mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-27 09:04:53 +08:00
Compare commits
1 Commits
add_op_tes
...
eager_mode
| Author | SHA1 | Date | |
|---|---|---|---|
| c56b575e61 |
@ -241,7 +241,7 @@ def wait_for_connection(addr, port, timeout=15, attempt_cnt=5):
|
|||||||
try:
|
try:
|
||||||
with socket.create_connection((addr, port), timeout=timeout):
|
with socket.create_connection((addr, port), timeout=timeout):
|
||||||
return
|
return
|
||||||
except (ConnectionRefusedError, TimeoutError): # noqa: PERF203
|
except (ConnectionRefusedError, socket.timeout): # noqa: PERF203
|
||||||
if i == attempt_cnt - 1:
|
if i == attempt_cnt - 1:
|
||||||
raise
|
raise
|
||||||
time.sleep(timeout)
|
time.sleep(timeout)
|
||||||
@ -1004,7 +1004,7 @@ if __name__ == "__main__":
|
|||||||
install_condaforge_python(host, args.python_version)
|
install_condaforge_python(host, args.python_version)
|
||||||
sys.exit(0)
|
sys.exit(0)
|
||||||
|
|
||||||
python_version = args.python_version if args.python_version is not None else "3.10"
|
python_version = args.python_version if args.python_version is not None else "3.9"
|
||||||
|
|
||||||
if args.use_torch_from_pypi:
|
if args.use_torch_from_pypi:
|
||||||
configure_system(host, compiler=args.compiler, python_version=python_version)
|
configure_system(host, compiler=args.compiler, python_version=python_version)
|
||||||
|
|||||||
@ -69,8 +69,7 @@ RUN bash ./install_cuda.sh 13.0
|
|||||||
ENV DESIRED_CUDA=13.0
|
ENV DESIRED_CUDA=13.0
|
||||||
|
|
||||||
FROM ${ROCM_IMAGE} as rocm
|
FROM ${ROCM_IMAGE} as rocm
|
||||||
ARG PYTORCH_ROCM_ARCH
|
ENV PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
|
||||||
ENV PYTORCH_ROCM_ARCH ${PYTORCH_ROCM_ARCH}
|
|
||||||
ADD ./common/install_mkl.sh install_mkl.sh
|
ADD ./common/install_mkl.sh install_mkl.sh
|
||||||
RUN bash ./install_mkl.sh && rm install_mkl.sh
|
RUN bash ./install_mkl.sh && rm install_mkl.sh
|
||||||
ENV MKLROOT /opt/intel
|
ENV MKLROOT /opt/intel
|
||||||
|
|||||||
@ -36,12 +36,6 @@ case ${DOCKER_TAG_PREFIX} in
|
|||||||
;;
|
;;
|
||||||
rocm*)
|
rocm*)
|
||||||
BASE_TARGET=rocm
|
BASE_TARGET=rocm
|
||||||
PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
|
|
||||||
# add gfx950 conditionally starting in ROCm 7.0
|
|
||||||
if [[ "$ROCM_VERSION" == *"7.0"* ]]; then
|
|
||||||
PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950"
|
|
||||||
fi
|
|
||||||
EXTRA_BUILD_ARGS="${EXTRA_BUILD_ARGS} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH}"
|
|
||||||
;;
|
;;
|
||||||
*)
|
*)
|
||||||
echo "ERROR: Unknown docker tag ${DOCKER_TAG_PREFIX}"
|
echo "ERROR: Unknown docker tag ${DOCKER_TAG_PREFIX}"
|
||||||
|
|||||||
@ -262,10 +262,13 @@ case "$tag" in
|
|||||||
TRITON_CPU=yes
|
TRITON_CPU=yes
|
||||||
;;
|
;;
|
||||||
pytorch-linux-jammy-linter)
|
pytorch-linux-jammy-linter)
|
||||||
PYTHON_VERSION=3.10
|
# TODO: Use 3.9 here because of this issue https://github.com/python/mypy/issues/13627.
|
||||||
|
# We will need to update mypy version eventually, but that's for another day. The task
|
||||||
|
# would be to upgrade mypy to 1.0.0 with Python 3.11
|
||||||
|
PYTHON_VERSION=3.9
|
||||||
;;
|
;;
|
||||||
pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter)
|
pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter)
|
||||||
PYTHON_VERSION=3.10
|
PYTHON_VERSION=3.9
|
||||||
CUDA_VERSION=12.8.1
|
CUDA_VERSION=12.8.1
|
||||||
;;
|
;;
|
||||||
pytorch-linux-jammy-aarch64-py3.10-gcc11)
|
pytorch-linux-jammy-aarch64-py3.10-gcc11)
|
||||||
|
|||||||
@ -1 +1 @@
|
|||||||
v2.28.3-1
|
v2.27.5-1
|
||||||
|
|||||||
@ -1 +1 @@
|
|||||||
v2.28.3-1
|
v2.27.7-1
|
||||||
|
|||||||
@ -1 +1 @@
|
|||||||
bbb06c0334a6772b92d24bde54956e675c8c6604
|
5ae38bdb0dc066c5823e34dc9797afb9de42c866
|
||||||
|
|||||||
@ -12,8 +12,8 @@ function do_install() {
|
|||||||
|
|
||||||
rocm_version_nodot=${rocm_version//./}
|
rocm_version_nodot=${rocm_version//./}
|
||||||
|
|
||||||
# https://github.com/icl-utk-edu/magma/pull/65
|
# Version 2.7.2 + ROCm related updates
|
||||||
MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
|
MAGMA_VERSION=a1625ff4d9bc362906bd01f805dbbe12612953f6
|
||||||
magma_archive="magma-rocm${rocm_version_nodot}-${MAGMA_VERSION}-1.tar.bz2"
|
magma_archive="magma-rocm${rocm_version_nodot}-${MAGMA_VERSION}-1.tar.bz2"
|
||||||
|
|
||||||
rocm_dir="/opt/rocm"
|
rocm_dir="/opt/rocm"
|
||||||
|
|||||||
@ -40,16 +40,12 @@ case ${DOCKER_TAG_PREFIX} in
|
|||||||
;;
|
;;
|
||||||
rocm*)
|
rocm*)
|
||||||
# we want the patch version of 6.4 instead
|
# we want the patch version of 6.4 instead
|
||||||
if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
|
if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
|
||||||
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
|
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
|
||||||
fi
|
fi
|
||||||
BASE_TARGET=rocm
|
BASE_TARGET=rocm
|
||||||
GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete
|
GPU_IMAGE=rocm/dev-ubuntu-22.04:${GPU_ARCH_VERSION}-complete
|
||||||
PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
|
PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
|
||||||
# add gfx950 conditionally starting in ROCm 7.0
|
|
||||||
if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
|
|
||||||
PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950"
|
|
||||||
fi
|
|
||||||
DOCKER_GPU_BUILD_ARG="--build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg ROCM_VERSION=${GPU_ARCH_VERSION}"
|
DOCKER_GPU_BUILD_ARG="--build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg ROCM_VERSION=${GPU_ARCH_VERSION}"
|
||||||
;;
|
;;
|
||||||
*)
|
*)
|
||||||
|
|||||||
@ -82,7 +82,7 @@ case ${image} in
|
|||||||
;;
|
;;
|
||||||
manylinux2_28-builder:rocm*)
|
manylinux2_28-builder:rocm*)
|
||||||
# we want the patch version of 6.4 instead
|
# we want the patch version of 6.4 instead
|
||||||
if [[ "$GPU_ARCH_VERSION" == *"6.4"* ]]; then
|
if [[ $(ver $GPU_ARCH_VERSION) -eq $(ver 6.4) ]]; then
|
||||||
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
|
GPU_ARCH_VERSION="${GPU_ARCH_VERSION}.2"
|
||||||
fi
|
fi
|
||||||
TARGET=rocm_final
|
TARGET=rocm_final
|
||||||
@ -90,10 +90,6 @@ case ${image} in
|
|||||||
DEVTOOLSET_VERSION="11"
|
DEVTOOLSET_VERSION="11"
|
||||||
GPU_IMAGE=rocm/dev-almalinux-8:${GPU_ARCH_VERSION}-complete
|
GPU_IMAGE=rocm/dev-almalinux-8:${GPU_ARCH_VERSION}-complete
|
||||||
PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
|
PYTORCH_ROCM_ARCH="gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
|
||||||
# add gfx950 conditionally starting in ROCm 7.0
|
|
||||||
if [[ "$GPU_ARCH_VERSION" == *"7.0"* ]]; then
|
|
||||||
PYTORCH_ROCM_ARCH="${PYTORCH_ROCM_ARCH};gfx950"
|
|
||||||
fi
|
|
||||||
DOCKER_GPU_BUILD_ARG="--build-arg ROCM_VERSION=${GPU_ARCH_VERSION} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg DEVTOOLSET_VERSION=${DEVTOOLSET_VERSION}"
|
DOCKER_GPU_BUILD_ARG="--build-arg ROCM_VERSION=${GPU_ARCH_VERSION} --build-arg PYTORCH_ROCM_ARCH=${PYTORCH_ROCM_ARCH} --build-arg DEVTOOLSET_VERSION=${DEVTOOLSET_VERSION}"
|
||||||
;;
|
;;
|
||||||
manylinux2_28-builder:xpu)
|
manylinux2_28-builder:xpu)
|
||||||
|
|||||||
@ -112,6 +112,8 @@ ninja==1.11.1.3
|
|||||||
#Pinned versions: 1.11.1.3
|
#Pinned versions: 1.11.1.3
|
||||||
#test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
|
#test that import: run_test.py, test_cpp_extensions_aot.py,test_determination.py
|
||||||
|
|
||||||
|
numba==0.49.0 ; python_version < "3.9" and platform_machine != "s390x"
|
||||||
|
numba==0.55.2 ; python_version == "3.9" and platform_machine != "s390x"
|
||||||
numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
|
numba==0.55.2 ; python_version == "3.10" and platform_machine != "s390x"
|
||||||
numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
|
numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
|
||||||
#Description: Just-In-Time Compiler for Numerical Functions
|
#Description: Just-In-Time Compiler for Numerical Functions
|
||||||
@ -132,7 +134,7 @@ numba==0.60.0 ; python_version == "3.12" and platform_machine != "s390x"
|
|||||||
#test_nn.py, test_namedtensor.py, test_linalg.py, test_jit_cuda_fuser.py,
|
#test_nn.py, test_namedtensor.py, test_linalg.py, test_jit_cuda_fuser.py,
|
||||||
#test_jit.py, test_indexing.py, test_datapipe.py, test_dataloader.py,
|
#test_jit.py, test_indexing.py, test_datapipe.py, test_dataloader.py,
|
||||||
#test_binary_ufuncs.py
|
#test_binary_ufuncs.py
|
||||||
numpy==1.22.4; python_version == "3.10"
|
numpy==1.22.4; python_version == "3.9" or python_version == "3.10"
|
||||||
numpy==1.26.2; python_version == "3.11" or python_version == "3.12"
|
numpy==1.26.2; python_version == "3.11" or python_version == "3.12"
|
||||||
numpy==2.1.2; python_version >= "3.13"
|
numpy==2.1.2; python_version >= "3.13"
|
||||||
|
|
||||||
@ -324,6 +326,8 @@ pywavelets==1.7.0 ; python_version >= "3.12"
|
|||||||
lxml==5.3.0
|
lxml==5.3.0
|
||||||
#Description: This is a requirement of unittest-xml-reporting
|
#Description: This is a requirement of unittest-xml-reporting
|
||||||
|
|
||||||
|
# Python-3.9 binaries
|
||||||
|
|
||||||
PyGithub==2.3.0
|
PyGithub==2.3.0
|
||||||
|
|
||||||
sympy==1.13.3
|
sympy==1.13.3
|
||||||
|
|||||||
@ -1,15 +1,8 @@
|
|||||||
sphinx==5.3.0
|
sphinx==5.3.0
|
||||||
#Description: This is used to generate PyTorch docs
|
#Description: This is used to generate PyTorch docs
|
||||||
#Pinned versions: 5.3.0
|
#Pinned versions: 5.3.0
|
||||||
|
-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@1657ad2fc1acdc98aa719eebecbb0128a7c13ce4#egg=pytorch_sphinx_theme2
|
||||||
|
|
||||||
standard-imghdr==3.13.0; python_version >= "3.13"
|
|
||||||
#Description: This is needed by Sphinx, so it needs to be added here.
|
|
||||||
# The reasons are as follows:
|
|
||||||
# 1) This module has been removed from the Python standard library since Python 3.13(https://peps.python.org/pep-0594/#imghdr);
|
|
||||||
# 2) The current version of Sphinx (5.3.0) is not compatible with Python 3.13.
|
|
||||||
# Once Sphinx is upgraded to a version compatible with Python 3.13 or later, we can remove this dependency.
|
|
||||||
|
|
||||||
-e git+https://github.com/pytorch/pytorch_sphinx_theme.git@d53b0ffb9b1cda68260693ea98f3483823c88d8e#egg=pytorch_sphinx_theme2
|
|
||||||
# TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
|
# TODO: sphinxcontrib.katex 0.9.0 adds a local KaTeX server to speed up pre-rendering
|
||||||
# but it doesn't seem to work and hangs around idly. The initial thought that it is probably
|
# but it doesn't seem to work and hangs around idly. The initial thought that it is probably
|
||||||
# something related to Docker setup. We can investigate this later.
|
# something related to Docker setup. We can investigate this later.
|
||||||
|
|||||||
@ -72,7 +72,7 @@ def sample_vllm_test_library():
|
|||||||
]
|
]
|
||||||
),
|
),
|
||||||
"pytest -v -s entrypoints/llm/test_generate.py",
|
"pytest -v -s entrypoints/llm/test_generate.py",
|
||||||
"pytest -v -s entrypoints/offline_mode",
|
"VLLM_USE_V1=0 pytest -v -s entrypoints/offline_mode",
|
||||||
],
|
],
|
||||||
},
|
},
|
||||||
"vllm_regression_test": {
|
"vllm_regression_test": {
|
||||||
|
|||||||
@ -1,11 +1,11 @@
|
|||||||
SHELL=/usr/bin/env bash
|
SHELL=/usr/bin/env bash
|
||||||
|
|
||||||
DOCKER_CMD ?= docker
|
DOCKER_CMD ?= docker
|
||||||
DESIRED_ROCM ?= 7.0
|
DESIRED_ROCM ?= 6.4
|
||||||
DESIRED_ROCM_SHORT = $(subst .,,$(DESIRED_ROCM))
|
DESIRED_ROCM_SHORT = $(subst .,,$(DESIRED_ROCM))
|
||||||
PACKAGE_NAME = magma-rocm
|
PACKAGE_NAME = magma-rocm
|
||||||
# inherit this from underlying docker image, do not pass this env var to docker
|
# inherit this from underlying docker image, do not pass this env var to docker
|
||||||
#PYTORCH_ROCM_ARCH ?= gfx900;gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201
|
#PYTORCH_ROCM_ARCH ?= gfx900;gfx906;gfx908;gfx90a;gfx942;gfx1030;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201
|
||||||
|
|
||||||
DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
|
DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
|
||||||
-v $(shell git rev-parse --show-toplevel)/.ci:/builder \
|
-v $(shell git rev-parse --show-toplevel)/.ci:/builder \
|
||||||
@ -16,7 +16,6 @@ DOCKER_RUN = set -eou pipefail; ${DOCKER_CMD} run --rm -i \
|
|||||||
magma-rocm/build_magma.sh
|
magma-rocm/build_magma.sh
|
||||||
|
|
||||||
.PHONY: all
|
.PHONY: all
|
||||||
all: magma-rocm70
|
|
||||||
all: magma-rocm64
|
all: magma-rocm64
|
||||||
all: magma-rocm63
|
all: magma-rocm63
|
||||||
|
|
||||||
@ -25,11 +24,6 @@ clean:
|
|||||||
$(RM) -r magma-*
|
$(RM) -r magma-*
|
||||||
$(RM) -r output
|
$(RM) -r output
|
||||||
|
|
||||||
.PHONY: magma-rocm70
|
|
||||||
magma-rocm70: DESIRED_ROCM := 7.0
|
|
||||||
magma-rocm70:
|
|
||||||
$(DOCKER_RUN)
|
|
||||||
|
|
||||||
.PHONY: magma-rocm64
|
.PHONY: magma-rocm64
|
||||||
magma-rocm64: DESIRED_ROCM := 6.4
|
magma-rocm64: DESIRED_ROCM := 6.4
|
||||||
magma-rocm64:
|
magma-rocm64:
|
||||||
|
|||||||
@ -6,8 +6,8 @@ set -eou pipefail
|
|||||||
# The script expects DESIRED_CUDA and PACKAGE_NAME to be set
|
# The script expects DESIRED_CUDA and PACKAGE_NAME to be set
|
||||||
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
|
ROOT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
|
||||||
|
|
||||||
# https://github.com/icl-utk-edu/magma/pull/65
|
# Version 2.7.2 + ROCm related updates
|
||||||
MAGMA_VERSION=d6e4117bc88e73f06d26c6c2e14f064e8fc3d1ec
|
MAGMA_VERSION=a1625ff4d9bc362906bd01f805dbbe12612953f6
|
||||||
|
|
||||||
# Folders for the build
|
# Folders for the build
|
||||||
PACKAGE_FILES=${ROOT_DIR}/magma-rocm/package_files # metadata
|
PACKAGE_FILES=${ROOT_DIR}/magma-rocm/package_files # metadata
|
||||||
@ -20,7 +20,7 @@ mkdir -p ${PACKAGE_DIR} ${PACKAGE_OUTPUT}/linux-64 ${PACKAGE_BUILD} ${PACKAGE_RE
|
|||||||
|
|
||||||
# Fetch magma sources and verify checksum
|
# Fetch magma sources and verify checksum
|
||||||
pushd ${PACKAGE_DIR}
|
pushd ${PACKAGE_DIR}
|
||||||
git clone https://github.com/jeffdaily/magma
|
git clone https://bitbucket.org/icl/magma.git
|
||||||
pushd magma
|
pushd magma
|
||||||
git checkout ${MAGMA_VERSION}
|
git checkout ${MAGMA_VERSION}
|
||||||
popd
|
popd
|
||||||
|
|||||||
@ -58,7 +58,7 @@ time python tools/setup_helpers/generate_code.py \
|
|||||||
|
|
||||||
# Build the docs
|
# Build the docs
|
||||||
pushd docs/cpp
|
pushd docs/cpp
|
||||||
time make VERBOSE=1 html
|
time make VERBOSE=1 html -j
|
||||||
|
|
||||||
popd
|
popd
|
||||||
popd
|
popd
|
||||||
|
|||||||
@ -55,7 +55,7 @@ test_python_shard() {
|
|||||||
|
|
||||||
setup_test_python
|
setup_test_python
|
||||||
|
|
||||||
time python test/run_test.py --verbose --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests --shard "$1" "$NUM_TEST_SHARDS"
|
time python test/run_test.py --verbose --exclude-jit-executor --exclude-distributed-tests --shard "$1" "$NUM_TEST_SHARDS"
|
||||||
|
|
||||||
assert_git_not_dirty
|
assert_git_not_dirty
|
||||||
}
|
}
|
||||||
|
|||||||
@ -322,29 +322,23 @@ test_python_shard() {
|
|||||||
|
|
||||||
# modify LD_LIBRARY_PATH to ensure it has the conda env.
|
# modify LD_LIBRARY_PATH to ensure it has the conda env.
|
||||||
# This set of tests has been shown to be buggy without it for the split-build
|
# This set of tests has been shown to be buggy without it for the split-build
|
||||||
time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests $INCLUDE_CLAUSE --shard "$1" "$NUM_TEST_SHARDS" --verbose $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
|
time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests $INCLUDE_CLAUSE --shard "$1" "$NUM_TEST_SHARDS" --verbose $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
|
||||||
|
|
||||||
assert_git_not_dirty
|
assert_git_not_dirty
|
||||||
}
|
}
|
||||||
|
|
||||||
test_python() {
|
test_python() {
|
||||||
# shellcheck disable=SC2086
|
# shellcheck disable=SC2086
|
||||||
time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests $INCLUDE_CLAUSE --verbose $PYTHON_TEST_EXTRA_OPTION
|
time python test/run_test.py --exclude-jit-executor --exclude-distributed-tests $INCLUDE_CLAUSE --verbose $PYTHON_TEST_EXTRA_OPTION
|
||||||
assert_git_not_dirty
|
assert_git_not_dirty
|
||||||
}
|
}
|
||||||
|
|
||||||
test_python_smoke() {
|
test_python_smoke() {
|
||||||
# Smoke tests for H100/B200
|
# Smoke tests for H100
|
||||||
time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
|
time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 inductor/test_max_autotune $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
|
||||||
assert_git_not_dirty
|
assert_git_not_dirty
|
||||||
}
|
}
|
||||||
|
|
||||||
test_python_smoke_b200() {
|
|
||||||
# Targeted smoke tests for B200 - staged approach to avoid too many failures
|
|
||||||
time python test/run_test.py --include test_matmul_cuda inductor/test_fp8 $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
|
|
||||||
assert_git_not_dirty
|
|
||||||
}
|
|
||||||
|
|
||||||
test_h100_distributed() {
|
test_h100_distributed() {
|
||||||
# Distributed tests at H100
|
# Distributed tests at H100
|
||||||
time python test/run_test.py --include distributed/_composable/test_composability/test_pp_composability.py $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
|
time python test/run_test.py --include distributed/_composable/test_composability/test_pp_composability.py $PYTHON_TEST_EXTRA_OPTION --upload-artifacts-while-running
|
||||||
@ -390,7 +384,6 @@ test_dynamo_wrapped_shard() {
|
|||||||
--exclude-distributed-tests \
|
--exclude-distributed-tests \
|
||||||
--exclude-torch-export-tests \
|
--exclude-torch-export-tests \
|
||||||
--exclude-aot-dispatch-tests \
|
--exclude-aot-dispatch-tests \
|
||||||
--exclude-quantization-tests \
|
|
||||||
--shard "$1" "$NUM_TEST_SHARDS" \
|
--shard "$1" "$NUM_TEST_SHARDS" \
|
||||||
--verbose \
|
--verbose \
|
||||||
--upload-artifacts-while-running
|
--upload-artifacts-while-running
|
||||||
@ -568,6 +561,43 @@ else
|
|||||||
DYNAMO_BENCHMARK_FLAGS+=(--device cuda)
|
DYNAMO_BENCHMARK_FLAGS+=(--device cuda)
|
||||||
fi
|
fi
|
||||||
|
|
||||||
|
# Validate backend availability for dynamo_eager configs
|
||||||
|
if [[ "${TEST_CONFIG}" == *dynamo_eager* ]]; then
|
||||||
|
echo "Validating eager backend availability for TEST_CONFIG: ${TEST_CONFIG}"
|
||||||
|
if ! python -c "import torch; backends = torch._dynamo.list_backends(); print('Available backends:', backends); assert 'eager' in backends, f'eager backend not available. Available: {backends}'"; then
|
||||||
|
echo "ERROR: eager backend not available in this environment"
|
||||||
|
echo "This might be due to missing dependencies or incorrect PyTorch installation"
|
||||||
|
exit 1
|
||||||
|
fi
|
||||||
|
echo "eager backend validation successful"
|
||||||
|
|
||||||
|
# Additional validation: test that torch.compile works with eager backend
|
||||||
|
echo "Testing torch.compile with eager backend..."
|
||||||
|
if ! python -c "
|
||||||
|
import torch
|
||||||
|
import torch._dynamo as dynamo
|
||||||
|
|
||||||
|
def test_func(x):
|
||||||
|
return x * 2
|
||||||
|
|
||||||
|
# Test that eager backend works
|
||||||
|
try:
|
||||||
|
compiled_func = torch.compile(test_func, backend='eager')
|
||||||
|
result = compiled_func(torch.tensor([1.0, 2.0]))
|
||||||
|
print('torch.compile with eager backend test successful')
|
||||||
|
except Exception as e:
|
||||||
|
print(f'ERROR: torch.compile with eager backend failed: {e}')
|
||||||
|
exit(1)
|
||||||
|
"; then
|
||||||
|
echo "ERROR: torch.compile with eager backend failed"
|
||||||
|
exit 1
|
||||||
|
fi
|
||||||
|
fi
|
||||||
|
|
||||||
|
# Debug logging for backend selection
|
||||||
|
echo "TEST_CONFIG: ${TEST_CONFIG}"
|
||||||
|
echo "DYNAMO_BENCHMARK_FLAGS: ${DYNAMO_BENCHMARK_FLAGS[*]}"
|
||||||
|
|
||||||
test_cachebench() {
|
test_cachebench() {
|
||||||
TEST_REPORTS_DIR=$(pwd)/test/test-reports
|
TEST_REPORTS_DIR=$(pwd)/test/test-reports
|
||||||
mkdir -p "$TEST_REPORTS_DIR"
|
mkdir -p "$TEST_REPORTS_DIR"
|
||||||
@ -629,6 +659,16 @@ test_perf_for_dashboard() {
|
|||||||
shift
|
shift
|
||||||
|
|
||||||
local backend=inductor
|
local backend=inductor
|
||||||
|
# Allow surfacing eager metrics in CI by switching backend based on TEST_CONFIG
|
||||||
|
if [[ "${TEST_CONFIG}" == *dynamo_eager* ]]; then
|
||||||
|
backend=eager
|
||||||
|
elif [[ "${TEST_CONFIG}" == *aot_eager* ]]; then
|
||||||
|
backend=aot_eager
|
||||||
|
fi
|
||||||
|
|
||||||
|
# Debug logging for backend selection in test_perf_for_dashboard
|
||||||
|
echo "test_perf_for_dashboard: TEST_CONFIG=${TEST_CONFIG}, selected backend=${backend}"
|
||||||
|
echo "DASHBOARD_TAG=${DASHBOARD_TAG}"
|
||||||
local modes=()
|
local modes=()
|
||||||
if [[ "$DASHBOARD_TAG" == *training-true* ]]; then
|
if [[ "$DASHBOARD_TAG" == *training-true* ]]; then
|
||||||
modes+=(training)
|
modes+=(training)
|
||||||
@ -682,20 +722,37 @@ test_perf_for_dashboard() {
|
|||||||
fi
|
fi
|
||||||
|
|
||||||
if [[ "$DASHBOARD_TAG" == *default-true* ]]; then
|
if [[ "$DASHBOARD_TAG" == *default-true* ]]; then
|
||||||
$TASKSET python "benchmarks/dynamo/$suite.py" \
|
echo "Running benchmark: ${backend}_no_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
|
echo "Command: $TASKSET python benchmarks/dynamo/$suite.py ${target_flag[*]} --$mode --$dtype --backend $backend --disable-cudagraphs $*"
|
||||||
|
if ! $TASKSET python "benchmarks/dynamo/$suite.py" \
|
||||||
"${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" --disable-cudagraphs "$@" \
|
"${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" --disable-cudagraphs "$@" \
|
||||||
--output "$TEST_REPORTS_DIR/${backend}_no_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}.csv"
|
--output "$TEST_REPORTS_DIR/${backend}_no_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}.csv"; then
|
||||||
|
echo "ERROR: Benchmark failed for ${backend}_no_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
|
echo "This might indicate an issue with the eager backend or benchmark configuration"
|
||||||
|
exit 1
|
||||||
|
fi
|
||||||
|
echo "Benchmark completed successfully: ${backend}_no_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
fi
|
fi
|
||||||
if [[ "$DASHBOARD_TAG" == *cudagraphs-true* ]]; then
|
if [[ "$DASHBOARD_TAG" == *cudagraphs-true* ]]; then
|
||||||
$TASKSET python "benchmarks/dynamo/$suite.py" \
|
echo "Running benchmark: ${backend}_with_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
|
if ! $TASKSET python "benchmarks/dynamo/$suite.py" \
|
||||||
"${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" "$@" \
|
"${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" "$@" \
|
||||||
--output "$TEST_REPORTS_DIR/${backend}_with_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}.csv"
|
--output "$TEST_REPORTS_DIR/${backend}_with_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}.csv"; then
|
||||||
|
echo "ERROR: Benchmark failed for ${backend}_with_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
|
exit 1
|
||||||
|
fi
|
||||||
|
echo "Benchmark completed successfully: ${backend}_with_cudagraphs_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
fi
|
fi
|
||||||
if [[ "$DASHBOARD_TAG" == *dynamic-true* ]]; then
|
if [[ "$DASHBOARD_TAG" == *dynamic-true* ]]; then
|
||||||
$TASKSET python "benchmarks/dynamo/$suite.py" \
|
echo "Running benchmark: ${backend}_dynamic_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
|
if ! $TASKSET python "benchmarks/dynamo/$suite.py" \
|
||||||
"${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" --dynamic-shapes \
|
"${target_flag[@]}" --"$mode" --"$dtype" --backend "$backend" --dynamic-shapes \
|
||||||
--dynamic-batch-only "$@" \
|
--dynamic-batch-only "$@" \
|
||||||
--output "$TEST_REPORTS_DIR/${backend}_dynamic_${suite}_${dtype}_${mode}_${device}_${target}.csv"
|
--output "$TEST_REPORTS_DIR/${backend}_dynamic_${suite}_${dtype}_${mode}_${device}_${target}.csv"; then
|
||||||
|
echo "ERROR: Benchmark failed for ${backend}_dynamic_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
|
exit 1
|
||||||
|
fi
|
||||||
|
echo "Benchmark completed successfully: ${backend}_dynamic_${suite}_${dtype}_${mode}_${device}_${target}"
|
||||||
fi
|
fi
|
||||||
if [[ "$DASHBOARD_TAG" == *cppwrapper-true* ]]; then
|
if [[ "$DASHBOARD_TAG" == *cppwrapper-true* ]]; then
|
||||||
TORCHINDUCTOR_CPP_WRAPPER=1 $TASKSET python "benchmarks/dynamo/$suite.py" \
|
TORCHINDUCTOR_CPP_WRAPPER=1 $TASKSET python "benchmarks/dynamo/$suite.py" \
|
||||||
@ -1163,12 +1220,6 @@ test_distributed() {
|
|||||||
fi
|
fi
|
||||||
}
|
}
|
||||||
|
|
||||||
test_quantization() {
|
|
||||||
echo "Testing quantization"
|
|
||||||
|
|
||||||
python test/test_quantization.py
|
|
||||||
}
|
|
||||||
|
|
||||||
test_rpc() {
|
test_rpc() {
|
||||||
echo "Testing RPC C++ tests"
|
echo "Testing RPC C++ tests"
|
||||||
# NB: the ending test_rpc must match the current function name for the current
|
# NB: the ending test_rpc must match the current function name for the current
|
||||||
@ -1586,7 +1637,7 @@ test_executorch() {
|
|||||||
test_linux_aarch64() {
|
test_linux_aarch64() {
|
||||||
python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
|
python test/run_test.py --include test_modules test_mkldnn test_mkldnn_fusion test_openmp test_torch test_dynamic_shapes \
|
||||||
test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
|
test_transformers test_multiprocessing test_numpy_interop test_autograd test_binary_ufuncs test_complex test_spectral_ops \
|
||||||
test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops profiler/test_memory_profiler \
|
test_foreach test_reductions test_unary_ufuncs test_tensor_creation_ops test_ops \
|
||||||
distributed/elastic/timer/api_test distributed/elastic/timer/local_timer_example distributed/elastic/timer/local_timer_test \
|
distributed/elastic/timer/api_test distributed/elastic/timer/local_timer_example distributed/elastic/timer/local_timer_test \
|
||||||
--shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
|
--shard "$SHARD_NUMBER" "$NUM_TEST_SHARDS" --verbose
|
||||||
|
|
||||||
@ -1630,25 +1681,6 @@ test_operator_benchmark() {
|
|||||||
--expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
|
--expected "expected_ci_operator_benchmark_eager_float32_cpu.csv"
|
||||||
}
|
}
|
||||||
|
|
||||||
test_operator_microbenchmark() {
|
|
||||||
TEST_REPORTS_DIR=$(pwd)/test/test-reports
|
|
||||||
mkdir -p "$TEST_REPORTS_DIR"
|
|
||||||
TEST_DIR=$(pwd)
|
|
||||||
|
|
||||||
cd benchmarks/operator_benchmark/pt_extension
|
|
||||||
python -m pip install .
