[ { "path": ".clang-format", "content": "---\nAccessModifierOffset: -1\nAlignAfterOpenBracket: AlwaysBreak\nAlignConsecutiveAssignments: false\nAlignConsecutiveDeclarations: false\nAlignEscapedNewlinesLeft: true\nAlignOperands: false\nAlignTrailingComments: false\nAllowAllParametersOfDeclarationOnNextLine: false\nAllowShortBlocksOnASingleLine: false\nAllowShortCaseLabelsOnASingleLine: false\nAllowShortFunctionsOnASingleLine: Empty\nAllowShortIfStatementsOnASingleLine: false\nAllowShortLoopsOnASingleLine: false\nAlwaysBreakAfterReturnType: None\nAlwaysBreakBeforeMultilineStrings: true\nAlwaysBreakTemplateDeclarations: true\nBinPackArguments: false\nBinPackParameters: false\nBraceWrapping:\n AfterClass: false\n AfterControlStatement: false\n AfterEnum: false\n AfterFunction: false\n AfterNamespace: false\n AfterObjCDeclaration: false\n AfterStruct: false\n AfterUnion: false\n BeforeCatch: false\n BeforeElse: false\n IndentBraces: false\nBreakBeforeBinaryOperators: None\nBreakBeforeBraces: Attach\nBreakBeforeTernaryOperators: true\nBreakConstructorInitializersBeforeComma: false\nBreakAfterJavaFieldAnnotations: false\nBreakStringLiterals: false\nColumnLimit: 80\nCommentPragmas: '^ IWYU pragma:'\n#CompactNamespaces: false\nConstructorInitializerAllOnOneLineOrOnePerLine: true\nConstructorInitializerIndentWidth: 4\nContinuationIndentWidth: 4\nCpp11BracedListStyle: true\nDerivePointerAlignment: false\nDisableFormat: false\nForEachMacros: [ FOR_EACH_RANGE, FOR_EACH, ]\nIncludeCategories:\n - Regex: '^<.*\\.h(pp)?>'\n Priority: 1\n - Regex: '^<.*'\n Priority: 2\n - Regex: '.*'\n Priority: 3\nIndentCaseLabels: true\nIndentWidth: 2\nIndentWrappedFunctionNames: false\nKeepEmptyLinesAtTheStartOfBlocks: false\nMacroBlockBegin: ''\nMacroBlockEnd: ''\nMaxEmptyLinesToKeep: 1\nNamespaceIndentation: None\nObjCBlockIndentWidth: 2\nObjCSpaceAfterProperty: false\nObjCSpaceBeforeProtocolList: false\nPenaltyBreakBeforeFirstCallParameter: 1\nPenaltyBreakComment: 300\nPenaltyBreakFirstLessLess: 120\nPenaltyBreakString: 1000\nPenaltyExcessCharacter: 1000000\nPenaltyReturnTypeOnItsOwnLine: 2000000\nPointerAlignment: Left\nReflowComments: true\nSortIncludes: true\nSpaceAfterCStyleCast: false\nSpaceBeforeAssignmentOperators: true\nSpaceBeforeParens: ControlStatements\nSpaceInEmptyParentheses: false\nSpacesBeforeTrailingComments: 1\nSpacesInAngles: false\nSpacesInContainerLiterals: true\nSpacesInCStyleCastParentheses: false\nSpacesInParentheses: false\nSpacesInSquareBrackets: false\nStandard: Cpp11\nTabWidth: 8\nUseTab: Never\n...\n" }, { "path": ".coveragerc", "content": "[run]\nomit =\n docs/*\n tests/*\n setup.py\n xformers/benchmarks/*\n xformers/triton/k_*\n stubs/*\n third_party/*\n" }, { "path": ".editorconfig", "content": "root = true\n\n[*.py]\ncharset = utf-8\ntrim_trailing_whitespace = true\nend_of_line = lf\ninsert_final_newline = true\nindent_style = space\nindent_size = 4\n\n[*.md]\ntrim_trailing_whitespace = false\n" }, { "path": ".flake8", "content": "[flake8]\nexclude =\n .git\n ,.github/run-clang-format.py\n ,third_party\nmax-line-length = 140\ncopyright-check = True\nselect = E,F,W,C\ncopyright-regexp=Copyright \\(c\\) Facebook, Inc. and its affiliates. All Rights Reserved\nignore=W503,E203,E704\n" }, { "path": ".github/ISSUE_TEMPLATE/bug-report.md", "content": "---\nname: \"\\U0001F41B Bug Report\"\nabout: Submit a bug report to help us improve xFormers\n\n---\n\n# ๐Ÿ› Bug\n\n\n\n## Command\n\n## To Reproduce\n\nSteps to reproduce the behavior:\n\n\n\n1.\n2.\n3.\n\n\n\n## Expected behavior\n\n\n\n## Environment\n\nPlease copy and paste the output from the\nenvironment collection script from PyTorch\n(or fill out the checklist below manually).\n\nYou can run the script with:\n\n```bash\n# For security purposes, please check the contents of collect_env.py before running it.\npython -m torch.utils.collect_env\n```\n\n- PyTorch Version (e.g., 1.0):\n- OS (e.g., Linux):\n- How you installed PyTorch (`conda`, `pip`, source):\n- Build command you used (if compiling from source):\n- Python version:\n- CUDA/cuDNN version:\n- GPU models and configuration:\n- Any other relevant information:\n\n## Additional context\n\n\n" }, { "path": ".github/ISSUE_TEMPLATE/feature-request.md", "content": "---\nname: \"\\U0001F680Feature Request\"\nabout: Submit a proposal/request for a new xFormers feature\n\n---\n\n# ๐Ÿš€ Feature\n\n\n\n## Motivation\n\n\n\n## Pitch\n\n\n\n## Alternatives\n\n\n\n## Additional context\n\n\n" }, { "path": ".github/ISSUE_TEMPLATE/questions-help-support.md", "content": "---\nname: \"โ“Questions/Help/Support\"\nabout: Do you need support?\n\n---\n\n# โ“ Questions and Help\n" }, { "path": ".github/PULL_REQUEST_TEMPLATE.md", "content": "## What does this PR do?\nFixes # (issue).\n\n## Before submitting\n\n- [ ] Did you have fun?\n - Make sure you had fun coding ๐Ÿ™ƒ\n- [ ] Did you read the [contributor guideline](https://github.com/facebookresearch/xformers/blob/master/CONTRIBUTING.md)?\n- [ ] Was this discussed/approved via a Github issue? (no need for typos, doc improvements)\n - [ ] N/A\n- [ ] Did you make sure to update the docs?\n - [ ] N/A\n- [ ] Did you write any new necessary tests?\n - [ ] N/A\n- [ ] Did you update the [changelog](https://github.com/facebookresearch/xformers/blob/master/CHANGELOG.md)? (if needed)\n - [ ] N/A\n\n\n## PR review\nAnyone in the community is free to review the PR once the tests have passed.\nIf we didn't discuss your PR in Github issues there's a high chance it will not be merged.\n" }, { "path": ".github/actions/setup-build-cuda/action.yml", "content": "name: Set up Runner for build\n\ninputs:\n toolkit_type:\n description: cuda or rocm\n type: string\n toolkit_short_version:\n required: true\n type: string\n description: \"Example: 117 for 11.7\"\n python:\n description: Python version to install\n type: string\n default: \"3.10\"\n\nruns:\n using: composite\n steps:\n - id: cuda_info\n shell: python3 \"{0}\"\n run: |\n import os\n import sys\n print(sys.version)\n cushort = \"${{ inputs.toolkit_short_version }}\"\n # Version uploaded to pypi (rather than PyTorch s3)\n TORCH_CUDA_DEFAULT = \"128\" # since pytorch 2.9.0\n # https://github.com/Jimver/cuda-toolkit/blob/master/src/links/linux-links.ts\n full_version, install_script = {\n \"130\": (\"13.0.1\", \"https://developer.download.nvidia.com/compute/cuda/13.0.1/local_installers/cuda_13.0.1_580.82.07_linux.run\"),\n \"129\": (\"12.9.1\", \"https://developer.download.nvidia.com/compute/cuda/12.9.1/local_installers/cuda_12.9.1_575.57.08_linux.run\"),\n \"128\": (\"12.8.1\", \"https://developer.download.nvidia.com/compute/cuda/12.8.1/local_installers/cuda_12.8.1_570.124.06_linux.run\"),\n # (Build with nvcc 12.8 on linux even when building for 12.6 to avoid seg fault in Flash3 build)\n \"126\": (\"12.8.1\", \"https://developer.download.nvidia.com/compute/cuda/12.8.1/local_installers/cuda_12.8.1_570.124.06_linux.run\"),\n \"118\": (\"11.8.0\", \"https://developer.download.nvidia.com/compute/cuda/11.8.0/local_installers/cuda_11.8.0_520.61.05_linux.run\"),\n \"6.0\": (\"6.0.2\", \"https://repo.radeon.com/amdgpu-install/6.0.2/rhel/8.9/amdgpu-install-6.0.60002-1.el8.noarch.rpm\"),\n \"6.1\": (\"6.1.3\", \"https://repo.radeon.com/amdgpu-install/6.1.3/rhel/8.9/amdgpu-install-6.1.60103-1.el8.noarch.rpm\"),\n \"6.2.4\": (\"6.2.4\", \"https://repo.radeon.com/amdgpu-install/6.2.4/rhel/8.9/amdgpu-install-6.2.60204-1.el8.noarch.rpm\"),\n \"6.3\": (\"6.3.1\", \"https://repo.radeon.com/amdgpu-install/6.3.1/rhel/8.9/amdgpu-install-6.3.60301-1.el8.noarch.rpm\"),\n \"6.4\": (\"6.4.2\", \"https://repo.radeon.com/amdgpu-install/6.4.2/rhel/8.9/amdgpu-install-6.4.60402-1.el8.noarch.rpm\"),\n \"7.0\": (\"7.0.3\", \"https://repo.radeon.com/amdgpu-install/7.0.3/rhel/8/amdgpu-install-7.0.3.70003-1.el8.noarch.rpm\"),\n \"7.1\": (\"7.1.0\", \"https://repo.radeon.com/amdgpu-install/7.1/rhel/8/amdgpu-install-7.1.70100-1.el8.noarch.rpm\"),\n }[cushort]\n with open(os.environ['GITHUB_OUTPUT'], \"r+\") as fp:\n fp.write(\"CUDA_VERSION=\" + full_version + \"\\n\")\n if cushort == TORCH_CUDA_DEFAULT:\n fp.write(\"CUDA_VERSION_SUFFIX=\\n\")\n else:\n fp.write(\"CUDA_VERSION_SUFFIX=+\" + (\"cu\" if \"cuda\" == \"${{ inputs.toolkit_type }}\" else \"rocm\") + cushort + \"\\n\")\n fp.write(\"CUDA_INSTALL_SCRIPT=\" + install_script + \"\\n\")\n - run: echo \"CUDA_VERSION_SUFFIX=${{ steps.cuda_info.outputs.CUDA_VERSION_SUFFIX }}\" >> ${GITHUB_ENV}\n shell: bash\n\n # WINDOWS STEPS\n - name: Install cuda\n if: runner.os == 'Windows' && inputs.toolkit_type == 'cuda'\n id: cuda-toolkit\n # Using N-Storm fork until https://github.com/Jimver/cuda-toolkit/issues/395 is resolved\n uses: N-Storm/cuda-toolkit@v0.2.28\n with:\n cuda: ${{ steps.cuda_info.outputs.CUDA_VERSION }}\n method: network\n - if: runner.os == 'Windows' && inputs.toolkit_type == 'cuda'\n shell: bash\n run: |\n echo \"Installed cuda version is: ${{ steps.cuda-toolkit.outputs.cuda }}\"\n echo \"Cuda install location: ${{ steps.cuda-toolkit.outputs.CUDA_PATH }}\"\n echo \"CUDA_HOME=${{ steps.cuda-toolkit.outputs.CUDA_PATH }}\" >> ${GITHUB_ENV}\n cat ${GITHUB_ENV}\n\n - name: Install python\n if: runner.os == 'Windows'\n uses: actions/setup-python@v4\n with:\n python-version: ${{ inputs.python }}\n\n - name: Setup MSVC\n if: runner.os == 'Windows'\n uses: ilammy/msvc-dev-cmd@v1\n\n # really unfortunate: https://github.com/ilammy/msvc-dev-cmd#name-conflicts-with-shell-bash\n - name: Remove link.exe\n if: runner.os == 'Windows'\n shell: bash\n run: rm /usr/bin/link\n\n # LINUX STEPS\n - if: ${{ runner.os == 'Linux' && !