mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-31 20:34:54 +08:00
Compare commits
2 Commits
ciflow/tru
...
csl/xml_st
| Author | SHA1 | Date | |
|---|---|---|---|
| 25244e6544 | |||
| 04f788e46b |
@ -49,20 +49,12 @@ if [ -n "$ANACONDA_PYTHON_VERSION" ]; then
|
||||
export SYSROOT_DEP="sysroot_linux-64=2.17"
|
||||
fi
|
||||
|
||||
# Install correct Python version
|
||||
# Also ensure sysroot is using a modern GLIBC to match system compilers
|
||||
if [ "$ANACONDA_PYTHON_VERSION" = "3.14" ]; then
|
||||
as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
|
||||
python="3.14.0" \
|
||||
${SYSROOT_DEP} \
|
||||
-c conda-forge
|
||||
else
|
||||
# Install correct Python version
|
||||
# Also ensure sysroot is using a modern GLIBC to match system compilers
|
||||
as_jenkins conda create -n py_$ANACONDA_PYTHON_VERSION -y\
|
||||
python="$ANACONDA_PYTHON_VERSION" \
|
||||
${SYSROOT_DEP}
|
||||
fi
|
||||
|
||||
# libstdcxx from conda default channels are too old, we need GLIBCXX_3.4.30
|
||||
# which is provided in libstdcxx 12 and up.
|
||||
conda_install libstdcxx-ng=12.3.0 --update-deps -c conda-forge
|
||||
|
||||
@ -10,7 +10,7 @@ else
|
||||
arch_path='sbsa'
|
||||
fi
|
||||
|
||||
NVSHMEM_VERSION=3.4.5
|
||||
NVSHMEM_VERSION=3.3.24
|
||||
|
||||
function install_cuda {
|
||||
version=$1
|
||||
|
||||
@ -40,7 +40,11 @@ EOF
|
||||
|
||||
# Default url values
|
||||
rocm_baseurl="http://repo.radeon.com/rocm/apt/${ROCM_VERSION}"
|
||||
amdgpu_baseurl="https://repo.radeon.com/amdgpu/${ROCM_VERSION}/ubuntu"
|
||||
|
||||
# Add amdgpu repository
|
||||
UBUNTU_VERSION_NAME=`cat /etc/os-release | grep UBUNTU_CODENAME | awk -F= '{print $2}'`
|
||||
echo "deb [arch=amd64] ${amdgpu_baseurl} ${UBUNTU_VERSION_NAME} main" > /etc/apt/sources.list.d/amdgpu.list
|
||||
|
||||
# Add rocm repository
|
||||
wget -qO - http://repo.radeon.com/rocm/rocm.gpg.key | apt-key add -
|
||||
|
||||
@ -138,12 +138,10 @@ numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
|
||||
#test_binary_ufuncs.py
|
||||
numpy==1.22.4; python_version == "3.10"
|
||||
numpy==1.26.2; python_version == "3.11" or python_version == "3.12"
|
||||
numpy==2.1.2; python_version >= "3.13" and python_version < "3.14"
|
||||
numpy==2.3.4; python_version >= "3.14"
|
||||
numpy==2.1.2; python_version >= "3.13"
|
||||
|
||||
pandas==2.0.3; python_version < "3.13"
|
||||
pandas==2.2.3; python_version >= "3.13" and python_version < "3.14"
|
||||
pandas==2.3.3; python_version >= "3.14"
|
||||
pandas==2.2.3; python_version >= "3.13"
|
||||
|
||||
#onnxruntime
|
||||
#Description: scoring engine for Open Neural Network Exchange (ONNX) models
|
||||
@ -155,8 +153,7 @@ opt-einsum==3.3
|
||||
#Pinned versions: 3.3
|
||||
#test that import: test_linalg.py
|
||||
|
||||
optree==0.13.0 ; python_version < "3.14"
|
||||
optree==0.17.0 ; python_version >= "3.14"
|
||||
optree==0.13.0
|
||||
#Description: A library for tree manipulation
|
||||
#Pinned versions: 0.13.0
|
||||
#test that import: test_vmap.py, test_aotdispatch.py, test_dynamic_shapes.py,
|
||||
@ -255,8 +252,7 @@ scikit-image==0.22.0
|
||||
#test that import:
|
||||
|
||||
scipy==1.10.1 ; python_version <= "3.11"
|
||||
scipy==1.14.1 ; python_version > "3.11" and python_version < "3.14"
|
||||
scipy==1.16.2 ; python_version >= "3.14"
|
||||
scipy==1.14.1 ; python_version >= "3.12"
|
||||
# Pin SciPy because of failing distribution tests (see #60347)
|
||||
#Description: scientific python
|
||||
#Pinned versions: 1.10.1
|
||||
@ -328,8 +324,7 @@ pywavelets==1.7.0 ; python_version >= "3.12"
|
||||
#Pinned versions: 1.4.1
|
||||
#test that import:
|
||||
|
||||
lxml==5.3.0 ; python_version < "3.14"
|
||||
lxml==6.0.2 ; python_version >= "3.14"
|
||||
lxml==5.3.0
|
||||
#Description: This is a requirement of unittest-xml-reporting
|
||||
|
||||
PyGithub==2.3.0
|
||||
@ -339,9 +334,7 @@ sympy==1.13.3
|
||||
#Pinned versions:
|
||||
#test that import:
|
||||
|
||||
onnx==1.19.1 ; python_version < "3.14"
|
||||
# Unpin once Python 3.14 is supported. See onnxruntime issue 26309.
|
||||
onnx==1.18.0 ; python_version == "3.14"
|
||||
onnx==1.19.1
|
||||
#Description: Required by onnx tests, and mypy and test_public_bindings.py when checking torch.onnx._internal
|
||||
#Pinned versions:
|
||||
#test that import:
|
||||
@ -366,7 +359,7 @@ pwlf==2.2.1
|
||||
#test that import: test_sac_estimator.py
|
||||
|
||||
# To build PyTorch itself
|
||||
pyyaml==6.0.3
|
||||
pyyaml==6.0.2
|
||||
pyzstd
|
||||
setuptools==78.1.1
|
||||
packaging==23.1
|
||||
|
||||
@ -6,7 +6,7 @@ dependencies = [
|
||||
"GitPython==3.1.45",
|
||||
"docker==7.1.0",
|
||||
"pytest==7.3.2",
|
||||
"uv==0.9.6"
|
||||
"uv==0.9.5"
|
||||
]
|
||||
|
||||
[tool.setuptools]
|
||||
|
||||
4
.github/actions/diskspace-cleanup/action.yml
vendored
4
.github/actions/diskspace-cleanup/action.yml
vendored
@ -27,9 +27,7 @@ runs:
|
||||
docker system prune -af
|
||||
diskspace_new=$(df -H --output=pcent ${docker_root_dir} | sed -n 2p | sed 's/%//' | sed 's/ //')
|
||||
if [[ "$diskspace_new" -gt "$diskspace_cutoff" ]] ; then
|
||||
diskspace_cutoff_int=$((diskspace_cutoff + 0))
|
||||
difference=$((100 - diskspace_cutoff_int))
|
||||
echo "Error: Available diskspace is less than $difference percent. Not enough diskspace."
|
||||
echo "Error: Available diskspace is less than $diskspace_cutoff percent. Not enough diskspace."
|
||||
echo "$msg"
|
||||
exit 1
|
||||
else
|
||||
|
||||
2
.github/ci_commit_pins/audio.txt
vendored
2
.github/ci_commit_pins/audio.txt
vendored
@ -1 +1 @@
|
||||
3b0e7a6f192ca2715e7e6cbe5db007aea7165fe2
|
||||
69bbe7363897764f9e758d851cd0340147d27f94
|
||||
|
||||
2
.github/ci_commit_pins/vision.txt
vendored
2
.github/ci_commit_pins/vision.txt
vendored
@ -1 +1 @@
|
||||
218d2ab791d437309f91e0486eb9fa7f00badc17
|
||||
1752fe6809b74921644866275ab80244b96e80bc
|
||||
|
||||
20
.github/merge_rules.yaml
vendored
20
.github/merge_rules.yaml
vendored
@ -540,26 +540,6 @@
|
||||
- Lint
|
||||
- pull
|
||||
|
||||
- name: PrivateUse1
|
||||
patterns:
|
||||
- torch/accelerator/**
|
||||
- torch/utils/backend_registration.py
|
||||
- torch/csrc/acc/**
|
||||
- torch/csrc/DeviceAccelerator.*
|
||||
- torch/csrc/profiler/standalone/privateuse1_observer.*
|
||||
- aten/src/ATen/DeviceAccelerator.*
|
||||
- aten/src/ATen/core/GeneratorForPrivateuseone.*
|
||||
- aten/src/ATen/detail/PrivateUse1HooksInterface.*
|
||||
- docs/source/accelerator/**
|
||||
- test/cpp_extensions/open_registration_extension/torch_openreg/**
|
||||
approved_by:
|
||||
- albanD
|
||||
- fffrog
|
||||
mandatory_checks_name:
|
||||
- EasyCLA
|
||||
- Lint
|
||||
- pull
|
||||
|
||||
- name: superuser
|
||||
patterns:
|
||||
- '*'
|
||||
|
||||
1
.github/pytorch-probot.yml
vendored
1
.github/pytorch-probot.yml
vendored
@ -26,7 +26,6 @@ ciflow_push_tags:
|
||||
- ciflow/nightly
|
||||
- ciflow/op-benchmark
|
||||
- ciflow/periodic
|
||||
- ciflow/periodic-rocm-mi200
|
||||
- ciflow/periodic-rocm-mi300
|
||||
- ciflow/pull
|
||||
- ciflow/quantization-periodic
|
||||
|
||||
@ -56,7 +56,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
|
||||
"nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
|
||||
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
|
||||
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
|
||||
"nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
|
||||
"nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
|
||||
"nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
|
||||
"nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
|
||||
"nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'"
|
||||
@ -73,7 +73,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
|
||||
"nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
|
||||
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
|
||||
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
|
||||
"nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
|
||||
"nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
|
||||
"nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
|
||||
"nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
|
||||
"nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'"
|
||||
@ -90,7 +90,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
|
||||
"nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | "
|
||||
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
|
||||
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
|
||||
"nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | "
|
||||
"nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | "
|
||||
"nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | "
|
||||
"nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | "
|
||||
"nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'"
|
||||
@ -107,7 +107,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
|
||||
"nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | "
|
||||
"nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
|
||||
"nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
|
||||
"nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | "
|
||||
"nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
|
||||
"nvidia-nvtx==13.0.85; platform_system == 'Linux' | "
|
||||
"nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | "
|
||||
"nvidia-cufile==1.15.1.6; platform_system == 'Linux'"
|
||||
|
||||
2
.github/workflows/build-almalinux-images.yml
vendored
2
.github/workflows/build-almalinux-images.yml
vendored
@ -36,7 +36,7 @@ jobs:
|
||||
runs-on: linux.9xlarge.ephemeral
|
||||
strategy:
|
||||
matrix:
|
||||
tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.4", "rocm7.0", "rocm7.1", "cpu"]
|
||||
tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.4", "rocm7.0", "cpu"]
|
||||
steps:
|
||||
- name: Build docker image
|
||||
uses: pytorch/pytorch/.github/actions/binary-docker-build@main
|
||||
|
||||
1
.github/workflows/build-libtorch-images.yml
vendored
1
.github/workflows/build-libtorch-images.yml
vendored
@ -54,7 +54,6 @@ jobs:
|
||||
{ tag: "cuda12.6" },
|
||||
{ tag: "rocm6.4" },
|
||||
{ tag: "rocm7.0" },
|
||||
{ tag: "rocm7.1" },
|
||||
{ tag: "cpu" },
|
||||
]
|
||||
steps:
|
||||
|
||||
1
.github/workflows/build-manywheel-images.yml
vendored
1
.github/workflows/build-manywheel-images.yml
vendored
@ -56,7 +56,6 @@ jobs:
|
||||
{ name: "manylinuxaarch64-builder", tag: "cuda12.6", runner: "linux.arm64.2xlarge.ephemeral" },
|
||||
{ name: "manylinux2_28-builder", tag: "rocm6.4", runner: "linux.9xlarge.ephemeral" },
|
||||
{ name: "manylinux2_28-builder", tag: "rocm7.0", runner: "linux.9xlarge.ephemeral" },
|
||||
{ name: "manylinux2_28-builder", tag: "rocm7.1", runner: "linux.9xlarge.ephemeral" },
|
||||
{ name: "manylinux2_28-builder", tag: "cpu", runner: "linux.9xlarge.ephemeral" },
|
||||
{ name: "manylinux2_28_aarch64-builder", tag: "cpu-aarch64", runner: "linux.arm64.2xlarge.ephemeral" },
|
||||
{ name: "manylinux2_28-builder", tag: "xpu", runner: "linux.9xlarge.ephemeral" },
|
||||
|
||||
1
.github/workflows/docker-builds.yml
vendored
1
.github/workflows/docker-builds.yml
vendored
@ -57,7 +57,6 @@ jobs:
|
||||
pytorch-linux-jammy-cuda12.4-cudnn9-py3-gcc11,
|
||||
pytorch-linux-jammy-py3.10-clang12,
|
||||
pytorch-linux-jammy-py3.13-clang12,
|
||||
pytorch-linux-jammy-py3.14-clang12,
|
||||
pytorch-linux-jammy-rocm-n-py3,
|
||||
pytorch-linux-noble-rocm-n-py3,
|
||||
pytorch-linux-jammy-rocm-n-py3-benchmarks,
|
||||
|
||||
56
.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
generated
vendored
56
.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
generated
vendored
@ -132,7 +132,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_10-cuda-aarch64-12_6
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -178,7 +178,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_10-cuda-aarch64-12_8
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -224,7 +224,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_10-cuda-aarch64-12_9
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -270,7 +270,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_10-cuda-aarch64-13_0
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -381,7 +381,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_11-cuda-aarch64-12_6
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -427,7 +427,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_11-cuda-aarch64-12_8
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -473,7 +473,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_11-cuda-aarch64-12_9
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -519,7 +519,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_11-cuda-aarch64-13_0
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -630,7 +630,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_12-cuda-aarch64-12_6
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -676,7 +676,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_12-cuda-aarch64-12_8
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -722,7 +722,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_12-cuda-aarch64-12_9
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -768,7 +768,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_12-cuda-aarch64-13_0
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -879,7 +879,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_13-cuda-aarch64-12_6
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -925,7 +925,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_13-cuda-aarch64-12_8
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -971,7 +971,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_13-cuda-aarch64-12_9
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1017,7 +1017,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_13-cuda-aarch64-13_0
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1128,7 +1128,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_13t-cuda-aarch64-12_6
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1174,7 +1174,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_13t-cuda-aarch64-12_8
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1220,7 +1220,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_13t-cuda-aarch64-12_9
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1266,7 +1266,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_13t-cuda-aarch64-13_0
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1377,7 +1377,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_14-cuda-aarch64-12_6
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1423,7 +1423,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_14-cuda-aarch64-12_8
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1469,7 +1469,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_14-cuda-aarch64-12_9
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1515,7 +1515,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_14-cuda-aarch64-13_0
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1626,7 +1626,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_14t-cuda-aarch64-12_6
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1672,7 +1672,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_14t-cuda-aarch64-12_8
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1718,7 +1718,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_14t-cuda-aarch64-12_9
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
@ -1764,7 +1764,7 @@ jobs:
|
||||
ALPINE_IMAGE: "arm64v8/alpine"
|
||||
build_name: manywheel-py3_14t-cuda-aarch64-13_0
|
||||
build_environment: linux-aarch64-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
timeout-minutes: 420
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
|
||||
56
.github/workflows/generated-linux-binary-manywheel-nightly.yml
generated
vendored
56
.github/workflows/generated-linux-binary-manywheel-nightly.yml
generated
vendored
@ -127,7 +127,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
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-cuda12_9
|
||||
build_environment: linux-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_10-cuda12_9-test: # Testing
|
||||
@ -325,7 +325,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_10-cuda13_0-test: # Testing
|
||||
@ -793,7 +793,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_11-cuda12_6-test: # Testing
|
||||
@ -859,7 +859,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_11-cuda12_8-test: # Testing
|
||||
@ -925,7 +925,7 @@ jobs:
|
||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||
build_name: manywheel-py3_11-cuda12_9
|
||||
build_environment: linux-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_11-cuda12_9-test: # Testing
|
||||
@ -991,7 +991,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_11-cuda13_0-test: # Testing
|
||||
@ -1459,7 +1459,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_12-cuda12_6-test: # Testing
|
||||
@ -1525,7 +1525,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_12-cuda12_8-test: # Testing
|
||||
@ -1591,7 +1591,7 @@ jobs:
|
||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||
build_name: manywheel-py3_12-cuda12_9
|
||||
build_environment: linux-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_12-cuda12_9-test: # Testing
|
||||
@ -1657,7 +1657,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_12-cuda13_0-test: # Testing
|
||||
@ -2125,7 +2125,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_13-cuda12_6-test: # Testing
|
||||
@ -2191,7 +2191,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_13-cuda12_8-test: # Testing
|
||||
@ -2257,7 +2257,7 @@ jobs:
|
||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||
build_name: manywheel-py3_13-cuda12_9
|
||||
build_environment: linux-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_13-cuda12_9-test: # Testing
|
||||
@ -2323,7 +2323,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_13-cuda13_0-test: # Testing
|
||||
@ -2791,7 +2791,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_13t-cuda12_6-test: # Testing
|
||||
@ -2857,7 +2857,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_13t-cuda12_8-test: # Testing
|
||||
@ -2923,7 +2923,7 @@ jobs:
|
||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||
build_name: manywheel-py3_13t-cuda12_9
|
||||
build_environment: linux-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_13t-cuda12_9-test: # Testing
|
||||
@ -2989,7 +2989,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_13t-cuda13_0-test: # Testing
|
||||
@ -3457,7 +3457,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_14-cuda12_6-test: # Testing
|
||||
@ -3523,7 +3523,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_14-cuda12_8-test: # Testing
|
||||
@ -3589,7 +3589,7 @@ jobs:
|
||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||
build_name: manywheel-py3_14-cuda12_9
|
||||
build_environment: linux-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_14-cuda12_9-test: # Testing
|
||||
@ -3655,7 +3655,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_14-cuda13_0-test: # Testing
|
||||
@ -4123,7 +4123,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.6.77; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.6.80; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.6.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.0.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.7.77; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_14t-cuda12_6-test: # Testing
|
||||
@ -4189,7 +4189,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.8.90; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.8.90; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.8.4.1; platform_system == 'Linux' | nvidia-cufft-cu12==11.3.3.83; platform_system == 'Linux' | nvidia-curand-cu12==10.3.9.90; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_14t-cuda12_8-test: # Testing
|
||||
@ -4255,7 +4255,7 @@ jobs:
|
||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||
build_name: manywheel-py3_14t-cuda12_9
|
||||
build_environment: linux-binary-manywheel
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.4.5; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.9.86; platform_system == 'Linux' | nvidia-cuda-runtime-cu12==12.9.79; platform_system == 'Linux' | nvidia-cuda-cupti-cu12==12.9.79; platform_system == 'Linux' | nvidia-cudnn-cu12==9.10.2.21; platform_system == 'Linux' | nvidia-cublas-cu12==12.9.1.4; platform_system == 'Linux' | nvidia-cufft-cu12==11.4.1.4; platform_system == 'Linux' | nvidia-curand-cu12==10.3.10.19; platform_system == 'Linux' | nvidia-cusolver-cu12==11.7.5.82; platform_system == 'Linux' | nvidia-cusparse-cu12==12.5.10.65; platform_system == 'Linux' | nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.20; platform_system == 'Linux' | nvidia-nvtx-cu12==12.9.79; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.9.86; platform_system == 'Linux' | nvidia-cufile-cu12==1.14.1.1; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_14t-cuda12_9-test: # Testing
|
||||
@ -4321,7 +4321,7 @@ jobs:
|
||||
runner_prefix: "${{ 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==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.4.5; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.88; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.96; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.85; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.1.0.3; platform_system == 'Linux' | nvidia-cufft==12.0.0.61; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.4.66; platform_system == 'Linux' | nvidia-cusparse==12.6.3.3; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.85; platform_system == 'Linux' | nvidia-nvjitlink==13.0.88; platform_system == 'Linux' | nvidia-cufile==1.15.1.6; platform_system == 'Linux'
|
||||
secrets:
|
||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||
manywheel-py3_14t-cuda13_0-test: # Testing
|
||||
|
||||
84
.github/workflows/periodic-rocm-mi200.yml
vendored
84
.github/workflows/periodic-rocm-mi200.yml
vendored
@ -1,84 +0,0 @@
|
||||
name: periodic-rocm-mi200
|
||||
|
||||
on:
|
||||
schedule:
|
||||
# We have several schedules so jobs can check github.event.schedule to activate only for a fraction of the runs.
|
||||
# Also run less frequently on weekends.
|
||||
- cron: 45 0,8,16 * * 1-5
|
||||
- cron: 45 4 * * 0,6
|
||||
- cron: 45 4,12,20 * * 1-5
|
||||
- cron: 45 12 * * 0,6
|
||||
- cron: 29 8 * * * # about 1:29am PDT, for mem leak check and rerun disabled tests
|
||||
push:
|
||||
tags:
|
||||
- ciflow/periodic/*
|
||||
- ciflow/periodic-rocm-mi200/*
|
||||
branches:
|
||||
- release/*
|
||||
workflow_dispatch:
|
||||
|
||||
concurrency:
|
||||
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}-${{ github.event.schedule }}
|
||||
cancel-in-progress: true
|
||||
|
||||
permissions:
|
||||
id-token: write
|
||||
contents: read
|
||||
|
||||
jobs:
|
||||
llm-td:
|
||||
if: github.repository_owner == 'pytorch'
|
||||
name: before-test
|
||||
uses: ./.github/workflows/llm_td_retrieval.yml
|
||||
permissions:
|
||||
id-token: write
|
||||
contents: read
|
||||
|
||||
target-determination:
|
||||
name: before-test
|
||||
uses: ./.github/workflows/target_determination.yml
|
||||
needs: llm-td
|
||||
permissions:
|
||||
id-token: write
|
||||
contents: read
|
||||
|
||||
get-label-type:
|
||||
name: get-label-type
|
||||
uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
|
||||
if: (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch'
|
||||
with:
|
||||
triggering_actor: ${{ github.triggering_actor }}
|
||||
issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
|
||||
curr_branch: ${{ github.head_ref || github.ref_name }}
|
||||
curr_ref_type: ${{ github.ref_type }}
|
||||
|
||||
linux-jammy-rocm-py3_10-build:
|
||||
name: linux-jammy-rocm-py3.10
|
||||
uses: ./.github/workflows/_linux-build.yml
|
||||
needs: get-label-type
|
||||
with:
|
||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||
build-environment: linux-jammy-rocm-py3.10
|
||||
docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
|
||||
test-matrix: |
|
||||
{ include: [
|
||||
{ config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
|
||||
{ config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
|
||||
{ config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.4", owners: ["module:rocm", "oncall:distributed"] },
|
||||
]}
|
||||
secrets: inherit
|
||||
|
||||
linux-jammy-rocm-py3_10-test:
|
||||
permissions:
|
||||
id-token: write
|
||||
contents: read
|
||||
name: linux-jammy-rocm-py3.10
|
||||
uses: ./.github/workflows/_rocm-test.yml
|
||||
needs:
|
||||
- linux-jammy-rocm-py3_10-build
|
||||
- target-determination
|
||||
with:
|
||||
build-environment: linux-jammy-rocm-py3.10
|
||||
docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
|
||||
test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
|
||||
secrets: inherit
|
||||
31
.github/workflows/periodic.yml
vendored
31
.github/workflows/periodic.yml
vendored
@ -204,6 +204,37 @@ jobs:
|
||||
test-matrix: ${{ needs.linux-jammy-cuda13_0-py3_10-gcc11-build.outputs.test-matrix }}
|
||||
secrets: inherit
|
||||
|
||||
linux-jammy-rocm-py3_10-build:
|
||||
name: linux-jammy-rocm-py3.10
|
||||
uses: ./.github/workflows/_linux-build.yml
|
||||
needs: get-label-type
|
||||
with:
|
||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||
build-environment: linux-jammy-rocm-py3.10
|
||||
docker-image-name: ci-image:pytorch-linux-jammy-rocm-n-py3
|
||||
test-matrix: |
|
||||
{ include: [
|
||||
{ config: "distributed", shard: 1, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
|
||||
{ config: "distributed", shard: 2, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
|
||||
{ config: "distributed", shard: 3, num_shards: 3, runner: "linux.rocm.gpu.mi250.4", owners: ["module:rocm", "oncall:distributed"] },
|
||||
]}
|
||||
secrets: inherit
|
||||
|
||||
linux-jammy-rocm-py3_10-test:
|
||||
permissions:
|
||||
id-token: write
|
||||
contents: read
|
||||
name: linux-jammy-rocm-py3.10
|
||||
uses: ./.github/workflows/_rocm-test.yml
|
||||
needs:
|
||||
- linux-jammy-rocm-py3_10-build
|
||||
- target-determination
|
||||
with:
|
||||
build-environment: linux-jammy-rocm-py3.10
|
||||
docker-image: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.docker-image }}
|
||||
test-matrix: ${{ needs.linux-jammy-rocm-py3_10-build.outputs.test-matrix }}
|
||||
secrets: inherit
|
||||
|
||||
linux-jammy-cuda12_8-py3-gcc11-slow-gradcheck-build:
|
||||
name: linux-jammy-cuda12.8-py3-gcc11-slow-gradcheck
|
||||
uses: ./.github/workflows/_linux-build.yml
|
||||
|
||||
1
.github/workflows/upload-test-stats.yml
vendored
1
.github/workflows/upload-test-stats.yml
vendored
@ -6,7 +6,6 @@ on:
|
||||
- pull
|
||||
- trunk
|
||||
- periodic
|
||||
- periodic-rocm-mi200
|
||||
- periodic-rocm-mi300
|
||||
- inductor
|
||||
- unstable
|
||||
|
||||
20
.github/workflows/xpu.yml
vendored
20
.github/workflows/xpu.yml
vendored
@ -59,18 +59,14 @@ jobs:
|
||||
runner: linux.c7i.12xlarge
|
||||
test-matrix: |
|
||||
{ include: [
|
||||
{ config: "default", shard: 1, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 2, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 3, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 4, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 5, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 6, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 7, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 8, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 9, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 10, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 11, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 12, num_shards: 12, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 1, num_shards: 8, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 2, num_shards: 8, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 3, num_shards: 8, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 4, num_shards: 8, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 5, num_shards: 8, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 6, num_shards: 8, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 7, num_shards: 8, runner: "linux.idc.xpu" },
|
||||
{ config: "default", shard: 8, num_shards: 8, runner: "linux.idc.xpu" },
|
||||
]}
|
||||
secrets: inherit
|
||||
|
||||
|
||||
@ -374,7 +374,7 @@ cmake_dependent_option(
|
||||
"Build the lazy Torchscript backend, not compatible with mobile builds" ON
|
||||
"NOT INTERN_BUILD_MOBILE" OFF)
|
||||
cmake_dependent_option(BUILD_FUNCTORCH "Build Functorch" ON "BUILD_PYTHON" OFF)
|
||||
cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin folder"
|
||||
cmake_dependent_option(BUILD_BUNDLE_PTXAS "Bundle PTX into torch/bin fodler"
|
||||
OFF "USE_CUDA" OFF)
|
||||
cmake_dependent_option(USE_KLEIDIAI "Use KleidiAI for the ARM CPU & AARCH64 architecture." ON
|
||||
"CPU_AARCH64" OFF)
|
||||
|
||||
@ -260,7 +260,7 @@ IF(USE_FBGEMM_GENAI)
|
||||
if(USE_CUDA)
|
||||
# To avoid increasing the build time/binary size unnecessarily, use an allow-list of kernels to build.
|
||||
# If you want to integrate a kernel from FBGEMM into torch, you have to add it here.
|
||||
set(FBGEMM_CUTLASS_KERNELS_REGEX ".*(mx8mx8bf16_grouped|f4f4bf16_grouped|f4f4bf16).*")
|
||||
set(FBGEMM_CUTLASS_KERNELS_REGEX ".*mx8mx8bf16_grouped.*")
|
||||
file(GLOB_RECURSE fbgemm_genai_native_cuda_cu
|
||||
"${FBGEMM_GENAI_SRCS}/cutlass_extensions/*.cu"
|
||||
"${FBGEMM_GENAI_SRCS}/cutlass_extensions/**/*.cu")
|
||||
@ -291,7 +291,6 @@ IF(USE_FBGEMM_GENAI)
|
||||
|
||||
set(fbgemm_genai_cuh
|
||||
"${FBGEMM_GENAI_SRCS}/cutlass_extensions/mx8mx8bf16_grouped/"
|
||||
"${FBGEMM_GENAI_SRCS}/cutlass_extensions/f4f4bf16_grouped/"
|
||||
"${FBGEMM_GENAI_SRCS}/"
|
||||
)
|
||||
|
||||
|
||||
@ -677,8 +677,8 @@ struct CachingHostAllocatorImpl {
|
||||
// size. This allows us to quickly find a free block of the right size.
