revert expectedFailure on TestFSDPWithEP::test_e2e

test_gather_object* require exact world size (4) to succeed
Likely, all NVSHMEM Triton tests would fail. Skipping all for now.

.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_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/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/pytorch-docstring.md

@@ -1,354 +0,0 @@
-# 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.

.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/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
+1b013f5b5a87a1882eb143c26d79d091150d6a37

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-1752fe6809b74921644866275ab80244b96e80bc
+faffd5cf673615583da6517275e361cb3dbc77e6

.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

@@ -33,7 +33,6 @@ 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

@@ -79,21 +79,21 @@ 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==13.0.48; platform_system == 'Linux' | "
@@ -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/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 }}
@@ -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 }}
@@ -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 }}
@@ -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 }}
@@ -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 }}
@@ -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 }}
@@ -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 }}

.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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_10-cuda12_9-test:  # Testing
@@ -925,7 +925,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_11-cuda12_9-test:  # Testing
@@ -1591,7 +1591,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_12-cuda12_9-test:  # Testing
@@ -2257,7 +2257,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13-cuda12_9-test:  # Testing
@@ -2923,7 +2923,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_13t-cuda12_9-test:  # Testing
@@ -3589,7 +3589,7 @@ jobs:
       runner_prefix: "${{ 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: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14-cuda12_9-test:  # Testing
@@ -4255,7 +4255,7 @@ jobs:
       runner_prefix: "${{ 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

.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

@@ -30,9 +30,9 @@ 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
@@ -43,36 +43,12 @@ jobs:
         ]}
     secrets: inherit
 
-  x86-opbenchmark-test:
-    name: x86-opbenchmark-test
+  opbenchmark-test:
+    name: opbenchmark-test
     uses: ./.github/workflows/_linux-test.yml
-    needs: x86-opbenchmark-build
+    needs: 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
-    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
-      test-matrix: |
-        { include: [
-          { config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.arm64.m8g.4xlarge" },
-        ]}
-    secrets: inherit
-
-  aarch64-opbenchmark-test:
-    name: aarch64-opbenchmark-test
-    uses: ./.github/workflows/_linux-test.yml
-    needs: aarch64-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 }}
+      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 = [
@@ -1138,8 +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

aten/src/ATen/CMakeLists.txt

@@ -289,15 +289,14 @@ 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
       ${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
     )
@@ -314,14 +313,13 @@ IF(USE_FBGEMM_GENAI)
 
     # 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)
-    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)
-      endif()
 
     # Only compile for gfx942 for now.
     # This is rather hacky, I could not figure out a clean solution :(

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/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_;
 };
@@ -122,7 +122,7 @@ 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_
@@ -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/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));

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/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:
@@ -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);
   }
 
@@ -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/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/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

@@ -5,114 +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
-CONVERT_TEMPLATE(float16_t, uint8_t)
-CONVERT_TEMPLATE(float16_t, int8_t)
-CONVERT_TEMPLATE(float16_t, int16_t)
-CONVERT_TEMPLATE(float16_t, int32_t)
-CONVERT_TEMPLATE(float16_t, int64_t)
-CONVERT_TEMPLATE(float16_t, float16_t)
-CONVERT_TEMPLATE(float16_t, float)
-CONVERT_TEMPLATE(float16_t, double)
-CONVERT_TEMPLATE(uint8_t, float16_t)
-CONVERT_TEMPLATE(int8_t, float16_t)
-CONVERT_TEMPLATE(int16_t, float16_t)
-CONVERT_TEMPLATE(int32_t, float16_t)
-CONVERT_TEMPLATE(int64_t, float16_t)
-CONVERT_TEMPLATE(float, float16_t)
-CONVERT_TEMPLATE(double, float16_t)
-#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

@@ -540,6 +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,

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

@@ -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_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/cuda/CUDABlas.cpp

@@ -16,8 +16,6 @@
 #include <c10/util/irange.h>
 #include <c10/core/ScalarType.h>
 
-#include <ATen/cuda/detail/BLASConstants.h>
-
 #ifdef USE_ROCM
 #include <c10/cuda/CUDAStream.h>
 #include <hipblaslt/hipblaslt-ext.hpp>
@@ -1956,15 +1954,13 @@ void scaled_gemm(
     const void *result_scale_ptr,
     int64_t result_ld,
     ScalarType result_dtype,
-    bool use_fast_accum,
-    const std::optional<Tensor>& alpha) {
+    bool use_fast_accum) {
   // Note: see `cublasCommonArgs` for various non-intuitive manupulations
   // of input arguments to this function.
   const auto computeType = CUBLAS_COMPUTE_32F;
   const auto scaleType = CUDA_R_32F;
-  // Note: alpha_val may change later depending on user-passed argument
-  float alpha_val = 1.0;
-  float beta_val = 0.0;
+  const float alpha_val = 1.0;
+  const float beta_val = 0.0;
   CuBlasLtMatmulDescriptor computeDesc(computeType, scaleType);
   computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_TRANSA, _cublasOpFromChar(transa));
   computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_TRANSB, _cublasOpFromChar(transb));
@@ -2035,33 +2031,6 @@ void scaled_gemm(
     computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_EPILOGUE, CUBLASLT_EPILOGUE_BIAS);
     computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_BIAS_DATA_TYPE, ScalarTypeToCudaDataType(bias_dtype));
   }
-
-  // Handle user-passed alpha
-  float *alpha_ptr = &alpha_val;
-  float *beta_ptr = &beta_val;
-
-  if (alpha.has_value()) {
-    auto& a = alpha.value();
-
-    // if device-tensor
-    if (a.is_cuda()) {
-      // NOTE: there are lifetime requirements on device-side pointers for alpha/beta -- the value must be
-      //       valid & correct until the cublas call finishes (not is scheduled like host-side values). Thus
-      //       we need to use allocations for alpha/beta that have some guarantees on lifetime - a statically
-      //       managed 4B buffer for alpha that we'll copy the passed alpha value into, and constant memory
-      //       for beta respectively.
-      float *user_alpha_ptr = at::cuda::detail::get_user_alpha_ptr();
-      at::Tensor user_alpha = at::from_blob(user_alpha_ptr, {1}, TensorOptions().device(kCUDA).dtype(kFloat));
-      user_alpha.copy_(a);
-      // Tell cublasLt we're using device-side pointers for alpha/beta
-      auto pointer_mode = CUBLASLT_POINTER_MODE_DEVICE;
-      computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_POINTER_MODE, pointer_mode);
-      alpha_ptr = user_alpha.data_ptr<float>();
-      beta_ptr = at::cuda::detail::get_cublas_device_zero();
-    } else {
-      alpha_val = a.item<float>();
-    }
-  }
     // For other data types, use the get_scale_mode function based on scaling type
     // The SCALE_MODE attrs only exist in cuBLAS 12.8+/ROCm 7.0 or in recent hipblaslt,
     // but we must invoke get_scale_mode anyways to trigger the version checks.
@@ -2079,7 +2048,6 @@ void scaled_gemm(
   cublasLtMatmulHeuristicResult_t heuristicResult = {};
   int returnedResult = 0;
   cublasLtHandle_t ltHandle = at::cuda::getCurrentCUDABlasLtHandle();
-
   TORCH_CUDABLAS_CHECK(cublasLtMatmulAlgoGetHeuristic(
       ltHandle,
       computeDesc.descriptor(),
@@ -2120,10 +2088,10 @@ void scaled_gemm(
         auto is_valid_status = hipblaslt_ext::matmulIsAlgoSupported(
                 ltHandle,
                 computeDesc.descriptor(),
-                alpha_ptr,
+                &alpha_val,
                 Adesc.descriptor(),
                 Bdesc.descriptor(),
-                beta_ptr,
+                &beta_val,
                 Cdesc.descriptor(),
                 Ddesc.descriptor(),
                 all_algos[i].algo,
@@ -2142,14 +2110,17 @@ void scaled_gemm(
   cublasStatus_t cublasStatus = cublasLtMatmul(
       ltHandle,
       computeDesc.descriptor(),
-      alpha_ptr,
+      &alpha_val,
       mat1_ptr,
       Adesc.descriptor(),
       mat2_ptr,
       Bdesc.descriptor(),
-      beta_ptr,
-      // NOTE: always use result_ptr here, because cuBLASLt w/device beta=0 can't handle nullptr either
+      &beta_val,
+#ifdef USE_ROCM
       result_ptr, // unused, since beta_val is 0, but hipblaslt can't handle nullptr
+#else
+      nullptr,
+#endif // ifdef USE_ROCM
       Cdesc.descriptor(),
       result_ptr,
       Ddesc.descriptor(),

aten/src/ATen/cuda/CUDABlas.h

@@ -161,8 +161,7 @@ void scaled_gemm(
     const void* result_scale_ptr,
     int64_t result_ld,
     ScalarType result_dtype,
-    bool use_fast_accum,
-    const std::optional<Tensor>& alpha);
+    bool use_fast_accum);
 