(contains(inputs.toolkit_type, 'cuda') && fromJSON(inputs.toolkit_short_version) > 124) }}\n shell: bash\n run: |\n # Use GCC11 for ROCM / cu118 / cu124\n yum list installed\n yum install gcc-toolset-11-gcc gcc-toolset-11-gcc-c++ gcc-toolset-11-libstdc++-devel wget git -y\n echo \"source /opt/rh/gcc-toolset-11/enable\" >> ~/.profile\n\n - if: ${{ runner.os == 'Linux' && contains(inputs.toolkit_type, 'cuda') && fromJSON(inputs.toolkit_short_version) > 124 }}\n shell: bash\n run: |\n # Use GCC13 for cu126+\n yum list installed\n yum install gcc-toolset-13-gcc gcc-toolset-13-gcc-c++ gcc-toolset-13-libstdc++-devel wget git -y\n echo \"source /opt/rh/gcc-toolset-13/enable\" >> ~/.profile\n\n - if: runner.os == 'Linux'\n shell: bash -l {0}\n run: |\n yum list installed\n yum install wget git -y\n which g++\n g++ --version\n\n - if: runner.os == 'Linux' && contains(inputs.toolkit_type, 'cuda')\n name: (Linux) install cuda\n shell: bash -l {0}\n run: |\n wget -q \"${{ steps.cuda_info.outputs.CUDA_INSTALL_SCRIPT }}\" -O cuda.run && \\\n sh ./cuda.run --silent --toolkit && \\\n rm ./cuda.run\n echo \"CUDA_HOME=/usr/local/cuda\" >> ${GITHUB_ENV}\n\n - if: runner.os == 'Linux' && contains(inputs.toolkit_type, 'cuda')\n name: (Linux) print cuda setup info\n shell: bash -l {0}\n run: |\n echo \"CUDA_HOME=$CUDA_HOME\"\n echo \"###############################\"\n echo \"############ NVCC ############\"\n echo \"###############################\"\n $CUDA_HOME/bin/nvcc --version\n md5sum $CUDA_HOME/bin/nvcc\n echo \"###############################\"\n echo \"############ PTXAS ############\"\n echo \"###############################\"\n $CUDA_HOME/bin/ptxas --version\n md5sum $CUDA_HOME/bin/ptxas\n\n - if: runner.os == 'Linux' && contains(inputs.toolkit_type, 'rocm')\n name: (Linux) install rocm\n shell: bash\n run: |\n yum install -y libzstd\n yum install -y ${{ steps.cuda_info.outputs.CUDA_INSTALL_SCRIPT }}\n amdgpu-install -y --usecase=rocm --no-dkms\n echo \"ROCM_PATH=/opt/rocm\" >> ${GITHUB_ENV}\n echo \"PATH=$PATH:/opt/rocm/bin\" >> ${GITHUB_ENV}\n echo \"MAX_JOBS=16\" >> ${GITHUB_ENV}\n\n # host compiler is too new for cuda 12.1 :(\n - run: echo \"NVCC_FLAGS=-allow-unsupported-compiler\" >> $GITHUB_ENV\n shell: bash\n" }, { "path": ".github/actions/setup-env-build/action.yml", "content": "name: Install env + build\ninputs:\n arch:\n description: 'GPU architecture'\n required: true\n python:\n description: 'Python version'\n required: false\n default: \"3.11\"\n\nruns:\n using: composite\n steps:\n - name: Cleanup\n shell: bash\n run: rm -f ~/.profile ~/.bashrc\n - id: prepare_conda_env_paths\n shell: python\n run: |\n import os\n import subprocess\n import hashlib\n import glob\n import datetime\n from pathlib import Path\n\n CONDA_INSTALL_CMD = \"micromamba create python=${{ inputs.python }} zlib pip ninja ccache=4.8 -c conda-forge -q -y\"\n\n conda_env_key = CONDA_INSTALL_CMD + \"[cu130][v2]\"\n for file in sorted(glob.glob(\"requirement*.txt\")):\n conda_env_key += f\"\\n########## {file}\\n\"\n conda_env_key += Path(file).read_text()\n env_name_key = hashlib.sha224(conda_env_key.encode(\"ascii\")).hexdigest()[:8]\n env_name_key += \"-${{ inputs.arch }}\"\n # Nightly or Test, update every week\n env_name_key += \"-\"+datetime.date.today().strftime(\"%Y-week%W\")\n shared_dir = os.environ.get(\"GHRUNNER_SHARED_DIR\", os.getcwd())\n env_path = os.path.join(shared_dir, \"tmp\", \"${{ inputs.arch }}\", os.environ[\"GITHUB_RUN_ID\"])\n final_env = Path(shared_dir) / f\"env_{env_name_key}.txt\"\n pkg_dir = Path(shared_dir) / \"pkgs-sm${{ inputs.arch }}\"\n (Path(shared_dir) / f\"env_{env_name_key}_content.txt\").write_text(conda_env_key)\n CONDA_INSTALL_CMD += \" -p \" + env_path\n env_already_built = False\n # If environment is already built\n if final_env.is_file():\n final_env_link = final_env.read_text().strip()\n if (Path(final_env_link) / \"bin\" / \"python\").is_file():\n print(\"Found valid env - skipping env setup\")\n CONDA_INSTALL_CMD = \"true\"\n env_already_built = True\n env_path = final_env_link\n else:\n print(\"Invalid env\")\n with open(os.environ['GITHUB_ENV'], \"r+\") as fp:\n fp.write(\"CONDA_ENV_LINK=\" + str(final_env) + \"\\n\")\n fp.write(\"CONDA_PREFIX=\" + env_path + \"\\n\")\n fp.write(\"CONDA_PKGS_DIRS=\" + str(pkg_dir) + \"\\n\")\n fp.write(\"CONDA_INSTALL_CMD=\" + CONDA_INSTALL_CMD + \"\\n\")\n fp.write(\"CONDA_ENV_HASH=\" + env_name_key + \"\\n\")\n fp.write(\"PY=\" + os.path.join(env_path, \"bin\", \"python\") + \"\\n\")\n fp.write(\"PIP=\" + os.path.join(env_path, \"bin\", \"pip\") + \"\\n\")\n with open(os.environ['GITHUB_OUTPUT'], \"r+\") as fp:\n fp.write(f\"ENV_CACHED={int(env_already_built)}\\n\")\n - name: Print conda commands\n shell: bash -l {0}\n run: |\n echo \"CONDA_PREFIX=$CONDA_PREFIX\"\n echo \"CONDA_INSTALL_CMD=$CONDA_INSTALL_CMD\"\n echo \"CONDA_ENV_HASH=$CONDA_ENV_HASH\"\n echo \"PY=$PY\"\n - name: Install micromamba\n shell: bash -l {0}\n run: |\n set -ex\n curl -Ls https://micro.mamba.pm/api/micromamba/linux-64/latest -o micromamba.tar.bz2\n tar --extract --verbose --bzip2 --file=micromamba.tar.bz2 bin/micromamba\n echo \"eval \\\"\\$($(pwd)/bin/micromamba shell hook --shell bash)\\\"\" >> ~/.profile\n - name: Conda/pip setup\n shell: bash -l {0}\n if: steps.prepare_conda_env_paths.outputs.ENV_CACHED == 0\n run: |\n set -ex\n micromamba config set channel_priority strict\n # Retry if failed after removing downloaded packages cache\n $CONDA_INSTALL_CMD || (rm -rf $CONDA_PKGS_DIRS && rm -rf $CONDA_PREFIX && $CONDA_INSTALL_CMD)\n $PY -m pip install cmake\n $PY -m pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu128\n $PY -m pip install -r requirements-benchmark.txt --progress-bar off\n - name: Activate environment\n shell: bash -l {0}\n run: |\n echo \"micromamba activate $CONDA_PREFIX\" >> ~/.profile\n echo \"==== .profile =====\"\n cat ~/.profile\n - run: which python\n shell: bash -l {0}\n - name: Setup ccache nvcc\n shell: bash -l {0}\n if: steps.prepare_conda_env_paths.outputs.ENV_CACHED == 0\n run: |\n echo \"#!/bin/bash\" > $CONDA_PREFIX/bin/nvcc-ccache\n echo \"ccache nvcc \\\"\\$@\\\"\" >> $CONDA_PREFIX/bin/nvcc-ccache\n cat $CONDA_PREFIX/bin/nvcc-ccache\n chmod +x $CONDA_PREFIX/bin/nvcc-ccache\n which nvcc\n ccache --version\n\n - name: Setup ccache g++\n shell: bash -l {0}\n if: steps.prepare_conda_env_paths.outputs.ENV_CACHED == 0\n run: |\n echo \"#!/bin/bash\" > $CONDA_PREFIX/bin/g++-ccache\n echo \"ccache g++ \\\"\\$@\\\"\" >> $CONDA_PREFIX/bin/g++-ccache\n cat $CONDA_PREFIX/bin/g++-ccache\n chmod +x $CONDA_PREFIX/bin/g++-ccache\n which g++-ccache\n\n - name: Patch for https://github.com/pytorch/pytorch/issues/114962\n shell: bash -l {0}\n run: |\n CPP_EXTENSIONS_PY=$(python -c \"import torch.utils.cpp_extension; print(torch.utils.cpp_extension.__file__)\")\n echo \"Patching $CPP_EXTENSIONS_PY\"\n sed -i \"/generate-dependencies-with-compile/d\" $CPP_EXTENSIONS_PY\n - name: Check NVIDIA libs\n shell: bash -l {0}\n run: |\n ldconfig -p | grep libcuda.so\n ls /.singularity.d/libs/\n - name: Mark env as ready\n shell: bash -l {0}\n if: steps.prepare_conda_env_paths.outputs.ENV_CACHED == 0\n run: echo $CONDA_PREFIX > $CONDA_ENV_LINK\n - name: Setup ccache\n shell: bash -l {0}\n run: |\n export CCACHE_DIR=$GHRUNNER_SHARED_DIR/ccache\n echo \"CCACHE_DIR=$CCACHE_DIR\" >> ${GITHUB_ENV}\n mkdir -p $CCACHE_DIR\n ccache -s\n - name: Build\n shell: bash -l {0}\n run: |\n PYTORCH_NVCC=\"$CONDA_PREFIX/bin/nvcc-ccache\" CXX=\"g++-ccache\" TORCH_CUDA_ARCH_LIST=${{ inputs.arch }} python -m pip install -v --no-build-isolation -e .\n - name: Check for PyTorch stable symbols\n shell: bash -l {0}\n run: |\n bad_symbols=$(nm --dynamic --undefined-only --demangle xformers/_C.so | grep --extended-regexp \"(torch|at|c10|c10d)::\" || true)\n if [[ $bad_symbols != \"\" ]]; then echo \"These non-stable PyTorch symbols made it into the xFormers shared library:\"; echo $bad_symbols; exit 1; fi\n - name: Build info\n run: |\n printenv\n python -m xformers.info\n python xformers/_triton_version_fairinternal.py\n ccache -s\n shell: bash -l {0}\n" }, { "path": ".github/compute_wheel_version.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.\n#\n# This source code is licensed under the BSD license found in the\n# LICENSE file in the root directory of this source tree.\nimport argparse\nimport subprocess\nfrom pathlib import Path\nfrom typing import Optional\n\n# TODO: consolidate with the code in build_conda.py\nTHIS_PATH = Path(__file__).resolve()\nversion_from_file = (THIS_PATH.parents[1] / \"version.txt\").read_text().strip()\n\n\ndef get_tagged_version() -> Optional[str]:\n \"\"\"\n Return whether we are at an exact version (namely the version variable).\n \"\"\"\n try:\n tag = subprocess.check_output(\n [\"git\", \"describe\", \"--tags\", \"--exact-match\", \"HEAD\"],\n text=True,\n stderr=subprocess.DEVNULL,\n ).strip()\n except subprocess.CalledProcessError: # no tag\n return None\n\n if not tag.startswith(\"v\"):\n return None\n return tag[1:]\n\n\ndef get_dev_version() -> str:\n assert \".dev\" not in version_from_file\n num_commits = subprocess.check_output(\n [\"git\", \"rev-list\", \"--count\", \"HEAD\"], text=True\n ).strip()\n return f\"{version_from_file}.dev{num_commits}\"\n\n\nif __name__ == \"__main__\":\n parser = argparse.ArgumentParser()\n parser.add_argument(\n \"--source\", choices=[\"tag\", \"dev\", \"tag,dev\"], required=False, default=\"tag,dev\"\n )\n args = parser.parse_args()\n\n if \"tag\" in args.source:\n tagged_version = get_tagged_version()\n if args.source == \"tag\" and tagged_version is None:\n raise ValueError(\"No tag found\")\n else:\n tagged_version = None\n if tagged_version is not None:\n print(tagged_version, end=\"\")\n else:\n print(get_dev_version(), end=\"\")\n" }, { "path": ".github/gpu_benchmark_diff.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.\n#\n# This source code is licensed under the BSD license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport glob\nimport os\nimport subprocess\n\nimport xformers.benchmarks.utils as utils\n\n\nclass NamedObject:\n def __init__(self, name) -> None:\n self.__name__ = name\n\n\ndef git_file_at(filename: str, ref: str) -> str:\n try:\n return subprocess.check_output(\n [\"git\", \"show\", f\"{ref}:{filename}\"], text=True\n ).strip()\n except subprocess.CalledProcessError:\n return \"\" # File does not exist in that revision\n\n\nGITHUB_BASE_REF = subprocess.check_output(\n [\"git\", \"rev-parse\", \"origin/\" + os.environ[\"GITHUB_BASE_REF\"]], text=True\n).strip()\nXFORMERS_BENCHMARKS_CACHE = os.environ[\"XFORMERS_BENCHMARKS_CACHE\"]\nGITHUB_CURRENT_REF = subprocess.check_output(\n [\"git\", \"rev-parse\", \"HEAD\"], text=True\n).strip()\n\nfor f in glob.glob(os.path.join(XFORMERS_BENCHMARKS_CACHE, \"*\", \"*.csv\")):\n before = git_file_at(f, ref=GITHUB_BASE_REF)\n now = git_file_at(f, ref=GITHUB_CURRENT_REF)\n if before == \"\" or before == now:\n continue\n benchmark_name = os.path.basename(os.path.dirname(f))\n\n print(\"#\" * 100)\n print(f\"# UPDATED: {f}\")\n print(\"#\" * 100)\n\n filename_before = f.replace(\"reference\", \"before\")\n filename_now = f.replace(\"reference\", \"now\")\n with open(filename_before, \"w+\") as fd:\n fd.write(before)\n with open(filename_now, \"w+\") as fd:\n fd.write(now)\n utils.benchmark_run_and_compare(\n benchmark_fn=NamedObject(benchmark_name),\n cases=[],\n compare=[\n os.path.basename(filename_before)[: -len(\".csv\")],\n os.path.basename(filename_now)[: -len(\".csv\")],\n ],\n )\n" }, { "path": ".github/run-clang-format.py", "content": "#!/usr/bin/env python3\n\"\"\"\nMIT License\nCopyright (c) 2017 Guillaume Papin\nPermission is hereby granted, free of charge, to any person obtaining a copy\nof this software and associated documentation files (the \"Software\"), to deal\nin the Software without restriction, including without limitation the rights\nto use, copy, modify, merge, publish, distribute, sublicense, and/or sell\ncopies of the Software, and to permit persons to whom the Software is\nfurnished to do so, subject to the following conditions:\nThe above copyright notice and this permission notice shall be included in all\ncopies or substantial portions of the Software.