|
|
||||||
|
|
||||||
cd "${TEST_DIR}"/benchmarks/operator_benchmark
|
|
||||||
|
|
||||||
for OP_BENCHMARK_TESTS in matmul mm addmm bmm; do
|
|
||||||
$TASKSET python -m pt.${OP_BENCHMARK_TESTS}_test --tag-filter long \
|
|
||||||
--output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_microbenchmark_${OP_BENCHMARK_TESTS}_compile.json" \
|
|
||||||
--benchmark-name "PyTorch operator microbenchmark" --use-compile
|
|
||||||
$TASKSET python -m pt.${OP_BENCHMARK_TESTS}_test --tag-filter long \
|
|
||||||
--output-json-for-dashboard "${TEST_REPORTS_DIR}/operator_microbenchmark_${OP_BENCHMARK_TESTS}.json" \
|
|
||||||
--benchmark-name "PyTorch operator microbenchmark"
|
|
||||||
done
|
|
||||||
}
|
|
||||||
|
|
||||||
if ! [[ "${BUILD_ENVIRONMENT}" == *libtorch* || "${BUILD_ENVIRONMENT}" == *-bazel-* ]]; then
|
if ! [[ "${BUILD_ENVIRONMENT}" == *libtorch* || "${BUILD_ENVIRONMENT}" == *-bazel-* ]]; then
|
||||||
(cd test && python -c "import torch; print(torch.__config__.show())")
|
(cd test && python -c "import torch; print(torch.__config__.show())")
|
||||||
@ -1681,8 +1713,6 @@ elif [[ "${TEST_CONFIG}" == *executorch* ]]; then
|
|||||||
test_executorch
|
test_executorch
|
||||||
elif [[ "$TEST_CONFIG" == 'jit_legacy' ]]; then
|
elif [[ "$TEST_CONFIG" == 'jit_legacy' ]]; then
|
||||||
test_python_legacy_jit
|
test_python_legacy_jit
|
||||||
elif [[ "$TEST_CONFIG" == 'quantization' ]]; then
|
|
||||||
test_quantization
|
|
||||||
elif [[ "${BUILD_ENVIRONMENT}" == *libtorch* ]]; then
|
elif [[ "${BUILD_ENVIRONMENT}" == *libtorch* ]]; then
|
||||||
# TODO: run some C++ tests
|
# TODO: run some C++ tests
|
||||||
echo "no-op at the moment"
|
echo "no-op at the moment"
|
||||||
@ -1705,8 +1735,6 @@ elif [[ "${TEST_CONFIG}" == *operator_benchmark* ]]; then
|
|||||||
test_operator_benchmark cpu ${TEST_MODE}
|
test_operator_benchmark cpu ${TEST_MODE}
|
||||||
|
|
||||||
fi
|
fi
|
||||||
elif [[ "${TEST_CONFIG}" == *operator_microbenchmark* ]]; then
|
|
||||||
test_operator_microbenchmark
|
|
||||||
elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
|
elif [[ "${TEST_CONFIG}" == *inductor_distributed* ]]; then
|
||||||
test_inductor_distributed
|
test_inductor_distributed
|
||||||
elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
|
elif [[ "${TEST_CONFIG}" == *inductor-halide* ]]; then
|
||||||
@ -1809,14 +1837,10 @@ elif [[ "${BUILD_ENVIRONMENT}" == *xpu* ]]; then
|
|||||||
test_xpu_bin
|
test_xpu_bin
|
||||||
elif [[ "${TEST_CONFIG}" == smoke ]]; then
|
elif [[ "${TEST_CONFIG}" == smoke ]]; then
|
||||||
test_python_smoke
|
test_python_smoke
|
||||||
elif [[ "${TEST_CONFIG}" == smoke_b200 ]]; then
|
|
||||||
test_python_smoke_b200
|
|
||||||
elif [[ "${TEST_CONFIG}" == h100_distributed ]]; then
|
elif [[ "${TEST_CONFIG}" == h100_distributed ]]; then
|
||||||
test_h100_distributed
|
test_h100_distributed
|
||||||
elif [[ "${TEST_CONFIG}" == "h100-symm-mem" ]]; then
|
elif [[ "${TEST_CONFIG}" == "h100-symm-mem" ]]; then
|
||||||
test_h100_symm_mem
|
test_h100_symm_mem
|
||||||
elif [[ "${TEST_CONFIG}" == "b200-symm-mem" ]]; then
|
|
||||||
test_h100_symm_mem
|
|
||||||
elif [[ "${TEST_CONFIG}" == h100_cutlass_backend ]]; then
|
elif [[ "${TEST_CONFIG}" == h100_cutlass_backend ]]; then
|
||||||
test_h100_cutlass_backend
|
test_h100_cutlass_backend
|
||||||
else
|
else
|
||||||
|
|||||||
@ -25,7 +25,7 @@ echo Copying over test times file
|
|||||||
robocopy /E "%PYTORCH_FINAL_PACKAGE_DIR_WIN%\.additional_ci_files" "%PROJECT_DIR_WIN%\.additional_ci_files"
|
robocopy /E "%PYTORCH_FINAL_PACKAGE_DIR_WIN%\.additional_ci_files" "%PROJECT_DIR_WIN%\.additional_ci_files"
|
||||||
|
|
||||||
echo Run nn tests
|
echo Run nn tests
|
||||||
python run_test.py --exclude-jit-executor --exclude-distributed-tests --exclude-quantization-tests --shard "%SHARD_NUMBER%" "%NUM_TEST_SHARDS%" --verbose
|
python run_test.py --exclude-jit-executor --exclude-distributed-tests --shard "%SHARD_NUMBER%" "%NUM_TEST_SHARDS%" --verbose
|
||||||
if ERRORLEVEL 1 goto fail
|
if ERRORLEVEL 1 goto fail
|
||||||
|
|
||||||
popd
|
popd
|
||||||
|
|||||||
@ -63,7 +63,7 @@ if errorlevel 1 exit /b 1
|
|||||||
call %CONDA_HOME%\condabin\activate.bat testenv
|
call %CONDA_HOME%\condabin\activate.bat testenv
|
||||||
if errorlevel 1 exit /b 1
|
if errorlevel 1 exit /b 1
|
||||||
|
|
||||||
call conda install -y -q -c conda-forge libuv=1.51
|
call conda install -y -q -c conda-forge libuv=1.39
|
||||||
call conda install -y -q intel-openmp
|
call conda install -y -q intel-openmp
|
||||||
|
|
||||||
echo "install and test libtorch"
|
echo "install and test libtorch"
|
||||||
|
|||||||
@ -69,8 +69,6 @@ readability-string-compare,
|
|||||||
'
|
'
|
||||||
HeaderFilterRegex: '^(aten/|c10/|torch/).*$'
|
HeaderFilterRegex: '^(aten/|c10/|torch/).*$'
|
||||||
WarningsAsErrors: '*'
|
WarningsAsErrors: '*'
|
||||||
LineFilter:
|
|
||||||
- name: '/usr/include/.*'
|
|
||||||
CheckOptions:
|
CheckOptions:
|
||||||
cppcoreguidelines-special-member-functions.AllowSoleDefaultDtor: true
|
cppcoreguidelines-special-member-functions.AllowSoleDefaultDtor: true
|
||||||
cppcoreguidelines-special-member-functions.AllowImplicitlyDeletedCopyOrMove: true
|
cppcoreguidelines-special-member-functions.AllowImplicitlyDeletedCopyOrMove: true
|
||||||
|
|||||||
3
.github/actionlint.yaml
vendored
3
.github/actionlint.yaml
vendored
@ -22,9 +22,6 @@ self-hosted-runner:
|
|||||||
- linux.arm64.m7g.4xlarge
|
- linux.arm64.m7g.4xlarge
|
||||||
- linux.arm64.m7g.4xlarge.ephemeral
|
- linux.arm64.m7g.4xlarge.ephemeral
|
||||||
- linux.arm64.r7g.12xlarge.memory
|
- linux.arm64.r7g.12xlarge.memory
|
||||||
- linux.aws.h100
|
|
||||||
- linux.aws.h100.4
|
|
||||||
- linux.aws.h100.8
|
|
||||||
- linux.4xlarge.nvidia.gpu
|
- linux.4xlarge.nvidia.gpu
|
||||||
- linux.8xlarge.nvidia.gpu
|
- linux.8xlarge.nvidia.gpu
|
||||||
- linux.16xlarge.nvidia.gpu
|
- linux.16xlarge.nvidia.gpu
|
||||||
|
|||||||
2
.github/actions/setup-win/action.yml
vendored
2
.github/actions/setup-win/action.yml
vendored
@ -59,7 +59,7 @@ runs:
|
|||||||
set -x
|
set -x
|
||||||
|
|
||||||
# Create new py_tmp env with python-version
|
# Create new py_tmp env with python-version
|
||||||
${CONDA} create -y -n py_tmp python=${PYTHON_VERSION} intel-openmp libuv
|
${CONDA} create -y -n py_tmp python=${PYTHON_VERSION} intel-openmp
|
||||||
|
|
||||||
PYTHON3=$(${CONDA_RUN} -n py_tmp which python3)
|
PYTHON3=$(${CONDA_RUN} -n py_tmp which python3)
|
||||||
EXIT_CODE=$?
|
EXIT_CODE=$?
|
||||||
|
|||||||
2
.github/ci_commit_pins/vllm.txt
vendored
2
.github/ci_commit_pins/vllm.txt
vendored
@ -1 +1 @@
|
|||||||
da63274d9f3d06ba5815b5c8786a7194923a0234
|
367a480bd3534edf27a8dac3c6f7ea8af9d1ed45
|
||||||
|
|||||||
15
.github/merge_rules.yaml
vendored
15
.github/merge_rules.yaml
vendored
@ -525,21 +525,6 @@
|
|||||||
- Lint
|
- Lint
|
||||||
- pull
|
- pull
|
||||||
|
|
||||||
- name: typechecking
|
|
||||||
patterns:
|
|
||||||
- 'pyrefly.toml'
|
|
||||||
- 'mypy.ini'
|
|
||||||
- 'mypy-strict.ini'
|
|
||||||
approved_by:
|
|
||||||
- lolpack
|
|
||||||
- maggiemoss
|
|
||||||
- ndmitchell
|
|
||||||
- kinto0
|
|
||||||
mandatory_checks_name:
|
|
||||||
- EasyCLA
|
|
||||||
- Lint
|
|
||||||
- pull
|
|
||||||
|
|
||||||
- name: superuser
|
- name: superuser
|
||||||
patterns:
|
patterns:
|
||||||
- '*'
|
- '*'
|
||||||
|
|||||||
39
.github/pytorch-probot.yml
vendored
39
.github/pytorch-probot.yml
vendored
@ -1,44 +1,41 @@
|
|||||||
tracking_issue: 24422
|
tracking_issue: 24422
|
||||||
ciflow_tracking_issue: 64124
|
ciflow_tracking_issue: 64124
|
||||||
ciflow_push_tags:
|
ciflow_push_tags:
|
||||||
- ciflow/b200
|
|
||||||
- ciflow/b200-symm-mem
|
|
||||||
- ciflow/binaries
|
- ciflow/binaries
|
||||||
- ciflow/binaries_libtorch
|
- ciflow/binaries_libtorch
|
||||||
- ciflow/binaries_wheel
|
- ciflow/binaries_wheel
|
||||||
- ciflow/h100
|
- ciflow/triton_binaries
|
||||||
- ciflow/h100-cutlass-backend
|
|
||||||
- ciflow/h100-distributed
|
|
||||||
- ciflow/h100-symm-mem
|
|
||||||
- ciflow/inductor
|
- ciflow/inductor
|
||||||
- ciflow/inductor-cu126
|
|
||||||
- ciflow/inductor-micro-benchmark
|
|
||||||
- ciflow/inductor-micro-benchmark-cpu-x86
|
|
||||||
- ciflow/inductor-perf-compare
|
|
||||||
- ciflow/inductor-perf-test-nightly-rocm
|
|
||||||
- ciflow/inductor-perf-test-nightly-x86-zen
|
|
||||||
- ciflow/inductor-periodic
|
- ciflow/inductor-periodic
|
||||||
- ciflow/inductor-rocm
|
- ciflow/inductor-rocm
|
||||||
|
- ciflow/inductor-perf-test-nightly-rocm
|
||||||
|
- ciflow/inductor-perf-compare
|
||||||
|
- ciflow/inductor-micro-benchmark
|
||||||
|
- ciflow/inductor-micro-benchmark-cpu-x86
|
||||||
|
- ciflow/inductor-perf-test-nightly-x86-zen
|
||||||
|
- ciflow/inductor-cu126
|
||||||
- ciflow/linux-aarch64
|
- ciflow/linux-aarch64
|
||||||
- ciflow/mps
|
- ciflow/mps
|
||||||
- ciflow/nightly
|
- ciflow/nightly
|
||||||
- ciflow/op-benchmark
|
|
||||||
- ciflow/periodic
|
- ciflow/periodic
|
||||||
- ciflow/periodic-rocm-mi300
|
- ciflow/periodic-rocm-mi300
|
||||||
- ciflow/pull
|
|
||||||
- ciflow/quantization-periodic
|
|
||||||
- ciflow/riscv64
|
|
||||||
- ciflow/rocm
|
- ciflow/rocm
|
||||||
- ciflow/rocm-mi300
|
- ciflow/rocm-mi300
|
||||||
- ciflow/s390
|
- ciflow/s390
|
||||||
|
- ciflow/riscv64
|
||||||
- ciflow/slow
|
- ciflow/slow
|
||||||
- ciflow/torchbench
|
|
||||||
- ciflow/triton_binaries
|
|
||||||
- ciflow/trunk
|
- ciflow/trunk
|
||||||
- ciflow/unstable
|
- ciflow/unstable
|
||||||
- ciflow/vllm
|
|
||||||
- ciflow/win-arm64
|
|
||||||
- ciflow/xpu
|
- ciflow/xpu
|
||||||
|
- ciflow/vllm
|
||||||
|
- ciflow/torchbench
|
||||||
|
- ciflow/op-benchmark
|
||||||
|
- ciflow/pull
|
||||||
|
- ciflow/h100
|
||||||
|
- ciflow/h100-distributed
|
||||||
|
- ciflow/win-arm64
|
||||||
|
- ciflow/h100-symm-mem
|
||||||
|
- ciflow/h100-cutlass-backend
|
||||||
retryable_workflows:
|
retryable_workflows:
|
||||||
- pull
|
- pull
|
||||||
- trunk
|
- trunk
|
||||||
@ -47,4 +44,4 @@ retryable_workflows:
|
|||||||
- inductor-A100-perf-nightly
|
- inductor-A100-perf-nightly
|
||||||
labeler_config: labeler.yml
|
labeler_config: labeler.yml
|
||||||
label_to_label_config: label_to_label.yml
|
label_to_label_config: label_to_label.yml
|
||||||
mergebot: true
|
mergebot: True
|
||||||
|
|||||||
@ -30,7 +30,7 @@ CUDA_ARCHES_CUDNN_VERSION = {
|
|||||||
}
|
}
|
||||||
|
|
||||||
# NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
|
# NOTE: Please also update the ROCm sources in `PIP_SOURCES` in tools/nightly.py when changing this
|
||||||
ROCM_ARCHES = ["6.4", "7.0"]
|
ROCM_ARCHES = ["6.3", "6.4"]
|
||||||
|
|
||||||
XPU_ARCHES = ["xpu"]
|
XPU_ARCHES = ["xpu"]
|
||||||
|
|
||||||
@ -53,7 +53,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
|
|||||||
"nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
|
"nvidia-cusolver-cu12==11.7.1.2; platform_system == 'Linux' | "
|
||||||
"nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
|
"nvidia-cusparse-cu12==12.5.4.2; platform_system == 'Linux' | "
|
||||||
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
|
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
|
||||||
"nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | "
|
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
|
||||||
"nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
|
"nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
|
||||||
"nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
|
"nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | "
|
||||||
"nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
|
"nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | "
|
||||||
@ -70,7 +70,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
|
|||||||
"nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
|
"nvidia-cusolver-cu12==11.7.3.90; platform_system == 'Linux' | "
|
||||||
"nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
|
"nvidia-cusparse-cu12==12.5.8.93; platform_system == 'Linux' | "
|
||||||
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
|
"nvidia-cusparselt-cu12==0.7.1; platform_system == 'Linux' | "
|
||||||
"nvidia-nccl-cu12==2.28.3; platform_system == 'Linux' | "
|
"nvidia-nccl-cu12==2.27.5; platform_system == 'Linux' | "
|
||||||
"nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
|
"nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | "
|
||||||
"nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
|
"nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | "
|
||||||
"nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
|
"nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | "
|
||||||
@ -87,7 +87,7 @@ PYTORCH_EXTRA_INSTALL_REQUIREMENTS = {
|
|||||||
"nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
|
"nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | "
|
||||||
"nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
|
"nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | "
|
||||||
"nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
|
"nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | "
|
||||||
"nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | "
|
"nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | "
|
||||||
"nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
|
"nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | "
|
||||||
"nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
|
"nvidia-nvtx==13.0.39; platform_system == 'Linux' | "
|
||||||
"nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
|
"nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | "
|
||||||
|
|||||||
2
.github/scripts/generate_ci_workflows.py
vendored
2
.github/scripts/generate_ci_workflows.py
vendored
@ -155,7 +155,7 @@ LINUX_BINARY_SMOKE_WORKFLOWS = [
|
|||||||
package_type="manywheel",
|
package_type="manywheel",
|
||||||
build_configs=generate_binary_build_matrix.generate_wheels_matrix(
|
build_configs=generate_binary_build_matrix.generate_wheels_matrix(
|
||||||
OperatingSystem.LINUX,
|
OperatingSystem.LINUX,
|
||||||
arches=["13.0"],
|
arches=["12.8"],
|
||||||
python_versions=["3.12"],
|
python_versions=["3.12"],
|
||||||
),
|
),
|
||||||
branches="main",
|
branches="main",
|
||||||
|
|||||||
@ -71,15 +71,12 @@ jobs:
|
|||||||
with:!{{ upload.binary_env_as_input(config) }}
|
with:!{{ upload.binary_env_as_input(config) }}
|
||||||
{%- if "aarch64" in build_environment %}
|
{%- if "aarch64" in build_environment %}
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
{%- elif "s390x" in build_environment %}
|
{%- elif "s390x" in build_environment %}
|
||||||
runs_on: linux.s390x
|
runs_on: linux.s390x
|
||||||
ALPINE_IMAGE: "docker.io/s390x/alpine"
|
ALPINE_IMAGE: "docker.io/s390x/alpine"
|
||||||
timeout-minutes: 420
|
timeout-minutes: 420
|
||||||
{%- elif config["gpu_arch_type"] == "rocm" %}
|
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
|
||||||
timeout-minutes: 300
|
|
||||||
{%- elif "conda" in build_environment and config["gpu_arch_type"] == "cuda" %}
|
{%- elif "conda" in build_environment and config["gpu_arch_type"] == "cuda" %}
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.24xlarge.ephemeral
|
runs_on: linux.24xlarge.ephemeral
|
||||||
|
|||||||
2
.github/workflows/_docs.yml
vendored
2
.github/workflows/_docs.yml
vendored
@ -67,7 +67,7 @@ jobs:
|
|||||||
# an OOM issue when running the job, so this upgrades the runner from 4xlarge
|
# an OOM issue when running the job, so this upgrades the runner from 4xlarge
|
||||||
# to the next available tier of 12xlarge. So much memory just to generate cpp
|
# to the next available tier of 12xlarge. So much memory just to generate cpp
|
||||||
# doc
|
# doc
|
||||||
runner: ${{ inputs.runner_prefix }}linux.12xlarge.memory
|
runner: ${{ inputs.runner_prefix }}linux.12xlarge
|
||||||
# TODO: Nightly cpp docs take longer and longer to finish (more than 3h now)
|
# TODO: Nightly cpp docs take longer and longer to finish (more than 3h now)
|
||||||
# Let's try to figure out how this can be improved
|
# Let's try to figure out how this can be improved
|
||||||
timeout-minutes: 360
|
timeout-minutes: 360
|
||||||
|
|||||||
28
.github/workflows/_get-changed-files.yml
vendored
28
.github/workflows/_get-changed-files.yml
vendored
@ -2,12 +2,6 @@ name: Get Changed Files
|
|||||||
|
|
||||||
on:
|
on:
|
||||||
workflow_call:
|
workflow_call:
|
||||||
inputs:
|
|
||||||
all_files:
|
|
||||||
description: "Whether to return all files instead of just changed files"
|
|
||||||
required: false
|
|
||||||
type: boolean
|
|
||||||
default: false
|
|
||||||
outputs:
|
outputs:
|
||||||
changed-files:
|
changed-files:
|
||||||
description: "List of changed files (space-separated) or '*' if not in a PR"
|
description: "List of changed files (space-separated) or '*' if not in a PR"
|
||||||
@ -32,23 +26,17 @@ jobs:
|
|||||||
# Get the PR number from the github context
|
# Get the PR number from the github context
|
||||||
PR_NUMBER="${{ github.event.number }}"
|
PR_NUMBER="${{ github.event.number }}"
|
||||||
|
|
||||||
# Check if all_files is requested
|
# Use gh CLI to get changed files in the PR with explicit repo
|
||||||
if [ "${{ inputs.all_files }}" = "true" ]; then
|
CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
|
||||||
echo "all_files input is true, returning all files"
|
|
||||||
echo "changed-files=*" >> "$GITHUB_OUTPUT"
|
|
||||||
else
|
|
||||||
# Use gh CLI to get changed files in the PR with explicit repo
|
|
||||||
CHANGED_FILES=$(gh api repos/${{ github.repository }}/pulls/$PR_NUMBER/files --paginate --jq '.[] | select(.status != "removed") | .filename' | tr '\n' ' ' | sed 's/ $//')
|
|
||||||
|
|
||||||
if [ -z "$CHANGED_FILES" ]; then
|
if [ -z "$CHANGED_FILES" ]; then
|
||||||
echo "No changed files found, setting to '*'"
|
echo "No changed files found, setting to '*'"
|
||||||
CHANGED_FILES="*"
|
CHANGED_FILES="*"
|
||||||
fi
|
|
||||||
|
|
||||||
echo "Changed files: $CHANGED_FILES"
|
|
||||||
echo "changed-files=$CHANGED_FILES" >> "$GITHUB_OUTPUT"
|
|
||||||
fi
|
fi
|
||||||
|
|
||||||
|
echo "Changed files: $CHANGED_FILES"
|
||||||
|
echo "changed-files=$CHANGED_FILES" >> "$GITHUB_OUTPUT"
|
||||||
|
|
||||||
else
|
else
|
||||||
echo "Not in PR context, setting changed files to '*'"
|
echo "Not in PR context, setting changed files to '*'"
|
||||||
echo "changed-files=*" >> "$GITHUB_OUTPUT"
|
echo "changed-files=*" >> "$GITHUB_OUTPUT"
|
||||||
|
|||||||
2
.github/workflows/_linux-test.yml
vendored
2
.github/workflows/_linux-test.yml
vendored
@ -273,8 +273,6 @@ jobs:
|
|||||||
TEST_CONFIG: ${{ matrix.config }}
|
TEST_CONFIG: ${{ matrix.config }}
|
||||||
SHARD_NUMBER: ${{ matrix.shard }}
|
SHARD_NUMBER: ${{ matrix.shard }}
|
||||||
NUM_TEST_SHARDS: ${{ matrix.num_shards }}
|
NUM_TEST_SHARDS: ${{ matrix.num_shards }}
|
||||||
EXTRA_FLAGS: ${{ matrix.extra_flags || '' }}
|
|
||||||
OP_BENCHMARK_TESTS: ${{ matrix.op_benchmark_tests }}
|
|
||||||
REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
|
REENABLED_ISSUES: ${{ steps.keep-going.outputs.reenabled-issues }}
|
||||||
CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
|
CONTINUE_THROUGH_ERROR: ${{ steps.keep-going.outputs.keep-going }}
|
||||||
VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
|
VERBOSE_TEST_LOGS: ${{ steps.keep-going.outputs.ci-verbose-test-logs }}
|
||||||
|
|||||||
60
.github/workflows/b200-symm-mem.yml
vendored
60
.github/workflows/b200-symm-mem.yml
vendored
@ -1,60 +0,0 @@
|
|||||||
name: Limited CI for symmetric memory tests on B200
|
|
||||||
|
|
||||||
on:
|
|
||||||
pull_request:
|
|
||||||
paths:
|
|
||||||
- .github/workflows/b200-symm-mem.yml
|
|
||||||
workflow_dispatch:
|
|
||||||
push:
|
|
||||||
tags:
|
|
||||||
- ciflow/b200-symm-mem/*
|
|
||||||
schedule:
|
|
||||||
- cron: 22 8 * * * # about 1:22am PDT
|
|
||||||
|
|
||||||
concurrency:
|
|
||||||
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
|
|
||||||
cancel-in-progress: true
|
|
||||||
|
|
||||||
permissions:
|
|
||||||
id-token: write
|
|
||||||
contents: read
|
|
||||||
|
|
||||||
jobs:
|
|
||||||
|
|
||||||
get-label-type:
|
|
||||||
if: github.repository_owner == 'pytorch'
|
|
||||||
name: get-label-type
|
|
||||||
uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
|
|
||||||
with:
|
|
||||||
triggering_actor: ${{ github.triggering_actor }}
|
|
||||||
issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
|
|
||||||
curr_branch: ${{ github.head_ref || github.ref_name }}
|
|
||||||
curr_ref_type: ${{ github.ref_type }}
|
|
||||||
|
|
||||||
linux-jammy-cuda12_8-py3_10-gcc11-sm100-build-symm:
|
|
||||||
name: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
|
|
||||||
uses: ./.github/workflows/_linux-build.yml
|
|
||||||
needs: get-label-type
|
|
||||||
with:
|
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
|
||||||
runner: linux.12xlarge.memory
|
|
||||||
build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
|
|
||||||
docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
|
|
||||||
cuda-arch-list: '10.0'
|
|
||||||
test-matrix: |
|
|
||||||
{ include: [
|
|
||||||
{ config: "b200-symm-mem", shard: 1, num_shards: 1, runner: "linux.dgx.b200.8" },
|
|
||||||
]}
|
|
||||||
secrets: inherit
|
|
||||||
|
|
||||||
linux-jammy-cuda12_8-py3_10-gcc11-sm100-test:
|
|
||||||
name: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
|
|
||||||
uses: ./.github/workflows/_linux-test.yml
|
|
||||||
needs:
|
|
||||||
- linux-jammy-cuda12_8-py3_10-gcc11-sm100-build-symm
|
|
||||||
with:
|
|
||||||
build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100-symm
|
|
||||||
docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build-symm.outputs.docker-image }}
|
|
||||||
test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build-symm.outputs.test-matrix }}
|
|
||||||
aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
|
|
||||||
secrets: inherit
|
|
||||||
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
|
runs-on: linux.9xlarge.ephemeral
|
||||||
strategy:
|
strategy:
|
||||||
matrix:
|
matrix:
|
||||||
tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.3", "rocm6.4", "rocm7.0", "cpu"]
|
tag: ["cuda12.6", "cuda12.8", "cuda12.9", "cuda13.0", "rocm6.3", "rocm6.4", "cpu"]
|
||||||
steps:
|
steps:
|
||||||
- name: Build docker image
|
- name: Build docker image
|
||||||
uses: pytorch/pytorch/.github/actions/binary-docker-build@main
|
uses: pytorch/pytorch/.github/actions/binary-docker-build@main
|
||||||
|
|||||||
2
.github/workflows/build-libtorch-images.yml
vendored
2
.github/workflows/build-libtorch-images.yml
vendored
@ -52,8 +52,8 @@ jobs:
|
|||||||
{ tag: "cuda12.9" },
|
{ tag: "cuda12.9" },
|
||||||
{ tag: "cuda12.8" },
|
{ tag: "cuda12.8" },
|
||||||
{ tag: "cuda12.6" },
|
{ tag: "cuda12.6" },
|
||||||
|
{ tag: "rocm6.3" },
|
||||||
{ tag: "rocm6.4" },
|
{ tag: "rocm6.4" },
|
||||||
{ tag: "rocm7.0" },
|
|
||||||
{ tag: "cpu" },
|
{ tag: "cpu" },
|
||||||
]
|
]
|
||||||
steps:
|
steps:
|
||||||
|
|||||||
2
.github/workflows/build-magma-rocm-linux.yml
vendored
2
.github/workflows/build-magma-rocm-linux.yml
vendored
@ -34,7 +34,7 @@ jobs:
|
|||||||
id-token: write
|
id-token: write
|
||||||
strategy:
|
strategy:
|
||||||
matrix:
|
matrix:
|
||||||
rocm_version: ["70", "64"]
|
rocm_version: ["64", "63"]
|
||||||
steps:
|
steps:
|
||||||
- name: Checkout PyTorch
|
- name: Checkout PyTorch
|
||||||
uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
|
uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
|
||||||
|
|||||||
2
.github/workflows/build-manywheel-images.yml
vendored
2
.github/workflows/build-manywheel-images.yml
vendored
@ -52,8 +52,8 @@ jobs:
|
|||||||
{ name: "manylinuxaarch64-builder", tag: "cuda13.0", runner: "linux.arm64.2xlarge.ephemeral" },
|
{ name: "manylinuxaarch64-builder", tag: "cuda13.0", runner: "linux.arm64.2xlarge.ephemeral" },
|
||||||
{ name: "manylinuxaarch64-builder", tag: "cuda12.8", runner: "linux.arm64.2xlarge.ephemeral" },
|
{ name: "manylinuxaarch64-builder", tag: "cuda12.8", runner: "linux.arm64.2xlarge.ephemeral" },
|
||||||
{ name: "manylinuxaarch64-builder", tag: "cuda12.6", runner: "linux.arm64.2xlarge.ephemeral" },
|
{ name: "manylinuxaarch64-builder", tag: "cuda12.6", runner: "linux.arm64.2xlarge.ephemeral" },
|
||||||
|
{ name: "manylinux2_28-builder", tag: "rocm6.3", runner: "linux.9xlarge.ephemeral" },
|
||||||
{ name: "manylinux2_28-builder", tag: "rocm6.4", runner: "linux.9xlarge.ephemeral" },
|
{ name: "manylinux2_28-builder", tag: "rocm6.4", runner: "linux.9xlarge.ephemeral" },
|
||||||
{ name: "manylinux2_28-builder", tag: "rocm7.0", runner: "linux.9xlarge.ephemeral" },
|
|
||||||
{ name: "manylinux2_28-builder", tag: "cpu", 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_aarch64-builder", tag: "cpu-aarch64", runner: "linux.arm64.2xlarge.ephemeral" },
|
||||||
{ name: "manylinuxcxx11-abi-builder", tag: "cpu-cxx11-abi", runner: "linux.9xlarge.ephemeral" },
|
{ name: "manylinuxcxx11-abi-builder", tag: "cpu-cxx11-abi", runner: "linux.9xlarge.ephemeral" },
|
||||||
|
|||||||
9
.github/workflows/build-triton-wheel.yml
vendored
9
.github/workflows/build-triton-wheel.yml
vendored
@ -50,12 +50,12 @@ jobs:
|
|||||||
strategy:
|
strategy:
|
||||||
fail-fast: false
|
fail-fast: false
|
||||||
matrix:
|
matrix:
|
||||||
py_vers: [ "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
|
py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
|
||||||
device: ["cuda", "rocm", "xpu", "aarch64"]
|
device: ["cuda", "rocm", "xpu", "aarch64"]
|
||||||
docker-image: ["pytorch/manylinux2_28-builder:cpu"]
|
docker-image: ["pytorch/manylinux2_28-builder:cpu"]
|
||||||
include:
|
include:
|
||||||
- device: "rocm"
|
- device: "rocm"
|
||||||
rocm_version: "7.0"
|
rocm_version: "6.4"
|
||||||
runs_on: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge"
|
runs_on: "${{ needs.get-label-type.outputs.label-type }}linux.4xlarge"
|
||||||
- device: "cuda"
|
- device: "cuda"
|
||||||
rocm_version: ""
|
rocm_version: ""
|
||||||
@ -108,6 +108,9 @@ jobs:
|
|||||||
|
|
||||||
# Determine python executable for given version
|
# Determine python executable for given version
|
||||||
case $PY_VERS in
|
case $PY_VERS in
|
||||||
|
3.9)
|
||||||
|
PYTHON_EXECUTABLE=/opt/python/cp39-cp39/bin/python
|
||||||
|
;;
|
||||||
3.10)
|
3.10)
|
||||||
PYTHON_EXECUTABLE=/opt/python/cp310-cp310/bin/python
|
PYTHON_EXECUTABLE=/opt/python/cp310-cp310/bin/python
|
||||||
;;
|
;;
|
||||||
@ -191,7 +194,7 @@ jobs:
|
|||||||
strategy:
|
strategy:
|
||||||
fail-fast: false
|
fail-fast: false
|
||||||
matrix:
|
matrix:
|
||||||
py_vers: [ "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
|
py_vers: [ "3.9", "3.10", "3.11", "3.12", "3.13", "3.13t", "3.14", "3.14t" ]
|
||||||
device: ["xpu"]
|
device: ["xpu"]
|
||||||
timeout-minutes: 40
|
timeout-minutes: 40
|
||||||
env:
|
env:
|
||||||
|
|||||||
59
.github/workflows/create_release.yml
vendored
59
.github/workflows/create_release.yml
vendored
@ -35,7 +35,6 @@ jobs:
|
|||||||
contents: write
|
contents: write
|
||||||
outputs:
|
outputs:
|
||||||
pt_release_name: ${{ steps.release_name.outputs.pt_release_name }}
|
pt_release_name: ${{ steps.release_name.outputs.pt_release_name }}
|
||||||
pt_pep517_release_name: ${{ steps.release_name.outputs.pt_pep517_release_name }}
|
|
||||||
steps:
|
steps:
|
||||||
- uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
|
- uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
|
||||||
with:
|
with:
|
||||||
@ -54,12 +53,8 @@ jobs:
|
|||||||
tag_or_branch="${tag_or_branch#refs/heads/}"
|
tag_or_branch="${tag_or_branch#refs/heads/}"
|
||||||
# replace directory separators with _ in branch name
|
# replace directory separators with _ in branch name
|
||||||
tag_or_branch="${tag_or_branch//\//_}"
|
tag_or_branch="${tag_or_branch//\//_}"
|
||||||
torch_version="$(python -c 'from tools.generate_torch_version import get_torch_version; print(get_torch_version())')"
|
echo "PT_RELEASE_NAME=pytorch-$tag_or_branch" >> "$GITHUB_ENV"
|
||||||
{
|
echo "PT_RELEASE_FILE=pytorch-$tag_or_branch.tar.gz" >> "$GITHUB_ENV"
|
||||||
echo "PT_RELEASE_NAME=pytorch-$tag_or_branch";
|
|
||||||
echo "PT_RELEASE_FILE=pytorch-$tag_or_branch.tar.gz";
|
|
||||||
echo "PT_PEP517_RELEASE_FILE=torch-${torch_version}.tar.gz";
|
|
||||||
} >> "$GITHUB_ENV"
|
|
||||||
- name: Checkout optional submodules
|
- name: Checkout optional submodules
|
||||||
run: python3 tools/optional_submodules.py
|
run: python3 tools/optional_submodules.py
|
||||||
- name: Copy docs requirements for inclusion
|
- name: Copy docs requirements for inclusion
|
||||||
@ -69,47 +64,30 @@ jobs:
|
|||||||
cp .ci/docker/requirements-docs.txt docs/requirements.txt
|
cp .ci/docker/requirements-docs.txt docs/requirements.txt
|
||||||
- name: Create source distribution
|
- name: Create source distribution
|
||||||
run: |
|
run: |
|
||||||
# Create new folder with specified name so extracting the archive yields that
|
# Create new folder with specified name so extracting the archive yields that
|
||||||
rm -rf "/tmp/$PT_RELEASE_NAME"
|
rm -rf "/tmp/$PT_RELEASE_NAME"
|
||||||
cp -r "$PWD" "/tmp/$PT_RELEASE_NAME"
|
cp -r "$PWD" "/tmp/$PT_RELEASE_NAME"
|
||||||
mv "/tmp/$PT_RELEASE_NAME" .
|
mv "/tmp/$PT_RELEASE_NAME" .