|
||||
// We use deque to store per size free list and guard the list with its own
|
||||
// mutex.
|
||||
alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>>
|
||||
free_list_{MAX_SIZE_INDEX};
|
||||
alignas(hardware_destructive_interference_size) std::vector<FreeBlockList<B>> free_list_ =
|
||||
std::vector<FreeBlockList<B>>(MAX_SIZE_INDEX);
|
||||
|
||||
alignas(hardware_destructive_interference_size) std::mutex events_mutex_;
|
||||
std::deque<std::pair<E, B*>> events_; // event queue paired with block
|
||||
|
||||
@ -354,9 +354,47 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
|
||||
|
||||
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(abs)
|
||||
Vectorized frac() const;
|
||||
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
|
||||
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(trunc)
|
||||
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(sqrt)
|
||||
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
Vectorized<c10::BFloat16> neg() const {
|
||||
return -values;
|
||||
}
|
||||
Vectorized<c10::BFloat16> reciprocal() const {
|
||||
return 1.0f / values;
|
||||
}
|
||||
Vectorized<c10::BFloat16> operator==(
|
||||
const Vectorized<c10::BFloat16>& other) const {
|
||||
return values == other.values;
|
||||
}
|
||||
|
||||
Vectorized<c10::BFloat16> operator!=(
|
||||
const Vectorized<c10::BFloat16>& other) const {
|
||||
return values != other.values;
|
||||
}
|
||||
|
||||
Vectorized<c10::BFloat16> operator<(
|
||||
const Vectorized<c10::BFloat16>& other) const {
|
||||
return values < other.values;
|
||||
}
|
||||
|
||||
Vectorized<c10::BFloat16> operator<=(
|
||||
const Vectorized<c10::BFloat16>& other) const {
|
||||
return values <= other.values;
|
||||
}
|
||||
|
||||
Vectorized<c10::BFloat16> operator>(
|
||||
const Vectorized<c10::BFloat16>& other) const {
|
||||
return values > other.values;
|
||||
}
|
||||
|
||||
Vectorized<c10::BFloat16> operator>=(
|
||||
const Vectorized<c10::BFloat16>& other) const {
|
||||
return values >= other.values;
|
||||
}
|
||||
#else
|
||||
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(neg)
|
||||
DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD(reciprocal)
|
||||
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator==)
|
||||
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator!=)
|
||||
@ -364,6 +402,7 @@ class Vectorized<c10::BFloat16> : public Vectorized16<
|
||||
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator<=)
|
||||
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator>)
|
||||
DEFINE_BINARY_COMPARISON_OPERATOR_VIA_FLOAT_METHOD(operator>=)
|
||||
#endif
|
||||
|
||||
#undef DEFINE_UNARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD
|
||||
#undef DEFINE_BINARY_ELEMENTWISE_FUNC_VIA_FLOAT_METHOD
|
||||
@ -412,28 +451,52 @@ template <>
|
||||
Vectorized<c10::BFloat16> inline operator+(
|
||||
const Vectorized<c10::BFloat16>& a,
|
||||
const Vectorized<c10::BFloat16>& b) {
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
bfloat16x8_t x = a;
|
||||
bfloat16x8_t y = b;
|
||||
return x + y;
|
||||
#else
|
||||
return binary_operator_via_float(std::plus<Vectorized<float>>(), a, b);
|
||||
#endif
|
||||
}
|
||||
|
||||
template <>
|
||||
Vectorized<c10::BFloat16> inline operator-(
|
||||
const Vectorized<c10::BFloat16>& a,
|
||||
const Vectorized<c10::BFloat16>& b) {
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
bfloat16x8_t x = a;
|
||||
bfloat16x8_t y = b;
|
||||
return x - y;
|
||||
#else
|
||||
return binary_operator_via_float(std::minus<Vectorized<float>>(), a, b);
|
||||
#endif
|
||||
}
|
||||
|
||||
template <>
|
||||
Vectorized<c10::BFloat16> inline operator*(
|
||||
const Vectorized<c10::BFloat16>& a,
|
||||
const Vectorized<c10::BFloat16>& b) {
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
bfloat16x8_t x = a;
|
||||
bfloat16x8_t y = b;
|
||||
return x * y;
|
||||
#else
|
||||
return binary_operator_via_float(std::multiplies<Vectorized<float>>(), a, b);
|
||||
#endif
|
||||
}
|
||||
|
||||
template <>
|
||||
Vectorized<c10::BFloat16> inline operator/(
|
||||
const Vectorized<c10::BFloat16>& a,
|
||||
const Vectorized<c10::BFloat16>& b) {
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
bfloat16x8_t x = a;
|
||||
bfloat16x8_t y = b;
|
||||
return x / y;
|
||||
#else
|
||||
return binary_operator_via_float(std::divides<Vectorized<float>>(), a, b);
|
||||
#endif
|
||||
}
|
||||
|
||||
// frac. Implement this here so we can use subtraction
|
||||
@ -544,12 +607,19 @@ Vectorized<c10::BFloat16> inline fmadd(
|
||||
const Vectorized<c10::BFloat16>& a,
|
||||
const Vectorized<c10::BFloat16>& b,
|
||||
const Vectorized<c10::BFloat16>& c) {
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
bfloat16x8_t x = a;
|
||||
bfloat16x8_t y = b;
|
||||
bfloat16x8_t z = c;
|
||||
return x * y + z;
|
||||
#else
|
||||
// NOTE [BF16 FMA]: There isn't an FMA that accumulates into BF16! Also,
|
||||
// vbfmlalbq_f32 and vbfmlaltq_f32 take the even and odd-numbered
|
||||
// elements, not the bottom and top half, so they don't seem
|
||||
// particularly useful here. Ideally we would include dot product in
|
||||
// the Vectorized interface...
|
||||
return a * b + c;
|
||||
#endif
|
||||
}
|
||||
|
||||
template <>
|
||||
@ -557,8 +627,15 @@ Vectorized<c10::BFloat16> inline fnmadd(
|
||||
const Vectorized<c10::BFloat16>& a,
|
||||
const Vectorized<c10::BFloat16>& b,
|
||||
const Vectorized<c10::BFloat16>& c) {
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
bfloat16x8_t x = a;
|
||||
bfloat16x8_t y = b;
|
||||
bfloat16x8_t z = c;
|
||||
return (-x) * y + z;
|
||||
#else
|
||||
// See NOTE [BF16 FMA] above.
|
||||
return -a * b + c;
|
||||
#endif
|
||||
}
|
||||
|
||||
template <>
|
||||
@ -566,8 +643,15 @@ Vectorized<c10::BFloat16> inline fmsub(
|
||||
const Vectorized<c10::BFloat16>& a,
|
||||
const Vectorized<c10::BFloat16>& b,
|
||||
const Vectorized<c10::BFloat16>& c) {
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
bfloat16x8_t x = a;
|
||||
bfloat16x8_t y = b;
|
||||
bfloat16x8_t z = c;
|
||||
return x * y - z;
|
||||
#else
|
||||
// See NOTE [BF16 FMA] above.
|
||||
return a * b - c;
|
||||
#endif
|
||||
}
|
||||
|
||||
template <>
|
||||
@ -575,8 +659,15 @@ Vectorized<c10::BFloat16> inline fnmsub(
|
||||
const Vectorized<c10::BFloat16>& a,
|
||||
const Vectorized<c10::BFloat16>& b,
|
||||
const Vectorized<c10::BFloat16>& c) {
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
bfloat16x8_t x = a;
|
||||
bfloat16x8_t y = b;
|
||||
bfloat16x8_t z = c;
|
||||
return (-x) * y - z;
|
||||
#else
|
||||
// See NOTE [BF16 FMA] above.
|
||||
return -a * b - c;
|
||||
#endif
|
||||
}
|
||||
|
||||
#endif // !defined(C10_MOBILE) && defined(__aarch64__)
|
||||
|
||||
@ -21,46 +21,12 @@ inline void convertImpl(
|
||||
}
|
||||
}
|
||||
|
||||
template <typename to_type>
|
||||
inline void convertFromBool(
|
||||
const bool* __restrict src,
|
||||
to_type* __restrict dst,
|
||||
int64_t n) {
|
||||
const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
|
||||
uint64_t len = static_cast<uint64_t>(n);
|
||||
for (uint64_t i = 0; i < len; i++) {
|
||||
dst[i] = srcPtr[i] != 0 ? static_cast<to_type>(1) : static_cast<to_type>(0);
|
||||
}
|
||||
}
|
||||
|
||||
template <typename from_type>
|
||||
inline void convertToBool(
|
||||
const from_type* __restrict src,
|
||||
bool* __restrict dst,
|
||||
int64_t n) {
|
||||
uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
|
||||
uint64_t len = static_cast<uint64_t>(n);
|
||||
for (uint64_t i = 0; i < len; i++) {
|
||||
dstPtr[i] = src[i] != static_cast<from_type>(0) ? 1 : 0;
|
||||
}
|
||||
}
|
||||
|
||||
#define CONVERT_TEMPLATE(from_type, to_type) \
|
||||
template <> \
|
||||
inline void convert(const from_type* src, to_type* dst, int64_t n) { \
|
||||
return convertImpl<from_type, to_type>(src, dst, n); \
|
||||
}
|
||||
|
||||
#define CONVERT_FROM_BOOL_TEMPLATE(to_type) \
|
||||
inline void convert(const bool* src, to_type* dst, int64_t n) { \
|
||||
return convertFromBool<to_type>(src, dst, n); \
|
||||
}
|
||||
|
||||
#define CONVERT_TO_BOOL_TEMPLATE(from_type) \
|
||||
inline void convert(const from_type* src, bool* dst, int64_t n) { \
|
||||
return convertToBool<from_type>(src, dst, n); \
|
||||
}
|
||||
|
||||
CONVERT_TEMPLATE(uint8_t, uint8_t)
|
||||
CONVERT_TEMPLATE(uint8_t, int8_t)
|
||||
CONVERT_TEMPLATE(uint8_t, int16_t)
|
||||
@ -68,7 +34,6 @@ CONVERT_TEMPLATE(uint8_t, int32_t)
|
||||
CONVERT_TEMPLATE(uint8_t, int64_t)
|
||||
CONVERT_TEMPLATE(uint8_t, float)
|
||||
CONVERT_TEMPLATE(uint8_t, double)
|
||||
CONVERT_TO_BOOL_TEMPLATE(uint8_t)
|
||||
CONVERT_TEMPLATE(int8_t, uint8_t)
|
||||
CONVERT_TEMPLATE(int8_t, int8_t)
|
||||
CONVERT_TEMPLATE(int8_t, int16_t)
|
||||
@ -76,7 +41,6 @@ CONVERT_TEMPLATE(int8_t, int32_t)
|
||||
CONVERT_TEMPLATE(int8_t, int64_t)
|
||||
CONVERT_TEMPLATE(int8_t, float)
|
||||
CONVERT_TEMPLATE(int8_t, double)
|
||||
CONVERT_TO_BOOL_TEMPLATE(int8_t)
|
||||
CONVERT_TEMPLATE(int16_t, uint8_t)
|
||||
CONVERT_TEMPLATE(int16_t, int8_t)
|
||||
CONVERT_TEMPLATE(int16_t, int16_t)
|
||||
@ -84,7 +48,6 @@ CONVERT_TEMPLATE(int16_t, int32_t)
|
||||
CONVERT_TEMPLATE(int16_t, int64_t)
|
||||
CONVERT_TEMPLATE(int16_t, float)
|
||||
CONVERT_TEMPLATE(int16_t, double)
|
||||
CONVERT_TO_BOOL_TEMPLATE(int16_t)
|
||||
CONVERT_TEMPLATE(int32_t, uint8_t)
|
||||
CONVERT_TEMPLATE(int32_t, int8_t)
|
||||
CONVERT_TEMPLATE(int32_t, int16_t)
|
||||
@ -92,7 +55,6 @@ CONVERT_TEMPLATE(int32_t, int32_t)
|
||||
CONVERT_TEMPLATE(int32_t, int64_t)
|
||||
CONVERT_TEMPLATE(int32_t, float)
|
||||
CONVERT_TEMPLATE(int32_t, double)
|
||||
CONVERT_TO_BOOL_TEMPLATE(int32_t)
|
||||
CONVERT_TEMPLATE(int64_t, uint8_t)
|
||||
CONVERT_TEMPLATE(int64_t, int8_t)
|
||||
CONVERT_TEMPLATE(int64_t, int16_t)
|
||||
@ -100,7 +62,6 @@ CONVERT_TEMPLATE(int64_t, int32_t)
|
||||
CONVERT_TEMPLATE(int64_t, int64_t)
|
||||
CONVERT_TEMPLATE(int64_t, float)
|
||||
CONVERT_TEMPLATE(int64_t, double)
|
||||
CONVERT_TO_BOOL_TEMPLATE(int64_t)
|
||||
CONVERT_TEMPLATE(float, uint8_t)
|
||||
CONVERT_TEMPLATE(float, int8_t)
|
||||
CONVERT_TEMPLATE(float, int16_t)
|
||||
@ -108,7 +69,6 @@ CONVERT_TEMPLATE(float, int32_t)
|
||||
CONVERT_TEMPLATE(float, int64_t)
|
||||
CONVERT_TEMPLATE(float, float)
|
||||
CONVERT_TEMPLATE(float, double)
|
||||
CONVERT_TO_BOOL_TEMPLATE(float)
|
||||
CONVERT_TEMPLATE(double, uint8_t)
|
||||
CONVERT_TEMPLATE(double, int8_t)
|
||||
CONVERT_TEMPLATE(double, int16_t)
|
||||
@ -116,14 +76,6 @@ CONVERT_TEMPLATE(double, int32_t)
|
||||
CONVERT_TEMPLATE(double, int64_t)
|
||||
CONVERT_TEMPLATE(double, float)
|
||||
CONVERT_TEMPLATE(double, double)
|
||||
CONVERT_TO_BOOL_TEMPLATE(double)
|
||||
CONVERT_FROM_BOOL_TEMPLATE(uint8_t)
|
||||
CONVERT_FROM_BOOL_TEMPLATE(int8_t)
|
||||
CONVERT_FROM_BOOL_TEMPLATE(int16_t)
|
||||
CONVERT_FROM_BOOL_TEMPLATE(int32_t)
|
||||
CONVERT_FROM_BOOL_TEMPLATE(int64_t)
|
||||
CONVERT_FROM_BOOL_TEMPLATE(float)
|
||||
CONVERT_FROM_BOOL_TEMPLATE(double)
|
||||
#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
|
||||
|
||||
#define CONVERT_FROM_FP16_TEMPLATE(to_type) \
|
||||
@ -155,41 +107,6 @@ CONVERT_TO_FP16_TEMPLATE(int32_t)
|
||||
CONVERT_TO_FP16_TEMPLATE(int64_t)
|
||||
CONVERT_TO_FP16_TEMPLATE(float)
|
||||
CONVERT_TO_FP16_TEMPLATE(double)
|
||||
|
||||
inline void convertBoolToFp16Impl(
|
||||
const bool* __restrict src,
|
||||
at::Half* __restrict dst,
|
||||
int64_t n) {
|
||||
const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
|
||||
float16_t* dstPtr = reinterpret_cast<float16_t*>(dst);
|
||||
uint64_t len = static_cast<uint64_t>(n);
|
||||
for (uint64_t i = 0; i < len; i++) {
|
||||
dstPtr[i] = srcPtr[i] != 0 ? 1.0 : 0;
|
||||
}
|
||||
}
|
||||
|
||||
template <>
|
||||
inline void convert(const bool* src, at::Half* dst, int64_t n) {
|
||||
return convertBoolToFp16Impl(src, dst, n);
|
||||
}
|
||||
|
||||
inline void convertFp16ToBoolImpl(
|
||||
const at::Half* __restrict src,
|
||||
bool* __restrict dst,
|
||||
int64_t n) {
|
||||
const float16_t* srcPtr = reinterpret_cast<const float16_t*>(src);
|
||||
uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
|
||||
uint64_t len = static_cast<uint64_t>(n);
|
||||
for (uint64_t i = 0; i < len; i++) {
|
||||
dstPtr[i] = srcPtr[i] != 0.0 ? 1 : 0;
|
||||
}
|
||||
}
|
||||
|
||||
template <>
|
||||
inline void convert(const at::Half* src, bool* dst, int64_t n) {
|
||||
return convertFp16ToBoolImpl(src, dst, n);
|
||||
}
|
||||
|
||||
#endif
|
||||
#ifdef __ARM_FEATURE_BF16
|
||||
CONVERT_TEMPLATE(bfloat16_t, uint8_t)
|
||||
@ -207,44 +124,6 @@ CONVERT_TEMPLATE(int32_t, bfloat16_t)
|
||||
CONVERT_TEMPLATE(int64_t, bfloat16_t)
|
||||
CONVERT_TEMPLATE(float, bfloat16_t)
|
||||
CONVERT_TEMPLATE(double, bfloat16_t)
|
||||
|
||||
inline void convertBoolToBfloat16Impl(
|
||||
const bool* __restrict src,
|
||||
c10::BFloat16* __restrict dst,
|
||||
int64_t n) {
|
||||
const uint8_t* srcPtr = reinterpret_cast<const uint8_t*>(src);
|
||||
uint16_t* dstPtr = reinterpret_cast<uint16_t*>(dst);
|
||||
uint64_t len = static_cast<uint64_t>(n);
|
||||
constexpr uint16_t kBf16One = 0x3f80; // 1.0 in bfloat16
|
||||
for (uint64_t i = 0; i < len; i++) {
|
||||
dstPtr[i] = srcPtr[i] != 0 ? kBf16One : 0;
|
||||
}
|
||||
}
|
||||
|
||||
template <>
|
||||
inline void convert(const bool* src, c10::BFloat16* dst, int64_t n) {
|
||||
return convertBoolToBfloat16Impl(src, dst, n);
|
||||
}
|
||||
|
||||
inline void convertBfloat16ToBoolImpl(
|
||||
const c10::BFloat16* __restrict src,
|
||||
bool* __restrict dst,
|
||||
int64_t n) {
|
||||
uint8_t* dstPtr = reinterpret_cast<uint8_t*>(dst);
|
||||
const uint16_t* srcPtr = reinterpret_cast<const uint16_t*>(src);
|
||||
uint64_t len = static_cast<uint64_t>(n);
|
||||
for (uint64_t i = 0; i < len; i++) {
|
||||
// Check if all non-sign bits are 0
|
||||
bool isBf16Zero = (srcPtr[i] & 0x7fff) == 0;
|
||||
dstPtr[i] = isBf16Zero ? 0 : 1;
|
||||
}
|
||||
}
|
||||
|
||||
template <>
|
||||
inline void convert(const c10::BFloat16* src, bool* dst, int64_t n) {
|
||||
return convertBfloat16ToBoolImpl(src, dst, n);
|
||||
}
|
||||
|
||||
#endif
|
||||
|
||||
#endif
|
||||
|
||||
@ -309,7 +309,7 @@ class Vectorized<float> {
|
||||
DEFINE_SLEEF_COMPATIBLE_UNARY_ELEMENTWISE_FUNC(expm1)
|
||||
// Implementation copied from Arm Optimized Routine
|
||||
// https://github.com/ARM-software/optimized-routines/blob/master/math/aarch64/advsimd/expf.c
|
||||
inline Vectorized<float> vexpq_f32_u20() const {
|
||||
Vectorized<float> exp_u20() const {
|
||||
// bail out to sleef if it's a special case:
|
||||
// i.e. there's an input s.t. |input| > 87.3....