 #define CUDABLAS_BGEMM_ARGTYPES(Dtype)  CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYPE(Dtype, Dtype)
 

aten/src/ATen/cuda/CUDAGeneratorImpl.cpp

@@ -325,9 +325,9 @@ uint64_t CUDAGeneratorImpl::seed() {
  */
 c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
   // The RNG state comprises the seed, and an offset used for Philox.
-  constexpr size_t seed_size = sizeof(uint64_t);
-  constexpr size_t offset_size = sizeof(int64_t);
-  constexpr size_t total_size = seed_size + offset_size;
+  static const size_t seed_size = sizeof(uint64_t);
+  static const size_t offset_size = sizeof(int64_t);
+  static const size_t total_size = seed_size + offset_size;
 
   auto state_tensor = at::detail::empty_cpu({(int64_t)total_size}, ScalarType::Byte, std::nullopt, std::nullopt, std::nullopt, std::nullopt);
   auto rng_state = state_tensor.data_ptr<uint8_t>();
@@ -346,9 +346,9 @@ c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
  * and size of the internal state.
  */
 void CUDAGeneratorImpl::set_state(const c10::TensorImpl& new_state) {
-  constexpr size_t seed_size = sizeof(uint64_t);
-  constexpr size_t offset_size = sizeof(int64_t);
-  constexpr size_t total_size = seed_size + offset_size;
+  static const size_t seed_size = sizeof(uint64_t);
+  static const size_t offset_size = sizeof(int64_t);
+  static const size_t total_size = seed_size + offset_size;
 
   detail::check_rng_state(new_state);
 

aten/src/ATen/cuda/CUDAGreenContext.cpp

@@ -1,192 +0,0 @@
-#include <ATen/cuda/CUDAGreenContext.h>
-
-namespace at::cuda {
-  GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
-#if CUDA_HAS_GREEN_CONTEXT
-    int driver_version;
-    C10_CUDA_CHECK(cudaDriverGetVersion(&driver_version));
-    TORCH_CHECK(
-        driver_version >= 12080, "cuda driver too old to use green context!");
-    CUcontext pctx = nullptr;
-    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&pctx));
-    if (C10_UNLIKELY(!pctx)) {
-      TORCH_WARN(
-          "Attempted to create a green context but"
-          " there was no primary context! Creating a primary context...");
-
-      cudaFree(0);
-    }
-
-    CUdevice device;
-    device_id_ = device_id;
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuDeviceGet_(&device, device_id));
-
-    // Get device resources
-    CUdevResource device_resource;
-    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuDeviceGetDevResource_(
-        device, &device_resource, CU_DEV_RESOURCE_TYPE_SM));
-
-    // Split resources
-    std::vector<CUdevResource> result(1);
-    auto result_data = result.data();
-    unsigned int nb_groups = 1;
-    CUdevResource remaining;
-
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuDevSmResourceSplitByCount_(
-            result_data,
-            &nb_groups,
-            &device_resource,
-            &remaining,
-            0, // default flags
-            num_sms));
-
-    TORCH_CHECK(nb_groups == 1, "Failed to create single resource group");
-
-    // Generate resource descriptor
-    CUdevResourceDesc desc;
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuDevResourceGenerateDesc_(
-            &desc, result_data, 1));
-
-    // Create green context
-    // CU_GREEN_CTX_DEFAULT_STREAM is required per docs:
-    // https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__GREEN__CONTEXTS.html
-    C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuGreenCtxCreate_(
-        &green_ctx_, desc, device, CU_GREEN_CTX_DEFAULT_STREAM));
-
-    // Convert to regular context
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuCtxFromGreenCtx_(&context_, green_ctx_));
-    TORCH_CHECK(context_, "Green ctx conversion to regular ctx failed!");
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  std::unique_ptr<GreenContext> GreenContext::create(
-      uint32_t num_sms,
-      std::optional<uint32_t> device_id) {
-#if CUDA_HAS_GREEN_CONTEXT
-    if (!device_id.has_value()) {
-      device_id = at::cuda::current_device();
-    }
-    return std::make_unique<GreenContext>(device_id.value(), num_sms);
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  // Implement move operations
-  GreenContext::GreenContext(GreenContext&& other) noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
-    device_id_ = std::exchange(other.device_id_, -1);
-    green_ctx_ = std::exchange(other.green_ctx_, nullptr);
-    context_ = std::exchange(other.context_, nullptr);
-    parent_stream_ = std::exchange(other.parent_stream_, nullptr);
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  GreenContext& GreenContext::operator=(GreenContext&& other) noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
-    if (this != &other) {
-      // Clean up current resources
-      if (green_ctx_) {
-        CUcontext current = nullptr;
-        C10_CUDA_DRIVER_CHECK(
-            c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&current));
-        if (current == context_) {
-          TORCH_CHECK(
-              false,
-              "attempting to overwrite current green ctx "
-              "when it is active!");
-        }
-        C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuGreenCtxDestroy_(green_ctx_));
-      }
-
-      // Take ownership of other's resources
-      device_id_ = std::exchange(other.device_id_, -1);
-      green_ctx_ = std::exchange(other.green_ctx_, nullptr);
-      context_ = std::exchange(other.context_, nullptr);
-      parent_stream_ = std::exchange(other.parent_stream_, nullptr);
-    }
-    return *this;
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  GreenContext::~GreenContext() noexcept{
-#if CUDA_HAS_GREEN_CONTEXT
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuGreenCtxDestroy_(green_ctx_));
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  // Get the underlying CUDA context
-  CUcontext GreenContext::getContext() const {
-#if CUDA_HAS_GREEN_CONTEXT
-    return context_;
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  // Get the underlying green context
-#if CUDA_HAS_GREEN_CONTEXT
-  CUgreenCtx GreenContext::getGreenContext() const {
-    return green_ctx_;
-  }
-#endif
-
-  // Make this context current
-  void GreenContext::setContext() {
-#if CUDA_HAS_GREEN_CONTEXT
-    auto current_stream = c10::cuda::getCurrentCUDAStream();
-    parent_stream_ = current_stream.stream();
-
-    at::cuda::CUDAEvent ev;
-    ev.record(current_stream);
-
-    CUcontext current = nullptr;
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&current));
-    if (!current) {
-      C10_CUDA_DRIVER_CHECK(
-          c10::cuda::DriverAPI::get()->cuCtxSetCurrent_(context_));
-    } else {
-      C10_CUDA_DRIVER_CHECK(
-          c10::cuda::DriverAPI::get()->cuCtxPushCurrent_(context_));
-    }
-    // currently hardcodes the new green context to use the default stream
-    // TODO(eqy): consider creating a new stream if e.g., it allows interop
-    // with CUDA Graph captures etc.
-    auto default_stream = c10::cuda::getDefaultCUDAStream();
-    ev.block(default_stream);
-    c10::cuda::setCurrentCUDAStream(default_stream);
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-
-  void GreenContext::popContext() {
-#if CUDA_HAS_GREEN_CONTEXT
-    // see above note about stream being hardcoded to the default stream
-    at::cuda::CUDAEvent ev;
-    ev.record(c10::cuda::getCurrentCUDAStream());
-    CUcontext popped;
-    C10_CUDA_DRIVER_CHECK(
-        c10::cuda::DriverAPI::get()->cuCtxPopCurrent_(&popped));
-    TORCH_INTERNAL_ASSERT(
-        popped == context_, "expected popped context to be the current ctx");
-    ev.block(c10::cuda::getStreamFromExternal(parent_stream_, device_id_));
-#else
-    TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
-#endif
-  }
-} // namespace at::cuda