\nTHE SOFTWARE IS PROVIDED \"AS IS\", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR\nIMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,\nFITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE\nAUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER\nLIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,\nOUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE\nSOFTWARE.\n\"\"\"\n\n\"\"\"A wrapper script around clang-format, suitable for linting multiple files\nand to use for continuous integration.\nThis is an alternative API for the clang-format command line.\nIt runs over multiple files and directories in parallel.\nA diff output is produced and a sensible exit code is returned.\n\"\"\"\n\nimport argparse # noqa: E402\nimport difflib # noqa: E402\nimport fnmatch # noqa: E402\nimport io # noqa: E402\nimport multiprocessing # noqa: E402\nimport os # noqa: E402\nimport signal # noqa: E402\nimport subprocess # noqa: E402\nimport sys # noqa: E402\nimport traceback # noqa: E402\nfrom functools import partial # noqa: E402\nfrom subprocess import DEVNULL # noqa: E402\n\nDEFAULT_EXTENSIONS = \"c,h,C,H,cpp,hpp,cc,hh,c++,h++,cxx,hxx,cu\"\n\n\nclass ExitStatus:\n SUCCESS = 0\n DIFF = 1\n TROUBLE = 2\n\n\ndef list_files(files, recursive=False, extensions=None, exclude=None):\n if extensions is None:\n extensions = []\n if exclude is None:\n exclude = []\n\n out = []\n for file in files:\n if recursive and os.path.isdir(file):\n for dirpath, dnames, fnames in os.walk(file):\n fpaths = [os.path.join(dirpath, fname) for fname in fnames]\n for pattern in exclude:\n # os.walk() supports trimming down the dnames list\n # by modifying it in-place,\n # to avoid unnecessary directory listings.\n dnames[:] = [\n x\n for x in dnames\n if not fnmatch.fnmatch(os.path.join(dirpath, x), pattern)\n ]\n fpaths = [x for x in fpaths if not fnmatch.fnmatch(x, pattern)]\n for f in fpaths:\n ext = os.path.splitext(f)[1][1:]\n if ext in extensions:\n out.append(f)\n else:\n out.append(file)\n return out\n\n\ndef make_diff(file, original, reformatted):\n return list(\n difflib.unified_diff(\n original,\n reformatted,\n fromfile=\"a/{}\\t(original)\".format(file),\n tofile=\"b/{}\\t(reformatted)\".format(file),\n n=3,\n )\n )\n\n\nclass DiffError(Exception):\n def __init__(self, message, errs=None):\n super(DiffError, self).__init__(message)\n self.errs = errs or []\n\n\nclass UnexpectedError(Exception):\n def __init__(self, message, exc=None):\n super(UnexpectedError, self).__init__(message)\n self.formatted_traceback = traceback.format_exc()\n self.exc = exc\n\n\ndef run_clang_format_diff_wrapper(args, file):\n try:\n ret = run_clang_format_diff(args, file)\n return ret\n except DiffError:\n raise\n except Exception as e:\n raise UnexpectedError(\"{}: {}: {}\".format(file, e.__class__.__name__, e), e)\n\n\ndef run_clang_format_diff(args, file):\n try:\n with io.open(file, \"r\", encoding=\"utf-8\") as f:\n original = f.readlines()\n except IOError as exc:\n raise DiffError(str(exc))\n invocation = [args.clang_format_executable, file]\n\n # Use of utf-8 to decode the process output.\n #\n # Hopefully, this is the correct thing to do.\n #\n # It's done due to the following assumptions (which may be incorrect):\n # - clang-format will returns the bytes read from the files as-is,\n # without conversion, and it is already assumed that the files use utf-8.\n # - if the diagnostics were internationalized, they would use utf-8:\n # > Adding Translations to Clang\n # >\n # > Not possible yet!\n # > Diagnostic strings should be written in UTF-8,\n # > the client can translate to the relevant code page if needed.\n # > Each translation completely replaces the format string\n # > for the diagnostic.\n # > -- http://clang.llvm.org/docs/InternalsManual.html#internals-diag-translation\n\n try:\n proc = subprocess.Popen(\n invocation,\n stdout=subprocess.PIPE,\n stderr=subprocess.PIPE,\n universal_newlines=True,\n encoding=\"utf-8\",\n )\n except OSError as exc:\n raise DiffError(\n \"Command '{}' failed to start: {}\".format(\n subprocess.list2cmdline(invocation), exc\n )\n )\n proc_stdout = proc.stdout\n proc_stderr = proc.stderr\n\n # hopefully the stderr pipe won't get full and block the process\n outs = list(proc_stdout.readlines())\n errs = list(proc_stderr.readlines())\n proc.wait()\n if proc.returncode:\n raise DiffError(\n \"Command '{}' returned non-zero exit status {}\".format(\n subprocess.list2cmdline(invocation), proc.returncode\n ),\n errs,\n )\n return make_diff(file, original, outs), errs\n\n\ndef bold_red(s):\n return \"\\x1b[1m\\x1b[31m\" + s + \"\\x1b[0m\"\n\n\ndef colorize(diff_lines):\n def bold(s):\n return \"\\x1b[1m\" + s + \"\\x1b[0m\"\n\n def cyan(s):\n return \"\\x1b[36m\" + s + \"\\x1b[0m\"\n\n def green(s):\n return \"\\x1b[32m\" + s + \"\\x1b[0m\"\n\n def red(s):\n return \"\\x1b[31m\" + s + \"\\x1b[0m\"\n\n for line in diff_lines:\n if line[:4] in [\"--- \", \"+++ \"]:\n yield bold(line)\n elif line.startswith(\"@@ \"):\n yield cyan(line)\n elif line.startswith(\"+\"):\n yield green(line)\n elif line.startswith(\"-\"):\n yield red(line)\n else:\n yield line\n\n\ndef print_diff(diff_lines, use_color):\n if use_color:\n diff_lines = colorize(diff_lines)\n sys.stdout.writelines(diff_lines)\n\n\ndef print_trouble(prog, message, use_colors):\n error_text = \"error:\"\n if use_colors:\n error_text = bold_red(error_text)\n print(\"{}: {} {}\".format(prog, error_text, message), file=sys.stderr)\n\n\ndef main():\n parser = argparse.ArgumentParser(description=__doc__)\n parser.add_argument(\n \"--clang-format-executable\",\n metavar=\"EXECUTABLE\",\n help=\"path to the clang-format executable\",\n default=\"clang-format\",\n )\n parser.add_argument(\n \"--extensions\",\n help=\"comma separated list of file extensions (default: {})\".format(\n DEFAULT_EXTENSIONS\n ),\n default=DEFAULT_EXTENSIONS,\n )\n parser.add_argument(\n \"-r\",\n \"--recursive\",\n action=\"store_true\",\n help=\"run recursively over directories\",\n )\n parser.add_argument(\"files\", metavar=\"file\", nargs=\"+\")\n parser.add_argument(\"-q\", \"--quiet\", action=\"store_true\")\n parser.add_argument(\n \"-j\",\n metavar=\"N\",\n type=int,\n default=0,\n help=\"run N clang-format jobs in parallel\" \" (default number of cpus + 1)\",\n )\n parser.add_argument(\n \"--color\",\n default=\"auto\",\n choices=[\"auto\", \"always\", \"never\"],\n help=\"show colored diff (default: auto)\",\n )\n parser.add_argument(\n \"-e\",\n \"--exclude\",\n metavar=\"PATTERN\",\n action=\"append\",\n default=[],\n help=\"exclude paths matching the given glob-like pattern(s)\"\n \" from recursive search\",\n )\n\n args = parser.parse_args()\n\n # use default signal handling, like diff return SIGINT value on ^C\n # https://bugs.python.org/issue14229#msg156446\n signal.signal(signal.SIGINT, signal.SIG_DFL)\n try:\n signal.SIGPIPE\n except AttributeError:\n # compatibility, SIGPIPE does not exist on Windows\n pass\n else:\n signal.signal(signal.SIGPIPE, signal.SIG_DFL)\n\n colored_stdout = False\n colored_stderr = False\n if args.color == \"always\":\n colored_stdout = True\n colored_stderr = True\n elif args.color == \"auto\":\n colored_stdout = sys.stdout.isatty()\n colored_stderr = sys.stderr.isatty()\n\n version_invocation = [args.clang_format_executable, str(\"--version\")]\n try:\n subprocess.check_call(version_invocation, stdout=DEVNULL)\n except subprocess.CalledProcessError as e:\n print_trouble(parser.prog, str(e), use_colors=colored_stderr)\n return ExitStatus.TROUBLE\n except OSError as e:\n print_trouble(\n parser.prog,\n \"Command '{}' failed to start: {}\".format(\n subprocess.list2cmdline(version_invocation), e\n ),\n use_colors=colored_stderr,\n )\n return ExitStatus.TROUBLE\n\n retcode = ExitStatus.SUCCESS\n files = list_files(\n args.files,\n recursive=args.recursive,\n exclude=args.exclude,\n extensions=args.extensions.split(\",\"),\n )\n\n if not files:\n return\n\n njobs = args.j\n if njobs == 0:\n njobs = multiprocessing.cpu_count() + 1\n njobs = min(len(files), njobs)\n\n if njobs == 1:\n # execute directly instead of in a pool,\n # less overhead, simpler stacktraces\n it = (run_clang_format_diff_wrapper(args, file) for file in files)\n pool = None\n else:\n pool = multiprocessing.Pool(njobs)\n it = pool.imap_unordered(partial(run_clang_format_diff_wrapper, args), files)\n while True:\n try:\n outs, errs = next(it)\n except StopIteration:\n break\n except DiffError as e:\n print_trouble(parser.prog, str(e), use_colors=colored_stderr)\n retcode = ExitStatus.TROUBLE\n sys.stderr.writelines(e.errs)\n except UnexpectedError as e:\n print_trouble(parser.prog, str(e), use_colors=colored_stderr)\n sys.stderr.write(e.formatted_traceback)\n retcode = ExitStatus.TROUBLE\n # stop at the first unexpected error,\n # something could be very wrong,\n # don't process all files unnecessarily\n if pool:\n pool.terminate()\n break\n else:\n sys.stderr.writelines(errs)\n if outs == []:\n continue\n if not args.quiet:\n print_diff(outs, use_color=colored_stdout)\n if retcode == ExitStatus.SUCCESS:\n retcode = ExitStatus.DIFF\n return retcode\n\n\nif __name__ == \"__main__\":\n sys.exit(main())\n" }, { "path": ".github/run_benchmark_wrapper.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.\n#\n# This source code is licensed under the BSD license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport glob\nimport os\nimport shlex\nimport subprocess\nimport sys\n\nimport torch\n\nimport xformers\n\n# Build failed - return early\nif not xformers._has_cpp_library:\n print(\"xFormers wasn't built correctly - can't run benchmarks\")\n sys.exit(0)\n\nbenchmark_script = os.path.join(\"xformers\", \"benchmarks\", sys.argv[1])\nbenchmark_fn = sys.argv[2]\nlabel = subprocess.check_output([\"git\", \"rev-parse\", \"HEAD\"], text=True).strip()[:8]\ncmd = [\n sys.executable,\n benchmark_script,\n \"--label\",\n label,\n \"--fn\",\n benchmark_fn,\n \"--fail_if_regression\",\n \"--quiet\",\n]\nenv = (\n torch.cuda.get_device_name(torch.cuda.current_device())\n .replace(\" \", \"_\")\n .replace(\"-\", \"_\")\n .replace(\".\", \"_\")\n)\n\n# Figure out the name of the baseline\npattern = os.path.join(os.environ[\"XFORMERS_BENCHMARKS_CACHE\"], benchmark_fn, \"*.csv\")\nref_names = glob.glob(pattern)\nbaseline_names = set(\n os.path.basename(s)[: -len(\".csv\")]\n for s in ref_names\n # Only compare to benchmark data on same hardware\n if env in os.path.basename(s)\n)\nif baseline_names:\n if len(baseline_names) > 1:\n raise RuntimeError(\n f\"Supplied more than one reference for this benchmark: {','.join(baseline_names)}\"\n )\n cmd += [\"--compare\", \",\".join(baseline_names)]\n\nprint(\"EXEC:\", shlex.join(cmd))\n\nretcode = 0\ntry:\n subprocess.check_call(cmd)\nexcept subprocess.CalledProcessError as e:\n retcode = e.returncode\n\n# Remove original benchmark files\nfor f in ref_names:\n os.remove(f)\n# Rename new ones as 'ref'\nfor f in glob.glob(pattern):\n os.rename(f, f.replace(label, \"reference\"))\n\nsys.exit(retcode)\n" }, { "path": ".github/selective_ci/requirements.txt", "content": "GitPython\n" }, { "path": ".github/selective_ci/selective_ci.py", "content": "# Copyright (c) Facebook, Inc. and its affiliates. All rights reserved.