add test print

.ci/docker/build.sh

@@ -195,16 +195,13 @@ case "$tag" in
     NINJA_VERSION=1.9.0
     TRITON=yes
     ;;
-  pytorch-linux-jammy-xpu-n-py3 | pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks)
+  pytorch-linux-jammy-xpu-n-py3)
     ANACONDA_PYTHON_VERSION=3.10
     GCC_VERSION=11
     VISION=yes
     XPU_VERSION=2025.2
     NINJA_VERSION=1.9.0
     TRITON=yes
-    if [[ $tag =~ "benchmarks" ]]; then
-      INDUCTOR_BENCHMARKS=yes
-    fi
     ;;
   pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
     ANACONDA_PYTHON_VERSION=3.10

.ci/docker/common/install_acl.sh

@@ -3,7 +3,7 @@
 
 set -eux
 
-ACL_VERSION=${ACL_VERSION:-"v52.6.0"}
+ACL_VERSION=${ACL_VERSION:-"v25.02"}
 ACL_INSTALL_DIR="/acl"
 
 # Clone ACL

.ci/docker/common/install_conda.sh

@@ -49,20 +49,12 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
     export SYSROOT_DEP="sysroot_linux-64=2.17"
   fi
 
-# Install correct Python version
-# Also ensure sysroot is using a modern GLIBC to match system compilers
-if [ "$ANACONDA_PYTHON_VERSION" = "3.14" ]; then
-  as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
-             python="3.14.0" \
-             ${SYSROOT_DEP} \
-             -c conda-forge
-else
   # Install correct Python version
   # Also ensure sysroot is using a modern GLIBC to match system compilers
   as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
              python="$ANACONDA_PYTHON_VERSION" \
              ${SYSROOT_DEP}
-fi
+
   # libstdcxx from conda default channels are too old, we need GLIBCXX_3.4.30
   # which is provided in libstdcxx 12 and up.
   conda_install libstdcxx-ng=12.3.0 --update-deps -c conda-forge

.ci/docker/common/install_cuda.sh

@@ -10,7 +10,7 @@ else
   arch_path='sbsa'
 fi
 
-NVSHMEM_VERSION=3.4.5
+NVSHMEM_VERSION=3.3.24
 
 function install_cuda {
   version=$1
@@ -150,7 +150,7 @@ function install_130 {
   CUDNN_VERSION=9.13.0.50
   echo "Installing CUDA 13.0 and cuDNN ${CUDNN_VERSION} and NVSHMEM and NCCL and cuSparseLt-0.7.1"
   # install CUDA 13.0 in the same container
-  install_cuda 13.0.2 cuda_13.0.2_580.95.05_linux
+  install_cuda 13.0.0 cuda_13.0.0_580.65.06_linux
 
   # cuDNN license: https://developer.nvidia.com/cudnn/license_agreement
   install_cudnn 13 $CUDNN_VERSION

.ci/docker/common/install_rocm.sh

@@ -40,7 +40,11 @@ EOF
 
     # Default url values
     rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
+    amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu"
+
+    # Add amdgpu repository
     UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
+    echo "deb [arch=amd64] ${amdgpu_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
 
     # Add rocm repository
     wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -

.ci/docker/common/install_rocm_magma.sh

@@ -12,8 +12,8 @@ function do_install() {
 
     rocm_version_nodot=${rocm_version//./}
 
-    # post merge of https://github.com/icl-utk-edu/magma/pull/65
-    MAGMA_VERSION=c0792ae825fb36872784892ea643dd6f3456bc5f
+    # https://github.com/icl-utk-edu/magma/pull/65
+    MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
     magma_archive="magma-rocm${rocm_version_nodot}-${MAGMA_VERSION}-1.tar.bz2"
 
     rocm_dir="/opt/rocm"

.ci/docker/manywheel/build.sh

@@ -97,7 +97,7 @@ case ${image} in
     manylinux2_28-builder:xpu)
         TARGET=xpu_final
         GPU_IMAGE=amd64/almalinux:8
-        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=13"
+        DOCKER_GPU_BUILD_ARG=" --build-arg DEVTOOLSET_VERSION=11"
         MANY_LINUX_VERSION="2_28"
         ;;
     *)

.ci/docker/requirements-ci.txt

@@ -138,12 +138,10 @@ numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
 #test_binary_ufuncs.py
 numpy==1.22.4; python_version == "3.10"
 numpy==1.26.2; python_version == "3.11" or python_version == "3.12"
-numpy==2.1.2; python_version >= "3.13" and python_version < "3.14"
-numpy==2.3.4; python_version >= "3.14"
+numpy==2.1.2; python_version >= "3.13"
 
 pandas==2.0.3; python_version < "3.13"
-pandas==2.2.3; python_version >= "3.13" and python_version < "3.14"
-pandas==2.3.3; python_version >= "3.14"
+pandas==2.2.3; python_version >= "3.13"
 
 #onnxruntime
 #Description: scoring engine for Open Neural Network Exchange (ONNX) models
@@ -155,8 +153,7 @@ opt-einsum==3.3
 #Pinned versions: 3.3
 #test that import: test_linalg.py
 
-optree==0.13.0 ; python_version < "3.14"
-optree==0.17.0 ; python_version >= "3.14"
+optree==0.13.0
 #Description: A library for tree manipulation
 #Pinned versions: 0.13.0
 #test that import: test_vmap.py, test_aotdispatch.py, test_dynamic_shapes.py,
@@ -255,8 +252,7 @@ scikit-image==0.22.0
 #test that import:
 
 scipy==1.10.1 ; python_version <= "3.11"
-scipy==1.14.1 ; python_version > "3.11" and python_version < "3.14"
-scipy==1.16.2 ; python_version >= "3.14"
+scipy==1.14.1 ; python_version >= "3.12"
 # Pin SciPy because of failing distribution tests (see #60347)
 #Description: scientific python
 #Pinned versions: 1.10.1
@@ -328,8 +324,7 @@ pywavelets==1.7.0 ; python_version >= "3.12"
 #Pinned versions: 1.4.1
 #test that import:
 
-lxml==5.3.0 ; python_version < "3.14"
-lxml==6.0.2 ; python_version >= "3.14"
+lxml==5.3.0
 #Description: This is a requirement of unittest-xml-reporting
 
 PyGithub==2.3.0
@@ -339,9 +334,7 @@ sympy==1.13.3
 #Pinned versions:
 #test that import:
 
-onnx==1.19.1 ; python_version < "3.14"
-# Unpin once Python 3.14 is supported. See  onnxruntime issue 26309.
-onnx==1.18.0 ; python_version == "3.14"
+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:
@@ -366,7 +359,7 @@ pwlf==2.2.1
 #test that import: test_sac_estimator.py
 
 # To build PyTorch itself
-pyyaml==6.0.3
+pyyaml==6.0.2
 pyzstd
 setuptools==78.1.1
 packaging==23.1

.ci/docker/ubuntu-xpu/Dockerfile

@@ -54,15 +54,12 @@ ENV OPENSSL_DIR /opt/openssl
 RUN rm install_openssl.sh
 
 ARG INDUCTOR_BENCHMARKS
-ARG ANACONDA_PYTHON_VERSION
-ENV ANACONDA_PYTHON_VERSION=$ANACONDA_PYTHON_VERSION
 COPY ./common/install_inductor_benchmark_deps.sh install_inductor_benchmark_deps.sh
 COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
-COPY ci_commit_pins/torchbench.txt torchbench.txt
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
-RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
+RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt
 
 # Install XPU Dependencies
 ARG XPU_VERSION

.ci/docker/ubuntu/Dockerfile

@@ -100,8 +100,6 @@ COPY ./common/common_utils.sh common_utils.sh
 COPY ci_commit_pins/huggingface-requirements.txt huggingface-requirements.txt
 COPY ci_commit_pins/timm.txt timm.txt
 COPY ci_commit_pins/torchbench.txt torchbench.txt
-# Only build aoti cpp tests when INDUCTOR_BENCHMARKS is set to True
-ENV BUILD_AOT_INDUCTOR_TEST ${INDUCTOR_BENCHMARKS}
 RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
 RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
 

.ci/lumen_cli/pyproject.toml

@@ -6,7 +6,7 @@ dependencies = [
     "GitPython==3.1.45",
     "docker==7.1.0",
     "pytest==7.3.2",
-    "uv==0.9.6"
+    "uv==0.9.5"
 ]
 
 [tool.setuptools]

.ci/magma-rocm/Makefile

@@ -1,7 +1,7 @@
 SHELL=/usr/bin/env bash
 
 DOCKER_CMD ?= docker
-DESIRED_ROCM ?= 7.1
+DESIRED_ROCM ?= 7.0
 DESIRED_ROCM_SHORT = $(subst .,,$(DESIRED_ROCM))
 PACKAGE_NAME = magma-rocm
 # inherit this from underlying docker image, do not pass this env var to docker
@@ -16,7 +16,6 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
 	magma-rocm/build_magma.sh
 
 .PHONY: all
-all: magma-rocm71
 all: magma-rocm70
 all: magma-rocm64
 
@@ -25,11 +24,6 @@ clean:
 	$(RM) -r magma-*
 	$(RM) -r output
 
-.PHONY: magma-rocm71
-magma-rocm71: DESIRED_ROCM := 7.1
-magma-rocm71:
-	$(DOCKER_RUN)
-
 .PHONY: magma-rocm70
 magma-rocm70: DESIRED_ROCM := 7.0
 magma-rocm70:

.ci/magma-rocm/build_magma.sh

@@ -6,8 +6,8 @@ set -eou pipefail
 # The script expects DESIRED_CUDA and PACKAGE_NAME to be set
 ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
 
-# post merge of https://github.com/icl-utk-edu/magma/pull/65
-MAGMA_VERSION=c0792ae825fb36872784892ea643dd6f3456bc5f
+# https://github.com/icl-utk-edu/magma/pull/65
+MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
 
 # Folders for the build
 PACKAGE_FILES=${ROOT_DIR}/magma-rocm/package_files # metadata
@@ -20,7 +20,7 @@ mkdir -p ${PACKAGE_DIR} ${PACKAGE_OUTPUT}/linux-64 ${PACKAGE_BUILD} ${PACKAGE_RE
 
 # Fetch magma sources and verify checksum
 pushd ${PACKAGE_DIR}
-git clone https://github.com/icl-utk-edu/magma
+git clone https://github.com/jeffdaily/magma
 pushd magma
 git checkout ${MAGMA_VERSION}
 popd

