Update on "[FSDP][Replicate] final version integrating 1D device mesh replicate into fsdp"

**Summary:** I have created a new composable replicate api that's integrated into FSDP's codebase with minimal changes. The key changes I made are when we use DDPMeshInfo, we use Replicate placements, prevent initial sharding of parameters, set worldsize to 1 to skip allgathers and reducescatter. 

**Test Cases**
1. pytest test/distributed/_composable/test_replicate_training.py
2. pytest test_pp_composability.py
3. pytest test_replicate_with_fsdp.py




cc H-Huang awgu wanchaol fegin fduwjj wz337 wconstab d4l3k pragupta msaroufim dcci

[ghstack-poisoned]

.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/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:

.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.

.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/vision.txt

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

.github/ci_commit_pins/xla.txt

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

.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/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/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

@@ -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

.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

.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',

SECURITY.md

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

aten/src/ATen/CMakeLists.txt

@@ -38,7 +38,7 @@ set_bool(AT_HIPSPARSELT_ENABLED CAFFE2_USE_HIPSPARSELT)
 
 configure_file(Config.h.in "${CMAKE_CURRENT_SOURCE_DIR}/Config.h")
 # TODO: Do not generate CUDAConfig.h for ROCm BUILDS
-# At the moment, `jit_macors.h` include CUDAConfig.h for both CUDA and HIP builds
+# At the moment, `jit_macros.h` include CUDAConfig.h for both CUDA and HIP builds
 if(USE_CUDA OR USE_ROCM)
   configure_file(cuda/CUDAConfig.h.in "${CMAKE_CURRENT_SOURCE_DIR}/cuda/CUDAConfig.h")
 endif()
@@ -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
     )

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/FunctionalTensorWrapper.cpp

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

aten/src/ATen/TensorIterator.cpp

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

aten/src/ATen/core/CachingHostAllocator.h

@@ -94,11 +94,11 @@ struct PinnedReserveSegment {
 struct TORCH_API HostStats {
   // COUNT: total allocations (active)
   Stat active_requests;
-  // SUM: bytes allocated/reserved by this memory alocator. (active)
+  // SUM: bytes allocated/reserved by this memory allocator. (active)
   Stat active_bytes;
   // COUNT: total allocations (active + free)
   Stat allocations;
-  // SUM: bytes allocated/reserved by this memory alocator. This accounts
+  // SUM: bytes allocated/reserved by this memory allocator. This accounts
   // for both free and in-use blocks.
   Stat allocated_bytes;
 
@@ -127,7 +127,7 @@ struct alignas(hardware_destructive_interference_size) HostStatsStaged {
   // COUNT: total allocations (active + free)
   // LOCK: access to this stat is protected by the allocator's blocks_mutex_
   Stat allocations;
-  // SUM: bytes allocated/reserved by this memory alocator. This accounts
+  // SUM: bytes allocated/reserved by this memory allocator. This accounts
   // for both free and in-use blocks.
   Stat allocated_bytes;
   // COUNT: number of allocations per bucket (active)
@@ -455,7 +455,7 @@ struct CachingHostAllocatorImpl {
   }
 
   void resetAccumulatedStats() {
-    // Reseting accumulated memory stats requires concurrently holding both the
+    // Resetting accumulated memory stats requires concurrently holding both the
     // free list mutexes and the blocks mutex. Previously, this was only done in
     // empty_cache function.
     for (size_t i = 0; i < free_list_.size(); ++i) {
@@ -482,7 +482,7 @@ struct CachingHostAllocatorImpl {
   }
 
   void resetPeakStats() {
-    // Reseting peak memory stats requires concurrently holding both the
+    // Resetting peak memory stats requires concurrently holding both the
     // free list mutexes and the blocks mutex. Previously, this was only done in
     // empty_cache function.
     for (size_t i = 0; i < free_list_.size(); ++i) {

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/class_type.h

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

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

@@ -442,11 +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));
@@ -531,7 +537,7 @@ int64_t Dispatcher::sequenceNumberForRunningRecordFunction(DispatchKey dispatchK
 
   // Note: this records a sequence number for both Autograd keys, and for
   // non-Autograd keys where the dispatchKeySet still contains an autograd key.
-  // This means that we might collect the same sequence nubmer two different
+  // This means that we might collect the same sequence number two different
   // events if they all occurred above Autograd and still had the Autograd
   // dispatch key in the dispatch key set.
   // However, this usually doesn't happen: normally the first call will