|
||||||
# Cleanup
|
# Cleanup
|
||||||
rm -rf "$PT_RELEASE_NAME"/{.circleci,.ci}
|
rm -rf "$PT_RELEASE_NAME"/{.circleci,.ci}
|
||||||
find "$PT_RELEASE_NAME" -name '.git*' -exec rm -rv {} \; || true
|
find "$PT_RELEASE_NAME" -name '.git*' -exec rm -rv {} \; || true
|
||||||
# Create archive
|
# Create archive
|
||||||
tar -czf "$PT_RELEASE_FILE" "$PT_RELEASE_NAME"
|
tar -czf "$PT_RELEASE_FILE" "$PT_RELEASE_NAME"
|
||||||
echo "Created source archive $PT_RELEASE_FILE with content: $(ls -a "$PT_RELEASE_NAME")"
|
echo "Created source archive $PT_RELEASE_FILE with content: $(ls -a "$PT_RELEASE_NAME")"
|
||||||
- name: Create PEP 517 compatible source distribution
|
|
||||||
run: |
|
|
||||||
pip install build==1.2.2.post1 || exit 1
|
|
||||||
python -m build --sdist || exit 1
|
|
||||||
cd dist || exit 1
|
|
||||||
- name: Upload source distribution for release
|
- name: Upload source distribution for release
|
||||||
if: ${{ github.event_name == 'release' }}
|
if: ${{ github.event_name == 'release' }}
|
||||||
uses: softprops/action-gh-release@da05d552573ad5aba039eaac05058a918a7bf631 # v2.2.2
|
uses: softprops/action-gh-release@da05d552573ad5aba039eaac05058a918a7bf631 # v2.2.2
|
||||||
with:
|
with:
|
||||||
files: |
|
files: ${{env.PT_RELEASE_FILE}}
|
||||||
${{ env.PT_RELEASE_FILE }}
|
- name: Upload source distribution to GHA artifacts for release tags
|
||||||
${{ env.PT_PEP517_RELEASE_FILE }}
|
|
||||||
- name: Upload source distribution to GHA artifacts # for release tags
|
|
||||||
if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
|
if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
|
||||||
uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
|
uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
|
||||||
with:
|
with:
|
||||||
name: ${{ env.PT_RELEASE_FILE }}
|
name: ${{ env.PT_RELEASE_FILE }}
|
||||||
path: ${{ env.PT_RELEASE_FILE }}
|
path: ${{ env.PT_RELEASE_FILE }}
|
||||||
- name: Upload PEP 517 source distribution to GHA artifacts # for release tags
|
|
||||||
if: ${{ github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
|
|
||||||
uses: actions/upload-artifact@50769540e7f4bd5e21e526ee35c689e35e0d6874 # v4.4.0
|
|
||||||
with:
|
|
||||||
name: ${{ env.PT_PEP517_RELEASE_FILE }}
|
|
||||||
path: dist/${{ env.PT_PEP517_RELEASE_FILE }}
|
|
||||||
- name: Set output
|
- name: Set output
|
||||||
id: release_name
|
id: release_name
|
||||||
run: |
|
run: echo "pt_release_name=${{ env.PT_RELEASE_NAME }}.tar.gz" >> "${GITHUB_OUTPUT}"
|
||||||
{
|
|
||||||
echo "pt_release_name=${{ env.PT_RELEASE_FILE }}";
|
|
||||||
echo "pt_pep517_release_name=${{ env.PT_PEP517_RELEASE_FILE }}";
|
|
||||||
} >> "${GITHUB_OUTPUT}"
|
|
||||||
|
|
||||||
upload_source_code_to_s3:
|
upload_source_code_to_s3:
|
||||||
if: ${{ github.repository == 'pytorch/pytorch' && github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
|
if: ${{ github.repository == 'pytorch/pytorch' && github.event_name == 'push' && startsWith(github.ref, 'refs/tags/v') && contains(github.ref, 'rc') }}
|
||||||
@ -125,9 +103,6 @@ jobs:
|
|||||||
- uses: actions/download-artifact@65a9edc5881444af0b9093a5e628f2fe47ea3b2e # v4.1.7
|
- uses: actions/download-artifact@65a9edc5881444af0b9093a5e628f2fe47ea3b2e # v4.1.7
|
||||||
with:
|
with:
|
||||||
name: ${{ needs.release.outputs.pt_release_name }}
|
name: ${{ needs.release.outputs.pt_release_name }}
|
||||||
- uses: actions/download-artifact@65a9edc5881444af0b9093a5e628f2fe47ea3b2e # v4.1.7
|
|
||||||
with:
|
|
||||||
name: ${{ needs.release.outputs.pt_pep517_release_name }}
|
|
||||||
- name: Configure AWS credentials(PyTorch account)
|
- name: Configure AWS credentials(PyTorch account)
|
||||||
uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
|
uses: aws-actions/configure-aws-credentials@ececac1a45f3b08a01d2dd070d28d111c5fe6722 # v4.1.0
|
||||||
with:
|
with:
|
||||||
@ -138,9 +113,7 @@ jobs:
|
|||||||
s3-bucket: pytorch
|
s3-bucket: pytorch
|
||||||
s3-prefix: source_code/test
|
s3-prefix: source_code/test
|
||||||
if-no-files-found: warn
|
if-no-files-found: warn
|
||||||
path: |
|
path: ${{ needs.release.outputs.pt_release_name }}
|
||||||
${{ needs.release.outputs.pt_release_name }}
|
|
||||||
${{ needs.release.outputs.pt_pep517_release_name }}
|
|
||||||
|
|
||||||
concurrency:
|
concurrency:
|
||||||
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name }}
|
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name }}
|
||||||
|
|||||||
2
.github/workflows/docker-builds.yml
vendored
2
.github/workflows/docker-builds.yml
vendored
@ -70,7 +70,7 @@ jobs:
|
|||||||
pytorch-linux-jammy-py3-clang18-asan,
|
pytorch-linux-jammy-py3-clang18-asan,
|
||||||
pytorch-linux-jammy-py3-clang12-onnx,
|
pytorch-linux-jammy-py3-clang12-onnx,
|
||||||
pytorch-linux-jammy-linter,
|
pytorch-linux-jammy-linter,
|
||||||
pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter,
|
pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter,
|
||||||
pytorch-linux-jammy-py3-clang12-executorch,
|
pytorch-linux-jammy-py3-clang12-executorch,
|
||||||
pytorch-linux-jammy-py3.12-triton-cpu,
|
pytorch-linux-jammy-py3.12-triton-cpu,
|
||||||
pytorch-linux-noble-riscv64-py3.12-gcc14
|
pytorch-linux-noble-riscv64-py3.12-gcc14
|
||||||
|
|||||||
98
.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
generated
vendored
98
.github/workflows/generated-linux-aarch64-binary-manywheel-nightly.yml
generated
vendored
@ -62,7 +62,7 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
||||||
DESIRED_PYTHON: "3.10"
|
DESIRED_PYTHON: "3.10"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_10-cpu-aarch64
|
build_name: manywheel-py3_10-cpu-aarch64
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
@ -128,11 +128,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
||||||
DESIRED_PYTHON: "3.10"
|
DESIRED_PYTHON: "3.10"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_10-cuda-aarch64-12_6
|
build_name: manywheel-py3_10-cuda-aarch64-12_6
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -174,11 +174,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.10"
|
DESIRED_PYTHON: "3.10"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_10-cuda-aarch64-12_8
|
build_name: manywheel-py3_10-cuda-aarch64-12_8
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -220,11 +220,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
||||||
DESIRED_PYTHON: "3.10"
|
DESIRED_PYTHON: "3.10"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_10-cuda-aarch64-13_0
|
build_name: manywheel-py3_10-cuda-aarch64-13_0
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
||||||
timeout-minutes: 420
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -265,7 +265,7 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
||||||
DESIRED_PYTHON: "3.11"
|
DESIRED_PYTHON: "3.11"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_11-cpu-aarch64
|
build_name: manywheel-py3_11-cpu-aarch64
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
@ -331,11 +331,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
||||||
DESIRED_PYTHON: "3.11"
|
DESIRED_PYTHON: "3.11"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_11-cuda-aarch64-12_6
|
build_name: manywheel-py3_11-cuda-aarch64-12_6
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -377,11 +377,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.11"
|
DESIRED_PYTHON: "3.11"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_11-cuda-aarch64-12_8
|
build_name: manywheel-py3_11-cuda-aarch64-12_8
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -423,11 +423,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
||||||
DESIRED_PYTHON: "3.11"
|
DESIRED_PYTHON: "3.11"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_11-cuda-aarch64-13_0
|
build_name: manywheel-py3_11-cuda-aarch64-13_0
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
||||||
timeout-minutes: 420
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -468,7 +468,7 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
||||||
DESIRED_PYTHON: "3.12"
|
DESIRED_PYTHON: "3.12"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_12-cpu-aarch64
|
build_name: manywheel-py3_12-cpu-aarch64
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
@ -534,11 +534,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
||||||
DESIRED_PYTHON: "3.12"
|
DESIRED_PYTHON: "3.12"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_12-cuda-aarch64-12_6
|
build_name: manywheel-py3_12-cuda-aarch64-12_6
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -580,11 +580,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.12"
|
DESIRED_PYTHON: "3.12"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_12-cuda-aarch64-12_8
|
build_name: manywheel-py3_12-cuda-aarch64-12_8
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -626,11 +626,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
||||||
DESIRED_PYTHON: "3.12"
|
DESIRED_PYTHON: "3.12"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_12-cuda-aarch64-13_0
|
build_name: manywheel-py3_12-cuda-aarch64-13_0
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
||||||
timeout-minutes: 420
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -671,7 +671,7 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
||||||
DESIRED_PYTHON: "3.13"
|
DESIRED_PYTHON: "3.13"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_13-cpu-aarch64
|
build_name: manywheel-py3_13-cpu-aarch64
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
@ -737,11 +737,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
||||||
DESIRED_PYTHON: "3.13"
|
DESIRED_PYTHON: "3.13"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_13-cuda-aarch64-12_6
|
build_name: manywheel-py3_13-cuda-aarch64-12_6
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -783,11 +783,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.13"
|
DESIRED_PYTHON: "3.13"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_13-cuda-aarch64-12_8
|
build_name: manywheel-py3_13-cuda-aarch64-12_8
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -829,11 +829,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
||||||
DESIRED_PYTHON: "3.13"
|
DESIRED_PYTHON: "3.13"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_13-cuda-aarch64-13_0
|
build_name: manywheel-py3_13-cuda-aarch64-13_0
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
||||||
timeout-minutes: 420
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -874,7 +874,7 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
||||||
DESIRED_PYTHON: "3.13t"
|
DESIRED_PYTHON: "3.13t"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_13t-cpu-aarch64
|
build_name: manywheel-py3_13t-cpu-aarch64
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
@ -940,11 +940,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
||||||
DESIRED_PYTHON: "3.13t"
|
DESIRED_PYTHON: "3.13t"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_13t-cuda-aarch64-12_6
|
build_name: manywheel-py3_13t-cuda-aarch64-12_6
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -986,11 +986,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.13t"
|
DESIRED_PYTHON: "3.13t"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_13t-cuda-aarch64-12_8
|
build_name: manywheel-py3_13t-cuda-aarch64-12_8
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -1032,11 +1032,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
||||||
DESIRED_PYTHON: "3.13t"
|
DESIRED_PYTHON: "3.13t"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_13t-cuda-aarch64-13_0
|
build_name: manywheel-py3_13t-cuda-aarch64-13_0
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
||||||
timeout-minutes: 420
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -1077,7 +1077,7 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
||||||
DESIRED_PYTHON: "3.14"
|
DESIRED_PYTHON: "3.14"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_14-cpu-aarch64
|
build_name: manywheel-py3_14-cpu-aarch64
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
@ -1143,11 +1143,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
||||||
DESIRED_PYTHON: "3.14"
|
DESIRED_PYTHON: "3.14"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_14-cuda-aarch64-12_6
|
build_name: manywheel-py3_14-cuda-aarch64-12_6
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -1189,11 +1189,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.14"
|
DESIRED_PYTHON: "3.14"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_14-cuda-aarch64-12_8
|
build_name: manywheel-py3_14-cuda-aarch64-12_8
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -1235,11 +1235,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
||||||
DESIRED_PYTHON: "3.14"
|
DESIRED_PYTHON: "3.14"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_14-cuda-aarch64-13_0
|
build_name: manywheel-py3_14-cuda-aarch64-13_0
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
||||||
timeout-minutes: 420
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -1280,7 +1280,7 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
DOCKER_IMAGE_TAG_PREFIX: cpu-aarch64
|
||||||
DESIRED_PYTHON: "3.14t"
|
DESIRED_PYTHON: "3.14t"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_14t-cpu-aarch64
|
build_name: manywheel-py3_14t-cpu-aarch64
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
@ -1346,11 +1346,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.6
|
||||||
DESIRED_PYTHON: "3.14t"
|
DESIRED_PYTHON: "3.14t"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_14t-cuda-aarch64-12_6
|
build_name: manywheel-py3_14t-cuda-aarch64-12_6
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.6.77; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.6.85; platform_system == 'Linux' | nvidia-cufile-cu12==1.11.1.6; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.6.77; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -1392,11 +1392,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.14t"
|
DESIRED_PYTHON: "3.14t"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_14t-cuda-aarch64-12_8
|
build_name: manywheel-py3_14t-cuda-aarch64-12_8
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
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==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu12==3.3.24; platform_system == 'Linux' | nvidia-nvtx-cu12==12.8.90; platform_system == 'Linux' | nvidia-nvjitlink-cu12==12.8.93; platform_system == 'Linux' | nvidia-cufile-cu12==1.13.1.3; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-cu12==12.8.93; platform_system == 'Linux' | 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
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
@ -1438,11 +1438,11 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
||||||
DESIRED_PYTHON: "3.14t"
|
DESIRED_PYTHON: "3.14t"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.arm64.r7g.12xlarge.memory
|
runs_on: linux.arm64.m7g.4xlarge.ephemeral
|
||||||
ALPINE_IMAGE: "arm64v8/alpine"
|
ALPINE_IMAGE: "arm64v8/alpine"
|
||||||
build_name: manywheel-py3_14t-cuda-aarch64-13_0
|
build_name: manywheel-py3_14t-cuda-aarch64-13_0
|
||||||
build_environment: linux-aarch64-binary-manywheel
|
build_environment: linux-aarch64-binary-manywheel
|
||||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.27.7; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
||||||
timeout-minutes: 420
|
timeout-minutes: 420
|
||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
|
|||||||
230
.github/workflows/generated-linux-binary-libtorch-nightly.yml
generated
vendored
230
.github/workflows/generated-linux-binary-libtorch-nightly.yml
generated
vendored
@ -316,6 +316,120 @@ jobs:
|
|||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
uses: ./.github/workflows/_binary-upload.yml
|
uses: ./.github/workflows/_binary-upload.yml
|
||||||
|
|
||||||
|
libtorch-rocm6_3-shared-with-deps-release-build:
|
||||||
|
if: ${{ github.repository_owner == 'pytorch' }}
|
||||||
|
uses: ./.github/workflows/_binary-build-linux.yml
|
||||||
|
needs: get-label-type
|
||||||
|
with:
|
||||||
|
PYTORCH_ROOT: /pytorch
|
||||||
|
PACKAGE_TYPE: libtorch
|
||||||
|
# TODO: This is a legacy variable that we eventually want to get rid of in
|
||||||
|
# favor of GPU_ARCH_VERSION
|
||||||
|
DESIRED_CUDA: rocm6.3
|
||||||
|
GPU_ARCH_VERSION: "6.3"
|
||||||
|
GPU_ARCH_TYPE: rocm
|
||||||
|
DOCKER_IMAGE: libtorch-cxx11-builder
|
||||||
|
DOCKER_IMAGE_TAG_PREFIX: rocm6.3
|
||||||
|
LIBTORCH_CONFIG: release
|
||||||
|
LIBTORCH_VARIANT: shared-with-deps
|
||||||
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
|
build_name: libtorch-rocm6_3-shared-with-deps-release
|
||||||
|
build_environment: linux-binary-libtorch
|
||||||
|
secrets:
|
||||||
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
|
libtorch-rocm6_3-shared-with-deps-release-test: # Testing
|
||||||
|
if: ${{ github.repository_owner == 'pytorch' }}
|
||||||
|
needs:
|
||||||
|
- libtorch-rocm6_3-shared-with-deps-release-build
|
||||||
|
- get-label-type
|
||||||
|
runs-on: linux.rocm.gpu.mi250
|
||||||
|
timeout-minutes: 240
|
||||||
|
env:
|
||||||
|
PYTORCH_ROOT: /pytorch
|
||||||
|
PACKAGE_TYPE: libtorch
|
||||||
|
# TODO: This is a legacy variable that we eventually want to get rid of in
|
||||||
|
# favor of GPU_ARCH_VERSION
|
||||||
|
DESIRED_CUDA: rocm6.3
|
||||||
|
GPU_ARCH_VERSION: "6.3"
|
||||||
|
GPU_ARCH_TYPE: rocm
|
||||||
|
SKIP_ALL_TESTS: 1
|
||||||
|
DOCKER_IMAGE: libtorch-cxx11-builder
|
||||||
|
DOCKER_IMAGE_TAG_PREFIX: rocm6.3
|
||||||
|
LIBTORCH_CONFIG: release
|
||||||
|
LIBTORCH_VARIANT: shared-with-deps
|
||||||
|
steps:
|
||||||
|
- name: Setup ROCm
|
||||||
|
uses: ./.github/actions/setup-rocm
|
||||||
|
- uses: actions/download-artifact@v4.1.7
|
||||||
|
name: Download Build Artifacts
|
||||||
|
with:
|
||||||
|
name: libtorch-rocm6_3-shared-with-deps-release
|
||||||
|
path: "${{ runner.temp }}/artifacts/"
|
||||||
|
- name: Checkout PyTorch
|
||||||
|
uses: actions/checkout@v4
|
||||||
|
with:
|
||||||
|
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
|
||||||
|
submodules: recursive
|
||||||
|
path: pytorch
|
||||||
|
show-progress: false
|
||||||
|
- name: Clean PyTorch checkout
|
||||||
|
run: |
|
||||||
|
# Remove any artifacts from the previous checkouts
|
||||||
|
git clean -fxd
|
||||||
|
working-directory: pytorch
|
||||||
|
- name: ROCm set GPU_FLAG
|
||||||
|
run: |
|
||||||
|
echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
|
||||||
|
- name: configure aws credentials
|
||||||
|
id: aws_creds
|
||||||
|
if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
|
||||||
|
uses: aws-actions/configure-aws-credentials@v4
|
||||||
|
with:
|
||||||
|
role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
|
||||||
|
aws-region: us-east-1
|
||||||
|
role-duration-seconds: 18000
|
||||||
|
- name: Calculate docker image
|
||||||
|
id: calculate-docker-image
|
||||||
|
uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
|
||||||
|
with:
|
||||||
|
docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
|
||||||
|
docker-image-name: libtorch-cxx11-builder
|
||||||
|
custom-tag-prefix: rocm6.3
|
||||||
|
docker-build-dir: .ci/docker
|
||||||
|
working-directory: pytorch
|
||||||
|
- name: Pull Docker image
|
||||||
|
uses: pytorch/test-infra/.github/actions/pull-docker-image@main
|
||||||
|
with:
|
||||||
|
docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
|
||||||
|
- name: Test Pytorch binary
|
||||||
|
uses: ./pytorch/.github/actions/test-pytorch-binary
|
||||||
|
env:
|
||||||
|
DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
|
||||||
|
- name: Teardown ROCm
|
||||||
|
uses: ./.github/actions/teardown-rocm
|
||||||
|
libtorch-rocm6_3-shared-with-deps-release-upload: # Uploading
|
||||||
|
if: ${{ github.repository_owner == 'pytorch' }}
|
||||||
|
permissions:
|
||||||
|
id-token: write
|
||||||
|
contents: read
|
||||||
|
needs: libtorch-rocm6_3-shared-with-deps-release-test
|
||||||
|
with:
|
||||||
|
PYTORCH_ROOT: /pytorch
|
||||||
|
PACKAGE_TYPE: libtorch
|
||||||
|
# TODO: This is a legacy variable that we eventually want to get rid of in
|
||||||
|
# favor of GPU_ARCH_VERSION
|
||||||
|
DESIRED_CUDA: rocm6.3
|
||||||
|
GPU_ARCH_VERSION: "6.3"
|
||||||
|
GPU_ARCH_TYPE: rocm
|
||||||
|
DOCKER_IMAGE: libtorch-cxx11-builder
|
||||||
|
DOCKER_IMAGE_TAG_PREFIX: rocm6.3
|
||||||
|
LIBTORCH_CONFIG: release
|
||||||
|
LIBTORCH_VARIANT: shared-with-deps
|
||||||
|
build_name: libtorch-rocm6_3-shared-with-deps-release
|
||||||
|
secrets:
|
||||||
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
|
uses: ./.github/workflows/_binary-upload.yml
|
||||||
|
|
||||||
libtorch-rocm6_4-shared-with-deps-release-build:
|
libtorch-rocm6_4-shared-with-deps-release-build:
|
||||||
if: ${{ github.repository_owner == 'pytorch' }}
|
if: ${{ github.repository_owner == 'pytorch' }}
|
||||||
uses: ./.github/workflows/_binary-build-linux.yml
|
uses: ./.github/workflows/_binary-build-linux.yml
|
||||||
@ -333,7 +447,6 @@ jobs:
|
|||||||
LIBTORCH_CONFIG: release
|
LIBTORCH_CONFIG: release
|
||||||
LIBTORCH_VARIANT: shared-with-deps
|
LIBTORCH_VARIANT: shared-with-deps
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
timeout-minutes: 300
|
|
||||||
build_name: libtorch-rocm6_4-shared-with-deps-release
|
build_name: libtorch-rocm6_4-shared-with-deps-release
|
||||||
build_environment: linux-binary-libtorch
|
build_environment: linux-binary-libtorch
|
||||||
secrets:
|
secrets:
|
||||||
@ -430,118 +543,3 @@ jobs:
|
|||||||
secrets:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
uses: ./.github/workflows/_binary-upload.yml
|
uses: ./.github/workflows/_binary-upload.yml
|
||||||
|
|
||||||
libtorch-rocm7_0-shared-with-deps-release-build:
|
|
||||||
if: ${{ github.repository_owner == 'pytorch' }}
|
|
||||||
uses: ./.github/workflows/_binary-build-linux.yml
|
|
||||||
needs: get-label-type
|
|
||||||
with:
|
|
||||||
PYTORCH_ROOT: /pytorch
|
|
||||||
PACKAGE_TYPE: libtorch
|
|
||||||
# TODO: This is a legacy variable that we eventually want to get rid of in
|
|
||||||
# favor of GPU_ARCH_VERSION
|
|
||||||
DESIRED_CUDA: rocm7.0
|
|
||||||
GPU_ARCH_VERSION: "7.0"
|
|
||||||
GPU_ARCH_TYPE: rocm
|
|
||||||
DOCKER_IMAGE: libtorch-cxx11-builder
|
|
||||||
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
|
|
||||||
LIBTORCH_CONFIG: release
|
|
||||||
LIBTORCH_VARIANT: shared-with-deps
|
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
|
||||||
timeout-minutes: 300
|
|
||||||
build_name: libtorch-rocm7_0-shared-with-deps-release
|
|
||||||
build_environment: linux-binary-libtorch
|
|
||||||
secrets:
|
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
|
||||||
libtorch-rocm7_0-shared-with-deps-release-test: # Testing
|
|
||||||
if: ${{ github.repository_owner == 'pytorch' }}
|
|
||||||
needs:
|
|
||||||
- libtorch-rocm7_0-shared-with-deps-release-build
|
|
||||||
- get-label-type
|
|
||||||
runs-on: linux.rocm.gpu.mi250
|
|
||||||
timeout-minutes: 240
|
|
||||||
env:
|
|
||||||
PYTORCH_ROOT: /pytorch
|
|
||||||
PACKAGE_TYPE: libtorch
|
|
||||||
# TODO: This is a legacy variable that we eventually want to get rid of in
|
|
||||||
# favor of GPU_ARCH_VERSION
|
|
||||||
DESIRED_CUDA: rocm7.0
|
|
||||||
GPU_ARCH_VERSION: "7.0"
|
|
||||||
GPU_ARCH_TYPE: rocm
|
|
||||||
SKIP_ALL_TESTS: 1
|
|
||||||
DOCKER_IMAGE: libtorch-cxx11-builder
|
|
||||||
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
|
|
||||||
LIBTORCH_CONFIG: release
|
|
||||||
LIBTORCH_VARIANT: shared-with-deps
|
|
||||||
steps:
|
|
||||||
- name: Setup ROCm
|
|
||||||
uses: ./.github/actions/setup-rocm
|
|
||||||
- uses: actions/download-artifact@v4.1.7
|
|
||||||
name: Download Build Artifacts
|
|
||||||
with:
|
|
||||||
name: libtorch-rocm7_0-shared-with-deps-release
|
|
||||||
path: "${{ runner.temp }}/artifacts/"
|
|
||||||
- name: Checkout PyTorch
|
|
||||||
uses: actions/checkout@v4
|
|
||||||
with:
|
|
||||||
ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
|
|
||||||
submodules: recursive
|
|
||||||
path: pytorch
|
|
||||||
show-progress: false
|
|
||||||
- name: Clean PyTorch checkout
|
|
||||||
run: |
|
|
||||||
# Remove any artifacts from the previous checkouts
|
|
||||||
git clean -fxd
|
|
||||||
working-directory: pytorch
|
|
||||||
- name: ROCm set GPU_FLAG
|
|
||||||
run: |
|
|
||||||
echo "GPU_FLAG=--device=/dev/mem --device=/dev/kfd --device=/dev/dri --group-add video --group-add daemon" >> "${GITHUB_ENV}"
|
|
||||||
- name: configure aws credentials
|
|
||||||
id: aws_creds
|
|
||||||
if: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') }}
|
|
||||||
uses: aws-actions/configure-aws-credentials@v4
|
|
||||||
with:
|
|
||||||
role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
|
|
||||||
aws-region: us-east-1
|
|
||||||
role-duration-seconds: 18000
|
|
||||||
- name: Calculate docker image
|
|
||||||
id: calculate-docker-image
|
|
||||||
uses: pytorch/test-infra/.github/actions/calculate-docker-image@main
|
|
||||||
with:
|
|
||||||
docker-registry: ${{ startsWith(github.event.ref, 'refs/tags/ciflow/') && '308535385114.dkr.ecr.us-east-1.amazonaws.com' || 'docker.io' }}
|
|
||||||
docker-image-name: libtorch-cxx11-builder
|
|
||||||
custom-tag-prefix: rocm7.0
|
|
||||||
docker-build-dir: .ci/docker
|
|
||||||
working-directory: pytorch
|
|
||||||
- name: Pull Docker image
|
|
||||||
uses: pytorch/test-infra/.github/actions/pull-docker-image@main
|
|
||||||
with:
|
|
||||||
docker-image: ${{ steps.calculate-docker-image.outputs.docker-image }}
|
|
||||||
- name: Test Pytorch binary
|
|
||||||
uses: ./pytorch/.github/actions/test-pytorch-binary
|
|
||||||
env:
|
|
||||||
DOCKER_IMAGE: ${{ steps.calculate-docker-image.outputs.docker-image }}
|
|
||||||
- name: Teardown ROCm
|
|
||||||
uses: ./.github/actions/teardown-rocm
|
|
||||||
libtorch-rocm7_0-shared-with-deps-release-upload: # Uploading
|
|
||||||
if: ${{ github.repository_owner == 'pytorch' }}
|
|
||||||
permissions:
|
|
||||||
id-token: write
|
|
||||||
contents: read
|
|
||||||
needs: libtorch-rocm7_0-shared-with-deps-release-test
|
|
||||||
with:
|
|
||||||
PYTORCH_ROOT: /pytorch
|
|
||||||
PACKAGE_TYPE: libtorch
|
|
||||||
# TODO: This is a legacy variable that we eventually want to get rid of in
|
|
||||||
# favor of GPU_ARCH_VERSION
|
|
||||||
DESIRED_CUDA: rocm7.0
|
|
||||||
GPU_ARCH_VERSION: "7.0"
|
|
||||||
GPU_ARCH_TYPE: rocm
|
|
||||||
DOCKER_IMAGE: libtorch-cxx11-builder
|
|
||||||
DOCKER_IMAGE_TAG_PREFIX: rocm7.0
|
|
||||||
LIBTORCH_CONFIG: release
|
|
||||||
LIBTORCH_VARIANT: shared-with-deps
|
|
||||||
build_name: libtorch-rocm7_0-shared-with-deps-release
|
|
||||||
secrets:
|
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
|
||||||
uses: ./.github/workflows/_binary-upload.yml
|
|
||||||
|
|||||||
24
.github/workflows/generated-linux-binary-manywheel-main.yml
generated
vendored
24
.github/workflows/generated-linux-binary-manywheel-main.yml
generated
vendored
@ -42,7 +42,7 @@ jobs:
|
|||||||
issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
|
issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
|
||||||
curr_branch: ${{ github.head_ref || github.ref_name }}
|
curr_branch: ${{ github.head_ref || github.ref_name }}
|
||||||
curr_ref_type: ${{ github.ref_type }}
|
curr_ref_type: ${{ github.ref_type }}
|
||||||
manywheel-py3_12-cuda13_0-build:
|
manywheel-py3_12-cuda12_8-build:
|
||||||
if: ${{ github.repository_owner == 'pytorch' }}
|
if: ${{ github.repository_owner == 'pytorch' }}
|
||||||
uses: ./.github/workflows/_binary-build-linux.yml
|
uses: ./.github/workflows/_binary-build-linux.yml
|
||||||
needs: get-label-type
|
needs: get-label-type
|
||||||
@ -51,22 +51,22 @@ jobs:
|
|||||||
PACKAGE_TYPE: manywheel
|
PACKAGE_TYPE: manywheel
|
||||||
# TODO: This is a legacy variable that we eventually want to get rid of in
|
# TODO: This is a legacy variable that we eventually want to get rid of in
|
||||||
# favor of GPU_ARCH_VERSION
|
# favor of GPU_ARCH_VERSION
|
||||||
DESIRED_CUDA: cu130
|
DESIRED_CUDA: cu128
|
||||||
GPU_ARCH_VERSION: "13.0"
|
GPU_ARCH_VERSION: "12.8"
|
||||||
GPU_ARCH_TYPE: cuda
|
GPU_ARCH_TYPE: cuda
|
||||||
DOCKER_IMAGE: manylinux2_28-builder
|
DOCKER_IMAGE: manylinux2_28-builder
|
||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.12"
|