aten/src/ATen/cuda/CUDAGreenContext.h

@@ -1,53 +0,0 @@
-#pragma once
-#include <ATen/cuda/CUDAEvent.h>
-
-#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
-#include <c10/cuda/driver_api.h>
-#include <cuda.h>
-#include <memory>
-#include <stdexcept>
-#include <vector>
-#define CUDA_HAS_GREEN_CONTEXT 1
-#else
-#define CUDA_HAS_GREEN_CONTEXT 0
-#endif
-
-namespace at::cuda {
-
-class TORCH_CUDA_CPP_API GreenContext {
- public:
-  GreenContext(uint32_t device_id, uint32_t num_sms);
-
-  static std::unique_ptr<GreenContext> create(uint32_t num_sms, std::optional<uint32_t> device_id);
-
-  // Delete copy constructor and assignment
-  GreenContext(const GreenContext&) = delete;
-  GreenContext& operator=(const GreenContext&) = delete;
-
-  // Implement move operations
-  GreenContext(GreenContext&& other) noexcept;
-  GreenContext& operator=(GreenContext&& other) noexcept;
-  ~GreenContext() noexcept;
-
-  // Get the underlying CUDA context
-  CUcontext getContext() const;
-
-  // Get the underlying green context
-#if CUDA_HAS_GREEN_CONTEXT
-  CUgreenCtx getGreenContext() const;
-#endif
-
-  // Make this context current
-  void setContext();
-
-  void popContext();
-
- private:
-#if CUDA_HAS_GREEN_CONTEXT
-  int32_t device_id_ = -1;
-  CUgreenCtx green_ctx_ = nullptr;
-  CUcontext context_ = nullptr;
-  cudaStream_t parent_stream_ = nullptr;
-#endif
-};
-} // namespace at::cuda

aten/src/ATen/cuda/CUDAScaledBlas.cpp

@@ -1,270 +0,0 @@
-#include <cstdint>
-#include <c10/util/typeid.h>
-#include <c10/util/Exception.h>
-#include <c10/util/SmallVector.h>
-#include <c10/core/Scalar.h>
-#include <c10/core/ScalarType.h>
-#include <c10/util/Exception.h>
-#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
-#include <ATen/core/Tensor.h>
-#include <ATen/core/NamedTensor.h>
-#include <ATen/Dispatch.h>
-#include <ATen/ExpandUtils.h>
-#include <ATen/OpMathType.h>
-#include <ATen/TensorUtils.h>
-#include <ATen/cuda/CUDABlas.h>
-#include <ATen/cuda/tunable/Tunable.h>
-#include <ATen/cuda/tunable/TunableGemm.h>
-#include <ATen/native/Resize.h>
-#include <c10/util/MaybeOwned.h>
-#include <ATen/native/GroupedMMUtils.h>
-#include <ATen/native/cuda/RowwiseScaledMM.h>
-#include <ATen/native/cuda/ScaledGroupMM.h>
-#include <ATen/native/cuda/GroupMM.h>
-#include <ATen/ceil_div.h>
-
-#ifdef USE_FBGEMM_GENAI
-#include <fbgemm_gpu/torch_ops.h>
-#endif
-
-#ifndef AT_PER_OPERATOR_HEADERS
-#include <ATen/Functions.h>
-#include <ATen/NativeFunctions.h>
-#else
-#include <ATen/ops/_addmm_activation_native.h>
-#include <ATen/ops/_efficientzerotensor.h>
-#include <ATen/ops/_scaled_mm_native.h>
-#include <ATen/ops/_unsafe_view_native.h>
-#include <ATen/ops/abs.h>
-#include <ATen/ops/addmm_native.h>
-#include <ATen/ops/addmv_native.h>
-#include <ATen/ops/baddbmm_native.h>
-#include <ATen/ops/bmm_native.h>
-#include <ATen/ops/copy_native.h>
-#include <ATen/ops/dot_native.h>
-#include <ATen/ops/empty.h>
-#include <ATen/ops/empty_strided.h>
-#include <ATen/ops/gelu.h>
-#include <ATen/ops/max.h>
-#include <ATen/ops/mm_native.h>
-#include <ATen/ops/mul.h>
-#include <ATen/ops/relu.h>
-#include <ATen/ops/ones.h>
-#include <ATen/ops/scalar_tensor_native.h>
-#include <ATen/ops/vdot_native.h>
-#endif
-
-using at::blas::ScalingType;
-using at::blas::SwizzleType;
-
-namespace at::cuda::scaled {
-
-/**
- * Both inputs must be fp8,
- * Each needs a single scale, {Tensorwise (float)}
- */
-bool check_tensorwise_recipe(c10::ScalarType type_a,
-                             std::vector<ScalingType>& recipe_a,
-                             ArrayRef<Tensor>& scales_a,
-                             c10::ScalarType type_b,
-                             std::vector<ScalingType>& recipe_b,
-                             ArrayRef<Tensor>& scales_b) {
-  // both types must be fp8
-  if (!isFloat8Type(type_a) || !isFloat8Type(type_b)) {
-    return false;
-  }
-
-  // 1 scale each, {Tensorwise, float}
-  if (scales_a.size() != 1 || recipe_a.size() != 1 || scales_b.size() != 1 || recipe_b.size() != 1) {
-    return false;
-  }
-  // Need {Blockwise_1x32, e8m0} for A & B
-  if (recipe_a[0] != ScalingType::TensorWise) return false;
-  if (scales_a[0].scalar_type() != ScalarType::Float) return false;
-  if (recipe_b[0] != ScalingType::TensorWise) return false;
-  if (scales_b[0].scalar_type() != ScalarType::Float) return false;
-
-  return true;
-}
-
-/**
- * Both inputs must be fp8,
- * Each needs scales, {Rowwise (float)}
- */
-bool check_rowwise_recipe(c10::ScalarType type_a,
-                             std::vector<ScalingType>& recipe_a,
-                             ArrayRef<Tensor>& scales_a,
-                             c10::ScalarType type_b,
-                             std::vector<ScalingType>& recipe_b,
-                             ArrayRef<Tensor>& scales_b) {
-  // both types must be fp8
-  if (!isFloat8Type(type_a) || !isFloat8Type(type_b)) {
-    return false;
-  }
-
-  // 1 scale each, {Tensorwise, float}
-  if (scales_a.size() != 1 || recipe_a.size() != 1 || scales_b.size() != 1 || recipe_b.size() != 1) {
-    return false;
-  }
-
-  // Need {RowWise, dp32} for A & B
-  if (recipe_a[0] != ScalingType::RowWise) return false;
-  if (scales_a[0].scalar_type() != ScalarType::Float) return false;
-  if (recipe_b[0] != ScalingType::RowWise) return false;
-  if (scales_b[0].scalar_type() != ScalarType::Float) return false;
-
-  return true;
-}
-
-
-/**
- * Two-level scaling, canonical NVFP4
- * Both inputs must be fp4
- * A, B need 2 scales, {Blockwise_1x16 (e4m3), Tensorwise (fp32)}
- */
-bool check_nvfp4_recipe(c10::ScalarType type_a,
-                        std::vector<ScalingType>& recipe_a,
-                        ArrayRef<Tensor>& scales_a,
-                        c10::ScalarType type_b,
-                        std::vector<ScalingType>& recipe_b,
-                        ArrayRef<Tensor>& scales_b) {
-  // both types must be fp4
-  if (type_a != ScalarType::Float4_e2m1fn_x2 || type_b != ScalarType::Float4_e2m1fn_x2) {
-    return false;
-  }
-
-  // 2 scales, 2 recipes for each input
-  if (scales_a.size() != 2 || recipe_a.size() != 2 || scales_b.size() != 2 || recipe_b.size() != 2) {
-    return false;
-  }
-
-  // Need {Blockwise_1x16, e4m3 for scale[0], Tensorwise, fp32 for scale[1]}
-  if (recipe_a[0] != ScalingType::BlockWise1x16 || recipe_a[1] != ScalingType::TensorWise) return false;
-  if (scales_a[0].scalar_type() != ScalarType::Float8_e4m3fn || scales_a[1].scalar_type() != ScalarType::Float) return false;
-  if (recipe_b[0] != ScalingType::BlockWise1x16 || recipe_b[1] != ScalingType::TensorWise) return false;
-  if (scales_b[0].scalar_type() != ScalarType::Float8_e4m3fn || scales_b[1].scalar_type() != ScalarType::Float) return false;
-
-  return true;
-}
-
-/**
- * Single-level scaling, what PyT currently understands
- * Both inputs must be fp4
- * A, B need 1 scale, {Blockwise_1x16 (e4m3)}
- */
-bool check_nvfp4_recipe_single_scale
-                       (c10::ScalarType type_a,
-                        std::vector<ScalingType>& recipe_a,
-                        ArrayRef<Tensor>& scales_a,
-                        c10::ScalarType type_b,
-                        std::vector<ScalingType>& recipe_b,
-                        ArrayRef<Tensor>& scales_b) {
-  // both types must be fp4
-  if (type_a != ScalarType::Float4_e2m1fn_x2 || type_b != ScalarType::Float4_e2m1fn_x2) {
-    return false;
-  }
-
-  // 2 scales, 2 recipes for each input
-  if (scales_a.size() != 1 || recipe_a.size() != 1 || scales_b.size() != 1 || recipe_b.size() != 1) {
-    return false;
-  }
-
-  // Need {Blockwise_1x16, e4m3 for scale[0], Tensorwise, fp32 for scale[1]}
-  if (recipe_a[0] != ScalingType::BlockWise1x16) return false;
-  if (scales_a[0].scalar_type() != ScalarType::Float8_e4m3fn) return false;
-  if (recipe_b[0] != ScalingType::BlockWise1x16) return false;
-  if (scales_b[0].scalar_type() != ScalarType::Float8_e4m3fn) return false;
-
-  return true;
-}
-
-/**
- * Both inputs must be fp8
- * A, B must only have 1 scale each, A: {Blockwise_1x128 (float), B: {Blockwise_128x128 (float)
- */
-bool check_deepseek_recipe(ScalingType expected_recipe_a,
-                           ScalingType expected_recipe_b,
-                           c10::ScalarType type_a,
-                           std::vector<ScalingType>& recipe_a,
-                           ArrayRef<Tensor>& scales_a,
-                           c10::ScalarType type_b,
-                           std::vector<ScalingType>& recipe_b,
-                           ArrayRef<Tensor>& scales_b) {
-  // both types must be fp8
-  if (type_a != ScalarType::Float8_e4m3fn || type_b != ScalarType::Float8_e4m3fn) {
-    return false;
-  }
-
-  // 1 scales, 1 recipes for each input
-  if (scales_a.size() != 1 || recipe_a.size() != 1 || scales_b.size() != 1 || recipe_b.size() != 1) {
-    return false;
-  }
-
-  // Need {Blockwise_1x128, float} for A, {Blockwise_128x128, float} for B
-  if (recipe_a[0] != expected_recipe_a) return false;
-  if (scales_a[0].scalar_type() != ScalarType::Float) return false;
-  if (recipe_b[0] != expected_recipe_b) return false;
-  if (scales_b[0].scalar_type() != ScalarType::Float) return false;
-
-  return true;
-}
-
-/**
- * Both inputs must be fp8
- * A, B must have 1 scale each, {Blockwise_1x32, e8m0}
- */
-bool check_mxfp8_recipe(c10::ScalarType type_a,
-                        std::vector<ScalingType>& recipe_a,
-                        ArrayRef<Tensor>& scales_a,
-                        c10::ScalarType type_b,
-                        std::vector<ScalingType>& recipe_b,
-                        ArrayRef<Tensor>& scales_b) {
-  // both types must be fp8
-  if (type_a != ScalarType::Float8_e4m3fn || type_b != ScalarType::Float8_e4m3fn) {
-    return false;
-  }
-
-  // 1 scales, 1 recipes for each input
-  if (scales_a.size() != 1 || recipe_a.size() != 1 || scales_b.size() != 1 || recipe_b.size() != 1) {
-    return false;
-  }
-
-  // Need {Blockwise_1x32, e8m0} for A & B
-  if (recipe_a[0] != ScalingType::BlockWise1x32) return false;
-  if (scales_a[0].scalar_type() != ScalarType::Float8_e8m0fnu) return false;
-  if (recipe_b[0] != ScalingType::BlockWise1x32) return false;
-  if (scales_b[0].scalar_type() != ScalarType::Float8_e8m0fnu) return false;
-
-  return true;
-}
-
-/**
- * Both inputs must be fp4
- * A, B must have 1 scale each, {Blockwise_1x32, e8m0}
- */
-bool check_mxfp4_recipe(c10::ScalarType type_a,
-                        std::vector<ScalingType>& recipe_a,
-                        ArrayRef<Tensor>& scales_a,
-                        c10::ScalarType type_b,
-                        std::vector<ScalingType>& recipe_b,
-                        ArrayRef<Tensor>& scales_b) {
-  // both types must be fp4
-  if (type_a != ScalarType::Float4_e2m1fn_x2 || type_b != ScalarType::Float4_e2m1fn_x2) {
-    return false;
-  }
-
-  // 1 scales, 1 recipes for each input
-  if (scales_a.size() != 1 || recipe_a.size() != 1 || scales_b.size() != 1 || recipe_b.size() != 1) {
-    return false;
-  }
-
-  // Need {Blockwise_1x32, e8m0} for A & B
-  if (recipe_a[0] != ScalingType::BlockWise1x32) return false;
-  if (scales_a[0].scalar_type() != ScalarType::Float8_e8m0fnu) return false;
-  if (recipe_b[0] != ScalingType::BlockWise1x32) return false;
-  if (scales_b[0].scalar_type() != ScalarType::Float8_e8m0fnu) return false;
-
-  return true;
-}
-
-} // namespace at::native::cuda::blas::scaled

