frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 12:54:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

Update (base update)

Browse Source 
[ghstack-poisoned]
This commit is contained in:
Xuehai Pan
2025-10-19 18:28:29 +08:00
parent ecd4542830 57ba575242
commit 98aff4e90e
1212 changed files with 29636 additions and 9583 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
.ci/aarch64_linux/aarch64_ci_build.sh
View File

@ -8,6 +8,8 @@ if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
export TORCH_CUDA_ARCH_LIST="8.0;9.0"
elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
elif [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
fi

2
.ci/docker/build.sh
View File

@ -113,6 +113,7 @@ case "$tag" in
UCX_COMMIT=${_UCX_COMMIT}
UCC_COMMIT=${_UCC_COMMIT}
TRITON=yes
INSTALL_MINGW=yes
;;
pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11)
CUDA_VERSION=13.0.0
@ -361,6 +362,7 @@ docker build \
--build-arg "OPENBLAS=${OPENBLAS:-}" \
--build-arg "SKIP_SCCACHE_INSTALL=${SKIP_SCCACHE_INSTALL:-}" \
--build-arg "SKIP_LLVM_SRC_BUILD_INSTALL=${SKIP_LLVM_SRC_BUILD_INSTALL:-}" \
--build-arg "INSTALL_MINGW=${INSTALL_MINGW:-}" \
-f $(dirname ${DOCKERFILE})/Dockerfile \
-t "$tmp_tag" \
"$@" \

10
.ci/docker/common/install_mingw.sh Normal file
View File

@ -0,0 +1,10 @@
#!/bin/bash
set -ex
# Install MinGW-w64 for Windows cross-compilation
apt-get update
apt-get install -y g++-mingw-w64-x86-64-posix
echo "MinGW-w64 installed successfully"
x86_64-w64-mingw32-g++ --version

2
.ci/docker/common/install_onnx.sh
View File

@ -20,7 +20,7 @@ pip_install \
pip_install coloredlogs packaging
pip_install onnxruntime==1.23.0
pip_install onnxscript==0.5.3
pip_install onnxscript==0.5.4
# Cache the transformers model to be used later by ONNX tests. We need to run the transformers
# package to download the model. By default, the model is cached at ~/.cache/huggingface/hub/

6
.ci/docker/libtorch/build.sh
View File

@ -39,9 +39,13 @@ case ${DOCKER_TAG_PREFIX} in
DOCKER_GPU_BUILD_ARG=""
;;
rocm*)
# we want the patch version of 7.0 instead
if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
fi
# we want the patch version of 6.4 instead
if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
fi
BASE_TARGET=rocm
GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete

6
.ci/docker/manywheel/build.sh
View File

@ -75,9 +75,13 @@ case ${image} in
DOCKERFILE_SUFFIX="_cuda_aarch64"
;;
manylinux2_28-builder:rocm*)
# we want the patch version of 7.0 instead
if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
fi
# we want the patch version of 6.4 instead
if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.4"
fi
TARGET=rocm_final
MANY_LINUX_VERSION="2_28"

5
.ci/docker/ubuntu/Dockerfile
View File

@ -103,6 +103,11 @@ COPY ci_commit_pins/torchbench.txt torchbench.txt
RUN if [ -n "${INDUCTOR_BENCHMARKS}" ]; then bash ./install_inductor_benchmark_deps.sh; fi
RUN rm install_inductor_benchmark_deps.sh common_utils.sh timm.txt huggingface-requirements.txt torchbench.txt
ARG INSTALL_MINGW
COPY ./common/install_mingw.sh install_mingw.sh
RUN if [ -n "${INSTALL_MINGW}" ]; then bash ./install_mingw.sh; fi
RUN rm install_mingw.sh
ARG TRITON
ARG TRITON_CPU

4
.ci/lumen_cli/cli/lib/common/git_helper.py
View File

@ -57,8 +57,8 @@ def clone_external_repo(target: str, repo: str, dst: str = "", update_submodules
logger.info("Successfully cloned %s", target)
return r, commit
except GitCommandError as e:
logger.error("Git operation failed: %s", e)
except GitCommandError:
logger.exception("Git operation failed")
raise

2
.ci/lumen_cli/cli/lib/core/vllm/lib.py
View File

@ -143,7 +143,7 @@ def sample_vllm_test_library():
"pytest -v -s compile/test_decorator.py",
],
},
"vllm_languagde_model_test_extended_generation_28_failure_test": {
"vllm_language_model_test_extended_generation_28_failure_test": {
"title": "Language Models Test (Extended Generation) 2.8 release failure",
"id": "vllm_languagde_model_test_extended_generation_28_failure_test",
"package_install": [

2
.ci/lumen_cli/cli/lib/core/vllm/vllm_build.py
View File

@ -63,7 +63,7 @@ class VllmBuildParameters:
# DOCKERFILE_PATH: path to Dockerfile used when use_local_dockerfile is True"
use_local_dockerfile: bool = env_bool_field("USE_LOCAL_DOCKERFILE", True)
dockerfile_path: Path = env_path_field(
"DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
"DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile"
)
# the cleaning script to remove torch dependencies from pip

7
.ci/manywheel/build_cuda.sh
View File

@ -187,19 +187,22 @@ if [[ $CUDA_VERSION == 12* || $CUDA_VERSION == 13* ]]; then
export USE_CUFILE=0
else
DEPS_LIST+=(
"/usr/local/cuda/lib64/libnvToolsExt.so.1"
"/usr/local/cuda/lib64/libcublas.so.12"
"/usr/local/cuda/lib64/libcublasLt.so.12"
"/usr/local/cuda/lib64/libcudart.so.12"
"/usr/local/cuda/lib64/libnvrtc.so.12"
"/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12")
DEPS_SONAME+=(
"libnvToolsExt.so.1"
"libcublas.so.12"
"libcublasLt.so.12"
"libcudart.so.12"
"libnvrtc.so.12"
"libcupti.so.12")
if [[ $CUDA_VERSION != 12.9* ]]; then
DEPS_LIST+=("/usr/local/cuda/lib64/libnvToolsExt.so.1")
DEPS_SONAME+=("libnvToolsExt.so.1")
fi
fi
else
echo "Using nvidia libs from pypi."

27
.ci/pytorch/test.sh
View File

@ -485,6 +485,22 @@ test_inductor_aoti() {
/usr/bin/env "${TEST_ENVS[@]}" python test/run_test.py --cpp --verbose -i cpp/test_aoti_abi_check cpp/test_aoti_inference cpp/test_vec_half_AVX2 -dist=loadfile
}
test_inductor_aoti_cross_compile_for_windows() {
TEST_REPORTS_DIR=$(pwd)/test/test-reports
mkdir -p "$TEST_REPORTS_DIR"
# Set WINDOWS_CUDA_HOME environment variable
WINDOWS_CUDA_HOME="$(pwd)/win-torch-wheel-extracted"
export WINDOWS_CUDA_HOME
echo "WINDOWS_CUDA_HOME is set to: $WINDOWS_CUDA_HOME"
echo "Contents:"
ls -lah "$(pwd)/win-torch-wheel-extracted/lib/x64/" || true
python test/inductor/test_aoti_cross_compile_windows.py -k compile --package-dir "$TEST_REPORTS_DIR" --win-torch-lib-dir "$(pwd)/win-torch-wheel-extracted/torch/lib"
}
test_inductor_cpp_wrapper_shard() {
if [[ -z "$NUM_TEST_SHARDS" ]]; then
echo "NUM_TEST_SHARDS must be defined to run a Python test shard"
@ -900,7 +916,7 @@ test_inductor_set_cpu_affinity(){
export LD_PRELOAD="$JEMALLOC_LIB":"$LD_PRELOAD"
export MALLOC_CONF="oversize_threshold:1,background_thread:true,metadata_thp:auto,dirty_decay_ms:-1,muzzy_decay_ms:-1"
if [[ "${TEST_CONFIG}" != *aarch64* ]]; then
if [[ "$(uname -m)" != "aarch64" ]]; then
# Use Intel OpenMP for x86
IOMP_LIB="$(dirname "$(which python)")/../lib/libiomp5.so"
export LD_PRELOAD="$IOMP_LIB":"$LD_PRELOAD"
@ -914,7 +930,7 @@ test_inductor_set_cpu_affinity(){
cores=$((cpus / thread_per_core))
# Set number of cores to 16 on aarch64 for performance runs
if [[ "${TEST_CONFIG}" == *aarch64* && $cores -gt 16 ]]; then
if [[ "$(uname -m)" == "aarch64" && $cores -gt 16 ]]; then
cores=16
fi
export OMP_NUM_THREADS=$cores
@ -1615,6 +1631,7 @@ test_operator_benchmark() {
TEST_REPORTS_DIR=$(pwd)/test/test-reports
mkdir -p "$TEST_REPORTS_DIR"
TEST_DIR=$(pwd)
ARCH=$(uname -m)
test_inductor_set_cpu_affinity
@ -1629,7 +1646,7 @@ test_operator_benchmark() {
pip_install pandas
python check_perf_csv.py \
--actual "${TEST_REPORTS_DIR}/operator_benchmark_eager_float32_cpu.csv" \
--expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
--expected "${ARCH}_expected_ci_operator_benchmark_eager_float32_cpu.csv"
}
test_operator_microbenchmark() {
@ -1666,7 +1683,7 @@ if [[ "${TEST_CONFIG}" == *numpy_2* ]]; then
python -m pip install --pre numpy==2.0.2 scipy==1.13.1 numba==0.60.0
fi
python test/run_test.py --include dynamo/test_functions.py dynamo/test_unspec.py test_binary_ufuncs.py test_fake_tensor.py test_linalg.py test_numpy_interop.py test_tensor_creation_ops.py test_torch.py torch_np/test_basic.py
elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" != *perf_cpu_aarch64* ]]; then
elif [[ "${BUILD_ENVIRONMENT}" == *aarch64* && "${TEST_CONFIG}" == 'default' ]]; then
test_linux_aarch64
elif [[ "${TEST_CONFIG}" == *backward* ]]; then
test_forward_backward_compatibility
@ -1717,6 +1734,8 @@ elif [[ "${TEST_CONFIG}" == *inductor-triton-cpu* ]]; then
test_inductor_triton_cpu
elif [[ "${TEST_CONFIG}" == *inductor-micro-benchmark* ]]; then
test_inductor_micro_benchmark
elif [[ "${TEST_CONFIG}" == *aoti_cross_compile_for_windows* ]]; then
test_inductor_aoti_cross_compile_for_windows
elif [[ "${TEST_CONFIG}" == *huggingface* ]]; then
install_torchvision
id=$((SHARD_NUMBER-1))

4
.flake8
View File

@ -13,10 +13,6 @@ ignore =
EXE001,
# these ignores are from flake8-bugbear; please fix!
B007,B008,B017,B019,B023,B028,B903,B905,B906,B907,B908,B910
# these ignores are from flake8-comprehensions; please fix!
C407,
# these ignores are from flake8-logging-format; please fix!
G100,G101,G200
# these ignores are from flake8-simplify. please fix or ignore with commented reason
SIM105,SIM108,SIM110,SIM111,SIM113,SIM114,SIM115,SIM116,SIM117,SIM118,SIM119,SIM12,
# SIM104 is already covered by pyupgrade ruff

1
.github/ISSUE_TEMPLATE/ci-sev.md vendored
View File

@ -8,6 +8,7 @@ assignees: ''
---
> NOTE: Remember to label this issue with "`ci: sev`"
> If you want autorevert to be disabled, keep the ci: disable-autorevert label
<!-- Add the `merge blocking` label to this PR to prevent PRs from being merged while this issue is open -->

4
.github/ISSUE_TEMPLATE/disable-autorevert.md vendored
View File

@ -1,7 +1,7 @@
---
name: DISABLE AUTOREVERT
name: "D❌​\U0001F519​ ISABLE AUTOREVERT"
about: Disables autorevert when open
title: "❌​\U0001F519​ [DISABLE AUTOREVERT]"
title: "[DISABLE AUTOREVERT]"
labels: 'ci: disable-autorevert'
assignees: ''

2
.github/actions/build-external-packages/action.yml vendored
View File

@ -65,7 +65,7 @@ runs:
cd .ci/lumen_cli
python3 -m pip install -e .
)
MAX_JOBS="$(nproc --ignore=6)"
MAX_JOBS="$(nproc --ignore=10)"
export MAX_JOBS
# Split the comma-separated list and build each target

13
.github/actions/setup-rocm/action.yml vendored
View File

@ -111,3 +111,16 @@ runs:
# This video group ID maps to subgid 1 inside the docker image due to the /etc/subgid entries.
# The group name corresponding to group ID 1 can change depending on the OS, so both are necessary.
echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd $DEVICE_FLAG --group-add video --group-add $render_gid --group-add daemon --group-add bin --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --network=host" >> "${GITHUB_ENV}"
- name: configure aws credentials
id: aws_creds
uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
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: Login to Amazon ECR
id: login-ecr
continue-on-error: true
uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1

2
.github/ci_commit_pins/audio.txt vendored
View File

@ -1 +1 @@
87ff22e49ed0e92576c4935ccb8c143daac4a3cd
69bbe7363897764f9e758d851cd0340147d27f94

2
.github/ci_commit_pins/vision.txt vendored
View File

@ -1 +1 @@
966da7e46f65d6d49df3e31214470a4fe5cc8e66
faffd5cf673615583da6517275e361cb3dbc77e6

2
.github/ci_commit_pins/vllm.txt vendored
View File

@ -1 +1 @@
0ad9951c416d33c5da4f7a504fb162cbe62386f5
e5192819208c4d68194844b7dfafbc00020d0dea

2
.github/ci_commit_pins/xla.txt vendored
View File

@ -1 +1 @@
2a9138a26ee257fef05310ad3fecf7c55fe80d73
0fa6e3129e61143224663e1ec67980d12b7ec4eb

275
.github/ci_configs/vllm/Dockerfile.tmp_vllm → .github/ci_configs/vllm/Dockerfile vendored
View File

