.default_branch }}
 
@@ -81,9 +76,10 @@ jobs:
     strategy:
       matrix: ${{ fromJSON(inputs.test-matrix) }}
       fail-fast: false
-    runs-on: ${{ matrix.runner }}
     timeout-minutes: ${{ matrix.mem_leak_check == 'mem_leak_check' && 600 || inputs.timeout-minutes }}
+    runs-on: ${{ matrix.runner }}
     steps:
+      # [see note: pytorch repo ref]
       - name: Checkout PyTorch
         uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
         with:
@@ -135,9 +131,6 @@ jobs:
 
       - name: Start monitoring script
         id: monitor-script
-        if: ${{ !inputs.disable-monitor }}
-        shell: bash
-        continue-on-error: true
         env:
           JOB_ID: ${{ steps.get-job-id.outputs.job-id }}
           JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
@@ -145,6 +138,9 @@ jobs:
           WORKFLOW_RUN_ID: ${{github.run_id}}
           MONITOR_LOG_INTERVAL: ${{ inputs.monitor-log-interval }}
           MONITOR_DATA_COLLECT_INTERVAL: ${{ inputs.monitor-data-collect-interval }}
+        if: ${{ !inputs.disable-monitor }}
+        shell: bash
+        continue-on-error: true
         run: |
           python3 -m pip install psutil==5.9.8 dataclasses_json==0.6.7
           python3 -m tools.stats.monitor --log-interval "$MONITOR_LOG_INTERVAL" --data-collect-interval "$MONITOR_DATA_COLLECT_INTERVAL" > usage_log.txt 2>&1 &
@@ -182,12 +178,6 @@ jobs:
         run: |
           echo "timeout=$((JOB_TIMEOUT-30))" >> "${GITHUB_OUTPUT}"
 
-      - name: Preserve github env variables for use in docker
-        shell: bash
-        run: |
-          env | grep '^GITHUB' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-          env | grep '^CI' >> "/tmp/github_env_${GITHUB_RUN_ID}"
-
       - name: Test
         id: test
         env:
@@ -203,22 +193,20 @@ jobs:
           JOB_NAME: ${{ steps.get-job-id.outputs.job-name }}
           BRANCH: ${{ steps.parse-ref.outputs.branch }}
           SHA1: ${{ github.event.pull_request.head.sha || github.sha }}
-          BASE_SHA: ${{ github.event.pull_request.base.sha || github.sha }}
-          TEST_CONFIG: ${{ matrix.config }}
-          SHARD_NUMBER: ${{ matrix.shard }}
-          NUM_TEST_SHARDS: ${{ matrix.num_shards }}
-          REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
           CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
           VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
           TEST_SHOWLOCALS: ${{ steps.keep-going.outputs.ci-test-showlocals }}
           NO_TEST_TIMEOUT: ${{ steps.keep-going.outputs.ci-no-test-timeout }}
           NO_TD: ${{ steps.keep-going.outputs.ci-no-td }}
+          TEST_CONFIG: ${{ matrix.config }}
+          SHARD_NUMBER: ${{ matrix.shard }}
+          NUM_TEST_SHARDS: ${{ matrix.num_shards }}
+          REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
           DOCKER_IMAGE: ${{ inputs.docker-image }}
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
           PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
           TESTS_TO_INCLUDE: ${{ inputs.tests-to-include }}
           DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
         timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
         run: |
           set -x
@@ -248,7 +236,6 @@ jobs:
             -e GITHUB_RUN_ATTEMPT \
             -e JOB_ID \
             -e JOB_NAME \
-            -e BASE_SHA \
             -e BRANCH \
             -e SHA1 \
             -e AWS_DEFAULT_REGION \
@@ -266,12 +253,10 @@ jobs:
             -e PYTORCH_TEST_CUDA_MEM_LEAK_CHECK \
             -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
             -e TESTS_TO_INCLUDE \
-            -e HUGGING_FACE_HUB_TOKEN \
             -e DASHBOARD_TAG \
             --env-file="${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}" \
             --ulimit stack=10485760:83886080 \
             --ulimit core=0 \
-            --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
             --security-opt seccomp=unconfined \
             --cap-add=SYS_PTRACE \
             --shm-size="8g" \

.github/workflows/_win-build.yml

@@ -77,7 +77,6 @@ jobs:
         run: |
           git config --global core.longpaths true
           git config --global core.symlinks true
-          git config --global core.ignorecase false
 
           # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
           # the directory on Windows and prevent GHA from checking out as reported
@@ -151,7 +150,7 @@ jobs:
           BUILD_WHEEL: 1
           MAX_JOBS: 8
           CUDA_VERSION: ${{ inputs.cuda-version }}
-          PYTHON_VERSION: "3.10"
+          PYTHON_VERSION: "3.9"
           SCCACHE_BUCKET: "ossci-compiler-cache"
           SCCACHE_S3_KEY_PREFIX: ${{ github.workflow }}
           SCCACHE_REGION: us-east-1

.github/workflows/_win-test.yml

@@ -70,7 +70,6 @@ jobs:
         run: |
           git config --global core.longpaths true
           git config --global core.symlinks true
-          git config --global core.ignorecase false
 
           # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
           # the directory on Windows and prevent GHA from checking out as reported
@@ -184,7 +183,7 @@ jobs:
         env:
           USE_CUDA: ${{ inputs.cuda-version != 'cpu' && '1' || '0' }}
           INSTALL_WINDOWS_SDK: 1
-          PYTHON_VERSION: "3.10"
+          PYTHON_VERSION: 3.9
           CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
           VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
           TEST_SHOWLOCALS: ${{ steps.keep-going.outputs.ci-test-showlocals }}

.github/workflows/_xpu-test.yml

@@ -275,7 +275,7 @@ jobs:
       - name: Change permissions
         if: ${{ always() && steps.test.conclusion }}
         run: |
-          docker exec -t "${{ env.CONTAINER_NAME }}" sh -c "sudo chown -R 1000:1000 test"
+          docker exec -t "${{ env.CONTAINER_NAME }}" sh -c "sudo chown -R 1001:1001 test"
 
       - name: Print remaining test logs
         shell: bash

.github/workflows/build-manywheel-images.yml

@@ -47,11 +47,12 @@ jobs:
       matrix:
         include: [
           { name: "manylinux2_28-builder",          tag: "cuda13.0",         runner: "linux.9xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "cuda12.9",         runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cuda12.8",          runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cuda12.6",          runner: "linux.9xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda13.0",          runner: "linux.arm64.2xlarge.ephemeral" },
+          { name: "manylinuxaarch64-builder",       tag: "cuda12.9",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.8",          runner: "linux.arm64.2xlarge.ephemeral" },
-          { name: "manylinuxaarch64-builder",       tag: "cuda12.6",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm6.3",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cpu",               runner: "linux.9xlarge.ephemeral" },

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

@@ -145,7 +145,7 @@ jobs:
           fi
 
           docker exec -t "${container_name}" yum install -y zlib-devel zip
-          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}"  -m pip install -U setuptools==78.1.0 pybind11==3.0.1 auditwheel wheel
+          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}"  -m pip install -U setuptools==78.1.0 pybind11==2.13.1 auditwheel wheel
           set +e
           docker exec -t "${container_name}" command -v pip
           has_pip=$?