.ci/pytorch/build.sh

@@ -426,7 +426,7 @@ fi
 if [[ "$BUILD_ENVIRONMENT" != *libtorch* && "$BUILD_ENVIRONMENT" != *bazel* ]]; then
   # export test times so that potential sharded tests that'll branch off this build will use consistent data
   # don't do this for libtorch as libtorch is C++ only and thus won't have python tests run on its build
-  PYTHONPATH=. python tools/stats/export_test_times.py
+  python tools/stats/export_test_times.py
 fi
 # don't do this for bazel or s390x or riscv64 as they don't use sccache
 if [[ "$BUILD_ENVIRONMENT" != *s390x* && "$BUILD_ENVIRONMENT" != *riscv64* && "$BUILD_ENVIRONMENT" != *-bazel-* ]]; then

.ci/pytorch/test.sh

@@ -460,18 +460,28 @@ test_inductor_shard() {
     --verbose
 }
 
-test_inductor_aoti_cpp() {
+test_inductor_aoti() {
+  # docker build uses bdist_wheel which does not work with test_aot_inductor
+  # TODO: need a faster way to build
   if [[ "$BUILD_ENVIRONMENT" == *rocm* ]]; then
     # We need to hipify before building again
     python3 tools/amd_build/build_amd.py
   fi
   if [[ "$BUILD_ENVIRONMENT" == *sm86* ]]; then
+    BUILD_COMMAND=(TORCH_CUDA_ARCH_LIST=8.6 USE_FLASH_ATTENTION=OFF python -m pip install --no-build-isolation -v -e .)
     # TODO: Replace me completely, as one should not use conda libstdc++, nor need special path to TORCH_LIB
     TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="/opt/conda/envs/py_3.10/lib:${TORCH_LIB_DIR}:${LD_LIBRARY_PATH}")
   else
+    BUILD_COMMAND=(python -m pip install --no-build-isolation -v -e .)
     TEST_ENVS=(CPP_TESTS_DIR="${BUILD_BIN_DIR}" LD_LIBRARY_PATH="${TORCH_LIB_DIR}")
   fi
 
+  # aoti cmake custom command requires `torch` to be installed
+  # initialize the cmake build cache and install torch
+  /usr/bin/env "${BUILD_COMMAND[@]}"
+  # rebuild with the build cache with `BUILD_AOT_INDUCTOR_TEST` enabled
+  /usr/bin/env CMAKE_FRESH=1 BUILD_AOT_INDUCTOR_TEST=1 "${BUILD_COMMAND[@]}"
+
   /usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
 }
 
@@ -572,8 +582,6 @@ fi
 
 if [[ "${TEST_CONFIG}" == *cpu* ]]; then
   DYNAMO_BENCHMARK_FLAGS+=(--device cpu)
-elif [[ "${TEST_CONFIG}" == *xpu* ]]; then
-  DYNAMO_BENCHMARK_FLAGS+=(--device xpu)
 else
   DYNAMO_BENCHMARK_FLAGS+=(--device cuda)
 fi
@@ -667,8 +675,6 @@ test_perf_for_dashboard() {
     device=cuda_b200
   elif [[ "${TEST_CONFIG}" == *rocm* ]]; then
     device=rocm
-  elif [[ "${TEST_CONFIG}" == *xpu* ]]; then
-    device=xpu
   fi
 
   for mode in "${modes[@]}"; do
@@ -1761,7 +1767,7 @@ elif [[ "${TEST_CONFIG}" == *torchbench* ]]; then
   else
     # Do this after checkout_install_torchbench to ensure we clobber any
     # nightlies that torchbench may pull in
-    if [[ "${TEST_CONFIG}" != *cpu* && "${TEST_CONFIG}" != *xpu* ]]; then
+    if [[ "${TEST_CONFIG}" != *cpu* ]]; then
       install_torchrec_and_fbgemm
     fi
     PYTHONPATH=/torchbench test_dynamo_benchmark torchbench "$id"
@@ -1770,7 +1776,7 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
   install_torchvision
   PYTHONPATH=/torchbench test_inductor_cpp_wrapper_shard "$SHARD_NUMBER"
   if [[ "$SHARD_NUMBER" -eq "1" ]]; then
-    test_inductor_aoti_cpp
+    test_inductor_aoti
   fi
 elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
   install_torchvision

.ci/pytorch/windows/internal/install_python.bat

@@ -7,9 +7,12 @@ if "%DESIRED_PYTHON%" == "3.13t" (
     set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.13.0/python-3.13.0-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.13t"
+) else if "%DESIRED_PYTHON%"=="3.14" (
+    echo Python version is set to 3.14 or 3.14t
+    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
 ) else if "%DESIRED_PYTHON%"=="3.14t" (
     echo Python version is set to 3.14 or 3.14t
-    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0-amd64.exe"
+    set "PYTHON_INSTALLER_URL=https://www.python.org/ftp/python/3.14.0/python-3.14.0rc1-amd64.exe"
     set ADDITIONAL_OPTIONS="Include_freethreaded=1"
     set PYTHON_EXEC="python3.14t"
 ) else (

.claude/skills/pytorch-docstring.md

@@ -1,8 +1,3 @@
----
-name: docstring
-description: Write docstrings for PyTorch functions and methods following PyTorch conventions. Use when writing or updating docstrings in PyTorch code.
----
-
 # PyTorch Docstring Writing Guide
 
 This skill describes how to write docstrings for functions and methods in the PyTorch project, following the conventions in `torch/_tensor_docs.py` and `torch/nn/functional.py`.

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

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

.github/actions/diskspace-cleanup/action.yml

@@ -27,9 +27,7 @@ runs:
             docker system prune -af
             diskspace_new=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
             if [[ "$diskspace_new" -gt "$diskspace_cutoff" ]] ; then
-                diskspace_cutoff_int=$((diskspace_cutoff + 0))
-                difference=$((100 - diskspace_cutoff_int))
-                echo "Error: Available diskspace is less than $difference percent. Not enough diskspace."
+                echo "Error: Available diskspace is less than $diskspace_cutoff percent. Not enough diskspace."
                 echo "$msg"
                 exit 1
             else

.github/ci_commit_pins/audio.txt

@@ -1 +1 @@
-3b0e7a6f192ca2715e7e6cbe5db007aea7165fe2
+69bbe7363897764f9e758d851cd0340147d27f94

.github/ci_commit_pins/vision.txt

@@ -1 +1 @@
-218d2ab791d437309f91e0486eb9fa7f00badc17
+1752fe6809b74921644866275ab80244b96e80bc

.github/ci_commit_pins/xla.txt

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

.github/merge_rules.yaml

@@ -540,26 +540,6 @@
   - Lint
   - pull
 
-- name: PrivateUse1
-  patterns:
-  - torch/accelerator/**
-  - torch/utils/backend_registration.py
-  - torch/csrc/acc/**
-  - torch/csrc/DeviceAccelerator.*
-  - torch/csrc/profiler/standalone/privateuse1_observer.*
-  - aten/src/ATen/DeviceAccelerator.*
-  - aten/src/ATen/core/GeneratorForPrivateuseone.*
-  - aten/src/ATen/detail/PrivateUse1HooksInterface.*
-  - docs/source/accelerator/**
-  - test/cpp_extensions/open_registration_extension/torch_openreg/**
-  approved_by:
-  - albanD
-  - fffrog
-  mandatory_checks_name:
-  - EasyCLA
-  - Lint
-  - pull
-
 - name: superuser
   patterns:
   - '*'

.github/pytorch-probot.yml

@@ -19,7 +19,6 @@ ciflow_push_tags:
 - ciflow/inductor-perf-test-nightly-rocm-mi300
 - ciflow/inductor-perf-test-nightly-rocm-mi355
 - ciflow/inductor-perf-test-nightly-x86-zen
-- ciflow/inductor-perf-test-nightly-xpu
 - ciflow/inductor-periodic
 - ciflow/inductor-rocm
 - ciflow/linux-aarch64
@@ -27,7 +26,6 @@ ciflow_push_tags:
 - ciflow/nightly
 - ciflow/op-benchmark
 - ciflow/periodic
-- ciflow/periodic-rocm-mi200
 - ciflow/periodic-rocm-mi300
 - ciflow/pull
 - ciflow/quantization-periodic

.github/scripts/generate_binary_build_matrix.py

@@ -11,24 +11,18 @@ architectures:
     * Latest XPU
 """
 
-import json
 import os
-import re
-from pathlib import Path
 from typing import Optional
 
 
-SCRIPT_DIR = Path(__file__).absolute().parent
-REPO_ROOT = SCRIPT_DIR.parent.parent
-
-
+# NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
 CUDA_ARCHES = ["12.6", "12.8", "12.9", "13.0"]
 CUDA_STABLE = "12.8"
 CUDA_ARCHES_FULL_VERSION = {
     "12.6": "12.6.3",
     "12.8": "12.8.1",
     "12.9": "12.9.1",
-    "13.0": "13.0.2",
+    "13.0": "13.0.0",
 }
 CUDA_ARCHES_CUDNN_VERSION = {
     "12.6": "9",
@@ -37,7 +31,8 @@ CUDA_ARCHES_CUDNN_VERSION = {
     "13.0": "9",
 }
 
-ROCM_ARCHES = ["7.0", "7.1"]
+# NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
+ROCM_ARCHES = ["6.4", "7.0"]
 
 XPU_ARCHES = ["xpu"]
 
@@ -61,7 +56,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
         "nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
         "nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
         "nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
@@ -78,7 +73,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
         "nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
         "nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
         "nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
@@ -95,27 +90,27 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
         "nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
         "nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
+        "nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | "
         "nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | "
         "nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | "
         "nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
     ),
     "13.0": (
-        "nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | "
-        "nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | "
-        "nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | "
+        "nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
+        "nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
+        "nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | "
         "nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | "
-        "nvidia-cublas==13.1.0.3; platform_system == 'Linux' | "
-        "nvidia-cufft==12.0.0.61; platform_system == 'Linux' | "
+        "nvidia-cublas==13.0.0.19; platform_system == 'Linux' | "
+        "nvidia-cufft==12.0.0.15; platform_system == 'Linux' | "
         "nvidia-curand==10.4.0.35; platform_system == 'Linux' | "
-        "nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | "
-        "nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | "
+        "nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
+        "nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
         "nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
         "nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
-        "nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | "
-        "nvidia-nvtx==13.0.85; platform_system == 'Linux' | "
-        "nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | "
-        "nvidia-cufile==1.15.1.6; platform_system == 'Linux'"
+        "nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
+        "nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
+        "nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
+        "nvidia-cufile==1.15.0.42; platform_system == 'Linux'"
     ),
     "xpu": (
         "intel-cmplr-lib-rt==2025.2.1 | "
@@ -142,48 +137,9 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
 }
 