\n#\n# This source code is licensed under the BSD license found in the\n# LICENSE file in the root directory of this source tree.\n\nimport argparse\nimport fnmatch\nimport os\nfrom dataclasses import dataclass, field\nfrom pathlib import Path\n\nimport git\n\n\n@dataclass\nclass ComponentInfo:\n \"\"\"\n A component is deemed to have changed if any of its\n files or dependencies have changed.\n If it has not changed, its files will be removed.\n \"\"\"\n\n name: str\n # These files will be deleted if the component is not enabled\n files: list[str]\n dependencies: list[str]\n disable_set_env: dict[str, str] = field(default_factory=dict)\n\n\nCOMMON_PATTERNS = [\n # All components will be tested if something in there changes\n \"setup.py\",\n]\n\nCOMPONENTS = [\n ComponentInfo(\n name=\"attention\",\n files=[\n \"tests/test_mem_eff_attention.py\",\n \"tests/test_find_sparse_locations*.py\",\n \"tests/test_block_sparse_mem_eff_attention*.py\",\n \"tests/test_attention_patterns.py\",\n \"tests/test_rope_padded.py\",\n \"tests/test_tree_attention*.py\",\n \"tests/test_fmha*.py\",\n ],\n dependencies=[\n \"xformers/ops/fmha/*\",\n \"third_party/cutlass\",\n \"third_party/composable_kernel_tiled\",\n \"xformers/csrc/attention/*\",\n \"xformers/triton/*\",\n ],\n disable_set_env={\n \"XFORMERS_DISABLE_FLASH_ATTN\": \"1\",\n },\n ),\n ComponentInfo(\n name=\"sp24\",\n files=[\n \"tests/test_sparsity24.py\",\n \"xformers/csrc/sparse24/*\",\n ],\n dependencies=[\n \"xformers/ops/sp24.py\",\n ],\n ),\n ComponentInfo(\n name=\"sequence_parallel_fused\",\n files=[\n \"tests/test_seqpar.py\",\n \"tests/test_sequence_parallel_fused_ops.py\",\n \"tests/test_tiled_matmul.py\",\n ],\n dependencies=[\n \"tests/multiprocessing_utils.py\",\n \"xformers/ops/sequence_parallel_fused_ops.py\",\n ],\n ),\n]\n\nrepo_root_path = Path(__file__).parent.parent.parent.resolve().absolute()\nrepo = git.Repo(repo_root_path)\n\n\ndef list_files_in_commit(commit: git.Commit):\n file_list = []\n stack = [commit.tree]\n while len(stack) > 0:\n tree = stack.pop()\n # enumerate blobs (files) at this level\n for b in tree.blobs:\n file_list.append(str(Path(b.path).absolute().relative_to(repo_root_path)))\n for subtree in tree.trees:\n stack.append(subtree)\n # you can return dir_list if you want directories too\n return file_list\n\n\ndef check_patterns_are_valid(patterns):\n # Only check patterns in `fairinternal` repo\n if os.environ.get(\"GITHUB_REPOSITORY\", \"\") != \"fairinternal/xformers\":\n return\n found_patterns = set()\n for f in all_files:\n for pattern in patterns:\n if fnmatch.fnmatch(f, pattern):\n found_patterns.add(pattern)\n for pattern in patterns:\n if pattern not in found_patterns:\n assert False, f\"Pattern does not match any file: `{pattern}`\"\n\n\nparser = argparse.ArgumentParser(\"xFormers selective CI\")\nparser.add_argument(\"--base_commit\", default=\"origin/main\")\nargs = parser.parse_args()\n\nbase_commit = repo.rev_parse(args.base_commit)\nall_files = list_files_in_commit(repo.head.commit) + [sm.path for sm in repo.submodules]\nall_modified_files = set()\nfor item in repo.head.commit.diff(base_commit):\n if item.a_path is not None:\n all_modified_files.add(item.a_path)\n if item.b_path is not None:\n all_modified_files.add(item.b_path)\n\ncheck_patterns_are_valid(COMMON_PATTERNS)\nfor component in COMPONENTS:\n # Sanity check that all files exist\n check_patterns_are_valid(component.files + component.dependencies)\n\n # Check if module is updated\n skip_module = True\n for pattern in COMMON_PATTERNS + component.files + component.dependencies:\n for f in all_modified_files:\n if fnmatch.fnmatch(f, pattern):\n skip_module = False\n break\n print(component.name, \"SKIP\" if skip_module else \"TEST\")\n if not skip_module:\n continue\n\n # Delete component files\n for f in all_files:\n for pattern in component.files:\n if fnmatch.fnmatch(f, pattern):\n if Path(f).exists():\n Path(f).unlink()\n\n # Set env variable\n for env_k, env_v in component.disable_set_env.items():\n if \"GITHUB_ENV\" not in os.environ:\n print(f\"{env_k}={env_v}\")\n continue\n with open(os.environ[\"GITHUB_ENV\"], \"a\") as fd:\n fd.write(f\"{env_k}={env_v}\\n\")\n" }, { "path": ".github/workflows/gh-pages.yml", "content": "name: Build & deploy documentation\n\non:\n push:\n branches:\n - main\n pull_request:\n\njobs:\n deploy:\n runs-on: ubuntu-24.04\n concurrency:\n group: ${{ github.workflow }}-${{ github.ref }}\n\n steps:\n - uses: actions/checkout@v2\n\n - name: Setup Python\n uses: actions/setup-python@v2\n with:\n python-version: '3.9'\n\n - name: Upgrade pip\n run: |\n # install pip=>20.1 to use \"pip cache dir\"\n python3 -m pip install --upgrade pip\n\n - name: Get pip cache dir\n id: pip-cache\n run: echo \"::set-output name=dir::$(pip cache dir)\"\n\n - name: Cache dependencies\n uses: actions/cache@v4\n with:\n path: ${{ steps.pip-cache.outputs.dir }}\n key: ${{ runner.os }}-pip-${{ hashFiles('**/requirements.txt') }}\n restore-keys: |\n ${{ runner.os }}-pip-\n\n - name: Build docs\n\n run: |\n cd docs\n pip install --progress-bar off -r requirements.txt\n make help\n make html\n\n - name: Deploy\n uses: peaceiris/actions-gh-pages@v3\n with:\n github_token: ${{ secrets.GITHUB_TOKEN }}\n publish_dir: docs/build/html\n if: github.event_name != 'pull_request'\n" }, { "path": ".github/workflows/gpu_test_gh.yml", "content": "name: gpu_test_gh\n\non:\n workflow_dispatch: {}\n pull_request:\n paths:\n - \"xformers/**\"\n - \"!xformers/benchmarks/**\"\n - \"!xformers/version.txt\"\n - \".github/workflows/gpu_test_gh*\"\n - \"tests/**\"\n - \"setup.py\"\n - \"requirements*.txt\"\n - \"third_party/**\"\n push:\n branches:\n - main\n\nenv:\n XFORMERS_BUILD_TYPE: \"Release\"\n CI: \"1\"\n TORCHINDUCTOR_COMPILE_THREADS: \"1\"\n\njobs:\n gpu_test_gh:\n strategy:\n fail-fast: false\n matrix:\n gpu:\n - runner: \"h100-256GB\"\n sm: \"9.0a\"\n - runner: \"4-core-ubuntu-gpu-t4\"\n sm: \"7.5\"\n python: [3.11]\n\n name: test_sm${{ matrix.gpu.sm }}\n runs-on: ${{ matrix.gpu.runner }}\n\n timeout-minutes: 360\n defaults:\n run:\n shell: bash -l {0}\n steps:\n - name: Recursive checkout\n uses: actions/checkout@v3\n with:\n submodules: recursive\n path: \".\"\n fetch-depth: 0 # We need commits history as well\n - run: nvidia-smi\n - name: Install micromamba\n run: |\n set -ex\n curl -Ls https://micro.mamba.pm/api/micromamba/linux-64/latest | tar -xvj bin/micromamba\n echo \"eval \\\"\\$($(pwd)/bin/micromamba shell hook --shell bash)\\\"\" >> ~/.profile\n cat ~/.profile\n - name: Create environment\n run: |\n set -ex\n micromamba config set channel_priority strict\n micromamba create -n env python=${{ matrix.python }} \\\n zlib pip ninja ccache=4.8 cuda-toolkit \\\n -c \"nvidia/label/cuda-12.6\" -c conda-forge -q -y\n - name: Activate environment\n shell: bash -l {0}\n run: |\n echo \"micromamba activate env\" >> ~/.profile\n echo \"==== .profile =====\"\n cat ~/.profile\n - name: Selective build/tests\n if: github.event_name == 'pull_request'\n run: |\n pip install -r .github/selective_ci/requirements.txt\n python .github/selective_ci/selective_ci.py --base_commit ${{ github.event.pull_request.base.sha }}\n - name: Setup test requirements\n run: |\n which python\n which nvcc\n pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu126\n pip install --pre flash_attn_3 --index-url https://download.pytorch.org/whl/cu126\n pip install -r requirements-test.txt --progress-bar off\n - run: TORCH_CUDA_ARCH_LIST=${{ matrix.gpu.sm }} python -m pip install -v --no-build-isolation -e .\n env:\n TORCH_DONT_CHECK_COMPILER_ABI: 1\n - run: python -m xformers.info\n - name: xFormers import should not init cuda context\n run: |\n # NOTE: we check GPU version by default to determine if triton should be used\n # and this initializes CUDA context, unless we set `XFORMERS_ENABLE_TRITON`\n XFORMERS_ENABLE_TRITON=1 python -c \"import xformers; import xformers.ops; import torch; assert not torch.cuda.is_initialized()\"\n - name: Check for PyTorch stable symbols\n run: |\n bad_symbols=$(nm --dynamic --undefined-only --demangle xformers/_C.so | grep --extended-regexp \"(torch|at|c10|c10d)::\" || true)\n if [[ $bad_symbols != \"\" ]]; then echo \"These non-stable PyTorch symbols made it into the xFormers shared library:\"; echo $bad_symbols; exit 1; fi\n - name: Unit tests\n run: |\n python -m pytest --verbose --random-order-bucket=global --maxfail=20 --junitxml=test-results/junit.xml --cov-report=xml --cov=./ tests\n - name: Publish Test Report\n uses: mikepenz/action-junit-report@v3\n if: success() || failure() # always run even if the previous step fails\n with:\n report_paths: 'test-results/*.xml'\n" }, { "path": ".github/workflows/linters.yml", "content": "on:\n pull_request: {}\n push:\n branches:\n - main\n\njobs:\n repo:\n uses: ./.github/workflows/linters_reusable.yml\n" }, { "path": ".github/workflows/linters_reusable.yml", "content": "name: lint\n\non:\n workflow_call:\n inputs:\n pre-script:\n type: string\n\njobs:\n linters:\n runs-on: ubuntu-22.04\n\n steps:\n - uses: actions/checkout@v3\n with:\n fetch-depth: 0\n - name: Setup Python\n uses: actions/setup-python@v5\n with:\n python-version: '3.10'\n - name: Cleanup host\n run: |\n # Github's ubuntu-latest comes with a ton of stuff;\n # https://carlosbecker.com/posts/github-actions-disk-space suggests\n # this hotfix:\n df -h\n sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL\n sudo docker image prune --all --force\n sudo docker builder prune -a\n df -h\n - name: Run pre-script\n if: ${{ inputs.pre-script }}\n run: ${{ inputs.pre-script }}\n # Triton is too slow to install, and beside it's not needed\n - run: sed -i '/triton/d' requirements-test.txt\n - name: Install deps\n run: pip install -r requirements-test.txt\n - name: ufmt\n if: success() || failure()\n run: ufmt check\n - name: mypy\n if: success() || failure()\n run: |\n python -m mypy --version\n python -m mypy --ignore-missing-imports --scripts-are-modules --pretty --exclude \"(build|stubs|third_party|docs|examples|setup.py)\" .\n - name: flake8\n if: success() || failure()\n run: python -m flake8 --config .flake8 --show-source --statistics\n - name: clang-format\n if: success() || failure()\n run: |\n pip install clang-format\n clang-format --version\n\n # apply to our files - excluding autogenerated files\n ./.github/run-clang-format.py -e \"*fmha/autogen\" -r xformers/csrc\n - name: PyTorch stable API includes\n if: success() || failure()\n run: |\n bad_files=$(git grep --extended-regex -e \"#\\s*include\\s*<.