Enable B028

Signed-off-by: Yuanyuan Chen <cyyever@outlook.com>

.ci/docker/common/install_cpython.sh

@@ -83,10 +83,6 @@ 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_onnx.sh

@@ -19,7 +19,7 @@ pip_install \
   transformers==4.36.2
 
 pip_install coloredlogs packaging
-pip_install onnxruntime==1.23.0
+pip_install onnxruntime==1.23.1
 pip_install onnxscript==0.5.4
 
 # Cache the transformers model to be used later by ONNX tests. We need to run the transformers

.ci/docker/libtorch/build.sh

@@ -39,9 +39,13 @@ 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="${GPU_ARCH_VERSION}.2"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
         fi
         BASE_TARGET=rocm
         GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete

.ci/docker/manywheel/build.sh

@@ -75,9 +75,13 @@ 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="${GPU_ARCH_VERSION}.2"
+            GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
         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==1.18.0
+onnx==1.19.1
 #Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:
 
-onnxscript==0.5.3
+onnxscript==0.5.4
 #Description: Required by mypy and test_public_bindings.py when checking torch.onnx._internal
 #Pinned versions:
 #test that import:

.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 as e:
-        logger.error("Git operation failed: %s", e)
+    except GitCommandError:
+        logger.exception("Git operation failed")
         raise
 
 

.ci/lumen_cli/pyproject.toml

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

.circleci/scripts/binary_populate_env.sh

@@ -163,8 +163,13 @@ if [[ "$(uname)" != Darwin ]]; then
   MEMORY_LIMIT_MAX_JOBS=12
   NUM_CPUS=$(( $(nproc) - 2 ))
 
-  # 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} : ${NUM_CPUS} ))}
+  if [[ "$(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
 
   cat >>"$envfile" <<EOL
   export MAX_JOBS="${MAX_JOBS}"

.claude/skills/pytorch-docstring.md

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

@@ -7,16 +7,12 @@ 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,E721,E741,F405,F841,F999,W503,W504,C408,E302,W291,E303,F824,
+    E203,E305,E402,E501,E704,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
+    B007,B008,B017,B019,B023,B903,B905,B906,B907,B908,B910
     # 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

@@ -124,3 +124,10 @@ 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 @@
-1b013f5b5a87a1882eb143c26d79d091150d6a37
+69bbe7363897764f9e758d851cd0340147d27f94

.github/ci_commit_pins/vision.txt

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

.github/ci_configs/vllm/Dockerfile

@@ -283,6 +283,9 @@ 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
@@ -295,6 +298,8 @@ 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

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

.github/labeler.yml

@@ -133,3 +133,32 @@
 
 "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

@@ -33,6 +33,7 @@ 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' 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'"
+        "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'"
     ),
     "13.0": (
         "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "

.github/templates/macos_binary_build_workflow.yml.j2

@@ -26,9 +26,8 @@ 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 '3.14.0-rc.2' }}"
+          python-version: "!{{ py_ver.strip('t') + ('.4' if '3.14' not in py_ver else '.0') }}"
           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: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     {%- else %}
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge.nonephemeral"
     {%- endif %}
     {%- endif %}
     timeout-minutes: !{{ common.timeout_minutes_windows_binary }}

.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' 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'
+      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'
       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' 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'
+      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'
       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' 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'
+      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'
       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' 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'
+      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'
       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' 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'
+      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'
       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' 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'
+      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'
       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' 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'
+      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'
       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' 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'
+      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'
     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' 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'
+      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'
     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' 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'
+      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'
     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' 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'
+      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'
     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' 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'
+      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'
     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' 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'
+      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'
     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' 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'
+      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'
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
   manywheel-py3_14t-cuda12_9-test:  # Testing

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

@@ -63,7 +63,6 @@ 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

@@ -59,7 +59,6 @@ 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
@@ -169,7 +168,6 @@ 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
@@ -279,7 +277,6 @@ 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
@@ -389,7 +386,6 @@ 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
@@ -499,7 +495,6 @@ 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
@@ -609,9 +604,8 @@ 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-rc.2"
+          python-version: "3.14.0"
           freethreaded: false
       - name: Checkout PyTorch
         uses: actions/checkout@v4
@@ -719,9 +713,8 @@ 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-rc.2"
+          python-version: "3.14.0"
           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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -791,7 +791,7 @@ jobs:
   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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -791,7 +791,7 @@ jobs:
   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.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch

.github/workflows/generated-windows-binary-wheel-nightly.yml

@@ -44,7 +44,7 @@ jobs:
   wheel-py3_10-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -279,7 +279,7 @@ jobs:
   wheel-py3_10-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -517,7 +517,7 @@ jobs:
   wheel-py3_10-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -755,7 +755,7 @@ jobs:
   wheel-py3_10-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -993,7 +993,7 @@ jobs:
   wheel-py3_10-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -1229,7 +1229,7 @@ jobs:
   wheel-py3_11-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -1464,7 +1464,7 @@ jobs:
   wheel-py3_11-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -1702,7 +1702,7 @@ jobs:
   wheel-py3_11-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -1940,7 +1940,7 @@ jobs:
   wheel-py3_11-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -2178,7 +2178,7 @@ jobs:
   wheel-py3_11-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -2414,7 +2414,7 @@ jobs:
   wheel-py3_12-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -2649,7 +2649,7 @@ jobs:
   wheel-py3_12-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -2887,7 +2887,7 @@ jobs:
   wheel-py3_12-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -3125,7 +3125,7 @@ jobs:
   wheel-py3_12-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -3363,7 +3363,7 @@ jobs:
   wheel-py3_12-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -3599,7 +3599,7 @@ jobs:
   wheel-py3_13-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -3834,7 +3834,7 @@ jobs:
   wheel-py3_13-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -4072,7 +4072,7 @@ jobs:
   wheel-py3_13-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -4310,7 +4310,7 @@ jobs:
   wheel-py3_13-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -4548,7 +4548,7 @@ jobs:
   wheel-py3_13-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -4784,7 +4784,7 @@ jobs:
   wheel-py3_13t-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5019,7 +5019,7 @@ jobs:
   wheel-py3_13t-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5257,7 +5257,7 @@ jobs:
   wheel-py3_13t-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5495,7 +5495,7 @@ jobs:
   wheel-py3_13t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5733,7 +5733,7 @@ jobs:
   wheel-py3_13t-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -5969,7 +5969,7 @@ jobs:
   wheel-py3_14-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -6204,7 +6204,7 @@ jobs:
   wheel-py3_14-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -6442,7 +6442,7 @@ jobs:
   wheel-py3_14-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -6680,7 +6680,7 @@ jobs:
   wheel-py3_14-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -6918,7 +6918,7 @@ jobs:
   wheel-py3_14-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -7154,7 +7154,7 @@ jobs:
   wheel-py3_14t-cpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -7389,7 +7389,7 @@ jobs:
   wheel-py3_14t-cuda12_6-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -7627,7 +7627,7 @@ jobs:
   wheel-py3_14t-cuda12_8-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -7865,7 +7865,7 @@ jobs:
   wheel-py3_14t-cuda13_0-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch
@@ -8103,7 +8103,7 @@ jobs:
   wheel-py3_14t-xpu-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     needs: get-label-type
-    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge"
+    runs-on: "${{ needs.get-label-type.outputs.label-type }}windows.12xlarge"
     timeout-minutes: 360
     env:
       PYTORCH_ROOT: ${{ github.workspace }}/pytorch

.github/workflows/inductor-periodic.yml

@@ -88,7 +88,6 @@ 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/periodic.yml

@@ -147,15 +147,16 @@ 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.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"] },
+          { 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"] },
         ]}
     secrets: inherit
 

.github/workflows/pull.yml

@@ -347,7 +347,8 @@ jobs:
     uses: ./.github/workflows/_linux-build.yml
     needs: get-label-type
     with:
-      sync-tag: linux-xpu-n-build
+      # This should sync with the build in xpu.yml but xpu uses a larger runner
+      # 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

@@ -45,7 +45,6 @@ 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

@@ -42,7 +42,6 @@ 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" },

.github/workflows/rocm-navi31.yml

@@ -0,0 +1,75 @@
+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

@@ -26,11 +26,23 @@ 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
@@ -59,29 +71,3 @@ 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

@@ -58,8 +58,10 @@ jobs:
           else
             COMMIT_SHA="${{ github.sha }}"
           fi
-          echo "sha=${COMMIT_SHA}" >> "${GITHUB_OUTPUT}"
-          echo "tag_name=trunk/${COMMIT_SHA}" >> "${GITHUB_OUTPUT}"
+          {
+            echo "sha=${COMMIT_SHA}"
+            echo "tag_name=trunk/${COMMIT_SHA}"
+          } >> "${GITHUB_OUTPUT}"
 
       - name: Validate commit SHA
         run: |
@@ -87,7 +89,7 @@ jobs:
             echo "✅ Commit ${COMMIT_SHA} is valid (automatic push trigger)"
           fi
 
-      - name: Create and push tag with retry
+      - name: Create and push tag(s) with retry
         id: check_tag
         env:
           TAG_NAME: ${{ steps.commit.outputs.tag_name }}
@@ -112,14 +114,23 @@ jobs:
             return 1
           }
 
-          # 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
+          # Counters for summary reporting
+          created_count=0
+          skipped_count=0
+          failed_count=0
 
-          echo "Tag ${TAG_NAME} does not exist, proceeding with creation"
+          # Always write outputs once on exit
+          finish() {
+            set +e
+            if [ -n "${GITHUB_OUTPUT:-}" ]; then
+              {
+                echo "created_count=${created_count}"
+                echo "skipped_count=${skipped_count}"
+                echo "failed_count=${failed_count}"
+              } >> "${GITHUB_OUTPUT}"
+            fi
+          }
+          trap finish EXIT
 
           # Retry configuration
           MAX_RETRIES=5
@@ -194,31 +205,111 @@ jobs:
             }
           }
 
-          # Execute with retry
-          if retry_with_backoff "tag_with_retry" "Creating tag ${TAG_NAME} for commit ${COMMIT_SHA}"; then
-            echo "exists=false" >> "${GITHUB_OUTPUT}"
+          # 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
             exit 0
           else
-            echo "Tag creation failed after all retry attempts"
-            exit 1
+            # 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
           fi
 
       - name: Tag creation summary
         if: always()
         run: |
-          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 ${{ steps.commit.outputs.sha }}"
+          if [ "${{ 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
           else
-            echo "❌ Failed to create tag ${{ steps.commit.outputs.tag_name }} for commit ${{ steps.commit.outputs.sha }}"
-          fi
+            if [ "${{ 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 }}"
+            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
           fi

.github/workflows/trunk.yml

@@ -190,6 +190,40 @@ 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

.gitignore

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

.lintrunner.toml

@@ -833,8 +833,7 @@ exclude_patterns = [
 command = [
     'python3',
     'tools/linter/adapters/grep_linter.py',
-    '--pattern=cudaSetDevice(',
-    '--pattern=cudaGetDevice(',
+    '--pattern=(cudaSetDevice|cudaGetDevice)\\(',
     '--linter-name=RAWCUDADEVICE',
     '--error-name=raw CUDA API usage',
     """--error-description=\
@@ -1138,11 +1137,8 @@ command = [
 [[linter]]
 code = 'WORKFLOWSYNC'
 include_patterns = [
-    '.github/workflows/pull.yml',
-    '.github/workflows/trunk.yml',
-    '.github/workflows/periodic.yml',
-    '.github/workflows/mac-mps.yml',
-    '.github/workflows/slow.yml',
+    '.github/workflows/*.yml',
+    '.github/workflows/*.yaml',
 ]
 command = [
     'python3',

CODEOWNERS

@@ -201,3 +201,17 @@ 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

@@ -289,14 +289,15 @@ IF(USE_FBGEMM_GENAI)
 
     set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
 
-    set(fbgemm_genai_mx8mx8bf16_grouped
+    set(fbgemm_genai_cuh
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
+      "${FBGEMM_GENAI_SRCS}/"
     )
 
     target_include_directories(fbgemm_genai PRIVATE
       ${FBGEMM_THIRD_PARTY}/cutlass/include
       ${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
-      ${fbgemm_genai_mx8mx8bf16_grouped}
+      ${fbgemm_genai_cuh}
       ${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
     )
@@ -313,13 +314,14 @@ 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

@@ -19,6 +19,7 @@
 #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>
@@ -88,6 +89,8 @@ 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,
@@ -196,7 +199,7 @@ class TORCH_API Context {
     return c10::impl::hasDeviceGuardImpl(c10::DeviceType::IPU);
   }
   static bool hasXLA() {
-    return c10::impl::hasDeviceGuardImpl(c10::DeviceType::XLA);
+    return detail::getXLAHooks().hasXLA();
   }
   static bool hasXPU() {
     return detail::getXPUHooks().hasXPU();

aten/src/ATen/core/CachingHostAllocator.h

@@ -39,7 +39,7 @@ struct HostBlock {
 };
 
 template <typename B>
-struct alignas(64) FreeBlockList {
+struct alignas(hardware_destructive_interference_size) 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(64) HostStatsStaged {
+struct alignas(hardware_destructive_interference_size) 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(64) std::mutex blocks_mutex_;
+  alignas(hardware_destructive_interference_size) 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(64) std::vector<FreeBlockList<B>> free_list_ =
+  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>> free_list_ =
       std::vector<FreeBlockList<B>>(MAX_SIZE_INDEX);
 