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

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

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

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

aten/src/ATen/core/function_schema.cpp

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

aten/src/ATen/core/interned_strings.cpp

@@ -68,11 +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/ivalue.h

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

aten/src/ATen/core/ivalue_inl.h

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

aten/src/ATen/core/jit_type.h

@@ -8,6 +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:
@@ -374,7 +373,7 @@ struct TORCH_API SymbolicShape {
   // Unranked shape constructor.
   SymbolicShape() : dims_(std::nullopt) {}
 
-  // Known rank but unknown dimentions.
+  // Known rank but unknown dimensions.
   SymbolicShape(std::optional<size_t> rank) : dims_(std::nullopt) {
     if(!rank) {
       return;
@@ -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);
   }
 
@@ -891,9 +884,9 @@ struct TORCH_API ListType
 
   // global singleton
   // Given an inner type T and an identifier,
-  // this function wil return the global singleton type pointer
+  // this function will return the global singleton type pointer
   // the type List<T>.
-  // The extra "identifier" argument is needed beccause we have multiple container types
+  // The extra "identifier" argument is needed because we have multiple container types
   // that all re-use this function (List<T>, array<T, N>, etc.)
   static TypePtr get(const std::string& identifier, TypePtr inner);
 
@@ -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/op_registration/op_registration.h

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

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

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

aten/src/ATen/core/tensor_type.cpp

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

aten/src/ATen/core/type.cpp

@@ -8,6 +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

@@ -6,6 +6,7 @@
 #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>

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

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

@@ -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,
@@ -632,8 +634,7 @@ inline Vectorized<float> Vectorized<float>::erf() const {
   // - exp(- x * x)
   auto pow_2 = (*this) * (*this);
   auto neg_pow_2 = pow_2 ^ neg_zero_vec;
-  auto tmp4 = neg_pow_2.map(
-      std::exp); // This can be swapped for a faster implementation of exp.
+  auto tmp4 = neg_pow_2.exp();
   auto tmp5 = tmp4 ^ neg_zero_vec;
   // erf(x) = sign(x) * (1 - r * t * exp(- x * x))
   auto tmp6 = t * tmp5;

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

@@ -234,7 +234,7 @@ class Vectorized<c10::Half> : public Vectorized16<
         vshlq_u16(vandq_u16(is_zero_vec, vdupq_n_u16(1)), shift);
     return vaddvq_u16(bits_vec);
 #else // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
-    // use known working implmentation.
+    // use known working implementation.
     __at_align__ value_type tmp[size()];
     store(tmp);
     int mask = 0;
@@ -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/vec256/vec256_int.h

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

@@ -634,7 +634,7 @@ struct Vectorized {
   }
   Vectorized<T> neg() const {
     // NB: the trailing return type is needed because we need to coerce the
-    // return value back to T in the case of unary operator- incuring a
+    // return value back to T in the case of unary operator- incurring a
     // promotion
     return map([](T x) -> T { return -x; });
   }

aten/src/ATen/cuda/CUDABlas.cpp

@@ -1958,7 +1958,7 @@ void scaled_gemm(
     ScalarType result_dtype,
     bool use_fast_accum,
     const std::optional<Tensor>& alpha) {
-  // Note: see `cublasCommonArgs` for various non-intuitive manupulations
+  // Note: see `cublasCommonArgs` for various non-intuitive manipulations
   // of input arguments to this function.
   const auto computeType = CUBLAS_COMPUTE_32F;
   const auto scaleType = CUDA_R_32F;

aten/src/ATen/cuda/CUDAGraph.cpp

@@ -168,11 +168,9 @@ void CUDAGraph::instantiate() {
   // https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__GRAPH.html#group__CUDART__GRAPH_1g1accfe1da0c605a577c22d9751a09597
   // cudaGraphInstantiateWithFlags
   // https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__GRAPH.html#group__CUDART__GRAPH_1ga2c652a24ba93e52b99a47bec0888233
-#if !defined(USE_ROCM) || ROCM_VERSION >= 60200
   int version = 0;
   AT_CUDA_CHECK(cudaDriverGetVersion(&version));
   if (version < 11040) {
-#endif
     // Trailing NULL, NULL, 0 arguments were recommended by Cuda driver people,
     // who prefer not to report error message through these arguments moving forward
     // (they prefer return value, or errors on api calls internal to the capture)
@@ -183,13 +181,11 @@ void CUDAGraph::instantiate() {
 #endif
 //Since ROCm 6.2, we want to go down this path as hipGraphExecDestroy in the destructor will not immediately free the memory.
 //It will wait for the next sync operation. cudaGraphInstantiateFlagAutoFreeOnLaunch will add async frees after graph launch.
-#if !defined(USE_ROCM) || ROCM_VERSION >= 60200
   } else {
     AT_CUDA_CHECK(cudaGraphInstantiateWithFlags(&graph_exec_,
                                                 graph_,
                                                 cudaGraphInstantiateFlagAutoFreeOnLaunch));
   }
-#endif
   has_graph_exec_ = true;
 }
 