|
||||
const float32x4_t special_bound = vdupq_n_f32(0x1.5d5e2ap+6f);
|
||||
@ -348,9 +348,6 @@ class Vectorized<float> {
|
||||
|
||||
return vfmaq_f32(scale, poly, scale);
|
||||
}
|
||||
Vectorized<float> exp_u20() const {
|
||||
return vexpq_f32_u20();
|
||||
}
|
||||
Vectorized<float> fexp_u20() const {
|
||||
return exp_u20();
|
||||
}
|
||||
@ -637,7 +634,7 @@ inline Vectorized<float> Vectorized<float>::erf() const {
|
||||
// - exp(- x * x)
|
||||
auto pow_2 = (*this) * (*this);
|
||||
auto neg_pow_2 = pow_2 ^ neg_zero_vec;
|
||||
auto tmp4 = neg_pow_2.vexpq_f32_u20();
|
||||
auto tmp4 = neg_pow_2.exp();
|
||||
auto tmp5 = tmp4 ^ neg_zero_vec;
|
||||
// erf(x) = sign(x) * (1 - r * t * exp(- x * x))
|
||||
auto tmp6 = t * tmp5;
|
||||
|
||||
@ -1,90 +1,78 @@
|
||||
#include <ATen/cuda/CUDAGreenContext.h>
|
||||
|
||||
#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
|
||||
#include <c10/cuda/driver_api.h>
|
||||
#include <stdexcept>
|
||||
#include <vector>
|
||||
#define HAS_CUDA_GREEN_CONTEXT() 1
|
||||
#else
|
||||
#define HAS_CUDA_GREEN_CONTEXT() 0
|
||||
// Suppress unsued private field warnings as this class is not supposed to be called
|
||||
C10_DIAGNOSTIC_PUSH_AND_IGNORED_IF_DEFINED("-Wunused-private-field")
|
||||
#endif
|
||||
|
||||
namespace at::cuda {
|
||||
GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
int driver_version;
|
||||
C10_CUDA_CHECK(cudaDriverGetVersion(&driver_version));
|
||||
TORCH_CHECK(
|
||||
driver_version >= 12080, "cuda driver too old to use green context!");
|
||||
CUcontext pctx = nullptr;
|
||||
C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&pctx));
|
||||
if (C10_UNLIKELY(!pctx)) {
|
||||
TORCH_WARN(
|
||||
"Attempted to create a green context but"
|
||||
" there was no primary context! Creating a primary context...");
|
||||
|
||||
GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
|
||||
#if HAS_CUDA_GREEN_CONTEXT()
|
||||
int driver_version;
|
||||
C10_CUDA_CHECK(cudaDriverGetVersion(&driver_version));
|
||||
TORCH_CHECK(
|
||||
driver_version >= 12080, "cuda driver too old to use green context!");
|
||||
CUcontext pctx = nullptr;
|
||||
C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuCtxGetCurrent_(&pctx));
|
||||
if (C10_UNLIKELY(!pctx)) {
|
||||
TORCH_WARN(
|
||||
"Attempted to create a green context but"
|
||||
" there was no primary context! Creating a primary context...");
|
||||
cudaFree(0);
|
||||
}
|
||||
|
||||
cudaFree(0);
|
||||
}
|
||||
CUdevice device;
|
||||
device_id_ = device_id;
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuDeviceGet_(&device, device_id));
|
||||
|
||||
CUdevice device;
|
||||
device_id_ = device_id;
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuDeviceGet_(&device, device_id));
|
||||
// Get device resources
|
||||
CUdevResource device_resource;
|
||||
C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuDeviceGetDevResource_(
|
||||
device, &device_resource, CU_DEV_RESOURCE_TYPE_SM));
|
||||
|
||||
// Get device resources
|
||||
CUdevResource device_resource;
|
||||
C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuDeviceGetDevResource_(
|
||||
device, &device_resource, CU_DEV_RESOURCE_TYPE_SM));
|
||||
// Split resources
|
||||
std::vector<CUdevResource> result(1);
|
||||
auto result_data = result.data();
|
||||
unsigned int nb_groups = 1;
|
||||
CUdevResource remaining;
|
||||
|
||||
// Split resources
|
||||
std::vector<CUdevResource> result(1);
|
||||
auto result_data = result.data();
|
||||
unsigned int nb_groups = 1;
|
||||
CUdevResource remaining;
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuDevSmResourceSplitByCount_(
|
||||
result_data,
|
||||
&nb_groups,
|
||||
&device_resource,
|
||||
&remaining,
|
||||
0, // default flags
|
||||
num_sms));
|
||||
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuDevSmResourceSplitByCount_(
|
||||
result_data,
|
||||
&nb_groups,
|
||||
&device_resource,
|
||||
&remaining,
|
||||
0, // default flags
|
||||
num_sms));
|
||||
TORCH_CHECK(nb_groups == 1, "Failed to create single resource group");
|
||||
|
||||
TORCH_CHECK(nb_groups == 1, "Failed to create single resource group");
|
||||
// Generate resource descriptor
|
||||
CUdevResourceDesc desc;
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuDevResourceGenerateDesc_(
|
||||
&desc, result_data, 1));
|
||||
|
||||
// Generate resource descriptor
|
||||
CUdevResourceDesc desc;
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuDevResourceGenerateDesc_(
|
||||
&desc, result_data, 1));
|
||||
// Create green context
|
||||
// CU_GREEN_CTX_DEFAULT_STREAM is required per docs:
|
||||
// https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__GREEN__CONTEXTS.html
|
||||
C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuGreenCtxCreate_(
|
||||
&green_ctx_, desc, device, CU_GREEN_CTX_DEFAULT_STREAM));
|
||||
|
||||
// Create green context
|
||||
// CU_GREEN_CTX_DEFAULT_STREAM is required per docs:
|
||||
// https://docs.nvidia.com/cuda/cuda-driver-api/group__CUDA__GREEN__CONTEXTS.html
|
||||
C10_CUDA_DRIVER_CHECK(c10::cuda::DriverAPI::get()->cuGreenCtxCreate_(
|
||||
&green_ctx_, desc, device, CU_GREEN_CTX_DEFAULT_STREAM));
|
||||
|
||||
// Convert to regular context
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuCtxFromGreenCtx_(&context_, green_ctx_));
|
||||
TORCH_CHECK(context_, "Green ctx conversion to regular ctx failed!");
|
||||
// Convert to regular context
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuCtxFromGreenCtx_(&context_, green_ctx_));
|
||||
TORCH_CHECK(context_, "Green ctx conversion to regular ctx failed!");
|
||||
#else
|
||||
TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
|
||||
TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
|
||||
#endif
|
||||
}
|
||||
|
||||
std::unique_ptr<GreenContext> GreenContext::create(
|
||||
uint32_t num_sms,
|
||||
std::optional<uint32_t> device_id) {
|
||||
#if HAS_CUDA_GREEN_CONTEXT()
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
if (!device_id.has_value()) {
|
||||
device_id = at::cuda::current_device();
|
||||
}
|
||||
return std::unique_ptr<GreenContext>(new GreenContext(device_id.value(), num_sms));
|
||||
return std::make_unique<GreenContext>(device_id.value(), num_sms);
|
||||
#else
|
||||
TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
|
||||
#endif
|
||||
@ -92,7 +80,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
|
||||
|
||||
// Implement move operations
|
||||
GreenContext::GreenContext(GreenContext&& other) noexcept{
|
||||
#if HAS_CUDA_GREEN_CONTEXT()
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
device_id_ = std::exchange(other.device_id_, -1);
|
||||
green_ctx_ = std::exchange(other.green_ctx_, nullptr);
|
||||
context_ = std::exchange(other.context_, nullptr);
|
||||
@ -103,7 +91,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
|
||||
}
|
||||
|
||||
GreenContext& GreenContext::operator=(GreenContext&& other) noexcept{
|
||||
#if HAS_CUDA_GREEN_CONTEXT()
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
if (this != &other) {
|
||||
// Clean up current resources
|
||||
if (green_ctx_) {
|
||||
@ -132,7 +120,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
|
||||
}
|
||||
|
||||
GreenContext::~GreenContext() noexcept{
|
||||
#if HAS_CUDA_GREEN_CONTEXT()
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
C10_CUDA_DRIVER_CHECK(
|
||||
c10::cuda::DriverAPI::get()->cuGreenCtxDestroy_(green_ctx_));
|
||||
#else
|
||||
@ -140,9 +128,25 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
|
||||
#endif
|
||||
}
|
||||
|
||||
// Get the underlying CUDA context
|
||||
CUcontext GreenContext::getContext() const {
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
return context_;
|
||||
#else
|
||||
TORCH_CHECK(false, "Green Context is only supported on CUDA 12.8+!");
|
||||
#endif
|
||||
}
|
||||
|
||||
// Get the underlying green context
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
CUgreenCtx GreenContext::getGreenContext() const {
|
||||
return green_ctx_;
|
||||
}
|
||||
#endif
|
||||
|
||||
// Make this context current
|
||||
void GreenContext::setContext() {
|
||||
#if HAS_CUDA_GREEN_CONTEXT()
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
auto current_stream = c10::cuda::getCurrentCUDAStream();
|
||||
parent_stream_ = current_stream.stream();
|
||||
|
||||
@ -171,7 +175,7 @@ GreenContext::GreenContext(uint32_t device_id, uint32_t num_sms) {
|
||||
}
|
||||
|
||||
void GreenContext::popContext() {
|
||||
#if HAS_CUDA_GREEN_CONTEXT()
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
// see above note about stream being hardcoded to the default stream
|
||||
at::cuda::CUDAEvent ev;
|
||||
ev.record(c10::cuda::getCurrentCUDAStream());
|
||||
|
||||
@ -1,38 +1,53 @@
|
||||
#pragma once
|
||||
#include <ATen/cuda/CUDAEvent.h>
|
||||
#include <cuda.h>
|
||||
|
||||
// Forward declare green context as opaque ptr
|
||||
typedef struct CUgreenCtx_st* CUgreenCtx;
|
||||
#if defined(CUDA_VERSION) && !defined(USE_ROCM) && defined(PYTORCH_C10_DRIVER_API_SUPPORTED)
|
||||
#include <c10/cuda/driver_api.h>
|
||||
#include <cuda.h>
|
||||
#include <memory>
|
||||
#include <stdexcept>
|
||||
#include <vector>
|
||||
#define CUDA_HAS_GREEN_CONTEXT 1
|
||||
#else
|
||||
#define CUDA_HAS_GREEN_CONTEXT 0
|
||||
#endif
|
||||
|
||||
namespace at::cuda {
|
||||
|
||||
class TORCH_CUDA_CPP_API GreenContext {
|
||||
public:
|
||||
// Green context creation
|
||||
static std::unique_ptr<GreenContext> create(
|
||||
uint32_t num_sms,
|
||||
std::optional<uint32_t> device_id);
|
||||
~GreenContext() noexcept;
|
||||
GreenContext(uint32_t device_id, uint32_t num_sms);
|
||||
|
||||
static std::unique_ptr<GreenContext> create(uint32_t num_sms, std::optional<uint32_t> device_id);
|
||||
|
||||
// Delete copy constructor and assignment
|
||||
GreenContext(const GreenContext&) = delete;
|
||||
GreenContext& operator=(const GreenContext&) = delete;
|
||||
|
||||
// Implement move operations
|
||||
GreenContext(GreenContext&& other) noexcept;
|
||||
GreenContext& operator=(GreenContext&& other) noexcept;
|
||||
~GreenContext() noexcept;
|
||||
|
||||
// Get the underlying CUDA context
|
||||
CUcontext getContext() const;
|
||||
|
||||
// Get the underlying green context
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
CUgreenCtx getGreenContext() const;
|
||||
#endif
|
||||
|
||||
// Make this context current
|
||||
void setContext();
|
||||
|
||||
void popContext();
|
||||
|
||||
private:
|
||||
GreenContext(uint32_t device_id, uint32_t num_sms);
|
||||
// Implement move operations
|
||||
GreenContext(GreenContext&& other) noexcept;
|
||||
GreenContext& operator=(GreenContext&& other) noexcept;
|
||||
|
||||
#if CUDA_HAS_GREEN_CONTEXT
|
||||
int32_t device_id_ = -1;
|
||||
CUgreenCtx green_ctx_ = nullptr;
|
||||
CUcontext context_ = nullptr;
|
||||
cudaStream_t parent_stream_ = nullptr;
|
||||
#endif
|
||||
};
|
||||
} // namespace at::cuda
|
||||
|
||||
@ -7,6 +7,17 @@
|
||||
#endif
|
||||
|
||||
|
||||
#if defined(USE_ROCM)
|
||||
// hipSparse const API added in v2.4.0
|
||||
#if HIPSPARSE_VERSION >= 200400
|
||||
#define AT_USE_HIPSPARSE_GENERIC_API() 1
|
||||
#else
|
||||
#define AT_USE_HIPSPARSE_GENERIC_API() 1
|
||||
#endif
|
||||
#else // USE_ROCM
|
||||
#define AT_USE_HIPSPARSE_GENERIC_API() 0
|
||||
#endif // USE_ROCM
|
||||
|
||||
// cuSparse Generic API spsv function was added in CUDA 11.3.0
|
||||
#if defined(CUDART_VERSION) && defined(CUSPARSE_VERSION) && (CUSPARSE_VERSION >= 11500)
|
||||
#define AT_USE_CUSPARSE_GENERIC_SPSV() 1
|
||||
|
||||
@ -1,7 +1,6 @@
|
||||
#include <ATen/cuda/CUDAContextLight.h>
|
||||
#include <ATen/cuda/Sleep.h>
|
||||
|
||||
#include <c10/cuda/CUDACachingAllocator.h>
|
||||
#include <c10/cuda/CUDAException.h>
|
||||
#include <c10/cuda/CUDAStream.h>
|
||||
|
||||
@ -25,22 +24,8 @@ __global__ void spin_kernel(int64_t cycles) {
|
||||
#endif
|
||||
}
|
||||
}
|
||||
|
||||
thread_local int *flag = nullptr;
|
||||
|
||||
__global__ void busy_wait_for_flag_kernel(int *flag) {
|
||||
atomicExch(flag, 1);
|
||||
while (atomicAdd(flag, 0) == 1) {
|
||||
// do nothing
|
||||
}
|
||||
}
|
||||
|
||||
__global__ void clear_flag_kernel(int *flag) {
|
||||
atomicExch(flag, 0);
|
||||
}
|
||||
|
||||
} // anonymous namespace
|
||||
|
||||
void sleep(int64_t cycles) {
|
||||
dim3 grid(1);
|
||||
dim3 block(1);
|
||||
@ -48,26 +33,6 @@ void sleep(int64_t cycles) {
|
||||
C10_CUDA_KERNEL_LAUNCH_CHECK();
|
||||
}
|
||||
|
||||
void busy_wait_for_flag() {
|
||||
if (!flag) {
|
||||
flag = (int*)c10::cuda::CUDACachingAllocator::raw_alloc(sizeof(int));
|
||||
}
|
||||
dim3 grid(1);
|
||||
dim3 block(1);
|
||||
busy_wait_for_flag_kernel<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>(flag);
|
||||
C10_CUDA_KERNEL_LAUNCH_CHECK();
|
||||
}
|
||||
|
||||
void clear_flag() {
|
||||
if (!flag) {
|
||||
flag = (int*)c10::cuda::CUDACachingAllocator::raw_alloc(sizeof(int));
|
||||
}
|
||||
dim3 grid(1);
|
||||
dim3 block(1);
|
||||
clear_flag_kernel<<<grid, block, 0, c10::cuda::getCurrentCUDAStream()>>>(flag);
|
||||
C10_CUDA_KERNEL_LAUNCH_CHECK();
|
||||
}
|
||||
|
||||
#ifdef USE_ROCM
|
||||
__global__ void flush_icache_kernel()
|
||||
{
|
||||
|
||||
@ -7,11 +7,6 @@ namespace at::cuda {
|
||||
// enqueues a kernel that spins for the specified number of cycles
|
||||
TORCH_CUDA_CU_API void sleep(int64_t cycles);
|
||||
|
||||
// enqueues a kernel that spins until a flag is cleared by a
|
||||
// corresponding call to clear_flag()
|
||||
TORCH_CUDA_CU_API void busy_wait_for_flag();
|
||||
TORCH_CUDA_CU_API void clear_flag();
|
||||
|
||||
// flushes instruction cache for ROCm; no-op for CUDA
|
||||
TORCH_CUDA_CU_API void flush_icache();
|
||||
|
||||
|
||||
@ -580,7 +580,7 @@ std::ofstream& TuningContext::GetUntunedFile(){
|
||||
filename.append(device);
|
||||
}
|
||||
|
||||
untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::app);
|
||||
untuned_file_ = std::ofstream(filename, std::ios::out | std::ios::trunc);
|
||||
}
|
||||
return untuned_file_;
|
||||
}
|
||||
|
||||
@ -1,6 +1,5 @@
|
||||
#pragma once
|
||||
|
||||
#include <c10/core/CachingDeviceAllocator.h>
|
||||
#include <c10/core/Device.h>
|
||||
#include <c10/util/Exception.h>
|
||||
|
||||
@ -152,36 +151,6 @@ struct TORCH_API MTIAHooksInterface : AcceleratorHooksInterface {
|
||||
}
|
||||
|
||||
virtual bool isAvailable() const override;
|
||||
|
||||
/* MTIAGraph related APIs */
|
||||
virtual int64_t mtiagraphCreate(bool keep_graph = false) const {
|
||||
FAIL_MTIAHOOKS_FUNC(__func__);
|
||||
return -1;
|
||||
}
|
||||
|
||||
virtual void mtiagraphCaptureBegin(int64_t handle, MempoolId_t pool) const {
|
||||
FAIL_MTIAHOOKS_FUNC(__func__);
|
||||
}
|
||||
|
||||
virtual void mtiagraphCaptureEnd(int64_t handle) const {
|
||||
FAIL_MTIAHOOKS_FUNC(__func__);
|
||||
}
|
||||
|
||||
virtual void mtiagraphInstantiate(int64_t handle) const {
|
||||
FAIL_MTIAHOOKS_FUNC(__func__);
|
||||
}
|
||||
|
||||
virtual void mtiagraphReplay(int64_t handle) const {
|
||||
FAIL_MTIAHOOKS_FUNC(__func__);
|
||||
}
|
||||
|
||||
virtual void mtiagraphReset(int64_t handle) const {
|
||||
FAIL_MTIAHOOKS_FUNC(__func__);
|
||||
}
|
||||
|
||||
virtual MempoolId_t mtiagraphPool(int64_t handle) const {
|
||||
FAIL_MTIAHOOKS_FUNC(__func__);
|
||||
}
|
||||
};
|
||||
|
||||
struct TORCH_API MTIAHooksArgs {};
|
||||
|
||||
@ -410,8 +410,8 @@ struct ConvParams {
|
||||
return false;
|
||||
}
|
||||
static long cudnn_version = detail::getCUDAHooks().versionCuDNN();
|
||||
// broken on cuDNN 9.8 - 9.14
|
||||
if (cudnn_version >= 90800 && cudnn_version < 91500) {
|
||||
// broken on cuDNN 9.8
|
||||
if (cudnn_version >= 90800) {
|
||||
if (cudnn_conv_suggest_memory_format(input, weight) == at::MemoryFormat::Contiguous &&
|
||||
(input.scalar_type() == at::kBFloat16 || input.scalar_type() == at::kHalf) &&
|
||||
weight.dim() == 5) {
|
||||
@ -689,10 +689,6 @@ static void check_shape_forward(const at::Tensor& input,
|
||||
", but got bias of size ", at::symint::sizes<T>(bias), " instead");
|
||||
|
||||
for (const auto i : c10::irange(2, k)) {
|
||||
// T could be int64_t or SymInt, Specialized numeric_limts<SymInt> in c10/core/SymInt.h
|
||||
TORCH_CHECK(padding[i-2] <= (std::numeric_limits<T>::max() - padding[i-2]),
|
||||
"Given padding=", padding[i-2], " at dimension ", i-2, " , expected padding to be at most ",
|
||||
(std::numeric_limits<T>::max() / 2));
|
||||
input_shape.push_back(at::symint::size<T>(input, i) + 2 * padding[i-2]);
|
||||
// log new kernel size considering dilation
|
||||
kernel_shape.push_back(dilation[i-2] * (weight_sizes[i]-1) + 1);
|
||||
@ -719,11 +715,6 @@ static void check_shape_forward(const at::Tensor& input,
|
||||
"Kernel size: (", kernel_ss.str(), "). Kernel size can't be greater than actual input size");
|
||||
}
|
||||
} else { // transposed
|
||||
for (const auto i : c10::irange(2, k)) {
|
||||
TORCH_CHECK(padding[i-2] <= (std::numeric_limits<T>::max() - padding[i-2]),
|
||||
"Given padding=", padding[i-2], " at dimension ", i-2, " , expected padding to be at most ",
|
||||
(std::numeric_limits<T>::max() / 2));
|
||||
}
|
||||
TORCH_CHECK(at::symint::size<T>(input, 1) == weight_sizes[0],
|
||||
"Given transposed=", transposed, ", weight of size ", weight_sizes,
|
||||
", expected input", at::symint::sizes<T>(input), " to have ", weight_sizes[0],
|
||||
|
||||
@ -52,7 +52,8 @@ Tensor conv_tbc(const Tensor& self, const Tensor& weight, const Tensor& bias, in
|
||||
for (const auto k : c10::irange(kw)) {
|
||||
int iShift = std::max(0, static_cast<int>(k - real_pad));
|
||||
int oShift = std::max(0, static_cast<int>(real_pad - k));
|
||||
long t = std::min(ilen + real_pad - k, olen) - oShift;
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
int t = std::min(ilen + real_pad - k, olen) - oShift;
|
||||
// Note: gemm assumes column-major matrices
|
||||
// input is l*m (row-major)
|
||||
// weight is m*r (row-major)
|
||||
|
||||
@ -16,7 +16,8 @@ bool canUse32BitIndexMath(const TensorBase& t, int64_t max_elem) {
|
||||
auto linearId = elements - 1;
|
||||
|
||||
// NOTE: Assumes all strides are positive, which is true for now
|
||||
for (auto i = t.dim() - 1; i >= 0; --i) {
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
for (int i = t.dim() - 1; i >= 0; --i) {
|
||||
auto curDimIndex = linearId % t.sym_size(i);
|
||||
auto curDimOffset = curDimIndex * t.sym_stride(i);
|
||||
offset += curDimOffset;
|
||||
|
||||
@ -68,6 +68,7 @@ Tensor fbgemm_linear_int8_weight_fp32_activation(
|
||||
const float* input_ptr = input_contig.const_data_ptr<float>();
|
||||
|
||||
TORCH_CHECK(input.dim() >= 2);
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
const int64_t M = size_to_dim_(input.dim() - 1, input.sizes());
|
||||
const int64_t K = input.size(input.dim() - 1);
|
||||
TORCH_CHECK(weight.dim() == 2);
|
||||
|
||||
@ -160,9 +160,10 @@ struct Dist {
|
||||
// value of k.
|
||||
parallel_for(0, combs, internal::GRAIN_SIZE / (16 * m), [p, self_start, self_end, n, m, res_start](int64_t k, int64_t end) {
|
||||
const Vec pvec(p);
|
||||
double n2 = static_cast<double>(n) - .5;
|
||||
double n2 = n - .5;
|
||||
// The -1 accounts for floating point truncation issues
|
||||
int64_t i = static_cast<int64_t>((n2 - std::sqrt(n2 * n2 - 2.0 * static_cast<double>(k) - 1.0)));
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
int64_t i = static_cast<int64_t>((n2 - std::sqrt(n2 * n2 - 2 * k - 1)));
|
||||
int64_t j = k - n * i + i * (i + 1) / 2 + i + 1;
|
||||
|
||||
const scalar_t * self_i = self_start + i * m;
|
||||
|
||||
@ -170,14 +170,10 @@ static bool isInputCompliesAddmmCudaLt(Tensor& result, const Tensor& self, const
|
||||
#if defined(CUDA_VERSION) || defined(USE_ROCM)
|
||||
const auto scalar_type = mat1.scalar_type();
|
||||
return (beta.toComplexDouble() == 1.0
|
||||
// self.dim() == 1 && result.dim() == 2 && self.sizes()[0] == mat2_sizes[1]
|
||||
// is to use lt interface only when self is bias.
|
||||
&& self.dim() == 1 && self.sizes()[0] == mat2_sizes[1] && self.is_contiguous()
|
||||
&& result.dim() == 2 && result.is_contiguous()
|
||||
// Conditions for bias to be fusable
|
||||
&& (
|
||||
self.is_contiguous() &&
|
||||
// NOTE: fine to have 1-len dims to the left from the right-most one
|
||||
(self.dim() == 1 || self.squeeze().dim() == 1) &&
|
||||
self.sizes().back() == mat2_sizes[1]
|
||||
)
|
||||
&& ( // some dtype restrictions
|
||||
#ifndef USE_ROCM
|
||||
scalar_type == at::ScalarType::Double ||
|
||||
|
||||
@ -208,62 +208,6 @@ _f8_f8_bf16_rowwise_grouped_mm(
|
||||
#endif
|
||||
}
|
||||
|
||||
Tensor&
|
||||
_f4_f4_bf16_grouped_mm_fbgemm(
|
||||
const Tensor& mat_a,
|
||||
const Tensor& mat_b,
|
||||
const Tensor& scale_a,
|
||||
const std::optional<Tensor>& global_scale_a,
|
||||
const Tensor& scale_b,
|
||||
const std::optional<Tensor>& global_scale_b,
|
||||
const std::optional<Tensor>& offs,
|
||||
const std::optional<Tensor>& bias,
|
||||
Tensor& out) {
|
||||
#if !defined(USE_ROCM) && defined(USE_FBGEMM_GENAI)
|
||||
// Typing checks
|
||||
TORCH_CHECK_VALUE(mat_a.scalar_type() == at::kFloat4_e2m1fn_x2,
|
||||
"mat_a must be Float4_e2n1fn_2, got: ", mat_a.scalar_type());
|
||||
TORCH_CHECK_VALUE(mat_b.scalar_type() == at::kFloat4_e2m1fn_x2,
|
||||
"mat_b must be Float4_e2n1fn_2, got: ", mat_b.scalar_type());
|
||||
|
||||
std::optional<Tensor> combined_global_scale = std::nullopt;
|
||||
if (global_scale_a.has_value() || global_scale_b.has_value()) {
|
||||
// NVFP4
|
||||
TORCH_CHECK_VALUE(global_scale_a.has_value() && global_scale_b.has_value(),
|
||||
"For NVFP4 grouped gemm both of global_scale_{a,b} must have values")
|
||||
TORCH_CHECK_VALUE(scale_a.scalar_type() == at::kFloat8_e4m3fn,
|
||||
"scale_a must be Float8_e4m3fn, got: ", scale_a.scalar_type());
|
||||
TORCH_CHECK_VALUE(scale_b.scalar_type() == at::kFloat8_e4m3fn,
|
||||
"scale_b must be Float8_e4m3fn, got: ", scale_b.scalar_type());
|
||||
TORCH_CHECK_VALUE(global_scale_a.value().scalar_type() == at::kFloat,
|
||||
"global_scale_a must be Float, got: ", global_scale_a.value().scalar_type());
|
||||
TORCH_CHECK_VALUE(global_scale_b.value().scalar_type() == at::kFloat,
|
||||
"global_scale_b must be Float, got: ", global_scale_b.value().scalar_type());
|
||||
combined_global_scale = global_scale_a.value().mul(global_scale_b.value());
|
||||
} else {
|
||||
// MXFP4
|
||||
TORCH_CHECK_VALUE(scale_a.scalar_type() == at::kFloat8_e8m0fnu,
|
||||
"scale_a must be Float8_e8m0fnu, got: ", scale_a.scalar_type());
|
||||
TORCH_CHECK_VALUE(scale_b.scalar_type() == at::kFloat8_e8m0fnu,
|
||||
"scale_b must be Float8_e8m0fnu, got: ", scale_b.scalar_type());
|
||||
}
|
||||
|
||||
auto o = fbgemm_gpu::f4f4bf16_grouped_mm(
|
||||
mat_a,
|
||||
mat_b,
|
||||
scale_a,
|
||||
scale_b,
|
||||
offs.value(),
|
||||
out,
|
||||
combined_global_scale
|
||||
);
|
||||
#else
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(false, "nvfp4 grouped gemm is not supported without USE_FBGEMM_GENAI, and only for CUDA")
|
||||
#endif
|
||||
|
||||
return out;
|
||||
}
|
||||
|
||||
void _check_scales_fp8_rowwise(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx, const int scale_multiplier=1) {
|
||||
// Checks scales for 2d or 3d target tensors (`mat`).
|
||||
if (mat.dim() == 2) {
|
||||
@ -301,15 +245,7 @@ void _check_scales_fp8_rowwise(const Tensor& mat, const Tensor& scale, const int
|
||||
}
|
||||
}
|
||||
|
||||
void _check_scales_blocked(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx) {
|
||||
// if {mx,nv}fp4, will need to modify K later
|
||||
bool is_fp4 = (mat.scalar_type() == kFloat4_e2m1fn_x2);
|
||||
int blocksize = 32;
|
||||
// check for nvfp4 vs. mxfp4 to fix blocksize
|
||||
if (is_fp4 && scale.scalar_type() == kFloat8_e4m3fn) {
|
||||
blocksize = 16;
|
||||
}
|
||||
|
||||
void _check_scales_mxfp8(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx) {
|
||||
// Checks scales for 2d or 3d target tensors (`mat`).
|
||||
if (mat.dim() == 2) {
|
||||
// For MXFP8, 2d tensors have variable size groups represented as subtensors,
|
||||
@ -317,19 +253,17 @@ void _check_scales_blocked(const Tensor& mat, const Tensor& scale, const int dim
|
||||
// so we can't check the scale sizes without doing a d2h sync to get the group sizes here.
|
||||
TORCH_CHECK(
|
||||
scale.dim() == mat.dim(),
|
||||
"for block-scaled, scale must have same number of dimensions as parent tensor, but got mat.dim() = ", mat.dim(),
|
||||
" and scale.dim() = ", scale.dim(), " for arg ", arg_idx
|
||||
);
|
||||
"for mxfp8, scale must have same number of dimensions as parent tensor, but got mat.dim() = ", mat.dim(), " and scale.dim() = ", scale.dim(), " for arg ", arg_idx);
|
||||
|
||||
// LHS mat shape (M, total_K) -> scale shape (rounded_up(M, 128), rounded_up_per_group(K/blocksize, 4))
|
||||
// RHS mat shape (total_K, N) -> scale shape (rounded_up(N, 128), rounded_up_per_group(K/blocksize, 4))
|
||||
// LHS mat shape (M, total_K) -> scale shape (rounded_up(M, 128), rounded_up_per_group(K/32, 4))
|
||||
// RHS mat shape (total_K, N) -> scale shape (rounded_up(N, 128), rounded_up_per_group(K/32, 4))
|
||||
// * weight is transposed prior to the call, scale stays non-transposed.
|
||||
bool LHS = arg_idx == 0;
|
||||
int scale_dim_to_check = 0;
|
||||
int mat_dim_to_check = LHS ? 0 : 1;
|
||||
TORCH_CHECK(
|
||||
scale.size(scale_dim_to_check) >= mat.size(mat_dim_to_check),
|
||||
"for block-scaled, arg ", arg_idx, " tensor shape (", mat.size(0), ", ", mat.size(1), ") ",
|
||||
"for mxfp8, arg ", arg_idx, " tensor shape (", mat.size(0), ", ", mat.size(1), ") ",
|
||||
"must have scale.shape[", scale_dim_to_check, "] >= ", mat.size(mat_dim_to_check), " but got scale.shape=(", scale.size(0), ", ", scale.size(1), ")");
|
||||
} else {
|
||||
// For MXFP8, 3d tensors have static group sizes (stack of 2d tensors),
|
||||
@ -339,40 +273,32 @@ void _check_scales_blocked(const Tensor& mat, const Tensor& scale, const int dim
|
||||
};
|
||||
|
||||
// TODO: this is for 3d tensor in 2d-3d case specifically.
|
||||
// We'll need to support 3d-3d and 3d-2d cases once mxfp8/nvfp4 grouped gemm supports them.
|
||||
// We'll need to support 3d-3d and 3d-2d cases once mxfp8 grouped gemm supports them.
|
||||
int64_t G = mat.size(0);
|
||||
int64_t K = mat.size(1);
|
||||
if (is_fp4) {
|
||||
// FP4 packs 2 values into a single 8b word - the "real" K is 2x the
|
||||
// reported K. Reverse that adjustment.
|
||||
const int fp4_elems_per_byte = 2;
|
||||
K *= fp4_elems_per_byte;
|
||||
}
|
||||
int64_t N = mat.size(2);
|
||||
int64_t blocked_scale_K = round_up(K/blocksize, 4);
|
||||
int64_t blocked_scale_K = round_up(K/32, 4);
|
||||
int64_t blocked_scale_N = round_up(N, 128);
|
||||
|
||||
// fbgemm expects stack of flattened blocked scales for 3d tensor, shape (G, blocked_scale_K * blocked_scale_N).
