Compare commits

..

13 Commits

Author SHA1 Message Date
9c701f03ee update json 2025-09-08 22:16:33 -07:00
c193ed6c84 Merge remote-tracking branch 'origin/main' into add_compile_benchmarking 2025-09-08 12:42:06 -07:00
eab7bd0d4c Remove mm_bwd_test.py 2025-09-08 11:34:58 -07:00
199318f978 Remove cpu benchmarking 2025-09-08 11:28:30 -07:00
9b226b2ce4 Add cpu memory calculation 2025-09-08 00:10:10 -07:00
6357d4e05a Add cpu memory calculation 2025-09-08 00:03:50 -07:00
162e7d3c20 Updates 2025-09-07 22:02:53 -07:00
ada9c165dd Lint fixes 2025-09-04 13:12:34 -07:00
461c7ad698 Enable bwd pass 2025-09-03 21:51:42 -07:00
819159610d Add fixes 2025-09-03 20:47:09 -07:00
d257ebf9c7 Add peak memory calculation 2025-09-03 11:10:16 -07:00
aab478833d Make jit and compile mutually exclusive 2025-08-27 14:21:37 -07:00
ba1319f414 Update the op benchmarking, to benchmark using torch.compile 2025-08-25 00:15:50 -07:00
871 changed files with 4367 additions and 15791 deletions

View File

@ -3,13 +3,12 @@ set -eux -o pipefail
GPU_ARCH_VERSION=${GPU_ARCH_VERSION:-}
# Set CUDA architecture lists to match x86 build_cuda.sh
if [[ "$GPU_ARCH_VERSION" == *"12.6"* ]]; then
export TORCH_CUDA_ARCH_LIST="8.0;9.0"
elif [[ "$GPU_ARCH_VERSION" == *"12.8"* ]]; then
if [[ "$GPU_ARCH_VERSION" == *"12.9"* ]]; then
export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;12.0"
elif [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0+PTX"
fi
if [[ "$GPU_ARCH_VERSION" == *"13.0"* ]]; then
export TORCH_CUDA_ARCH_LIST="8.0;9.0;10.0;11.0;12.0"
fi
# Compress the fatbin with -compress-mode=size for CUDA 13
@ -28,7 +27,7 @@ cd /
# on the mounted pytorch repo
git config --global --add safe.directory /pytorch
pip install -r /pytorch/requirements.txt
pip install auditwheel==6.2.0 wheel
pip install auditwheel==6.2.0
if [ "$DESIRED_CUDA" = "cpu" ]; then
echo "BASE_CUDA_VERSION is not set. Building cpu wheel."
#USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
@ -36,16 +35,6 @@ if [ "$DESIRED_CUDA" = "cpu" ]; then
else
echo "BASE_CUDA_VERSION is set to: $DESIRED_CUDA"
export USE_SYSTEM_NCCL=1
# Check if we should use NVIDIA libs from PyPI (similar to x86 build_cuda.sh logic)
if [[ -z "$PYTORCH_EXTRA_INSTALL_REQUIREMENTS" ]]; then
echo "Bundling CUDA libraries with wheel for aarch64."
else
echo "Using nvidia libs from pypi for aarch64."
echo "Updated PYTORCH_EXTRA_INSTALL_REQUIREMENTS for aarch64: $PYTORCH_EXTRA_INSTALL_REQUIREMENTS"
export USE_NVIDIA_PYPI_LIBS=1
fi
#USE_PRIORITIZED_TEXT_FOR_LD for enable linker script optimization https://github.com/pytorch/pytorch/pull/121975/files
USE_PRIORITIZED_TEXT_FOR_LD=1 python /pytorch/.ci/aarch64_linux/aarch64_wheel_ci_build.py --enable-mkldnn --enable-cuda
fi

View File

@ -69,186 +69,83 @@ def replace_tag(filename) -> None:
f.writelines(lines)
def patch_library_rpath(
folder: str,
lib_name: str,
use_nvidia_pypi_libs: bool = False,
desired_cuda: str = "",
) -> None:
"""Apply patchelf to set RPATH for a library in torch/lib"""
lib_path = f"{folder}/tmp/torch/lib/{lib_name}"
if use_nvidia_pypi_libs:
# For PyPI NVIDIA libraries, construct CUDA RPATH
cuda_rpaths = [
"$ORIGIN/../../nvidia/cudnn/lib",
"$ORIGIN/../../nvidia/nvshmem/lib",
"$ORIGIN/../../nvidia/nccl/lib",
"$ORIGIN/../../nvidia/cusparselt/lib",
]
if "130" in desired_cuda:
cuda_rpaths.append("$ORIGIN/../../nvidia/cu13/lib")
else:
cuda_rpaths.extend(
[
"$ORIGIN/../../nvidia/cublas/lib",
"$ORIGIN/../../nvidia/cuda_cupti/lib",
"$ORIGIN/../../nvidia/cuda_nvrtc/lib",
"$ORIGIN/../../nvidia/cuda_runtime/lib",
"$ORIGIN/../../nvidia/cufft/lib",
"$ORIGIN/../../nvidia/curand/lib",
"$ORIGIN/../../nvidia/cusolver/lib",
"$ORIGIN/../../nvidia/cusparse/lib",
"$ORIGIN/../../nvidia/nvtx/lib",
"$ORIGIN/../../nvidia/cufile/lib",
]
)
# Add $ORIGIN for local torch libs
rpath = ":".join(cuda_rpaths) + ":$ORIGIN"
else:
# For bundled libraries, just use $ORIGIN
rpath = "$ORIGIN"
if os.path.exists(lib_path):
os.system(
f"cd {folder}/tmp/torch/lib/; "
f"patchelf --set-rpath '{rpath}' --force-rpath {lib_name}"
)
def copy_and_patch_library(
src_path: str,
folder: str,
use_nvidia_pypi_libs: bool = False,
desired_cuda: str = "",
) -> None:
"""Copy a library to torch/lib and patch its RPATH"""
if os.path.exists(src_path):
lib_name = os.path.basename(src_path)
shutil.copy2(src_path, f"{folder}/tmp/torch/lib/{lib_name}")
patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
def package_cuda_wheel(wheel_path, desired_cuda) -> None:
"""
Package the cuda wheel libraries
"""
folder = os.path.dirname(wheel_path)
wheelname = os.path.basename(wheel_path)
os.mkdir(f"{folder}/tmp")
os.system(f"unzip {wheel_path} -d {folder}/tmp")
# Delete original wheel since it will be repackaged
os.system(f"rm {wheel_path}")
# Common libraries for all CUDA versions
common_libs = [
# Non-NVIDIA system libraries
"/lib64/libgomp.so.1",
"/usr/lib64/libgfortran.so.5",
"/acl/build/libarm_compute.so",
"/acl/build/libarm_compute_graph.so",
# Common CUDA libraries (same for all versions)
"/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
"/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
"/usr/local/lib/libnvpl_lapack_core.so.0",
"/usr/local/lib/libnvpl_blas_core.so.0",
"/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
"/usr/local/cuda/lib64/libcudnn.so.9",
"/usr/local/cuda/lib64/libcusparseLt.so.0",
"/usr/local/cuda/lib64/libcurand.so.10",
"/usr/local/cuda/lib64/libnccl.so.2",
"/usr/local/cuda/lib64/libnvshmem_host.so.3",
"/usr/local/cuda/lib64/libcudnn_adv.so.9",
"/usr/local/cuda/lib64/libcudnn_cnn.so.9",
"/usr/local/cuda/lib64/libcudnn_graph.so.9",
"/usr/local/cuda/lib64/libcudnn_ops.so.9",
"/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9",
"/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9",
"/usr/local/cuda/lib64/libcudnn_heuristic.so.9",
"/usr/local/cuda/lib64/libcufile.so.0",
"/usr/local/cuda/lib64/libcufile_rdma.so.1",
"/usr/local/cuda/lib64/libcusparse.so.12",
]
# Check if we should use PyPI NVIDIA libraries or bundle system libraries
use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
if use_nvidia_pypi_libs:
print("Using nvidia libs from pypi - skipping CUDA library bundling")
# For PyPI approach, we don't bundle CUDA libraries - they come from PyPI packages
# We only need to bundle non-NVIDIA libraries
minimal_libs_to_copy = [
"/lib64/libgomp.so.1",
"/usr/lib64/libgfortran.so.5",
"/acl/build/libarm_compute.so",
"/acl/build/libarm_compute_graph.so",
"/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
"/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
"/usr/local/lib/libnvpl_lapack_core.so.0",
"/usr/local/lib/libnvpl_blas_core.so.0",
# CUDA version-specific libraries
if "130" in desired_cuda:
version_specific_libs = [
"/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
"/usr/local/cuda/lib64/libcublas.so.13",
"/usr/local/cuda/lib64/libcublasLt.so.13",
"/usr/local/cuda/lib64/libcudart.so.13",
"/usr/local/cuda/lib64/libcufft.so.12",
"/usr/local/cuda/lib64/libcusolver.so.12",
"/usr/local/cuda/lib64/libnvJitLink.so.13",
"/usr/local/cuda/lib64/libnvrtc.so.13",
"/usr/local/cuda/lib64/libnvrtc-builtins.so.13.0",
]
elif "12" in desired_cuda:
# Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
minor_version = desired_cuda[-1]
version_specific_libs = [
"/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
"/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/libcufft.so.11",
"/usr/local/cuda/lib64/libcusolver.so.11",
"/usr/local/cuda/lib64/libnvJitLink.so.12",
"/usr/local/cuda/lib64/libnvrtc.so.12",
f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
]
# Copy minimal libraries to unzipped_folder/torch/lib
for lib_path in minimal_libs_to_copy:
copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
# Combine all libraries
libs_to_copy = common_libs + version_specific_libs
# Patch torch libraries used for searching libraries
torch_libs_to_patch = [
"libtorch.so",
"libtorch_cpu.so",
"libtorch_cuda.so",
"libtorch_cuda_linalg.so",
"libtorch_global_deps.so",
"libtorch_python.so",
"libtorch_nvshmem.so",
"libc10.so",
"libc10_cuda.so",
"libcaffe2_nvrtc.so",
"libshm.so",
]
for lib_name in torch_libs_to_patch:
patch_library_rpath(folder, lib_name, use_nvidia_pypi_libs, desired_cuda)
else:
print("Bundling CUDA libraries with wheel")
# Original logic for bundling system CUDA libraries
# Common libraries for all CUDA versions
common_libs = [
# Non-NVIDIA system libraries
"/lib64/libgomp.so.1",
"/usr/lib64/libgfortran.so.5",
"/acl/build/libarm_compute.so",
"/acl/build/libarm_compute_graph.so",
# Common CUDA libraries (same for all versions)
"/usr/local/lib/libnvpl_lapack_lp64_gomp.so.0",
"/usr/local/lib/libnvpl_blas_lp64_gomp.so.0",
"/usr/local/lib/libnvpl_lapack_core.so.0",
"/usr/local/lib/libnvpl_blas_core.so.0",
"/usr/local/cuda/extras/CUPTI/lib64/libnvperf_host.so",
"/usr/local/cuda/lib64/libcudnn.so.9",
"/usr/local/cuda/lib64/libcusparseLt.so.0",
"/usr/local/cuda/lib64/libcurand.so.10",
"/usr/local/cuda/lib64/libnccl.so.2",
"/usr/local/cuda/lib64/libnvshmem_host.so.3",
"/usr/local/cuda/lib64/libcudnn_adv.so.9",
"/usr/local/cuda/lib64/libcudnn_cnn.so.9",
"/usr/local/cuda/lib64/libcudnn_graph.so.9",
"/usr/local/cuda/lib64/libcudnn_ops.so.9",
"/usr/local/cuda/lib64/libcudnn_engines_runtime_compiled.so.9",
"/usr/local/cuda/lib64/libcudnn_engines_precompiled.so.9",
"/usr/local/cuda/lib64/libcudnn_heuristic.so.9",
"/usr/local/cuda/lib64/libcufile.so.0",
"/usr/local/cuda/lib64/libcufile_rdma.so.1",
"/usr/local/cuda/lib64/libcusparse.so.12",
]
# CUDA version-specific libraries
if "13" in desired_cuda:
minor_version = desired_cuda[-1]
version_specific_libs = [
"/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.13",
"/usr/local/cuda/lib64/libcublas.so.13",
"/usr/local/cuda/lib64/libcublasLt.so.13",
"/usr/local/cuda/lib64/libcudart.so.13",
"/usr/local/cuda/lib64/libcufft.so.12",
"/usr/local/cuda/lib64/libcusolver.so.12",
"/usr/local/cuda/lib64/libnvJitLink.so.13",
"/usr/local/cuda/lib64/libnvrtc.so.13",
f"/usr/local/cuda/lib64/libnvrtc-builtins.so.13.{minor_version}",
]
elif "12" in desired_cuda:
# Get the last character for libnvrtc-builtins version (e.g., "129" -> "9")
minor_version = desired_cuda[-1]
version_specific_libs = [
"/usr/local/cuda/extras/CUPTI/lib64/libcupti.so.12",
"/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/libcufft.so.11",
"/usr/local/cuda/lib64/libcusolver.so.11",
"/usr/local/cuda/lib64/libnvJitLink.so.12",
"/usr/local/cuda/lib64/libnvrtc.so.12",
f"/usr/local/cuda/lib64/libnvrtc-builtins.so.12.{minor_version}",
]
else:
raise ValueError(f"Unsupported CUDA version: {desired_cuda}.")
# Combine all libraries
libs_to_copy = common_libs + version_specific_libs
# Copy libraries to unzipped_folder/torch/lib
for lib_path in libs_to_copy:
copy_and_patch_library(lib_path, folder, use_nvidia_pypi_libs, desired_cuda)
# Copy libraries to unzipped_folder/a/lib
for lib_path in libs_to_copy:
lib_name = os.path.basename(lib_path)
shutil.copy2(lib_path, f"{folder}/tmp/torch/lib/{lib_name}")
os.system(
f"cd {folder}/tmp/torch/lib/; "
f"patchelf --set-rpath '$ORIGIN' --force-rpath {folder}/tmp/torch/lib/{lib_name}"
)
# Make sure the wheel is tagged with manylinux_2_28
for f in os.scandir(f"{folder}/tmp/"):
@ -256,8 +153,14 @@ def package_cuda_wheel(wheel_path, desired_cuda) -> None:
replace_tag(f"{f.path}/WHEEL")
break
os.system(f"wheel pack {folder}/tmp/ -d {folder}")
os.system(f"rm -rf {folder}/tmp/")
os.mkdir(f"{folder}/cuda_wheel")
os.system(f"cd {folder}/tmp/; zip -r {folder}/cuda_wheel/{wheelname} *")
shutil.move(
f"{folder}/cuda_wheel/{wheelname}",
f"{folder}/{wheelname}",
copy_function=shutil.copy2,
)
os.system(f"rm -rf {folder}/tmp/ {folder}/cuda_wheel/")
def complete_wheel(folder: str) -> str:
@ -280,7 +183,14 @@ def complete_wheel(folder: str) -> str:
f"/{folder}/dist/{repaired_wheel_name}",
)
else:
repaired_wheel_name = list_dir(f"/{folder}/dist")[0]
repaired_wheel_name = wheel_name.replace(
"linux_aarch64", "manylinux_2_28_aarch64"
)
print(f"Renaming {wheel_name} wheel to {repaired_wheel_name}")
os.rename(
f"/{folder}/dist/{wheel_name}",
f"/{folder}/dist/{repaired_wheel_name}",
)
print(f"Copying {repaired_wheel_name} to artifacts")
shutil.copy2(
@ -322,16 +232,6 @@ if __name__ == "__main__":
if enable_cuda:
build_vars += "MAX_JOBS=5 "
# Handle PyPI NVIDIA libraries vs bundled libraries
use_nvidia_pypi_libs = os.getenv("USE_NVIDIA_PYPI_LIBS", "0") == "1"
if use_nvidia_pypi_libs:
print("Configuring build for PyPI NVIDIA libraries")
# Configure for dynamic linking (matching x86 logic)
build_vars += "ATEN_STATIC_CUDA=0 USE_CUDA_STATIC_LINK=0 USE_CUPTI_SO=1 "
else:
print("Configuring build for bundled NVIDIA libraries")
# Keep existing static linking approach - already configured above
override_package_version = os.getenv("OVERRIDE_PACKAGE_VERSION")
desired_cuda = os.getenv("DESIRED_CUDA")
if override_package_version is not None:

View File

@ -214,7 +214,8 @@ case "$tag" in
TRITON=yes
;;
pytorch-linux-jammy-py3-gcc11-inductor-benchmarks)
ANACONDA_PYTHON_VERSION=3.10
# TODO (huydhn): Upgrade this to Python >= 3.10
ANACONDA_PYTHON_VERSION=3.9
GCC_VERSION=11
VISION=yes
KATEX=yes

View File

@ -56,13 +56,9 @@ ENV INSTALLED_VISION ${VISION}
# Install rocm
ARG ROCM_VERSION
RUN mkdir ci_commit_pins
COPY ./common/common_utils.sh common_utils.sh
COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
COPY ./common/install_rocm.sh install_rocm.sh
RUN bash ./install_rocm.sh
RUN rm install_rocm.sh common_utils.sh
RUN rm -r ci_commit_pins
RUN rm install_rocm.sh
COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
RUN rm install_rocm_magma.sh

View File

@ -1 +0,0 @@
7fe50dc3da2069d6645d9deb8c017a876472a977

View File

@ -1 +1 @@
5ae38bdb0dc066c5823e34dc9797afb9de42c866
fccfc522864cf8bc172abe0cd58ae5581e2d44b9

View File

@ -2,11 +2,6 @@
set -ex
# for pip_install function
source "$(dirname "${BASH_SOURCE[0]}")/common_utils.sh"
ROCM_COMPOSABLE_KERNEL_VERSION="$(cat $(dirname $0)/../ci_commit_pins/rocm-composable-kernel.txt)"
ver() {
printf "%3d%03d%03d%03d" $(echo "$1" | tr '.' ' ');
}
@ -118,8 +113,6 @@ EOF
rm -rf HIP clr
fi
pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
# Cleanup
apt-get autoclean && apt-get clean
rm -rf /var/lib/apt/lists/* /tmp/* /var/tmp/*
@ -183,8 +176,6 @@ install_centos() {
sqlite3 $kdb "PRAGMA journal_mode=off; PRAGMA VACUUM;"
done
pip_install "git+https://github.com/rocm/composable_kernel@$ROCM_COMPOSABLE_KERNEL_VERSION"
# Cleanup
yum clean all
rm -rf /var/cache/yum

View File

@ -52,13 +52,9 @@ ENV INSTALLED_VISION ${VISION}
# Install rocm
ARG ROCM_VERSION
RUN mkdir ci_commit_pins
COPY ./common/common_utils.sh common_utils.sh
COPY ./ci_commit_pins/rocm-composable-kernel.txt ci_commit_pins/rocm-composable-kernel.txt
COPY ./common/install_rocm.sh install_rocm.sh
RUN bash ./install_rocm.sh
RUN rm install_rocm.sh common_utils.sh
RUN rm -r ci_commit_pins
RUN rm install_rocm.sh
COPY ./common/install_rocm_magma.sh install_rocm_magma.sh
RUN bash ./install_rocm_magma.sh ${ROCM_VERSION}
RUN rm install_rocm_magma.sh