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

@@ -1,236 +0,0 @@
-name: Build vLLM wheels
-
-on:
-  push:
-    branches:
-      - main
-    paths:
-      - .github/workflows/build-vllm-wheel.yml
-      - .github/ci_commit_pins/vllm.txt
-  workflow_dispatch:
-  pull_request:
-    paths:
-      - .github/workflows/build-vllm-wheel.yml
-      - .github/ci_commit_pins/vllm.txt
-  schedule:
-    # every morning at 01:30PM UTC, 9:30AM EST, 6:30AM PST
-    - cron: 30 13 * * *
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
-  cancel-in-progress: true
-
-jobs:
-  build-wheel:
-    if: github.repository_owner == 'pytorch'
-    strategy:
-      fail-fast: false
-      matrix:
-        python-version: [ '3.12' ]
-        # TODO (huydhn): Add cu130 after https://github.com/vllm-project/vllm/issues/24464 is resolved
-        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129' ]
-        include:
-          - platform: manylinux_2_28_x86_64
-            device: cu128
-            manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.8'
-            runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_x86_64
-            device: cu129
-            manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
-            runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_aarch64
-            device: cu128
-            manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.8'
-            runner: linux.arm64.r7g.12xlarge.memory
-          - platform: manylinux_2_28_aarch64
-            device: cu129
-            manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.9'
-            runner: linux.arm64.r7g.12xlarge.memory
-    name: "Build ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
-    runs-on: ${{ matrix.runner }}
-    timeout-minutes: 480
-    env:
-      PY_VERS: ${{ matrix.python-version }}
-      MANYLINUX_IMAGE: ${{ matrix.manylinux-image }}
-      PLATFORM: ${{ matrix.platform }}
-      BUILD_DEVICE: ${{ matrix.device }}
-    steps:
-      - name: Setup SSH (Click me for login details)
-        uses: pytorch/test-infra/.github/actions/setup-ssh@main
-        with:
-          github-secret: ${{ secrets.GITHUB_TOKEN }}
-
-      - name: Checkout PyTorch
-        uses: pytorch/pytorch/.github/actions/checkout-pytorch@main
-        with:
-          submodules: false
-
-      - name: Setup Linux
-        uses: ./.github/actions/setup-linux
-
-      - name: Get latest PyTorch nightly
-        shell: bash
-        run: |
-          set -eux
-
-          # Determine python executable for given version (copied from build-triton-wheel)
-          case $PY_VERS in
-          3.10)
-            PYTHON_EXECUTABLE=/opt/python/cp310-cp310/bin/python
-            ;;
-          3.11)
-            PYTHON_EXECUTABLE=/opt/python/cp311-cp311/bin/python
-            ;;
-          3.12)
-            PYTHON_EXECUTABLE=/opt/python/cp312-cp312/bin/python
-            ;;
-          3.13)
-            PYTHON_EXECUTABLE=/opt/python/cp313-cp313/bin/python
-            ;;
-          3.13t)
-            PYTHON_EXECUTABLE=/opt/python/cp313-cp313t/bin/python
-            ;;
-          3.14)
-            PYTHON_EXECUTABLE=/opt/python/cp314-cp314/bin/python
-            ;;
-          3.14t)
-            PYTHON_EXECUTABLE=/opt/python/cp314-cp314t/bin/python
-            ;;
-          *)
-            echo "Unsupported python version ${PY_VERS}"
-            exit 1
-            ;;
-          esac
-
-          # Keep PyTorch nightly wheel here so that we can install it later during
-          # vLLM build process
-          mkdir -p "${RUNNER_TEMP}/artifacts/"
-
-          container_name=$(docker run \
-            --tty \
-            --detach \
-            -e PLATFORM \
-            -e PYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" \
-            -v "${GITHUB_WORKSPACE}:/pytorch" \
-            -v "${RUNNER_TEMP}/artifacts:/artifacts" \
-            -w /artifacts/ \
-            "${MANYLINUX_IMAGE}"
-          )
-
-          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}" -mpip install \
-            --pre torch torchvision torchaudio \
-            --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
-
-          # I wonder if there is a command to both download and install the wheels
-          # in one go
-          docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}" -mpip download \
-            --pre torch torchvision torchaudio \
-            --index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
-
-          # Save this for later
-          echo "container_name=${container_name}" >> "$GITHUB_ENV"
-
-      - name: Build vLLM wheel
-        uses: ./.github/actions/build-external-packages
-        with:
-          build-targets: vllm
-          docker-image: ${{ env.MANYLINUX_IMAGE }}
-          cuda-arch-list: '8.0;8.9;9.0;10.0;12.0'
-          torch-wheel-dir: ${{ runner.temp }}/artifacts
-          output-dir: ${{ runner.temp }}/artifacts/externals
-
-      - name: Prepare vLLM wheel
-        shell: bash
-        run: |
-          set -eux
-
-          # Get these wheels ready, the vllm renaming logic is copied from its .buildkite/scripts/upload-wheels.sh
-          docker exec -t "${container_name}" bash -c /pytorch/.github/scripts/prepare_vllm_wheels.sh
-          docker exec -t "${container_name}" chown -R 1000:1000 /artifacts
-
-      - uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
-        with:
-          name: vllm-wheel-${{ matrix.device }}-${{ matrix.platform }}-${{ matrix.python-version }}
-          if-no-files-found: error
-          path: ${{ runner.temp }}/artifacts/externals/vllm/wheels/*.whl
-
-      - name: Teardown Linux
-        uses: pytorch/test-infra/.github/actions/teardown-linux@main
-        if: always()
-
-  # Copied from build-triton-wheel workflow (mostly)
-  upload-wheel:
-    name: "Upload ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
-    needs:
-      - build-wheel
-    runs-on: ubuntu-latest
-    strategy:
-      fail-fast: false
-      matrix:
-        platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129' ]
-    env:
-      PLATFORM: ${{ matrix.platform }}
-      BUILD_DEVICE: ${{ matrix.device }}
-    permissions:
-      id-token: write
-      contents: read
-    container:
-      image: continuumio/miniconda3:4.12.0
-    environment: ${{ ((github.event_name == 'push' && github.event.ref == 'refs/heads/main') || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch') && 'nightly-wheel-upload' || '' }}
-    steps:
-      - uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
-
-      - name: Configure AWS credentials(PyTorch account) for main
-        if: ${{ (github.event_name == 'push' && github.event.ref == 'refs/heads/main') || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
-        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-        with:
-          role-to-assume: arn:aws:iam::749337293305:role/gha_workflow_nightly_build_wheels
-          aws-region: us-east-1
-
-      - name: Configure AWS credentials(PyTorch account) for RC builds
-        if: ${{ github.event_name == 'push' &&  (startsWith(github.event.ref, 'refs/tags/') && !startsWith(github.event.ref, 'refs/tags/ciflow/')) }}
-        uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
-        with:
-          role-to-assume: arn:aws:iam::749337293305:role/gha_workflow_test_build_wheels
-          aws-region: us-east-1
-
-      - name: Download Build Artifacts
-        uses: actions/download-artifact@65a9edc5881444af0b9093a5e628f2fe47ea3b2e # v4.1.7
-        with:
-          # Download all available artifacts
-          path: ${{ runner.temp }}/artifacts-all
-
-      - name: Select Wheel Artifacts
-        shell: bash
-        run: |
-          set -eux
-          mkdir -p "${RUNNER_TEMP}/artifacts/"
-          mv "${RUNNER_TEMP}"/artifacts-all/vllm-wheel-"${BUILD_DEVICE}"-"${PLATFORM}"-*/* "${RUNNER_TEMP}/artifacts/"
-
-      - name: Set DRY_RUN
-        if: ${{ (github.event_name == 'push' && (github.event.ref == 'refs/heads/main' || startsWith(github.event.ref, 'refs/tags/v'))) || github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' }}
-        shell: bash
-        run: |
-          echo "DRY_RUN=disabled" >> "$GITHUB_ENV"
-
-      - name: Set UPLOAD_CHANNEL
-        if: ${{ github.event_name == 'push' && startsWith(github.event.ref, 'refs/tags/v') }}
-        shell: bash
-        run: |
-          set -ex
-
-          if [[ "${GITHUB_REF_NAME}" = *-rc[0-9]* ]]; then
-            echo "UPLOAD_CHANNEL=test" >> "$GITHUB_ENV"
-          fi
-
-      - name: Upload binaries
-        env:
-          PACKAGE_TYPE: wheel
-          UPLOAD_SUBFOLDER: ${{ env.BUILD_DEVICE }}
-          PKG_DIR: ${{ runner.temp }}/artifacts
-        shell: bash
-        run: |
-          set -ex
-          bash .circleci/scripts/binary_upload.sh