@@ -311,7 +307,7 @@ CUDAGraph::~CUDAGraph() {
 // There are recent HIP changes where hipGraphExecDestroy doesn't immediately free memory.
 // They wait for next sync point in order to free the memory, this is to ensure that all
 // hipGraphLaunch are finished before we release any memory. This feature was enabled in rocm6.2.
-// We need to ensure all async opreations finish before deleting the object.
+// We need to ensure all async operations finish before deleting the object.
 #if (defined(USE_ROCM) && ROCM_VERSION >= 60200)
   if (capture_dev_ != UNDEFINED_DEVICE) // check if capture_dev_ contains the real device id
   {

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

@@ -137,7 +137,7 @@ struct CUDACachingHostAllocatorImpl
   void free_block_slowpath(Block* block) {
     auto start = std::chrono::steady_clock::now();
     // Users may change the allocator config at will. torch unit tests do this.
-    // However, allocations using cudaHostRegister should use corresonding
+    // However, allocations using cudaHostRegister should use corresponding
     // cudaHostUnregister and similarly for cudaHostAlloc / cudaFreeHost.
     void* ptr = block->ptr_;
     bool use_register = false;

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> {
@@ -203,36 +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
@@ -277,25 +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,
@@ -303,7 +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
 }
@@ -555,16 +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,
@@ -572,7 +504,6 @@ 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
 }

aten/src/ATen/cuda/cub.h

@@ -4,7 +4,7 @@
 #include <ATen/cuda/CUDAConfig.h>
 
 // NOTE: These templates are intentionally not defined in this header,
-// which aviods re-compiling them for each translation unit. If you get
+// which avoids re-compiling them for each translation unit. If you get
 // a link error, you need to add an explicit instantiation for your
 // types in cub.cu
 

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
@@ -28,14 +20,6 @@
 #define USE_GLOBAL_CUB_WRAPPED_NAMESPACE() false
 #endif
 
-// cub support for cub::FutureValue is added to cub 1.15 in:
-// https://github.com/NVIDIA/cub/pull/305
-#if CUB_VERSION >= 101500
-#define CUB_SUPPORTS_FUTURE_VALUE() true
-#else
-#define CUB_SUPPORTS_FUTURE_VALUE() false
-#endif
-
 // There were many bc-breaking changes in major version release of CCCL v3.0.0
 // Please see https://nvidia.github.io/cccl/cccl/3.0_migration_guide.html
 #if CUB_VERSION >= 200800