@ -1,59 +1,71 @@
# TODO(elainwy): remove this file after the torch nightly dockerfile is in sync in vllm repo
# The vLLM Dockerfile is used to construct vLLM image against torch nightly and torch main that can be directly used for testing
ARG CUDA_VERSION=12.8.1
ARG PYTHON_VERSION=3.12
# BUILD_BASE_IMAGE: used to setup python build xformers, and vllm wheels, It can be replaced with a different base image from local machine,
# by default, it uses the torch-nightly-base stage from this docker image
ARG BUILD_BASE_IMAGE=torch-nightly-base
# FINAL_BASE_IMAGE: used to set up vllm-instaled environment and build flashinfer,
# by default, it uses devel-ubuntu22.04 official image.
ARG FINAL_BASE_IMAGE=nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04
# The logic is copied from https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile
ARG GET_PIP_URL="https://bootstrap.pypa.io/get-pip.py"
#################### TORCH NIGHTLY BASE IMAGE ####################
# A base image for building vLLM with devel ubuntu 22.04, this is mainly used to build vllm in vllm builtkite ci
FROM nvidia/cuda:${CUDA_VERSION}-devel-ubuntu22.04 as torch-nightly-base
ARG CUDA_VERSION
ARG PYTHON_VERSION
ARG GET_PIP_URL
# Install Python and other dependencies
# Install system dependencies and uv, then create Python virtual environment
RUN apt-get update -y \
&& apt-get install -y ccache software-properties-common git curl wget sudo vim \
&& add-apt-repository -y ppa:deadsnakes/ppa \
&& apt-get update -y \
&& apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
&& update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
&& update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
&& ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
&& curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION} \
&& apt-get install -y ccache software-properties-common git curl sudo vim python3-pip \
&& curl -LsSf https://astral.sh/uv/install.sh | sh \
&& $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
&& rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
&& ln -s /opt/venv/bin/python3 /usr/bin/python3 \
&& ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
&& ln -s /opt/venv/bin/pip /usr/bin/pip \
&& python3 --version && python3 -m pip --version
# Upgrade to GCC 10 to avoid https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92519
# as it was causing spam when compiling the CUTLASS kernels
# Ensure gcc >= 10 to avoid CUTLASS issues (bug 92519)
RUN current_gcc_version=$(gcc -dumpversion | cut -f1 -d.) && \
if command -v apt-get >/dev/null; then \
if [ "$current_gcc_version" -lt 10 ]; then \
echo "GCC version is $current_gcc_version, installing gcc-10..."; \
apt-get update \
&& apt-get install -y gcc-10 g++-10 \
&& update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 100 \
&& update-alternatives --install /usr/bin/g++ g++ /usr/bin/g++-10 100; \
else \
echo "GCC version is $current_gcc_version, no need to install gcc-10."; \
fi \
fi \
&& gcc --version && g++ --version
RUN apt-get install -y gcc-10 g++-10
RUN update-alternatives --install /usr/bin/gcc gcc /usr/bin/gcc-10 110 --slave /usr/bin/g++ g++ /usr/bin/g++-10
RUN <<EOF
gcc --version
EOF
# install uv for faster pip installs
# Install uv for faster pip installs
RUN --mount=type=cache,target=/root/.cache/uv \
python3 -m pip install uv==0.8.4
ENV UV_HTTP_TIMEOUT=500
ENV UV_INDEX_STRATEGY="unsafe-best-match"
# Use copy mode to avoid hardlink failures with Docker cache mounts
ENV UV_LINK_MODE=copy
#################### TORCH NIGHTLY BASE IMAGE ####################
#################### BASE BUILD IMAGE ####################
FROM ${BUILD_BASE_IMAGE} AS base
USER root
ARG CUDA_VERSION
ARG PYTHON_VERSION
# Only work with PyTorch manylinux builder
ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
# Install some system dependencies and double check python version
RUN if command -v apt-get >/dev/null; then \
apt-get update -y \
&& apt-get install -y ccache software-properties-common git wget sudo vim; \
else \
dnf install -y git wget sudo; \
fi \
&& python3 --version && python3 -m pip --version
# Install uv for faster pip installs if not existed
RUN --mount=type=cache,target=/root/.cache/uv \
python3 -m pip install uv==0.8.4
@ -62,51 +74,17 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
# Use copy mode to avoid hardlink failures with Docker cache mounts
ENV UV_LINK_MODE=copy
#################### TORCH NIGHTLY BASE IMAGE ####################
#################### BASE BUILD IMAGE ####################
# A base image for building vLLM with torch nightly or torch wheels
# prepare basic build environment
FROM ${BUILD_BASE_IMAGE} AS base
USER root
ARG CUDA_VERSION
ARG PYTHON_VERSION
# TODO (huydhn): Only work with PyTorch manylinux builder
ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
# Install some system dependencies and double check python version
RUN if command -v apt-get >/dev/null; then \
apt-get update -y \
&& apt-get install -y ccache software-properties-common git curl wget sudo vim; \
else \
dnf install -y git curl wget sudo; \
fi \
&& python3 --version && python3 -m pip --version
# Install uv for faster pip installs if not existed
RUN --mount=type=cache,target=/root/.cache/uv \
if ! python3 -m uv --version >/dev/null 2>&1; then \
python3 -m pip install uv==0.8.4; \
fi
ENV UV_HTTP_TIMEOUT=500
ENV UV_INDEX_STRATEGY="unsafe-best-match"
# Use copy mode to avoid hardlink failures with Docker cache mounts
ENV UV_LINK_MODE=copy
WORKDIR /workspace
# install build and runtime dependencies
# Install build and runtime dependencies
COPY requirements/common.txt requirements/common.txt
COPY use_existing_torch.py use_existing_torch.py
COPY pyproject.toml pyproject.toml
# install build and runtime dependencies without stable torch version
# Install build and runtime dependencies without stable torch version
RUN python3 use_existing_torch.py
# default mount file as placeholder, this just avoid the mount error
# Default mount file as placeholder, this just avoid the mount error
# change to a different vllm folder if this does not exist anymore
ARG TORCH_WHEELS_PATH="./requirements"
ARG PINNED_TORCH_VERSION
@ -138,56 +116,36 @@ RUN --mount=type=cache,target=/root/.cache/uv \
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -r requirements/common.txt
# Must put before installing xformers, so it can install the correct version of xfomrers.
ARG xformers_cuda_arch_list='7.5;8.0+PTX;9.0a'
ENV TORCH_CUDA_ARCH_LIST=${xformers_cuda_arch_list}
ARG max_jobs=16
ENV MAX_JOBS=${max_jobs}
RUN echo ${TORCH_CUDA_ARCH_LIST}
RUN echo ${MAX_JOBS}
RUN pip freeze | grep -E 'ninja'
RUN --mount=type=cache,target=/root/.cache/uv bash - <<'BASH'
export TORCH_CUDA_ARCH_LIST='7.5 8.0+PTX 9.0a'
git clone https://github.com/facebookresearch/xformers.git
# Build xformers with cuda and torch nightly/wheel
# following official xformers guidance: https://github.com/facebookresearch/xformers#build
# sha for https://github.com/facebookresearch/xformers/tree/v0.0.32.post2
ARG XFORMERS_COMMIT=5d4b92a5e5a9c6c6d4878283f47d82e17995b468
ENV CCACHE_DIR=/root/.cache/ccache
pushd xformers
git checkout v0.0.32.post2
git submodule update --init --recursive
python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose
popd
RUN --mount=type=cache,target=/root/.cache/ccache \
--mount=type=cache,target=/root/.cache/uv \
echo 'git clone xformers...' \
&& git clone https://github.com/facebookresearch/xformers.git --recursive \
&& cd xformers \
&& git checkout ${XFORMERS_COMMIT} \
&& git submodule update --init --recursive \
&& echo 'finish git clone xformers...' \
&& rm -rf build \
&& python3 setup.py bdist_wheel --dist-dir=../xformers-dist --verbose \
&& cd .. \
&& rm -rf xformers
rm -rf xformers
BASH
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system xformers-dist/*.whl --verbose
uv pip install --system xformers-dist/*.whl
# Build can take a long time, and the torch nightly version fetched from url can be different in next docker stage.
# track the nightly torch version used in the build, when we set up runtime environment we can make sure the version is the same
RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio' > torch_build_versions.txt
RUN cat torch_build_versions.txt
RUN pip freeze | grep -E 'torch|xformers|torchvision|torchaudio'
#################### BASE BUILD IMAGE ####################
#################### WHEEL BUILD IMAGE ####################
# Image used to build vllm wheel
FROM base AS build
ARG TARGETPLATFORM
COPY . .
RUN python3 use_existing_torch.py
RUN --mount=type=cache,target=/root/.cache/uv \
@ -197,20 +155,17 @@ ARG GIT_REPO_CHECK=0
RUN --mount=type=bind,source=.git,target=.git \
if [ "$GIT_REPO_CHECK" != "0" ]; then bash tools/check_repo.sh ; fi
# Max jobs used by Ninja to build extensions
ARG max_jobs=16
ENV MAX_JOBS=${max_jobs}
ARG nvcc_threads=4
ARG nvcc_threads=8
ENV NVCC_THREADS=$nvcc_threads
ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
ARG USE_SCCACHE
ARG SCCACHE_BUCKET_NAME=vllm-build-sccache
ARG SCCACHE_REGION_NAME=us-west-2
ARG SCCACHE_S3_NO_CREDENTIALS=0
# if USE_SCCACHE is set, use sccache to speed up compilation
# Use sccache to speed up compilation
RUN --mount=type=cache,target=/root/.cache/uv \
--mount=type=bind,source=.git,target=.git \
if [ "$USE_SCCACHE" = "1" ]; then \
@ -235,6 +190,9 @@ RUN --mount=type=cache,target=/root/.cache/uv \
&& sccache --show-stats; \
fi
ARG torch_cuda_arch_list='8.0 8.6 8.9 9.0'
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
ARG vllm_target_device="cuda"
ENV VLLM_TARGET_DEVICE=${vllm_target_device}
ENV CCACHE_DIR=/root/.cache/ccache
@ -248,17 +206,10 @@ RUN --mount=type=cache,target=/root/.cache/ccache \
export VLLM_DOCKER_BUILD_CONTEXT=1 && \
python3 setup.py bdist_wheel --dist-dir=vllm-dist --py-limited-api=cp38; \
fi
RUN echo "[INFO] Listing current directory:" && \
ls -al && \
echo "[INFO] Showing torch_build_versions.txt content:" && \
cat torch_build_versions.txt
#################### WHEEL BUILD IMAGE ####################
################### VLLM INSTALLED IMAGE ####################
# Setup clean environment for vLLM for test and api server using ubuntu22.04 with AOT flashinfer
FROM ${FINAL_BASE_IMAGE} AS vllm-base
USER root
@ -266,7 +217,7 @@ ARG CUDA_VERSION
ARG PYTHON_VERSION
ARG GET_PIP_URL
# TODO (huydhn): Only work with PyTorch manylinux builder
# Only work with PyTorch manylinux builder
ENV PATH="/opt/python/cp312-cp312/bin:${PATH}"
# prepare for environment starts
@ -275,20 +226,19 @@ WORKDIR /workspace
# Install Python and other dependencies
RUN if command -v apt-get >/dev/null; then \
apt-get update -y \
&& apt-get install -y ccache software-properties-common git curl wget sudo vim \
&& add-apt-repository -y ppa:deadsnakes/ppa \
&& apt-get update -y \
&& apt-get install -y python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv \
&& update-alternatives --install /usr/bin/python3 python3 /usr/bin/python${PYTHON_VERSION} 1 \
&& update-alternatives --set python3 /usr/bin/python${PYTHON_VERSION} \
&& ln -sf /usr/bin/python${PYTHON_VERSION}-config /usr/bin/python3-config \
&& curl -sS ${GET_PIP_URL} | python${PYTHON_VERSION}; \
&& apt-get install -y ccache software-properties-common git sudo vim python3-pip; \
else \
dnf install -y git curl wget sudo; \
dnf install -y git wget sudo; \
fi \
&& curl -LsSf https://astral.sh/uv/install.sh | sh \
&& $HOME/.local/bin/uv venv /opt/venv --python ${PYTHON_VERSION} \
&& rm -f /usr/bin/python3 /usr/bin/python3-config /usr/bin/pip \
&& ln -s /opt/venv/bin/python3 /usr/bin/python3 \
&& ln -s /opt/venv/bin/python3-config /usr/bin/python3-config \
&& ln -s /opt/venv/bin/pip /usr/bin/pip \
&& python3 --version && python3 -m pip --version
# Get the torch versions, and whls used in previous stagtes for consistency
# Get the torch versions, and whls used in previous stage
COPY --from=base /workspace/torch_build_versions.txt ./torch_build_versions.txt
COPY --from=base /workspace/xformers-dist /wheels/xformers
COPY --from=build /workspace/vllm-dist /wheels/vllm
@ -297,33 +247,29 @@ RUN echo "[INFO] Listing current directory before torch install step:" && \
echo "[INFO] Showing torch_build_versions.txt content:" && \
cat torch_build_versions.txt
# Install build and runtime dependencies, this is needed for flashinfer install
COPY requirements/build.txt requirements/build.txt
COPY use_existing_torch.py use_existing_torch.py
RUN python3 use_existing_torch.py
RUN cat requirements/build.txt
# Install uv for faster pip installs if not existed
RUN --mount=type=cache,target=/root/.cache/uv \
if ! python3 -m uv --version > /dev/null 2>&1; then \
python3 -m pip install uv==0.8.4; \
fi
python3 -m pip install uv==0.8.4
ENV UV_HTTP_TIMEOUT=500
ENV UV_INDEX_STRATEGY="unsafe-best-match"
# Use copy mode to avoid hardlink failures with Docker cache mounts
ENV UV_LINK_MODE=copy
# Install build and runtime dependencies, this is needed for flashinfer install
COPY requirements/build.txt requirements/build.txt
COPY use_existing_torch.py use_existing_torch.py
RUN python3 use_existing_torch.py
RUN cat requirements/build.txt
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -r requirements/build.txt
# Default mount file as placeholder, this just avoid the mount error
ARG TORCH_WHEELS_PATH="./requirements"
# Install torch, torchaudio and torchvision
# if TORCH_WHEELS_PATH is default "./requirements", it will pull the nightly versions using pip using torch_build_versions.txt
# otherwise, it will use the whls from TORCH_WHEELS_PATH from the host machine
# Install torch, torchaudio and torchvision. If TORCH_WHEELS_PATH is default
# to ./requirements, it will pull the nightly versions using pip. Otherwise,
# it will use the local wheels from TORCH_WHEELS_PATH
RUN --mount=type=bind,source=${TORCH_WHEELS_PATH},target=/dist \
--mount=type=cache,target=/root/.cache/uv \
if [ -n "$TORCH_WHEELS_PATH" ] && [ "$TORCH_WHEELS_PATH" != "./requirements" ] && [ -d "/dist" ] && ls /dist/torch*.whl >/dev/null 2>&1; then \
@ -344,18 +290,14 @@ RUN --mount=type=cache,target=/root/.cache/uv \
# Install xformers wheel from previous stage
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system /wheels/xformers/*.whl --verbose
# Build flashinfer from source.
# Build FlashInfer from source
ARG torch_cuda_arch_list='8.0;8.9;9.0a;10.0a;12.0'
# install package for build flashinfer
# see issue: https://github.com/flashinfer-ai/flashinfer/issues/738
RUN pip freeze | grep -E 'setuptools|packaging|build'
ENV TORCH_CUDA_ARCH_LIST=${torch_cuda_arch_list}
# Build flashinfer for torch nightly from source around 10 mins
ARG FLASHINFER_GIT_REPO="https://github.com/flashinfer-ai/flashinfer.git"
# Keep this in sync with https://github.com/vllm-project/vllm/blob/main/requirements/cuda.txt
ARG FLASHINFER_GIT_REF="v0.2.14.post1"
RUN --mount=type=cache,target=/root/.cache/uv \
git clone --depth 1 --recursive --shallow-submodules \
--branch ${FLASHINFER_GIT_REF} \
@ -367,7 +309,7 @@ RUN --mount=type=cache,target=/root/.cache/uv \
&& cd .. \
&& rm -rf flashinfer
# install flashinfer python
# Install FlashInfer
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system wheels/flashinfer/*.whl --verbose
@ -377,49 +319,6 @@ RUN uv pip freeze | grep -i '^torch\|^torchvision\|^torchaudio\|^xformers\|^vllm
################### VLLM INSTALLED IMAGE ####################
#################### UNITTEST IMAGE #############################
FROM vllm-base as test
ENV UV_HTTP_TIMEOUT=500
ENV UV_INDEX_STRATEGY="unsafe-best-match"
# Use copy mode to avoid hardlink failures with Docker cache mounts
ENV UV_LINK_MODE=copy
COPY tests/ tests/
COPY examples examples
COPY benchmarks benchmarks
COPY ./vllm/collect_env.py .
COPY requirements/common.txt requirements/common.txt
COPY use_existing_torch.py use_existing_torch.py
COPY pyproject.toml pyproject.toml
# Install build and runtime dependencies without stable torch version
COPY requirements/nightly_torch_test.txt requirements/nightly_torch_test.txt
RUN python3 use_existing_torch.py
# install packages
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -r requirements/common.txt
# enable fast downloads from hf (for testing)
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system hf_transfer
ENV HF_HUB_ENABLE_HF_TRANSFER 1
# install development dependencies (for testing)
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -e tests/vllm_test_utils
RUN --mount=type=cache,target=/root/.cache/uv \
uv pip install --system -r requirements/nightly_torch_test.txt
# Logging to confirm the torch versions
RUN pip freeze | grep -E 'torch|xformers|vllm|flashinfer'
# Logging to confirm all the packages are installed
RUN pip freeze
#################### UNITTEST IMAGE #############################
#################### EXPORT STAGE ####################
FROM scratch as export-wheels

29
.github/labeler.yml vendored
View File

@ -133,3 +133,32 @@
"ciflow/vllm":
- .github/ci_commit_pins/vllm.txt
"ciflow/b200":
- test/test_matmul_cuda.py
- test/test_scaled_matmul_cuda.py
- test/inductor/test_fp8.py
- aten/src/ATen/native/cuda/Blas.cpp
- torch/**/*cublas*
- torch/_inductor/kernel/mm.py
- test/inductor/test_max_autotune.py
- third_party/fbgemm
"ciflow/h100":
- test/test_matmul_cuda.py
- test/test_scaled_matmul_cuda.py
- test/inductor/test_fp8.py
- aten/src/ATen/native/cuda/Blas.cpp
- torch/**/*cublas*
- torch/_inductor/kernel/mm.py
- test/inductor/test_max_autotune.py
- third_party/fbgemm
"ciflow/rocm":
- test/test_matmul_cuda.py
- test/test_scaled_matmul_cuda.py
- test/inductor/test_fp8.py
- aten/src/ATen/native/cuda/Blas.cpp
- torch/_inductor/kernel/mm.py
- test/inductor/test_max_autotune.py
- third_party/fbgemm

4
.github/pytorch-probot.yml vendored
View File

@ -3,6 +3,7 @@ ciflow_tracking_issue: 64124
ciflow_push_tags:
- ciflow/b200
- ciflow/b200-symm-mem
- ciflow/b200-distributed
- ciflow/binaries
- ciflow/binaries_libtorch
- ciflow/binaries_wheel
@ -15,7 +16,8 @@ ciflow_push_tags:
- ciflow/inductor-micro-benchmark
- ciflow/inductor-micro-benchmark-cpu-x86
- ciflow/inductor-perf-compare
- ciflow/inductor-perf-test-nightly-rocm
- ciflow/inductor-perf-test-nightly-rocm-mi300
- ciflow/inductor-perf-test-nightly-rocm-mi355
- ciflow/inductor-perf-test-nightly-x86-zen
- ciflow/inductor-periodic
- ciflow/inductor-rocm

2
.github/scripts/filter_test_configs.py vendored
View File

@ -512,6 +512,8 @@ def perform_misc_tasks(
"keep-going",
branch == MAIN_BRANCH
or bool(tag and re.match(r"^trunk/[a-f0-9]{40}$", tag))
# Pattern for tags created via manual run on HUD
or bool(tag and re.match(r"^ciflow/[^/]+/[a-f0-9]{40}$", tag))
or check_for_setting(labels, pr_body, "keep-going"),
)
set_output(

38
.github/scripts/generate_binary_build_matrix.py vendored
View File

@ -16,16 +16,18 @@ from typing import Optional
# NOTE: Please also update the CUDA sources in `PIP_SOURCES` in tools/nightly.py when changing this
CUDA_ARCHES = ["12.6", "12.8", "13.0"]
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": "13.0.0",
}
CUDA_ARCHES_CUDNN_VERSION = {
"12.6": "9",
"12.8": "9",
"12.9": "9",
"13.0": "9",
}
@ -38,7 +40,7 @@ CPU_AARCH64_ARCH = ["cpu-aarch64"]
CPU_S390X_ARCH = ["cpu-s390x"]
CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "13.0-aarch64"]
CUDA_AARCH64_ARCHES = ["12.6-aarch64", "12.8-aarch64", "12.9-aarch64", "13.0-aarch64"]
PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
@ -76,6 +78,23 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
"nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
"nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
),
"12.9": (
"nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | "
"nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | "
"nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | "
"nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
"nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | "
"nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | "
"nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | "
"nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | "
"nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | "
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
"nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | "
"nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | "
"nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | "
"nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
),
"13.0": (
"nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
"nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
@ -222,7 +241,11 @@ def generate_libtorch_matrix(
arches += CUDA_ARCHES
arches += ROCM_ARCHES
elif os == "windows":
arches += CUDA_ARCHES
# TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
# in 2.10
windows_cuda_arches = CUDA_ARCHES.copy()
windows_cuda_arches.remove("12.9")
arches += windows_cuda_arches
if libtorch_variants is None:
libtorch_variants = [
"shared-with-deps",
@ -286,7 +309,11 @@ def generate_wheels_matrix(
if os == "linux":
arches += CUDA_ARCHES + ROCM_ARCHES + XPU_ARCHES
elif os == "windows":
arches += CUDA_ARCHES + XPU_ARCHES
# TODO (huydhn): Only build CUDA 12.9 for Linux. This logic is to be cleaned up
# in 2.10
windows_cuda_arches = CUDA_ARCHES.copy()
windows_cuda_arches.remove("12.9")
arches += windows_cuda_arches + XPU_ARCHES
elif os == "linux-aarch64":
# Separate new if as the CPU type is different and
# uses different build/test scripts
@ -322,7 +349,7 @@ def generate_wheels_matrix(
# cuda linux wheels require PYTORCH_EXTRA_INSTALL_REQUIREMENTS to install
if (
arch_version in ["13.0", "12.8", "12.6"]
arch_version in ["13.0", "12.9", "12.8", "12.6"]
and os == "linux"
or arch_version in CUDA_AARCH64_ARCHES
):
@ -386,5 +413,6 @@ def generate_wheels_matrix(
validate_nccl_dep_consistency("13.0")
validate_nccl_dep_consistency("12.9")
validate_nccl_dep_consistency("12.8")
validate_nccl_dep_consistency("12.6")

6
.github/scripts/trymerge.py vendored
View File

@ -1092,7 +1092,7 @@ class GitHubPR:
editor = node["editor"]
return GitHubComment(
body_text=node["bodyText"],
created_at=node["createdAt"] if "createdAt" in node else "",
created_at=node.get("createdAt", ""),
author_login=node["author"]["login"],
author_url=node["author"].get("url", None),
author_association=node["authorAssociation"],
@ -2042,10 +2042,6 @@ def validate_revert(
f"[{', '.join(allowed_reverters)}], but instead is {author_association}."
)
# Raises exception if matching rule is not found, but ignores all status checks
find_matching_merge_rule(
pr, repo, skip_mandatory_checks=True, skip_internal_checks=True
)
commit_sha = get_pr_commit_sha(repo, pr)
return (author_login, commit_sha)

3
.github/templates/linux_binary_build_workflow.yml.j2 vendored
View File

@ -177,6 +177,9 @@ jobs:
runs-on: linux.rocm.gpu.mi250
timeout-minutes: !{{ common.timeout_minutes }}
!{{ upload.binary_env(config) }}
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm

2
.github/workflows/_docs.yml vendored
View File

@ -72,7 +72,7 @@ jobs:
# Let's try to figure out how this can be improved
timeout-minutes: 360
- docs_type: python
runner: ${{ inputs.runner_prefix }}linux.2xlarge
runner: ${{ inputs.runner_prefix }}linux.c7i.2xlarge
# It takes less than 30m to finish python docs unless there are issues
timeout-minutes: 30
# Set a fixed name for this job instead of using the current matrix-generated name, i.e. build-docs (cpp, linux.12xlarge, 180)

2
.github/workflows/_linux-build.yml vendored
View File

@ -37,7 +37,7 @@ on:
runner:
required: false
type: string
default: "linux.2xlarge"
default: "linux.c7i.2xlarge"
description: |
Label of the runner this job should run on.
test-matrix:

40
.github/workflows/_linux-test.yml vendored
View File

@ -224,6 +224,46 @@ jobs:
continue-on-error: true
uses: ./.github/actions/download-td-artifacts
- name: Download Windows torch wheel for cross-compilation
if: matrix.win_torch_wheel_artifact != ''
uses: seemethere/download-artifact-s3@1da556a7aa0a088e3153970611f6c432d58e80e6 # v4.2.0
with:
name: ${{ matrix.win_torch_wheel_artifact }}
path: win-torch-wheel
- name: Extract Windows wheel and setup CUDA libraries
if: matrix.win_torch_wheel_artifact != ''
shell: bash
run: |
set -x
# Find the wheel file
WHEEL_FILE=$(find win-torch-wheel -name "*.whl" -type f | head -n 1)
if [ -z "$WHEEL_FILE" ]; then
echo "Error: No wheel file found in win-torch-wheel directory"
exit 1
fi
echo "Found wheel file: $WHEEL_FILE"
# Unzip the wheel file
unzip -q "$WHEEL_FILE" -d win-torch-wheel-extracted
echo "Extracted wheel contents"
# Setup CUDA libraries (cuda.lib and cudart.lib) directory
mkdir -p win-torch-wheel-extracted/lib/x64
if [ -f "win-torch-wheel/cuda.lib" ]; then
mv win-torch-wheel/cuda.lib win-torch-wheel-extracted/lib/x64/
echo "Moved cuda.lib to win-torch-wheel-extracted/lib/x64/"
fi
if [ -f "win-torch-wheel/cudart.lib" ]; then
mv win-torch-wheel/cudart.lib win-torch-wheel-extracted/lib/x64/
echo "Moved cudart.lib to win-torch-wheel-extracted/lib/x64/"
fi
# Verify CUDA libraries are present
echo "CUDA libraries:"
ls -la win-torch-wheel-extracted/lib/x64/ || echo "No CUDA libraries found"
- name: Parse ref
id: parse-ref
run: .github/scripts/parse_ref.py

13
.github/workflows/_rocm-test.yml vendored
View File

@ -102,19 +102,6 @@ jobs:
exit 1
fi
- name: configure aws credentials
id: aws_creds
uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
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: Login to Amazon ECR
id: login-ecr
continue-on-error: true
uses: aws-actions/amazon-ecr-login@062b18b96a7aff071d4dc91bc00c4c1a7945b076 # v2.0.1
- name: Calculate docker image
id: calculate-docker-image
uses: pytorch/test-infra/.github/actions/calculate-docker-image@main

25
.github/workflows/_win-build.yml vendored
View File

@ -168,6 +168,31 @@ jobs:
run: |
.ci/pytorch/win-build.sh
# Collect Windows torch libs and CUDA libs for cross-compilation
- name: Collect Windows CUDA libs for cross-compilation
if: steps.build.outcome != 'skipped' && inputs.cuda-version != 'cpu'
shell: bash
run: |
set -ex
# Create directory structure if does not exist
mkdir -p /c/${{ github.run_id }}/build-results
# Copy CUDA libs
CUDA_PATH="/c/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v${{ inputs.cuda-version }}"
if [ -f "${CUDA_PATH}/lib/x64/cuda.lib" ]; then
cp "${CUDA_PATH}/lib/x64/cuda.lib" /c/${{ github.run_id }}/build-results/
fi
if [ -f "${CUDA_PATH}/lib/x64/cudart.lib" ]; then
cp "${CUDA_PATH}/lib/x64/cudart.lib" /c/${{ github.run_id }}/build-results/
fi
# List collected files
echo "Collected CUDA libs:"
ls -lah /c/${{ github.run_id }}/build-results/*.lib
# Upload to github so that people can click and download artifacts
- name: Upload artifacts to s3
if: steps.build.outcome != 'skipped'