 
-# Used by tools/nightly.py
-PYTORCH_NIGHTLY_PIP_INDEX_URL = "https://download.pytorch.org/whl/nightly"
-NIGHTLY_SOURCE_MATRIX = {
-    "cpu": dict(
-        name="cpu",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/cpu",
-        supported_platforms=["Linux", "macOS", "Windows"],
-        accelerator="cpu",
-    )
-}
-CUDA_NIGHTLY_SOURCE_MATRIX = {
-    f"cuda-{major}.{minor}": dict(
-        name=f"cuda-{major}.{minor}",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/cu{major}{minor}",
-        supported_platforms=["Linux", "Windows"],
-        accelerator="cuda",
-    )
-    for major, minor in (map(int, version.split(".")) for version in CUDA_ARCHES)
-}
-ROCM_NIGHTLY_SOURCE_MATRIX = {
-    f"rocm-{major}.{minor}": dict(
-        name=f"rocm-{major}.{minor}",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/rocm{major}.{minor}",
-        supported_platforms=["Linux"],
-        accelerator="rocm",
-    )
-    for major, minor in (map(int, version.split(".")) for version in ROCM_ARCHES)
-}
-XPU_NIGHTLY_SOURCE_MATRIX = {
-    "xpu": dict(
-        name="xpu",
-        index_url=f"{PYTORCH_NIGHTLY_PIP_INDEX_URL}/xpu",
-        supported_platforms=["Linux"],
-        accelerator="xpu",
-    )
-}
-NIGHTLY_SOURCE_MATRIX.update(CUDA_NIGHTLY_SOURCE_MATRIX)
-NIGHTLY_SOURCE_MATRIX.update(ROCM_NIGHTLY_SOURCE_MATRIX)
-NIGHTLY_SOURCE_MATRIX.update(XPU_NIGHTLY_SOURCE_MATRIX)
-
-
 def get_nccl_wheel_version(arch_version: str) -> str:
+    import re
+
     requirements = map(
         str.strip, re.split("[;|]", PYTORCH_EXTRA_INSTALL_REQUIREMENTS[arch_version])
     )
@@ -191,14 +147,17 @@ def get_nccl_wheel_version(arch_version: str) -> str:
 
 
 def read_nccl_pin(arch_version: str) -> str:
-    nccl_pin_path = (
-        REPO_ROOT
-        / ".ci"
-        / "docker"
-        / "ci_commit_pins"
-        / f"nccl-cu{arch_version[:2]}.txt"
+    from pathlib import Path
+
+    nccl_pin_path = os.path.join(
+        Path(__file__).absolute().parents[2],
+        ".ci",
+        "docker",
+        "ci_commit_pins",
+        f"nccl-cu{arch_version[:2]}.txt",
     )
-    return nccl_pin_path.read_text().strip()
+    with open(nccl_pin_path) as f:
+        return f.read().strip()
 
 
 def validate_nccl_dep_consistency(arch_version: str) -> None:
@@ -206,8 +165,7 @@ def validate_nccl_dep_consistency(arch_version: str) -> None:
     wheel_ver = get_nccl_wheel_version(arch_version)
     if not nccl_release_tag.startswith(f"v{wheel_ver}"):
         raise RuntimeError(
-            f"{arch_version} NCCL release tag version {nccl_release_tag} "
-            f"does not correspond to wheel version {wheel_ver}"
+            f"{arch_version} NCCL release tag version {nccl_release_tag} does not correspond to wheel version {wheel_ver}"
         )
 
 
@@ -454,14 +412,7 @@ def generate_wheels_matrix(
     return ret
 
 
-arch_version = ""
-for arch_version in CUDA_ARCHES:
-    validate_nccl_dep_consistency(arch_version)
-del arch_version
-
-
-if __name__ == "__main__":
-    # Used by tools/nightly.py
-    (SCRIPT_DIR / "nightly_source_matrix.json").write_text(
-        json.dumps(NIGHTLY_SOURCE_MATRIX, indent=4) + "\n"
-    )
+validate_nccl_dep_consistency("13.0")
+validate_nccl_dep_consistency("12.9")
+validate_nccl_dep_consistency("12.8")
+validate_nccl_dep_consistency("12.6")

.github/workflows/_xpu-test.yml

@@ -38,10 +38,6 @@ on:
         default: ""
         description: |
           List of tests to include (empty string implies default list)
-      dashboard-tag:
-        required: false
-        type: string
-        default: ""
       disable-monitor:
         description: |
           [Experimental] Disable utilization monitoring for tests.
@@ -62,11 +58,6 @@ on:
         required: false
         type: number
         default: 1
-    secrets:
-      HUGGING_FACE_HUB_TOKEN:
-        required: false
-        description: |
-          HF Auth token to avoid rate limits when downloading models or datasets from hub
 permissions:
   id-token: write
   contents: read
@@ -205,8 +196,6 @@ jobs:
           PYTORCH_TEST_CUDA_MEM_LEAK_CHECK: ${{ matrix.mem_leak_check && '1' || '0' }}
           PYTORCH_TEST_RERUN_DISABLED_TESTS: ${{ matrix.rerun_disabled_tests && '1' || '0' }}
           TESTS_TO_INCLUDE: ${{ inputs.tests-to-include }}
-          DASHBOARD_TAG: ${{ inputs.dashboard-tag }}
-          HUGGING_FACE_HUB_TOKEN: ${{ secrets.HUGGING_FACE_HUB_TOKEN }}
         timeout-minutes: ${{ fromJson(steps.test-timeout.outputs.timeout) }}
         run: |
           # Fetch aws credential from IMDs
@@ -257,8 +246,6 @@ jobs:
             -e PYTORCH_TEST_RERUN_DISABLED_TESTS \
             -e TESTS_TO_INCLUDE \
             -e ZE_AFFINITY_MASK \
-            -e HUGGING_FACE_HUB_TOKEN \
-            -e DASHBOARD_TAG \
             --env-file="/tmp/github_env_${GITHUB_RUN_ID}" \
             --ulimit stack=10485760:83886080 \
             --ulimit core=0 \

.github/workflows/build-almalinux-images.yml

@@ -36,7 +36,7 @@ jobs:
     runs-on: linux.9xlarge.ephemeral
     strategy:
       matrix:
-        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm7.0", "rocm7.1", "cpu"]
+        tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.4", "rocm7.0", "cpu"]
     steps:
       - name: Build docker image
         uses: pytorch/pytorch/.github/actions/binary-docker-build@main

.github/workflows/build-libtorch-images.yml

@@ -52,8 +52,8 @@ jobs:
           { tag: "cuda12.9" },
           { tag: "cuda12.8" },
           { tag: "cuda12.6" },
+          { tag: "rocm6.4"  },
           { tag: "rocm7.0"  },
-          { tag: "rocm7.1"  },
           { tag: "cpu"      },
         ]
     steps:

.github/workflows/build-magma-rocm-linux.yml

@@ -34,7 +34,7 @@ jobs:
       id-token: write
     strategy:
       matrix:
-        rocm_version: ["71", "70"]
+        rocm_version: ["70", "64"]
     steps:
       - name: Checkout PyTorch
         uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2

.github/workflows/build-manywheel-images.yml

@@ -54,8 +54,8 @@ jobs:
           { name: "manylinuxaarch64-builder",       tag: "cuda12.9",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.8",          runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinuxaarch64-builder",       tag: "cuda12.6",          runner: "linux.arm64.2xlarge.ephemeral" },
+          { name: "manylinux2_28-builder",          tag: "rocm6.4",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "rocm7.0",           runner: "linux.9xlarge.ephemeral" },
-          { name: "manylinux2_28-builder",          tag: "rocm7.1",           runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "cpu",               runner: "linux.9xlarge.ephemeral" },
           { name: "manylinux2_28_aarch64-builder",  tag: "cpu-aarch64",       runner: "linux.arm64.2xlarge.ephemeral" },
           { name: "manylinux2_28-builder",          tag: "xpu",               runner: "linux.9xlarge.ephemeral" },

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

@@ -55,7 +55,7 @@ jobs:
         docker-image: ["pytorch/manylinux2_28-builder:cpu"]
         include:
           - device: "rocm"
-            rocm_version: "7.1"
+            rocm_version: "7.0"
             runs_on: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge"
           - device: "cuda"
             rocm_version: ""
@@ -159,7 +159,12 @@ jobs:
             WITH_CLANG_LDD="--with-clang-ldd"
           fi
 
-          docker exec -t "${container_name}" bash -c "${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE $WITH_CLANG_LDD"
+          if [[ "${BUILD_DEVICE}" == xpu ]]; then
+            docker exec -t "${container_name}" bash -c "dnf install -y gcc-toolset-13-gcc-c++"
+            docker exec -t "${container_name}" bash -c "source /opt/rh/gcc-toolset-13/enable && ${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE"
+          else
+            docker exec -t "${container_name}" bash -c "${PYTHON_EXECUTABLE} /pytorch/.github/scripts/build_triton_wheel.py --device=$BUILD_DEVICE $RELEASE $WITH_CLANG_LDD"
+          fi
 
           if [[ ("${{ matrix.device }}" == "cuda" || "${{ matrix.device }}" == "xpu") ]]; then
             docker exec -t "${container_name}"  bash -c "auditwheel repair --plat ${PLATFORM} //artifacts/*.whl"

.github/workflows/docker-builds.yml

@@ -57,7 +57,6 @@ jobs:
           pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
           pytorch-linux-jammy-py3.10-clang12,
           pytorch-linux-jammy-py3.13-clang12,
-          pytorch-linux-jammy-py3.14-clang12,
           pytorch-linux-jammy-rocm-n-py3,
           pytorch-linux-noble-rocm-n-py3,
           pytorch-linux-jammy-rocm-n-py3-benchmarks,
@@ -67,7 +66,6 @@ jobs:
           pytorch-linux-jammy-py3.12-halide,
           pytorch-linux-jammy-xpu-n-1-py3,
           pytorch-linux-jammy-xpu-n-py3,
-          pytorch-linux-jammy-xpu-n-py3-inductor-benchmarks,
           pytorch-linux-jammy-py3-clang18-asan,
           pytorch-linux-jammy-py3-clang12-onnx,
           pytorch-linux-jammy-linter,

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

@@ -132,7 +132,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -178,7 +178,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -224,7 +224,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -270,7 +270,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_10-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -381,7 +381,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -427,7 +427,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -473,7 +473,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -519,7 +519,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_11-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -630,7 +630,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -676,7 +676,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -722,7 +722,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -768,7 +768,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_12-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -879,7 +879,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -925,7 +925,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -971,7 +971,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1017,7 +1017,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1128,7 +1128,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1174,7 +1174,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1220,7 +1220,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1266,7 +1266,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_13t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1377,7 +1377,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1423,7 +1423,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1469,7 +1469,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1515,7 +1515,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1626,7 +1626,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_6
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1672,7 +1672,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_8
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1718,7 +1718,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-12_9
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
@@ -1764,7 +1764,7 @@ jobs:
       ALPINE_IMAGE: "arm64v8/alpine"
       build_name: manywheel-py3_14t-cuda-aarch64-13_0
       build_environment: linux-aarch64-binary-manywheel
-      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
+      PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
       timeout-minutes: 420
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}