*(torch|ATen|c10)\" --and --not -e \"#\\s*include\\s*= 120 && fromJSON(inputs.toolkit_short_version) < 130\n run: |\n echo \"TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST 8.0 9.0a\" >> ${GITHUB_ENV}\n\n - if: contains(inputs.toolkit_type, 'cuda') && fromJSON(inputs.toolkit_short_version) >= 130\n run: |\n echo \"TORCH_CUDA_ARCH_LIST=$TORCH_CUDA_ARCH_LIST 8.0 9.0a 10.0a 10.3a 11.0a 12.0a 12.1a\" >> ${GITHUB_ENV}\n\n - if: runner.os == 'Windows'\n run: git config --system core.longpaths true\n - name: Recursive checkout\n uses: actions/checkout@v4\n with:\n submodules: recursive\n path: \".\"\n fetch-depth: 0 # for tags\n\n - name: HACKFIX for cutlass compiler bug\n if: runner.os == 'Windows'\n run: |\n # See https://github.com/NVIDIA/cutlass/issues/1732\n rm -f third_party/cutlass/include/cutlass/gemm/kernel/sm90_gemm_tma_warpspecialized_pingpong.hpp\n touch third_party/cutlass/include/cutlass/gemm/kernel/sm90_gemm_tma_warpspecialized_pingpong.hpp\n - name: Setup Runner\n uses: ./.github/actions/setup-build-cuda\n with:\n toolkit_type: ${{ inputs.toolkit_type }}\n toolkit_short_version: ${{ inputs.toolkit_short_version }}\n python: ${{ inputs.python }}\n - if: runner.os == 'Linux'\n run: printenv\n\n - if: runner.os != 'Windows'\n name: (Linux) Setup venv for linux\n shell: bash -l {0}\n run: |\n $PY -m venv venv\n . ./venv/bin/activate\n which pip\n echo \"PY=$(which python)\" >> ${GITHUB_ENV}\n echo \"PATH=$PATH\" >> ${GITHUB_ENV}\n git config --global --add safe.directory \"*\"\n pip install packaging ninja wheel setuptools twine\n\n - name: Define version\n id: xformers_version\n env:\n VERSION_SOURCE: ${{ github.ref_type == 'tag' && 'tag' || 'dev' }}\n run: |\n set -Eeuo pipefail\n git config --global --add safe.directory \"*\"\n version=`python .github/compute_wheel_version.py --source $VERSION_SOURCE`\n echo $version > version.txt\n echo \"BUILD_VERSION=$version${{ steps.cuda_info.outputs.CUDA_VERSION_SUFFIX }}\" >> ${GITHUB_ENV}\n echo \"BUILD_VERSION=$version${{ steps.cuda_info.outputs.CUDA_VERSION_SUFFIX }}\" >> ${GITHUB_OUTPUT}\n which ninja\n ninja --version\n cat ${GITHUB_ENV}\n - run: echo \"xformers-${BUILD_VERSION}\"\n - run: echo \"release version (will upload to PyTorch)\"\n if: ${{ !contains(steps.xformers_version.outputs.BUILD_VERSION, '.dev') }}\n\n - name: Install corresponding PyTorch\n run: |\n PYTORCH_INDEX_URL=\"https://download.pytorch.org/whl/${{ contains(inputs.toolkit_type, 'cuda') && 'cu' || 'rocm' }}${{ inputs.toolkit_short_version }}\"\n $PY -m pip install wheel -r requirements.txt --extra-index-url $PYTORCH_INDEX_URL\n\n - name: Build wheel\n shell: bash -l {0}\n run: |\n $PY setup.py bdist_wheel -d dist/ -k $PLAT_ARG\n env:\n PLAT_ARG: ${{ contains(inputs.os, 'ubuntu') && '--plat-name manylinux_2_28_x86_64' || '' }}\n\n - run: du -h dist/*\n - uses: actions/upload-artifact@v4\n with:\n name: ${{ inputs.os }}-py${{ inputs.python }}-torch${{ inputs.torch_version }}+${{ contains(inputs.toolkit_type, 'cuda') && 'cu' || 'rocm' }}${{ inputs.toolkit_short_version }}_${{ inputs.artifact_tag }}\n path: dist/*.whl\n# Note: it might be helpful to have additional steps that test if the built wheels actually work\n" }, { "path": ".github/workflows/wheels_upload_pip.yml", "content": "name: wheels_upload_pip\n\non:\n workflow_call:\n secrets:\n twine_password:\n required: true\n inputs:\n twine_username:\n required: true\n type: string\n pypirc:\n required: false\n type: string\n filter:\n required: true\n type: string\n description: Filter which runs to upload. Example '*+cu121*'\n execute:\n required: true\n type: boolean\n description: Actually upload the wheels. Dry-run if false\n artifact_tag:\n default: \"facebookresearch\"\n type: string\n\nenv:\n TWINE_USERNAME: __token__\n\njobs:\n wheels_upload_pip:\n name: wheels_upload_pip\n runs-on: ubuntu-24.04\n\n timeout-minutes: 360\n defaults:\n run:\n shell: bash\n steps:\n - name: Recursive checkout\n uses: actions/checkout@v3\n with:\n submodules: recursive\n path: \".\"\n fetch-depth: 0 # for tags\n\n # inspired by https://github.com/jlumbroso/free-disk-space/blob/main/action.yml\n - name: Free disk space\n run: |\n sudo rm -rf /usr/local/lib/android || true\n sudo rm -rf /usr/share/dotnet || true\n\n - name: Setup twine config\n if: inputs.pypirc\n run: |\n echo \"${{ inputs.pypirc }}\" > ~/.pypirc\n cat ~/.pypirc\n\n - uses: actions/download-artifact@v4\n with:\n path: dist\n\n # Filter builds (eg vN+cu118 for instance)\n - run: ls -R dist/\n - name: Extract builds to upload\n run: |\n set -ex\n mv dist all-dist\n mkdir dist\n for f in all-dist/${{ inputs.filter }}_${{ inputs.artifact_tag }}/*.whl; do\n cp $f dist/\n done;\n - run: ls -R dist/\n\n - name: Setup venv\n run: |\n python3 -m venv venv\n . ./venv/bin/activate\n which pip\n # (we need pytorch to create a source distr...)\n pip install torch packaging twine\n echo \"PY=$(which python)\" >> ${GITHUB_ENV}\n echo \"PATH=$PATH\" >> ${GITHUB_ENV}\n\n - name: Create source distribution\n env:\n VERSION_SOURCE: ${{ github.ref_type == 'tag' && 'tag' || 'dev' }}\n run: |\n version=`$PY .github/compute_wheel_version.py --source $VERSION_SOURCE`\n echo $version > version.txt\n cat version.txt\n\n BUILD_VERSION=$version $PY setup.py sdist -d sdist/\n\n - run: ls -R sdist/\n - name: Upload wheel to PyPi\n if: inputs.execute\n run: $PY -m twine upload --skip-existing dist/*.whl sdist/*\n env:\n TWINE_USERNAME: ${{ inputs.twine_username }}\n TWINE_PASSWORD: ${{ secrets.twine_password }}\n" }, { "path": ".github/workflows/wheels_upload_s3.yml", "content": "name: wheels_upload_s3\n\non:\n workflow_call:\n inputs:\n aws_role:\n required: true\n type: string\n s3_path:\n required: true\n type: string\n description: Example 's3://bucket/path/xformers/'\n aws_s3_cp_extra_args:\n required: false\n type: string\n default: ''\n description: Example '--acl public-read'\n filter:\n required: true\n type: string\n description: Filter which runs to upload. Example '*+cu121*'\n execute:\n required: true\n type: boolean\n description: Actually upload the wheels. Dry-run if false\n artifact_tag:\n default: \"facebookresearch\"\n type: string\n\njobs:\n wheels_upload_s3:\n permissions:\n id-token: write # Needed to assume AWS role\n pull-requests: read\n contents: read\n name: ${{ inputs.s3_path }}\n runs-on: ubuntu-24.04\n\n timeout-minutes: 360\n defaults:\n run:\n shell: bash\n steps:\n - name: Recursive checkout\n uses: actions/checkout@v3\n with:\n submodules: recursive\n path: \".\"\n fetch-depth: 0 # for tags\n\n # inspired by https://github.com/jlumbroso/free-disk-space/blob/main/action.yml\n - name: Free disk space\n run: |\n sudo rm -rf /usr/local/lib/android || true\n sudo rm -rf /usr/share/dotnet || true\n\n - uses: actions/download-artifact@v4\n with:\n path: dist\n # Filter builds (eg vN+cu118 for instance)\n - run: ls -R dist/\n - name: Extract builds to upload\n run: |\n set -ex\n mv dist all-dist\n mkdir dist\n for f in all-dist/${{ inputs.filter }}_${{ inputs.artifact_tag }}/*.whl; do\n cp $f dist/\n done;\n - run: ls -R dist/\n\n - name: configure aws credentials\n if: inputs.execute\n uses: aws-actions/configure-aws-credentials@v1.7.0\n with:\n role-to-assume: ${{ inputs.aws_role }}\n role-session-name: GitHub_CI\n aws-region: \"us-east-1\"\n\n - name: Sts GetCallerIdentity\n if: inputs.execute\n run: |\n aws sts get-caller-identity\n\n - name: Upload wheels to ${{ inputs.s3_path }}\n if: inputs.execute\n run: |\n set -ex\n for f in dist/*.whl; do\n echo $f;\n aws s3 cp $f ${{ inputs.s3_path }} ${{ inputs.aws_s3_cp_extra_args }}\n done;\n aws s3 ls ${{ inputs.s3_path }}\n" }, { "path": ".github/workflows/win-build.yml", "content": "name: win-build\n\non:\n pull_request:\n paths:\n - \"third_party/**\"\n - \"xformers/csrc/**\"\n - \".github/workflows/win-build.yml\"\n - \".github/actions/setup-build-cuda/action.yml\"\n - \"setup.py\"\n - \"requirements*.txt\"\n\nenv:\n FORCE_CUDA: 1\n MAX_JOBS: 6\n DISTUTILS_USE_SDK: 1 # otherwise distutils will complain on windows about multiple versions of msvc\n XFORMERS_BUILD_TYPE: \"Release\"\n TMPDIR: \"./x\"\n\njobs:\n win_build:\n strategy:\n fail-fast: false\n matrix:\n arch:\n - \"8.0\"\n name: win-build-${{ matrix.arch }}\n runs-on: windows-8-core\n env:\n PY: python3\n TORCH_CUDA_ARCH_LIST: ${{ matrix.arch }}\n\n timeout-minutes: 360\n defaults:\n run:\n shell: bash\n steps:\n - name: Workarounds for longpaths - git-config\n run: |\n git config --system core.longpaths true\n - name: Recursive checkout\n uses: actions/checkout@v3\n with:\n submodules: recursive\n path: \".\"\n\n - name: Workarounds for longpaths - TMPDIR\n run: |\n mkdir x\n python -c \"import tempfile; print(tempfile.gettempdir())\"\n python -c \"import tempfile; assert(len(tempfile.gettempdir()) < 30)\"\n\n - name: HACKFIX for cutlass compiler bug\n if: runner.os == 'Windows'\n run: |\n # See https://github.com/NVIDIA/cutlass/issues/1732\n rm -f third_party/cutlass/include/cutlass/gemm/kernel/sm90_gemm_tma_warpspecialized_pingpong.hpp\n touch third_party/cutlass/include/cutlass/gemm/kernel/sm90_gemm_tma_warpspecialized_pingpong.hpp\n\n - name: Setup Runner\n uses: ./.github/actions/setup-build-cuda\n with:\n toolkit_type: \"cuda\"\n toolkit_short_version: \"130\"\n python: \"3.10\"\n\n - name: Remove internal code\n run: |\n mkdir -p .github/sync.fairinternal/\n touch .github/sync.fairinternal/ossify.sh\n chmod +x .github/sync.fairinternal/ossify.sh\n .github/sync.fairinternal/ossify.sh\n\n - name: Install build dependencies\n run: |\n $PY -m pip install wheel setuptools ninja -r requirements.txt --extra-index-url https://download.pytorch.org/whl/cu130\n git config --global --add safe.directory \"*\"\n $PY -c \"import torch; print('torch', torch.__version__)\"\n $PY -c \"import torch; print('torch.cuda', torch.version.cuda)\"\n ninja --version\n\n - name: Create sdist\n run: $PY setup.py sdist\n\n - name: Build from sdist\n shell: bash -l {0}\n run: |\n $PY -m pip install -v --no-build-isolation dist/*\n\n - name: Info\n run: |\n cd ../../ # So we don't have a folder named `xformers`\n XFORMERS_MORE_DETAILS=1 $PY -m xformers.info\n\n # - name: Open an SSH session on failure to debug\n # if: ${{ failure() }}\n # uses: mxschmitt/action-tmate@v3\n" }, { "path": ".gitignore", "content": "*~\n*.swp\n\n*.pyc\n*.pyo\n*.so\n\n.mypy_cache/\n*.egg-info/\n\nbuild/\ndist/\n\n# for autocomplete\ncompile_commands.json\n\n# Pytest verbose output\ntest-results/\n\n# Coverage reports\n.coverage\n.coverage.