62
.github/workflows/b200-distributed.yml vendored Normal file
View File

@ -0,0 +1,62 @@
name: CI for distributed tests on B200
on:
pull_request:
paths:
- .github/workflows/b200-distributed.yml
workflow_dispatch:
push:
tags:
- ciflow/b200-distributed/*
schedule:
- cron: 46 8 * * * # about 1:46am PDT
concurrency:
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
cancel-in-progress: true
permissions:
id-token: write
contents: read
jobs:
get-label-type:
if: github.repository_owner == 'pytorch'
name: get-label-type
uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
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: ${{ github.ref_type }}
linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200:
name: linux-jammy-cuda12.8-py3.10-gcc11-build-distributed-b200
uses: ./.github/workflows/_linux-build.yml
needs: get-label-type
with:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runner: linux.12xlarge.memory
build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
cuda-arch-list: '10.0'
test-matrix: |
{ include: [
{ config: "distributed", shard: 1, num_shards: 2, runner: "linux.dgx.b200.8" },
{ config: "distributed", shard: 2, num_shards: 2, runner: "linux.dgx.b200.8" },
]}
secrets: inherit
linux-jammy-cuda12_8-py3_10-gcc11-test-distributed-b200:
name: linux-jammy-cuda12.8-py3.10-gcc11-test-b200
uses: ./.github/workflows/_linux-test.yml
needs:
- linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200
with:
timeout-minutes: 1200
build-environment: linux-jammy-cuda12.8-py3.10-gcc11-distributed-b200
docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200.outputs.docker-image }}
test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build-distributed-b200.outputs.test-matrix }}
aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
secrets: inherit

4
.github/workflows/build-manywheel-images.yml vendored
View File

@ -46,10 +46,12 @@ jobs:
fail-fast: false
matrix:
include: [
{ name: "manylinux2_28-builder", tag: "cuda13.0", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinux2_28-builder", tag: "cuda13.0", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinux2_28-builder", tag: "cuda12.8", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinux2_28-builder", tag: "cuda12.9", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinux2_28-builder", tag: "cuda12.6", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinuxaarch64-builder", tag: "cuda13.0", runner: "linux.arm64.2xlarge.ephemeral" },
{ 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" },

19
.github/workflows/build-vllm-wheel.yml vendored
View File

@ -27,9 +27,8 @@ jobs:
fail-fast: false
matrix:
python-version: [ '3.12' ]
# TODO (huydhn): Add cu130 after https://github.com/vllm-project/vllm/issues/24464 is resolved
platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
device: [ 'cu128', 'cu129' ]
device: [ 'cu128', 'cu129', 'cu130' ]
include:
- platform: manylinux_2_28_x86_64
device: cu128
@ -39,6 +38,10 @@ jobs:
device: cu129
manylinux-image: 'pytorch/manylinux2_28-builder:cuda12.9'
runner: linux.12xlarge.memory
- platform: manylinux_2_28_x86_64
device: cu130
manylinux-image: 'pytorch/manylinux2_28-builder:cuda13.0'
runner: linux.12xlarge.memory
- platform: manylinux_2_28_aarch64
device: cu128
manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.8'
@ -47,6 +50,11 @@ jobs:
device: cu129
manylinux-image: 'pytorch/manylinuxaarch64-builder:cuda12.9'
runner: linux.arm64.r7g.12xlarge.memory
exclude:
# TODO (huydhn): Add cu130 aarch64 once PyTorch is on 2.9+ and
# xformers is update to support 13.0
- platform: manylinux_2_28_aarch64
device: cu130
name: "Build ${{ matrix.device }} vLLM wheel on ${{ matrix.platform }}"
runs-on: ${{ matrix.runner }}
timeout-minutes: 480
@ -169,7 +177,12 @@ jobs:
fail-fast: false
matrix:
platform: [ 'manylinux_2_28_x86_64', 'manylinux_2_28_aarch64' ]
device: [ 'cu128', 'cu129' ]
device: [ 'cu128', 'cu129', 'cu130' ]
exclude:
# TODO (huydhn): Add cu130 aarch64 once PyTorch is on 2.9+ and
# xformers is update to support 13.0
- platform: manylinux_2_28_aarch64
device: cu130
env:
PLATFORM: ${{ matrix.platform }}
BUILD_DEVICE: ${{ matrix.device }}

322
.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml generated vendored
View File

@ -204,6 +204,52 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_10-cuda-aarch64-12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.10"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.r7g.12xlarge.memory
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_10-cuda-aarch64-12_9
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_10-cuda-aarch64-12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_10-cuda-aarch64-12_9-build
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.10"
build_name: manywheel-py3_10-cuda-aarch64-12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_10-cuda-aarch64-13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -407,6 +453,52 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_11-cuda-aarch64-12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.11"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.r7g.12xlarge.memory
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_11-cuda-aarch64-12_9
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_11-cuda-aarch64-12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_11-cuda-aarch64-12_9-build
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.11"
build_name: manywheel-py3_11-cuda-aarch64-12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_11-cuda-aarch64-13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -610,6 +702,52 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_12-cuda-aarch64-12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.12"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.r7g.12xlarge.memory
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_12-cuda-aarch64-12_9
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda-aarch64-12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_12-cuda-aarch64-12_9-build
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.12"
build_name: manywheel-py3_12-cuda-aarch64-12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_12-cuda-aarch64-13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -813,6 +951,52 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13-cuda-aarch64-12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.r7g.12xlarge.memory
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_13-cuda-aarch64-12_9
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13-cuda-aarch64-12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_13-cuda-aarch64-12_9-build
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13"
build_name: manywheel-py3_13-cuda-aarch64-12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13-cuda-aarch64-13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -1016,6 +1200,52 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13t-cuda-aarch64-12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13t"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.r7g.12xlarge.memory
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_13t-cuda-aarch64-12_9
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13t-cuda-aarch64-12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_13t-cuda-aarch64-12_9-build
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13t"
build_name: manywheel-py3_13t-cuda-aarch64-12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13t-cuda-aarch64-13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -1219,6 +1449,52 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14-cuda-aarch64-12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.r7g.12xlarge.memory
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_14-cuda-aarch64-12_9
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14-cuda-aarch64-12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_14-cuda-aarch64-12_9-build
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14"
build_name: manywheel-py3_14-cuda-aarch64-12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14-cuda-aarch64-13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -1422,6 +1698,52 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14t-cuda-aarch64-12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14t"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.r7g.12xlarge.memory
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_14t-cuda-aarch64-12_9
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14t-cuda-aarch64-12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_14t-cuda-aarch64-12_9-build
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14t"
build_name: manywheel-py3_14t-cuda-aarch64-12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14t-cuda-aarch64-13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml

74
.github/workflows/generated-linux-binary-libtorch-nightly.yml generated vendored
View File

@ -248,6 +248,74 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
libtorch-cuda12_9-shared-with-deps-release-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: libtorch
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: libtorch-cxx11-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
LIBTORCH_CONFIG: release
LIBTORCH_VARIANT: shared-with-deps
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: libtorch-cuda12_9-shared-with-deps-release
build_environment: linux-binary-libtorch
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
libtorch-cuda12_9-shared-with-deps-release-test: # Testing
if: ${{ github.repository_owner == 'pytorch' }}
needs:
- libtorch-cuda12_9-shared-with-deps-release-build
- get-label-type
uses: ./.github/workflows/_binary-test-linux.yml
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: libtorch
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: libtorch-cxx11-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
LIBTORCH_CONFIG: release
LIBTORCH_VARIANT: shared-with-deps
build_name: libtorch-cuda12_9-shared-with-deps-release
build_environment: linux-binary-libtorch
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
libtorch-cuda12_9-shared-with-deps-release-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: libtorch-cuda12_9-shared-with-deps-release-test
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: libtorch
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: libtorch-cxx11-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
LIBTORCH_CONFIG: release
LIBTORCH_VARIANT: shared-with-deps
build_name: libtorch-cuda12_9-shared-with-deps-release
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
libtorch-cuda13_0-shared-with-deps-release-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -358,6 +426,9 @@ jobs:
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
@ -473,6 +544,9 @@ jobs:
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
LIBTORCH_CONFIG: release
LIBTORCH_VARIANT: shared-with-deps
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm

504
.github/workflows/generated-linux-binary-manywheel-nightly.yml generated vendored
View File

@ -241,6 +241,72 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_10-cuda12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.10"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_10-cuda12_9
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_10-cuda12_9-test: # Testing
if: ${{ github.repository_owner == 'pytorch' }}
needs:
- manywheel-py3_10-cuda12_9-build
- get-label-type
uses: ./.github/workflows/_binary-test-linux.yml
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.10"
build_name: manywheel-py3_10-cuda12_9
build_environment: linux-binary-manywheel
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_10-cuda12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_10-cuda12_9-test
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.10"
build_name: manywheel-py3_10-cuda12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_10-cuda13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -347,6 +413,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
DESIRED_PYTHON: "3.10"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -459,6 +528,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
DESIRED_PYTHON: "3.10"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -835,6 +907,72 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_11-cuda12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.11"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_11-cuda12_9
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_11-cuda12_9-test: # Testing
if: ${{ github.repository_owner == 'pytorch' }}
needs:
- manywheel-py3_11-cuda12_9-build
- get-label-type
uses: ./.github/workflows/_binary-test-linux.yml
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.11"
build_name: manywheel-py3_11-cuda12_9
build_environment: linux-binary-manywheel
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_11-cuda12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_11-cuda12_9-test
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.11"
build_name: manywheel-py3_11-cuda12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_11-cuda13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -941,6 +1079,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
DESIRED_PYTHON: "3.11"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -1053,6 +1194,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
DESIRED_PYTHON: "3.11"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -1429,6 +1573,72 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_12-cuda12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.12"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_12-cuda12_9
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda12_9-test: # Testing
if: ${{ github.repository_owner == 'pytorch' }}
needs:
- manywheel-py3_12-cuda12_9-build
- get-label-type
uses: ./.github/workflows/_binary-test-linux.yml
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.12"
build_name: manywheel-py3_12-cuda12_9
build_environment: linux-binary-manywheel
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_12-cuda12_9-test
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.12"
build_name: manywheel-py3_12-cuda12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_12-cuda13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -1535,6 +1745,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
DESIRED_PYTHON: "3.12"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -1647,6 +1860,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
DESIRED_PYTHON: "3.12"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -2023,6 +2239,72 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13-cuda12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_13-cuda12_9
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13-cuda12_9-test: # Testing
if: ${{ github.repository_owner == 'pytorch' }}
needs:
- manywheel-py3_13-cuda12_9-build
- get-label-type
uses: ./.github/workflows/_binary-test-linux.yml
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13"
build_name: manywheel-py3_13-cuda12_9
build_environment: linux-binary-manywheel
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13-cuda12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_13-cuda12_9-test
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13"
build_name: manywheel-py3_13-cuda12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13-cuda13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -2129,6 +2411,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
DESIRED_PYTHON: "3.13"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -2241,6 +2526,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
DESIRED_PYTHON: "3.13"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -2617,6 +2905,72 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13t-cuda12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13t"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_13t-cuda12_9
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13t-cuda12_9-test: # Testing
if: ${{ github.repository_owner == 'pytorch' }}
needs:
- manywheel-py3_13t-cuda12_9-build
- get-label-type
uses: ./.github/workflows/_binary-test-linux.yml
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13t"
build_name: manywheel-py3_13t-cuda12_9
build_environment: linux-binary-manywheel
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13t-cuda12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_13t-cuda12_9-test
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.13t"
build_name: manywheel-py3_13t-cuda12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13t-cuda13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -2723,6 +3077,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
DESIRED_PYTHON: "3.13t"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -2835,6 +3192,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
DESIRED_PYTHON: "3.13t"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -3211,6 +3571,72 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14-cuda12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_14-cuda12_9
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14-cuda12_9-test: # Testing
if: ${{ github.repository_owner == 'pytorch' }}
needs:
- manywheel-py3_14-cuda12_9-build
- get-label-type
uses: ./.github/workflows/_binary-test-linux.yml
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14"
build_name: manywheel-py3_14-cuda12_9
build_environment: linux-binary-manywheel
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14-cuda12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_14-cuda12_9-test
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14"
build_name: manywheel-py3_14-cuda12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14-cuda13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -3317,6 +3743,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
DESIRED_PYTHON: "3.14"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -3429,6 +3858,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
DESIRED_PYTHON: "3.14"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -3805,6 +4237,72 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14t-cuda12_9-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14t"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_14t-cuda12_9
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14t-cuda12_9-test: # Testing
if: ${{ github.repository_owner == 'pytorch' }}
needs:
- manywheel-py3_14t-cuda12_9-build
- get-label-type
uses: ./.github/workflows/_binary-test-linux.yml
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14t"
build_name: manywheel-py3_14t-cuda12_9
build_environment: linux-binary-manywheel
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14t-cuda12_9-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_14t-cuda12_9-test
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu129
GPU_ARCH_VERSION: "12.9"
GPU_ARCH_TYPE: cuda
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.9
DESIRED_PYTHON: "3.14t"
build_name: manywheel-py3_14t-cuda12_9
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14t-cuda13_0-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
@ -3911,6 +4409,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
DESIRED_PYTHON: "3.14t"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm
@ -4023,6 +4524,9 @@ jobs:
DOCKER_IMAGE: manylinux2_28-builder
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
DESIRED_PYTHON: "3.14t"
permissions:
id-token: write
contents: read
steps:
- name: Setup ROCm
uses: ./.github/actions/setup-rocm

2
.github/workflows/inductor-perf-test-nightly-h100.yml vendored
View File

@ -130,7 +130,7 @@ jobs:
name: test-periodically
uses: ./.github/workflows/_linux-test.yml
needs: build
if: github.event.schedule == '15 0,12 * * 1-6'
if: github.event.schedule == '15 0 * * 1-6'
with:
build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm90
dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true

132
.github/workflows/inductor-perf-test-nightly-rocm-mi300.yml vendored Normal file
View File

@ -0,0 +1,132 @@
name: inductor-perf-nightly-rocm-mi300
on:
push:
tags:
- ciflow/inductor-perf-test-nightly-rocm-mi300/*
schedule:
- cron: 15 0 * * *
# NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
# out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
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: true
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_rocm_mi300,inductor_timm_perf_rocm_mi300,inductor_torchbench_perf_rocm_mi300
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
linux-jammy-rocm-py3_10-inductor-benchmark-build:
if: github.repository_owner == 'pytorch'
name: rocm-py3_10-inductor-benchmark-build
uses: ./.github/workflows/_linux-build.yml
with:
build-environment: linux-jammy-rocm-py3_10
docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
test-matrix: |
{ include: [
{ config: "inductor_huggingface_perf_rocm_mi300", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_huggingface_perf_rocm_mi300", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_huggingface_perf_rocm_mi300", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_huggingface_perf_rocm_mi300", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_huggingface_perf_rocm_mi300", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm_mi300", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm_mi300", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm_mi300", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm_mi300", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm_mi300", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm_mi300", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm_mi300", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm_mi300", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.gfx942.1" },
]}
secrets: inherit
linux-jammy-rocm-py3_10-inductor-benchmark-test:
permissions:
id-token: write
contents: read
name: rocm-py3_10-inductor-benchmark-test
uses: ./.github/workflows/_rocm-test.yml
needs: linux-jammy-rocm-py3_10-inductor-benchmark-build
with:
build-environment: linux-jammy-rocm-py3_10
dashboard-tag: training-true-inference-true-default-true-dynamic-true-cudagraphs-true-cppwrapper-true-aotinductor-true-freezing_cudagraphs-true-cudagraphs_low_precision-true
docker-image: ${{ needs.linux-jammy-rocm-py3_10-inductor-benchmark-build.outputs.docker-image }}
test-matrix: ${{ needs.linux-jammy-rocm-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

46
.github/workflows/inductor-perf-test-nightly-rocm.yml → .github/workflows/inductor-perf-test-nightly-rocm-mi355.yml vendored
View File

@ -1,11 +1,11 @@
name: inductor-perf-nightly-rocm
name: inductor-perf-nightly-rocm-mi355
on:
push:
tags:
- ciflow/inductor-perf-test-nightly-rocm/*
- ciflow/inductor-perf-test-nightly-rocm-mi355/*
schedule:
- cron: 0 7 * * 0,3
- cron: 15 0 * * *
# NB: GitHub has an upper limit of 10 inputs here, so before we can sort it
# out, let try to run torchao cudagraphs_low_precision as part of cudagraphs
workflow_dispatch:
@ -59,7 +59,7 @@ on:
description: The list of configs used the benchmark
required: false
type: string
default: inductor_huggingface_perf_rocm,inductor_timm_perf_rocm,inductor_torchbench_perf_rocm
default: inductor_huggingface_perf_rocm_mi355,inductor_timm_perf_rocm_mi355,inductor_torchbench_perf_rocm_mi355
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' }}
@ -88,23 +88,27 @@ jobs:
docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3-benchmarks
test-matrix: |
{ include: [
{ config: "inductor_huggingface_perf_rocm", shard: 1, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_huggingface_perf_rocm", shard: 2, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_huggingface_perf_rocm", shard: 3, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_huggingface_perf_rocm", shard: 4, num_shards: 4, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_timm_perf_rocm", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm", shard: 1, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm", shard: 2, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm", shard: 3, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm", shard: 4, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm", shard: 5, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm", shard: 6, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm", shard: 7, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_torchbench_perf_rocm", shard: 8, num_shards: 8, runner: "linux.rocm.gpu.gfx942.1" },
{ config: "inductor_huggingface_perf_rocm_mi355", shard: 1, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_huggingface_perf_rocm_mi355", shard: 2, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_huggingface_perf_rocm_mi355", shard: 3, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_huggingface_perf_rocm_mi355", shard: 4, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_huggingface_perf_rocm_mi355", shard: 5, num_shards: 5, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_timm_perf_rocm_mi355", shard: 1, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_timm_perf_rocm_mi355", shard: 2, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_timm_perf_rocm_mi355", shard: 3, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_timm_perf_rocm_mi355", shard: 4, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_timm_perf_rocm_mi355", shard: 5, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_timm_perf_rocm_mi355", shard: 6, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_timm_perf_rocm_mi355", shard: 7, num_shards: 7, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 1, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 2, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 3, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 4, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 5, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 6, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 7, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 8, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
{ config: "inductor_torchbench_perf_rocm_mi355", shard: 9, num_shards: 9, runner: "linux.rocm.gpu.mi355.1" },
]}
secrets: inherit

11
.github/workflows/lint.yml vendored
View File

@ -12,6 +12,7 @@ on:
- landchecks/*
tags:
- ciflow/pull/*
- ciflow/trunk/*
workflow_dispatch:
permissions: read-all
@ -32,10 +33,12 @@ jobs:
name: Get changed files
uses: ./.github/workflows/_get-changed-files.yml
with:
all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') }}
all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') || github.event_name == 'push' }}
lintrunner-clang:
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
# Needed to prevent deduping on HUD
name: lintrunner-clang-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
needs: [get-label-type, get-changed-files]
# Only run if there are changed files relevant to clangtidy / clangformat
if: |
@ -75,6 +78,7 @@ jobs:
# fails to find types when it should
lintrunner-mypy:
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
name: lintrunner-mypy-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
needs: [get-label-type, get-changed-files]
# Only run if there are changed files relevant to mypy
if: |
@ -99,6 +103,7 @@ jobs:
lintrunner-noclang:
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
name: lintrunner-noclang-${{ needs.get-changed-files.outputs.changed-files == '*' && 'all' || 'partial' }}
needs: [get-label-type, get-changed-files]
with:
timeout: 120
@ -113,9 +118,9 @@ jobs:
CHANGED_FILES="${{ needs.get-changed-files.outputs.changed-files }}"
echo "Running all other linters"
if [ "$CHANGED_FILES" = '*' ]; then
ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT --all-files" .github/scripts/lintrunner.sh
ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY --all-files" .github/scripts/lintrunner.sh
else
ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT ${CHANGED_FILES}" .github/scripts/lintrunner.sh
ADDITIONAL_LINTRUNNER_ARGS="--skip CLANGTIDY,CLANGFORMAT,MYPY,MYPYSTRICT,PYREFLY ${CHANGED_FILES}" .github/scripts/lintrunner.sh
fi
quick-checks:

61
.github/workflows/operator_benchmark.yml vendored
View File

@ -7,9 +7,11 @@ on:
workflow_dispatch:
inputs:
test_mode:
required: false
type: string
default: 'short'
type: choice
options:
- 'short'
- 'long'
- 'all'
description: tag filter for operator benchmarks, options from long, short, all
schedule:
# Run at 07:00 UTC every Sunday
@ -28,38 +30,49 @@ permissions:
contents: read
jobs:
opbenchmark-build:
x86-opbenchmark-build:
if: github.repository_owner == 'pytorch'
name: opbenchmark-build
name: x86-opbenchmark-build
uses: ./.github/workflows/_linux-build.yml
with:
build-environment: linux-jammy-py3.10-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
test-matrix: |
{ include: [
{ config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
{ config: "cpu_operator_benchmark_${{ inputs.test_mode || 'short' }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
]}
secrets: inherit
opbenchmark-on-demand-build:
if: ${{ github.event_name == 'workflow_dispatch' && github.repository_owner == 'pytorch' }}
name: opbenchmark-on-demand-build
uses: ./.github/workflows/_linux-build.yml
with:
build-environment: linux-jammy-py3.10-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
test-matrix: |
{ include: [
{ config: "cpu_operator_benchmark_${{ inputs.test_mode }}", shard: 1, num_shards: 1, runner: "linux.12xlarge" },
]}
secrets: inherit
opbenchmark-test:
name: opbenchmark-test
x86-opbenchmark-test:
name: x86-opbenchmark-test
uses: ./.github/workflows/_linux-test.yml
needs: opbenchmark-build
needs: x86-opbenchmark-build
with:
build-environment: linux-jammy-py3.10-gcc11-build
docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
docker-image: ${{ needs.x86-opbenchmark-build.outputs.docker-image }}
test-matrix: ${{ needs.x86-opbenchmark-build.outputs.test-matrix }}
secrets: inherit
aarch64-opbenchmark-build:
if: github.repository_owner == 'pytorch'
name: aarch64-opbenchmark-build
uses: ./.github/workflows/_linux-build.yml
with:
build-environment: linux-jammy-aarch64-py3.10
runner: linux.arm64.m7g.4xlarge
docker-image-name: ci-image:pytorch-linux-jammy-aarch64-py3.10-gcc11
test-matrix: |
{ include: [
{ config: "cpu_operator_benchmark_short", shard: 1, num_shards: 1, runner: "linux.arm64.m8g.4xlarge" },
]}
secrets: inherit
aarch64-opbenchmark-test:
name: aarch64-opbenchmark-test
uses: ./.github/workflows/_linux-test.yml
needs: aarch64-opbenchmark-build
with:
build-environment: linux-jammy-aarch64-py3.10
docker-image: ${{ needs.aarch64-opbenchmark-build.outputs.docker-image }}
test-matrix: ${{ needs.aarch64-opbenchmark-build.outputs.test-matrix }}
secrets: inherit

10
.github/workflows/periodic.yml vendored
View File

@ -182,11 +182,11 @@ jobs:
docker-image-name: ci-image:pytorch-linux-jammy-cuda13.0-cudnn9-py3-gcc11
test-matrix: |
{ include: [
{ config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
{ config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
{ config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
{ config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
{ config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
{ config: "nogpu_AVX512", shard: 1, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
{ config: "nogpu_AVX512", shard: 2, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
{ config: "nogpu_AVX512", shard: 3, num_shards: 3, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
{ config: "nogpu_NO_AVX2", shard: 1, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
{ config: "nogpu_NO_AVX2", shard: 2, num_shards: 2, runner: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge" },
{ config: "jit_legacy", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g4dn.4xlarge.nvidia.gpu" },
]}
secrets: inherit

1
.github/workflows/pull.yml vendored
View File

@ -127,6 +127,7 @@ jobs:
uses: ./.github/workflows/_linux-build.yml
needs: get-label-type
with:
runner: linux.2xlarge.memory
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build-environment: linux-jammy-py3.10-clang18-asan
docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan

12
.github/workflows/rocm-mi355.yml vendored
View File

@ -45,12 +45,12 @@ jobs:
sync-tag: rocm-build
test-matrix: |
{ include: [
{ config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
{ config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
{ config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
{ config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
{ config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
{ config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.2" },
{ config: "default", shard: 1, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
{ config: "default", shard: 2, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
{ config: "default", shard: 3, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
{ config: "default", shard: 4, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
{ config: "default", shard: 5, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
{ config: "default", shard: 6, num_shards: 6, runner: "linux.rocm.gpu.mi355.1" },
]}
secrets: inherit

1
.github/workflows/slow.yml vendored
View File

@ -140,6 +140,7 @@ jobs:
uses: ./.github/workflows/_linux-build.yml
needs: get-label-type
with:
runner: linux.2xlarge.memory
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build-environment: linux-jammy-py3.10-clang18-asan
docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan

25
.github/workflows/trunk.yml vendored
View File

@ -180,13 +180,13 @@ jobs:
disable-monitor: false
secrets: inherit
win-vs2022-cuda12_6-py3-build:
name: win-vs2022-cuda12.6-py3
win-vs2022-cuda12_8-py3-build:
name: win-vs2022-cuda12.8-py3
uses: ./.github/workflows/_win-build.yml
needs: get-label-type
with:
build-environment: win-vs2022-cuda12.6-py3
cuda-version: "12.6"
build-environment: win-vs2022-cuda12.8-py3
cuda-version: "12.8"
runner: "${{ needs.get-label-type.outputs.label-type }}windows.4xlarge.nonephemeral"
secrets: inherit
@ -200,6 +200,23 @@ jobs:
cuda-arch-list: '8.0'
secrets: inherit
# Test cross-compiled models with Windows libs extracted from wheel
cross-compile-linux-test:
name: cross-compile-linux-test
uses: ./.github/workflows/_linux-test.yml
needs:
- linux-jammy-cuda12_8-py3_10-gcc11-build
- get-label-type
- win-vs2022-cuda12_8-py3-build
with:
build-environment: linux-jammy-cuda12.8-py3.10-gcc11
docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-build.outputs.docker-image }}
test-matrix: |
{ include: [
{ config: "aoti_cross_compile_for_windows", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu", win_torch_wheel_artifact: "win-vs2022-cuda12.8-py3" },
]}
secrets: inherit
verify-cachebench-cpu-build:
name: verify-cachebench-cpu-build
uses: ./.github/workflows/_linux-build.yml

4
.github/workflows/vllm.yml vendored
View File

@ -46,7 +46,7 @@ jobs:
runner: linux.24xlarge.memory
test-matrix: |
{ include: [
{ config: "vllm_basic_correctness_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_basic_correctness_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_basic_models_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_entrypoints_test", shard: 1, num_shards: 1,runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_regression_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
@ -54,7 +54,7 @@ jobs:
{ config: "vllm_pytorch_compilation_unit_tests", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_lora_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_multi_model_test_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
{ config: "vllm_languagde_model_test_extended_generation_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
{ config: "vllm_language_model_test_extended_generation_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu"},
{ config: "vllm_distributed_test_2_gpu_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_lora_test", shard: 0, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_lora_test", shard: 1, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },

2
.gitignore vendored
View File

@ -374,6 +374,7 @@ third_party/ruy/
third_party/glog/
# Virtualenv
.venv/
venv/
# Log files
@ -395,3 +396,4 @@ android/pytorch_android_torchvision/.cxx
CLAUDE.local.md
/test_*.py
/debug_*.py
CLAUDE_CONTEXT/

40
.lintrunner.toml
View File

@ -209,6 +209,46 @@ command = [
'@{{PATHSFILE}}'
]
[[linter]]
code = 'PYREFLY'
include_patterns = [
'torch/**/*.py',
'torch/**/*.pyi',
'torchgen/**/*.py',
'torchgen/**/*.pyi',
'functorch/**/*.py',
'functorch/**/*.pyi',
]
exclude_patterns = []
command = [
'python3',
'tools/linter/adapters/pyrefly_linter.py',
'--config=pyrefly.toml',
]
init_command = [
'python3',
'tools/linter/adapters/pip_init.py',
'--dry-run={{DRYRUN}}',
'numpy==2.1.0 ; python_version >= "3.12"',
'expecttest==0.3.0',
'pyrefly==0.36.2',
'sympy==1.13.3',
'types-requests==2.27.25',
'types-pyyaml==6.0.2',
'types-tabulate==0.8.8',
'types-protobuf==5.29.1.20250403',
'types-setuptools==79.0.0.20250422',
'types-jinja2==2.11.9',
'types-colorama==0.4.6',
'filelock==3.18.0',
'junitparser==2.1.1',
'rich==14.1.0',
'optree==0.17.0',
'types-openpyxl==3.1.5.20250919',
'types-python-dateutil==2.9.0.20251008'
]
[[linter]]
code = 'CLANGTIDY'
include_patterns = [

14
CODEOWNERS
View File

@ -201,3 +201,17 @@ torch/backends/cudnn/ @eqy @syed-ahmed @Aidyn-A
/torch/csrc/stable/ @janeyx99 @mikaylagawarecki
/torch/headeronly/ @janeyx99
/torch/header_only_apis.txt @janeyx99
# FlexAttention
/torch/nn/attention/flex_attention.py @drisspg
/torch/_higher_order_ops/flex_attention.py @drisspg
/torch/_inductor/kernel/flex/ @drisspg
/torch/_inductor/codegen/cpp_flex_attention_template.py @drisspg
/test/inductor/test_flex_attention.py @drisspg
/test/inductor/test_flex_decoding.py @drisspg
# Low Precision GEMMs
/aten/src/ATen/native/cuda/Blas.cpp @drisspg @slayton58
/aten/src/ATen/cuda/CUDABlas.cpp @drisspg @slayton58
/aten/src/ATen/cuda/CUDABlas.h @drisspg @slayton58
/test/test_scaled_matmul_cuda.py @drisspg @slayton58

102
aten/src/ATen/CMakeLists.txt
View File

@ -256,6 +256,7 @@ endif()
IF(USE_FBGEMM_GENAI)
set(FBGEMM_THIRD_PARTY ${PROJECT_SOURCE_DIR}/third_party/fbgemm/external/)
set(FBGEMM_GENAI_SRCS ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize)
if(USE_CUDA)
# To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
# If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
@ -288,62 +289,69 @@ IF(USE_FBGEMM_GENAI)
set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
set(fbgemm_genai_mx8mx8bf16_grouped
set(fbgemm_genai_cuh
"${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
"${FBGEMM_GENAI_SRCS}/"
)
target_include_directories(fbgemm_genai PUBLIC
target_include_directories(fbgemm_genai PRIVATE
${FBGEMM_THIRD_PARTY}/cutlass/include
${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
${fbgemm_genai_mx8mx8bf16_grouped}
${fbgemm_genai_cuh}
${FBGEMM_GENAI_SRCS}/common/include/ # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
${FBGEMM_GENAI_SRCS}/include/ # includes fbgemm_gpu/torch_ops.h
)
else()
if(USE_ROCM)
# Only include the kernels we want to build to avoid increasing binary size.
file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
"${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
"${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
# Add additional HIPCC compiler flags for performance
set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-mllvm
-amdgpu-coerce-illegal-types=1
-mllvm
-enable-post-misched=0
-mllvm
-greedy-reverse-local-assignment=1
-fhip-new-launch-api)
# Add FBGEMM_GENAI include directories for torch_ops.h
list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
list(APPEND ATen_CUDA_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
elseif(USE_ROCM)
# Only include the kernels we want to build to avoid increasing binary size.
file(GLOB_RECURSE fbgemm_genai_native_rocm_hip
"${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/kernels/fp8_rowwise_grouped*.hip"
"${FBGEMM_GENAI_SRCS}/ck_extensions/fp8_rowwise_grouped/fp8_rowwise_grouped_gemm.hip")
set_source_files_properties(${fbgemm_genai_native_rocm_hip} PROPERTIES HIP_SOURCE_PROPERTY_FORMAT 1)
# Only compile for gfx942 for now.
# This is rather hacky, I could not figure out a clean solution :(
set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
endif()
set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
# Add additional HIPCC compiler flags for performance
set(FBGEMM_GENAI_EXTRA_HIPCC_FLAGS
-mllvm
-amdgpu-coerce-illegal-types=1
-mllvm
-enable-post-misched=0
-mllvm
-greedy-reverse-local-assignment=1
-fhip-new-launch-api)
hip_add_library(
fbgemm_genai STATIC
${fbgemm_genai_native_rocm_hip}
HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
target_include_directories(fbgemm_genai PUBLIC
# FBGEMM version of Composable Kernel is used due to some customizations
${FBGEMM_THIRD_PARTY}/composable_kernel/include
${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
${FBGEMM_THIRD_PARTY}/cutlass/include
${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
${FBGEMM_GENAI_SRCS}/common/include/ # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
${FBGEMM_GENAI_SRCS}/include/ # includes fbgemm_gpu/torch_ops.h
)
# Only compile for gfx942 for now.
# This is rather hacky, I could not figure out a clean solution :(
set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
endif()
set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
hip_add_library(
fbgemm_genai STATIC
${fbgemm_genai_native_rocm_hip}
HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
target_include_directories(fbgemm_genai PRIVATE
# FBGEMM version of Composable Kernel is used due to some customizations
${FBGEMM_THIRD_PARTY}/composable_kernel/include
${FBGEMM_THIRD_PARTY}/composable_kernel/library/include
${FBGEMM_THIRD_PARTY}/cutlass/include
${FBGEMM_THIRD_PARTY}/cutlass/tools/util/include
${FBGEMM_GENAI_SRCS}/common/include/ # includes fbgemm_gpu/quantize/utils.h, fbgemm_gpu/quantize/tuning_cache.hpp
${FBGEMM_GENAI_SRCS}/include/ # includes fbgemm_gpu/torch_ops.h
)
# Add FBGEMM_GENAI include directories for torch_ops.h
list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
list(APPEND ATen_HIP_INCLUDE ${PROJECT_SOURCE_DIR}/third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
endif()
endif()
@ -692,12 +700,6 @@ if(USE_CUDA AND NOT USE_ROCM)
list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/include)
list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/cutlass/tools/util/include)
# Add FBGEMM_GENAI include directories for torch_ops.h
if(USE_FBGEMM_GENAI)
list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/include)
list(APPEND ATen_CUDA_INCLUDE ${CMAKE_CURRENT_SOURCE_DIR}/../../../third_party/fbgemm/fbgemm_gpu/experimental/gen_ai/src/quantize/common/include)
endif()
if($ENV{ATEN_STATIC_CUDA})
if(CUDA_VERSION VERSION_LESS_EQUAL 12.9)
list(APPEND ATen_CUDA_DEPENDENCY_LIBS

29
aten/src/ATen/DLConvertor.cpp
View File

@ -389,37 +389,16 @@ void fillVersion<DLManagedTensorVersioned>(
// constructed out of ATen tensor
template <class T>
T* toDLPackImpl(const Tensor& src) {
auto view = src;
// Detect whether there is need to normalize the strides
// Background: gh-83069
//
// However, normalizing strides can come at a high-cost
// to slow down toDLPack conversion 3x, so we
// only normalize if needed.
//
// The following code detects whether the src follows
// a continuous pattern. If the src follows such pattern (common-case)
// then we do not need to normalize the strides.
bool need_normalize_strides = src.dim() == 1 && src.size(0) == 1 && src.stride(0) != 1;
// less common case, try normalizing the strides
if (need_normalize_strides) {
// create a new tensor with possibly normalized strides
// gh-83069
auto shape = src.sizes();
view = src.as_strided(shape, {1}, src.storage_offset());
}
ATenDLMTensor<T>* atDLMTensor(new ATenDLMTensor<T>);
atDLMTensor->handle = view;
atDLMTensor->handle = src;
atDLMTensor->tensor.manager_ctx = atDLMTensor;
atDLMTensor->tensor.deleter = &deleter<T>;
atDLMTensor->tensor.dl_tensor.data = view.data_ptr();
atDLMTensor->tensor.dl_tensor.data = src.data_ptr();
atDLMTensor->tensor.dl_tensor.device = torchDeviceToDLDevice(src.device());
atDLMTensor->tensor.dl_tensor.ndim = static_cast<int32_t>(src.dim());
atDLMTensor->tensor.dl_tensor.dtype = getDLDataType(src);
atDLMTensor->tensor.dl_tensor.shape = const_cast<int64_t*>(view.sizes().data());
atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(view.strides().data());
atDLMTensor->tensor.dl_tensor.shape = const_cast<int64_t*>(src.sizes().data());
atDLMTensor->tensor.dl_tensor.strides = const_cast<int64_t*>(src.strides().data());
atDLMTensor->tensor.dl_tensor.byte_offset = 0;
fillVersion(&atDLMTensor->tensor);

8
aten/src/ATen/DLConvertor.h
View File

@ -52,16 +52,16 @@ struct DLPackTraits {};
template <>
struct DLPackTraits<DLManagedTensor> {
inline static const char* capsule = "dltensor";
inline static const char* used = "used_dltensor";
inline static constexpr const char* capsule = "dltensor";
inline static constexpr const char* used = "used_dltensor";
inline static auto toDLPack = at::toDLPack;
inline static auto fromDLPack = at::fromDLPack;
};
template <>
struct DLPackTraits<DLManagedTensorVersioned> {
inline static const char* capsule = "dltensor_versioned";
inline static const char* used = "used_dltensor_versioned";
inline static constexpr const char* capsule = "dltensor_versioned";
inline static constexpr const char* used = "used_dltensor_versioned";
inline static auto toDLPack = at::toDLPackVersioned;
inline static auto fromDLPack = at::fromDLPackVersioned;
};

5
aten/src/ATen/InferSize.h
View File

@ -4,6 +4,7 @@
#include <c10/core/ScalarType.h>
#include <c10/core/SymIntArrayRef.h>
#include <c10/util/DimVector.h>
#include <c10/util/Exception.h>
#include <optional>
#include <sstream>
#include <vector>
@ -26,9 +27,7 @@ inline void infer_size_impl(
std::optional<int64_t> infer_dim;
for (int64_t dim = 0, ndim = shape.size(); dim != ndim; dim++) {
if (TORCH_GUARD_OR_FALSE(sym_eq(shape[dim], -1))) {
if (infer_dim) {
throw std::runtime_error("only one dimension can be inferred");
}
TORCH_CHECK(!infer_dim, "only one dimension can be inferred");
infer_dim = dim;
} else {
// in case of unbacked shape[dim] we assume it's not -1 and add a runtime

10
aten/src/ATen/LegacyBatchingRegistrations.cpp
View File

@ -58,7 +58,7 @@ namespace at {
namespace{
// PyTorch allows operations to specify dim 0 and dim -1 on a scalar tensor.
static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
return dim == 0 || dim == -1;
}
@ -365,7 +365,7 @@ Tensor select_batching_rule(const Tensor& self, int64_t dim, int64_t index) {
return self_physical.getPhysicalToLogicalMap().apply(result);
}
static int64_t getGradInputPhysicalDim(int64_t dim, IntArrayRef input_sizes, int64_t num_batch_dims) {
int64_t getGradInputPhysicalDim(int64_t dim, IntArrayRef input_sizes, int64_t num_batch_dims) {
return maybe_wrap_dim(dim, static_cast<int64_t>(input_sizes.size())) + num_batch_dims;
}
@ -488,7 +488,7 @@ Tensor view_as_complex_batching_rule(const Tensor& self) {
// Checks that the smallest batch stride is greater than the largest example
// stride. This is something we can support but we choose not to because it's
// potentially error prone.
static void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t num_batch_dims) {
void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t num_batch_dims) {
auto smallest_batch_stride = std::min_element(
physical_strides.begin(), physical_strides.begin() + num_batch_dims);
auto largest_example_stride = std::max_element(
@ -508,7 +508,7 @@ static void checkBatchDimsAtFrontInLayout(IntArrayRef physical_strides, int64_t
// given (sizes, strides, storage_offset) returns the maximum location that
// can be indexed (or nullopt if such a location doesn't exist, e.g., tensors
// with zero-size dims).
static std::optional<int64_t> maximum_indexable_location(
std::optional<int64_t> maximum_indexable_location(
IntArrayRef sizes, IntArrayRef strides, int64_t storage_offset) {
auto result = native::storage_size_for(sizes, strides);
if (result == 0) {
@ -521,7 +521,7 @@ static std::optional<int64_t> maximum_indexable_location(
// This checks that the range of possible memory locations accessible by
// x.as_strided(sizes, strides, maybe_storage_offset)
// are within the bounds of possible memory locations accessible by x.
static void checkBasicAsStridedValidForSlice(
void checkBasicAsStridedValidForSlice(
const Tensor& physical_tensor,
int64_t num_batch_dims,
IntArrayRef sizes,

10
aten/src/ATen/PythonTorchFunctionTLS.cpp
View File

@ -42,8 +42,14 @@ const PythonTorchFunctionTLS& PythonTorchFunctionTLS::get_state() {
}
bool torch_function_mode_enabled() {
return PythonTorchFunctionTLS::get_disabled_state() != TorchFunctionDisabledState::ALL_DISABLED &&
PythonTorchFunctionTLS::stack_len() > 0;
// Manually flatten because gcc is refusing to inline here. Note
// that we are still calling __tls_get_addr twice here with GCC,
// presumably because of
// https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81501 (which says
// the fix ships in GCC 16), but forcing inlining still improves
// performance.
const auto& ptfs = pythonTorchFunctionState;
return ptfs.disabled_state_ != TorchFunctionDisabledState::ALL_DISABLED && !ptfs.stack_.empty();
}
// This is needed to disambiguate the ternary torch function disabled states

1
aten/src/ATen/PythonTorchFunctionTLS.h
View File

@ -27,6 +27,7 @@ struct TORCH_API PythonTorchFunctionTLS {
TorchFunctionDisabledState disabled_state_ =
TorchFunctionDisabledState::ENABLED;
std::vector<std::shared_ptr<c10::SafePyObject>> stack_;
friend TORCH_API bool torch_function_mode_enabled();
};
TORCH_API bool torch_function_mode_enabled();