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

@@ -384,6 +384,124 @@ jobs:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
 
+  libtorch-rocm6_4-shared-with-deps-release-build:
+    if: ${{ github.repository_owner == 'pytorch' }}
+    uses: ./.github/workflows/_binary-build-linux.yml
+    needs: get-label-type
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
+      timeout-minutes: 300
+      build_name: libtorch-rocm6_4-shared-with-deps-release
+      build_environment: linux-binary-libtorch
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+  libtorch-rocm6_4-shared-with-deps-release-test:  # Testing
+    if: ${{ github.repository_owner == 'pytorch' }}
+    needs:
+      - libtorch-rocm6_4-shared-with-deps-release-build
+      - get-label-type
+    runs-on: linux.rocm.gpu.mi250
+    timeout-minutes: 240
+    env:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      SKIP_ALL_TESTS: 1
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - name: Setup ROCm
+        uses: ./.github/actions/setup-rocm
+      - uses: actions/download-artifact@v4.1.7
+        name: Download Build Artifacts
+        with:
+          name: libtorch-rocm6_4-shared-with-deps-release
+          path: "${{ runner.temp }}/artifacts/"
+      - name: Checkout PyTorch
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          submodules: recursive
+          path: pytorch
+          show-progress: false
+      - name: Clean PyTorch checkout
+        run: |
+          # Remove any artifacts from the previous checkouts
+          git clean -fxd
+        working-directory: pytorch
+      - name: ROCm set GPU_FLAG
+        run: |
+          echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
+      - name: configure aws credentials
+        id: aws_creds
+        if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
+          aws-region: us-east-1
+          role-duration-seconds: 18000
+      - name: Calculate docker image
+        id: calculate-docker-image
+        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
+        with:
+          docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
+          docker-image-name: libtorch-cxx11-builder
+          custom-tag-prefix: rocm6.4
+          docker-build-dir: .ci/docker
+          working-directory: pytorch
+      - name: Pull Docker image
+        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
+        with:
+          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
+      - name: Test Pytorch binary
+        uses: ./pytorch/.github/actions/test-pytorch-binary
+        env:
+          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
+      - name: Teardown ROCm
+        uses: ./.github/actions/teardown-rocm
+  libtorch-rocm6_4-shared-with-deps-release-upload:  # Uploading
+    if: ${{ github.repository_owner == 'pytorch' }}
+    permissions:
+      id-token: write
+      contents: read
+    needs: libtorch-rocm6_4-shared-with-deps-release-test
+    with:
+      PYTORCH_ROOT: /pytorch
+      PACKAGE_TYPE: libtorch
+      # TODO: This is a legacy variable that we eventually want to get rid of in
+      #       favor of GPU_ARCH_VERSION
+      DESIRED_CUDA: rocm6.4
+      GPU_ARCH_VERSION: "6.4"
+      GPU_ARCH_TYPE: rocm
+      DOCKER_IMAGE: libtorch-cxx11-builder
+      DOCKER_IMAGE_TAG_PREFIX: rocm6.4
+      LIBTORCH_CONFIG: release
+      LIBTORCH_VARIANT: shared-with-deps
+      build_name: libtorch-rocm6_4-shared-with-deps-release
+    secrets:
+      github-token: ${{ secrets.GITHUB_TOKEN }}
+    uses: ./.github/workflows/_binary-upload.yml
+
   libtorch-rocm7_0-shared-with-deps-release-build:
     if: ${{ github.repository_owner == 'pytorch' }}
     uses: ./.github/workflows/_binary-build-linux.yml
@@ -501,121 +619,3 @@ jobs:
     secrets:
       github-token: ${{ secrets.GITHUB_TOKEN }}
     uses: ./.github/workflows/_binary-upload.yml
-
-  libtorch-rocm7_1-shared-with-deps-release-build:
-    if: ${{ github.repository_owner == 'pytorch' }}
-    uses: ./.github/workflows/_binary-build-linux.yml
-    needs: get-label-type
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
-      timeout-minutes: 300
-      build_name: libtorch-rocm7_1-shared-with-deps-release
-      build_environment: linux-binary-libtorch
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-  libtorch-rocm7_1-shared-with-deps-release-test:  # Testing
-    if: ${{ github.repository_owner == 'pytorch' }}
-    needs:
-      - libtorch-rocm7_1-shared-with-deps-release-build
-      - get-label-type
-    runs-on: linux.rocm.gpu.mi250
-    timeout-minutes: 240
-    env:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      SKIP_ALL_TESTS: 1
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-    permissions:
-      id-token: write
-      contents: read
-    steps:
-      - name: Setup ROCm
-        uses: ./.github/actions/setup-rocm
-      - uses: actions/download-artifact@v4.1.7
-        name: Download Build Artifacts
-        with:
-          name: libtorch-rocm7_1-shared-with-deps-release
-          path: "${{ runner.temp }}/artifacts/"
-      - name: Checkout PyTorch
-        uses: actions/checkout@v4
-        with:
-          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
-          submodules: recursive
-          path: pytorch
-          show-progress: false
-      - name: Clean PyTorch checkout
-        run: |
-          # Remove any artifacts from the previous checkouts
-          git clean -fxd
-        working-directory: pytorch
-      - name: ROCm set GPU_FLAG
-        run: |
-          echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
-      - name: configure aws credentials
-        id: aws_creds
-        if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
-        uses: aws-actions/configure-aws-credentials@v4
-        with:
-          role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
-          aws-region: us-east-1
-          role-duration-seconds: 18000
-      - name: Calculate docker image
-        id: calculate-docker-image
-        uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
-        with:
-          docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
-          docker-image-name: libtorch-cxx11-builder
-          custom-tag-prefix: rocm7.1
-          docker-build-dir: .ci/docker
-          working-directory: pytorch
-      - name: Pull Docker image
-        uses: pytorch/test-infra/.github/actions/pull-docker-image@main
-        with:
-          docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
-      - name: Test Pytorch binary
-        uses: ./pytorch/.github/actions/test-pytorch-binary
-        env:
-          DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
-      - name: Teardown ROCm
-        uses: ./.github/actions/teardown-rocm
-  libtorch-rocm7_1-shared-with-deps-release-upload:  # Uploading
-    if: ${{ github.repository_owner == 'pytorch' }}
-    permissions:
-      id-token: write
-      contents: read
-    needs: libtorch-rocm7_1-shared-with-deps-release-test
-    with:
-      PYTORCH_ROOT: /pytorch
-      PACKAGE_TYPE: libtorch
-      # TODO: This is a legacy variable that we eventually want to get rid of in
-      #       favor of GPU_ARCH_VERSION
-      DESIRED_CUDA: rocm7.1
-      GPU_ARCH_VERSION: "7.1"
-      GPU_ARCH_TYPE: rocm
-      DOCKER_IMAGE: libtorch-cxx11-builder
-      DOCKER_IMAGE_TAG_PREFIX: rocm7.1
-      LIBTORCH_CONFIG: release
-      LIBTORCH_VARIANT: shared-with-deps
-      build_name: libtorch-rocm7_1-shared-with-deps-release
-    secrets:
-      github-token: ${{ secrets.GITHUB_TOKEN }}
-    uses: ./.github/workflows/_binary-upload.yml

.github/workflows/inductor-perf-test-nightly-xpu.yml

@@ -1,148 +0,0 @@
-name: inductor-perf-nightly-xpu
-
-on:
-  push:
-    tags:
-      - ciflow/inductor-perf-test-nightly-xpu/*
-  schedule:
-    - cron: 30 17 * * *
-  workflow_dispatch:
-    inputs:
-      training:
-        description: Run training (on by default)?
-        required: false
-        type: boolean
-        default: true
-      inference:
-        description: Run inference (on by default)?
-        required: false
-        type: boolean
-        default: true
-      default:
-        description: Run inductor_default?
-        required: false
-        type: boolean
-        default: false
-      dynamic:
-        description: Run inductor_dynamic_shapes?
-        required: false
-        type: boolean
-        default: false
-      cppwrapper:
-        description: Run inductor_cpp_wrapper?
-        required: false
-        type: boolean
-        default: false
-      cudagraphs:
-        description: Run inductor_cudagraphs?
-        required: false
-        type: boolean
-        default: false
-      freezing_cudagraphs:
-        description: Run inductor_cudagraphs with freezing for inference?
-        required: false
-        type: boolean
-        default: false
-      aotinductor:
-        description: Run aot_inductor for inference?
-        required: false
-        type: boolean
-        default: false
-      maxautotune:
-        description: Run inductor_max_autotune?
-        required: false
-        type: boolean
-        default: false
-      benchmark_configs:
-        description: The list of configs used the benchmark
-        required: false
-        type: string
-        default: inductor_huggingface_perf,inductor_timm_perf,inductor_torchbench_perf,cachebench
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
-  cancel-in-progress: true
-
-permissions: read-all
-
-jobs:
-  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: ${{ github.ref_type }}
-      opt_out_experiments: lf
-
-  xpu-n-py3_10-inductor-benchmark-build:
-    name: xpu-n-py3.10-inductor-benchmark
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ 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-inductor-benchmarks
-      runner: linux.c7i.12xlarge
-      test-matrix: |
-        { include: [
-          { config: "inductor_huggingface_perf_xpu", shard: 1, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 2, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 3, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 4, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_huggingface_perf_xpu", shard: 5, num_shards: 5, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 2, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 3, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 4, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 5, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_timm_perf_xpu", shard: 6, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 1, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 2, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 3, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 4, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 5, num_shards: 6, runner: "linux.idc.xpu" },
-          { config: "inductor_torchbench_perf_xpu", shard: 6, num_shards: 6, runner: "linux.idc.xpu" },
-        ]}
-    secrets: inherit
-
-  xpu-n-py3_10-inductor-benchmark-test-nightly:
-    permissions:
-      id-token: write
-      contents: read
-    if: github.event_name != 'workflow_dispatch'
-    name: xpu-n-py3.10-inductor-benchmark
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: xpu-n-py3_10-inductor-benchmark-build
-    with:
-      build-environment: linux-jammy-xpu-n-py3.10
-      dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-false-cppwrapper-true-aotinductor-true-freezing_cudagraphs-false-cudagraphs_low_precision-false
-      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
-      timeout-minutes: 720
-      # Disable monitor in perf tests for more investigation
-      disable-monitor: true
-      monitor-log-interval: 10
-      monitor-data-collect-interval: 2
-    secrets: inherit
-
-  xpu-n-py3_10-inductor-benchmark-test:
-    permissions:
-      id-token: write
-      contents: read
-    if: github.event_name == 'workflow_dispatch'
-    name: xpu-n-py3.10-inductor-test
-    uses: ./.github/workflows/_xpu-test.yml
-    needs: xpu-n-py3_10-inductor-benchmark-build
-    with:
-      build-environment: linux-jammy-xpu-n-py3.10
-      dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cudagraphs-${{ inputs.cudagraphs }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-maxautotune-${{ inputs.maxautotune }}-freezing_cudagraphs-${{ inputs.freezing_cudagraphs }}-cudagraphs_low_precision-${{ inputs.cudagraphs }}
-      docker-image: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.docker-image }}
-      test-matrix: ${{ needs.xpu-n-py3_10-inductor-benchmark-build.outputs.test-matrix }}
-      timeout-minutes: 720
-      disable-monitor: false
-      monitor-log-interval: 15
-      monitor-data-collect-interval: 4
-    secrets: inherit

