Update base for Update on "[prototype] aot_export_partitioned_graphs_v2"

Improvements
* No need of dynamo gm state dict monkeypatching
* Handle constant tensors in the fx graphs
* No need to worry about the signature of forward - params/buffers are
  passed on internally

cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta msaroufim dcci voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx chenyang78 kadeng chauhang amjames Lucaskabela

[ghstack-poisoned]

.ci/docker/build.sh

@@ -113,7 +113,6 @@ case "$tag" in
     UCX_COMMIT=${_UCX_COMMIT}
     UCC_COMMIT=${_UCC_COMMIT}
     TRITON=yes
-    INSTALL_MINGW=yes
     ;;
   pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
     CUDA_VERSION=13.0.0
@@ -362,7 +361,6 @@ docker build \
        --build-arg "OPENBLAS=${OPENBLAS:-}" \
        --build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
        --build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
-       --build-arg "INSTALL_MINGW=${INSTALL_MINGW:-}" \
        -f $(dirname ${DOCKERFILE})/Dockerfile \
        -t "$tmp_tag" \
        "$@" \

.ci/docker/common/install_cpython.sh

@@ -83,6 +83,10 @@ 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
+    if [ "$py_suffix" == "3.14.0" ]; then
+        py_suffix="3.14.0rc2"
+    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

@@ -150,7 +150,7 @@ function install_130 {
   CUDNN_VERSION=9.13.0.50
   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 13.0.2 cuda_13.0.2_580.95.05_linux
+  install_cuda 13.0.0 cuda_13.0.0_580.65.06_linux
 
   # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
   install_cudnn 13 $CUDNN_VERSION

.ci/docker/common/install_mingw.sh

@@ -1,10 +0,0 @@
-#!/bin/bash
-
-set -ex
-
-# Install MinGW-w64 for Windows cross-compilation
-apt-get update
-apt-get install -y g++-mingw-w64-x86-64-posix
-
-echo "MinGW-w64 installed successfully"
-x86_64-w64-mingw32-g++ --version

.ci/docker/common/install_onnx.sh

@@ -19,8 +19,8 @@ pip_install \
   transformers==4.36.2
 
 pip_install coloredlogs packaging
-pip_install onnxruntime==1.23.1
-pip_install onnxscript==0.5.4
+pip_install onnxruntime==1.23.0
+pip_install onnxscript==0.5.3
 
 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers
 # package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/

.ci/docker/libtorch/build.sh

@@ -39,13 +39,9 @@ case ${DOCKER_TAG_PREFIX} in
         DOCKER_GPU_BUILD_ARG=""
         ;;
     rocm*)
-        # we want the patch version of 7.0 instead
-        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
-        fi
         # we want the patch version of 6.4 instead
         if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete

.ci/docker/manywheel/build.sh

@@ -75,13 +75,9 @@ case ${image} in
         DOCKERFILE_SUFFIX="_cuda_aarch64"
         ;;
     manylinux2_28-builder:rocm*)
-        # we want the patch version of 7.0 instead
-        if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
-        fi
         # we want the patch version of 6.4 instead
         if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
-            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
         fi
         TARGET=rocm_final
         MANY_LINUX_VERSION="2_28"

.ci/docker/requirements-ci.txt

@@ -334,12 +334,12 @@ sympy==1.13.3
 #Pinned versions:
 #test that import:
 
-onnx==1.19.1
+onnx==1.18.0
 #Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
 
-onnxscript==0.5.4
+onnxscript==0.5.3
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:

.ci/docker/ubuntu/Dockerfile

@@ -103,11 +103,6 @@ COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
 RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
 
-ARG INSTALL_MINGW
-COPY ./common/install_mingw.sh install_mingw.sh
-RUN if [ -n "${INSTALL_MINGW}" ]; then bash ./install_mingw.sh; fi
-RUN rm install_mingw.sh
-
 ARG TRITON
 ARG TRITON_CPU
 

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

@@ -57,8 +57,8 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
         logger.info("Successfully cloned %s", target)
         return r, commit
 
-    except GitCommandError:
-        logger.exception("Git operation failed")
+    except GitCommandError as e:
+        logger.error("Git operation failed: %s", e)
         raise
 
 

.ci/lumen_cli/pyproject.toml

@@ -6,7 +6,7 @@ dependencies = [
     "GitPython==3.1.45",
     "docker==7.1.0",
     "pytest==7.3.2",
-    "uv==0.9.5"
+    "uv==0.8.6"
 ]
 
 [tool.setuptools]

.ci/manywheel/build_cuda.sh

@@ -187,22 +187,19 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
             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
         fi
     else
         echo "Using nvidia libs from pypi."

.ci/pytorch/test.sh

@@ -485,22 +485,6 @@ test_inductor_aoti() {
   /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
 }
 
-test_inductor_aoti_cross_compile_for_windows() {
-
-  TEST_REPORTS_DIR=$(pwd)/test/test-reports
-  mkdir -p "$TEST_REPORTS_DIR"
-
-  # Set WINDOWS_CUDA_HOME environment variable
-  WINDOWS_CUDA_HOME="$(pwd)/win-torch-wheel-extracted"
-  export WINDOWS_CUDA_HOME
-
-  echo "WINDOWS_CUDA_HOME is set to: $WINDOWS_CUDA_HOME"
-  echo "Contents:"
-  ls -lah "$(pwd)/win-torch-wheel-extracted/lib/x64/" || true
-
-  python test/inductor/test_aoti_cross_compile_windows.py -k compile --package-dir "$TEST_REPORTS_DIR" --win-torch-lib-dir "$(pwd)/win-torch-wheel-extracted/torch/lib"
-}
-
 test_inductor_cpp_wrapper_shard() {
   if [[ -z "$NUM_TEST_SHARDS" ]]; then
     echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
@@ -916,7 +900,7 @@ test_inductor_set_cpu_affinity(){
   export LD_PRELOAD="$JEMALLOC_LIB":"$LD_PRELOAD"
   export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:-1,muzzy_decay_ms:-1"
 
-  if [[ "$(uname -m)" != "aarch64" ]]; then
+  if [[ "${TEST_CONFIG}" != *aarch64* ]]; then
     # Use Intel OpenMP for x86
     IOMP_LIB="$(dirname "$(which python)")/../lib/libiomp5.so"
     export LD_PRELOAD="$IOMP_LIB":"$LD_PRELOAD"
@@ -930,7 +914,7 @@ test_inductor_set_cpu_affinity(){
   cores=$((cpus / thread_per_core))
 
   # Set number of cores to 16 on aarch64 for performance runs
-  if [[ "$(uname -m)" == "aarch64" && $cores -gt 16 ]]; then
+  if [[ "${TEST_CONFIG}" == *aarch64* && $cores -gt 16 ]]; then
     cores=16
   fi
   export OMP_NUM_THREADS=$cores
@@ -1631,7 +1615,6 @@ test_operator_benchmark() {
   TEST_REPORTS_DIR=$(pwd)/test/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
   TEST_DIR=$(pwd)
-  ARCH=$(uname -m)
 
   test_inductor_set_cpu_affinity
 
@@ -1646,7 +1629,7 @@ test_operator_benchmark() {
   pip_install pandas
   python check_perf_csv.py \
       --actual "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv" \
-      --expected "${ARCH}_expected_ci_operator_benchmark_eager_float32_cpu.csv"
+      --expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
 }
 
 test_operator_microbenchmark() {
@@ -1683,7 +1666,7 @@ if [[ "${TEST_CONFIG}" == *numpy_2* ]]; then
     python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
   fi
   python test/run_test.py --include dynamo/test_functions.py dynamo/test_unspec.py test_binary_ufuncs.py test_fake_tensor.py test_linalg.py test_numpy_interop.py test_tensor_creation_ops.py test_torch.py torch_np/test_basic.py
-elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" == 'default' ]]; then
+elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" != *perf_cpu_aarch64* ]]; then
   test_linux_aarch64
 elif [[ "${TEST_CONFIG}" == *backward* ]]; then
   test_forward_backward_compatibility
@@ -1734,8 +1717,6 @@ elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
   test_inductor_triton_cpu
 elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
   test_inductor_micro_benchmark
-elif [[ "${TEST_CONFIG}" == *aoti_cross_compile_for_windows* ]]; then
-  test_inductor_aoti_cross_compile_for_windows
 elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
   install_torchvision
   id=$((SHARD_NUMBER-1))

.circleci/scripts/binary_populate_env.sh

@@ -163,13 +163,8 @@ if [[ "$(uname)" != Darwin ]]; then
   MEMORY_LIMIT_MAX_JOBS=12
   NUM_CPUS=$(( $(nproc) - 2 ))
 
-  if [[ "$(uname)" == Linux ]]; then
-    # Defaults here for **binary** linux builds so they can be changed in one place
-    export MAX_JOBS=${MAX_JOBS:-$(( ${NUM_CPUS} > ${MEMORY_LIMIT_MAX_JOBS} ? ${MEMORY_LIMIT_MAX_JOBS} : ${NUM_CPUS} ))}
-  else
-    # For other builds
-    export MAX_JOBS=${NUM_CPUS}
-  fi
+  # Defaults here for **binary** linux builds so they can be changed in one place
+  export MAX_JOBS=${MAX_JOBS:-$(( ${NUM_CPUS} > ${MEMORY_LIMIT_MAX_JOBS} ? ${MEMORY_LIMIT_MAX_JOBS} : ${NUM_CPUS} ))}
 
   cat >>"$envfile" <<EOL
   export MAX_JOBS="${MAX_JOBS}"

.claude/skills/docstring/SKILL.md

@@ -1,359 +0,0 @@
----
-name: docstring
-description: Write docstrings for PyTorch functions and methods following PyTorch conventions. Use when writing or updating docstrings in PyTorch code.
----
-
-# PyTorch Docstring Writing Guide
-
-This skill describes how to write docstrings for functions and methods in the PyTorch project, following the conventions in `torch/_tensor_docs.py` and `torch/nn/functional.py`.
-
-## General Principles
-
-- Use **raw strings** (`r"""..."""`) for all docstrings to avoid issues with LaTeX/math backslashes
-- Follow **Sphinx/reStructuredText** (reST) format for documentation
-- Be **concise but complete** - include all essential information
-- Always include **examples** when possible
-- Use **cross-references** to related functions/classes
-
-## Docstring Structure
-
-### 1. Function Signature (First Line)
-
-Start with the function signature showing all parameters:
-
-```python
-r"""function_name(param1, param2, *, kwarg1=default1, kwarg2=default2) -> ReturnType
-```
-
-**Notes:**
-- Include the function name
-- Show positional and keyword-only arguments (use `*` separator)
-- Include default values
-- Show return type annotation
-- This line should NOT end with a period
-
-### 2. Brief Description
-
-Provide a one-line description of what the function does:
-
-```python
-r"""conv2d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1) -> Tensor
-
-Applies a 2D convolution over an input image composed of several input
-planes.
-```
-
-### 3. Mathematical Formulas (if applicable)
-
-Use Sphinx math directives for mathematical expressions:
-
-```python
-.. math::
-    \text{Softmax}(x_{i}) = \frac{\exp(x_i)}{\sum_j \exp(x_j)}
-```
-
-Or inline math: `:math:\`x^2\``
-
-### 4. Cross-References
-
-Link to related classes and functions using Sphinx roles:
-
-- `:class:\`~torch.nn.ModuleName\`` - Link to a class
-- `:func:\`torch.function_name\`` - Link to a function
-- `:meth:\`~Tensor.method_name\`` - Link to a method
-- `:attr:\`attribute_name\`` - Reference an attribute
-- The `~` prefix shows only the last component (e.g., `Conv2d` instead of `torch.nn.Conv2d`)
-
-**Example:**
-```python
-See :class:`~torch.nn.Conv2d` for details and output shape.
-```
-
-### 5. Notes and Warnings
-
-Use admonitions for important information:
-
-```python
-.. note::
-    This function doesn't work directly with NLLLoss,
-    which expects the Log to be computed between the Softmax and itself.
-    Use log_softmax instead (it's faster and has better numerical properties).
-
-.. warning::
-    :func:`new_tensor` always copies :attr:`data`. If you have a Tensor
-    ``data`` and want to avoid a copy, use :func:`torch.Tensor.requires_grad_`
-    or :func:`torch.Tensor.detach`.
-```
-
-### 6. Args Section
-
-Document all parameters with type annotations and descriptions:
-
-```python
-Args:
-    input (Tensor): input tensor of shape :math:`(\text{minibatch} , \text{in\_channels} , iH , iW)`
-    weight (Tensor): filters of shape :math:`(\text{out\_channels} , kH , kW)`
-    bias (Tensor, optional): optional bias tensor of shape :math:`(\text{out\_channels})`. Default: ``None``
-    stride (int or tuple): the stride of the convolving kernel. Can be a single number or a
-      tuple `(sH, sW)`. Default: 1
-```
-
-**Formatting rules:**
-- Parameter name in **lowercase**
-- Type in parentheses: `(Type)`, `(Type, optional)` for optional parameters
-- Description follows the type
-- For optional parameters, include "Default: ``value``" at the end
-- Use double backticks for inline code: ``` ``None`` ```
-- Indent continuation lines by 2 spaces
-
-### 7. Keyword Args Section (if applicable)
-
-Sometimes keyword arguments are documented separately:
-
-```python
-Keyword args:
-    dtype (:class:`torch.dtype`, optional): the desired type of returned tensor.
-        Default: if None, same :class:`torch.dtype` as this tensor.
-    device (:class:`torch.device`, optional): the desired device of returned tensor.
-        Default: if None, same :class:`torch.device` as this tensor.
-    requires_grad (bool, optional): If autograd should record operations on the
-        returned tensor. Default: ``False``.
-```
-
-### 8. Returns Section (if needed)
-
-Document the return value:
-
-```python
-Returns:
-    Tensor: Sampled tensor of same shape as `logits` from the Gumbel-Softmax distribution.
-        If ``hard=True``, the returned samples will be one-hot, otherwise they will
-        be probability distributions that sum to 1 across `dim`.
-```
-
-Or simply include it in the function signature line if obvious from context.
-
-### 9. Examples Section
-
-Always include examples when possible:
-
-```python
-Examples::
-
-    >>> inputs = torch.randn(33, 16, 30)
-    >>> filters = torch.randn(20, 16, 5)
-    >>> F.conv1d(inputs, filters)
-
-    >>> # With square kernels and equal stride
-    >>> filters = torch.randn(8, 4, 3, 3)
-    >>> inputs = torch.randn(1, 4, 5, 5)
-    >>> F.conv2d(inputs, filters, padding=1)
-```
-
-**Formatting rules:**
-- Use `Examples::` with double colon
-- Use `>>>` prompt for Python code
-- Include comments with `#` when helpful
-- Show actual output when it helps understanding (indent without `>>>`)
-
-### 10. External References
-
-Link to papers or external documentation:
-
-```python
-.. _Link Name:
-    https://arxiv.org/abs/1611.00712
-```
-
-Reference them in text: ```See `Link Name`_```
-
-## Method Types
-
-### Native Python Functions
-
-For regular Python functions, use a standard docstring:
-
-```python
-def relu(input: Tensor, inplace: bool = False) -> Tensor:
-    r"""relu(input, inplace=False) -> Tensor
-
-    Applies the rectified linear unit function element-wise. See
-    :class:`~torch.nn.ReLU` for more details.
-    """
-    # implementation
-```
-
-### C-Bound Functions (using add_docstr)
-
-For C-bound functions, use `_add_docstr`:
-
-```python
-conv1d = _add_docstr(
-    torch.conv1d,
-    r"""
-conv1d(input, weight, bias=None, stride=1, padding=0, dilation=1, groups=1) -> Tensor
-
-Applies a 1D convolution over an input signal composed of several input
-planes.
-
-See :class:`~torch.nn.Conv1d` for details and output shape.
-
-Args:
-    input: input tensor of shape :math:`(\text{minibatch} , \text{in\_channels} , iW)`
-    weight: filters of shape :math:`(\text{out\_channels} , kW)`
-    ...
-""",
-)
-```
-
-### In-Place Variants
-
-For in-place operations (ending with `_`), reference the original:
-
-```python
-add_docstr_all(
-    "abs_",
-    r"""
-abs_() -> Tensor
-
-In-place version of :meth:`~Tensor.abs`
-""",
-)
-```
-
-### Alias Functions
-
-For aliases, simply reference the original:
-
-```python
-add_docstr_all(
-    "absolute",
-    r"""
-absolute() -> Tensor
-
-Alias for :func:`abs`
-""",
-)
-```
-
-## Common Patterns
-
-### Shape Documentation
-
-Use LaTeX math notation for tensor shapes:
-
-```python
-:math:`(\text{minibatch} , \text{in\_channels} , iH , iW)`
-```
-
-### Reusable Argument Definitions
-
-For commonly used arguments, define them once and reuse:
-
-```python
-common_args = parse_kwargs(
-    """
-    dtype (:class:`torch.dtype`, optional): the desired type of returned tensor.
-        Default: if None, same as this tensor.
-"""
-)
-
-# Then use with .format():
-r"""
-...
-
-Keyword args:
-    {dtype}
-    {device}
-""".format(**common_args)
-```
-
-### Template Insertion
-
-Insert reproducibility notes or other common text:
-
-```python
-r"""
-{tf32_note}
-
-{cudnn_reproducibility_note}
-""".format(**reproducibility_notes, **tf32_notes)
-```
-
-## Complete Example
-
-Here's a complete example showing all elements:
-
-```python
-def gumbel_softmax(
-    logits: Tensor,
-    tau: float = 1,
-    hard: bool = False,
-    eps: float = 1e-10,
-    dim: int = -1,
-) -> Tensor:
-    r"""
-    Sample from the Gumbel-Softmax distribution and optionally discretize.
-
-    Args:
-        logits (Tensor): `[..., num_features]` unnormalized log probabilities
-        tau (float): non-negative scalar temperature
-        hard (bool): if ``True``, the returned samples will be discretized as one-hot vectors,
-              but will be differentiated as if it is the soft sample in autograd. Default: ``False``
-        dim (int): A dimension along which softmax will be computed. Default: -1
-
-    Returns:
-        Tensor: Sampled tensor of same shape as `logits` from the Gumbel-Softmax distribution.
-            If ``hard=True``, the returned samples will be one-hot, otherwise they will
-            be probability distributions that sum to 1 across `dim`.
-
-    .. note::
-        This function is here for legacy reasons, may be removed from nn.Functional in the future.
-
-    Examples::
-        >>> logits = torch.randn(20, 32)
-        >>> # Sample soft categorical using reparametrization trick:
-        >>> F.gumbel_softmax(logits, tau=1, hard=False)
-        >>> # Sample hard categorical using "Straight-through" trick:
-        >>> F.gumbel_softmax(logits, tau=1, hard=True)
-
-    .. _Link 1:
-        https://arxiv.org/abs/1611.00712
-    """
-    # implementation
-```
-
-## Quick Checklist
-
-When writing a PyTorch docstring, ensure:
-
-- [ ] Use raw string (`r"""`)
-- [ ] Include function signature on first line
-- [ ] Provide brief description
-- [ ] Document all parameters in Args section with types
-- [ ] Include default values for optional parameters
-- [ ] Use Sphinx cross-references (`:func:`, `:class:`, `:meth:`)
-- [ ] Add mathematical formulas if applicable
-- [ ] Include at least one example in Examples section
-- [ ] Add warnings/notes for important caveats
-- [ ] Link to related module class with `:class:`
-- [ ] Use proper math notation for tensor shapes
-- [ ] Follow consistent formatting and indentation
-
-## Common Sphinx Roles Reference
-
-- `:class:\`~torch.nn.Module\`` - Class reference
-- `:func:\`torch.function\`` - Function reference
-- `:meth:\`~Tensor.method\`` - Method reference
-- `:attr:\`attribute\`` - Attribute reference
-- `:math:\`equation\`` - Inline math
-- `:ref:\`label\`` - Internal reference
-- ``` ``code`` ``` - Inline code (use double backticks)
-
-## Additional Notes
-
-- **Indentation**: Use 4 spaces for code, 2 spaces for continuation of parameter descriptions
-- **Line length**: Try to keep lines under 100 characters when possible
-- **Periods**: End sentences with periods, but not the signature line
-- **Backticks**: Use double backticks for code: ``` ``True`` ``None`` ``False`` ```
-- **Types**: Common types are `Tensor`, `int`, `float`, `bool`, `str`, `tuple`, `list`, etc.