2
aten/src/ATen/SavedTensorHooks.cpp
View File

@ -13,7 +13,7 @@ namespace {
// and left at true for the rest of the execution.
// It's an optimization so that users who never use default hooks don't need to
// read the thread_local variables pack_hook_ and unpack_hook_.
static bool is_initialized(false);
bool is_initialized(false);
}
static void assertSavedTensorHooksNotDisabled() {

2
aten/src/ATen/TensorIterator.cpp
View File

@ -56,7 +56,7 @@ inline void get_strides(int64_t* strides, ArrayRef<OperandInfo> operands, int64_
}
}
static OptionalTensorRef make_otr(const TensorBase &tensor) {
OptionalTensorRef make_otr(const TensorBase &tensor) {
if (tensor.defined()) {
return OptionalTensorRef(tensor);
} else {

2
aten/src/ATen/autocast_mode.cpp
View File

@ -36,7 +36,7 @@ namespace {
using weakref_type = c10::weak_intrusive_ptr<TensorImpl, UndefinedTensorImpl>;
using val_type = std::tuple<weakref_type, Tensor>;
static ska::flat_hash_map<TensorImpl*, val_type>& get_cached_casts() {
ska::flat_hash_map<TensorImpl*, val_type>& get_cached_casts() {
static ska::flat_hash_map<TensorImpl*, val_type> cached_casts;
return cached_casts;
}

4
aten/src/ATen/core/CachingHostAllocator.cpp
View File

@ -6,9 +6,9 @@ namespace at {
namespace {
static std::array<HostAllocator*, at::COMPILE_TIME_MAX_DEVICE_TYPES>
std::array<HostAllocator*, at::COMPILE_TIME_MAX_DEVICE_TYPES>
allocator_array{};
static std::array<uint8_t, at::COMPILE_TIME_MAX_DEVICE_TYPES>
std::array<uint8_t, at::COMPILE_TIME_MAX_DEVICE_TYPES>
allocator_priority{};
} // anonymous namespace

8
aten/src/ATen/core/PhiloxRNGEngine.h
View File

@ -229,10 +229,10 @@ private:
}
static const uint32_t kPhilox10A = 0x9E3779B9;
static const uint32_t kPhilox10B = 0xBB67AE85;
static const uint32_t kPhiloxSA = 0xD2511F53;
static const uint32_t kPhiloxSB = 0xCD9E8D57;
static constexpr uint32_t kPhilox10A = 0x9E3779B9;
static constexpr uint32_t kPhilox10B = 0xBB67AE85;
static constexpr uint32_t kPhiloxSA = 0xD2511F53;
static constexpr uint32_t kPhiloxSB = 0xCD9E8D57;
};
typedef philox_engine Philox4_32;

7
aten/src/ATen/core/ivalue.h
View File

@ -624,7 +624,14 @@ struct TORCH_API IValue final {
IValue(const c10::SymBool& i) {
if (auto mi = i.maybe_as_bool()) {
tag = Tag::Bool;
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__
payload.u.as_int = *mi;
#elif __BYTE_ORDER__ == __ORDER_BIG_ENDIAN__
/* due to byteorder if value assigned as_int, as_bool actually is not set correctly */
payload.u.as_bool = *mi;
#else
#error Unexpected or undefined __BYTE_ORDER__
#endif
} else {
tag = Tag::SymBool;
payload.u.as_intrusive_ptr = i.toSymNodeImpl().release();

1
aten/src/ATen/cpu/vec/vec128/vec128.h
View File

@ -8,6 +8,7 @@
#include <ATen/cpu/vec/vec128/vec128_bfloat16_neon.h>
#include <ATen/cpu/vec/vec128/vec128_float_neon.h>
#include <ATen/cpu/vec/vec128/vec128_half_neon.h>
#include <ATen/cpu/vec/vec128/vec128_int_aarch64.h>
#endif
#include <ATen/cpu/vec/vec128/vec128_convert.h>

794
aten/src/ATen/cpu/vec/vec128/vec128_int_aarch64.h Normal file
View File

@ -0,0 +1,794 @@
#pragma once
#include <ATen/cpu/vec/intrinsics.h>
#include <ATen/cpu/vec/vec_base.h>
#include <c10/macros/Macros.h>
#include <c10/util/irange.h>
namespace at::vec {
// Note [CPU_CAPABILITY namespace]
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// This header, and all of its subheaders, will be compiled with
// different architecture flags for each supported set of vector
// intrinsics. So we need to make sure they aren't inadvertently
// linked together. We do this by declaring objects in an `inline
// namespace` which changes the name mangling, but can still be
// accessed as `at::vec`.
inline namespace CPU_CAPABILITY {
#define VEC_INT_NEON_TEMPLATE(vl, bit) \
template <> \
struct is_vec_specialized_for<int##bit##_t> : std::bool_constant<true> {}; \
\
template <> \
class Vectorized<int##bit##_t> { \
using neon_type = int##bit##x##vl##_t; \
\
private: \
neon_type values; \
\
public: \
using value_type = int##bit##_t; \
using size_type = int; \
static constexpr size_type size() { \
return vl; \
} \
Vectorized() { \
values = vdupq_n_s##bit(0); \
} \
Vectorized(neon_type v) : values(v) {} \
Vectorized(int##bit##_t val); \
template < \
typename... Args, \
typename = std::enable_if_t<(sizeof...(Args) == size())>> \
Vectorized(Args... vals) { \
__at_align__ int##bit##_t buffer[size()] = {vals...}; \
values = vld1q_s##bit(buffer); \
} \
operator neon_type() const { \
return values; \
} \
static Vectorized<int##bit##_t> loadu( \
const void* ptr, \
int64_t count = size()); \
void store(void* ptr, int64_t count = size()) const; \
template <int64_t mask> \
static Vectorized<int##bit##_t> blend( \
const Vectorized<int##bit##_t>& a, \
const Vectorized<int##bit##_t>& b); \
static Vectorized<int##bit##_t> blendv( \
const Vectorized<int##bit##_t>& a, \
const Vectorized<int##bit##_t>& b, \
const Vectorized<int##bit##_t>& mask_) { \
return vbslq_s##bit(vreinterpretq_u##bit##_s##bit(mask_.values), b, a); \
} \
template <typename step_t> \
static Vectorized<int##bit##_t> arange( \
value_type base = 0, \
step_t step = static_cast<step_t>(1)); \
static Vectorized<int##bit##_t> set( \
const Vectorized<int##bit##_t>& a, \
const Vectorized<int##bit##_t>& b, \
int64_t count = size()); \
const int##bit##_t& operator[](int idx) const = delete; \
int##bit##_t& operator[](int idx) = delete; \
Vectorized<int##bit##_t> abs() const { \
return vabsq_s##bit(values); \
} \
Vectorized<int##bit##_t> real() const { \
return values; \
} \
Vectorized<int##bit##_t> imag() const { \
return vdupq_n_s##bit(0); \
} \
Vectorized<int##bit##_t> conj() const { \
return values; \
} \
Vectorized<int##bit##_t> neg() const { \
return vnegq_s##bit(values); \
} \
int##bit##_t reduce_add() const { \
return vaddvq_s##bit(values); \
} \
int##bit##_t reduce_max() const; \
Vectorized<int##bit##_t> operator==( \
const Vectorized<int##bit##_t>& other) const { \
return Vectorized<value_type>( \
vreinterpretq_s##bit##_u##bit(vceqq_s##bit(values, other.values))); \
} \
Vectorized<int##bit##_t> operator!=( \
const Vectorized<int##bit##_t>& other) const; \
Vectorized<int##bit##_t> operator<( \
const Vectorized<int##bit##_t>& other) const { \
return Vectorized<value_type>( \
vreinterpretq_s##bit##_u##bit(vcltq_s##bit(values, other.values))); \
} \
Vectorized<int##bit##_t> operator<=( \
const Vectorized<int##bit##_t>& other) const { \
return Vectorized<value_type>( \
vreinterpretq_s##bit##_u##bit(vcleq_s##bit(values, other.values))); \
} \
Vectorized<int##bit##_t> operator>( \
const Vectorized<int##bit##_t>& other) const { \
return Vectorized<value_type>( \
vreinterpretq_s##bit##_u##bit(vcgtq_s##bit(values, other.values))); \
} \
Vectorized<int##bit##_t> operator>=( \
const Vectorized<int##bit##_t>& other) const { \
return Vectorized<value_type>( \
vreinterpretq_s##bit##_u##bit(vcgeq_s##bit(values, other.values))); \
} \
Vectorized<int##bit##_t> eq(const Vectorized<int##bit##_t>& other) const; \
Vectorized<int##bit##_t> ne(const Vectorized<int##bit##_t>& other) const; \
Vectorized<int##bit##_t> gt(const Vectorized<int##bit##_t>& other) const; \
Vectorized<int##bit##_t> ge(const Vectorized<int##bit##_t>& other) const; \
Vectorized<int##bit##_t> lt(const Vectorized<int##bit##_t>& other) const; \
Vectorized<int##bit##_t> le(const Vectorized<int##bit##_t>& other) const; \
}; \
template <> \
Vectorized<int##bit##_t> inline operator+( \
const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
return vaddq_s##bit(a, b); \
} \
template <> \
Vectorized<int##bit##_t> inline operator-( \
const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
return vsubq_s##bit(a, b); \
} \
template <> \
Vectorized<int##bit##_t> inline operator&( \
const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
return vandq_s##bit(a, b); \
} \
template <> \
Vectorized<int##bit##_t> inline operator|( \
const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
return vorrq_s##bit(a, b); \
} \
template <> \
Vectorized<int##bit##_t> inline operator^( \
const Vectorized<int##bit##_t>& a, const Vectorized<int##bit##_t>& b) { \
return veorq_s##bit(a, b); \
} \
Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::eq( \
const Vectorized<int##bit##_t>& other) const { \
return (*this == other) & Vectorized<int##bit##_t>(1); \
} \
Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::ne( \
const Vectorized<int##bit##_t>& other) const { \
return (*this != other) & Vectorized<int##bit##_t>(1); \
} \
Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::gt( \
const Vectorized<int##bit##_t>& other) const { \
return (*this > other) & Vectorized<int##bit##_t>(1); \
} \
Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::ge( \
const Vectorized<int##bit##_t>& other) const { \
return (*this >= other) & Vectorized<int##bit##_t>(1); \
} \
Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::lt( \
const Vectorized<int##bit##_t>& other) const { \
return (*this < other) & Vectorized<int##bit##_t>(1); \
} \
Vectorized<int##bit##_t> inline Vectorized<int##bit##_t>::le( \
const Vectorized<int##bit##_t>& other) const { \
return (*this <= other) & Vectorized<int##bit##_t>(1); \
}
VEC_INT_NEON_TEMPLATE(2, 64)
VEC_INT_NEON_TEMPLATE(4, 32)
VEC_INT_NEON_TEMPLATE(8, 16)
VEC_INT_NEON_TEMPLATE(16, 8)
inline int32_t Vectorized<int32_t>::reduce_max() const {
return vmaxvq_s32(values);
}
inline int16_t Vectorized<int16_t>::reduce_max() const {
return vmaxvq_s16(values);
}
inline int8_t Vectorized<int8_t>::reduce_max() const {
return vmaxvq_s8(values);
}
template <>
Vectorized<int32_t> inline operator*(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& b) {
return vmulq_s32(a, b);
}
template <>
Vectorized<int16_t> inline operator*(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& b) {
return vmulq_s16(a, b);
}
template <>
Vectorized<int8_t> inline operator*(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& b) {
return vmulq_s8(a, b);
}
template <>
inline Vectorized<int64_t> operator~(const Vectorized<int64_t>& a) {
int64x2_t val = a;
return ~val;
}
template <>
inline Vectorized<int32_t> operator~(const Vectorized<int32_t>& a) {
return vmvnq_s32(a);
}
template <>
inline Vectorized<int16_t> operator~(const Vectorized<int16_t>& a) {
return vmvnq_s16(a);
}
template <>
inline Vectorized<int8_t> operator~(const Vectorized<int8_t>& a) {
return vmvnq_s8(a);
}
inline Vectorized<int64_t> Vectorized<int64_t>::operator!=(
const Vectorized<int64_t>& other) const {
return ~(*this == other);
}
inline Vectorized<int32_t> Vectorized<int32_t>::operator!=(
const Vectorized<int32_t>& other) const {
return ~(*this == other);
}
inline Vectorized<int16_t> Vectorized<int16_t>::operator!=(
const Vectorized<int16_t>& other) const {
return ~(*this == other);
}
inline Vectorized<int8_t> Vectorized<int8_t>::operator!=(
const Vectorized<int8_t>& other) const {
return ~(*this == other);
}
template <>
Vectorized<int32_t> inline minimum(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& b) {
return vminq_s32(a, b);
}
template <>
Vectorized<int16_t> inline minimum(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& b) {
return vminq_s16(a, b);
}
template <>
Vectorized<int8_t> inline minimum(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& b) {
return vminq_s8(a, b);
}
template <>
Vectorized<int32_t> inline maximum(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& b) {
return vmaxq_s32(a, b);
}
template <>
Vectorized<int16_t> inline maximum(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& b) {
return vmaxq_s16(a, b);
}
template <>
Vectorized<int8_t> inline maximum(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& b) {
return vmaxq_s8(a, b);
}
template <int64_t mask>
Vectorized<int64_t> Vectorized<int64_t>::blend(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& b) {
// Build an array of flags: each bit of element is 1 if the corresponding bit
// in 'mask' is set, 0 otherwise.
uint64x2_t maskArray = {
(mask & 1LL) ? 0xFFFFFFFFFFFFFFFF : 0,
(mask & 2LL) ? 0xFFFFFFFFFFFFFFFF : 0};
// Use BSL to select elements from b where the mask is 1, else from a
return vbslq_s64(maskArray, b.values, a.values);
}
template <int64_t mask>
Vectorized<int32_t> Vectorized<int32_t>::blend(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& b) {
// Build an array of flags: each bit of element is 1 if the corresponding bit
// in 'mask' is set, 0 otherwise.
uint32x4_t maskArray = {
(mask & 1LL) ? 0xFFFFFFFF : 0,
(mask & 2LL) ? 0xFFFFFFFF : 0,
(mask & 4LL) ? 0xFFFFFFFF : 0,
(mask & 8LL) ? 0xFFFFFFFF : 0};
// Use BSL to select elements from b where the mask is 1, else from a
return vbslq_s32(maskArray, b.values, a.values);
}
template <int64_t mask>
Vectorized<int16_t> Vectorized<int16_t>::blend(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& b) {
// Build an array of flags: each bit of element is 1 if the corresponding bit
// in 'mask' is set, 0 otherwise.
uint16x8_t maskArray = {
(mask & 1LL) ? 0xFFFF : 0,
(mask & 2LL) ? 0xFFFF : 0,
(mask & 4LL) ? 0xFFFF : 0,
(mask & 8LL) ? 0xFFFF : 0,
(mask & 16LL) ? 0xFFFF : 0,
(mask & 32LL) ? 0xFFFF : 0,
(mask & 64LL) ? 0xFFFF : 0,
(mask & 128LL) ? 0xFFFF : 0};
// Use BSL to select elements from b where the mask is 1, else from a
return vbslq_s16(maskArray, b.values, a.values);
}
template <int64_t mask>
Vectorized<int8_t> Vectorized<int8_t>::blend(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& b) {
// Build an array of flags: each bit of element is 1 if the corresponding bit
// in 'mask' is set, 0 otherwise.
uint8x16_t maskArray = {
(mask & 1LL) ? 0xFF : 0,
(mask & 2LL) ? 0xFF : 0,
(mask & 4LL) ? 0xFF : 0,
(mask & 8LL) ? 0xFF : 0,
(mask & 16LL) ? 0xFF : 0,
(mask & 32LL) ? 0xFF : 0,
(mask & 64LL) ? 0xFF : 0,
(mask & 128LL) ? 0xFF : 0,
(mask & 256LL) ? 0xFF : 0,
(mask & 512LL) ? 0xFF : 0,
(mask & 1024LL) ? 0xFF : 0,
(mask & 2048LL) ? 0xFF : 0,
(mask & 4096LL) ? 0xFF : 0,
(mask & 8192LL) ? 0xFF : 0,
(mask & 16384LL) ? 0xFF : 0,
(mask & 32768LL) ? 0xFF : 0};
// Use BSL to select elements from b where the mask is 1, else from a
return vbslq_s8(maskArray, b.values, a.values);
}
#define VEC_INT_NEON_OPS(vl, bit) \
inline Vectorized<int##bit##_t>::Vectorized(int##bit##_t val) { \
values = vdupq_n_s##bit(val); \
} \
inline Vectorized<int##bit##_t> Vectorized<int##bit##_t>::loadu( \
const void* ptr, int64_t count) { \
if (count == size()) { \
return vld1q_s##bit(reinterpret_cast<const int##bit##_t*>(ptr)); \
} else { \
__at_align__ int##bit##_t tmp_values[size()]; \
for (const auto i : c10::irange(size())) { \
tmp_values[i] = 0; \
} \
std::memcpy( \
tmp_values, \
reinterpret_cast<const int##bit##_t*>(ptr), \
count * sizeof(int##bit##_t)); \
return vld1q_s##bit(reinterpret_cast<const int##bit##_t*>(tmp_values)); \
} \
} \
inline void Vectorized<int##bit##_t>::store(void* ptr, int64_t count) \
const { \
if (count == size()) { \
vst1q_s##bit(reinterpret_cast<int##bit##_t*>(ptr), values); \
} else { \
int##bit##_t tmp_values[size()]; \
vst1q_s##bit(reinterpret_cast<int##bit##_t*>(tmp_values), values); \
std::memcpy(ptr, tmp_values, count * sizeof(int##bit##_t)); \
} \
}
VEC_INT_NEON_OPS(2, 64)
VEC_INT_NEON_OPS(4, 32)
VEC_INT_NEON_OPS(8, 16)
VEC_INT_NEON_OPS(16, 8)
template <>
Vectorized<int64_t> inline operator*(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& b) {
int64x2_t x = a;
int64x2_t y = b;
return x * y;
}
template <>
Vectorized<int64_t> inline operator/(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& b) {
int64x2_t x = a;
int64x2_t y = b;
return x / y;
}
template <>
Vectorized<int32_t> inline operator/(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& b) {
int32x4_t x = a;
int32x4_t y = b;
return x / y;
}
inline int64_t Vectorized<int64_t>::reduce_max() const {
return std::max(values[0], values[1]);
}
template <>
Vectorized<int64_t> inline minimum(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& b) {
int64x2_t x = a;
int64x2_t y = b;
return {std::min(x[0], y[0]), std::min(x[1], y[1])};
}
template <>
Vectorized<int64_t> inline maximum(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& b) {
int64x2_t x = a;
int64x2_t y = b;
return {std::max(x[0], y[0]), std::max(x[1], y[1])};
}
template <typename step_t>
inline Vectorized<int64_t> Vectorized<int64_t>::arange(
int64_t base,
step_t step) {
const Vectorized<int64_t> base_vec(base);
const Vectorized<int64_t> step_vec(step);
const int64x2_t step_sizes = {0, 1};
return base_vec.values + step_sizes * step_vec.values;
}
template <typename step_t>
inline Vectorized<int32_t> Vectorized<int32_t>::arange(
int32_t base,
step_t step) {
const Vectorized<int32_t> base_vec(base);
const Vectorized<int32_t> step_vec(step);
const int32x4_t step_sizes = {0, 1, 2, 3};
return vmlaq_s32(base_vec, step_sizes, step_vec);
}
template <typename step_t>
inline Vectorized<int16_t> Vectorized<int16_t>::arange(
int16_t base,
step_t step) {
const Vectorized<int16_t> base_vec(base);
const Vectorized<int16_t> step_vec(step);
const int16x8_t step_sizes = {0, 1, 2, 3, 4, 5, 6, 7};
return vmlaq_s16(base_vec, step_sizes, step_vec);
}
template <typename step_t>
inline Vectorized<int8_t> Vectorized<int8_t>::arange(int8_t base, step_t step) {
const Vectorized<int8_t> base_vec(base);
const Vectorized<int8_t> step_vec(step);
const int8x16_t step_sizes = {
0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15};
return vmlaq_s8(base_vec, step_sizes, step_vec);
}
template <>
Vectorized<int64_t> inline operator>>(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& b) {
int64x2_t x = a;
int64x2_t y = b;
uint64x2_t u = vreinterpretq_u64_s64(y);
uint64x2_t z = {std::min(u[0], (uint64_t)63), std::min(u[1], (uint64_t)63)};
return x >> vreinterpretq_s64_u64(z);
}
template <>
Vectorized<int32_t> inline operator>>(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& b) {
int32x4_t x = a;
int32x4_t y = b;
uint32x4_t bound = vdupq_n_u32(31);
uint32x4_t z = vminq_u32(vreinterpretq_u32_s32(y), bound);
return x >> vreinterpretq_s32_u32(z);
}
template <>
Vectorized<int16_t> inline operator>>(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& b) {
int16x8_t x = a;
int16x8_t y = b;
uint16x8_t bound = vdupq_n_u16(15);
uint16x8_t z = vminq_u16(vreinterpretq_u16_s16(y), bound);
return x >> vreinterpretq_s16_u16(z);
}
template <>
Vectorized<int8_t> inline operator>>(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& b) {
int8x16_t x = a;
int8x16_t y = b;
uint8x16_t bound = vdupq_n_u8(7);
int8x16_t z = vreinterpretq_s8_u8(vminq_u8(vreinterpretq_u8_s8(y), bound));
return x >> z;
}
template <>
Vectorized<int64_t> inline operator<<(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& b) {
int64x2_t y = b;
uint64x2_t u = vreinterpretq_u64_s64(y);
uint64x2_t z = {std::min(u[0], (uint64_t)64), std::min(u[1], (uint64_t)64)};
return vshlq_s64(a, vreinterpretq_s64_u64(z));
}
template <>
Vectorized<int32_t> inline operator<<(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& b) {
int32x4_t y = b;
uint32x4_t bound = vdupq_n_u32(32);
uint32x4_t z = vminq_u32(vreinterpretq_u32_s32(y), bound);
return vshlq_s32(a, vreinterpretq_s32_u32(z));
}
template <>
Vectorized<int16_t> inline operator<<(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& b) {
int16x8_t y = b;
uint16x8_t bound = vdupq_n_u16(16);
uint16x8_t z = vminq_u16(vreinterpretq_u16_s16(y), bound);
return vshlq_s16(a, vreinterpretq_s16_u16(z));
}
template <>
Vectorized<int8_t> inline operator<<(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& b) {
int8x16_t y = b;
uint8x16_t bound = vdupq_n_u8(8);
int8x16_t z = vreinterpretq_s8_u8(vminq_u8(vreinterpretq_u8_s8(y), bound));
return vshlq_s8(a, z);
}
inline Vectorized<int64_t> Vectorized<int64_t>::set(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& b,
int64_t count) {
if (count == 0) {
return a;
} else if (count >= 2) {
return b;
} else {
int64x2_t c = {b.values[0], a.values[1]};
return c;
}
}
inline Vectorized<int32_t> Vectorized<int32_t>::set(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& b,
int64_t count) {
if (count == 0) {
return a;
} else if (count >= 4) {
return b;
} else {
// Build an array of flags: each bit of element is 1 if the corresponding
// bit in 'mask' is set, 0 otherwise.
uint32x4_t maskArray = {
(count >= 1LL) ? 0xFFFFFFFF : 0,
(count >= 2LL) ? 0xFFFFFFFF : 0,
(count >= 3LL) ? 0xFFFFFFFF : 0,
0};
// Use BSL to select elements from b where the mask is 1, else from a
return vbslq_s32(maskArray, b.values, a.values);
}
}
inline Vectorized<int16_t> Vectorized<int16_t>::set(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& b,
int64_t count) {
if (count == 0) {
return a;
} else if (count >= 8) {
return b;
} else {
// Build an array of flags: each bit of element is 1 if the corresponding
// bit in 'mask' is set, 0 otherwise.
uint16x8_t maskArray = {
static_cast<uint16_t>((count >= 1LL) ? 0xFFFF : 0),
static_cast<uint16_t>((count >= 2LL) ? 0xFFFF : 0),
static_cast<uint16_t>((count >= 3LL) ? 0xFFFF : 0),
static_cast<uint16_t>((count >= 4LL) ? 0xFFFF : 0),
static_cast<uint16_t>((count >= 5LL) ? 0xFFFF : 0),
static_cast<uint16_t>((count >= 6LL) ? 0xFFFF : 0),
static_cast<uint16_t>((count >= 7LL) ? 0xFFFF : 0),
0};
// Use BSL to select elements from b where the mask is 1, else from a
return vbslq_s16(maskArray, b.values, a.values);
}
}
inline Vectorized<int8_t> Vectorized<int8_t>::set(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& b,
int64_t count) {
if (count == 0) {
return a;
} else if (count >= 16) {
return b;
} else {
// Build an array of flags: each bit of element is 1 if the corresponding
// bit in 'mask' is set, 0 otherwise.
uint8x16_t maskArray = {
static_cast<uint8_t>((count >= 1LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 2LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 3LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 4LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 5LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 6LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 7LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 8LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 9LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 10LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 11LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 12LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 13LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 14LL) ? 0xFF : 0),
static_cast<uint8_t>((count >= 15LL) ? 0xFF : 0),
0};
// Use BSL to select elements from b where the mask is 1, else from a
return vbslq_s8(maskArray, b.values, a.values);
}
}
template <>
Vectorized<int16_t> inline operator/(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& b) {
Vectorized<int32_t> highBitsA = vmovl_high_s16(a);
Vectorized<int32_t> highBitsB = vmovl_high_s16(b);
Vectorized<int32_t> lowBitsA = vmovl_s16(vget_low_s16(a));
Vectorized<int32_t> lowBitsB = vmovl_s16(vget_low_s16(b));
int32x4_t highBitsResult = highBitsA / highBitsB;
int32x4_t lowBitsResult = lowBitsA / lowBitsB;
return vuzp1q_s16(
vreinterpretq_s16_s32(lowBitsResult),
vreinterpretq_s16_s32(highBitsResult));
}
template <>
Vectorized<int8_t> inline operator/(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& b) {
Vectorized<int16_t> highBitsA = vmovl_high_s8(a);
Vectorized<int16_t> highBitsB = vmovl_high_s8(b);
Vectorized<int16_t> lowBitsA = vmovl_s8(vget_low_s8(a));
Vectorized<int16_t> lowBitsB = vmovl_s8(vget_low_s8(b));
int16x8_t highBitsResult = highBitsA / highBitsB;
int16x8_t lowBitsResult = lowBitsA / lowBitsB;
return vuzp1q_s8(
vreinterpretq_s8_s16(lowBitsResult),
vreinterpretq_s8_s16(highBitsResult));
}
template <>
Vectorized<int64_t> inline clamp(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& min,
const Vectorized<int64_t>& max) {
return minimum(max, maximum(min, a));
}
template <>
Vectorized<int32_t> inline clamp(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& min,
const Vectorized<int32_t>& max) {
return minimum(max, maximum(min, a));
}
template <>
Vectorized<int16_t> inline clamp(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& min,
const Vectorized<int16_t>& max) {
return minimum(max, maximum(min, a));
}
template <>
Vectorized<int8_t> inline clamp(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& min,
const Vectorized<int8_t>& max) {
return minimum(max, maximum(min, a));
}
template <>
Vectorized<int64_t> inline clamp_max(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& max) {
return minimum(max, a);
}
template <>
Vectorized<int32_t> inline clamp_max(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& max) {
return minimum(max, a);
}
template <>
Vectorized<int16_t> inline clamp_max(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& max) {
return minimum(max, a);
}
template <>
Vectorized<int8_t> inline clamp_max(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& max) {
return minimum(max, a);
}
template <>
Vectorized<int64_t> inline clamp_min(
const Vectorized<int64_t>& a,
const Vectorized<int64_t>& min) {
return maximum(min, a);
}
template <>
Vectorized<int32_t> inline clamp_min(
const Vectorized<int32_t>& a,
const Vectorized<int32_t>& min) {
return maximum(min, a);
}
template <>
Vectorized<int16_t> inline clamp_min(
const Vectorized<int16_t>& a,
const Vectorized<int16_t>& min) {
return maximum(min, a);
}
template <>
Vectorized<int8_t> inline clamp_min(
const Vectorized<int8_t>& a,
const Vectorized<int8_t>& min) {
return maximum(min, a);
}
} // namespace CPU_CAPABILITY
} // namespace at::vec