|
||||
TORCH_CHECK(
|
||||
scale.dim() == mat.dim() - 1,
|
||||
"for block-scaled 2d-3d grouped GEMM, the 3d tensor of shape (G,K,N) must have a 2d scale of shape (G, blocked_scale_K * blocked_scale_N),",
|
||||
"but scale is ", scale.dim(), "D for arg ", arg_idx
|
||||
"for mxfp8 2d-3d grouped GEMM, the 3d tensor of shape (G,K,N) must have a 2d scale of shape (G, blocked_scale_K * blocked_scale_N), but scale is ", scale.dim(), "D for arg ", arg_idx
|
||||
);
|
||||
TORCH_CHECK(
|
||||
scale.size(0) == G && scale.size(1) == blocked_scale_K * blocked_scale_N,
|
||||
"for block-scaled grouped GEMM, the tensor shape (", G, ", ", K, ", ", N, ") must have scale shape (", G, ",", blocked_scale_K, ",", blocked_scale_N, ")",
|
||||
" for arg ", arg_idx, ", got: ", scale.size(0), ", ", scale.size(1)
|
||||
"for mxfp8, the tensor shape (", G, ", ", K, ", ", N, ") must have scale shape (", G, ",", blocked_scale_K, ",", blocked_scale_N, ") for arg ", arg_idx
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
void check_scale(const Tensor& mat, const Tensor& scale, const int dim, const int arg_idx, const int scale_multiplier=1) {
|
||||
bool using_fp8_rowwise = scale.scalar_type() == kFloat;
|
||||
bool using_mx = scale.scalar_type() == at::kFloat8_e8m0fnu;
|
||||
bool using_mxfp8 = scale.scalar_type() == at::kFloat8_e8m0fnu;
|
||||
if (using_fp8_rowwise) {
|
||||
_check_scales_fp8_rowwise(mat, scale, dim, arg_idx, scale_multiplier);
|
||||
} else if (using_mx) {
|
||||
_check_scales_blocked(mat, scale, dim, arg_idx);
|
||||
} else if (using_mxfp8) {
|
||||
_check_scales_mxfp8(mat, scale, dim, arg_idx);
|
||||
} else {
|
||||
TORCH_CHECK(false, "scale must be float32 or float8_e8m0fnu, but got ", scale.dtype());
|
||||
}
|
||||
@ -485,11 +411,9 @@ namespace {
|
||||
|
||||
using acceptance_fn = std::function<bool(c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&, c10::ScalarType, std::vector<ScalingType>&, ArrayRef<Tensor>&)>;
|
||||
|
||||
std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 4> scale_grouped_kernel_dispatch = {{
|
||||
std::array<std::tuple<std::string, acceptance_fn, ScaledGemmImplementation>, 2> scale_grouped_kernel_dispatch = {{
|
||||
{ "rowwise_rowwise", scaled_blas::check_rowwise_recipe, ScaledGemmImplementation::ROWWISE_ROWWISE},
|
||||
{ "mxfp8_mxfp8", scaled_blas::check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8},
|
||||
{ "mxfp4_mxfp4", scaled_blas::check_mxfp4_recipe, ScaledGemmImplementation::MXFP4_MXFP4},
|
||||
{ "nvfp4_nvfp4", scaled_blas::check_nvfp4_recipe, ScaledGemmImplementation::NVFP4_NVFP4}}};
|
||||
{ "mxfp8_mxfp8", scaled_blas::check_mxfp8_recipe, ScaledGemmImplementation::MXFP8_MXFP8}}};
|
||||
|
||||
} // anonymous namespace
|
||||
|
||||
@ -601,9 +525,8 @@ _scaled_grouped_mm_cuda_v2(
|
||||
out);
|
||||
}
|
||||
case ScaledGemmImplementation::MXFP8_MXFP8: {
|
||||
// scale shape checks
|
||||
_check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
|
||||
_check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
|
||||
_check_scales_mxfp8(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
|
||||
_check_scales_mxfp8(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
|
||||
return _mx8_mx8_bf16_grouped_mm_fbgemm(
|
||||
mat_a,
|
||||
mat_b,
|
||||
@ -614,36 +537,6 @@ _scaled_grouped_mm_cuda_v2(
|
||||
offs.value(),
|
||||
out);
|
||||
}
|
||||
case ScaledGemmImplementation::MXFP4_MXFP4: {
|
||||
// scale shape checks
|
||||
_check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
|
||||
_check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
|
||||
return _f4_f4_bf16_grouped_mm_fbgemm(
|
||||
mat_a,
|
||||
mat_b,
|
||||
scale_a[0], /* block-scale A */
|
||||
std::nullopt, /* global-scale A */
|
||||
scale_b[0], /* block-scale B */
|
||||
std::nullopt, /* global-scale B */
|
||||
offs.value(),
|
||||
std::nullopt, /* bias */
|
||||
out);
|
||||
}
|
||||
case ScaledGemmImplementation::NVFP4_NVFP4: {
|
||||
// scale shape checks
|
||||
_check_scales_blocked(mat_a, scale_a[0], 0 /* dim */, 0 /* arg_idx */);
|
||||
_check_scales_blocked(mat_b, scale_b[0], 1 /* dim */, 1 /* arg_idx */);
|
||||
return _f4_f4_bf16_grouped_mm_fbgemm(
|
||||
mat_a,
|
||||
mat_b,
|
||||
scale_a[0], /* block-scale A */
|
||||
scale_a[1], /* global-scale A */
|
||||
scale_b[0], /* block-scale B */
|
||||
scale_b[1], /* global-scale B */
|
||||
offs.value(),
|
||||
std::nullopt, /* bias */
|
||||
out);
|
||||
}
|
||||
default:
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(false,
|
||||
"_scaled_grouped_mm_cuda_v2 is in an inconsistent state - should never reach here");
|
||||
|
||||
@ -13,7 +13,7 @@ __global__ void vectorized_gather_kernel(char * out, char * inp, index_t * idx,
|
||||
if (allow_neg_indices) {
|
||||
ind = (ind < 0) ? ind + ind_dim_size : ind;
|
||||
}
|
||||
CUDA_KERNEL_ASSERT_VERBOSE(ind >=0 && ind < ind_dim_size && "vectorized gather kernel index out of bounds", "Expected 0 <= index < ind_dim_size(%ld), but got index = %ld", ind_dim_size, ind);
|
||||
CUDA_KERNEL_ASSERT(ind >=0 && ind < ind_dim_size && "vectorized gather kernel index out of bounds");
|
||||
int32_t off = (blockDim.x * blockIdx.y + threadIdx.x) * Alignment; // off is guaranteed to be within int32 limits
|
||||
if (off >= slice_size) return;
|
||||
auto vec = at::native::memory::ld_vec<Alignment>(inp + ind * inp_stride + off);
|
||||
|
||||
@ -59,22 +59,6 @@
|
||||
// forward declare
|
||||
class cublasCommonArgs;
|
||||
|
||||
namespace fbgemm_gpu {
|
||||
|
||||
// NOTE(slayton58): FBGemm_GPU kernels come from <fbgemm_gpu/torch_ops.h> within the FBGemm repo.
|
||||
// To update supported ops means a submodule bump, which is.. painful. Instead, we
|
||||
// can simply forward-declare the methods we want to use.. Works at least as a short-term
|
||||
// thing, but should still be fixed somewhere/somehow.
|
||||
at::Tensor f4f4bf16(
|
||||
at::Tensor,
|
||||
at::Tensor,
|
||||
at::Tensor,
|
||||
at::Tensor,
|
||||
std::optional<at::Tensor>,
|
||||
bool use_mx);
|
||||
|
||||
} // namespace fbgemm_gpu
|
||||
|
||||
using at::blas::ScalingType;
|
||||
using at::blas::SwizzleType;
|
||||
|
||||
@ -810,24 +794,6 @@ void _check_deepseek_scale_stride(const Tensor& scale, const Tensor& t, const Sc
|
||||
}
|
||||
}
|
||||
|
||||
void
|
||||
_check_deepseek_support() {
|
||||
#ifndef USE_ROCM
|
||||
auto dprops = at::cuda::getCurrentDeviceProperties();
|
||||
if (dprops->major != 9) {
|
||||
// Only on Hopper GPUs
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(
|
||||
dprops->major == 9,
|
||||
"DeepSeek style (1x128, 128x128) scaling only supported in CUDA for SM90")
|
||||
}
|
||||
// Only in cublasLt >= 12.9
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(
|
||||
CUBLAS_VERSION < 120900 || cublasLtGetVersion() < 120900,
|
||||
"DeepSeek style (1x128, 128x128) scaling requires cublasLt >= 12.9"
|
||||
);
|
||||
#endif
|
||||
}
|
||||
|
||||
Tensor&
|
||||
_scaled_block1x128_block1x128(
|
||||
const Tensor& mat_a, const Tensor& mat_b,
|
||||
@ -836,12 +802,8 @@ _scaled_block1x128_block1x128(
|
||||
const c10::ScalarType out_dtype,
|
||||
const bool use_fast_accum,
|
||||
Tensor& out) {
|
||||
#ifndef USE_ROCM
|
||||
// Restrictions:
|
||||
// A, B are FP8, scales are fp32, shape K//128
|
||||
// CUDA: Only Hopper GPUs
|
||||
_check_deepseek_support();
|
||||
|
||||
TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
|
||||
mat_a.scalar_type(), mat_b.scalar_type());
|
||||
TORCH_CHECK_VALUE(scale_a.sizes()[0] == mat_a.sizes()[0] && scale_a.sizes()[1] == mat_a.sizes()[1] / 128 && scale_a.scalar_type() == kFloat,
|
||||
@ -859,12 +821,6 @@ _scaled_block1x128_block1x128(
|
||||
_scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);
|
||||
|
||||
return out;
|
||||
#else
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(
|
||||
false,
|
||||
"1x128 and 128x128 scaling not available with ROCm"
|
||||
);
|
||||
#endif
|
||||
}
|
||||
|
||||
Tensor&
|
||||
@ -875,12 +831,10 @@ _scaled_block128x128_block1x128(
|
||||
const c10::ScalarType out_dtype,
|
||||
const bool use_fast_accum,
|
||||
Tensor& out) {
|
||||
#ifndef USE_ROCM
|
||||
// Restrictions:
|
||||
// A, B are FP8, scales are fp32, shape K//128
|
||||
// CUDA: Only Hopper GPUs
|
||||
_check_deepseek_support();
|
||||
|
||||
std::cout << "mat_b: " << mat_b.dim() << ", " << mat_b.sizes() << ", " << mat_b.strides() << std::endl;
|
||||
std::cout << "scale_b: " << scale_b.dim() << ", " << scale_b.sizes() << ", " << scale_b.strides() << std::endl;
|
||||
TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
|
||||
mat_a.scalar_type(), mat_b.scalar_type());
|
||||
TORCH_CHECK_VALUE(scale_a.sizes()[0] == ceil_div<int64_t>(mat_a.sizes()[0], 128) && scale_a.sizes()[1] == ceil_div<int64_t>(mat_a.sizes()[1], 128) && scale_a.scalar_type() == kFloat,
|
||||
@ -898,12 +852,6 @@ _scaled_block128x128_block1x128(
|
||||
_scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);
|
||||
|
||||
return out;
|
||||
#else
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(
|
||||
false,
|
||||
"1x128 and 128x128 scaling not available with ROCm"
|
||||
);
|
||||
#endif
|
||||
}
|
||||
|
||||
Tensor&
|
||||
@ -914,12 +862,8 @@ _scaled_block1x128_block128x128(
|
||||
const c10::ScalarType out_dtype,
|
||||
const bool use_fast_accum,
|
||||
Tensor& out) {
|
||||
#ifndef USE_ROCM
|
||||
// Restrictions:
|
||||
// A, B are FP8, scales are fp32, A: shape K//128, B: K//128, N//128
|
||||
// CUDA: Only Hopper GPUs
|
||||
_check_deepseek_support();
|
||||
|
||||
TORCH_CHECK_VALUE(isFloat8Type(mat_a.scalar_type()) && isFloat8Type(mat_b.scalar_type()), "mat_a and mat_b must be fp8 types, got: ",
|
||||
mat_a.scalar_type(), mat_b.scalar_type());
|
||||
TORCH_CHECK_VALUE(scale_a.sizes()[0] == mat_a.sizes()[0] && scale_a.sizes()[1] == mat_a.sizes()[1] / 128 && scale_a.scalar_type() == kFloat,
|
||||
@ -937,12 +881,6 @@ _scaled_block1x128_block128x128(
|
||||
_scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, use_fast_accum, out);
|
||||
|
||||
return out;
|
||||
#else
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(
|
||||
false,
|
||||
"1x128 and 128x128 scaling not available with ROCm"
|
||||
);
|
||||
#endif
|
||||
}
|
||||
|
||||
Tensor&
|
||||
@ -1013,47 +951,26 @@ _scaled_mxfp4_mxfp4(
|
||||
const std::optional<Tensor>& bias,
|
||||
const c10::ScalarType out_dtype,
|
||||
Tensor& out) {
|
||||
#if !defined(USE_ROCM) && !defined(USE_FBGEMM_GENAI)
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(false, "MXFP4 scaling supported on ROCM and CUDA+FBGEMM_GENAI only");
|
||||
#ifndef USE_ROCM
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(false, "MXFP4 scaling supported on ROCM only");
|
||||
#endif
|
||||
// Restrictions:
|
||||
// A, B are FP4, scales are e8m0, A: shape K//32, B: K, N//32
|
||||
TORCH_CHECK_VALUE(mat_a.scalar_type() == at::kFloat4_e2m1fn_x2 && mat_b.scalar_type() == at::kFloat4_e2m1fn_x2, "mat_a and mat_b must be fp4 types, got: ",
|
||||
mat_a.scalar_type(), mat_b.scalar_type());
|
||||
|
||||
// Packed FP4 format means actual-K = 2 * reported-K -- adjust
|
||||
auto K_multiplier = 2;
|
||||
#ifdef USE_ROCM
|
||||
// AMD
|
||||
auto scale_a_elems = ceil_div<int64_t>(K_multiplier * mat_a.size(0), 32) * mat_a.size(1);
|
||||
auto scale_b_elems = ceil_div<int64_t>(K_multiplier * mat_b.size(1), 32) * mat_b.size(0);
|
||||
#else
|
||||
// NVIDIA
|
||||
auto scale_a_elems = round_up<int64_t>(mat_a.size(0), 128) * round_up<int64_t>(ceil_div<int64_t>(K_multiplier * mat_a.size(1), 32), 4);
|
||||
auto scale_b_elems = round_up<int64_t>(mat_b.size(1), 128) * round_up<int64_t>(ceil_div<int64_t>(K_multiplier * mat_b.size(0), 32), 4);
|
||||
#endif
|
||||
auto scale_a_elems = ceil_div<int64_t>(2 * mat_a.size(0), 32) * mat_a.size(1);
|
||||
auto scale_b_elems = ceil_div<int64_t>(2 * mat_b.size(1), 32) * mat_b.size(0);
|
||||
TORCH_CHECK_VALUE(scale_a_elems == scale_a.numel(),
|
||||
"For Blockwise scaling scale_a should have ", scale_a_elems, " elements, got: ", scale_a.numel());
|
||||
TORCH_CHECK_VALUE(scale_b_elems == scale_b.numel(),
|
||||
"For Blockwise scaling scale_b should have ", scale_b_elems, " elements, got: ", scale_b.numel());
|
||||
|
||||
#ifdef USE_ROCM
|
||||
// AMD
|
||||
TORCH_CHECK_VALUE(swizzle_a == SwizzleType::NO_SWIZZLE, "scale_a must not be swizzled (NO_SWIZZLE format)");
|
||||
TORCH_CHECK_VALUE(swizzle_b == SwizzleType::NO_SWIZZLE, "scale_b must not be swizzled (NO_SWIZZLE format)");
|
||||
#else
|
||||
// NVIDIA
|
||||
TORCH_CHECK_VALUE(swizzle_a == SwizzleType::SWIZZLE_32_4_4, "scale_a must be swizzled to SWIZZLE_32_4_4 format");
|
||||
TORCH_CHECK_VALUE(swizzle_b == SwizzleType::SWIZZLE_32_4_4, "scale_b must be swizzled to SWIZZLE_32_4_4 format");
|
||||
#endif
|
||||
|
||||
TORCH_CHECK_VALUE(scale_a.is_contiguous() && scale_b.is_contiguous(),
|
||||
"For Blockwise scaling both scales should be contiguous");
|
||||
|
||||
TORCH_CHECK_VALUE(out.scalar_type() == out_dtype, "expected out.scalar_type() to be ", out_dtype, ", but got ", out_dtype);
|
||||
|
||||
#ifdef USE_ROCM
|
||||
// AMD
|
||||
auto scaling_choice_a = ScalingType::BlockWise1x32;
|
||||
auto scaling_choice_b = ScalingType::BlockWise1x32;
|
||||
|
||||
@ -1068,29 +985,11 @@ _scaled_mxfp4_mxfp4(
|
||||
TORCH_CHECK_VALUE(out.scalar_type() == ScalarType::BFloat16 ||
|
||||
out.scalar_type() == ScalarType::Half,
|
||||
"Block-wise scaling only supports BFloat16 or Half output types");
|
||||
#else
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(false, "Block-wise scaling for Float8_e8m0fnu requires ROCm 7.0 or later");
|
||||
#endif
|
||||
|
||||
return _scaled_gemm(mat_a, mat_b, scale_a, scale_b, scaling_choice_a, scaling_choice_b, bias, false /* use_fast_accum */, out);
|
||||
#else
|
||||
// NVIDIA
|
||||
// NOTE(slayton58): fbgemm_gpu::f4f4bf16 does *not* allow passing an output tensor,
|
||||
// but we have one we need to use. Two clear options are to copy into
|
||||
// our output (slow), or use a move-assignment-operator (faster).
|
||||
// However, the compiler can complain about the explicit move preventing
|
||||
// copy elision because the return from f4f4bf16 is a temporary object.
|
||||
// So we don't explicitly move, and trust the compiler here...
|
||||
// In the longer term this should be fixed on the FBGemm side.
|
||||
out = fbgemm_gpu::f4f4bf16(
|
||||
mat_a,
|
||||
mat_b.transpose(-2, -1),
|
||||
scale_a,
|
||||
scale_b,
|
||||
std::nullopt, /* global_scale */
|
||||
true /* use_mx */
|
||||
);
|
||||
|
||||
return out;
|
||||
#endif
|
||||
}
|
||||
|
||||
Tensor&
|
||||
@ -1215,20 +1114,17 @@ _scaled_mm_cuda_v2_out(
|
||||
mat_a.size(0), "x", mat_a.size(1), " and ", mat_b.size(0), "x", mat_b.size(1), ")");
|
||||
}
|
||||
|
||||
// Handle fp4 packed-K dimension
|
||||
int K_multiplier = (mat_a.scalar_type() == ScalarType::Float4_e2m1fn_x2) ? 2 : 1;
|
||||
|
||||
TORCH_CHECK_VALUE(!bias || bias->numel() == mat_b.sizes()[1], "Bias must be size ", mat_b.sizes()[1],
|
||||
" but got ", bias->numel());
|
||||
TORCH_CHECK_VALUE(
|
||||
K_multiplier * mat_a.sizes()[1] % 16 == 0,
|
||||
mat_a.sizes()[1] % 16 == 0,
|
||||
"Expected trailing dimension of mat1 to be divisible by 16 ",
|
||||
"but got mat1 shape: (",
|
||||
mat_a.sizes()[0],
|
||||
"x",
|
||||
K_multiplier * mat_a.sizes()[1],
|
||||
mat_a.sizes()[1],
|
||||
").");
|
||||
TORCH_CHECK_VALUE(K_multiplier * mat_b.sizes()[0] % 16 == 0 && mat_b.sizes()[1] % 16 == 0, "mat2 shape (", mat_b.sizes()[0], "x",
|
||||
TORCH_CHECK_VALUE(mat_b.sizes()[0] % 16 == 0 && mat_b.sizes()[1] % 16 == 0, "mat2 shape (", mat_b.sizes()[0], "x",
|
||||
mat_b.sizes()[1], ") must be divisible by 16");
|
||||
|
||||
// TODO(slayton): Existing checks, not sure if they should really be here.
|
||||
|
||||
@ -160,8 +160,8 @@ struct _cuda_scatter_gather_internal_kernel {
|
||||
auto offsets = offset_calc.get(i);
|
||||
|
||||
int64_t idx_dim = *(index_t*)(index_ptr + offsets[2]);
|
||||
CUDA_KERNEL_ASSERT_VERBOSE(idx_dim >= 0 && idx_dim < index_size
|
||||
&& "scatter gather kernel index out of bounds", "Expected 0 <= idx_dim < index_size (%ld), but got idx_dim = %ld", index_size, idx_dim);
|
||||
CUDA_KERNEL_ASSERT(idx_dim >= 0 && idx_dim < index_size
|
||||
&& "scatter gather kernel index out of bounds");
|
||||
|
||||
f(
|
||||
(scalar_t*)(self_ptr + offsets[0]),
|
||||
@ -406,8 +406,9 @@ struct _cuda_scatter_fill_internal_kernel {
|
||||
auto offsets = offset_calc.get(i);
|
||||
|
||||
int64_t idx_dim = *(index_t*)(index_ptr + offsets[1]);
|
||||
CUDA_KERNEL_ASSERT_VERBOSE(idx_dim >= 0 && idx_dim < index_size
|
||||
&& "index out of bounds", "Expected 0 <= idx_dim < index_size (%ld), but got idx_dim = %ld", index_size, idx_dim);
|
||||
CUDA_KERNEL_ASSERT(idx_dim >= 0 && idx_dim < index_size
|
||||
&& "index out of bounds"
|
||||
);
|
||||
|
||||
f(
|
||||
(scalar_t*)(self_ptr + offsets[0]),
|
||||
|
||||
@ -12,15 +12,14 @@
|
||||
|
||||
namespace at::native {
|
||||
|
||||
#if 0 && AT_USE_JITERATOR()
|
||||
#if AT_USE_JITERATOR()
|
||||
constexpr char tan_name[] = "tan_impl";
|
||||
#endif
|
||||
|
||||
void tan_kernel_cuda(TensorIteratorBase& iter) {
|
||||
auto common_dtype = iter.common_dtype();
|
||||
if (at::isComplexType(common_dtype)) {
|
||||
// Disabled due to accuracy issues
|
||||
#if 0 && AT_USE_JITERATOR()
|
||||
#if AT_USE_JITERATOR()
|
||||
static const auto tan_string = jiterator_stringify(
|
||||
template <typename T> T tan_impl(T a) { return std::tan(a); });
|
||||
AT_DISPATCH_COMPLEX_TYPES_AND(
|
||||
|
||||
@ -12,15 +12,14 @@
|
||||
|
||||
namespace at::native {
|
||||
|
||||
#if 0 && AT_USE_JITERATOR()
|
||||
#if AT_USE_JITERATOR()
|
||||
constexpr char tanh_name[] = "tanh_impl";
|
||||
#endif
|
||||
|
||||
void tanh_kernel_cuda(TensorIteratorBase& iter) {
|
||||
auto common_dtype = iter.common_dtype();
|
||||
if (at::isComplexType(common_dtype)) {
|
||||
// Disabled due to accuracy issues
|
||||
#if 0 && AT_USE_JITERATOR()
|
||||
#if AT_USE_JITERATOR()
|
||||
static const auto tanh_string = jiterator_stringify(
|
||||
template <typename T> T tanh_impl(T a) { return std::tanh(a); });
|
||||
AT_DISPATCH_COMPLEX_TYPES_AND(
|
||||
|
||||
@ -141,8 +141,7 @@ WelfordDataLN cuWelfordOnlineSum(
|
||||
if constexpr (!rms_norm){
|
||||
U delta = val - curr_sum.mean;
|
||||
U new_count = curr_sum.count + 1.f;
|
||||
//Due to low CU count, we run into accuracy issues on gfx90a with `__builtin_amdgcn_rcpf`
|
||||
#if defined(USE_ROCM) && !defined(__gfx90a__) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
|
||||
#if defined(USE_ROCM) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
|
||||
U new_mean = curr_sum.mean + delta * __builtin_amdgcn_rcpf(new_count);
|
||||
#else
|
||||
U new_mean = curr_sum.mean + delta * (1.f/new_count); //proper division is slow, this is less accurate but noticeably faster
|
||||
@ -164,8 +163,7 @@ WelfordDataLN cuWelfordCombine(
|
||||
U count = dataA.count + dataB.count;
|
||||
U mean, sigma2;
|
||||
if (count > decltype(dataB.count){0}) {
|
||||
//Due to low CU count, we run into accuracy issues on gfx90a with `__builtin_amdgcn_rcpf`
|
||||
#if defined(USE_ROCM) && !defined(__gfx90a__) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
|
||||
#if defined(USE_ROCM) && defined(USE_LAYERNORM_FAST_RECIPROCAL)
|
||||
auto coef = __builtin_amdgcn_rcpf(count);
|
||||
#else
|
||||
auto coef = 1.f/count; //NB we don't use --use_fast_math, but this is emulation, 1./count goes to intrinsic, `* coef` is multiplication, instead of slow fp division
|
||||
|
||||
@ -617,9 +617,6 @@ Tensor& index_select_out_mps(const Tensor& self, int64_t dim, const Tensor& inde
|
||||
TORCH_CHECK(self.scalar_type() == output.scalar_type(),
|
||||
"index_select(): self and output must have the same scalar type");
|
||||
TORCH_CHECK(dim == 0 || dim < self.dim(), "index_select(): Indexing dim ", dim, " is out of bounds of tensor");
|
||||
at::assert_no_internal_overlap(output);
|
||||
at::assert_no_overlap(output, self);
|
||||
at::assert_no_overlap(output, index);
|
||||
auto output_size = self.sizes().vec();
|
||||
if (self.dim() > 0) {
|
||||
output_size[dim] = num_indices;
|
||||
|
||||
@ -1028,18 +1028,15 @@ TORCH_IMPL_FUNC(prod_out_mps)
|
||||
}
|
||||
|
||||
TORCH_IMPL_FUNC(amax_out_mps)(const Tensor& input_t, IntArrayRef dim, bool keepdim, const Tensor& output_t) {
|
||||
TORCH_CHECK(!c10::isComplexType(input_t.scalar_type()), "amax is not defined for complex types");
|
||||
reduction_out_mps(input_t, dim, keepdim, std::nullopt, output_t, MPSReductionType::AMAX, "amax_out_mps");
|
||||
}
|
||||
|
||||
TORCH_IMPL_FUNC(amin_out_mps)(const Tensor& input_t, IntArrayRef dim, bool keepdim, const Tensor& output_t) {
|
||||
TORCH_CHECK(!c10::isComplexType(input_t.scalar_type()), "amin is not defined for complex types");
|
||||
reduction_out_mps(input_t, dim, keepdim, std::nullopt, output_t, MPSReductionType::AMIN, "amin_out_mps");
|
||||
}
|
||||
|
||||
TORCH_IMPL_FUNC(aminmax_out_mps)
|
||||
(const Tensor& input_t, std::optional<int64_t> dim_opt, bool keepdim, const Tensor& min_t, const Tensor& max_t) {
|
||||
TORCH_CHECK(!c10::isComplexType(input_t.scalar_type()), "aminmax is not defined for complex types");
|
||||
reduction_out_mps(input_t,
|
||||
dim_opt.has_value() ? OptionalIntArrayRef({*dim_opt}) : std::nullopt,
|
||||
keepdim,
|
||||
|
||||
@ -31,7 +31,6 @@ void kthvalue_out_mps_impl(const Tensor& self, int64_t k, int64_t dim, Tensor& v
|
||||
indices.copy_(values.toType(at::ScalarType::Long));
|
||||
return;
|
||||
}
|
||||
TORCH_CHECK_NOT_IMPLEMENTED(!c10::isComplexType(self.scalar_type()), "kthvalue is not implemented for complex types");
|
||||
// issue #154890, raising error to prevent crash within MPSGraph until
|
||||
// workaround is implemented.