View File

@ -96,24 +96,14 @@ def sample_vllm_test_library():
"num_gpus": 4,
"steps": [
"pytest -v -s -x lora/test_chatglm3_tp.py",
"echo $VLLM_WORKER_MULTIPROC_METHOD",
"pytest -v -s -x lora/test_llama_tp.py",
"pytest -v -s -x lora/test_llm_with_multi_loras.py",
"pytest -v -s -x lora/test_multi_loras_with_tp.py",
],
},
"vllm_distributed_test_28_failure_test": {
"title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
"id": "vllm_distributed_test_28_failure_test",
"env_vars": {
"VLLM_WORKER_MULTIPROC_METHOD": "spawn",
},
"num_gpus": 4,
"steps": [
"pytest -v -s distributed/test_sequence_parallel.py",
],
},
"vllm_lora_28_failure_test": {
"title": "LoRA pytorch 2.8 failure test",
"id": "vllm_lora_28_failure_test",
"vllm_lora_280_failure_test": {
"title": "LoRA 280 failure test",
"id": "vllm_lora_280_failure_test",
"steps": ["pytest -v lora/test_quant_model.py"],
},
"vllm_multi_model_processor_test": {
@ -124,15 +114,6 @@ def sample_vllm_test_library():
"pytest -v -s models/multimodal/processing --ignore models/multimodal/processing/test_tensor_schema.py",
],
},
"vllm_multi_model_test_28_failure_test": {
"title": "Multi-Model Test (Failed 2.8 release)",
"id": "vllm_multi_model_test_28_failure_test",
"package_install": ["git+https://github.com/TIGER-AI-Lab/Mantis.git"],
"steps": [
"pytest -v -s models/multimodal/generation/test_voxtral.py",
"pytest -v -s models/multimodal/pooling",
],
},
"vllm_pytorch_compilation_unit_tests": {
"title": "PyTorch Compilation Unit Tests",
"id": "vllm_pytorch_compilation_unit_tests",
@ -147,28 +128,6 @@ def sample_vllm_test_library():
"pytest -v -s compile/test_decorator.py",
],
},
"vllm_languagde_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": [
"--no-build-isolation",
"git+https://github.com/Dao-AILab/causal-conv1d@v1.5.0.post8",
],
"steps": [
"pytest -v -s models/language/generation/test_mistral.py",
],
},
"vllm_distributed_test_2_gpu_28_failure_test": {
"title": "Distributed Tests (2 GPUs) pytorch 2.8 release failure",
"id": "vllm_distributed_test_2_gpu_28_failure_test",
"env_vars": {
"VLLM_WORKER_MULTIPROC_METHOD": "spawn",
},
"num_gpus": 4,
"steps": [
"pytest -v -s distributed/test_sequence_parallel.py",
],
},
# TODO(elainewy):need to add g6 with 4 gpus to run this test
"vllm_lora_test": {
"title": "LoRA Test %N",

View File

@ -66,11 +66,6 @@ class VllmBuildParameters:
"DOCKERFILE_PATH", ".github/ci_configs/vllm/Dockerfile.tmp_vllm"
)
# the cleaning script to remove torch dependencies from pip
cleaning_script: Path = env_path_field(
"cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
)
# OUTPUT_DIR: where docker buildx (local exporter) will write artifacts
output_dir: Path = env_path_field("OUTPUT_DIR", "external/vllm")
@ -165,7 +160,6 @@ class VllmBuildRunner(BaseRunner):
logger.info("Running vllm build with inputs: %s", inputs)
vllm_commit = clone_vllm()
self.cp_torch_cleaning_script(inputs)
self.cp_dockerfile_if_exist(inputs)
# cp torch wheels from root direct to vllm workspace if exist
self.cp_torch_whls_if_exist(inputs)
@ -211,11 +205,6 @@ class VllmBuildRunner(BaseRunner):
copy(inputs.torch_whls_path, tmp_dir)
return tmp_dir
def cp_torch_cleaning_script(self, inputs: VllmBuildParameters):
script = get_path(inputs.cleaning_script, resolve=True)
vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
copy(script, vllm_script)
def cp_dockerfile_if_exist(self, inputs: VllmBuildParameters):
if not inputs.use_local_dockerfile:
logger.info("using vllm default dockerfile.torch_nightly for build")

View File

@ -11,7 +11,7 @@ from typing import Any
from cli.lib.common.cli_helper import BaseRunner
from cli.lib.common.envs_helper import env_path_field, env_str_field, get_env
from cli.lib.common.path_helper import copy, get_path, remove_dir
from cli.lib.common.path_helper import copy, remove_dir
from cli.lib.common.pip_helper import (
pip_install_first_match,
pip_install_packages,
@ -43,10 +43,6 @@ class VllmTestParameters:
torch_cuda_arch_list: str = env_str_field("TORCH_CUDA_ARCH_LIST", "8.9")
cleaning_script: Path = env_path_field(
"cleaning_script", ".github/ci_configs/vllm/use_existing_torch.py"
)
def __post_init__(self):
if not self.torch_whls_path.exists():
raise ValueError("missing torch_whls_path")
@ -96,13 +92,11 @@ class VllmTestRunner(BaseRunner):
self._set_envs(params)
clone_vllm(dst=self.work_directory)
self.cp_torch_cleaning_script(params)
with working_directory(self.work_directory):
remove_dir(Path("vllm"))
self._install_wheels(params)
self._install_dependencies()
# verify the torches are not overridden by test dependencies
check_versions()
def run(self):
@ -110,31 +104,20 @@ class VllmTestRunner(BaseRunner):
main function to run vllm test
"""
self.prepare()
try:
with working_directory(self.work_directory):
if self.test_type == TestInpuType.TEST_PLAN:
if self.num_shards > 1:
run_test_plan(
self.test_plan,
"vllm",
sample_vllm_test_library(),
self.shard_id,
self.num_shards,
)
else:
run_test_plan(
self.test_plan, "vllm", sample_vllm_test_library()
)
with working_directory(self.work_directory):
if self.test_type == TestInpuType.TEST_PLAN:
if self.num_shards > 1:
run_test_plan(
self.test_plan,
"vllm",
sample_vllm_test_library(),
self.shard_id,
self.num_shards,
)
else:
raise ValueError(f"Unknown test type {self.test_type}")
finally:
# double check the torches are not overridden by other packages
check_versions()
def cp_torch_cleaning_script(self, params: VllmTestParameters):
script = get_path(params.cleaning_script, resolve=True)
vllm_script = Path(f"./{self.work_directory}/use_existing_torch.py")
copy(script, vllm_script)
run_test_plan(self.test_plan, "vllm", sample_vllm_test_library())
else:
raise ValueError(f"Unknown test type {self.test_type}")
def _install_wheels(self, params: VllmTestParameters):
logger.info("Running vllm test with inputs: %s", params)

View File

@ -258,19 +258,11 @@ function install_torchrec_and_fbgemm() {
git clone --recursive https://github.com/pytorch/fbgemm
pushd fbgemm/fbgemm_gpu
git checkout "${fbgemm_commit}" --recurse-submodules
# until the fbgemm_commit includes the tbb patch
patch <<'EOF'
--- a/FbgemmGpu.cmake
+++ b/FbgemmGpu.cmake
@@ -184,5 +184,6 @@ gpu_cpp_library(
fbgemm_gpu_tbe_cache
fbgemm_gpu_tbe_optimizers
fbgemm_gpu_tbe_utils
+ tbb
DESTINATION
fbgemm_gpu)
EOF
python setup.py bdist_wheel --build-variant=rocm
python setup.py bdist_wheel \
--build-variant=rocm \
-DHIP_ROOT_DIR="${ROCM_PATH}" \
-DCMAKE_C_FLAGS="-DTORCH_USE_HIP_DSA" \
-DCMAKE_CXX_FLAGS="-DTORCH_USE_HIP_DSA"
popd
# Save the wheel before cleaning up

View File

@ -386,8 +386,8 @@ def smoke_test_compile(device: str = "cpu") -> None:
def smoke_test_nvshmem() -> None:
if not torch.cuda.is_available() or target_os == "windows":
print("Windows platform or CUDA is not available, skipping NVSHMEM test")
if not torch.cuda.is_available():
print("CUDA is not available, skipping NVSHMEM test")
return
# Check if NVSHMEM is compiled in current build
@ -396,9 +396,7 @@ def smoke_test_nvshmem() -> None:
except ImportError:
# Not built with NVSHMEM support.
# torch is not compiled with NVSHMEM prior to 2.9
from torch.torch_version import TorchVersion
if TorchVersion(torch.__version__) < (2, 9):
if torch.__version__ < "2.9":
return
else:
# After 2.9: NVSHMEM is expected to be compiled in current build

View File

@ -1721,6 +1721,11 @@ elif [[ "${TEST_CONFIG}" == *inductor_cpp_wrapper* ]]; then
elif [[ "${TEST_CONFIG}" == *inductor* ]]; then
install_torchvision
test_inductor_shard "${SHARD_NUMBER}"
if [[ "${SHARD_NUMBER}" == 1 ]]; then
if [[ "${BUILD_ENVIRONMENT}" != linux-jammy-py3.9-gcc11-build ]]; then
test_inductor_distributed
fi
fi
elif [[ "${TEST_CONFIG}" == *einops* ]]; then
test_einops
elif [[ "${TEST_CONFIG}" == *dynamo_wrapped* ]]; then

View File

@ -1,9 +1,9 @@
set WIN_DRIVER_VN=580.88
set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe" & REM @lint-ignore
curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe
set WIN_DRIVER_VN=528.89
set "DRIVER_DOWNLOAD_LINK=https://ossci-windows.s3.amazonaws.com/%WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe" & REM @lint-ignore
curl --retry 3 -kL %DRIVER_DOWNLOAD_LINK% --output %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe
if errorlevel 1 exit /b 1
start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe -s -noreboot
start /wait %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe -s -noreboot
if errorlevel 1 exit /b 1
del %WIN_DRIVER_VN%-data-center-tesla-desktop-win10-win11-64bit-dch-international.exe || ver > NUL
del %WIN_DRIVER_VN%-data-center-tesla-desktop-winserver-2016-2019-2022-dch-international.exe || ver > NUL

View File

@ -85,7 +85,7 @@ mkdir -p "$PYTORCH_FINAL_PACKAGE_DIR" || true
# Create an isolated directory to store this builds pytorch checkout and conda
# installation
if [[ -z "$MAC_PACKAGE_WORK_DIR" ]]; then
MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_${DESIRED_PYTHON}_$(date +%H%M%S)"
MAC_PACKAGE_WORK_DIR="$(pwd)/tmp_wheel_conda_${DESIRED_PYTHON}_$(date +%H%M%S)"
fi
mkdir -p "$MAC_PACKAGE_WORK_DIR" || true
if [[ -n ${GITHUB_ACTIONS} ]]; then
@ -96,11 +96,11 @@ fi
whl_tmp_dir="${MAC_PACKAGE_WORK_DIR}/dist"
mkdir -p "$whl_tmp_dir"
mac_version='macosx-11_0-arm64'
mac_version='macosx_11_0_arm64'
libtorch_arch='arm64'
# Create a consistent wheel package name to rename the wheel to
wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version//[-,]/_}.whl"
wheel_filename_new="${TORCH_PACKAGE_NAME}-${build_version}${build_number_prefix}-cp${python_nodot}-none-${mac_version}.whl"
###########################################################
@ -125,6 +125,7 @@ popd
export TH_BINARY_BUILD=1
export INSTALL_TEST=0 # dont install test binaries into site-packages
export MACOSX_DEPLOYMENT_TARGET=11.0
export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
@ -132,19 +133,25 @@ RENAME_WHEEL=true
case $desired_python in
3.14t)
echo "Using 3.14 deps"
mac_version='macosx-11.0-arm64'
NUMPY_PINNED_VERSION="==2.1.0"
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
RENAME_WHEEL=false
;;
3.14)
echo "Using 3.14t deps"
mac_version='macosx-11.0-arm64'
NUMPY_PINNED_VERSION="==2.1.0"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
RENAME_WHEEL=false
;;
3.13t)
echo "Using 3.13 deps"
NUMPY_PINNED_VERSION="==2.1.0"
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
RENAME_WHEEL=false
;;
3.13)
@ -169,12 +176,17 @@ case $desired_python in
;;
esac
# Install into a fresh env
tmp_env_name="wheel_py$python_nodot"
conda create ${EXTRA_CONDA_INSTALL_FLAGS} -yn "$tmp_env_name" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS}
source activate "$tmp_env_name"
PINNED_PACKAGES=(
"numpy${NUMPY_PINNED_VERSION}"
)
python -mvenv ~/${desired_python}-build
source ~/${desired_python}-build/bin/activate
retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements.txt"
retry pip install "${PINNED_PACKAGES[@]}" -r "${pytorch_rootdir}/requirements-build.txt"
pip install requests ninja typing-extensions
retry pip install -r "${pytorch_rootdir}/requirements.txt" || true
retry brew install libomp
# For USE_DISTRIBUTED=1 on macOS, this enables gloo, which needs libuv, which
@ -188,7 +200,7 @@ export BUILD_TEST=OFF
pushd "$pytorch_rootdir"
echo "Calling setup.py bdist_wheel at $(date)"
_PYTHON_HOST_PLATFORM=${mac_version} ARCHFLAGS="-arch arm64" python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name "${mac_version//[-.]/_}"
python setup.py bdist_wheel -d "$whl_tmp_dir" --plat-name ${mac_version}
echo "Finished setup.py bdist_wheel at $(date)"

View File

@ -73,7 +73,7 @@ exclude =
./docs/src,
./functorch/docs,
./functorch/examples,
./functorch/docs/source/tutorials,
./functorch/notebooks,
./scripts,
./test/generated_type_hints_smoketest.py,
./third_party,

View File

@ -1 +1 @@
caba63f0fa29ef9e3d566699f32f11c07c8bda4e
3f90600fc287b276979ff2c8550a61d5d896bb8d

View File

@ -1 +1 @@
08ae0af1395c8d8471f4025deb6af9aef90b342f
7f1de94a4c2d14f59ad4ca84538c36084ea6b2c8

View File

@ -1 +1 @@
f510715882304796a96e33028b4f6de1b026c2c7
4172235ab78b09989fb56edaf734dbee283dda3e

View File

@ -1,17 +0,0 @@
import glob
requires_files = glob.glob("requirements/*.txt")
requires_files += ["pyproject.toml"]
for file in requires_files:
print(f">>> cleaning {file}")
with open(file) as f:
lines = f.readlines()
if "torch" in "".join(lines).lower():
print("removed:")
with open(file, "w") as f:
for line in lines:
if "torch" not in line.lower():
f.write(line)
print(f"<<< done cleaning {file}")
print()

View File

@ -15,7 +15,7 @@ optree==0.13.0
packaging==23.1
parameterized==0.8.1
pillow==10.3.0
protobuf==5.29.5
protobuf==5.29.4
psutil==5.9.8
pygments==2.15.0
pytest-cpp==2.3.0
@ -26,7 +26,7 @@ pytest-xdist==3.3.1
pytest==7.3.2
pyyaml==6.0.2
scipy==1.12.0
setuptools==78.1.1
setuptools==72.1.0
sympy==1.13.3
tlparse==0.4.0
tensorboard==2.13.0

View File

@ -39,9 +39,7 @@ def main() -> None:
pull_request_label_names = [label.name for label in pull_request_labels]
issue_label_names = [label.name for label in issue_labels]
labels_to_add = [
label
for label in issue_label_names
if label not in pull_request_label_names and label != "actionable"
label for label in issue_label_names if label not in pull_request_label_names
]
if not labels_to_add:
print("The pull request already has the same labels.")

View File

@ -38,60 +38,60 @@ 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 = ["13.0-aarch64"]
PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
"12.6": (
"nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | "
"nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | "
"nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | "
"nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
"nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | "
"nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | "
"nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | "
"nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
"nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
"nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
"nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
"nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
"nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
"nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'"
),
"12.8": (
"nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | "
"nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | "
"nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | "
"nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | "
"nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | "
"nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | "
"nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | "
"nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
"nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
"nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
"nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
"nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
"nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
"nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'"
),
"13.0": (
"nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | "
"nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | "
"nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | "
"nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | "
"nvidia-cublas==13.0.0.19; platform_system == 'Linux' | "
"nvidia-cufft==12.0.0.15; platform_system == 'Linux' | "
"nvidia-curand==10.4.0.35; platform_system == 'Linux' | "
"nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
"nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
"nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
"nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
"nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
"nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
"nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
"nvidia-cufile==1.15.0.42; platform_system == 'Linux'"
"nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | "
"nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'"
),
"xpu": (
"intel-cmplr-lib-rt==2025.2.1 | "

View File

@ -1,91 +0,0 @@
#!/usr/bin/env bash
set -eux
torch_version=$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
nightly=$(echo ${torch_version} | cut -d'.' -f4)
# Copied from .ci/manywheel/build_common.sh
make_wheel_record() {
fpath=$1
if echo $fpath | grep RECORD >/dev/null 2>&1; then
echo "$fpath,,"
else
fhash=$(openssl dgst -sha256 -binary $fpath | openssl base64 | sed -e 's/+/-/g' | sed -e 's/\//_/g' | sed -e 's/=//g')
fsize=$(ls -nl $fpath | awk '{print $5}')
echo "$fpath,sha256=$fhash,$fsize"
fi
}
change_wheel_version() {
local package=$1
local wheel=$2
local f_version=$3
local t_version=$4
# Extract the wheel
${PYTHON_EXECUTABLE} -mwheel unpack $wheel
mv "${package}-${f_version}" "${package}-${t_version}"
# Change the version from f_version to t_version in the dist-info dir
pushd "${package}-${t_version}"
mv "${package}-${f_version}.dist-info" "${package}-${t_version}.dist-info"
pushd "${package}-${t_version}.dist-info"
sed -i "s/${package}-${f_version}.dist-info/${package}-${t_version}.dist-info/g" RECORD
# Update the version in METADATA and its SHA256 hash
sed -i "s/Version: ${f_version}/Version: ${t_version}/g" METADATA
# then add PyTorch nightly dependency of vLLM
if [[ "${package}" == vllm ]] || [[ "${package}" == xformers ]]; then
sed -i "/License-File/a\Requires-Dist: torch==${torch_version}" METADATA
fi
sed -i '/METADATA,sha256/d' RECORD
popd
make_wheel_record "${package}-${t_version}.dist-info/METADATA" >> "${package}-${t_version}.dist-info/RECORD"
popd
# Repack the wheel
${PYTHON_EXECUTABLE} -mwheel pack "${package}-${t_version}"
# Clean up
rm -rf "${package}-${t_version}"
}
repackage_wheel() {
local package=$1
pushd $package
local orig_wheel=$(find . -name *${package//-/_}*)
local orig_version=$(unzip -p $orig_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
local version=""
if [[ "${package}" == vllm ]]; then
# Copied from vllm/.buildkite/scripts/upload-wheels.sh
version=1.0.0
else
version=$(echo $orig_version | tr '.+' '.' | cut -d'.' -f1-3)
fi
local nightly_version=$version.$nightly
# Use nightly version
change_wheel_version ${package//-/_} $orig_wheel $orig_version $nightly_version
# Clean up
rm "${orig_wheel}"
auditwheel repair --plat $PLATFORM *.whl \
--exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
local repair_wheel=$(find wheelhouse -name *${PLATFORM}*)
local repair_wheel=$(basename ${repair_wheel})
popd
cp ${package}/wheelhouse/${repair_wheel} .
rm -rf $package
}
pushd externals/vllm/wheels
for package in xformers flashinfer-python vllm; do
repackage_wheel $package
done
popd