4
aten/src/ATen/cpu/vec/vec256/vec256_qint.h
View File

@ -1377,7 +1377,7 @@ Vectorized<c10::quint8> inline maximum(
#if (defined(__aarch64__) && !defined(CPU_CAPABILITY_SVE256))
std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
at::vec::Vectorized<int8_t> src) {
auto s8x8 = vld1_s8(src.operator const int8_t*());
auto s8x8 = vget_low_s8(src);
auto s16x8 = vmovl_s8(s8x8);
auto s32x4_hi = vmovl_s16(vget_high_s16(s16x8));
@ -1402,7 +1402,7 @@ std::pair<Vectorized<float>, Vectorized<float>> inline convert_int8_to_float(
Vectorized<float> inline convert_int8_half_register_to_float(
at::vec::Vectorized<int8_t> src) {
auto s8x8 = vld1_s8(src.operator const int8_t*());
auto s8x8 = vget_low_s8(src);
auto s16x8 = vmovl_s8(s8x8);
auto s32x4_lo = vmovl_s16(vget_low_s16(s16x8));

57
aten/src/ATen/cuda/CUDABlas.cpp
View File

@ -16,6 +16,8 @@
#include <c10/util/irange.h>
#include <c10/core/ScalarType.h>
#include <ATen/cuda/detail/BLASConstants.h>
#ifdef USE_ROCM
#include <c10/cuda/CUDAStream.h>
#include <hipblaslt/hipblaslt-ext.hpp>
@ -108,7 +110,7 @@ static hipblasStatus_t rocBLASStatusToHIPStatus(rocblas_status error)
namespace {
static cublasOperation_t _cublasOpFromChar(char op) {
cublasOperation_t _cublasOpFromChar(char op) {
// NOLINTNEXTLINE(bugprone-switch-missing-default-case)
switch (op) {
case 'n':
@ -128,7 +130,7 @@ static cublasOperation_t _cublasOpFromChar(char op) {
"_cublasOpFromChar input should be 't', 'n' or 'c' but got `", op, "`");
}
static void _cublasAdjustLdLevel2(int64_t m, int64_t n, int64_t* lda) {
void _cublasAdjustLdLevel2(int64_t m, int64_t n, int64_t* lda) {
// Note: leading dimensions generally are checked that they are > 0
// and at least as big the result requires (even if the value won't
// be used).
@ -142,7 +144,7 @@ static void _cublasAdjustLdLevel2(int64_t m, int64_t n, int64_t* lda) {
*lda = std::max<int64_t>(m, 1);
}
static void _cublasAdjustLdLevel3(
void _cublasAdjustLdLevel3(
char transa,
char transb,
int64_t m,
@ -1954,13 +1956,15 @@ void scaled_gemm(
const void *result_scale_ptr,
int64_t result_ld,
ScalarType result_dtype,
bool use_fast_accum) {
bool use_fast_accum,
const std::optional<Tensor>& alpha) {
// Note: see `cublasCommonArgs` for various non-intuitive manipulations
// of input arguments to this function.
const auto computeType = CUBLAS_COMPUTE_32F;
const auto scaleType = CUDA_R_32F;
const float alpha_val = 1.0;
const float beta_val = 0.0;
// Note: alpha_val may change later depending on user-passed argument
float alpha_val = 1.0;
float beta_val = 0.0;
CuBlasLtMatmulDescriptor computeDesc(computeType, scaleType);
computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_TRANSA, _cublasOpFromChar(transa));
computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_TRANSB, _cublasOpFromChar(transb));
@ -2031,6 +2035,33 @@ void scaled_gemm(
computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_EPILOGUE, CUBLASLT_EPILOGUE_BIAS);
computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_BIAS_DATA_TYPE, ScalarTypeToCudaDataType(bias_dtype));
}
// Handle user-passed alpha
float *alpha_ptr = &alpha_val;
float *beta_ptr = &beta_val;
if (alpha.has_value()) {
auto& a = alpha.value();
// if device-tensor
if (a.is_cuda()) {
// NOTE: there are lifetime requirements on device-side pointers for alpha/beta -- the value must be
// valid & correct until the cublas call finishes (not is scheduled like host-side values). Thus
// we need to use allocations for alpha/beta that have some guarantees on lifetime - a statically
// managed 4B buffer for alpha that we'll copy the passed alpha value into, and constant memory
// for beta respectively.
float *user_alpha_ptr = at::cuda::detail::get_user_alpha_ptr();
at::Tensor user_alpha = at::from_blob(user_alpha_ptr, {1}, TensorOptions().device(kCUDA).dtype(kFloat));
user_alpha.copy_(a);
// Tell cublasLt we're using device-side pointers for alpha/beta
auto pointer_mode = CUBLASLT_POINTER_MODE_DEVICE;
computeDesc.setAttribute(CUBLASLT_MATMUL_DESC_POINTER_MODE, pointer_mode);
alpha_ptr = user_alpha.data_ptr<float>();
beta_ptr = at::cuda::detail::get_cublas_device_zero();
} else {
alpha_val = a.item<float>();
}
}
// For other data types, use the get_scale_mode function based on scaling type
// The SCALE_MODE attrs only exist in cuBLAS 12.8+/ROCm 7.0 or in recent hipblaslt,
// but we must invoke get_scale_mode anyways to trigger the version checks.
@ -2048,6 +2079,7 @@ void scaled_gemm(
cublasLtMatmulHeuristicResult_t heuristicResult = {};
int returnedResult = 0;
cublasLtHandle_t ltHandle = at::cuda::getCurrentCUDABlasLtHandle();
TORCH_CUDABLAS_CHECK(cublasLtMatmulAlgoGetHeuristic(
ltHandle,
computeDesc.descriptor(),
@ -2088,10 +2120,10 @@ void scaled_gemm(
auto is_valid_status = hipblaslt_ext::matmulIsAlgoSupported(
ltHandle,
computeDesc.descriptor(),
&alpha_val,
alpha_ptr,
Adesc.descriptor(),
Bdesc.descriptor(),
&beta_val,
beta_ptr,
Cdesc.descriptor(),
Ddesc.descriptor(),
all_algos[i].algo,
@ -2110,17 +2142,14 @@ void scaled_gemm(
cublasStatus_t cublasStatus = cublasLtMatmul(
ltHandle,
computeDesc.descriptor(),
&alpha_val,
alpha_ptr,
mat1_ptr,
Adesc.descriptor(),
mat2_ptr,
Bdesc.descriptor(),
&beta_val,
#ifdef USE_ROCM
beta_ptr,
// NOTE: always use result_ptr here, because cuBLASLt w/device beta=0 can't handle nullptr either
result_ptr, // unused, since beta_val is 0, but hipblaslt can't handle nullptr
#else
nullptr,
#endif // ifdef USE_ROCM
Cdesc.descriptor(),
result_ptr,
Ddesc.descriptor(),

3
aten/src/ATen/cuda/CUDABlas.h
View File

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

20
aten/src/ATen/cuda/CUDAGeneratorImpl.cpp
View File

@ -15,19 +15,19 @@ namespace cuda::detail {
namespace {
// Total number of gpus in the system.
static int64_t num_gpus;
int64_t num_gpus;
// Ensures default_gens_cuda is initialized once.
static std::deque<c10::once_flag> cuda_gens_init_flag;
std::deque<c10::once_flag> cuda_gens_init_flag;
// Default, global CUDA generators, one per GPU.
static std::vector<Generator> default_gens_cuda;
std::vector<Generator> default_gens_cuda;
/*
* Populates the global variables related to CUDA generators
* Warning: this function must only be called once!
*/
static void initCUDAGenVector() {
void initCUDAGenVector() {
// Ensures we only call cudaGetDeviceCount only once.
static bool num_gpu_init_flag [[maybe_unused]] = []() {
num_gpus = static_cast<int32_t>(c10::cuda::device_count());
@ -325,9 +325,9 @@ uint64_t CUDAGeneratorImpl::seed() {
*/
c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
// The RNG state comprises the seed, and an offset used for Philox.
static const size_t seed_size = sizeof(uint64_t);
static const size_t offset_size = sizeof(int64_t);
static const size_t total_size = seed_size + offset_size;
constexpr size_t seed_size = sizeof(uint64_t);
constexpr size_t offset_size = sizeof(int64_t);
constexpr size_t total_size = seed_size + offset_size;
auto state_tensor = at::detail::empty_cpu({(int64_t)total_size}, ScalarType::Byte, std::nullopt, std::nullopt, std::nullopt, std::nullopt);
auto rng_state = state_tensor.data_ptr<uint8_t>();
@ -346,9 +346,9 @@ c10::intrusive_ptr<c10::TensorImpl> CUDAGeneratorImpl::get_state() const {
* and size of the internal state.
*/
void CUDAGeneratorImpl::set_state(const c10::TensorImpl& new_state) {
static const size_t seed_size = sizeof(uint64_t);
static const size_t offset_size = sizeof(int64_t);
static const size_t total_size = seed_size + offset_size;
constexpr size_t seed_size = sizeof(uint64_t);
constexpr size_t offset_size = sizeof(int64_t);
constexpr size_t total_size = seed_size + offset_size;
detail::check_rng_state(new_state);

4
aten/src/ATen/cuda/cub.cuh
View File

@ -177,7 +177,6 @@ inline void segmented_sort_pairs(
}
}
#if CUB_SUPPORTS_UNIQUE_BY_KEY()
template <typename KeysInputIteratorT, typename ValuesInputIteratorT, typename ValuesOutputIteratorT, typename NumSelectedIteratorT>
inline void unique_by_key(
KeysInputIteratorT keys_in, ValuesInputIteratorT values_in,
@ -193,7 +192,6 @@ inline void unique_by_key(
CUB_WRAPPER(NO_ROCM(at_cuda_detail)::cub::DeviceSelect::UniqueByKey,
keys_in, values_in, keys_out_, values_out, num_selected, num_input_items, c10::cuda::getCurrentCUDAStream());
}
#endif
namespace impl {
@ -579,7 +577,6 @@ inline void exclusive_scan(InputIteratorT input, OutputIteratorT output, ScanOpT
#endif
}
#if CUB_SUPPORTS_SCAN_BY_KEY()
template <typename KeysInputIteratorT, typename ValuesInputIteratorT, typename ValuesOutputIteratorT>
inline void inclusive_sum_by_key(KeysInputIteratorT keys, ValuesInputIteratorT input, ValuesOutputIteratorT output, int64_t num_items) {
@ -607,7 +604,6 @@ inline void inclusive_scan_by_key(KeysInputIteratorT keys, ValuesInputIteratorT
#endif
}
#endif
template <typename InputIteratorT, typename OutputIteratorT, typename NumSelectedIteratorT>
void unique(InputIteratorT input, OutputIteratorT output,

16
aten/src/ATen/cuda/cub_definitions.cuh
View File

@ -28,22 +28,6 @@
#define USE_GLOBAL_CUB_WRAPPED_NAMESPACE() false
#endif
// cub support for UniqueByKey is added to cub 1.16 in:
// https://github.com/NVIDIA/cub/pull/405
#if CUB_VERSION >= 101600
#define CUB_SUPPORTS_UNIQUE_BY_KEY() true
#else
#define CUB_SUPPORTS_UNIQUE_BY_KEY() false
#endif
// cub support for scan by key is added to cub 1.15
// in https://github.com/NVIDIA/cub/pull/376
#if CUB_VERSION >= 101500
#define CUB_SUPPORTS_SCAN_BY_KEY() 1
#else
#define CUB_SUPPORTS_SCAN_BY_KEY() 0
#endif
// cub support for cub::FutureValue is added to cub 1.15 in:
// https://github.com/NVIDIA/cub/pull/305
#if CUB_VERSION >= 101500