|
||||
TORCH_CHECK(self.dim() - dim <= 4, "On-going issue on MPSGraph topk when ndims() - axis > 4, see issue #154890");
|
||||
|
||||
@ -73,7 +73,8 @@ void upsample_bilinear2d_out_frame(
|
||||
const auto rwidth = area_pixel_compute_scale<float>(
|
||||
input_width, output_width, align_corners, scales_w);
|
||||
|
||||
float output_scale = static_cast<float>(output.q_scale() / input.q_scale());
|
||||
// NOLINTNEXTLINE(cppcoreguidelines-narrowing-conversions,bugprone-narrowing-conversions)
|
||||
float output_scale = output.q_scale() / input.q_scale();
|
||||
|
||||
const int64_t input_q_zero_point = input.q_zero_point();
|
||||
const int64_t output_q_zero_point = output.q_zero_point();
|
||||
|
||||
@ -148,7 +148,7 @@ Tensor qcat_nhwc_kernel(
|
||||
// Vectorized loop
|
||||
if (c + VLEN <= curr_C) {
|
||||
auto curr_scale_vec = Vectorized<float>(curr_scale);
|
||||
auto curr_zero_pt_vec = Vectorized<float>(curr_zero_pt);
|
||||
auto curr_zero_pt_vec = Vectorized<float>((float)curr_zero_pt);
|
||||
auto scale_neg_zp_premul = curr_scale_vec * curr_zero_pt_vec.neg();
|
||||
for (; c + VLEN <= curr_C; c += VLEN) {
|
||||
auto inp_vec = Vec::loadu(iptr + c);
|
||||
@ -174,7 +174,7 @@ Tensor qcat_nhwc_kernel(
|
||||
int64_t elem_size = curr_C - c;
|
||||
if ((VLEN == 4 * kVLEN) && elem_size >= kVLEN) {
|
||||
auto curr_scale_vec = Vectorized<float>(curr_scale);
|
||||
auto curr_zero_pt_vec = Vectorized<float>(curr_zero_pt);
|
||||
auto curr_zero_pt_vec = Vectorized<float>((float)curr_zero_pt);
|
||||
auto scale_neg_zp_premul = curr_scale_vec * curr_zero_pt_vec.neg();
|
||||
int64_t vec_num = elem_size / kVLEN;
|
||||
std::array<typename scalar_t::underlying, VLEN> buf_in{};
|
||||
@ -611,10 +611,12 @@ void qrelu_kernel(const Tensor& qx, Tensor& qy) {
|
||||
void leaky_qrelu_out_kernel(Tensor& out, const Tensor& qx,
|
||||
const Scalar& negval_) {
|
||||
int64_t i_zp = qx.q_zero_point();
|
||||
float i_scale = static_cast<float>(qx.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float i_scale = qx.q_scale();
|
||||
|
||||
int64_t o_zp = out.q_zero_point();
|
||||
float o_scale = static_cast<float>(out.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float o_scale = out.q_scale();
|
||||
float o_inv_scale = 1.0f / o_scale;
|
||||
|
||||
float negval = negval_.to<float>();
|
||||
@ -625,8 +627,8 @@ void leaky_qrelu_out_kernel(Tensor& out, const Tensor& qx,
|
||||
Vec zero_vec = Vec(0.0f);
|
||||
Vec one_vec = Vec(1.0f);
|
||||
|
||||
Vec i_scale_vec = Vec(i_scale);
|
||||
Vec i_zp_vec = Vec(i_zp);
|
||||
Vec i_scale_vec = Vec((float)i_scale);
|
||||
Vec i_zp_vec = Vec((float)i_zp);
|
||||
Vec i_scale_zp_neg_premul_vec = i_scale_vec * i_zp_vec.neg();
|
||||
|
||||
Vec negval_vec = Vec(negval);
|
||||
@ -736,9 +738,10 @@ void qprelu_out_kernel(Tensor& out,
|
||||
|
||||
void qgelu_kernel(const Tensor& qx, Tensor& qy, GeluType approximate) {
|
||||
int64_t zero_point = qx.q_zero_point();
|
||||
float scale = static_cast<float>(qx.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float scale = qx.q_scale();
|
||||
auto scale_vec = Vectorized<float>(scale);
|
||||
auto zero_point_vec = Vectorized<float>(zero_point);
|
||||
auto zero_point_vec = Vectorized<float>((float)zero_point);
|
||||
auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
|
||||
int64_t output_zero_point = zero_point;
|
||||
float output_scale = scale;
|
||||
@ -825,9 +828,10 @@ void qgelu_kernel(const Tensor& qx, Tensor& qy, GeluType approximate) {
|
||||
void qsigmoid_kernel(
|
||||
const Tensor& qx, Tensor& qy, double output_scale, int64_t output_zero_point ) {
|
||||
int64_t zero_point = qx.q_zero_point();
|
||||
float scale = static_cast<float>(qx.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float scale = qx.q_scale();
|
||||
auto scale_vec = Vectorized<float>(scale);
|
||||
auto zero_point_vec = Vectorized<float>(zero_point);
|
||||
auto zero_point_vec = Vectorized<float>((float)zero_point);
|
||||
|
||||
AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qsigmoid", [&]() {
|
||||
float inv_output_scale = 1.0 / output_scale;
|
||||
@ -866,9 +870,10 @@ void qsigmoid_kernel(
|
||||
|
||||
void qhardsigmoid_kernel(const Tensor& qx, Tensor& qy) {
|
||||
int64_t zero_point = qx.q_zero_point();
|
||||
float scale = static_cast<float>(qx.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float scale = qx.q_scale();
|
||||
auto scale_vec = Vectorized<float>(scale);
|
||||
auto zero_point_vec = Vectorized<float>(zero_point);
|
||||
auto zero_point_vec = Vectorized<float>((float)zero_point);
|
||||
auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
|
||||
|
||||
AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qhardsigmoid", [&]() {
|
||||
@ -1024,10 +1029,13 @@ void qthreshold_kernel(
|
||||
|
||||
// defines input and output scales and zero_points
|
||||
int64_t input_zero_point = qx.q_zero_point();
|
||||
float input_scale = static_cast<float>(qx.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float input_scale = qx.q_scale();
|
||||
int64_t output_zero_point = qy.q_zero_point();
|
||||
float output_scale = static_cast<float>(qy.q_scale());
|
||||
float inv_output_scale = static_cast<float>(1.0 / output_scale);
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float output_scale = qy.q_scale();
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float inv_output_scale = 1.0 / output_scale;
|
||||
|
||||
AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qthreshold", [&]() {
|
||||
qy = at::_empty_affine_quantized(
|
||||
@ -1088,7 +1096,8 @@ void qhardswish_kernel(const Tensor& qx, Tensor& qy) {
|
||||
|
||||
const auto o_scale = qy.q_scale();
|
||||
const auto o_zero_point = qy.q_zero_point();
|
||||
const float o_inv_scale = static_cast<float>(1.0 / o_scale);
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
const float o_inv_scale = 1.0 / o_scale;
|
||||
|
||||
using fVec = Vectorized<float>;
|
||||
fVec i_scale_vec(i_scale);
|
||||
@ -1126,9 +1135,10 @@ void qhardswish_kernel(const Tensor& qx, Tensor& qy) {
|
||||
|
||||
void qtanh_kernel(const Tensor& qx, Tensor& qy) {
|
||||
int64_t zero_point = qx.q_zero_point();
|
||||
float scale = static_cast<float>(qx.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float scale = qx.q_scale();
|
||||
auto scale_vec = Vectorized<float>(scale);
|
||||
auto zero_point_vec = Vectorized<float>(zero_point);
|
||||
auto zero_point_vec = Vectorized<float>((float)zero_point);
|
||||
auto scale_neg_zp_premul_vec = scale_vec * zero_point_vec.neg();
|
||||
|
||||
AT_DISPATCH_QINT_TYPES(qx.scalar_type(), "qtanh", [&]() {
|
||||
@ -1188,13 +1198,16 @@ void qelu_kernel(
|
||||
// they are NOT related to the quantization scale term
|
||||
|
||||
int64_t i_zp = qx.q_zero_point();
|
||||
float i_scale = static_cast<float>(qx.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float i_scale = qx.q_scale();
|
||||
|
||||
// In a future PR, we can improve on output scale and zero_point
|
||||
// selection.
|
||||
int64_t o_zp = qy.q_zero_point();
|
||||
float o_scale = static_cast<float>(qy.q_scale());
|
||||
float inv_o_scale = static_cast<float>(1.0 / o_scale);
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float o_scale = qy.q_scale();
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float inv_o_scale = 1.0 / o_scale;
|
||||
|
||||
float alpha_float = alpha.to<float>();
|
||||
float scale_coef = scale.to<float>();
|
||||
@ -1214,7 +1227,7 @@ void qelu_kernel(
|
||||
Vec scale_coef_vec = Vec(scale_coef);
|
||||
Vec input_scale_coef_vec = Vec(input_scale_coef);
|
||||
Vec i_scale_vec = Vec(i_scale);
|
||||
Vec i_zero_point_vec = Vec(i_zp);
|
||||
Vec i_zero_point_vec = Vec((float)i_zp);
|
||||
Vec i_scale_neg_zp_premul_vec = i_scale_vec * i_zero_point_vec.neg();
|
||||
|
||||
cpu_kernel_vec(
|
||||
@ -1313,20 +1326,23 @@ void qadd_scalar_kernel(Tensor& out, const Tensor& self, const Scalar& other) {
|
||||
template <bool ReLUFused = false>
|
||||
void qadd_kernel(Tensor& out, const Tensor& self, const Tensor& other) {
|
||||
int64_t zero_point = out.q_zero_point();
|
||||
float scale = static_cast<float>(out.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float scale = out.q_scale();
|
||||
float inv_scale = 1.0f / scale;
|
||||
int64_t self_zero_point = self.q_zero_point();
|
||||
float self_scale = static_cast<float>(self.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float self_scale = self.q_scale();
|
||||
int64_t other_zero_point = other.q_zero_point();
|
||||
float other_scale = static_cast<float>(other.q_scale());
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
float other_scale = other.q_scale();
|
||||
|
||||
// Broadcast out the parameters here to amortize out that cost across
|
||||
// loop iterations.
|
||||
// TODO: we can optimize dequantization by doing a premultiplication
|
||||
// of the zero point by scale and doing FMA on scale*x_q - (scale*zero_point)
|
||||
auto self_zero_point_vec = Vectorized<float>(self_zero_point);
|
||||
auto self_zero_point_vec = Vectorized<float>((float)self_zero_point);
|
||||
auto self_scale_vec = Vectorized<float>(self_scale);
|
||||
auto other_zero_point_vec = Vectorized<float>(other_zero_point);
|
||||
auto other_zero_point_vec = Vectorized<float>((float)other_zero_point);
|
||||
auto other_scale_vec = Vectorized<float>(other_scale);
|
||||
|
||||
auto self_scale_neg_zp_premul_vec = self_scale_vec * self_zero_point_vec.neg();
|
||||
@ -2949,7 +2965,7 @@ void quantized_normalize_kernel(
|
||||
const bool beta_null = beta_data == nullptr;
|
||||
int64_t x_zp = X.q_zero_point();
|
||||
float x_scale = X.q_scale();
|
||||
fVec x_zp_vec(x_zp);
|
||||
fVec x_zp_vec((float)x_zp);
|
||||
fVec one_vec(1.0f);
|
||||
fVec zero_vec(0.0f);
|
||||
float x_fake_scale = 1.0f;
|
||||
@ -3237,7 +3253,7 @@ void quantized_groupnorm_nhwc_kernel(
|
||||
const bool beta_null = beta_data == nullptr;
|
||||
int64_t x_zp = X.q_zero_point();
|
||||
float x_scale = X.q_scale();
|
||||
fVec x_zp_vec(x_zp);
|
||||
fVec x_zp_vec((float)x_zp);
|
||||
fVec one_vec(1.0f);
|
||||
fVec zero_vec(0.0f);
|
||||
float x_fake_scale = 1.0f;
|
||||
|
||||
@ -414,6 +414,7 @@ at::Tensor& PackedLinearWeightFp16::apply_dynamic_impl(
|
||||
TORCH_CHECK(input.size(input.dim() - 1) == packed_weight_fp16.numRows())
|
||||
TORCH_CHECK(input.dim() >= 2);
|
||||
|
||||
// NOLINTNEXTLINE(bugprone-narrowing-conversions,cppcoreguidelines-narrowing-conversions)
|
||||
const int64_t M = size_to_dim_(input.dim() - 1, input.sizes());
|
||||
const int64_t N = packed_weight_fp16.numCols();
|
||||
std::vector<int64_t> output_sizes = input.sizes().vec();
|
||||
|
||||
@ -22,7 +22,6 @@
|
||||
#else
|
||||
#include <ATen/ops/empty.h>
|
||||
#include <ATen/ops/empty_like.h>
|
||||
#include <ATen/ops/zeros_like.h>
|
||||
#include <ATen/ops/reshape.h>
|
||||
#include <ATen/ops/scalar_tensor.h>
|
||||
#include <ATen/ops/sum.h>
|
||||
@ -43,6 +42,7 @@ C10_DIAGNOSTIC_POP()
|
||||
#include <static_switch.h>
|
||||
#include <ATen/native/transformers/cuda/flash_attn/flash_api.h>
|
||||
|
||||
|
||||
#include <c10/util/Exception.h>
|
||||
|
||||
namespace FLASH_NAMESPACE {
|
||||
@ -417,26 +417,6 @@ mha_fwd(const at::Tensor &q, // batch_size x seqlen_q x num_heads x head
|
||||
const int head_size_og = sizes[3];
|
||||
const int seqlen_k = k.size(1);
|
||||
const int num_heads_k = k.size(2);
|
||||
|
||||
if (batch_size == 0) {
|
||||
auto opts = q.options();
|
||||
at::Tensor out = at::empty({0, seqlen_q, num_heads, head_size_og}, opts);
|
||||
at::Tensor q_padded = at::empty({0, seqlen_q, num_heads, head_size_og}, opts);
|
||||
at::Tensor k_padded = at::empty({0, seqlen_k, num_heads_k, head_size_og}, opts);
|
||||
at::Tensor v_padded = at::empty({0, seqlen_k, num_heads_k, head_size_og}, opts);
|
||||
at::Tensor softmax_lse = at::empty({0, num_heads, seqlen_q}, opts.dtype(at::kFloat));
|
||||
at::Tensor rng_state = at::empty({2}, at::dtype(c10::kUInt64).device(at::kCUDA));
|
||||
at::Tensor _unused = at::empty({}, at::dtype(c10::kUInt64).device(at::kCUDA));
|
||||
at::Tensor p = at::empty({0}, opts);
|
||||
if (return_softmax) {
|
||||
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
|
||||
const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
|
||||
const int seqlen_k_rounded = round_multiple(seqlen_k, 128);
|
||||
p = at::empty({0, num_heads, seqlen_q_rounded, seqlen_k_rounded}, opts);
|
||||
}
|
||||
return {std::move(out), std::move(q_padded), std::move(k_padded), std::move(v_padded), std::move(softmax_lse), std::move(rng_state), _unused, std::move(p)};
|
||||
}
|
||||
|
||||
TORCH_CHECK(batch_size > 0, "batch size must be positive");
|
||||
TORCH_CHECK(head_size_og % 8 == 0, "head_size must be a multiple of 8, this is ensured by padding!");
|
||||
TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
|
||||
@ -567,7 +547,7 @@ mha_fwd(const at::Tensor &q, // batch_size x seqlen_q x num_heads x head
|
||||
q_padded = q_padded.transpose(1, 2).reshape({batch_size, 1, num_heads_k * seqlen_q, head_size_og});
|
||||
softmax_lse = softmax_lse.reshape({batch_size, num_heads_k * seqlen_q, 1});
|
||||
}
|
||||
return {std::move(out), std::move(q_padded), std::move(k_padded), std::move(v_padded), std::move(softmax_lse), std::move(rng_state), std::move(_unused), std::move(p)};
|
||||
return {out, q_padded, k_padded, v_padded, softmax_lse, rng_state, _unused, p};
|
||||
}
|
||||
|
||||
std::tuple<at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor, at::Tensor>
|
||||
@ -872,6 +852,7 @@ mha_bwd(const at::Tensor &dout, // batch_size x seqlen_q x num_heads, x head_si
|
||||
TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
|
||||
TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
|
||||
TORCH_CHECK(out.stride(-1) == 1, "out tensor must have contiguous last dimension");
|
||||
TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
|
||||
|
||||
const auto sizes = q.sizes();
|
||||
|
||||
@ -882,20 +863,6 @@ mha_bwd(const at::Tensor &dout, // batch_size x seqlen_q x num_heads, x head_si
|
||||
const int head_size = sizes[3];
|
||||
const int seqlen_k = k.size(1);
|
||||
const int num_heads_k = k.size(2);
|
||||
|
||||
if (batch_size == 0) {
|
||||
auto opts = q.options();
|
||||
at::Tensor dq = at::empty_like(q);
|
||||
at::Tensor dk = at::empty_like(k);
|
||||
at::Tensor dv = at::empty_like(v);
|
||||
auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
|
||||
const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
|
||||
at::Tensor softmax_d = at::empty({0, num_heads, seqlen_q_rounded}, opts.dtype(at::kFloat));
|
||||
return {dq, dk, dv, softmax_d};
|
||||
}
|
||||
|
||||
TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");
|
||||
|
||||
TORCH_CHECK(batch_size > 0, "batch size must be positive");
|
||||
TORCH_CHECK(head_size % 8 == 0, "head_size should be a multiple of 8");
|
||||
TORCH_CHECK(head_size_og % 8 == 0, "head_size_og should be a multiple of 8, this is ensured by padding!");
|
||||
|
||||
@ -1,157 +0,0 @@
|
||||
"""Configuration utilities for parsing JSON and YAML config files."""
|
||||
|
||||
import json
|
||||
import re
|
||||
|
||||
|
||||
def heads_input_type(s: str) -> tuple[int, int]:
|
||||
"""Convert string format 'Hq,Hkv' to tuple (Hq, Hkv)."""
|
||||
try:
|
||||
hq, hkv = map(int, s.split(","))
|
||||
return hq, hkv
|
||||
except Exception as e:
|
||||
raise ValueError("Heads must be Hq,Hkv") from e
|
||||
|
||||
|
||||
default_config = {
|
||||
"dynamic": False,
|
||||
"calculate_bwd": False,
|
||||
"dtype": "bfloat16",
|
||||
"b": [2, 8, 16],
|
||||
"nh": ["16,16", "16,2"],
|
||||
"s": [512, 1024, 4096],
|
||||
"d": [64, 128],
|
||||
"mods": ["noop", "causal", "alibi", "sliding_window"],
|
||||
"backend": ["efficient"],
|
||||
"max_autotune": False,
|
||||
"decoding": False,
|
||||
"kv_size": None,
|
||||
"throughput": True,
|
||||
"save_path": None,
|
||||
"output_json_for_dashboard": None,
|
||||
"benchmark_name": "PyTorch operator microbenchmark",
|
||||
}
|
||||
|
||||
|
||||
def load_config_file(config_path: str) -> dict:
|
||||
"""Load configuration from JSON or YAML file.
|
||||
|
||||
Automatically converts 'nh' field from strings to tuples.
|
||||
|
||||
Args:
|
||||
config_path: Path to the configuration file
|
||||
|
||||
Returns:
|
||||
Dictionary containing the configuration
|
||||
|
||||
Raises:
|
||||
FileNotFoundError: If config file doesn't exist
|
||||
ValueError: If config file format is invalid
|
||||
"""
|
||||
with open(config_path) as f:
|
||||
config_str = f.read()
|
||||
|
||||
# Try to load as JSON first
|
||||
try:
|
||||
config = json.loads(config_str)
|
||||
except json.JSONDecodeError:
|
||||
# Fall back to YAML parsing
|
||||
config = _parse_simple_yaml(config_str)
|
||||
|
||||
# Apply automatic conversions for 'nh' field
|
||||
if "nh" in config and isinstance(config["nh"], list):
|
||||
config["nh"] = [
|
||||
heads_input_type(h) if isinstance(h, str) else h for h in config["nh"]
|
||||
]
|
||||
|
||||
return config
|
||||
|
||||
|
||||
def _parse_simple_yaml(yaml_str: str) -> dict:
|
||||
"""Simple YAML parser for basic configs (without external dependencies).
|
||||
|
||||
Supports:
|
||||
- key: value pairs
|
||||
- booleans (true/false)
|
||||
- null values
|
||||
- integers and floats
|
||||
- strings (quoted and unquoted)
|
||||
- lists in JSON format [item1, item2, ...]
|
||||
- comments (lines starting with # or after #)
|
||||
|
||||
Args:
|
||||
yaml_str: YAML content as string
|
||||
|
||||
Returns:
|
||||
Dictionary containing parsed YAML content
|
||||
"""
|
||||
config = {}
|
||||
|
||||
for line in yaml_str.split("\n"):
|
||||
# Remove comments
|
||||
line = line.split("#")[0].strip()
|
||||
|
||||
if not line or ":" not in line:
|
||||
continue
|
||||
|
||||
key, value = line.split(":", 1)
|
||||
key = key.strip()
|
||||
value = value.strip()
|
||||
|
||||
# Parse value based on type
|
||||
if value.lower() == "true":
|
||||
config[key] = True
|
||||
elif value.lower() == "false":
|
||||
config[key] = False
|
||||
elif value.lower() in ("null", "none", ""):
|
||||
config[key] = None
|
||||
elif value.startswith("[") and value.endswith("]"):
|
||||
# Parse list - handle quoted strings properly
|
||||
pattern = r'"([^"]+)"|\'([^\']+)\'|([^,\[\]\s]+)'
|
||||
matches = re.findall(pattern, value[1:-1]) # Remove [ ]
|
||||
parsed_items = []
|
||||
for match in matches:
|
||||
# match is a tuple of (double_quoted, single_quoted, unquoted)
|
||||
item = match[0] or match[1] or match[2]
|
||||
item = item.strip()
|
||||
if item:
|
||||
try:
|
||||
parsed_items.append(int(item))
|
||||
except ValueError:
|
||||
parsed_items.append(item)
|
||||
config[key] = parsed_items
|
||||
elif value.startswith(('"', "'")):
|
||||
config[key] = value.strip("\"'")
|
||||
else:
|
||||
# Try to parse as number
|
||||
try:
|
||||
config[key] = int(value)
|
||||
except ValueError:
|
||||
try:
|
||||
config[key] = float(value)
|
||||
except ValueError:
|
||||
config[key] = value
|
||||
|
||||
return config
|
||||
|
||||
|
||||
def print_default_config(output_format: str) -> None:
|
||||
"""Print a default configuration template in JSON or YAML format.
|
||||
|
||||
Args:
|
||||
output_format: Either "json" or "yaml"
|
||||
"""
|
||||
if output_format == "json":
|
||||
print(json.dumps(default_config, indent=2))
|
||||
else: # yaml
|
||||
for key, value in default_config.items():
|
||||
if value is None:
|
||||
print(f"{key}: null")
|
||||
elif isinstance(value, bool):
|
||||
print(f"{key}: {str(value).lower()}")
|
||||
elif isinstance(value, str):
|
||||
print(f'{key}: "{value}"')
|
||||
elif isinstance(value, list):
|
||||
print(f"{key}: {json.dumps(value)}")
|
||||
else:
|
||||
print(f"{key}: {value}")
|
||||
@ -1,29 +0,0 @@
|
||||
# Basic benchmark configuration for PyTorch transformer benchmarks
|
||||
# Usage: python score_mod.py --config config_basic.yaml
|
||||
|
||||
# Core parameters
|
||||
dynamic: false
|
||||
calculate_bwd: true
|
||||
dtype: "bfloat16"
|
||||
|
||||
# Shape parameters - larger sweep
|
||||
b: [1, 2, 4, 8, 16] # batch sizes
|
||||
nh: ["16,16", "16,2", "32,32", "32,4"] # [query_heads,key_value_heads]
|
||||
s: [512, 1024, 2048, 4096, 8192] # sequence lengths
|
||||
d: [64, 128] # head dimensions (limited to 128 for Flash Attention/cuDNN compatibility)
|
||||
|
||||
# All attention types
|
||||
mods: ["noop", "causal", "rel", "head_bias", "alibi", "sliding_window", "prefix_lm", "softcap"]
|
||||
|
||||
# Multiple backends for comparison (SDPA + Flash Attention) - flex is always included internally
|
||||
backend: ["efficient", "math", "cudnn", "fav2"]
|
||||
max_autotune: true # Enable torch.compile with max-autotune for optimal performance
|
||||
|
||||
# Decoding and cache settings
|
||||
decoding: false
|
||||
kv_size: null
|
||||
|
||||
# Metrics and output
|
||||
throughput: true # Calculate memory bandwidth & TFLOPS
|
||||
save_path: "comprehensive_results.csv" # Save to CSV
|
||||
output_json_for_dashboard: "attn_bench_basic.json"
|
||||
@ -1,19 +1,15 @@
|
||||
import argparse
|
||||
import csv
|
||||
import gc
|
||||
import itertools
|
||||
import json
|
||||
import random
|
||||
import sys
|
||||
from collections import defaultdict
|
||||
from collections.abc import Callable
|
||||
from contextlib import nullcontext
|
||||
from dataclasses import asdict, dataclass
|
||||
from functools import partial, wraps
|
||||
from typing import Literal, Optional, Union
|
||||
from functools import partial
|
||||
from typing import Optional, Union
|
||||
|
||||
import numpy as np
|
||||
from config_utils import heads_input_type, load_config_file, print_default_config
|
||||
from tabulate import tabulate
|
||||
from tqdm import tqdm
|
||||
|
||||
@ -37,96 +33,6 @@ torch._dynamo.config.recompile_limit = 1000
|
||||
from torch._inductor.runtime.benchmarking import benchmarker
|
||||
|
||||
|
||||
def cleanup_memory():
|
||||
"""Aggressively free GPU memory"""
|
||||
torch.cuda.empty_cache()
|
||||
gc.collect()
|
||||
if torch.cuda.is_available():
|
||||
torch.cuda.synchronize()
|
||||
|
||||
|
||||
def safe_backend(backend_name=None, return_dict=False):
|
||||
"""Decorator that wraps backend functions with error handling
|
||||
|
||||
Args:
|
||||
backend_name: Name of the backend for error messages
|
||||
return_dict: If True, returns dict of results for all backends (for run_single_experiment)
|
||||
If False, returns single ExperimentResults (for individual backend functions)
|
||||
"""
|
||||
|
||||
def decorator(func):
|
||||
@wraps(func)
|
||||
def wrapper(config, *args, **kwargs):
|
||||
try:
|
||||
return func(config, *args, **kwargs)
|
||||
except torch.OutOfMemoryError:
|
||||
print(
|
||||
f"[SKIP] OOM for {backend_name or func.__name__} with shape {config.shape}"
|
||||
)
|
||||
cleanup_memory()
|
||||
except RuntimeError as e:
|
||||
error_msg = str(e)
|
||||
if "out of resource" in error_msg or "OutOfMemoryError" in error_msg:
|
||||
print(
|
||||
f"[SKIP] Triton OOM for {backend_name or func.__name__} with shape {config.shape}"
|
||||
)
|
||||
cleanup_memory()
|
||||
elif "No valid triton configs" in error_msg:
|
||||
print(
|
||||
f"[SKIP] No valid Triton config for {backend_name or func.__name__} with shape {config.shape}"
|
||||
)
|
||||
else:
|
||||
print(
|
||||
f"[SKIP] Runtime error for {backend_name or func.__name__} with shape {config.shape}: {str(e)[:100]}"
|
||||
)
|
||||
except Exception as e:
|
||||
print(
|
||||
f"[SKIP] Error for {backend_name or func.__name__} with shape {config.shape}: {str(e)[:100]}"
|
||||
)
|
||||
|
||||
# Return appropriate NaN result based on function type
|
||||
if return_dict:
|
||||
# For run_single_experiment: return dict with NaN for all backends
|
||||
nan_result = ExperimentResults(
|
||||
fwd_time=float("nan"),
|
||||
bwd_time=float("nan") if config.calculate_bwd_time else None,
|
||||
)
|
||||
results = dict.fromkeys(config.backends, nan_result)
|
||||
results["flex"] = ExperimentResults(
|
||||
fwd_time=float("nan"),
|
||||
bwd_time=float("nan") if config.calculate_bwd_time else None,
|
||||
sparsity=None,
|
||||
)
|
||||
return results
|
||||
else:
|
||||
# For individual backend functions: return single ExperimentResults
|
||||
return ExperimentResults(
|
||||
fwd_time=float("nan"),
|
||||
bwd_time=float("nan") if config.calculate_bwd_time else None,
|
||||
)
|
||||
|
||||
return wrapper
|
||||
|
||||
return decorator
|
||||
|
||||
|
||||
# Type definitions
|
||||
Backend = Literal["math", "efficient", "cudnn", "fav2", "fav3", "fakv", "og-eager"]
|
||||
AttentionType = Literal[
|
||||
"noop",
|
||||
"causal",
|
||||
"rel",
|
||||
"head_bias",
|
||||
"alibi",
|
||||
"sliding_window",
|
||||
"document_mask",
|
||||
"prefix_lm",
|
||||
"softcap",
|
||||
]
|
||||
DtypeString = Literal["bfloat16", "float16", "float32"]
|
||||
SpeedupType = Literal["fwd", "bwd"]
|
||||
|
||||
|
||||
def benchmark_torch_function_in_microseconds(func: Callable, *args, **kwargs) -> float:
|
||||
# warmup
|
||||
for _ in range(5):
|
||||
@ -142,7 +48,6 @@ class ExperimentConfig:
|
||||
calculate_bwd_time: bool
|
||||
cal_bandwidth: bool
|
||||
backends: list[str]
|
||||
max_autotune: bool
|
||||
|
||||
def __post_init__(self):
|
||||
assert len(self.shape) == 6, (
|
||||
@ -157,7 +62,6 @@ class ExperimentConfig:
|
||||
d.pop("cal_bandwidth", None)
|
||||
d["shape(B,Hq,M,Hkv,N,D)"] = d.pop("shape")
|
||||
d.pop("backends", None)
|
||||
d.pop("max_autotune", False)
|
||||
return d
|
||||
|
||||
|
||||
@ -305,7 +209,6 @@ def query_key_value_clones(
|
||||
return query_ref, key_ref, value_ref
|
||||
|
||||
|
||||
@safe_backend("SDPA")
|
||||
def run_single_backend_sdpa(
|
||||
config: ExperimentConfig,
|
||||
query: torch.Tensor,
|
||||
@ -320,7 +223,6 @@ def run_single_backend_sdpa(
|
||||
backend_context = get_backend_context(backend)
|
||||
with backend_context:
|
||||
_device = torch.device("cuda")
|
||||
|
||||
eager_sdpa = generate_eager_sdpa(
|
||||
config.attn_type, config.shape, config.dtype, block_mask, score_mod
|
||||
)
|
||||
@ -388,7 +290,6 @@ def run_single_backend_sdpa(
|
||||
)
|
||||
|
||||
|
||||
@safe_backend("FlashAttention")
|
||||
def run_single_backend_FA(
|
||||
config: ExperimentConfig,
|
||||
query: torch.Tensor,
|
||||
@ -400,9 +301,9 @@ def run_single_backend_FA(
|
||||
mask_kwargs,
|
||||
backend: str,
|
||||
) -> ExperimentResults:
|
||||
assert backend in ["fav3", "fakv"]
|
||||
assert backend in ["fav2", "fav3", "fakv"]