View File

@ -22,16 +22,6 @@ name: !{{ build_environment }}
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
{%- endmacro %}
{%- macro setup_python(py_ver) -%}
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "!{{ (py_ver.strip('t') + '.4') if '3.14' not in py_ver else '3.14.0-rc.2' }}"
freethreaded: !{{ "true" if py_ver.endswith('t') else "false" }}
{%- endmacro %}
on:
# TODO: Migrate to new ciflow trigger, reference https://github.com/pytorch/pytorch/pull/70321
push:
@ -71,13 +61,23 @@ jobs:
{%- endif %}
steps:
!{{ set_runner_specific_vars() }}
!{{ setup_python(config.get("python_version", "3.10")) }}
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
!{{ common.checkout(deep_clone=False, directory="pytorch") }}
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -94,6 +94,8 @@ jobs:
{%- if config["package_type"] == "wheel" %}
- name: Test PyTorch wheel
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -104,9 +106,33 @@ jobs:
SMOKE_TEST_PARAMS=""
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
# shellcheck disable=SC2153
case $DESIRED_PYTHON in
3.14t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.14)
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.13t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
;;
*)
# shellcheck disable=SC2153
desired_python=${DESIRED_PYTHON}
;;
esac
# shellcheck disable=SC2086
python -mvenv test_venv
source test_venv/bin/activate
conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
conda activate test_conda_env
pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
# shellcheck disable=SC2086

View File

@ -47,11 +47,12 @@ jobs:
matrix:
include: [
{ name: "manylinux2_28-builder", tag: "cuda13.0", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinux2_28-builder", tag: "cuda12.9", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinux2_28-builder", tag: "cuda12.8", 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: "rocm6.3", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinux2_28-builder", tag: "rocm6.4", runner: "linux.9xlarge.ephemeral" },
{ name: "manylinux2_28-builder", tag: "cpu", runner: "linux.9xlarge.ephemeral" },

View File

@ -59,6 +59,20 @@ jobs:
run: |
set -eux
# Keep PyTorch nightly wheel here so that we can install it later during
# vLLM build process
mkdir -p "${RUNNER_TEMP}/artifacts/"
container_name=$(docker run \
--tty \
--detach \
-e PLATFORM \
-v "${GITHUB_WORKSPACE}:/pytorch" \
-v "${RUNNER_TEMP}/artifacts:/artifacts" \
-w /artifacts/ \
"${MANYLINUX_IMAGE}"
)
# Determine python executable for given version (copied from build-triton-wheel)
case $PY_VERS in
3.10)
@ -88,21 +102,6 @@ jobs:
;;
esac
# Keep PyTorch nightly wheel here so that we can install it later during
# vLLM build process
mkdir -p "${RUNNER_TEMP}/artifacts/"
container_name=$(docker run \
--tty \
--detach \
-e PLATFORM \
-e PYTHON_EXECUTABLE="${PYTHON_EXECUTABLE}" \
-v "${GITHUB_WORKSPACE}:/pytorch" \
-v "${RUNNER_TEMP}/artifacts:/artifacts" \
-w /artifacts/ \
"${MANYLINUX_IMAGE}"
)
docker exec -t "${container_name}" "${PYTHON_EXECUTABLE}" -mpip install \
--pre torch torchvision torchaudio \
--index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
@ -114,6 +113,7 @@ jobs:
--index-url "https://download.pytorch.org/whl/nightly/${BUILD_DEVICE}"
# Save this for later
echo "PYTHON_EXECUTABLE=${PYTHON_EXECUTABLE}" >> "$GITHUB_ENV"
echo "container_name=${container_name}" >> "$GITHUB_ENV"
- name: Build vLLM wheel
@ -131,7 +131,36 @@ jobs:
set -eux
# Get these wheels ready, the vllm renaming logic is copied from its .buildkite/scripts/upload-wheels.sh
docker exec -t "${container_name}" bash -c /pytorch/.github/scripts/prepare_vllm_wheels.sh
docker exec -t "${container_name}" bash -c "
set -eux
nightly=\$(unzip -p torch-* '**/METADATA' | grep '^Version: ' | cut -d' ' -f2 | cut -d'.' -f4)
pushd externals/vllm/wheels
for package in xformers flashinfer-python vllm; do
pushd \$package
auditwheel repair --plat \$PLATFORM *.whl \
--exclude libc10* --exclude libtorch* --exclude libcu* --exclude libnv*
repair_wheel=\$(find wheelhouse -name *\${PLATFORM}*)
repair_wheel=\$(basename \${repair_wheel})
popd
cp \${package}/wheelhouse/\${repair_wheel} .
version=\$(unzip -p \$repair_wheel '**/METADATA' | grep '^Version: ' | cut -d' ' -f2)
if [[ \$package == vllm ]]; then
new_wheel=\${repair_wheel/\$version/1.0.0.\$nightly}
else
major_version=\$(echo \$version | tr '.+' '.' | cut -d'.' -f1-3)
new_wheel=\${repair_wheel/\$version/\$major_version.\$nightly}
fi
mv -- \$repair_wheel \$new_wheel
rm -rf \$package
done
popd
"
docker exec -t "${container_name}" chown -R 1000:1000 /artifacts
- uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0

View File

@ -112,98 +112,6 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_10-cuda-aarch64-12_6-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.10"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_10-cuda-aarch64-12_6
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_10-cuda-aarch64-12_6-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_10-cuda-aarch64-12_6-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: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.10"
build_name: manywheel-py3_10-cuda-aarch64-12_6
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_10-cuda-aarch64-12_8-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.10"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_10-cuda-aarch64-12_8
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_10-cuda-aarch64-12_8-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_10-cuda-aarch64-12_8-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: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.10"
build_name: manywheel-py3_10-cuda-aarch64-12_8
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
@ -224,7 +132,7 @@ jobs:
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_10-cuda-aarch64-13_0
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
@ -315,98 +223,6 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_11-cuda-aarch64-12_6-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.11"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_11-cuda-aarch64-12_6
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_11-cuda-aarch64-12_6-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_11-cuda-aarch64-12_6-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: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.11"
build_name: manywheel-py3_11-cuda-aarch64-12_6
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_11-cuda-aarch64-12_8-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.11"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_11-cuda-aarch64-12_8
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_11-cuda-aarch64-12_8-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_11-cuda-aarch64-12_8-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: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.11"
build_name: manywheel-py3_11-cuda-aarch64-12_8
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
@ -427,7 +243,7 @@ jobs:
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_11-cuda-aarch64-13_0
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
@ -518,98 +334,6 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_12-cuda-aarch64-12_6-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.12"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_12-cuda-aarch64-12_6
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda-aarch64-12_6-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_12-cuda-aarch64-12_6-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: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.12"
build_name: manywheel-py3_12-cuda-aarch64-12_6
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_12-cuda-aarch64-12_8-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.12"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_12-cuda-aarch64-12_8
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda-aarch64-12_8-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_12-cuda-aarch64-12_8-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: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.12"
build_name: manywheel-py3_12-cuda-aarch64-12_8
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
@ -630,7 +354,7 @@ jobs:
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_12-cuda-aarch64-13_0
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
@ -721,98 +445,6 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13-cuda-aarch64-12_6-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.13"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_13-cuda-aarch64-12_6
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13-cuda-aarch64-12_6-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_13-cuda-aarch64-12_6-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: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.13"
build_name: manywheel-py3_13-cuda-aarch64-12_6
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13-cuda-aarch64-12_8-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.13"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_13-cuda-aarch64-12_8
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13-cuda-aarch64-12_8-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_13-cuda-aarch64-12_8-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: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.13"
build_name: manywheel-py3_13-cuda-aarch64-12_8
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
@ -833,7 +465,7 @@ jobs:
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_13-cuda-aarch64-13_0
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
@ -924,98 +556,6 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13t-cuda-aarch64-12_6-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.13t"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_13t-cuda-aarch64-12_6
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13t-cuda-aarch64-12_6-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_13t-cuda-aarch64-12_6-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: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.13t"
build_name: manywheel-py3_13t-cuda-aarch64-12_6
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_13t-cuda-aarch64-12_8-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.13t"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_13t-cuda-aarch64-12_8
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13t-cuda-aarch64-12_8-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_13t-cuda-aarch64-12_8-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: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.13t"
build_name: manywheel-py3_13t-cuda-aarch64-12_8
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
@ -1036,7 +576,7 @@ jobs:
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_13t-cuda-aarch64-13_0
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1127,98 +667,6 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14-cuda-aarch64-12_6-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.14"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_14-cuda-aarch64-12_6
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14-cuda-aarch64-12_6-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_14-cuda-aarch64-12_6-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: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.14"
build_name: manywheel-py3_14-cuda-aarch64-12_6
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14-cuda-aarch64-12_8-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.14"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_14-cuda-aarch64-12_8
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14-cuda-aarch64-12_8-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_14-cuda-aarch64-12_8-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: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.14"
build_name: manywheel-py3_14-cuda-aarch64-12_8
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
@ -1239,7 +687,7 @@ jobs:
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_14-cuda-aarch64-13_0
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
@ -1330,98 +778,6 @@ jobs:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14t-cuda-aarch64-12_6-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.14t"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_14t-cuda-aarch64-12_6
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14t-cuda-aarch64-12_6-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_14t-cuda-aarch64-12_6-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: cu126
GPU_ARCH_VERSION: "12.6-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
DESIRED_PYTHON: "3.14t"
build_name: manywheel-py3_14t-cuda-aarch64-12_6
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
uses: ./.github/workflows/_binary-upload.yml
manywheel-py3_14t-cuda-aarch64-12_8-build:
if: ${{ github.repository_owner == 'pytorch' }}
uses: ./.github/workflows/_binary-build-linux.yml
needs: get-label-type
with:
PYTORCH_ROOT: /pytorch
PACKAGE_TYPE: manywheel
# TODO: This is a legacy variable that we eventually want to get rid of in
# favor of GPU_ARCH_VERSION
DESIRED_CUDA: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.14t"
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
runs_on: linux.arm64.m7g.4xlarge.ephemeral
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_14t-cuda-aarch64-12_8
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14t-cuda-aarch64-12_8-upload: # Uploading
if: ${{ github.repository_owner == 'pytorch' }}
permissions:
id-token: write
contents: read
needs: manywheel-py3_14t-cuda-aarch64-12_8-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: cu128
GPU_ARCH_VERSION: "12.8-aarch64"
GPU_ARCH_TYPE: cuda-aarch64
DOCKER_IMAGE: manylinuxaarch64-builder
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
DESIRED_PYTHON: "3.14t"
build_name: manywheel-py3_14t-cuda-aarch64-12_8
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
@ -1442,7 +798,7 @@ jobs:
ALPINE_IMAGE: "arm64v8/alpine"
build_name: manywheel-py3_14t-cuda-aarch64-13_0
build_environment: linux-aarch64-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
timeout-minutes: 420
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}

View File

@ -60,7 +60,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_12-cuda12_8
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda12_8-test: # Testing

View File

@ -127,7 +127,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_10-cuda12_6
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_10-cuda12_6-test: # Testing
@ -193,7 +193,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_10-cuda12_8
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_10-cuda12_8-test: # Testing
@ -259,7 +259,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_10-cuda13_0
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_10-cuda13_0-test: # Testing
@ -719,7 +719,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_11-cuda12_6
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_11-cuda12_6-test: # Testing
@ -785,7 +785,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_11-cuda12_8
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_11-cuda12_8-test: # Testing
@ -851,7 +851,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_11-cuda13_0
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_11-cuda13_0-test: # Testing
@ -1311,7 +1311,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_12-cuda12_6
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda12_6-test: # Testing
@ -1377,7 +1377,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_12-cuda12_8
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda12_8-test: # Testing
@ -1443,7 +1443,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_12-cuda13_0
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_12-cuda13_0-test: # Testing
@ -1903,7 +1903,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_13-cuda12_6
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13-cuda12_6-test: # Testing
@ -1969,7 +1969,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_13-cuda12_8
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13-cuda12_8-test: # Testing
@ -2035,7 +2035,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_13-cuda13_0
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13-cuda13_0-test: # Testing
@ -2495,7 +2495,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_13t-cuda12_6
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13t-cuda12_6-test: # Testing
@ -2561,7 +2561,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_13t-cuda12_8
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13t-cuda12_8-test: # Testing
@ -2627,7 +2627,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_13t-cuda13_0
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_13t-cuda13_0-test: # Testing
@ -3087,7 +3087,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_14-cuda12_6
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14-cuda12_6-test: # Testing
@ -3153,7 +3153,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_14-cuda12_8
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14-cuda12_8-test: # Testing
@ -3219,7 +3219,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_14-cuda13_0
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14-cuda13_0-test: # Testing
@ -3679,7 +3679,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_14t-cuda12_6
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14t-cuda12_6-test: # Testing
@ -3745,7 +3745,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_14t-cuda12_8
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14t-cuda12_8-test: # Testing
@ -3811,7 +3811,7 @@ jobs:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build_name: manywheel-py3_14t-cuda13_0
build_environment: linux-binary-manywheel
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-curand==10.4.0.35; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvtx==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' and platform_machine == 'x86_64' | nvidia-cufile==1.15.0.42; platform_system == 'Linux' and platform_machine == 'x86_64'
secrets:
github-token: ${{ secrets.GITHUB_TOKEN }}
manywheel-py3_14t-cuda13_0-test: # Testing

View File

@ -60,13 +60,13 @@ jobs:
echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
# shellcheck disable=SC2129
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "3.10.4"
freethreaded: false
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
- name: Checkout PyTorch
uses: actions/checkout@v4
with:
@ -81,9 +81,13 @@ jobs:
working-directory: pytorch
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"

View File

@ -56,13 +56,13 @@ jobs:
echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
# shellcheck disable=SC2129
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "3.10.4"
freethreaded: false
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
- name: Checkout PyTorch
uses: actions/checkout@v4
with:
@ -77,9 +77,13 @@ jobs:
working-directory: pytorch
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -95,6 +99,8 @@ jobs:
"${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
- name: Test PyTorch wheel
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -105,9 +111,33 @@ jobs:
SMOKE_TEST_PARAMS=""
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
# shellcheck disable=SC2153
case $DESIRED_PYTHON in
3.14t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.14)
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.13t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
;;
*)
# shellcheck disable=SC2153
desired_python=${DESIRED_PYTHON}
;;
esac
# shellcheck disable=SC2086
python -mvenv test_venv
source test_venv/bin/activate
conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
conda activate test_conda_env
pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
# shellcheck disable=SC2086
@ -166,13 +196,13 @@ jobs:
echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
# shellcheck disable=SC2129
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "3.11.4"
freethreaded: false
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
- name: Checkout PyTorch
uses: actions/checkout@v4
with:
@ -187,9 +217,13 @@ jobs:
working-directory: pytorch
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -205,6 +239,8 @@ jobs:
"${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
- name: Test PyTorch wheel
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -215,9 +251,33 @@ jobs:
SMOKE_TEST_PARAMS=""
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
# shellcheck disable=SC2153
case $DESIRED_PYTHON in
3.14t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.14)
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.13t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
;;
*)
# shellcheck disable=SC2153
desired_python=${DESIRED_PYTHON}
;;
esac
# shellcheck disable=SC2086
python -mvenv test_venv
source test_venv/bin/activate
conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
conda activate test_conda_env
pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
# shellcheck disable=SC2086
@ -276,13 +336,13 @@ jobs:
echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
# shellcheck disable=SC2129
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "3.12.4"
freethreaded: false
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
- name: Checkout PyTorch
uses: actions/checkout@v4
with:
@ -297,9 +357,13 @@ jobs:
working-directory: pytorch
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -315,6 +379,8 @@ jobs:
"${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
- name: Test PyTorch wheel
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -325,9 +391,33 @@ jobs:
SMOKE_TEST_PARAMS=""
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
# shellcheck disable=SC2153
case $DESIRED_PYTHON in
3.14t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.14)
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.13t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
;;
*)
# shellcheck disable=SC2153
desired_python=${DESIRED_PYTHON}
;;
esac
# shellcheck disable=SC2086
python -mvenv test_venv
source test_venv/bin/activate
conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
conda activate test_conda_env
pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
# shellcheck disable=SC2086
@ -386,13 +476,13 @@ jobs:
echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
# shellcheck disable=SC2129
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "3.13.4"
freethreaded: false
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
- name: Checkout PyTorch
uses: actions/checkout@v4
with:
@ -407,9 +497,13 @@ jobs:
working-directory: pytorch
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -425,6 +519,8 @@ jobs:
"${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
- name: Test PyTorch wheel
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -435,9 +531,33 @@ jobs:
SMOKE_TEST_PARAMS=""
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
# shellcheck disable=SC2153
case $DESIRED_PYTHON in
3.14t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.14)
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.13t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
;;
*)
# shellcheck disable=SC2153
desired_python=${DESIRED_PYTHON}
;;
esac
# shellcheck disable=SC2086
python -mvenv test_venv
source test_venv/bin/activate
conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
conda activate test_conda_env
pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
# shellcheck disable=SC2086
@ -496,13 +616,13 @@ jobs:
echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
# shellcheck disable=SC2129
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "3.13.4"
freethreaded: true
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
- name: Checkout PyTorch
uses: actions/checkout@v4
with:
@ -517,9 +637,13 @@ jobs:
working-directory: pytorch
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -535,6 +659,8 @@ jobs:
"${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
- name: Test PyTorch wheel
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -545,9 +671,33 @@ jobs:
SMOKE_TEST_PARAMS=""
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
# shellcheck disable=SC2153
case $DESIRED_PYTHON in
3.14t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.14)
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.13t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
;;
*)
# shellcheck disable=SC2153
desired_python=${DESIRED_PYTHON}
;;
esac
# shellcheck disable=SC2086
python -mvenv test_venv
source test_venv/bin/activate
conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
conda activate test_conda_env
pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
# shellcheck disable=SC2086
@ -606,13 +756,13 @@ jobs:
echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
# shellcheck disable=SC2129
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "3.14.0-rc.2"
freethreaded: false
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
- name: Checkout PyTorch
uses: actions/checkout@v4
with:
@ -627,9 +777,13 @@ jobs:
working-directory: pytorch
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -645,6 +799,8 @@ jobs:
"${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
- name: Test PyTorch wheel
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -655,9 +811,33 @@ jobs:
SMOKE_TEST_PARAMS=""
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
# shellcheck disable=SC2153
case $DESIRED_PYTHON in
3.14t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.14)
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.13t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
;;
*)
# shellcheck disable=SC2153
desired_python=${DESIRED_PYTHON}
;;
esac
# shellcheck disable=SC2086
python -mvenv test_venv
source test_venv/bin/activate
conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
conda activate test_conda_env
pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
# shellcheck disable=SC2086
@ -716,13 +896,13 @@ jobs:
echo "PYTORCH_FINAL_PACKAGE_DIR=${RUNNER_TEMP}/artifacts" >> "${GITHUB_ENV}"
# shellcheck disable=SC2129
echo "MAC_PACKAGE_WORK_DIR=${RUNNER_TEMP}" >> "${GITHUB_ENV}"
- name: Setup Python
uses: actions/setup-python@v6
with:
# TODO: Removeme once 3.14 is out
# .4 version is min minor for 3.10, and also no-gil version of 3.13 needs at least 3.13.3
python-version: "3.14.0-rc.2"
freethreaded: true
- name: Install conda and dependencies
run: |
# Install conda, setup-miniconda messes with the path that messes with the ruby stuff we do later on
curl --retry 3 --retry-all-errors -o "${RUNNER_TEMP}/conda.sh" "https://repo.anaconda.com/miniconda/Miniconda3-py310_23.5.2-0-MacOSX-$(uname -m).sh"
chmod +x "${RUNNER_TEMP}/conda.sh"
/bin/bash "${RUNNER_TEMP}/conda.sh" -b -p "${RUNNER_TEMP}/anaconda"
echo "${RUNNER_TEMP}/anaconda/bin" >> "${GITHUB_PATH}"
- name: Checkout PyTorch
uses: actions/checkout@v4
with:
@ -737,9 +917,13 @@ jobs:
working-directory: pytorch
- name: Populate binary env
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
"${PYTORCH_ROOT}/.circleci/scripts/binary_populate_env.sh"
- name: Build PyTorch binary
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -755,6 +939,8 @@ jobs:
"${PYTORCH_ROOT}/.ci/wheel/build_wheel.sh"
- name: Test PyTorch wheel
run: |
# shellcheck disable=SC1091
source "${RUNNER_TEMP}/anaconda/bin/activate"
set -eux -o pipefail
# shellcheck disable=SC1090
source "${BINARY_ENV_FILE:-/Users/distiller/project/env}"
@ -765,9 +951,33 @@ jobs:
SMOKE_TEST_PARAMS=""
EXTRA_CONDA_INSTALL_FLAGS=""
CONDA_ENV_CREATE_FLAGS=""
# shellcheck disable=SC2153
case $DESIRED_PYTHON in
3.14t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.14)
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge/label/python_rc -c conda-forge"
desired_python="3.14.0rc1"
;;
3.13t)
CONDA_ENV_CREATE_FLAGS="python-freethreading"
EXTRA_CONDA_INSTALL_FLAGS="-c conda-forge"
desired_python="3.13"
;;
*)
# shellcheck disable=SC2153
desired_python=${DESIRED_PYTHON}
;;
esac
# shellcheck disable=SC2086
python -mvenv test_venv
source test_venv/bin/activate
conda create -yn "test_conda_env" python="$desired_python" ${CONDA_ENV_CREATE_FLAGS} ${EXTRA_CONDA_INSTALL_FLAGS}
conda activate test_conda_env
pip install "$PYTORCH_FINAL_PACKAGE_DIR"/*.whl numpy -v
# shellcheck disable=SC2086