.github/workflows/periodic-rocm-mi200.yml

@@ -1,84 +0,0 @@
-name: periodic-rocm-mi200
-
-on:
-  schedule:
-    # We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
-    # Also run less frequently on weekends.
-    - cron: 45 0,8,16 * * 1-5
-    - cron: 45 4 * * 0,6
-    - cron: 45 4,12,20 * * 1-5
-    - cron: 45 12 * * 0,6
-    - cron: 29 8 * * *  # about 1:29am PDT, for mem leak check and rerun disabled tests
-  push:
-    tags:
-      - ciflow/periodic/*
-      - ciflow/periodic-rocm-mi200/*
-    branches:
-      - release/*
-  workflow_dispatch:
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}-${{ github.event.schedule }}
-  cancel-in-progress: true
-
-permissions:
-  id-token: write
-  contents: read
-
-jobs:
-  llm-td:
-    if: github.repository_owner == 'pytorch'
-    name: before-test
-    uses: ./.github/workflows/llm_td_retrieval.yml
-    permissions:
-      id-token: write
-      contents: read
-
-  target-determination:
-    name: before-test
-    uses: ./.github/workflows/target_determination.yml
-    needs: llm-td
-    permissions:
-      id-token: write
-      contents: read
-
-  get-label-type:
-    name: get-label-type
-    uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
-    if: (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch'
-    with:
-      triggering_actor: ${{ github.triggering_actor }}
-      issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
-      curr_branch: ${{ github.head_ref || github.ref_name }}
-      curr_ref_type: ${{ github.ref_type }}
-
-  linux-jammy-rocm-py3_10-build:
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_linux-build.yml
-    needs: get-label-type
-    with:
-      runner_prefix: "${{ 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
-      test-matrix: |
-        { include: [
-          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
-        ]}
-    secrets: inherit
-
-  linux-jammy-rocm-py3_10-test:
-    permissions:
-      id-token: write
-      contents: read
-    name: linux-jammy-rocm-py3.10
-    uses: ./.github/workflows/_rocm-test.yml
-    needs:
-      - linux-jammy-rocm-py3_10-build
-      - target-determination
-    with:
-      build-environment: linux-jammy-rocm-py3.10
-      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
-      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
-    secrets: inherit

.github/workflows/periodic.yml

@@ -204,6 +204,37 @@ jobs:
       test-matrix: ${{ needs.linux-jammy-cuda13_0-py3_10-gcc11-build.outputs.test-matrix }}
     secrets: inherit
 
+  linux-jammy-rocm-py3_10-build:
+    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
+      test-matrix: |
+        { include: [
+          { config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+          { config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
+        ]}
+    secrets: inherit
+
+  linux-jammy-rocm-py3_10-test:
+    permissions:
+      id-token: write
+      contents: read
+    name: linux-jammy-rocm-py3.10
+    uses: ./.github/workflows/_rocm-test.yml
+    needs:
+      - linux-jammy-rocm-py3_10-build
+      - target-determination
+    with:
+      build-environment: linux-jammy-rocm-py3.10
+      docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
+      test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
+    secrets: inherit
+
   linux-jammy-cuda12_8-py3-gcc11-slow-gradcheck-build:
     name: linux-jammy-cuda12.8-py3-gcc11-slow-gradcheck
     uses: ./.github/workflows/_linux-build.yml

.github/workflows/upload-test-stats.yml

@@ -6,7 +6,6 @@ on:
       - pull
       - trunk
       - periodic
-      - periodic-rocm-mi200
       - periodic-rocm-mi300
       - inductor
       - unstable

.github/workflows/xpu.yml

@@ -59,18 +59,14 @@ jobs:
       runner: linux.c7i.12xlarge
       test-matrix: |
         { include: [
-          { config: "default", shard: 1, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 2, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 3, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 4, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 5, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 6, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 7, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 8, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 9, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 10, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 11, num_shards: 12, runner: "linux.idc.xpu" },
-          { config: "default", shard: 12, num_shards: 12, runner: "linux.idc.xpu" },
+          { config: "default", shard: 1, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 2, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 3, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 4, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 5, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 6, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 7, num_shards: 8, runner: "linux.idc.xpu" },
+          { config: "default", shard: 8, num_shards: 8, runner: "linux.idc.xpu" },
         ]}
     secrets: inherit
 

.gitignore

@@ -143,7 +143,6 @@ scripts/release_notes/*.json
 sccache-stats*.json
 lint.json
 merge_record.json
-.github/scripts/nightly_source_matrix.json
 
 # These files get copied over on invoking setup.py
 torchgen/packaged/*

.lintrunner.toml

@@ -833,7 +833,8 @@ exclude_patterns = [
 command = [
     'python3',
     'tools/linter/adapters/grep_linter.py',
-    '--pattern=(cudaSetDevice|cudaGetDevice)\\(',
+    '--pattern=cudaSetDevice(',
+    '--pattern=cudaGetDevice(',
     '--linter-name=RAWCUDADEVICE',
     '--error-name=raw CUDA API usage',
     """--error-description=\

CMakeLists.txt

@@ -374,7 +374,7 @@ cmake_dependent_option(
   "Build the lazy Torchscript backend, not compatible with mobile builds" ON
   "NOT INTERN_BUILD_MOBILE" OFF)
 cmake_dependent_option(BUILD_FUNCTORCH "Build Functorch" ON "BUILD_PYTHON" OFF)
-cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin folder"
+cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
                        OFF "USE_CUDA" OFF)
 cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
                         "CPU_AARCH64" OFF)

README.md

@@ -1,4 +1,4 @@
-![PyTorch Logo](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/pytorch-logo-dark.png)
+![PyTorch Logo](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/pytorch-logo-dark.png)
 
 --------------------------------------------------------------------------------
 
@@ -72,7 +72,7 @@ Elaborating Further:
 
 If you use NumPy, then you have used Tensors (a.k.a. ndarray).
 
-![Tensor illustration](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/tensor_illustration.png)
+![Tensor illustration](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/tensor_illustration.png)
 
 PyTorch provides Tensors that can live either on the CPU or the GPU and accelerates the
 computation by a huge amount.
@@ -99,7 +99,7 @@ from several research papers on this topic, as well as current and past work suc
 While this technique is not unique to PyTorch, it's one of the fastest implementations of it to date.
 You get the best of speed and flexibility for your crazy research.
 
-![Dynamic graph](https://github.com/pytorch/pytorch/raw/main/docs/source/_static/img/dynamic_graph.gif)
+![Dynamic graph](https://github.com/pytorch/pytorch/blob/9708fcf92db88b80b9010c68662d634434da3106/docs/source/_static/img/dynamic_graph.gif)
 
 ### Python First
 

SECURITY.md

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

aten/src/ATen/CMakeLists.txt

@@ -38,7 +38,7 @@ set_bool(AT_HIPSPARSELT_ENABLED CAFFE2_USE_HIPSPARSELT)
 
 configure_file(Config.h.in "${CMAKE_CURRENT_SOURCE_DIR}/Config.h")
 # TODO: Do not generate CUDAConfig.h for ROCm BUILDS
-# At the moment, `jit_macros.h` include CUDAConfig.h for both CUDA and HIP builds
+# At the moment, `jit_macors.h` include CUDAConfig.h for both CUDA and HIP builds
 if(USE_CUDA OR USE_ROCM)
   configure_file(cuda/CUDAConfig.h.in "${CMAKE_CURRENT_SOURCE_DIR}/cuda/CUDAConfig.h")
 endif()
@@ -260,7 +260,7 @@ IF(USE_FBGEMM_GENAI)
   if(USE_CUDA)
     # To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
     # If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
-    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*(mx8mx8bf16_grouped|f4f4bf16_grouped).*")
+    set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
     file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
@@ -291,7 +291,6 @@ IF(USE_FBGEMM_GENAI)
 
     set(fbgemm_genai_cuh
       "${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
-      "${FBGEMM_GENAI_SRCS}/cutlass_extensions/f4f4bf16_grouped/"
       "${FBGEMM_GENAI_SRCS}/"
     )
 

aten/src/ATen/Context.cpp

@@ -825,14 +825,6 @@ void Context::setDisplayVmapFallbackWarnings(bool enabled) {
   display_vmap_fallback_warnings_ = enabled;
 }
 
-bool Context::warnOnAccumulateGradStreamMismatch() const {
-  return warn_on_accumulate_grad_stream_mismatch_;
-}
-
-void Context::setWarnOnAccumulateGradStreamMismatch(bool enabled) {
-  warn_on_accumulate_grad_stream_mismatch_ = enabled;
-}
-
 bool Context::isDefaultMobileCPUAllocatorSet() {
   return prev_allocator_ptr_ != nullptr;
 }

aten/src/ATen/Context.h

@@ -404,9 +404,6 @@ class TORCH_API Context {
   void setDisplayVmapFallbackWarnings(bool enabled);
   bool areVmapFallbackWarningsEnabled() const;
 
-  void setWarnOnAccumulateGradStreamMismatch(bool enabled);
-  bool warnOnAccumulateGradStreamMismatch() const;
-
   bool isDefaultMobileCPUAllocatorSet();
   void setDefaultMobileCPUAllocator();
   void unsetDefaultMobileCPUAllocator();
@@ -497,7 +494,6 @@ class TORCH_API Context {
   bool release_original_weights = false;
 #endif
   bool display_vmap_fallback_warnings_ = false;
-  bool warn_on_accumulate_grad_stream_mismatch_ = true;
   std::atomic<at::QEngine> quantized_engine = at::QEngine::NoQEngine;
   bool enable_sparse_tensor_invariant_checks = false;
   bool allow_fp16_reduction_cpu = false;

aten/src/ATen/FunctionalTensorWrapper.cpp

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

aten/src/ATen/TensorIterator.cpp

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

aten/src/ATen/core/CachingHostAllocator.h

@@ -94,11 +94,11 @@ struct PinnedReserveSegment {
 struct TORCH_API HostStats {
   // COUNT: total allocations (active)
   Stat active_requests;
-  // SUM: bytes allocated/reserved by this memory allocator. (active)
+  // SUM: bytes allocated/reserved by this memory alocator. (active)
   Stat active_bytes;
   // COUNT: total allocations (active + free)
   Stat allocations;
-  // SUM: bytes allocated/reserved by this memory allocator. This accounts
+  // SUM: bytes allocated/reserved by this memory alocator. This accounts
   // for both free and in-use blocks.
   Stat allocated_bytes;
 