aten/src/ATen/cuda/CUDAScaledBlas.h

@@ -1,174 +0,0 @@
-#include <cstdint>
-#include <c10/util/typeid.h>
-#include <c10/util/Exception.h>
-#include <c10/util/SmallVector.h>
-#include <c10/core/Scalar.h>
-#include <c10/core/ScalarType.h>
-#include <c10/util/Exception.h>
-#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
-#include <ATen/core/Tensor.h>
-#include <ATen/core/NamedTensor.h>
-#include <ATen/Dispatch.h>
-#include <ATen/ExpandUtils.h>
-#include <ATen/OpMathType.h>
-#include <ATen/TensorUtils.h>
-#include <ATen/cuda/CUDABlas.h>
-#include <ATen/cuda/tunable/Tunable.h>
-#include <ATen/cuda/tunable/TunableGemm.h>
-#include <ATen/native/Resize.h>
-#include <c10/util/MaybeOwned.h>
-#include <ATen/native/GroupedMMUtils.h>
-#include <ATen/native/cuda/RowwiseScaledMM.h>
-#include <ATen/native/cuda/ScaledGroupMM.h>
-#include <ATen/native/cuda/GroupMM.h>
-#include <ATen/ceil_div.h>
-
-#ifdef USE_FBGEMM_GENAI
-#include <fbgemm_gpu/torch_ops.h>
-#endif
-
-#ifndef AT_PER_OPERATOR_HEADERS
-#include <ATen/Functions.h>
-#include <ATen/NativeFunctions.h>
-#else
-#include <ATen/ops/_addmm_activation_native.h>
-#include <ATen/ops/_efficientzerotensor.h>
-#include <ATen/ops/_scaled_mm_native.h>
-#include <ATen/ops/_unsafe_view_native.h>
-#include <ATen/ops/abs.h>
-#include <ATen/ops/addmm_native.h>
-#include <ATen/ops/addmv_native.h>
-#include <ATen/ops/baddbmm_native.h>
-#include <ATen/ops/bmm_native.h>
-#include <ATen/ops/copy_native.h>
-#include <ATen/ops/dot_native.h>
-#include <ATen/ops/empty.h>
-#include <ATen/ops/empty_strided.h>
-#include <ATen/ops/gelu.h>
-#include <ATen/ops/max.h>
-#include <ATen/ops/mm_native.h>
-#include <ATen/ops/mul.h>
-#include <ATen/ops/relu.h>
-#include <ATen/ops/ones.h>
-#include <ATen/ops/scalar_tensor_native.h>
-#include <ATen/ops/vdot_native.h>
-#endif
-
-using at::blas::ScalingType;
-using at::blas::SwizzleType;
-
-namespace at::cuda::scaled {
-
-static bool _scaled_mm_allowed_device(bool sm90_only=false, bool sm100_only=false) {
-#ifdef USE_ROCM
-    static const std::vector<std::string> archs = {
-        "gfx942",
-#if ROCM_VERSION >= 60300
-        "gfx1200", "gfx1201",
-#endif
-#if ROCM_VERSION >= 60500
-        "gfx950"
-#endif
-    };
-    return at::detail::getCUDAHooks().isGPUArch(archs);
-#else
-    auto dprops = at::cuda::getCurrentDeviceProperties();
-
-    if (sm90_only || sm100_only) {
-      return (sm90_only && dprops->major == 9) || (sm100_only && dprops->major == 10);
-    } else {
-      return dprops->major >= 9 || (dprops->major == 8 && dprops->minor == 9);
-    }
-#endif
-}
-
-#ifdef USE_ROCM
-static bool _scaled_mm_is_fnuz() {
-    return at::detail::getCUDAHooks().isGPUArch({"gfx942"});
-}
-#endif
-/**
- * Track concrete implementations available
- */
-enum class ScaledGemmImplementation {
-  NONE = 0,
-  TENSORWISE_TENSORWISE = 1,
-  ROWWISE_ROWWISE = 2,
-  BLOCK_128x128_1x128 = 3,
-  BLOCK_1x128_128x128 = 4,
-  BLOCK_1x128_1x128 = 5,
-  MXFP8_MXFP8 = 6,
-  NVFP4_NVFP4 = 7,
-  NVFP4_NVFP4_SINGLE_SCALE = 8,
-  MXFP4_MXFP4 = 9,
-};
-
-/**
- * Convert passed int (enum) from python back into a
- * strictly-typed enum
- */
-template <class EnumType, class ArrayType>
-std::vector<EnumType> convert_int_to_enum(ArrayType& v) {
-  std::vector<EnumType> converted;
-  converted.reserve(v.size());
-
-  for (auto vi : v) {
-    converted.push_back(static_cast<EnumType>(vi));
-  }
-  return converted;
-}
-
-bool check_tensorwise_recipe(c10::ScalarType,
-                             std::vector<ScalingType>&,
-                             ArrayRef<Tensor>&,
-                             c10::ScalarType,
-                             std::vector<ScalingType>&,
-                             ArrayRef<Tensor>&);
-
-
-bool check_rowwise_recipe(c10::ScalarType,
-                             std::vector<ScalingType>&,
-                             ArrayRef<Tensor>&,
-                             c10::ScalarType,
-                             std::vector<ScalingType>&,
-                             ArrayRef<Tensor>&);
-
-bool check_nvfp4_recipe(c10::ScalarType,
-                        std::vector<ScalingType>&,
-                        ArrayRef<Tensor>&,
-                        c10::ScalarType,
-                        std::vector<ScalingType>&,
-                        ArrayRef<Tensor>&);
-
-bool check_nvfp4_recipe_single_scale
-                       (c10::ScalarType,
-                        std::vector<ScalingType>&,
-                        ArrayRef<Tensor>&,
-                        c10::ScalarType,
-                        std::vector<ScalingType>&,
-                        ArrayRef<Tensor>&);
-
-bool check_deepseek_recipe(ScalingType,
-                           ScalingType,
-                           c10::ScalarType,
-                           std::vector<ScalingType>&,
-                           ArrayRef<Tensor>&,
-                           c10::ScalarType,
-                           std::vector<ScalingType>&,
-                           ArrayRef<Tensor>&);
-
-bool check_mxfp8_recipe(c10::ScalarType,
-                        std::vector<ScalingType>&,
-                        ArrayRef<Tensor>&,
-                        c10::ScalarType,
-                        std::vector<ScalingType>&,
-                        ArrayRef<Tensor>&);
-
-bool check_mxfp4_recipe(c10::ScalarType,
-                        std::vector<ScalingType>&,
-                        ArrayRef<Tensor>&,
-                        c10::ScalarType,
-                        std::vector<ScalingType>&,
-                        ArrayRef<Tensor>&);
-
-} // namespace at::native::cuda::blas::scaled