.github/workflows/docker-builds.yml

@@ -50,28 +50,30 @@ jobs:
         runner: [linux.12xlarge]
         docker-image-name: [
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11,
-          pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc11-vllm,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9-inductor-benchmarks,
+          pytorch-linux-jammy-cuda12.8-cudnn9-py3.12-gcc9-inductor-benchmarks,
+          pytorch-linux-jammy-cuda12.8-cudnn9-py3.13-gcc9-inductor-benchmarks,
           pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9,
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
-          pytorch-linux-jammy-py3.10-clang12,
+          pytorch-linux-jammy-py3.9-clang12,
           pytorch-linux-jammy-py3.13-clang12,
           pytorch-linux-jammy-rocm-n-py3,
           pytorch-linux-noble-rocm-n-py3,
           pytorch-linux-noble-rocm-alpha-py3,
           pytorch-linux-jammy-rocm-n-py3-benchmarks,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-clang12,
-          pytorch-linux-jammy-py3.10-gcc11,
-          pytorch-linux-jammy-py3-gcc11-inductor-benchmarks,
+          pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-clang12,
+          pytorch-linux-jammy-py3.9-gcc11,
+          pytorch-linux-jammy-py3.9-gcc11-inductor-benchmarks,
           pytorch-linux-jammy-py3.12-halide,
-          pytorch-linux-jammy-xpu-n-1-py3,
-          pytorch-linux-jammy-xpu-n-py3,
+          pytorch-linux-jammy-xpu-2025.0-py3,
+          pytorch-linux-jammy-xpu-2025.1-py3,
           pytorch-linux-jammy-py3-clang18-asan,
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,
-          pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter,
-          pytorch-linux-jammy-py3-clang12-executorch,
+          pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter,
+          # Executorch pin needs update
+          # pytorch-linux-jammy-py3-clang12-executorch,
           pytorch-linux-jammy-py3.12-triton-cpu,
           pytorch-linux-noble-riscv64-py3.12-gcc14
         ]
@@ -122,7 +124,7 @@ jobs:
           GHCR_PAT: ${{ secrets.GHCR_PAT }}
         with:
           shell: bash
-          timeout_minutes: 60
+          timeout_minutes: 30
           max_attempts: 5
           retry_wait_seconds: 90
           command: |