@@ -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 allocator. This accounts
+  // SUM: bytes allocated/reserved by this memory alocator. This accounts
   // for both free and in-use blocks.
   Stat allocated_bytes;
   // COUNT: number of allocations per bucket (active)
@@ -455,7 +455,7 @@ struct CachingHostAllocatorImpl {
   }
 
   void resetAccumulatedStats() {
-    // Resetting accumulated memory stats requires concurrently holding both the
+    // Reseting accumulated memory stats requires concurrently holding both the
     // free list mutexes and the blocks mutex. Previously, this was only done in
     // empty_cache function.
     for (size_t i = 0; i < free_list_.size(); ++i) {
@@ -482,7 +482,7 @@ struct CachingHostAllocatorImpl {
   }
 
   void resetPeakStats() {
-    // Resetting peak memory stats requires concurrently holding both the
+    // Reseting peak memory stats requires concurrently holding both the
     // free list mutexes and the blocks mutex. Previously, this was only done in
     // empty_cache function.
     for (size_t i = 0; i < free_list_.size(); ++i) {
@@ -677,8 +677,8 @@ struct CachingHostAllocatorImpl {
   // size. This allows us to quickly find a free block of the right size.
   // We use deque to store per size free list and guard the list with its own
   // mutex.
-  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>>
-      free_list_{MAX_SIZE_INDEX};
+  alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>> free_list_ =
+      std::vector<FreeBlockList<B>>(MAX_SIZE_INDEX);
 
   alignas(hardware_destructive_interference_size) std::mutex events_mutex_;
   std::deque<std::pair<E, B*>> events_; // event queue paired with block

aten/src/ATen/core/class_type.h

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

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

@@ -537,7 +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 number two different
+  // This means that we might collect the same sequence nubmer two different
   // events if they all occurred above Autograd and still had the Autograd
   // dispatch key in the dispatch key set.
   // However, this usually doesn't happen: normally the first call will

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

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

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

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

aten/src/ATen/core/function_schema.cpp

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

aten/src/ATen/core/ivalue.h

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

aten/src/ATen/core/ivalue_inl.h

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

aten/src/ATen/core/jit_type.h

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

aten/src/ATen/core/jit_type_base.h

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

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

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

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

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

aten/src/ATen/core/tensor_type.cpp

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

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

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

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

@@ -19,13 +19,6 @@ inline namespace CPU_CAPABILITY {
 #error "Big endian is not supported."
 #endif
 
-// GCC does not properly optimize bf16 operators
-#if defined(__ARM_FEATURE_BF16) && (__clang_major__ >= 19)
-#define BF16_ARITHMETIC_SUPPORTED() 1
-#else
-#define BF16_ARITHMETIC_SUPPORTED() 0
-#endif
-
 // Unlike the float16_t family of types, bfloat16_t is not available
 // when we're not targeting bfloat16 hardware support on some
 // platforms (but not Mac, so we have to be careful not to shadow the
@@ -359,35 +352,18 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
         other, &Vectorized<float>::name);                        \
   }
 
+  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
   Vectorized frac() const;
   DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(trunc)
   DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(sqrt)
 
 #ifdef __ARM_FEATURE_BF16
-  // Flip sign bit
   Vectorized<c10::BFloat16> neg() const {
-    return vreinterpretq_bf16_s16(vreinterpretq_s16_bf16(values) ^ (-32768));
+    return -values;
   }
-  // Fast reciprocal is fine because we are truncating results
   Vectorized<c10::BFloat16> reciprocal() const {
-    auto x = vcvtq_low_f32_bf16(values);
-    auto y = vcvtq_high_f32_bf16(values);
-    x = vrecpeq_f32(x);
-    y = vrecpeq_f32(y);
-    return vcvtq_high_bf16_f32(vcvtq_low_bf16_f32(x), y);
+    return 1.0f / values;
   }
-  // Clearing the sign bit
-  Vectorized<c10::BFloat16> abs() const {
-    return vreinterpretq_bf16_u16(vreinterpretq_u16_bf16(values) & 0x7FFF);
-  }
-#else
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
-  DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(reciprocal)
-#endif
-
-// These functions are optimized on clang-21+
-#if BF16_ARITHMETIC_SUPPORTED() && (__clang_major__ >= 21)
   Vectorized<c10::BFloat16> operator==(
       const Vectorized<c10::BFloat16>& other) const {
     return values == other.values;
@@ -418,6 +394,8 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
     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!=)
   DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator<)
@@ -473,7 +451,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator+(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   return x + y;
@@ -486,7 +464,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator-(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   return x - y;
@@ -499,7 +477,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator*(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   return x * y;
@@ -512,7 +490,7 @@ template <>
 Vectorized<c10::BFloat16> inline operator/(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   return x / y;
@@ -629,7 +607,7 @@ Vectorized<c10::BFloat16> inline fmadd(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   bfloat16x8_t z = c;
@@ -649,7 +627,7 @@ Vectorized<c10::BFloat16> inline fnmadd(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   bfloat16x8_t z = c;
@@ -665,7 +643,7 @@ Vectorized<c10::BFloat16> inline fmsub(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   bfloat16x8_t z = c;
@@ -681,7 +659,7 @@ Vectorized<c10::BFloat16> inline fnmsub(
     const Vectorized<c10::BFloat16>& a,
     const Vectorized<c10::BFloat16>& b,
     const Vectorized<c10::BFloat16>& c) {
-#if BF16_ARITHMETIC_SUPPORTED()
+#ifdef __ARM_FEATURE_BF16
   bfloat16x8_t x = a;
   bfloat16x8_t y = b;
   bfloat16x8_t z = c;

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

@@ -5,250 +5,6 @@
 namespace at::vec {
 inline namespace CPU_CAPABILITY {
 #if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
-
-// Enable auto-vectorization for clang-17+
-// GCC-12 has a bug: gcc.gnu.org/bugzilla/show_bug.cgi?id=117001
-#if 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]);
-  }
-}
-
-template <typename to_type>
-inline void convertFromBool(
-    const bool* __restrict src,
-    to_type* __restrict dst,
-    int64_t n) {
-  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    dst[i] = srcPtr[i] != 0 ? static_cast<to_type>(1) : static_cast<to_type>(0);
-  }
-}
-
-template <typename from_type>
-inline void convertToBool(
-    const from_type* __restrict src,
-    bool* __restrict dst,
-    int64_t n) {
-  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    dstPtr[i] = src[i] != static_cast<from_type>(0) ? 1 : 0;
-  }
-}
-
-#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);               \
-  }
-
-#define CONVERT_FROM_BOOL_TEMPLATE(to_type)                       \
-  inline void convert(const bool* src, to_type* dst, int64_t n) { \
-    return convertFromBool<to_type>(src, dst, n);                 \
-  }
-
-#define CONVERT_TO_BOOL_TEMPLATE(from_type)                         \
-  inline void convert(const from_type* src, bool* dst, int64_t n) { \
-    return convertToBool<from_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_TO_BOOL_TEMPLATE(uint8_t)
-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_TO_BOOL_TEMPLATE(int8_t)
-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_TO_BOOL_TEMPLATE(int16_t)
-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_TO_BOOL_TEMPLATE(int32_t)
-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_TO_BOOL_TEMPLATE(int64_t)
-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_TO_BOOL_TEMPLATE(float)
-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)
-CONVERT_TO_BOOL_TEMPLATE(double)
-CONVERT_FROM_BOOL_TEMPLATE(uint8_t)
-CONVERT_FROM_BOOL_TEMPLATE(int8_t)
-CONVERT_FROM_BOOL_TEMPLATE(int16_t)
-CONVERT_FROM_BOOL_TEMPLATE(int32_t)
-CONVERT_FROM_BOOL_TEMPLATE(int64_t)
-CONVERT_FROM_BOOL_TEMPLATE(float)
-CONVERT_FROM_BOOL_TEMPLATE(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)
-
-inline void convertBoolToFp16Impl(
-    const bool* __restrict src,
-    at::Half* __restrict dst,
-    int64_t n) {
-  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
-  float16_t* dstPtr = reinterpret_cast<float16_t*>(dst);
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    dstPtr[i] = srcPtr[i] != 0 ? 1.0 : 0;
-  }
-}
-
-template <>
-inline void convert(const bool* src, at::Half* dst, int64_t n) {
-  return convertBoolToFp16Impl(src, dst, n);
-}
-
-inline void convertFp16ToBoolImpl(
-    const at::Half* __restrict src,
-    bool* __restrict dst,
-    int64_t n) {
-  const float16_t* srcPtr = reinterpret_cast<const float16_t*>(src);
-  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    dstPtr[i] = srcPtr[i] != 0.0 ? 1 : 0;
-  }
-}
-
-template <>
-inline void convert(const at::Half* src, bool* dst, int64_t n) {
-  return convertFp16ToBoolImpl(src, dst, n);
-}
-
-#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)
-
-inline void convertBoolToBfloat16Impl(
-    const bool* __restrict src,
-    c10::BFloat16* __restrict dst,
-    int64_t n) {
-  const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
-  uint16_t* dstPtr = reinterpret_cast<uint16_t*>(dst);
-  uint64_t len = static_cast<uint64_t>(n);
-  constexpr uint16_t kBf16One = 0x3f80; // 1.0 in bfloat16
-  for (uint64_t i = 0; i < len; i++) {
-    dstPtr[i] = srcPtr[i] != 0 ? kBf16One : 0;
-  }
-}
-
-template <>
-inline void convert(const bool* src, c10::BFloat16* dst, int64_t n) {
-  return convertBoolToBfloat16Impl(src, dst, n);
-}
-
-inline void convertBfloat16ToBoolImpl(
-    const c10::BFloat16* __restrict src,
-    bool* __restrict dst,
-    int64_t n) {
-  uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
-  const uint16_t* srcPtr = reinterpret_cast<const uint16_t*>(src);
-  uint64_t len = static_cast<uint64_t>(n);
-  for (uint64_t i = 0; i < len; i++) {
-    // Check if all non-sign bits are 0
-    bool isBf16Zero = (srcPtr[i] & 0x7fff) == 0;
-    dstPtr[i] = isBf16Zero ? 0 : 1;
-  }
-}
-
-template <>
-inline void convert(const c10::BFloat16* src, bool* dst, int64_t n) {
-  return convertBfloat16ToBoolImpl(src, dst, n);
-}
-
-#endif
-
-#endif
-
 template <typename src_t>
 struct VecConvert<
     float,