.claude/skills/skill-writer/SKILL.md

@@ -1,385 +0,0 @@
----
-name: skill-writer
-description: Guide users through creating Agent Skills for Claude Code. Use when the user wants to create, write, author, or design a new Skill, or needs help with SKILL.md files, frontmatter, or skill structure.
----
-
-# Skill Writer
-
-This Skill helps you create well-structured Agent Skills for Claude Code that follow best practices and validation requirements.
-
-## When to use this Skill
-
-Use this Skill when:
-- Creating a new Agent Skill
-- Writing or updating SKILL.md files
-- Designing skill structure and frontmatter
-- Troubleshooting skill discovery issues
-- Converting existing prompts or workflows into Skills
-
-## Instructions
-
-### Step 1: Determine Skill scope
-
-First, understand what the Skill should do:
-
-1. **Ask clarifying questions**:
-   - What specific capability should this Skill provide?
-   - When should Claude use this Skill?
-   - What tools or resources does it need?
-   - Is this for personal use or team sharing?
-
-2. **Keep it focused**: One Skill = one capability
-   - Good: "PDF form filling", "Excel data analysis"
-   - Too broad: "Document processing", "Data tools"
-
-### Step 2: Choose Skill location
-
-Determine where to create the Skill:
-
-**Personal Skills** (`~/.claude/skills/`):
-- Individual workflows and preferences
-- Experimental Skills
-- Personal productivity tools
-
-**Project Skills** (`.claude/skills/`):
-- Team workflows and conventions
-- Project-specific expertise
-- Shared utilities (committed to git)
-
-### Step 3: Create Skill structure
-
-Create the directory and files:
-
-```bash
-# Personal
-mkdir -p ~/.claude/skills/skill-name
-
-# Project
-mkdir -p .claude/skills/skill-name
-```
-
-For multi-file Skills:
-```
-skill-name/
-├── SKILL.md (required)
-├── reference.md (optional)
-├── examples.md (optional)
-├── scripts/
-│   └── helper.py (optional)
-└── templates/
-    └── template.txt (optional)
-```
-
-### Step 4: Write SKILL.md frontmatter
-
-Create YAML frontmatter with required fields:
-
-```yaml
----
-name: skill-name
-description: Brief description of what this does and when to use it
----
-```
-
-**Field requirements**:
-
-- **name**:
-  - Lowercase letters, numbers, hyphens only
-  - Max 64 characters
-  - Must match directory name
-  - Good: `pdf-processor`, `git-commit-helper`
-  - Bad: `PDF_Processor`, `Git Commits!`
-
-- **description**:
-  - Max 1024 characters
-  - Include BOTH what it does AND when to use it
-  - Use specific trigger words users would say
-  - Mention file types, operations, and context
-
-**Optional frontmatter fields**:
-
-- **allowed-tools**: Restrict tool access (comma-separated list)
-  ```yaml
-  allowed-tools: Read, Grep, Glob
-  ```
-  Use for:
-  - Read-only Skills
-  - Security-sensitive workflows
-  - Limited-scope operations
-
-### Step 5: Write effective descriptions
-
-The description is critical for Claude to discover your Skill.
-
-**Formula**: `[What it does] + [When to use it] + [Key triggers]`
-
-**Examples**:
-
-✅ **Good**:
-```yaml
-description: Extract text and tables from PDF files, fill forms, merge documents. Use when working with PDF files or when the user mentions PDFs, forms, or document extraction.
-```
-
-✅ **Good**:
-```yaml
-description: Analyze Excel spreadsheets, create pivot tables, and generate charts. Use when working with Excel files, spreadsheets, or analyzing tabular data in .xlsx format.
-```
-
-❌ **Too vague**:
-```yaml
-description: Helps with documents
-description: For data analysis
-```
-
-**Tips**:
-- Include specific file extensions (.pdf, .xlsx, .json)
-- Mention common user phrases ("analyze", "extract", "generate")
-- List concrete operations (not generic verbs)
-- Add context clues ("Use when...", "For...")
-
-### Step 6: Structure the Skill content
-
-Use clear Markdown sections:
-
-```markdown
-# Skill Name
-
-Brief overview of what this Skill does.
-
-## Quick start
-
-Provide a simple example to get started immediately.
-
-## Instructions
-
-Step-by-step guidance for Claude:
-1. First step with clear action
-2. Second step with expected outcome
-3. Handle edge cases
-
-## Examples
-
-Show concrete usage examples with code or commands.
-
-## Best practices
-
-- Key conventions to follow
-- Common pitfalls to avoid
-- When to use vs. not use
-
-## Requirements
-
-List any dependencies or prerequisites:
-```bash
-pip install package-name
-```
-
-## Advanced usage
-
-For complex scenarios, see [reference.md](reference.md).
-```
-
-### Step 7: Add supporting files (optional)
-
-Create additional files for progressive disclosure:
-
-**reference.md**: Detailed API docs, advanced options
-**examples.md**: Extended examples and use cases
-**scripts/**: Helper scripts and utilities
-**templates/**: File templates or boilerplate
-
-Reference them from SKILL.md:
-```markdown
-For advanced usage, see [reference.md](reference.md).
-
-Run the helper script:
-\`\`\`bash
-python scripts/helper.py input.txt
-\`\`\`
-```
-
-### Step 8: Validate the Skill
-
-Check these requirements:
-
-✅ **File structure**:
-- [ ] SKILL.md exists in correct location
-- [ ] Directory name matches frontmatter `name`
-
-✅ **YAML frontmatter**:
-- [ ] Opening `---` on line 1
-- [ ] Closing `---` before content
-- [ ] Valid YAML (no tabs, correct indentation)
-- [ ] `name` follows naming rules
-- [ ] `description` is specific and < 1024 chars
-
-✅ **Content quality**:
-- [ ] Clear instructions for Claude
-- [ ] Concrete examples provided
-- [ ] Edge cases handled
-- [ ] Dependencies listed (if any)
-
-✅ **Testing**:
-- [ ] Description matches user questions
-- [ ] Skill activates on relevant queries
-- [ ] Instructions are clear and actionable
-
-### Step 9: Test the Skill
-
-1. **Restart Claude Code** (if running) to load the Skill
-
-2. **Ask relevant questions** that match the description:
-   ```
-   Can you help me extract text from this PDF?
-   ```
-
-3. **Verify activation**: Claude should use the Skill automatically
-
-4. **Check behavior**: Confirm Claude follows the instructions correctly
-
-### Step 10: Debug if needed
-
-If Claude doesn't use the Skill:
-
-1. **Make description more specific**:
-   - Add trigger words
-   - Include file types
-   - Mention common user phrases
-
-2. **Check file location**:
-   ```bash
-   ls ~/.claude/skills/skill-name/SKILL.md
-   ls .claude/skills/skill-name/SKILL.md
-   ```
-
-3. **Validate YAML**:
-   ```bash
-   cat SKILL.md | head -n 10
-   ```
-
-4. **Run debug mode**:
-   ```bash
-   claude --debug
-   ```
-
-## Common patterns
-
-### Read-only Skill
-
-```yaml
----
-name: code-reader
-description: Read and analyze code without making changes. Use for code review, understanding codebases, or documentation.
-allowed-tools: Read, Grep, Glob
----
-```
-
-### Script-based Skill
-
-```yaml
----
-name: data-processor
-description: Process CSV and JSON data files with Python scripts. Use when analyzing data files or transforming datasets.
----
-
-# Data Processor
-
-## Instructions
-
-1. Use the processing script:
-\`\`\`bash
-python scripts/process.py input.csv --output results.json
-\`\`\`
-
-2. Validate output with:
-\`\`\`bash
-python scripts/validate.py results.json
-\`\`\`
-```
-
-### Multi-file Skill with progressive disclosure
-
-```yaml
----
-name: api-designer
-description: Design REST APIs following best practices. Use when creating API endpoints, designing routes, or planning API architecture.
----
-
-# API Designer
-
-Quick start: See [examples.md](examples.md)
-
-Detailed reference: See [reference.md](reference.md)
-
-## Instructions
-
-1. Gather requirements
-2. Design endpoints (see examples.md)
-3. Document with OpenAPI spec
-4. Review against best practices (see reference.md)
-```
-
-## Best practices for Skill authors
-
-1. **One Skill, one purpose**: Don't create mega-Skills
-2. **Specific descriptions**: Include trigger words users will say
-3. **Clear instructions**: Write for Claude, not humans
-4. **Concrete examples**: Show real code, not pseudocode
-5. **List dependencies**: Mention required packages in description
-6. **Test with teammates**: Verify activation and clarity
-7. **Version your Skills**: Document changes in content
-8. **Use progressive disclosure**: Put advanced details in separate files
-
-## Validation checklist
-
-Before finalizing a Skill, verify:
-
-- [ ] Name is lowercase, hyphens only, max 64 chars
-- [ ] Description is specific and < 1024 chars
-- [ ] Description includes "what" and "when"
-- [ ] YAML frontmatter is valid
-- [ ] Instructions are step-by-step
-- [ ] Examples are concrete and realistic
-- [ ] Dependencies are documented
-- [ ] File paths use forward slashes
-- [ ] Skill activates on relevant queries
-- [ ] Claude follows instructions correctly
-
-## Troubleshooting
-
-**Skill doesn't activate**:
-- Make description more specific with trigger words
-- Include file types and operations in description
-- Add "Use when..." clause with user phrases
-
-**Multiple Skills conflict**:
-- Make descriptions more distinct
-- Use different trigger words
-- Narrow the scope of each Skill
-
-**Skill has errors**:
-- Check YAML syntax (no tabs, proper indentation)
-- Verify file paths (use forward slashes)
-- Ensure scripts have execute permissions
-- List all dependencies
-
-## Examples
-
-See the documentation for complete examples:
-- Simple single-file Skill (commit-helper)
-- Skill with tool permissions (code-reviewer)
-- Multi-file Skill (pdf-processing)
-
-## Output format
-
-When creating a Skill, I will:
-
-1. Ask clarifying questions about scope and requirements
-2. Suggest a Skill name and location
-3. Create the SKILL.md file with proper frontmatter
-4. Include clear instructions and examples
-5. Add supporting files if needed
-6. Provide testing instructions
-7. Validate against all requirements
-
-The result will be a complete, working Skill that follows all best practices and validation rules.

.flake8

@@ -7,12 +7,16 @@ max-line-length = 120
 # C408 ignored because we like the dict keyword argument syntax
 # E501 is not flexible enough, we're using B950 instead
 ignore =
-    E203,E305,E402,E501,E704,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
+    E203,E305,E402,E501,E704,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
     # shebang has extra meaning in fbcode lints, so I think it's not worth trying
     # to line this up with executable bit
     EXE001,
     # these ignores are from flake8-bugbear; please fix!
     B007,B008,B017,B019,B023,B028,B903,B905,B906,B907,B908,B910
+    # these ignores are from flake8-comprehensions; please fix!
+    C407,
+    # these ignores are from flake8-logging-format; please fix!
+    G100,G101,G200
     # these ignores are from flake8-simplify. please fix or ignore with commented reason
     SIM105,SIM108,SIM110,SIM111,SIM113,SIM114,SIM115,SIM116,SIM117,SIM118,SIM119,SIM12,
     # SIM104 is already covered by pyupgrade ruff

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

@@ -65,7 +65,7 @@ runs:
           cd .ci/lumen_cli
           python3 -m pip install -e .
         )
-        MAX_JOBS="$(nproc --ignore=10)"
+        MAX_JOBS="$(nproc --ignore=6)"
         export MAX_JOBS
 
         # Split the comma-separated list and build each target

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

@@ -124,10 +124,3 @@ runs:
       id: login-ecr
       continue-on-error: true
       uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
-
-    - name: Preserve github env variables for use in docker
-      shell: bash
-      run: |
-        env | grep '^GITHUB' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
-        env | grep '^CI' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"
-        env | grep '^RUNNER' >> "${RUNNER_TEMP}/github_env_${GITHUB_RUN_ID}"

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-69bbe7363897764f9e758d851cd0340147d27f94
+8ad2aa5d354d1bf432339113860185d5a5d1abbd

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-1752fe6809b74921644866275ab80244b96e80bc
+f5c6c2ec6490455e86f67b2a25c10390d60a27f7

.github/ci_commit_pins/xla.txt

@@ -1 +1 @@
-df6798dfb931ce7c7fe5bed2447cd1092a5981af
+0fa6e3129e61143224663e1ec67980d12b7ec4eb

.github/ci_configs/vllm/Dockerfile

@@ -283,9 +283,6 @@ RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
         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 '.'); \
     fi
 
-RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install --system --pre apache-tvm-ffi==0.1.0b15
-
 # Install the vllm wheel from previous stage
 RUN --mount=type=cache,target=/root/.cache/uv \
     uv pip install --system /wheels/vllm/*.whl --verbose
@@ -298,8 +295,6 @@ RUN --mount=type=cache,target=/root/.cache/uv \
 ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
 ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
 
-# TODO(elainewy): remove this once vllm commit is updated, and install flashinfer from pip
-# see https://github.com/pytorch/pytorch/pull/165274#issuecomment-3408531784
 ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
 ARG FLASHINFER_GIT_REF="v0.2.14.post1"
 

.github/label_to_label.yml

@@ -15,11 +15,6 @@
   - "module: reinplacing"
   then:
   - "module: pt2-dispatcher"
-- any:
-  - "vllm-compile"
-  then:
-  - "module: vllm"
-  - "oncall: pt2"
 - any:
   - "module: vmap"
   then:
@@ -32,6 +27,10 @@
   - "module: pt2 optimizer"
   then:
   - "module: dynamo"
+- any:
+  - "module: flex attention"
+  then:
+  - "module: higher order operators"
 - any:
   - "module: aotinductor"
   then:

.github/labeler.yml

@@ -133,32 +133,3 @@
 
 "ciflow/vllm":
 - .github/ci_commit_pins/vllm.txt
-
-"ciflow/b200":
-- test/test_matmul_cuda.py
-- test/test_scaled_matmul_cuda.py
-- test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/Blas.cpp
-- torch/**/*cublas*
-- torch/_inductor/kernel/mm.py
-- test/inductor/test_max_autotune.py
-- third_party/fbgemm
-
-"ciflow/h100":
-- test/test_matmul_cuda.py
-- test/test_scaled_matmul_cuda.py
-- test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/Blas.cpp
-- torch/**/*cublas*
-- torch/_inductor/kernel/mm.py
-- test/inductor/test_max_autotune.py
-- third_party/fbgemm
-
-"ciflow/rocm":
-- test/test_matmul_cuda.py
-- test/test_scaled_matmul_cuda.py
-- test/inductor/test_fp8.py
-- aten/src/ATen/native/cuda/Blas.cpp
-- torch/_inductor/kernel/mm.py
-- test/inductor/test_max_autotune.py
-- third_party/fbgemm

.github/pytorch-probot.yml

@@ -3,7 +3,6 @@ ciflow_tracking_issue: 64124
 ciflow_push_tags:
 - ciflow/b200
 - ciflow/b200-symm-mem
-- ciflow/b200-distributed
 - ciflow/binaries
 - ciflow/binaries_libtorch
 - ciflow/binaries_wheel
@@ -33,7 +32,6 @@ ciflow_push_tags:
 - ciflow/rocm
 - ciflow/rocm-mi300
 - ciflow/rocm-mi355
-- ciflow/rocm-navi31
 - ciflow/s390
 - ciflow/slow
 - ciflow/torchbench

.github/scripts/generate_binary_build_matrix.py

@@ -22,7 +22,7 @@ CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
     "12.9": "12.9.1",
-    "13.0": "13.0.2",
+    "13.0": "13.0.0",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
@@ -79,38 +79,38 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
     ),
     "12.9": (
-        "nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
-        "nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | "
-        "nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | "
-        "nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | "
-        "nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | "
-        "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | "
-        "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
-        "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | "
-        "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | "
-        "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | "
-        "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
+        "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==13.0.88; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | "
+        "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==13.1.0.3; platform_system == 'Linux' | "
-        "nvidia-cufft==12.0.0.61; 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==12.0.4.66; platform_system == 'Linux' | "
-        "nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | "
-        "nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | "
-        "nvidia-cufile==1.15.1.6; 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'"
     ),
     "xpu": (
         "intel-cmplr-lib-rt==2025.2.1 | "
@@ -241,11 +241,7 @@ def generate_libtorch_matrix(
             arches += CUDA_ARCHES
             arches += ROCM_ARCHES
         elif os == "windows":
-            # TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
-            # in 2.10
-            windows_cuda_arches = CUDA_ARCHES.copy()
-            windows_cuda_arches.remove("12.9")
-            arches += windows_cuda_arches
+            arches += CUDA_ARCHES
     if libtorch_variants is None:
         libtorch_variants = [
             "shared-with-deps",
@@ -309,11 +305,7 @@ def generate_wheels_matrix(
         if os == "linux":
             arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
         elif os == "windows":
-            # TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
-            # in 2.10
-            windows_cuda_arches = CUDA_ARCHES.copy()
-            windows_cuda_arches.remove("12.9")
-            arches += windows_cuda_arches + XPU_ARCHES
+            arches += CUDA_ARCHES + XPU_ARCHES
         elif os == "linux-aarch64":
             # Separate new if as the CPU type is different and
             # uses different build/test scripts

.github/scripts/trymerge.py

@@ -1092,7 +1092,7 @@ class GitHubPR:
         editor = node["editor"]
         return GitHubComment(
             body_text=node["bodyText"],
-            created_at=node.get("createdAt", ""),
+            created_at=node["createdAt"] if "createdAt" in node else "",
             author_login=node["author"]["login"],
             author_url=node["author"].get("url", None),
             author_association=node["authorAssociation"],

.github/templates/macos_binary_build_workflow.yml.j2

@@ -26,8 +26,9 @@ name: !{{ build_environment }}
       - 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 '.0') }}"
+          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 %}
 

.github/templates/windows_binary_build_workflow.yml.j2

@@ -79,9 +79,9 @@ jobs:
     runs-on: "windows-11-arm64-preview"
     {%- else %}
     {%- if branches == "nightly" %}
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     {%- else %}
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge.nonephemeral"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
     {%- endif %}
     {%- endif %}
     timeout-minutes: !{{ common.timeout_minutes_windows_binary }}

.github/workflows/_linux-build.yml

@@ -37,7 +37,7 @@ on:
       runner:
         required: false
         type: string
-        default: "linux.c7i.2xlarge"
+        default: "linux.2xlarge"
         description: |
           Label of the runner this job should run on.
       test-matrix:

.github/workflows/_linux-test.yml

@@ -224,46 +224,6 @@ jobs:
         continue-on-error: true
         uses: ./.github/actions/download-td-artifacts
 
-      - name: Download Windows torch wheel for cross-compilation
-        if: matrix.win_torch_wheel_artifact != ''
-        uses: seemethere/download-artifact-s3@1da556a7aa0a088e3153970611f6c432d58e80e6 # v4.2.0
-        with:
-          name: ${{ matrix.win_torch_wheel_artifact }}
-          path: win-torch-wheel
-
-      - name: Extract Windows wheel and setup CUDA libraries
-        if: matrix.win_torch_wheel_artifact != ''
-        shell: bash
-        run: |
-          set -x
-
-          # Find the wheel file
-          WHEEL_FILE=$(find win-torch-wheel -name "*.whl" -type f | head -n 1)
-          if [ -z "$WHEEL_FILE" ]; then
-            echo "Error: No wheel file found in win-torch-wheel directory"
-            exit 1
-          fi
-          echo "Found wheel file: $WHEEL_FILE"
-
-          # Unzip the wheel file
-          unzip -q "$WHEEL_FILE" -d win-torch-wheel-extracted
-          echo "Extracted wheel contents"
-
-          # Setup CUDA libraries (cuda.lib and cudart.lib) directory
-          mkdir -p win-torch-wheel-extracted/lib/x64
-          if [ -f "win-torch-wheel/cuda.lib" ]; then
-            mv win-torch-wheel/cuda.lib win-torch-wheel-extracted/lib/x64/
-            echo "Moved cuda.lib to win-torch-wheel-extracted/lib/x64/"
-          fi
-          if [ -f "win-torch-wheel/cudart.lib" ]; then
-            mv win-torch-wheel/cudart.lib win-torch-wheel-extracted/lib/x64/
-            echo "Moved cudart.lib to win-torch-wheel-extracted/lib/x64/"
-          fi
-
-          # Verify CUDA libraries are present
-          echo "CUDA libraries:"
-          ls -la win-torch-wheel-extracted/lib/x64/ || echo "No CUDA libraries found"
-
       - name: Parse ref
         id: parse-ref
         run: .github/scripts/parse_ref.py

.github/workflows/_win-build.yml

@@ -168,31 +168,6 @@ jobs:
         run: |
           .ci/pytorch/win-build.sh
 
-      # Collect Windows torch libs and CUDA libs for cross-compilation
-      - name: Collect Windows CUDA libs for cross-compilation
-        if: steps.build.outcome != 'skipped' && inputs.cuda-version != 'cpu'
-        shell: bash
-        run: |
-          set -ex
-
-          # Create directory structure if does not exist
-          mkdir -p /c/${{ github.run_id }}/build-results
-
-          # Copy CUDA libs
-          CUDA_PATH="/c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v${{ inputs.cuda-version }}"
-
-          if [ -f "${CUDA_PATH}/lib/x64/cuda.lib" ]; then
-            cp "${CUDA_PATH}/lib/x64/cuda.lib" /c/${{ github.run_id }}/build-results/
-          fi
-
-          if [ -f "${CUDA_PATH}/lib/x64/cudart.lib" ]; then
-            cp "${CUDA_PATH}/lib/x64/cudart.lib" /c/${{ github.run_id }}/build-results/
-          fi
-
-          # List collected files
-          echo "Collected CUDA libs:"
-          ls -lah /c/${{ github.run_id }}/build-results/*.lib
-
       # Upload to github so that people can click and download artifacts
       - name: Upload artifacts to s3
         if: steps.build.outcome != 'skipped'

.github/workflows/b200-distributed.yml

@@ -1,62 +0,0 @@
-name: CI for distributed tests on B200
-
-on:
-  pull_request:
-    paths:
-      - .github/workflows/b200-distributed.yml
-  workflow_dispatch:
-  push:
-    tags:
-      - ciflow/b200-distributed/*
-  schedule:
-    - cron: 46 8 * * *  # about 1:46am PDT
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-
-  get-label-type:
-    if: github.repository_owner == 'pytorch'
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-build-distributed-b200
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      runner: linux.12xlarge.memory
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
-      docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
-      cuda-arch-list: '10.0'
-      test-matrix: |
-        { include: [
-          { config: "distributed", shard: 1, num_shards: 2, runner: "linux.dgx.b200.8" },
-          { config: "distributed", shard: 2, num_shards: 2, runner: "linux.dgx.b200.8" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-cuda12_8-py3_10-gcc11-test-distributed-b200:
-    name: linux-jammy-cuda12.8-py3.10-gcc11-test-b200
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200
-    with:
-      timeout-minutes: 1200
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200.outputs.test-matrix }}
-      aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-    secrets: inherit

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

@@ -27,8 +27,9 @@ jobs:
       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', 'cu130' ]
+        device: [ 'cu128', 'cu129' ]
         include:
           - platform: manylinux_2_28_x86_64
             device: cu128
@@ -38,10 +39,6 @@ jobs:
             device: cu129
             manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
             runner: linux.12xlarge.memory
-          - platform: manylinux_2_28_x86_64
-            device: cu130
-            manylinux-image: 'pytorch/manylinux2_28-builder:cuda13.0'
-            runner: linux.12xlarge.memory
           - platform: manylinux_2_28_aarch64
             device: cu128
             manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.8'
@@ -50,11 +47,6 @@ jobs:
             device: cu129
             manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.9'
             runner: linux.arm64.r7g.12xlarge.memory
-        exclude:
-          # TODO (huydhn): Add cu130 aarch64 once PyTorch is on 2.9+ and
-          # xformers is update to support 13.0
-          - platform: manylinux_2_28_aarch64
-            device: cu130
     name: "Build ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
     runs-on: ${{ matrix.runner }}
     timeout-minutes: 480
@@ -177,12 +169,7 @@ jobs:
       fail-fast: false
       matrix:
         platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
-        device: [ 'cu128', 'cu129', 'cu130' ]
-        exclude:
-          # TODO (huydhn): Add cu130 aarch64 once PyTorch is on 2.9+ and
-          # xformers is update to support 13.0
-          - platform: manylinux_2_28_aarch64
-            device: cu130
+        device: [ 'cu128', 'cu129' ]
     env:
       PLATFORM: ${{ matrix.platform }}
       BUILD_DEVICE: ${{ matrix.device }}

.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml

@@ -224,7 +224,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -270,7 +270,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -473,7 +473,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -519,7 +519,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -722,7 +722,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -768,7 +768,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -971,7 +971,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1017,7 +1017,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1220,7 +1220,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1266,7 +1266,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1469,7 +1469,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1515,7 +1515,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1718,7 +1718,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1764,7 +1764,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

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

@@ -259,7 +259,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_9-test:  # Testing
@@ -325,7 +325,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_10-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda13_0-test:  # Testing
@@ -925,7 +925,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_9-test:  # Testing
@@ -991,7 +991,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_11-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda13_0-test:  # Testing
@@ -1591,7 +1591,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_9-test:  # Testing
@@ -1657,7 +1657,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_12-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda13_0-test:  # Testing
@@ -2257,7 +2257,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_9-test:  # Testing
@@ -2323,7 +2323,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda13_0-test:  # Testing
@@ -2923,7 +2923,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_9-test:  # Testing
@@ -2989,7 +2989,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_13t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda13_0-test:  # Testing
@@ -3589,7 +3589,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_9-test:  # Testing
@@ -3655,7 +3655,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda13_0-test:  # Testing
@@ -4255,7 +4255,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda12_9
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_9-test:  # Testing
@@ -4321,7 +4321,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build_name: manywheel-py3_14t-cuda13_0
       build_environment: linux-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; 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==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: 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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda13_0-test:  # Testing

.github/workflows/generated-macos-arm64-binary-libtorch-release-nightly.yml

@@ -63,6 +63,7 @@ jobs:
       - 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: "3.10.4"
           freethreaded: false

.github/workflows/generated-macos-arm64-binary-wheel-nightly.yml

@@ -59,6 +59,7 @@ jobs:
       - 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: "3.10.4"
           freethreaded: false
@@ -168,6 +169,7 @@ jobs:
       - 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: "3.11.4"
           freethreaded: false
@@ -277,6 +279,7 @@ jobs:
       - 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: "3.12.4"
           freethreaded: false
@@ -386,6 +389,7 @@ jobs:
       - 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: "3.13.4"
           freethreaded: false
@@ -495,6 +499,7 @@ jobs:
       - 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: "3.13.4"
           freethreaded: true
@@ -604,8 +609,9 @@ jobs:
       - 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: "3.14.0"
+          python-version: "3.14.0-rc.2"
           freethreaded: false
       - name: Checkout PyTorch
         uses: actions/checkout@v4
@@ -713,8 +719,9 @@ jobs:
       - 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: "3.14.0"
+          python-version: "3.14.0-rc.2"
           freethreaded: true
       - name: Checkout PyTorch
         uses: actions/checkout@v4

.github/workflows/generated-windows-binary-libtorch-debug-nightly.yml

@@ -44,7 +44,7 @@ jobs:
   libtorch-cpu-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -291,7 +291,7 @@ jobs:
   libtorch-cuda12_6-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -541,7 +541,7 @@ jobs:
   libtorch-cuda12_8-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -788,10 +788,260 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+  libtorch-cuda12_9-shared-with-deps-debug-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs: get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-debug
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+
+  libtorch-cuda12_9-shared-with-deps-debug-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-debug-build
+      - get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/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: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # 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"
+    steps:
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-debug
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Test PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_test.sh"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+  libtorch-cuda12_9-shared-with-deps-debug-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-debug-test
+    with:
+      PYTORCH_ROOT: ${{ github.workspace }}/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: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      LIBTORCH_CONFIG: debug
+      LIBTORCH_VARIANT: shared-with-deps
+      # 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"
+      build_name: libtorch-cuda12_9-shared-with-deps-debug
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   libtorch-cuda13_0-shared-with-deps-debug-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch

.github/workflows/generated-windows-binary-libtorch-release-nightly.yml

@@ -44,7 +44,7 @@ jobs:
   libtorch-cpu-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -291,7 +291,7 @@ jobs:
   libtorch-cuda12_6-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -541,7 +541,7 @@ jobs:
   libtorch-cuda12_8-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -788,10 +788,260 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
+  libtorch-cuda12_9-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs: get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # This is a dummy value for libtorch to work correctly with our batch scripts
+      # without this value pip does not get installed for some reason
+      DESIRED_PYTHON: "3.10"
+    steps:
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Build PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_build.sh"
+      - uses: actions/upload-artifact@v4.4.0
+        if: always()
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-release
+          retention-days: 14
+          if-no-files-found: error
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+
+  libtorch-cuda12_9-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-cuda12_9-shared-with-deps-release-build
+      - get-label-type
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.g4dn.xlarge"
+    timeout-minutes: 360
+    env:
+      PYTORCH_ROOT: ${{ github.workspace }}/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: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      SKIP_ALL_TESTS: 1
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # 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"
+    steps:
+      - name: Display EC2 information
+        shell: bash
+        run: |
+          set -euo pipefail
+          function get_ec2_metadata() {
+            # Pulled from instance metadata endpoint for EC2
+            # see https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/instancedata-data-retrieval.html
+            category=$1
+            curl -H "X-aws-ec2-metadata-token: $(curl -s -X PUT "http://169.254.169.254/latest/api/token" -H "X-aws-ec2-metadata-token-ttl-seconds: 30")" -fsSL "http://169.254.169.254/latest/meta-data/${category}"
+          }
+          echo "ami-id: $(get_ec2_metadata ami-id)"
+          echo "instance-id: $(get_ec2_metadata instance-id)"
+          echo "instance-type: $(get_ec2_metadata instance-type)"
+          echo "system info $(uname -a)"
+      - name: "[FB EMPLOYEES] Enable SSH (Click me for login details)"
+        uses: pytorch/test-infra/.github/actions/setup-ssh@main
+        continue-on-error: true
+        with:
+          github-secret: ${{ secrets.GITHUB_TOKEN }}
+      - name: Enable git long paths and symlinks on Windows and disable fsmonitor daemon
+        shell: bash
+        run: |
+          git config --global core.longpaths true
+          git config --global core.symlinks true
+
+          # https://git-scm.com/docs/git-fsmonitor--daemon.  The daemon could lock
+          # the directory on Windows and prevent GHA from checking out as reported
+          # in https://github.com/actions/checkout/issues/1018
+          git config --global core.fsmonitor false
+      # Needed for binary builds, see: https://github.com/pytorch/pytorch/issues/73339#issuecomment-1058981560
+      - name: Enable long paths on Windows
+        shell: powershell
+        run: |
+          Set-ItemProperty -Path "HKLM:\\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1
+      # Since it's just a defensive command, the workflow should continue even the command fails. This step can be
+      # removed once Windows Defender is removed from the AMI
+      - name: Disables Windows Defender scheduled and real-time scanning for files in directories used by PyTorch
+        continue-on-error: true
+        shell: powershell
+        run: |
+          Add-MpPreference -ExclusionPath $(Get-Location).tostring(),$Env:TEMP -ErrorAction Ignore
+          # Let's both exclude the path and disable Windows Defender completely just to be sure
+          # that it doesn't interfere
+          Set-MpPreference -DisableRealtimeMonitoring $True -ErrorAction Ignore
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      # NOTE: These environment variables are put here so that they can be applied on every job equally
+      #       They are also here because setting them at a workflow level doesn't give us access to the
+      #       runner.temp variable, which we need.
+      - name: Populate binary env
+        shell: bash
+        run: |
+          echo "BINARY_ENV_FILE=${RUNNER_TEMP}/env" >> "${GITHUB_ENV}"
+          echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
+          echo "WIN_PACKAGE_WORK_DIR=${RUNNER_TEMP}"
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-cuda12_9-shared-with-deps-release
+          path: "${{ env.PYTORCH_FINAL_PACKAGE_DIR }}"
+      - name: Populate binary env
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
+      - name: Test PyTorch binary
+        shell: bash
+        run: |
+          "${PYTORCH_ROOT}/.circleci/scripts/binary_windows_test.sh"
+      - name: Wait until all sessions have drained
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        timeout-minutes: 120
+        run: |
+          .github\scripts\wait_for_ssh_to_drain.ps1
+      - name: Kill active ssh sessions if still around (Useful if workflow was cancelled)
+        shell: powershell
+        working-directory: pytorch
+        if: always()
+        run: |
+          .github\scripts\kill_active_ssh_sessions.ps1
+  libtorch-cuda12_9-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-cuda12_9-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: ${{ github.workspace }}/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: cu129
+      GPU_ARCH_VERSION: "12.9"
+      GPU_ARCH_TYPE: cuda
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      # 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"
+      build_name: libtorch-cuda12_9-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
   libtorch-cuda13_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch

.github/workflows/inductor-perf-test-nightly-rocm-mi355.yml

@@ -88,27 +88,27 @@ jobs:
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
       test-matrix: |
         { include: [
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_huggingface_perf_rocm_mi355", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_timm_perf_rocm_mi355", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "inductor_torchbench_perf_rocm_mi355", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_huggingface_perf_rocm_mi355", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_timm_perf_rocm_mi355", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "inductor_torchbench_perf_rocm_mi355", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.mi355.2" },
         ]}
     secrets: inherit
 

.github/workflows/inductor-periodic.yml

@@ -88,6 +88,7 @@ jobs:
     with:
       build-environment: linux-jammy-rocm-py3_10
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
+      sync-tag: rocm-build
       test-matrix: |
         { include: [
           { config: "dynamo_eager_torchbench", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },

.github/workflows/lint.yml

@@ -118,9 +118,9 @@ jobs:
         CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
         echo "Running all other linters"
         if [ "$CHANGED_FILES" = '*' ]; then
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY --all-files" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT --all-files" .github/scripts/lintrunner.sh
         else
-          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
+          ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT ${CHANGED_FILES}" .github/scripts/lintrunner.sh
         fi
 
   quick-checks:

.github/workflows/operator_benchmark.yml

@@ -7,11 +7,9 @@ on:
   workflow_dispatch:
     inputs:
       test_mode:
-        type: choice
-        options:
-          - 'short'
-          - 'long'
-          - 'all'
+        required: false
+        type: string
+        default: 'short'
         description: tag filter for operator benchmarks, options from long, short, all
   schedule:
     # Run at 07:00 UTC every Sunday
@@ -30,49 +28,38 @@ permissions:
   contents: read
 
 jobs:
-  x86-opbenchmark-build:
+  opbenchmark-build:
     if: github.repository_owner == 'pytorch'
-    name: x86-opbenchmark-build
+    name: opbenchmark-build
     uses: ./.github/workflows/_linux-build.yml
     with:
       build-environment: linux-jammy-py3.10-gcc11-build
       docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
-          { config: "cpu_operator_benchmark_${{ inputs.test_mode || 'short' }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
+          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
         ]}
     secrets: inherit
 
-  x86-opbenchmark-test:
-    name: x86-opbenchmark-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs: x86-opbenchmark-build
-    with:
-      build-environment: linux-jammy-py3.10-gcc11-build
-      docker-image: ${{ needs.x86-opbenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.x86-opbenchmark-build.outputs.test-matrix }}
-    secrets: inherit
-
-  aarch64-opbenchmark-build:
-    if: github.repository_owner == 'pytorch'
-    name: aarch64-opbenchmark-build
+  opbenchmark-on-demand-build:
+    if: ${{ github.event_name == 'workflow_dispatch' && github.repository_owner == 'pytorch' }}
+    name: opbenchmark-on-demand-build
     uses: ./.github/workflows/_linux-build.yml
     with:
-      build-environment: linux-jammy-aarch64-py3.10
-      runner: linux.arm64.m7g.4xlarge
-      docker-image-name: ci-image:pytorch-linux-jammy-aarch64-py3.10-gcc11
+      build-environment: linux-jammy-py3.10-gcc11-build
+      docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
       test-matrix: |
         { include: [
-          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.arm64.m8g.4xlarge" },
+          { config: "cpu_operator_benchmark_${{ inputs.test_mode }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
         ]}
     secrets: inherit
 
-  aarch64-opbenchmark-test:
-    name: aarch64-opbenchmark-test
+  opbenchmark-test:
+    name: opbenchmark-test
     uses: ./.github/workflows/_linux-test.yml
-    needs: aarch64-opbenchmark-build
+    needs: opbenchmark-build
     with:
-      build-environment: linux-jammy-aarch64-py3.10
-      docker-image: ${{ needs.aarch64-opbenchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.aarch64-opbenchmark-build.outputs.test-matrix }}
+      build-environment: linux-jammy-py3.10-gcc11-build
+      docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
+      test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
     secrets: inherit

.github/workflows/periodic.yml

@@ -147,16 +147,15 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-cuda12.8-py3.10-gcc9-debug
       docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc9
-      cuda-arch-list: 8.9
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 2, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 3, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 4, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 5, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 6, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
-          { config: "default", shard: 7, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 1, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 2, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 3, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 4, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 5, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 6, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
+          { config: "default", shard: 7, num_shards: 7, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge.nvidia.gpu", owners: ["oncall:debug-build"] },
         ]}
     secrets: inherit
 

.github/workflows/pull.yml

@@ -347,8 +347,7 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      # This should sync with the build in xpu.yml but xpu uses a larger runner
-      # sync-tag: linux-xpu-n-build
+      sync-tag: linux-xpu-n-build
       runner_prefix: ${{ needs.get-label-type.outputs.label-type }}
       build-environment: linux-jammy-xpu-n-py3.10
       docker-image-name: ci-image:pytorch-linux-jammy-xpu-n-py3

.github/workflows/rocm-mi300.yml

@@ -45,6 +45,7 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-noble-rocm-py3.12-mi300
       docker-image-name: ci-image:pytorch-linux-noble-rocm-n-py3
+      sync-tag: rocm-build
       test-matrix: |
         { include: [
           { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.gfx942.1" },

.github/workflows/rocm-mi355.yml

@@ -42,14 +42,15 @@ jobs:
       runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-noble-rocm-py3.12-mi355
       docker-image-name: ci-image:pytorch-linux-noble-rocm-n-py3
+      sync-tag: rocm-build
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
-          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
+          { config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
+          { config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
         ]}
     secrets: inherit
 

.github/workflows/rocm-navi31.yml

@@ -1,75 +0,0 @@
-name: rocm-navi31
-
-on:
-  push:
-    tags:
-      - ciflow/rocm-navi31/*
-  workflow_dispatch:
-  schedule:
-    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
-    # Also run less frequently on weekends.
-    - cron: 45 */2 * * 1-5
-    - cron: 45 4,12 * * 0,6
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions: read-all
-
-jobs:
-  target-determination:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      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 }}
-
-  linux-jammy-rocm-py3_10-build:
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
-          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3_10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-      tests-to-include: >-
-         ${{ github.event_name == 'schedule' && 'test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs
-         test_autograd inductor/test_torchinductor inductor/test_kernel_benchmark
-         inductor/test_pad_mm inductor/test_benchmark_fusion inductor/test_aot_inductor
-         inductor/test_torchinductor inductor/test_decompose_mem_bound_mm
-         inductor/test_flex_attention inductor/test_max_autotune' || '' }}
-    secrets: inherit

.github/workflows/rocm.yml

@@ -26,23 +26,11 @@ jobs:
       id-token: write
       contents: read
 
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      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 }}
-
   linux-jammy-rocm-py3_10-build:
     if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
     name: linux-jammy-rocm-py3.10
     uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
     with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
       build-environment: linux-jammy-rocm-py3.10
       docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
       sync-tag: rocm-build
@@ -71,3 +59,29 @@ jobs:
       docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
       test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
     secrets: inherit
+
+  linux-jammy-rocm-py3_10-gfx1100-test:
+    if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/main' }}
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3_10-gfx1100
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: |
+        { include: [
+          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
+          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx1100" },
+        ]}
+      tests-to-include: >
+         test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs
+         test_autograd inductor/test_torchinductor inductor/test_kernel_benchmark
+         inductor/test_pad_mm inductor/test_benchmark_fusion inductor/test_aot_inductor
+         inductor/test_torchinductor inductor/test_decompose_mem_bound_mm
+         inductor/test_flex_attention inductor/test_max_autotune
+    secrets: inherit

.github/workflows/trunk-tagging.yml

@@ -58,10 +58,8 @@ jobs:
           else
             COMMIT_SHA="${{ github.sha }}"
           fi
-          {
-            echo "sha=${COMMIT_SHA}"
-            echo "tag_name=trunk/${COMMIT_SHA}"
-          } >> "${GITHUB_OUTPUT}"
+          echo "sha=${COMMIT_SHA}" >> "${GITHUB_OUTPUT}"
+          echo "tag_name=trunk/${COMMIT_SHA}" >> "${GITHUB_OUTPUT}"
 
       - name: Validate commit SHA
         run: |
@@ -89,7 +87,7 @@ jobs:
             echo "✅ Commit ${COMMIT_SHA} is valid (automatic push trigger)"
           fi
 
-      - name: Create and push tag(s) with retry
+      - name: Create and push tag with retry
         id: check_tag
         env:
           TAG_NAME: ${{ steps.commit.outputs.tag_name }}
@@ -114,23 +112,14 @@ jobs:
             return 1
           }
 
-          # Counters for summary reporting
-          created_count=0
-          skipped_count=0
-          failed_count=0
+          # Exit early if tag already exists
+          if check_tag_exists; then
+            echo "✅ Tag already exists - no action needed"
+            echo "exists=true" >> "${GITHUB_OUTPUT}"
+            exit 0
+          fi
 
-          # Always write outputs once on exit
-          finish() {
-            set +e
-            if [ -n "${GITHUB_OUTPUT:-}" ]; then
-              {
-                echo "created_count=${created_count}"
-                echo "skipped_count=${skipped_count}"
-                echo "failed_count=${failed_count}"
-              } >> "${GITHUB_OUTPUT}"
-            fi
-          }
-          trap finish EXIT
+          echo "Tag ${TAG_NAME} does not exist, proceeding with creation"
 
           # Retry configuration
           MAX_RETRIES=5
@@ -205,111 +194,31 @@ jobs:
             }
           }
 
-          # New behavior for push events: enumerate commits in the push and tag each one.
-          # For workflow_dispatch, retain existing single-SHA behavior.
-
-          # Always fetch tags once up front to improve idempotency in loops
-          git fetch origin --tags --quiet || true
-
-          if [ "${{ github.event_name }}" = "push" ]; then
-            BEFORE_SHA="${{ github.event.before }}"
-            AFTER_SHA="${{ github.sha }}"  # same as event.after
-
-            # List commits introduced by this push (old..new), oldest first for stable ordering
-            commits_file="$(mktemp)"
-            git rev-list --reverse "${BEFORE_SHA}..${AFTER_SHA}" > "${commits_file}"
-
-            if [ ! -s "${commits_file}" ]; then
-              echo "No new commits found between ${BEFORE_SHA}..${AFTER_SHA}; nothing to tag."
-              rm -f "${commits_file}"
-              exit 0
-            fi
-
-            commit_count="$(wc -l < "${commits_file}" | tr -d ' ')"
-            echo "Found ${commit_count} commit(s) to tag for push:"
-            while IFS= read -r sha; do
-              printf '  %s\n' "${sha}"
-            done < "${commits_file}"
-
-            while IFS= read -r sha; do
-              TAG_NAME="trunk/${sha}"
-              COMMIT_SHA="${sha}"
-
-              # If tag already exists locally or remotely, skip (idempotent)
-              if check_tag_exists; then
-                echo "✅ Tag ${TAG_NAME} already exists - skipping"
-                skipped_count=$((skipped_count + 1))
-                continue
-              fi
-
-              echo "Tag ${TAG_NAME} does not exist, proceeding with creation"
-
-              if retry_with_backoff "tag_with_retry" "Creating tag ${TAG_NAME} for commit ${COMMIT_SHA}"; then
-                created_count=$((created_count + 1))
-              else
-                echo "Tag creation failed after all retry attempts for ${TAG_NAME}"
-                failed_count=$((failed_count + 1))
-              fi
-            done < "${commits_file}"
-
-            rm -f "${commits_file}"
-
-            if [ "${failed_count}" -gt 0 ]; then
-              exit 1
-            fi
+          # Execute with retry
+          if retry_with_backoff "tag_with_retry" "Creating tag ${TAG_NAME} for commit ${COMMIT_SHA}"; then
+            echo "exists=false" >> "${GITHUB_OUTPUT}"
             exit 0
           else
-            # workflow_dispatch path (single SHA tagging preserved)
-
-            # Exit early if tag already exists
-            if check_tag_exists; then
-              echo "✅ Tag already exists - no action needed"
-              skipped_count=1
-              exit 0
-            fi
-
-            echo "Tag ${TAG_NAME} does not exist, proceeding with creation"
-
-            if retry_with_backoff "tag_with_retry" "Creating tag ${TAG_NAME} for commit ${COMMIT_SHA}"; then
-              created_count=1
-              exit 0
-            else
-              echo "Tag creation failed after all retry attempts"
-              failed_count=1
-              exit 1
-            fi
+            echo "Tag creation failed after all retry attempts"
+            exit 1
           fi
 
       - name: Tag creation summary
         if: always()
         run: |
-          if [ "${{ github.event_name }}" = "push" ]; then
-            echo "Trigger: push on main"
-            echo "Created: ${{ steps.check_tag.outputs.created_count }}"
-            echo "Skipped (already existed): ${{ steps.check_tag.outputs.skipped_count }}"
-            echo "Failed: ${{ steps.check_tag.outputs.failed_count }}"
-            if [ "${{ steps.check_tag.outputs.failed_count }}" = "0" ]; then
-              echo "✅ Completed tagging for push range ${{ github.event.before }}..${{ github.sha }}"
-            else
-              echo "❌ Some tags failed to create for push range ${{ github.event.before }}..${{ github.sha }}"
-            fi
+          if [ "${{ steps.check_tag.outputs.exists }}" = "true" ]; then
+            echo "✅ Tag ${{ steps.commit.outputs.tag_name }} already existed - no action needed"
+          elif [ "${{ job.status }}" = "success" ]; then
+            echo "✅ Successfully created tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
           else
-            if [ "${{ steps.check_tag.outputs.failed_count }}" = "0" ]; then
-              if [ "${{ steps.check_tag.outputs.created_count }}" = "0" ]; then
-                echo "✅ Tag ${{ steps.commit.outputs.tag_name }} already existed - no action needed"
-              else
-                echo "✅ Successfully created tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
-              fi
-            else
-              echo "❌ Failed to create tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
-            fi
-
-            echo ""
-            echo "Tag details:"
-            echo "  Name: ${{ steps.commit.outputs.tag_name }}"
-            echo "  Commit: ${{ steps.commit.outputs.sha }}"
-            echo "  Trigger: ${{ github.event_name }}"
-            if [ -n "${{ github.event.inputs.commit_sha }}" ]; then
-              echo "  Manual commit: ${{ github.event.inputs.commit_sha }}"
-            fi
+            echo "❌ Failed to create tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
+          fi
+
+          echo ""
+          echo "Tag details:"
+          echo "  Name: ${{ steps.commit.outputs.tag_name }}"
+          echo "  Commit: ${{ steps.commit.outputs.sha }}"
+          echo "  Trigger: ${{ github.event_name }}"
+          if [ -n "${{ github.event.inputs.commit_sha }}" ]; then
+            echo "  Manual commit: ${{ github.event.inputs.commit_sha }}"
           fi

.github/workflows/trunk.yml

@@ -190,40 +190,6 @@ jobs:
       runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
     secrets: inherit
 
-  linux-jammy-rocm-py3_10-build:
-    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
-      sync-tag: rocm-build
-      test-matrix: |
-        { include: [
-          { config: "default", shard: 1, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-          { config: "default", shard: 2, num_shards: 2, runner: "linux.rocm.gpu.gfx942.1" },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/trunk') }}
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-      tests-to-include: "test_nn test_torch test_cuda test_ops test_unary_ufuncs test_binary_ufuncs test_autograd inductor/test_torchinductor"
-    secrets: inherit
-
   inductor-build:
     name: inductor-build
     uses: ./.github/workflows/_linux-build.yml
@@ -234,23 +200,6 @@ jobs:
       cuda-arch-list: '8.0'
     secrets: inherit
 
-  # Test cross-compiled models with Windows libs extracted from wheel
-  cross-compile-linux-test:
-    name: cross-compile-linux-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs:
-      - linux-jammy-cuda12_8-py3_10-gcc11-build
-      - get-label-type
-      - win-vs2022-cuda12_8-py3-build
-    with:
-      build-environment: linux-jammy-cuda12.8-py3.10-gcc11
-      docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
-      test-matrix: |
-        { include: [
-          { config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
-        ]}
-    secrets: inherit
-
   verify-cachebench-cpu-build:
     name: verify-cachebench-cpu-build
     uses: ./.github/workflows/_linux-build.yml

.gitignore

@@ -374,7 +374,6 @@ third_party/ruy/
 third_party/glog/
 
 # Virtualenv
-.venv/
 venv/
 
 # Log files

.lintrunner.toml

@@ -209,46 +209,6 @@ command = [
     '@{{PATHSFILE}}'
 ]
 
-
-[[linter]]
-code = 'PYREFLY'
-include_patterns = [
-    'torch/**/*.py',
-    'torch/**/*.pyi',
-    'torchgen/**/*.py',
-    'torchgen/**/*.pyi',
-    'functorch/**/*.py',
-    'functorch/**/*.pyi',
-]
-exclude_patterns = []
-command = [
-    'python3',
-    'tools/linter/adapters/pyrefly_linter.py',
-    '--config=pyrefly.toml',
-]
-init_command = [
-    'python3',
-    'tools/linter/adapters/pip_init.py',
-    '--dry-run={{DRYRUN}}',
-    'numpy==2.1.0 ; python_version >= "3.12"',
-    'expecttest==0.3.0',
-    'pyrefly==0.36.2',
-    'sympy==1.13.3',
-    'types-requests==2.27.25',
-    'types-pyyaml==6.0.2',
-    'types-tabulate==0.8.8',
-    'types-protobuf==5.29.1.20250403',
-    'types-setuptools==79.0.0.20250422',
-    'types-jinja2==2.11.9',
-    'types-colorama==0.4.6',
-    'filelock==3.18.0',
-    'junitparser==2.1.1',
-    'rich==14.1.0',
-    'optree==0.17.0',
-    'types-openpyxl==3.1.5.20250919',
-    'types-python-dateutil==2.9.0.20251008'
-]
-
 [[linter]]
 code = 'CLANGTIDY'
 include_patterns = [
@@ -833,7 +793,8 @@ exclude_patterns = [
 command = [
     'python3',
     'tools/linter/adapters/grep_linter.py',
-    '--pattern=(cudaSetDevice|cudaGetDevice)\\(',
+    '--pattern=cudaSetDevice(',
+    '--pattern=cudaGetDevice(',
     '--linter-name=RAWCUDADEVICE',
     '--error-name=raw CUDA API usage',
     """--error-description=\
@@ -1137,8 +1098,11 @@ command = [
 [[linter]]
 code = 'WORKFLOWSYNC'
 include_patterns = [
-    '.github/workflows/*.yml',
-    '.github/workflows/*.yaml',
+    '.github/workflows/pull.yml',
+    '.github/workflows/trunk.yml',
+    '.github/workflows/periodic.yml',
+    '.github/workflows/mac-mps.yml',
+    '.github/workflows/slow.yml',
 ]
 command = [
     'python3',

CODEOWNERS

@@ -201,17 +201,3 @@ torch/backends/cudnn/ @eqy @syed-ahmed @Aidyn-A
 /torch/csrc/stable/ @janeyx99 @mikaylagawarecki
 /torch/headeronly/ @janeyx99
 /torch/header_only_apis.txt @janeyx99
-
-# FlexAttention
-/torch/nn/attention/flex_attention.py @drisspg
-/torch/_higher_order_ops/flex_attention.py @drisspg
-/torch/_inductor/kernel/flex/ @drisspg
-/torch/_inductor/codegen/cpp_flex_attention_template.py @drisspg
-/test/inductor/test_flex_attention.py @drisspg
-/test/inductor/test_flex_decoding.py @drisspg
-
-# Low Precision GEMMs
-/aten/src/ATen/native/cuda/Blas.cpp @drisspg @slayton58
-/aten/src/ATen/cuda/CUDABlas.cpp @drisspg @slayton58
-/aten/src/ATen/cuda/CUDABlas.h @drisspg @slayton58
-/test/test_scaled_matmul_cuda.py @drisspg @slayton58

SECURITY.md

@@ -31,9 +31,9 @@ Be careful when running untrusted models. This classification includes models cr
 
 **Prefer to execute untrusted models within a secure, isolated environment such as a sandbox** (e.g., containers, virtual machines). This helps protect your system from potentially malicious code. You can find further details and instructions in [this page](https://developers.google.com/code-sandboxing).
 
-**Be mindful of risky model formats**. Give preference to share and load weights with the appropriate format for your use case. [safetensors](https://huggingface.co/docs/safetensors/en/index) gives the most safety but is the most restricted in what it supports. [`torch.load`](https://pytorch.org/docs/stable/generated/torch.load.html#torch.load) has a significantly larger surface of attack but is more flexible in what it can serialize. See the documentation for more details.
+**Be mindful of risky model formats**. Give preference to share and load weights with the appropriate format for your use case. [safetensors](https://huggingface.co/docs/safetensors/en/index) gives the most safety but is the most restricted in what it supports. [`torch.load`](https://pytorch.org/docs/stable/generated/torch.load.html#torch.load) with `weights_only=True` is also secure to our knowledge even though it offers significantly larger surface of attack. Loading un-trusted checkpoint with `weights_only=False` MUST never be done.
+
 
-Even for more secure serialization formats, unexpected inputs to the downstream system can cause diverse security threats (e.g. denial of service, out of bound reads/writes) and thus we recommend extensive validation of any untrusted inputs.
 