aten/src/ATen/cuda/tunable/README.md

@@ -38,7 +38,7 @@ GemmTunableOp_float_NT,nt_25088_4096_64,1219,1.262
 GemmTunableOp_float_NT,nt_4096_4096_64,1216,0.033
 ```
 
-Note the "Validator" lines. If you change a library verison, or ROCm version, or PyTorch version, TunableOp will detect
+Note the "Validator" lines. If you change a library version, or ROCm version, or PyTorch version, TunableOp will detect
 this and reject the tunings file because the prior tunings are likely affected by other software changes.
 
 The remaining lines are the tuned solutions for each TunableOp encountered during your execution. Each line consists of

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

@@ -235,7 +235,7 @@ class TunableOp {
       // numeric check option is controlled by non-static env var, so check it once per tuned operator
       bool do_numerics_check = ctx->IsNumericsCheckEnabled();
 
-      // calcaulte a reference answer for numerical check
+      // calculate a reference answer for numerical check
       if (do_numerics_check) {
         reference_params = params->DeepCopy(false);
         TORCH_CHECK(ops_[ResultEntry::Default()]->Call(reference_params) == OK);

aten/src/ATen/detail/AcceleratorHooksInterface.h

@@ -12,7 +12,7 @@ namespace at {
 
 // AcceleratorHooksInterface is a shared interface provided by all
 // accelerators to allow generic code.
-// This inferface is hook-based as it corresponds to all the functions
+// This interface is hook-based as it corresponds to all the functions
 // that are going to be called in a generic way from the CPU code.
 
 struct TORCH_API AcceleratorHooksInterface {

aten/src/ATen/detail/PrivateUse1HooksInterface.h

@@ -38,7 +38,7 @@ struct TORCH_API PrivateUse1HooksInterface : AcceleratorHooksInterface {
 
   Generator getNewGenerator(
       [[maybe_unused]] DeviceIndex device_index = -1) const override {
-    // TODO(FFFrog): Perserved for BC and will be removed in the future.
+    // TODO(FFFrog): Preserved for BC and will be removed in the future.
     if (at::GetGeneratorPrivate().has_value())
       return at::GetGeneratorForPrivateuse1(device_index);
 

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/BatchRulesHelper.h

@@ -283,7 +283,7 @@ inline void boxed_existing_bdim_all_batch_rule(
 // Use when all tensors arguments accept one (normal) batch dim.
 // This batching rule expands the batch dim on all Tensors, reshapes it into
 // dim 0, calls the op, and then reshapes the batch dim out of dim 0.
-// This is not the most efficient thing; if there are alternatives, plese try
+// This is not the most efficient thing; if there are alternatives, please try
 // to use them. Use this only as a last resort.
 #define EXISTING_BDIM_ALL_BOXED(op) \
   m.impl(#op, torch::CppFunction::makeFromBoxedFunction<boxed_existing_bdim_all_batch_rule>());

aten/src/ATen/functorch/BatchRulesLinearAlgebra.cpp

@@ -384,7 +384,7 @@ fourOutputs solve_ex_batch_rule(
 
   // NOTE [ solve_ex Batch Rule Contiguity ]
   // A determines whether or not linalg_solve takes an optimized path. We need the check on A_ to match the one run on
-  // A as BatchedTensor since it might have been saved by autograd (specifically by the jvp) and the autograd behvaior
+  // A as BatchedTensor since it might have been saved by autograd (specifically by the jvp) and the autograd behavior
   // differs based on whether or not the optimized path was taken
   const auto batched_A_was_contiguous = A_bdim.has_value() ? at::select(A, *A_bdim, 0).is_contiguous() : A.is_contiguous();
   if (batched_A_was_contiguous && !A.is_complex()) {

aten/src/ATen/functorch/BatchRulesReduceOps.cpp

@@ -282,7 +282,7 @@ static std::tuple<Tensor, std::optional<int64_t>> _softmax_backward_batch_rule(
 
   dim = getPhysicalDim(output_, /*has_batch_dim*/true, dim);
 
-  // Not sure why output_ needs to be marked as .contiguous(). Someting must
+  // Not sure why output_ needs to be marked as .contiguous(). Something must
   // have changed in PyTorch (and output of softmax is probably always contiguous)
   return std::make_tuple(at::_softmax_backward_data(grad_output_, output_.contiguous(), dim, input_dtype), 0);
 }

aten/src/ATen/functorch/BatchedFallback.cpp

@@ -224,7 +224,7 @@ static Tensor safeStack(TensorList tensors) {