*\n\n# Environments\n.env\n.venv\nenv/\nvenv/\nENV/\nenv.bak/\nvenv.bak/\n.vscode/*\nxformers/benchmarks/LRA/datasets\nxformers/benchmarks/LRA/logs\n\nmy_runs.md\n\n# Triton cache\n.cache\n\n# JetBrains PyCharm IDE\n.idea/\n\n# Pyre cache\n.pyre/\n\n# Watchman config files\n.watchmanconfig\n\n# examples demo files\nexamples/input.txt\nexamples/lightning_logs\nexamples/data\n\n# Hydra default output dir\nmultirun\noutputs\n\n.benchmarks\nxformers/version.py\nxformers/cpp_lib.json\n\n## temporary files\nxformers/csrc/attention/hip_fmha/*.cu\nxformers/csrc/attention/hip_fmha/*.hip\nxformers/csrc/attention/hip_fmha/*_hip.h\nxformers/csrc/attention/hip_fmha/instances/*.cu\nxformers/csrc/attention/hip_fmha/instances/*.hip\nxformers/csrc/attention/hip_fmha/instances/*_hip.h\nxformers/csrc/attention/hip_decoder/*.cu\nxformers/csrc/attention/hip_decoder/*.hip\nxformers/csrc/attention/hip_decoder/*_hip.h\n" }, { "path": ".gitmodules", "content": "[submodule \"third_party/cutlass\"]\n\tpath = third_party/cutlass\n\turl = https://github.com/NVIDIA/cutlass.git\n[submodule \"third_party/composable_kernel_tiled\"]\n\tpath = third_party/composable_kernel_tiled\n\turl = https://github.com/ROCm/composable_kernel.git\n\tbranch = develop\n" }, { "path": ".isort.cfg", "content": "[settings]\nknown_third_party =fvcore,hydra,input_pipeline,matplotlib,numpy,omegaconf,pandas,pl_bolts,pyre_extensions,pytest,pytorch_lightning,ragged_inference,recommonmark,seaborn,setuptools,sklearn,submitit,tensorflow,timm,torch,torchmetrics,torchvision,tqdm,triton,typing_extensions\nskip_glob=third_party/*\n" }, { "path": ".markdownlint.json", "content": "{\n \"MD013\": false,\n \"MD033\": false\n}\n" }, { "path": ".pre-commit-config.yaml", "content": "exclude: 'build|stubs'\n\ndefault_language_version:\n python: python3\n\nrepos:\n- repo: https://github.com/pre-commit/pre-commit-hooks\n rev: v3.4.0\n hooks:\n - id: trailing-whitespace\n - id: check-ast\n - id: check-merge-conflict\n - id: no-commit-to-branch\n args: ['--branch=master']\n - id: check-added-large-files\n args: ['--maxkb=500']\n - id: end-of-file-fixer\n\n- repo: https://github.com/omnilib/ufmt\n rev: v2.8.0\n hooks:\n - id: ufmt\n additional_dependencies:\n - black == 26.3.1\n - usort == 1.0.8.post1\n\n- repo: https://github.com/pycqa/flake8\n rev: 6.1.0\n hooks:\n - id: flake8\n additional_dependencies: [flake8-copyright]\n\n- repo: https://github.com/pre-commit/mirrors-mypy\n rev: 'v1.10.0'\n hooks:\n - id: mypy\n" }, { "path": ".pyre_configuration", "content": "{\n \"ignore_all_errors\": [\"xformers/benchmarks/\"],\n \"python_version\": \"3.9\",\n \"source_directories\": [\n \"stubs\",\n {\"import_root\": \".\", \"source\": \"xformers\"}\n ]\n}\n" }, { "path": "CHANGELOG.md", "content": "# Changelog\nAll notable changes to this project will be documented in this file.\n\nThe format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),\nand this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).\n\n## [0.0.36] - 2026-??-??\n\n\n## [0.0.35] - 2026-02-20\nPre-built binary wheels are available for PyTorch 2.10.0 (and later).\n\n### Improved\n- Supported free-threading Python.\n\n### Removed\n- Stopped bundling pre-built versions of Flash-Attention 3, and instead started relying on the wheels provided by the PyTorch indices.\n\n\n## [0.0.34] - 2026-01-23\nPre-built binary wheels are available for PyTorch 2.10.0 (and later).\n\n### Improved\n- Migrated xFormers to the PyTorch stable API/ABI, which means that binary builds targeting PyTorch 2.10+ will be compatible with any later version\n\n### Removed\n- Removed optimized fast-path of SwiGLU (which was only available for A100 GPUs)\n- Removed most legacy components\n\n\n## [0.0.33.post2] - 2025-12-03\nPre-built binary wheels are available for PyTorch 2.9.1.\n\n\n## [0.0.33.post1] - 2025-11-13\nFixed issues with wheel upload to PyPI\n\n\n## [0.0.33] - 2025-11-12\nPre-built binary wheels are available for PyTorch 2.9.0.\n\n### Added\n- cutlass fmha Op for Blackwell GPUs\n- Support flash-attention package up to 2.8.3\n- expose FA3 deterministic mode\n- FW+BW pass overlap for DeepSeek-like comms/compute overlap\n\n### Improved\n- merge_attentions support for irregular head dimension\n\n\n## [0.0.32] - 2025-08-13\nPre-built binary wheels are available for PyTorch 2.8.0.\n\n### Added\n- Support flash-attention package up to 2.8.2\n- Speed improvements to `python -m xformers.profiler.find_slowest`\n\n### Removed\n- Removed autograd backward pass for merge_attentions as it is easy to use incorrectly.\n- Attention biases are no longer `torch.Tensor` subclasses. This is no longer\nnecessary for torch.compile to work, and was adding more complexity\n\n\n## [0.0.31] - 2025-06-25\nPre-built binary wheels are available for PyTorch 2.7.1.\n### Added\n- xFormers wheels are now python-version agnostic: this means that the same wheel can be used for python 3.9, 3.10, ... 3.13\n- Added support for Flash-Attention 3 on Ampere GPUs\n### Removed\n- We will no longer support V100 or older GPUs, following PyTorch (pytorch/pytorch#147607)\n- Deprecated support for building Flash-Attention 2 as part of xFormers. For Ampere GPUs, we now use Flash-Attention 3 on windows, and Flash-Attention 2 can still be used through PyTorch on linux.\n\n## [0.0.30] - 2025-04-28\nPre-built binary wheels are available for PyTorch 2.7.0. Following PyTorch, we build wheels for CUDA 11.8, 12.6, and 12.8 only (we no longer build for CUDA 12.4).\nxFormers now requires PyTorch >= 2.7\n### Added\n- [fMHA] Added support for local attention on the Flash3 backend (H100)\n- [fMHA] Added a new paged gappy attention bias\n### Improved\n- [fMHA] The FlashAttention3 backend now ships with more head dimensions to support MLA, and with a FLOPs formula in order to be compatible with PyTorch's partitioner-base automatic activation checkpointing\n- The fused operators for sequence parallelism were migrated to PyTorch's SymmetricMemory\n- The profiler prepends the traces' filenames with the rank of the process when doing distributed training\n### Removed\n- Removed documentation for legacy unmaintained components\n\n## [0.0.29.post2] - 2025-01-31\nPre-built binary wheels are available for PyTorch 2.6.0. Following PyTorch, we build wheels for CUDA 11.8, 12.4, and 12.6 only (we no longer build for CUDA 12.1).\nxFormers now requires PyTorch >= 2.6\n\n\n## [0.0.29] - 2024-12-27\n### Improved:\n- [fMHA] Creating a `LowerTriangularMask` no longer creates a CUDA tensor\n- [fMHA] Updated Flash-Attention to `v2.7.2.post1`\n- [fMHA] Flash-Attention v3 will now be used by `memory_efficient_attention` by default when available, unless the operator is enforced with the `op` keyword-argument. Switching from Flash2 to Flash3 can make transformer trainings ~10% faster end-to-end on H100s\n- [fMHA] Fixed a performance regression with the `cutlass` backend for the backward pass (facebookresearch/xformers#1176) - mostly used on older GPUs (eg V100)\n- Fixed swiglu operator compatibility with torch-compile with PyTorch 2.6\n- Fix activation checkpointing of SwiGLU when AMP is enabled (facebookresearch/xformers#1152)\n### Removed:\n- Following PyTorch, xFormers no longer builds binaries for conda. Pip is now the only recommended way to get xFormers\n- Removed unmaintained/deprecated components in `xformers.components.*` (see facebookresearch/xformers#848)\n\n## [0.0.28.post3] - 2024-10-30\nPre-built binary wheels require PyTorch 2.5.1\n\n## [0.0.28.post2] - 2024-10-18\nPre-built binary wheels require PyTorch 2.5.0\n\n## [0.0.28.post1] - 2024-09-13\nProperly upload wheels for cuda 12.4\n\n## [0.0.28] - 2024-09-12\nPre-built binary wheels require PyTorch 2.4.1\n### Added\n- Added wheels for cuda 12.4\n- Added conda builds for python 3.11\n- Added wheels for rocm 6.1\n### Improved\n- Profiler: Fix computation of FLOPS for the attention when using xFormers\n- Profiler: Fix MFU/HFU calculation when multiple dtypes are used\n- Profiler: Trace analysis to compute MFU & HFU is now much faster\n- fMHA/splitK: Fixed `nan` in the output when using a `torch.Tensor` bias where a lot of consecutive keys are masked with `-inf`\n- Update Flash-Attention version to `v2.6.3` *when building from scratch*\n- When using the most recent version of Flash-Attention, it is no longer possible to mix it with the cutlass backend. In other words, it is no longer possible to use the cutlass Fw with the flash Bw.\n### Removed\n- fMHA: Removed `decoder` and `small_k` backends\n- profiler: Removed `DetectSlowOpsProfiler` profiler\n- Removed compatibility with PyTorch < 2.4\n- Removed conda builds for python 3.11\n- Removed windows pip wheels for cuda 12.1 and 11.8\n\n## [0.0.27.post2] - 2024-07-26\nPre-built binary wheels require PyTorch 2.4.0\n\n## [0.0.27.post1] - 2024-07-25\nPre-built binary wheels require PyTorch 2.4.0\n\n## [0.0.27] - 2024-07-10\nPre-built binary wheels require PyTorch 2.3.1\n### Added\n- fMHA: `PagedBlockDiagonalGappyKeysMask`\n- fMHA: heterogeneous queries in `triton_splitk`\n- fMHA: support for paged attention in flash\n- fMHA: Added backwards pass for `merge_attentions`\n- fMHA: Added `torch.compile` support for 3 biases (`LowerTriangularMask`, `LowerTriangularMaskWithTensorBias` and `BlockDiagonalMask`) - some might require PyTorch 2.4\n- fMHA: Added `torch.compile` support in `memory_efficient_attention` when passing the flash operator explicitely (eg `memory_efficient_attention(..., op=(flash.FwOp, flash.BwOp))`)\n- fMHA: `memory_efficient_attention` now expects its `attn_bias` argument to be on the same device as the other input tensor. Previously, it would convert the bias to the right device.\n- fMHA: `AttentionBias` subclasses are now constructed by default on the `cuda` device if available - they used to be created on the CPU device\n- 2:4 sparsity: Added `xformers.ops.sp24.sparsify24_ste` for Straight Through Estimator (STE) with options to rescale the gradient differently for masked out/kept values\n### Improved\n- fMHA: Fixed out-of-bounds reading for Split-K triton implementation\n- Profiler: fix bug with modules that take a single tuple as argument\n- Profiler: Added manual trigger for a profiling step, by creating a `trigger` file in the profiling directory\n### Removed\n- Removed support for PyTorch version older than 2.2\n\n## [0.0.26] - 2024-04-29\nPre-built binary wheels require PyTorch 2.3.0\n### Added\n- [2:4 sparsity] Added support for Straight-Through Estimator for `sparsify24` gradient (`GRADIENT_STE`)\n- [2:4 sparsity] `sparsify24_like` now supports the cuSparseLt backend, and the STE gradient\n- Basic support for `torch.compile` for the `memory_efficient_attention` operator. Currently only supports Flash-Attention, and without any bias provided. We want to expand this coverage progressively.\n### Improved\n- merge_attentions no longer needs inputs to be stacked.\n- fMHA: triton_splitk now supports additive bias\n- fMHA: benchmark cleanup\n\n## [0.0.25.post1] - 2024-03-29\nPre-built binary wheels require PyTorch 2.2.2\n\n## [0.0.25] - 2024-03-14\nPre-built binary wheels require PyTorch 2.2.1\n### Added\n- New `merge_attentions` function\n- fMHA: New gappy attention biases.\n### Improved\n- fMHA: Updated Flash-Attention to v2.5.6: this has a performance improvement for multiquery.\n- fMHA: triton_splitk changed and expanded. Now amalgamates using LSE. Can autotune, supports causal with a small number of queries - not just 1. Experimental support for paged attention.\n- `rope_padded`: Fixed CUDA error with many queries (more than 65k)\n- `rmsnorm`: Fixed CUDA error with large inputs (enables 512k+ sequence length on Llama2 70B)\n### Removed\n- fMHA: Removed triton operator (`fmha.triton.