 Important Note: The trustworthiness of a model is not binary. You must always determine the proper level of caution depending on the specific model and how it matches your use case and risk tolerance.
 

aten/src/ATen/CMakeLists.txt

@@ -38,7 +38,7 @@ set_bool(AT_HIPSPARSELT_ENABLED CAFFE2_USE_HIPSPARSELT)
 
 configure_file(Config.h.in "${CMAKE_CURRENT_SOURCE_DIR}/Config.h")
 # TODO: Do not generate CUDAConfig.h for ROCm BUILDS
-# At the moment, `jit_macros.h` include CUDAConfig.h for both CUDA and HIP builds
+# At the moment, `jit_macors.h` include CUDAConfig.h for both CUDA and HIP builds
 if(USE_CUDA OR USE_ROCM)
   configure_file(cuda/CUDAConfig.h.in "${CMAKE_CURRENT_SOURCE_DIR}/cuda/CUDAConfig.h")
 endif()
@@ -256,7 +256,6 @@ endif()
 IF(USE_FBGEMM_GENAI)
   set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
   set(FBGEMM_GENAI_SRCS ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
-
   if(USE_CUDA)
     # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
     # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
@@ -289,70 +288,62 @@ IF(USE_FBGEMM_GENAI)
 
     set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
-    set(fbgemm_genai_cuh
+    set(fbgemm_genai_mx8mx8bf16_grouped
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
-      "${FBGEMM_GENAI_SRCS}/"
     )
 
-    target_include_directories(fbgemm_genai PRIVATE
+    target_include_directories(fbgemm_genai PUBLIC
       ${FBGEMM_THIRD_PARTY}/cutlass/include
       ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-      ${fbgemm_genai_cuh}
+      ${fbgemm_genai_mx8mx8bf16_grouped}
       ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
       ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
     )
+  else()
+    if(USE_ROCM)
+      # Only include the kernels we want to build to avoid increasing binary size.
+      file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
+        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
+        "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
+      set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
 
-    # Add FBGEMM_GENAI include directories for torch_ops.h
-    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
-    list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
-  elseif(USE_ROCM)
-    # Only include the kernels we want to build to avoid increasing binary size.
-    file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
-      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
-      "${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
-    set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
+      # Add additional HIPCC compiler flags for performance
+      set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
+        -mllvm
+        -amdgpu-coerce-illegal-types=1
+        -mllvm
+        -enable-post-misched=0
+        -mllvm
+        -greedy-reverse-local-assignment=1
+        -fhip-new-launch-api)
 
-    # Add additional HIPCC compiler flags for performance
-    set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-      -mllvm
-      -enable-post-misched=0
-      -mllvm
-      -greedy-reverse-local-assignment=1
-      -fhip-new-launch-api)
-    if(DEFINED ROCM_VERSION_DEV AND ROCM_VERSION_DEV VERSION_LESS "7.2.0")
-        list(PREPEND FBGEMM_GENAI_EXTRA_HIPCC_FLAGS -mllvm -amdgpu-coerce-illegal-types=1)
+      # Only compile for gfx942 for now.
+      # This is rather hacky, I could not figure out a clean solution :(
+      set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
+      string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
+      if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
+        list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
       endif()
+      set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
 
-    # Only compile for gfx942 for now.
-    # This is rather hacky, I could not figure out a clean solution :(
-    set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
-    string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
-    if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
-      list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
+      hip_add_library(
+        fbgemm_genai STATIC
+        ${fbgemm_genai_native_rocm_hip}
+        HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
+      set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
+      set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
+      target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
+
+      target_include_directories(fbgemm_genai PUBLIC
+        # FBGEMM version of Composable Kernel is used due to some customizations
+        ${FBGEMM_THIRD_PARTY}/composable_kernel/include
+        ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
+        ${FBGEMM_THIRD_PARTY}/cutlass/include
+        ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
+        ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
+        ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
+      )
     endif()
-    set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
-
-    hip_add_library(
-      fbgemm_genai STATIC
-      ${fbgemm_genai_native_rocm_hip}
-      HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
-    set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
-    set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
-    target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
-
-    target_include_directories(fbgemm_genai PRIVATE
-      # FBGEMM version of Composable Kernel is used due to some customizations
-      ${FBGEMM_THIRD_PARTY}/composable_kernel/include
-      ${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
-      ${FBGEMM_THIRD_PARTY}/cutlass/include
-      ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-      ${FBGEMM_GENAI_SRCS}/common/include/   # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
-      ${FBGEMM_GENAI_SRCS}/include/          # includes fbgemm_gpu/torch_ops.h
-    )
-
-    # Add FBGEMM_GENAI include directories for torch_ops.h
-    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
-    list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
   endif()
 endif()
 
@@ -701,6 +692,12 @@ if(USE_CUDA AND NOT USE_ROCM)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
   list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
 
+  # Add FBGEMM_GENAI include directories for torch_ops.h
+  if(USE_FBGEMM_GENAI)
+    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
+    list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
+  endif()
+
   if($ENV{ATEN_STATIC_CUDA})
     if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
       list(APPEND ATen_CUDA_DEPENDENCY_LIBS

aten/src/ATen/Context.h

@@ -19,7 +19,6 @@
 #include <ATen/detail/MPSHooksInterface.h>
 #include <ATen/detail/MTIAHooksInterface.h>
 #include <ATen/detail/PrivateUse1HooksInterface.h>
-#include <ATen/detail/XLAHooksInterface.h>
 #include <ATen/detail/XPUHooksInterface.h>
 #include <c10/core/QEngine.h>
 #include <c10/core/impl/DeviceGuardImplInterface.h>
@@ -89,8 +88,6 @@ class TORCH_API Context {
       return at::detail::getHIPHooks();
     } else if (opt_device_type == at::kHPU) {
       return at::detail::getHPUHooks();
-    } else if (opt_device_type == at::kXLA) {
-      return at::detail::getXLAHooks();
     } else {
       TORCH_CHECK(
           false,
@@ -199,7 +196,7 @@ class TORCH_API Context {
     return c10::impl::hasDeviceGuardImpl(c10::DeviceType::IPU);
   }
   static bool hasXLA() {
-    return detail::getXLAHooks().hasXLA();
+    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::XLA);
   }
   static bool hasXPU() {
     return detail::getXPUHooks().hasXPU();

aten/src/ATen/FunctionalTensorWrapper.cpp

@@ -122,7 +122,7 @@ void FunctionalTensorWrapper::freeze_storage() const {
 //          |   have their own storages, but backends like functorch      |
 //         \/   are allowed to re-alias underneath the pass               \/
 // . - - - - - - - - - - - - - .                             . - - - - - - - - - - - - - - - .
-// |    underlying_storage     |                             |      underlying_storage       |
+// |    underyling_storage     |                             |      underyling_storage       |
 // . - - - - - - - - - - - - - .                             . - - - - - - - - - - - - - - - .
 //
 // This constructor is only used by view ops.

aten/src/ATen/TensorIterator.cpp

@@ -1534,7 +1534,7 @@ void TensorIteratorBase::build(TensorIteratorConfig& config) {
 
   // XLA and lazy tensors don't have storage, so they don't have an underlying data pointer.
   // Nothing beyond this point is important for meta functions, so it's fine to exit early here.
-  // Extend the condition to MAIA tensors as MAIA tensors also don't have storage.
+  // Extend the condition to MAIA tesnors as MAIA tensors also don't have storage.
   if (privateuse1_without_storage  ||
       common_device_.type() == DeviceType::XLA  ||
       common_device_.type() == DeviceType::IPU  ||

aten/src/ATen/core/CachingHostAllocator.h

@@ -39,7 +39,7 @@ struct HostBlock {
 };
 
 template <typename B>
-struct alignas(hardware_destructive_interference_size) FreeBlockList {
+struct alignas(64) FreeBlockList {
   std::mutex mutex_;
   std::deque<B*> list_;
 };
@@ -94,11 +94,11 @@ struct PinnedReserveSegment {
 struct TORCH_API HostStats {
   // COUNT: total allocations (active)
   Stat active_requests;
-  // SUM: bytes allocated/reserved by this memory allocator. (active)
+  // SUM: bytes allocated/reserved by this memory alocator. (active)
   Stat active_bytes;
   // COUNT: total allocations (active + free)
   Stat allocations;
-  // SUM: bytes allocated/reserved by this memory allocator. This accounts
+  // SUM: bytes allocated/reserved by this memory alocator. This accounts
   // for both free and in-use blocks.
   Stat allocated_bytes;
 
@@ -122,12 +122,12 @@ struct TORCH_API HostStats {
 // Struct containing memory allocator summary statistics for host, as they
 // are staged for reporting. This is a temporary struct that is used to
 // avoid locking the allocator while collecting stats.
-struct alignas(hardware_destructive_interference_size) HostStatsStaged {
+struct alignas(64) HostStatsStaged {
   std::mutex timing_mutex_;
   // COUNT: total allocations (active + free)
   // LOCK: access to this stat is protected by the allocator's blocks_mutex_
   Stat allocations;
-  // SUM: bytes allocated/reserved by this memory allocator. This accounts
+  // SUM: bytes allocated/reserved by this memory alocator. This accounts
   // for both free and in-use blocks.
   Stat allocated_bytes;
   // COUNT: number of allocations per bucket (active)
@@ -455,7 +455,7 @@ struct CachingHostAllocatorImpl {
   }
 
   void resetAccumulatedStats() {
-    // Resetting accumulated memory stats requires concurrently holding both the
+    // Reseting accumulated memory stats requires concurrently holding both the
     // free list mutexes and the blocks mutex. Previously, this was only done in
     // empty_cache function.
     for (size_t i = 0; i < free_list_.size(); ++i) {
@@ -482,7 +482,7 @@ struct CachingHostAllocatorImpl {
   }
 
   void resetPeakStats() {
-    // Resetting peak memory stats requires concurrently holding both the
+    // Reseting peak memory stats requires concurrently holding both the
     // free list mutexes and the blocks mutex. Previously, this was only done in
     // empty_cache function.
     for (size_t i = 0; i < free_list_.size(); ++i) {
@@ -669,7 +669,7 @@ struct CachingHostAllocatorImpl {
     TORCH_CHECK_NOT_IMPLEMENTED(false, "Not implemented for query_event");
   }
 
-  alignas(hardware_destructive_interference_size) std::mutex blocks_mutex_;
+  alignas(64) std::mutex blocks_mutex_;
   ska::flat_hash_set<B*> blocks_; // block list
   ska::flat_hash_map<void*, B*> ptr_to_block_;
 
@@ -677,17 +677,17 @@ struct CachingHostAllocatorImpl {
   // size. This allows us to quickly find a free block of the right size.
   // We use deque to store per size free list and guard the list with its own
   // mutex.
-  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>> free_list_ =
+  alignas(64) std::vector<FreeBlockList<B>> free_list_ =
       std::vector<FreeBlockList<B>>(MAX_SIZE_INDEX);
 
-  alignas(hardware_destructive_interference_size) std::mutex events_mutex_;
+  alignas(64) std::mutex events_mutex_;
   std::deque<std::pair<E, B*>> events_; // event queue paired with block
 
   // Indicates whether the object is active.
   // Set to false in the destructor to signal background threads to stop.
   std::atomic<bool> active_{true};
 protected:
-  alignas(hardware_destructive_interference_size) HostStatsStaged stats_;
+  alignas(64) HostStatsStaged stats_;
 };
 
 struct TORCH_API HostAllocator : public at::Allocator {

aten/src/ATen/core/Generator.h

@@ -59,7 +59,9 @@ struct TORCH_API Generator {
 
   explicit Generator(c10::intrusive_ptr<c10::GeneratorImpl> gen_impl)
    : impl_(std::move(gen_impl)) {
-    TORCH_CHECK(impl_.get(), "GeneratorImpl with nullptr is not supported");
+    if (impl_.get() == nullptr) {
+      throw std::runtime_error("GeneratorImpl with nullptr is not supported");
+    }
   }
 
   bool operator==(const Generator& rhs) const {

aten/src/ATen/core/PhiloxRNGEngine.h

@@ -229,10 +229,10 @@ private:
   }
 
 
-  static constexpr uint32_t kPhilox10A = 0x9E3779B9;
-  static constexpr uint32_t kPhilox10B = 0xBB67AE85;
-  static constexpr uint32_t kPhiloxSA = 0xD2511F53;
-  static constexpr uint32_t kPhiloxSB = 0xCD9E8D57;
+  static const uint32_t kPhilox10A = 0x9E3779B9;
+  static const uint32_t kPhilox10B = 0xBB67AE85;
+  static const uint32_t kPhiloxSA = 0xD2511F53;
+  static const uint32_t kPhiloxSB = 0xCD9E8D57;
 };
 
 typedef philox_engine Philox4_32;

aten/src/ATen/core/TensorBase.h

@@ -111,7 +111,9 @@ class TORCH_API TensorBase {
   explicit TensorBase(
       c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> tensor_impl)
       : impl_(std::move(tensor_impl)) {
-    TORCH_CHECK(impl_.get(), "TensorImpl with nullptr is not supported");
+    if (impl_.get() == nullptr) {
+      throw std::runtime_error("TensorImpl with nullptr is not supported");
+    }
   }
   TensorBase(const TensorBase&) = default;
   TensorBase(TensorBase&&) noexcept = default;

aten/src/ATen/core/VariableFallbackKernel.cpp

@@ -109,10 +109,6 @@ TORCH_LIBRARY_IMPL(_, AutogradHPU, m) {
   m.fallback(AUTOGRAD_FALLBACK);
 }
 
-TORCH_LIBRARY_IMPL(_, AutogradPrivateUse1, m) {
-  m.fallback(AUTOGRAD_FALLBACK);
-}
-
 #undef AUTOGRAD_FALLBACK
 
 } // namespace

aten/src/ATen/core/class_type.h

@@ -148,7 +148,7 @@ struct TORCH_API ClassType : public NamedType {
 
   void checkNotExist(const std::string& name, const std::string& what) const;
 
-  // Attributes are stored in a specific slot at runtime for efficiency.
+  // Attributes are stored in a specific slot at runtime for effiency.
   // When emitting instructions we specify the slot so that attribute access is
   // a constant lookup
   std::optional<size_t> findAttributeSlot(const std::string& name) const {
@@ -412,7 +412,7 @@ struct TORCH_API ClassType : public NamedType {
   // Holds method attributes
   std::weak_ptr<CompilationUnit> compilation_unit_;
 
-  // Holds all attributes, attribute details are found on ClassAttribute
+  // Holds all atrributes, attribute details are found on ClassAttribute
   std::vector<ClassAttribute> attributes_;
   // Construct mirroring attributes_, only around due to the fact that `containedTypes()` method returns an ArrayRef.
   // Never fill this without using the appropriate provideNewClassAttribute method

aten/src/ATen/core/dispatch/Dispatcher.cpp

@@ -442,17 +442,11 @@ RegistrationHandleRAII Dispatcher::registerFallback(DispatchKey dispatchKey, Ker
 
   auto idx = getDispatchTableIndexForDispatchKey(dispatchKey);
   TORCH_CHECK(idx >= 0 && static_cast<uint64_t>(idx) < backendFallbackKernels_.size(), "idx=", idx);
-  // NB: Perserve BC for registering fallback for AutogradPrivateUse1 multiple time,
-  // refer to https://github.com/pytorch/pytorch/issues/163979 for more informations.
   TORCH_CHECK(
-      dispatchKey == DispatchKey::AutogradPrivateUse1 ||
-          !backendFallbackKernels_[idx].kernel.isValid(),
-      "Tried to register multiple backend fallbacks for the same dispatch key ",
-      dispatchKey,
-      "; previous registration ",
-      backendFallbackKernels_[idx].debug,
-      ", new registration ",
-      debug);
+    !backendFallbackKernels_[idx].kernel.isValid(),
+    "Tried to register multiple backend fallbacks for the same dispatch key ", dispatchKey, "; previous registration ",
+    backendFallbackKernels_[idx].debug, ", new registration ", debug
+  );
   // NB: inferred function schema is always nullptr for fallbacks, as fallbacks
   // cannot be unboxed
   backendFallbackKernels_[idx] = impl::AnnotatedKernel(std::move(kernel), nullptr, std::move(debug));
@@ -537,7 +531,7 @@ int64_t Dispatcher::sequenceNumberForRunningRecordFunction(DispatchKey dispatchK
 
   // Note: this records a sequence number for both Autograd keys, and for
   // non-Autograd keys where the dispatchKeySet still contains an autograd key.
-  // This means that we might collect the same sequence number two different
+  // This means that we might collect the same sequence nubmer two different
   // events if they all occurred above Autograd and still had the Autograd
   // dispatch key in the dispatch key set.
   // However, this usually doesn't happen: normally the first call will

aten/src/ATen/core/dispatch/Dispatcher.h

@@ -585,7 +585,7 @@ class TORCH_API OperatorHandle {
 
   // We need to store this iterator in order to make
   // Dispatcher::cleanup() fast -- it runs a lot on program
-  // termination (and presumably library unloading).
+  // termination (and presuambly library unloading).
   std::list<Dispatcher::OperatorDef>::iterator operatorIterator_;
 };
 

aten/src/ATen/core/dispatch/OperatorEntry.cpp

@@ -365,7 +365,7 @@ std::pair<const AnnotatedKernel&, const char*> OperatorEntry::computeDispatchTab
   //          For autograd keys, we only use kernel from CompositeImplicitAutograd when there's no direct registration
   //          to its corresponding backend key or CompositeExplicitAutograd. See Note [CompositeExplicitAutograd and CompositeImplicitAutograd].
   //          For AutogradOther, we eagerly return ambiguousAutogradOtherKernel() if there's registration to any of
-  //          its backends and ask backend extender to request a dedicated Autograd key for the backend.
+  //          its backends and ask backend extender to request a decicated Autograd key for the backend.
   //          See Note [Ambiguity in AutogradOther kernel] for more details.
   //          A CompositeExplicitAutograd kernel prevents CompositeImplicitAutograd kernel being used for Autograd keys, but it doesn't
   //          cause confusion for AutogradOther. It's pretty straightforward to use Autograd (if available)

aten/src/ATen/core/function_schema.cpp

@@ -261,7 +261,7 @@ std::ostream& operator<<(std::ostream& out, const FunctionSchema& schema) {
     //
     // There are 2 cases
     // 1. something like 'aten::items.str(Dict(str, t) self) -> ((str, t)[])'.
-    // without the extra parenthesis, the c++ scheme parser can not parse it.
+    // without the extra parenthesis, the c++ schem parser can not parse it.
     // 2. something like '-> ((str, str))'. Need extra parenthesis so the return
     // type is a single tuple rather than two strings.
     // PR (https://github.com/pytorch/pytorch/pull/23204) has more context about

aten/src/ATen/core/interned_strings.cpp

@@ -68,7 +68,11 @@ Symbol InternedStrings::_symbol(const std::string& s) {
     return it->second;
 
   auto pos = s.find("::");
-  TORCH_CHECK(pos != std::string::npos, "all symbols must have a namespace, <namespace>::<string>, but found: ", s);
+  if (pos == std::string::npos) {
+    std::stringstream ss;
+    ss << "all symbols must have a namespace, <namespace>::<string>, but found: " << s;
+    throw std::runtime_error(ss.str());
+  }
   Symbol ns = _symbol("namespaces::" + s.substr(0, pos));
 
   Symbol sym(sym_to_info_.size());
@@ -117,7 +121,12 @@ std::string Symbol::domainString() const {
 }
 
 Symbol Symbol::fromDomainAndUnqualString(const std::string & d, const std::string & s) {
-  TORCH_CHECK(d.compare(0, domain_prefix().size(), domain_prefix()) == 0, "Symbol: domain string is expected to be prefixed with '", domain_prefix(), "', e.g. 'org.pytorch.aten'");
+  if (d.compare(0, domain_prefix().size(), domain_prefix()) != 0) {
+    std::ostringstream ss;
+    ss << "Symbol: domain string is expected to be prefixed with '"
+       << domain_prefix() << "', e.g. 'org.pytorch.aten'";
+    throw std::runtime_error(ss.str());
+  }
   std::string qualString = d.substr(domain_prefix().size()) + "::" + s;
   return fromQualString(qualString);
 }

aten/src/ATen/core/ivalue.cpp

@@ -7,7 +7,6 @@
 #include <ATen/core/jit_type.h>
 #include <ATen/core/stack.h>
 #include <ATen/core/type_factory.h>
-#include <c10/util/Exception.h>
 #include <c10/util/StringUtil.h>
 #include <c10/util/hash.h>
 #include <c10/util/irange.h>
@@ -413,7 +412,7 @@ size_t IValue::hash(const IValue& v) {
     case Tag::Enum:
     case Tag::Stream:
     case Tag::Uninitialized:
-      TORCH_CHECK(false,
+      throw std::runtime_error(
           "unhashable type: '" + v.type()->repr_str() + "'");
   }
   // the above switch should be exhaustive

aten/src/ATen/core/ivalue.h

@@ -1176,7 +1176,7 @@ struct TORCH_API IValue final {
   using HashIdentityIValueMap =
       std::unordered_map<IValue, IValue, HashIdentityIValue, CompIdentityIValues>;
 
-  // Checks if this and rhs has a subvalues in common.
+  // Chechs if this and rhs has a subvalues in common.
   // [t1,t2] and [t2, t3] returns true.
   bool overlaps(const IValue& rhs) const;
 

aten/src/ATen/core/ivalue_inl.h

@@ -1501,7 +1501,7 @@ struct C10_EXPORT ivalue::Object final : c10::intrusive_ptr_target {
   // However, the CompilationUnit holds ownership of the type's graphs, so
   // inserting a constant object into a Graph would create a reference cycle if
   // that constant object held a shared_ptr to its CU. For these objects we
-  // instantiate them with non-owning references to its CU
+  // instatiate them with non-owning references to its CU
   Object(WeakOrStrongTypePtr type, size_t numSlots) : type_(std::move(type)) {
     slots_.resize(numSlots);
   }

aten/src/ATen/core/jit_type.h

@@ -8,7 +8,6 @@
 #include <ATen/core/type_factory.h>
 #include <ATen/core/qualified_name.h>
 #include <c10/util/TypeList.h>
-#include <c10/util/Exception.h>
 #include <optional>
 #include <c10/core/SymFloat.h>
 #include <c10/core/SymBool.h>
@@ -117,8 +116,10 @@ struct SingleElementType : public SharedType {
 
  protected:
   SingleElementType(TypePtr elem) : SharedType(Kind), elem(std::move(elem)) {
-    TORCH_CHECK(this->elem, c10::str(
+    if (!this->elem) {
+      throw std::runtime_error(c10::str(
             "Can not create ", typeKindToString(Kind), " with None type"));
+    }
   }
 
  private:
@@ -373,7 +374,7 @@ struct TORCH_API SymbolicShape {
   // Unranked shape constructor.
   SymbolicShape() : dims_(std::nullopt) {}
 
-  // Known rank but unknown dimensions.
+  // Known rank but unknown dimentions.
   SymbolicShape(std::optional<size_t> rank) : dims_(std::nullopt) {
     if(!rank) {
       return;
@@ -415,12 +416,16 @@ struct TORCH_API SymbolicShape {
   }
 
   ShapeSymbol operator[](size_t i) const {
-    TORCH_CHECK(dims_, "Rank isn't fixed");
+    if (!dims_) {
+      throw std::runtime_error("Rank isn't fixed");
+    }
     return (*dims_).at(i);
   }
 
   ShapeSymbol at(size_t i) const {
-    TORCH_CHECK(dims_, "Rank isn't fixed");
+    if (!dims_) {
+      throw std::runtime_error("Rank isn't fixed");
+    }
     return (*dims_).at(i);
   }
 
@@ -515,7 +520,9 @@ struct VaryingShape {
   }
 
   const std::optional<T> &operator[](size_t i) const {
-    TORCH_CHECK(dims_, "Rank isn't fixed");
+    if (!dims_) {
+      throw std::runtime_error("Rank isn't fixed");
+    }
     return (*dims_).at(i);
   }
 
@@ -884,9 +891,9 @@ struct TORCH_API ListType
 
   // global singleton
   // Given an inner type T and an identifier,
-  // this function will return the global singleton type pointer
+  // this function wil return the global singleton type pointer
   // the type List<T>.
-  // The extra "identifier" argument is needed because we have multiple container types
+  // The extra "identifier" argument is needed beccause we have multiple container types
   // that all re-use this function (List<T>, array<T, N>, etc.)
   static TypePtr get(const std::string& identifier, TypePtr inner);
 
@@ -950,7 +957,9 @@ struct TORCH_API DictType : public SharedType {
 
   TypePtr createWithContained(
       std::vector<TypePtr> contained_types) const override {
-    TORCH_CHECK(contained_types.size() == 2, "Expected 2 contained types");
+    if (contained_types.size() != 2) {
+      throw std::runtime_error("Expected 2 contained types");
+    }
     return create(std::move(contained_types.at(0)), std::move(contained_types.at(1)));
   }
 

aten/src/ATen/core/jit_type_base.h

@@ -185,11 +185,11 @@ struct TORCH_API Type {
         : repr_(nullptr) {}
 
     /* implicit */ SingletonOrSharedTypePtr(SingletonTypePtr<T> p)
-        : repr_(makeSingletonSharedPtr(p.get())) {}
+        : repr_(p) {}
 
     template <typename U, std::enable_if_t<std::is_convertible_v<U*, T*>, bool> = true>
     /* implicit */ SingletonOrSharedTypePtr(SingletonTypePtr<U> p)
-        : repr_(makeSingletonSharedPtr(static_cast<T*>(p.get()))) {}
+        : repr_(SingletonTypePtr<T>(p.get())) {}
 