|
||||
# Generate callable for specific backend.
|
||||
if backend in ["fav3"]:
|
||||
if backend in ["fav2", "fav3"]:
|
||||
FA = generate_FA_callable(
|
||||
config.attn_type, config.shape, config.dtype, backend, **mask_kwargs
|
||||
)
|
||||
@ -453,10 +354,10 @@ def run_single_backend_FA(
|
||||
)
|
||||
|
||||
|
||||
@safe_backend("flex_attention", return_dict=True)
|
||||
def run_single_experiment(
|
||||
config: ExperimentConfig,
|
||||
dynamic=False,
|
||||
max_autotune=False,
|
||||
) -> dict[str, ExperimentResults]:
|
||||
device = torch.device("cuda")
|
||||
batch_size, q_heads, q_seq_len, kv_heads, kv_seq_len, head_dim = config.shape
|
||||
@ -476,7 +377,7 @@ def run_single_experiment(
|
||||
block_mask, mask_kwargs = generate_block_mask(config.attn_type, config.shape)
|
||||
kernel_options = get_kernel_options(config.attn_type, config.shape)
|
||||
|
||||
if config.max_autotune:
|
||||
if max_autotune:
|
||||
compiled_sdpa = torch.compile(
|
||||
flex_attention, dynamic=dynamic, mode="max-autotune-no-cudagraphs"
|
||||
)
|
||||
@ -506,7 +407,7 @@ def run_single_experiment(
|
||||
|
||||
results = {}
|
||||
for backend in config.backends:
|
||||
if backend in ["fav3", "fakv"]:
|
||||
if backend in ["fav2", "fav3", "fakv"]:
|
||||
results[backend] = run_single_backend_FA(
|
||||
config,
|
||||
query,
|
||||
@ -518,7 +419,7 @@ def run_single_experiment(
|
||||
mask_kwargs,
|
||||
backend,
|
||||
)
|
||||
else: # sdpa (also supports fav2)
|
||||
else: # sdpa
|
||||
results[backend] = run_single_backend_sdpa(
|
||||
config,
|
||||
query,
|
||||
@ -539,7 +440,7 @@ def run_single_experiment(
|
||||
sparsity = block_mask.sparsity() / 100.0 if block_mask is not None else 0.0
|
||||
sparsity = sparsity if config.attn_type != "document_mask" else 0.5
|
||||
|
||||
results["flex"] = ExperimentResults(
|
||||
results["compiled"] = ExperimentResults(
|
||||
fwd_time=forward_compiled_time,
|
||||
bwd_time=backward_compile_time if config.calculate_bwd_time else None,
|
||||
sparsity=sparsity,
|
||||
@ -600,15 +501,15 @@ def calculate_tflops(config: ExperimentConfig, results: ExperimentResults) -> fl
|
||||
softmax_flops = M * N * 2 # Not counting online softmax overhead
|
||||
o_flops = M * D * N * 2
|
||||
# Not counting split k overhead
|
||||
sparsity = results.sparsity if results.sparsity is not None else 0.0
|
||||
total_flops = B * Hq * (qk_flops + softmax_flops + o_flops) * (1 - sparsity)
|
||||
total_flops = B * Hq * (qk_flops + softmax_flops + o_flops) * (1 - results.sparsity)
|
||||
return total_flops / results.fwd_time / 1e6 # in TFLOPs/
|
||||
|
||||
|
||||
def get_average_speedups(results: list[Experiment], type: str, backend: str):
|
||||
# Calculate speedups
|
||||
speedups = [
|
||||
calculate_speedup(r.results["flex"], r.results[backend], type) for r in results
|
||||
calculate_speedup(r.results["compiled"], r.results[backend], type)
|
||||
for r in results
|
||||
]
|
||||
|
||||
# Find indices of max and min speedups
|
||||
@ -636,7 +537,7 @@ def get_average_speedups(results: list[Experiment], type: str, backend: str):
|
||||
def print_results(results: list[Experiment], save_path: Optional[str] = None):
|
||||
table_data = defaultdict(list)
|
||||
for experiment in results:
|
||||
backends = experiment.config.backends + ["flex"]
|
||||
backends = experiment.config.backends + ["compiled"]
|
||||
for key, value in experiment.asdict().items():
|
||||
if key in backends:
|
||||
if value.fwd_time:
|
||||
@ -649,43 +550,45 @@ def print_results(results: list[Experiment], save_path: Optional[str] = None):
|
||||
# Calculate speedups
|
||||
for backend in results[0].config.backends:
|
||||
fwd_speedups = [
|
||||
calculate_speedup(r.results["flex"], r.results[backend], type="fwd")
|
||||
calculate_speedup(r.results["compiled"], r.results[backend], type="fwd")
|
||||
for r in results
|
||||
]
|
||||
table_data[f"fwd_speedup_flex_over_{backend}"] = fwd_speedups
|
||||
table_data[f"fwd_{backend}_speedup"] = fwd_speedups
|
||||
|
||||
if results[0].config.calculate_bwd_time:
|
||||
for backend in results[0].config.backends:
|
||||
bwd_speedups = [
|
||||
calculate_speedup(r.results["flex"], r.results[backend], type="bwd")
|
||||
calculate_speedup(r.results["compiled"], r.results[backend], type="bwd")
|
||||
for r in results
|
||||
]
|
||||
table_data[f"bwd_speedup_flex_over_{backend}"] = bwd_speedups
|
||||
table_data[f"bwd_{backend}_speedup"] = bwd_speedups
|
||||
|
||||
# Calculate mem + computational throughput
|
||||
if results[0].config.cal_bandwidth:
|
||||
fwd_bandwidth = [
|
||||
calculate_bandwidth(r.config, r.results["flex"], type="fwd")
|
||||
calculate_bandwidth(r.config, r.results["compiled"], type="fwd")
|
||||
for r in results
|
||||
]
|
||||
table_data["fwd_mem_bw (TB/s)"] = fwd_bandwidth
|
||||
fwd_tflops = [calculate_tflops(r.config, r.results["flex"]) for r in results]
|
||||
fwd_tflops = [
|
||||
calculate_tflops(r.config, r.results["compiled"]) for r in results
|
||||
]
|
||||
table_data["TFlops/s"] = fwd_tflops
|
||||
|
||||
print(tabulate(table_data, headers="keys", tablefmt="github", floatfmt=".3f"))
|
||||
|
||||
for backend in results[0].config.backends:
|
||||
if np.isnan(table_data[f"fwd_speedup_flex_over_{backend}"]).all():
|
||||
if np.isnan(table_data[f"fwd_{backend}_speedup"]).all():
|
||||
continue
|
||||
print("\n")
|
||||
print(f"FWD Speedup of Flex over {backend}".center(125, "="))
|
||||
print(f"FWD Speedups vs. {backend}".center(125, "="))
|
||||
print("\n")
|
||||
average_data = get_average_speedups(results, type="fwd", backend=backend)
|
||||
print(tabulate(average_data, headers="keys", tablefmt="github", floatfmt=".3f"))
|
||||
|
||||
if results[0].config.calculate_bwd_time:
|
||||
print("\n")
|
||||
print(f"BWD Speedup of Flex over {backend}".center(125, "="))
|
||||
print(f"BWD Speedups vs. {backend}".center(125, "="))
|
||||
print("\n")
|
||||
average_data = get_average_speedups(results, type="bwd", backend=backend)
|
||||
print(
|
||||
@ -888,14 +791,14 @@ def get_backend_context(backend: str):
|
||||
Returns a context manager for the specified backend.
|
||||
Args:
|
||||
backend (str): The name of the backend to use.
|
||||
Valid options are 'math', 'efficient', 'cudnn', 'fav2', 'fav3', 'fakv', 'og-eager'.
|
||||
Valid options are 'fav2', 'cudnn', 'math', 'efficient', 'fav3', 'fakv', 'og-eager'.
|
||||
Returns:
|
||||
A context manager for the specified backend.
|
||||
Raises:
|
||||
ValueError: If an invalid backend is specified.
|
||||
"""
|
||||
backends = {
|
||||
"fav2": sdpa_kernel(SDPBackend.FLASH_ATTENTION),
|
||||
"fav2": nullcontext(),
|
||||
"cudnn": sdpa_kernel(SDPBackend.CUDNN_ATTENTION),
|
||||
"math": sdpa_kernel(SDPBackend.MATH),
|
||||
"efficient": sdpa_kernel(SDPBackend.EFFICIENT_ATTENTION),
|
||||
@ -917,7 +820,15 @@ def generate_FA_callable(
|
||||
) -> Callable | None:
|
||||
if dtype not in [torch.float16, torch.bfloat16]:
|
||||
return None
|
||||
if backend == "fav3":
|
||||
if backend == "fav2":
|
||||
try:
|
||||
from flash_attn import flash_attn_func, flash_attn_varlen_func
|
||||
except ImportError:
|
||||
print(
|
||||
"Flash attention 2 is not installed. Please install it to run fav2 backend. "
|
||||
)
|
||||
raise
|
||||
elif backend == "fav3":
|
||||
try:
|
||||
from flash_attn.flash_attn_interface import (
|
||||
flash_attn_func,
|
||||
@ -1123,7 +1034,6 @@ def generate_experiment_configs(
|
||||
kv_cache_size: list[int],
|
||||
cal_bandwidth: bool,
|
||||
backends: list[str],
|
||||
max_autotune: bool,
|
||||
) -> list[ExperimentConfig]:
|
||||
assert not (calculate_bwd and decoding), "Decoding does not support backward"
|
||||
|
||||
@ -1167,333 +1077,52 @@ def generate_experiment_configs(
|
||||
calculate_bwd_time=calculate_bwd,
|
||||
cal_bandwidth=cal_bandwidth,
|
||||
backends=backends,
|
||||
max_autotune=max_autotune,
|
||||
)
|
||||
)
|
||||
|
||||
return all_configs
|
||||
|
||||
|
||||
def _output_json_for_dashboard(
|
||||
experiments,
|
||||
output_file,
|
||||
benchmark_name="PyTorch operator microbenchmark",
|
||||
):
|
||||
"""
|
||||
Write the result into JSON format for PyTorch OSS dashboard.
|
||||
The JSON format is defined at
|
||||
https://github.com/pytorch/pytorch/wiki/How-to-integrate-with-PyTorch-OSS-benchmark-database
|
||||
|
||||
Args:
|
||||
experiments: List of experiment results
|
||||
output_file: Path to output JSON file
|
||||
benchmark_name: Name of the benchmark
|
||||
"""
|
||||
if not experiments:
|
||||
return
|
||||
|
||||
import math
|
||||
import platform
|
||||
from dataclasses import asdict, dataclass
|
||||
from typing import Any, Optional
|
||||
|
||||
# Prepare headers and records for JSON output
|
||||
records = []
|
||||
for experiment in experiments:
|
||||
config = experiment.config
|
||||
results_dict = (
|
||||
experiment.results
|
||||
) # This is a dict: backend -> ExperimentResults
|
||||
|
||||
# Process each backend result
|
||||
for backend, results in results_dict.items():
|
||||
# Skip backends that were not run (NaN results)
|
||||
if math.isnan(results.fwd_time):
|
||||
continue
|
||||
|
||||
# Extract data from experiment
|
||||
test_name = f"{backend}_{config.attn_type}_"
|
||||
input_config = f"shape: {config.shape}, dtype: {config.dtype}"
|
||||
|
||||
# Determine mode based on backward pass
|
||||
mode = "training" if config.calculate_bwd_time else "inference"
|
||||
|
||||
# Extract dtype
|
||||
dtype = (
|
||||
str(config.dtype).split(".")[1]
|
||||
if "." in str(config.dtype)
|
||||
else str(config.dtype)
|
||||
)
|
||||
|
||||
# Determine device
|
||||
device = "cuda"
|
||||
|
||||
# Get device architecture
|
||||
device_arch = (
|
||||
torch.cuda.get_device_name(0)
|
||||
if device == "cuda"
|
||||
else platform.processor()
|
||||
if device == "cpu"
|
||||
else "unknown"
|
||||
)
|
||||
|
||||
# Create dataclasses for JSON structure
|
||||
@dataclass
|
||||
class BenchmarkInfo:
|
||||
name: str
|
||||
mode: Optional[str]
|
||||
dtype: str
|
||||
extra_info: dict[str, Any]
|
||||
|
||||
@dataclass
|
||||
class ModelInfo:
|
||||
name: str
|
||||
type: str
|
||||
origins: list[str]
|
||||
extra_info: dict[str, Any]
|
||||
|
||||
@dataclass
|
||||
class MetricInfo:
|
||||
name: str
|
||||
unit: str
|
||||
benchmark_values: list[float]
|
||||
target_value: Optional[float]
|
||||
|
||||
@dataclass
|
||||
class BenchmarkRecord:
|
||||
benchmark: BenchmarkInfo
|
||||
model: ModelInfo
|
||||
metric: MetricInfo
|
||||
|
||||
# Benchmark extra info
|
||||
benchmark_extra_info = {
|
||||
"input_config": input_config,
|
||||
"device": device,
|
||||
"arch": device_arch,
|
||||
"operator_name": backend,
|
||||
"attn_type": config.attn_type,
|
||||
"shape": str(config.shape),
|
||||
"max_autotune": config.max_autotune,
|
||||
}
|
||||
# Add record for forward latency
|
||||
record_fwd_latency = BenchmarkRecord(
|
||||
benchmark=BenchmarkInfo(
|
||||
name=benchmark_name,
|
||||
mode=mode,
|
||||
dtype=dtype,
|
||||
extra_info=benchmark_extra_info,
|
||||
),
|
||||
model=ModelInfo(
|
||||
name=test_name + str(config.shape),
|
||||
type="attention-benchmark",
|
||||
origins=["pytorch"],
|
||||
extra_info={
|
||||
"operator_name": backend,
|
||||
"attn_type": config.attn_type,
|
||||
},
|
||||
),
|
||||
metric=MetricInfo(
|
||||
name="forward latency",
|
||||
unit="us",
|
||||
benchmark_values=[results.fwd_time],
|
||||
target_value=None,
|
||||
),
|
||||
)
|
||||
records.append(asdict(record_fwd_latency))
|
||||
|
||||
# Add record for forward memory bandwidth (if available)
|
||||
if config.cal_bandwidth:
|
||||
record_fwd_bandwidth = BenchmarkRecord(
|
||||
benchmark=BenchmarkInfo(
|
||||
name=benchmark_name,
|
||||
mode=mode,
|
||||
dtype=dtype,
|
||||
extra_info=benchmark_extra_info,
|
||||
),
|
||||
model=ModelInfo(
|
||||
name=test_name + str(config.shape),
|
||||
type="attention-benchmark",
|
||||
origins=["pytorch"],
|
||||
extra_info={
|
||||
"operator_name": backend,
|
||||
},
|
||||
),
|
||||
metric=MetricInfo(
|
||||
name="memory bandwidth",
|
||||
unit="TB/s",
|
||||
benchmark_values=[calculate_bandwidth(config, results, "fwd")],
|
||||
target_value=None,
|
||||
),
|
||||
)
|
||||
records.append(asdict(record_fwd_bandwidth))
|
||||
|
||||
# Add record for forward TFLOPS (if available)
|
||||
if config.cal_bandwidth:
|
||||
record_fwd_tflops = BenchmarkRecord(
|
||||
benchmark=BenchmarkInfo(
|
||||
name=benchmark_name,
|
||||
mode=mode,
|
||||
dtype=dtype,
|
||||
extra_info=benchmark_extra_info,
|
||||
),
|
||||
model=ModelInfo(
|
||||
name=test_name + str(config.shape),
|
||||
type="attention-benchmark",
|
||||
origins=["pytorch"],
|
||||
extra_info={
|
||||
"operator_name": backend,
|
||||
},
|
||||
),
|
||||
metric=MetricInfo(
|
||||
name="tflops",
|
||||
unit="TFLOPS/s",
|
||||
benchmark_values=[calculate_tflops(config, results)],
|
||||
target_value=None,
|
||||
),
|
||||
)
|
||||
records.append(asdict(record_fwd_tflops))
|
||||
|
||||
# Add record for backward latency (if available and not NaN)
|
||||
if (
|
||||
config.calculate_bwd_time
|
||||
and results.bwd_time is not None
|
||||
and not math.isnan(results.bwd_time)
|
||||
):
|
||||
record_bwd_latency = BenchmarkRecord(
|
||||
benchmark=BenchmarkInfo(
|
||||
name=benchmark_name,
|
||||
mode=mode,
|
||||
dtype=dtype,
|
||||
extra_info=benchmark_extra_info,
|
||||
),
|
||||
model=ModelInfo(
|
||||
name=test_name + str(config.shape),
|
||||
type="attention-benchmark",
|
||||
origins=["pytorch"],
|
||||
extra_info={
|
||||
"operator_name": backend,
|
||||
},
|
||||
),
|
||||
metric=MetricInfo(
|
||||
name="backward latency",
|
||||
unit="us",
|
||||
benchmark_values=[results.bwd_time],
|
||||
target_value=None,
|
||||
),
|
||||
)
|
||||
records.append(asdict(record_bwd_latency))
|
||||
|
||||
# Write all records to the output file
|
||||
with open(output_file, "w", encoding="utf-8") as f:
|
||||
json.dump(records, f, indent=2)
|
||||
|
||||
|
||||
def main(
|
||||
dynamic: bool = False,
|
||||
calculate_bwd: bool = False,
|
||||
dtype: DtypeString = "bfloat16",
|
||||
b: list[int] | None = None,
|
||||
nh: list[str] | None = None,
|
||||
s: list[int] | None = None,
|
||||
d: list[int] | None = None,
|
||||
mods: list[AttentionType] | None = None,
|
||||
backend: list[Backend] | None = None,
|
||||
max_autotune: bool = False,
|
||||
decoding: bool = False,
|
||||
kv_size: Optional[list[int]] = None,
|
||||
throughput: bool = True,
|
||||
save_path: Optional[str] = None,
|
||||
output_json_for_dashboard: Optional[str] = None,
|
||||
benchmark_name: str = "PyTorch operator microbenchmark",
|
||||
) -> None:
|
||||
"""Run sweep over sizes and score mods for flex attention.
|
||||
|
||||
Usage Examples:
|
||||
# Use a yml config file
|
||||
python score_mod.py --config basic_config.yaml
|
||||
|
||||
# Use a json config file
|
||||
python score_mod.py --config my_config.json
|
||||
|
||||
# Generate a config template
|
||||
python score_mod.py --print-config json > my_config.json # For a json config
|
||||
python score_mod.py --print-config yaml > my_config.yaml # For a yaml config
|
||||
|
||||
# Override config with CLI args
|
||||
python score_mod.py --config my_config.json -dtype float16 --max-autotune
|
||||
|
||||
# Pure CLI usage
|
||||
python score_mod.py -b 4 8 -s 1024 2048 -mods causal alibi --backend efficient
|
||||
|
||||
Args:
|
||||
dynamic: Runs a dynamic shapes version of compiled flex attention
|
||||
calculate_bwd: Calculate backward pass times
|
||||
dtype: Data type for tensors (bfloat16, float16, float32)
|
||||
b: Batch sizes to benchmark
|
||||
nh: Number of query and key/value heads in format "Hq,Hkv"
|
||||
s: Sequence lengths to benchmark
|
||||
d: Head dimensions to benchmark
|
||||
mods: Score modifications: noop, causal, rel, head_bias, alibi, sliding_window, document_mask, prefix_lm, softcap
|
||||
backend: Backends for attention computation: math, efficient, cudnn, fav2, fav3, fakv, og-eager
|
||||
max_autotune: Turn on max-autotune optimization
|
||||
decoding: Benchmark decoding mode (query sequence length = 1)
|
||||
kv_size: Key/value cache size in MiB (ignores batch size if specified)
|
||||
throughput: Calculate kernel memory bandwidth & computational throughput (always True)
|
||||
save_path: Path to save the results CSV file
|
||||
output_json_for_dashboard: Path to save results in JSON format for PyTorch OSS dashboard
|
||||
benchmark_name: Name of the benchmark for dashboard output
|
||||
"""
|
||||
# Convert dtype string to torch dtype (if not already converted)
|
||||
import torch
|
||||
|
||||
if isinstance(dtype, str):
|
||||
dtype = getattr(torch, dtype)
|
||||
|
||||
# Always calculate throughput
|
||||
throughput = True
|
||||
print("Backend: ", backend)
|
||||
def main(args):
|
||||
seed = 123
|
||||
np.random.seed(seed)
|
||||
torch.manual_seed(seed)
|
||||
results = []
|
||||
for experiment_count, config in enumerate(
|
||||
tqdm(
|
||||
generate_experiment_configs(
|
||||
calculate_bwd,
|
||||
dtype,
|
||||
b,
|
||||
nh,
|
||||
s,
|
||||
d,
|
||||
mods,
|
||||
decoding,
|
||||
kv_size,
|
||||
throughput,
|
||||
backend,
|
||||
max_autotune,
|
||||
)
|
||||
),
|
||||
start=1,
|
||||
for config in tqdm(
|
||||
generate_experiment_configs(
|
||||
args.calculate_bwd,
|
||||
args.dtype,
|
||||
args.b,
|
||||
args.nh,
|
||||
args.s,
|
||||
args.d,
|
||||
args.mods,
|
||||
args.decoding,
|
||||
args.kv_size,
|
||||
args.throughput,
|
||||
args.backend,
|
||||
)
|
||||
):
|
||||
results.append(
|
||||
Experiment(
|
||||
config,
|
||||
run_single_experiment(
|
||||
config,
|
||||
dynamic=dynamic,
|
||||
dynamic=args.dynamic,
|
||||
max_autotune=args.max_autotune,
|
||||
),
|
||||
)
|
||||
)
|
||||
|
||||
# Periodic memory cleanup every 50 experiments
|
||||
if experiment_count % 50 == 0:
|
||||
cleanup_memory()
|
||||
print_results(results, args.save_path)
|
||||
|
||||
print_results(results, save_path)
|
||||
|
||||
# Output JSON for dashboard if requested
|
||||
if output_json_for_dashboard:
|
||||
_output_json_for_dashboard(results, output_json_for_dashboard, benchmark_name)
|
||||
def heads_input_type(s):
|
||||
try:
|
||||
hq, hkv = map(int, s.split(","))
|
||||
return hq, hkv
|
||||
except Exception as e:
|
||||
raise argparse.ArgumentTypeError("Heads must be Hq,Hkv") from e
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
@ -1501,12 +1130,6 @@ if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser(
|
||||
description="Run sweep over sizes and score mods for flex attention"
|
||||
)
|
||||
parser.add_argument(
|
||||
"--config",
|
||||
type=str,
|
||||
help="Path to JSON config file. CLI args override config file values.",
|
||||
default=None,
|
||||
)
|
||||
parser.add_argument(
|
||||
"--dynamic",
|
||||
action="store_true",
|
||||
@ -1576,49 +1199,8 @@ Ignores -b batch size and calculate batch size from kv size instead when specifi
|
||||
default=["efficient"],
|
||||
help="Backend to use for attention computation",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--output-json-for-dashboard",
|
||||
type=str,
|
||||
help="Path to save results in JSON format for PyTorch OSS dashboard",
|
||||
default=None,
|
||||
)
|
||||
parser.add_argument(
|
||||
"--benchmark-name",
|
||||
type=str,
|
||||
help="Name of the benchmark for dashboard output",
|
||||
default="PyTorch operator microbenchmark",
|
||||
)
|
||||
parser.add_argument(
|
||||
"--print-config",
|
||||
type=str,
|
||||
choices=["json", "yaml"],
|
||||
help="Print a default config template in JSON or YAML format and exit",
|
||||
default=None,
|
||||
)
|
||||
# Parse arguments
|
||||
args = parser.parse_args()
|
||||
args.dtype = getattr(torch, args.dtype)
|
||||
|
||||
# Handle --print-config
|
||||
if args.print_config:
|
||||
print_default_config(args.print_config)
|
||||
sys.exit(0)
|
||||
|
||||
# Load and merge config if provided
|
||||
if args.config:
|
||||
config = load_config_file(args.config)
|
||||
|
||||
# Merge config with CLI args (CLI args take precedence)
|
||||
json_args = argparse.Namespace()
|
||||
json_args.__dict__ = config
|
||||
args = parser.parse_args(namespace=json_args)
|
||||
|
||||
# Convert dtype string to torch dtype (only if it's still a string)
|
||||
if isinstance(args.dtype, str):
|
||||
args.dtype = getattr(torch, args.dtype)
|
||||
|
||||
# Remove config and print_config from args before passing to main
|
||||
args_dict = vars(args)
|
||||
args_dict.pop("config", None)
|
||||
args_dict.pop("print_config", None)
|
||||
|
||||
main(**args_dict)
|
||||
main(args)
|
||||
|
||||
@ -482,7 +482,6 @@ inductor_core_resources = [
|
||||
"torch/csrc/inductor/aoti_torch/oss_proxy_executor.cpp",
|
||||
"torch/csrc/inductor/inductor_ops.cpp",
|
||||
"torch/csrc/jit/serialization/pickle.cpp",
|
||||
"torch/csrc/shim_common.cpp",
|
||||
]
|
||||
|
||||
libtorch_core_sources = sorted(
|
||||
|
||||
@ -556,26 +556,3 @@ inline SymBool sym_ge(const SymInt& a, const SymInt& b) {
|
||||
}
|
||||
|
||||
} // namespace c10
|
||||
|
||||
#include <limits>
|
||||
|
||||
namespace std {
|
||||
|
||||
template <>
|
||||
class numeric_limits<c10::SymInt> {
|
||||
public:
|
||||
static constexpr bool is_specialized = true;
|
||||
|
||||
static constexpr int64_t max() noexcept {
|
||||
return std::numeric_limits<int64_t>::max();
|
||||
}
|
||||
|
||||
static constexpr int64_t min() noexcept {
|
||||
return std::numeric_limits<int64_t>::min();
|
||||
}
|
||||
|
||||
static constexpr bool is_signed = true;
|
||||
static constexpr bool is_integer = true;
|
||||
};
|
||||
|
||||
} // namespace std
|
||||
|
||||
@ -1,4 +1,4 @@
|
||||
// Implementation of special math functions for Metal
|
||||
// Implementation of specal math functions for Metal
|
||||
#pragma once
|
||||
#include <c10/metal/expm1f.h>
|
||||
#include <c10/metal/igamma.h>
|
||||
|
||||
@ -34,7 +34,7 @@ struct MemEvent {
|
||||
bool overlaps(const MemBlock& a, const MemBlock& b) {
|
||||
// two blocks dont overlap if
|
||||
// |---a--------|--------------b--------|
|
||||
// start_a end_a <= start_b end_b
|
||||
// strat_a end_a <= start_b end_b
|
||||
return !(
|
||||
(a.end_offset <= b.start_offset) || (b.end_offset <= a.start_offset));
|
||||
}
|
||||
|
||||
@ -33,7 +33,7 @@ struct bitset final {
|
||||
constexpr bitset() noexcept = default;
|
||||
constexpr bitset(const bitset&) noexcept = default;
|
||||
constexpr bitset(bitset&&) noexcept = default;
|
||||
// there is an issue for gcc 5.3.0 when define default function as constexpr
|
||||
// there is an issure for gcc 5.3.0 when define default function as constexpr
|
||||
// see https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68754.