54
aten/src/ATen/cuda/detail/BLASConstants.cu Normal file
View File

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

11
aten/src/ATen/cuda/detail/BLASConstants.h Normal file
View File

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

52
aten/src/ATen/cuda/tunable/GemmCommon.h
View File

@ -13,6 +13,7 @@
#include <c10/core/ScalarType.h>
#include <ATen/cuda/tunable/TunableOp.h>
#include <ATen/cuda/tunable/Tunable.h>
#include <ATen/cuda/CUDABlas.h>
#include <ATen/cuda/Exceptions.h>
#include <c10/util/StringUtil.h>
@ -150,6 +151,7 @@ inline std::string ScalarTypeToBLASType(c10::ScalarType scalar_type) {
BLASType = "unknown";
}
return BLASType;
}
// Similar to Compute Type in GemmRocblas.h
@ -244,33 +246,25 @@ inline std::string to_string_epilogue(const at::cuda::blas::GEMMAndBiasActivatio
namespace detail {
static bool NumericalCheck(ScalarType dtype, void* c, void* other_c, int64_t size) {
static bool NumericalCheck(ScalarType dtype, void* c, void* other_c, int64_t size, const NumericalCheckConfig& config) {
if (!config.enabled) {
return true; // skip when disabled
}
auto options = at::TensorOptions().dtype(dtype).device(at::kCUDA);
// comparison done as 1D tensor
at::Tensor ref = at::from_blob(c, {size}, options);
at::Tensor oth = at::from_blob(other_c, {size}, options);
at::Tensor ref_float = ref.to(at::kFloat);
at::Tensor oth_float = oth.to(at::kFloat);
std::vector<double> atols{1e-1, 1e-2, 1e-3, 1e-4, 1e-5};
std::vector<double> rtols{1e-1, 1e-2, 1e-3, 1e-4, 1e-5};
double last_succeed_atol = 1;
double last_succeed_rtol = 1;
for (auto& atol : atols) {
for (auto& rtol : rtols) {
if (at::allclose(ref_float, oth_float, rtol, atol)) {
last_succeed_atol = atol;
last_succeed_rtol = rtol;
}
}
}
if (last_succeed_atol == 1) {
return false;
}
else {
TUNABLE_LOG3("├──verify numerics: atol=", last_succeed_atol, ", rtol=", last_succeed_rtol);
}
return true;
const bool ok = at::allclose(ref_float, oth_float, config.rtol, config.atol);
if (ok) {
TUNABLE_LOG3("├──verify numerics: PASSED with atol=", config.atol, ", rtol=", config.rtol);
} else {
TUNABLE_LOG3("├──verify numerics: FAILED with atol=", config.atol, ", rtol=", config.rtol);
}
return ok;
}
}
@ -355,8 +349,10 @@ struct GemmParams : OpParams {
}
TuningStatus NumericalCheck(GemmParams<T> *other) {
auto* ctx = getTuningContext();
auto cfg = ctx->GetNumericalCheckConfig();
auto c_dtype = c10::CppTypeToScalarType<T>::value;
return detail::NumericalCheck(c_dtype, c, other->c, GetSizeC()/sizeof(T)) ? OK : FAIL;
return detail::NumericalCheck(c_dtype, c, other->c, GetSizeC()/sizeof(T), cfg) ? OK : FAIL;
}
char transa{};
@ -449,8 +445,10 @@ struct GemmAndBiasParams : OpParams {
}
TuningStatus NumericalCheck(GemmAndBiasParams<T> *other) {
auto* ctx = getTuningContext();
auto cfg = ctx->GetNumericalCheckConfig();
auto c_dtype = c10::CppTypeToScalarType<T>::value;
return detail::NumericalCheck(c_dtype, c, other->c, GetSizeC()/sizeof(T)) ? OK : FAIL;
return detail::NumericalCheck(c_dtype, c, other->c, GetSizeC()/sizeof(T), cfg) ? OK : FAIL;
}
char transa{};
@ -546,8 +544,10 @@ struct GemmStridedBatchedParams : OpParams {
}
TuningStatus NumericalCheck(GemmStridedBatchedParams<T> *other) {
auto* ctx = getTuningContext();
auto cfg = ctx->GetNumericalCheckConfig();
auto c_dtype = c10::CppTypeToScalarType<C_Dtype>::value;
return detail::NumericalCheck(c_dtype, c, other->c, GetSizeC()/sizeof(T)) ? OK : FAIL;
return detail::NumericalCheck(c_dtype, c, other->c, GetSizeC()/sizeof(T), cfg) ? OK : FAIL;
}
char transa{};
@ -663,7 +663,9 @@ struct ScaledGemmParams : OpParams {
}
TuningStatus NumericalCheck(ScaledGemmParams<T> *other) {
return detail::NumericalCheck(c_dtype, c, other->c, GetSizeC()/sizeof(T)) ? OK : FAIL;
auto* ctx = getTuningContext();
auto cfg = ctx->GetNumericalCheckConfig();
return detail::NumericalCheck(c_dtype, c, other->c, GetSizeC()/sizeof(T), cfg) ? OK : FAIL;
}
char transa{};

5
aten/src/ATen/cuda/tunable/README.md
View File

@ -145,7 +145,7 @@ programmatically since the settings become fixed. Use the C++ or Python APIs ins
| PYTORCH_TUNABLEOP_VERBOSE | Default is 0. Set to 1 to enable basic logging. 2 for basic tuning status. 3 for full trace. |
| PYTORCH_TUNABLEOP_VERBOSE_FILENAME | Default is "err" for stderr. Set to "out" for stdout or a filename for capturing verbose logging. |
| PYTORCH_TUNABLEOP_FILENAME | Default is 'tunableop_results.csv'. |
| PYTORCH_TUNABLEOP_NUMERICAL_CHECK | Default is 0. Set to 1 to enable. |
| PYTORCH_TUNABLEOP_NUMERICAL_CHECK | Default is off. Set 'atol_rtol' to enable, for example "1e-5_1e-5". |
| PYTORCH_TUNABLEOP_ROCBLAS_ENABLED | Default is 1. Set to 0 to disable rocblas being considered during tuning. |
| PYTORCH_TUNABLEOP_HIPBLASLT_ENABLED | Default is 1. Set to 0 to disable hipblaslt being considered during tuning. |
| PYTORCH_TUNABLEOP_MAX_TUNING_DURATION_MS | Default is 30. Unit is milliseconds. |
@ -173,10 +173,9 @@ All python APIs exist in the `torch.cuda.tunable` module.
| get_max_tuning_iterations() -> int | |
| set_filename(filename: str, insert_device_ordinal: bool = False) -> None | |
| get_filename() -> str | |
| set_numerical_check_tolerances(enable: bool, atol: float, rtol: float) -> None | Enable or disable numerical checking; atol and rtol default to 1e-5.
| get_results() -> Tuple[str, str, str, float] | |
| get_validators() -> Tuple[str, str] | |
| write_file_on_exit(val: bool) -> None | Default is True. |
| write_file(filename: Optional[str] = None) -> None | If filename not given, it will call get_filename(). |
| read_file(filename: Optional[str] = None) -> None | If filename not given, it will call get_filename(). |
| tune_gemm_in_file(filename: str) -> None | read an untuned file and tune GEMMs in it. |
| mgpu_tune_gemm_in_file(filename_pattern: str, num_gpus: int) -> None: -> None | read one or more untuned files and tune all unique GEMMs on one or more GPUs. |

190
aten/src/ATen/cuda/tunable/Tunable.cpp
View File

@ -107,14 +107,30 @@ void TuningResultsManager::AddImpl(const std::string& op_signature,
}
void TuningResultsManager::Add(const std::string& op_signature, const std::string& params_signature, ResultEntry best) {
std::scoped_lock l{lock_};
bool is_new = false;
ResultEntry inserted = ResultEntry::Null();
auto it = results_.find(op_signature);
if (it == results_.end()) {
it = results_.insert({op_signature, {}}).first;
// ---- mutate maps under results lock ----
{
std::scoped_lock l{lock_};
auto& km = results_[op_signature]; // creates if missing
is_new = (km.find(params_signature) == km.end());
AddImpl(op_signature, params_signature, std::move(best), km);
if (is_new) {
inserted = km.at(params_signature); // snapshot for I/O after unlocking
}
}
if (!is_new) return; // only write once per unique (op, params)
TuningContext* ctx = getTuningContext();
if (ctx->IsTuningEnabled() && !ctx->IsRecordUntunedEnabled()) {
InitRealtimeAppend(ctx->GetFilename(), ctx->GetTuningResultsValidator().GetAllValidators());
if (is_new && realtime_out_ && realtime_out_->good()) {
AppendResultLine(op_signature, params_signature, inserted);
}
}
AddImpl(op_signature, params_signature, std::move(best), it->second);
}
void TuningResultsManager::RecordUntuned( std::ofstream& untuned_file, const std::string& op_signature,
@ -150,6 +166,77 @@ void TuningResultsManager::RecordUntuned( std::ofstream& untuned_file, const std
}
}
void TuningResultsManager::InitRealtimeAppend(const std::string& filename, const std::unordered_map<std::string, std::string>& validators) {
std::scoped_lock fl{realtime_file_mutex_};
if (realtime_out_ && realtime_out_->good() && realtime_filename_ == filename) {
return;
}
if (realtime_out_ && realtime_filename_ != filename) {
realtime_out_->flush();
realtime_out_->close();
realtime_out_.reset();
validators_written_ = false;
}
bool file_exists = false;
bool file_empty = true;
{
std::ifstream check_file(filename);
if (check_file.good()) {
file_exists = true;
file_empty = (check_file.peek() == std::ifstream::traits_type::eof());
}
}
realtime_out_ = std::make_unique<std::ofstream>(filename, std::ios::out | std::ios::app);
if (!realtime_out_->good()) {
TORCH_WARN("TunableOp realtime append: failed to open '", filename,"'");
realtime_out_.reset();
return;
}
if(!file_exists || file_empty) {
for(const auto& [key, val] : validators) {
(*realtime_out_) << "Validator," << key << "," << val << std::endl;
realtime_out_->flush();
}
validators_written_ = true;
TUNABLE_LOG2("Wrote validators to realtime output file");
}
realtime_filename_ = filename;
}
void TuningResultsManager::AppendResultLine(const std::string& op_sig, const std::string& param_sig, const ResultEntry& result) {
std::scoped_lock fl{realtime_file_mutex_};
if(!realtime_out_ || !realtime_out_->good()) {
return;
}
(*realtime_out_) << op_sig << "," << param_sig << "," << result << std::endl;
realtime_out_->flush(); //ensure immediate write to disk
TUNABLE_LOG3("Realtime append: ", op_sig, "(", param_sig, ") -> ", result);
}
void TuningResultsManager::CloseRealtimeAppend() {
std::scoped_lock fl{realtime_file_mutex_};
if(realtime_out_) {
realtime_out_->flush();
realtime_out_->close();
realtime_out_.reset();
TUNABLE_LOG2("Closed realtime output file");
}
}
void TuningResultsManager::Delete(const std::string& op_signature, const std::string& params_signature) {
std::scoped_lock l{lock_};
@ -396,7 +483,6 @@ TuningContext::TuningContext() :
tuning_enable_{true},
record_untuned_enable_{false},
manager_initialized_{false},
write_file_on_exit_{true},
numerics_check_enable_{false},
max_tuning_duration_ms_{30},
max_tuning_iterations_{100},
@ -417,20 +503,8 @@ TuningContext::~TuningContext() {
// but doesn't do any computation itself.
return;
}
auto filename = GetFilename();
if (IsTunableOpEnabled() && IsTuningEnabled() && !filename.empty() && write_file_on_exit_) {
if (results_count_from_input_file_ < GetTuningResultsManager().GetSize()) {
if (results_count_from_input_file_ > 0) {
TUNABLE_LOG1("additional tuning results available, rewriting file ", filename);
}
else {
TUNABLE_LOG1("writing file ", filename);
}
if (!WriteFile(filename)) {
TUNABLE_LOG1("failed to write file ", filename);
}
}
}
TUNABLE_LOG1("Closing File");
GetTuningResultsManager().CloseRealtimeAppend(); // Since, we do instant logging by default now.
if (untuned_file_.good()) {
untuned_file_.close();
@ -511,20 +585,54 @@ std::ofstream& TuningContext::GetUntunedFile(){
return untuned_file_;
}
void TuningContext::WriteFileOnExit(bool value) {
write_file_on_exit_ = value;
}
void TuningContext::EnableNumericsCheck(bool value) {
numerics_check_enable_ = value;
}
bool TuningContext::IsNumericsCheckEnabled() const {
const auto env = c10::utils::get_env("PYTORCH_TUNABLEOP_NUMERICAL_CHECK");
if (env == "1") {
return true;
NumericalCheckConfig TuningContext::GetNumericalCheckConfig() const {
const auto env_opt = c10::utils::get_env("PYTORCH_TUNABLEOP_NUMERICAL_CHECK");
if (!env_opt.has_value()) {
return numerics_cfg_;
}
return numerics_check_enable_;
const std::string& env = env_opt.value();
if (env == "0") {
return NumericalCheckConfig(false, 1e-5, 1e-5);
}
const size_t underscore = env.find('_');
TORCH_CHECK(
underscore != std::string::npos,
"Invalid PYTORCH_TUNABLEOP_NUMERICAL_CHECK format. "
"Expected 'atol_rtol', got: ",
env);
double atol = 0.0;
double rtol = 0.0;
try {
atol = std::stod(env.substr(0, underscore));
rtol = std::stod(env.substr(underscore + 1));
} catch (const std::exception& e) {
TORCH_CHECK(false, "Failed to parse PYTORCH_TUNABLEOP_NUMERICAL_CHECK: ", e.what());
}
TORCH_CHECK( atol > 0.0 && rtol > 0.0, "Tolerance values must be positive. atol=", atol, ", rtol=", rtol);
return NumericalCheckConfig(true, atol, rtol);
}
void TuningContext::SetNumericalCheckConfig(bool enabled, double atol, double rtol) {
TORCH_CHECK(atol > 0.0 && rtol > 0.0, "Numerical check tolerances must be positive");
numerics_cfg_ = {enabled, atol, rtol};
}
bool TuningContext::IsNumericsCheckEnabled() const {
const auto cfg = GetNumericalCheckConfig();
return cfg.enabled || numerics_check_enable_;
}
void TuningContext::SetMaxTuningDurationMs(int max_duration_ms) {
@ -634,11 +742,6 @@ TuningResultsManager& TuningContext::GetTuningResultsManager() {
auto filename = GetFilename();
if (!filename.empty() && !IsRecordUntunedEnabled()) {
ReadFile(filename);
// attempt immediately to open file for writing to catch errors early
std::ofstream file(filename, std::ios::out | std::ios::app);
if (!file.good()) {
TORCH_WARN("failed to open file '", filename, "' for writing; your tuning results will not be saved");
}
}
});
return manager_;
@ -744,27 +847,6 @@ bool TuningContext::ReadFile(const std::string& filename_) {
return true;
}
bool TuningContext::WriteFile(const std::string& filename_) {
std::string filename = filename_.empty() ? GetFilename() : filename_;
std::ofstream file(filename, std::ios::out | std::ios::trunc);
if (!file.good()) {
TUNABLE_LOG1("error opening tuning results file for writing ", filename);
return false;
}
auto validators = GetTuningResultsValidator().GetAllValidators();
for (const auto& [key, val] : validators) {
file << "Validator," << key << "," << val << std::endl;
}
auto results = GetTuningResultsManager().Dump();
for (const auto& [op_sig, kernelmap] : results) {
for (const auto& [param_sig, result] : kernelmap) {
file << op_sig << "," << param_sig << "," << result << std::endl;
}
}
file.close();
return true;
}
namespace {
struct MaybeDelete {

32
aten/src/ATen/cuda/tunable/Tunable.h
View File

@ -103,10 +103,24 @@ class TORCH_CUDA_CPP_API TuningResultsManager {
void RecordUntuned( std::ofstream& untuned_file, const std::string& op_signature,
const std::string& params_signature, const std::string& blas_signature);
void InitRealtimeAppend(
const std::string& filename,
const std::unordered_map<std::string, std::string>& validators);
void AppendResultLine(const std::string& op_sig,
const std::string& param_sig,
const ResultEntry& result);
void CloseRealtimeAppend(); // For clean shutdown
private:
std::mutex lock_;
std::mutex realtime_file_mutex_;
std::unique_ptr<std::ofstream> realtime_out_;
std::string realtime_filename_;
ResultsMap results_;
UntunedMap untuned_results_;
bool validators_written_ = false;
};
@ -134,6 +148,16 @@ class TORCH_CUDA_CPP_API TuningResultsValidator {
GetValidateFuncs validators_;
};
struct NumericalCheckConfig {
bool enabled{false};
double atol{1e-5};
double rtol{1e-5};
NumericalCheckConfig() = default;
NumericalCheckConfig(bool e, double a, double r) : enabled(e), atol(a), rtol(r) {}
};
class TORCH_CUDA_CPP_API TuningContext {
public:
TuningContext();
@ -155,6 +179,8 @@ class TORCH_CUDA_CPP_API TuningContext {
void EnableNumericsCheck(bool value);
bool IsNumericsCheckEnabled() const;
void SetNumericalCheckConfig(bool enabled, double atol, double rtol);
NumericalCheckConfig GetNumericalCheckConfig() const;
void SetMaxTuningDurationMs(int max_duration_ms);
int GetMaxTuningDurationMs() const;
@ -185,10 +211,7 @@ class TORCH_CUDA_CPP_API TuningContext {
void SetFilename(const std::string& filename, bool insert_device_ordinal=false);
std::string GetFilename() const;
void WriteFileOnExit(bool value);
bool ReadFile(const std::string& filename={});
bool WriteFile(const std::string& filename={});
template<class... Types>
void Log(int level, Types... args) {
@ -207,7 +230,6 @@ class TORCH_CUDA_CPP_API TuningContext {
bool tuning_enable_;
bool record_untuned_enable_;
bool manager_initialized_;
bool write_file_on_exit_;
bool numerics_check_enable_;
int max_tuning_duration_ms_;
int max_tuning_iterations_;
@ -222,6 +244,8 @@ class TORCH_CUDA_CPP_API TuningContext {
std::ofstream untuned_file_;
size_t results_count_from_input_file_;
bool is_shutting_down_;
NumericalCheckConfig numerics_cfg_{};
};
TORCH_CUDA_CPP_API TuningContext* getTuningContext();

3
aten/src/ATen/cuda/tunable/TunableGemm.h
View File

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

41
aten/src/ATen/cuda/tunable/TunableOp.h
View File

@ -267,27 +267,10 @@ class TunableOp {
for (size_t i = 0; i < op_names_.size(); i++) {
auto* candidate = ops_[op_names_[i]].get(); // borrow pointer
if (do_numerics_check) {
ParamsT* numerical_params = params->DeepCopy(false);
auto status = candidate->Call(numerical_params);
if (status != OK) {
numerical_params->Delete();
TUNABLE_LOG3("├──unsupported id=", i, ", ", op_sig, '(', params_sig, ") ", op_names_[i]);
continue;
}
status = reference_params->NumericalCheck(numerical_params);
numerical_params->Delete();
if (status != OK) {
TUNABLE_LOG3("├──numerics check failed for id=", i, ", ", op_sig, '(', params_sig, ") ", op_names_[i]);
continue;
}
}
else {
auto status = candidate->Call(reusable_params[0]);
if (status != OK) {
TUNABLE_LOG3("├──unsupported id=", i, ", ", op_sig, '(', params_sig, ") ", op_names_[i]);
continue;
}
auto status = candidate->Call(reusable_params[0]);
if (status != OK) {
TUNABLE_LOG3("├──unsupported id=", i, ", ", op_sig, '(', params_sig, ") ", op_names_[i]);
continue;
}
// collect a small profile
@ -310,6 +293,22 @@ class TunableOp {
continue;
}
if (do_numerics_check) {
ParamsT* numerical_params = params->DeepCopy(false);
auto status = candidate->Call(numerical_params);
if (status != OK) {
numerical_params->Delete();
TUNABLE_LOG3("├──unsupported id=", i, ", ", op_sig, '(', params_sig, ") ", op_names_[i]);
continue;
}
status = reference_params->NumericalCheck(numerical_params);
numerical_params->Delete();
if (status != OK) {
TUNABLE_LOG3("├──numerics check failed for id=", i, ", ", op_sig, '(', params_sig, ") ", op_names_[i]);
continue;
}
}
// for warmup does user set max duration, max iters, or both?
// warmup is skipped by default, i.e. warmup_iter = 0
// warmup will be set to the non-zero value of max_warmup_duration

9
aten/src/ATen/cudnn/Types.cpp
View File

@ -2,6 +2,8 @@
#include <ATen/ATen.h>
#include <c10/util/Exception.h>
namespace at::native {
cudnnDataType_t getCudnnDataTypeFromScalarType(const at::ScalarType dtype) {
@ -20,9 +22,10 @@ cudnnDataType_t getCudnnDataTypeFromScalarType(const at::ScalarType dtype) {
} else if (dtype == at::kByte) {
return CUDNN_DATA_UINT8;
}
std::string msg("getCudnnDataTypeFromScalarType() not supported for ");
msg += toString(dtype);
throw std::runtime_error(msg);
TORCH_CHECK(false,
"getCudnnDataTypeFromScalarType() not supported for ",
toString(dtype)
);
}
cudnnDataType_t getCudnnDataType(const at::Tensor& tensor) {

12
aten/src/ATen/functorch/BatchRulesLinearAlgebra.cpp
View File

@ -39,7 +39,7 @@ Tensor vdot_decomp(const Tensor& A, const Tensor& B) {
// NB: I wrote this like this because we *might* want its for a future matmul
// batch rule that isn't decomposed...
// "tv" = tensor @ vector
static std::tuple<Tensor, std::optional<int64_t>> tv_batch_rule(
std::tuple<Tensor, std::optional<int64_t>> tv_batch_rule(
const Tensor& self, std::optional<int64_t> self_bdim,
const Tensor& other, std::optional<int64_t> other_bdim) {
if (self_bdim && other_bdim) {
@ -66,7 +66,7 @@ static std::tuple<Tensor, std::optional<int64_t>> tv_batch_rule(
TORCH_INTERNAL_ASSERT(false, "can't get here");
}
static std::tuple<Tensor, std::optional<int64_t>> mv_batch_rule(
std::tuple<Tensor, std::optional<int64_t>> mv_batch_rule(
const Tensor& self, std::optional<int64_t> self_bdim,
const Tensor& other, std::optional<int64_t> other_bdim) {
auto self_logical_rank = rankWithoutBatchDim(self, self_bdim);
@ -79,7 +79,7 @@ static std::tuple<Tensor, std::optional<int64_t>> mv_batch_rule(
return tv_batch_rule(self, self_bdim, other, other_bdim);
}
static std::tuple<Tensor, std::optional<int64_t>> mm_batch_rule(
std::tuple<Tensor, std::optional<int64_t>> mm_batch_rule(
const Tensor& self, std::optional<int64_t> self_bdim,
const Tensor& other, std::optional<int64_t> other_bdim) {
auto self_logical_rank = rankWithoutBatchDim(self, self_bdim);
@ -94,7 +94,7 @@ static std::tuple<Tensor, std::optional<int64_t>> mm_batch_rule(
return std::make_tuple( at::matmul(self_, other_), 0 );
}
static std::tuple<Tensor, std::optional<int64_t>> bmm_batch_rule(
std::tuple<Tensor, std::optional<int64_t>> bmm_batch_rule(
const Tensor& self, std::optional<int64_t> self_bdim,
const Tensor& other, std::optional<int64_t> other_bdim) {
auto self_logical_rank = rankWithoutBatchDim(self, self_bdim);
@ -250,7 +250,7 @@ struct LinalgCheckMatrixBinaryRuleHelper<op_name, F, Func, typelist<A, B, T...>>
}
};
static void expect_at_least_rank(
void expect_at_least_rank(
const Tensor& tensor,
std::optional<int64_t> tensor_bdim,
int64_t expected_rank,
@ -472,7 +472,7 @@ atol_rtol_tensor_batch_rule(
return std::make_tuple(Func(input_, atol_, rtol_, hermitian), 0);
}
static std::tuple<Tensor, std::optional<int64_t>>
std::tuple<Tensor, std::optional<int64_t>>
pinv_batch_rule(
const Tensor& input, std::optional<int64_t> input_bdim, const std::optional<Tensor>& atol,
const std::optional<int64_t> atol_bdim, const std::optional<Tensor>& rtol,