-  alignas(64) std::mutex events_mutex_;
+  alignas(hardware_destructive_interference_size) 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(64) HostStatsStaged stats_;
+  alignas(hardware_destructive_interference_size) HostStatsStaged stats_;
 };
 
 struct TORCH_API HostAllocator : public at::Allocator {

aten/src/ATen/core/Generator.h

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

aten/src/ATen/core/TensorBase.h

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

aten/src/ATen/core/VariableFallbackKernel.cpp

@@ -109,6 +109,10 @@ 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

@@ -442,11 +442,17 @@ 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(
-    !backendFallbackKernels_[idx].kernel.isValid(),
-    "Tried to register multiple backend fallbacks for the same dispatch key ", dispatchKey, "; previous registration ",
-    backendFallbackKernels_[idx].debug, ", new registration ", debug
-  );
+      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);
   // 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

@@ -68,11 +68,7 @@ Symbol InternedStrings::_symbol(const std::string& s) {
     return it->second;
 
   auto pos = s.find("::");
-  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());
-  }
+  TORCH_CHECK(pos != std::string::npos, "all symbols must have a namespace, <namespace>::<string>, but found: ", s);
   Symbol ns = _symbol("namespaces::" + s.substr(0, pos));
 
   Symbol sym(sym_to_info_.size());
@@ -121,12 +117,7 @@ std::string Symbol::domainString() const {
 }
 
 Symbol Symbol::fromDomainAndUnqualString(const std::string & d, const std::string & s) {
-  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());
-  }
+  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'");
   std::string qualString = d.substr(domain_prefix().size()) + "::" + s;
   return fromQualString(qualString);
 }

aten/src/ATen/core/ivalue.cpp

@@ -7,6 +7,7 @@
 #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>
@@ -412,7 +413,7 @@ size_t IValue::hash(const IValue& v) {
     case Tag::Enum:
     case Tag::Stream:
     case Tag::Uninitialized:
-      throw std::runtime_error(
+      TORCH_CHECK(false,
           "unhashable type: '" + v.type()->repr_str() + "'");
   }
   // the above switch should be exhaustive

aten/src/ATen/core/jit_type.h

@@ -8,6 +8,7 @@
 #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>
@@ -116,10 +117,8 @@ struct SingleElementType : public SharedType {
 
  protected:
   SingleElementType(TypePtr elem) : SharedType(Kind), elem(std::move(elem)) {
-    if (!this->elem) {
-      throw std::runtime_error(c10::str(
+    TORCH_CHECK(this->elem, c10::str(
             "Can not create ", typeKindToString(Kind), " with None type"));
-    }
   }
 
  private:
@@ -416,16 +415,12 @@ struct TORCH_API SymbolicShape {
   }
 
   ShapeSymbol operator[](size_t i) const {
-    if (!dims_) {
-      throw std::runtime_error("Rank isn't fixed");
-    }
+    TORCH_CHECK(dims_, "Rank isn't fixed");
     return (*dims_).at(i);
   }
 
   ShapeSymbol at(size_t i) const {
-    if (!dims_) {
-      throw std::runtime_error("Rank isn't fixed");
-    }
+    TORCH_CHECK(dims_, "Rank isn't fixed");
     return (*dims_).at(i);
   }
 
@@ -520,9 +515,7 @@ struct VaryingShape {
   }
 
   const std::optional<T> &operator[](size_t i) const {
-    if (!dims_) {
-      throw std::runtime_error("Rank isn't fixed");
-    }
+    TORCH_CHECK(dims_, "Rank isn't fixed");
     return (*dims_).at(i);
   }
 
@@ -957,9 +950,7 @@ struct TORCH_API DictType : public SharedType {
 
   TypePtr createWithContained(
       std::vector<TypePtr> contained_types) const override {
-    if (contained_types.size() != 2) {
-      throw std::runtime_error("Expected 2 contained types");
-    }
+    TORCH_CHECK(contained_types.size() == 2, "Expected 2 contained types");
     return create(std::move(contained_types.at(0)), std::move(contained_types.at(1)));
   }
 

aten/src/ATen/core/jit_type_base.h

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

aten/src/ATen/core/type.cpp

@@ -8,6 +8,7 @@
 #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>
@@ -826,9 +827,7 @@ 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) {
-        if (!v) {
-          throw std::runtime_error("Can not create tuple with None type");
-        }
+        TORCH_CHECK(v, "Can not create tuple with None type");
         return v->hasFreeVariables();
       })), schema_(std::move(schema)) {
 

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

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

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

@@ -6,9 +6,11 @@
 #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

@@ -354,9 +354,47 @@ 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!=)
@@ -364,6 +402,7 @@ 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
@@ -412,28 +451,52 @@ 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
@@ -544,12 +607,19 @@ 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 <>
@@ -557,8 +627,15 @@ 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 <>
@@ -566,8 +643,15 @@ 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 <>
@@ -575,8 +659,15 @@ 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

@@ -5,6 +5,114 @@
 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

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

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

@@ -307,11 +307,49 @@ class Vectorized<float> {
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(exp)
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(exp2)
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(expm1)
+  // Implementation copied from Arm Optimized Routine
+  // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/expf.c
   Vectorized<float> exp_u20() const {
-    return exp();
+    // bail out to sleef if it's a special case:
+    // i.e. there's an input s.t. |input| > 87.3....
+    const float32x4_t special_bound = vdupq_n_f32(0x1.5d5e2ap+6f);
+    uint32x4_t cmp = vcagtq_f32(values, special_bound);
+    if (vpaddd_u64(vreinterpretq_u64_u32(cmp)) != 0) {
+      return exp();
+    }
+
+    const float32x4_t inv_ln2 = vdupq_n_f32(0x1.715476p+0f);
+    const float ln2_hi = 0x1.62e4p-1f;
+    const float ln2_lo = 0x1.7f7d1cp-20f;
+    const float c0 = 0x1.0e4020p-7f;
+    const float c2 = 0x1.555e66p-3f;
+    const float32x4_t ln2_c02 = {ln2_hi, ln2_lo, c0, c2};
+
+    const uint32x4_t exponent_bias = vdupq_n_u32(0x3f800000);
+    const float32x4_t c1 = vdupq_n_f32(0x1.573e2ep-5f);
+    const float32x4_t c3 = vdupq_n_f32(0x1.fffdb6p-2f);
+    const float32x4_t c4 = vdupq_n_f32(0x1.ffffecp-1f);
+
+    /* exp(x) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)]
+      x = ln2*n + r, with r in [-ln2/2, ln2/2].  */
+
+    float32x4_t n = vrndaq_f32(vmulq_f32(values, inv_ln2));
+    float32x4_t r = vfmsq_laneq_f32(values, n, ln2_c02, 0);
+    r = vfmsq_laneq_f32(r, n, ln2_c02, 1);
+    uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_s32(vcvtq_s32_f32(n)), 23);
+    float32x4_t scale = vreinterpretq_f32_u32(vaddq_u32(e, exponent_bias));
+
+    float32x4_t r2 = vmulq_f32(r, r);
+    float32x4_t p = vfmaq_laneq_f32(c1, r, ln2_c02, 2);
+    float32x4_t q = vfmaq_laneq_f32(c3, r, ln2_c02, 3);
+    q = vfmaq_f32(q, p, r2);
+    p = vmulq_f32(c4, r);
+    float32x4_t poly = vfmaq_f32(p, q, r2);
+
+    return vfmaq_f32(scale, poly, scale);
   }
   Vectorized<float> fexp_u20() const {
-    return exp();
+    return exp_u20();
   }
   DEFINE_SLEEF_COMPATIBLE_BINARY_ELEMENTWISE_FUNC_WITH_SLEEF_NAME(
       fmod,
@@ -540,42 +578,6 @@ 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

@@ -569,46 +569,6 @@ 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_uint_aarch64.h

@@ -0,0 +1,378 @@
+#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

@@ -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 = vld1_u8(src.operator const uint8_t*());
+  auto u8x8 = vget_low_u8(src);
   auto u16x8 = vmovl_u8(u8x8);
   auto u32x4_hi = vmovl_u16(vget_high_u16(u16x8));
   auto u32x4_lo = vmovl_u16(vget_low_u16(u16x8));
@@ -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 = vld1_u8(src.operator const uint8_t*());
+  auto u8x8 = vget_low_u8(src);
   auto u16x8 = vmovl_u8(u8x8);
   auto u32x4_lo = vmovl_u16(vget_low_u16(u16x8));
 

aten/src/ATen/cuda/CUDAGreenContext.cpp

@@ -0,0 +1,192 @@
+#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

@@ -0,0 +1,53 @@
+#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

@@ -0,0 +1,270 @@
+#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

@@ -0,0 +1,174 @@
+#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

@@ -183,11 +183,6 @@ 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

@@ -70,11 +70,7 @@
 #define ATEN_CUB_MAXIMUM() NO_ROCM(at_cuda_detail)ROCM_HIPCUB(::cub)::Max()
 #endif
 
-#if (!defined(USE_ROCM) && !CUB_SUPPORTS_NV_BFLOAT16()) || defined(USE_ROCM)
-
-#if !defined(USE_ROCM)
-namespace at_cuda_detail {
-#endif
+#if defined(USE_ROCM)
 
 // backport https://github.com/NVIDIA/cub/pull/306 for c10::BFloat16
 
@@ -96,10 +92,6 @@ 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)
@@ -121,7 +113,7 @@ struct cuda_type<c10::Half> {
   using type = __half;
 };
 
-#if !defined(USE_ROCM) && CUB_SUPPORTS_NV_BFLOAT16()
+#if !defined(USE_ROCM)
 
 template<>
 struct cuda_type<c10::BFloat16> {
@@ -177,7 +169,6 @@ 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,
@@ -193,7 +184,6 @@ 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 {
 
@@ -205,36 +195,6 @@ __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
@@ -279,25 +239,6 @@ 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,
@@ -305,7 +246,6 @@ 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
 }
@@ -557,16 +497,6 @@ 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,
@@ -574,12 +504,10 @@ 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) {
@@ -607,7 +535,6 @@ 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

@@ -10,14 +10,6 @@
 #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
@@ -28,30 +20,6 @@
 #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/detail/XLAHooksInterface.cpp

@@ -0,0 +1,23 @@
+#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

@@ -0,0 +1,79 @@
+#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

@@ -160,6 +160,10 @@ 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/Convolution.cpp

@@ -658,6 +658,7 @@ 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/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 (const std::exception& e) {
+    } catch ([[maybe_unused]] const std::exception& e) {
       TORCH_WARN(func_name, " failed, switching to BLAS gemm: ", e.what());
     }
   }

aten/src/ATen/native/PixelShuffle.h

@@ -11,6 +11,8 @@ 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/cpu/DepthwiseConvKernel.cpp

@@ -259,11 +259,20 @@ 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);
-
-  *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);
+  /* 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;
 }
 
 inline void winograd_f2k3_output_transform_inplace__rvv(
@@ -277,9 +286,15 @@ 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);
-
-  *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);
+  /* 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;
 }
 
 inline vfloat32m1_t
@@ -300,11 +315,17 @@ 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);
-
-  *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);
+  /* 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;
 }
 
 inline vfloat32m1x4_t v4f_transpose4x4__rvv(const vfloat32m1x4_t m) {

aten/src/ATen/native/cuda/Embedding.cu

@@ -15,9 +15,7 @@
 #include <ATen/native/cuda/block_reduce.cuh>
 #include <ATen/native/cuda/thread_constants.h>
 
-#if CUB_SUPPORTS_SCAN_BY_KEY()
 #include <thrust/iterator/reverse_iterator.h>
-#endif
 
 #ifndef AT_PER_OPERATOR_HEADERS
 #include <ATen/Functions.h>
@@ -240,10 +238,6 @@ __global__ void renorm_kernel(
 
 } // anonymous namespace
 
-#if !CUB_SUPPORTS_SCAN_BY_KEY()
-template<typename index_t>
-void embedding_dense_backward_cuda_scan(Tensor &sorted_indices, Tensor &count);
-#endif
 