   // is possible for the backward function to return an undefined grad for some
   // grad_input for each example. In that case, we return an undefined grad.
   //
-  // It is theoretically posssible for *some* of the examples to produce an
+  // It is theoretically possible for *some* of the examples to produce an
   // undefined grad (a kernel could peek at the gradient values and return an
   // undefined tensor if it determines the gradient is full of zeros). We
   // could handle this by treating the undefined grad as a zero-filled tensor

aten/src/ATen/functorch/BatchedTensorImpl.cpp

@@ -113,7 +113,7 @@ SymIntArrayRef BatchedTensorImpl::sym_sizes_custom() const {
   return sym_sizes_default();
 }
 
-// The following are publically exposed as methods of Tensor
+// The following are publicly exposed as methods of Tensor
 
 IntArrayRef BatchedTensorImpl::strides_custom() const {
   return strides_default();

aten/src/ATen/functorch/DynamicLayer.h

@@ -37,7 +37,7 @@ namespace at::functorch  {
 // how to perform the transform.
 //
 // TODO: we can excise DynamicLayer in favor of Interpreter,
-// But I am going to leave it for now as a compatiblity shim to avoid
+// But I am going to leave it for now as a compatibility shim to avoid
 // needing to refactor a lot of callsites...
 struct TORCH_API DynamicLayer {
   explicit DynamicLayer(

aten/src/ATen/functorch/Interpreter.h

@@ -88,7 +88,7 @@ std::ostream& operator<<(std::ostream& os, const TransformType& t);
 // >>> VmapInterpreterPtr(&interpreter).batchSize()
 //
 // Finally, Interpreter::process switches on the type of the interpreter
-// and calls one of {Transform}Intepreter::processImpl under the hood.
+// and calls one of {Transform}Interpreter::processImpl under the hood.
 // Same for Interpreter::sendToNextInterpreter :)
 
 struct VmapInterpreterMeta {

aten/src/ATen/functorch/LegacyBatchingRegistrations.cpp

@@ -733,7 +733,7 @@ TORCH_LIBRARY_IMPL(_, FuncTorchBatched, m) {
 }
 
 TORCH_LIBRARY_IMPL(aten, FuncTorchBatched, m) {
-  // still legacy b/c teturns multiple tensors
+  // still legacy b/c returns multiple tensors
   m.impl("split.Tensor", split_batching_rule);
   m.impl("split_with_sizes", split_with_sizes_batching_rule);
   m.impl("split_with_sizes_copy", split_with_sizes_copy_batching_rule);

aten/src/ATen/mps/MPSStream.mm

@@ -158,7 +158,7 @@ void MPSStream::fill(id<MTLBuffer> buffer, uint8_t value, size_t length, size_t
       endKernelCoalescing();
       id<MTLBlitCommandEncoder> blitEncoder = [commandBuffer() blitCommandEncoder];
 
-      // For some reason fillBufferfor stopped working for lengh > 4Gb on MacOS 26
+      // For some reason fillBufferfor stopped working for length > 4Gb on MacOS 26
       // See https://github.com/pytorch/pytorch/issues/163962
       // Workaround by batching copy commands into 4Gb chunks
       constexpr size_t max_copy_size = 0x100000000; // 4GB

aten/src/ATen/native/LinearAlgebraUtils.h

@@ -148,7 +148,7 @@ inline void checkInputsSolver(const Tensor& A,
 
 inline bool is_row_or_column_contiguous(const Tensor& t) {
   // This could be made more general, similar to how it's checked in matmul, which would allow to
-  // ellide the copy with strides such as (6, 12, 1, 3) or (3, 1, 9), but this is quite tricky.
+  // elide the copy with strides such as (6, 12, 1, 3) or (3, 1, 9), but this is quite tricky.
   // We choose to be conservative for simplicity
   return t.is_contiguous() || t.transpose(-2, -1).is_contiguous();
 }

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/SpectralOpsUtils.h

@@ -21,7 +21,7 @@ enum class fft_norm_mode {
 // NOTE [ Fourier Transform Conjugate Symmetry ]
 //
 // Real-to-complex Fourier transform satisfies the conjugate symmetry. That is,
-// assuming X is the transformed K-dimensionsal signal, we have
+// assuming X is the transformed K-dimensional signal, we have
 //
 //     X[i_1, ..., i_K] = X[j_i, ..., j_K]*,
 //

aten/src/ATen/native/ao_sparse/quantized/cpu/qlinear.cpp

@@ -128,7 +128,7 @@ at::Tensor PackedLinearWeight::apply_impl(
   auto* input_tr_ptr =
       reinterpret_cast<uint8_t*>(input_tr.data_ptr<c10::quint8>());
   // TODO: Activation transpose before and after the kernel can be removed if we
-  // keep activation tensor always tranposed.
+  // keep activation tensor always transposed.
   fbgemm::transpose_simd<uint8_t>(