DESIRED_PYTHON: "3.12"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
build_name: manywheel-py3_12-cuda13_0
|
build_name: manywheel-py3_12-cuda12_8
|
||||||
build_environment: linux-binary-manywheel
|
build_environment: linux-binary-manywheel
|
||||||
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc==13.0.48; platform_system == 'Linux' | nvidia-cuda-runtime==13.0.48; platform_system == 'Linux' | nvidia-cuda-cupti==13.0.48; platform_system == 'Linux' | nvidia-cudnn-cu13==9.13.0.50; platform_system == 'Linux' | nvidia-cublas==13.0.0.19; platform_system == 'Linux' | nvidia-cufft==12.0.0.15; platform_system == 'Linux' | nvidia-curand==10.4.0.35; platform_system == 'Linux' | nvidia-cusolver==12.0.3.29; platform_system == 'Linux' | nvidia-cusparse==12.6.2.49; platform_system == 'Linux' | nvidia-cusparselt-cu13==0.8.0; platform_system == 'Linux' | nvidia-nccl-cu13==2.28.3; platform_system == 'Linux' | nvidia-nvshmem-cu13==3.3.24; platform_system == 'Linux' | nvidia-nvtx==13.0.39; platform_system == 'Linux' | nvidia-nvjitlink==13.0.39; platform_system == 'Linux' | nvidia-cufile==1.15.0.42; platform_system == 'Linux'
|
PYTORCH_EXTRA_INSTALL_REQUIREMENTS: nvidia-cuda-nvrtc-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:
|
secrets:
|
||||||
github-token: ${{ secrets.GITHUB_TOKEN }}
|
github-token: ${{ secrets.GITHUB_TOKEN }}
|
||||||
manywheel-py3_12-cuda13_0-test: # Testing
|
manywheel-py3_12-cuda12_8-test: # Testing
|
||||||
if: ${{ github.repository_owner == 'pytorch' }}
|
if: ${{ github.repository_owner == 'pytorch' }}
|
||||||
needs:
|
needs:
|
||||||
- manywheel-py3_12-cuda13_0-build
|
- manywheel-py3_12-cuda12_8-build
|
||||||
- get-label-type
|
- get-label-type
|
||||||
uses: ./.github/workflows/_binary-test-linux.yml
|
uses: ./.github/workflows/_binary-test-linux.yml
|
||||||
with:
|
with:
|
||||||
@ -74,13 +74,13 @@ jobs:
|
|||||||
PACKAGE_TYPE: manywheel
|
PACKAGE_TYPE: manywheel
|
||||||
# TODO: This is a legacy variable that we eventually want to get rid of in
|
# TODO: This is a legacy variable that we eventually want to get rid of in
|
||||||
# favor of GPU_ARCH_VERSION
|
# favor of GPU_ARCH_VERSION
|
||||||
DESIRED_CUDA: cu130
|
DESIRED_CUDA: cu128
|
||||||
GPU_ARCH_VERSION: "13.0"
|
GPU_ARCH_VERSION: "12.8"
|
||||||
GPU_ARCH_TYPE: cuda
|
GPU_ARCH_TYPE: cuda
|
||||||
DOCKER_IMAGE: manylinux2_28-builder
|
DOCKER_IMAGE: manylinux2_28-builder
|
||||||
DOCKER_IMAGE_TAG_PREFIX: cuda13.0
|
DOCKER_IMAGE_TAG_PREFIX: cuda12.8
|
||||||
DESIRED_PYTHON: "3.12"
|
DESIRED_PYTHON: "3.12"
|
||||||
build_name: manywheel-py3_12-cuda13_0
|
build_name: manywheel-py3_12-cuda12_8
|
||||||
build_environment: linux-binary-manywheel
|
build_environment: linux-binary-manywheel
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
|
runs_on: linux.g4dn.4xlarge.nvidia.gpu # 12.8+ builds need sm_70+ runner
|
||||||
|
|||||||
1610
.github/workflows/generated-linux-binary-manywheel-nightly.yml
generated
vendored
1610
.github/workflows/generated-linux-binary-manywheel-nightly.yml
generated
vendored
File diff suppressed because it is too large
Load Diff
1
.github/workflows/generated-linux-binary-manywheel-rocm-main.yml
generated
vendored
1
.github/workflows/generated-linux-binary-manywheel-rocm-main.yml
generated
vendored
@ -60,7 +60,6 @@ jobs:
|
|||||||
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
|
DOCKER_IMAGE_TAG_PREFIX: rocm6.4
|
||||||
DESIRED_PYTHON: "3.10"
|
DESIRED_PYTHON: "3.10"
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
timeout-minutes: 300
|
|
||||||
build_name: manywheel-py3_10-rocm6_4
|
build_name: manywheel-py3_10-rocm6_4
|
||||||
build_environment: linux-binary-manywheel-rocm
|
build_environment: linux-binary-manywheel-rocm
|
||||||
secrets:
|
secrets:
|
||||||
|
|||||||
19
.github/workflows/inductor-perf-test-nightly.yml
vendored
19
.github/workflows/inductor-perf-test-nightly.yml
vendored
@ -57,7 +57,7 @@ on:
|
|||||||
description: The list of configs used the benchmark
|
description: The list of configs used the benchmark
|
||||||
required: false
|
required: false
|
||||||
type: string
|
type: string
|
||||||
default: inductor_huggingface_perf,inductor_timm_perf,inductor_torchbench_perf,cachebench
|
default: inductor_huggingface_perf,inductor_timm_perf,inductor_torchbench_perf,dynamo_eager_huggingface_perf,dynamo_eager_timm_perf,dynamo_eager_torchbench_perf,cachebench
|
||||||
|
|
||||||
concurrency:
|
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' }}
|
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' }}
|
||||||
@ -97,18 +97,35 @@ jobs:
|
|||||||
{ config: "inductor_huggingface_perf", shard: 3, num_shards: 5, runner: "linux.aws.a100" },
|
{ config: "inductor_huggingface_perf", shard: 3, num_shards: 5, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_huggingface_perf", shard: 4, num_shards: 5, runner: "linux.aws.a100" },
|
{ config: "inductor_huggingface_perf", shard: 4, num_shards: 5, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_huggingface_perf", shard: 5, num_shards: 5, runner: "linux.aws.a100" },
|
{ config: "inductor_huggingface_perf", shard: 5, num_shards: 5, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_huggingface_perf", shard: 1, num_shards: 5, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_huggingface_perf", shard: 2, num_shards: 5, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_huggingface_perf", shard: 3, num_shards: 5, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_huggingface_perf", shard: 4, num_shards: 5, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_huggingface_perf", shard: 5, num_shards: 5, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_timm_perf", shard: 1, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_timm_perf", shard: 1, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_timm_perf", shard: 2, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_timm_perf", shard: 2, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_timm_perf", shard: 3, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_timm_perf", shard: 3, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_timm_perf", shard: 4, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_timm_perf", shard: 4, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_timm_perf", shard: 5, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_timm_perf", shard: 5, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_timm_perf", shard: 6, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_timm_perf", shard: 6, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_timm_perf", shard: 1, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_timm_perf", shard: 2, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_timm_perf", shard: 3, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_timm_perf", shard: 4, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_timm_perf", shard: 5, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_timm_perf", shard: 6, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_torchbench_perf", shard: 1, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_torchbench_perf", shard: 1, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_torchbench_perf", shard: 2, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_torchbench_perf", shard: 2, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_torchbench_perf", shard: 3, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_torchbench_perf", shard: 3, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_torchbench_perf", shard: 4, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_torchbench_perf", shard: 4, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_torchbench_perf", shard: 5, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_torchbench_perf", shard: 5, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "inductor_torchbench_perf", shard: 6, num_shards: 6, runner: "linux.aws.a100" },
|
{ config: "inductor_torchbench_perf", shard: 6, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_torchbench_perf", shard: 1, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_torchbench_perf", shard: 2, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_torchbench_perf", shard: 3, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_torchbench_perf", shard: 4, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_torchbench_perf", shard: 5, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
|
{ config: "dynamo_eager_torchbench_perf", shard: 6, num_shards: 6, runner: "linux.aws.a100" },
|
||||||
{ config: "cachebench", shard: 1, num_shards: 2, runner: "linux.aws.a100" },
|
{ config: "cachebench", shard: 1, num_shards: 2, runner: "linux.aws.a100" },
|
||||||
{ config: "cachebench", shard: 2, num_shards: 2, runner: "linux.aws.a100" },
|
{ config: "cachebench", shard: 2, num_shards: 2, runner: "linux.aws.a100" },
|
||||||
]}
|
]}
|
||||||
|
|||||||
8
.github/workflows/lint.yml
vendored
8
.github/workflows/lint.yml
vendored
@ -31,8 +31,6 @@ jobs:
|
|||||||
if: github.repository_owner == 'pytorch'
|
if: github.repository_owner == 'pytorch'
|
||||||
name: Get changed files
|
name: Get changed files
|
||||||
uses: ./.github/workflows/_get-changed-files.yml
|
uses: ./.github/workflows/_get-changed-files.yml
|
||||||
with:
|
|
||||||
all_files: ${{ contains(github.event.pull_request.labels.*.name, 'lint-all-files') || contains(github.event.pull_request.labels.*.name, 'Reverted') }}
|
|
||||||
|
|
||||||
lintrunner-clang:
|
lintrunner-clang:
|
||||||
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
|
uses: pytorch/test-infra/.github/workflows/linux_job_v2.yml@main
|
||||||
@ -55,7 +53,7 @@ jobs:
|
|||||||
with:
|
with:
|
||||||
timeout: 120
|
timeout: 120
|
||||||
runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
|
runner: "${{ needs.get-label-type.outputs.label-type }}linux.2xlarge"
|
||||||
docker-image: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3.10-linter
|
docker-image: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3.9-linter
|
||||||
# NB: A shallow checkout won't work here because calculate-docker-image requires a full checkout
|
# NB: A shallow checkout won't work here because calculate-docker-image requires a full checkout
|
||||||
# to run git rev-parse HEAD~:.ci/docker when a new image is needed
|
# to run git rev-parse HEAD~:.ci/docker when a new image is needed
|
||||||
fetch-depth: 0
|
fetch-depth: 0
|
||||||
@ -266,10 +264,10 @@ jobs:
|
|||||||
with:
|
with:
|
||||||
submodules: false
|
submodules: false
|
||||||
fetch-depth: 1
|
fetch-depth: 1
|
||||||
- name: Setup Python 3.10
|
- name: Setup Python 3.9
|
||||||
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
|
uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5.6.0
|
||||||
with:
|
with:
|
||||||
python-version: '3.10'
|
python-version: '3.9'
|
||||||
architecture: x64
|
architecture: x64
|
||||||
cache: pip
|
cache: pip
|
||||||
- name: Install dependencies
|
- name: Install dependencies
|
||||||
|
|||||||
46
.github/workflows/operator_microbenchmark.yml
vendored
46
.github/workflows/operator_microbenchmark.yml
vendored
@ -1,46 +0,0 @@
|
|||||||
name: operator_microbenchmark
|
|
||||||
|
|
||||||
on:
|
|
||||||
push:
|
|
||||||
tags:
|
|
||||||
- ciflow/op-benchmark/*
|
|
||||||
workflow_dispatch:
|
|
||||||
schedule:
|
|
||||||
# Run at 06:00 UTC everyday
|
|
||||||
- cron: 0 6 * * *
|
|
||||||
|
|
||||||
concurrency:
|
|
||||||
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
|
|
||||||
cancel-in-progress: true
|
|
||||||
|
|
||||||
permissions:
|
|
||||||
id-token: write
|
|
||||||
contents: read
|
|
||||||
|
|
||||||
jobs:
|
|
||||||
opmicrobenchmark-build:
|
|
||||||
if: github.repository_owner == 'pytorch'
|
|
||||||
name: opmicrobenchmark-build
|
|
||||||
uses: ./.github/workflows/_linux-build.yml
|
|
||||||
with:
|
|
||||||
runner: linux.12xlarge.memory
|
|
||||||
build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
|
|
||||||
docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
|
|
||||||
cuda-arch-list: '8.0 9.0'
|
|
||||||
test-matrix: |
|
|
||||||
{ include: [
|
|
||||||
{ config: "operator_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.aws.h100" },
|
|
||||||
{ config: "operator_microbenchmark_test", shard: 1, num_shards: 1, runner: "linux.aws.a100" },
|
|
||||||
]}
|
|
||||||
secrets: inherit
|
|
||||||
|
|
||||||
opmicrobenchmark-test:
|
|
||||||
name: opmicrobenchmark-test
|
|
||||||
uses: ./.github/workflows/_linux-test.yml
|
|
||||||
needs: opmicrobenchmark-build
|
|
||||||
with:
|
|
||||||
timeout-minutes: 500
|
|
||||||
build-environment: linux-jammy-cuda12.8-py3.10-gcc9-sm80
|
|
||||||
docker-image: ${{ needs.opmicrobenchmark-build.outputs.docker-image }}
|
|
||||||
test-matrix: ${{ needs.opmicrobenchmark-build.outputs.test-matrix }}
|
|
||||||
secrets: inherit
|
|
||||||
2
.github/workflows/pull.yml
vendored
2
.github/workflows/pull.yml
vendored
@ -127,8 +127,6 @@ jobs:
|
|||||||
uses: ./.github/workflows/_linux-build.yml
|
uses: ./.github/workflows/_linux-build.yml
|
||||||
needs: get-label-type
|
needs: get-label-type
|
||||||
with:
|
with:
|
||||||
# More memory is needed to build with asan
|
|
||||||
runner: linux.2xlarge.memory
|
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
build-environment: linux-jammy-py3.10-clang18-asan
|
build-environment: linux-jammy-py3.10-clang18-asan
|
||||||
docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
|
docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
|
||||||
|
|||||||
54
.github/workflows/quantization-periodic.yml
vendored
54
.github/workflows/quantization-periodic.yml
vendored
@ -1,54 +0,0 @@
|
|||||||
name: quantization-periodic
|
|
||||||
|
|
||||||
on:
|
|
||||||
push:
|
|
||||||
tags:
|
|
||||||
- ciflow/quantization-periodic/*
|
|
||||||
workflow_dispatch:
|
|
||||||
schedule:
|
|
||||||
# run weekly
|
|
||||||
- cron: "45 0 * * 0"
|
|
||||||
|
|
||||||
concurrency:
|
|
||||||
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.ref_name }}-${{ github.ref_type == 'branch' && github.sha }}-${{ github.event_name == 'workflow_dispatch' }}
|
|
||||||
cancel-in-progress: true
|
|
||||||
|
|
||||||
permissions:
|
|
||||||
id-token: write
|
|
||||||
contents: read
|
|
||||||
|
|
||||||
jobs:
|
|
||||||
get-default-label-prefix:
|
|
||||||
name: get-default-label-prefix
|
|
||||||
uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
|
|
||||||
if: ${{ (github.event_name != 'schedule' || github.repository == 'pytorch/pytorch') && github.repository_owner == 'pytorch' }}
|
|
||||||
with:
|
|
||||||
triggering_actor: ${{ github.triggering_actor }}
|
|
||||||
issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
|
|
||||||
curr_branch: ${{ github.head_ref || github.ref_name }}
|
|
||||||
curr_ref_type: ${{ github.ref_type }}
|
|
||||||
opt_out_experiments: lf
|
|
||||||
|
|
||||||
periodic-quantization-build:
|
|
||||||
name: periodic-quantization-build
|
|
||||||
uses: ./.github/workflows/_linux-build.yml
|
|
||||||
needs: get-default-label-prefix
|
|
||||||
with:
|
|
||||||
runner_prefix: "${{ needs.get-default-label-prefix.outputs.label-type }}"
|
|
||||||
build-environment: linux-jammy-cuda12.8-cudnn9-py3-gcc11
|
|
||||||
docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
|
|
||||||
cuda-arch-list: '8.9'
|
|
||||||
test-matrix: |
|
|
||||||
{ include: [
|
|
||||||
{ config: "quantization", shard: 1, num_shards: 1, runner: "${{ needs.get-default-label-prefix.outputs.label-type }}linux.g6.4xlarge.experimental.nvidia.gpu" },
|
|
||||||
]}
|
|
||||||
secrets: inherit
|
|
||||||
periodic-test-quantization:
|
|
||||||
name: periodic-test-quantization
|
|
||||||
uses: ./.github/workflows/_linux-test.yml
|
|
||||||
needs: periodic-quantization-build
|
|
||||||
with:
|
|
||||||
build-environment: linux-jammy-cuda12.8-cudnn9-py3-gcc11
|
|
||||||
docker-image: ${{ needs.periodic-quantization-build.outputs.docker-image }}
|
|
||||||
test-matrix: ${{ needs.periodic-quantization-build.outputs.test-matrix }}
|
|
||||||
secrets: inherit
|
|
||||||
2
.github/workflows/slow.yml
vendored
2
.github/workflows/slow.yml
vendored
@ -140,8 +140,6 @@ jobs:
|
|||||||
uses: ./.github/workflows/_linux-build.yml
|
uses: ./.github/workflows/_linux-build.yml
|
||||||
needs: get-label-type
|
needs: get-label-type
|
||||||
with:
|
with:
|
||||||
# More memory is needed to build with asan
|
|
||||||
runner: linux.2xlarge.memory
|
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
build-environment: linux-jammy-py3.10-clang18-asan
|
build-environment: linux-jammy-py3.10-clang18-asan
|
||||||
docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
|
docker-image-name: ci-image:pytorch-linux-jammy-py3-clang18-asan
|
||||||
|
|||||||
76
.github/workflows/test-b200.yml
vendored
76
.github/workflows/test-b200.yml
vendored
@ -1,76 +0,0 @@
|
|||||||
# B200 Smoke Tests CI Workflow
|
|
||||||
#
|
|
||||||
# This workflow runs smoke tests on B200 hardware
|
|
||||||
#
|
|
||||||
# Flow:
|
|
||||||
# 1. Builds PyTorch with CUDA 12.8+ and sm100 architecture for B200
|
|
||||||
# 2. Runs smoke tests on linux.dgx.b200 runner
|
|
||||||
# 3. Tests executed are defined in .ci/pytorch/test.sh -> test_python_smoke() function
|
|
||||||
#
|
|
||||||
# Triggered by:
|
|
||||||
# - Pull requests modifying this workflow file
|
|
||||||
# - Manual dispatch
|
|
||||||
# - Schedule (every 6 hours)
|
|
||||||
# - Adding ciflow/b200 label to a PR (creates ciflow/b200/* tag)
|
|
||||||
|
|
||||||
name: B200 Smoke Tests
|
|
||||||
|
|
||||||
on:
|
|
||||||
pull_request:
|
|
||||||
paths:
|
|
||||||
- .github/workflows/test-b200.yml
|
|
||||||
workflow_dispatch:
|
|
||||||
schedule:
|
|
||||||
- cron: 0 4,10,16,22 * * * # every 6 hours
|
|
||||||
push:
|
|
||||||
tags:
|
|
||||||
- ciflow/b200/*
|
|
||||||
|
|
||||||
concurrency:
|
|
||||||
group: ${{ github.workflow }}-${{ github.event.pull_request.number || github.sha }}-${{ github.event_name == 'workflow_dispatch' }}-${{ github.event_name == 'schedule' }}
|
|
||||||
cancel-in-progress: true
|
|
||||||
|
|
||||||
permissions:
|
|
||||||
id-token: write
|
|
||||||
contents: read
|
|
||||||
|
|
||||||
jobs:
|
|
||||||
|
|
||||||
get-label-type:
|
|
||||||
if: github.repository_owner == 'pytorch'
|
|
||||||
name: get-label-type
|
|
||||||
uses: pytorch/pytorch/.github/workflows/_runner-determinator.yml@main
|
|
||||||
with:
|
|
||||||
triggering_actor: ${{ github.triggering_actor }}
|
|
||||||
issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
|
|
||||||
curr_branch: ${{ github.head_ref || github.ref_name }}
|
|
||||||
curr_ref_type: ${{ github.ref_type }}
|
|
||||||
|
|
||||||
linux-jammy-cuda12_8-py3_10-gcc11-sm100-build:
|
|
||||||
name: linux-jammy-cuda12.8-py3.10-gcc11-sm100
|
|
||||||
uses: ./.github/workflows/_linux-build.yml
|
|
||||||
needs: get-label-type
|
|
||||||
with:
|
|
||||||
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
|
||||||
runner: linux.12xlarge.memory
|
|
||||||
build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
|
|
||||||
docker-image-name: ci-image:pytorch-linux-jammy-cuda12.8-cudnn9-py3-gcc11
|
|
||||||
cuda-arch-list: '10.0'
|
|
||||||
test-matrix: |
|
|
||||||
{ include: [
|
|
||||||
{ config: "smoke_b200", shard: 1, num_shards: 1, runner: "linux.dgx.b200" },
|
|
||||||
]}
|
|
||||||
# config: "smoke_b200" maps to test_python_smoke_b200() in .ci/pytorch/test.sh
|
|
||||||
secrets: inherit
|
|
||||||
|
|
||||||
linux-jammy-cuda12_8-py3_10-gcc11-sm100-test:
|
|
||||||
name: linux-jammy-cuda12.8-py3.10-gcc11-sm100
|
|
||||||
uses: ./.github/workflows/_linux-test.yml
|
|
||||||
needs:
|
|
||||||
- linux-jammy-cuda12_8-py3_10-gcc11-sm100-build
|
|
||||||
with:
|
|
||||||
build-environment: linux-jammy-cuda12.8-py3.10-gcc11-sm100
|
|
||||||
docker-image: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build.outputs.docker-image }}
|
|
||||||
test-matrix: ${{ needs.linux-jammy-cuda12_8-py3_10-gcc11-sm100-build.outputs.test-matrix }}
|
|
||||||
aws-role-to-assume: arn:aws:iam::308535385114:role/gha_workflow_s3_and_ecr_read_only
|
|
||||||
secrets: inherit
|
|
||||||
24
.github/workflows/unstable.yml
vendored
24
.github/workflows/unstable.yml
vendored
@ -53,3 +53,27 @@ jobs:
|
|||||||
issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
|
issue_owner: ${{ github.event.pull_request.user.login || github.event.issue.user.login }}
|
||||||
curr_branch: ${{ github.head_ref || github.ref_name }}
|
curr_branch: ${{ github.head_ref || github.ref_name }}
|
||||||
curr_ref_type: ${{ github.ref_type }}
|
curr_ref_type: ${{ github.ref_type }}
|
||||||
|
|
||||||
|
linux-jammy-py3_9-clang9-xla-build:
|
||||||
|
name: linux-jammy-py3_9-clang9-xla
|
||||||
|
uses: ./.github/workflows/_linux-build.yml
|
||||||
|
needs: get-label-type
|
||||||
|
with:
|
||||||
|
runner_prefix: "${{ needs.get-label-type.outputs.label-type }}"
|
||||||
|
build-environment: linux-jammy-py3.9-clang9-xla
|
||||||
|
docker-image-name: 308535385114.dkr.ecr.us-east-1.amazonaws.com/pytorch/xla_base:v1.3-lite
|
||||||
|
test-matrix: |
|
||||||
|
{ include: [
|
||||||
|
{ config: "xla", shard: 1, num_shards: 1, runner: "${{ needs.get-label-type.outputs.label-type }}linux.12xlarge" },
|
||||||
|
]}
|
||||||
|
secrets: inherit
|
||||||
|
|
||||||
|
linux-jammy-py3_9-clang9-xla-test:
|
||||||
|
name: linux-jammy-py3_9-clang9-xla
|
||||||
|
uses: ./.github/workflows/_linux-test.yml
|
||||||
|
needs: linux-jammy-py3_9-clang9-xla-build
|
||||||
|
with:
|
||||||
|
build-environment: linux-jammy-py3.9-clang9-xla
|
||||||
|
docker-image: ${{ needs.linux-jammy-py3_9-clang9-xla-build.outputs.docker-image }}
|
||||||
|
test-matrix: ${{ needs.linux-jammy-py3_9-clang9-xla-build.outputs.test-matrix }}
|
||||||
|
secrets: inherit
|
||||||
|
|||||||
1
.gitignore
vendored
1
.gitignore
vendored
@ -82,7 +82,6 @@ torch/return_types.pyi
|
|||||||
torch/nn/functional.pyi
|
torch/nn/functional.pyi
|
||||||
torch/utils/data/datapipes/datapipe.pyi
|
torch/utils/data/datapipes/datapipe.pyi
|
||||||
torch/csrc/autograd/generated/*
|
torch/csrc/autograd/generated/*
|
||||||
torch/csrc/functionalization/generated/*
|
|
||||||
torch/csrc/lazy/generated/*.[!m]*
|
torch/csrc/lazy/generated/*.[!m]*
|
||||||
torch_compile_debug/
|
torch_compile_debug/
|
||||||
# Listed manually because some files in this directory are not generated
|
# Listed manually because some files in this directory are not generated
|
||||||
|
|||||||
@ -49,7 +49,7 @@ init_command = [
|
|||||||
'mccabe==0.7.0',
|
'mccabe==0.7.0',
|
||||||
'pycodestyle==2.14.0',
|
'pycodestyle==2.14.0',
|
||||||
'pyflakes==3.4.0',
|
'pyflakes==3.4.0',
|
||||||
'torchfix==0.4.0 ; python_version >= "3.10" and python_version < "3.13"',
|
'torchfix==0.4.0 ; python_version >= "3.9" and python_version < "3.13"',
|
||||||
]
|
]
|
||||||
|
|
||||||
|
|
||||||
@ -153,7 +153,7 @@ init_command = [
|
|||||||
'python3',
|
'python3',
|
||||||
'tools/linter/adapters/pip_init.py',
|
'tools/linter/adapters/pip_init.py',
|
||||||
'--dry-run={{DRYRUN}}',
|
'--dry-run={{DRYRUN}}',
|
||||||
'numpy==1.26.4 ; python_version >= "3.10" and python_version <= "3.11"',
|
'numpy==1.26.4 ; python_version >= "3.9" and python_version <= "3.11"',
|
||||||
'numpy==2.1.0 ; python_version >= "3.12"',
|
'numpy==2.1.0 ; python_version >= "3.12"',
|
||||||
'expecttest==0.3.0',
|
'expecttest==0.3.0',
|
||||||
'mypy==1.16.0',
|
'mypy==1.16.0',
|
||||||
@ -196,7 +196,6 @@ exclude_patterns = [
|
|||||||
'tools/test/gen_operators_yaml_test.py',
|
'tools/test/gen_operators_yaml_test.py',
|
||||||
'tools/test/gen_oplist_test.py',
|
'tools/test/gen_oplist_test.py',
|
||||||
'tools/test/test_selective_build.py',
|
'tools/test/test_selective_build.py',
|
||||||
'tools/experimental/dynamic_shapes/torchfuzz/**',
|
|
||||||
]
|
]
|
||||||
command = [
|
command = [
|
||||||
'python3',
|
'python3',
|
||||||
@ -1453,7 +1452,7 @@ init_command = [
|
|||||||
'--dry-run={{DRYRUN}}',
|
'--dry-run={{DRYRUN}}',
|
||||||
'usort==1.0.8.post1',
|
'usort==1.0.8.post1',
|
||||||
'isort==6.0.1',
|
'isort==6.0.1',
|
||||||
'ruff==0.13.1', # sync with RUFF
|
'ruff==0.12.9', # sync with RUFF
|
||||||
]
|
]
|
||||||
is_formatter = true
|
is_formatter = true
|
||||||
|
|
||||||
@ -1587,7 +1586,7 @@ init_command = [
|
|||||||
'python3',
|
'python3',
|
||||||
'tools/linter/adapters/pip_init.py',
|
'tools/linter/adapters/pip_init.py',
|
||||||
'--dry-run={{DRYRUN}}',
|
'--dry-run={{DRYRUN}}',
|
||||||
'ruff==0.13.1', # sync with PYFMT
|
'ruff==0.12.9', # sync with PYFMT
|
||||||
]
|
]
|
||||||
is_formatter = true
|
is_formatter = true
|
||||||
|
|
||||||
|
|||||||
34
BUILD.bazel
34
BUILD.bazel
@ -91,8 +91,6 @@ generated_cpu_cpp = [
|
|||||||
"aten/src/ATen/NativeMetaFunctions.h",
|
"aten/src/ATen/NativeMetaFunctions.h",
|
||||||
"aten/src/ATen/RegistrationDeclarations.h",
|
"aten/src/ATen/RegistrationDeclarations.h",
|
||||||
"aten/src/ATen/VmapGeneratedPlumbing.h",
|
"aten/src/ATen/VmapGeneratedPlumbing.h",
|
||||||
"aten/src/ATen/ViewMetaClasses.h",
|
|
||||||
"aten/src/ATen/ViewMetaClasses.cpp",
|
|
||||||
"aten/src/ATen/core/aten_interned_strings.h",
|
"aten/src/ATen/core/aten_interned_strings.h",
|
||||||
"aten/src/ATen/core/enum_tag.h",
|
"aten/src/ATen/core/enum_tag.h",
|
||||||
"aten/src/ATen/core/TensorBody.h",
|
"aten/src/ATen/core/TensorBody.h",
|
||||||
@ -835,6 +833,36 @@ pybind_extension(
|
|||||||
],
|
],
|
||||||
)
|
)
|
||||||
|
|
||||||
|
cc_library(
|
||||||
|
name = "functorch",
|
||||||
|
hdrs = glob([
|
||||||
|
"functorch/csrc/dim/*.h",
|
||||||
|
]),
|
||||||
|
srcs = glob([
|
||||||
|
"functorch/csrc/dim/*.cpp",
|
||||||
|
]),
|
||||||
|
deps = [
|
||||||
|
":aten_nvrtc",
|
||||||
|
":torch_python",
|
||||||
|
"@pybind11",
|
||||||
|
],
|
||||||
|
)
|
||||||
|
|
||||||
|
pybind_extension(
|
||||||
|
name = "functorch/_C",
|
||||||
|
copts=[
|
||||||
|
"-DTORCH_EXTENSION_NAME=_C"
|
||||||
|
],
|
||||||
|
srcs = [
|
||||||
|
"functorch/csrc/init_dim_only.cpp",
|
||||||
|
],
|
||||||
|
deps = [
|
||||||
|
":functorch",
|
||||||
|
":torch_python",
|
||||||
|
":aten_nvrtc",
|
||||||
|
],
|
||||||
|
)
|
||||||
|
|
||||||
cc_binary(
|
cc_binary(
|
||||||
name = "torch/bin/torch_shm_manager",
|
name = "torch/bin/torch_shm_manager",
|
||||||
srcs = [
|
srcs = [
|
||||||
@ -875,6 +903,7 @@ py_library(
|
|||||||
],
|
],
|
||||||
data = [
|
data = [
|
||||||
":torch/_C.so",
|
":torch/_C.so",
|
||||||
|
":functorch/_C.so",
|
||||||
":torch/bin/torch_shm_manager",
|
":torch/bin/torch_shm_manager",
|
||||||
],
|
],
|
||||||
)
|
)
|
||||||
@ -1077,7 +1106,6 @@ test_suite(
|
|||||||
"aten/src/ATen/templates/LazyNonNativeIr.h",
|
"aten/src/ATen/templates/LazyNonNativeIr.h",
|
||||||
"aten/src/ATen/templates/RegisterDispatchKey.cpp",
|
"aten/src/ATen/templates/RegisterDispatchKey.cpp",
|
||||||
"aten/src/ATen/templates/RegisterDispatchDefinitions.ini",
|
"aten/src/ATen/templates/RegisterDispatchDefinitions.ini",
|
||||||
"aten/src/ATen/templates/ViewMetaClassesPythonBinding.cpp",
|
|
||||||
"aten/src/ATen/native/native_functions.yaml",
|
"aten/src/ATen/native/native_functions.yaml",
|
||||||
"aten/src/ATen/native/tags.yaml",
|
"aten/src/ATen/native/tags.yaml",
|
||||||
"aten/src/ATen/native/ts_native_functions.yaml",
|
"aten/src/ATen/native/ts_native_functions.yaml",
|
||||||
|
|||||||
@ -1,4 +1,5 @@
|
|||||||
cmake_minimum_required(VERSION 3.27 FATAL_ERROR)
|
cmake_minimum_required(VERSION 3.27 FATAL_ERROR)
|
||||||
|
# cmake_policy(SET CMP0022 NEW) cmake_policy(SET CMP0023 NEW)
|
||||||
|
|
||||||
# Use compiler ID "AppleClang" instead of "Clang" for XCode. Not setting this
|
# Use compiler ID "AppleClang" instead of "Clang" for XCode. Not setting this
|
||||||
# sometimes makes XCode C compiler gets detected as "Clang", even when the C++
|
# sometimes makes XCode C compiler gets detected as "Clang", even when the C++
|
||||||
@ -442,7 +443,7 @@ if(WIN32)
|
|||||||
message(
|
message(
|
||||||
WARNING
|
WARNING
|
||||||
"Libuv is not installed in current conda env. Set USE_DISTRIBUTED to OFF. "
|
"Libuv is not installed in current conda env. Set USE_DISTRIBUTED to OFF. "
|
||||||
"Please run command 'conda install -c conda-forge libuv=1.51' to install libuv."
|
"Please run command 'conda install -c conda-forge libuv=1.39' to install libuv."