*`, `xformers.ops.MemoryEfficientAttentionTritonFwdFlashBwOp`, `xformers.ops.TritonFlashAttentionOp`), as it has correctness issues under some conditions, and is slower than other implementations.\n\n## [0.0.24] - 2024-01-31\nPre-built binary wheels require PyTorch 2.2.0\n### Added\n- Added components for model/sequence parallelism, as near-drop-in replacements for FairScale/Megatron Column&RowParallelLinear modules. They support fusing communication and computation for sequence parallelism, thus making the communication effectively free. [Read more](https://twitter.com/d_haziza/status/1753030654118211593)\n- Added kernels for training models with 2:4-sparsity. We introduced a very fast kernel for converting a matrix A into 24-sparse format, which can be used during training to sparsify weights dynamically, activations etc... xFormers also provides an API that is compatible with torch-compile, see `xformers.ops.sparsify24`.\n### Improved\n- Make selective activation checkpointing be compatible with torch.compile.\n### Removed\n- Triton kernels now require a GPU with compute capability 8.0 at least (A100 or newer). This is due to newer versions of triton not supporting older GPUs correctly\n- Removed support for PyTorch version older than 2.1.0\n\n## [0.0.23] - 2023-12-05\nPre-built binary wheels require PyTorch 2.1.1 (xFormers `0.0.23`) or PyTorch 2.1.2 (xFormers `0.0.23.post1`).\n### Fixed\n- fMHA: Fixed a bug in cutlass backend forward pass where the logsumexp was not correctly calculated, resulting in wrong results in the BW pass. This would happen with MQA when one sequence has a query with `length%64 == 1`\n- fMHA: Updated Flash-Attention to v2.3.6 - this fixes a performance regression in causal backward passes, and now supports `BlockDiagonalCausalWithOffsetPaddedKeysMask`\n### Added\n- fMHA: Added `LocalAttentionFromBottomRightMask` (local)\n- fMHA: Added `LowerTriangularFromBottomRightMask` (causal)\n- fMHA: Added `LowerTriangularFromBottomRightLocalAttentionMask` (local + causal)\n### Removed\n- Removed `xformers.triton.sum_strided`\n\n## [0.0.22] - 2023-09-27\n### Fixed\n- fMHA: Backward pass now works in PyTorch deterministic mode (although slower)\n### Added\n- fMHA: Added experimental support for Multi-Query Attention and Grouped-Query Attention. This is handled by passing 5-dimensional inputs to `memory_efficient_attention`, see the documentation for more details\n- fMHA: Added experimental support for Local Attention biases to `memory_efficient_attention`\n- Added an example of efficient [LLaMa decoding](https://github.com/facebookresearch/xformers/tree/main/examples/llama_inference) using xformers operators\n- Added Flash-Decoding for faster attention during Large Language Model (LLM) decoding - up to 50x faster for long sequences (token decoding up to 8x faster end-to-end)\n- Added an efficient rope implementation in triton, to be used in LLM decoding\n- Added selective activation checkpointing, which gives fine-grained control of which activations to keep and which activations to recompute\n- `xformers.info` now indicates the Flash-Attention version used\n### Removed\n- fMHA: Removed `smallK` backend support for CPU. `memory_efficient_attention` only works for CUDA/GPU tensors now\n- **DEPRECATION**: Many classes in `xformers.factory`, `xformers.triton` and `xformers.components` have been or will be deprecated soon (see tracking issue facebookresearch/xformers#848)\n\n## [0.0.21] - 2023-08-18\n### Improved\n- fMHA: Updated [flash-attention](https://github.com/Dao-AILab/flash-attention) to v2, with massive performance improvements for both the forward pass and backward pass. This implementation is now used by default when it's available\n### Bug fixes\n- fMHA/cutlass: Fix potential race condition in the FW/BW passes\n- fMHA/cutlass: Fix `attn_bias` stride overflow for very long sequences (>32k)\n- `LowerTriangularMask` is now backward compatible with older xformers versions\n### Breaking changes\n- `memory_efficient_attention` now expects the `attn_bias` argument to have a head dimension\n- `memory_efficient_attention` no longer broadcasts the batch/head dimensions of `attn_bias`. Please use `.expand` if you need to broadcast the bias\n- Remove `causal_diagonal` argument from `BlockDiagonalCausalWithOffsetPaddedKeysMask`\n### Added\n- Binary wheels on pypi/conda now contain H100 kernels\n- fMHA: Added backend specialized for decoding that does not use TensorCores - useful when not using multiquery\n\n**NOTE**: Binary wheels are now provided only for PyTorch 2 with cuda 11.8. It is still possible to use xFormers with older versions of PyTorch by building from source or using conda.\n\n\n## [0.0.20] - 2023-05-23\n### Improved\n- fMHA/cutlass (backward): Massive performance improvements when `batch_size * num_heads` is low (10x+)\n- fMHA/cutlass: Further performance improvements for both the forward & backward kernels\n- fMHA (backward): Now dispatching to cutlass when `embed_dim>64`\n- fMHA: Updated Flash-Attention to `v1.0.5`\n### Added\n- fMHA now runs on H100 (support is experimental)\n\n## [0.0.19] - 2023-04-28\n### Added\n- Display `nvcc` version used to compile `xformers` in `python -m xformers.info`\n\n### Fixed\n- Fixed performance regression with `nvcc>11.6` (facebookresearch/xformers#712)\n- fMHA/cutlass: Fixed `nan` in the output when using a `torch.Tensor` with `-inf` prefixes as `attn_bias` (facebookresearch/xformers#722)\n- fMHA/cutlass: Fixed `nan` in the output when the sequence length is larger than `2 ** 15` (facebookresearch/xformers#719)\n- fMHA/cutlass: Significative performance improvements (up to 2x) for both the forward pass and backward pass\n- fMHA/cutlass: The kernel are now deterministic\n- fMHA/cutlass: Fixed backward pass correctness when using dropout (facebookresearch/xformers#724)\n\n## [0.0.18] - 2023-03-31\n### Added\n- Added `xformers.ops.index_select_cat` and `xformers.ops.scaled_index_add` - those are experimental functions that only work with a few shapes, and can be used to write efficient stochastic depth in transformer architectures for instance\n\n### Fixed\n- fMHA: `memory_efficient_attention` now accepts `torch.Tensor` as attention bias for any seqlen, although there are still requirements on the alignment of the bias tensor (see facebookresearch/xformers#683)\n\n## [0.0.17] - 2023-03-28\n### Fixed\n- fMHA: Fixed BW pass on Sm86/Sm89 GPUs when `K > 64` (RTX 3090, RTX 4090, A6000, ..) [facebookresearch/xformers#631]\n\n### Added\n- fMHA/CUTLASS: Added tensor attn bias support [facebookresearch/xformers#587] - contribution from [@jfc4050](https://github.com/jfc4050)\n- fMHA/CUTLASS: Added tensor attn bias grad support [facebookresearch/xformers#587] - contribution from [@jfc4050](https://github.com/jfc4050)\n- fMHA/CUTLASS: Added dropout support [facebookresearch/xformers#587] - contribution from [@jfc4050](https://github.com/jfc4050)\n- fMHA: Added support for varying sequence lengths [facebookresearch/xformers#500]\n\n\n## [0.0.16] - 2023-01-31\n### Fixed\n- Updated triton dependency [facebookresearch/xformers#418]\n- Stripe lineinfo from binaries, reducing the binary size [facebookresearch/xformers#549]\n- Added support for pip wheels [facebookresearch/xformers#588, facebookresearch/xformers#573, facebookresearch/xformers#534, facebookresearch/xformers#523, ...] big thanks to [@AbdBarho](https://github.com/AbdBarho)!\n- Fixed compatibility with Python 3.7 [facebookresearch/xformers#541] - thanks to [@susumuota](https://github.com/susumuota)\n- fMHA: Fixed strides for QKV gradients for cutlass attention [facebookresearch/xformers#535]\n- fMHA: Stricter inputs validation to avoid CUDA errors for unsupported inputs [facebookresearch/xformers#592]\n- fMHA/Flash-Attention: Updated to https://github.com/HazyResearch/flash-attention/commit/a1f49a2b92b6fa022379bbebafed9d7f5e96a675 with multiple changes from [@TriDao](https://github.com/tridao) that make the operator up to 20% faster\n- fMHA/Flash-Attention: Fixed backward pass wrapper, where non-contiguous gradients could give the wrong result [facebookresearch/xformers#548]\n- fMHA: Separate each operator into forward and backward operators. It's now possible to use any combination of forward+backward (for instance Triton forward and Flash-Attention backward) [facebookresearch/xformers#560]\n\n### Added\n- fMHA: Added Triton operator for forward pass from [Flash-Attention](https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attn_triton.py) authored by [@TriDao](https://github.com/tridao), will be automatically used on A100 when compatible\n- fMHA: Added [`xformers.ops.memory_efficient_attention_forward`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention_forward), [`xformers.ops.memory_efficient_attention_forward_requires_grad`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention_forward_requires_grad), [`xformers.ops.memory_efficient_attention_backward`](https://facebookresearch.github.io/xformers/components/ops.html#xformers.ops.memory_efficient_attention_backward) for power-users who write custom autograd functions [facebookresearch/xformers#560]\n- fMHA: Support for custom scaling for the CUTLASS-based kernel [facebookresearch/xformers#530] - contribution from [@comaniac](https://github.com/comaniac)\n\n## [0.0.15] - Skipped\n\n## [0.0.14] - 2022-11-10\n### Fixed\n- fMHA/CUTLASS: The current CUDA stream is now used by the kernel [facebookresearch/xformers#491]\n- fMHA/CUTLASS: Improve overall performance\n\n### Added\n- SwiGLU: Added `xformers.ops.SwiGLU` and its functional counterpart (`xformers.ops.swiglu`) [facebookresearch/xformers#490]\n- fMHA: Possible to combine CUTLASS's forward with flash-attention's backward pass [facebookresearch/xformers#469] - improves performance on A100 for K = 128\n- fMHA: Add custom `xformers.ops.unbind` operator to avoid a cat in the attention block [facebookresearch/xformers#458]\n\n## [0.0.13] - 2022-09-26\n### Added\n- fMHA: Added CUTLASS-based kernel for `xformers.ops.memory_efficient_attention`. This kernel is automatically depending on the inputs, and works on any GPU after P100 [facebookresearch/xformers#362]\n\n## [0.0.12] - 2022-08-08\n### Fixed\n- Removed duplicated biases in the FusedMLP layers [facebookresearch/xformers#317]\n- Rotary embeddings respecting input types [facebookresearch/xformers#326]\n- Poolformer style instantiating useless projection layers [facebookresearch/xformers#349]\n- Fix layer position not being properly tracked, causing extra layernorms for programmatic xformers [facebookresearch/xformers#348]\n- Pass use_triton flag to LayerNorm module [facebookresearch/xformers#336]\n\n### Added\n- Four blocksparsity layouts from DeepSpeed [facebookresearch/xformers#320]\n- Support several initialization options [facebookresearch/xformers#312]\n- Conv2DFeedforward feedforward part [facebookresearch/xformers#321]\n- VisualAttention [facebookresearch/xformers#329]\n- Automatic