       batch_size, K, input_ptr, K, input_tr_ptr, batch_size);
 

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

@@ -520,7 +520,7 @@ cpu_adaptive_avg_pool3d_channels_last(
       scalar_t* out = output_data + i * channels;
       int64_t size = channels;
 
-      // Note: For oridinary usage scenario, each out lane should
+      // Note: For ordinary usage scenario, each out lane should
       //   fit in L1 cache; otherwise consider block dim C.
       // Pass I: zero the out lane
       int64_t d1 = 0;

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/cpu/DistanceOpsKernel.cpp

@@ -34,7 +34,7 @@ struct Dist {
   //     finish :   This tells what to do with the aggregated value to compute
   //                the norm. Generally this is the result of val ^ (1 / p).
   //     backward : This is the gradient for that norm. Arguments are pretty
-  //                self explanitory.
+  //                self explanatory.
   //
   // There are a few cases where these aren't used. The 0 norm has no backward,
   // because it's always 0, so that's shortcircuited earlier. There's a special

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

@@ -30,7 +30,7 @@ vec::Vectorized<scalar_t> is_nan_vec(vec::Vectorized<scalar_t> vec) {
   return vec.isnan();
 }
 
-// TODO: use is_integeral/is_same to check the scalar_t and simplify the implementation
+// TODO: use is_integral/is_same to check the scalar_t and simplify the implementation
 // currently it does not work
 template <>
 vec::Vectorized<unsigned char> is_nan_vec<unsigned char>(vec::Vectorized<unsigned char> vec) {

aten/src/ATen/native/cpu/README.md

@@ -74,7 +74,7 @@ it to sum up the entire array into a single value.
 
 `ReduceOpsKernel.cpp` uses the `CPU_CAPABILITY_*` macros to "know" under which
 compiler flags it is currently compiled. This allows the programmer to write
-generic code, which will be compiled under multipled compilation settings.
+generic code, which will be compiled under multiplied compilation settings.
 
 `../ReduceOps.cpp` now includes the header `ReduceOpsKernel.h`, which contains
 a generic definition of `sumImplAll`. This function allows the user to reduce

aten/src/ATen/native/cpu/UpSampleKernelAVXAntialias.h

@@ -889,7 +889,7 @@ void ImagingResampleHorizontalConvolution8u(
             _mm_loadu_si128((__m128i *) (lineIn_min + stride * i))),
             _mm_loadu_si128((__m128i *) (lineIn_min + stride * (i + 4))), 1);
 
-        // Extract lower part of each lane, cast to epi16 and reoder RGBARGBA -> RRGGBBAA
+        // Extract lower part of each lane, cast to epi16 and reorder RGBARGBA -> RRGGBBAA
         // RGBA: pix1 = [
         //   r0 0 r1 0  g0 0 g1 0  b0 0 b1 0  a0 0 a1 0
         //   r4 0 r5 0  g4 0 g5 0  b4 0 b5 0  a4 0 a5 0

aten/src/ATen/native/cpu/avx_mathfun.h

@@ -240,7 +240,7 @@ _PS256_CONST(coscof_p2,  4.166664568298827E-002);
 _PS256_CONST(cephes_FOPI, 1.27323954473516); // 4 / M_PI
 
 
-/* evaluation of 8 sines at onces using AVX intrinsics
+/* evaluation of 8 sines at once using AVX intrinsics
 
    The code is the exact rewriting of the cephes sinf function.
    Precision is excellent as long as x < 8192 (I did not bother to

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

@@ -311,7 +311,7 @@ void GroupNormKernelImplChannelsLastInternal(
   const bool gamma_null = (gamma_data == nullptr);
   const bool beta_null = beta_data == nullptr;
 
-  // NB: About algorithm choosen:
+  // NB: About algorithm chosen:
   //
   // On channels last, GroupNorm has a input shape of {N, H, W, GD},
   // Mean and rstd are collected per each n and g, which involves reduction

aten/src/ATen/native/cpu/int4mm_kernel.cpp

@@ -930,7 +930,7 @@ void ref_dyn_quant_matmul_4bit_channelwise_kernel(
         }
       };
 
-  // Dynamically Quantize the float32 input to 8 bit assymetric
+  // Dynamically Quantize the float32 input to 8 bit asymmetric
   input_quant_pack_8bit_channelwise(m, k, lhs_f32, (int8_t*)lhs_qa8dx);
 
   const size_t lhs_stride =
@@ -1163,7 +1163,7 @@ void dyn_quant_matmul_4bit_kernel(
   const int64_t weight_packed_size =
       kleidiai::kai_pack_rhs_int4_size(N, K, block_size);
   if (weight_packed_size == packed_weights.numel()) {
-    // KleidiAI interface intenally handles the Channelwise and groupwise