 
     // We need to support construction from T* for pybind. The problem
@@ -202,8 +202,8 @@ struct TORCH_API Type {
     // Case 2: if T is exactly Type, we need to do a dynamic_cast to
     // check if it's a SharedType and do the right thing.
     //
-    // Case 3: Otherwise, T is not a SharedType. Use a singleton
-    // pointer.
+    // Case 3: Otherwise, T is not a SharedType. (debug-check this
+    // assumption!) Use a singleton pointer.
 
     template <typename U = T, std::enable_if_t<std::is_base_of_v<SharedType, U>, bool> = true>
     /* implicit */ SingletonOrSharedTypePtr(T* p) : SingletonOrSharedTypePtr(static_cast<typename detail::as_shared_type<U>::type>(p)->shared_from_this()) {}
@@ -211,15 +211,15 @@ struct TORCH_API Type {
     template <typename U = T, std::enable_if_t<std::is_same_v<Type, U>, bool> = true>
     /* implicit */ SingletonOrSharedTypePtr(T* p) {
       if (auto* shared_p = dynamic_cast<typename detail::as_shared_type<U>::type>(p)) {
-        repr_ = shared_p->shared_from_this();
+        repr_ = Repr(shared_p->shared_from_this());
       } else {
-        repr_ = makeSingletonSharedPtr(p);
+        repr_ = Repr(p);
       }
     }
 
     template <typename U = T, std::enable_if_t<!std::is_same_v<Type, U> && !std::is_base_of_v<SharedType, U>, bool> = true>
     /* implicit */ SingletonOrSharedTypePtr(T* p)
-        : repr_(makeSingletonSharedPtr(p)) {
+        : repr_(p) {
       TORCH_INTERNAL_ASSERT_DEBUG_ONLY(dynamic_cast<typename detail::as_shared_type<U>::type>(p) == nullptr);
     }
 
@@ -230,19 +230,19 @@ struct TORCH_API Type {
     ~SingletonOrSharedTypePtr() = default;
 
     T* get() const {
-      return repr_.get();
+      return repr_.isSharedAndNonNull() ? repr_.shared_.repr_.get() : static_cast<T*>(repr_.rawRepr().first);
     }
 
     operator bool() const {
-      return repr_ != nullptr;
+      return repr_.isNonNull();
     }
 
     bool operator==(std::nullptr_t) const {
-      return repr_ == nullptr;
+      return !repr_.isNonNull();
     }
 
     bool operator!=(std::nullptr_t) const {
-      return repr_ != nullptr;
+      return repr_.isNonNull();
     }
 
     template <typename U = T, std::enable_if_t<!std::is_same_v<std::remove_const_t<U>, void>, bool> = true>
@@ -255,14 +255,138 @@ struct TORCH_API Type {
     }
 
   private:
-    // Use shared_ptr's aliasing constructor to create a non-owning pointer
-    // to a singleton. The lifetime is tied to the null shared_ptr, so there's
-    // no reference counting overhead for the singleton itself.
-    static std::shared_ptr<T> makeSingletonSharedPtr(T* ptr) {
-      return std::shared_ptr<T>(std::shared_ptr<T>(), ptr);
-    }
+    // NOTE: SharedPtrWrapper exists to work around a baffling bug in
+    // nvcc; see comment in destroy() below.
+    struct SharedPtrWrapper {
+      SharedPtrWrapper(std::shared_ptr<T> &&x)
+          : repr_(std::move(x)) {}
+      std::shared_ptr<T> repr_;
+    };
+    union Repr {
+      Repr() : Repr(nullptr) {}
 
-    std::shared_ptr<T> repr_;
+      explicit Repr(std::shared_ptr<T> x)
+          : shared_(std::move(x)) {}
+
+      explicit Repr(std::nullptr_t)
+          : singletonRepr_(nullptr) {}
+
+      explicit Repr(SingletonTypePtr<T> p)
+          : singletonRepr_(p.get()) {}
+
+      ~Repr() {
+        destroy();
+      }
+
+      // NOTE: the only non-UB way to access our null state is through
+      // rawRepr(), because our copy operation doesn't preserve which
+      // union member is active for null pointers.
+      Repr(const Repr& rhs) {
+        if (rhs.isSharedAndNonNull()) {
+          new (&shared_) SharedPtrWrapper(rhs.shared_);
+        } else {
+          singletonRepr_.singleton_ = static_cast<T*>(rhs.rawRepr().first);
+          TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.singletonRepr_.unused_ == nullptr);
+          singletonRepr_.unused_ = nullptr;
+        }
+      }
+
+      Repr(Repr&& rhs) noexcept {
+        if (rhs.isSharedAndNonNull()) {
+          new (&shared_) SharedPtrWrapper(std::move(rhs.shared_));
+        } else {
+          singletonRepr_.singleton_ = static_cast<T*>(rhs.rawRepr().first);
+          TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.singletonRepr_.unused_ == nullptr);
+          singletonRepr_.unused_ = nullptr;
+        }
+      }
+
+      Repr& operator=(const Repr& rhs) {
+        if (&rhs == this) {
+          return *this;
+        }
+        if (rhs.isSharedAndNonNull()) {
+          if (isSharedAndNonNull()) {
+            shared_ = rhs.shared_;
+          } else {
+            new (&shared_) SharedPtrWrapper(rhs.shared_);
+          }
+        } else {
+          if (isSharedAndNonNull()) {
+            destroy();
+          }
+          singletonRepr_.singleton_ = static_cast<T*>(rhs.rawRepr().first);
+          TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.rawRepr().nullIfSingleton_ == nullptr);
+          singletonRepr_.unused_ = nullptr;
+        }
+        return *this;
+      }
+
+      Repr& operator=(Repr&& rhs) noexcept {
+        if (&rhs == this) {
+          return *this;
+        }
+        if (rhs.isSharedAndNonNull()) {
+          if (isSharedAndNonNull()) {
+            shared_ = std::move(rhs.shared_);
+          } else {
+            new (&shared_) SharedPtrWrapper(std::move(rhs.shared_));
+          }
+        } else {
+          if (isSharedAndNonNull()) {
+            destroy();
+          }
+          singletonRepr_.singleton_ = static_cast<T*>(rhs.rawRepr().first);
+          TORCH_INTERNAL_ASSERT_DEBUG_ONLY(rhs.rawRepr().nullIfSingleton_ == nullptr);
+          singletonRepr_.unused_ = nullptr;
+        }
+        return *this;
+      }
+
+      SharedPtrWrapper shared_;
+
+      struct SingletonRepr {
+        explicit SingletonRepr(T* s) : singleton_(s) {}
+        T* singleton_;
+        void* unused_ = nullptr;
+      } singletonRepr_;
+      struct RawRepr {
+        void* first;
+        void* nullIfSingleton_;
+      };
+
+      // It is UB to read the singleton part of Repr if it was
+      // constructed as a shared_ptr and vice versa, but memcpying out
+      // the representation is always OK, so here's an accessor to obey
+      // the letter of the law.
+      RawRepr rawRepr() const {
+        RawRepr repr{};
+        memcpy(&repr, reinterpret_cast<const char *>(this), sizeof(RawRepr));
+        return repr;
+      }
+
+      bool isNonNull() const {
+        auto repr = rawRepr();
+        TORCH_INTERNAL_ASSERT_DEBUG_ONLY(repr.nullIfSingleton_ == nullptr || repr.first != nullptr);
+        return repr.first != nullptr;
+      }
+
+      bool isSharedAndNonNull() const {
+        return rawRepr().nullIfSingleton_ != nullptr;
+      }
+
+     private:
+      void destroy() {
+        if (isSharedAndNonNull()) {
+          // Without SharedPtrWrapper, this line would read
+          // `shared_.~shared_ptr()` and nvcc would complain with
+          // "error: expected primary-expression before '>' token"
+          // referring to the "t" in "shared_ptr". SharedPtrWrapper
+          // exists to work around this compiler bug.
+          shared_.~SharedPtrWrapper();
+        }
+      }
+    } repr_;
   };
 
   using TypePtr = SingletonOrSharedTypePtr<Type>;

aten/src/ATen/core/op_registration/op_registration.h

@@ -21,7 +21,7 @@ namespace c10 {
 
 namespace detail {
 // The first argument of the schema might be of type DispatchKeySet, in which case we remove it.
-// We do this because every argument in a function schema is expected to be convertible
+// We do this because every argument in a function schema is expected to be convertable
 // to an ivalue, but DispatchKeySet is not a type we want the jit to be aware of.
 // See Note [Plumbing Keys Through The Dispatcher]
 template<class KernelFunctor>

aten/src/ATen/core/op_registration/op_registration_test.cpp

@@ -251,7 +251,7 @@ TEST(OperatorRegistrationTest, whenRegisteringCPUTensorType_thenCanOnlyCallUnbox
   callOpUnboxedWithPrecomputedDispatchKeySet<void, Tensor>(*op, c10::DispatchKeySet(c10::DispatchKey::CPU), dummyTensor(c10::DispatchKey::CUDA));
   EXPECT_TRUE(called_kernel_cpu);
 
-  // Ensure that dispatch key from tensor is not used here.
+  // Ensure that disptach key from tensor is not used here.
   called_kernel_cpu = false;
   expectThrows<c10::Error>([&] {
     callOpUnboxedWithPrecomputedDispatchKeySet<void, Tensor>(*op, c10::DispatchKeySet(c10::DispatchKey::CUDA), dummyTensor(c10::DispatchKey::CPU));

aten/src/ATen/core/tensor_type.cpp

@@ -172,7 +172,7 @@ VaryingShape<Stride> TensorType::computeStrideProps(
   // The logic below follows what TensorIterator uses in its logic:
   //   1. Fast_set_up is the short-cut to identify a. channels_last and
   //      b. contiguous format, which is what we have in the below logic.
-  //   2. In more general cases, it does best effort to preserve permutatoin.
+  //   2. In more generla cases, it does best effort to preserve permutatoin.
   if (is_channels_last_strides_2d(sizes, strides) || is_channels_last_strides_3d(sizes, strides)) {
     // case 1.a. short cut channels last
     std::iota(stride_indices.rbegin() + 1, stride_indices.rend() - 1, 2);

aten/src/ATen/core/type.cpp

@@ -8,7 +8,6 @@
 #include <ATen/core/jit_type.h>
 #include <c10/macros/Macros.h>
 #include <c10/util/env.h>
-#include <c10/util/Exception.h>
 #include <c10/util/flat_hash_map.h>
 #include <c10/util/irange.h>
 #include <array>
@@ -827,7 +826,9 @@ TupleType::TupleType(
     : NamedType(TypeKind::TupleType, std::move(name)),
       elements_(std::move(elements)),
       has_free_variables_(std::any_of(elements_.begin(), elements_.end(), [](const TypePtr& v) {
-        TORCH_CHECK(v, "Can not create tuple with None type");
+        if (!v) {
+          throw std::runtime_error("Can not create tuple with None type");
+        }
         return v->hasFreeVariables();
       })), schema_(std::move(schema)) {
 

aten/src/ATen/cpu/vec/sve/vec_float.h

@@ -104,6 +104,71 @@ class Vectorized<float> {
     }
     return b;
   }
+  // Implementation is picked from
+  // https://github.com/ARM-software/ComputeLibrary/blob/v25.01/src/core/NEON/SVEMath.inl#L105
+  inline svfloat32_t svexp_f32_z(svbool_t pg, svfloat32_t x) const {
+    const auto c1 =
+        svreinterpret_f32_u32(svdup_n_u32(0x3f7ffff6)); // x^1: 0x1.ffffecp-1f
+    const auto c2 =
+        svreinterpret_f32_u32(svdup_n_u32(0x3efffedb)); // x^2: 0x1.fffdb6p-2f
+    const auto c3 =
+        svreinterpret_f32_u32(svdup_n_u32(0x3e2aaf33)); // x^3: 0x1.555e66p-3f
+    const auto c4 =
+        svreinterpret_f32_u32(svdup_n_u32(0x3d2b9f17)); // x^4: 0x1.573e2ep-5f
+    const auto c5 =
+        svreinterpret_f32_u32(svdup_n_u32(0x3c072010)); // x^5: 0x1.0e4020p-7f
+    const auto shift = svreinterpret_f32_u32(
+        svdup_n_u32(0x4b00007f)); // 2^23 + 127 = 0x1.0000fep23f
+    const auto inv_ln2 = svreinterpret_f32_u32(
+        svdup_n_u32(0x3fb8aa3b)); // 1 / ln(2) = 0x1.715476p+0f
+    const auto neg_ln2_hi = svreinterpret_f32_u32(svdup_n_u32(
+        0xbf317200)); // -ln(2) from bits  -1 to -19: -0x1.62e400p-1f
+    const auto neg_ln2_lo = svreinterpret_f32_u32(svdup_n_u32(
+        0xb5bfbe8e)); // -ln(2) from bits -20 to -42: -0x1.7f7d1cp-20f
+    const auto inf = svdup_n_f32(std::numeric_limits<float>::infinity());
+    const auto max_input = svdup_n_f32(88.37f); // Approximately ln(2^127.5)
+    const auto zero = svdup_n_f32(0.f);
+    const auto min_input = svdup_n_f32(-86.64f); // Approximately ln(2^-125)
+    // Range reduction:
+    //   e^x = 2^n * e^r
+    // where:
+    //   n = floor(x / ln(2))
+    //   r = x - n * ln(2)
+    //
+    // By adding x / ln(2) with 2^23 + 127 (shift):
+    //   * As FP32 fraction part only has 23-bits, the addition of 2^23 + 127
+    //   forces decimal part
+    //     of x / ln(2) out of the result. The integer part of x / ln(2) (i.e.
+    //     n) + 127 will occupy the whole fraction part of z in FP32 format.
+    //     Subtracting 2^23 + 127 (shift) from z will result in the integer part
+    //     of x / ln(2) (i.e. n) because the decimal part has been pushed out
+    //     and lost.
+    //   * The addition of 127 makes the FP32 fraction part of z ready to be
+    //   used as the exponent
+    //     in FP32 format. Left shifting z by 23 bits will result in 2^n.
+    const auto z = svmla_f32_z(pg, shift, x, inv_ln2);
+    const auto n = svsub_f32_z(pg, z, shift);
+    const auto scale = svreinterpret_f32_u32(
+        svlsl_n_u32_z(pg, svreinterpret_u32_f32(z), 23)); // 2^n
+    // The calculation of n * ln(2) is done using 2 steps to achieve accuracy
+    // beyond FP32. This outperforms longer Taylor series (3-4 tabs) both in
+    // term of accuracy and performance.
+    const auto r_hi = svmla_f32_z(pg, x, n, neg_ln2_hi);
+    const auto r = svmla_f32_z(pg, r_hi, n, neg_ln2_lo);
+    // Compute the truncated Taylor series of e^r.
+    //   poly = scale * (1 + c1 * r + c2 * r^2 + c3 * r^3 + c4 * r^4 + c5 * r^5)
+    const auto r2 = svmul_f32_z(pg, r, r);
+    const auto p1 = svmul_f32_z(pg, c1, r);
+    const auto p23 = svmla_f32_z(pg, c2, c3, r);
+    const auto p45 = svmla_f32_z(pg, c4, c5, r);
+    const auto p2345 = svmla_f32_z(pg, p23, p45, r2);
+    const auto p12345 = svmla_f32_z(pg, p1, p2345, r2);
+    auto poly = svmla_f32_z(pg, scale, p12345, scale);
+    // Handle underflow and overflow.
+    poly = svsel_f32(svcmplt_f32(pg, x, min_input), zero, poly);
+    poly = svsel_f32(svcmpgt_f32(pg, x, max_input), inf, poly);
+    return poly;
+  }
   static Vectorized<float> loadu(const void* ptr, int64_t count = size()) {
     if (count == size())
       return svld1_f32(ptrue, reinterpret_cast<const float*>(ptr));
@@ -248,41 +313,11 @@ class Vectorized<float> {
     return USE_SLEEF(
         Vectorized<float>(Sleef_expm1fx_u10sve(values)), map(std::expm1));
   }
-  // Implementation copied from Arm Optimized Routines:
-  // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/sve/expf.c
   Vectorized<float> exp_u20() const {
-    // special case to handle special inputs that are too large or too small
-    // i.e. where there's at least one element x, s.t. |x| >= 87.3...
-    svbool_t is_special_case = svacgt(svptrue_b32(), values, 0x1.5d5e2ap+6f);
-    if (svptest_any(svptrue_b32(), is_special_case)) {
-      return exp();
-    }
-    const svfloat32_t ln2_hi = svdup_n_f32(0x1.62e4p-1f);
-    const svfloat32_t ln2_lo = svdup_n_f32(0x1.7f7d1cp-20f);
-    const svfloat32_t c1 = svdup_n_f32(0.5f);
-    const svfloat32_t inv_ln2 = svdup_n_f32(0x1.715476p+0f);
-
-    const float shift = 0x1.803f8p17f;
-
-    /* n = round(x/(ln2/N)).  */
-    svfloat32_t z = svmad_x(svptrue_b32(), inv_ln2, values, shift);
-    svfloat32_t n = svsub_x(svptrue_b32(), z, shift);
-
-    /* r = x - n*ln2/N.  */
-    svfloat32_t r = values;
-    r = svmls_x(svptrue_b32(), r, n, ln2_hi);
-    r = svmls_x(svptrue_b32(), r, n, ln2_lo);
-
-    /* scale = 2^(n/N).  */
-    svfloat32_t scale = svexpa(svreinterpret_u32(z));
-
-    /* poly(r) = exp(r) - 1 ~= r + 0.5 r^2.  */
-    svfloat32_t r2 = svmul_x(svptrue_b32(), r, r);
-    svfloat32_t poly = svmla_x(svptrue_b32(), r, r2, c1);
-    return svmla_x(svptrue_b32(), scale, scale, poly);
+    return exp();
   }
   Vectorized<float> fexp_u20() const {
-    return exp_u20();
+    return exp();
   }
   Vectorized<float> fmod(const Vectorized<float>& q) const {USE_SLEEF(
       { return Vectorized<float>(Sleef_fmodfx_sve(values, q)); },
@@ -418,11 +453,9 @@ class Vectorized<float> {
         ptrue, svmax_f32_z(ptrue, values, CONST_MIN_TANH), CONST_MAX_TANH);
 
     // Step 2: Calculate exp(2 * x), where x is the clamped value.
-    // svmul_f32_z computes 2 * x, and exp_u20() computes the exponential of
-    // the result (via Vectorized<float>, then auto-converts back to
-    // svfloat32_t).
-    svfloat32_t exp2x =
-        Vectorized<float>(svmul_f32_z(ptrue, CONST_2, x)).exp_u20();
+    // svmul_f32_z computes 2 * x, and svexp_f32_z computes the exponential of
+    // the result.
+    svfloat32_t exp2x = svexp_f32_z(ptrue, svmul_f32_z(ptrue, CONST_2, x));
 
     // Step 3: Calculate the numerator of the tanh function, which is exp(2x)
     // - 1.

aten/src/ATen/cpu/vec/vec128/vec128.h

@@ -6,11 +6,8 @@
 #ifdef __aarch64__
 #if !defined(CPU_CAPABILITY_SVE)
 #include <ATen/cpu/vec/vec128/vec128_bfloat16_neon.h>
-#include <ATen/cpu/vec/vec128/vec128_double_neon.h>
 #include <ATen/cpu/vec/vec128/vec128_float_neon.h>
 #include <ATen/cpu/vec/vec128/vec128_half_neon.h>
-#include <ATen/cpu/vec/vec128/vec128_int_aarch64.h>
-#include <ATen/cpu/vec/vec128/vec128_uint_aarch64.h>
 #endif
 
 #include <ATen/cpu/vec/vec128/vec128_convert.h>

aten/src/ATen/cpu/vec/vec128/vec128_bfloat16_neon.h

@@ -354,47 +354,9 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
 
   DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
   Vectorized frac() const;
+  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
   DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(trunc)
   DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(sqrt)
-
-#ifdef __ARM_FEATURE_BF16
-  Vectorized<c10::BFloat16> neg() const {
-    return -values;
-  }
-  Vectorized<c10::BFloat16> reciprocal() const {
-    return 1.0f / values;
-  }
-  Vectorized<c10::BFloat16> operator==(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values == other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator!=(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values != other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator<(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values < other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator<=(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values <= other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator>(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values > other.values;
-  }
-
-  Vectorized<c10::BFloat16> operator>=(
-      const Vectorized<c10::BFloat16>& other) const {
-    return values >= other.values;
-  }
-#else
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
   DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(reciprocal)
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator==)
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator!=)
@@ -402,7 +364,6 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator<=)
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator>)
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator>=)
-#endif
 
 #undef DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD
 #undef DEFINE_BINARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD
@@ -451,52 +412,28 @@ template <>
 Vectorized<c10::BFloat16> inline operator+(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  return x + y;
-#else
   return binary_operator_via_float(std::plus<Vectorized<float>>(), a, b);
-#endif
 }
 
 template <>
 Vectorized<c10::BFloat16> inline operator-(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  return x - y;
-#else
   return binary_operator_via_float(std::minus<Vectorized<float>>(), a, b);
-#endif
 }
 
 template <>
 Vectorized<c10::BFloat16> inline operator*(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  return x * y;
-#else
   return binary_operator_via_float(std::multiplies<Vectorized<float>>(), a, b);
-#endif
 }
 
 template <>
 Vectorized<c10::BFloat16> inline operator/(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  return x / y;
-#else
   return binary_operator_via_float(std::divides<Vectorized<float>>(), a, b);
-#endif
 }
 
 // frac. Implement this here so we can use subtraction
@@ -607,19 +544,12 @@ Vectorized<c10::BFloat16> inline fmadd(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  bfloat16x8_t z = c;
-  return x * y + z;
-#else
   // NOTE [BF16 FMA]: There isn't an FMA that accumulates into BF16!  Also,
   // vbfmlalbq_f32 and vbfmlaltq_f32 take the even and odd-numbered
   // elements, not the bottom and top half, so they don't seem
   // particularly useful here. Ideally we would include dot product in
   // the Vectorized interface...
   return a * b + c;
-#endif
 }
 
 template <>
@@ -627,15 +557,8 @@ Vectorized<c10::BFloat16> inline fnmadd(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  bfloat16x8_t z = c;
-  return (-x) * y + z;
-#else
   // See NOTE [BF16 FMA] above.
   return -a * b + c;
-#endif
 }
 
 template <>
@@ -643,15 +566,8 @@ Vectorized<c10::BFloat16> inline fmsub(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  bfloat16x8_t z = c;
-  return x * y - z;
-#else
   // See NOTE [BF16 FMA] above.
   return a * b - c;
-#endif
 }
 
 template <>
@@ -659,15 +575,8 @@ Vectorized<c10::BFloat16> inline fnmsub(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#ifdef __ARM_FEATURE_BF16
-  bfloat16x8_t x = a;
-  bfloat16x8_t y = b;
-  bfloat16x8_t z = c;
-  return (-x) * y - z;
-#else
   // See NOTE [BF16 FMA] above.
   return -a * b - c;
-#endif
 }
 