|
||||||
)
|
)
|
||||||
else()
|
else()
|
||||||
set(ENV{libuv_ROOT} ${libuv_tmp_LIBRARY}/../../)
|
set(ENV{libuv_ROOT} ${libuv_tmp_LIBRARY}/../../)
|
||||||
@ -1390,6 +1391,10 @@ endif()
|
|||||||
include(cmake/Summary.cmake)
|
include(cmake/Summary.cmake)
|
||||||
caffe2_print_configuration_summary()
|
caffe2_print_configuration_summary()
|
||||||
|
|
||||||
|
if(BUILD_FUNCTORCH)
|
||||||
|
add_subdirectory(functorch)
|
||||||
|
endif()
|
||||||
|
|
||||||
# Parse custom debug info
|
# Parse custom debug info
|
||||||
if(DEFINED USE_CUSTOM_DEBINFO)
|
if(DEFINED USE_CUSTOM_DEBINFO)
|
||||||
string(REPLACE ";" " " SOURCE_FILES "${USE_CUSTOM_DEBINFO}")
|
string(REPLACE ";" " " SOURCE_FILES "${USE_CUSTOM_DEBINFO}")
|
||||||
@ -1481,4 +1486,4 @@ else()
|
|||||||
To do so please export USE_PRIORITIZED_TEXT_FOR_LD=1
|
To do so please export USE_PRIORITIZED_TEXT_FOR_LD=1
|
||||||
]])
|
]])
|
||||||
endif()
|
endif()
|
||||||
endif()
|
endif()
|
||||||
105
MANIFEST.in
105
MANIFEST.in
@ -1,61 +1,20 @@
|
|||||||
# Reference: https://setuptools.pypa.io/en/latest/userguide/miscellaneous.html
|
# Reference: https://setuptools.pypa.io/en/latest/userguide/miscellaneous.html
|
||||||
|
|
||||||
# Include individual top-level files
|
# Include source files in SDist
|
||||||
include CITATION.cff
|
include CMakeLists.txt
|
||||||
include CODEOWNERS
|
include *.bzl *.bazel .bazel* BUILD *.BUILD BUILD.* WORKSPACE
|
||||||
include Dockerfile
|
include BUCK BUCK.*
|
||||||
include LICENSE
|
include requirements*.txt
|
||||||
include MANIFEST.in
|
include version.txt
|
||||||
include Makefile
|
include [Mm]akefile *.[Mm]akefile [Mm]akefile.*
|
||||||
include NOTICE
|
include [Dd]ockerfile *.[Dd]ockerfile [Dd]ockerfile.* .dockerignore
|
||||||
include .bc-linter.yml
|
|
||||||
include .clang-format .clang-tidy
|
|
||||||
include .cmakelintrc
|
|
||||||
include .coveragerc
|
|
||||||
include .dockerignore
|
|
||||||
include .editorconfig
|
|
||||||
include .flake8
|
|
||||||
include .gdbinit
|
|
||||||
include .lintrunner.toml
|
|
||||||
include .lldbinit
|
|
||||||
include codex_setup.sh
|
|
||||||
include docker.Makefile
|
|
||||||
include pyrefly.toml
|
|
||||||
include ubsan.supp
|
|
||||||
|
|
||||||
# Include bazel and BUCK related files
|
|
||||||
include BUILD.bazel BUCK.oss
|
|
||||||
include WORKSPACE
|
|
||||||
include *.bzl
|
|
||||||
include .bazelignore .bazelrc .bazelversion
|
|
||||||
|
|
||||||
# Include general configuration files
|
|
||||||
include *.ini
|
|
||||||
# Include important top-level information
|
|
||||||
include *.md
|
|
||||||
# Include technical text files at the moment, comprises
|
|
||||||
# version.txt, CMakeLists.txt, requirements.txt
|
|
||||||
include *.txt
|
|
||||||
|
|
||||||
# Include ctags configuration
|
|
||||||
include .ctags.d/*.ctags
|
|
||||||
|
|
||||||
# Include subfolders completely
|
|
||||||
graft .devcontainer
|
|
||||||
graft .vscode
|
|
||||||
graft android
|
graft android
|
||||||
graft aten
|
graft aten
|
||||||
graft benchmarks
|
|
||||||
graft binaries
|
graft binaries
|
||||||
graft c10
|
graft c10
|
||||||
graft caffe2
|
graft caffe2
|
||||||
graft cmake
|
graft cmake
|
||||||
graft docs
|
|
||||||
graft functorch
|
graft functorch
|
||||||
graft ios
|
|
||||||
graft mypy_plugins
|
|
||||||
graft scripts
|
|
||||||
graft test
|
|
||||||
graft third_party
|
graft third_party
|
||||||
graft tools
|
graft tools
|
||||||
graft torch
|
graft torch
|
||||||
@ -63,37 +22,29 @@ graft torchgen
|
|||||||
# FIXME: torch-xla build during codegen will fail if include this file in wheel
|
# FIXME: torch-xla build during codegen will fail if include this file in wheel
|
||||||
exclude torchgen/BUILD.bazel
|
exclude torchgen/BUILD.bazel
|
||||||
|
|
||||||
# The following exclusions omit parts from third-party dependencies that
|
# Misc files and directories in SDist
|
||||||
# contain invalid symlinks[1] and that are not needed for pytorch, such as
|
include *.md
|
||||||
# bindings for unused languages
|
include CITATION.cff
|
||||||
prune third_party/flatbuffers/java
|
include LICENSE NOTICE
|
||||||
prune third_party/flatbuffers/kotlin
|
include mypy*.ini
|
||||||
prune third_party/ittapi/rust
|
graft benchmarks
|
||||||
prune third_party/nccl/pkg/debian
|
graft docs
|
||||||
prune third_party/opentelemetry-cpp/third_party/prometheus-cpp/cmake/project-import-*
|
graft mypy_plugins
|
||||||
|
graft scripts
|
||||||
# The following document is also an invalid symlink[1] and superfluous
|
|
||||||
exclude third_party/flatbuffers/docs/source/CONTRIBUTING.md
|
|
||||||
|
|
||||||
# Omit autogenerated code
|
|
||||||
prune torchgen/packaged
|
|
||||||
|
|
||||||
# Omit caches, compiled, and scm related content
|
|
||||||
prune */__pycache__
|
|
||||||
prune **/.github
|
|
||||||
prune **/.gitlab
|
|
||||||
global-exclude *.o *.obj *.so *.dylib *.a *.pxd *.dll *.lib
|
|
||||||
global-exclude *.py[cod] *.swp *~
|
|
||||||
global-exclude .git .git-blame-ignore-revs .gitattributes .gitignore .gitmodules
|
|
||||||
global-exclude .gitlab-ci.yml
|
|
||||||
|
|
||||||
# Misc files needed for custom setuptools command
|
# Misc files needed for custom setuptools command
|
||||||
include .gitignore
|
include .gitignore
|
||||||
include .gitmodules
|
include .gitmodules
|
||||||
|
|
||||||
# [1] Invalid symlinks for the purposes of Python source distributions are,
|
# Include test suites in SDist
|
||||||
# according to the source distribution format[2] links pointing outside the
|
graft test
|
||||||
# destination directory or links with a `..` component, which is those of
|
include pytest.ini
|
||||||
# concern here.
|
include .coveragerc
|
||||||
|
|
||||||
# [2] https://packaging.python.org/en/latest/specifications/source-distribution-format/#source-distribution-archive-features
|
# Prune generated/compiled files
|
||||||
|
prune torchgen/packaged
|
||||||
|
prune */__pycache__
|
||||||
|
global-exclude *.o *.obj *.so *.a *.dylib *.pxd *.dll *.lib *.py[cod]
|
||||||
|
|
||||||
|
prune */.git
|
||||||
|
global-exclude .git *~ *.swp
|
||||||
|
|||||||
@ -161,7 +161,7 @@ They require JetPack 4.2 and above, and [@dusty-nv](https://github.com/dusty-nv)
|
|||||||
|
|
||||||
#### Prerequisites
|
#### Prerequisites
|
||||||
If you are installing from source, you will need:
|
If you are installing from source, you will need:
|
||||||
- Python 3.10 or later
|
- Python 3.9 or later
|
||||||
- A compiler that fully supports C++17, such as clang or gcc (gcc 9.4.0 or newer is required, on Linux)
|
- A compiler that fully supports C++17, such as clang or gcc (gcc 9.4.0 or newer is required, on Linux)
|
||||||
- Visual Studio or Visual Studio Build Tool (Windows only)
|
- Visual Studio or Visual Studio Build Tool (Windows only)
|
||||||
|
|
||||||
@ -275,7 +275,7 @@ conda install pkg-config libuv
|
|||||||
pip install mkl-static mkl-include
|
pip install mkl-static mkl-include
|
||||||
# Add these packages if torch.distributed is needed.
|
# Add these packages if torch.distributed is needed.
|
||||||
# Distributed package support on Windows is a prototype feature and is subject to changes.
|
# Distributed package support on Windows is a prototype feature and is subject to changes.
|
||||||
conda install -c conda-forge libuv=1.51
|
conda install -c conda-forge libuv
|
||||||
```
|
```
|
||||||
|
|
||||||
#### Install PyTorch
|
#### Install PyTorch
|
||||||
|
|||||||
@ -317,20 +317,10 @@ IF(USE_FBGEMM_GENAI)
|
|||||||
-greedy-reverse-local-assignment=1
|
-greedy-reverse-local-assignment=1
|
||||||
-fhip-new-launch-api)
|
-fhip-new-launch-api)
|
||||||
|
|
||||||
# Only compile for gfx942 for now.
|
|
||||||
# This is rather hacky, I could not figure out a clean solution :(
|
|
||||||
set(HIP_CLANG_FLAGS_ORIGINAL ${HIP_CLANG_FLAGS})
|
|
||||||
string(REGEX REPLACE "--offload-arch=[^ ]*" "" FILTERED_HIP_CLANG_FLAGS "${HIP_CLANG_FLAGS}")
|
|
||||||
if("gfx942" IN_LIST PYTORCH_ROCM_ARCH)
|
|
||||||
list(APPEND FILTERED_HIP_CLANG_FLAGS --offload-arch=gfx942;)
|
|
||||||
endif()
|
|
||||||
set(HIP_CLANG_FLAGS ${FILTERED_HIP_CLANG_FLAGS})
|
|
||||||
|
|
||||||
hip_add_library(
|
hip_add_library(
|
||||||
fbgemm_genai STATIC
|
fbgemm_genai STATIC
|
||||||
${fbgemm_genai_native_rocm_hip}
|
${fbgemm_genai_native_rocm_hip}
|
||||||
HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
|
HIPCC_OPTIONS ${HIP_HCC_FLAGS} ${FBGEMM_GENAI_EXTRA_HIPCC_FLAGS})
|
||||||
set(HIP_CLANG_FLAGS ${HIP_CLANG_FLAGS_ORIGINAL})
|
|
||||||
set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
set_target_properties(fbgemm_genai PROPERTIES POSITION_INDEPENDENT_CODE ON)
|
||||||
target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
|
target_compile_definitions(fbgemm_genai PRIVATE FBGEMM_GENAI_NO_EXTENDED_SHAPES)
|
||||||
|
|
||||||
|
|||||||
@ -401,13 +401,30 @@ T* toDLPackImpl(const Tensor& src) {
|
|||||||
// The following code detects whether the src follows
|
// The following code detects whether the src follows
|
||||||
// a continuous pattern. If the src follows such pattern (common-case)
|
// a continuous pattern. If the src follows such pattern (common-case)
|
||||||
// then we do not need to normalize the strides.
|
// then we do not need to normalize the strides.
|
||||||
bool need_normalize_strides = src.dim() == 1 && src.size(0) == 1 && src.stride(0) != 1;
|
bool need_normalize_strides = false;
|
||||||
|
int64_t expected_stride = 1;
|
||||||
|
for (int i = src.dim() - 1; i >= 0; i--) {
|
||||||
|
// detect if we do not meet continuous pattern
|
||||||
|
// and the size is 1, so there is opportunity to normalize
|
||||||
|
if (src.stride(i) != expected_stride && src.size(i) == 1) {
|
||||||
|
need_normalize_strides = true;
|
||||||
|
break;
|
||||||
|
}
|
||||||
|
expected_stride *= src.size(i);
|
||||||
|
}
|
||||||
|
|
||||||
// less common case, try normalizing the strides
|
// less common case, try normalizing the strides
|
||||||
if (need_normalize_strides) {
|
if (need_normalize_strides) {
|
||||||
// create a new tensor with possibly normalized strides
|
// create a new tensor with possibly normalized strides
|
||||||
// gh-83069
|
// gh-83069
|
||||||
auto shape = src.sizes();
|
auto shape = src.sizes();
|
||||||
view = src.as_strided(shape, {1}, src.storage_offset());
|
auto strides = src.strides().vec();
|
||||||
|
for (int i = 0; i < src.dim(); i++) {
|
||||||
|
if (shape[i] < 2) {
|
||||||
|
strides[i] = 1;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
view = src.as_strided(shape, strides, src.storage_offset());
|
||||||
}
|
}
|
||||||
|
|
||||||
ATenDLMTensor<T>* atDLMTensor(new ATenDLMTensor<T>);
|
ATenDLMTensor<T>* atDLMTensor(new ATenDLMTensor<T>);
|
||||||
|
|||||||
@ -468,7 +468,7 @@ inline Tensor _sum_to(
|
|||||||
// if we assume no reduction due to unbacked we ensure that at runtime.
|
// if we assume no reduction due to unbacked we ensure that at runtime.
|
||||||
TORCH_MAYBE_SYM_CHECK(
|
TORCH_MAYBE_SYM_CHECK(
|
||||||
sym_eq(shape[i - leading_dims], sizes[i]),
|
sym_eq(shape[i - leading_dims], sizes[i]),
|
||||||
"non-reduction path was assumed due to unbacked symbols expected those two sizes to be the same:",
|
"non-reduction path was assumed due to unabcked symbols expected those two sizes to be the same:",
|
||||||
shape[i - leading_dims],
|
shape[i - leading_dims],
|
||||||
", ",
|
", ",
|
||||||
sizes[i])
|
sizes[i])
|
||||||
|
|||||||
@ -9,6 +9,11 @@
|
|||||||
|
|
||||||
namespace at::functionalization {
|
namespace at::functionalization {
|
||||||
|
|
||||||
|
ViewMeta ViewMeta::to_out_idx(int64_t out_idx) {
|
||||||
|
if (out_idx == this->out_index) return *this;
|
||||||
|
return ViewMeta(forward_fn, reverse_fn, has_symbolic_inputs, is_multi_output, is_as_strided, out_idx);
|
||||||
|
}
|
||||||
|
|
||||||
// Note [Functionalization: Alias Removal Part 2]
|
// Note [Functionalization: Alias Removal Part 2]
|
||||||
// See Note [Functionalization: Alias Removal] for more details.
|
// See Note [Functionalization: Alias Removal] for more details.
|
||||||
// This function applies a single update from one of the views to the StorageImpl.
|
// This function applies a single update from one of the views to the StorageImpl.
|
||||||
@ -37,12 +42,12 @@ namespace at::functionalization {
|
|||||||
static const Tensor apply_update(const FunctionalStorageImpl::Update& update, const Tensor& base) {
|
static const Tensor apply_update(const FunctionalStorageImpl::Update& update, const Tensor& base) {
|
||||||
at::Tensor t = update.new_val;
|
at::Tensor t = update.new_val;
|
||||||
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
|
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
|
||||||
if (update.view_metas.empty()) { return t; }
|
if (update.view_metas.empty()) return t;
|
||||||
|
|
||||||
std::vector<at::Tensor> tmp_values({base});
|
std::vector<at::Tensor> tmp_values({base});
|
||||||
tmp_values.reserve(update.view_metas.size());
|
tmp_values.reserve(update.view_metas.size());
|
||||||
for (size_t i = 0; i < update.view_metas.size() - 1; ++i) {
|
for (size_t i = 0; i < update.view_metas.size() - 1; ++i) {
|
||||||
at::Tensor next_view = update.view_metas[i]->forward(tmp_values.back());
|
at::Tensor next_view = update.view_metas[i].forward_fn(tmp_values.back(), update.view_metas[i].out_index);
|
||||||
// NB: We only actually need tmp_values for ops like select/slice/diagonal/squeeze/as_strided
|
// NB: We only actually need tmp_values for ops like select/slice/diagonal/squeeze/as_strided
|
||||||
// All of these ops require additional information to recover the sizes of the original tensor.
|
// All of these ops require additional information to recover the sizes of the original tensor.
|
||||||
// If need to, we could probably apply this optimization and only bother computing tmp_values
|
// If need to, we could probably apply this optimization and only bother computing tmp_values
|
||||||
@ -50,8 +55,9 @@ static const Tensor apply_update(const FunctionalStorageImpl::Update& update, co
|
|||||||
tmp_values.push_back(std::move(next_view));
|
tmp_values.push_back(std::move(next_view));
|
||||||
}
|
}
|
||||||
for(int64_t i = static_cast<int64_t>(update.view_metas.size()) - 1; i >= 0; --i) {
|
for(int64_t i = static_cast<int64_t>(update.view_metas.size()) - 1; i >= 0; --i) {
|
||||||
|
int64_t out_idx = update.view_metas[i].out_index;
|
||||||
// Each view inverse is implemented in ViewInverses.cpp.
|
// Each view inverse is implemented in ViewInverses.cpp.
|
||||||
t = update.view_metas[i]->reverse(tmp_values[i], t);
|
t = update.view_metas[i].reverse_fn(tmp_values[i], t, out_idx);
|
||||||
}
|
}
|
||||||
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
|
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
|
||||||
return t;
|
return t;
|
||||||
@ -105,13 +111,13 @@ FunctionalStorageImpl::FunctionalStorageImpl(const Tensor& base)
|
|||||||
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(base_));
|
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(base_));
|
||||||
}
|
}
|
||||||
|
|
||||||
void FunctionalStorageImpl::add_update(const Tensor& updated_val, const std::vector<std::shared_ptr<ViewMeta>>& metas) {
|
void FunctionalStorageImpl::add_update(const Tensor& updated_val, const std::vector<ViewMeta>& metas) {
|
||||||
TORCH_CHECK(!frozen_, "cannot mutate tensors with frozen storage");
|
TORCH_CHECK(!frozen_, "cannot mutate tensors with frozen storage");
|
||||||
|
|
||||||
if (metas.size() > 1) {
|
if (metas.size() > 1) {
|
||||||
for (size_t i = 1; i < metas.size(); ++i) {
|
for (size_t i = 1; i < metas.size(); ++i) {
|
||||||
// Skipping this check for XLA. Would be good to add it back, but it is failing XLA CI
|
// Skipping this check for XLA. Would be good to add it back, but it is failing XLA CI
|
||||||
TORCH_CHECK(updated_val.device().type() == c10::DeviceType::XLA || !metas[i]->is_as_strided,
|
TORCH_CHECK(updated_val.device().type() == c10::DeviceType::XLA || !metas[i].is_as_strided,
|
||||||
"During torch.compile, encountered a mutation on a view chain of length ", metas.size(), ", where view ", i,
|
"During torch.compile, encountered a mutation on a view chain of length ", metas.size(), ", where view ", i,
|
||||||
" was an as_strided() call. as_strided() is non-compositional, and therefore is not possible to functionalize properly today,"
|
" was an as_strided() call. as_strided() is non-compositional, and therefore is not possible to functionalize properly today,"
|
||||||
"so this behavior is banned in compile. As a workaround, you can either remove the mutation from the model code, or you "
|
"so this behavior is banned in compile. As a workaround, you can either remove the mutation from the model code, or you "
|
||||||
|
|||||||
@ -8,89 +8,44 @@ namespace at::functionalization {
|
|||||||
|
|
||||||
// See Note [Functionalization Pass In Core]
|
// See Note [Functionalization Pass In Core]
|
||||||
|
|
||||||
enum class InverseReturnMode {
|
|
||||||
/// Specifies that functional inverses should always return a view.
|
|
||||||
AlwaysView,
|
|
||||||
/// Specifies that functional inverses should always return a non-view / copy.
|
|
||||||
NeverView,
|
|
||||||
/// Specifies that functional inverses should return a view unless a (copying)
|
|
||||||
/// scatter
|
|
||||||
/// inverse exists, in which case that will be used instead.
|
|
||||||
/// This avoids as_strided() calls that can be difficult for subclasses to
|
|
||||||
/// handle.
|
|
||||||
ViewOrScatterInverse,
|
|
||||||
};
|
|
||||||
|
|
||||||
#define FUNCTIONALIZATION_VIEWMETA_NAME(TYPE) \
|
|
||||||
static const char* name() { \
|
|
||||||
return #TYPE; \
|
|
||||||
}
|
|
||||||
|
|
||||||
#define FUNCTIONALIZATION_VIEWMETA_SERIALIZABLE_TUPLE(...) \
|
|
||||||
using SerializableTuple = std::tuple<__VA_ARGS__>
|
|
||||||
|
|
||||||
// ViewMeta is a class used by the functionalization pass to navigate between
|
// ViewMeta is a class used by the functionalization pass to navigate between
|
||||||
// a base tensor and a view tensor.
|
// a base tensor and a view tensor.
|
||||||
// For example, if I call `b = a.view1(...)`
|
// For example, if I call `b = a.view1(...)`
|
||||||
// the functionalization pass will generate and store a ViewMeta specialization
|
// the functionalization pass will generate and store a ViewMeta on b that looks
|
||||||
// for `view1` operation on b that looks like:
|
// like:
|
||||||
//
|
//
|
||||||
// struct TORCH_API view1_ViewMeta : public ViewMeta {
|
// ViewMeta(
|
||||||
// FUNCTIONALIZATION_VIEWMETA_NAME(view1_ViewMeta);
|
// [<captures>](const Tensor& base, int64_t mutated_view_idx) {
|
||||||
// FUNCTIONALIZATION_VIEWMETA_SERIALIZABLE_TUPLE(
|
// return base.view1(...);
|
||||||
// bool /* reapply_views */,
|
// },
|
||||||
// const std::vector<int64_t>&);
|
// [<captures>](const at::Tensor& base, const at::Tensor& mutated_view,
|
||||||
//
|
// int64_t mutated_view_idx) -> at::Tensor {
|
||||||
// view1_ViewMeta(const SerializableTuple& tpl)
|
// return at::functionalization::impl::view1_inverse(base, mutated_view,
|
||||||
// : view1_ViewMeta(std::get<0>(tpl), std::get<1>(tpl)) {}
|
// ...);
|
||||||
//
|
|
||||||
// view1_ViewMeta(bool reapply_views, const std::vector<int64_t>& size)
|
|
||||||
// : ViewMeta(/*has_symbolic_inputs=*/false),
|
|
||||||
// reapply_views(reapply_views),
|
|
||||||
// size(size) {}
|
|
||||||
//
|
|
||||||
// Tensor forward(const Tensor& base) override {
|
|
||||||
// return base.view1(...);
|
|
||||||
// }
|
// }
|
||||||
//
|
//
|
||||||
// Tensor reverse(const Tensor& base, const Tensor& mutated_view) override {
|
// The forward_fn lambda describes how to replay view1 on a tensor.
|
||||||
// return at::functionalization::impl::view1_inverse(base, mutated_view,
|
|
||||||
// ...);
|
|
||||||
// }
|
|
||||||
//
|
//
|
||||||
// SerializableTuple to_serializable_tuple() {
|
// The reverse_fn lambda describes how, given a tensor that is already a view,
|
||||||
// return std::make_tuple(reapply_views, size);
|
|
||||||
// }
|
|
||||||
//
|
|
||||||
// bool reapply_views;
|
|
||||||
// std::vector<int64_t> size;
|
|
||||||
// };
|
|
||||||
//
|
|
||||||
// The forward function describes how to replay view1 on a tensor.
|
|
||||||
//
|
|
||||||
// The reverse function describes how, given a tensor that is already a view,
|
|
||||||
// how to get the corresponding base tensor. See Note [Functionalization Pass:
|
// how to get the corresponding base tensor. See Note [Functionalization Pass:
|
||||||
// View Inverses] for details.
|
// View Inverses] for details.
|
||||||
//
|
|
||||||
// `SerializedTuple` is a typedef that defines an `std::tuple<...>` type
|
|
||||||
// representing the `ViewMeta` instance state. Methods that take in/return such
|
|
||||||
// a type are used for supporting pickle serialization.
|
|
||||||
struct ViewMeta {
|
struct ViewMeta {
|
||||||
ViewMeta(
|
ViewMeta(
|
||||||
|
std::function<Tensor(const Tensor&, int64_t)> forward,
|
||||||
|
std::function<Tensor(const Tensor&, const Tensor&, int64_t)> reverse,
|
||||||
bool has_symbolic_inputs,
|
bool has_symbolic_inputs,
|
||||||
bool is_multi_output = false,
|
bool is_multi_output = false,
|
||||||
bool is_as_strided = false,
|
bool is_as_strided = false,
|
||||||
int64_t out_idx = 0)
|
int64_t out_idx = 0)
|
||||||
: out_index(out_idx),
|
: forward_fn(std::move(forward)),
|
||||||
|
reverse_fn(std::move(reverse)),
|
||||||
|
out_index(out_idx),
|
||||||
is_multi_output(is_multi_output),
|
is_multi_output(is_multi_output),
|
||||||
is_as_strided(is_as_strided),
|
is_as_strided(is_as_strided),
|
||||||
has_symbolic_inputs(has_symbolic_inputs) {}
|
has_symbolic_inputs(has_symbolic_inputs) {}
|
||||||
|
|
||||||
virtual ~ViewMeta() = default;
|
std::function<Tensor(const Tensor&, int64_t)> forward_fn;
|
||||||
|
std::function<Tensor(const Tensor&, const Tensor&, int64_t)> reverse_fn;
|
||||||
virtual Tensor forward(const Tensor& base) = 0;
|
|
||||||
virtual Tensor reverse(const Tensor& base, const Tensor& mutated_view) = 0;
|
|
||||||
|
|
||||||
// See Note [out_idx in ViewMeta]
|
// See Note [out_idx in ViewMeta]
|
||||||
int64_t out_index;
|
int64_t out_index;
|
||||||
|
|
||||||
@ -102,17 +57,10 @@ struct ViewMeta {
|
|||||||
// Tells us if this view operation has any symbolic inputs
|
// Tells us if this view operation has any symbolic inputs
|
||||||
bool has_symbolic_inputs;
|
bool has_symbolic_inputs;
|
||||||
|
|
||||||
// Returns a new ViewMeta with the same forward/reverse
|
// Returns a copy of the current ViewMeta, if out_idx matches the current
|
||||||
|
// out_index. Otherwise, returns a new ViewMeta with the same forward/reverse
|
||||||
// functions, but a new out index.
|
// functions, but a new out index.
|
||||||
//
|
ViewMeta to_out_idx(int64_t out_idx);
|
||||||
// This method should be implemented by those `ViewMeta` that have more than
|
|
||||||
// one output.
|
|
||||||
virtual std::shared_ptr<ViewMeta> to_out_index(int64_t out_index) {
|
|
||||||
TORCH_CHECK_NOT_IMPLEMENTED(
|
|
||||||
false,
|
|
||||||
"ViewMeta::to_out_index not implemented. ",
|
|
||||||
"Likely because there's only one output.");
|
|
||||||
}
|
|
||||||
};
|
};
|
||||||
|
|
||||||
// FunctionalStorageImpl is a subclass of StorageImpl used by the
|
// FunctionalStorageImpl is a subclass of StorageImpl used by the
|
||||||
@ -145,14 +93,14 @@ struct TORCH_API FunctionalStorageImpl : public c10::StorageImpl {
|
|||||||
// NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
|
// NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
|
||||||
const at::Tensor new_val;
|
const at::Tensor new_val;
|
||||||
// NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
|
// NOLINTNEXTLINE(cppcoreguidelines-avoid-const-or-ref-data-members)
|
||||||
const std::vector<std::shared_ptr<ViewMeta>> view_metas;
|
const std::vector<ViewMeta> view_metas;
|
||||||
};
|
};
|
||||||
|
|
||||||
explicit FunctionalStorageImpl(const Tensor& value);
|
explicit FunctionalStorageImpl(const Tensor& value);
|
||||||
|
|
||||||
void add_update(
|
void add_update(
|
||||||
const Tensor& updated_val,
|
const Tensor& updated_val,
|
||||||
const std::vector<std::shared_ptr<ViewMeta>>& view_metas);
|
const std::vector<ViewMeta>& view_metas);
|
||||||
bool apply_updates();
|
bool apply_updates();
|
||||||
const Tensor& base() {
|
const Tensor& base() {
|
||||||
return base_;
|
return base_;
|
||||||
|
|||||||
@ -129,19 +129,17 @@ void FunctionalTensorWrapper::freeze_storage() const {
|
|||||||
// - view_value: The output tensor that we need to wrap.
|
// - view_value: The output tensor that we need to wrap.
|
||||||
// - base: The "base" of the view that `view_value` was generated from.
|
// - base: The "base" of the view that `view_value` was generated from.
|
||||||
// See Note [Functionalization: Alias Removal Part 2] for more details on the mutation replay logic.
|
// See Note [Functionalization: Alias Removal Part 2] for more details on the mutation replay logic.
|
||||||
FunctionalTensorWrapper::FunctionalTensorWrapper(
|
FunctionalTensorWrapper::FunctionalTensorWrapper(const Tensor& view_value, const FunctionalTensorWrapper* base, const functionalization::ViewMeta& meta)
|
||||||
const Tensor& view_value,
|
: c10::TensorImpl(
|
||||||
const FunctionalTensorWrapper* base,
|
c10::DispatchKeySet(DispatchKey::Functionalize),
|
||||||
const std::shared_ptr<functionalization::ViewMeta>& meta)
|
view_value.dtype(),
|
||||||
: c10::TensorImpl(
|
base->storage().data_ptr().device()
|
||||||
c10::DispatchKeySet(DispatchKey::Functionalize),
|
),
|
||||||
view_value.dtype(),
|
value_(view_value),
|
||||||
base->storage().data_ptr().device()),
|
is_multi_output_view_(base->is_multi_output_view_ || meta.is_multi_output),
|
||||||
value_(view_value),
|
was_storage_changed_(base->was_storage_changed_),
|
||||||
is_multi_output_view_(
|
is_symbolic_(base->is_symbolic_)
|
||||||
base->is_multi_output_view_ || meta->is_multi_output),
|
{
|
||||||
was_storage_changed_(base->was_storage_changed_),
|
|
||||||
is_symbolic_(base->is_symbolic_) {
|
|
||||||
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(value_));
|
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(value_));
|
||||||
TORCH_INTERNAL_ASSERT(!value_.key_set().has(c10::DispatchKey::Functionalize));
|
TORCH_INTERNAL_ASSERT(!value_.key_set().has(c10::DispatchKey::Functionalize));
|
||||||
set_constructor_metadata();
|
set_constructor_metadata();
|
||||||
@ -150,10 +148,11 @@ FunctionalTensorWrapper::FunctionalTensorWrapper(
|
|||||||
view_metas_ = base->view_metas_; // copy
|
view_metas_ = base->view_metas_; // copy
|
||||||
}
|
}
|
||||||
view_metas_.push_back(meta);
|
view_metas_.push_back(meta);
|
||||||
maybe_mark_symbolic(meta.get());
|
maybe_mark_symbolic(meta);
|
||||||
storage_ = base->storage_; // alias this tensor's storage with the base tensor's
|
storage_ = base->storage_; // alias this tensor's storage with the base tensor's
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
||||||
functionalization::FunctionalStorageImpl* FunctionalTensorWrapper::functional_storage_impl() const {
|
functionalization::FunctionalStorageImpl* FunctionalTensorWrapper::functional_storage_impl() const {
|
||||||
return static_cast<functionalization::FunctionalStorageImpl*>(storage_.unsafeGetStorageImpl());
|
return static_cast<functionalization::FunctionalStorageImpl*>(storage_.unsafeGetStorageImpl());
|
||||||
}
|
}
|
||||||
@ -177,18 +176,18 @@ bool FunctionalTensorWrapper::is_up_to_date() const {
|
|||||||
}
|
}
|
||||||
|
|
||||||
// See Note [Functionalization Pass - Inplace View Ops]
|
// See Note [Functionalization Pass - Inplace View Ops]
|
||||||
void FunctionalTensorWrapper::mutate_view_meta(const std::shared_ptr<at::functionalization::ViewMeta>& meta) {
|
void FunctionalTensorWrapper::mutate_view_meta(const at::functionalization::ViewMeta& meta) {
|
||||||
view_metas_.push_back(meta);
|
view_metas_.push_back(meta);
|
||||||
// Manually track the fact that this tensor received a metadata mutation!
|
// Manually track the fact that this tensor received a metadata mutation!
|
||||||
has_metadata_mutation_ = true;
|
has_metadata_mutation_ = true;
|
||||||
// Mark this tensor as being symbolic if there are any symbolic inputs used by the view operation.
|
// Mark this tensor as being symbolic if there are any symbolic inputs used by the view operation.
|
||||||
maybe_mark_symbolic(meta.get());
|
maybe_mark_symbolic(meta);
|
||||||
// Note [Functionalization Pass - Inplace View Ops]
|
// Note [Functionalization Pass - Inplace View Ops]
|
||||||
// So, these ops are special - they're mutation AND view ops. They get special codegen.
|
// So, these ops are special - they're mutation AND view ops. They get special codegen.
|
||||||
// An example is transpose_, e.g. `a.transpose_()`
|
// An example is transpose_, e.g. `a.transpose_()`
|
||||||
// Calling transpose_() should ensure that a gets an alias, and append the new ViewMeta to a's current list of ViewMetas.
|
// Calling transpose_() should ensure that a gets an alias, and append the new ViewMeta to a's current list of ViewMetas.
|
||||||
at::AutoDispatchSkipFunctionalize guard;
|
at::AutoDispatchSkipFunctionalize guard;
|
||||||
value_ = meta->forward(value_);
|
value_ = meta.forward_fn(value_, meta.out_index);
|
||||||
TORCH_INTERNAL_ASSERT(!value_.key_set().has(c10::DispatchKey::Functionalize));
|
TORCH_INTERNAL_ASSERT(!value_.key_set().has(c10::DispatchKey::Functionalize));
|
||||||
}
|
}
|
||||||
|
|
||||||
@ -369,8 +368,15 @@ void FunctionalTensorWrapper::sync_() {
|
|||||||
regenerate_from_base();
|
regenerate_from_base();
|
||||||
}
|
}
|
||||||
|
|
||||||
const std::vector<std::shared_ptr<functionalization::ViewMeta>>& FunctionalTensorWrapper::view_metas() const {
|
Tensor FunctionalTensorWrapper::apply_view_metas(const Tensor& base) {
|
||||||
return view_metas_;
|
auto t = base;
|
||||||
|
|
||||||
|
// Reapply views to get the viewed tensor from the base in alias_
|
||||||
|
for (auto& view_meta: view_metas_) {
|
||||||
|
t = view_meta.forward_fn(t, view_meta.out_index);
|
||||||
|
}
|
||||||
|
|
||||||
|
return t;
|
||||||
}
|
}
|
||||||
|
|
||||||
void FunctionalTensorWrapper::regenerate_from_base() {
|
void FunctionalTensorWrapper::regenerate_from_base() {
|
||||||
@ -379,7 +385,7 @@ void FunctionalTensorWrapper::regenerate_from_base() {
|
|||||||
auto t = storage_impl->base();
|
auto t = storage_impl->base();
|
||||||
|
|
||||||
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
|
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
|
||||||
t = at::functionalization::impl::apply_view_meta_sequence(t, view_metas_);
|
t = apply_view_metas(t);
|
||||||
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
|
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(t));
|
||||||
|
|
||||||
replace_(t, /*from_lazy_regenerate=*/true);
|
replace_(t, /*from_lazy_regenerate=*/true);
|
||||||
@ -721,11 +727,11 @@ bool isFunctionalTensor(const std::optional<Tensor>& t) {
|
|||||||
}
|
}
|
||||||
|
|
||||||
bool isFunctionalTensor(const c10::List<::std::optional<Tensor>>& t_list) {
|
bool isFunctionalTensor(const c10::List<::std::optional<Tensor>>& t_list) {
|
||||||
if (t_list.empty()) { return false; }
|
if (t_list.empty()) return false;
|
||||||
auto functional_count = 0;
|
auto functional_count = 0;
|
||||||
for (const auto i : c10::irange(t_list.size())) {
|
for (const auto i : c10::irange(t_list.size())) {
|
||||||
auto const & e= t_list[i];
|
auto const & e= t_list[i];
|
||||||
if (!e.has_value() || !e->defined()) { continue; }
|
if (!e.has_value() || !e->defined()) continue;
|
||||||
if (isFunctionalTensor(e)) {
|
if (isFunctionalTensor(e)) {
|
||||||
++functional_count;
|
++functional_count;
|
||||||
}
|
}
|
||||||
@ -735,10 +741,10 @@ bool isFunctionalTensor(const c10::List<::std::optional<Tensor>>& t_list) {
|
|||||||
|
|
||||||
template <typename T>
|
template <typename T>
|
||||||
static bool isFunctionalTensorIListRef(c10::IListRef<T> list) {
|
static bool isFunctionalTensorIListRef(c10::IListRef<T> list) {
|
||||||
if (list.size() == 0) { return false; }
|
if (list.size() == 0) return false;
|
||||||
auto functional_count = 0;
|
auto functional_count = 0;
|
||||||
for (const auto& tensor : list) {
|
for (const auto& tensor : list) {
|
||||||
if (!tensor.defined()) { continue; }
|
if (!tensor.defined()) continue;
|
||||||
if (isFunctionalTensor(tensor)) {
|
if (isFunctionalTensor(tensor)) {
|
||||||
++functional_count;
|
++functional_count;
|
||||||
}
|
}
|
||||||
@ -756,28 +762,20 @@ void freeze_functional_tensor(const Tensor& tensor) {
|
|||||||
functional_base_impl->freeze_storage();
|
functional_base_impl->freeze_storage();
|
||||||
}
|
}
|
||||||
|
|
||||||
Tensor create_functional_tensor_with_view_meta(
|
Tensor create_functional_tensor_with_view_meta(const at::Tensor& view_to_wrap, const at::Tensor& base, functionalization::ViewMeta meta, int64_t out_idx) {
|
||||||
const at::Tensor& view_to_wrap,
|
|
||||||
const at::Tensor& base,
|
|
||||||
const std::shared_ptr<functionalization::ViewMeta>& meta,
|
|
||||||
int64_t out_idx) {
|
|
||||||
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(view_to_wrap));
|
TORCH_INTERNAL_ASSERT(!at::functionalization::impl::isFunctionalTensor(view_to_wrap));
|
||||||
TORCH_INTERNAL_ASSERT(at::functionalization::impl::isFunctionalTensor(base));
|
TORCH_INTERNAL_ASSERT(at::functionalization::impl::isFunctionalTensor(base));
|
||||||
auto functional_base_impl = at::functionalization::impl::unsafeGetFunctionalWrapper(base);
|
auto functional_base_impl = at::functionalization::impl::unsafeGetFunctionalWrapper(base);
|
||||||
auto meta_ = meta;
|
|
||||||
if (out_idx != 0) {
|
if (out_idx != 0) {
|
||||||
// Note [out_idx in ViewMeta]
|
// Note [out_idx in ViewMeta]
|
||||||
// When a view op outputs multiple tensors, each output needs its own separate ViewMeta.
