Merge remote-tracking branch 'upstream/main' into rocm7.1_internal_testing_IFU_2025-09-24

# Conflicts:
#	.ci/docker/ci_commit_pins/triton.txt
#	.ci/docker/common/install_rocm.sh
#	.ci/docker/requirements-ci.txt
#	CMakeLists.txt
#	aten/src/ATen/native/Normalization.cpp
#	aten/src/ATen/native/miopen/BatchNorm_miopen.cpp
#	requirements-build.txt
#	test/nn/test_convolution.py
#	test/test_binary_ufuncs.py
#	test/test_nn.py
#	torch/_inductor/runtime/triton_heuristics.py
#	torch/testing/_internal/common_utils.py

Author: github-actions[bot]

.ci/aarch64_linux/aarch64_ci_build.sh

@@ -5,9 +5,9 @@ GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
 
 # Set CUDA architecture lists to match x86 build_cuda.sh
 if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="5.0;6.0;7.0;8.0;9.0"
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0"
 elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
-    export TORCH_CUDA_ARCH_LIST="7.0;8.0;9.0;10.0;12.0"
+    export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
 elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
     export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
 fi
@@ -31,8 +31,7 @@ pip install -r /pytorch/requirements.txt
 pip install auditwheel==6.2.0 wheel
 if [ "$DESIRED_CUDA" = "cpu" ]; then
     echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
-    #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
+    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn
 else
     echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
     export USE_SYSTEM_NCCL=1
@@ -42,13 +41,9 @@ else
         echo "Bundling CUDA libraries with wheel for aarch64."
     else
         echo "Using nvidia libs from pypi for aarch64."
-        # Fix platform constraints in PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64
-        # Replace 'platform_machine == "x86_64"' with 'platform_machine == "aarch64"'
-        export PYTORCH_EXTRA_INSTALL_REQUIREMENTS="${PYTORCH_EXTRA_INSTALL_REQUIREMENTS//platform_machine == \'x86_64\'/platform_machine == \'aarch64\'}"
         echo "Updated PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64: $PYTORCH_EXTRA_INSTALL_REQUIREMENTS"
         export USE_NVIDIA_PYPI_LIBS=1
     fi
 
-    #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
+    python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
 fi

.ci/aarch64_linux/aarch64_wheel_ci_build.py

@@ -138,6 +138,8 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
     folder = os.path.dirname(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}")
 
     # Check if we should use PyPI NVIDIA libraries or bundle system libraries
     use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
@@ -211,7 +213,8 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
         ]
 
         # CUDA version-specific libraries
-        if "130" in desired_cuda:
+        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",
@@ -221,7 +224,7 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
                 "/usr/local/cuda/lib64/libcusolver.so.12",
                 "/usr/local/cuda/lib64/libnvJitLink.so.13",
                 "/usr/local/cuda/lib64/libnvrtc.so.13",
-                "/usr/local/cuda/lib64/libnvrtc-builtins.so.13.0",
+                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")
@@ -237,6 +240,8 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
                 "/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
@@ -275,14 +280,7 @@ def complete_wheel(folder: str) -> str:
             f"/{folder}/dist/{repaired_wheel_name}",
         )
     else:
-        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}",
-        )
+        repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
 
     print(f"Copying {repaired_wheel_name} to artifacts")
     shutil.copy2(
@@ -319,7 +317,7 @@ def parse_arguments():
     ).decode()
 
     print("Building PyTorch wheel")
-    build_vars = "CMAKE_SHARED_LINKER_FLAGS=-Wl,-z,max-page-size=0x10000 "
+    build_vars = ""
     # MAX_JOB=5 is not required for CPU backend (see commit 465d98b)
     if enable_cuda:
         build_vars += "MAX_JOBS=5 "

.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, socket.timeout):  # noqa: PERF203
+        except (ConnectionRefusedError, TimeoutError):  # noqa: PERF203
             if i == attempt_cnt - 1:
                 raise
             time.sleep(timeout)
@@ -1004,7 +1004,7 @@ def parse_arguments():
         install_condaforge_python(host, args.python_version)
         sys.exit(0)
 
-    python_version = args.python_version if args.python_version is not None else "3.9"
+    python_version = args.python_version if args.python_version is not None else "3.10"
 
     if args.use_torch_from_pypi:
         configure_system(host, compiler=args.compiler, python_version=python_version)

.ci/docker/build.sh

@@ -214,8 +214,7 @@ case "$tag" in
     TRITON=yes
     ;;
   pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
-    # TODO (huydhn): Upgrade this to Python >= 3.10
-    ANACONDA_PYTHON_VERSION=3.9
+    ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
     KATEX=yes
@@ -263,13 +262,10 @@ case "$tag" in
     TRITON_CPU=yes
     ;;
   pytorch-linux-jammy-linter)
-    # 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
+    PYTHON_VERSION=3.10
     ;;
-  pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter)
-    PYTHON_VERSION=3.9
+  pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter)
+    PYTHON_VERSION=3.10
     CUDA_VERSION=12.8.1
     ;;
   pytorch-linux-jammy-aarch64-py3.10-gcc11)

.ci/docker/centos-rocm/Dockerfile

@@ -59,9 +59,13 @@ 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
+RUN rm install_rocm.sh common_utils.sh
+RUN rm -r ci_commit_pins
 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 @@
-56392aa978594cc155fa8af48cd949f5b5f1823a
+e0dda9059d082537cee36be6c5e4fe3b18c880c0

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

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

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

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

.ci/docker/ci_commit_pins/triton.txt

@@ -1 +1,5 @@
+<<<<<<< HEAD
 6193b30becb1ac7be704cf87b8cb9bf13e7f9689
+=======
+bbb06c0334a6772b92d24bde54956e675c8c6604
+>>>>>>> upstream/main

.ci/docker/common/install_executorch.sh

@@ -42,22 +42,27 @@ 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"
+  export CMAKE_ARGS="-DEXECUTORCH_BUILD_PYBIND=ON -DEXECUTORCH_BUILD_XNNPACK=ON -DEXECUTORCH_BUILD_KERNELS_QUANTIZED=ON -DEXECUTORCH_BUILD_TESTS=ON"
 
   as_jenkins .ci/scripts/setup-linux.sh --build-tool cmake || true
-  popd
 }
 
-clone_executorch
-install_buck2
-install_conda_dependencies
-install_pip_dependencies
-setup_executorch
+if [ $# -eq 0 ]; then
+  clone_executorch
+  install_buck2
+  install_conda_dependencies
+  install_pip_dependencies
+  pushd executorch
+  setup_executorch
+  popd
+else
+  "$@"
+fi