 #endif // !defined(C10_MOBILE) && defined(__aarch64__)

aten/src/ATen/cpu/vec/vec128/vec128_convert.h

@@ -5,129 +5,6 @@
 namespace at::vec {
 inline namespace CPU_CAPABILITY {
 #if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
-
-// Enable auto-vectorization for GCC-13+ and clang-17+
-// GCC-12 has a bug: gcc.gnu.org/bugzilla/show_bug.cgi?id=117001
-#if __GNUC__ > 12 || (defined(__clang__) && (__clang_major__ >= 17))
-
-template <typename from_type, typename to_type>
-inline void convertImpl(
-    const from_type* __restrict src,
-    to_type* __restrict dst,
-    int64_t n) {
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    dst[i] = static_cast<to_type>(src[i]);
-  }
-}
-
-#define CONVERT_TEMPLATE(from_type, to_type)                           \
-  template <>                                                          \
-  inline void convert(const from_type* src, to_type* dst, int64_t n) { \
-    return convertImpl<from_type, to_type>(src, dst, n);               \
-  }
-
-CONVERT_TEMPLATE(uint8_t, uint8_t)
-CONVERT_TEMPLATE(uint8_t, int8_t)
-CONVERT_TEMPLATE(uint8_t, int16_t)
-CONVERT_TEMPLATE(uint8_t, int32_t)
-CONVERT_TEMPLATE(uint8_t, int64_t)
-CONVERT_TEMPLATE(uint8_t, float)
-CONVERT_TEMPLATE(uint8_t, double)
-CONVERT_TEMPLATE(int8_t, uint8_t)
-CONVERT_TEMPLATE(int8_t, int8_t)
-CONVERT_TEMPLATE(int8_t, int16_t)
-CONVERT_TEMPLATE(int8_t, int32_t)
-CONVERT_TEMPLATE(int8_t, int64_t)
-CONVERT_TEMPLATE(int8_t, float)
-CONVERT_TEMPLATE(int8_t, double)
-CONVERT_TEMPLATE(int16_t, uint8_t)
-CONVERT_TEMPLATE(int16_t, int8_t)
-CONVERT_TEMPLATE(int16_t, int16_t)
-CONVERT_TEMPLATE(int16_t, int32_t)
-CONVERT_TEMPLATE(int16_t, int64_t)
-CONVERT_TEMPLATE(int16_t, float)
-CONVERT_TEMPLATE(int16_t, double)
-CONVERT_TEMPLATE(int32_t, uint8_t)
-CONVERT_TEMPLATE(int32_t, int8_t)
-CONVERT_TEMPLATE(int32_t, int16_t)
-CONVERT_TEMPLATE(int32_t, int32_t)
-CONVERT_TEMPLATE(int32_t, int64_t)
-CONVERT_TEMPLATE(int32_t, float)
-CONVERT_TEMPLATE(int32_t, double)
-CONVERT_TEMPLATE(int64_t, uint8_t)
-CONVERT_TEMPLATE(int64_t, int8_t)
-CONVERT_TEMPLATE(int64_t, int16_t)
-CONVERT_TEMPLATE(int64_t, int32_t)
-CONVERT_TEMPLATE(int64_t, int64_t)
-CONVERT_TEMPLATE(int64_t, float)
-CONVERT_TEMPLATE(int64_t, double)
-CONVERT_TEMPLATE(float, uint8_t)
-CONVERT_TEMPLATE(float, int8_t)
-CONVERT_TEMPLATE(float, int16_t)
-CONVERT_TEMPLATE(float, int32_t)
-CONVERT_TEMPLATE(float, int64_t)
-CONVERT_TEMPLATE(float, float)
-CONVERT_TEMPLATE(float, double)
-CONVERT_TEMPLATE(double, uint8_t)
-CONVERT_TEMPLATE(double, int8_t)
-CONVERT_TEMPLATE(double, int16_t)
-CONVERT_TEMPLATE(double, int32_t)
-CONVERT_TEMPLATE(double, int64_t)
-CONVERT_TEMPLATE(double, float)
-CONVERT_TEMPLATE(double, double)
-#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-
-#define CONVERT_FROM_FP16_TEMPLATE(to_type)                            \
-  template <>                                                          \
-  inline void convert(const at::Half* src, to_type* dst, int64_t n) {  \
-    const float16_t* srcPtr = reinterpret_cast<const float16_t*>(src); \
-    return convertImpl<float16_t, to_type>(srcPtr, dst, n);            \
-  }
-
-#define CONVERT_TO_FP16_TEMPLATE(from_type)                             \
-  template <>                                                           \
-  inline void convert(const from_type* src, at::Half* dst, int64_t n) { \
-    float16_t* dstPtr = reinterpret_cast<float16_t*>(dst);              \
-    return convertImpl<from_type, float16_t>(src, dstPtr, n);           \
-  }
-
-CONVERT_FROM_FP16_TEMPLATE(uint8_t)
-CONVERT_FROM_FP16_TEMPLATE(int8_t)
-CONVERT_FROM_FP16_TEMPLATE(int16_t)
-CONVERT_FROM_FP16_TEMPLATE(int32_t)
-CONVERT_FROM_FP16_TEMPLATE(int64_t)
-CONVERT_FROM_FP16_TEMPLATE(float16_t)
-CONVERT_FROM_FP16_TEMPLATE(float)
-CONVERT_FROM_FP16_TEMPLATE(double)
-CONVERT_TO_FP16_TEMPLATE(uint8_t)
-CONVERT_TO_FP16_TEMPLATE(int8_t)
-CONVERT_TO_FP16_TEMPLATE(int16_t)
-CONVERT_TO_FP16_TEMPLATE(int32_t)
-CONVERT_TO_FP16_TEMPLATE(int64_t)
-CONVERT_TO_FP16_TEMPLATE(float)
-CONVERT_TO_FP16_TEMPLATE(double)
-#endif
-#ifdef __ARM_FEATURE_BF16
-CONVERT_TEMPLATE(bfloat16_t, uint8_t)
-CONVERT_TEMPLATE(bfloat16_t, int8_t)
-CONVERT_TEMPLATE(bfloat16_t, int16_t)
-CONVERT_TEMPLATE(bfloat16_t, int32_t)
-CONVERT_TEMPLATE(bfloat16_t, int64_t)
-CONVERT_TEMPLATE(bfloat16_t, bfloat16_t)
-CONVERT_TEMPLATE(bfloat16_t, float)
-CONVERT_TEMPLATE(bfloat16_t, double)
-CONVERT_TEMPLATE(uint8_t, bfloat16_t)
-CONVERT_TEMPLATE(int8_t, bfloat16_t)
-CONVERT_TEMPLATE(int16_t, bfloat16_t)
-CONVERT_TEMPLATE(int32_t, bfloat16_t)
-CONVERT_TEMPLATE(int64_t, bfloat16_t)
-CONVERT_TEMPLATE(float, bfloat16_t)
-CONVERT_TEMPLATE(double, bfloat16_t)
-#endif
-
-#endif
-
 template <typename src_t>
 struct VecConvert<
     float,

aten/src/ATen/cpu/vec/vec128/vec128_double_neon.h

@@ -1,586 +0,0 @@
-#pragma once
-
-#include <ATen/cpu/vec/intrinsics.h>
-#include <ATen/cpu/vec/vec_base.h>
-#include <c10/macros/Macros.h>
-#include <c10/util/irange.h>
-#include <cmath>
-
-namespace at::vec {
-// Note [CPU_CAPABILITY namespace]
-// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-// This header, and all of its subheaders, will be compiled with
-// different architecture flags for each supported set of vector
-// intrinsics. So we need to make sure they aren't inadvertently
-// linked together. We do this by declaring objects in an `inline
-// namespace` which changes the name mangling, but can still be
-// accessed as `at::vec`.
-inline namespace CPU_CAPABILITY {
-
-template <>
-struct is_vec_specialized_for<double> : std::bool_constant<true> {};
-
-template <>
-class Vectorized<double> {
- private:
-  float64x2_t values;
-
- public:
-  using value_type = double;
-  using size_type = int;
-  static constexpr size_type size() {
-    return 2;
-  }
-  Vectorized() {
-    values = vdupq_n_f64(0.0);
-  }
-  Vectorized(float64x2_t v) : values(v) {}
-  Vectorized(double val) {
-    values = vdupq_n_f64(val);
-  }
-  template <
-      typename... Args,
-      typename = std::enable_if_t<(sizeof...(Args) == size())>>
-  Vectorized(Args... vals) {
-    __at_align__ double buffer[size()] = {vals...};
-    values = vld1q_f64(buffer);
-  }
-  operator float64x2_t() const {
-    return values;
-  }
-  template <int64_t mask>
-  static Vectorized<double> blend(
-      const Vectorized<double>& a,
-      const Vectorized<double>& b) {
-    // Build an array of flags: each bit of element is 1 if the corresponding
-    // bit in 'mask' is set, 0 otherwise.
-    uint64x2_t maskArray = {
-        (mask & 1ULL) ? 0xFFFFFFFFFFFFFFFF : 0,
-        (mask & 2ULL) ? 0xFFFFFFFFFFFFFFFF : 0};
-    // Use BSL to select elements from b where the mask is 1, else from a
-    return vbslq_f64(maskArray, b.values, a.values);
-  }
-  static Vectorized<double> blendv(
-      const Vectorized<double>& a,
-      const Vectorized<double>& b,
-      const Vectorized<double>& mask_) {
-    return vbslq_f64(vreinterpretq_u64_f64(mask_.values), b.values, a.values);
-  }
-  template <typename step_t>
-  static Vectorized<double> arange(
-      double base = 0.,
-      step_t step = static_cast<step_t>(1)) {
-    return {base, base + static_cast<double>(step)};
-  }
-  static inline Vectorized<double> set(
-      const Vectorized<double>& a,
-      const Vectorized<double>& b,
-      int64_t count = size()) {
-    if (count == 0) {
-      return a;
-    } else if (count >= 2) {
-      return b;
-    } else {
-      float64x2_t c = {b.values[0], a.values[1]};
-      return c;
-    }
-  }
-  static Vectorized<double> loadu(const void* ptr, int64_t count = size()) {
-    if (count == size()) {
-      return vld1q_f64(reinterpret_cast<const double*>(ptr));
-    } else if (count == 1) {
-      float64x1_t x = vld1_f64(reinterpret_cast<const double*>(ptr));
-      float64x1_t z = {0.0};
-      return vcombine_f64(x, z);
-    } else {
-      return vdupq_n_f64(0.0);
-    }
-  }
-  void store(void* ptr, int64_t count = size()) const {
-    if (count == size()) {
-      vst1q_f64(reinterpret_cast<double*>(ptr), values);
-    } else if (count == 1) {
-      vst1_f64(reinterpret_cast<double*>(ptr), vget_low_f64(values));
-    }
-  }
-  const double& operator[](int idx) const = delete;
-  double& operator[](int idx) = delete;
-  int64_t zero_mask() const {
-    // returns an integer mask where all zero elements are translated to 1-bit
-    // and others are translated to 0-bit
-    uint64x2_t cmpReg = vceqzq_f64(values);
-    uint64x2_t mask = {1, 2};
-    uint64x2_t res = vandq_u64(cmpReg, mask);
-    return res[0] | res[1];
-  }
-  Vectorized<double> isnan() const {
-    // NaN check
-    return vreinterpretq_f64_u32(
-        vmvnq_u32(vreinterpretq_u32_u64(vceqq_f64(values, values))));
-  }
-  bool has_inf_nan() const {
-    Vectorized<double> x = vsubq_f64(values, values);
-    float64x2_t r = x.isnan();
-    uint64x2_t u = vreinterpretq_u64_f64(r);
-    return u[0] | u[1];
-  }
-  Vectorized<double> map(double (*f)(double)) const {
-    float64x2_t result;
-    result[0] = f(values[0]);
-    result[1] = f(values[1]);
-    return result;
-  }
-  Vectorized<double> map2(
-      const Vectorized<double>& second,
-      double (*const f)(double, double)) const {
-    float64x2_t result;
-    result[0] = f(values[0], second.values[0]);
-    result[1] = f(values[1], second.values[1]);
-    return result;
-  }
-  Vectorized<double> abs() const {
-    return vabsq_f64(values);
-  }
-  Vectorized<double> angle() const {
-    auto zero = Vectorized<double>(0.0);
-    auto pi = Vectorized<double>(c10::pi<double>);
-    auto tmp = blendv(zero, pi, vreinterpretq_f64_u64(vcltzq_f64(values)));
-    return blendv(tmp, *this, isnan());
-  }
-  Vectorized<double> real() const {
-    return *this;
-  }
-  Vectorized<double> imag() const {
-    return Vectorized<double>(0.0);
-  }
-  Vectorized<double> conj() const {
-    return *this;
-  }
-  Vectorized<double> acos() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_acosd2_u10(values)), map(std::acos));
-  }
-  Vectorized<double> acosh() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_acoshd2_u10(values)), map(std::acosh));
-  }
-  Vectorized<double> asin() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_asind2_u10(values)), map(std::asin));
-  }
-  Vectorized<double> asinh() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_asinhd2_u10(values)), map(std::asinh));
-  }
-  Vectorized<double> atan() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_atand2_u10(values)), map(std::atan));
-  }
-  Vectorized<double> atanh() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_atanhd2_u10(values)), map(std::atanh));
-  }
-  Vectorized<double> atan2(const Vectorized<double>& b) const {USE_SLEEF(
-      { return Vectorized<double>(Sleef_atan2d2_u10(values, b)); },
-      {
-        __at_align__ double tmp[size()];
-        __at_align__ double tmp_b[size()];
-        store(tmp);
-        b.store(tmp_b);
-        for (int64_t i = 0; i < size(); i++) {
-          tmp[i] = std::atan2(tmp[i], tmp_b[i]);
-        }
-        return loadu(tmp);
-      })} Vectorized<double> copysign(const Vectorized<double>& sign) const {
-      USE_SLEEF(
-          { return Vectorized<double>(Sleef_copysignd2(values, sign)); },
-          {
-            __at_align__ double tmp[size()];
-            __at_align__ double tmp_sign[size()];
-            store(tmp);
-            sign.store(tmp_sign);
-            for (int64_t i = 0; i < size(); i++) {
-              tmp[i] = std::copysign(tmp[i], tmp_sign[i]);
-            }
-            return loadu(tmp);
-          })} Vectorized<double> erf() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_erfd2_u10(values)), map(std::erf));
-  }
-  Vectorized<double> erfc() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_erfcd2_u15(values)), map(std::erfc));
-  }
-  Vectorized<double> exp() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_expd2_u10(values)), map(std::exp));
-  }
-  Vectorized<double> exp2() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_exp2d2_u10(values)), map(std::exp2));
-  }
-  Vectorized<double> expm1() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_expm1d2_u10(values)), map(std::expm1));
-  }
-  Vectorized<double> fmod(const Vectorized<double>& q) const {USE_SLEEF(
-      { return Vectorized<double>(Sleef_fmodd2(values, q)); },
-      {
-        __at_align__ double tmp[size()];
-        __at_align__ double tmp_q[size()];
-        store(tmp);
-        q.store(tmp_q);
-        for (int64_t i = 0; i < size(); i++) {
-          tmp[i] = std::fmod(tmp[i], tmp_q[i]);
-        }
-        return loadu(tmp);
-      })} Vectorized<double> hypot(const Vectorized<double>& b) const {
-      USE_SLEEF(
-          { return Vectorized<double>(Sleef_hypotd2_u05(values, b)); },
-          {
-            __at_align__ double tmp[size()];
-            __at_align__ double tmp_b[size()];
-            store(tmp);
-            b.store(tmp_b);
-            for (int64_t i = 0; i < size(); i++) {
-              tmp[i] = std::hypot(tmp[i], tmp_b[i]);
-            }
-            return loadu(tmp);
-          })} Vectorized<double> i0() const {
-    return map(calc_i0);
-  }
-  Vectorized<double> nextafter(const Vectorized<double>& b) const {USE_SLEEF(
-      { return Vectorized<double>(Sleef_nextafterd2(values, b)); },
-      {
-        __at_align__ double tmp[size()];
-        __at_align__ double tmp_b[size()];
-        store(tmp);
-        b.store(tmp_b);
-        for (int64_t i = 0; i < size(); ++i) {
-          tmp[i] = std::nextafter(tmp[i], tmp_b[i]);
-        }
-        return loadu(tmp);
-      })} Vectorized<double> log() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_logd2_u10(values)), map(std::log));
-  }
-  Vectorized<double> log2() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_log2d2_u10(values)), map(std::log2));
-  }
-  Vectorized<double> log10() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_log10d2_u10(values)), map(std::log10));
-  }
-  Vectorized<double> log1p() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_log1pd2_u10(values)), map(std::log1p));
-  }
-  Vectorized<double> frac() const;
-  Vectorized<double> sin() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_sind2_u10(values)), map(std::sin));
-  }
-  Vectorized<double> sinh() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_sinhd2_u10(values)), map(std::sinh));
-  }
-  Vectorized<double> cos() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_cosd2_u10(values)), map(std::cos));
-  }
-  Vectorized<double> cosh() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_coshd2_u10(values)), map(std::cosh));
-  }
-  Vectorized<double> pow(const Vectorized<double>& b) const {USE_SLEEF(
-      { return Vectorized<double>(Sleef_powd2_u10(values, b)); },
-      {
-        __at_align__ double tmp[size()];
-        __at_align__ double tmp_b[size()];
-        store(tmp);
-        b.store(tmp_b);
-        for (int64_t i = 0; i < size(); i++) {
-          tmp[i] = std::pow(tmp[i], tmp_b[i]);
-        }
-        return loadu(tmp);
-      })} // Comparison using the _CMP_**_OQ predicate.
-          //   `O`: get false if an operand is NaN
-          //   `Q`: do not raise if an operand is NaN
-  Vectorized<double> tan() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_tand2_u10(values)), map(std::tan));
-  }
-  Vectorized<double> tanh() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_tanhd2_u10(values)), map(std::tanh));
-  }
-  Vectorized<double> lgamma() const {
-    return USE_SLEEF(
-        Vectorized<double>(Sleef_lgammad2_u10(values)), map(std::lgamma));
-  }
-  Vectorized<double> erfinv() const {
-    return map(calc_erfinv);
-  }
-  Vectorized<double> exp_u20() const {
-    return exp();
-  }
-  Vectorized<double> fexp_u20() const {
-    return exp();
-  }
-  Vectorized<double> i0e() const {
-    return map(calc_i0e);
-  }
-  Vectorized<double> digamma() const {
-    return map(calc_digamma);
-  }
-  Vectorized<double> igamma(const Vectorized<double>& x) const {
-    __at_align__ double tmp[size()];
-    __at_align__ double tmp_x[size()];
-    store(tmp);
-    x.store(tmp_x);
-    for (int64_t i = 0; i < size(); i++) {
-      tmp[i] = calc_igamma(tmp[i], tmp_x[i]);
-    }
-    return loadu(tmp);
-  }
-  Vectorized<double> igammac(const Vectorized<double>& x) const {
-    __at_align__ double tmp[size()];
-    __at_align__ double tmp_x[size()];
-    store(tmp);
-    x.store(tmp_x);
-    for (int64_t i = 0; i < size(); i++) {
-      tmp[i] = calc_igammac(tmp[i], tmp_x[i]);
-    }
-    return loadu(tmp);
-  }
-  Vectorized<double> ceil() const {
-    return vrndpq_f64(values);
-  }
-  Vectorized<double> floor() const {
-    return vrndmq_f64(values);
-  }
-  Vectorized<double> neg() const {
-    return vnegq_f64(values);
-  }
-  Vectorized<double> round() const {
-    return vrndiq_f64(values);
-  }
-  Vectorized<double> trunc() const {
-    return vrndq_f64(values);
-  }
-  Vectorized<double> sqrt() const {
-    return vsqrtq_f64(values);
-  }
-  Vectorized<double> reciprocal() const {
-    return vdivq_f64(vdupq_n_f64(1.0), values);
-  }
-  Vectorized<double> rsqrt() const {
-    return vdivq_f64(vdupq_n_f64(1.0), vsqrtq_f64(values));
-  }
-  double reduce_add() const {
-    return vaddvq_f64(values);
-  }
-  double reduce_max() const {
-    return vmaxvq_f64(values);
-  }
-  Vectorized<double> operator==(const Vectorized<double>& other) const {
-    return Vectorized<double>(
-        vreinterpretq_f64_u64(vceqq_f64(values, other.values)));
-  }
-
-  Vectorized<double> operator!=(const Vectorized<double>& other) const {
-    float64x2_t r0 = vreinterpretq_f64_u32(
-        vmvnq_u32(vreinterpretq_u32_u64(vceqq_f64(values, other.values))));
-    return Vectorized<double>(r0);
-  }
-
-  Vectorized<double> operator<(const Vectorized<double>& other) const {
-    return Vectorized<double>(
-        vreinterpretq_f64_u64(vcltq_f64(values, other.values)));
-  }
-
-  Vectorized<double> operator<=(const Vectorized<double>& other) const {
-    return Vectorized<double>(
-        vreinterpretq_f64_u64(vcleq_f64(values, other.values)));
-  }
-
-  Vectorized<double> operator>(const Vectorized<double>& other) const {
-    return Vectorized<double>(
-        vreinterpretq_f64_u64(vcgtq_f64(values, other.values)));
-  }
-
-  Vectorized<double> operator>=(const Vectorized<double>& other) const {
-    return Vectorized<double>(
-        vreinterpretq_f64_u64(vcgeq_f64(values, other.values)));
-  }
-
-  Vectorized<double> eq(const Vectorized<double>& other) const;
-  Vectorized<double> ne(const Vectorized<double>& other) const;
-  Vectorized<double> gt(const Vectorized<double>& other) const;
-  Vectorized<double> ge(const Vectorized<double>& other) const;
-  Vectorized<double> lt(const Vectorized<double>& other) const;
-  Vectorized<double> le(const Vectorized<double>& other) const;
-};
-
-template <>
-Vectorized<double> inline operator+(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vaddq_f64(a, b);
-}
-
-template <>
-Vectorized<double> inline operator-(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vsubq_f64(a, b);
-}
-
-template <>
-Vectorized<double> inline operator*(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vmulq_f64(a, b);
-}
-
-template <>
-Vectorized<double> inline operator/(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vdivq_f64(a, b);
-}
-
-// frac. Implement this here so we can use subtraction
-Vectorized<double> inline Vectorized<double>::frac() const {
-  return *this - this->trunc();
-}
-
-// Implements the IEEE 754 201X `maximum` operation, which propagates NaN if
-// either input is a NaN.
-template <>
-Vectorized<double> inline maximum(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vmaxq_f64(a, b);
-}
-
-// Implements the IEEE 754 201X `minimum` operation, which propagates NaN if
-// either input is a NaN.
-template <>
-Vectorized<double> inline minimum(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vminq_f64(a, b);
-}
-
-template <>
-Vectorized<double> inline clamp(
-    const Vectorized<double>& a,
-    const Vectorized<double>& min,
-    const Vectorized<double>& max) {
-  return vminq_f64(max, vmaxq_f64(min, a));
-}
-
-template <>
-Vectorized<double> inline clamp_max(
-    const Vectorized<double>& a,
-    const Vectorized<double>& max) {
-  return vminq_f64(max, a);
-}
-
-template <>
-Vectorized<double> inline clamp_min(
-    const Vectorized<double>& a,
-    const Vectorized<double>& min) {
-  return vmaxq_f64(min, a);
-}
-
-template <>
-Vectorized<double> inline operator&(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vreinterpretq_f64_u64(
-      vandq_u64(vreinterpretq_u64_f64(a), vreinterpretq_u64_f64(b)));
-}
-
-template <>
-Vectorized<double> inline operator|(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vreinterpretq_f64_u64(
-      vorrq_u64(vreinterpretq_u64_f64(a), vreinterpretq_u64_f64(b)));
-}
-
-template <>
-Vectorized<double> inline operator^(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b) {
-  return vreinterpretq_f64_u64(
-      veorq_u64(vreinterpretq_u64_f64(a), vreinterpretq_u64_f64(b)));
-}
-
-inline Vectorized<double> Vectorized<double>::eq(
-    const Vectorized<double>& other) const {
-  return (*this == other) & Vectorized<double>(1.0);
-}
-
-inline Vectorized<double> Vectorized<double>::ne(
-    const Vectorized<double>& other) const {
-  return (*this != other) & Vectorized<double>(1.0);
-}
-
-inline Vectorized<double> Vectorized<double>::gt(
-    const Vectorized<double>& other) const {
-  return (*this > other) & Vectorized<double>(1.0);
-}
-
-inline Vectorized<double> Vectorized<double>::ge(
-    const Vectorized<double>& other) const {
-  return (*this >= other) & Vectorized<double>(1.0);
-}
-
-inline Vectorized<double> Vectorized<double>::lt(
-    const Vectorized<double>& other) const {
-  return (*this < other) & Vectorized<double>(1.0);
-}
-
-inline Vectorized<double> Vectorized<double>::le(
-    const Vectorized<double>& other) const {
-  return (*this <= other) & Vectorized<double>(1.0);
-}
-
-template <>
-Vectorized<double> inline fmadd(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return vfmaq_f64(c, a, b);
-}
-
-template <>
-Vectorized<double> inline fnmadd(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return vfmsq_f64(c, a, b);
-}
-
-template <>
-Vectorized<double> inline fmsub(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return vfmaq_f64(vnegq_f64(c), a, b);
-}
-
-template <>
-Vectorized<double> inline fnmsub(
-    const Vectorized<double>& a,
-    const Vectorized<double>& b,
-    const Vectorized<double>& c) {
-  return vfmsq_f64(vnegq_f64(c), a, b);
-}
-
-} // namespace CPU_CAPABILITY
-} // namespace at::vec

aten/src/ATen/cpu/vec/vec128/vec128_float_neon.h

@@ -307,49 +307,11 @@ class Vectorized<float> {
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(exp)
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(exp2)
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(expm1)
-  // Implementation copied from Arm Optimized Routine
-  // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/expf.c
   Vectorized<float> exp_u20() const {
-    // bail out to sleef if it's a special case:
-    // i.e. there's an input s.t. |input| > 87.3....
-    const float32x4_t special_bound = vdupq_n_f32(0x1.5d5e2ap+6f);
-    uint32x4_t cmp = vcagtq_f32(values, special_bound);
-    if (vpaddd_u64(vreinterpretq_u64_u32(cmp)) != 0) {
-      return exp();
-    }
-
-    const float32x4_t inv_ln2 = vdupq_n_f32(0x1.715476p+0f);
-    const float ln2_hi = 0x1.62e4p-1f;
-    const float ln2_lo = 0x1.7f7d1cp-20f;
-    const float c0 = 0x1.0e4020p-7f;
-    const float c2 = 0x1.555e66p-3f;
-    const float32x4_t ln2_c02 = {ln2_hi, ln2_lo, c0, c2};
-
-    const uint32x4_t exponent_bias = vdupq_n_u32(0x3f800000);
-    const float32x4_t c1 = vdupq_n_f32(0x1.573e2ep-5f);
-    const float32x4_t c3 = vdupq_n_f32(0x1.fffdb6p-2f);
-    const float32x4_t c4 = vdupq_n_f32(0x1.ffffecp-1f);
-
-    /* exp(x) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)]
-      x = ln2*n + r, with r in [-ln2/2, ln2/2].  */
-
-    float32x4_t n = vrndaq_f32(vmulq_f32(values, inv_ln2));
-    float32x4_t r = vfmsq_laneq_f32(values, n, ln2_c02, 0);
-    r = vfmsq_laneq_f32(r, n, ln2_c02, 1);
-    uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_s32(vcvtq_s32_f32(n)), 23);
-    float32x4_t scale = vreinterpretq_f32_u32(vaddq_u32(e, exponent_bias));
-
-    float32x4_t r2 = vmulq_f32(r, r);
-    float32x4_t p = vfmaq_laneq_f32(c1, r, ln2_c02, 2);
-    float32x4_t q = vfmaq_laneq_f32(c3, r, ln2_c02, 3);
-    q = vfmaq_f32(q, p, r2);
-    p = vmulq_f32(c4, r);
-    float32x4_t poly = vfmaq_f32(p, q, r2);
-
-    return vfmaq_f32(scale, poly, scale);
+    return exp();
   }
   Vectorized<float> fexp_u20() const {
-    return exp_u20();
+    return exp();
   }
   DEFINE_SLEEF_COMPATIBLE_BINARY_ELEMENTWISE_FUNC_WITH_SLEEF_NAME(
       fmod,
@@ -578,6 +540,42 @@ inline Vectorized<float> Vectorized<float>::le(
   return (*this <= other) & Vectorized<float>(1.0f);
 }
 