 Tensor embedding_dense_backward_cuda(const Tensor & grad_, const Tensor & indices_,
                                int64_t num_weights, int64_t padding_idx,
@@ -306,7 +300,6 @@ Tensor embedding_dense_backward_cuda(const Tensor & grad_, const Tensor & indice
 
   if (scale_grad_by_freq) {
     count = at::empty_like(indices, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
-#if CUB_SUPPORTS_SCAN_BY_KEY()
     AT_DISPATCH_INDEX_TYPES(indices.scalar_type(), "embedding_dense_backward_cuda", [&] () {
       cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
@@ -333,11 +326,6 @@ Tensor embedding_dense_backward_cuda(const Tensor & grad_, const Tensor & indice
         num_indices
       );
     });
-#else
-    AT_DISPATCH_INDEX_TYPES(indices.scalar_type(), "embedding_dense_backward_cuda", [&] () {
-      embedding_dense_backward_cuda_scan<index_t>(sorted_indices, count);
-    });
-#endif
   }
 
   return embedding_backward_cuda_kernel(grad, orig_indices,

aten/src/ATen/native/cuda/EmbeddingBackwardKernel.cu

@@ -10,9 +10,7 @@
 
 #include <c10/macros/Macros.h>
 
-#if CUB_SUPPORTS_UNIQUE_BY_KEY()
 #include <thrust/iterator/counting_iterator.h>
-#endif
 
 #ifndef AT_PER_OPERATOR_HEADERS
 #include <ATen/Functions.h>
@@ -196,18 +194,9 @@ __global__ void compute_num_of_partial_segments(const index_t *partials_per_segm
             partials_per_segment_offset[num_of_segments-1];
 }
 
-#if !CUB_SUPPORTS_UNIQUE_BY_KEY()
-__global__ void write_num_of_segments_for_legacy_thrust_path(int64_t *num_of_segments_ptr, int64_t num_of_segments) {
-  *num_of_segments_ptr = num_of_segments;
-}
-#endif
 
 } // anon namespace
 
-#if !CUB_SUPPORTS_UNIQUE_BY_KEY()
-template<typename index_t>
-int64_t embedding_backward_cuda_kernel_unique_by_key(const Tensor &sorted_indices, Tensor &segment_offsets);
-#endif
 
 Tensor embedding_backward_cuda_kernel(
         const Tensor &grad,
@@ -234,20 +223,12 @@ Tensor embedding_backward_cuda_kernel(
   auto segment_offsets = at::empty({numel}, orig_indices.options());
   auto num_of_segments_tensor = at::empty({}, grad.options().dtype(kLong));
   int64_t *num_of_segments_ptr = num_of_segments_tensor.mutable_data_ptr<int64_t>();
-#if !CUB_SUPPORTS_UNIQUE_BY_KEY()
-  AT_DISPATCH_INDEX_TYPES(orig_indices.scalar_type(), "embedding_backward_cuda_kernel", [&] () {
-    int64_t num_of_segments = embedding_backward_cuda_kernel_unique_by_key<index_t>(sorted_indices, segment_offsets);
-    write_num_of_segments_for_legacy_thrust_path<<<1, 1, 0, c10::cuda::getCurrentCUDAStream()>>>(num_of_segments_ptr, num_of_segments);
-    C10_CUDA_KERNEL_LAUNCH_CHECK();
-  });
-#else
   AT_DISPATCH_INDEX_TYPES(orig_indices.scalar_type(), "embedding_backward_cuda_kernel", [&] () {
     cuda::cub::unique_by_key(
       sorted_indices.const_data_ptr<index_t>(), thrust::make_counting_iterator(0),
       segment_offsets.mutable_data_ptr<index_t>(),
       num_of_segments_ptr, sorted_indices.numel());
   });
-#endif
 
   int64_t max_segments = std::min<int64_t>(numel, num_weights);
 

aten/src/ATen/native/cuda/EmbeddingBag.cu

@@ -31,16 +31,10 @@
 
 #include <c10/macros/Macros.h>
 
-#if CUB_SUPPORTS_SCAN_BY_KEY()
 #include <thrust/iterator/reverse_iterator.h>
-#endif
 
 namespace at::native {
 
-#if !CUB_SUPPORTS_SCAN_BY_KEY()
-template<typename index_t>
-void embedding_dense_backward_cuda_scan(Tensor &sorted_indices, Tensor &count);
-#endif
 
 namespace {
 
@@ -199,7 +193,6 @@ Tensor embedding_bag_backward_cuda_sum_avg(
 
   if (scale_grad_by_freq) {
     count = at::empty_like(indices, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
-#if CUB_SUPPORTS_SCAN_BY_KEY()
     AT_DISPATCH_INDEX_TYPES(indices.scalar_type(), "embedding_bag_backward_cuda_sum_avg", [&] () {
       cudaStream_t stream = at::cuda::getCurrentCUDAStream();
 
@@ -226,11 +219,6 @@ Tensor embedding_bag_backward_cuda_sum_avg(
         num_indices
       );
     });
-#else
-    AT_DISPATCH_INDEX_TYPES(indices.scalar_type(), "embedding_bag_backward_cuda_sum_avg", [&] () {
-      embedding_dense_backward_cuda_scan<index_t>(sorted_indices, count);
-    });
-#endif
   }
   return embedding_backward_cuda_kernel(grad, orig_indices, sorted_indices,
       count, num_weights, padding_idx, mode == EmbeddingBagMode::MEAN, offset2bag,