aten/src/ATen/cuda/CachingHostAllocator.cpp

@@ -183,6 +183,11 @@ struct CUDACachingHostAllocatorImpl
     return true;
   }
 
+  bool pinned_use_background_threads() override {
+    return c10::cuda::CUDACachingAllocator::CUDAAllocatorConfig::
+        pinned_use_background_threads();
+  }
+
   EventPool::Event create_event_internal(DeviceIndex idx) {
     // Leak the event pool to avoid shutdown issue.
     static auto* event_pool = new EventPool();

aten/src/ATen/cuda/cub.cuh

@@ -70,7 +70,11 @@
 #define ATEN_CUB_MAXIMUM() NO_ROCM(at_cuda_detail)ROCM_HIPCUB(::cub)::Max()
 #endif
 
-#if defined(USE_ROCM)
+#if (!defined(USE_ROCM) && !CUB_SUPPORTS_NV_BFLOAT16()) || defined(USE_ROCM)
+
+#if !defined(USE_ROCM)
+namespace at_cuda_detail {
+#endif
 
 // backport https://github.com/NVIDIA/cub/pull/306 for c10::BFloat16
 
@@ -92,6 +96,10 @@ template <>
 struct ROCM_HIPCUB(cub)::NumericTraits<c10::BFloat16>:
        ROCM_HIPCUB(cub)::BaseTraits<ROCM_HIPCUB(cub)::FLOATING_POINT, true, false, unsigned short, c10::BFloat16> {};
 
+#if !defined(USE_ROCM)
+} // namespace at_cuda_detail
+#endif
+
 #endif
 
 #if !defined(USE_ROCM)
@@ -113,7 +121,7 @@ struct cuda_type<c10::Half> {
   using type = __half;
 };
 
-#if !defined(USE_ROCM)
+#if !defined(USE_ROCM) && CUB_SUPPORTS_NV_BFLOAT16()
 
 template<>
 struct cuda_type<c10::BFloat16> {
@@ -169,6 +177,7 @@ inline void segmented_sort_pairs(
   }
 }
 
+#if CUB_SUPPORTS_UNIQUE_BY_KEY()
 template <typename KeysInputIteratorT, typename ValuesInputIteratorT, typename ValuesOutputIteratorT, typename NumSelectedIteratorT>
 inline void unique_by_key(
   KeysInputIteratorT keys_in, ValuesInputIteratorT values_in,
@@ -184,6 +193,7 @@ inline void unique_by_key(
   CUB_WRAPPER(NO_ROCM(at_cuda_detail)::cub::DeviceSelect::UniqueByKey,
     keys_in, values_in, keys_out_, values_out, num_selected, num_input_items, c10::cuda::getCurrentCUDAStream());
 }
+#endif
 
 namespace impl {
 
@@ -195,6 +205,36 @@ __global__ void transform_vals(InputIteratorT1 a, InputIteratorT2 b, OutputItera
   *out = scan_op(static_cast<acc_t>(*a), static_cast<acc_t>(*b));
 }
 
+#if !CUB_SUPPORTS_FUTURE_VALUE()
+template<typename ValueT, typename InputIteratorT>
+struct chained_iterator {
+  using iterator_category = std::random_access_iterator_tag;
+  using difference_type   = std::ptrdiff_t;
+  using value_type        = ValueT;
+  using pointer           = ValueT*;
+  using reference         = ValueT&;
+
+  InputIteratorT iter;
+  ValueT *first;
+  difference_type offset = 0;
+
+  __device__ ValueT operator[](difference_type i) {
+    i +=  offset;
+    if (i == 0) {
+      return *first;
+    } else {
+      return ValueT(iter[i - 1]);
+    }
+  }
+  __device__ chained_iterator operator+(difference_type i) {
+    return chained_iterator{iter, first, i};
+  }
+  __device__ ValueT operator*() {
+    return (*this)[0];
+  }
+};
+#endif
+
 // even though cub is supposed to support tensors with int_max elements, in reality it doesn't,
 // so split at int_max/2
 constexpr int max_cub_size = std::numeric_limits<int>::max() / 2 + 1; // 2**30
@@ -239,6 +279,25 @@ inline void inclusive_scan(InputIteratorT input, OutputIteratorT output, ScanOpT
         first_elem_ptr,
         scan_op);
     C10_CUDA_KERNEL_LAUNCH_CHECK();
+#if !CUB_SUPPORTS_FUTURE_VALUE()
+    using ArgIndexInputIterator = NO_ROCM(at_cuda_detail)::cub::ArgIndexInputIterator<InputIteratorT>;
+    using tuple = typename ArgIndexInputIterator::value_type;
+    auto input_iter_transform = [=] __device__ (const tuple &x)->input_t  {
+      if (x.key == 0) {
+        return *first_elem_ptr;
+      } else {
+        return x.value;
+      }
+    };
+    auto input_ = ATEN_CUB_TRANSFORM_ITERATOR(input_t, decltype(input_iter_transform), ArgIndexInputIterator)(
+      ArgIndexInputIterator(input + i), input_iter_transform);
+    CUB_WRAPPER(NO_ROCM(at_cuda_detail)::cub::DeviceScan::InclusiveScan,
+        input_,
+        output + i,
+        scan_op,
+        size_cub,
+        at::cuda::getCurrentCUDAStream());
+#else
     CUB_WRAPPER(NO_ROCM(at_cuda_detail)::cub::DeviceScan::ExclusiveScan,
         input + i + 1,
         output + i,
@@ -246,6 +305,7 @@ inline void inclusive_scan(InputIteratorT input, OutputIteratorT output, ScanOpT
         ::at_cuda_detail::cub::FutureValue<input_t>(first_elem_ptr),
         size_cub,
         at::cuda::getCurrentCUDAStream());
+#endif
   }
 #endif
 }
@@ -497,6 +557,16 @@ inline void exclusive_scan(InputIteratorT input, OutputIteratorT output, ScanOpT
         first_elem_ptr,
         scan_op);
     C10_CUDA_KERNEL_LAUNCH_CHECK();
+#if !CUB_SUPPORTS_FUTURE_VALUE()
+    auto input_ = impl::chained_iterator<InitValueT, InputIteratorT>{
+      input + i, first_elem_ptr};
+    CUB_WRAPPER(NO_ROCM(at_cuda_detail)::cub::DeviceScan::InclusiveScan,
+        input_,
+        output + i,
+        scan_op,
+        size_cub,
+        at::cuda::getCurrentCUDAStream());
+#else
     CUB_WRAPPER(NO_ROCM(at_cuda_detail)::cub::DeviceScan::ExclusiveScan,
         input + i,
         output + i,
@@ -504,10 +574,12 @@ inline void exclusive_scan(InputIteratorT input, OutputIteratorT output, ScanOpT
         ::at_cuda_detail::cub::FutureValue<InitValueT>(first_elem_ptr),
         size_cub,
         at::cuda::getCurrentCUDAStream());
+#endif
   }
 #endif
 }
 