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

@@ -307,52 +307,11 @@ class Vectorized<float> {
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(exp)
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(exp2)
   DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(expm1)
-  // Implementation copied from Arm Optimized Routine
-  // https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/expf.c
-  inline Vectorized<float> vexpq_f32_u20() const {
-    // bail out to sleef if it's a special case:
-    // i.e. there's an input s.t. |input| > 87.3....
-    const float32x4_t special_bound = vdupq_n_f32(0x1.5d5e2ap+6f);
-    uint32x4_t cmp = vcagtq_f32(values, special_bound);
-    if (vpaddd_u64(vreinterpretq_u64_u32(cmp)) != 0) {
-      return exp();
-    }
-
-    const float32x4_t inv_ln2 = vdupq_n_f32(0x1.715476p+0f);
-    const float ln2_hi = 0x1.62e4p-1f;
-    const float ln2_lo = 0x1.7f7d1cp-20f;
-    const float c0 = 0x1.0e4020p-7f;
-    const float c2 = 0x1.555e66p-3f;
-    const float32x4_t ln2_c02 = {ln2_hi, ln2_lo, c0, c2};
-
-    const uint32x4_t exponent_bias = vdupq_n_u32(0x3f800000);
-    const float32x4_t c1 = vdupq_n_f32(0x1.573e2ep-5f);
-    const float32x4_t c3 = vdupq_n_f32(0x1.fffdb6p-2f);
-    const float32x4_t c4 = vdupq_n_f32(0x1.ffffecp-1f);
-
-    /* exp(x) = 2^n (1 + poly(r)), with 1 + poly(r) in [1/sqrt(2),sqrt(2)]
-      x = ln2*n + r, with r in [-ln2/2, ln2/2].  */
-
-    float32x4_t n = vrndaq_f32(vmulq_f32(values, inv_ln2));
-    float32x4_t r = vfmsq_laneq_f32(values, n, ln2_c02, 0);
-    r = vfmsq_laneq_f32(r, n, ln2_c02, 1);
-    uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_s32(vcvtq_s32_f32(n)), 23);
-    float32x4_t scale = vreinterpretq_f32_u32(vaddq_u32(e, exponent_bias));
-
-    float32x4_t r2 = vmulq_f32(r, r);
-    float32x4_t p = vfmaq_laneq_f32(c1, r, ln2_c02, 2);
-    float32x4_t q = vfmaq_laneq_f32(c3, r, ln2_c02, 3);
-    q = vfmaq_f32(q, p, r2);
-    p = vmulq_f32(c4, r);
-    float32x4_t poly = vfmaq_f32(p, q, r2);
-
-    return vfmaq_f32(scale, poly, scale);
-  }
   Vectorized<float> exp_u20() const {
-    return vexpq_f32_u20();
+    return exp();
   }
   Vectorized<float> fexp_u20() const {
-    return exp_u20();
+    return exp();
   }
   DEFINE_SLEEF_COMPATIBLE_BINARY_ELEMENTWISE_FUNC_WITH_SLEEF_NAME(
       fmod,
@@ -581,6 +540,42 @@ inline Vectorized<float> Vectorized<float>::le(
   return (*this <= other) & Vectorized<float>(1.0f);
 }
 
+template <>
+inline void convert(const float* src, int32_t* dst, int64_t n) {
+  int64_t i;
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (i = 0; i <= (n - Vectorized<float>::size());
+       i += Vectorized<float>::size()) {
+    vst1q_s32(dst + i, vcvtq_s32_f32(vld1q_f32(src + i)));
+  }
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (; i < n; i++) {
+    dst[i] = static_cast<int32_t>(src[i]);
+  }
+}
+
+template <>
+inline void convert(const int32_t* src, float* dst, int64_t n) {
+  int64_t i;
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (i = 0; i <= (n - Vectorized<float>::size());
+       i += Vectorized<float>::size()) {
+    vst1q_f32(dst + i, vcvtq_f32_s32(vld1q_s32(src + i)));
+  }
+#ifndef __msvc_cl__
+#pragma unroll
+#endif
+  for (; i < n; i++) {
+    dst[i] = static_cast<float>(src[i]);
+  }
+}
+
 template <>
 Vectorized<float> inline fmadd(
     const Vectorized<float>& a,
@@ -637,7 +632,8 @@ inline Vectorized<float> Vectorized<float>::erf() const {
   // - exp(- x * x)
   auto pow_2 = (*this) * (*this);
   auto neg_pow_2 = pow_2 ^ neg_zero_vec;
-  auto tmp4 = neg_pow_2.vexpq_f32_u20();
+  auto tmp4 = neg_pow_2.map(
+      std::exp); // This can be swapped for a faster implementation of exp.
   auto tmp5 = tmp4 ^ neg_zero_vec;
   // erf(x) = sign(x) * (1 - r * t * exp(- x * x))
   auto tmp6 = t * tmp5;

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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- incurring a
+    // return value back to T in the case of unary operator- incuring 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 manipulations
+  // Note: see `cublasCommonArgs` for various non-intuitive manupulations
   // of input arguments to this function.
   const auto computeType = CUBLAS_COMPUTE_32F;
   const auto scaleType = CUDA_R_32F;

aten/src/ATen/cuda/CUDAEvent.h

@@ -2,10 +2,10 @@
 
 #include <ATen/cuda/ATenCUDAGeneral.h>
 #include <ATen/cuda/CUDAContext.h>
-#include <ATen/cuda/Exceptions.h>
 #include <c10/core/impl/GPUTrace.h>
-#include <c10/cuda/CUDAGuard.h>
 #include <c10/cuda/CUDAStream.h>
+#include <c10/cuda/CUDAGuard.h>
+#include <ATen/cuda/Exceptions.h>
 #include <c10/util/Exception.h>
 
 #include <cuda_runtime_api.h>
@@ -246,79 +246,4 @@ private:
   }
 };
 
-// EventPool - Thread-safe pool of CUDA events to avoid expensive cudaEventCreate
-// calls. cudaEventCreate when concurrently invoked from multiple threads can be
-// very expensive (especially on certain device/driver combinations).
-using CUDAEventPtr =
-    std::unique_ptr<CUDAEvent, std::function<void(CUDAEvent*)>>;
-
-class EventPool {
- public:
-  EventPool() : pools_(at::cuda::device_count()) {}
-
-  CUDAEventPtr get(const DeviceIndex device) {
-    // If the device is invalid, return a default event and no pooling
-    if (device < 0 || device >= (DeviceIndex)pools_.size()) {
-      auto deleter = [](CUDAEvent* event) {
-        delete event;
-      };
-      return CUDAEventPtr(
-        std::make_unique<CUDAEvent>(cudaEventDisableTiming).release(), deleter);
-    }
-
-    auto& pool = pools_[device];
-
-    // Create a destructor that returns the event to the appropriate device pool
-    auto destructor = [&pool](CUDAEvent* event) noexcept {
-      if (event != nullptr) {
-        std::lock_guard<std::mutex> lock(pool.mutex_);
-        pool.event_pool_.emplace_back(event);
-      }
-    };
-
-    {
-      std::lock_guard<std::mutex> lock(pool.mutex_);
-      if (!pool.event_pool_.empty()) {
-        auto event = std::move(pool.event_pool_.back());
-        pool.event_pool_.pop_back();
-        return CUDAEventPtr(event.release(), destructor);
-      }
-    }
-
-    return CUDAEventPtr(
-        std::make_unique<CUDAEvent>(cudaEventDisableTiming).release(),
-        destructor);
-  }
-
-  void empty_cache() {
-    for (auto& pool : pools_) {
-      std::lock_guard<std::mutex> lock(pool.mutex_);
-      pool.event_pool_.clear();
-    }
-  }
-
-  void init_num_events(const size_t num_events) {
-    for (DeviceIndex device_idx = 0; device_idx < at::cuda::device_count(); ++device_idx) {
-        CUDAGuard device_guard(device_idx);
-        std::vector<CUDAEventPtr> temp_events;
-        temp_events.reserve(num_events);
-        for (size_t i = 0; i < num_events; ++i) {
-          auto event = get(device_idx);
-          // Record the event to ensure it's properly initialized
-          event->record();
-          temp_events.emplace_back(std::move(event));
-        }
-        // Events will be returned to pool when temp_events is destroyed
-    }
-  }
-
- private:
-  struct alignas(64) PerDevicePool {
-    alignas(64) std::mutex mutex_;
-    std::vector<std::unique_ptr<CUDAEvent>> event_pool_;
-  };
-
-  std::vector<PerDevicePool> pools_;
-};
-
 } // namespace at::cuda

aten/src/ATen/cuda/CUDAGraph.cpp

@@ -168,9 +168,11 @@ 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)
@@ -181,11 +183,13 @@ 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;
 }
 
@@ -307,7 +311,7 @@ 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 operations finish before deleting the object.
+// We need to ensure all async opreations 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

@@ -1,90 +1,78 @@
 #include <ATen/cuda/CUDAGreenContext.h>
 
-#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
-#include <c10/cuda/driver_api.h>
-#include <stdexcept>
-#include <vector>
-#define HAS_CUDA_GREEN_CONTEXT() 1
-#else
-#define HAS_CUDA_GREEN_CONTEXT() 0
-// Suppress unsued private field warnings as this class is not supposed to be called
-C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-private-field")
-#endif
-
 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...");
 
-GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
-#if HAS_CUDA_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);
+    }
 
-    cudaFree(0);
-  }
+    CUdevice device;
+    device_id_ = device_id;
+    C10_CUDA_DRIVER_CHECK(
+        c10::cuda::DriverAPI::get()->cuDeviceGet_(&device, device_id));
 
-   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));
 
-  // 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;
 
-  // 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));
 
-  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");
 
-  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));
 
-  // 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));
 
-  // 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!");
+    // 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+!");
+    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 HAS_CUDA_GREEN_CONTEXT()
+#if CUDA_HAS_GREEN_CONTEXT
     if (!device_id.has_value()) {
       device_id = at::cuda::current_device();
     }
-    return std::unique_ptr<GreenContext>(new GreenContext(device_id.value(), num_sms));
+    return std::make_unique<GreenContext>(device_id.value(), num_sms);
 #else
     TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
 #endif
@@ -92,7 +80,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
 
   // Implement move operations
   GreenContext::GreenContext(GreenContext&& other) noexcept{
-#if HAS_CUDA_GREEN_CONTEXT()
+#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);
@@ -103,7 +91,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
   }
 
   GreenContext& GreenContext::operator=(GreenContext&& other) noexcept{
-#if HAS_CUDA_GREEN_CONTEXT()
+#if CUDA_HAS_GREEN_CONTEXT
     if (this != &other) {
       // Clean up current resources
       if (green_ctx_) {
@@ -132,7 +120,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
   }
 