aten/src/ATen/native/cuda/GroupedBlas.cpp

@@ -0,0 +1,574 @@
+#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/CUDAScaledBlas.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 scaled_blas = at::cuda::scaled;
+using scaled_blas::ScaledGemmImplementation;
+using scaled_blas::convert_int_to_enum;
+using scaled_blas::_scaled_mm_allowed_device;
+
+namespace at::native {
+
+namespace {
+
+// 2d-2d and 2d-3d
+// scaling=MXFP8
+// CUDA-only
+Tensor&
+_mx8_mx8_bf16_grouped_mm_fbgemm(
+        const Tensor& mat_a,
+        const Tensor& mat_b,
+        const Tensor& scale_a,
+        const SwizzleType& swizzle_a,
+        const Tensor& scale_b,
+        const SwizzleType& swizzle_b,
+        const std::optional<at::Tensor>& offs,
+        Tensor& out) {
+    const bool a_is_2d = mat_a.dim() == 2;
+    const bool b_is_2d = mat_b.dim() == 2;
+    bool b_is_3d = mat_b.dim() == 3;
+    bool is_2d_2d = a_is_2d && b_is_2d;
+    bool is_2d_3d = a_is_2d && b_is_3d;
+    TORCH_CHECK_VALUE(is_2d_2d || is_2d_3d, "MXFP8 grouped GEMM currently only supports 2d-2d and 2d-3d cases");
+    TORCH_CHECK_VALUE(offs.has_value(), "MXFP8 2d-2d and 2d-3d grouped GEMMs requires offsets");
+    TORCH_CHECK_VALUE(out.scalar_type() == at::kBFloat16, "Only bf16 out_dtype is supported for MXFP8 grouped gemm");
+    // MXFP8 expects float8_e8m0fnu scales.
+    TORCH_CHECK_VALUE(scale_a.scalar_type() == at::kFloat8_e8m0fnu && scale_b.scalar_type() == at::kFloat8_e8m0fnu,
+        "For MXFP8 grouped gemm, both scales must be float8_e8m0fnu tensors.");
+#ifdef USE_ROCM
+    TORCH_CHECK_VALUE(swizzle_a == SwizzleType::NO_SWIZZLE && swizzle_b == SwizzleType::NO_SWIZZLE,
+        "For ROCM MXFP8 grouped gemm, both scale swizzle types must be SWIZZLE_NONE");
+#else
+    TORCH_CHECK_VALUE(swizzle_a == SwizzleType::SWIZZLE_32_4_4 && swizzle_b == SwizzleType::SWIZZLE_32_4_4,
+        "For CUDA MXFP8 grouped gemm, both scale swizzle types must be SWIZZLE_32_4_4");
+#endif
+
+#if defined(USE_FBGEMM_GENAI) and !defined(USE_ROCM)
+    fbgemm_gpu::mx8mx8bf16_grouped_mm(
+        mat_a,
+        mat_b,
+        scale_a,
+        scale_b,
+        offs.value(),
+        out);
+#else
+    TORCH_CHECK_NOT_IMPLEMENTED(false, "mxfp8_mxfp8 grouped gemm requires compile with USE_FBGEMM_GENAI");
+#endif
+    return out;
+}
+
+// 2d-2d and 2d-3d cases
+// scaling=rowwise
+// CUDA-only
+Tensor&
+_f8_f8_bf16_rowwise_grouped_mm_cuda(
+          const Tensor& mat_a,
+          const Tensor& mat_b,
+          const Tensor& scale_a,
+          const Tensor& scale_b,
+          const std::optional<Tensor>& offs,
+          const std::optional<Tensor>& bias,
+          const bool use_fast_accum,
+          Tensor& out) {
+  TORCH_CHECK_VALUE(mat_a.dtype() == at::kFloat8_e4m3fn, "Expected mat_a to be Float8_e4m3 matrix got ", mat_a.scalar_type());
+  TORCH_CHECK_VALUE(mat_b.dtype() == at::kFloat8_e4m3fn, "Expected mat_a to be Float8_e4m3 matrix got ", mat_b.scalar_type());
+
+  at::cuda::detail::f8f8bf16_grouped_mm(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      offs,
+      bias,
+      use_fast_accum,
+      out);
+    return out;
+}
+
+// 2d-2d and 2d-3d cases
+// scaling=rowwise
+// only being called for rocm
+Tensor&
+_f8_f8_bf16_rowwise_grouped_mm_rocm(
+      const Tensor& mat_a,
+      const Tensor& mat_b,
+      const Tensor& scale_a,
+      const Tensor& scale_b,
+      const std::optional<Tensor>& offs,
+      Tensor& out) {
+  TORCH_CHECK_VALUE(mat_a.dtype() == at::kFloat8_e4m3fnuz, "Expected mat_a to be Float8_e4m3fnuz matrix got ", mat_a.scalar_type());
+  TORCH_CHECK_VALUE(mat_b.dtype() == at::kFloat8_e4m3fnuz, "Expected mat_a to be Float8_e4m3fnuz matrix got ", mat_b.scalar_type());
+
+#if defined(USE_FBGEMM_GENAI) && defined(USE_ROCM)
+  fbgemm_gpu::f8f8bf16_rowwise_grouped_mm(
+      mat_a,
+      // FBGEMM expects B matrix shape to be (.., N, K)
+      mat_b.transpose(-2, -1),
+      scale_a,
+      scale_b,
+      offs,
+      out);
+#else
+  TORCH_CHECK_NOT_IMPLEMENTED(false, "grouped gemm is not supported without USE_FBGEMM_GENAI on ROCM")
+#endif
+  return out;
+
+}
+
+// Dispatch f8 x f8 -> bf16 row-wise scaled to rocm/cuda
+Tensor&
+_f8_f8_bf16_rowwise_grouped_mm(
+      const Tensor& mat_a,
+      const Tensor& mat_b,
+      const Tensor& scale_a,
+      const Tensor& scale_b,
+      const std::optional<Tensor>& offs,
+      const std::optional<Tensor>& bias,
+      bool use_fast_accum,
+      Tensor& out) {
+  // FP8 per-tensor and per-row scaling expect fp32 scales.
+  TORCH_CHECK_VALUE(scale_a.scalar_type() == kFloat && scale_b.scalar_type() == kFloat,
+      "For grouped FP8 rowwise, both scales must be float32 tensors");
+#ifndef USE_ROCM
+  return _f8_f8_bf16_rowwise_grouped_mm_cuda(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      offs,
+      bias,
+      use_fast_accum,
+      out);
+#else
+  // NOTE: ignore use_fast_accum
+  TORCH_CHECK_VALUE(!bias.has_value(), "ROCM grouped gemm does not support bias")
+  return _f8_f8_bf16_rowwise_grouped_mm_rocm(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      offs,
+      out);
+#endif
+}
+
+void _check_scales_fp8_rowwise(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx, const int scale_multiplier=1) {
+  // Checks scales for 2d or 3d target tensors (`mat`).
+  if (mat.dim() == 2) {
+    TORCH_CHECK(
+        scale.dim() == 1,
+        "scale must be a 1D tensor, but got ",
+        scale.dim(),
+        "D, arg ",
+        arg_idx);
+    TORCH_CHECK(
+        scale.is_contiguous(), "scale must be contiguous for arg ", arg_idx);
+    TORCH_CHECK(
+        scale.size(0) == mat.size(dim) * scale_multiplier,
+        "scale must have the same length as mat for arg ",
+        arg_idx);
+  } else {
+    TORCH_CHECK(
+        scale.dim() == 2,
+        "scale must be a 2D tensor, but got ",
+        scale.dim(),
+        "D for arg ",
+        arg_idx);
+    TORCH_CHECK(
+        scale.stride(1) == 1,
+        "scale must be contiguous in the last dimension for arg ",
+        arg_idx);
+    TORCH_CHECK(
+        scale.size(0) == mat.size(0),
+        "scale must have the same batch dimension as mat for arg ",
+        arg_idx);
+    TORCH_CHECK(
+        scale.size(1) == mat.size(1 + dim),
+        "scale must have the same first dimension as mat for arg ",
+        arg_idx);
+  }
+}
+
+void _check_scales_mxfp8(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx) {
+  // Checks scales for 2d or 3d target tensors (`mat`).
+  if (mat.dim() == 2) {
+    // For MXFP8, 2d tensors have variable size groups represented as subtensors,
+    // that are converted to blocked padded format individually,
+    // so we can't check the scale sizes without doing a d2h sync to get the group sizes here.
+    TORCH_CHECK(
+      scale.dim() == mat.dim(),
+      "for mxfp8, scale must have same number of dimensions as parent tensor, but got mat.dim() = ", mat.dim(), " and scale.dim() = ", scale.dim(), " for arg ", arg_idx);
+
+    // LHS mat shape (M, total_K) -> scale shape (rounded_up(M, 128), rounded_up_per_group(K/32, 4))
+    // RHS mat shape (total_K, N) -> scale shape (rounded_up(N, 128), rounded_up_per_group(K/32, 4))
+    //   * weight is transposed prior to the call, scale stays non-transposed.
+    bool LHS = arg_idx == 0;
+    int scale_dim_to_check = 0;
+    int mat_dim_to_check = LHS ? 0 : 1;
+    TORCH_CHECK(
+        scale.size(scale_dim_to_check) >= mat.size(mat_dim_to_check),
+        "for mxfp8, arg ", arg_idx, " tensor shape (", mat.size(0), ", ", mat.size(1), ") ",
+        "must have scale.shape[", scale_dim_to_check, "] >= ", mat.size(mat_dim_to_check), " but got scale.shape=(", scale.size(0), ", ", scale.size(1), ")");
+  } else {
+    // For MXFP8, 3d tensors have static group sizes (stack of 2d tensors),
+    // so we can check the exact expected scale sizes here without a d2h sync.
+    auto round_up = [](auto x, auto y) {
+        return ((x + y - 1) / y) * y;
+    };
+
+    // TODO: this is for 3d tensor in 2d-3d case specifically.
+    // We'll need to support 3d-3d and 3d-2d cases once mxfp8 grouped gemm supports them.
+    int64_t G = mat.size(0);
+    int64_t K = mat.size(1);
+    int64_t N = mat.size(2);
+    int64_t blocked_scale_K = round_up(K/32, 4);
+    int64_t blocked_scale_N = round_up(N, 128);
+
+    // fbgemm expects stack of flattened blocked scales for 3d tensor, shape (G, blocked_scale_K * blocked_scale_N).
+    TORCH_CHECK(
+      scale.dim() == mat.dim() - 1,
+      "for mxfp8 2d-3d grouped GEMM, the 3d tensor of shape (G,K,N) must have a 2d scale of shape (G, blocked_scale_K * blocked_scale_N), but scale is ", scale.dim(), "D for arg ", arg_idx
+    );
+    TORCH_CHECK(
+      scale.size(0) == G && scale.size(1) == blocked_scale_K * blocked_scale_N,
+      "for mxfp8, the tensor shape (", G, ", ", K, ", ", N, ") must have scale shape (", G, ",", blocked_scale_K, ",", blocked_scale_N, ") for arg ", arg_idx
+    );
+  }
+}
+
+void check_scale(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx, const int scale_multiplier=1) {
+  bool using_fp8_rowwise = scale.scalar_type() == kFloat;
+  bool using_mxfp8 = scale.scalar_type() == at::kFloat8_e8m0fnu;
+  if (using_fp8_rowwise) {
+    _check_scales_fp8_rowwise(mat, scale, dim, arg_idx, scale_multiplier);
+  } else if (using_mxfp8) {
+    _check_scales_mxfp8(mat, scale, dim, arg_idx);
+  } else {
+    TORCH_CHECK(false, "scale must be float32 or float8_e8m0fnu, but got ", scale.dtype());
+  }
+}
+
+} // namespace
+
+Tensor
+_scaled_grouped_mm_cuda(
+        const Tensor& mat_a,
+        const Tensor& mat_b,
+        const Tensor& scale_a,
+        const Tensor& scale_b,
+        const std::optional<at::Tensor>& offs,
+        const std::optional<at::Tensor>& bias,
+        const std::optional<at::Tensor>& scale_result,
+        std::optional<c10::ScalarType> out_dtype,
+        bool use_fast_accum) {
+  bool allowed_device = _scaled_mm_allowed_device(/*sm90_only*/true, /*sm100_only*/true);
+  TORCH_CHECK_VALUE(allowed_device, "torch._scaled_grouped_mm is only supported on CUDA devices with compute capability = [9.0, 10.0], or ROCm MI300+");
+
+  TORCH_CHECK_VALUE(!check_valid_strides_and_return_transposed(mat_a), "Expected mat1 to not be transposed");
+  TORCH_CHECK_VALUE(check_valid_strides_and_return_transposed(mat_b), "Expected mat2 to be transposed");
+  TORCH_CHECK_VALUE(mat_a.dim() == 2 || mat_a.dim() == 3, "mat_a has to be 2 or 3d");
+  TORCH_CHECK_VALUE(mat_b.dim() == 2 || mat_b.dim() == 3, "mat_b has to be 2 or 3d");
+  const bool a_is_2d = mat_a.dim() == 2;
+  const bool b_is_2d = mat_b.dim() == 2;
+
+  // NOTE(slayton): For sub-1B formats want contraction_dim argument?
+  if (!a_is_2d || !b_is_2d) {
+    TORCH_CHECK_VALUE(mat_a.size(-1) == mat_b.size(-2), "contraction dimension of mat_a and mat_b must match");
+  }
+  TORCH_CHECK_VALUE(
+    mat_a.size(-1) % 16 == 0,
+    "Expected trailing dimension of mat_a to be divisible by 16 ",
+    "but got mat1 shape: (",
+    mat_a.sizes(),
+    ").");
+  TORCH_CHECK_VALUE(mat_b.size(-2) % 16 == 0 && mat_b.size(-1) % 16 == 0,
+    "Expected mat_b shape to be divisible by 16 ",
+    "but got mat_b shape: (",
+    mat_b.sizes(),
+    ").");
+
+
+  TORCH_CHECK_VALUE(!bias.has_value(), "Bias not supported yet");
+  TORCH_CHECK_VALUE(!scale_result.has_value(), "Scale result not supported yet");
+  TORCH_CHECK_VALUE(offs.has_value() ==  (a_is_2d || b_is_2d), "Have to provide offsets if there is a 2d matrix");
+
+  // NOTE: mxfp8 x mxfp8 requires (and asserts later) that offsets is present.
+  //       for rowwise, no offsets implies 3d-3d and is handled by lower-level
+  //       routines
+  if (offs.has_value()) {
+    TORCH_CHECK_VALUE(offs->dim() == 1, "offs has to be 1D");
+    TORCH_CHECK_VALUE(offs->dtype() == at::kInt, "Offsets have to be int32");
+  }
+  // FP8 per-tensor and per-row scaling expect fp32 scales.
+  // MXFP8 expects float8_e8m0fnu scales.
+  TORCH_CHECK_VALUE(
+      (scale_a.scalar_type() == kFloat && scale_b.scalar_type() == kFloat) ||
+      (scale_a.scalar_type() == at::kFloat8_e8m0fnu && scale_b.scalar_type() == at::kFloat8_e8m0fnu),
+      "For FP8 tensorwise and rowwise, both scales must both be float32 tensors. For MXFP8, scales must both be float8_e8m0fnu tensors.");
+
+  const int scale_multiplier = (mat_a.dim() == 2 && mat_b.dim() == 2) ? offs->size(0) : 1;
+  check_scale(mat_a, scale_a, 0 ,0, scale_multiplier);
+  check_scale(mat_b, scale_b, 1, 1, scale_multiplier);
+
+  const auto out_dtype_ = out_dtype.value_or(kBFloat16);
+  TORCH_CHECK_VALUE(out_dtype_ == kBFloat16, "Only bf16 high precision output types are supported for grouped gemm");
+
+  Tensor out = create_grouped_gemm_output_tensor(mat_a, mat_b, offs, out_dtype_);
+
+#if defined(USE_FBGEMM_GENAI) && defined(USE_CUDA) && !defined(USE_ROCM)
+  // MXFP8 grouped GEMM dispatching
+  bool is_mx8mx8bf16 = (
+    mat_a.scalar_type() == at::kFloat8_e4m3fn && mat_b.scalar_type() == at::kFloat8_e4m3fn &&
+    scale_a.scalar_type() == at::kFloat8_e8m0fnu && scale_b.scalar_type() == at::kFloat8_e8m0fnu
+  );
+#else
+  bool is_mx8mx8bf16 = false;
+#endif
+
+  if (is_mx8mx8bf16) {
+    // Note: Passing implied SwizzleType here, correctness of scale previously checked
+    //       in `check_scale` call
+    return _mx8_mx8_bf16_grouped_mm_fbgemm(
+        mat_a,
+        mat_b,
+        scale_a,
+        SwizzleType::SWIZZLE_32_4_4,
+        scale_b,
+        SwizzleType::SWIZZLE_32_4_4,
+        offs.value(),
+        out);
+  }
+
+  // If we're not MXFP8, then we're row-wise scaling.
+  return _f8_f8_bf16_rowwise_grouped_mm(
+      mat_a,
+      mat_b,
+      scale_a,
+      scale_b,
+      offs,
+      bias,
+      use_fast_accum,
+      out);
+}
+
+namespace {
+
+using acceptance_fn = std::function<bool(c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&, c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&)>;
+
+std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 2> scale_grouped_kernel_dispatch = {{
+  { "rowwise_rowwise", scaled_blas::check_rowwise_recipe, ScaledGemmImplementation::ROWWISE_ROWWISE},
+  { "mxfp8_mxfp8", scaled_blas::check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8}}};
+
+} // anonymous namespace
+
+Tensor
+_scaled_grouped_mm_cuda_v2(
+          const Tensor& mat_a, const Tensor& mat_b,
+          ArrayRef<Tensor> scale_a,
+          IntArrayRef scale_recipe_a,
+          IntArrayRef swizzle_a,
+          ArrayRef<Tensor> scale_b,
+          IntArrayRef scale_recipe_b,
+          IntArrayRef swizzle_b,
+          const std::optional<Tensor>& offs,
+          const std::optional<Tensor>& bias,
+          const std::optional<c10::ScalarType> out_dtype,
+          IntArrayRef contraction_dim,
+          bool use_fast_accum) {
+  bool allowed_device = _scaled_mm_allowed_device(/*sm90_only*/true, /*sm100_only*/true);
+  TORCH_CHECK_VALUE(allowed_device, "torch._scaled_grouped_mm is only supported on CUDA devices with compute capability = [9.0, 10.0], or ROCm MI300+");
+
+  TORCH_CHECK_VALUE(!check_valid_strides_and_return_transposed(mat_a), "Expected mat1 to not be transposed");
+  TORCH_CHECK_VALUE(check_valid_strides_and_return_transposed(mat_b), "Expected mat2 to be transposed");
+  TORCH_CHECK_VALUE(mat_a.dim() == 2 || mat_a.dim() == 3, "mat_a has to be 2 or 3d");
+  TORCH_CHECK_VALUE(mat_b.dim() == 2 || mat_b.dim() == 3, "mat_b has to be 2 or 3d");
+  const bool a_is_2d = mat_a.dim() == 2;
+  const bool b_is_2d = mat_b.dim() == 2;
+
+  // NOTE(slayton): For sub-1B formats want contraction_dim argument?
+  if (!a_is_2d || !b_is_2d) {
+    if (contraction_dim.size() > 0) {
+      const int dim_a = contraction_dim[0], dim_b = mat_b.size(contraction_dim[1]);
+      TORCH_CHECK_VALUE(mat_a.size(dim_a) == mat_b.size(dim_b),
+          "Contraction dimensions (", dim_a, ",", dim_b, ") of mat_a and mat_b must match, got: ", mat_a.size(dim_a), " and ",
+          mat_b.size(dim_b));
+      // Note: only (-1, -2) is currently supported
+      TORCH_CHECK_VALUE(dim_a == -1 && dim_b == -2, "Curently contraction dims must be (-1, -2) only");
+    } else {
+      TORCH_CHECK_VALUE(mat_a.size(-1) == mat_b.size(-2), "contraction dimension of mat_a and mat_b must match");
+    }
+  }
+  TORCH_CHECK_VALUE(
+    mat_a.size(-1) % 16 == 0,
+    "Expected trailing dimension of mat_a to be divisible by 16 ",
+    "but got mat1 shape: (",
+    mat_a.sizes(),
+    ").");
+  TORCH_CHECK_VALUE(mat_b.size(-2) % 16 == 0 && mat_b.size(-1) % 16 == 0,
+    "Expected mat_b shape to be divisible by 16 ",
+    "but got mat_b shape: (",
+    mat_b.sizes(),
+    ").");
+
+  TORCH_CHECK_VALUE(!bias.has_value(), "Bias not supported yet");
+  TORCH_CHECK_VALUE(offs.has_value() ==  (a_is_2d || b_is_2d), "Have to provide offsets if there is a 2d matrix");
+
+  // NOTE: mxfp8 x mxfp8 requires (and asserts later) that offsets is present.
+  //       for rowwise, no offsets implies 3d-3d and is handled by lower-level
+  //       routines
+  if (offs.has_value()) {
+    TORCH_CHECK_VALUE(offs->dim() == 1, "offs has to be 1D");
+    TORCH_CHECK_VALUE(offs->dtype() == at::kInt, "Offsets have to be int32");
+  }
+
+  const auto out_dtype_ = out_dtype.value_or(kBFloat16);
+  TORCH_CHECK_VALUE(out_dtype_ == kBFloat16, "Only bf16 high precision output types are supported for grouped gemm");
+
+  Tensor out = create_grouped_gemm_output_tensor(mat_a, mat_b, offs, out_dtype_);
+
+  // Conversion of implicitly-defined enums to explicit
+  auto scale_recipe_a_enum = convert_int_to_enum<ScalingType>(scale_recipe_a);
+  auto swizzle_a_enum = convert_int_to_enum<SwizzleType>(swizzle_a);
+  auto scale_recipe_b_enum = convert_int_to_enum<ScalingType>(scale_recipe_b);
+  auto swizzle_b_enum = convert_int_to_enum<SwizzleType>(swizzle_b);
+
+  // at this point we can start working out what we want to be doing
+  // Try to do as few steps as possible.
+  // NOTE: support is deliberately sparse, can explicitly enumerate all combinations allowed.
+  // Do this via a list of defined (name, acceptance, concrete_impl) tuples.
+  ScaledGemmImplementation gemm_impl = ScaledGemmImplementation::NONE;
+  for (const auto& fn_entry : scale_grouped_kernel_dispatch) {
+    const auto [name, accept_fn, scaled_gemm_impl] = fn_entry;
+    bool ok = accept_fn(mat_a.scalar_type(),
+                        scale_recipe_a_enum,
+                        scale_a,
+                        mat_b.scalar_type(),
+                        scale_recipe_b_enum,
+                        scale_b);
+    if (ok) {
+      gemm_impl = scaled_gemm_impl;
+      break;
+    }
+  }
+  TORCH_CHECK_VALUE(gemm_impl != ScaledGemmImplementation::NONE,
+      "No gemm implementation was found");
+
+  switch (gemm_impl) {
+    case ScaledGemmImplementation::ROWWISE_ROWWISE: {
+      const int scale_multiplier = (mat_a.dim() == 2 && mat_b.dim() == 2) ? offs->size(0) : 1;
+      _check_scales_fp8_rowwise(mat_a, scale_a[0], 0 /* dim */ , 0 /* arg_idx */, scale_multiplier);
+      _check_scales_fp8_rowwise(mat_b, scale_b[0], 1 /* dim */ , 1 /* arg_idx */, scale_multiplier);
+      return _f8_f8_bf16_rowwise_grouped_mm(
+          mat_a,
+          mat_b,
+          scale_a[0],
+          scale_b[0],
+          offs,
+          bias,
+          use_fast_accum,
+          out);
+    }
+    case ScaledGemmImplementation::MXFP8_MXFP8: {
+      _check_scales_mxfp8(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
+      _check_scales_mxfp8(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
+      return _mx8_mx8_bf16_grouped_mm_fbgemm(
+          mat_a,
+          mat_b,
+          scale_a[0],
+          swizzle_a_enum[0],
+          scale_b[0],
+          swizzle_b_enum[0],
+          offs.value(),
+          out);
+    }
+    default:
+      TORCH_CHECK_NOT_IMPLEMENTED(false,
+          "_scaled_grouped_mm_cuda_v2 is in an inconsistent state - should never reach here");
+  }
+}
+
+Tensor _grouped_mm_cuda(const Tensor& mat_a, const Tensor& mat_b,
+const std::optional<at::Tensor>& offs,
+const std::optional<at::Tensor>& bias,
+std::optional<c10::ScalarType> out_dtype) {
+  _grouped_mm_validate_inputs(mat_a, mat_b, offs, bias, out_dtype);
+  bool a_b_and_out_are_bf16 = (
+    mat_a.dtype() == at::kBFloat16 &&
+    mat_b.dtype() == at::kBFloat16 &&
+    out_dtype.value_or(at::kBFloat16) == at::kBFloat16
+  );
+#ifndef USE_ROCM
+  bool use_fast_path = _scaled_mm_allowed_device(/*sm90_only*/true, /*sm100_only*/true) && a_b_and_out_are_bf16;
+#else
+  // _scaled_mm_allowed_device is used here within _grouped_mm_cuda which seems incorrect since scale is not used.
+  // the _grouped_mm_fallback should be safe for any ROCm GPU since it's just calling typical mm/bmm
+  bool use_fast_path = false;
+#endif
+  const auto out_dtype_ = _resolve_grouped_mm_out_dtype(mat_a, mat_b, out_dtype);
+  Tensor out = create_grouped_gemm_output_tensor(mat_a, mat_b, offs, out_dtype_);
+  if (use_fast_path) {
+    // fast path, no d2h sync needed
+    at::cuda::detail::bf16bf16_grouped_mm(mat_a, mat_b, offs, bias, out);
+  } else {
+    _grouped_mm_fallback(mat_a, mat_b, offs, bias, out_dtype, out);
+  }
+  return out;
+}
+
+} // namespace at::native