View File

@ -37,7 +37,7 @@ jobs:
uses: ./.github/workflows/_linux-build.yml
needs: get-default-label-prefix
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
test-matrix: |
@ -56,7 +56,7 @@ jobs:
uses: ./.github/workflows/_linux-test.yml
needs: nightly-dynamo-benchmarks-build
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image: ${{ needs.nightly-dynamo-benchmarks-build.outputs.docker-image }}
test-matrix: ${{ needs.nightly-dynamo-benchmarks-build.outputs.test-matrix }}
timeout-minutes: 720

View File

@ -75,7 +75,7 @@ jobs:
needs: get-label-type
with:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
test-matrix: |
{ include: [
@ -101,7 +101,7 @@ jobs:
needs: inductor-build
if: github.event.schedule == '0 7 * * *'
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true
docker-image: ${{ needs.inductor-build.outputs.docker-image }}
test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
@ -118,7 +118,7 @@ jobs:
needs: inductor-build
if: github.event_name == 'workflow_dispatch'
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}
docker-image: ${{ needs.inductor-build.outputs.docker-image }}
test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}

View File

@ -80,7 +80,7 @@ jobs:
needs: get-label-type
with:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
test-matrix: |
{ include: [
@ -107,7 +107,7 @@ jobs:
needs: inductor-build
if: github.event.schedule == '0 7 * * *'
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
dashboard-tag: training-false-inference-true-default-true-dynamic-true-cppwrapper-true-aotinductor-true-freezing-true
docker-image: ${{ needs.inductor-build.outputs.docker-image }}
test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}
@ -124,7 +124,7 @@ jobs:
needs: inductor-build
if: github.event_name == 'workflow_dispatch'
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
dashboard-tag: training-${{ inputs.training }}-inference-${{ inputs.inference }}-default-${{ inputs.default }}-dynamic-${{ inputs.dynamic }}-cppwrapper-${{ inputs.cppwrapper }}-aotinductor-${{ inputs.aotinductor }}-freezing-${{ inputs.freezing }}
docker-image: ${{ needs.inductor-build.outputs.docker-image }}
test-matrix: ${{ needs.inductor-build.outputs.test-matrix }}

View File

@ -154,7 +154,7 @@ jobs:
uses: ./.github/workflows/_linux-build.yml
needs: get-default-label-prefix
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
test-matrix: |
@ -200,7 +200,7 @@ jobs:
uses: ./.github/workflows/_linux-test.yml
needs: periodic-dynamo-benchmarks-cpu-build
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image: ${{ needs.periodic-dynamo-benchmarks-cpu-build.outputs.docker-image }}
test-matrix: ${{ needs.periodic-dynamo-benchmarks-cpu-build.outputs.test-matrix }}
secrets: inherit

View File

@ -3,10 +3,18 @@ name: inductor-rocm
on:
push:
branches:
- main
#- main
- release/*
tags:
- ciflow/inductor-rocm/*
schedule:
# We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
# Also run less frequently on weekends.
- cron: 45 0,8,16 * * 1-5
- cron: 45 4 * * 0,6
- cron: 45 4,12,20 * * 1-5
- cron: 45 12 * * 0,6
- cron: 29 8 * * * # about 1:29am PDT, for mem leak check and rerun disabled tests
workflow_dispatch:
concurrency:

View File

@ -110,7 +110,7 @@ jobs:
uses: ./.github/workflows/_linux-build.yml
needs: get-label-type
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
test-matrix: |
@ -127,7 +127,7 @@ jobs:
uses: ./.github/workflows/_linux-test.yml
needs: inductor-cpu-build
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image: ${{ needs.inductor-cpu-build.outputs.docker-image }}
test-matrix: ${{ needs.inductor-cpu-build.outputs.test-matrix }}
secrets: inherit

View File

@ -79,7 +79,7 @@ jobs:
uses: ./.github/workflows/_linux-build.yml
needs: get-label-type
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
test-matrix: |
@ -101,7 +101,7 @@ jobs:
uses: ./.github/workflows/_linux-test.yml
needs: inductor-cpu-build
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image: ${{ needs.inductor-cpu-build.outputs.docker-image }}
test-matrix: ${{ needs.inductor-cpu-build.outputs.test-matrix }}
secrets: inherit

View File

@ -54,7 +54,7 @@ jobs:
- get-label-type
with:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build-environment: linux-jammy-py3.10-gcc11
build-environment: linux-jammy-py3.9-gcc11
docker-image: ${{ needs.docs-build.outputs.docker-image }}
push: ${{ github.event_name == 'schedule' || github.event_name == 'workflow_dispatch' || startsWith(github.event.ref, 'refs/tags/v') }}
run-doxygen: true

View File

@ -14,10 +14,6 @@ on:
schedule:
# Run at 07:00 UTC every Sunday
- cron: 0 7 * * 0
pull_request:
paths:
- benchmarks/operator_benchmark/**
- .github/workflows/operator_benchmark.yml
concurrency:
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
@ -33,7 +29,7 @@ jobs:
name: opbenchmark-build
uses: ./.github/workflows/_linux-build.yml
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
test-matrix: |
{ include: [
@ -46,7 +42,7 @@ jobs:
name: opbenchmark-on-demand-build
uses: ./.github/workflows/_linux-build.yml
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
test-matrix: |
{ include: [
@ -59,7 +55,7 @@ jobs:
uses: ./.github/workflows/_linux-test.yml
needs: opbenchmark-build
with:
build-environment: linux-jammy-py3.10-gcc11-build
build-environment: linux-jammy-py3.9-gcc11-build
docker-image: ${{ needs.opbenchmark-build.outputs.docker-image }}
test-matrix: ${{ needs.opbenchmark-build.outputs.test-matrix }}
secrets: inherit

View File

@ -127,8 +127,6 @@ jobs:
uses: ./.github/workflows/_linux-build.yml
needs: get-label-type
with:
# More memory is needed to build with asan
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

View File

@ -3,13 +3,19 @@ name: rocm
on:
push:
branches:
- main
# - main
- release/*
tags:
- ciflow/rocm/*
workflow_dispatch:
schedule:
- cron: 29 8 * * * # about 1:29am PDT
# We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
# Also run less frequently on weekends.
- cron: 45 0,8,16 * * 1-5
- cron: 45 4 * * 0,6
- cron: 45 4,12,20 * * 1-5
- cron: 45 12 * * 0,6
- cron: 29 8 * * * # about 1:29am PDT, for mem leak check and rerun disabled tests
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' }}

View File

@ -140,8 +140,6 @@ jobs:
uses: ./.github/workflows/_linux-build.yml
needs: get-label-type
with:
# More memory is needed to build with asan
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

View File

@ -240,7 +240,7 @@ jobs:
needs: get-label-type
with:
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
build-environment: linux-jammy-py3.10-gcc11
build-environment: linux-jammy-py3.9-gcc11
docker-image-name: ci-image:pytorch-linux-jammy-py3-gcc11-inductor-benchmarks
test-matrix: |
{ include: [
@ -255,7 +255,7 @@ jobs:
- verify-cachebench-cpu-build
- target-determination
with:
build-environment: linux-jammy-py3.10-gcc11
build-environment: linux-jammy-py3.9-gcc11
docker-image: ${{ needs.verify-cachebench-cpu-build.outputs.docker-image }}
test-matrix: ${{ needs.verify-cachebench-cpu-build.outputs.test-matrix }}
secrets: inherit

View File

@ -2,9 +2,6 @@ name: vllm-test
on:
push:
branches:
- main
- release/*
tags:
- ciflow/vllm/*
workflow_dispatch:
@ -48,18 +45,14 @@ jobs:
{ 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" },
{ config: "vllm_lora_280_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_multi_model_processor_test", shard: 1, num_shards: 1, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ 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_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" },
{ config: "vllm_lora_test", shard: 2, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_lora_test", shard: 3, num_shards: 4, runner: "linux.g6.4xlarge.experimental.nvidia.gpu" },
{ config: "vllm_lora_tp_test_distributed", shard: 1, num_shards: 1, runner: "linux.g6.12xlarge.nvidia.gpu"},
{ config: "vllm_distributed_test_28_failure_test", shard: 1, num_shards: 1, runner: "linux.g6.12xlarge.nvidia.gpu"}
{ config: "vllm_lora_tp_test_distributed", shard: 1, num_shards: 1, runner: "linux.aws.h100.4"},
]}
secrets: inherit

View File

@ -13,7 +13,7 @@ exclude_patterns = [
'**/fb/**',
'functorch/docs/**',
'functorch/examples/**',
'functorch/docs/source/tutorials/**',
'functorch/notebooks/**',
'torch/_inductor/fx_passes/serialized_patterns/**',
'torch/_inductor/autoheuristic/artifacts/**',
'scripts/**',
@ -1568,6 +1568,7 @@ include_patterns = [
exclude_patterns = [
'caffe2/**',
'functorch/docs/**',
'functorch/notebooks/**',
'torch/_inductor/fx_passes/serialized_patterns/**',
'torch/_inductor/autoheuristic/artifacts/**',
'test/dynamo/cpython/**',

View File

@ -810,7 +810,7 @@ cc_library(
name = "torch_python",
srcs = libtorch_python_core_sources
+ if_cuda(libtorch_python_cuda_sources)
+ libtorch_python_distributed_sources
+ if_cuda(libtorch_python_distributed_sources)
+ GENERATED_AUTOGRAD_PYTHON,
hdrs = glob([
"torch/csrc/generic/*.cpp",

View File

@ -234,7 +234,6 @@ cmake_dependent_option(INSTALL_TEST "Install test binaries if BUILD_TEST is on"
option(USE_CPP_CODE_COVERAGE "Compile C/C++ with code coverage flags" OFF)
option(USE_COLORIZE_OUTPUT "Colorize output during compilation" ON)
option(USE_ASAN "Use Address+Undefined Sanitizers" OFF)
option(USE_LSAN "Use Leak Sanitizer" OFF)
option(USE_TSAN "Use Thread Sanitizer" OFF)
option(USE_CUDA "Use CUDA" ON)
option(USE_XPU "Use XPU" ON)
@ -874,7 +873,7 @@ cmake_dependent_option(
"Whether to build the flash_attention kernel for scaled dot product attention.\
Will be disabled if not supported by the platform"
ON
"USE_CUDA OR USE_ROCM"
"USE_CUDA OR USE_ROCM;NOT MSVC"
OFF)
cmake_dependent_option(
@ -890,9 +889,9 @@ IF(USE_FBGEMM_GENAI AND USE_ROCM AND NOT "gfx942" IN_LIST PYTORCH_ROCM_ARCH)
set(USE_FBGEMM_GENAI off)
endif()
# Set USE_FBGEMM_GENAI to ON for CUDA build on SM100.
if(USE_CUDA AND "$ENV{TORCH_CUDA_ARCH_LIST}" MATCHES "10.0" AND CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.8)
message(STATUS "Setting USE_FBGEMM_GENAI to ON, doing CUDA build for SM100a")
# Set USE_FBGEMM_GENAI to ON for CUDA build on SM100
if(USE_CUDA AND "$ENV{TORCH_CUDA_ARCH_LIST}" MATCHES "10.0a")
message(WARNING "Setting USE_FBGEMM_GENAI to ON for CUDA build on SM100")
set(USE_FBGEMM_GENAI ON)
endif()
@ -909,7 +908,7 @@ cmake_dependent_option(
# USE_FLASH_ATTENTION -> USE_ROCM -> Dependencies.cmake -> aotriton.cmake
#
if(USE_ROCM)
if(USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION)
if(UNIX AND (USE_FLASH_ATTENTION OR USE_MEM_EFF_ATTENTION))
include(cmake/External/aotriton.cmake)
endif()
endif()

View File

@ -50,7 +50,6 @@ Following is the Release Compatibility Matrix for PyTorch releases:
| PyTorch version | Python | C++ | Stable CUDA | Experimental CUDA | Stable ROCm |
| --- | --- | --- | --- | --- | --- |
| 2.9 | >=3.10, <=(3.14, 3.14t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21) | CUDA 13.0 (CUDNN 9.13.0.50) | ROCm 6.4 |
| 2.8 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 12.6 (CUDNN 9.10.2.21), CUDA 12.8 (CUDNN 9.10.2.21) | CUDA 12.9 (CUDNN 9.10.2.21) | ROCm 6.4 |
| 2.7 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 11.8 (CUDNN 9.1.0.70), CUDA 12.6 (CUDNN 9.5.1.17) | CUDA 12.8 (CUDNN 9.7.1.26) | ROCm 6.3 |
| 2.6 | >=3.9, <=3.13, (3.13t experimental) | C++17 | CUDA 11.8, CUDA 12.4 (CUDNN 9.1.0.70) | CUDA 12.6 (CUDNN 9.5.1.17) | ROCm 6.2.4 |

View File

@ -16,8 +16,6 @@ However, if you believe you have found a security vulnerability in PyTorch, we e
Please report security issues using https://github.com/pytorch/pytorch/security/advisories/new
All reports submitted thru the security advisories mechanism would **either be made public or dismissed by the team within 90 days of the submission**. If advisory has been closed on the grounds that it is not a security issue, please do not hesitate to create an [new issue](https://github.com/pytorch/pytorch/issues/new?template=bug-report.yml) as it is still likely a valid issue within the framework.
Please refer to the following page for our responsible disclosure policy, reward guidelines, and those things that should not be reported:
https://www.facebook.com/whitehat

View File

@ -265,14 +265,6 @@ IF(USE_FBGEMM_GENAI)
"${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
list(FILTER fbgemm_genai_native_cuda_cu INCLUDE REGEX ${FBGEMM_CUTLASS_KERNELS_REGEX})
# PyTorch is not built for 10.0a in CI, due to lack of portability,
# so we need to explicitly build these files for 10.0a.
foreach(cu_file ${fbgemm_genai_native_cuda_cu})
_BUILD_FOR_ADDITIONAL_ARCHS(
"${cu_file}"
"100a")
endforeach()
file(GLOB_RECURSE fbgemm_genai_native_cuda_cpp
"${FBGEMM_GENAI_SRCS}/common/*.cpp"
)

View File

@ -133,12 +133,12 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
"resize_ called on tensor with symbolic shape")
TORCH_CHECK(
sparse_dim + dense_dim == static_cast<int64_t>(size.size()),
"'len(size) == sparse_dim + dense_dim' is not satisfied: len(size) = ",
size.size(),
", sparse_dim = ",
"number of dimensions must be sparse_dim (",
sparse_dim,
", dense_dim = ",
dense_dim);
") + dense_dim (",
dense_dim,
"), but got ",
size.size());
if (nnz() > 0) {
[[maybe_unused]] auto constexpr alt_options_msg =
"You could try the following options:\n\
@ -254,12 +254,12 @@ struct TORCH_API SparseTensorImpl : public TensorImpl {
"resize_and_clear_ called on tensor with symbolic shape")
TORCH_CHECK(
sparse_dim + dense_dim == static_cast<int64_t>(size.size()),
"'len(size) == sparse_dim + dense_dim' is not satisfied: len(size) = ",
size.size(),
", sparse_dim = ",
"number of dimensions must be sparse_dim (",
sparse_dim,
", dense_dim = ",
dense_dim);
") + dense_dim (",
dense_dim,
"), but got ",
size.size());
set_sizes_and_strides(size, std::vector<int64_t>(size.size()));
sparse_dim_ = sparse_dim;