|
||||
bitset& operator=(const bitset&) noexcept = default;
|
||||
bitset& operator=(bitset&&) noexcept = default;
|
||||
|
||||
@ -123,8 +123,6 @@ class DeviceCachingAllocator {
|
||||
ska::flat_hash_map<xpu::XPUStream, std::deque<std::pair<sycl::event, Block*>>>
|
||||
xpu_events;
|
||||
DeviceIndex device_index;
|
||||
size_t allowed_memory_maximum = 0;
|
||||
bool set_fraction = false;
|
||||
|
||||
size_t try_merge_blocks(Block* dst, Block* src, BlockPool& pool) {
|
||||
if (!src || src->allocated || src->event_count > 0 ||
|
||||
@ -247,12 +245,6 @@ class DeviceCachingAllocator {
|
||||
if (isRetry) {
|
||||
stats.num_alloc_retries += 1;
|
||||
}
|
||||
if (set_fraction &&
|
||||
stats.reserved_bytes[static_cast<size_t>(StatType::AGGREGATE)].current +
|
||||
size >
|
||||
allowed_memory_maximum) {
|
||||
return false;
|
||||
}
|
||||
void* ptr = sycl::aligned_alloc_device(
|
||||
kDeviceAlignment,
|
||||
size,
|
||||
@ -443,11 +435,6 @@ class DeviceCachingAllocator {
|
||||
device_free =
|
||||
raw_device.get_info<sycl::ext::intel::info::device::free_memory>();
|
||||
}
|
||||
std::string allowed_info;
|
||||
if (set_fraction) {
|
||||
allowed_info = format_size(allowed_memory_maximum) + " allowed; ";
|
||||
}
|
||||
|
||||
auto allocated_bytes =
|
||||
stats.allocated_bytes[static_cast<size_t>(StatType::AGGREGATE)]
|
||||
.current;
|
||||
@ -472,9 +459,7 @@ class DeviceCachingAllocator {
|
||||
format_size(device_total),
|
||||
" of which ",
|
||||
format_size(device_free),
|
||||
" is free. ",
|
||||
allowed_info,
|
||||
"Of the allocated memory ",
|
||||
" is free. Of the allocated memory ",
|
||||
format_size(allocated_bytes),
|
||||
" is allocated by PyTorch, and ",
|
||||
format_size(reserved_bytes - allocated_bytes),
|
||||
@ -553,25 +538,6 @@ class DeviceCachingAllocator {
|
||||
stats.requested_bytes[statType].reset_peak();
|
||||
}
|
||||
}
|
||||
|
||||
double getMemoryFraction() {
|
||||
if (!set_fraction) {
|
||||
return 1.0;
|
||||
}
|
||||
|
||||
c10::xpu::DeviceProp device_prop;
|
||||
c10::xpu::get_device_properties(&device_prop, device_index);
|
||||
return static_cast<double>(allowed_memory_maximum) /
|
||||
static_cast<double>(device_prop.global_mem_size);
|
||||
}
|
||||
|
||||
void setMemoryFraction(double fraction) {
|
||||
c10::xpu::DeviceProp device_prop;
|
||||
c10::xpu::get_device_properties(&device_prop, device_index);
|
||||
auto device_total = device_prop.global_mem_size;
|
||||
allowed_memory_maximum = static_cast<size_t>(fraction * device_total);
|
||||
set_fraction = true;
|
||||
}
|
||||
};
|
||||
|
||||
static void local_raw_delete(void* ptr);
|
||||
@ -734,21 +700,6 @@ class XPUAllocator : public DeviceAllocator {
|
||||
assertValidDevice(device);
|
||||
device_allocators[device]->resetAccumulatedStats();
|
||||
}
|
||||
|
||||
double getMemoryFraction(DeviceIndex device) {
|
||||
assertValidDevice(device);
|
||||
return device_allocators[device]->getMemoryFraction();
|
||||
}
|
||||
|
||||
void setMemoryFraction(double fraction, DeviceIndex device) {
|
||||
assertValidDevice(device);
|
||||
TORCH_CHECK_VALUE(
|
||||
0 < fraction && fraction <= 1,
|
||||
"invalid fraction:",
|
||||
fraction,
|
||||
". Please set within (0, 1].");
|
||||
device_allocators[device]->setMemoryFraction(fraction);
|
||||
}
|
||||
};
|
||||
|
||||
static XPUAllocator allocator;
|
||||
@ -793,14 +744,6 @@ void recordStream(const DataPtr& dataPtr, XPUStream stream) {
|
||||
return allocator.recordStream(dataPtr, stream);
|
||||
}
|
||||
|
||||
double getMemoryFraction(DeviceIndex device) {
|
||||
return allocator.getMemoryFraction(device);
|
||||
}
|
||||
|
||||
void setMemoryFraction(double fraction, DeviceIndex device) {
|
||||
return allocator.setMemoryFraction(fraction, device);
|
||||
}
|
||||
|
||||
REGISTER_ALLOCATOR(kXPU, &allocator)
|
||||
|
||||
} // namespace c10::xpu::XPUCachingAllocator
|
||||
|
||||
@ -25,8 +25,4 @@ C10_XPU_API void raw_delete(void* ptr);
|
||||
|
||||
C10_XPU_API void recordStream(const DataPtr& dataPtr, XPUStream stream);
|
||||
|
||||
C10_XPU_API double getMemoryFraction(DeviceIndex device);
|
||||
|
||||
C10_XPU_API void setMemoryFraction(double fraction, DeviceIndex device);
|
||||
|
||||
} // namespace c10::xpu::XPUCachingAllocator
|
||||
|
||||
@ -38,7 +38,7 @@ uint32_t crc32_combine (uint32_t crcA, uint32_t crcB, size_t lengthB);
|
||||
|
||||
/// compute CRC32 (bitwise algorithm)
|
||||
uint32_t crc32_bitwise (const void* data, size_t length, uint32_t previousCrc32 = 0);
|
||||
/// compute CRC32 (half-byte algorithm)
|
||||
/// compute CRC32 (half-byte algoritm)
|
||||
uint32_t crc32_halfbyte(const void* data, size_t length, uint32_t previousCrc32 = 0);
|
||||
|
||||
#ifdef CRC32_USE_LOOKUP_TABLE_BYTE
|
||||
@ -96,7 +96,7 @@ uint32_t crc32_16bytes_prefetch(const void* data, size_t length, uint32_t previo
|
||||
#define __BIG_ENDIAN 4321
|
||||
#endif
|
||||
|
||||
// define endianness and some integer data types
|
||||
// define endianess and some integer data types
|
||||
#if defined(_MSC_VER) || defined(__MINGW32__)
|
||||
// Windows always little endian
|
||||
#define __BYTE_ORDER __LITTLE_ENDIAN
|
||||
@ -168,7 +168,7 @@ namespace
|
||||
/// zlib's CRC32 polynomial
|
||||
const uint32_t Polynomial = 0xEDB88320;
|
||||
|
||||
/// swap endianness
|
||||
/// swap endianess
|
||||
static inline uint32_t swap(uint32_t x)
|
||||
{
|
||||
#if defined(__GNUC__) || defined(__clang__)
|
||||
@ -229,7 +229,7 @@ uint32_t crc32_bitwise(const void* data, size_t length, uint32_t previousCrc32)
|
||||
}
|
||||
|
||||
|
||||
/// compute CRC32 (half-byte algorithm)
|
||||
/// compute CRC32 (half-byte algoritm)
|
||||
uint32_t crc32_halfbyte(const void* data, size_t length, uint32_t previousCrc32)
|
||||
{
|
||||
uint32_t crc = ~previousCrc32; // same as previousCrc32 ^ 0xFFFFFFFF
|
||||
@ -662,7 +662,7 @@ uint32_t crc32_combine(uint32_t crcA, uint32_t crcB, size_t lengthB)
|
||||
// - if you append length(B) zeros to A and call it A' (think of it as AAAA000)
|
||||
// and prepend length(A) zeros to B and call it B' (think of it as 0000BBB)
|
||||
// then exists a C' = A' ^ B'
|
||||
// - remember: if you XOR something with zero, it remains unchanged: X ^ 0 = X
|
||||
// - remember: if you XOR someting with zero, it remains unchanged: X ^ 0 = X
|
||||
// - that means C' = A concat B so that crc(A concat B) = crc(C') = crc(A') ^ crc(B')
|
||||
// - the trick is to compute crc(A') based on crc(A)
|
||||
// and crc(B') based on crc(B)
|
||||
|
||||
@ -76,7 +76,7 @@ typedef struct mz_zip_archive mz_zip_archive;
|
||||
// 2) Writing with 1-pass sequential access
|
||||
// -> We must take care not to require updating values that have already
|
||||
// been written. We place the variable-length index at the end and do
|
||||
// not put any index into the header to fulfill this constraint.
|
||||
// not put any indicies into the header to fulfill this constraint.
|
||||
|
||||
// The model.json, which contains all the metadata information,
|
||||
// should be written as the last file. One reason is that the size of tensor
|
||||
|
||||
@ -519,7 +519,7 @@ TEST(PyTorchStreamWriterAndReader, SaveAndLoadWithAllocator) {
|
||||
std::tie(data_ptr, size) = reader.getRecord("key1", &overrideAllocator);
|
||||
EXPECT_EQ(overrideAllocator.getAllocatedBytes(), kBytes1);
|
||||
EXPECT_EQ(baseAllocator.getAllocatedBytes(), allocBytes);
|
||||
// allocate with base allocator
|
||||
// allcoate with base allocator
|
||||
std::tie(data_ptr, size) = reader.getRecord("key1");
|
||||
EXPECT_EQ(overrideAllocator.getAllocatedBytes(), kBytes1);
|
||||
EXPECT_EQ(baseAllocator.getAllocatedBytes(), allocBytes + kBytes1);
|
||||
|
||||
@ -383,7 +383,7 @@ function(torch_compile_options libname)
|
||||
-Wno-strict-aliasing
|
||||
)
|
||||
if(CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
|
||||
list(APPEND private_compile_options -Wredundant-move -Wno-interference-size)
|
||||
list(APPEND private_compile_options -Wredundant-move)
|
||||
endif()
|
||||
if(CMAKE_CXX_COMPILER_ID MATCHES "Clang")
|
||||
list(APPEND private_compile_options -Wextra-semi -Wmove)
|
||||
|
||||
@ -14,7 +14,7 @@ Combining, these building blocks form a research and
|
||||
production ready C++ library for tensor computation and dynamic neural
|
||||
networks with strong emphasis on GPU acceleration as well as fast CPU
|
||||
performance. It is currently in use at Facebook in research and
|
||||
production; we are looking forward to welcoming more users of the PyTorch C++ API.
|
||||
production; we are looking forward to welcome more users of the PyTorch C++ API.
|
||||
|
||||
.. warning::
|
||||
|
||||
|
||||
@ -64,7 +64,7 @@ users should pay additional attention to:
|
||||
|
||||
- Both guards affects tensor execution process to skip work not related to inference, but ``InferenceMode``
|
||||
also affects tensor creation while ``AutoNonVariableTypeMode`` doesn't. In other words, tensors created
|
||||
inside ``InferenceMode`` are marked as inference tensors so that certain limitations can be applied after
|
||||
inside ``InferenceMode`` are marked as inference tensors so that certain limitation can be applied after
|
||||
exiting ``InferenceMode``.
|
||||
- Enabled/disabled ``InferenceMode`` states can be nested while ``AutoNonVariableTypeMode`` only allows enabled state.
|
||||
|
||||
|
||||
@ -17,7 +17,7 @@ restoring the RNG state during each checkpoint.
|
||||
The stashing logic saves and restores the RNG state for CPU and another
|
||||
device type (infer the device type from Tensor arguments excluding CPU
|
||||
tensors by `_infer_device_type`) to the `run_fn`. If there are multiple
|
||||
devices, device state will only be saved for devices of a single device type,
|
||||
device, device state will only be saved for devices of a single device type,
|
||||
and the remaining devices will be ignored. Consequently, if any checkpointed
|
||||
functions involve randomness, this may result in incorrect gradients. (Note
|
||||
that if CUDA devices are among the devices detected, it will be prioritized;
|
||||
|
||||
@ -1066,8 +1066,6 @@ coverage_ignore_functions = [
|
||||
"set_current_meta",
|
||||
"set_grad_fn_seq_nr",
|
||||
"set_stack_trace",
|
||||
"set_current_replay_node",
|
||||
"get_current_replay_node",
|
||||
# torch.jit.annotations
|
||||
"ann_to_type",
|
||||
"check_fn",
|
||||
|
||||
@ -59,14 +59,14 @@ MPI supports CUDA only if the implementation used to build PyTorch supports it.
|
||||
|
||||
### Backends that come with PyTorch
|
||||
|
||||
PyTorch distributed package supports Linux (stable), macOS (stable), and Windows (prototype).
|
||||
PyTorch distributed package supports Linux (stable), MacOS (stable), and Windows (prototype).
|
||||
By default for Linux, the Gloo and NCCL backends are built and included in PyTorch
|
||||
distributed (NCCL only when building with CUDA). MPI is an optional backend that can only be
|
||||
included if you build PyTorch from source. (e.g. building PyTorch on a host that has MPI
|
||||
installed.)
|
||||
|
||||
:::{note}
|
||||
As of PyTorch v1.8, Windows supports all collective communications backends but NCCL,
|
||||
As of PyTorch v1.8, Windows supports all collective communications backend but NCCL,
|
||||
If the `init_method` argument of {func}`init_process_group` points to a file it must adhere
|
||||
to the following schema:
|
||||
|
||||
|
||||
@ -99,12 +99,6 @@ DTensor supports the following types of {class}`Placement` on each {class}`Devic
|
||||
:undoc-members:
|
||||
```
|
||||
|
||||
```{eval-rst}
|
||||
.. autoclass:: MaskPartial
|
||||
:members:
|
||||
:undoc-members:
|
||||
```
|
||||
|
||||
```{eval-rst}
|
||||
.. autoclass:: Placement
|
||||
:members:
|
||||
|
||||
@ -1,6 +1,6 @@
|
||||
# torch.mtia
|
||||
|
||||
The MTIA backend is implemented out of the tree, only interfaces are defined here.
|
||||
The MTIA backend is implemented out of the tree, only interfaces are be defined here.
|
||||
|
||||
```{eval-rst}
|
||||
.. automodule:: torch.mtia
|
||||
|
||||
@ -1,6 +1,6 @@
|
||||
# torch.mtia.memory
|
||||
|
||||
The MTIA backend is implemented out of the tree, only interfaces are defined here.
|
||||
The MTIA backend is implemented out of the tree, only interfaces are be defined here.
|
||||
|
||||
```{eval-rst}
|
||||
.. automodule:: torch.mtia.memory
|
||||
|
||||
@ -122,38 +122,12 @@ The above is relevant in two places:
|
||||
}
|
||||
```
|
||||
|
||||
2. `torch_call_dispatcher`
|
||||
2. `aoti_torch_call_dispatcher`
|
||||
This API allows you to call the PyTorch dispatcher from C/C++ code. It has the following signature:
|
||||
|
||||
```cpp
|
||||
torch_call_dispatcher(const char* opName, const char* overloadName, StableIValue* stack, uint64_t extension_build_version);
|
||||
aoti_torch_call_dispatcher(const char* opName, const char* overloadName, StableIValue* stack);
|
||||
```
|
||||
|
||||
`torch_call_dispatcher` will call the op overload defined by a given `opName`, `overloadName`, a stack of
|
||||
StableIValues and the `TORCH_ABI_VERSION` of the user extension. This call will populate any return values of the
|
||||
op into the stack in their StableIValue form, with `ret0` at index 0, `ret1` at index 1, and so on.
|
||||
|
||||
We caution against using this API to call functions that have been registered to the dispatcher by other extensions
|
||||
unless the caller can guarantee that the signature they expect matches that which the custom extension has
|
||||
registered.
|
||||
|
||||
### Versioning and Forward/Backward compatibility guarantees
|
||||
|
||||
We provide a `TORCH_ABI_VERSION` macro in `torch/headeronly/version.h` of the form
|
||||
|
||||
```
|
||||
[ byte ][ byte ][ byte ][ byte ][ byte ][ byte ][ byte ][ byte ]
|
||||
[MAJ ][ MIN ][PATCH ][ ABI TAG ]
|
||||
```
|
||||
|
||||
In the present phase of development, APIs in the C-shim will be versioned based on major.minor.patch release that they are first introduced in, with 2.10 being the first release where this will be enforced. The ABI tag is reserved for future use.
|
||||
|
||||
Extensions can select the minimum abi version to be compatible with using:
|
||||
|
||||
```
|
||||
#define TORCH_TARGET_VERSION (((0ULL + major) << 56) | ((0ULL + minor) << 48))
|
||||
```
|
||||
|
||||
before including any stable headers or by passing the equivalent `-D` option to the compiler. Otherwise, the default will be the current `TORCH_ABI_VERSION`.
|
||||
|
||||
The above ensures that if a user defines `TORCH_TARGET_VERSION` to be 0x0209000000000000 (2.9) and attempts to use a C shim API `foo` that was introduced in version 2.10, a compilation error will be raised. Similarly, the C++ wrapper APIs in `torch/csrc/stable` are compatible with older libtorch binaries up to the TORCH_ABI_VERSION they are exposed in and forward compatible with newer libtorch binaries.
|
||||
`aoti_torch_call_dispatcher` will call the op overload defined by a given `opName`, `overloadName`, and a stack of
|
||||
StableIValues. This call will populate any return values of the op into the stack in their StableIValue form,
|
||||
with `ret0` at index 0, `ret1` at index 1, and so on.
|
||||
|
||||
@ -76,7 +76,6 @@
|
||||
:nosignatures:
|
||||
|
||||
empty_cache
|
||||
get_per_process_memory_fraction
|
||||
max_memory_allocated
|
||||
max_memory_reserved
|
||||
mem_get_info
|
||||
@ -86,7 +85,6 @@
|
||||
memory_stats_as_nested_dict
|
||||
reset_accumulated_memory_stats
|
||||
reset_peak_memory_stats
|
||||
set_per_process_memory_fraction
|
||||
```
|
||||
|
||||
```{eval-rst}
|
||||
|
||||
2
setup.py
2
setup.py
@ -1106,7 +1106,7 @@ class build_ext(setuptools.command.build_ext.build_ext):
|
||||
continue
|
||||
self.copy_file(source_lib, target_lib)
|
||||
# Delete old rpath and add @loader_lib to the rpath
|
||||
# This should prevent deallocate from attempting to package another instance
|
||||
# This should prevent delocate from attempting to package another instance
|
||||
# of OpenMP library in torch wheel as well as loading two libomp.dylib into
|
||||
# the address space, as libraries are cached by their unresolved names
|
||||
install_name_tool_args = [
|
||||
|
||||
@ -238,7 +238,7 @@ def pytest_pycollect_makemodule(module_path, path, parent) -> Module:
|
||||
|
||||
@pytest.hookimpl(hookwrapper=True)
|
||||
def pytest_report_teststatus(report, config):
|
||||
# Add the test time to the verbose output, unfortunately I don't think this
|
||||
# Add the test time to the verbose output, unforunately I don't think this
|
||||
# includes setup or teardown
|
||||
pluggy_result = yield
|
||||
if not isinstance(report, pytest.TestReport):
|
||||
|
||||
@ -584,7 +584,7 @@ TEST(CustomAutogradTest, MarkDirty) {
|
||||
}
|
||||
};
|
||||
|
||||
// Clone here because modifying leaves inplace is not allowed
|
||||
// Clone here because modifying leafs inplace is not allowed
|
||||
auto x = torch::randn({5, 5}, torch::requires_grad()).clone();
|
||||
auto version_before = x._version();
|
||||
auto out = MyFunction::apply(x);
|
||||
|
||||
@ -264,7 +264,7 @@ TEST_F(ParallelTest, DataParallelNumericalEquivalence_MultiCUDA) {
|
||||
input += i;
|
||||
input_dp += i;
|
||||
|
||||
// non-parallel training
|
||||
// non-prallel training
|
||||
torch::optim::SGD optim(model->parameters(), torch::optim::SGDOptions(0.1));
|
||||
auto output = model->forward(input);
|
||||
auto loss = torch::mse_loss(output, torch::zeros_like(output));
|
||||
|
||||
@ -149,8 +149,8 @@ When `import torch`, installed accelerators (such as `torch_openreg`) will be au
|
||||
### Installation
|
||||
|
||||
```python
|
||||
python -m pip install --no-build-isolation -e . # for develop
|
||||
python -m pip install --no-build-isolation . # for install
|
||||
pip3 install --no-build-isolation -e . # for develop
|
||||
pip3 install --no-build-isolation . # for install
|
||||
```
|
||||
|
||||
### Usage Example
|
||||
@ -188,7 +188,7 @@ Please refer to [this](https://docs.pytorch.org/docs/main/accelerator/index.html
|
||||
- Device-agnostic APIs
|
||||
- Memory Management
|
||||
- Generator
|
||||
- Distributed
|
||||
- Distrubuted
|
||||
- Custom Tensor&Storage
|
||||
- ...
|
||||
- **Improve Tests**: Add more test cases related to the integration mechanism.
|
||||
|
||||
@ -5,7 +5,6 @@ from torch.testing._internal.common_utils import run_tests, skipIfTorchDynamo, T
|
||||
|
||||
|
||||
class TestStream(TestCase):
|
||||
@skipIfTorchDynamo()
|
||||
def test_stream_create(self):
|
||||
stream = torch.Stream(device="openreg")
|
||||
self.assertEqual(stream.device_index, torch.openreg.current_device())
|
||||
@ -25,7 +24,6 @@ class TestStream(TestCase):
|
||||
)
|
||||
self.assertEqual(stream, stream1)
|
||||
|
||||
@skipIfTorchDynamo()
|
||||
def test_stream_context(self):
|
||||
with torch.Stream(device="openreg:1") as stream:
|
||||
self.assertEqual(torch.accelerator.current_stream(), stream)
|
||||
@ -42,7 +40,6 @@ class TestStream(TestCase):
|
||||
current_stream = torch.accelerator.current_stream()
|
||||
self.assertEqual(current_stream, stream2)
|
||||
|
||||
@skipIfTorchDynamo()
|
||||
def test_stream_synchronize(self):
|
||||
stream = torch.Stream(device="openreg:1")
|
||||
self.assertEqual(True, stream.query())
|
||||
@ -52,14 +49,12 @@ class TestStream(TestCase):
|
||||
stream.synchronize()
|
||||
self.assertEqual(True, stream.query())
|
||||
|
||||
@skipIfTorchDynamo()
|
||||
def test_stream_repr(self):
|
||||
stream = torch.Stream(device="openreg:1")
|
||||
self.assertTrue(
|
||||
"torch.Stream device_type=openreg, device_index=1" in repr(stream)
|
||||
)
|
||||
|
||||
@skipIfTorchDynamo()
|
||||
def test_stream_wait_stream(self):
|
||||
stream_1 = torch.Stream(device="openreg:0")
|
||||
stream_2 = torch.Stream(device="openreg:1")
|
||||
|
||||
@ -827,7 +827,7 @@ class TestFullyShardShardPlacementFnMultiProcess(FSDPTest):
|
||||
|
||||
torch.manual_seed(42 + self.rank)
|
||||
inp = torch.randint(0, model_args.vocab_size, (2, 16), device=device_type.type)
|
||||
for _ in range(5):
|
||||
for iter_idx in range(5):
|
||||
ref_loss = ref_model(inp).sum()
|
||||
loss = model(inp).sum()
|
||||
self.assertEqual(ref_loss, loss)
|
||||
|
||||
@ -218,7 +218,7 @@ class TestFullyShard2DTraining(FSDPTest):
|
||||
|
||||
torch.manual_seed(42 + global_mesh.get_local_rank("dp"))
|
||||
inp = torch.randint(0, model_args.vocab_size, (2, 16), device=device_type)
|
||||
for _ in range(5):
|
||||
for iter_idx in range(5):
|
||||
ref_loss = ref_model(inp).sum()
|
||||
loss = model(inp).sum()
|
||||
self.assertEqual(ref_loss, loss)
|
||||
@ -238,7 +238,9 @@ class TestFullyShard2DTraining(FSDPTest):
|
||||
# runs its reduce-scatter
|
||||
self.assertIsInstance(model.pos_embeddings.weight.placements[1], Shard)
|
||||
self.assertIsInstance(model.pos_embeddings.weight.grad.placements[1], Shard)
|
||||
for ref_param, param in zip(ref_model.parameters(), model.parameters()):
|
||||
for ref_param, (param_name, param) in zip(
|
||||
ref_model.parameters(), model.named_parameters()
|
||||
):
|
||||
full_grad = param.grad.full_tensor()
|
||||
self.assertEqual(ref_param.grad, full_grad)
|
||||
|
||||
|
||||
@ -336,7 +336,7 @@ class ComposabilityTest(MultiProcessTestCase):
|
||||
for model in partial_models
|
||||
]
|
||||
|
||||
for _train_step in range(5):
|
||||
for train_step in range(5):
|
||||
for optimizer in optimizers:
|
||||
optimizer.zero_grad()
|
||||
inputs = torch.rand((num_microbatches, dim), device=self.device)
|
||||
@ -517,7 +517,7 @@ class ComposabilityTest(MultiProcessTestCase):
|
||||
for model in ref_partial_models
|
||||
]
|
||||
|
||||
for _train_step in range(5):
|
||||
for train_step in range(5):
|
||||
for optimizer in optimizers:
|
||||
optimizer.zero_grad()
|
||||
for ref_optimizer in ref_optimizers:
|
||||
|
||||
@ -216,7 +216,7 @@ class TestSavePlan(TestCase):
|
||||
# Number of plans should remain unchanged
|
||||
self.assertEqual(len(all_plans), len(deduped_plans))
|
||||
|
||||
# Number of items in the deduped plans should be less than the original plans
|
||||
# Numer of items in the deduped plans should be less than the original plans
|
||||
for new_plan, old_plan in zip(deduped_plans, all_plans):
|
||||
self.assertFalse(_compare_save_plans(new_plan, old_plan))
|
||||
self.assertTrue(len(new_plan.items) < len(old_plan.items))
|
||||
|
||||
@ -4,7 +4,7 @@ import copy
|
||||
import functools
|
||||
import sys
|
||||
from collections.abc import Callable
|
||||
from itertools import chain, product
|
||||
from itertools import chain
|
||||
from typing import Union
|
||||
|
||||
import torch
|
||||
@ -708,43 +708,29 @@ class TestStateDict(DTensorTestBase, VerifyStateDictMixin):
|
||||
@with_comms
|
||||
@skip_if_lt_x_gpu(2)
|
||||
def test_flattened_osd(self) -> None:
|
||||
"""
|
||||
Test flattened optimizer state dictionaries with different combinations of
|
||||
flatten_optimizer_state_dict flag for saving and loading.