|
// When a view op outputs multiple tensors, each output needs its own separate ViewMeta.
|
||||||
// Each ViewMeta also tracks the index of the particular output tensor, which is needed in the reverse function.
|
// Each ViewMeta also tracks the index of the particular output tensor, which is needed in the reverse function.
|
||||||
meta_ = meta->to_out_index(out_idx);
|
meta = meta.to_out_idx(out_idx);
|
||||||
}
|
}
|
||||||
return at::detail::make_tensor<FunctionalTensorWrapper>(view_to_wrap, functional_base_impl, meta_);
|
return at::detail::make_tensor<FunctionalTensorWrapper>(view_to_wrap, functional_base_impl, meta);
|
||||||
}
|
}
|
||||||
|
|
||||||
std::vector<Tensor> create_functional_tensor_with_view_meta(
|
std::vector<Tensor> create_functional_tensor_with_view_meta(ITensorListRef view_to_wrap, const at::Tensor& base, const functionalization::ViewMeta& meta) {
|
||||||
ITensorListRef view_to_wrap,
|
|
||||||
const at::Tensor& base,
|
|
||||||
const std::shared_ptr<functionalization::ViewMeta>& meta) {
|
|
||||||
std::vector<Tensor> outputs(view_to_wrap.size());
|
std::vector<Tensor> outputs(view_to_wrap.size());
|
||||||
int64_t i = 0;
|
int64_t i = 0;
|
||||||
for (const auto& tensor : view_to_wrap) {
|
for (const auto& tensor : view_to_wrap) {
|
||||||
@ -787,22 +785,12 @@ std::vector<Tensor> create_functional_tensor_with_view_meta(
|
|||||||
return outputs;
|
return outputs;
|
||||||
}
|
}
|
||||||
|
|
||||||
void mutate_view_meta(const at::Tensor& self, const std::shared_ptr<functionalization::ViewMeta>& meta) {
|
void mutate_view_meta(const at::Tensor& self, const functionalization::ViewMeta& meta) {
|
||||||
TORCH_INTERNAL_ASSERT(at::functionalization::impl::isFunctionalTensor(self));
|
TORCH_INTERNAL_ASSERT(at::functionalization::impl::isFunctionalTensor(self));
|
||||||
auto self_impl = at::functionalization::impl::unsafeGetFunctionalWrapper(self);
|
auto self_impl = at::functionalization::impl::unsafeGetFunctionalWrapper(self);
|
||||||
self_impl->mutate_view_meta(meta);
|
self_impl->mutate_view_meta(meta);
|
||||||
}
|
}
|
||||||
|
|
||||||
Tensor apply_view_meta_sequence(
|
|
||||||
const Tensor& base,
|
|
||||||
const std::vector<std::shared_ptr<functionalization::ViewMeta>>& sequence) {
|
|
||||||
Tensor r = base;
|
|
||||||
for (auto& vm : sequence) {
|
|
||||||
r = vm->forward(r);
|
|
||||||
}
|
|
||||||
return r;
|
|
||||||
}
|
|
||||||
|
|
||||||
// Note [Propagating strides in the functionalization pass]
|
// Note [Propagating strides in the functionalization pass]
|
||||||
// In order to properly compute stride information, the functionalization pass
|
// In order to properly compute stride information, the functionalization pass
|
||||||
// calls each {view} reference implementations with meta tensors.
|
// calls each {view} reference implementations with meta tensors.
|
||||||
@ -896,7 +884,7 @@ void functionalize_op_helper(const c10::OperatorHandle& op, torch::jit::Stack* s
|
|||||||
const auto& ivalue = returns[idx];
|
const auto& ivalue = returns[idx];
|
||||||
if (ivalue.isTensor()) {
|
if (ivalue.isTensor()) {
|
||||||
const auto& t = ivalue.toTensor();
|
const auto& t = ivalue.toTensor();
|
||||||
if (!t.defined()) { continue; }
|
if (!t.defined()) continue;
|
||||||
at::functionalization::impl::sync(t);
|
at::functionalization::impl::sync(t);
|
||||||
auto t_new = c10::IValue(at::functionalization::impl::from_functional_tensor(t));
|
auto t_new = c10::IValue(at::functionalization::impl::from_functional_tensor(t));
|
||||||
(*stack)[returns_begin + idx] = t_new;
|
(*stack)[returns_begin + idx] = t_new;
|
||||||
|
|||||||
@ -56,7 +56,7 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
|
|||||||
explicit FunctionalTensorWrapper(
|
explicit FunctionalTensorWrapper(
|
||||||
const Tensor& view_value,
|
const Tensor& view_value,
|
||||||
const FunctionalTensorWrapper* base,
|
const FunctionalTensorWrapper* base,
|
||||||
const std::shared_ptr<functionalization::ViewMeta>& meta);
|
const functionalization::ViewMeta& meta);
|
||||||
|
|
||||||
// Get the underlying, actual tensor, that doesn't know anything about
|
// Get the underlying, actual tensor, that doesn't know anything about
|
||||||
// functionalization.
|
// functionalization.
|
||||||
@ -99,17 +99,17 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
|
|||||||
->are_all_mutations_under_no_grad_or_inference_mode();
|
->are_all_mutations_under_no_grad_or_inference_mode();
|
||||||
}
|
}
|
||||||
|
|
||||||
void maybe_mark_symbolic(functionalization::ViewMeta* meta) {
|
void maybe_mark_symbolic(const functionalization::ViewMeta& meta) {
|
||||||
is_symbolic_ = is_symbolic_ | meta->has_symbolic_inputs;
|
is_symbolic_ = is_symbolic_ | meta.has_symbolic_inputs;
|
||||||
}
|
}
|
||||||
|
|
||||||
bool is_symbolic() const {
|
bool is_symbolic() const {
|
||||||
return is_symbolic_;
|
return is_symbolic_;
|
||||||
}
|
}
|
||||||
|
|
||||||
// Retrieves the ViewMeta sequence of this tensor.
|
// Runs the forward_fn of every ViewMeta collected in the current instance
|
||||||
const std::vector<std::shared_ptr<functionalization::ViewMeta>>& view_metas()
|
// to some other base.
|
||||||
const;
|
Tensor apply_view_metas(const Tensor& base);
|
||||||
|
|
||||||
// Sync's the underlying tensor with its alias, if it's out of date. This
|
// Sync's the underlying tensor with its alias, if it's out of date. This
|
||||||
// involves two steps: 1) Apply any pending updates/mutations to the alias 2)
|
// involves two steps: 1) Apply any pending updates/mutations to the alias 2)
|
||||||
@ -146,8 +146,7 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
|
|||||||
// from the base tensor. This method is used by inplace-view ops like
|
// from the base tensor. This method is used by inplace-view ops like
|
||||||
// transpose_. It appends a ViewMeta to the existing stack, and refreshes the
|
// transpose_. It appends a ViewMeta to the existing stack, and refreshes the
|
||||||
// tensor by replaying the views off of the alias.
|
// tensor by replaying the views off of the alias.
|
||||||
void mutate_view_meta(
|
void mutate_view_meta(const at::functionalization::ViewMeta& meta);
|
||||||
const std::shared_ptr<at::functionalization::ViewMeta>& meta);
|
|
||||||
|
|
||||||
// Custom implementation of self.set_(src)
|
// Custom implementation of self.set_(src)
|
||||||
void set__impl(const FunctionalTensorWrapper* other);
|
void set__impl(const FunctionalTensorWrapper* other);
|
||||||
@ -286,7 +285,7 @@ struct TORCH_API FunctionalTensorWrapper : public c10::TensorImpl {
|
|||||||
bool is_symbolic_ = false;
|
bool is_symbolic_ = false;
|
||||||
|
|
||||||
size_t generation_ = 0;
|
size_t generation_ = 0;
|
||||||
std::vector<std::shared_ptr<at::functionalization::ViewMeta>> view_metas_;
|
std::vector<at::functionalization::ViewMeta> view_metas_;
|
||||||
|
|
||||||
protected:
|
protected:
|
||||||
static void copy_tensor_metadata(
|
static void copy_tensor_metadata(
|
||||||
@ -378,20 +377,16 @@ TORCH_API void propagate_xla_data_direct(
|
|||||||
Tensor create_functional_tensor_with_view_meta(
|
Tensor create_functional_tensor_with_view_meta(
|
||||||
const Tensor& view_to_wrap,
|
const Tensor& view_to_wrap,
|
||||||
const Tensor& base,
|
const Tensor& base,
|
||||||
const std::shared_ptr<functionalization::ViewMeta>& meta,
|
functionalization::ViewMeta meta,
|
||||||
int64_t out_idx = 0);
|
int64_t out_idx = 0);
|
||||||
std::vector<Tensor> create_functional_tensor_with_view_meta(
|
std::vector<Tensor> create_functional_tensor_with_view_meta(
|
||||||
ITensorListRef view_to_wrap,
|
ITensorListRef view_to_wrap,
|
||||||
const Tensor& base,
|
const Tensor& base,
|
||||||
const std::shared_ptr<functionalization::ViewMeta>& meta);
|
const functionalization::ViewMeta& meta);
|
||||||
|
|
||||||
void mutate_view_meta(
|
void mutate_view_meta(
|
||||||
const Tensor& self,
|
const Tensor& self,
|
||||||
const std::shared_ptr<functionalization::ViewMeta>& meta);
|
const functionalization::ViewMeta& meta);
|
||||||
|
|
||||||
TORCH_API Tensor apply_view_meta_sequence(
|
|
||||||
const Tensor& base,
|
|
||||||
const std::vector<std::shared_ptr<functionalization::ViewMeta>>& sequence);
|
|
||||||
|
|
||||||
void set_sizes_strides_offset(const Tensor& out, const Tensor& meta_out);
|
void set_sizes_strides_offset(const Tensor& out, const Tensor& meta_out);
|
||||||
void set_sizes_strides_offset(
|
void set_sizes_strides_offset(
|
||||||
|
|||||||
@ -1,5 +1,3 @@
|
|||||||
#include <ATen/FunctionalizeFallbackKernel.h>
|
|
||||||
|
|
||||||
#include <ATen/core/dispatch/Dispatcher.h>
|
#include <ATen/core/dispatch/Dispatcher.h>
|
||||||
#include <ATen/core/LegacyTypeDispatch.h>
|
#include <ATen/core/LegacyTypeDispatch.h>
|
||||||
#include <ATen/EmptyTensor.h>
|
#include <ATen/EmptyTensor.h>
|
||||||
@ -9,6 +7,7 @@
|
|||||||
#include <torch/library.h>
|
#include <torch/library.h>
|
||||||
#include <c10/util/irange.h>
|
#include <c10/util/irange.h>
|
||||||
#include <c10/util/strides.h>
|
#include <c10/util/strides.h>
|
||||||
|
#include <ATen/EmptyTensor.h>
|
||||||
|
|
||||||
#ifndef AT_PER_OPERATOR_HEADERS
|
#ifndef AT_PER_OPERATOR_HEADERS
|
||||||
#include <ATen/ATen.h>
|
#include <ATen/ATen.h>
|
||||||
@ -29,31 +28,6 @@
|
|||||||
#include <utility>
|
#include <utility>
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
namespace at::functionalization {
|
|
||||||
|
|
||||||
Tensor resize__ViewMeta::forward(const Tensor& base) {
|
|
||||||
if (reapply_views) {
|
|
||||||
return base.as_strided(size, c10::contiguous_strides(size));
|
|
||||||
} else {
|
|
||||||
return at::as_strided_copy(base, size, c10::contiguous_strides(size));
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
Tensor resize__ViewMeta::reverse(const Tensor& base, const Tensor& mutated_view) {
|
|
||||||
return base.as_strided_scatter(
|
|
||||||
mutated_view, size, c10::contiguous_strides(size));
|
|
||||||
}
|
|
||||||
|
|
||||||
Tensor _unsafe_view_ViewMeta::forward(const Tensor& base) {
|
|
||||||
return at::_unsafe_view_symint(base, size);
|
|
||||||
}
|
|
||||||
|
|
||||||
Tensor _unsafe_view_ViewMeta::reverse(const Tensor& base, const Tensor& mutated_view) {
|
|
||||||
return at::_unsafe_view_symint(mutated_view, base.sym_sizes());
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace at::functionalization
|
|
||||||
|
|
||||||
namespace {
|
namespace {
|
||||||
void functionalizeFallback(const c10::OperatorHandle& op, c10::DispatchKeySet dispatchKeySet [[maybe_unused]], torch::jit::Stack* stack) {
|
void functionalizeFallback(const c10::OperatorHandle& op, c10::DispatchKeySet dispatchKeySet [[maybe_unused]], torch::jit::Stack* stack) {
|
||||||
const auto& schema = op.schema();
|
const auto& schema = op.schema();
|
||||||
@ -132,9 +106,7 @@ namespace {
|
|||||||
const auto& ivalue = returns[idx];
|
const auto& ivalue = returns[idx];
|
||||||
if (ivalue.isTensor() && should_wrap_outputs) {
|
if (ivalue.isTensor() && should_wrap_outputs) {
|
||||||
const auto& t = ivalue.toTensor();
|
const auto& t = ivalue.toTensor();
|
||||||
if (!t.defined()) {
|
if (!t.defined()) continue;
|
||||||
continue;
|
|
||||||
}
|
|
||||||
auto t_new = c10::IValue(at::functionalization::impl::to_functional_tensor(t));
|
auto t_new = c10::IValue(at::functionalization::impl::to_functional_tensor(t));
|
||||||
(*stack)[returns_begin + idx] = t_new;
|
(*stack)[returns_begin + idx] = t_new;
|
||||||
} else if (ivalue.isTensorList() && should_wrap_outputs) {
|
} else if (ivalue.isTensorList() && should_wrap_outputs) {
|
||||||
@ -197,8 +169,19 @@ static const at::Tensor & resize__functionalization(c10::DispatchKeySet dispatch
|
|||||||
// The output of resizing is equivalent to taking a slice of a larger tensor.
|
// The output of resizing is equivalent to taking a slice of a larger tensor.
|
||||||
// We have to emulate this "slicing" with an as_strided call.
|
// We have to emulate this "slicing" with an as_strided call.
|
||||||
auto reapply_views = at::functionalization::impl::getFunctionalizationReapplyViewsTLS();
|
auto reapply_views = at::functionalization::impl::getFunctionalizationReapplyViewsTLS();
|
||||||
auto view_meta = std::make_shared<at::functionalization::resize__ViewMeta>(
|
at::functionalization::ViewMeta view_meta = at::functionalization::ViewMeta(
|
||||||
reapply_views, size.vec());
|
[reapply_views = reapply_views, size = size.vec()](const at::Tensor & base, int64_t mutated_view_idx [[maybe_unused]]) -> at::Tensor {
|
||||||
|
if (reapply_views) {
|
||||||
|
return base.as_strided(size, c10::contiguous_strides(size));
|
||||||
|
} else {
|
||||||
|
return at::as_strided_copy(base, size, c10::contiguous_strides(size));
|
||||||
|
}
|
||||||
|
},
|
||||||
|
[size = size.vec()](const at::Tensor & base, const at::Tensor & mutated_view, int64_t mutated_view_idx [[maybe_unused]]) -> at::Tensor {
|
||||||
|
return base.as_strided_scatter(mutated_view, size, c10::contiguous_strides(size));
|
||||||
|
},
|
||||||
|
/*has_symbolic_inputs=*/false
|
||||||
|
);
|
||||||
at::functionalization::impl::mutate_view_meta(self, view_meta);
|
at::functionalization::impl::mutate_view_meta(self, view_meta);
|
||||||
return self;
|
return self;
|
||||||
}
|
}
|
||||||
@ -317,11 +300,17 @@ static at::Tensor _unsafe_view_functionalize(const at::Tensor & self, at::SymInt
|
|||||||
tmp_output = at::_unsafe_view_symint(self_, size);
|
tmp_output = at::_unsafe_view_symint(self_, size);
|
||||||
}
|
}
|
||||||
|
|
||||||
bool has_symbolic_inputs = std::any_of(
|
bool has_symbolic_inputs = std::any_of(size.begin(), size.end(), [=](auto& s) { return s.is_symbolic(); });
|
||||||
size.begin(), size.end(), [=](auto& s) { return s.is_symbolic(); });
|
|
||||||
auto view_meta =
|
at::functionalization::ViewMeta view_meta = at::functionalization::ViewMeta(
|
||||||
std::make_shared<at::functionalization::_unsafe_view_ViewMeta>(
|
[size = size.vec()](const at::Tensor & base, int64_t mutated_view_idx [[maybe_unused]]) -> at::Tensor {
|
||||||
has_symbolic_inputs, size.vec());
|
return at::_unsafe_view_symint(base, size);
|
||||||
|
},
|
||||||
|
[size = size.vec()](const at::Tensor & base, const at::Tensor & mutated_view, int64_t mutated_view_idx [[maybe_unused]]) -> at::Tensor {
|
||||||
|
return at::_unsafe_view_symint(mutated_view, base.sym_sizes());
|
||||||
|
},
|
||||||
|
/*has_symbolic_inputs=*/has_symbolic_inputs
|
||||||
|
);
|
||||||
|
|
||||||
auto out = at::functionalization::impl::create_functional_tensor_with_view_meta(tmp_output, self, std::move(view_meta));
|
auto out = at::functionalization::impl::create_functional_tensor_with_view_meta(tmp_output, self, std::move(view_meta));
|
||||||
// See Note [Propagating strides in the functionalization pass]
|
// See Note [Propagating strides in the functionalization pass]
|
||||||
|
|||||||
@ -1,58 +0,0 @@
|
|||||||
#pragma once
|
|
||||||
|
|
||||||
#include <ATen/FunctionalStorageImpl.h>
|
|
||||||
|
|
||||||
namespace at::functionalization {
|
|
||||||
|
|
||||||
// `ViewMeta` implementation for `resize_` operation.
|
|
||||||
struct TORCH_API resize__ViewMeta : public ViewMeta {
|
|
||||||
FUNCTIONALIZATION_VIEWMETA_NAME(resize__ViewMeta)
|
|
||||||
FUNCTIONALIZATION_VIEWMETA_SERIALIZABLE_TUPLE(
|
|
||||||
bool /* reapply_views */,
|
|
||||||
const std::vector<int64_t>&);
|
|
||||||
|
|
||||||
resize__ViewMeta(const SerializableTuple& tpl)
|
|
||||||
: resize__ViewMeta(std::get<0>(tpl), std::get<1>(tpl)) {}
|
|
||||||
|
|
||||||
resize__ViewMeta(bool reapply_views, const std::vector<int64_t>& size)
|
|
||||||
: ViewMeta(/*has_symbolic_inputs=*/false),
|
|
||||||
reapply_views(reapply_views),
|
|
||||||
size(size) {}
|
|
||||||
|
|
||||||
Tensor forward(const Tensor& base) override;
|
|
||||||
Tensor reverse(const Tensor& base, const Tensor& mutated_view) override;
|
|
||||||
|
|
||||||
SerializableTuple to_serializable_tuple() {
|
|
||||||
return std::make_tuple(reapply_views, size);
|
|
||||||
}
|
|
||||||
|
|
||||||
bool reapply_views;
|
|
||||||
std::vector<int64_t> size;
|
|
||||||
};
|
|
||||||
|
|
||||||
// `ViewMeta` implementation for `_unsafe_view` operation.
|
|
||||||
struct TORCH_API _unsafe_view_ViewMeta : public ViewMeta {
|
|
||||||
FUNCTIONALIZATION_VIEWMETA_NAME(_unsafe_view_ViewMeta)
|
|
||||||
FUNCTIONALIZATION_VIEWMETA_SERIALIZABLE_TUPLE(
|
|
||||||
bool /* has_symbolic_inputs */,
|
|
||||||
const std::vector<c10::SymInt>&);
|
|
||||||
|
|
||||||
_unsafe_view_ViewMeta(const SerializableTuple& tpl)
|
|
||||||
: _unsafe_view_ViewMeta(std::get<0>(tpl), std::get<1>(tpl)) {}
|
|
||||||
|
|
||||||
_unsafe_view_ViewMeta(
|
|
||||||
bool has_symbolic_inputs,
|
|
||||||
const std::vector<c10::SymInt>& size)
|
|
||||||
: ViewMeta(has_symbolic_inputs), size(size) {}
|
|
||||||
|
|
||||||
Tensor forward(const Tensor& base) override;
|
|
||||||
Tensor reverse(const Tensor& base, const Tensor& mutated_view) override;
|
|
||||||
|
|
||||||
SerializableTuple to_serializable_tuple() {
|
|
||||||
return std::make_tuple(has_symbolic_inputs, size);
|
|
||||||
}
|
|
||||||
|
|
||||||
std::vector<c10::SymInt> size;
|
|
||||||
};
|
|
||||||
|
|
||||||
} // namespace at::functionalization
|
|
||||||
@ -45,39 +45,7 @@ inline void infer_size_impl(
|
|||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
if (infer_dim) {
|
auto set_infer_dim = [&]() {
|
||||||
// numel is the product of known sizes, it has to be divisible by newsize.
|
|
||||||
// and newsize should be positive unless newsize == numel (we throw
|
|
||||||
// different) error message in that case.
|
|
||||||
if constexpr (std::is_same_v<NumelType, c10::SymInt>) {
|
|
||||||
auto v = newsize.maybe_as_int();
|
|
||||||
if (v and *v == 0) {
|
|
||||||
// Avoid div by 0 when sym_eq(numel % newsize, 0) is constructed!
|
|
||||||
// which may happen when newsize is not a symbol! if its a symbol
|
|
||||||
// division won't happen anyway during compile.
|
|
||||||
TORCH_MAYBE_SYM_CHECK(
|
|
||||||
numel == newsize,
|
|
||||||
"shape '",
|
|
||||||
shape,
|
|
||||||
"' is invalid for input of size ",
|
|
||||||
numel);
|
|
||||||
} else {
|
|
||||||
auto cond = sym_gt(newsize, 0)
|
|
||||||
.sym_and(sym_eq(numel % newsize, 0))
|
|
||||||
.sym_or(sym_eq(numel, newsize));
|
|
||||||
TORCH_MAYBE_SYM_CHECK(
|
|
||||||
cond, "shape '", shape, "' is invalid for input of size ", numel);
|
|
||||||
}
|
|
||||||
|
|
||||||
} else {
|
|
||||||
TORCH_CHECK(
|
|
||||||
(newsize > 0 && (numel % newsize == 0)) || numel == newsize,
|
|
||||||
"shape '",
|
|
||||||
shape,
|
|
||||||
"' is invalid for input of size ",
|
|
||||||
numel);
|
|
||||||
}
|
|
||||||
|
|
||||||
// We have a degree of freedom here to select the dimension size; follow
|
// We have a degree of freedom here to select the dimension size; follow
|
||||||
// NumPy semantics and just bail. However, a nice error message is needed
|
// NumPy semantics and just bail. However, a nice error message is needed
|
||||||
// because users often use `view` as a way to flatten & unflatten
|
// because users often use `view` as a way to flatten & unflatten
|
||||||
@ -86,15 +54,19 @@ inline void infer_size_impl(
|
|||||||
// works yet
|
// works yet
|
||||||
// empty_tensor.view(-1, 0)
|
// empty_tensor.view(-1, 0)
|
||||||
// doesn't.
|
// doesn't.
|
||||||
TORCH_MAYBE_SYM_CHECK(
|
TORCH_CHECK(
|
||||||
newsize != 0,
|
newsize != 0,
|
||||||
"cannot reshape tensor of 0 elements into shape ",
|
"cannot reshape tensor of 0 elements into shape ",
|
||||||
shape,
|
shape,
|
||||||
" because the unspecified dimension size -1 can be any "
|
" because the unspecified dimension size -1 can be any "
|
||||||
"value and is ambiguous");
|
"value and is ambiguous");
|
||||||
|
|
||||||
res[*infer_dim] = numel / newsize;
|
res[*infer_dim] = numel / newsize;
|
||||||
return;
|
return;
|
||||||
|
};
|
||||||
|
|
||||||
|
if (infer_dim && newsize > 0 && numel % newsize == 0) {
|
||||||
|
set_infer_dim();
|
||||||
|
return;
|
||||||
}
|
}
|
||||||
|
|
||||||
TORCH_MAYBE_SYM_CHECK(
|
TORCH_MAYBE_SYM_CHECK(
|
||||||
@ -103,6 +75,9 @@ inline void infer_size_impl(
|
|||||||
shape,
|
shape,
|
||||||
"' is invalid for input of size ",
|
"' is invalid for input of size ",
|
||||||
numel);
|
numel);
|
||||||
|
if (infer_dim) {
|
||||||
|
set_infer_dim();
|
||||||
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
inline std::vector<int64_t> infer_size(IntArrayRef shape, int64_t numel) {
|
inline std::vector<int64_t> infer_size(IntArrayRef shape, int64_t numel) {
|
||||||
|
|||||||
@ -1,22 +1,32 @@
|
|||||||
#include <ATen/core/PythonOpRegistrationTrampoline.h>
|
#include <ATen/core/PythonOpRegistrationTrampoline.h>
|
||||||
#include <c10/core/impl/PyInterpreterHooks.h>
|
|
||||||
|
|
||||||
// TODO: delete this
|
|
||||||
namespace at::impl {
|
namespace at::impl {
|
||||||
|
|
||||||
c10::impl::PyInterpreter* PythonOpRegistrationTrampoline::interpreter_ = nullptr;
|
// The strategy is that all python interpreters attempt to register themselves
|
||||||
|
// as the main interpreter, but only one wins. Only that interpreter is
|
||||||
|
// allowed to interact with the C++ dispatcher. Furthermore, when we execute
|
||||||
|
// logic on that interpreter, we do so hermetically, never setting pyobj field
|
||||||
|
// on Tensor.
|
||||||
|
|
||||||
|
std::atomic<c10::impl::PyInterpreter*>
|
||||||
|
PythonOpRegistrationTrampoline::interpreter_{nullptr};
|
||||||
|
|
||||||
c10::impl::PyInterpreter* PythonOpRegistrationTrampoline::getInterpreter() {
|
c10::impl::PyInterpreter* PythonOpRegistrationTrampoline::getInterpreter() {
|
||||||
return c10::impl::getGlobalPyInterpreter();
|
return PythonOpRegistrationTrampoline::interpreter_.load();
|
||||||
}
|
}
|
||||||
|
|
||||||
bool PythonOpRegistrationTrampoline::registerInterpreter(
|
bool PythonOpRegistrationTrampoline::registerInterpreter(
|
||||||
c10::impl::PyInterpreter* interp) {
|
c10::impl::PyInterpreter* interp) {
|
||||||
if (interpreter_ != nullptr) {
|
c10::impl::PyInterpreter* expected = nullptr;
|
||||||
|
interpreter_.compare_exchange_strong(expected, interp);
|
||||||
|
if (expected != nullptr) {
|
||||||
|
// This is the second (or later) Python interpreter, which means we need
|
||||||
|
// non-trivial hermetic PyObject TLS
|
||||||
|
c10::impl::HermeticPyObjectTLS::init_state();
|
||||||
return false;
|
return false;
|
||||||
|
} else {
|
||||||
|
return true;
|
||||||
}
|
}
|
||||||
interpreter_ = interp;
|
|
||||||
return true;
|
|
||||||
}
|
}
|
||||||
|
|
||||||
} // namespace at::impl
|
} // namespace at::impl
|
||||||
|
|||||||
@ -2,21 +2,19 @@
|
|||||||
|
|
||||||
#include <ATen/core/dispatch/Dispatcher.h>
|
#include <ATen/core/dispatch/Dispatcher.h>
|
||||||
|
|
||||||
// TODO: We can get rid of this
|
// TODO: this can probably live in c10
|
||||||
|
|
||||||
|
|
||||||
namespace at::impl {
|
namespace at::impl {
|
||||||
|
|
||||||
// Manages the single Python interpreter instance for PyTorch.
|
|
||||||
class TORCH_API PythonOpRegistrationTrampoline final {
|
class TORCH_API PythonOpRegistrationTrampoline final {
|
||||||
static c10::impl::PyInterpreter* interpreter_;
|
static std::atomic<c10::impl::PyInterpreter*> interpreter_;
|
||||||
|
|
||||||
public:
|
public:
|
||||||
// Register the Python interpreter. Returns true on first registration,
|
// Returns true if you successfully registered yourself (that means
|
||||||
// false if an interpreter was already registered.
|
// you are in the hot seat for doing the operator registrations!)
|
||||||
static bool registerInterpreter(c10::impl::PyInterpreter*);
|
static bool registerInterpreter(c10::impl::PyInterpreter*);
|
||||||
|
|
||||||
// Returns the registered interpreter via the global PyInterpreter hooks.
|
|
||||||
// Returns nullptr if no interpreter has been registered yet.
|
// Returns nullptr if no interpreter has been registered yet.
|
||||||
static c10::impl::PyInterpreter* getInterpreter();
|
static c10::impl::PyInterpreter* getInterpreter();
|
||||||
};
|
};
|
||||||
|
|||||||
@ -149,105 +149,5 @@ static inline void pack_vnni4(
|
|||||||
#endif
|
#endif
|
||||||
}
|
}
|
||||||
|
|
||||||
// This is a helper function for transpose_pack_vnni4
|
|
||||||
// Transform a [4, 16] block (with incontiguous output)
|
|
||||||
// Src:
|
|
||||||
// a1 a2 a3 a4 a5 a6 a7 a8 a9 a10 a11 a12 a13 a14 a15 a16
|
|
||||||
// b1 b2 b3 b4 b5 b6 b7 b8 b9 b10 b11 b12 b13 b14 b15 b16
|
|
||||||
// c1 c2 c3 c4 c5 c6 c7 c8 c9 c10 c11 c12 c13 c14 c15 c16
|
|
||||||
// d1 d2 d3 d4 d5 d6 d7 d8 d9 d10 d11 d12 d13 d14 d15 d16
|
|
||||||
// Dst:
|
|
||||||
// a1 a2 a3 a4 b1 b2 b3 b4 c1 c2 c3 c4 d1 d2 d3 d4
|
|
||||||
// a5 a6 a7 a8 b5 b6 b7 b8 c5 c6 c7 c8 d5 d6 d7 d8
|
|
||||||
// a9 a10 a11 a12 b9 b10 b11 b12 c9 c10 c11 c12 d9 d10 d11 d12
|
|
||||||
// a13 a14 a15 a16 b13 b14 b15 b16 c13 c14 c15 c16 d13 d14 d15 d16
|
|
||||||
template <typename scalar_t, typename = std::enable_if_t<sizeof(scalar_t) == 1>>
|
|
||||||
static inline void transpose_vnni4_pad_4x16_block(
|
|
||||||
const scalar_t* src,
|
|
||||||
scalar_t* dst,
|
|
||||||
int64_t ld_src,
|
|
||||||
int64_t ld_dst,
|
|
||||||
int krem = 4) {
|
|
||||||
#if defined(CPU_CAPABILITY_AVX512)
|
|
||||||
__m128i r[4];
|
|
||||||
for (int i = 0; i < krem; ++i) {
|
|
||||||
r[i] = _mm_loadu_si128(reinterpret_cast<const __m128i*>(src + i * ld_src));
|
|
||||||
}
|
|
||||||
for (int i = krem; i < 4; ++i) {
|
|
||||||
r[i] = _mm_setzero_si128();
|
|
||||||
}
|
|
||||||
|
|
||||||
// Transpose 4x16 bytes using unpack and shuffle
|
|
||||||
__m128i t0 = _mm_unpacklo_epi32(r[0], r[1]);
|
|
||||||
__m128i t1 = _mm_unpackhi_epi32(r[0], r[1]);
|
|
||||||
__m128i t2 = _mm_unpacklo_epi32(r[2], r[3]);
|
|
||||||
__m128i t3 = _mm_unpackhi_epi32(r[2], r[3]);
|
|
||||||
|
|
||||||
__m128i r0 = _mm_unpacklo_epi64(t0, t2);
|
|
||||||
__m128i r1 = _mm_unpackhi_epi64(t0, t2);
|
|
||||||
__m128i r2 = _mm_unpacklo_epi64(t1, t3);
|
|
||||||
__m128i r3 = _mm_unpackhi_epi64(t1, t3);
|
|
||||||
|
|
||||||
// Store output
|
|
||||||
if (krem == 4) {
|
|
||||||
// normal case
|
|
||||||
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst), r0);
|
|
||||||
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst + ld_dst), r1);
|
|
||||||
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst + ld_dst * 2), r2);
|
|
||||||
_mm_storeu_si128(reinterpret_cast<__m128i*>(dst + ld_dst * 3), r3);
|
|
||||||
} else {
|
|
||||||
// masked case
|
|
||||||
__mmask16 mask = (1ULL << (krem * 4)) - 1;
|
|
||||||
_mm_mask_storeu_epi8(dst, mask, r0);
|
|
||||||
_mm_mask_storeu_epi8(reinterpret_cast<__m128i*>(dst + ld_dst), mask, r1);
|
|
||||||
_mm_mask_storeu_epi8(
|
|
||||||
reinterpret_cast<__m128i*>(dst + ld_dst * 2), mask, r2);
|
|
||||||
_mm_mask_storeu_epi8(
|
|
||||||
reinterpret_cast<__m128i*>(dst + ld_dst * 3), mask, r3);
|
|
||||||
}
|
|
||||||
#else
|
|
||||||
TORCH_CHECK(
|
|
||||||
false,
|
|
||||||
"transpose_vnni4_pad_4x16_block is only supported when AVX-512 is supported")
|
|
||||||
#endif
|
|
||||||
}
|
|
||||||
|
|
||||||
// Do the transpose packing fusion with VNNI4
|
|
||||||
// Reorder [K, N] → [N/4, K, 4] (VNNI4-style layout for bit8)
|
|
||||||
template <typename scalar_t, typename = std::enable_if_t<sizeof(scalar_t) == 1>>
|
|
||||||
static inline void transpose_pack_vnni4(
|
|
||||||
const scalar_t* src,
|
|
||||||
scalar_t* dst,
|
|
||||||
int64_t ld_src,
|
|
||||||
int64_t K,
|
|
||||||
int64_t N) {
|
|
||||||
#if defined(CPU_CAPABILITY_AVX512)
|
|
||||||
TORCH_CHECK(
|
|
||||||
N % 16 == 0, "N needs to be multiple of 16 for transpose_pack_vnni4");
|
|
||||||
int64_t bk = 0;
|
|
||||||
int64_t _K = K / 4 * 4;
|
|
||||||
for (; bk < _K; bk += 4) {
|
|
||||||
int64_t bn = 0;
|
|
||||||
for (; bn < N; bn += 16) {
|
|
||||||
transpose_vnni4_pad_4x16_block(
|
|
||||||
src + bk * ld_src + bn, dst + bn * K + bk * 4, ld_src, K * 4);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
// Handle leftover K rows (< 4)
|
|
||||||
if (K % 4 != 0) {
|
|
||||||
int krem = K - bk;
|
|
||||||
int64_t bn = 0;
|
|
||||||
for (; bn < N; bn += 16) {
|
|
||||||
transpose_vnni4_pad_4x16_block(
|
|
||||||
src + bk * ld_src + bn, dst + bn * K + bk * 4, ld_src, K * 4, krem);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
#else
|
|
||||||
TORCH_CHECK(
|
|
||||||
false, "transpose_pack_vnni4 is only supported when AVX-512 is supported")
|
|
||||||
#endif
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace CPU_CAPABILITY
|
} // namespace CPU_CAPABILITY
|
||||||
} // namespace at::vec
|
} // namespace at::vec
|
||||||
|
|||||||
@ -1637,7 +1637,9 @@ bool gemm_and_bias(
|
|||||||
if (activation == GEMMAndBiasActivationEpilogue::RELU) {
|
if (activation == GEMMAndBiasActivationEpilogue::RELU) {
|
||||||
epilogue = CUBLASLT_EPILOGUE_RELU_BIAS;
|
epilogue = CUBLASLT_EPILOGUE_RELU_BIAS;
|
||||||
} else if (activation == GEMMAndBiasActivationEpilogue::GELU) {
|
} else if (activation == GEMMAndBiasActivationEpilogue::GELU) {
|
||||||
|
#if CUDA_VERSION >= 11040 || defined(USE_ROCM)
|
||||||
epilogue = CUBLASLT_EPILOGUE_GELU_BIAS;
|
epilogue = CUBLASLT_EPILOGUE_GELU_BIAS;
|
||||||
|
#endif
|
||||||
}
|
}
|
||||||
|
|
||||||
if (bias != nullptr) {
|
if (bias != nullptr) {
|
||||||
@ -1929,6 +1931,7 @@ void scaled_gemm(
|
|||||||
bool use_fast_accum) {
|
bool use_fast_accum) {
|
||||||
// Note: see `cublasCommonArgs` for various non-intuitive manupulations
|
// Note: see `cublasCommonArgs` for various non-intuitive manupulations
|
||||||
// of input arguments to this function.