View File

@ -64,7 +64,6 @@ constexpr DynamicTypeBits kDynamicClassTypeBit = DYNAMIC_TYPE_BIT(10);
_(ScalarType, kDynamicIntTypeBit, 1) \
_(Layout, kDynamicIntTypeBit, 1) \
_(SymInt, kDynamicIntTypeBit, 1) \
_(SymBool, kDynamicIntTypeBit, 1) \
_(MemoryFormat, kDynamicIntTypeBit, 1)
#define FORWARD_DECL_TYPE(NAME, _, __) struct NAME ## Type;

View File

@ -644,8 +644,6 @@ inline void bgemm_internal_cublas_half_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYP
void * beta_ptr = &fbeta;
#ifdef USE_ROCM
int flag = 0;
rocblas_datatype c_type = std::is_same<C_Dtype, float>::value ? rocblas_datatype_f32_r : rocblas_datatype_f16_r;
rocblas_datatype d_type = c_type;
#if USE_GEMM_FLAGS_FP16_ALT_IMPL
flag = at::ROCmBackwardPassGuard::is_backward_pass() ? rocblas_gemm_flags_fp16_alt_impl : 0;
#endif
@ -654,8 +652,8 @@ inline void bgemm_internal_cublas_half_helper(CUDABLAS_BGEMM_ARGTYPES_AND_C_DTYP
hipOperationToRocOperation(opb), (int)m, (int)n, (int)k,
(void*)alpha_ptr, a, rocblas_datatype_f16_r, (int)lda, stridea,
b, rocblas_datatype_f16_r, (int)ldb, strideb,
(void*)beta_ptr, c, c_type, (int)ldc, stridec,
c, d_type, (int)ldc, stridec,
(void*)beta_ptr, c, rocblas_datatype_f16_r, (int)ldc, stridec,
c, rocblas_datatype_f16_r, (int)ldc, stridec,
(int) num_batches, rocblas_datatype_f32_r, rocblas_gemm_algo_standard,
0, flag)));
#else
@ -1098,8 +1096,6 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(
GEMM_CHECK_ARGVALUES(at::Half);
#ifdef USE_ROCM
int flag = 0;
rocblas_datatype c_type = std::is_same<C_Dtype, float>::value ? rocblas_datatype_f32_r : rocblas_datatype_f16_r;
rocblas_datatype d_type = c_type;
#if USE_GEMM_FLAGS_FP16_ALT_IMPL
flag = at::ROCmBackwardPassGuard::is_backward_pass() ? rocblas_gemm_flags_fp16_alt_impl : 0;
#endif
@ -1119,10 +1115,10 @@ inline void gemm_internal_cublas_half_helper(CUDABLAS_GEMM_ARGTYPES_AND_C_DTYPE(
ldb,
beta_ptr,
c,
c_type,
rocblas_datatype_f16_r,
ldc,
c,
d_type,
rocblas_datatype_f16_r,
ldc,
rocblas_datatype_f32_r,
rocblas_gemm_algo_standard,

View File

@ -45,24 +45,6 @@ struct OffsetCalculator {
C10_HOST_DEVICE offset_type get(index_t linear_idx) const {
offset_type offsets;
#if defined(USE_ROCM)
if ((dims > 0) && (dims <= 2)) {
auto divmod = sizes_[0].divmod(linear_idx);
#pragma unroll
for (int arg = 0; arg < NARGS; arg++)
offsets[arg] = divmod.mod * strides_[0][arg];
if (dims >= 2) {
divmod = sizes_[1].divmod(divmod.div);
#pragma unroll
for (int arg = 0; arg < NARGS; arg++)
offsets[arg] += divmod.mod * strides_[1][arg];
}
// [...]
return offsets;
}
#endif
#pragma unroll
for (int arg = 0; arg < NARGS; arg++) {
offsets[arg] = 0;

View File

@ -457,9 +457,24 @@ void gemm(
return;
}
#endif
// for the fallback path, first compute gemm with beta = 0,
// and then add c in full precision.
int64_t c_size = n * m;
std::vector<float> float_c(c_size, 0.f);
gemm_no_downcast_stub(
at::kCPU, at::kBFloat16,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float_c.data(), m);
for (const auto j : c10::irange(n)) {
for (const auto i : c10::irange(m)) {
auto offset = j * ldc + i;
// beta == 0 won't propagate NaN from C
if (beta == 0.f) {
c[offset] = float_c[j * m + i];
} else {
c[offset] = beta * c[offset] + float_c[j * m + i];
}
}
}
}
void gemm(
@ -478,9 +493,24 @@ void gemm(
return;
}
#endif
// for the fallback path, first compute gemm with beta = 0,
// and then add c in full precision.
int64_t c_size = n * m;
std::vector<float> float_c(c_size, 0.f);
gemm_no_downcast_stub(
at::kCPU, at::kHalf,
transa, transb, m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
transa, transb, m, n, k, alpha, a, lda, b, ldb, 0.f, float_c.data(), m);
for (const auto j : c10::irange(n)) {
for (const auto i : c10::irange(m)) {
auto offset = j * ldc + i;
// beta == 0 won't propagate NaN from C
if (beta == 0.f) {
c[offset] = float_c[j * m + i];
} else {
c[offset] = beta * c[offset] + float_c[j * m + i];
}
}
}
}
void gemm(

View File

@ -14,7 +14,6 @@
#include <c10/util/accumulate.h>
#include <c10/util/irange.h>
#include <c10/macros/Macros.h>
#include <algorithm>
#include <limits>
#include <utility>
@ -301,50 +300,67 @@ struct ConvParams {
bool allow_tf32{};
bool is_strided() const {
return std::any_of(
stride.cbegin(), stride.cend(), [](const T& s) { return s != 1; });
bool is_strided = false;
for (const auto& s : stride) {
is_strided |= (s != 1);
}
return is_strided;
}
bool is_dilated() const {
return std::any_of(
dilation.cbegin(), dilation.cend(), [](const T& d) { return d != 1; });
bool is_dilated = false;
for (const auto& d : dilation) {
is_dilated |= (d != 1);
}
return is_dilated;
}
bool is_padded() const {
return std::any_of(
padding.cbegin(), padding.cend(), [](const T& p) { return p != 0; });
bool is_padded = false;
for (auto p : padding) {
is_padded |= (p != 0);
}
return is_padded;
}
bool is_output_padding_neg() const {
return std::any_of(
output_padding.cbegin(),
output_padding.cend(),
[](const T& p) { return p < 0; });
bool is_non_neg = false;
for (const auto& p : output_padding) {
is_non_neg |= (p < 0);
}
return is_non_neg;
}
bool is_output_padding_big() const {
// Revisit this with std::views::zip at C++20.
bool is_big = false;
for (auto i: c10::irange(output_padding.size())) {
if (output_padding[i] >= stride[i]) {
return true;
}
is_big |= (output_padding[i] >= stride[i]);
}
return false;
return is_big;
}
bool is_padding_neg() const {
return std::any_of(
padding.cbegin(), padding.cend(), [](const T& p) { return p < 0; });
bool is_non_neg = false;
for (const auto& p : padding) {
is_non_neg |= (p < 0);
}
return is_non_neg;
}
bool is_dilation_neg() const {
return std::any_of(
dilation.cbegin(), dilation.cend(), [](const T& d) { return d < 0; });
bool is_non_neg = false;
for (const auto& p : dilation) {
is_non_neg |= (p < 0);
}
return is_non_neg;
}
bool is_stride_nonpos() const {
return std::any_of(
stride.cbegin(), stride.cend(), [](const T& s) { return s <= 0; });
bool is_nonpos = false;
for (const auto& s : stride) {
is_nonpos |= (s <= 0);
}
return is_nonpos;
}
void view1d_as_2d() {

View File

@ -1360,8 +1360,7 @@ Tensor outer(const Tensor& self, const Tensor& vec2) {
#endif
#if !defined(__aarch64__) || AT_MKLDNN_ACL_ENABLED()
// Used by default on x86 platforms and on AArch64+ACL
#if defined(__aarch64__) && AT_MKLDNN_ACL_ENABLED()
static inline int64_t get_mkldnn_matmul_min_dim() {
static auto value = [&] {
const int64_t default_min_dim = [&] {
@ -1396,6 +1395,8 @@ static inline bool apply_mkldnn_matmul_heur(int64_t m, int64_t k, int64_t n) {
return at::globalContext().userEnabledMkldnn() && m > min_dim && k > min_dim && n > min_dim && m * k * n > min_size;
}
#endif
static void addmm_impl_cpu_(
Tensor &result, const Tensor &self, Tensor m1, Tensor m2, const Scalar& beta, const Scalar& alpha) {
TORCH_INTERNAL_ASSERT(self.dim() == 2 && m1.dim() == 2 && m2.dim() == 2);
@ -1771,8 +1772,8 @@ static inline void bmm_out_or_baddbmm_(const Tensor& self_or_result_, const Tens
return (strides[2] == 1 && (sizes[1] == 1 || strides[1] >= sizes[2])) ||
(strides[1] == 1 && (sizes[2] == 1 || strides[2] >= sizes[1]));
};
#if !defined(__aarch64__) || AT_MKLDNN_ACL_ENABLED()
// Always apply mkldnn heuristic on x86 platform, but on ARM only if compiled with ACL
#if defined(__aarch64__) && AT_MKLDNN_ACL_ENABLED()
bool apply_heur = apply_mkldnn_matmul_heur(batch1.sizes()[1], batch1.sizes()[2], batch2.sizes()[2]);
if (apply_heur && use_mkldnn_matmul(batch1, batch2, self_or_result)) {
try {
@ -1784,6 +1785,7 @@ static inline void bmm_out_or_baddbmm_(const Tensor& self_or_result_, const Tens
}
}
#endif
if (contraction_size * res_rows * res_cols < 400) {
if (is_bmm_out) {
AT_DISPATCH_ALL_TYPES_AND_COMPLEX_AND2(kBFloat16, kHalf, batch1.scalar_type(), "bmm", [&] {

View File

@ -624,9 +624,7 @@ std::tuple<Tensor, Tensor, Tensor, Tensor, int64_t> _batch_norm_impl_index(
if (backend == BatchNormBackend::Miopen) {
return std::tuple_cat(
at::miopen_batch_norm(
input.contiguous(input.suggest_memory_format()),
weight.contiguous(),
bias.contiguous(),
input.contiguous(), weight.contiguous(), bias.contiguous(),
running_mean.defined() ? running_mean.contiguous() : running_mean,
running_var.defined() ? running_var.contiguous() : running_var,
training, momentum, eps),

View File

@ -18,7 +18,6 @@
#include <ATen/ops/is_set_to_native.h>
#include <ATen/ops/size_native.h>
#include <ATen/ops/stride_native.h>
#include <ATen/ops/sym_is_contiguous_native.h>
#include <ATen/ops/sym_numel_native.h>
#include <ATen/ops/sym_size_native.h>
#include <ATen/ops/sym_storage_offset_native.h>
@ -58,12 +57,6 @@ c10::SymInt sym_size(const Tensor& self, int64_t dim) {
return self.sym_size(dim);
}
c10::SymBool sym_is_contiguous(
const Tensor& self,
c10::MemoryFormat memory_format) {
return self.sym_is_contiguous(memory_format);
}
c10::SymInt sym_stride(const Tensor& self, int64_t dim) {
return self.sym_stride(dim);
}

View File

@ -36,7 +36,7 @@ void hardsigmoid_kernel(TensorIteratorBase& iter) {
[zero, one_sixth, three, six] GPU_LAMBDA(
scalar_t self_val) -> scalar_t {
opmath_t x = static_cast<opmath_t>(self_val);
return std::min<opmath_t>(std::max<opmath_t>(x + three, zero), six) * one_sixth;
return std::min(std::max(x + three, zero), six) * one_sixth;
});
});
}

View File

@ -1080,6 +1080,16 @@ static bool _scaled_mm_allowed_device(bool sm90_only=false, bool sm100_only=fals
#endif
}
static bool _grouped_mm_allowed_device() {
#ifdef USE_ROCM
return false;
#else
auto dprops = at::cuda::getCurrentDeviceProperties();
// CUDA capability 8.0 and greater
return dprops->major >= 8;
#endif
}
#ifdef USE_ROCM
static bool _scaled_mm_is_fnuz() {
return at::detail::getCUDAHooks().isGPUArch({"gfx942"});
@ -1776,19 +1786,14 @@ Tensor _grouped_mm_cuda(const Tensor& mat_a, const Tensor& mat_b,
const std::optional<at::Tensor>& offs,
const std::optional<at::Tensor>& bias,
std::optional<c10::ScalarType> out_dtype) {
#ifndef USE_ROCM
_grouped_mm_validate_inputs(mat_a, mat_b, offs, bias, out_dtype);
bool a_b_and_out_are_bf16 = (
mat_a.dtype() == at::kBFloat16 &&
mat_b.dtype() == at::kBFloat16 &&
out_dtype.value_or(at::kBFloat16) == at::kBFloat16
);
#ifndef USE_ROCM
bool use_fast_path = _scaled_mm_allowed_device(/*sm90_only*/true, /*sm100_only*/true) && a_b_and_out_are_bf16;
#else
// _scaled_mm_allowed_device is used here within _grouped_mm_cuda which seems incorrect since scale is not used.
// the _grouped_mm_fallback should be safe for any ROCm GPU since it's just calling typical mm/bmm
bool use_fast_path = false;
#endif
const auto out_dtype_ = _resolve_grouped_mm_out_dtype(mat_a, mat_b, out_dtype);
Tensor out = create_grouped_gemm_output_tensor(mat_a, mat_b, offs, out_dtype_);
if (use_fast_path) {
@ -1798,6 +1803,9 @@ std::optional<c10::ScalarType> out_dtype) {
_grouped_mm_fallback(mat_a, mat_b, offs, bias, out_dtype, out);
}
return out;
#else
TORCH_CHECK(false, "grouped gemm is not supported on ROCM")
#endif
}
Tensor _bmm_dtype_cuda(const Tensor& batch1, const Tensor& batch2, const at::ScalarType out_dtype) {

View File

@ -482,7 +482,7 @@ auto build_graph(
auto scaled_dot_product_flash_attention_options =
fe::graph::SDPA_attributes()
.set_name("CUDNN_SDPA")
.set_generate_stats(return_softmaxstats)
.set_is_inference(return_softmaxstats == false)
.set_causal_mask(is_causal)
.set_attn_scale(attn_scale);
if (use_ragged_in_dense(q, k, v, o, attn_bias.has_value())) {
@ -702,7 +702,7 @@ auto build_graph_nestedtensor(
auto scaled_dot_product_flash_attention_options =
fe::graph::SDPA_attributes()
.set_name("CUDNN_SDPA_NESTEDTENSOR")
.set_generate_stats(return_softmaxstats)
.set_is_inference(return_softmaxstats == false)
.set_causal_mask(is_causal)
.set_attn_scale(attn_scale)
.set_seq_len_q(SEQ_LEN_Q_)

View File

@ -7,7 +7,6 @@
#include <ATen/NativeFunctions.h>
#else
#include <ATen/ops/empty.h>
#include <ATen/ops/empty_like.h>
#include <ATen/ops/miopen_batch_norm_native.h>
#include <ATen/ops/miopen_batch_norm_backward_native.h>
#endif
@ -103,7 +102,7 @@ std::tuple<Tensor, Tensor, Tensor> miopen_batch_norm(
mode = miopenBNSpatial;
}
auto output_t = at::empty_like(input_t, input_t.options(), input_t.suggest_memory_format());
auto output_t = at::empty(input->sizes(), input->options());
TensorArg output{ output_t, "output", 0 };
auto handle = getMiopenHandle();
@ -171,15 +170,20 @@ std::tuple<Tensor, Tensor, Tensor> miopen_batch_norm_backward(
const std::optional<Tensor>& save_var_t_opt,
double epsilon) {
// See [Note: hacky wrapper removal for optional tensor]
const Tensor& save_mean_t = save_mean_t_opt.value_or(Tensor());
const Tensor& save_var_t = save_var_t_opt.value_or(Tensor());
const Tensor& running_mean =
running_mean_opt.value_or(Tensor());
const Tensor& running_var =
running_var_opt.value_or(Tensor());
const Tensor& save_mean_t =
save_mean_t_opt.value_or(Tensor());
const Tensor& save_var_t =
save_var_t_opt.value_or(Tensor());
auto grad_output_contig =
grad_output_t.contiguous(input_t.suggest_memory_format());
TensorArg input{input_t, "input", 1},
grad_output{grad_output_contig, "grad_output", 2},
weight{weight_t, "weight", 3}, save_mean{save_mean_t, "save_mean", 4},
save_var{save_var_t, "save_var", 5};
TensorArg input{ input_t, "input", 1 },
grad_output{ grad_output_t, "grad_output", 2 },
weight{ weight_t, "weight", 3 },
save_mean{ save_mean_t, "save_mean", 4 },
save_var{ save_var_t, "save_var", 5 };
CheckedFrom c = "miopen_batch_norm_backward";
checkAllDefined(c, {input, grad_output, weight, save_mean, save_var});
@ -191,11 +195,7 @@ std::tuple<Tensor, Tensor, Tensor> miopen_batch_norm_backward(
}
checkAllSameType(c, {input, grad_output});
checkAllSameType(c, {weight, save_mean, save_var});
// TODO: is weight required to be contiguous?
checkAllContiguous(c, {save_mean, save_var});
// TODO: TensorArg check should start handle memory format
TORCH_CHECK(input->is_contiguous(input->suggest_memory_format()));
TORCH_CHECK(grad_output->is_contiguous(input->suggest_memory_format()));
checkAllContiguous(c, {input, grad_output, save_mean, save_var});
checkDimRange(c, input, 2, 6 /* exclusive */);
checkSameSize(c, input, grad_output);
auto num_features = input->size(1);
@ -210,7 +210,7 @@ std::tuple<Tensor, Tensor, Tensor> miopen_batch_norm_backward(
mode = miopenBNSpatial;
}
auto grad_input_t = at::empty(input->sizes(), input->options(), input->suggest_memory_format());
auto grad_input_t = at::empty(input->sizes(), input->options());
auto grad_weight_t = at::empty(weight->sizes(), weight->options());
auto grad_bias_t = at::empty(weight->sizes(), weight->options());