|
||||
|
||||
This test verifies that:
|
||||
1. We can save optimizer state dict with/without flattening
|
||||
2. We can load optimizer state dict with/without flattening
|
||||
3. The resulting optimizer state is equivalent regardless of flattening options
|
||||
"""
|
||||
for flatten_to_save, flatten_to_load in product([True, False], repeat=2):
|
||||
device_mesh = init_device_mesh(device_type, (self.world_size,))
|
||||
model = CompositeParamModel(device=torch.device(device_type))
|
||||
fsdp_model = fully_shard(copy.deepcopy(model), mesh=device_mesh)
|
||||
fsdp_optim = torch.optim.AdamW(fsdp_model.parameters())
|
||||
batch = torch.rand(8, 100, device=device_type)
|
||||
fsdp_model(batch).sum().backward()
|
||||
fsdp_optim.step()
|
||||
fsdp_optim.zero_grad()
|
||||
|
||||
# Get optimizer state dict with/without flattening option
|
||||
osd = get_optimizer_state_dict(
|
||||
fsdp_model,
|
||||
fsdp_optim,
|
||||
options=StateDictOptions(flatten_optimizer_state_dict=flatten_to_save),
|
||||
)
|
||||
|
||||
# Create a new optimizer and load the state from osd
|
||||
fsdp_optim2 = torch.optim.AdamW(fsdp_model.parameters())
|
||||
set_optimizer_state_dict(
|
||||
fsdp_model,
|
||||
optimizers=fsdp_optim2,
|
||||
optim_state_dict=osd,
|
||||
options=StateDictOptions(flatten_optimizer_state_dict=flatten_to_load),
|
||||
)
|
||||
|
||||
# Verify the loaded optimizer state matches the original
|
||||
self.assertEqual(fsdp_optim.state_dict(), fsdp_optim2.state_dict())
|
||||
device_mesh = init_device_mesh(device_type, (self.world_size,))
|
||||
model = CompositeParamModel(device=torch.device(device_type))
|
||||
fsdp_model = fully_shard(copy.deepcopy(model), mesh=device_mesh)
|
||||
fsdp_optim = torch.optim.AdamW(fsdp_model.parameters())
|
||||
batch = torch.rand(8, 100, device=device_type)
|
||||
fsdp_model(batch).sum().backward()
|
||||
fsdp_optim.step()
|
||||
fsdp_optim.zero_grad()
|
||||
osd1 = get_optimizer_state_dict(fsdp_model, fsdp_optim)
|
||||
osd2 = get_optimizer_state_dict(
|
||||
fsdp_model,
|
||||
fsdp_optim,
|
||||
options=StateDictOptions(flatten_optimizer_state_dict=True),
|
||||
)
|
||||
fsdp_optim2 = torch.optim.AdamW(fsdp_model.parameters())
|
||||
set_optimizer_state_dict(
|
||||
fsdp_model, optimizers=fsdp_optim2, optim_state_dict=osd2
|
||||
)
|
||||
self.assertEqual(fsdp_optim.state_dict(), fsdp_optim2.state_dict())
|
||||
set_optimizer_state_dict(
|
||||
fsdp_model, optimizers=fsdp_optim2, optim_state_dict=osd1
|
||||
)
|
||||
self.assertEqual(fsdp_optim.state_dict(), fsdp_optim2.state_dict())
|
||||
|
||||
def _test_deprecate_partial(self) -> None:
|
||||
model = CompositeParamModel(device=torch.device(device_type))
|
||||
|
||||
@ -31,17 +31,17 @@ if TEST_WITH_DEV_DBG_ASAN:
|
||||
sys.exit(0)
|
||||
|
||||
|
||||
_DISTRIBUTED_STATE_DICT_IMPLS = {
|
||||
_DISTRIBUTED_STATE_DICT_IMPLS = (
|
||||
StateDictType.LOCAL_STATE_DICT,
|
||||
StateDictType.SHARDED_STATE_DICT,
|
||||
}
|
||||
)
|
||||
|
||||
|
||||
class TestDistributedCheckpoint(FSDPTest):
|
||||
@property
|
||||
def world_size(self):
|
||||
if torch.accelerator.is_available():
|
||||
gpu_cnt = torch.accelerator.device_count()
|
||||
if torch.cuda.is_available():
|
||||
gpu_cnt = torch.cuda.device_count()
|
||||
if gpu_cnt < 2:
|
||||
return gpu_cnt
|
||||
return 2
|
||||
@ -93,9 +93,7 @@ class TestDistributedCheckpoint(FSDPTest):
|
||||
# TODO: add resharding test case.
|
||||
|
||||
|
||||
devices = ("cuda", "hpu", "xpu")
|
||||
instantiate_device_type_tests(
|
||||
TestDistributedCheckpoint, globals(), only_for=devices, allow_xpu=True
|
||||
)
|
||||
devices = ("cuda", "hpu")
|
||||
instantiate_device_type_tests(TestDistributedCheckpoint, globals(), only_for=devices)
|
||||
if __name__ == "__main__":
|
||||
run_tests()
|
||||
|
||||
@ -36,8 +36,8 @@ device_type = torch.device(get_devtype())
|
||||
class TestApply(FSDPTest):
|
||||
@property
|
||||
def world_size(self):
|
||||
if torch.accelerator.is_available():
|
||||
gpu_cnt = torch.accelerator.device_count()
|
||||
if torch.cuda.is_available():
|
||||
gpu_cnt = torch.cuda.device_count()
|
||||
if gpu_cnt < 2:
|
||||
return gpu_cnt
|
||||
return 2
|
||||
|
||||
@ -2,6 +2,7 @@
|
||||
# Owner(s): ["oncall: distributed"]
|
||||
|
||||
import sys
|
||||
from pathlib import Path
|
||||
|
||||
import torch
|
||||
import torch.distributed as dist
|
||||
@ -44,19 +45,53 @@ class TestInstantiator(TestCase):
|
||||
self.assertEqual(return_type_str, "Tuple[Tensor, int, str]")
|
||||
|
||||
def test_instantiate_scripted_remote_module_template(self):
|
||||
dir_path = Path(instantiator.INSTANTIATED_TEMPLATE_DIR_PATH)
|
||||
|
||||
# Cleanup.
|
||||
file_paths = dir_path.glob(f"{instantiator._FILE_PREFIX}*.py")
|
||||
for file_path in file_paths:
|
||||
file_path.unlink()
|
||||
|
||||
# Check before run.
|
||||
file_paths = dir_path.glob(f"{instantiator._FILE_PREFIX}*.py")
|
||||
num_files_before = len(list(file_paths))
|
||||
self.assertEqual(num_files_before, 0)
|
||||
|
||||
generated_module = instantiator.instantiate_scriptable_remote_module_template(
|
||||
MyModuleInterface
|
||||
)
|
||||
self.assertTrue(hasattr(generated_module, "_remote_forward"))
|
||||
self.assertTrue(hasattr(generated_module, "_generated_methods"))
|
||||
|
||||
# Check after run.
|
||||
file_paths = dir_path.glob(f"{instantiator._FILE_PREFIX}*.py")
|
||||
num_files_after = len(list(file_paths))
|
||||
self.assertEqual(num_files_after, 1)
|
||||
|
||||
def test_instantiate_non_scripted_remote_module_template(self):
|
||||
dir_path = Path(instantiator.INSTANTIATED_TEMPLATE_DIR_PATH)
|
||||
|
||||
# Cleanup.
|
||||
file_paths = dir_path.glob(f"{instantiator._FILE_PREFIX}*.py")
|
||||
for file_path in file_paths:
|
||||
file_path.unlink()
|
||||
|
||||
# Check before run.
|
||||
file_paths = dir_path.glob(f"{instantiator._FILE_PREFIX}*.py")
|
||||
num_files_before = len(list(file_paths))
|
||||
self.assertEqual(num_files_before, 0)
|
||||
|
||||
generated_module = (
|
||||
instantiator.instantiate_non_scriptable_remote_module_template()
|
||||
)
|
||||
self.assertTrue(hasattr(generated_module, "_remote_forward"))
|
||||
self.assertTrue(hasattr(generated_module, "_generated_methods"))
|
||||
|
||||
# Check after run.
|
||||
file_paths = dir_path.glob(f"{instantiator._FILE_PREFIX}*.py")
|
||||
num_files_after = len(list(file_paths))
|
||||
self.assertEqual(num_files_after, 1)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
run_tests()
|
||||
|
||||
@ -64,10 +64,6 @@ class TestDTensorDebugMode(TestCase):
|
||||
self.assertTrue(isinstance(debug_mode.operators[2], _RedistributeCall))
|
||||
self.assertEqual(next(iter(debug_mode.operators[1])), torch.ops.aten.mm.default)
|
||||
|
||||
# check stringification
|
||||
self.assertTrue(hasattr(debug_mode.operators[0], "args_str"))
|
||||
self.assertFalse(hasattr(debug_mode.operators[0], "args"))
|
||||
|
||||
def test_debug_string_inside_context(self):
|
||||
mesh = DeviceMesh(self.device_type, list(range(self.world_size)))
|
||||
|
||||
@ -271,7 +267,6 @@ class TestDTensorDebugMode(TestCase):
|
||||
record_torchfunction=True,
|
||||
record_faketensor=True,
|
||||
record_tensor_attributes=["a1", "a2"],
|
||||
store_original_args=True,
|
||||
) as debug_mode:
|
||||
torch.matmul(y, x)
|
||||
|
||||
@ -284,9 +279,6 @@ class TestDTensorDebugMode(TestCase):
|
||||
aten::_unsafe_view(t: f32[64, 8], [8, 8, 8])""",
|
||||
)
|
||||
|
||||
self.assertTrue(hasattr(debug_mode.operators[0], "args"))
|
||||
self.assertEqual(id(debug_mode.operators[0].args[0]), id(y))
|
||||
|
||||
@parametrize("has_inner_mode", [True, False])
|
||||
@parametrize("has_outer_mode", [True, False])
|
||||
def test_nested_debug_mode(self, has_inner_mode, has_outer_mode):
|
||||
|
||||
@ -20,18 +20,18 @@ from torch.distributed.tensor.experimental._attention import (
|
||||
_cp_options,
|
||||
_disable_context_parallel_dispatcher,
|
||||
_enable_context_parallel_dispatcher,
|
||||
_HeadTailLoadBalancer,
|
||||
_is_causal_behavior,
|
||||
_LoadBalancer,
|
||||
_PerDocumentHeadTailLoadBalancer,
|
||||
_PTRRLoadBalancer,
|
||||
_RotateMethod,
|
||||
context_parallel,
|
||||
context_parallel_unshard,
|
||||
set_rotate_method,
|
||||
)
|
||||
from torch.distributed.tensor.experimental._context_parallel._cp_custom_ops import (
|
||||
flex_cp_allgather,
|
||||
from torch.distributed.tensor.experimental._cp_custom_ops import flex_cp_allgather
|
||||
from torch.distributed.tensor.experimental._load_balancer import (
|
||||
_HeadTailLoadBalancer,
|
||||
_LoadBalancer,
|
||||
_PerDocumentHeadTailLoadBalancer,
|
||||
_PTRRLoadBalancer,
|
||||
)
|
||||
from torch.distributed.tensor.parallel import parallelize_module
|
||||
from torch.nn.attention import sdpa_kernel, SDPBackend
|
||||
@ -52,9 +52,7 @@ from torch.testing._internal.common_cuda import (
|
||||
from torch.testing._internal.common_distributed import skip_if_lt_x_gpu
|
||||
from torch.testing._internal.common_utils import run_tests, skipIfRocm
|
||||
from torch.testing._internal.distributed._tensor.common_dtensor import (
|
||||
create_local_tensor_test_class,
|
||||
DTensorTestBase,
|
||||
map_local_tensor_for_rank,
|
||||
with_comms,
|
||||
)
|
||||
|
||||
@ -160,7 +158,7 @@ class RingAttentionTest(DTensorTestBase):
|
||||
# parameters because when require_grad is True, resize_ is not
|
||||
# allowed. But requires_grad of cp_q, cp_k, and cp_v are False
|
||||
# now. So we can just use context_parallel() to shard q, k, v.
|
||||
# In reality, context_parallel() should be used to shard the input.
|
||||
# In reality, context_paralle() should be used to shard the input.
|
||||
# In reality, context_parallel() should only be used to shard
|
||||
# the model inputs (batch).
|
||||
|
||||
@ -703,7 +701,7 @@ class CPFlexAttentionTest(DTensorTestBase):
|
||||
)
|
||||
|
||||
# TODO: change this for-loop to run_subtests
|
||||
# Use a for-loop instead of run_subtests because we need to initialize the mask
|
||||
# Use a for-loop instead of run_subtests because we need to intialize the mask
|
||||
# for each subtest. This can be baked into self._test_cp_flex_attention as
|
||||
# a str argument denoting mask type.
|
||||
for batch_size, max_seq_len, lb_type in itertools.product(
|
||||
@ -802,47 +800,11 @@ class TestSharding(DTensorTestBase):
|
||||
chunks = freqs_cis.chunk(self.world_size * 2)
|
||||
self.assertEqual(
|
||||
freqs_cis_shard,
|
||||
map_local_tensor_for_rank(
|
||||
chunks,
|
||||
self.rank,
|
||||
lambda chunks, rank: torch.cat(
|
||||
[chunks[rank], chunks[self.world_size * 2 - rank - 1]],
|
||||
dim=0,
|
||||
),
|
||||
torch.cat(
|
||||
[chunks[self.rank], chunks[self.world_size * 2 - self.rank - 1]], dim=0
|
||||
),
|
||||
)
|
||||
|
||||
|
||||
RingAttentionTestWithLocalTensor = create_local_tensor_test_class(
|
||||
RingAttentionTest,
|
||||
skipped_tests=[
|
||||
# Need to make attention implementation local tensor friendly, e.g.
|
||||
# rewrite "rank local" logic
|
||||
"test_ring_attention_sdpa",
|
||||
],
|
||||
)
|
||||
|
||||
CPFlexAttentionTestWithLocalTensor = create_local_tensor_test_class(
|
||||
CPFlexAttentionTest,
|
||||
skipped_tests=[
|
||||
# Missing support for batched tensors
|
||||
"test_cp_flex_attention_causal_mask",
|
||||
"test_cp_flex_attention_document_mask",
|
||||
],
|
||||
)
|
||||
|
||||
TestCPCustomOpsWithLocalTensor = create_local_tensor_test_class(
|
||||
TestCPCustomOps,
|
||||
skipped_tests=[
|
||||
# Missing support for fake tensors
|
||||
"test_flex_cp_custom_op",
|
||||
],
|
||||
)
|
||||
|
||||
TestShardingWithLocalTensor = create_local_tensor_test_class(
|
||||
TestSharding,
|
||||
)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
run_tests()
|
||||
|
||||
@ -16,7 +16,6 @@ from torch.distributed.tensor import (
|
||||
from torch.nn import functional as F
|
||||
from torch.testing._internal.common_utils import run_tests
|
||||
from torch.testing._internal.distributed._tensor.common_dtensor import (
|
||||
create_local_tensor_test_class,
|
||||
DTensorTestBase,
|
||||
skip_if_lt_x_gpu,
|
||||
with_comms,
|
||||
@ -204,42 +203,34 @@ class DistConvolutionOpsTest(DTensorTestBase):
|
||||
self.assertTrue(b_dt.grad is not None)
|
||||
self.assertTrue(x_dt.grad is None)
|
||||
|
||||
def _run_single_arg_fwd(self, model, arg) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""Given model and arg, runs fwd model local and distbuted given device_mesh"""
|
||||
device_mesh = self.build_device_mesh()
|
||||
model_copy = copy.deepcopy(model).to(device=self.device_type)
|
||||
dist_model = distribute_module(model, device_mesh, _conv_fn)
|
||||
arg_dt = DTensor.from_local(arg, device_mesh, [Replicate()])
|
||||
out_dt = dist_model(arg_dt.to(device=self.device_type))
|
||||
out = model_copy(arg)
|
||||
return (out_dt.full_tensor(), out)
|
||||
|
||||
@with_comms
|
||||
def test_conv1d(self):
|
||||
device_mesh = self.build_device_mesh()
|
||||
model = nn.Conv1d(64, 64, 3, padding=1)
|
||||
x = torch.randn(1, 64, 8, device=self.device_type)
|
||||
out_dt, out = self._run_single_arg_fwd(model, x)
|
||||
model_gt = copy.deepcopy(model)
|
||||
x = torch.randn(1, 64, 8)
|
||||
x_dt = DTensor.from_local(x, device_mesh, [Replicate()])
|
||||
model_dt = distribute_module(
|
||||
model, device_mesh, _conv_fn, input_fn=None, output_fn=None
|
||||
)
|
||||
out_dt = model_dt(x_dt)
|
||||
out = model_gt(x)
|
||||
self.assertEqual(out_dt.shape, out.shape)
|
||||
|
||||
@with_comms
|
||||
def test_conv3d(self):
|
||||
device_mesh = self.build_device_mesh()
|
||||
model = nn.Conv3d(64, 64, 3, padding=1)
|
||||
model_gt = copy.deepcopy(model).to(device=self.device_type)
|
||||
x = torch.randn(1, 64, 8, 8, 8, device=self.device_type)
|
||||
out_dt, out = self._run_single_arg_fwd(model, x)
|
||||
x_dt = DTensor.from_local(x, device_mesh, [Replicate()])
|
||||
model_dt = distribute_module(
|
||||
model, device_mesh, _conv_fn, input_fn=None, output_fn=None
|
||||
)
|
||||
out_dt = model_dt(x_dt)
|
||||
out = model_gt(x)
|
||||
self.assertEqual(out_dt.shape, out.shape)
|
||||
|
||||
|
||||
DistConvolutionOpsTestWithLocalTensor = create_local_tensor_test_class(
|
||||
DistConvolutionOpsTest,
|
||||
# Send / recv ops are not supported
|
||||
skipped_tests=[
|
||||
"test_conv1d",
|
||||
"test_conv3d",
|
||||
"test_conv_backward_none_grad_inp",
|
||||
"test_depthwise_convolution",
|
||||
"test_downsampling_convolution",
|
||||
],
|
||||
)
|
||||
|
||||
if __name__ == "__main__":
|
||||
run_tests()
|
||||
|
||||
@ -464,6 +464,25 @@ def forward(self, b_parametrizations_buffer_original0, x):
|
||||
run(g, 64, 8)
|
||||
self.assertEqual(cnt.frame_count, 2)
|
||||
|
||||
def test_dtensor_requires_grad_recompile(self):
|
||||
cnt = torch._dynamo.testing.CompileCounterWithBackend("aot_eager")
|
||||
mesh = DeviceMesh(self.device_type, torch.arange(self.world_size))
|
||||
|
||||
@torch.compile(backend=cnt, fullgraph=True)
|
||||
def f(x):
|
||||
y = x * x
|
||||
return y.to_local()
|
||||
|
||||
full_x = torch.randn(8, 8, requires_grad=False)
|
||||
x = distribute_tensor(full_x, mesh, [Shard(0)])
|
||||
f(x)
|
||||
|
||||
full_x = torch.randn(8, 8, requires_grad=True)
|
||||
x = distribute_tensor(full_x, mesh, [Shard(0)])
|
||||
f(x)
|
||||
|
||||
self.assertEqual(cnt.frame_count, 2)
|
||||
|
||||
def test_dtensor_attribute_access_on_intermediate(self):
|
||||
mesh = DeviceMesh(self.device_type, torch.arange(self.world_size))
|
||||
|
||||
|
||||
@ -22,11 +22,7 @@ from torch.distributed.tensor.parallel import (
|
||||
parallelize_module,
|
||||
RowwiseParallel,
|
||||
)
|
||||
from torch.nn.attention.flex_attention import (
|
||||
BlockMask,
|
||||
create_block_mask,
|
||||
flex_attention,
|
||||
)
|
||||
from torch.nn.attention.flex_attention import create_block_mask, flex_attention
|
||||
from torch.testing._internal.common_utils import (
|
||||
instantiate_parametrized_tests,
|
||||
parametrize,
|
||||
@ -36,7 +32,6 @@ from torch.testing._internal.common_utils import (
|
||||
)
|
||||
from torch.testing._internal.distributed._tensor.common_dtensor import MLPModule
|
||||
from torch.testing._internal.distributed.fake_pg import FakeStore
|
||||
from torch.utils._pytree import register_pytree_node
|
||||
|
||||
|
||||
class SimpleModel(torch.nn.Module):
|
||||
@ -181,15 +176,6 @@ def _count_op(gm, target):
|
||||
return sum(1 for node in gm.graph.nodes if node.target == target)
|
||||
|
||||
|
||||
register_pytree_node(
|
||||
BlockMask,
|
||||
BlockMask._flatten,
|
||||
BlockMask._unflatten,
|
||||
flatten_with_keys_fn=BlockMask._flatten_with_keys,
|
||||
serialized_type_name="torch.nn.attention.flex_attention.BlockMask",
|
||||
)
|
||||
|
||||
|
||||
@requires_cuda
|
||||
class DTensorExportTest(TestCase):
|
||||
def tearDown(self):
|
||||
@ -520,21 +506,6 @@ class DTensorExportTest(TestCase):
|
||||
2,
|
||||
)
|
||||
|
||||
def test_union_typed_annotation(self):
|
||||
def fn(leaf: torch.Tensor | DTensor):
|
||||
def nest_fn(leaf: torch.Tensor | DTensor):
|
||||
# def nest_fn(leaf: Union[torch.Tensor, DTensor]): # this works
|
||||
if isinstance(leaf, DTensor):
|
||||
leaf = leaf.to_local()
|
||||
return leaf
|
||||
|
||||
return nest_fn(leaf) + 1
|
||||
|
||||
z = torch.randn(16, 16)
|
||||
gm = graph_capture_and_aot_export_joint_with_descriptors(fn, (z,))
|
||||
|
||||
self.assertEqual(fn(z), gm(z)[0])
|
||||
|
||||
|
||||
instantiate_parametrized_tests(DTensorExportTest)
|
||||
|
||||
|
||||
@ -168,7 +168,7 @@ class TestEmbeddingOp(DTensorTestBase):
|
||||
self._run_embedding_op_test(mesh, 0, [6, 7, 6], 13, 22)
|
||||
self._run_embedding_op_test(mesh, 0, [34], 15, 14, padding_idx=10)
|
||||
|
||||
from torch.distributed.tensor.placement_types import MaskPartial
|
||||
from torch.distributed.tensor._ops._embedding_ops import _MaskPartial
|
||||
|
||||
# test collectives
|
||||
embedding_mod = torch.nn.Embedding(10, 20, device=self.device_type)
|
||||
@ -176,7 +176,7 @@ class TestEmbeddingOp(DTensorTestBase):
|
||||
inp = torch.randint(0, 10, (8, 8), device=self.device_type)
|
||||
replicated_inp = DTensor.from_local(inp, mesh, [Replicate()], run_check=False)
|
||||
output = sharded_embedding(replicated_inp)
|
||||
self.assertIsInstance(output.placements[0], MaskPartial)
|
||||
self.assertIsInstance(output.placements[0], _MaskPartial)
|
||||
|
||||
comm_mode = CommDebugMode()
|
||||
|
||||
@ -192,9 +192,9 @@ class TestEmbeddingOp(DTensorTestBase):
|
||||
inp = torch.randint(0, 10, (4, 4), device=self.device_type)
|
||||
replicated_inp = DTensor.from_local(inp, mesh, [Replicate()], run_check=False)
|
||||
|
||||
from torch.distributed.tensor.placement_types import MaskPartial
|
||||
from torch.distributed.tensor._ops._embedding_ops import _MaskPartial
|
||||
|
||||
# case 1: two embeddings with the same shape, thus sharing the underlying MaskPartial
|
||||
# case 1: two embeddings with the same shape, thus sharing the underlying _MaskPartial
|
||||
# and MaskBuffer, because of cache hit from sharding propagation
|
||||
|
||||
emb1 = torch.nn.Embedding(10, 23, device=self.device_type)
|
||||
@ -206,23 +206,23 @@ class TestEmbeddingOp(DTensorTestBase):
|
||||
output2 = sharded_emb2(replicated_inp)
|
||||
|
||||
partial_placement1 = output1.placements[0]
|
||||
self.assertIsInstance(partial_placement1, MaskPartial)
|
||||
self.assertIsInstance(partial_placement1, _MaskPartial)
|
||||
output1.full_tensor()
|
||||
|
||||
partial_placement2 = output2.placements[0]
|
||||
self.assertIsInstance(partial_placement2, MaskPartial)
|
||||
self.assertIsInstance(partial_placement2, _MaskPartial)
|
||||
output2.full_tensor()
|
||||
|
||||
self.assertTrue(id(partial_placement1), id(partial_placement2))
|
||||
|
||||
# case 2: two embeddings with the same logical_dim_size, but different logical_shape
|
||||
# thus they will have different MaskPartial placements (with no cache hit)
|
||||
# thus they will have different _MaskPartial placements (with no cache hit)
|
||||
|
||||
emb3 = torch.nn.Embedding(10, 29, device=self.device_type)
|
||||
sharded_emb3 = self._apply_sharding(emb3, 0, mesh)
|
||||
output3 = sharded_emb3(replicated_inp)
|
||||
partial_placement3 = output3.placements[0]
|
||||
self.assertIsInstance(partial_placement3, MaskPartial)
|
||||
self.assertIsInstance(partial_placement3, _MaskPartial)
|
||||
output2.full_tensor()
|
||||
|
||||
# not equal because of different logical_shape, despite of same logical_dim_size
|
||||
|
||||
@ -60,9 +60,9 @@ class DistMathOpsTest(DTensorTestBase):
|
||||
shard_spec = [Shard(0)]
|
||||
|
||||
tensor = torch.randn(12, 8, 8)
|
||||
if op_str in ("any", "all"):
|
||||
# Test bool tensor for any() and all() reduction ops
|
||||
# Previously all() had a bug using sum reduction instead of product
|
||||
# TODO: check `all` correctness and test `all` on a bool tensor
|
||||
if op_str in ("any"):
|
||||
# test out a bool tensor for any
|
||||
tensor = tensor < 0
|
||||
dtensor = distribute_tensor(tensor, device_mesh, shard_spec)
|
||||
|
||||
|
||||
@ -511,7 +511,7 @@ class DistTensorOpsTest(DTensorTestBase):
|
||||
# case 2 input sharding: input sharded, index replicated, output mask partial
|
||||
# only works when index has size 1 on the gather dimension and
|
||||
# input is sharded on the gather dimension
|
||||
from torch.distributed.tensor.placement_types import MaskPartial
|
||||
from torch.distributed.tensor._ops._embedding_ops import _MaskPartial
|
||||
|
||||
gather_dim = 1
|
||||
global_input = torch.randn(12, 8, 16)
|
||||
@ -522,7 +522,7 @@ class DistTensorOpsTest(DTensorTestBase):
|
||||
with comm_mode:
|
||||
output_dt = torch.gather(input_dt, gather_dim, index_dt)
|
||||
self.assertEqual(comm_mode.get_total_counts(), 0)
|
||||
self.assertIsInstance(output_dt.placements[0], MaskPartial)
|
||||
self.assertIsInstance(output_dt.placements[0], _MaskPartial)
|
||||
self.assertEqual(output_dt.full_tensor(), global_output)
|
||||
|
||||
# case 3 index sharding: input replicated, index sharded, output sharded
|
||||
|
||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user