|
// of input arguments to this function.
|
||||||
|
#if CUDA_VERSION >= 11080 || defined(USE_ROCM)
|
||||||
const auto computeType = CUBLAS_COMPUTE_32F;
|
const auto computeType = CUBLAS_COMPUTE_32F;
|
||||||
const auto scaleType = CUDA_R_32F;
|
const auto scaleType = CUDA_R_32F;
|
||||||
const float alpha_val = 1.0;
|
const float alpha_val = 1.0;
|
||||||
@ -2130,6 +2133,8 @@ void scaled_gemm(
|
|||||||
" scaleType ",
|
" scaleType ",
|
||||||
scaleType);
|
scaleType);
|
||||||
return;
|
return;
|
||||||
|
#endif // if CUDA_VERSION >= 11080 || defined(USE_ROCM)
|
||||||
|
TORCH_CHECK(false, "scaled_gemm is only supported for CUDA 11.8 and above");
|
||||||
}
|
}
|
||||||
|
|
||||||
void int8_gemm(
|
void int8_gemm(
|
||||||
|
|||||||
@ -281,9 +281,6 @@ bool CUDAHooks::compiledWithMIOpen() const {
|
|||||||
|
|
||||||
bool CUDAHooks::supportsDilatedConvolutionWithCuDNN() const {
|
bool CUDAHooks::supportsDilatedConvolutionWithCuDNN() const {
|
||||||
#if AT_CUDNN_ENABLED()
|
#if AT_CUDNN_ENABLED()
|
||||||
if (!hasCUDA()) {
|
|
||||||
return false;
|
|
||||||
}
|
|
||||||
// NOTE: extra parenthesis around numbers disable clang warnings about
|
// NOTE: extra parenthesis around numbers disable clang warnings about
|
||||||
// dead code
|
// dead code
|
||||||
return true;
|
return true;
|
||||||
@ -294,9 +291,6 @@ bool CUDAHooks::supportsDilatedConvolutionWithCuDNN() const {
|
|||||||
|
|
||||||
bool CUDAHooks::supportsDepthwiseConvolutionWithCuDNN() const {
|
bool CUDAHooks::supportsDepthwiseConvolutionWithCuDNN() const {
|
||||||
#if AT_CUDNN_ENABLED()
|
#if AT_CUDNN_ENABLED()
|
||||||
if (!hasCUDA()) {
|
|
||||||
return false;
|
|
||||||
}
|
|
||||||
cudaDeviceProp* prop = at::cuda::getCurrentDeviceProperties();
|
cudaDeviceProp* prop = at::cuda::getCurrentDeviceProperties();
|
||||||
// Check for Volta cores
|
// Check for Volta cores
|
||||||
if (prop->major >= 7) {
|
if (prop->major >= 7) {
|
||||||
@ -311,9 +305,6 @@ bool CUDAHooks::supportsDepthwiseConvolutionWithCuDNN() const {
|
|||||||
|
|
||||||
bool CUDAHooks::supportsBFloat16ConvolutionWithCuDNNv8() const {
|
bool CUDAHooks::supportsBFloat16ConvolutionWithCuDNNv8() const {
|
||||||
#if AT_CUDNN_ENABLED()
|
#if AT_CUDNN_ENABLED()
|
||||||
if (!hasCUDA()) {
|
|
||||||
return false;
|
|
||||||
}
|
|
||||||
cudaDeviceProp* prop = at::cuda::getCurrentDeviceProperties();
|
cudaDeviceProp* prop = at::cuda::getCurrentDeviceProperties();
|
||||||
// Check for Volta cores
|
// Check for Volta cores
|
||||||
if (prop->major >= 8) {
|
if (prop->major >= 8) {
|
||||||
|
|||||||
@ -465,11 +465,8 @@ inline bool mps_conv_use_channels_last(const at::Tensor& input, const at::Tensor
|
|||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
|
|
||||||
auto is_channel_last = [](const at::Tensor& t) {
|
auto fmt = input.suggest_memory_format();
|
||||||
auto fmt = t.suggest_memory_format();
|
return fmt == at::MemoryFormat::ChannelsLast || fmt == at::MemoryFormat::ChannelsLast3d;
|
||||||
return fmt == at::MemoryFormat::ChannelsLast || fmt == at::MemoryFormat::ChannelsLast3d;
|
|
||||||
};
|
|
||||||
return is_channel_last(input) || is_channel_last(weight);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
} // namespace at::native
|
} // namespace at::native
|
||||||
|
|||||||
@ -32,6 +32,10 @@
|
|||||||
#include <ATen/native/mkldnn/Utils.h>
|
#include <ATen/native/mkldnn/Utils.h>
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
|
#ifdef USE_MPS
|
||||||
|
#include <ATen/mps/MPSDevice.h>
|
||||||
|
#endif
|
||||||
|
|
||||||
#ifndef AT_PER_OPERATOR_HEADERS
|
#ifndef AT_PER_OPERATOR_HEADERS
|
||||||
#include <ATen/Functions.h>
|
#include <ATen/Functions.h>
|
||||||
#include <ATen/NativeFunctions.h>
|
#include <ATen/NativeFunctions.h>
|
||||||
@ -406,23 +410,11 @@ struct ConvParams {
|
|||||||
// cudnn and miopen are guaranteed not to be on mobile, and T102591915 / T110194934 suggest
|
// cudnn and miopen are guaranteed not to be on mobile, and T102591915 / T110194934 suggest
|
||||||
// that maybe the compiledWithCuDNN() check sometimes segfaults (though I can't imagine how)
|
// that maybe the compiledWithCuDNN() check sometimes segfaults (though I can't imagine how)
|
||||||
#if !defined(C10_MOBILE)
|
#if !defined(C10_MOBILE)
|
||||||
if (!detail::getCUDAHooks().compiledWithCuDNN() || !input.is_cuda() || !cudnn_enabled) {
|
if (!detail::getCUDAHooks().compiledWithCuDNN()) {
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
static long cudnn_version = detail::getCUDAHooks().versionCuDNN();
|
|
||||||
// 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) {
|
|
||||||
for (int i = 2; i < weight.dim(); i++) {
|
|
||||||
if (weight.size(i) != 1) {
|
|
||||||
return false;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
}
|
|
||||||
if (needs_64bit_indexing_no_split(input, weight)) {
|
if (needs_64bit_indexing_no_split(input, weight)) {
|
||||||
|
static long cudnn_version = detail::getCUDAHooks().versionCuDNN();
|
||||||
if (!(cudnn_version >= 90300 && at::native::cudnnv8_enabled_check_debug())) {
|
if (!(cudnn_version >= 90300 && at::native::cudnnv8_enabled_check_debug())) {
|
||||||
TORCH_WARN_ONCE("cuDNN cannot be used for large non-batch-splittable convolutions"
|
TORCH_WARN_ONCE("cuDNN cannot be used for large non-batch-splittable convolutions"
|
||||||
" if the V8 API is not enabled or before cuDNN version 9.3+."
|
" if the V8 API is not enabled or before cuDNN version 9.3+."
|
||||||
@ -430,6 +422,9 @@ struct ConvParams {
|
|||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
if (!input.is_cuda() || !cudnn_enabled) {
|
||||||
|
return false;
|
||||||
|
}
|
||||||
if (input.scalar_type() == at::kBFloat16 || weight.scalar_type() == at::kBFloat16) {
|
if (input.scalar_type() == at::kBFloat16 || weight.scalar_type() == at::kBFloat16) {
|
||||||
if (!(detail::getCUDAHooks().supportsBFloat16ConvolutionWithCuDNNv8() && at::native::cudnnv8_enabled_check_debug())) {
|
if (!(detail::getCUDAHooks().supportsBFloat16ConvolutionWithCuDNNv8() && at::native::cudnnv8_enabled_check_debug())) {
|
||||||
return false;
|
return false;
|
||||||
@ -448,19 +443,16 @@ struct ConvParams {
|
|||||||
|
|
||||||
// Use cudnn for FP16 depthwise convolutions
|
// Use cudnn for FP16 depthwise convolutions
|
||||||
bool use_cudnn_depthwise(const at::Tensor& input, const at::Tensor& weight) const {
|
bool use_cudnn_depthwise(const at::Tensor& input, const at::Tensor& weight) const {
|
||||||
if (!cudnn_enabled || !detail::getCUDAHooks().compiledWithCuDNN() || !input.is_cuda()) {
|
if (!detail::getCUDAHooks().compiledWithCuDNN()) {
|
||||||
return false;
|
return false;
|
||||||
}
|
}
|
||||||
|
if (cudnn_conv_suggest_memory_format(input, weight) != at::MemoryFormat::Contiguous && use_cudnn(input, weight)) {
|
||||||
|
// always use cudnn_depthwise for channels_last format
|
||||||
|
return true;
|
||||||
|
}
|
||||||
// native kernel doesn't support 64-bit non-splittable case
|
// native kernel doesn't support 64-bit non-splittable case
|
||||||
if (!(canUse32BitIndexMath(input) && canUse32BitIndexMath(weight))) {
|
if (cudnn_enabled && !(canUse32BitIndexMath(input) && canUse32BitIndexMath(weight))) {
|
||||||
static long cudnn_version = detail::getCUDAHooks().compiledWithCuDNN() ? detail::getCUDAHooks().versionCuDNN() : -1;
|
static long cudnn_version = detail::getCUDAHooks().compiledWithCuDNN() ? detail::getCUDAHooks().versionCuDNN() : -1;
|
||||||
// TODO(eqy): remove this once cuDNN fixes 64-bit depthwise support, first broken in 9.11x
|
|
||||||
if (cudnn_conv_suggest_memory_format(input, weight) != at::MemoryFormat::Contiguous) {
|
|
||||||
if (cudnn_version < 0 || cudnn_version > 91000) {
|
|
||||||
return false;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
|
|
||||||
if (!(cudnn_version >= 90300 && at::native::cudnnv8_enabled_check_debug())) {
|
if (!(cudnn_version >= 90300 && at::native::cudnnv8_enabled_check_debug())) {
|
||||||
TORCH_WARN_ONCE("cuDNN cannot be used for large non-batch-splittable convolutions"
|
TORCH_WARN_ONCE("cuDNN cannot be used for large non-batch-splittable convolutions"
|
||||||
" if the V8 API is not enabled or before cuDNN version 9.3+."
|
" if the V8 API is not enabled or before cuDNN version 9.3+."
|
||||||
@ -470,10 +462,6 @@ struct ConvParams {
|
|||||||
return true;
|
return true;
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
if (cudnn_conv_suggest_memory_format(input, weight) != at::MemoryFormat::Contiguous) {
|
|
||||||
// always use cudnn_depthwise for channels_last format
|
|
||||||
return true;
|
|
||||||
}
|
|
||||||
if (detail::getCUDAHooks().supportsDepthwiseConvolutionWithCuDNN()) {
|
if (detail::getCUDAHooks().supportsDepthwiseConvolutionWithCuDNN()) {
|
||||||
bool kernel_cond = (use_cudnn(input, weight) &&
|
bool kernel_cond = (use_cudnn(input, weight) &&
|
||||||
input.scalar_type() == kHalf && // only for FP16
|
input.scalar_type() == kHalf && // only for FP16
|
||||||
@ -1441,8 +1429,12 @@ static inline at::MemoryFormat determine_backend_memory_format(
|
|||||||
}
|
}
|
||||||
break;
|
break;
|
||||||
case ConvBackend::Mps:
|
case ConvBackend::Mps:
|
||||||
case ConvBackend::MpsTranspose:
|
|
||||||
if (mps_conv_use_channels_last(input, weight)) {
|
if (mps_conv_use_channels_last(input, weight)) {
|
||||||
|
#ifdef USE_MPS
|
||||||
|
if (!mps::is_macos_13_or_newer(mps::MacOSVersion::MACOS_VER_15_0_PLUS)) {
|
||||||
|
break;
|
||||||
|
}
|
||||||
|
#endif
|
||||||
backend_memory_format = (k == 5) ? MemoryFormat::ChannelsLast3d : MemoryFormat::ChannelsLast;
|
backend_memory_format = (k == 5) ? MemoryFormat::ChannelsLast3d : MemoryFormat::ChannelsLast;
|
||||||
}
|
}
|
||||||
break;
|
break;
|
||||||
|
|||||||
@ -9,7 +9,6 @@
|
|||||||
#include <ATen/native/TransposeType.h>
|
#include <ATen/native/TransposeType.h>
|
||||||
#include <ATen/native/Unfold3d.h>
|
#include <ATen/native/Unfold3d.h>
|
||||||
#include <c10/util/irange.h>
|
#include <c10/util/irange.h>
|
||||||
#include <c10/util/safe_numerics.h>
|
|
||||||
|
|
||||||
#ifndef AT_PER_OPERATOR_HEADERS
|
#ifndef AT_PER_OPERATOR_HEADERS
|
||||||
#include <ATen/Functions.h>
|
#include <ATen/Functions.h>
|
||||||
@ -175,23 +174,6 @@ static inline void slow_conv3d_shape_check(
|
|||||||
const int64_t input_height = input.size(dim_height);
|
const int64_t input_height = input.size(dim_height);
|
||||||
const int64_t input_width = input.size(dim_width);
|
const int64_t input_width = input.size(dim_width);
|
||||||
|
|
||||||
constexpr int64_t MAX_SAFE_PAD = (1LL << 61);
|
|
||||||
|
|
||||||
TORCH_CHECK_VALUE(
|
|
||||||
pad_height <= MAX_SAFE_PAD,
|
|
||||||
"Padding height too large: pad_height=",
|
|
||||||
pad_height);
|
|
||||||
|
|
||||||
TORCH_CHECK_VALUE(
|
|
||||||
pad_width <= MAX_SAFE_PAD,
|
|
||||||
"Padding width too large: pad_width=",
|
|
||||||
pad_width);
|
|
||||||
|
|
||||||
TORCH_CHECK_VALUE(
|
|
||||||
pad_depth <= MAX_SAFE_PAD,
|
|
||||||
"Padding depth too large: pad_depth=",
|
|
||||||
pad_depth);
|
|
||||||
|
|
||||||
const int64_t exact_input_depth = input_depth + 2 * pad_depth;
|
const int64_t exact_input_depth = input_depth + 2 * pad_depth;
|
||||||
const int64_t exact_input_height = input_height + 2 * pad_height;
|
const int64_t exact_input_height = input_height + 2 * pad_height;
|
||||||
const int64_t exact_input_width = input_width + 2 * pad_width;
|
const int64_t exact_input_width = input_width + 2 * pad_width;
|
||||||
@ -239,14 +221,6 @@ static inline void slow_conv3d_shape_check(
|
|||||||
output_width,
|
output_width,
|
||||||
"). Output size is too small");
|
"). Output size is too small");
|
||||||
|
|
||||||
uint64_t kernel_product;
|
|
||||||
TORCH_CHECK(
|
|
||||||
!c10::mul_overflows(kernel_height, kernel_width, &kernel_product),
|
|
||||||
"Kernel height x width product is too large: kernel_height=",
|
|
||||||
kernel_height,
|
|
||||||
", kernel_width=",
|
|
||||||
kernel_width);
|
|
||||||
|
|
||||||
if (weight.defined()) {
|
if (weight.defined()) {
|
||||||
int64_t n_input_plane = weight.size(1);
|
int64_t n_input_plane = weight.size(1);
|
||||||
if (weight.dim() == 2) {
|
if (weight.dim() == 2) {
|
||||||
|
|||||||
@ -97,38 +97,43 @@ Tensor& fill_diagonal_(Tensor& self, const Scalar& fill_value, bool wrap) {
|
|||||||
int64_t nDims = self.dim();
|
int64_t nDims = self.dim();
|
||||||
TORCH_CHECK(nDims >= 2, "dimensions must larger than 1");
|
TORCH_CHECK(nDims >= 2, "dimensions must larger than 1");
|
||||||
|
|
||||||
auto height = self.sym_size(0);
|
int64_t height = self.size(0);
|
||||||
auto width = self.sym_size(1);
|
int64_t width = self.size(1);
|
||||||
|
|
||||||
if (nDims > 2) {
|
if (nDims > 2) {
|
||||||
|
int64_t dim1 = height;
|
||||||
for (const auto i : c10::irange(1, nDims)) {
|
for (const auto i : c10::irange(1, nDims)) {
|
||||||
if (self.sym_size(i) != height) {
|
if (self.size(i) != dim1) {
|
||||||
TORCH_CHECK(false, "all dimensions of input must be of equal length");
|
TORCH_CHECK(false, "all dimensions of input must be of equal length");
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
auto storage_offset = self.sym_storage_offset();
|
int64_t storage_offset = self.storage_offset();
|
||||||
auto size = std::min(height, width);
|
std::vector<int64_t> sizes;
|
||||||
|
std::vector<int64_t> strides;
|
||||||
|
int64_t size = std::min(height, width);
|
||||||
|
|
||||||
int64_t stride = 0;
|
int64_t stride = 0;
|
||||||
for (const auto i : c10::irange(nDims)) {
|
for (const auto i : c10::irange(nDims)) {
|
||||||
stride += self.stride(i);
|
stride += self.stride(i);
|
||||||
}
|
}
|
||||||
std::vector<SymInt> strides{stride};
|
strides.push_back(stride);
|
||||||
std::vector<SymInt> sizes{size};
|
sizes.push_back(size);
|
||||||
|
|
||||||
auto main_diag = self.as_strided_symint(sizes, strides, storage_offset);
|
auto main_diag = self.as_strided(sizes, strides, storage_offset);
|
||||||
main_diag.fill_(fill_value);
|
main_diag.fill_(fill_value);
|
||||||
|
|
||||||
if (wrap && nDims == 2 && height > width + 1) {
|
if (wrap && nDims == 2 && height > width + 1) {
|
||||||
auto step = width + 1;
|
std::vector<int64_t> wrap_sizes;
|
||||||
auto wrap_size = ((self.numel() + step - 1) / step) - size;
|
|
||||||
std::vector<SymInt> wrap_sizes{wrap_size};
|
|
||||||
|
|
||||||
auto offset = self.stride(0) * (width + 1);
|
int64_t step = width + 1;
|
||||||
|
int64_t wrap_size = ((self.numel() + step - 1) / step) - size;
|
||||||
|
wrap_sizes.push_back(wrap_size);
|
||||||
|
|
||||||
auto wrap_diag = self.as_strided_symint(wrap_sizes, strides, storage_offset + offset);
|
int64_t offset = self.stride(0) * (width + 1);
|
||||||
|
|
||||||
|
auto wrap_diag = self.as_strided(wrap_sizes, strides, storage_offset + offset);
|
||||||
wrap_diag.fill_(fill_value);
|
wrap_diag.fill_(fill_value);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|||||||
@ -23,7 +23,6 @@
|
|||||||
#include <ATen/ops/linspace.h>
|
#include <ATen/ops/linspace.h>
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
#include <cmath>
|
|
||||||
#include <numeric>
|
#include <numeric>
|
||||||
#include <tuple>
|
#include <tuple>
|
||||||
#include <vector>
|
#include <vector>
|
||||||
@ -203,46 +202,6 @@ select_outer_bin_edges(const Tensor& input, std::optional<c10::ArrayRef<double>>
|
|||||||
return std::make_pair(leftmost_edges, rightmost_edges);
|
return std::make_pair(leftmost_edges, rightmost_edges);
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|
||||||
/* Bin edges correction based on the precision representation.
|
|
||||||
* To maintain the backward compatibility we take max(std::nextafter<>, +1)
|
|
||||||
* and min(std::nextafter<>, -1) for scalar types. For other types +/- 1 as usual.
|
|
||||||
*/
|
|
||||||
void bins_edges_correction(const ScalarType& t, double &leftmost_edge, double &rightmost_edge)
|
|
||||||
{
|
|
||||||
#define UPDATE_WITH_LIMIT(real_type, scalartype) \
|
|
||||||
case ScalarType::scalartype: \
|
|
||||||
leftmost_edge = std::min( \
|
|
||||||
static_cast<double>( \
|
|
||||||
std::nexttoward( \
|
|
||||||
static_cast<real_type>(leftmost_edge), \
|
|
||||||
std::numeric_limits<real_type>::lowest() \
|
|
||||||
) \
|
|
||||||
), \
|
|
||||||
leftmost_edge - 1. \
|
|
||||||
); \
|
|
||||||
rightmost_edge = std::max( \
|
|
||||||
static_cast<double>( \
|
|
||||||
std::nexttoward( \
|
|
||||||
static_cast<real_type>(rightmost_edge), \
|
|
||||||
std::numeric_limits<real_type>::max() \
|
|
||||||
) \
|
|
||||||
), \
|
|
||||||
rightmost_edge + 1. \
|
|
||||||
); \
|
|
||||||
break;
|
|
||||||
|
|
||||||
switch (t) {
|
|
||||||
UPDATE_WITH_LIMIT(double, Double)
|
|
||||||
UPDATE_WITH_LIMIT(float, Float)
|
|
||||||
default:
|
|
||||||
// Fallback to the default behavior for other types
|
|
||||||
leftmost_edge -= 1;
|
|
||||||
rightmost_edge += 1;
|
|
||||||
}
|
|
||||||
#undef UPDATE_WITH_LIMIT
|
|
||||||
}
|
|
||||||
|
|
||||||
/* histc's version of the logic for outermost bin edges.
|
/* histc's version of the logic for outermost bin edges.
|
||||||
*/
|
*/
|
||||||
std::pair<double, double> histc_select_outer_bin_edges(const Tensor& input,
|
std::pair<double, double> histc_select_outer_bin_edges(const Tensor& input,
|
||||||
@ -257,7 +216,8 @@ std::pair<double, double> histc_select_outer_bin_edges(const Tensor& input,
|
|||||||
}
|
}
|
||||||
|
|
||||||
if (leftmost_edge == rightmost_edge) {
|
if (leftmost_edge == rightmost_edge) {
|
||||||
bins_edges_correction(input.dtype().toScalarType(), leftmost_edge, rightmost_edge);
|
leftmost_edge -= 1;
|
||||||
|
rightmost_edge += 1;
|
||||||
}
|
}
|
||||||
|
|
||||||
TORCH_CHECK(!(std::isinf(leftmost_edge) || std::isinf(rightmost_edge) ||
|
TORCH_CHECK(!(std::isinf(leftmost_edge) || std::isinf(rightmost_edge) ||
|
||||||
|
|||||||
@ -23,6 +23,8 @@ Tensor& max_unpooling2d_forward_out_cpu(
|
|||||||
// Nondeterministic with duplicate indices
|
// Nondeterministic with duplicate indices
|
||||||
at::globalContext().alertNotDeterministic("max_unpooling2d_forward_out");
|
at::globalContext().alertNotDeterministic("max_unpooling2d_forward_out");
|
||||||
|
|
||||||
|
auto oheight = output_size[0];
|
||||||
|
auto owidth = output_size[1];
|
||||||
TORCH_CHECK(
|
TORCH_CHECK(
|
||||||
indices_.scalar_type() == at::ScalarType::Long,
|
indices_.scalar_type() == at::ScalarType::Long,
|
||||||
"elements in indices should be type int64 but got: ", indices_.scalar_type());
|
"elements in indices should be type int64 but got: ", indices_.scalar_type());
|
||||||
@ -43,9 +45,6 @@ Tensor& max_unpooling2d_forward_out_cpu(
|
|||||||
self_.sizes(), " with dimension ", i , " being empty.");
|
self_.sizes(), " with dimension ", i , " being empty.");
|
||||||
}
|
}
|
||||||
|
|
||||||
auto oheight = output_size[0];
|
|
||||||
auto owidth = output_size[1];
|
|
||||||
|
|
||||||
auto memory_format = self_.suggest_memory_format();
|
auto memory_format = self_.suggest_memory_format();
|
||||||
auto self = self_.contiguous(memory_format);
|
auto self = self_.contiguous(memory_format);
|
||||||
auto indices = indices_.contiguous(memory_format);
|
auto indices = indices_.contiguous(memory_format);
|
||||||
|
|||||||
@ -1,5 +1,3 @@
|
|||||||
#include <ATen/core/ATen_fwd.h>
|
|
||||||
#include <c10/core/ScalarType.h>
|
|
||||||
#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
|
#define TORCH_ASSERT_ONLY_METHOD_OPERATORS
|
||||||
#include <ATen/AccumulateType.h>
|
#include <ATen/AccumulateType.h>
|
||||||
#include <ATen/Dispatch.h>
|
#include <ATen/Dispatch.h>
|
||||||
@ -1880,18 +1878,19 @@ Tensor repeat(const Tensor& self, IntArrayRef repeats) {
|
|||||||
|
|
||||||
Tensor xtensor = self.expand(padded_size);
|
Tensor xtensor = self.expand(padded_size);
|
||||||
|
|
||||||
Tensor urtensor;
|
Tensor result;
|
||||||
if (self.is_quantized()) {
|
if (self.is_quantized()) {
|
||||||
urtensor = at::empty_quantized(target_size, self);
|
result = at::empty_quantized(target_size, self);
|
||||||
} else {
|
} else {
|
||||||
urtensor = at::empty(target_size, self.options());
|
result = at::empty(target_size, self.options());
|
||||||
}
|
}
|
||||||
|
|
||||||
// return an empty tensor if one of the repeat dimensions is zero
|
// return an empty tensor if one of the repeat dimensions is zero
|
||||||
if (zero_tensor) {
|
if (zero_tensor) {
|
||||||
return urtensor;
|
return result;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
Tensor urtensor = at::alias(result);
|
||||||
for (const auto i : c10::irange(xtensor.dim())) {
|
for (const auto i : c10::irange(xtensor.dim())) {
|
||||||
// can't unfold with step 0, so make sure step is at least 1
|
// can't unfold with step 0, so make sure step is at least 1
|
||||||
// (it doesn't matter what it is in that case, because the size is 0).
|
// (it doesn't matter what it is in that case, because the size is 0).
|
||||||
@ -1901,22 +1900,7 @@ Tensor repeat(const Tensor& self, IntArrayRef repeats) {
|
|||||||
|
|
||||||
urtensor.copy_(xtensor.expand_as(urtensor));
|
urtensor.copy_(xtensor.expand_as(urtensor));
|
||||||
|
|
||||||
// Combine the dimensions to produce the target_size.
|
return result;
|
||||||
// xtensor dims: [a0, ..., ad-1]
|
|
||||||
// urtensor dims: [a0, ..., ad-1, b0, ..., bd-1]
|
|
||||||
// b dims are produced by unfold.