+template <>
+inline void convert(const float* src, int32_t* dst, int64_t n) {
+  int64_t i;
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (i = 0; i <= (n - Vectorized<float>::size());
+       i += Vectorized<float>::size()) {
+    vst1q_s32(dst + i, vcvtq_s32_f32(vld1q_f32(src + i)));
+  }
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (; i < n; i++) {
+    dst[i] = static_cast<int32_t>(src[i]);
+  }
+}
+
+template <>
+inline void convert(const int32_t* src, float* dst, int64_t n) {
+  int64_t i;
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (i = 0; i <= (n - Vectorized<float>::size());
+       i += Vectorized<float>::size()) {
+    vst1q_f32(dst + i, vcvtq_f32_s32(vld1q_s32(src + i)));
+  }
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (; i < n; i++) {
+    dst[i] = static_cast<float>(src[i]);
+  }
+}
+
 template <>
 Vectorized<float> inline fmadd(
     const Vectorized<float>& a,
@@ -634,7 +632,8 @@ inline Vectorized<float> Vectorized<float>::erf() const {
   // - exp(- x * x)
   auto pow_2 = (*this) * (*this);
   auto neg_pow_2 = pow_2 ^ neg_zero_vec;
-  auto tmp4 = neg_pow_2.exp();
+  auto tmp4 = neg_pow_2.map(
+      std::exp); // This can be swapped for a faster implementation of exp.
   auto tmp5 = tmp4 ^ neg_zero_vec;
   // erf(x) = sign(x) * (1 - r * t * exp(- x * x))
   auto tmp6 = t * tmp5;

aten/src/ATen/cpu/vec/vec128/vec128_half_neon.h

@@ -234,7 +234,7 @@ class Vectorized<c10::Half> : public Vectorized16<
         vshlq_u16(vandq_u16(is_zero_vec, vdupq_n_u16(1)), shift);
     return vaddvq_u16(bits_vec);
 #else // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    // use known working implementation.
+    // use known working implmentation.
     __at_align__ value_type tmp[size()];
     store(tmp);
     int mask = 0;
@@ -569,6 +569,46 @@ inline Vectorized<c10::Half> Vectorized<c10::Half>::le(
   return (*this <= other) & Vectorized<c10::Half>(1);
 }
 
+// These are global functions, so the defaults in vec_base.h should
+// work fine if __ARM_FEATURE_FP16_VECTOR_ARITHMETIC is not available.
+#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+template <>
+inline void convert(const float16_t* src, int16_t* dst, int64_t n) {
+  int64_t i;
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (i = 0; i <= (n - Vectorized<c10::Half>::size());
+       i += Vectorized<c10::Half>::size()) {
+    vst1q_s16(dst + i, vcvtq_s16_f16(vld1q_f16(src + i)));
+  }
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (; i < n; i++) {
+    dst[i] = static_cast<int16_t>(src[i]);
+  }
+}
+
+template <>
+inline void convert(const int16_t* src, float16_t* dst, int64_t n) {
+  int64_t i;
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (i = 0; i <= (n - Vectorized<c10::Half>::size());
+       i += Vectorized<c10::Half>::size()) {
+    vst1q_f16(dst + i, vcvtq_f16_s16(vld1q_s16(src + i)));
+  }
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (; i < n; i++) {
+    dst[i] = static_cast<float16_t>(src[i]);
+  }
+}
+#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
+
 template <>
 Vectorized<c10::Half> inline fmadd(
     const Vectorized<c10::Half>& a,

aten/src/ATen/cpu/vec/vec128/vec128_int_aarch64.h

@@ -1,794 +0,0 @@
-#pragma once
-
-#include <ATen/cpu/vec/intrinsics.h>
-#include <ATen/cpu/vec/vec_base.h>
-#include <c10/macros/Macros.h>
-#include <c10/util/irange.h>
-
-namespace at::vec {
-// Note [CPU_CAPABILITY namespace]
-// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-// This header, and all of its subheaders, will be compiled with
-// different architecture flags for each supported set of vector
-// intrinsics. So we need to make sure they aren't inadvertently
-// linked together. We do this by declaring objects in an `inline
-// namespace` which changes the name mangling, but can still be
-// accessed as `at::vec`.
-inline namespace CPU_CAPABILITY {
-
-#define VEC_INT_NEON_TEMPLATE(vl, bit)                                        \
-  template <>                                                                 \
-  struct is_vec_specialized_for<int##bit##_t> : std::bool_constant<true> {};  \
-                                                                              \
-  template <>                                                                 \
-  class Vectorized<int##bit##_t> {                                            \
-    using neon_type = int##bit##x##vl##_t;                                    \
-                                                                              \
-   private:                                                                   \
-    neon_type values;                                                         \
-                                                                              \
-   public:                                                                    \
-    using value_type = int##bit##_t;                                          \
-    using size_type = int;                                                    \
-    static constexpr size_type size() {                                       \
-      return vl;                                                              \
-    }                                                                         \
-    Vectorized() {                                                            \
-      values = vdupq_n_s##bit(0);                                             \
-    }                                                                         \
-    Vectorized(neon_type v) : values(v) {}                                    \
-    Vectorized(int##bit##_t val);                                             \
-    template <                                                                \
-        typename... Args,                                                     \
-        typename = std::enable_if_t<(sizeof...(Args) == size())>>             \
-    Vectorized(Args... vals) {                                                \
-      __at_align__ int##bit##_t buffer[size()] = {vals...};                   \
-      values = vld1q_s##bit(buffer);                                          \
-    }                                                                         \
-    operator neon_type() const {                                              \
-      return values;                                                          \
-    }                                                                         \
-    static Vectorized<int##bit##_t> loadu(                                    \
-        const void* ptr,                                                      \
-        int64_t count = size());                                              \
-    void store(void* ptr, int64_t count = size()) const;                      \
-    template <int64_t mask>                                                   \
-    static Vectorized<int##bit##_t> blend(                                    \
-        const Vectorized<int##bit##_t>& a,                                    \
-        const Vectorized<int##bit##_t>& b);                                   \
-    static Vectorized<int##bit##_t> blendv(                                   \
-        const Vectorized<int##bit##_t>& a,                                    \
-        const Vectorized<int##bit##_t>& b,                                    \
-        const Vectorized<int##bit##_t>& mask_) {                              \
-      return vbslq_s##bit(vreinterpretq_u##bit##_s##bit(mask_.values), b, a); \
-    }                                                                         \
-    template <typename step_t>                                                \
-    static Vectorized<int##bit##_t> arange(                                   \
-        value_type base = 0,                                                  \
-        step_t step = static_cast<step_t>(1));                                \
-    static Vectorized<int##bit##_t> set(                                      \
-        const Vectorized<int##bit##_t>& a,                                    \
-        const Vectorized<int##bit##_t>& b,                                    \
-        int64_t count = size());                                              \
-    const int##bit##_t& operator[](int idx) const = delete;                   \
-    int##bit##_t& operator[](int idx) = delete;                               \
-    Vectorized<int##bit##_t> abs() const {                                    \
-      return vabsq_s##bit(values);                                            \
-    }                                                                         \
-    Vectorized<int##bit##_t> real() const {                                   \
-      return values;                                                          \
-    }                                                                         \
-    Vectorized<int##bit##_t> imag() const {                                   \
-      return vdupq_n_s##bit(0);                                               \
-    }                                                                         \
-    Vectorized<int##bit##_t> conj() const {                                   \
-      return values;                                                          \
-    }                                                                         \
-    Vectorized<int##bit##_t> neg() const {                                    \
-      return vnegq_s##bit(values);                                            \
-    }                                                                         \
-    int##bit##_t reduce_add() const {                                         \
-      return vaddvq_s##bit(values);                                           \
-    }                                                                         \
-    int##bit##_t reduce_max() const;                                          \
-    Vectorized<int##bit##_t> operator==(                                      \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vceqq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> operator!=(                                      \
-        const Vectorized<int##bit##_t>& other) const;                         \
-    Vectorized<int##bit##_t> operator<(                                       \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vcltq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> operator<=(                                      \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vcleq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> operator>(                                       \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vcgtq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> operator>=(                                      \
-        const Vectorized<int##bit##_t>& other) const {                        \
-      return Vectorized<value_type>(                                          \
-          vreinterpretq_s##bit##_u##bit(vcgeq_s##bit(values, other.values))); \
-    }                                                                         \
-    Vectorized<int##bit##_t> eq(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> ne(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> gt(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> ge(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> lt(const Vectorized<int##bit##_t>& other) const; \
-    Vectorized<int##bit##_t> le(const Vectorized<int##bit##_t>& other) const; \
-  };                                                                          \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator+(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return vaddq_s##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator-(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return vsubq_s##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator&(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return vandq_s##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator|(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return vorrq_s##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<int##bit##_t> inline operator^(                                  \
-      const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
-    return veorq_s##bit(a, b);                                                \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::eq(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this == other) & Vectorized<int##bit##_t>(1);                    \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::ne(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this != other) & Vectorized<int##bit##_t>(1);                    \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::gt(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this > other) & Vectorized<int##bit##_t>(1);                     \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::ge(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this >= other) & Vectorized<int##bit##_t>(1);                    \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::lt(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this < other) & Vectorized<int##bit##_t>(1);                     \
-  }                                                                           \
-  Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::le(               \
-      const Vectorized<int##bit##_t>& other) const {                          \
-    return (*this <= other) & Vectorized<int##bit##_t>(1);                    \
-  }
-
-VEC_INT_NEON_TEMPLATE(2, 64)
-VEC_INT_NEON_TEMPLATE(4, 32)
-VEC_INT_NEON_TEMPLATE(8, 16)
-VEC_INT_NEON_TEMPLATE(16, 8)
-
-inline int32_t Vectorized<int32_t>::reduce_max() const {
-  return vmaxvq_s32(values);
-}
-
-inline int16_t Vectorized<int16_t>::reduce_max() const {
-  return vmaxvq_s16(values);
-}
-
-inline int8_t Vectorized<int8_t>::reduce_max() const {
-  return vmaxvq_s8(values);
-}
-
-template <>
-Vectorized<int32_t> inline operator*(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  return vmulq_s32(a, b);
-}
-
-template <>
-Vectorized<int16_t> inline operator*(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  return vmulq_s16(a, b);
-}
-
-template <>
-Vectorized<int8_t> inline operator*(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  return vmulq_s8(a, b);
-}
-
-template <>
-inline Vectorized<int64_t> operator~(const Vectorized<int64_t>& a) {
-  int64x2_t val = a;
-  return ~val;
-}
-
-template <>
-inline Vectorized<int32_t> operator~(const Vectorized<int32_t>& a) {
-  return vmvnq_s32(a);
-}
-
-template <>
-inline Vectorized<int16_t> operator~(const Vectorized<int16_t>& a) {
-  return vmvnq_s16(a);
-}
-
-template <>
-inline Vectorized<int8_t> operator~(const Vectorized<int8_t>& a) {
-  return vmvnq_s8(a);
-}
-
-inline Vectorized<int64_t> Vectorized<int64_t>::operator!=(
-    const Vectorized<int64_t>& other) const {
-  return ~(*this == other);
-}
-
-inline Vectorized<int32_t> Vectorized<int32_t>::operator!=(
-    const Vectorized<int32_t>& other) const {
-  return ~(*this == other);
-}
-
-inline Vectorized<int16_t> Vectorized<int16_t>::operator!=(
-    const Vectorized<int16_t>& other) const {
-  return ~(*this == other);
-}
-
-inline Vectorized<int8_t> Vectorized<int8_t>::operator!=(
-    const Vectorized<int8_t>& other) const {
-  return ~(*this == other);
-}
-
-template <>
-Vectorized<int32_t> inline minimum(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  return vminq_s32(a, b);
-}
-
-template <>
-Vectorized<int16_t> inline minimum(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  return vminq_s16(a, b);
-}
-
-template <>
-Vectorized<int8_t> inline minimum(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  return vminq_s8(a, b);
-}
-
-template <>
-Vectorized<int32_t> inline maximum(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  return vmaxq_s32(a, b);
-}
-
-template <>
-Vectorized<int16_t> inline maximum(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  return vmaxq_s16(a, b);
-}
-
-template <>
-Vectorized<int8_t> inline maximum(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  return vmaxq_s8(a, b);
-}
-
-template <int64_t mask>
-Vectorized<int64_t> Vectorized<int64_t>::blend(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint64x2_t maskArray = {
-      (mask & 1LL) ? 0xFFFFFFFFFFFFFFFF : 0,
-      (mask & 2LL) ? 0xFFFFFFFFFFFFFFFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_s64(maskArray, b.values, a.values);
-}
-
-template <int64_t mask>
-Vectorized<int32_t> Vectorized<int32_t>::blend(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint32x4_t maskArray = {
-      (mask & 1LL) ? 0xFFFFFFFF : 0,
-      (mask & 2LL) ? 0xFFFFFFFF : 0,
-      (mask & 4LL) ? 0xFFFFFFFF : 0,
-      (mask & 8LL) ? 0xFFFFFFFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_s32(maskArray, b.values, a.values);
-}
-
-template <int64_t mask>
-Vectorized<int16_t> Vectorized<int16_t>::blend(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint16x8_t maskArray = {
-      (mask & 1LL) ? 0xFFFF : 0,
-      (mask & 2LL) ? 0xFFFF : 0,
-      (mask & 4LL) ? 0xFFFF : 0,
-      (mask & 8LL) ? 0xFFFF : 0,
-      (mask & 16LL) ? 0xFFFF : 0,
-      (mask & 32LL) ? 0xFFFF : 0,
-      (mask & 64LL) ? 0xFFFF : 0,
-      (mask & 128LL) ? 0xFFFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_s16(maskArray, b.values, a.values);
-}
-
-template <int64_t mask>
-Vectorized<int8_t> Vectorized<int8_t>::blend(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint8x16_t maskArray = {
-      (mask & 1LL) ? 0xFF : 0,
-      (mask & 2LL) ? 0xFF : 0,
-      (mask & 4LL) ? 0xFF : 0,
-      (mask & 8LL) ? 0xFF : 0,
-      (mask & 16LL) ? 0xFF : 0,
-      (mask & 32LL) ? 0xFF : 0,
-      (mask & 64LL) ? 0xFF : 0,
-      (mask & 128LL) ? 0xFF : 0,
-      (mask & 256LL) ? 0xFF : 0,
-      (mask & 512LL) ? 0xFF : 0,
-      (mask & 1024LL) ? 0xFF : 0,
-      (mask & 2048LL) ? 0xFF : 0,
-      (mask & 4096LL) ? 0xFF : 0,
-      (mask & 8192LL) ? 0xFF : 0,
-      (mask & 16384LL) ? 0xFF : 0,
-      (mask & 32768LL) ? 0xFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_s8(maskArray, b.values, a.values);
-}
-
-#define VEC_INT_NEON_OPS(vl, bit)                                             \
-  inline Vectorized<int##bit##_t>::Vectorized(int##bit##_t val) {             \
-    values = vdupq_n_s##bit(val);                                             \
-  }                                                                           \
-  inline Vectorized<int##bit##_t> Vectorized<int##bit##_t>::loadu(            \
-      const void* ptr, int64_t count) {                                       \
-    if (count == size()) {                                                    \
-      return vld1q_s##bit(reinterpret_cast<const int##bit##_t*>(ptr));        \
-    } else {                                                                  \
-      __at_align__ int##bit##_t tmp_values[size()];                           \
-      for (const auto i : c10::irange(size())) {                              \
-        tmp_values[i] = 0;                                                    \
-      }                                                                       \
-      std::memcpy(                                                            \
-          tmp_values,                                                         \
-          reinterpret_cast<const int##bit##_t*>(ptr),                         \
-          count * sizeof(int##bit##_t));                                      \
-      return vld1q_s##bit(reinterpret_cast<const int##bit##_t*>(tmp_values)); \
-    }                                                                         \
-  }                                                                           \
-  inline void Vectorized<int##bit##_t>::store(void* ptr, int64_t count)       \
-      const {                                                                 \
-    if (count == size()) {                                                    \
-      vst1q_s##bit(reinterpret_cast<int##bit##_t*>(ptr), values);             \
-    } else {                                                                  \
-      int##bit##_t tmp_values[size()];                                        \
-      vst1q_s##bit(reinterpret_cast<int##bit##_t*>(tmp_values), values);      \
-      std::memcpy(ptr, tmp_values, count * sizeof(int##bit##_t));             \
-    }                                                                         \
-  }
-
-VEC_INT_NEON_OPS(2, 64)
-VEC_INT_NEON_OPS(4, 32)
-VEC_INT_NEON_OPS(8, 16)
-VEC_INT_NEON_OPS(16, 8)
-
-template <>
-Vectorized<int64_t> inline operator*(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  return x * y;
-}
-
-template <>
-Vectorized<int64_t> inline operator/(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  return x / y;
-}
-
-template <>
-Vectorized<int32_t> inline operator/(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  int32x4_t x = a;
-  int32x4_t y = b;
-  return x / y;
-}
-
-inline int64_t Vectorized<int64_t>::reduce_max() const {
-  return std::max(values[0], values[1]);
-}
-
-template <>
-Vectorized<int64_t> inline minimum(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  return {std::min(x[0], y[0]), std::min(x[1], y[1])};
-}
-
-template <>
-Vectorized<int64_t> inline maximum(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  return {std::max(x[0], y[0]), std::max(x[1], y[1])};
-}
-
-template <typename step_t>
-inline Vectorized<int64_t> Vectorized<int64_t>::arange(
-    int64_t base,
-    step_t step) {
-  const Vectorized<int64_t> base_vec(base);
-  const Vectorized<int64_t> step_vec(step);
-  const int64x2_t step_sizes = {0, 1};
-  return base_vec.values + step_sizes * step_vec.values;
-}
-
-template <typename step_t>
-inline Vectorized<int32_t> Vectorized<int32_t>::arange(
-    int32_t base,
-    step_t step) {
-  const Vectorized<int32_t> base_vec(base);
-  const Vectorized<int32_t> step_vec(step);
-  const int32x4_t step_sizes = {0, 1, 2, 3};
-  return vmlaq_s32(base_vec, step_sizes, step_vec);
-}
-
-template <typename step_t>
-inline Vectorized<int16_t> Vectorized<int16_t>::arange(
-    int16_t base,
-    step_t step) {
-  const Vectorized<int16_t> base_vec(base);
-  const Vectorized<int16_t> step_vec(step);
-  const int16x8_t step_sizes = {0, 1, 2, 3, 4, 5, 6, 7};
-  return vmlaq_s16(base_vec, step_sizes, step_vec);
-}
-
-template <typename step_t>
-inline Vectorized<int8_t> Vectorized<int8_t>::arange(int8_t base, step_t step) {
-  const Vectorized<int8_t> base_vec(base);
-  const Vectorized<int8_t> step_vec(step);
-  const int8x16_t step_sizes = {
-      0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
-  return vmlaq_s8(base_vec, step_sizes, step_vec);
-}
-
-template <>
-Vectorized<int64_t> inline operator>>(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t x = a;
-  int64x2_t y = b;
-  uint64x2_t u = vreinterpretq_u64_s64(y);
-  uint64x2_t z = {std::min(u[0], (uint64_t)63), std::min(u[1], (uint64_t)63)};
-  return x >> vreinterpretq_s64_u64(z);
-}
-
-template <>
-Vectorized<int32_t> inline operator>>(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  int32x4_t x = a;
-  int32x4_t y = b;
-  uint32x4_t bound = vdupq_n_u32(31);
-  uint32x4_t z = vminq_u32(vreinterpretq_u32_s32(y), bound);
-  return x >> vreinterpretq_s32_u32(z);
-}
-
-template <>
-Vectorized<int16_t> inline operator>>(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  int16x8_t x = a;
-  int16x8_t y = b;
-  uint16x8_t bound = vdupq_n_u16(15);
-  uint16x8_t z = vminq_u16(vreinterpretq_u16_s16(y), bound);
-  return x >> vreinterpretq_s16_u16(z);
-}
-
-template <>
-Vectorized<int8_t> inline operator>>(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  int8x16_t x = a;
-  int8x16_t y = b;
-  uint8x16_t bound = vdupq_n_u8(7);
-  int8x16_t z = vreinterpretq_s8_u8(vminq_u8(vreinterpretq_u8_s8(y), bound));
-  return x >> z;
-}
-
-template <>
-Vectorized<int64_t> inline operator<<(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b) {
-  int64x2_t y = b;
-  uint64x2_t u = vreinterpretq_u64_s64(y);
-  uint64x2_t z = {std::min(u[0], (uint64_t)64), std::min(u[1], (uint64_t)64)};
-  return vshlq_s64(a, vreinterpretq_s64_u64(z));
-}
-
-template <>
-Vectorized<int32_t> inline operator<<(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b) {
-  int32x4_t y = b;
-  uint32x4_t bound = vdupq_n_u32(32);
-  uint32x4_t z = vminq_u32(vreinterpretq_u32_s32(y), bound);
-  return vshlq_s32(a, vreinterpretq_s32_u32(z));
-}
-
-template <>
-Vectorized<int16_t> inline operator<<(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  int16x8_t y = b;
-  uint16x8_t bound = vdupq_n_u16(16);
-  uint16x8_t z = vminq_u16(vreinterpretq_u16_s16(y), bound);
-  return vshlq_s16(a, vreinterpretq_s16_u16(z));
-}
-
-template <>
-Vectorized<int8_t> inline operator<<(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  int8x16_t y = b;
-  uint8x16_t bound = vdupq_n_u8(8);
-  int8x16_t z = vreinterpretq_s8_u8(vminq_u8(vreinterpretq_u8_s8(y), bound));
-  return vshlq_s8(a, z);
-}
-
-inline Vectorized<int64_t> Vectorized<int64_t>::set(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& b,
-    int64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 2) {
-    return b;
-  } else {
-    int64x2_t c = {b.values[0], a.values[1]};
-    return c;
-  }
-}
-
-inline Vectorized<int32_t> Vectorized<int32_t>::set(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& b,
-    int64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 4) {
-    return b;
-  } else {
-    // Build an array of flags: each bit of element is 1 if the corresponding
-    // bit in 'mask' is set, 0 otherwise.
-    uint32x4_t maskArray = {
-        (count >= 1LL) ? 0xFFFFFFFF : 0,
-        (count >= 2LL) ? 0xFFFFFFFF : 0,
-        (count >= 3LL) ? 0xFFFFFFFF : 0,
-        0};
-    // Use BSL to select elements from b where the mask is 1, else from a
-    return vbslq_s32(maskArray, b.values, a.values);
-  }
-}
-
-inline Vectorized<int16_t> Vectorized<int16_t>::set(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b,
-    int64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 8) {
-    return b;
-  } else {
-    // Build an array of flags: each bit of element is 1 if the corresponding
-    // bit in 'mask' is set, 0 otherwise.
-    uint16x8_t maskArray = {
-        static_cast<uint16_t>((count >= 1LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 2LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 3LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 4LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 5LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 6LL) ? 0xFFFF : 0),
-        static_cast<uint16_t>((count >= 7LL) ? 0xFFFF : 0),
-        0};
-    // Use BSL to select elements from b where the mask is 1, else from a
-    return vbslq_s16(maskArray, b.values, a.values);
-  }
-}
-
-inline Vectorized<int8_t> Vectorized<int8_t>::set(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b,
-    int64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 16) {
-    return b;
-  } else {
-    // Build an array of flags: each bit of element is 1 if the corresponding
-    // bit in 'mask' is set, 0 otherwise.
-    uint8x16_t maskArray = {
-        static_cast<uint8_t>((count >= 1LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 2LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 3LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 4LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 5LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 6LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 7LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 8LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 9LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 10LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 11LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 12LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 13LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 14LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 15LL) ? 0xFF : 0),
-        0};
-
-    // Use BSL to select elements from b where the mask is 1, else from a
-    return vbslq_s8(maskArray, b.values, a.values);
-  }
-}
-
-template <>
-Vectorized<int16_t> inline operator/(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& b) {
-  Vectorized<int32_t> highBitsA = vmovl_high_s16(a);
-  Vectorized<int32_t> highBitsB = vmovl_high_s16(b);
-  Vectorized<int32_t> lowBitsA = vmovl_s16(vget_low_s16(a));
-  Vectorized<int32_t> lowBitsB = vmovl_s16(vget_low_s16(b));
-  int32x4_t highBitsResult = highBitsA / highBitsB;
-  int32x4_t lowBitsResult = lowBitsA / lowBitsB;
-  return vuzp1q_s16(
-      vreinterpretq_s16_s32(lowBitsResult),
-      vreinterpretq_s16_s32(highBitsResult));
-}
-
-template <>
-Vectorized<int8_t> inline operator/(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& b) {
-  Vectorized<int16_t> highBitsA = vmovl_high_s8(a);
-  Vectorized<int16_t> highBitsB = vmovl_high_s8(b);
-  Vectorized<int16_t> lowBitsA = vmovl_s8(vget_low_s8(a));
-  Vectorized<int16_t> lowBitsB = vmovl_s8(vget_low_s8(b));
-  int16x8_t highBitsResult = highBitsA / highBitsB;
-  int16x8_t lowBitsResult = lowBitsA / lowBitsB;
-  return vuzp1q_s8(
-      vreinterpretq_s8_s16(lowBitsResult),
-      vreinterpretq_s8_s16(highBitsResult));
-}
-
-template <>
-Vectorized<int64_t> inline clamp(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& min,
-    const Vectorized<int64_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<int32_t> inline clamp(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& min,
-    const Vectorized<int32_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<int16_t> inline clamp(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& min,
-    const Vectorized<int16_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<int8_t> inline clamp(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& min,
-    const Vectorized<int8_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<int64_t> inline clamp_max(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<int32_t> inline clamp_max(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<int16_t> inline clamp_max(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<int8_t> inline clamp_max(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<int64_t> inline clamp_min(
-    const Vectorized<int64_t>& a,
-    const Vectorized<int64_t>& min) {
-  return maximum(min, a);
-}
-
-template <>
-Vectorized<int32_t> inline clamp_min(
-    const Vectorized<int32_t>& a,
-    const Vectorized<int32_t>& min) {
-  return maximum(min, a);
-}
-
-template <>
-Vectorized<int16_t> inline clamp_min(
-    const Vectorized<int16_t>& a,
-    const Vectorized<int16_t>& min) {
-  return maximum(min, a);
-}
-
-template <>
-Vectorized<int8_t> inline clamp_min(
-    const Vectorized<int8_t>& a,
-    const Vectorized<int8_t>& min) {
-  return maximum(min, a);
-}
-
-} // namespace CPU_CAPABILITY
-} // namespace at::vec