+#if CUB_SUPPORTS_SCAN_BY_KEY()
 
 template <typename KeysInputIteratorT, typename ValuesInputIteratorT, typename ValuesOutputIteratorT>
 inline void inclusive_sum_by_key(KeysInputIteratorT keys, ValuesInputIteratorT input, ValuesOutputIteratorT output, int64_t num_items) {
@@ -535,6 +607,7 @@ inline void inclusive_scan_by_key(KeysInputIteratorT keys, ValuesInputIteratorT
 #endif
 }
 
+#endif
 
 template <typename InputIteratorT, typename OutputIteratorT, typename NumSelectedIteratorT>
 void unique(InputIteratorT input, OutputIteratorT output,

aten/src/ATen/cuda/cub_definitions.cuh

@@ -10,6 +10,14 @@
 #define CUB_VERSION 200001
 #endif
 
+// cub sort support for __nv_bfloat16 is added to cub 1.13 in:
+// https://github.com/NVIDIA/cub/pull/306
+#if CUB_VERSION >= 101300
+#define CUB_SUPPORTS_NV_BFLOAT16() true
+#else
+#define CUB_SUPPORTS_NV_BFLOAT16() false
+#endif
+
 // cub support for CUB_WRAPPED_NAMESPACE is added to cub 1.13.1 in:
 // https://github.com/NVIDIA/cub/pull/326
 // CUB_WRAPPED_NAMESPACE is defined globally in cmake/Dependencies.cmake
@@ -20,6 +28,30 @@
 #define USE_GLOBAL_CUB_WRAPPED_NAMESPACE() false
 #endif
 
+// cub support for UniqueByKey is added to cub 1.16 in:
+// https://github.com/NVIDIA/cub/pull/405
+#if CUB_VERSION >= 101600
+#define CUB_SUPPORTS_UNIQUE_BY_KEY() true
+#else
+#define CUB_SUPPORTS_UNIQUE_BY_KEY() false
+#endif
+
+// cub support for scan by key is added to cub 1.15
+// in https://github.com/NVIDIA/cub/pull/376
+#if CUB_VERSION >= 101500
+#define CUB_SUPPORTS_SCAN_BY_KEY() 1
+#else
+#define CUB_SUPPORTS_SCAN_BY_KEY() 0
+#endif
+
+// cub support for cub::FutureValue is added to cub 1.15 in:
+// https://github.com/NVIDIA/cub/pull/305
+#if CUB_VERSION >= 101500
+#define CUB_SUPPORTS_FUTURE_VALUE() true
+#else
+#define CUB_SUPPORTS_FUTURE_VALUE() false
+#endif
+
 // There were many bc-breaking changes in major version release of CCCL v3.0.0
 // Please see https://nvidia.github.io/cccl/cccl/3.0_migration_guide.html
 #if CUB_VERSION >= 200800

aten/src/ATen/cuda/detail/BLASConstants.cu

@@ -1,54 +0,0 @@
-#include <ATen/Functions.h>
-#include <ATen/Tensor.h>
-#include <ATen/cuda/Exceptions.h>
-
-#include <mutex>
-
-namespace at {
-namespace cuda {
-namespace detail {
-
-__device__ __constant__ float cublas_one_device;
-__device__ __constant__ float cublas_zero_device;
-
-float *get_cublas_device_one() {
-  static c10::once_flag init_flag;
-
-  c10::call_once(init_flag, []() {
-    const float one = 1.f;
-    AT_CUDA_CHECK(cudaMemcpyToSymbol(cublas_one_device, &one, sizeof(float)));
-  });
-
-  float *ptr;
-  AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_one_device));
-  return ptr;
-}
-
-float *get_cublas_device_zero() {
-  static c10::once_flag init_flag;
-
-  c10::call_once(init_flag, []() {
-    const float zero = 0.f;
-    AT_CUDA_CHECK(cudaMemcpyToSymbol(cublas_zero_device, &zero, sizeof(float)));
-  });
-
-  float *ptr;
-  AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_zero_device));
-  return ptr;
-}
-
-float *get_user_alpha_ptr() {
-  static float *alpha_ptr;
-
-  static c10::once_flag init_flag;
-
-  c10::call_once(init_flag, []() {
-    AT_CUDA_CHECK(cudaMalloc(&alpha_ptr, sizeof(float)));
-  });
-
-  return alpha_ptr;
-}
-
-} // namespace detail
-} // namespace cuda
-} // namespace at

aten/src/ATen/cuda/detail/BLASConstants.h

@@ -1,11 +0,0 @@
-#pragma once
-
-#include <ATen/core/TensorBase.h>
-
-namespace at::cuda::detail {
-
-float *get_cublas_device_one();
-float *get_cublas_device_zero();
-float *get_user_alpha_ptr();
-
-} // namespace at::cuda::detail

aten/src/ATen/cuda/tunable/TunableGemm.h

@@ -109,8 +109,7 @@ class DefaultScaledGemmOp : public Callable<ScaledGemmParams<T>> {
           params->c_scale_ptr,
           params->ldc,
           params->c_dtype,
-          params->use_fast_accum,
-          std::nullopt /* alpha */);
+          params->use_fast_accum);
       return OK;
     }
 };

aten/src/ATen/detail/XLAHooksInterface.cpp

@@ -1,23 +0,0 @@
-#include <ATen/detail/XLAHooksInterface.h>
-
-namespace at {
-namespace detail {
-
-const XLAHooksInterface& getXLAHooks() {
-  auto create_impl = [] {
-    // Create XLA hooks using the registry
-    auto hooks = XLAHooksRegistry()->Create("torch_xla::detail::XLAHooks", XLAHooksArgs{});
-    if (hooks) {
-      return hooks;
-    }
-    // If hooks creation fails, fall back to default implementation
-    return std::make_unique<XLAHooksInterface>();
-  };
-  static auto hooks = create_impl();
-  return *hooks;
-}
-} // namespace detail
-
-C10_DEFINE_REGISTRY(XLAHooksRegistry, XLAHooksInterface, XLAHooksArgs)
-
-} // namespace at