aten/src/ATen/native/cuda/LegacyThrustHelpers.cu

@@ -1,90 +0,0 @@
-#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
-#include <ATen/core/Tensor.h>
-#include <ATen/native/cuda/SortingCommon.cuh>
-#include <ATen/cuda/cub_definitions.cuh>
-
-#ifndef AT_PER_OPERATOR_HEADERS
-#include <ATen/Functions.h>
-#else
-#include <ATen/ops/empty_like.h>
-#endif
-
-#include <ATen/cuda/ThrustAllocator.h>
-#include <thrust/device_ptr.h>
-#include <thrust/execution_policy.h>
-#include <thrust/sort.h>
-#include <thrust/unique.h>
-#include <thrust/device_ptr.h>
-#include <thrust/iterator/constant_iterator.h>
-
-namespace at::native {
-
-#if !CUB_SUPPORTS_SCAN_BY_KEY()
-
-template<typename index_t>
-void embedding_dense_backward_cuda_scan(Tensor &sorted_indices, Tensor &count) {
-  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
-  at::cuda::ThrustAllocator allocator;
-  auto policy = thrust::cuda::par(allocator).on(stream);
-
-  auto num_indices = count.numel();
-
-  // Compute an increasing sequence per unique item in sortedIndices:
-  // sorted: 2 5 5 5 7 7 8 9 9
-  //  count: 1 1 2 3 1 2 1 1 2
-  auto sorted_data = thrust::device_ptr<const index_t>(sorted_indices.const_data_ptr<index_t>());
-  auto count_data = thrust::device_ptr<index_t>(count.mutable_data_ptr<index_t>());
-  thrust::inclusive_scan_by_key(
-    policy,
-    sorted_data,
-    sorted_data + num_indices,
-    thrust::make_constant_iterator(1),
-    count_data
-  );
-
-  // Take the maximum of each count per unique key in reverse:
-  // sorted: 2 5 5 5 7 7 8 9 9
-  //  count: 1 3 3 3 2 2 1 2 2
-  thrust::inclusive_scan_by_key(
-    policy,
-    thrust::make_reverse_iterator(sorted_data + num_indices),
-    thrust::make_reverse_iterator(sorted_data),
-    thrust::make_reverse_iterator(count_data + num_indices),
-    thrust::make_reverse_iterator(count_data + num_indices),
-    thrust::equal_to<index_t>(),
-    thrust::maximum<index_t>()
-  );
-}
-
-template
-void embedding_dense_backward_cuda_scan<int>(Tensor &sorted_indices, Tensor &count);
-template
-void embedding_dense_backward_cuda_scan<int64_t>(Tensor &sorted_indices, Tensor &count);
-
-#endif
-
-template<typename index_t>
-int64_t embedding_backward_cuda_kernel_unique_by_key(const Tensor &sorted_indices, Tensor &segment_offsets) {
-  auto stream = at::cuda::getCurrentCUDAStream();
-  at::cuda::ThrustAllocator allocator;
-  auto policy = thrust::cuda::par(allocator).on(stream);
-  const ptrdiff_t numel = sorted_indices.numel();
-  auto sorted_indices_dev = thrust::device_ptr<const index_t>(sorted_indices.const_data_ptr<index_t>());
-  auto dummy = at::empty_like(sorted_indices, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
-  auto dummy_dev = thrust::device_ptr<index_t>(dummy.mutable_data_ptr<index_t>());
-  auto ends = thrust::unique_by_key_copy(
-          policy,
-          sorted_indices_dev,
-          sorted_indices_dev + numel,
-          thrust::make_counting_iterator(0),
-          dummy_dev,
-          thrust::device_ptr<index_t>(segment_offsets.mutable_data_ptr<index_t>()));
-  return thrust::get<0>(ends) - dummy_dev;
-}
-
-template
-int64_t embedding_backward_cuda_kernel_unique_by_key<int>(const Tensor &sorted_indices, Tensor &segment_offsets);
-template
-int64_t embedding_backward_cuda_kernel_unique_by_key<int64_t>(const Tensor &sorted_indices, Tensor &segment_offsets);
-
-} // namespace at::native

aten/src/ATen/native/cuda/Loops.cuh

@@ -1,18 +1,17 @@
 #pragma once
 
+#include <ATen/OpMathType.h>
+#include <ATen/cuda/detail/OffsetCalculator.cuh>
 #include <ATen/detail/FunctionTraits.h>
 #include <ATen/native/TensorIterator.h>
 #include <ATen/native/TensorIteratorDynamicCasting.h>
-#include <ATen/cuda/detail/OffsetCalculator.cuh>
-#include <ATen/OpMathType.h>
 #include <ATen/native/cuda/thread_constants.h>
-
-#include <thrust/tuple.h>
-
 #include <ATen/native/cuda/MemoryAccess.cuh>
 
 #include <tuple>
 
+
+
 namespace at::native {
 
 template<int N>
@@ -62,7 +61,11 @@ __device__ inline void elementwise_kernel_helper(func_t f, policy_t policy) {
   #pragma unroll
   for (int i = 0; i < elems_per_thread; i++) {
     if (policy.check_inbounds(i)) {
+#if defined(__HIP__)
       results[i] = c10::guts::apply(f, args[i]);
+#else
+      results[i] = std::apply(f, args[i]);
+#endif
     }
   }
 

aten/src/ATen/native/cuda/NLLLoss2d.cu

@@ -146,6 +146,7 @@ __global__ void nll_loss2d_backward_no_reduce_kernel(
   int64_t batch_size = target.size(0);
   int64_t H = target.size(1);
   int64_t W = target.size(2);
+  int64_t n_classes = grad_input.size(1);
 
   CUDA_KERNEL_LOOP(index, n_threads) {
     const int64_t b = index % batch_size;
@@ -156,6 +157,7 @@ __global__ void nll_loss2d_backward_no_reduce_kernel(
     if (cur_target == ignore_index) {
       continue;
     }
+    CUDA_KERNEL_ASSERT(cur_target >= 0 && cur_target < n_classes);
     scalar_t value = -(weight != nullptr ? weight[cur_target] : static_cast<scalar_t>(1));
     grad_input[b][cur_target][h][w] = value * grad_output[b][h][w];
   }

aten/src/ATen/native/cuda/Normalization.cuh

@@ -23,7 +23,7 @@ namespace at::native {
 
 // The maximum number of threads in a block
 #if defined(USE_ROCM)
-constexpr int MAX_BLOCK_SIZE = 256;
+constexpr int MAX_BLOCK_SIZE = 1024;
 #else
 constexpr int MAX_BLOCK_SIZE = 512;
 #endif
@@ -33,7 +33,7 @@ constexpr unsigned MAX_GRID_SIZE = 65535u;
 // Number of threads in a block given an input size up to MAX_BLOCK_SIZE
 static int getNumThreads(int nElem) {
 #if defined(USE_ROCM)
-  int threadSizes[5] = { 16, 32, 64, 128, MAX_BLOCK_SIZE };
+  int threadSizes[5] = { 64, 128, 256, 512, MAX_BLOCK_SIZE };
 #else
   int threadSizes[5] = { 32, 64, 128, 256, MAX_BLOCK_SIZE };
 #endif
@@ -115,9 +115,23 @@ __device__ scalar_t reduce(Op op, PTA tensor, int plane) {
   // first the reductions each thread does separately
   scalar_t sum = static_cast<scalar_t>(0);
   for (int batch = threadIdx.y; batch < tensor.size(0); batch += blockDim.y) {
+#if defined(USE_ROCM)
+    constexpr int UNRL = 4; // load deserilize factor
+    scalar_t tmp[UNRL];
+    for (int x = threadIdx.x; x < tensor.size(2); x += blockDim.x*UNRL) {
+#pragma unroll
+      for (int u = 0; u < UNRL; u++)
+        tmp[u] = op(batch, plane, std::min((int)tensor.size(2)-1, (int)(x+u*blockDim.x)));
+#pragma unroll
+      for (int u = 0; u < UNRL; u++)
+        if (x+u*blockDim.x < tensor.size(2))
+          sum += tmp[u];
+    }
+#else
     for (int x = threadIdx.x; x < tensor.size(2); x += blockDim.x) {
       sum += op(batch, plane, x);
     }
+#endif
   }
   __shared__ scalar_t shared[C10_WARP_SIZE];
   SumReduceOp<scalar_t> reduce_op;
@@ -292,6 +306,22 @@ __global__ void batch_norm_collect_statistics_kernel(
   stat_accscalar_t var_n = 0;
   int n = 0;
   for (int batch = threadIdx.y; batch < input.size(0); batch += blockDim.y) {
+#if defined(USE_ROCM)
+    constexpr int UNRL = 4;
+    stat_accscalar_t v_[UNRL];
+    for (int x = threadIdx.x; x < input.size(2); x += blockDim.x*UNRL) {
+      for (int u = 0; u < UNRL; u++)
+        v_[u] = input[batch][plane][min(x+u*blockDim.x, input.size(2)-1)];
+      for (int u = 0; u < UNRL; u++) {
+        if (x+u*blockDim.x < input.size(2)) {
+          stat_accscalar_t d1 = v_[u] - avg;
+          n++;
+          avg += d1 / n;
+          var_n += d1 * (v_[u] - avg);
+        }
+      }
+    }
+#else
     for (int x = threadIdx.x; x < input.size(2); x += blockDim.x) {
       stat_accscalar_t v = input[batch][plane][x];
       stat_accscalar_t d1 = v - avg;
@@ -299,6 +329,7 @@ __global__ void batch_norm_collect_statistics_kernel(
       avg += d1 / n;
       var_n += d1 * (v - avg);
     }
+#endif
   }
 