View File

@ -39,13 +39,6 @@ struct lerp_alpha_functor {
}
};
struct native_dropout_mask_and_scale_functor {
template <typename TI, typename TA>
inline TA operator()(const TI a, const TI b, const TA scale) {
return static_cast<TA>(a) * static_cast<TA>(b) * scale;
}
};
struct fmax_functor {
template <typename T>
inline T operator()(const T a, const T b) {
@ -434,10 +427,6 @@ REGISTER_BINARY_ALPHA_OP(lerp_alpha, uchar, uchar, uchar);
REGISTER_BINARY_ALPHA_OP(lerp_alpha, char, char, char);
REGISTER_BINARY_ALPHA_OP(lerp_alpha, bool, bool, bool);
REGISTER_BINARY_ALPHA_OP(native_dropout_mask_and_scale, float, float, float);
REGISTER_BINARY_ALPHA_OP(native_dropout_mask_and_scale, bfloat, bfloat, bfloat);
REGISTER_BINARY_ALPHA_OP(native_dropout_mask_and_scale, half, half, half);
REGISTER_BINARY_ALPHA_OP(add_alpha, bfloat, bfloat, bfloat);
REGISTER_BINARY_ALPHA_OP(sub_alpha, bfloat, bfloat, bfloat);
REGISTER_BINARY_ALPHA_OP(lerp_alpha, bfloat, bfloat, bfloat);

View File

@ -168,10 +168,6 @@ static void lerp_scalar_mps_kernel(at::TensorIteratorBase& iter, const Scalar& w
lib.exec_binary_kernel(iter, "lerp_alpha", weight);
}
static void native_dropout_mask_and_scale_mps_kernel(at::TensorIteratorBase& iter, const Scalar& scale) {
lib.exec_binary_kernel(iter, "native_dropout_mask_and_scale", scale);
}
static void mul_mps_kernel(TensorIteratorBase& iter) {
lib.exec_binary_kernel(iter, "mul");
}

View File

@ -1,45 +0,0 @@
#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
#include <ATen/TensorOperators.h>
#include <ATen/mps/MPSGeneratorImpl.h>
#include <ATen/native/Distributions.h>
#include <ATen/native/mps/OperationUtils.h>
#include <ATen/native/mps/operations/BinaryKernel.h>
#ifndef AT_PER_OPERATOR_HEADERS
#include <ATen/Functions.h>
#include <ATen/NativeFunctions.h>
#else
#include <ATen/ops/bernoulli.h>
#include <ATen/ops/empty_like.h>
#include <ATen/ops/native_dropout_backward_native.h>
#include <ATen/ops/native_dropout_native.h>
#include <ATen/ops/ones_like.h>
#endif
namespace at::native {
static Tensor native_dropout_mask_and_scale(const Tensor& input, const Tensor& mask, float scale) {
auto output = at::empty_like(input);
mps::binary_op_kernel("native_dropout_mask_and_scale", input, mask, output, scale);
return output;
}
std::tuple<Tensor, Tensor> native_dropout_mps(const Tensor& input, double p, std::optional<bool> train) {
if (input.numel() == 0 || !train.value_or(false) || p == 0) {
return {input.clone(), at::ones_like(input, input.options().dtype(c10::kBool))};
}
float p_comp = 1.0f - p;
Tensor mask = at::empty_like(input, input.options().dtype(c10::kBool));
mask.bernoulli_(p_comp);
auto scale = p_comp == 0 ? 0.0f : 1.0f / p_comp;
Tensor output = native_dropout_mask_and_scale(input, mask, scale);
return {std::move(output), std::move(mask)};
}
Tensor native_dropout_backward_mps(const Tensor& grad, const Tensor& mask, double scale) {
auto grad_float = isFloatingType(grad.scalar_type()) ? grad : grad.to(c10::kFloat);
return native_dropout_mask_and_scale(grad_float, mask, scale);
}
} // namespace at::native

View File

@ -617,7 +617,6 @@ static Tensor median_common_mps(const Tensor& input_t, bool nanmedian) {
// we allocate 1 here due to MacOS13 bug for gather MPSGraph op, look below for the error
Tensor output_t = at::empty({1}, input_t.scalar_type(), std::nullopt, kMPS, std::nullopt, std::nullopt);
if (output_t.numel() == 0 || num_in_elements == 0) {
output_t.fill_(std::numeric_limits<float>::quiet_NaN());
return output_t;
}

View File

@ -288,7 +288,6 @@
dispatch:
CPU: native_dropout_cpu
CUDA: native_dropout_cuda
MPS: native_dropout_mps
NestedTensorCPU, NestedTensorHPU, NestedTensorCUDA: native_dropout_nested
tags: [nondeterministic_seeded, core]
autogen: native_dropout.out
@ -297,7 +296,6 @@
dispatch:
CPU, NestedTensorCPU, NestedTensorHPU, NestedTensorCUDA: native_dropout_backward
CUDA: native_dropout_backward_cuda
MPS: native_dropout_backward_mps
autogen: native_dropout_backward.out
tags: pointwise
@ -1414,7 +1412,7 @@
- func: cat(Tensor[] tensors, int dim=0) -> Tensor
structured_delegate: cat.out
dispatch:
SparseCPU, SparseCUDA, SparseMPS: cat_sparse
SparseCPU, SparseCUDA: cat_sparse
QuantizedCPU: cat_quantized_cpu
NestedTensorCPU, NestedTensorHPU, NestedTensorCUDA: cat_nested
tags: core
@ -1798,7 +1796,7 @@
device_guard: False
dispatch:
MkldnnCPU: copy_mkldnn_
SparseCPU, SparseCUDA, SparseMPS: copy_sparse_wrapper_
SparseCPU, SparseCUDA: copy_sparse_wrapper_
CompositeExplicitAutograd: copy_
SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: copy_sparse_compressed_
NestedTensorCPU, NestedTensorHPU, NestedTensorCUDA: copy_nested_
@ -2160,7 +2158,7 @@
variants: function, method
structured_delegate: div.out
dispatch:
SparseCPU, SparseCUDA, SparseMPS: div_sparse
SparseCPU, SparseCUDA: div_sparse
ZeroTensor: div_zerotensor
NestedTensorCPU, NestedTensorHPU, NestedTensorCUDA: NestedTensor_div_Tensor
tags: [core, pointwise]
@ -2170,7 +2168,7 @@
variants: method
structured_delegate: div.out
dispatch:
SparseCPU, SparseCUDA, SparseMPS: div_sparse_
SparseCPU, SparseCUDA: div_sparse_
tags: pointwise
- func: div.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
@ -2179,7 +2177,7 @@
structured_inherits: TensorIteratorBase
dispatch:
CPU, CUDA, MPS, MTIA: div_out
SparseCPU, SparseCUDA, SparseMPS: div_out_sparse_zerodim
SparseCPU, SparseCUDA: div_out_sparse_zerodim
tags: pointwise
- func: div.Tensor_mode(Tensor self, Tensor other, *, str? rounding_mode) -> Tensor
@ -2187,7 +2185,7 @@
variants: function, method
structured_delegate: div.out_mode
dispatch:
SparseCPU, SparseCUDA, SparseMPS: div_sparse
SparseCPU, SparseCUDA: div_sparse
tags: [core, pointwise]
- func: div_.Tensor_mode(Tensor(a!) self, Tensor other, *, str? rounding_mode) -> Tensor(a!)
@ -2195,7 +2193,7 @@
variants: method
structured_delegate: div.out_mode
dispatch:
SparseCPU, SparseCUDA, SparseMPS: div_sparse_
SparseCPU, SparseCUDA: div_sparse_
tags: pointwise
- func: div.out_mode(Tensor self, Tensor other, *, str? rounding_mode, Tensor(a!) out) -> Tensor(a!)
@ -2204,7 +2202,7 @@
structured_inherits: TensorIteratorBase
dispatch:
CPU, CUDA, MPS: div_out_mode
SparseCPU, SparseCUDA, SparseMPS: div_out_sparse_zerodim
SparseCPU, SparseCUDA: div_out_sparse_zerodim
tags: pointwise
# For C++ only, until we have conversion from C++ numbers to Tensor
@ -2768,20 +2766,20 @@
variants: function, method
dispatch:
CPU, CUDA, MPS, MTIA: floor_divide
SparseCPU, SparseCUDA, SparseMPS: floor_divide_sparse
SparseCPU, SparseCUDA: floor_divide_sparse
- func: floor_divide_.Tensor(Tensor(a!) self, Tensor other) -> Tensor(a!)
device_check: NoCheck # TensorIterator
variants: method
dispatch:
CPU, CUDA, MPS: floor_divide_
SparseCPU, SparseCUDA, SparseMPS: floor_divide_sparse_
SparseCPU, SparseCUDA: floor_divide_sparse_
- func: floor_divide.out(Tensor self, Tensor other, *, Tensor(a!) out) -> Tensor(a!)
device_check: NoCheck # TensorIterator
dispatch:
CPU, CUDA, MPS: floor_divide_out
SparseCPU, SparseCUDA, SparseMPS: floor_divide_out_sparse_zerodim
SparseCPU, SparseCUDA: floor_divide_out_sparse_zerodim
- func: floor_divide.Scalar(Tensor self, Scalar other) -> Tensor
device_check: NoCheck # TensorIterator
@ -4273,7 +4271,7 @@
structured_delegate: mul.out
variants: function, method
dispatch:
SparseCPU, SparseCUDA, SparseMPS: mul_sparse
SparseCPU, SparseCUDA: mul_sparse
SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: mul_sparse_csr
MkldnnCPU: mkldnn_mul
ZeroTensor: mul_zerotensor
@ -4285,7 +4283,7 @@
structured_delegate: mul.out
variants: method
dispatch:
SparseCPU, SparseCUDA, SparseMPS: mul_sparse_
SparseCPU, SparseCUDA: mul_sparse_
SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: mul_sparse_csr_
MkldnnCPU: mkldnn_mul_
NestedTensorCPU, NestedTensorHPU, NestedTensorCUDA: NestedTensor_mul__Tensor
@ -4299,7 +4297,6 @@
CPU, CUDA, MPS, MTIA: mul_out
SparseCPU: mul_out_sparse_cpu
SparseCUDA: mul_out_sparse_cuda
SparseMPS: mul_out_sparse_mps
SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: mul_out_sparse_csr
MkldnnCPU: mkldnn_mul_out
tags: pointwise
@ -5514,13 +5511,6 @@
tags: core
manual_cpp_binding: True
- func: sym_is_contiguous(Tensor self, MemoryFormat memory_format=contiguous_format) -> SymBool
variants: function
device_check: NoCheck
device_guard: False
tags: core
manual_cpp_binding: True
- func: sym_numel(Tensor self) -> SymInt
variants: function
device_check: NoCheck
@ -5849,7 +5839,7 @@
variants: function, method
dispatch:
CompositeExplicitAutograd: sum
SparseCPU, SparseCUDA, SparseMPS, SparseMeta: sum_coo
SparseCPU, SparseCUDA, SparseMeta: sum_coo
SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: sum_csr
autogen: sum.out
@ -5860,7 +5850,7 @@
variants: function, method
dispatch:
NestedTensorCPU: NestedTensor_sum_dim_CPU
SparseCPU, SparseCUDA, SparseMPS: sum_sparse_coo
SparseCPU, SparseCUDA: sum_sparse_coo
SparseCsrCPU, SparseCsrCUDA, SparseCsrMeta: sum_sparse_compressed
tags: core
@ -6976,7 +6966,7 @@
CPU, CUDA: sub_out
MPS: sub_out_mps
MTIA: sub_out_mtia
SparseCPU, SparseCUDA, SparseMPS: sub_out_sparse
SparseCPU, SparseCUDA: sub_out_sparse
tags: pointwise
- func: sub.Tensor(Tensor self, Tensor other, *, Scalar alpha=1) -> Tensor
@ -6984,7 +6974,7 @@
variants: function, method
structured_delegate: sub.out
dispatch:
SparseCPU, SparseCUDA, SparseMPS: sub_sparse
SparseCPU, SparseCUDA: sub_sparse
ZeroTensor: sub_zerotensor
NestedTensorCPU, NestedTensorHPU, NestedTensorCUDA: NestedTensor_sub_Tensor
tags: [core, pointwise]
@ -6994,7 +6984,7 @@
variants: method
structured_delegate: sub.out
dispatch:
SparseCPU, SparseCUDA, SparseMPS: sub_sparse_
SparseCPU, SparseCUDA: sub_sparse_
tags: pointwise
# For C++ only, until we have conversion from C++ numbers to Tensor
@ -10343,7 +10333,7 @@
structured_inherits: TensorIteratorBase
dispatch:
CPU, CUDA: pow_Tensor_Scalar_out
SparseCPU, SparseCUDA, SparseMPS: pow_out_sparse_scalar
SparseCPU, SparseCUDA: pow_out_sparse_scalar
MPS: pow_tensor_scalar_out_mps
tags: pointwise
@ -10352,7 +10342,7 @@
structured_delegate: pow.Tensor_Scalar_out
variants: function, method
dispatch:
SparseCPU, SparseCUDA, SparseMPS: pow_sparse_scalar
SparseCPU, SparseCUDA: pow_sparse_scalar
tags: [core, pointwise]
- func: pow_.Scalar(Tensor(a!) self, Scalar exponent) -> Tensor(a!)

View File

@ -2,7 +2,6 @@
#include <ATen/core/Tensor.h>
#include <ATen/Config.h>
#include <ATen/Dispatch.h>
#include <ATen/AccumulateType.h>
#include <ATen/NamedTensorUtils.h>
#include <ATen/native/sparse/ParamUtils.h>
#include <ATen/native/SparseTensorUtils.h>
@ -296,7 +295,6 @@ void cpu_sparse_coo_softmax(Tensor output, const Tensor& input, const int64_t di
to exp functions as well as reuse of softmax implementation for
log_softmax.
*/
using accscalar_t = at::acc_type<scalar_t, false>;
auto sparse_dim = input.sparse_dim();
auto indices = input._indices().contiguous();
auto values = input._values().contiguous();
@ -342,14 +340,14 @@ void cpu_sparse_coo_softmax(Tensor output, const Tensor& input, const int64_t di
continue;
/* Prepare scratch space */
std::vector<accscalar_t> mx_row(nvalues, -std::numeric_limits<accscalar_t>::infinity());
std::vector<accscalar_t> exp_sums_row(nvalues, 0);
std::vector<scalar_t> mx_row(nvalues, -std::numeric_limits<scalar_t>::infinity());
std::vector<scalar_t> exp_sums_row(nvalues, 0);
/* Compute mx */
for (int64_t i : pool_indices) {
auto values_row = values_accessor[i];
for (const auto j : c10::irange(nvalues)) {
mx_row[j] = std::max(mx_row[j], accscalar_t(values_row[j]));
mx_row[j] = std::max(mx_row[j], values_row[j]);
}
}

View File

@ -391,13 +391,13 @@ void _validate_sparse_coo_tensor_args(
int64_t sparse_dim = indices.size(0);
int64_t dense_dim = values.dim() - 1;
TORCH_CHECK(
sparse_dim + dense_dim == static_cast<int64_t>(size.size()),
"'len(size) == sparse_dim + dense_dim' is not satisfied: len(size) = ",
size.size(),
", sparse_dim = ",
sparse_dim,
", dense_dim = ",
dense_dim);
static_cast<int64_t>(size.size()) == sparse_dim + dense_dim,
"number of dimensions must be sparse_dim (",
sparse_dim,
") + dense_dim (",
dense_dim,
"), but got ",
size.size());
if (check_pinning) {
TORCH_CHECK(