.github/workflows/generated-linux-binary-libtorch-nightly.yml

@@ -248,7 +248,7 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
-  libtorch-cuda13_0-shared-with-deps-release-build:
+  libtorch-cuda12_9-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -257,22 +257,22 @@ jobs:
       PACKAGE_TYPE: libtorch
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: libtorch-cuda13_0-shared-with-deps-release
+      build_name: libtorch-cuda12_9-shared-with-deps-release
       build_environment: linux-binary-libtorch
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-cuda13_0-shared-with-deps-release-test:  # Testing
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
     if: ${{ github.repository_owner == 'pytorch' }}
     needs:
-      - libtorch-cuda13_0-shared-with-deps-release-build
+      - libtorch-cuda12_9-shared-with-deps-release-build
       - get-label-type
     uses: ./.github/workflows/_binary-test-linux.yml
     with:
@@ -280,38 +280,38 @@ jobs:
       PACKAGE_TYPE: libtorch
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-cuda13_0-shared-with-deps-release
+      build_name: libtorch-cuda12_9-shared-with-deps-release
       build_environment: linux-binary-libtorch
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-cuda13_0-shared-with-deps-release-upload:  # Uploading
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
     if: ${{ github.repository_owner == 'pytorch' }}
     permissions:
       id-token: write
       contents: read
-    needs: libtorch-cuda13_0-shared-with-deps-release-test
+    needs: libtorch-cuda12_9-shared-with-deps-release-test
     with:
       PYTORCH_ROOT: /pytorch
       PACKAGE_TYPE: libtorch
       # TODO: This is a legacy variable that we eventually want to get rid of in
       #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: cu130
-      GPU_ARCH_VERSION: "13.0"
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
       GPU_ARCH_TYPE: cuda
       DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: cuda13.0
+      DOCKER_IMAGE_TAG_PREFIX: cuda12.9
       LIBTORCH_CONFIG: release
       LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-cuda13_0-shared-with-deps-release
+      build_name: libtorch-cuda12_9-shared-with-deps-release
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml

.github/workflows/generated-linux-binary-manywheel-main.yml

@@ -60,7 +60,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_8
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_8-test:  # Testing

.github/workflows/generated-linux-binary-manywheel-rocm-main.yml

@@ -44,7 +44,7 @@ jobs:
       issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
       curr_branch: ${{ github.head_ref || github.ref_name }}
       curr_ref_type: ${{ github.ref_type }}
-  manywheel-py3_10-rocm6_4-build:
+  manywheel-py3_9-rocm6_4-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
     needs: get-label-type
@@ -58,16 +58,16 @@ jobs:
       GPU_ARCH_TYPE: rocm
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build_name: manywheel-py3_10-rocm6_4
+      build_name: manywheel-py3_9-rocm6_4
       build_environment: linux-binary-manywheel-rocm
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
-  manywheel-py3_10-rocm6_4-test:  # Testing
+  manywheel-py3_9-rocm6_4-test:  # Testing
     if: ${{ github.repository_owner == 'pytorch' }}
     needs:
-      - manywheel-py3_10-rocm6_4-build
+      - manywheel-py3_9-rocm6_4-build
       - get-label-type
     runs-on: linux.rocm.gpu.mi250
     timeout-minutes: 240
@@ -82,14 +82,14 @@ jobs:
       SKIP_ALL_TESTS: 1
       DOCKER_IMAGE: manylinux2_28-builder
       DOCKER_IMAGE_TAG_PREFIX: rocm6.4
-      DESIRED_PYTHON: "3.10"
+      DESIRED_PYTHON: "3.9"
     steps:
       - name: Setup ROCm
         uses: ./.github/actions/setup-rocm
       - uses: actions/download-artifact@v4.1.7
         name: Download Build Artifacts
         with:
-          name: manywheel-py3_10-rocm6_4
+          name: manywheel-py3_9-rocm6_4
           path: "${{ runner.temp }}/artifacts/"
       - name: Checkout PyTorch
         uses: actions/checkout@v4