   // first warpSum to get one value per thread to

aten/src/ATen/native/cuda/Reduce.cuh

@@ -413,14 +413,12 @@ struct ReduceOp {
       value = thread_reduce<output_vec_size>(input_slice);
     }
 
-    if (config.should_block_y_reduce()) {
-      value = block_y_reduce<output_vec_size>(value, shared_memory);
-    }
-    __syncthreads();
     if (config.should_block_x_reduce()) {
       value = block_x_reduce<output_vec_size>(value, shared_memory);
     }
-
+    if (config.should_block_y_reduce()) {
+      value = block_y_reduce<output_vec_size>(value, shared_memory);
+    }
     using out_ptr_vec_t = std::array<out_scalar_t*, output_vec_size>;
     using offset_vec_t = std::array<index_t, output_vec_size>;
     offset_vec_t base_offsets;
@@ -657,8 +655,8 @@ struct ReduceOp {
     __syncthreads();
     // Intra-warp reduction, fix CUDA to have offset decreasing for better numerics
     // matching Triton, etc.
-    // todo for AMD
-    #ifdef USE_ROCM
+    // TODO(PaulZhang12): AMD and internal
+    #if defined(USE_ROCM) || defined(FBCODE_CAFFE2)
     for (int offset = 1; offset < dim_x; offset <<= 1) {
     #else
     for (int offset = dim_x >> 1; offset > 0; offset >>= 1) {

aten/src/ATen/native/cuda/ReflectionPad.cu

@@ -92,6 +92,16 @@ inline thrust::pair<int64_t, int64_t>  get_index_mapping2d(
     output_offset + output_y * output_dim_x + output_x);
 }
 
+__device__ __forceinline__ int64_t reflect_index(int64_t x, int64_t len) {
+  const int64_t two = (len - 1) * 2;
+  if (two <= 0) {
+    return 0;
+  }
+  int64_t m = x % two;
+  if (m < 0) m += two;
+  return (m < len) ? m : (two - m);
+}
+
 template<typename scalar_t>
 __global__ void reflection_pad1d_out_kernel(
     const scalar_t * input, scalar_t * output,
@@ -106,6 +116,28 @@ __global__ void reflection_pad1d_out_kernel(
   }
 }
 
+template <typename scalar_t>
+__global__ void reflection_pad1d_flat(
+    const scalar_t* __restrict__ input,
+    scalar_t* __restrict__ output,
+    int64_t input_w, int64_t pad_l, int64_t pad_r,
+    int64_t out_w, int64_t plane_count) {
+
+  const int64_t bx = blockDim.x;
+  const int64_t tx = threadIdx.x;
+
+  const int64_t total = plane_count * out_w;
+  const int64_t grid_stride = static_cast<int64_t>(bx) * gridDim.x;
+  int64_t linear = static_cast<int64_t>(blockIdx.x) * bx + tx;
+
+  for (; linear < total; linear += grid_stride) {
+    const int64_t plane = linear / out_w;
+    const int64_t x = linear - plane * out_w;
+    const int64_t j = reflect_index(x - pad_l, input_w);
+    output[plane * out_w + x] = input[plane * input_w + j];
+  }
+}
+
 template <typename scalar_t>
 __global__ void reflection_pad1d_backward_out_kernel(
     scalar_t * grad_input, const scalar_t * grad_output,
@@ -710,25 +742,44 @@ TORCH_IMPL_FUNC(reflection_pad1d_out_cuda)
   int64_t input_w = input_.size(dim_w);
   int64_t output_w = input_w + pad_l + pad_r;
 
-  dim3 block_size(output_w > 256 ? 256 : output_w);
-  dim3 grid_size((int)::ceil(output_w / 256.0), nplane, nbatch);
 
   Tensor input = input_.contiguous();
 
-  AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(
-      kHalf, kBFloat16, input.scalar_type(), "reflection_pad1d_out_template", [&] {
-        reflection_pad1d_out_kernel<<<
-            grid_size,
-            block_size,
-            0,
-            at::cuda::getCurrentCUDAStream()>>>(
-            input.const_data_ptr<scalar_t>(),
-            output.mutable_data_ptr<scalar_t>(),
-            input_w,
-            pad_l,
-            pad_r);
-        C10_CUDA_KERNEL_LAUNCH_CHECK();
-      });
+  const int block_x = static_cast<int>(std::min<int64_t>(256, std::max<int64_t>(1, output_w)));
+  const cudaDeviceProp* prop = at::cuda::getCurrentDeviceProperties();
+  const int max_x = prop->maxGridSize[0];
+  const int max_y = prop->maxGridSize[1];
+  const int max_z = prop->maxGridSize[2];
+
+  AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(kHalf, kBFloat16, input.scalar_type(), "reflection_pad1d_out", [&] {
+    auto stream = at::cuda::getCurrentCUDAStream();
+
+    const int64_t gx = at::ceil_div(output_w, static_cast<int64_t>(block_x));
+
+    const bool fits3d = (nplane <= max_y) && (nbatch <= max_z) && (gx <= max_x);
+
+    if (fits3d) {
+      dim3 block(block_x, 1, 1);
+      dim3 grid(gx, static_cast<unsigned>(nplane), static_cast<unsigned>(nbatch));
+      reflection_pad1d_out_kernel<scalar_t><<<grid, block, 0, stream>>>(
+          input.const_data_ptr<scalar_t>(),
+          output.mutable_data_ptr<scalar_t>(),
+          input_w, pad_l, pad_r);
+    } else {
+      dim3 block(block_x, 1, 1);
+      const int64_t plane_count = nplane * nbatch;
+      const int64_t total_blocks = at::ceil_div(plane_count * output_w, static_cast<int64_t>(block_x));
+      const int grid_x = static_cast<int>(std::min<int64_t>(max_x, std::max<int64_t>(1, total_blocks)));
+      dim3 grid(grid_x, 1, 1);
+
+      reflection_pad1d_flat<scalar_t><<<grid, block, 0, stream>>>(
+          input.const_data_ptr<scalar_t>(),
+          output.mutable_data_ptr<scalar_t>(),
+          input_w, pad_l, pad_r, output_w, plane_count);
+    }
+
+    C10_CUDA_KERNEL_LAUNCH_CHECK();
+  });
 }
 
 TORCH_IMPL_FUNC(reflection_pad1d_backward_out_cuda)(const Tensor& grad_output_,

aten/src/ATen/native/cuda/Sorting.cpp

@@ -43,6 +43,12 @@ std::tuple<Tensor&, Tensor&> kthvalue_out_impl_cuda(
   TORCH_CHECK(k >= 1 && k <= slicesize,
               "kthvalue(): selected number k out of range for dimension ", dim);
 
+  TORCH_CHECK(
+      slicesize <= std::numeric_limits<int32_t>::max(),
+      "kthvalue(): dimension ", dim, " is too large (", slicesize,
+      "). The current CUDA implementation supports dimension sizes up to ",
+      std::numeric_limits<int32_t>::max());
+
   at::assert_no_overlap(self, values);
 
   _reduction_with_indices_allocate_or_resize_output(
@@ -163,10 +169,6 @@ std::tuple<Tensor&, Tensor&> kthvalue_out_cuda(
     bool keepdim,
     Tensor& values,
     Tensor& indices) {
-  // See note [Writing Nondeterministic Operations]
-  // If there are duplicate elements of the kth value, the procedure for choosing which
-  // of the duplicates to use for the indices output is nondeterministic.
-  at::globalContext().alertNotDeterministic("kthvalue CUDA");
   auto result = [&]() {
     NoNamesGuard guard;
     // `kthvalue_out_impl_cuda` expects contiguous in input `self`.

aten/src/ATen/native/cuda/Sorting.cu

@@ -65,25 +65,34 @@ __global__ void gatherKthValue(
       &kValue);
 
   // Find the index of the k-th highest element
-  index_t kValueIndex = 0;
-  bool foundKValue = false;
+  __shared__ int32_t minIndexFound;
+
+  if (threadIdx.x == 0) {
+      minIndexFound = static_cast<int32_t>(inputSliceSize);
+  }
+  __syncthreads();
 
   for (index_t i = threadIdx.x; i < inputSliceSize; i += blockDim.x) {
-    bool inRange = (i < inputSliceSize);
-    scalar_t v = inRange ? doLdg(&inputSliceStart[i * inputWithinSliceStride])
-                         : static_cast<scalar_t>(0);
-    bool isKValue = inRange &&
-        ((v == kValue) || (at::_isnan(v) && at::_isnan(kValue)));
-    if (isKValue) {
-      kValueIndex = i;
-      foundKValue = true;
-      break;
-    }
+      // Early exit based on best-so-far
+      if (i >= minIndexFound) {
+          break;
+      }
+
+      scalar_t v = doLdg(&inputSliceStart[i * inputWithinSliceStride]);
+      bool isKValue =
+          ((v == kValue) || (at::_isnan(v) && at::_isnan(kValue)));
+
+      if (isKValue) {
+          atomicMin(&minIndexFound, static_cast<int32_t>(i));
+          break;
+      }
   }
 
-  if (foundKValue) {
-    kthValueSliceStart[0] = kValue;
-    indicesSliceStart[0] = kValueIndex;
+  __syncthreads();
+
+  if (threadIdx.x == 0) {
+      indicesSliceStart[0] = static_cast<index_t>(minIndexFound);
+      kthValueSliceStart[0] = kValue;
   }
 }
 

aten/src/ATen/native/cuda/TensorTopK.cpp

@@ -19,7 +19,6 @@
 
 namespace at::native {
 
-// TODO: remove this when CUDA <11.6 is no longer supported
 void topk_out_with_sort(
   const Tensor& self,
   int64_t k, int64_t dim, bool largest,
@@ -31,21 +30,12 @@ void topk_out_with_sort(
   indices.copy_(sorted_indices.narrow(dim, 0, k));
 }
 
-// TODO: remove this when CUDA <11.6 is no longer supported
-bool disable_sort_for_topk();
 bool should_use_sort(const Tensor& self, int64_t dim) {
 #if defined(USE_ROCM)
   if (self.dtype() == kBool) return false; // Bool sort not supported in ROCm: https://github.com/pytorch/pytorch/issues/139972
   return (self.numel() >= 10000 && self.numel() == self.size(dim)); // based on the experiments in https://github.com/pytorch/pytorch/pull/146387
 #else
-  if (disable_sort_for_topk()) return false;
-  // This heuristics is based on the experiment in https://github.com/pytorch/pytorch/pull/68632
-  if (self.dim() == 0) return false;
-  if (self.dtype() == kBool) return false; // Bool is not support by topk
-  int64_t slice_size = self.size(dim);
-  if (slice_size == 0) return false;
-  int64_t num_slices = self.numel() / slice_size;
-  return num_slices <= 10 && slice_size >= 100000;
+  return false;
 #endif
 }
 

aten/src/ATen/native/cuda/TensorTopK.cu

@@ -21,11 +21,6 @@ using namespace at::native;
 
 namespace at::native {
 
-// TODO: remove this when CUDA <11.6 is no longer supported
-bool disable_sort_for_topk() {
-  return CUB_SUPPORTS_SCAN_BY_KEY();
-}
-
 namespace sbtopk { // single_block_topk
 
 template <typename T>
@@ -418,10 +413,6 @@ __global__ void computeBlockwiseWithinKCounts(
   }
   __syncthreads();
 
-#if !CUB_SUPPORTS_SCAN_BY_KEY()
-  return;
-#endif
-
   Bitwise desired_digit = at::cuda::Bitfield<Bitwise>::getBitfield(desired, current_bit, RADIX_BITS);
 