View File

@ -10,7 +10,6 @@
#include <ATen/ops/_sparse_coo_tensor_unsafe_native.h>
#include <ATen/ops/cat.h>
#include <ATen/ops/add_native.h>
#include <ATen/ops/mul_native.h>
#include <ATen/ops/empty_native.h>
#include <ATen/ops/zeros_native.h>
#include <ATen/ops/result_type.h>
@ -21,265 +20,10 @@
namespace at::native {
using namespace at::sparse;
using namespace mps;
#ifndef PYTORCH_JIT_COMPILE_SHADERS
static auto& lib = MetalShaderLibrary::getBundledLibrary();
#else
#include <ATen/native/mps/Mul_metallib.h>
#endif
Tensor& add_out_dense_sparse_mps(Tensor& out, const Tensor& dense, const SparseTensor& sparse, const Scalar& alpha);
static SparseTensor& mul_out_dense_sparse_mps(
const Tensor& dense,
const Tensor& sparse,
SparseTensor& out) {
TORCH_CHECK(sparse.is_sparse(), "mul: expected 'sparse' to be sparse COO");
TORCH_CHECK(sparse.is_mps(), "mul: expected 'sparse' to be MPS, got ", sparse.device());
TORCH_CHECK(out.is_mps(), "mul: expected 'out' to be MPS, got ", out.device());
const bool scalar_like = (dense.dim() == 0) || (dense.numel() == 1);
TORCH_CHECK(dense.is_mps() || scalar_like,
"mul: expected 'dense' to be MPS or scalar-like, got ", dense.device());
const int64_t nnz = sparse._nnz();
out.resize_as_(sparse);
auto commonDtype = at::result_type(dense, sparse);
TORCH_CHECK(canCast(commonDtype, out.scalar_type()),
"Can't convert result type ", commonDtype, " to output ", out.scalar_type());
auto indices = sparse._indices().contiguous();
auto values = sparse._values().to(commonDtype).contiguous();
if (nnz == 0) {
auto empty_vals = values.narrow(0, 0, 0);
alias_into_sparse(out,
indices.narrow(1, 0, 0),
(out.scalar_type() == commonDtype) ? empty_vals
: empty_vals.to(out.scalar_type()));
out._coalesced_(sparse.is_coalesced());
return out;
}
if (scalar_like) {
auto scalar = dense;
if (dense.numel() == 1 && dense.dim() > 0) {
scalar = dense.view({});
}
scalar = scalar.to(values.options());
auto out_vals = values.mul(scalar);
if (out.scalar_type() != commonDtype) {
out_vals = out_vals.to(out.scalar_type());
}
alias_into_sparse(out, indices, out_vals);
out._coalesced_(sparse.is_coalesced());
return out;
}
TORCH_CHECK(dense.sizes().equals(sparse.sizes()),
"mul(dense, sparse): sizes must match exactly (no broadcasting): ",
dense.sizes(), " vs ", sparse.sizes());
const int64_t ndim_i = sparse.sparse_dim();
const int64_t ndim = dense.dim();
TORCH_CHECK(
ndim_i <= ndim,
"mul(dense, sparse): sparse_dim=", ndim_i, " exceeds dense.dim()=", ndim);
// Prepare shapes
int64_t view_rows = 1, view_cols = 1;
for (int64_t i = 0; i < ndim_i; ++i) view_rows *= sparse.size(i);
for (int64_t i = ndim_i; i < ndim; ++i) view_cols *= sparse.size(i);
auto dense_mps = dense.to(commonDtype).contiguous().reshape({view_rows, view_cols});
auto out_vals = at::empty_like(values, values.options());
const uint32_t u_view_cols = static_cast<uint32_t>(view_cols);
const uint32_t u_nnz = static_cast<uint32_t>(nnz);
const uint32_t u_ndim_i = static_cast<uint32_t>(ndim_i);
auto stream = getCurrentMPSStream();
dispatch_sync_with_rethrow(stream->queue(), ^() {
@autoreleasepool {
auto pso = lib.getPipelineStateForFunc("dense_sparse_mul_kernel_" + mps::scalarToMetalTypeString(values));
auto computeEncoder = stream->commandEncoder();
[computeEncoder setComputePipelineState:pso];
const uint32_t gridWidth = u_view_cols;
const uint32_t gridDepth = u_nnz;
MTLSize gridSize = MTLSizeMake(gridWidth, 1, gridDepth);
const uint32_t maxThreadsPerGroup = pso.maxTotalThreadsPerThreadgroup;
const uint32_t tew = pso.threadExecutionWidth;
uint32_t tgWidth = std::min(gridWidth, tew);
MTLSize threadgroupSize = MTLSizeMake(tgWidth, 1, 1);
mtl_setArgs(
computeEncoder,
dense_mps,
values,
out_vals,
indices,
sparse.sizes(),
std::array<uint32_t, 3>{u_nnz, u_ndim_i, u_view_cols}
);
[computeEncoder dispatchThreads:gridSize threadsPerThreadgroup:threadgroupSize];
}
});
Tensor final_vals = out_vals;
if (out.scalar_type() != commonDtype) {
final_vals = final_vals.to(out.scalar_type());
}
alias_into_sparse(out, indices, final_vals);
out._coalesced_(sparse.is_coalesced());
return out;
}
SparseTensor& mul_out_sparse_mps(const Tensor& t_, const Tensor& src_, SparseTensor& r_) {
TORCH_CHECK(r_.is_mps(), "mul: expected 'out' to be MPS, but got ", r_.device());
// Dense x sparse fallback (keep dense first)
if (!t_.is_sparse() || !src_.is_sparse()) {
const Tensor& dense = t_.is_sparse() ? src_ : t_;
const Tensor& sparse = t_.is_sparse() ? t_ : src_;
return mul_out_dense_sparse_mps(dense, sparse, r_);
}
TORCH_CHECK(t_.is_mps(), "mul: expected 'self' to be MPS, but got ", t_.device());
TORCH_CHECK(src_.is_mps(), "mul: expected 'other' to be MPS, but got ", src_.device());
TORCH_CHECK(t_.sparse_dim() == src_.sparse_dim(),
"mul(sparse, sparse): must have same sparse_dim, got ",
t_.sparse_dim(), " vs ", src_.sparse_dim());
TORCH_CHECK(t_.sizes().equals(src_.sizes()),
"mul(sparse, sparse): sizes must match exactly (no broadcasting).");
// Coalesce and early-exit on structurally empty operands
auto lhs = t_.coalesce();
auto rhs = src_.coalesce();
const int64_t lhs_nnz = lhs._nnz();
const int64_t rhs_nnz = rhs._nnz();
if (!lhs_nnz || !rhs_nnz) {
r_.resize_as_(lhs);
return r_.zero_();
}
// dtype checks and promotion
auto commonDtype = at::result_type(lhs, rhs);
TORCH_CHECK(canCast(commonDtype, r_.scalar_type()),
"Can't convert result type ", commonDtype, " to output ", r_.scalar_type());
const int64_t ndim_i = lhs.sparse_dim();
// ndim_i == 0, at most one structural entry
if (ndim_i == 0) {
r_.resize_as_(lhs);
const bool has = (lhs_nnz && rhs_nnz);
auto out_indices = lhs._indices().narrow(1, 0, has ? 1 : 0);
Tensor lhs_vals = lhs._values().to(commonDtype);
Tensor rhs_vals = rhs._values().to(commonDtype);
lhs_vals = lhs_vals.narrow(0, 0, has ? 1 : 0);
rhs_vals = rhs_vals.narrow(0, 0, has ? 1 : 0);
Tensor out_values = lhs_vals.mul(rhs_vals);
if (r_.scalar_type() != commonDtype) {
out_values = out_values.to(r_.scalar_type());
}
alias_into_sparse(r_, out_indices, out_values);
r_._coalesced_(true);
return r_;
}
// General path, intersect keys, then gather + multiply on GPU
const auto device = r_.device();
auto stream = getCurrentMPSStream();
auto lhs_indices = lhs._indices();
auto rhs_indices = rhs._indices();
auto lhs_values = lhs._values().to(commonDtype);
auto rhs_values = rhs._values().to(commonDtype);
// Flatten sparse indices to keys
auto lhs_keys = flatten_indices(lhs_indices, lhs.sizes());
auto rhs_keys = flatten_indices(rhs_indices, rhs.sizes());
// Intersect sorted keys (search the shorter in the longer)
const bool A_is_lhs = (lhs_nnz <= rhs_nnz);
const int64_t lenA = A_is_lhs ? lhs_nnz : rhs_nnz;
const int64_t lenB = A_is_lhs ? rhs_nnz : lhs_nnz;
auto A_keys = A_is_lhs ? lhs_keys : rhs_keys;
auto B_keys = A_is_lhs ? rhs_keys : lhs_keys;
auto outA_idx = at::empty({lenA}, at::device(device).dtype(kLong));
auto outB_idx = at::empty({lenA}, at::device(device).dtype(kLong));
auto counter = at::zeros({1}, at::device(device).dtype(kInt));
dispatch_sync_with_rethrow(stream->queue(), ^() {
@autoreleasepool {
auto pso = lib.getPipelineStateForFunc("intersect_binary_search");
auto enc = stream->commandEncoder();
[enc setComputePipelineState:pso];
mtl_setArgs(enc, A_keys, B_keys, outA_idx, outB_idx, counter,
static_cast<uint32_t>(lenB), A_is_lhs);
mtl_dispatch1DJob(enc, pso, static_cast<uint32_t>(lenA));
}
});
const uint32_t M = counter.item<int32_t>(); // number of structural matches
r_.resize_as_(lhs);
auto out_indices = at::empty({ndim_i, static_cast<int64_t>(M)}, at::device(device).dtype(at::kLong));
auto lhs_match = outA_idx.narrow(0, 0, M);
auto rhs_match = outB_idx.narrow(0, 0, M);
auto out_val_sizes = lhs_values.sizes().vec();
out_val_sizes[0] = static_cast<int64_t>(M);
auto out_values = at::empty(out_val_sizes, lhs_values.options());
const uint32_t cols = static_cast<uint32_t>(
lhs_values.numel() / std::max<int64_t>(1, lhs_nnz));
dispatch_sync_with_rethrow(stream->queue(), ^() {
@autoreleasepool {
auto pso = lib.getPipelineStateForFunc(
"fused_gather_mul_kernel_" + mps::scalarToMetalTypeString(lhs_values));
auto enc = stream->commandEncoder();
[enc setComputePipelineState:pso];
const uint32_t tew = pso.threadExecutionWidth;
uint32_t tgW = std::min(cols, tew);
MTLSize grid = MTLSizeMake(cols, 1, M);
MTLSize tgs = MTLSizeMake(tgW, 1, 1);
mtl_setArgs(enc,
lhs_values, rhs_values,
lhs_match, rhs_match,
lhs_indices, out_indices,
out_values,
std::array<uint32_t, 2>{static_cast<uint32_t>(ndim_i), static_cast<uint32_t>(lhs_nnz)},
std::array<uint32_t, 2>{M, cols});
[enc dispatchThreads:grid threadsPerThreadgroup:tgs];
}
});
if (r_.scalar_type() != commonDtype) {
out_values = out_values.to(r_.scalar_type());
}
alias_into_sparse(r_, out_indices, out_values);
r_._coalesced_(true);
return r_;
}
static Tensor& add_out_dense_sparse_mps(
Tensor& add_out_dense_sparse_mps(
Tensor& out,
const Tensor& dense,
const SparseTensor& sparse,

View File

@ -1,150 +0,0 @@
#include <metal_stdlib>
#include <c10/metal/indexing.h>
using namespace metal;
template <typename T>
kernel void dense_sparse_mul_kernel(
device const T* dense [[buffer(0)]],
device const T* values [[buffer(1)]],
device T* out_values [[buffer(2)]],
device const long* indices [[buffer(3)]],
device const long* sizes [[buffer(4)]],
constant uint3& sparse_params [[buffer(5)]],
uint3 gid [[thread_position_in_grid]])
{
uint col = gid.x;
uint i = gid.z;
uint nnz = sparse_params.x;
uint ndim_i = sparse_params.y;
uint view_cols = sparse_params.z;
long key = 0;
for (uint d = 0; d < ndim_i; ++d) {
long idx_d = indices[(ulong)d * (ulong)nnz + (ulong)i];
const auto sz_d = sizes[d];
key = key * sz_d + idx_d;
}
ulong dense_idx = (ulong)key * (ulong)view_cols + (ulong)col;
ulong val_idx = (ulong)i * (ulong)view_cols + (ulong)col;
const auto a = static_cast<float>(values[val_idx]);
const auto b = static_cast<float>(dense[dense_idx]);
out_values[val_idx] = static_cast<T>(a * b);
}
kernel void intersect_binary_search(
device const long* keysA [[buffer(0)]],
device const long* keysB [[buffer(1)]],
device long* outA_idx [[buffer(2)]],
device long* outB_idx [[buffer(3)]],
device atomic_uint* counter [[buffer(4)]],
constant uint& lenB [[buffer(5)]],
constant bool& A_is_lhs [[buffer(6)]],
uint3 tid_in_grid [[thread_position_in_grid]])
{
uint gid = tid_in_grid.x;
long key = keysA[gid];
// lower_bound in B
uint lo = 0;
uint hi = lenB;
while (lo < hi) {
uint mid = (lo + hi) >> 1;
long v = keysB[mid];
if (v < key) lo = mid + 1;
else hi = mid;
}
if (lo < lenB && keysB[lo] == key) {
uint pos = atomic_fetch_add_explicit(counter, 1u, memory_order_relaxed);
if (A_is_lhs) {
outA_idx[pos] = (long)gid;
outB_idx[pos] = (long)lo;
} else {
outA_idx[pos] = (long)lo;
outB_idx[pos] = (long)gid;
}
}
}
template <typename T>
kernel void fused_gather_mul_kernel(
device const T* lhs_vals [[buffer(0)]],
device const T* rhs_vals [[buffer(1)]],
device const long* lhs_sel [[buffer(2)]],
device const long* rhs_sel [[buffer(3)]],
device const long* lhs_indices [[buffer(4)]],
device long* out_indices [[buffer(5)]],
device T* out_vals [[buffer(6)]],
constant uint2& dims_input [[buffer(7)]],
constant uint2& dims_output [[buffer(8)]],
uint3 gid [[thread_position_in_grid]])
{
const uint col = gid.x;
const uint k = gid.z;
const uint n_dim_i = dims_input.x;
const uint L = dims_input.y;
const uint M = dims_output.x;
const uint view_cols = dims_output.y;
const long iL = lhs_sel[k];
const long iR = rhs_sel[k];
if (col < view_cols) {
const ulong offL = (ulong)iL * (ulong)view_cols + (ulong)col;
const ulong offR = (ulong)iR * (ulong)view_cols + (ulong)col;
const ulong offO = (ulong)k * (ulong)view_cols + (ulong)col;
const float a = (float)lhs_vals[offL];
const float b = (float)rhs_vals[offR];
out_vals[offO] = (T)(a * b);
}
// One thread per match copies the indices column
if (col == 0) {
const ulong uL = (ulong)L;
const ulong uM = (ulong)M;
const ulong src_col = (ulong)iL; // gather from lhs
for (uint d = 0; d < n_dim_i; ++d) {
const long v = lhs_indices[(ulong)d * uL + src_col];
out_indices[(ulong)d * uM + (ulong)k] = v;
}
}
}
#define INSTANTIATE_DENSE_SPARSE_MUL(DTYPE) \
template [[host_name("dense_sparse_mul_kernel_" #DTYPE)]] kernel void \
dense_sparse_mul_kernel<DTYPE>( \
device const DTYPE* dense [[buffer(0)]], \
device const DTYPE* values [[buffer(1)]], \
device DTYPE* out_values [[buffer(2)]], \
device const long* indices [[buffer(3)]], \
device const long* sizes [[buffer(4)]], \
constant uint3& sparse_params [[buffer(5)]], \
uint3 gid [[thread_position_in_grid]]);
INSTANTIATE_DENSE_SPARSE_MUL(float);
INSTANTIATE_DENSE_SPARSE_MUL(half);
INSTANTIATE_DENSE_SPARSE_MUL(bfloat);
#define INSTANTIATE_FUSED_GATHER_MUL(DTYPE) \
template [[host_name("fused_gather_mul_kernel_" #DTYPE)]] kernel void \
fused_gather_mul_kernel<DTYPE>( \
device const DTYPE* lhs_vals [[buffer(0)]], \
device const DTYPE* rhs_vals [[buffer(1)]], \
device const long* lhs_sel [[buffer(2)]], \
device const long* rhs_sel [[buffer(3)]], \
device const long* lhs_indices [[buffer(4)]], \
device long* out_indices [[buffer(5)]], \
device DTYPE* out_vals [[buffer(6)]], \
constant uint2& dims_input [[buffer(7)]], \
constant uint2& dims_output [[buffer(8)]], \
uint3 gid [[thread_position_in_grid]]);
INSTANTIATE_FUSED_GATHER_MUL(float);
INSTANTIATE_FUSED_GATHER_MUL(half);
INSTANTIATE_FUSED_GATHER_MUL(bfloat);

View File

@ -95,72 +95,6 @@
#endif
#endif
#if defined(USE_ROCM) && (defined(USE_FLASH_ATTENTION) || defined(USE_MEM_EFF_ATTENTION))
namespace pytorch_flash
{
std::tuple<
at::Tensor,
at::Tensor,
at::Tensor,
at::Tensor,
at::Tensor,
at::Tensor,
at::Tensor,
at::Tensor>
mha_fwd(
const at::Tensor& q, // batch_size x seqlen_q x num_heads x head_size
const at::Tensor& k, // batch_size x seqlen_k x num_heads_k x head_size
const at::Tensor& v, // batch_size x seqlen_k x num_heads_k x head_size
std::optional<at::Tensor>&
out_, // batch_size x seqlen_q x num_heads x head_size
std::optional<at::Tensor>&
alibi_slopes_, // num_heads or batch_size x num_heads
const float p_dropout,
const float softmax_scale,
bool is_causal,
std::optional<int64_t> window_size_left,
std::optional<int64_t> window_size_right,
const float softcap,
const bool return_softmax,
std::optional<at::Generator> gen_) {
#if defined(USE_ROCM_CK_SDPA)
if (at::globalContext().getROCmFAPreferredBackend() ==
at::ROCmFABackend::Ck) {
const int non_null_window_left = window_size_left.value_or(-1);
const int non_null_window_right = window_size_right.value_or(-1);
std::optional<at::Tensor> dummy_attn_bias = std::nullopt;
return mha_fwd_ck(
q,
k,
v,
out_,
p_dropout,
softmax_scale,
is_causal,
non_null_window_left,
non_null_window_right,
return_softmax,
gen_,
dummy_attn_bias); // Not used in flash attention
}
#endif
return mha_fwd_aot(
q,
k,
v,
out_,
alibi_slopes_,
p_dropout,
softmax_scale,
is_causal,
window_size_left,
window_size_right,
return_softmax,
gen_);
}
}
#endif
namespace at {
namespace cuda::philox {

View File

@ -270,7 +270,7 @@ std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor> mha_varle
#endif
TORCH_API
std::tuple<
inline std::tuple<
at::Tensor,
at::Tensor,
at::Tensor,
@ -294,7 +294,42 @@ mha_fwd(
std::optional<int64_t> window_size_right,
const float softcap,
const bool return_softmax,
std::optional<at::Generator> gen_);
std::optional<at::Generator> gen_) {
#if defined(USE_ROCM_CK_SDPA)
if (at::globalContext().getROCmFAPreferredBackend() ==
at::ROCmFABackend::Ck) {
const int non_null_window_left = window_size_left.value_or(-1);
const int non_null_window_right = window_size_right.value_or(-1);
std::optional<at::Tensor> dummy_attn_bias = std::nullopt;
return mha_fwd_ck(
q,
k,
v,
out_,
p_dropout,
softmax_scale,
is_causal,
non_null_window_left,
non_null_window_right,
return_softmax,
gen_,
dummy_attn_bias); // Not used in flash attention
}
#endif
return mha_fwd_aot(
q,
k,
v,
out_,
alibi_slopes_,
p_dropout,
softmax_scale,
is_causal,
window_size_left,
window_size_right,
return_softmax,
gen_);
}
inline std::tuple<
at::Tensor,

View File

@ -98,11 +98,11 @@ dlrm,pass,0
doctr_det_predictor,pass,3
doctr_det_predictor,pass,5
doctr_reco_predictor,pass,1
doctr_reco_predictor,pass,4

1 name accuracy graph_breaks
98
99
100
101
102
103
104
105
106
107
108

View File

@ -98,11 +98,11 @@ dlrm,pass,0
doctr_det_predictor,pass,3
doctr_det_predictor,pass,5
doctr_reco_predictor,pass,1
doctr_reco_predictor,pass,4

1 name accuracy graph_breaks
98
99
100
101
102
103
104
105
106
107
108

View File

@ -98,11 +98,11 @@ dlrm,pass,0
doctr_det_predictor,pass,3
doctr_det_predictor,pass,5
doctr_reco_predictor,pass,1
doctr_reco_predictor,pass,4

1 name accuracy graph_breaks
98
99
100
101
102
103
104
105
106
107
108

View File

@ -82,11 +82,11 @@ dlrm,pass,0
doctr_det_predictor,pass,3
doctr_det_predictor,pass,5
doctr_reco_predictor,pass,1
doctr_reco_predictor,pass,4

1 name accuracy graph_breaks
82 tts_angular pass 2
83 vgg16 pass 0
84 vision_maskrcnn pass 29
85 yolov3 pass 0
86
87
88
89
90
91
92

View File

@ -98,11 +98,11 @@ dlrm,pass,0
doctr_det_predictor,pass,3
doctr_det_predictor,pass,5
doctr_reco_predictor,pass,1
doctr_reco_predictor,pass,4

1 name accuracy graph_breaks
98
99
100
101
102
103
104
105
106
107
108

View File

@ -106,11 +106,11 @@ dlrm,pass,0
doctr_det_predictor,eager_fail_to_run,3
doctr_det_predictor,eager_fail_to_run,5
doctr_reco_predictor,eager_fail_to_run,1
doctr_reco_predictor,eager_fail_to_run,4

1 name accuracy graph_breaks
106
107
108
109
110
111
112
113
114
115
116

View File

@ -106,11 +106,11 @@ dlrm,pass,0
doctr_det_predictor,eager_fail_to_run,3
doctr_det_predictor,eager_fail_to_run,5
doctr_reco_predictor,eager_fail_to_run,1
doctr_reco_predictor,eager_fail_to_run,4

1 name accuracy graph_breaks
106
107
108
109
110
111
112
113
114
115
116

View File

@ -106,11 +106,11 @@ dlrm,pass,0
doctr_det_predictor,eager_fail_to_run,3
doctr_det_predictor,eager_fail_to_run,5
doctr_reco_predictor,eager_fail_to_run,1
doctr_reco_predictor,eager_fail_to_run,4