   GreenContext::~GreenContext() noexcept{
-#if HAS_CUDA_GREEN_CONTEXT()
+#if CUDA_HAS_GREEN_CONTEXT
     C10_CUDA_DRIVER_CHECK(
         c10::cuda::DriverAPI::get()->cuGreenCtxDestroy_(green_ctx_));
 #else
@@ -140,9 +128,25 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
 #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 HAS_CUDA_GREEN_CONTEXT()
+#if CUDA_HAS_GREEN_CONTEXT
     auto current_stream = c10::cuda::getCurrentCUDAStream();
     parent_stream_ = current_stream.stream();
 
@@ -171,7 +175,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
   }
 
   void GreenContext::popContext() {
-#if HAS_CUDA_GREEN_CONTEXT()
+#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());

aten/src/ATen/cuda/CUDAGreenContext.h

@@ -1,38 +1,53 @@
 #pragma once
 #include <ATen/cuda/CUDAEvent.h>
-#include <cuda.h>
 
-// Forward declare green context as opaque ptr
-typedef struct CUgreenCtx_st* CUgreenCtx;
+#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:
-  // Green context creation
-  static std::unique_ptr<GreenContext> create(
-      uint32_t num_sms,
-      std::optional<uint32_t> device_id);
-  ~GreenContext() noexcept;
+  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:
-  GreenContext(uint32_t device_id, uint32_t num_sms);
-  // Implement move operations
-  GreenContext(GreenContext&& other) noexcept;
-  GreenContext& operator=(GreenContext&& other) noexcept;
-
+#if CUDA_HAS_GREEN_CONTEXT
   int32_t device_id_ = -1;
   CUgreenCtx green_ctx_ = nullptr;
   CUcontext context_ = nullptr;
   cudaStream_t parent_stream_ = nullptr;
+#endif
 };
 } // namespace at::cuda

aten/src/ATen/cuda/CUDAScaledBlas.cpp

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

aten/src/ATen/cuda/CUDAScaledBlas.h

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

aten/src/ATen/cuda/CUDASparse.h

@@ -7,6 +7,17 @@
 #endif
 
 
+#if defined(USE_ROCM)
+// hipSparse const API added in v2.4.0
+#if HIPSPARSE_VERSION >= 200400
+#define AT_USE_HIPSPARSE_GENERIC_API() 1
+#else
+#define AT_USE_HIPSPARSE_GENERIC_API() 1
+#endif
+#else // USE_ROCM
+#define AT_USE_HIPSPARSE_GENERIC_API() 0
+#endif // USE_ROCM
+
 // cuSparse Generic API spsv function was added in CUDA 11.3.0
 #if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && (CUSPARSE_VERSION >= 11500)
 #define AT_USE_CUSPARSE_GENERIC_SPSV() 1

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 corresponding
+    // However, allocations using cudaHostRegister should use corresonding
     // cudaHostUnregister and similarly for cudaHostAlloc / cudaFreeHost.
     void* ptr = block->ptr_;
     bool use_register = false;

aten/src/ATen/cuda/Sleep.cu

@@ -1,7 +1,6 @@
 #include <ATen/cuda/CUDAContextLight.h>
 #include <ATen/cuda/Sleep.h>
 
-#include <c10/cuda/CUDACachingAllocator.h>
 #include <c10/cuda/CUDAException.h>
 #include <c10/cuda/CUDAStream.h>
 
@@ -25,22 +24,8 @@ __global__ void spin_kernel(int64_t cycles) {
 #endif
   }
 }
-
-thread_local int *flag = nullptr;
-
-__global__ void busy_wait_for_flag_kernel(int *flag) {
-  atomicExch(flag, 1);
-  while (atomicAdd(flag, 0) == 1) {
-    // do nothing
-  }
 }
 
-__global__ void clear_flag_kernel(int *flag) {
-  atomicExch(flag, 0);
-}
-
-} // anonymous namespace
-
 void sleep(int64_t cycles) {
   dim3 grid(1);
   dim3 block(1);
@@ -48,26 +33,6 @@ void sleep(int64_t cycles) {
   C10_CUDA_KERNEL_LAUNCH_CHECK();
 }
 
-void busy_wait_for_flag() {
-  if (!flag) {
-    flag = (int*)c10::cuda::CUDACachingAllocator::raw_alloc(sizeof(int));
-  }
-  dim3 grid(1);
-  dim3 block(1);
-  busy_wait_for_flag_kernel<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>(flag);
-  C10_CUDA_KERNEL_LAUNCH_CHECK();
-}
-
-void clear_flag() {
-  if (!flag) {
-    flag = (int*)c10::cuda::CUDACachingAllocator::raw_alloc(sizeof(int));
-  }
-  dim3 grid(1);
-  dim3 block(1);
-  clear_flag_kernel<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>(flag);
-  C10_CUDA_KERNEL_LAUNCH_CHECK();
-}
-
 #ifdef USE_ROCM
 __global__ void flush_icache_kernel()
 {

aten/src/ATen/cuda/Sleep.h

@@ -7,11 +7,6 @@ namespace at::cuda {
 // enqueues a kernel that spins for the specified number of cycles
 TORCH_CUDA_CU_API void sleep(int64_t cycles);
 
-// enqueues a kernel that spins until a flag is cleared by a
-// corresponding call to clear_flag()
-TORCH_CUDA_CU_API void busy_wait_for_flag();
-TORCH_CUDA_CU_API void clear_flag();
-
 // flushes instruction cache for ROCm; no-op for CUDA
 TORCH_CUDA_CU_API void flush_icache();
 

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 avoids re-compiling them for each translation unit. If you get
+// which aviods 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/detail/BLASConstants.cu

@@ -2,6 +2,8 @@
 #include <ATen/Tensor.h>
 #include <ATen/cuda/Exceptions.h>
 
+#include <mutex>
+
 namespace at {
 namespace cuda {
 namespace detail {
@@ -10,36 +12,39 @@ __device__ __constant__ float cublas_one_device;
 __device__ __constant__ float cublas_zero_device;
 
 float *get_cublas_device_one() {
-  static float *ptr = nullptr;
-  static auto init_flag = [&]() {
+  static c10::once_flag init_flag;
+
+  c10::call_once(init_flag, []() {
     const float one = 1.f;
     AT_CUDA_CHECK(cudaMemcpyToSymbol(cublas_one_device, &one, sizeof(float)));
-    AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_one_device));
-    return true;
-  }();
+  });
 
+  float *ptr;
+  AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_one_device));
   return ptr;
 }
 
 float *get_cublas_device_zero() {
-  static float *ptr = nullptr;
-  static auto init_flag = [&]() {
+  static c10::once_flag init_flag;
+
+  c10::call_once(init_flag, []() {
     const float zero = 0.f;
     AT_CUDA_CHECK(cudaMemcpyToSymbol(cublas_zero_device, &zero, sizeof(float)));
-    AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_zero_device));
-    return true;
-  }();
+  });
 
+  float *ptr;
+  AT_CUDA_CHECK(cudaGetSymbolAddress(reinterpret_cast<void**>(&ptr), cublas_zero_device));
   return ptr;
 }
 
 float *get_user_alpha_ptr() {
   static float *alpha_ptr;
 
-  static bool init_flag [[maybe_unused]] = []() {
+  static c10::once_flag init_flag;
+
+  c10::call_once(init_flag, []() {
     AT_CUDA_CHECK(cudaMalloc(&alpha_ptr, sizeof(float)));
-    return true;
-  }();
+  });
 
   return alpha_ptr;
 }

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 version, or ROCm version, or PyTorch version, TunableOp will detect
+Note the "Validator" lines. If you change a library verison, 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/Tunable.cpp

@@ -580,7 +580,7 @@ std::ofstream& TuningContext::GetUntunedFile(){
       filename.append(device);
     }
 
-    untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::app);
+    untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::trunc);
   }
   return untuned_file_;
 }

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();
 
-      // calculate a reference answer for numerical check
+      // calcaulte 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 interface is hook-based as it corresponds to all the functions
+// This inferface 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/MTIAHooksInterface.h

@@ -1,6 +1,5 @@
 #pragma once
 
-#include <c10/core/CachingDeviceAllocator.h>
 #include <c10/core/Device.h>
 #include <c10/util/Exception.h>
 
@@ -152,36 +151,6 @@ struct TORCH_API MTIAHooksInterface : AcceleratorHooksInterface {
   }
 
   virtual bool isAvailable() const override;
-
-  /* MTIAGraph related APIs */
-  virtual int64_t mtiagraphCreate(bool keep_graph = false) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-    return -1;
-  }
-
-  virtual void mtiagraphCaptureBegin(int64_t handle, MempoolId_t pool) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphCaptureEnd(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphInstantiate(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphReplay(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual void mtiagraphReset(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
-
-  virtual MempoolId_t mtiagraphPool(int64_t handle) const {
-    FAIL_MTIAHOOKS_FUNC(__func__);
-  }
 };
 
 struct TORCH_API MTIAHooksArgs {};

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): Preserved for BC and will be removed in the future.
+    // TODO(FFFrog): Perserved for BC and will be removed in the future.
     if (at::GetGeneratorPrivate().has_value())
       return at::GetGeneratorForPrivateuse1(device_index);
 

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, please try
+// This is not the most efficient thing; if there are alternatives, plese 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 behavior
+  // A as BatchedTensor since it might have been saved by autograd (specifically by the jvp) and the autograd behvaior
   // 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(). Something must
+  // Not sure why output_ needs to be marked as .contiguous(). Someting 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/BatchRulesViews.cpp

@@ -534,20 +534,20 @@ Tensor trace_decomp(const Tensor& tensor) {
 std::tuple<Tensor, std::optional<int64_t>> tril_batch_rule(
     const Tensor& self,
     std::optional<int64_t> self_bdim,
-    c10::SymInt diagonal = 0) {
+    int64_t diagonal = 0) {
   TORCH_CHECK(self.dim() >= 2, "tril: The input tensor must have at least 2 dimensions.");
   auto self_ = moveBatchDimToFront(self, self_bdim);
-  auto result = at::tril_symint(self_, std::move(diagonal));
+  auto result = at::tril(self_, diagonal);
   return std::make_tuple(std::move(result), 0);
 }
 
 std::tuple<Tensor, std::optional<int64_t>> triu_batch_rule(
     const Tensor& self,
     std::optional<int64_t> self_bdim,
-    c10::SymInt diagonal = 0) {
+    int64_t diagonal = 0) {
   TORCH_CHECK(self.dim() >= 2, "triu: The input tensor must have at least 2 dimensions.");
   auto self_ = moveBatchDimToFront(self, self_bdim);
-  auto result = at::triu_symint(self_, std::move(diagonal));
+  auto result = at::triu(self_, diagonal);
   return std::make_tuple(std::move(result), 0);
 }