+    // KleidiAI interface internally handles the Channelwise and groupwise
     // distinction
     kleidiai::kai_quant_pack_lhs_int4_mm(
         output, inp, packed_weights, M, N, K, block_size);

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

@@ -13,6 +13,7 @@
 #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>
@@ -360,7 +361,7 @@ static bool isInputCompliesAddmmCudaLt(Tensor& result, const Tensor& self, const
       // and the leading stride is at least max(1, other dim length), so we might
       // end up with contiguous cols but not rows (i.e. holes between different rows)
       // and vice versa.
-      mat2_sizes[0] < 65535 * 32 && mat2_sizes[1] < 65535 * 32 &&
+      && mat2_sizes[0] < 65535 * 32 && mat2_sizes[1] < 65535 * 32 &&
       mat1_sizes[0] < 65535 * 32 && mat1_sizes[1] < 65535 * 32 &&
       && (
         // filter by dtype
@@ -504,7 +505,7 @@ Tensor& addmm_out_cuda_impl(Tensor& result, const Tensor& self, const Tensor& ma
   // Handle whether to use the Lt interface {
   static bool persistent_disable_addmm_cuda_lt = isGloballyDisabledAddmmCudaLt(self.device());
   // if lt path fails, we recurse back into this function here and force the lt path to off
-  // we cannot update varible disable_addmm_cuda_lt from above since it is static and would be permanent
+  // we cannot update variable disable_addmm_cuda_lt from above since it is static and would be permanent
   bool disable_addmm_cuda_lt = persistent_disable_addmm_cuda_lt || disable_addmm_cuda_lt_override;
   #ifdef USE_ROCM
   // Conditioned on the device index, which is not persistent
@@ -1628,104 +1629,6 @@ _scaled_mm_out_cuda(const Tensor& mat1, const Tensor& mat2,
   return _scaled_gemm(mat1, mat2, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);
 }
 
-namespace {
-  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());
-    }
-  }
-}
-
 Tensor
 _scaled_mm_cuda(const Tensor& mat_a, const Tensor& mat_b,
           const Tensor& scale_a,
@@ -1740,261 +1643,26 @@ _scaled_mm_cuda(const Tensor& mat_a, const Tensor& mat_b,
   return _scaled_mm_out_cuda(mat_a, mat_b, scale_a, scale_b, bias, scale_result, out_dtype, use_fast_accum, out);
 }
 
-/**
- * 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;
-}
-
-/**
- * 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;
-}
-
 using acceptance_fn = std::function<bool(c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&, c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&)>;
 using namespace std::placeholders;
 
+namespace scaled_blas = at::cuda::scaled;
+using scaled_blas::ScaledGemmImplementation;
+using scaled_blas::convert_int_to_enum;
+
 std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 9> scale_kernel_dispatch = {{
-  { "tensorwise_tensorwise", check_tensorwise_recipe, ScaledGemmImplementation::TENSORWISE_TENSORWISE },
-  { "rowwise_rowwise", check_rowwise_recipe, ScaledGemmImplementation::ROWWISE_ROWWISE},
-  { "block_1x128_128x128", std::bind(check_deepseek_recipe, ScalingType::BlockWise1x128, ScalingType::BlockWise128x128, _1, _2, _3, _4, _5, _6),
+  { "tensorwise_tensorwise", scaled_blas::check_tensorwise_recipe, ScaledGemmImplementation::TENSORWISE_TENSORWISE },
+  { "rowwise_rowwise", scaled_blas::check_rowwise_recipe, ScaledGemmImplementation::ROWWISE_ROWWISE},
+  { "block_1x128_128x128", std::bind(scaled_blas::check_deepseek_recipe, ScalingType::BlockWise1x128, ScalingType::BlockWise128x128, _1, _2, _3, _4, _5, _6),
     ScaledGemmImplementation::BLOCK_1x128_128x128},
-  { "block_128x128_1x128", std::bind(check_deepseek_recipe, ScalingType::BlockWise128x128, ScalingType::BlockWise1x128, _1, _2, _3, _4, _5, _6),
+  { "block_128x128_1x128", std::bind(scaled_blas::check_deepseek_recipe, ScalingType::BlockWise128x128, ScalingType::BlockWise1x128, _1, _2, _3, _4, _5, _6),
     ScaledGemmImplementation::BLOCK_128x128_1x128},
-  { "block_1x128_1x128", std::bind(check_deepseek_recipe, ScalingType::BlockWise1x128, ScalingType::BlockWise1x128, _1, _2, _3, _4, _5, _6),
+  { "block_1x128_1x128", std::bind(scaled_blas::check_deepseek_recipe, ScalingType::BlockWise1x128, ScalingType::BlockWise1x128, _1, _2, _3, _4, _5, _6),
     ScaledGemmImplementation::BLOCK_1x128_1x128},
-  { "nvfp4_nvfp4", check_nvfp4_recipe, ScaledGemmImplementation::NVFP4_NVFP4},
-  { "nvfp4_nvfp4_single_scale", check_nvfp4_recipe_single_scale, ScaledGemmImplementation::NVFP4_NVFP4_SINGLE_SCALE },