aten/src/ATen/detail/XLAHooksInterface.h

@@ -1,79 +0,0 @@
-#pragma once
-
-#include <c10/core/Device.h>
-#include <c10/util/Exception.h>
-#include <c10/util/Registry.h>
-
-#include <ATen/detail/AcceleratorHooksInterface.h>
-
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-parameter")
-
-namespace at {
-
-constexpr const char* XLA_HELP =
-  "This error has occurred because you are trying "
-  "to use some XLA functionality, but the XLA library has not been "
-  "loaded by the dynamic linker. You must load xla libraries by `import torch_xla`";
-
-struct TORCH_API XLAHooksInterface : AcceleratorHooksInterface {
-  ~XLAHooksInterface() override = default;
-
-  void init() const override {
-    TORCH_CHECK(false, "Cannot initialize XLA without torch_xla library. ", XLA_HELP);
-  }
-
-  virtual bool hasXLA() const {
-    return false;
-  }
-
-  virtual std::string showConfig() const {
-    TORCH_CHECK(
-        false,
-        "Cannot query detailed XLA version without torch_xla library. ",
-        XLA_HELP);
-  }
-
-  const Generator& getDefaultGenerator(
-      [[maybe_unused]] DeviceIndex device_index = -1) const override {
-    TORCH_CHECK(
-        false, "Cannot get default XLA generator without torch_xla library. ", XLA_HELP);
-  }
-
-  Generator getNewGenerator(
-      [[maybe_unused]] DeviceIndex device_index = -1) const override {
-    TORCH_CHECK(false, "Cannot get XLA generator without torch_xla library. ", XLA_HELP);
-  }
-
-  virtual DeviceIndex getCurrentDevice() const override {
-    TORCH_CHECK(false, "Cannot get current XLA device without torch_xla library. ", XLA_HELP);
-  }
-
-  Device getDeviceFromPtr(void* /*data*/) const override {
-    TORCH_CHECK(false, "Cannot get device of pointer on XLA without torch_xla library. ", XLA_HELP);
-  }
-
-  Allocator* getPinnedMemoryAllocator() const override {
-    TORCH_CHECK(false, "Cannot get XLA pinned memory allocator without torch_xla library. ", XLA_HELP);
-  }
-
-  bool isPinnedPtr(const void* data) const override {
-    return false;
-  }
-
-  bool hasPrimaryContext(DeviceIndex device_index) const override {
-    TORCH_CHECK(false, "Cannot query primary context without torch_xla library. ", XLA_HELP);
-  }
-
-};
-
-struct TORCH_API XLAHooksArgs {};
-
-TORCH_DECLARE_REGISTRY(XLAHooksRegistry, XLAHooksInterface, XLAHooksArgs);
-#define REGISTER_XLA_HOOKS(clsname) \
-  C10_REGISTER_CLASS(XLAHooksRegistry, clsname, clsname)
-
-namespace detail {
-TORCH_API const XLAHooksInterface& getXLAHooks();
-} // namespace detail
-} // namespace at
-C10_DIAGNOSTIC_POP()

aten/src/ATen/functorch/BatchedTensorImpl.h

@@ -160,10 +160,6 @@ constexpr DispatchKeySet kKeysToPropagateToWrapper({
   DispatchKey::CUDA,
   DispatchKey::CPU,
   DispatchKey::PrivateUse1,
-  DispatchKey::SparseCPU,
-  DispatchKey::SparseCUDA,
-  DispatchKey::SparseCsrCPU,
-  DispatchKey::SparseCsrCUDA,
 });
 
 inline DispatchKeySet getKeysToPropagateToWrapper(const Tensor& tensor, DispatchKeySet to_propagate=kKeysToPropagateToWrapper) {

aten/src/ATen/native/Activation.cpp

@@ -240,8 +240,8 @@ TORCH_META_FUNC(gelu_backward) (
 
 namespace at::native {
 
-static constexpr double SELU_ALPHA = 1.6732632423543772848170429916717;
-static constexpr double SELU_SCALE = 1.0507009873554804934193349852946;
+static const double SELU_ALPHA = 1.6732632423543772848170429916717;
+static const double SELU_SCALE = 1.0507009873554804934193349852946;
 
 DEFINE_DISPATCH(elu_stub);
 DEFINE_DISPATCH(elu_backward_stub);

aten/src/ATen/native/BlasKernel.cpp

@@ -286,7 +286,7 @@ template void scal_fast_path<scalar_t>(int *n, scalar_t *a, scalar_t *x, int *in
 #if AT_BUILD_WITH_BLAS()
 template <>
 bool scal_use_fast_path<double>(int64_t n, int64_t incx) {
-  auto constexpr intmax = std::numeric_limits<int>::max();
+  auto intmax = std::numeric_limits<int>::max();
   return n <= intmax && incx <= intmax;
 }
 
@@ -315,7 +315,7 @@ bool gemv_use_fast_path<float>(
     int64_t incx,
     [[maybe_unused]] float beta,
     int64_t incy) {
-  auto constexpr intmax = std::numeric_limits<int>::max();
+  auto intmax = std::numeric_limits<int>::max();
   return (m <= intmax) && (n <= intmax) && (lda <= intmax) &&
          (incx > 0) && (incx <= intmax) && (incy > 0) && (incy <= intmax);
 }

aten/src/ATen/native/Convolution.cpp

@@ -658,7 +658,6 @@ static void check_shape_forward(const at::Tensor& input,
   TORCH_CHECK(!params.is_output_padding_neg(), "negative output_padding is not supported");
   TORCH_CHECK(!params.is_stride_nonpos(), "non-positive stride is not supported");
   TORCH_CHECK(!params.is_dilation_neg(), "dilation should be greater than zero");
-  TORCH_CHECK(groups > 0, "expected groups to be greater than 0, but got groups=", groups);
 
   TORCH_CHECK(weight_dim == k,
            "Expected ", weight_dim, "-dimensional input for ", weight_dim,

aten/src/ATen/native/Distributions.h

@@ -1,6 +1,5 @@
 #pragma once
 
-#include <array>
 #include <ATen/native/Math.h>
 #include <c10/macros/Macros.h>
 #include <c10/util/MathConstants.h>
@@ -128,7 +127,7 @@ C10_DEVICE scalar_t sample_gamma(scalar_t alpha, BaseSampler<accscalar_t, unifor
 
 template<typename scalar_t>
 C10_DEVICE scalar_t stirling_approx_tail(scalar_t k) {
-  constexpr static scalar_t kTailValues[] = {
+  const static scalar_t kTailValues[] = {
     0.0810614667953272,
     0.0413406959554092,
     0.0276779256849983,
@@ -140,7 +139,7 @@ C10_DEVICE scalar_t stirling_approx_tail(scalar_t k) {
     0.00925546218271273,
     0.00833056343336287
   };
-  if (k < std::size(kTailValues)) {
+  if (k <= 9) {
     return kTailValues[static_cast<size_t>(k)];
   }
   scalar_t kp1sq = (k + 1) * (k + 1);

aten/src/ATen/native/LinearAlgebra.cpp

@@ -3620,7 +3620,7 @@ Tensor& _int_mm_out_cpu(const Tensor& self, const Tensor& mat2, Tensor& result)
     try {
       mkldnn_matmul_i8i8i32(self, mat2, result);
       dispatched = true;
-    } catch ([[maybe_unused]] const std::exception& e) {
+    } catch (const std::exception& e) {
       TORCH_WARN(func_name, " failed, switching to BLAS gemm: ", e.what());
     }
   }

aten/src/ATen/native/Math.h

@@ -581,7 +581,7 @@ scalar_t ratevl(scalar_t x, const scalar_t num[], int64_t M,
 template <typename scalar_t>
 static scalar_t lanczos_sum_expg_scaled(scalar_t x) {
   // lanczos approximation
-  static constexpr scalar_t lanczos_sum_expg_scaled_num[13] = {
+  static const scalar_t lanczos_sum_expg_scaled_num[13] = {
     0.006061842346248906525783753964555936883222,
     0.5098416655656676188125178644804694509993,
     19.51992788247617482847860966235652136208,
@@ -596,7 +596,7 @@ static scalar_t lanczos_sum_expg_scaled(scalar_t x) {
     103794043.1163445451906271053616070238554,
     56906521.91347156388090791033559122686859
   };
-  static constexpr scalar_t lanczos_sum_expg_scaled_denom[13] = {
+  static const scalar_t lanczos_sum_expg_scaled_denom[13] = {
     1.,
     66.,
     1925.,
@@ -712,7 +712,7 @@ static scalar_t _igamc_helper_series(scalar_t a, scalar_t x) {
 template <typename scalar_t>
 static scalar_t _igam_helper_asymptotic_series(scalar_t a, scalar_t x, bool igam) {
   // Compute igam/igamc using DLMF 8.12.3/8.12.4 [igam1]
-  static constexpr scalar_t d[25][25] =
+  static const scalar_t d[25][25] =
     {{-3.3333333333333333e-1, 8.3333333333333333e-2, -1.4814814814814815e-2,
       1.1574074074074074e-3, 3.527336860670194e-4, -1.7875514403292181e-4,
       3.9192631785224378e-5, -2.1854485106799922e-6, -1.85406221071516e-6,

aten/src/ATen/native/Normalization.cpp

@@ -62,7 +62,7 @@
 #include <utility>
 #include <vector>
 
-static constexpr int MIOPEN_DIM_MAX = 5;
+static const int MIOPEN_DIM_MAX = 5;
 
 namespace at::meta {
 

aten/src/ATen/native/PixelShuffle.h

@@ -11,8 +11,6 @@ inline void check_pixel_shuffle_shapes(const Tensor& self, int64_t upscale_facto
               "pixel_shuffle expects a positive upscale_factor, but got ",
               upscale_factor);
   int64_t c = self.size(-3);
-  TORCH_CHECK_VALUE(upscale_factor <= std::numeric_limits<decltype(upscale_factor)>::max() / upscale_factor,
-        "upscale factor is too large, (upscale_factor)^2 overflowed: upscale_factor=", upscale_factor);
   int64_t upscale_factor_squared = upscale_factor * upscale_factor;
   TORCH_CHECK(c % upscale_factor_squared == 0,
               "pixel_shuffle expects its input's 'channel' dimension to be divisible by the square of "

aten/src/ATen/native/TensorAdvancedIndexing.cpp

@@ -1906,9 +1906,11 @@ Tensor& index_fill_(
         "This also applies to advanced indexing e.g. tensor[mask] = scalar");
   }
 