blocksparse for causal attention [facebookresearch/xformers#334]\n- Better hierarchical transformer generation [facebookresearch/xformers#345]\n- Fused operations with AOTAutograd/NVFuser, integration into MLP [facebookresearch/xformers#357]\n- Refactor LRA code to use Pytorch Lightning [facebookresearch/xformers#343]\n\n## [0.0.11] - 2022-05-30\n### Fixed\n- Fix some torchscriptability [facebookresearch/xformers#246]\n- Fix FourierMix being compatible with AMP [facebookresearch/xformers#258]\n- Better asserts on QKV dimensions [facebookresearch/xformers#264]\n- Better perfs for FusedMLP and FusedLinearLayer [facebookresearch/xformers#283]\n- Deepnorm init missing self-attention [facebookresearch/xformers#284]\n\n### Added\n- Simplicial Embeddings [facebookresearch/xformers#259]\n- Mem efficient attention, FW pass [facebookresearch/xformers#267]\n- MHA benchmark\n- MLP benchmark\n- Move all triton kernels to triton v2 [facebookresearch/xformers#272]\n- Mem efficient attention, BW pass [facebookresearch/xformers#281]\n- Metaformer support [facebookresearch/xformers#294]\n\n## [0.0.10] - 2022-03-14\n### Fixed\n- Expose bias flag for feedforwards, same default as Timm [facebookresearch/xformers#220]\n- Update eps value for layernorm, same default as torch [facebookresearch/xformers#221]\n- PreNorm bugfix, only one input was normalized [facebookresearch/xformers#233]\n- Fix bug where embedding dimensions that did not match model dim would lead to a crash [facebookresearch/xformers#244]\n\n### Added\n- Add DeepNet (DeepNorm) residual path and init [facebookresearch/xformers#227]\n\n## [0.0.9] - 2022-02-09\n### Added\n- Compositional Attention [facebookresearch/xformers#41]\n- Experimental Ragged attention [facebookresearch/xformers#189]\n- Mixture of Experts [facebookresearch/xformers#181]\n- BlockSparseTensor [facebookresearch/xformers#202]\n- Nd-tensor support for triton softmax [facebookresearch/xformers#210]\n\n### Fixed\n- Bugfix Favor, single feature map [facebookresearch/xformers#183]\n- Sanity check blocksparse settings [facebookresearch/xformers#207]\n- Fixed some picklability [facebookresearch/xformers#204]\n\n## [0.0.8] - 2022-01-07\n### Fixed\n- Much faster fused dropout [facebookresearch/xformers#164]\n- Fused dropout repeatability [facebookresearch/xformers#173]\n\n### Added\n- Embedding weight tying option [facebookresearch/xformers#172]\n\n## [0.0.7] - 2021-11-30\n### Fixed\n- Dropout setting not properly passed in many attentions [facebookresearch/xformers#123]\n\n## [0.0.6] - 2021-11-24\n### Fixed\n- Fix self attention optimization not being triggered, broken residual path [facebookresearch/xformers#119]\n- Improve speed by not using contiguous Tensors when not needed [facebookresearch/xformers#119]\n\n### Added\n- Attention mask wrapper [facebookresearch/xformers#113]\n- ViT comparison benchmark [facebookresearch/xformers#117]\n\n## [0.0.4] - 2021-11-16\n### Fixed\n- Homogenizing the masks, additive or bool [facebookresearch/xformers#79][facebookresearch/xformers#85][facebookresearch/xformers#86]\n- Fix causality flag not being respected [facebookresearch/xformers#103]\n- Enabling FusedLayerNorm by default in the factory if Triton is available\n- Fixing Favor with fp16\n- Fixing Favor trainability\n\n### Added\n- Fused dropout/bias/activation layer [facebookresearch/xformers#58]\n- Fused layernorm used by default in the factory [facebookresearch/xformers#92]\n\n\n## [0.0.3] - 2021-11-01\n### Fixed\n- Nystrom causal attention [facebookresearch/xformers#75]\n\n\n## [0.0.2] - 2021-11-01\n### Fixed\n- More robust blocksparse [facebookresearch/xformers#24]\n\n### Added\n- Rotary embeddings [facebookresearch/xformers#32]\n- More flexible layernorm [facebookresearch/xformers#50]\n" }, { "path": "CODE_OF_CONDUCT.md", "content": "# Code of Conduct\n\n## Our Pledge\n\nIn the interest of fostering an open and welcoming environment, we as\ncontributors and maintainers pledge to make participation in our project and\nour community a harassment-free experience for everyone, regardless of age, body\nsize, disability, ethnicity, sex characteristics, gender identity and expression,\nlevel of experience, education, socio-economic status, nationality, personal\nappearance, race, religion, or sexual identity and orientation.\n\n## Our Standards\n\nExamples of behavior that contributes to creating a positive environment\ninclude:\n\n* Using welcoming and inclusive language\n* Being respectful of differing viewpoints and experiences\n* Gracefully accepting constructive criticism\n* Focusing on what is best for the community\n* Showing empathy towards other community members\n\nExamples of unacceptable behavior by participants include:\n\n* The use of sexualized language or imagery and unwelcome sexual attention or\nadvances\n* Trolling, insulting/derogatory comments, and personal or political attacks\n* Public or private harassment\n* Publishing others' private information, such as a physical or electronic\naddress, without explicit permission\n* Other conduct which could reasonably be considered inappropriate in a\nprofessional setting\n\n## Our Responsibilities\n\nProject maintainers are responsible for clarifying the standards of acceptable\nbehavior and are expected to take appropriate and fair corrective action in\nresponse to any instances of unacceptable behavior.\n\nProject maintainers have the right and responsibility to remove, edit, or\nreject comments, commits, code, wiki edits, issues, and other contributions\nthat are not aligned to this Code of Conduct, or to ban temporarily or\npermanently any contributor for other behaviors that they deem inappropriate,\nthreatening, offensive, or harmful.\n\n## Scope\n\nThis Code of Conduct applies within all project spaces, and it also applies when\nan individual is representing the project or its community in public spaces.\nExamples of representing a project or community include using an official\nproject e-mail address, posting via an official social media account, or acting\nas an appointed representative at an online or offline event. Representation of\na project may be further defined and clarified by project maintainers.\n\nThis Code of Conduct also applies outside the project spaces when there is a\nreasonable belief that an individual's behavior may have a negative impact on\nthe project or its community.\n\n## Enforcement\n\nInstances of abusive, harassing, or otherwise unacceptable behavior may be\nreported by contacting the project team at . All\ncomplaints will be reviewed and investigated and will result in a response that\nis deemed necessary and appropriate to the circumstances. The project team is\nobligated to maintain confidentiality with regard to the reporter of an incident.\nFurther details of specific enforcement policies may be posted separately.\n\nProject maintainers who do not follow or enforce the Code of Conduct in good\nfaith may face temporary or permanent repercussions as determined by other\nmembers of the project's leadership.\n\n## Attribution\n\nThis Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,\navailable at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html\n\n[homepage]: https://www.contributor-covenant.org\n\nFor answers to common questions about this code of conduct, see\nhttps://www.contributor-covenant.org/faq\n" }, { "path": "CONTRIBUTING.md", "content": "# Contributing to the xFormers repo\n\nWe want to make contributing to this project as easy and transparent as\npossible.\n\n## Our Development Process\n\nMinor changes and improvements will be released on an ongoing basis. Larger\nchanges (e.g., changesets implementing a new paper) will be released on a\nmore periodic basis.\n\n## Pull Requests\n\nWe actively welcome your pull requests.\n\n1. Fork the repo and create your branch from `main`.\n2. If you've added code that should be tested, add tests.\n3. If you've changed APIs, update the documentation.\n4. Ensure the test suite passes.\n5. Make sure your code lints.\n6. If you haven't already, complete the Contributor License Agreement (\"CLA\").\n\n## Contributor License Agreement (\"CLA\")\n\nIn order to accept your pull request, we need you to submit a CLA. You only need\nto do this once to work on any of Facebook's open source projects.\n\nComplete your CLA here: \n\n## Issues\n\nWe use GitHub issues to track public bugs. Please ensure your description is\nclear and has sufficient instructions to be able to reproduce the issue.\n\nFacebook has a [bounty program](https://www.facebook.com/whitehat/) for the safe\ndisclosure of security bugs. In those cases, please go through the process\noutlined on that page and do not file a public issue.\n\n## Environment setup\n\n```bash\n~$ python3 -m venv venv2\n~$ source venv2/bin/activate\n(venv2) ~$ cd git/template/\n(venv2) ~/git/template $ pip3 install -r requirements-test.txt\n```\n\n## Coding Style\n\nIn your editor, install the [editorconfig](https://editorconfig.org/) extension\nwhich should ensure that you are following the same standards as us.\n\nTwo options to make sure that the code is formatted and linted properly:\n* either you run mypy and ufmt before opening up your PR.\n\n```bash\nufmt format\nflake8 --config .flake8\nmypy --ignore-missing-imports --scripts-are-modules --pretty --exclude build/ --exclude stubs/ .\n```\n\n* or you can just install [pre-commit](https://pre-commit.com/), which will make sure that all of the above is run automatically anytime you commit\nin that case, you would need to\n```bash\npip install pre-commit\n```\nthen (in the xformers repository, just once)\n```bash\npre-commit install\n```\n\nAfter these steps each of your commits will run the same linting and formatting routines as the xformers continuous integration, which greatly helps getting your PRs all green !\n\n_Read the [editorconfig](.editorconfig) file to understand the exact coding style preferences._\n\n## Testing\n\n### Static analysis\n\n```bash\nmypy --ignore-missing-imports --scripts-are-modules --pretty --exclude stubs/ .\n```\n\n### Unit tests\n\n```bash\npytest\n```\n\nor\n\n``` bash\npython -m pytest\n```\n\n### Check test coverage\n\n``` bash\npython -m pytest --cov-report term --cov=template tests\n```\n\n### CircleCI status\n\nFrom your PR page, you can expand on the CircleCI results. For GPU test, you should see\nwhat CI has run, like:\n\n``` bash\n...\n----- generated xml file: /home/circleci/template/test-results/junit.xml ------\n================== 217 passed, 2 xfailed in 218.74s (0:03:38) ==================\nCircleCI received exit code 0\n```\n\nThe number of passed and failed should give you an idea on whether your local\ntest was the same or not.\n\n## Commit Guidelines\n\nWe follow the same guidelines as AngularJS. Each commit message consists of a **header**,\na **body** and a **footer**. The header has a special format that includes a **type**,\nand a **subject**:\n\n```bash\n[] \n\n\n\n