-  { "mxfp8_mxfp8", check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8},
-  { "mxfp4_mxfp4", check_mxfp4_recipe, ScaledGemmImplementation::MXFP4_MXFP4}}};
+  { "nvfp4_nvfp4", scaled_blas::check_nvfp4_recipe, ScaledGemmImplementation::NVFP4_NVFP4},
+  { "nvfp4_nvfp4_single_scale", scaled_blas::check_nvfp4_recipe_single_scale, ScaledGemmImplementation::NVFP4_NVFP4_SINGLE_SCALE },
+  { "mxfp8_mxfp8", scaled_blas::check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8},
+  { "mxfp4_mxfp4", scaled_blas::check_mxfp4_recipe, ScaledGemmImplementation::MXFP4_MXFP4}}};
 
 Tensor&
 _scaled_tensorwise_tensorwise(
@@ -2596,410 +2264,6 @@ _scaled_mm_cuda_v2(
                       out);
 }
 
-// 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
-}
-
-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 {
-
-std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 2> scale_grouped_kernel_dispatch = {{
-  { "rowwise_rowwise", check_rowwise_recipe, ScaledGemmImplementation::ROWWISE_ROWWISE},
-  { "mxfp8_mxfp8", 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;
-}
-
 static void baddbmm_bmm_out_dtype_checks(const Tensor& batch1, const Tensor& batch2, const Scalar& beta, const Scalar& alpha, const at::ScalarType out_dtype, bool is_bmm, const std::optional<Tensor>& self_baddbmm = std::nullopt) {
   // ref ATen/native/LinearAlgebra.cpp common_checks_baddbmm_bmm
   TORCH_CHECK(batch1.dim() == 3, "batch1 must be a 3D tensor");

aten/src/ATen/native/cuda/DistributionTemplates.h

@@ -494,7 +494,7 @@ void uniform_kernel(TensorIteratorBase& iter, double from_, double to_, RNG gen)
       auto value = static_cast<scalar_t>(rand * range + from);
       // reverse the bounds of curand4 from (0, 1] to [0, 1)
       // Note that this method is from legacy THCTensorRandom and is likely to give
-      // you more 0-s, since, the probability of gettings 1-s is higher than 0-s and
+      // you more 0-s, since, the probability of getting 1-s is higher than 0-s and
       // by reversing the bounds, we are flipping the probabilities of 1-s and 0-s.
       // BEFORE TOUCHING THIS CODE READ: https://github.com/pytorch/pytorch/issues/16706
       auto reverse_bound_value = value == to ? from : value;

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/KernelUtils.cuh

@@ -6,7 +6,7 @@
 #endif
 
 // ROCm 6.3 is planned to have these functions, but until then here they are.
-#if defined(USE_ROCM) && ROCM_VERSION >= 60201
+#if defined(USE_ROCM)
 #include <device_functions.h>
 #include <hip/hip_fp16.h>
 #include <hip/hip_bf16.h>
@@ -115,9 +115,7 @@ __device__ __forceinline__ void fastSpecializedAtomicAdd(
     index_t index,
     const index_t numel,
     scalar_t value) {
-#if (                      \
-    (defined(USE_ROCM) && ROCM_VERSION < 60201) || \
-    (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 700)))
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 700))
   gpuAtomicAddNoReturn(
       reinterpret_cast<at::Half*>(tensor) + index,
       static_cast<at::Half>(value));
@@ -160,9 +158,7 @@ __device__ __forceinline__ void fastSpecializedAtomicAdd(
     index_t index,
     const index_t numel,
     scalar_t value) {
-#if (                      \
-    (defined(USE_ROCM) && ROCM_VERSION < 60201) || \
-    (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 800)))
+#if (defined(__CUDA_ARCH__) && (__CUDA_ARCH__ < 800))
   gpuAtomicAddNoReturn(
       reinterpret_cast<at::BFloat16*>(tensor) + index,
       static_cast<at::BFloat16>(value));

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

@@ -154,7 +154,7 @@ REGISTER_CUDA_DISPATCH(lstsq_stub, &lazy_lstsq_kernel)
 
 // Old style dispatches
 // torch_cuda_linalg dynamic library should have a global constructor
-// that calls regiserLinaglDispatch so in order ot lazy bind
+// that calls registerLinalgDispatch so in order ot lazy bind
 // old style dispatch all one have to do is to load library and call disp.func_name
 // Protect from infinite recursion by initializing dispatch to self and checking
 // that values are different after linalg library were loaded

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/Normalization.cuh

@@ -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][std::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