aten/src/ATen/cpu/vec/vec128/vec128_uint_aarch64.h

@@ -1,378 +0,0 @@
-#pragma once
-
-#include <ATen/cpu/vec/intrinsics.h>
-#include <ATen/cpu/vec/vec_base.h>
-#include <c10/macros/Macros.h>
-#include <c10/util/irange.h>
-
-namespace at::vec {
-// Note [CPU_CAPABILITY namespace]
-// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-// This header, and all of its subheaders, will be compiled with
-// different architecture flags for each supported set of vector
-// intrinsics. So we need to make sure they aren't inadvertently
-// linked together. We do this by declaring objects in an `inline
-// namespace` which changes the name mangling, but can still be
-// accessed as `at::vec`.
-inline namespace CPU_CAPABILITY {
-
-#define VEC_UINT_NEON_TEMPLATE(vl, bit)                                       \
-  template <>                                                                 \
-  struct is_vec_specialized_for<uint##bit##_t> : std::bool_constant<true> {}; \
-                                                                              \
-  template <>                                                                 \
-  class Vectorized<uint##bit##_t> {                                           \
-    using neon_type = uint##bit##x##vl##_t;                                   \
-                                                                              \
-   private:                                                                   \
-    neon_type values;                                                         \
-                                                                              \
-   public:                                                                    \
-    using value_type = uint##bit##_t;                                         \
-    using size_type = int;                                                    \
-    static constexpr size_type size() {                                       \
-      return vl;                                                              \
-    }                                                                         \
-    Vectorized() {                                                            \
-      values = vdupq_n_u##bit(0);                                             \
-    }                                                                         \
-    Vectorized(neon_type v) : values(v) {}                                    \
-    Vectorized(uint##bit##_t val);                                            \
-    template <                                                                \
-        typename... Args,                                                     \
-        typename = std::enable_if_t<(sizeof...(Args) == size())>>             \
-    Vectorized(Args... vals) {                                                \
-      __at_align__ uint##bit##_t buffer[size()] = {vals...};                  \
-      values = vld1q_u##bit(buffer);                                          \
-    }                                                                         \
-    operator neon_type() const {                                              \
-      return values;                                                          \
-    }                                                                         \
-    static Vectorized<uint##bit##_t> loadu(                                   \
-        const void* ptr,                                                      \
-        uint64_t count = size());                                             \
-    void store(void* ptr, uint64_t count = size()) const;                     \
-    template <uint64_t mask>                                                  \
-    static Vectorized<uint##bit##_t> blend(                                   \
-        const Vectorized<uint##bit##_t>& a,                                   \
-        const Vectorized<uint##bit##_t>& b);                                  \
-    static Vectorized<uint##bit##_t> blendv(                                  \
-        const Vectorized<uint##bit##_t>& a,                                   \
-        const Vectorized<uint##bit##_t>& b,                                   \
-        const Vectorized<uint##bit##_t>& mask_) {                             \
-      return vbslq_u##bit(mask_.values, b, a);                                \
-    }                                                                         \
-    template <typename step_t>                                                \
-    static Vectorized<uint##bit##_t> arange(                                  \
-        value_type base = 0,                                                  \
-        step_t step = static_cast<step_t>(1));                                \
-    static Vectorized<uint##bit##_t> set(                                     \
-        const Vectorized<uint##bit##_t>& a,                                   \
-        const Vectorized<uint##bit##_t>& b,                                   \
-        uint64_t count = size());                                             \
-    const uint##bit##_t& operator[](uint idx) const = delete;                 \
-    uint##bit##_t& operator[](uint idx) = delete;                             \
-    Vectorized<uint##bit##_t> abs() const {                                   \
-      return values;                                                          \
-    }                                                                         \
-    Vectorized<uint##bit##_t> real() const {                                  \
-      return values;                                                          \
-    }                                                                         \
-    Vectorized<uint##bit##_t> imag() const {                                  \
-      return vdupq_n_u##bit(0);                                               \
-    }                                                                         \
-    Vectorized<uint##bit##_t> conj() const {                                  \
-      return values;                                                          \
-    }                                                                         \
-    Vectorized<uint##bit##_t> neg() const {                                   \
-      return vreinterpretq_u##bit##_s##bit(                                   \
-          vnegq_s##bit(vreinterpretq_s##bit##_u##bit(values)));               \
-    }                                                                         \
-    uint##bit##_t reduce_add() const {                                        \
-      return vaddvq_u##bit(values);                                           \
-    }                                                                         \
-    uint##bit##_t reduce_max() const;                                         \
-    Vectorized<uint##bit##_t> operator==(                                     \
-        const Vectorized<uint##bit##_t>& other) const {                       \
-      return Vectorized<value_type>(vceqq_u##bit(values, other.values));      \
-    }                                                                         \
-    Vectorized<uint##bit##_t> operator!=(                                     \
-        const Vectorized<uint##bit##_t>& other) const;                        \
-    Vectorized<uint##bit##_t> operator<(                                      \
-        const Vectorized<uint##bit##_t>& other) const {                       \
-      return Vectorized<value_type>(vcltq_u##bit(values, other.values));      \
-    }                                                                         \
-    Vectorized<uint##bit##_t> operator<=(                                     \
-        const Vectorized<uint##bit##_t>& other) const {                       \
-      return Vectorized<value_type>(vcleq_u##bit(values, other.values));      \
-    }                                                                         \
-    Vectorized<uint##bit##_t> operator>(                                      \
-        const Vectorized<uint##bit##_t>& other) const {                       \
-      return Vectorized<value_type>(vcgtq_u##bit(values, other.values));      \
-    }                                                                         \
-    Vectorized<uint##bit##_t> operator>=(                                     \
-        const Vectorized<uint##bit##_t>& other) const {                       \
-      return Vectorized<value_type>(vcgeq_u##bit(values, other.values));      \
-    }                                                                         \
-    Vectorized<uint##bit##_t> eq(                                             \
-        const Vectorized<uint##bit##_t>& other) const;                        \
-    Vectorized<uint##bit##_t> ne(                                             \
-        const Vectorized<uint##bit##_t>& other) const;                        \
-    Vectorized<uint##bit##_t> gt(                                             \
-        const Vectorized<uint##bit##_t>& other) const;                        \
-    Vectorized<uint##bit##_t> ge(                                             \
-        const Vectorized<uint##bit##_t>& other) const;                        \
-    Vectorized<uint##bit##_t> lt(                                             \
-        const Vectorized<uint##bit##_t>& other) const;                        \
-    Vectorized<uint##bit##_t> le(                                             \
-        const Vectorized<uint##bit##_t>& other) const;                        \
-  };                                                                          \
-  template <>                                                                 \
-  Vectorized<uint##bit##_t> inline operator+(                                 \
-      const Vectorized<uint##bit##_t>& a,                                     \
-      const Vectorized<uint##bit##_t>& b) {                                   \
-    return vaddq_u##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<uint##bit##_t> inline operator-(                                 \
-      const Vectorized<uint##bit##_t>& a,                                     \
-      const Vectorized<uint##bit##_t>& b) {                                   \
-    return vsubq_u##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<uint##bit##_t> inline operator&(                                 \
-      const Vectorized<uint##bit##_t>& a,                                     \
-      const Vectorized<uint##bit##_t>& b) {                                   \
-    return vandq_u##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<uint##bit##_t> inline operator|(                                 \
-      const Vectorized<uint##bit##_t>& a,                                     \
-      const Vectorized<uint##bit##_t>& b) {                                   \
-    return vorrq_u##bit(a, b);                                                \
-  }                                                                           \
-  template <>                                                                 \
-  Vectorized<uint##bit##_t> inline operator^(                                 \
-      const Vectorized<uint##bit##_t>& a,                                     \
-      const Vectorized<uint##bit##_t>& b) {                                   \
-    return veorq_u##bit(a, b);                                                \
-  }                                                                           \
-  Vectorized<uint##bit##_t> inline Vectorized<uint##bit##_t>::eq(             \
-      const Vectorized<uint##bit##_t>& other) const {                         \
-    return (*this == other) & Vectorized<uint##bit##_t>(1);                   \
-  }                                                                           \
-  Vectorized<uint##bit##_t> inline Vectorized<uint##bit##_t>::ne(             \
-      const Vectorized<uint##bit##_t>& other) const {                         \
-    return (*this != other) & Vectorized<uint##bit##_t>(1);                   \
-  }                                                                           \
-  Vectorized<uint##bit##_t> inline Vectorized<uint##bit##_t>::gt(             \
-      const Vectorized<uint##bit##_t>& other) const {                         \
-    return (*this > other) & Vectorized<uint##bit##_t>(1);                    \
-  }                                                                           \
-  Vectorized<uint##bit##_t> inline Vectorized<uint##bit##_t>::ge(             \
-      const Vectorized<uint##bit##_t>& other) const {                         \
-    return (*this >= other) & Vectorized<uint##bit##_t>(1);                   \
-  }                                                                           \
-  Vectorized<uint##bit##_t> inline Vectorized<uint##bit##_t>::lt(             \
-      const Vectorized<uint##bit##_t>& other) const {                         \
-    return (*this < other) & Vectorized<uint##bit##_t>(1);                    \
-  }                                                                           \
-  Vectorized<uint##bit##_t> inline Vectorized<uint##bit##_t>::le(             \
-      const Vectorized<uint##bit##_t>& other) const {                         \
-    return (*this <= other) & Vectorized<uint##bit##_t>(1);                   \
-  }
-
-VEC_UINT_NEON_TEMPLATE(16, 8)
-
-inline uint8_t Vectorized<uint8_t>::reduce_max() const {
-  return vmaxvq_u8(values);
-}
-
-template <>
-Vectorized<uint8_t> inline operator*(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& b) {
-  return vmulq_u8(a, b);
-}
-
-template <>
-inline Vectorized<uint8_t> operator~(const Vectorized<uint8_t>& a) {
-  return vmvnq_u8(a);
-}
-
-inline Vectorized<uint8_t> Vectorized<uint8_t>::operator!=(
-    const Vectorized<uint8_t>& other) const {
-  return ~(*this == other);
-}
-
-template <>
-Vectorized<uint8_t> inline minimum(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& b) {
-  return vminq_u8(a, b);
-}
-
-template <>
-Vectorized<uint8_t> inline maximum(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& b) {
-  return vmaxq_u8(a, b);
-}
-
-template <uint64_t mask>
-Vectorized<uint8_t> Vectorized<uint8_t>::blend(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& b) {
-  // Build an array of flags: each bit of element is 1 if the corresponding bit
-  // in 'mask' is set, 0 otherwise.
-  uint8x16_t maskArray = {
-      (mask & 1LL) ? 0xFF : 0,
-      (mask & 2LL) ? 0xFF : 0,
-      (mask & 4LL) ? 0xFF : 0,
-      (mask & 8LL) ? 0xFF : 0,
-      (mask & 16LL) ? 0xFF : 0,
-      (mask & 32LL) ? 0xFF : 0,
-      (mask & 64LL) ? 0xFF : 0,
-      (mask & 128LL) ? 0xFF : 0,
-      (mask & 256LL) ? 0xFF : 0,
-      (mask & 512LL) ? 0xFF : 0,
-      (mask & 1024LL) ? 0xFF : 0,
-      (mask & 2048LL) ? 0xFF : 0,
-      (mask & 4096LL) ? 0xFF : 0,
-      (mask & 8192LL) ? 0xFF : 0,
-      (mask & 16384LL) ? 0xFF : 0,
-      (mask & 32768LL) ? 0xFF : 0};
-  // Use BSL to select elements from b where the mask is 1, else from a
-  return vbslq_u8(maskArray, b.values, a.values);
-}
-
-#define VEC_UINT_NEON_OPS(vl, bit)                                             \
-  inline Vectorized<uint##bit##_t>::Vectorized(uint##bit##_t val) {            \
-    values = vdupq_n_u##bit(val);                                              \
-  }                                                                            \
-  inline Vectorized<uint##bit##_t> Vectorized<uint##bit##_t>::loadu(           \
-      const void* ptr, uint64_t count) {                                       \
-    if (count == size()) {                                                     \
-      return vld1q_u##bit(reinterpret_cast<const uint##bit##_t*>(ptr));        \
-    } else {                                                                   \
-      __at_align__ uint##bit##_t tmp_values[size()];                           \
-      for (const auto i : c10::irange(size())) {                               \
-        tmp_values[i] = 0;                                                     \
-      }                                                                        \
-      std::memcpy(                                                             \
-          tmp_values,                                                          \
-          reinterpret_cast<const uint##bit##_t*>(ptr),                         \
-          count * sizeof(uint##bit##_t));                                      \
-      return vld1q_u##bit(reinterpret_cast<const uint##bit##_t*>(tmp_values)); \
-    }                                                                          \
-  }                                                                            \
-  inline void Vectorized<uint##bit##_t>::store(void* ptr, uint64_t count)      \
-      const {                                                                  \
-    if (count == size()) {                                                     \
-      vst1q_u##bit(reinterpret_cast<uint##bit##_t*>(ptr), values);             \
-    } else {                                                                   \
-      uint##bit##_t tmp_values[size()];                                        \
-      vst1q_u##bit(reinterpret_cast<uint##bit##_t*>(tmp_values), values);      \
-      std::memcpy(ptr, tmp_values, count * sizeof(uint##bit##_t));             \
-    }                                                                          \
-  }
-
-VEC_UINT_NEON_OPS(16, 8)
-
-template <typename step_t>
-inline Vectorized<uint8_t> Vectorized<uint8_t>::arange(
-    uint8_t base,
-    step_t step) {
-  const Vectorized<uint8_t> base_vec(base);
-  const Vectorized<uint8_t> step_vec(step);
-  const uint8x16_t step_sizes = {
-      0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
-  return vmlaq_u8(base_vec, step_sizes, step_vec);
-}
-
-template <>
-Vectorized<uint8_t> inline operator>>(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& b) {
-  uint8x16_t x = a;
-  uint8x16_t bound = vdupq_n_u8(8);
-  uint8x16_t z = vminq_u8(b, bound);
-  return x >> z;
-}
-
-template <>
-Vectorized<uint8_t> inline operator<<(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& b) {
-  uint8x16_t bound = vdupq_n_u8(8);
-  uint8x16_t z = vminq_u8(b, bound);
-  return vshlq_u8(a, vreinterpretq_s8_u8(z));
-}
-
-inline Vectorized<uint8_t> Vectorized<uint8_t>::set(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& b,
-    uint64_t count) {
-  if (count == 0) {
-    return a;
-  } else if (count >= 16) {
-    return b;
-  } else {
-    // Build an array of flags: each bit of element is 1 if the corresponding
-    // bit in 'mask' is set, 0 otherwise.
-    uint8x16_t maskArray = {
-        static_cast<uint8_t>((count >= 1LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 2LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 3LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 4LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 5LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 6LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 7LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 8LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 9LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 10LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 11LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 12LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 13LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 14LL) ? 0xFF : 0),
-        static_cast<uint8_t>((count >= 15LL) ? 0xFF : 0),
-        0};
-
-    // Use BSL to select elements from b where the mask is 1, else from a
-    return vbslq_u8(maskArray, b.values, a.values);
-  }
-}
-
-template <>
-Vectorized<uint8_t> inline operator/(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& b) {
-  uint8x16_t x = a;
-  uint8x16_t y = b;
-  return x / y;
-}
-
-template <>
-Vectorized<uint8_t> inline clamp(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& min,
-    const Vectorized<uint8_t>& max) {
-  return minimum(max, maximum(min, a));
-}
-
-template <>
-Vectorized<uint8_t> inline clamp_max(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& max) {
-  return minimum(max, a);
-}
-
-template <>
-Vectorized<uint8_t> inline clamp_min(
-    const Vectorized<uint8_t>& a,
-    const Vectorized<uint8_t>& min) {
-  return maximum(min, a);
-}
-
-} // namespace CPU_CAPABILITY
-} // namespace at::vec

aten/src/ATen/cpu/vec/vec256/vec256_int.h

@@ -1740,7 +1740,7 @@ Vectorized<int16_t> inline shift_256_16(
 
   // Control masks for shuffle operation, treating 256 bits as an
   // array of 16-bit elements, and considering pairs of neighboring
-  // elements.  Specifically, a mask named "ctl_M_N" (M,N in [0,1], and
+  // elements.  Specifially, a mask named "ctl_M_N" (M,N in [0,1], and
   // M!=N) is set so that shuffle will move element with index M from
   // input pair into element with index N in output pair, and element
   // with index M in output pair will be set to all 0s.
@@ -1875,7 +1875,7 @@ Vectorized<T> inline shift_256_8(
 
   // Control masks for shuffle operation, treating 256 bits as an
   // array of 8-bit elements, and considering quadruples of
-  // neighboring elements.  Specifically, a mask named "ctl_M_N" (M,N
+  // neighboring elements.  Specifially, a mask named "ctl_M_N" (M,N
   // in [0,1,2,3], and M!=N) is set so that shuffle will move element
   // with index M from input quadruple into element with index N in
   // output quadruple, and other elements in output quadruple will be

aten/src/ATen/cpu/vec/vec256/vec256_qint.h

@@ -1377,7 +1377,7 @@ Vectorized<c10::quint8> inline maximum(
 #if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
 std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
     at::vec::Vectorized<int8_t> src) {
-  auto s8x8 = vget_low_s8(src);
+  auto s8x8 = vld1_s8(src.operator const int8_t*());
   auto s16x8 = vmovl_s8(s8x8);
 
   auto s32x4_hi = vmovl_s16(vget_high_s16(s16x8));
@@ -1390,7 +1390,7 @@ std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
 
 std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
     at::vec::Vectorized<uint8_t> src) {
-  auto u8x8 = vget_low_u8(src);
+  auto u8x8 = vld1_u8(src.operator const uint8_t*());
   auto u16x8 = vmovl_u8(u8x8);
   auto u32x4_hi = vmovl_u16(vget_high_u16(u16x8));
   auto u32x4_lo = vmovl_u16(vget_low_u16(u16x8));
@@ -1402,7 +1402,7 @@ std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
 
 Vectorized<float> inline convert_int8_half_register_to_float(
     at::vec::Vectorized<int8_t> src) {
-  auto s8x8 = vget_low_s8(src);
+  auto s8x8 = vld1_s8(src.operator const int8_t*());
   auto s16x8 = vmovl_s8(s8x8);
 
   auto s32x4_lo = vmovl_s16(vget_low_s16(s16x8));
@@ -1412,7 +1412,7 @@ Vectorized<float> inline convert_int8_half_register_to_float(
 
 Vectorized<float> inline convert_int8_half_register_to_float(
     at::vec::Vectorized<uint8_t> src) {
-  auto u8x8 = vget_low_u8(src);
+  auto u8x8 = vld1_u8(src.operator const uint8_t*());
   auto u16x8 = vmovl_u8(u8x8);
   auto u32x4_lo = vmovl_u16(vget_low_u16(u16x8));
 

aten/src/ATen/cpu/vec/vec256/vsx/vec256_double_vsx.h

@@ -143,7 +143,7 @@ class Vectorized<double> {
       const Vectorized<double>& a,
       const Vectorized<double>& b,
       const Vectorized<double>& mask) {
-    // the mask used here returned by comparison of vec256
+    // the mask used here returned by comparision of vec256
 
     return {
         vec_sel(a._vec0, b._vec0, mask._vecb0),

aten/src/ATen/cpu/vec/vec256/vsx/vec256_float_vsx.h

@@ -142,7 +142,7 @@ class Vectorized<float> {
       const Vectorized<float>& a,
       const Vectorized<float>& b,
       const Vectorized<float>& mask) {
-    // the mask used here returned by comparison of vec256
+    // the mask used here returned by comparision of vec256
     // assuming this we can use the same mask directly with vec_sel
     return {
         vec_sel(a._vec0, b._vec0, mask._vecb0),

aten/src/ATen/cpu/vec/vec256/vsx/vec256_int16_vsx.h

@@ -202,7 +202,7 @@ class Vectorized<int16_t> {
       const Vectorized<int16_t>& a,
       const Vectorized<int16_t>& b,
       const Vectorized<int16_t>& mask) {
-    // the mask used here returned by comparison of vec256
+    // the mask used here returned by comparision of vec256
     // assuming this we can use the same mask directly with vec_sel
     // warning intel style mask will not work properly
     return {

aten/src/ATen/cpu/vec/vec256/vsx/vec256_int32_vsx.h

@@ -155,7 +155,7 @@ class Vectorized<int32_t> {
       const Vectorized<int32_t>& a,
       const Vectorized<int32_t>& b,
       const Vectorized<int32_t>& mask) {
-    // the mask used here returned by comparison of vec256
+    // the mask used here returned by comparision of vec256
     // assuming this we can use the same mask directly with vec_sel
     // warning intel style mask will not work properly
     return {

aten/src/ATen/cpu/vec/vec256/vsx/vec256_int64_vsx.h

@@ -119,7 +119,7 @@ class Vectorized<int64_t> {
       const Vectorized<int64_t>& a,
       const Vectorized<int64_t>& b,
       const Vectorized<int64_t>& mask) {
-    // the mask used here returned by comparison of vec256
+    // the mask used here returned by comparision of vec256
 
     return {
         vec_sel(a._vec0, b._vec0, mask._vecb0),

aten/src/ATen/cpu/vec/vec512/vec512.h

@@ -397,7 +397,7 @@ inline Vectorized<bool> operator&&(
   const __m512i* other_ = reinterpret_cast<const __m512i*>(other.as_bytes());
   __m512i out = _mm512_and_si512(*self_, *other_);
   Vectorized<bool> ret;
-  // We do not have a constructor that takes __m512i, so we need to memcpy
+  // We do not have a constructer that takes __m512i, so we need to memcpy
   std::memcpy(ret, &out, ret.size() * sizeof(bool));
   return ret;
 }

aten/src/ATen/cpu/vec/vec512/vec512_int.h

@@ -1852,7 +1852,7 @@ Vectorized<T> inline shift_512_8(
 
   // Control masks for shuffle operation, treating 512 bits as an
   // array of 8-bit elements, and considering pairs of neighboring
-  // elements.  Specifically, a mask named "ctl_M_N" (M,N in [0,1], and
+  // elements.  Specifially, a mask named "ctl_M_N" (M,N in [0,1], and
   // M!=N) is set so that shuffle will move element with index M from
   // input pair into element with index N in output pair, and element
   // with index M in output pair will be set to all 0s.