38
aten/src/ATen/functorch/BatchRulesModules.cpp
View File

@ -213,40 +213,22 @@ static cudnn_grid_sample_backward_batch_rule(
return grid_sample_backward_helper_out(std::move(bw_out), 0, 0, bdim_size);
}
// TODO: replace with targetable functionalization
// uses functional formulation for one_hot under vmap to be compatible with
// fakeTensor/dynamic shapes and compiled functorch transforms.
// mirrors the meta path in aten/src/ATen/native/Onehot.cpp,
// but requires explicit positive num_classes under vmap to avoid
// data-dependent output shapes.
static Tensor one_hot_decomposition_hack(const Tensor &self, int64_t num_classes) {
TORCH_CHECK(self.dtype() == kLong, "one_hot is only applicable to index tensor.");
auto shape = self.sym_sizes().vec();
// empty tensor could be converted to one hot representation,
// but shape inference is not possible.
if (self.sym_numel() == 0) {
if (num_classes <= 0) {
TORCH_CHECK(false, "Can not infer total number of classes from empty tensor.");
} else {
shape.emplace_back(num_classes);
return at::empty_symint(shape, self.options());
}
}
// disallow implicit inference under vmap; this would be data-dependent
// and is intentionally guarded by Dynamo in torch/_dynamo/variables/torch.py.
TORCH_CHECK(num_classes > 0, "When vmap-ing torch.nn.functional.one_hot, please "
"provide an explicit positive num_classes argument.");
// Disabling all of the following checks. This is OK because scatter has checks too.
// Maybe one_hot should be a primitive wrt autograd so we don't have to deal with this.
// // non-empty tensor
// if (self.device().type() != at::kCUDA) {
// //for cuda, rely on device assert thrown by scatter
// TORCH_CHECK(self.min().item().toLong() >= 0, "Class values must be non-negative.");
// }
// if (self.device().type() != at::kCUDA) {
// //rely on device asserts from scatter to avoid sync here
// TORCH_CHECK(num_classes > self.max().item().toLong(), "Class values must be smaller than num_classes.");
// }
shape.emplace_back(num_classes);
Tensor ret = at::zeros_symint(shape, self.options());
return ret.scatter(-1, self.unsqueeze(-1), 1);
const auto options = self.options();
at::Tensor index = at::arange(num_classes, options);
return at::eq(self.unsqueeze(-1), index).to(at::kLong);
}
template <typename A, A a, typename C>

26
aten/src/ATen/functorch/BatchRulesScatterOps.cpp
View File

@ -12,13 +12,14 @@
#include <ATen/native/IndexKernel.h>
#include <ATen/native/IndexingUtils.h>
#include <torch/library.h>
#include <c10/util/Exception.h>
// NOLINTBEGIN(bugprone-unchecked-optional-access)
namespace at::functorch {
namespace {
static bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
for (const auto& bdim : bdims) {
if (bdim.has_value()) {
return true;
@ -27,7 +28,7 @@ static bool any_has_value(ArrayRef<std::optional<int64_t>> bdims) {
return false;
}
static int64_t get_num_leading_nones(ArrayRef<std::optional<Tensor>> indices) {
int64_t get_num_leading_nones(ArrayRef<std::optional<Tensor>> indices) {
int64_t result = 0;
for (const auto& idx : indices) {
if (!idx.has_value() || !idx->defined()) {
@ -39,7 +40,7 @@ static int64_t get_num_leading_nones(ArrayRef<std::optional<Tensor>> indices) {
return result;
}
static int64_t get_max_index_logical_dim(
int64_t get_max_index_logical_dim(
ArrayRef<std::optional<Tensor>> indices,
ArrayRef<std::optional<int64_t>> indices_bdims) {
int64_t max_logical_dim = -1;
@ -56,7 +57,7 @@ static int64_t get_max_index_logical_dim(
return max_logical_dim;
}
static std::vector<std::optional<Tensor>> batchIndices(
std::vector<std::optional<Tensor>> batchIndices(
at::TensorOptions options,
ArrayRef<std::optional<Tensor>> indices,
ArrayRef<std::optional<int64_t>> indices_bdims,
@ -94,9 +95,10 @@ static std::vector<std::optional<Tensor>> batchIndices(
if (index.has_value() && index->sym_numel() != 0) {
const auto idx_bdim = indices_bdims[i];
indices_.emplace_back(maybePadToLogicalRank(moveBatchDimToFront(index.value(), idx_bdim), idx_bdim, maxLogicalRank));
if (index.value().dtype() == kBool && indices_bdims[i].has_value()) {
throw std::runtime_error("vmap: We do not support batching operators that can support dynamic shape. Attempting to batch over indexing with a boolean mask.");
}
TORCH_CHECK(
!(index.value().dtype() == kBool) || !indices_bdims[i].has_value(),
"vmap: We do not support batching operators that can support dynamic shape. Attempting to batch over indexing with a boolean mask."
);
} else {
indices_.push_back(index);
}
@ -124,7 +126,7 @@ static std::vector<std::optional<Tensor>> batchIndices(
// Define an "advanced index" to be a selection object that is
// a non-trivial Tensor (i.e. it does not represent :).
static bool is_advanced_index(const std::optional<Tensor>& idx) {
bool is_advanced_index(const std::optional<Tensor>& idx) {
if (!idx.has_value()) {
return false;
}
@ -135,7 +137,7 @@ static bool is_advanced_index(const std::optional<Tensor>& idx) {
}
// See NOTE: [advanced indices adjacent] for definition
static bool are_advanced_indices_adjacent(ArrayRef<std::optional<Tensor>> indices) {
bool are_advanced_indices_adjacent(ArrayRef<std::optional<Tensor>> indices) {
int64_t num_advanced_indices_regions = 0;
bool in_advanced_indices_region = false;
for (const auto& idx : indices) {
@ -163,7 +165,7 @@ static bool are_advanced_indices_adjacent(ArrayRef<std::optional<Tensor>> indice
// - result: Tensor[B, 4, 5, 6, 2, 3, 7, 8]
// ------- ----
// region2 region1
static Tensor swap_regions(const Tensor& tensor, int64_t first_region_size, int64_t second_region_size) {
Tensor swap_regions(const Tensor& tensor, int64_t first_region_size, int64_t second_region_size) {
VmapDimVector permutation(tensor.dim(), 0);
std::iota(permutation.begin(), permutation.end(), 0);
std::rotate(
@ -551,7 +553,7 @@ Tensor &_index_put_impl__plumbing(Tensor &self, const List<std::optional<Tensor>
return self;
}
static Tensor maybe_permute_values(
Tensor maybe_permute_values(
const Tensor& values,
ArrayRef<std::optional<Tensor>> orig_indices,
ArrayRef<std::optional<int64_t>> orig_indices_bdims) {
@ -1050,7 +1052,7 @@ std::tuple<Tensor, std::optional<int64_t>> index_add_batch_rule(
other, other_bdim, alpha, false);
}
static std::tuple<Tensor,Tensor> binary_pointwise_align(
std::tuple<Tensor,Tensor> binary_pointwise_align(
const Tensor & self,
std::optional<int64_t> self_bdim,
const Tensor & mask,

2
aten/src/ATen/functorch/BatchRulesViews.cpp
View File

@ -346,7 +346,7 @@ std::tuple<Tensor, std::optional<int64_t>> slice_batch_rule(
return std::make_tuple(std::move(result), 0);
}
static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
return dim == 0 || dim == -1;
}

4
aten/src/ATen/functorch/BatchedTensorImpl.h
View File

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

10
aten/src/ATen/functorch/Interpreter.h
View File

@ -3,6 +3,7 @@
#include <ATen/functorch/Macros.h>
#include <ATen/core/dispatch/Dispatcher.h>
#include <c10/core/impl/LocalDispatchKeySet.h>
#include <c10/util/Exception.h>
#include <optional>
#include <bitset>
#include <utility>
@ -106,9 +107,10 @@ struct VmapInterpreterMeta {
template <typename T>
friend void to_json(T& json_j, const VmapInterpreterMeta& json_t) {
if (json_t.batchSize_.is_heap_allocated()) {
throw std::runtime_error("Serialization for heap-allocated SymInt is not implemented yet");
}
TORCH_CHECK(
!json_t.batchSize_.is_heap_allocated(),
"Serialization for heap-allocated SymInt is not implemented yet"
);
json_j["batchSize"] = json_t.batchSize_.as_int_unchecked();
json_j["randomness"] = static_cast<int64_t>(json_t.randomness_);
}
@ -302,7 +304,7 @@ struct Interpreter {
} else if (meta.contains("Functionalize")) {
json_t.meta_.emplace<FunctionalizeInterpreterMeta>(meta["Functionalize"].template get<FunctionalizeInterpreterMeta>());
} else {
throw std::runtime_error("unknown interpreter metadata type");
TORCH_CHECK(false, "unknown interpreter metadata type");
}
}

12
aten/src/ATen/functorch/LegacyBatchingRegistrations.cpp
View File

@ -68,18 +68,18 @@ namespace at::functorch {
namespace{
// PyTorch allows operations to specify dim 0 and dim -1 on a scalar tensor.
static bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
bool is_allowed_dim_on_scalar_tensor(int64_t dim) {
return dim == 0 || dim == -1;
}
static int64_t get_current_level() {
int64_t get_current_level() {
auto maybe_level = maybeCurrentDynamicLayer();
TORCH_INTERNAL_ASSERT(maybe_level.has_value());
return maybe_level->layerId();
}
// This check should probably go into the dispatcher...
static bool participatesInCurrentLevel(const Tensor& self) {
bool participatesInCurrentLevel(const Tensor& self) {
auto current_level = get_current_level();
auto* maybe_batched_impl = maybeGetBatchedImpl(self);
if (!maybe_batched_impl) {
@ -90,7 +90,7 @@ static bool participatesInCurrentLevel(const Tensor& self) {
return self_level == current_level;
}
static bool participatesInCurrentLevel(ITensorListRef self) {
bool participatesInCurrentLevel(ITensorListRef self) {
for (const Tensor& tensor : self) {
if (participatesInCurrentLevel(tensor)) {
return true;
@ -285,7 +285,7 @@ std::vector<Tensor> unbind_batching_rule(const Tensor& self, int64_t dim) {
// given (sizes, strides, storage_offset) returns the maximum location that
// can be indexed (or nullopt if such a location doesn't exist, e.g., tensors
// with zero-size dims).
static std::optional<c10::SymInt> maximum_indexable_location(
std::optional<c10::SymInt> maximum_indexable_location(
c10::SymIntArrayRef sizes, c10::SymIntArrayRef strides, const c10::SymInt& storage_offset) {
auto result = native::storage_size_for(sizes, strides);
if (result == 0) {
@ -298,7 +298,7 @@ static std::optional<c10::SymInt> maximum_indexable_location(
// This checks that the range of possible memory locations accessible by
// x.as_strided(sizes, strides, maybe_storage_offset)
// are within the bounds of possible memory locations accessible by x.
static void checkBasicAsStridedValidForSlice(
void checkBasicAsStridedValidForSlice(
const Tensor& physical_tensor,
int64_t num_batch_dims,
c10::SymIntArrayRef sizes,

13
aten/src/ATen/functorch/PyTorchOperatorHacks.cpp
View File

@ -6,6 +6,7 @@
#include <ATen/functorch/BatchedTensorImpl.h>
#include <ATen/Dispatch.h>
#include <c10/util/irange.h>
#include <c10/util/Exception.h>
#include <ATen/NamedTensorUtils.h>
#include <ATen/native/LinearAlgebraUtils.h>
#include <ATen/native/xnnpack/Engine.h>
@ -70,7 +71,7 @@ Tensor linear_hack(const Tensor& input, const Tensor& weight, const std::optiona
return output;
}
static inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
inline at::Tensor apply_loss_reduction(const at::Tensor& unreduced, int64_t reduction) {
if (reduction == at::Reduction::Mean) {
return unreduced.mean();
} else if (reduction == at::Reduction::Sum) {
@ -108,9 +109,7 @@ Tensor binary_cross_entropy_with_logits_hack(
}
Tensor trace_backward_decomp(const Tensor& grad, IntArrayRef sizes) {
if (sizes.size() != 2) {
throw std::runtime_error("expected matrix input");
}
TORCH_CHECK(sizes.size() == 2, "expected matrix input");
auto grad_input = at::zeros(sizes[0] * sizes[1], grad.options());
auto indices = at::arange(0, grad_input.numel(), sizes[1] + 1, grad.options().dtype(at::kLong));
// Workaround using index_put instead of yet unsupported index_fill_
@ -128,7 +127,7 @@ namespace {
template<bool inplace>
using Ctype = std::conditional_t<inplace, Tensor&, Tensor>;
static Tensor make_feature_noise(const Tensor& input) {
Tensor make_feature_noise(const Tensor& input) {
auto input_sizes = input.sizes();
TORCH_CHECK(input.dim() >= 2, "Feature dropout requires at least 2 dimensions in the input");
std::vector<int64_t> sizes;
@ -142,7 +141,7 @@ static Tensor make_feature_noise(const Tensor& input) {
return at::empty(sizes, input.options());
}
static bool is_fused_kernel_acceptable(const Tensor& input, double p) {
bool is_fused_kernel_acceptable(const Tensor& input, double p) {
return (input.is_cuda() || input.is_xpu() || input.is_lazy() || input.is_privateuseone()) && p > 0 && p < 1 && input.numel() > 0;
}
@ -211,7 +210,7 @@ ALIAS_SPECIALIZATION(_feature_dropout, true, false)
ALIAS_SPECIALIZATION(_alpha_dropout, false, true )
ALIAS_SPECIALIZATION(_feature_alpha_dropout, true, true )
static Tensor dropout(const Tensor& input, double p, bool train) {
Tensor dropout(const Tensor& input, double p, bool train) {
auto result = [&]() {
NoNamesGuard guard;
if (train && is_fused_kernel_acceptable(input, p)) {

4
aten/src/ATen/native/Activation.cpp
View File

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

4
aten/src/ATen/native/AdaptiveAveragePooling3d.cpp
View File

@ -24,7 +24,7 @@ namespace at::native {
namespace {
template <typename scalar_t>
static void adaptive_avg_pool3d_out_frame(
void adaptive_avg_pool3d_out_frame(
const scalar_t* input_p,
scalar_t* output_p,
int64_t sizeD,
@ -176,7 +176,7 @@ void adaptive_avg_pool3d_out_cpu_template(
}
template <typename scalar_t>
static void adaptive_avg_pool3d_backward_out_frame(
void adaptive_avg_pool3d_backward_out_frame(
scalar_t* gradInput_p,
const scalar_t* gradOutput_p,
int64_t sizeD,

8
aten/src/ATen/native/AdaptiveMaxPooling3d.cpp
View File

@ -93,7 +93,7 @@ namespace {
// 5d tensor B x D x T x H x W
template <typename scalar_t>
static void adaptive_max_pool3d_single_out_frame(
void adaptive_max_pool3d_single_out_frame(
const scalar_t *input_p,
scalar_t *output_p,
int64_t *ind_p,
@ -170,7 +170,7 @@ static void adaptive_max_pool3d_single_out_frame(
}
template <typename scalar_t>
static void adaptive_max_pool3d_out_frame(
void adaptive_max_pool3d_out_frame(
const scalar_t *input_data,
scalar_t *output_data,
int64_t *indices_data,
@ -202,7 +202,7 @@ static void adaptive_max_pool3d_out_frame(
}
template <typename scalar_t>
static void adaptive_max_pool3d_backward_single_out_frame(
void adaptive_max_pool3d_backward_single_out_frame(
scalar_t *gradInput_p,
const scalar_t *gradOutput_p,
const int64_t *ind_p,
@ -241,7 +241,7 @@ static void adaptive_max_pool3d_backward_single_out_frame(
}
template <typename scalar_t>
static void adaptive_max_pool3d_backward_out_frame(
void adaptive_max_pool3d_backward_out_frame(
scalar_t *gradInput_data,
const scalar_t *gradOutput_data,
const int64_t *indices_data,

4
aten/src/ATen/native/AveragePool3d.cpp
View File

@ -153,7 +153,7 @@ namespace at::native {
namespace {
template <typename scalar_t>
static void avg_pool3d_out_frame(
void avg_pool3d_out_frame(
const scalar_t *input_p,
scalar_t *output_p,
int64_t nslices,
@ -333,7 +333,7 @@ TORCH_IMPL_FUNC(avg_pool3d_out_cpu) (
namespace {
template <typename scalar_t>
static void avg_pool3d_backward_out_frame(
void avg_pool3d_backward_out_frame(
scalar_t *gradInput_p,
const scalar_t *gradOutput_p,
int64_t nslices,

8
aten/src/ATen/native/BatchLinearAlgebraKernel.cpp
View File

@ -143,13 +143,13 @@ Tensor& cholesky_inverse_kernel_impl(Tensor& result, Tensor& infos, bool upper)
For more info see https://github.com/pytorch/pytorch/issues/145801#issuecomment-2631781776
*/
template <typename T>
static inline
inline
std::enable_if_t<std::is_floating_point_v<T>, int> lapack_work_to_int(const T val) {
const auto next_after = std::nextafter(val, std::numeric_limits<T>::infinity());
return std::max<int>(1, std::ceil(next_after));
}
template <typename T>
static inline
inline
std::enable_if_t<c10::is_complex<T>::value, int> lapack_work_to_int(const T val) {
return lapack_work_to_int(val.real());
}
@ -343,7 +343,7 @@ void linalg_eigh_kernel(const Tensor& eigenvalues, const Tensor& eigenvectors, c
For further details, please see the LAPACK documentation for GEQRF.
*/
template <typename scalar_t>
static void apply_geqrf(const Tensor& input, const Tensor& tau) {
void apply_geqrf(const Tensor& input, const Tensor& tau) {
#if !AT_BUILD_WITH_LAPACK()
TORCH_CHECK(
false,
@ -1039,7 +1039,7 @@ void lu_solve_kernel(const Tensor& LU, const Tensor& pivots, const Tensor& B, Tr
}
template <typename scalar_t>
static void apply_svd(const Tensor& A,
void apply_svd(const Tensor& A,
const bool full_matrices,
const bool compute_uv,
const Tensor& U,

4
aten/src/ATen/native/BlasKernel.cpp
View File

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

2
aten/src/ATen/native/Col2Im.cpp
View File

@ -71,7 +71,7 @@
namespace at::native {
namespace {
static void col2im_out_cpu_template(
void col2im_out_cpu_template(
Tensor& output,
const Tensor& input_,
IntArrayRef output_size,

1
aten/src/ATen/native/Convolution.cpp
View File

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

10
aten/src/ATen/native/ConvolutionMM2d.cpp
View File

@ -25,7 +25,7 @@ namespace at::native {
namespace {
static Tensor compute_columns2d(
Tensor compute_columns2d(
const Tensor& input,
IntArrayRef padding,
IntArrayRef stride,
@ -93,7 +93,7 @@ static Tensor compute_columns2d(
return columns.contiguous();
}
static inline void slow_conv2d_shape_check(
inline void slow_conv2d_shape_check(
const Tensor& input,
const Tensor& grad_output,
const Tensor& weight,
@ -205,7 +205,7 @@ static inline void slow_conv2d_shape_check(
}
}
static inline Tensor view_weight_2d(const Tensor& weight_,
inline Tensor view_weight_2d(const Tensor& weight_,
at::MemoryFormat memory_format = at::MemoryFormat::Contiguous) {
Tensor weight = weight_.contiguous(memory_format);
if (weight.dim() == 4) {
@ -220,7 +220,7 @@ static inline Tensor view_weight_2d(const Tensor& weight_,
}
template <typename scalar_t>
static void slow_conv2d_update_output_frame(
void slow_conv2d_update_output_frame(
TensorAccessor<const scalar_t, 3> input,
TensorAccessor<scalar_t, 3> output,
TensorAccessor<const scalar_t, 2> weight,
@ -480,7 +480,7 @@ void slow_conv2d_backward_weight_frame(
}
}
static void slow_conv2d_backward_weight_out_cpu_template(
void slow_conv2d_backward_weight_out_cpu_template(
Tensor& grad_weight,
const Tensor& input,
const Tensor& grad_output_,

Some files were not shown because too many files have changed in this diff Show More