   // if largest, then only threads that has tidx > desired_digit are active
@@ -477,7 +468,6 @@ __global__ void computeBlockwiseWithinKCounts(
   }
 }
 
-#if CUB_SUPPORTS_SCAN_BY_KEY()
 // Assumption: slice_size can not be larger than UINT32_MAX
 template <typename Bitwise>
 __global__ void computeBlockwiseKthCounts(
@@ -609,7 +599,6 @@ __global__ void gatherTopK(at::cuda::detail::TensorInfo<const T, IndexType> inpu
     }
   }
 }
-#endif
 
 int get_items_per_thread(uint64_t num_slices, uint64_t slice_size) {
   // occupancy of this kernel is limited by registers per threads
@@ -687,16 +676,12 @@ void launch(
   uint32_t* digit_cum_sum = reinterpret_cast<uint32_t*>(digit_cum_sum_buffer.get());
   AT_CUDA_CHECK(cudaMemsetAsync(digit_cum_sum, 0, numInputSlices * RADIX_DIGITS * sizeof(uint32_t), stream));
 
-#if CUB_SUPPORTS_SCAN_BY_KEY()
   auto withinKCounts_buffer = allocator.allocate(num_blocks * sizeof(uint32_t));
   uint32_t* withinKCounts = reinterpret_cast<uint32_t*>(withinKCounts_buffer.get());
   AT_CUDA_CHECK(cudaMemsetAsync(withinKCounts, 0, num_blocks * sizeof(uint32_t), stream));
 
   auto kthCounts_buffer = allocator.allocate(num_blocks * sizeof(uint32_t));
   uint32_t* kthCounts = reinterpret_cast<uint32_t*>(kthCounts_buffer.get());
-#else
-  uint32_t* withinKCounts = nullptr;
-#endif
 
   Bitwise desiredMask = 0;
   dim3 grid;
@@ -743,7 +728,6 @@ void launch(
   }
   desired = desired_in;
 
-#if CUB_SUPPORTS_SCAN_BY_KEY()
   computeBlockwiseKthCounts<Bitwise><<<std::min(((int64_t)numInputSlices + 255) / 256, (int64_t)1073741824), 256, 0, stream>>>(
     desired, counts, num_blocks, blocks_per_slice, kthCounts);
   C10_CUDA_KERNEL_LAUNCH_CHECK();
@@ -759,28 +743,6 @@ void launch(
     topK, topKWithinSliceStride, indices, indicesWithinSliceStride, items_per_thread,
     blocks_per_slice, kthValues, withinKCounts, kthCounts, num_blocks);
   C10_CUDA_KERNEL_LAUNCH_CHECK();
-#else
-  // Find topk values based on kth values
-  {
-    dim3 grid;
-    TORCH_INTERNAL_ASSERT(getGridFromTiles(numInputSlices, grid), "Too many slices for topk");
-    int warp_size = at::cuda::warp_size();
-    dim3 block(std::min(at::ceil_div((int64_t)inputSliceSize, (int64_t)warp_size) * (int64_t)warp_size, (int64_t)1024));
-    sbtopk::gatherTopK<T, IndexType, Dim, /* WithKthValues= */true><<<grid, block, 0, stream>>>(
-        input,
-        inputSliceSize,
-        outputSliceSize,
-        largest,
-        numInputSlices,
-        inputWithinSliceStride,
-        topK,
-        topKWithinSliceStride,
-        indices,
-        indicesWithinSliceStride,
-        kthValues);
-    C10_CUDA_KERNEL_LAUNCH_CHECK();
-  }
-#endif
 }
 
 } // namespace mbtopk
@@ -788,7 +750,6 @@ void launch(
 bool should_use_multiblock(int64_t num_slices, int64_t slice_size) {
   if (num_slices > std::numeric_limits<uint32_t>::max() ||
       slice_size > std::numeric_limits<uint32_t>::max()) return false;
-#if CUB_SUPPORTS_SCAN_BY_KEY()
   // This heuristics is based on the experiment in https://github.com/pytorch/pytorch/pull/74267
   return (num_slices <= 20 && slice_size >= 20000) ||
       (num_slices > 20 && num_slices <= 40 && slice_size >= 10000) ||
@@ -797,12 +758,6 @@ bool should_use_multiblock(int64_t num_slices, int64_t slice_size) {
       (num_slices >= 200 && num_slices < 800 && slice_size >= 3000) ||
       (num_slices >= 800 && num_slices <= 4000 && slice_size >= 800) ||
       (num_slices > 4000 && slice_size >= 400);
-#else
-  // This heuristics is based on the experiment in https://github.com/pytorch/pytorch/pull/71081
-  return (num_slices <= 400 && slice_size >= 5000) ||
-      (num_slices > 400 && num_slices < 4000 && slice_size >= 1000) ||
-      (num_slices >= 4000 && slice_size >= 300);
-#endif
 }
 
 void launch_gather_topk_kernel(

aten/src/ATen/native/cuda/TriangularOps.cu

@@ -44,7 +44,7 @@ __global__ void triu_tril_kernel(
     const int64_t k,
     const int64_t N_padded,
     const IndexType last_dim_padded) {
-  int64_t linear_idx = (blockIdx.x * blockDim.x + threadIdx.x) * elements_per_thread;
+  int64_t linear_idx = (((int64_t)blockIdx.x) * blockDim.x + threadIdx.x) * elements_per_thread;
   if (linear_idx >= N_padded) {
     return;
   }

aten/src/ATen/native/cuda/UpSampleBilinear2d.cu

@@ -145,8 +145,8 @@ __device__ __forceinline__ void compute_output_range(
       lo = (input_pos - static_cast<accscalar_t>(0.5)) / scale - static_cast<accscalar_t>(0.5);
       hi = (input_pos + static_cast<accscalar_t>(1.5)) / scale - static_cast<accscalar_t>(0.5);
   }
-  min_output = max(0, static_cast<int>(ceil(lo)));
-  max_output = min(output_size - 1, static_cast<int>(floor(hi)));
+  min_output = max(0, static_cast<int>(std::ceil(lo)));
+  max_output = min(output_size - 1, static_cast<int>(std::floor(hi)));
 }
 #endif
 

aten/src/ATen/native/cuda/fused_adagrad_utils.cuh

@@ -52,7 +52,7 @@ struct FusedAdagradMathFunctor {
   using opmath_t = at::opmath_type<scalar_t>;
 
   C10_DEVICE __forceinline__ void operator()(
-      int chunk_size,
+      int64_t chunk_size,
       FusedOptimizerTensorListMetadata<3>& tl,
       const float* lr_ptr,
       const double& lr,
@@ -133,4 +133,4 @@ struct FusedAdagradMathFunctor {
 
 } // namespace
 
-} // namespace at::native
+} // namespace at::native

aten/src/ATen/native/cuda/reduction_template.cuh

@@ -466,7 +466,7 @@ struct ReduceJitOp {
 
     __syncthreads();
 
-    #ifdef USE_ROCM
+    #if defined(USE_ROCM) || defined(FBCODE_CAFFE2)
     for (int offset = 1; offset < dim_x; offset <<= 1) {
     #else
     for (int offset = dim_x >> 1; offset > 0; offset >>= 1) {

aten/src/ATen/native/cudnn/MHA.cpp

@@ -487,9 +487,7 @@ std::unique_ptr<fe::graph::Graph> build_graph(
   auto scaled_dot_product_flash_attention_options =
       fe::graph::SDPA_attributes()
           .set_name("CUDNN_SDPA")
-          .set_is_inference(return_softmaxstats == false)
-          // TODO(eqy): switch to this API once cuDNN FE is upgraded
-          // .set_generate_stats(return_softmaxstats)
+          .set_generate_stats(return_softmaxstats)
           .set_causal_mask(is_causal)
           .set_attn_scale(attn_scale);
   if (use_ragged_in_dense(q, k, v, o, attn_bias.has_value())) {
@@ -707,9 +705,7 @@ std::unique_ptr<fe::graph::Graph> build_graph_nestedtensor(
   auto scaled_dot_product_flash_attention_options =
       fe::graph::SDPA_attributes()
           .set_name("CUDNN_SDPA_NESTEDTENSOR")
-          .set_is_inference(return_softmaxstats == false)
-          // TODO(eqy): switch to this API once cuDNN FE is upgraded
-          // .set_generate_stats(return_softmaxstats)
+          .set_generate_stats(return_softmaxstats)
           .set_causal_mask(is_causal)
           .set_attn_scale(attn_scale)
           .set_seq_len_q(SEQ_LEN_Q_)

aten/src/ATen/native/mkldnn/xpu/Blas.cpp

@@ -2,6 +2,7 @@
 #include <ATen/WrapDimUtilsMulti.h>
 #include <ATen/native/Resize.h>
 #include <ATen/native/mkldnn/xpu/detail/oneDNN.h>
+#include <ATen/native/xpu/Blas.h>
 #include <torch/library.h>
 #ifndef AT_PER_OPERATOR_HEADERS
 
@@ -50,9 +51,13 @@ Tensor& addmm_out(
       mat1.dtype(),
       " != ",
       mat2.dtype())
+
   // complex case
-  TORCH_CHECK(
-      !mat1.is_complex(), "Complex datatype matmul is not supported in oneDNN");
+  if (self.is_complex()) {
+    at::native::addmm_complex_out_xpu(self, mat1, mat2, beta, alpha, result);
+
+    return result;
+  }
 
   std::vector<int64_t> result_shape = {mat1.size(0), mat2.size(1)};
   result.resize_(result_shape);
@@ -167,8 +172,11 @@ Tensor& mm_out(const Tensor& self, const Tensor& mat2, Tensor& result) {
     return result;
   }
 
-  TORCH_CHECK(
-      !self.is_complex(), "Complex datatype matmul is not supported in oneDNN");
+  if (self.is_complex()) {
+    at::native::mm_complex_out_xpu(self, mat2, result);
+
+    return result;
+  }
 
   onednn::matmul(result, self, mat2, Tensor(), true, onednn::Attr());
   return result;
@@ -208,9 +216,12 @@ Tensor& baddbmm_out(
       input.sizes());
 
   // complex case
-  TORCH_CHECK(
-      !batch1.is_complex(),
-      "Complex datatype matmul is not supported in oneDNN");
+  if (input.is_complex()) {
+    at::native::baddbmm_complex_out_xpu(
+        input, batch1, batch2, beta, alpha, result);
+
+    return result;
+  }
 
   // general case
   onednn::Attr attr;
@@ -257,8 +268,13 @@ Tensor& bmm_out(const Tensor& self, const Tensor& batch2, Tensor& result) {
     return result;
   }
 
-  TORCH_CHECK(
-      !self.is_complex(), "Complex datatype matmul is not supported in oneDNN");
+  // complex case
+  if (self.is_complex()) {
+    at::native::bmm_complex_out_xpu(self, batch2, result);
+
+    return result;
+  }
+
   onednn::matmul(result, self, batch2, at::Tensor(), true, onednn::Attr());
   return result;
 }

aten/src/ATen/native/mps/kernels/BinaryKernel.metal

@@ -222,6 +222,13 @@ struct nextafter_functor {
   }
 };
 
+struct hypot_functor {
+  template <typename T>
+  inline T operator()(const T a, const T b) {
+    return static_cast<T>(precise::sqrt(float(a) * a + float(b) * b));
+  }
+};
+
 // Complex binary functors
 struct polar_functor {
   template <typename U>
@@ -362,6 +369,7 @@ struct igammac_functor {
   REGISTER_OPMATH_BINARY_OP(NAME, half, half);   \
   REGISTER_OPMATH_BINARY_OP(NAME, bfloat, bfloat)
 