-  TORCH_CHECK(
-      self.is_complex() || !source.isComplex(),
-      "index_fill_(): Converting complex Scalar to non-complex type is not supported");
+  if (!self.is_complex() && source.isComplex()) {
+    TORCH_CHECK(
+        false,
+        "index_fill_(): Converting complex Scalar to non-complex type is not supported");
+  }
 
   // Handle the case when `self` is 0-dim
   Tensor self_nonzero_dim = (self.dim() == 0) ? self.unsqueeze(-1) : self;

aten/src/ATen/native/TensorAdvancedIndexingUtils.h

@@ -77,7 +77,7 @@ inline AdvancedIndex make_info(Tensor self, IOptTensorListRef orig) {
   // next broadcast all index tensors together
   try {
     indices = expand_outplace(indices);
-  } catch (std::exception&) {
+  } catch (std::exception& e) {
     TORCH_CHECK_INDEX(
         false,
         "shape mismatch: indexing tensors could not be broadcast together"

aten/src/ATen/native/cpu/DepthwiseConvKernel.cpp

@@ -259,20 +259,11 @@ inline void winograd_f2k3_input_transform_inplace__rvv(
   const vfloat32m1_t wd1 = __riscv_vfadd_vv_f32m1(d1, d2, 4);
   const vfloat32m1_t wd2 = __riscv_vfsub_vv_f32m1(d2, d1, 4);
   const vfloat32m1_t wd3 = __riscv_vfsub_vv_f32m1(d1, d3, 4);
-  /* GCC 14.2 (RISC-V RVV) ICE workaround:
-   * Avoid single-statement read-modify-write on MEM_REF like:
-   *   *input_tile_val =
-   *     __riscv_vset_v_f32m1_f32m1x4(*input_tile_val, idx, val);
-   * This triggers an ICE during GIMPLE lower (gsi_replace / riscv_gimple_fold_builtin)
-   * with -march=rv64gcv. Use a temporary then write back.
-   * Do NOT refactor into the single-statement form. Clang is unaffected.
-   */
-  vfloat32m1x4_t tmp_input_tile_val = *input_tile_val;
-  tmp_input_tile_val = __riscv_vset_v_f32m1_f32m1x4(tmp_input_tile_val, 0, wd0);
-  tmp_input_tile_val = __riscv_vset_v_f32m1_f32m1x4(tmp_input_tile_val, 1, wd1);
-  tmp_input_tile_val = __riscv_vset_v_f32m1_f32m1x4(tmp_input_tile_val, 2, wd2);
-  tmp_input_tile_val = __riscv_vset_v_f32m1_f32m1x4(tmp_input_tile_val, 3, wd3);
-  *input_tile_val = tmp_input_tile_val;
+
+  *input_tile_val = __riscv_vset_v_f32m1_f32m1x4(*input_tile_val, 0, wd0);
+  *input_tile_val = __riscv_vset_v_f32m1_f32m1x4(*input_tile_val, 1, wd1);
+  *input_tile_val = __riscv_vset_v_f32m1_f32m1x4(*input_tile_val, 2, wd2);
+  *input_tile_val = __riscv_vset_v_f32m1_f32m1x4(*input_tile_val, 3, wd3);
 }
 
 inline void winograd_f2k3_output_transform_inplace__rvv(
@@ -286,15 +277,9 @@ inline void winograd_f2k3_output_transform_inplace__rvv(
   const vfloat32m1_t wm0 = __riscv_vfadd_vv_f32m1(m0_plus_m1, m2, 4);
   const vfloat32m1_t m1_sub_m2 = __riscv_vfsub_vv_f32m1(m1, m2, 4);
   const vfloat32m1_t wm1 = __riscv_vfsub_vv_f32m1(m1_sub_m2, m3, 4);
-  /* GCC 14.2 (RISC-V RVV) ICE workaround — see note above.
-   * Keep the temporary + write-back pattern to avoid ICE.
-   * Do NOT rewrite into:
-   *   *input_tile_val = __riscv_vset_v_f32m1_f32m1x4(*input_tile_val, idx, val);
-   */
-  vfloat32m1x4_t tmp_output_tile_val = *input_tile_val;
-  tmp_output_tile_val = __riscv_vset_v_f32m1_f32m1x4(tmp_output_tile_val, 0, wm0);
-  tmp_output_tile_val = __riscv_vset_v_f32m1_f32m1x4(tmp_output_tile_val, 1, wm1);
-  *input_tile_val = tmp_output_tile_val;
+
+  *input_tile_val = __riscv_vset_v_f32m1_f32m1x4(*input_tile_val, 0, wm0);
+  *input_tile_val = __riscv_vset_v_f32m1_f32m1x4(*input_tile_val, 1, wm1);
 }
 
 inline vfloat32m1_t
@@ -315,17 +300,11 @@ inline void winograd_f2k3_kernel_transform__rvv(
   const vfloat32m1_t const_half = __riscv_vfmv_v_f_f32m1(0.5f, 4);
   const vfloat32m1_t g0_plus_g2 = __riscv_vfadd_vv_f32m1(g0, g2, 4);
   vfloat32m1_t half_g0_plus_g2 =  __riscv_vfmul_vv_f32m1(const_half, g0_plus_g2, 4);
-  /* GCC 14.2 (RISC-V RVV) ICE workaround — see note above.
-   * Keep the temporary + write-back pattern to avoid ICE.
-   * Do NOT rewrite into:
-   *   *transform = __riscv_vset_v_f32m1_f32m1x4(*transform, idx, val);
-   */
-  vfloat32m1x4_t tmp_transform = *transform;
-  tmp_transform = __riscv_vset_v_f32m1_f32m1x4(tmp_transform, 0, g0);
-  tmp_transform = __riscv_vset_v_f32m1_f32m1x4(tmp_transform, 1, vmuladdq_f32(half_g0_plus_g2, const_half, g1));
-  tmp_transform = __riscv_vset_v_f32m1_f32m1x4(tmp_transform, 2, vmulsubq_f32(half_g0_plus_g2, const_half, g1));
-  tmp_transform = __riscv_vset_v_f32m1_f32m1x4(tmp_transform, 3, g2);
-  *transform = tmp_transform;
+
+  *transform = __riscv_vset_v_f32m1_f32m1x4(*transform, 0, g0);
+  *transform = __riscv_vset_v_f32m1_f32m1x4(*transform, 1, vmuladdq_f32(half_g0_plus_g2, const_half, g1));
+  *transform = __riscv_vset_v_f32m1_f32m1x4(*transform, 2, vmulsubq_f32(half_g0_plus_g2, const_half, g1));
+  *transform = __riscv_vset_v_f32m1_f32m1x4(*transform, 3, g2);
 }
 
 inline vfloat32m1x4_t v4f_transpose4x4__rvv(const vfloat32m1x4_t m) {

aten/src/ATen/native/cpu/UpSampleKernel.cpp

@@ -1038,7 +1038,7 @@ struct HelperInterpNearest : public HelperInterpBase {
   // We keep this structure for BC and consider as deprecated.
   // See HelperInterpNearestExact as replacement
 
-  static constexpr int interp_size = 1;
+  static const int interp_size = 1;
 
   static inline void init_indices_weights(
     at::ScalarType output_type,
@@ -1155,7 +1155,7 @@ struct HelperInterpNearestExact : public HelperInterpNearest {
 
 struct HelperInterpLinear : public HelperInterpBase {
 
-  static constexpr int interp_size = 2;
+  static const int interp_size = 2;
 
   // Compute indices and weights for each interpolated dimension
   // indices_weights = {
@@ -1275,7 +1275,7 @@ struct HelperInterpLinear : public HelperInterpBase {
 
 struct HelperInterpCubic : public HelperInterpBase {
 
-  static constexpr int interp_size = 4;
+  static const int interp_size = 4;
 
   // Compute indices and weights for each interpolated dimension
   // indices_weights = {