1 name accuracy graph_breaks
106
107
108
109
110
111
112
113
114
115
116

View File

@ -106,11 +106,11 @@ dlrm,pass,0
doctr_det_predictor,eager_fail_to_run,3
doctr_det_predictor,eager_fail_to_run,5
doctr_reco_predictor,eager_fail_to_run,1
doctr_reco_predictor,eager_fail_to_run,4

1 name accuracy graph_breaks
106
107
108
109
110
111
112
113
114
115
116

View File

@ -106,11 +106,11 @@ dlrm,pass,0
doctr_det_predictor,eager_fail_to_run,3
doctr_det_predictor,eager_fail_to_run,5
doctr_reco_predictor,eager_fail_to_run,1
doctr_reco_predictor,eager_fail_to_run,4

1 name accuracy graph_breaks
106
107
108
109
110
111
112
113
114
115
116

View File

@ -219,7 +219,9 @@ skip:
- timm_regnet
- timm_nfnet
cuda: []
cuda:
# Temporary until https://github.com/pytorch/pytorch/issues/162282 is fixed
- sam_fast
test:
training:

View File

@ -373,14 +373,9 @@ class BenchmarkRunner:
curr_test_total_time = 0
time_trace = []
peak_memory = 0
input_values = test_case.op_bench.inputs.values()
device, device_module = None, None
if input_values and isinstance(next(iter(input_values)), torch.Tensor):
# The device and device module information are crucial for memory metric calculation,
# In case of ops where inputs are integers (not tensor), memory metrics need not be calculated.
sample_input = next(iter(input_values))
device = sample_input.device
device_module = torch.get_device_module(device.type)
sample_input = next(iter(test_case.op_bench.inputs.values()))
device = sample_input.device
device_module = torch.get_device_module(device.type)
# TODO: add support for cpu memory measurement
while True:
if hasattr(device_module, "reset_peak_memory_stats"):

View File

@ -1,5 +1,5 @@
Benchmarking Framework,Benchmarking Module Name,Case Name,tag,run_backward,Execution Time
PyTorch,add,add_M1_N1_K1_cpu,short,FALSE,2.459
PyTorch,add,add_M1_N1_K1_cpu,short,FALSE,3.9497
PyTorch,add,add_M64_N64_K64_cpu,short,FALSE,14.3181
PyTorch,add,add_M64_N64_K128_cpu,short,FALSE,14.6826
PyTorch,add,add_M1_N1_K1_cpu_bwdall_BACKWARD,short,TRUE,58.1449
@ -376,10 +376,10 @@ PyTorch,relu6,"relu6_dims(3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint32",sho
PyTorch,relu6,"relu6_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,9.6588
PyTorch,relu6,"relu6_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint8",short,FALSE,9.5969
PyTorch,relu6,"relu6_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint32",short,FALSE,9.547
PyTorch,relu6,"relu6_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,50.21375
PyTorch,relu6,"relu6_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,68.739
PyTorch,relu6,"relu6_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.qint8",short,FALSE,45.14133333
PyTorch,relu6,"relu6_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.qint32",short,FALSE,52.6664
PyTorch,relu6,"relu6_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,51.49525
PyTorch,relu6,"relu6_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,69.1875
PyTorch,relu6,"relu6_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.qint8",short,FALSE,48.3458
PyTorch,relu6,"relu6_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.qint32",short,FALSE,62.0719
PyTorch,functional.hardtanh,"functional.hardtanh_dims(3,4,5)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,7.5728
@ -388,10 +388,10 @@ PyTorch,functional.hardtanh,"functional.hardtanh_dims(3,4,5)_contigFalse_inplace
PyTorch,functional.hardtanh,"functional.hardtanh_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,8.1647
PyTorch,functional.hardtanh,"functional.hardtanh_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint8",short,FALSE,8.1768
PyTorch,functional.hardtanh,"functional.hardtanh_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint32",short,FALSE,8.0619
PyTorch,functional.hardtanh,"functional.hardtanh_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,48.88475
PyTorch,functional.hardtanh,"functional.hardtanh_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,67.118
PyTorch,functional.hardtanh,"functional.hardtanh_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.qint8",short,FALSE,43.702
PyTorch,functional.hardtanh,"functional.hardtanh_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.qint32",short,FALSE,50.3613
PyTorch,functional.hardtanh,"functional.hardtanh_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,50.3995
PyTorch,functional.hardtanh,"functional.hardtanh_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,67.436
PyTorch,functional.hardtanh,"functional.hardtanh_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.qint8",short,FALSE,46.9813
PyTorch,functional.hardtanh,"functional.hardtanh_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.qint32",short,FALSE,59.2295
PyTorch,functional.hardsigmoid,"functional.hardsigmoid_dims(3,4,5)_contigFalse_inplaceFalse_dtypetorch.quint8",short,FALSE,6.5189
@ -1316,4 +1316,4 @@ PyTorch,where,"where_cond_shape(8,16,1)_input_shape(1,)_other_shape(1,)_cpu_dtyp
PyTorch,where,"where_cond_shape(8,16,1)_input_shape(16,1)_other_shape(8,16,1)_cpu_dtypetorch.float32",short,FALSE,5.763
PyTorch,where,"where_cond_shape(8,16,1)_input_shape(8,1,1)_other_shape(1,)_cpu_dtypetorch.float32",short,FALSE,5.744666667
PyTorch,clamp,clamp_M512_N512_cpu,short,FALSE,15.26233333
PyTorch,gelu,gelu_M512_N512_cpu,short,FALSE,31.33166667
PyTorch,gelu,gelu_M512_N512_cpu,short,FALSE,31.33166667
1 Benchmarking Framework Benchmarking Module Name Case Name tag run_backward Execution Time
2 PyTorch add add_M1_N1_K1_cpu short FALSE 2.459 3.9497
3 PyTorch add add_M64_N64_K64_cpu short FALSE 14.3181
4 PyTorch add add_M64_N64_K128_cpu short FALSE 14.6826
5 PyTorch add add_M1_N1_K1_cpu_bwdall_BACKWARD short TRUE 58.1449
376 PyTorch relu6 relu6_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.quint8 short FALSE 9.6588
377 PyTorch relu6 relu6_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint8 short FALSE 9.5969
378 PyTorch relu6 relu6_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint32 short FALSE 9.547
379 PyTorch relu6 relu6_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.quint8 short FALSE 50.21375 68.739
380 PyTorch relu6 relu6_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.qint8 short FALSE 45.14133333
381 PyTorch relu6 relu6_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.qint32 short FALSE 52.6664
382 PyTorch relu6 relu6_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.quint8 short FALSE 51.49525 69.1875
383 PyTorch relu6 relu6_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.qint8 short FALSE 48.3458
384 PyTorch relu6 relu6_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.qint32 short FALSE 62.0719
385 PyTorch functional.hardtanh functional.hardtanh_dims(3,4,5)_contigFalse_inplaceFalse_dtypetorch.quint8 short FALSE 7.5728
388 PyTorch functional.hardtanh functional.hardtanh_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.quint8 short FALSE 8.1647
389 PyTorch functional.hardtanh functional.hardtanh_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint8 short FALSE 8.1768
390 PyTorch functional.hardtanh functional.hardtanh_dims(2,3,4,5)_contigFalse_inplaceFalse_dtypetorch.qint32 short FALSE 8.0619
391 PyTorch functional.hardtanh functional.hardtanh_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.quint8 short FALSE 48.88475 67.118
392 PyTorch functional.hardtanh functional.hardtanh_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.qint8 short FALSE 43.702
393 PyTorch functional.hardtanh functional.hardtanh_dims(512,512)_contigFalse_inplaceFalse_dtypetorch.qint32 short FALSE 50.3613
394 PyTorch functional.hardtanh functional.hardtanh_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.quint8 short FALSE 50.3995 67.436
395 PyTorch functional.hardtanh functional.hardtanh_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.qint8 short FALSE 46.9813
396 PyTorch functional.hardtanh functional.hardtanh_dims(256,1024)_contigFalse_inplaceFalse_dtypetorch.qint32 short FALSE 59.2295
397 PyTorch functional.hardsigmoid functional.hardsigmoid_dims(3,4,5)_contigFalse_inplaceFalse_dtypetorch.quint8 short FALSE 6.5189
1316 PyTorch where where_cond_shape(8,16,1)_input_shape(16,1)_other_shape(8,16,1)_cpu_dtypetorch.float32 short FALSE 5.763
1317 PyTorch where where_cond_shape(8,16,1)_input_shape(8,1,1)_other_shape(1,)_cpu_dtypetorch.float32 short FALSE 5.744666667
1318 PyTorch clamp clamp_M512_N512_cpu short FALSE 15.26233333
1319 PyTorch gelu gelu_M512_N512_cpu short FALSE 31.33166667

View File

@ -156,7 +156,7 @@ ROOT = "//" if IS_OSS else "//xplat/caffe2"
# for targets in subfolders
ROOT_PATH = "//" if IS_OSS else "//xplat/caffe2/"
C10 = "//c10:c10" if IS_OSS else ("//xplat/caffe2/c10:c10_ovrsource" if is_arvr_mode() else "//xplat/caffe2/c10:c10")
C10 = "//c10:c10" if IS_OSS else "//xplat/caffe2/c10:c10"
# a dictionary maps third party library name to fbsource and oss target
THIRD_PARTY_LIBS = {

View File

@ -638,13 +638,10 @@ libtorch_nativert_sources = [
"torch/nativert/kernels/KernelHandlerRegistry.cpp",
"torch/nativert/kernels/TritonKernel.cpp",
"torch/nativert/executor/triton/CpuTritonKernelManager.cpp",
"torch/nativert/executor/AOTInductorDelegateExecutor.cpp",
"torch/nativert/kernels/ETCallDelegateKernel.cpp",
]
libtorch_nativert_cuda_sources = [
"torch/nativert/executor/triton/CudaTritonKernelManager.cpp",
"torch/nativert/executor/AOTInductorModelContainerCudaShim.cpp",
]
torch_mobile_tracer_sources = [

View File

@ -313,15 +313,8 @@ void TensorImpl::throw_data_ptr_access_error() const {
c10::SymBool TensorImpl::sym_is_contiguous_custom(
at::MemoryFormat memory_format) const {
if (C10_UNLIKELY(matches_python_custom(SizesStridesPolicy::CustomStrides))) {
// TO reduce BC breaking and reduce having to introduce
// sym_is_contiguous. call is_contiguous when tensor does not
if (C10_UNLIKELY(has_symbolic_sizes_strides_)) {
return pyobj_slot_.load_pyobj_interpreter()->sym_is_contiguous(
this, memory_format);
} else {
return pyobj_slot_.load_pyobj_interpreter()->is_contiguous(
this, memory_format);
}
return pyobj_slot_.load_pyobj_interpreter()->is_contiguous(
this, memory_format);
}
return sym_is_contiguous_default(memory_format);

View File

@ -3269,7 +3269,7 @@ class C10_TensorImpl_Size_Check_Dummy_Class : private TensorImpl {
is_le<sizeof(autograd_meta_), 16, FieldNameEnum::autograd_meta_>();
is_le<sizeof(extra_meta_), 16, FieldNameEnum::extra_meta_>();
are_equal<sizeof(version_counter_), 8, FieldNameEnum::version_counter_>();
are_equal<sizeof(pyobj_slot_), 8, FieldNameEnum::pyobj_slot_>();
are_equal<sizeof(pyobj_slot_), 16, FieldNameEnum::pyobj_slot_>();
are_equal<sizeof(sizes_and_strides_), 88, FieldNameEnum::sizes_and_strides_>();
are_equal<sizeof(storage_offset_), 8, FieldNameEnum::storage_offset_>();
are_equal<sizeof(numel_), 8, FieldNameEnum::numel_>();

View File

@ -60,10 +60,6 @@ struct NoopPyInterpreterVTable final : public PyInterpreterVTable {
bool is_contiguous(const TensorImpl* self, at::MemoryFormat) const override {
PANIC(is_contiguous);
}
c10::SymBool sym_is_contiguous(const TensorImpl* self, at::MemoryFormat)
const override {
PANIC(sym_is_contiguous);
}
bool is_strides_like(const TensorImpl* self, at::MemoryFormat)
const override {
PANIC(is_strides_like);

View File

@ -168,9 +168,6 @@ struct C10_API PyInterpreterVTable {
virtual bool is_contiguous(const TensorImpl* self, at::MemoryFormat)
const = 0;
virtual c10::SymBool sym_is_contiguous(
const TensorImpl* self,
at::MemoryFormat) const = 0;
virtual bool is_strides_like(const TensorImpl* self, at::MemoryFormat)
const = 0;
virtual bool is_non_overlapping_and_dense(const TensorImpl* self) const = 0;

View File

@ -13,10 +13,11 @@ struct C10_API PyInterpreterHooksInterface {
// Get the PyInterpreter instance
// Stub implementation throws error when Python is not available
// We return nullptr rather than throwing an error since there are bits of c10
// that expect an empty PyObjectSlot when python is not available.
virtual PyInterpreter* getPyInterpreter() const {
return nullptr;
TORCH_CHECK(
false,
"PyTorch was compiled without Python support. "
"Cannot access Python interpreter from C++.");
}
};

View File

@ -2,7 +2,7 @@
namespace c10::impl {
PyObjectSlot::PyObjectSlot() : pyobj_(nullptr) {}
PyObjectSlot::PyObjectSlot() : pyobj_interpreter_(nullptr), pyobj_(nullptr) {}
PyObjectSlot::~PyObjectSlot() {
maybe_destroy_pyobj();
@ -10,9 +10,9 @@ PyObjectSlot::~PyObjectSlot() {
void PyObjectSlot::maybe_destroy_pyobj() {
if (owns_pyobj()) {
TORCH_INTERNAL_ASSERT(getGlobalPyInterpreter() != nullptr);
TORCH_INTERNAL_ASSERT(pyobj_interpreter_ != nullptr);
TORCH_INTERNAL_ASSERT(pyobj_ != nullptr);
(*getGlobalPyInterpreter())
(*pyobj_interpreter_.load(std::memory_order_acquire))
->decref(_unchecked_untagged_pyobj(), /*has_pyobj_slot*/ true);
// NB: this destructor can only be entered when there are no
// references to this C++ object (obviously), NOR any references
@ -25,7 +25,7 @@ void PyObjectSlot::maybe_destroy_pyobj() {
}
PyInterpreter* PyObjectSlot::pyobj_interpreter() {
return getGlobalPyInterpreter();
return pyobj_interpreter_.load(std::memory_order_acquire);
}
PyObject* PyObjectSlot::_unchecked_untagged_pyobj() const {
@ -35,7 +35,7 @@ PyObject* PyObjectSlot::_unchecked_untagged_pyobj() const {
}
PyInterpreter& PyObjectSlot::load_pyobj_interpreter() const {
auto interpreter = getGlobalPyInterpreter();
auto interpreter = pyobj_interpreter_.load(std::memory_order_acquire);
if (interpreter) {
return *interpreter;
}

View File

@ -6,17 +6,10 @@
#include <c10/util/python_stub.h>
#include <optional>
#include <atomic>
namespace c10::impl {
// Function pointer type for getting the global interpreter
using GetPyInterpreterFn = PyInterpreter* (*)();
// Global function pointer (set by csrc initialization)
C10_API extern GetPyInterpreterFn g_get_pyinterpreter_fn;
// Helper function to get the global interpreter
C10_API PyInterpreter* getGlobalPyInterpreter();
struct C10_API PyObjectSlot {
public:
PyObjectSlot();
@ -33,6 +26,8 @@ struct C10_API PyObjectSlot {
// NB: THIS FUNCTION CAN RAISE AN EXCEPTION. Make sure to clean up after
// PyObject if necessary!
void init_pyobj(PyObject* pyobj) {
pyobj_interpreter_.store(
getGlobalPyInterpreter(), std::memory_order_relaxed);
pyobj_ = pyobj;
}
@ -60,15 +55,18 @@ struct C10_API PyObjectSlot {
// @todo alban: I'm not too sure what's going on here, we can probably delete
// it but it's worthwhile making sure
std::optional<PyObject*> check_pyobj() const {
impl::PyInterpreter* interpreter = getGlobalPyInterpreter();
if (interpreter == nullptr || pyobj_ == nullptr) {
std::optional<PyObject*> check_pyobj(bool ignore_hermetic_tls = false) const {
impl::PyInterpreter* interpreter =
pyobj_interpreter_.load(std::memory_order_acquire);
if (interpreter == nullptr) {
return std::nullopt;
}
if (c10::impl::HermeticPyObjectTLS::get_state()) {
if (!ignore_hermetic_tls && c10::impl::HermeticPyObjectTLS::get_state()) {
return std::nullopt;
} else {
return _unchecked_untagged_pyobj();
}
return _unchecked_untagged_pyobj();
}
PyInterpreter& load_pyobj_interpreter() const;
@ -78,6 +76,30 @@ struct C10_API PyObjectSlot {
void set_owns_pyobj(bool b);
private:
// This field contains the interpreter tag for this object. See
// Note [Python interpreter tag] for general context
//
// Note [Memory ordering on Python interpreter tag]
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
// What memory_order do we need when accessing this atomic? We don't
// need a single total modification order (as provided by
// memory_order_seq_cst) as pyobj_interpreter_ is monotonic: it can only
// transition from -1 to some positive integer and never changes afterwards.
// Because there is only one modification, it trivially already has a total
// modification order (e.g., we don't need fences or locked instructions on
// x86)
//
// In fact, one could make a reasonable argument that relaxed reads are OK,
// due to the presence of external locking (GIL) to ensure that interactions
// with other data structures are still correctly synchronized, so that
// we fall in the "Single-Location Data Structures" case as described in
// http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p2055r0.pdf
// However, on x86, it doesn't matter if I use acquire or relaxed on the load
// as I get the same assembly in both cases. So I just use the more
// conservative acquire (which will impede compiler optimizations but I don't
// care)
std::atomic<PyInterpreter*> pyobj_interpreter_;
// This field contains a reference to a PyObject representing this Tensor.
// If pyobj is nullptr, when we transfer Tensor to Python, we allocate a new
// PyObject for it and set this field. This field does not have to be

View File

@ -504,16 +504,7 @@ struct ExpandableSegment {
SegmentRange share(SegmentRange range, std::ostream& buf) {
auto begin = segmentLeft(range.ptr);
auto end = segmentRight(range.ptr + range.size);
// header.pid needs to be padded with 4 bytes and initialized with
// 0 values to avoid random padding of different bytes each time,
// thereby ensuring that the handle can be correctly matched in
// ipcMemHandle_to_devptr.
ShareHeader header{};
header.pid = getpid();
header.segment_size = segment_size_;
header.num_handles = end - begin;
ShareHeader header{getpid(), segment_size_, end - begin};
buf.write((const char*)&header, sizeof(ShareHeader));
for (auto i : c10::irange(begin, end)) {
// NOLINTNEXTLINE(bugprone-unchecked-optional-access)

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