+REGISTER_FLOAT_BINARY_OP(hypot);
 REGISTER_FLOAT_BINARY_OP(copysign);
 REGISTER_INT2FLOAT_BINARY_OP(copysign);
 REGISTER_FLOAT_BINARY_OP(fmax);

aten/src/ATen/native/mps/kernels/LinearAlgebra.h

@@ -0,0 +1,16 @@
+#pragma onces
+#include <c10/metal/common.h>
+
+template <unsigned N = c10::metal::max_ndim>
+struct OrgqrParams {
+  int32_t num_batch_dims;
+
+  uint32_t m;
+  uint32_t n;
+  uint32_t k;
+
+  ::c10::metal::array<uint32_t, N> A_strides;
+  ::c10::metal::array<uint32_t, N> tau_strides;
+  ::c10::metal::array<uint32_t, N> H_strides;
+  ::c10::metal::array<uint32_t, N> H_sizes;
+};

aten/src/ATen/native/mps/kernels/LinearAlgebra.metal

@@ -1,3 +1,4 @@
+#include <ATen/native/mps/kernels/LinearAlgebra.h>
 #include <c10/metal/utils.h>
 #include <metal_array>
 #include <metal_simdgroup>
@@ -441,7 +442,7 @@ kernel void applySYRK(
     uint3 tid [[thread_position_in_threadgroup]],
     uint3 tgid [[threadgroup_position_in_grid]],
     uint3 tpg [[threads_per_threadgroup]],
-    uint sgitg [[simdgroup_index_in_threadgroup]]) {
+    uint warp_id [[simdgroup_index_in_threadgroup]]) {
   const uint tx = tid.x;
   const uint ty = tid.y;
   const uint simdGroupsPerThreadgroup = (tpg.x * tpg.y + 31) / 32;
@@ -474,11 +475,8 @@ kernel void applySYRK(
       (actSize_j % 8 == 0) && (actSize_h % 8 == 0) && (actSize_k % 8 == 0);
 
   if (use_simdgroup) {
-    uint warp_id = sgitg;
-
     simdgroup_matrix<float, 8, 8> negative_identity =
         simdgroup_matrix<float, 8, 8>(-1.0);
-    simdgroup_matrix<float, 8, 8> identity = simdgroup_matrix<float, 8, 8>(1.0);
     simdgroup_matrix<float, 8, 8> Prod;
     simdgroup_matrix<float, 8, 8> Afrag;
     simdgroup_matrix<float, 8, 8> Bfrag;
@@ -521,8 +519,7 @@ kernel void applySYRK(
             /* transpose = */ upper);
 
         simdgroup_multiply(Prod, Afrag, Bfrag);
-        simdgroup_multiply(Prod, Prod, negative_identity);
-        simdgroup_multiply_accumulate(Cfrag, Cfrag, identity, Prod);
+        simdgroup_multiply_accumulate(Cfrag, Prod, negative_identity, Cfrag);
       }
 
       simdgroup_store(
@@ -644,6 +641,164 @@ kernel void applyPivots(
   }
 }
 
+template <typename T>
+static T bool_to_float(bool b) {
+  return static_cast<T>(b);
+}
+
+template <>
+half2 bool_to_float(bool b) {
+  return half2(b ? 1 : 0, 0);
+}
+
+template <>
+float2 bool_to_float(bool b) {
+  return float2(b ? 1 : 0, 0);
+}
+
+template <typename T>
+static T calc_H_irc(
+    device T* A,
+    uint32_t A_stride_r,
+    uint32_t A_stride_c,
+    constant T* tau,
+    uint32_t tau_stride,
+    uint32_t r,
+    uint32_t c,
+    uint32_t i) {
+  T I_val = bool_to_float<T>(r == c);
+  T tau_val = tau[i * tau_stride];
+
+  T A_ci = c10::metal::conj(A[c * A_stride_r + i * A_stride_c]);
+  T A_ri = A[r * A_stride_r + i * A_stride_c];
+
+  T c_eq_i = bool_to_float<T>(c == i);
+  T r_eq_i = bool_to_float<T>(r == i);
+
+  T A_ci_ = (c > i) ? A_ci : c_eq_i;
+  T A_ri_ = (r > i) ? A_ri : r_eq_i;
+
+  return I_val - c10::metal::mul(tau_val, c10::metal::mul(A_ci_, A_ri_));
+}
+
+// Calculate (A @ B)[r, c], the element in the r-th row and c-th column of the
+// result of matrix multiplying A and B together. A and B must be size m-by-m
+// and have the same strides. The formula for this operation, written in Python
+// syntax, is:
+//   (A @ B)[r, c] = A[r, :].dot(B[:, c])
+template <typename T>
+static T calc_matmul_rc(
+    device T* A,
+    device T* B,
+    uint32_t stride_r,
+    uint32_t stride_c,
+    uint32_t m,
+    uint32_t r,
+    uint32_t c) {
+  T AB_rc = 0;
+  auto A_row_offset = r * stride_r;
+  auto B_col_offset = c * stride_c;
+
+  uint32_t A_col_offset = 0;
+  uint32_t B_row_offset = 0;
+
+  for (uint32_t j = 0; j < m;
+       j++, A_col_offset += stride_c, B_row_offset += stride_r) {
+    AB_rc += c10::metal::mul(
+        A[A_row_offset + A_col_offset], B[B_row_offset + B_col_offset]);
+  }
+  return AB_rc;
+}
+
+template <typename T>
+kernel void orgqr(
+    device T* A [[buffer(0)]],
+    constant T* tau [[buffer(1)]],
+    device T* H [[buffer(2)]],
+    device T* H_prod [[buffer(3)]],
+    constant OrgqrParams<>& params [[buffer(4)]],
+    uint tid [[thread_position_in_grid]]) {
+  constant auto& A_strides = params.A_strides;
+  constant auto& tau_strides = params.tau_strides;
+  constant auto& H_strides = params.H_strides;
+  constant auto& H_sizes = params.H_sizes;
+
+  auto num_batch_dims = params.num_batch_dims;
+  auto m = params.m;
+  auto n = params.n;
+  auto k = params.k;
+
+  auto m2 = m * m;
+  auto batch_idx = tid / m2;
+
+  // Find the matrices for this thread's batch index
+  uint32_t A_offset = 0;
+  uint32_t tau_offset = 0;
+  uint32_t H_offset = 0;
+
+  for (auto dim = num_batch_dims - 1; dim >= 0; dim--) {
+    auto dim_size = H_sizes[dim];
+    auto dim_idx = batch_idx % dim_size;
+
+    A_offset += dim_idx * A_strides[dim];
+    tau_offset += dim_idx * tau_strides[dim];
+    H_offset += dim_idx * H_strides[dim];
+
+    batch_idx /= dim_size;
+  }
+
+  A += A_offset;
+  tau += tau_offset;
+  H += H_offset;
+  H_prod += H_offset;
+
+  auto matrix_idx = tid % m2;
+  auto r = matrix_idx / m;
+  auto c = matrix_idx % m;
+  auto A_stride_r = A_strides[num_batch_dims];
+  auto A_stride_c = A_strides[num_batch_dims + 1];
+  auto tau_stride = tau_strides[num_batch_dims];
+  auto H_stride_r = H_strides[num_batch_dims];
+  auto H_stride_c = H_strides[num_batch_dims + 1];
+
+  // Find the element of H and H_prod that this thread will calculate
+  device T* H_elem_ptr = H + (r * H_stride_r + c * H_stride_c);
+  device T* H_prod_elem_ptr = H_prod + (r * H_stride_r + c * H_stride_c);
+
+  for (uint32_t i = 0; i < k; i++) {
+    // Calculate and write H_i
+
+    T H_irc = calc_H_irc(A, A_stride_r, A_stride_c, tau, tau_stride, r, c, i);
+
+    // Calculate element [r, c] of prod(H_0, ..., H_i)
+    if (i == 0) {
+      *H_prod_elem_ptr = H_irc;
+    } else {
+      *H_elem_ptr = H_irc;
+
+      // Need this sync because the below matmul requires all threads to finish
+      // writing their entries to `H_prod` and `H`.
+      threadgroup_barrier(mem_flags::mem_threadgroup);
+
+      T H_prod_0_to_i_rc =
+          calc_matmul_rc(H_prod, H, H_stride_r, H_stride_c, m, r, c);
+
+      // Need this sync because the above matmul uses the current values in
+      // `H_prod`, and we don't want to overwrite those until all threads are
+      // finished using them.
+      threadgroup_barrier(mem_flags::mem_threadgroup);
+
+      *H_prod_elem_ptr = H_prod_0_to_i_rc;
+    }
+  }
+
+  device T* A_elem_ptr = A + (r * A_stride_r + c * A_stride_c);
+
+  if (c < n) {
+    *A_elem_ptr = *H_prod_elem_ptr;
+  }
+}
+
 #define INSTANTIATE_MM_OPS(DTYPE)                                           \
   template [[host_name("matmul_" #DTYPE)]] kernel void matmul<DTYPE>(       \
       constant DTYPE * mat1Data [[buffer(0)]],                              \
@@ -683,3 +838,19 @@ INSTANTIATE_MM_OPS(int);
 INSTANTIATE_MM_OPS(short);
 INSTANTIATE_MM_OPS(char);
 INSTANTIATE_MM_OPS(uchar);
+
+#define REGISTER_ORGQR(T)                            \
+  template [[host_name("orgqr_" #T)]]                \
+  kernel void orgqr<T>(                              \
+      device T * A [[buffer(0)]],                    \
+      constant T * tau [[buffer(1)]],                \
+      device T * H [[buffer(2)]],                    \
+      device T * H_prod [[buffer(3)]],               \
+      constant OrgqrParams<> & params [[buffer(4)]], \
+      uint tid [[thread_position_in_grid]]);
+
+REGISTER_ORGQR(float);
+REGISTER_ORGQR(half);
+REGISTER_ORGQR(bfloat);
+REGISTER_ORGQR(float2);
+REGISTER_ORGQR(half2);

aten/src/ATen/native/mps/kernels/UnaryKernel.metal

@@ -5,6 +5,21 @@
 using namespace metal;
 using namespace c10::metal;
 
+struct angle_functor {
+  template <typename T, enable_if_t<is_complex_v<T>, bool> = true>
+  inline T operator()(const T x) {
+    return T(atan2(x.y, x.x), 0);
+  }
+  template <typename T, enable_if_t<is_scalar_floating_point_v<T>, bool> = true>
+  inline T operator()(const T x) {
+    return T(isnan(x) ? x : x < 0 ? M_PI_F : 0.0);
+  }
+  template <typename T, enable_if_t<is_scalar_integral_v<T>, bool> = true>
+  inline float operator()(const T x) {
+    return x < 0 ? M_PI_F : 0.0;
+  }
+};
+
 // Implement exp wrapper for both real and complex types
 template <typename T, enable_if_t<is_scalar_floating_point_v<T>, bool> = true>
 inline T exp_(const T x) {
@@ -545,6 +560,7 @@ REGISTER_UNARY_OP(abs, float, float);
 REGISTER_UNARY_OP(abs, half, half);
 
 #define INSTANTIATE_UNARY_KERNELS2(DTYPE0, DTYPE1) \
+  REGISTER_UNARY_OP(angle, DTYPE1, DTYPE0);        \
   REGISTER_UNARY_OP(erf, DTYPE1, DTYPE0);          \
   REGISTER_UNARY_OP(erfc, DTYPE1, DTYPE0);         \
   REGISTER_UNARY_OP(erfinv, DTYPE1, DTYPE0);       \
@@ -583,6 +599,7 @@ INSTANTIATE_UNARY_KERNELS2(float, int);
 INSTANTIATE_UNARY_KERNELS2(float, long);
 
 #define INSTANTIATE_UNARY_KERNELS_VEC2(DTYPE)     \
+  REGISTER_UNARY_OP(angle, DTYPE##2, DTYPE##2);   \
   REGISTER_UNARY_OP(neg, DTYPE##2, DTYPE##2);     \
   REGISTER_UNARY_OP(exp, DTYPE##2, DTYPE##2);     \
   REGISTER_UNARY_OP(expm1, DTYPE##2, DTYPE##2);   \

aten/src/ATen/native/mps/operations/Attention.mm

@@ -92,13 +92,8 @@ static std::tuple<Tensor, Tensor> sdpa_general_mps(const Tensor& query,
           }
 
           // upcasting to float32 if needed to improve precision when multiplying by the scale factor
-          if ([maskedMM dataType] != MPSDataTypeFloat32) {
-            maskedMM = [mpsGraph castTensor:maskedMM toType:MPSDataTypeFloat32 name:nil];
-          }
+          maskedMM = castMPSTensor(mpsGraph, maskedMM, MPSDataTypeFloat32);
           maskedMM = [mpsGraph multiplicationWithPrimaryTensor:maskedMM secondaryTensor:scaleTensor name:nil];
-          if ([maskedMM dataType] != qTensor.dataType) {
-            maskedMM = [mpsGraph castTensor:maskedMM toType:qTensor.dataType name:nil];
-          }
 
           if (is_causal) {
             auto causalMask = [mpsGraph constantWithScalar:1.0f
@@ -112,7 +107,9 @@ static std::tuple<Tensor, Tensor> sdpa_general_mps(const Tensor& query,
                                                       name:nil];
           } else if (attn_mask) {
             graph->maskTensor = mpsGraphRankedPlaceHolder(mpsGraph, *attn_mask);
-            maskedMM = [mpsGraph additionWithPrimaryTensor:maskedMM secondaryTensor:graph->maskTensor name:nil];
+            maskedMM = [mpsGraph additionWithPrimaryTensor:maskedMM
+                                           secondaryTensor:castMPSTensor(mpsGraph, graph->maskTensor, maskedMM.dataType)
+                                                      name:nil];
           }
 
           // Account for case where all values were masked causing division by 0 in softmax (issue:#156707)
@@ -133,8 +130,8 @@ static std::tuple<Tensor, Tensor> sdpa_general_mps(const Tensor& query,
           graph->qTensor = qTensor;
           graph->kTensor = kTensor;
           graph->vTensor = vTensor;
-          graph->outputTensor = output;
-          graph->attnTensor = sm;
+          graph->outputTensor = castMPSTensor(mpsGraph, output, qTensor.dataType);
+          graph->attnTensor = castMPSTensor(mpsGraph, sm, qTensor.dataType);
         });
     auto qPlaceholder = Placeholder(cachedGraph->qTensor, query);
     auto kPlaceholder = Placeholder(cachedGraph->kTensor, key);