|
|
||||||
// Transform urtensor to [a0 * b0, ..., ad-1 * bd-1]
|
|
||||||
const int64_t n_dims = xtensor.dim();
|
|
||||||
auto range_a = at::arange(xtensor.dim(), at::TensorOptions(at::kLong));
|
|
||||||
auto range_b = range_a + n_dims;
|
|
||||||
auto stacked = stack({std::move(range_a), std::move(range_b)}, 1).flatten();
|
|
||||||
auto permutation = IntArrayRef(stacked.data_ptr<int64_t>(), n_dims * 2);
|
|
||||||
// Permute from [a0, ..., ad-1, b0, ..., bd-1] to [a0, b0, ..., ad-1, bd-1]
|
|
||||||
urtensor = urtensor.permute(permutation);
|
|
||||||
// Reshape from [a0, b0, ..., ad-1, bd-1] to [a0 * b0, ..., ad-1 * bd-1]
|
|
||||||
urtensor = urtensor.reshape(target_size);
|
|
||||||
|
|
||||||
return urtensor;
|
|
||||||
}
|
}
|
||||||
|
|
||||||
Tensor tile_symint(const Tensor& self, SymIntArrayRef reps) {
|
Tensor tile_symint(const Tensor& self, SymIntArrayRef reps) {
|
||||||
|
|||||||
@ -999,41 +999,12 @@ void gpu_kernel_impl(TensorIteratorBase& iter, const func_t& f) {
|
|||||||
dtypes[i] = iter.dtype(i);
|
dtypes[i] = iter.dtype(i);
|
||||||
}
|
}
|
||||||
auto offset_calc = ::make_offset_calculator<traits::arity + 1>(iter);
|
auto offset_calc = ::make_offset_calculator<traits::arity + 1>(iter);
|
||||||
#ifdef USE_ROCM
|
|
||||||
constexpr int grp_sz = 128;
|
|
||||||
launch_legacy_kernel_manual_unroll<grp_sz, 4>(numel, [=] GPU_LAMBDA(int idx, bool unrl) {
|
|
||||||
if (unrl) {
|
|
||||||
auto offsets0 = offset_calc.get(idx);
|
|
||||||
auto offsets1 = offset_calc.get(idx + grp_sz);
|
|
||||||
auto offsets2 = offset_calc.get(idx + grp_sz * 2);
|
|
||||||
auto offsets3 = offset_calc.get(idx + grp_sz * 3);
|
|
||||||
void* out0 = data[0] + offsets0[0];
|
|
||||||
void* out1 = data[0] + offsets1[0];
|
|
||||||
void* out2 = data[0] + offsets2[0];
|
|
||||||
void* out3 = data[0] + offsets3[0];
|
|
||||||
arg0_t result0 = invoke(f, &data[1], &offsets0[1], &dtypes[1], 1);
|
|
||||||
arg0_t result1 = invoke(f, &data[1], &offsets1[1], &dtypes[1], 1);
|
|
||||||
arg0_t result2 = invoke(f, &data[1], &offsets2[1], &dtypes[1], 1);
|
|
||||||
arg0_t result3 = invoke(f, &data[1], &offsets3[1], &dtypes[1], 1);
|
|
||||||
c10::cast_and_store<arg0_t>(dtypes[0], out0, result0);
|
|
||||||
c10::cast_and_store<arg0_t>(dtypes[0], out1, result1);
|
|
||||||
c10::cast_and_store<arg0_t>(dtypes[0], out2, result2);
|
|
||||||
c10::cast_and_store<arg0_t>(dtypes[0], out3, result3);
|
|
||||||
} else {
|
|
||||||
auto offsets = offset_calc.get(idx);
|
|
||||||
void* out = data[0] + offsets[0];
|
|
||||||
arg0_t result = invoke(f, &data[1], &offsets[1], &dtypes[1], 1);
|
|
||||||
c10::cast_and_store<arg0_t>(dtypes[0], out, result);
|
|
||||||
}
|
|
||||||
});
|
|
||||||
#else
|
|
||||||
launch_legacy_kernel<128, 4>(numel, [=] GPU_LAMBDA(int idx) {
|
launch_legacy_kernel<128, 4>(numel, [=] GPU_LAMBDA(int idx) {
|
||||||
auto offsets = offset_calc.get(idx);
|
auto offsets = offset_calc.get(idx);
|
||||||
void* out = data[0] + offsets[0];
|
void* out = data[0] + offsets[0];
|
||||||
arg0_t result = invoke(f, &data[1], &offsets[1], &dtypes[1], 1);
|
arg0_t result = invoke(f, &data[1], &offsets[1], &dtypes[1], 1);
|
||||||
c10::cast_and_store<arg0_t>(dtypes[0], out, result);
|
c10::cast_and_store<arg0_t>(dtypes[0], out, result);
|
||||||
});
|
});
|
||||||
#endif
|
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|||||||
@ -42,19 +42,6 @@ void bfloat16_copy_kernel_cuda(TensorIteratorBase &iter) {
|
|||||||
});
|
});
|
||||||
}
|
}
|
||||||
|
|
||||||
#ifdef USE_ROCM
|
|
||||||
void bfloat16tofloat32_copy_kernel_cuda(TensorIteratorBase &iter) {
|
|
||||||
gpu_kernel_nocast(iter, [] GPU_LAMBDA(at::BFloat16 value) {
|
|
||||||
return static_cast<float>(value);
|
|
||||||
});
|
|
||||||
}
|
|
||||||
void float16tofloat32_copy_kernel_cuda(TensorIteratorBase &iter) {
|
|
||||||
gpu_kernel_nocast(iter, [] GPU_LAMBDA(at::Half value) {
|
|
||||||
return static_cast<float>(value);
|
|
||||||
});
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
|
|
||||||
void float8_copy_kernel_cuda(TensorIteratorBase &iter) {
|
void float8_copy_kernel_cuda(TensorIteratorBase &iter) {
|
||||||
ScalarType dtype = iter.dtype(0);
|
ScalarType dtype = iter.dtype(0);
|
||||||
ScalarType other_dtype = iter.dtype(1);
|
ScalarType other_dtype = iter.dtype(1);
|
||||||
@ -200,17 +187,7 @@ void direct_copy_kernel_cuda(TensorIteratorBase &iter) {
|
|||||||
} else {
|
} else {
|
||||||
float16_copy_kernel_cuda(iter);
|
float16_copy_kernel_cuda(iter);
|
||||||
}
|
}
|
||||||
}
|
} else if (isBitsType(dtype)) {
|
||||||
#ifdef USE_ROCM
|
|
||||||
else if ((iter.dtype(1) == kBFloat16 || iter.dtype(1) == kHalf) && dtype == kFloat) {
|
|
||||||
if (iter.dtype(1) == kBFloat16) {
|
|
||||||
bfloat16tofloat32_copy_kernel_cuda(iter);
|
|
||||||
} else {
|
|
||||||
float16tofloat32_copy_kernel_cuda(iter);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
#endif
|
|
||||||
else if (isBitsType(dtype)) {
|
|
||||||
TORCH_CHECK(dtype == iter.dtype(1), "copy_() does not support casting "
|
TORCH_CHECK(dtype == iter.dtype(1), "copy_() does not support casting "
|
||||||
"bits types to different bits types. Source dtype is ", iter.dtype(1), "target dtype is ", dtype);
|
"bits types to different bits types. Source dtype is ", iter.dtype(1), "target dtype is ", dtype);
|
||||||
AT_DISPATCH_BIT_TYPES(dtype, "copy_", [&] {
|
AT_DISPATCH_BIT_TYPES(dtype, "copy_", [&] {
|
||||||
|
|||||||
@ -125,6 +125,8 @@ Tensor& max_unpooling2d_forward_out_cuda(const Tensor& self_,
|
|||||||
TORCH_CHECK(
|
TORCH_CHECK(
|
||||||
indices_.scalar_type() == at::ScalarType::Long,
|
indices_.scalar_type() == at::ScalarType::Long,
|
||||||
"elements in indices should be type int64 but got: ", indices_.scalar_type());
|
"elements in indices should be type int64 but got: ", indices_.scalar_type());
|
||||||
|
auto oheight = output_size[0];
|
||||||
|
auto owidth = output_size[1];
|
||||||
|
|
||||||
TensorArg output_arg{output, "output", 1}, self_arg{self_, "self_", 2},
|
TensorArg output_arg{output, "output", 1}, self_arg{self_, "self_", 2},
|
||||||
indices_arg{indices_, "indices_", 3};
|
indices_arg{indices_, "indices_", 3};
|
||||||
@ -147,9 +149,6 @@ Tensor& max_unpooling2d_forward_out_cuda(const Tensor& self_,
|
|||||||
output_size.size() == 2,
|
output_size.size() == 2,
|
||||||
"There should be exactly two elements (height, width) in output_size, but got ", output_size.size(), " elements.");
|
"There should be exactly two elements (height, width) in output_size, but got ", output_size.size(), " elements.");
|
||||||
|
|
||||||
auto oheight = output_size[0];
|
|
||||||
auto owidth = output_size[1];
|
|
||||||
|
|
||||||
int64_t dimw = 2;
|
int64_t dimw = 2;
|
||||||
int64_t dimh = 1;
|
int64_t dimh = 1;
|
||||||
int64_t numBatch = 1;
|
int64_t numBatch = 1;
|
||||||
@ -218,6 +217,9 @@ static void max_unpooling3d_shape_check(
|
|||||||
IntArrayRef stride,
|
IntArrayRef stride,
|
||||||
IntArrayRef padding,
|
IntArrayRef padding,
|
||||||
const char *fn_name) {
|
const char *fn_name) {
|
||||||
|
int64_t oT = output_size[0];
|
||||||
|
int64_t oH = output_size[1];
|
||||||
|
int64_t oW = output_size[2];
|
||||||
TORCH_CHECK(
|
TORCH_CHECK(
|
||||||
indices.scalar_type() == at::ScalarType::Long,
|
indices.scalar_type() == at::ScalarType::Long,
|
||||||
"elements in indices should be type int64 but got: ", indices.scalar_type());
|
"elements in indices should be type int64 but got: ", indices.scalar_type());
|
||||||
@ -248,10 +250,6 @@ static void max_unpooling3d_shape_check(
|
|||||||
"strides should be greater than zero, but got stride: ",
|
"strides should be greater than zero, but got stride: ",
|
||||||
stride);
|
stride);
|
||||||
|
|
||||||
int64_t oT = output_size[0];
|
|
||||||
int64_t oH = output_size[1];
|
|
||||||
int64_t oW = output_size[2];
|
|
||||||
|
|
||||||
int dimw = 3;
|
int dimw = 3;
|
||||||
int dimh = 2;
|
int dimh = 2;
|
||||||
int dimt = 1;
|
int dimt = 1;
|
||||||
@ -404,6 +402,8 @@ at::Tensor& max_unpooling2d_backward_out_cuda(const Tensor& grad_output_,
|
|||||||
const Tensor& indices_,
|
const Tensor& indices_,
|
||||||
IntArrayRef output_size,
|
IntArrayRef output_size,
|
||||||
Tensor& grad_input) {
|
Tensor& grad_input) {
|
||||||
|
int64_t oheight = output_size[0];
|
||||||
|
int64_t owidth = output_size[1];
|
||||||
TORCH_CHECK(grad_input.is_contiguous(), "grad_input must be contiguous");
|
TORCH_CHECK(grad_input.is_contiguous(), "grad_input must be contiguous");
|
||||||
TORCH_CHECK(
|
TORCH_CHECK(
|
||||||
indices_.scalar_type() == at::ScalarType::Long,
|
indices_.scalar_type() == at::ScalarType::Long,
|
||||||
@ -426,9 +426,6 @@ at::Tensor& max_unpooling2d_backward_out_cuda(const Tensor& grad_output_,
|
|||||||
|
|
||||||
TORCH_CHECK(output_size.size() == 2, "output_size must have two elements, got size: ", output_size.size());
|
TORCH_CHECK(output_size.size() == 2, "output_size must have two elements, got size: ", output_size.size());
|
||||||
|
|
||||||
int64_t oheight = output_size[0];
|
|
||||||
int64_t owidth = output_size[1];
|
|
||||||
|
|
||||||
int64_t nInputCols, nInputRows, nInputPlane;
|
int64_t nInputCols, nInputRows, nInputPlane;
|
||||||
|
|
||||||
int dimw = 2;
|
int dimw = 2;
|
||||||
@ -508,14 +505,13 @@ at::Tensor& max_unpooling3d_backward_out_cuda(const Tensor& grad_output_,
|
|||||||
IntArrayRef padding,
|
IntArrayRef padding,
|
||||||
Tensor& grad_input) {
|
Tensor& grad_input) {
|
||||||
TORCH_CHECK(grad_input.is_contiguous(), "grad_input must be contiguous");
|
TORCH_CHECK(grad_input.is_contiguous(), "grad_input must be contiguous");
|
||||||
|
|
||||||
max_unpooling3d_shape_check(
|
|
||||||
self_, grad_output_, indices_, output_size, stride, padding, "max_unpooling3d_backward_out_cuda()");
|
|
||||||
|
|
||||||
int64_t oT = output_size[0];
|
int64_t oT = output_size[0];
|
||||||
int64_t oH = output_size[1];
|
int64_t oH = output_size[1];
|
||||||
int64_t oW = output_size[2];
|
int64_t oW = output_size[2];
|
||||||
|
|
||||||
|
max_unpooling3d_shape_check(
|
||||||
|
self_, grad_output_, indices_, output_size, stride, padding, "max_unpooling3d_backward_out_cuda()");
|
||||||
|
|
||||||
int batchSize = 0;
|
int batchSize = 0;
|
||||||
int inputSlices = 0;
|
int inputSlices = 0;
|
||||||
int inputTime = 0;
|
int inputTime = 0;
|
||||||
|
|||||||
@ -300,6 +300,8 @@ void nonzero_static_cuda_out_impl(
|
|||||||
int64_t size,
|
int64_t size,
|
||||||
int64_t fill_value,
|
int64_t fill_value,
|
||||||
Tensor& out) {
|
Tensor& out) {
|
||||||
|
#if defined(CUDA_VERSION) || defined(USE_ROCM)
|
||||||
|
|
||||||
Tensor self_contiguous_ = self.contiguous();
|
Tensor self_contiguous_ = self.contiguous();
|
||||||
// see comment in nonzero_cuda_out_impl on reqs for out
|
// see comment in nonzero_cuda_out_impl on reqs for out
|
||||||
bool out_correct_size =
|
bool out_correct_size =
|
||||||
@ -375,6 +377,9 @@ void nonzero_static_cuda_out_impl(
|
|||||||
if (need_to_copy) {
|
if (need_to_copy) {
|
||||||
out.copy_(out_temp);
|
out.copy_(out_temp);
|
||||||
}
|
}
|
||||||
|
#else
|
||||||
|
TORCH_CHECK(false, "Nonzero_static is not supported for cuda <= 11.4");
|
||||||
|
#endif
|
||||||
}
|
}
|
||||||
|
|
||||||
Tensor& nonzero_out_cuda(const Tensor& self, Tensor& out) {
|
Tensor& nonzero_out_cuda(const Tensor& self, Tensor& out) {
|
||||||
|
|||||||
@ -221,9 +221,22 @@ static const Tensor& _exec_fft(Tensor& out, const Tensor& self, IntArrayRef out_
|
|||||||
std::optional<CuFFTConfig> uncached_plan;
|
std::optional<CuFFTConfig> uncached_plan;
|
||||||
const CuFFTConfig * config = nullptr;
|
const CuFFTConfig * config = nullptr;
|
||||||
|
|
||||||
|
// Workaround for gh-63152, gh-58724
|
||||||
|
// Bluestein plans in CUDA 11.1 (cufft 10.3) cannot be re-used
|
||||||
// Bluestein's algorithm is only used when a size has large prime factors,
|
// Bluestein's algorithm is only used when a size has large prime factors,
|
||||||
// sizes with only small prime factors can still be cached
|
// sizes with only small prime factors can still be cached
|
||||||
if (plan_cache.max_size() > 0) {
|
bool use_caching = true;
|
||||||
|
#ifdef CUFFT_VERSION
|
||||||
|
if constexpr (10300 <= CUFFT_VERSION && CUFFT_VERSION < 10400) {
|
||||||
|
// Only cache plans for transforms with small prime factors
|
||||||
|
use_caching = std::none_of(
|
||||||
|
signal_size.begin() + 1, signal_size.end(), [](int64_t dim_size) {
|
||||||
|
return has_large_prime_factor(dim_size);
|
||||||
|
});
|
||||||
|
}
|
||||||
|
#endif
|
||||||
|
|
||||||
|
if (use_caching && plan_cache.max_size() > 0) {
|
||||||
guard.lock();
|
guard.lock();
|
||||||
if (plan_cache.max_size() > 0) { // check again after acquiring the lock
|
if (plan_cache.max_size() > 0) { // check again after acquiring the lock
|
||||||
config = &plan_cache.lookup(Params);
|
config = &plan_cache.lookup(Params);
|
||||||
|
|||||||
@ -1238,7 +1238,7 @@ Tensor _cholesky_solve_helper_cuda_magma(const Tensor& self, const Tensor& A, bo
|
|||||||
// Todo: cusolverDn<T>potrsBatched only supports nrhs == 1 and does not have good performance.
|
// Todo: cusolverDn<T>potrsBatched only supports nrhs == 1 and does not have good performance.
|
||||||
// Batched cholesky_solve is dispatched to magma.
|
// Batched cholesky_solve is dispatched to magma.
|
||||||
Tensor _cholesky_solve_helper_cuda(const Tensor& self, const Tensor& A, bool upper) {
|
Tensor _cholesky_solve_helper_cuda(const Tensor& self, const Tensor& A, bool upper) {
|
||||||
#if defined(USE_LINALG_SOLVER)
|
#if defined(USE_LINALG_SOLVER) && !defined(USE_ROCM)
|
||||||
auto preferred_backend = at::globalContext().linalgPreferredBackend();
|
auto preferred_backend = at::globalContext().linalgPreferredBackend();
|
||||||
switch (preferred_backend) {
|
switch (preferred_backend) {
|
||||||
case at::LinalgBackend::Cusolver:
|
case at::LinalgBackend::Cusolver:
|
||||||
@ -1352,7 +1352,7 @@ void cholesky_helper_magma(const Tensor& input, bool upper, const Tensor& info)
|
|||||||
}
|
}
|
||||||
|
|
||||||
static void cholesky_kernel(const Tensor& input, const Tensor& info, bool upper) {
|
static void cholesky_kernel(const Tensor& input, const Tensor& info, bool upper) {
|
||||||
#if defined(USE_LINALG_SOLVER)
|
#if defined(USE_LINALG_SOLVER) && !defined(USE_ROCM)
|
||||||
auto preferred_backend = at::globalContext().linalgPreferredBackend();
|
auto preferred_backend = at::globalContext().linalgPreferredBackend();
|
||||||
switch (preferred_backend) {
|
switch (preferred_backend) {
|
||||||
case at::LinalgBackend::Cusolver:
|
case at::LinalgBackend::Cusolver:
|
||||||
@ -2709,7 +2709,7 @@ void linalg_lstsq_gels(const Tensor& A, const Tensor& B, const Tensor& /*infos*/
|
|||||||
}
|
}
|
||||||
|
|
||||||
void gels_looped(const Tensor& a, Tensor& b, Tensor& infos) {
|
void gels_looped(const Tensor& a, Tensor& b, Tensor& infos) {
|
||||||
#if defined(USE_LINALG_SOLVER)
|
#if defined(USE_LINALG_SOLVER) && !defined(USE_ROCM)
|
||||||
auto preferred_backend = at::globalContext().linalgPreferredBackend();
|
auto preferred_backend = at::globalContext().linalgPreferredBackend();
|
||||||
switch (preferred_backend) {
|
switch (preferred_backend) {
|
||||||
case at::LinalgBackend::Magma:
|
case at::LinalgBackend::Magma:
|
||||||
@ -2733,7 +2733,7 @@ void lstsq_kernel(const Tensor& a, Tensor& b, Tensor& /*rank*/, Tensor& /*singul
|
|||||||
// first handle the underdetermined case (m < n)
|
// first handle the underdetermined case (m < n)
|
||||||
// this case is not supported by MAGMA or cuBLAS
|
// this case is not supported by MAGMA or cuBLAS
|
||||||
if (m < n) {
|
if (m < n) {
|
||||||
#if defined(USE_LINALG_SOLVER)
|
#if defined(USE_LINALG_SOLVER) && !defined(USE_ROCM)
|
||||||
linalg_lstsq_gels(a, b, infos);
|
linalg_lstsq_gels(a, b, infos);
|
||||||
#else
|
#else
|
||||||
TORCH_CHECK(
|
TORCH_CHECK(
|
||||||
|
|||||||
@ -14,7 +14,7 @@ struct EmbeddingBagParams {
|
|||||||
::c10::metal::array<idx_type_t, 2> output_strides;
|
::c10::metal::array<idx_type_t, 2> output_strides;
|
||||||
::c10::metal::array<idx_type_t, 2> max_indices_strides;
|
::c10::metal::array<idx_type_t, 2> max_indices_strides;
|
||||||
|
|
||||||
idx_type_t per_sample_weights_stride;
|
idx_type_t per_sample_weights_strides;
|
||||||
|
|
||||||
idx_type_t num_indices;
|
idx_type_t num_indices;
|
||||||
idx_type_t num_bags;
|
idx_type_t num_bags;
|
||||||
|
|||||||
@ -23,72 +23,54 @@ struct ReductionOpInit<EmbeddingBagMode::MAX, T> {
|
|||||||
template <EmbeddingBagMode M, typename T>
|
template <EmbeddingBagMode M, typename T>
|
||||||
struct ReductionOp {
|
struct ReductionOp {
|
||||||
inline opmath_t<T> operator()(
|
inline opmath_t<T> operator()(
|
||||||
opmath_t<T> weight_val,
|
T weight_val,
|
||||||
opmath_t<T> out_val,
|
opmath_t<T> out_val,
|
||||||
bool is_first) {
|
uint32_t per_sample_weights_index,
|
||||||
return weight_val + out_val;
|
constant T* per_sample_weights,
|
||||||
|
uint32_t per_sample_weights_strides);
|
||||||
|
};
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
struct ReductionOp<EmbeddingBagMode::SUM, T> {
|
||||||
|
inline opmath_t<T> operator()(
|
||||||
|
T weight_val,
|
||||||
|
opmath_t<T> out_val,
|
||||||
|
uint32_t per_sample_weights_index,
|
||||||
|
constant T* per_sample_weights,
|
||||||
|
uint32_t per_sample_weights_strides) {
|
||||||
|
if (per_sample_weights_strides) {
|
||||||
|
T per_sample_weight = per_sample_weights
|
||||||
|
[per_sample_weights_strides * per_sample_weights_index];
|
||||||
|
return static_cast<opmath_t<T>>(per_sample_weight) *
|
||||||
|
static_cast<opmath_t<T>>(weight_val) +
|
||||||
|
out_val;
|
||||||
|
} else {
|
||||||
|
return static_cast<opmath_t<T>>(weight_val) + out_val;
|
||||||
|
}
|
||||||
|
}
|
||||||
|
};
|
||||||
|
|
||||||
|
template <typename T>
|
||||||
|
struct ReductionOp<EmbeddingBagMode::MEAN, T> {
|
||||||
|
inline opmath_t<T> operator()(
|
||||||
|
T weight_val,
|
||||||
|
opmath_t<T> out_val,
|
||||||
|
uint32_t,
|
||||||
|
constant T*,
|
||||||
|
uint32_t) {
|
||||||
|
return static_cast<opmath_t<T>>(weight_val) + out_val;
|
||||||
}
|
}
|
||||||
};
|
};
|
||||||
|
|
||||||
template <typename T>
|
template <typename T>
|
||||||
struct ReductionOp<EmbeddingBagMode::MAX, T> {
|
struct ReductionOp<EmbeddingBagMode::MAX, T> {
|
||||||
inline opmath_t<T> operator()(
|
inline opmath_t<T> operator()(
|
||||||
opmath_t<T> weight_val,
|
T weight_val,
|
||||||
opmath_t<T> out_val,
|
opmath_t<T> out_val,
|
||||||
bool is_first) {
|
uint32_t,
|
||||||
return (is_first || weight_val > out_val) ? weight_val : out_val;
|
constant T*,
|
||||||
}
|
uint32_t) {
|
||||||
};
|
return max(static_cast<opmath_t<T>>(weight_val), out_val);
|
||||||
|
|
||||||
template <EmbeddingBagMode M, typename T>
|
|
||||||
struct MaybeApplyPerSampleWeight {
|
|
||||||
inline opmath_t<T> operator()(
|
|
||||||
opmath_t<T> weight_val,
|
|
||||||
uint32_t per_sample_weights_index,
|
|
||||||
constant T* per_sample_weights,
|
|
||||||
uint32_t per_sample_weights_stride) {
|
|
||||||
return weight_val;
|
|
||||||
}
|
|
||||||
};
|
|
||||||
|
|
||||||
template <typename T>
|
|
||||||
struct MaybeApplyPerSampleWeight<EmbeddingBagMode::SUM, T> {
|
|
||||||
inline opmath_t<T> operator()(
|
|
||||||
opmath_t<T> weight_val,
|
|
||||||
uint32_t per_sample_weights_index,
|
|
||||||
constant T* per_sample_weights,
|
|
||||||
uint32_t per_sample_weights_stride) {
|
|
||||||
if (per_sample_weights_stride) {
|
|
||||||
T per_sample_weight = per_sample_weights
|
|
||||||
[per_sample_weights_stride * per_sample_weights_index];
|
|
||||||
return static_cast<opmath_t<T>>(per_sample_weight) * weight_val;
|
|
||||||
} else {
|
|
||||||
return weight_val;
|
|
||||||
}
|
|
||||||
}
|
|
||||||
};
|
|
||||||
|
|
||||||
template <EmbeddingBagMode M, typename T, typename I>
|
|
||||||
struct MaybeCalcMaxIndex {
|
|
||||||
inline void operator()(
|
|
||||||
opmath_t<T> weight_val,
|
|
||||||
opmath_t<T> out_val,
|
|
||||||
bool is_first,
|
|
||||||
thread I& max_idx,
|
|
||||||
I weight_idx,
|
|
||||||
bool pad) {}
|
|
||||||
};
|
|
||||||
|
|
||||||
template <typename T, typename I>
|
|
||||||
struct MaybeCalcMaxIndex<EmbeddingBagMode::MAX, T, I> {
|
|
||||||
inline void operator()(
|
|
||||||
opmath_t<T> weight_val,
|
|
||||||
opmath_t<T> out_val,
|
|
||||||
bool is_first,
|
|
||||||
thread I& max_idx,
|
|
||||||
I weight_idx,
|
|
||||||
bool pad) {
|
|
||||||
max_idx = !pad && (is_first || weight_val > out_val) ? weight_idx : max_idx;
|
|
||||||
}
|
}
|
||||||
};
|
};
|
||||||
|
|
||||||
@ -114,30 +96,6 @@ struct ReductionOpFinal<EmbeddingBagMode::MAX, T> {
|
|||||||
}
|
}
|
||||||
};
|
};
|
||||||
|
|
||||||
template <EmbeddingBagMode M, typename I>
|
|
||||||
struct MaybeWriteMaxIndex {
|
|
||||||
inline void operator()(
|
|
||||||
device I*,
|
|
||||||
const constant ::c10::metal::array<uint32_t, 2>&,
|
|
||||||
uint32_t,
|
|
||||||
uint32_t,
|
|
||||||
I) {}
|
|
||||||
};
|
|
||||||
|
|
||||||
template <typename I>
|
|
||||||
struct MaybeWriteMaxIndex<EmbeddingBagMode::MAX, I> {
|
|
||||||
inline void operator()(
|
|
||||||
device I* max_indices,
|
|
||||||
const constant ::c10::metal::array<uint32_t, 2>& max_indices_strides,
|
|
||||||
uint32_t bag_idx,
|
|
||||||
uint32_t feature_idx,
|
|
||||||
I max_idx) {
|
|
||||||
max_indices
|
|
||||||
[bag_idx * max_indices_strides[0] +
|
|
||||||
feature_idx * max_indices_strides[1]] = max_idx;
|
|
||||||
}
|
|
||||||
};
|
|
||||||
|
|
||||||
template <EmbeddingBagMode M, typename T, typename I>
|
template <EmbeddingBagMode M, typename T, typename I>
|
||||||
void embedding_bag_impl(
|
void embedding_bag_impl(
|
||||||
constant T* weight,
|
constant T* weight,
|
||||||
@ -154,7 +112,7 @@ void embedding_bag_impl(
|
|||||||
auto num_bags = params.num_bags;
|
auto num_bags = params.num_bags;
|
||||||
auto feature_size = params.feature_size;
|
auto feature_size = params.feature_size;
|
||||||
auto padding_idx = params.padding_idx;
|
auto padding_idx = params.padding_idx;
|
||||||
auto per_sample_weights_stride = params.per_sample_weights_stride;
|
auto per_sample_weights_strides = params.per_sample_weights_strides;
|
||||||
constant auto& output_strides = params.output_strides;
|
constant auto& output_strides = params.output_strides;
|
||||||
constant auto& weight_strides = params.weight_strides;
|
constant auto& weight_strides = params.weight_strides;
|
||||||
constant auto& max_indices_strides = params.max_indices_strides;
|
constant auto& max_indices_strides = params.max_indices_strides;
|
||||||
@ -162,6 +120,8 @@ void embedding_bag_impl(
|
|||||||
auto bag_idx = tid / feature_size;
|
auto bag_idx = tid / feature_size;
|
||||||
auto feature_idx = tid % feature_size;
|
auto feature_idx = tid % feature_size;
|
||||||
|
|
||||||
|
output += bag_idx * output_strides[0] + feature_idx * output_strides[1];
|
||||||
|
|
||||||
uint32_t offsets_end = min(bag_idx + 1, num_bags - 1);
|
uint32_t offsets_end = min(bag_idx + 1, num_bags - 1);
|
||||||
bool is_last_bag = bag_idx + 1 == num_bags;
|
bool is_last_bag = bag_idx + 1 == num_bags;
|
||||||
uint32_t indices_start = static_cast<uint32_t>(offsets[bag_idx]);
|
uint32_t indices_start = static_cast<uint32_t>(offsets[bag_idx]);
|
||||||
@ -171,37 +131,28 @@ void embedding_bag_impl(
|
|||||||
auto out_val = ReductionOpInit<M, T>()();
|
auto out_val = ReductionOpInit<M, T>()();
|
||||||
|
|
||||||
uint32_t bag_size_ = 0;
|
uint32_t bag_size_ = 0;
|
||||||
I max_idx = 0;
|
|
||||||
|
|
||||||
for (uint32_t indices_idx = indices_start; indices_idx < indices_end;
|
for (uint32_t indices_idx = indices_start; indices_idx < indices_end;
|
||||||
indices_idx++) {
|
indices_idx++) {
|
||||||
I weight_idx = indices[indices_idx];
|
I weight_idx = indices[indices_idx];
|
||||||
bool pad = (weight_idx == padding_idx);
|
bool pad = (weight_idx == padding_idx);
|
||||||
auto weight_val = static_cast<opmath_t<T>>(
|
T weight_val = weight
|
||||||
weight
|
[static_cast<uint32_t>(weight_idx) * weight_strides[0] +
|
||||||
[static_cast<uint32_t>(weight_idx) * weight_strides[0] +
|
feature_idx * weight_strides[1]];
|
||||||
feature_idx * weight_strides[1]]);
|
|
||||||
|
|
||||||
weight_val = MaybeApplyPerSampleWeight<M, T>()(
|
|
||||||
weight_val, indices_idx, per_sample_weights, per_sample_weights_stride);
|
|
||||||
|
|
||||||
auto new_out_val = ReductionOp<M, T>()(weight_val, out_val, bag_size_ == 0);
|
|
||||||
|
|
||||||
MaybeCalcMaxIndex<M, T, I>()(
|
|
||||||
weight_val, out_val, bag_size_ == 0, max_idx, weight_idx, pad);
|
|
||||||
|
|
||||||
out_val = pad ? out_val : new_out_val;
|
|
||||||
offset2bag[indices_idx] = bag_idx;
|
|
||||||
bag_size_ += static_cast<uint32_t>(!pad);
|
bag_size_ += static_cast<uint32_t>(!pad);
|
||||||
|
|
||||||
|
auto tmp_val = ReductionOp<M, T>()(
|
||||||
|
weight_val,
|
||||||
|
out_val,
|
||||||
|
indices_idx,
|
||||||
|
per_sample_weights,
|
||||||
|
per_sample_weights_strides);
|
||||||
|
|
||||||
|
out_val = pad ? out_val : tmp_val;
|
||||||
}
|
}
|
||||||
|
|
||||||
output[bag_idx * output_strides[0] + feature_idx * output_strides[1]] =
|
*output = ReductionOpFinal<M, T>()(out_val, bag_size_);
|
||||||
ReductionOpFinal<M, T>()(out_val, bag_size_);
|
|
||||||
|
|
||||||
bag_size[bag_idx] = bag_size_;
|
|
||||||
|
|
||||||
MaybeWriteMaxIndex<M, I>()(
|
|
||||||
max_indices, max_indices_strides, bag_idx, feature_idx, max_idx);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
#define DISPATCH_IMPL(MODE) \
|
#define DISPATCH_IMPL(MODE) \
|
||||||
|
|||||||
@ -223,6 +223,9 @@ void grid_sampler_single_element(
|
|||||||
auto input_size = input_sizes[input_dim];
|
auto input_size = input_sizes[input_dim];
|
||||||
auto coord = static_cast<opmath_t<T>>(coords[coord_dim]);
|
auto coord = static_cast<opmath_t<T>>(coords[coord_dim]);
|
||||||
|
|
||||||
|
// Interpret nan as -1
|
||||||
|
coord = isnan(coord) ? -1 : coord;
|
||||||
|
|
||||||
if (!align_corners) {
|
if (!align_corners) {
|
||||||
// Map unaligned grid space to aligned grid space
|
// Map unaligned grid space to aligned grid space
|
||||||
auto corner_alignment_factor = static_cast<opmath_t<T>>(input_size) /
|
auto corner_alignment_factor = static_cast<opmath_t<T>>(input_size) /
|
||||||
|
|||||||
@ -52,7 +52,9 @@ static void fill_depthwise_conv_desc(MPSGraphDepthwiseConvolution3DOpDescriptor*
|
|||||||
NSUInteger dilationRateInX,
|
NSUInteger dilationRateInX,
|
||||||
NSUInteger dilationRateInY,
|
NSUInteger dilationRateInY,
|
||||||
NSUInteger paddingHorizontal,
|
NSUInteger paddingHorizontal,
|
||||||
NSUInteger paddingVertical) {
|
NSUInteger paddingVertical,
|
||||||
|
c10::MemoryFormat memory_format,
|
||||||
|
NSUInteger groups) {
|
||||||
descriptor_.strides =
|
descriptor_.strides =
|
||||||
@[ @1, [[NSNumber alloc] initWithInteger:strideInY], [[NSNumber alloc] initWithInteger:strideInX] ];
|
@[ @1, [[NSNumber alloc] initWithInteger:strideInY], [[NSNumber alloc] initWithInteger:strideInX] ];
|
||||||
descriptor_.dilationRates =
|
descriptor_.dilationRates =
|
||||||
@ -101,7 +103,7 @@ static void fill_conv_desc(MPSGraphConvolution2DOpDescriptor* descriptor_,
|
|||||||
descriptor_.groups = groups;
|
descriptor_.groups = groups;
|
||||||
}
|
}
|
||||||
|
|
||||||
static Tensor _mps_convolution_impl(const Tensor& input_t,
|
static Tensor _mps_convolution_impl(const Tensor& input_t_,
|
||||||
const Tensor& weight_t,
|
const Tensor& weight_t,
|
||||||
const std::optional<Tensor>& bias_opt,
|
const std::optional<Tensor>& bias_opt,
|
||||||
IntArrayRef padding,
|
IntArrayRef padding,
|
||||||
@ -109,15 +111,12 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
IntArrayRef dilation,
|
IntArrayRef dilation,
|
||||||
int64_t groups,
|
int64_t groups,
|
||||||
std::optional<IntArrayRef> input_shape) {
|
std::optional<IntArrayRef> input_shape) {
|
||||||
constexpr auto kChannelsLast = MemoryFormat::ChannelsLast;
|
const bool is_macOS_15_0_or_newer = is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_0_PLUS);
|
||||||
constexpr auto kContiguous = MemoryFormat::Contiguous;
|
Tensor input_t = input_t_;
|
||||||
const bool is_macos_15_plus = is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_0_PLUS);
|
bool is3DConv = input_t.dim() == 5;
|
||||||
|
if (!is_macOS_15_0_or_newer || is3DConv) {
|
||||||
const bool is3DConv = input_t.dim() == 5;
|
input_t = input_t.contiguous();
|
||||||
const auto memory_format = input_t.suggest_memory_format();
|
}
|
||||||
const auto input_suggested_layout = memory_format == kChannelsLast && is_macos_15_plus ? kChannelsLast : kContiguous;
|
|
||||||
const bool is_channels_last = mps_conv_use_channels_last(input_t, weight_t) && !is3DConv;
|
|
||||||
const bool bias_defined = bias_opt ? bias_opt->defined() : false;
|
|
||||||
|
|
||||||
TORCH_CHECK(isFloatingType(input_t.scalar_type()), "Convolution is supported only for Floating types");
|
TORCH_CHECK(isFloatingType(input_t.scalar_type()), "Convolution is supported only for Floating types");
|
||||||
|
|
||||||
@ -127,6 +126,15 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
checkAllSameType(c, {input, weight});
|
checkAllSameType(c, {input, weight});
|
||||||
checkAllSameGPU(c, {input, weight});
|
checkAllSameGPU(c, {input, weight});
|
||||||
|
|
||||||
|
bool bias_defined;
|
||||||
|
|
||||||
|
if (bias_opt == std::nullopt)
|
||||||
|
bias_defined = false;
|
||||||
|
else
|
||||||
|
bias_defined = bias_opt->defined();
|
||||||
|
|
||||||
|
auto memory_format = input_t.suggest_memory_format();
|
||||||
|
bool is_channels_last = (memory_format == at::MemoryFormat::ChannelsLast) && !is3DConv;
|
||||||
auto output_t =
|
auto output_t =
|
||||||
at::empty(input_shape.has_value() ? input_shape.value()
|
at::empty(input_shape.has_value() ? input_shape.value()
|
||||||
: conv_output_size(input->sizes(), weight->sizes(), padding, stride, dilation),
|
: conv_output_size(input->sizes(), weight->sizes(), padding, stride, dilation),
|
||||||
@ -134,18 +142,12 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
std::nullopt,
|
std::nullopt,
|
||||||
kMPS,
|
kMPS,
|
||||||
std::nullopt,
|
std::nullopt,
|
||||||
is_channels_last ? kChannelsLast : kContiguous);
|
is_macOS_15_0_or_newer ? memory_format : MemoryFormat::Contiguous);
|
||||||
if (output_t.numel() == 0) {
|
if (output_t.numel() == 0) {
|
||||||
return output_t;
|
return output_t;
|
||||||
}
|
}
|
||||||
TensorArg output{output_t, "result", 0};
|
TensorArg output{output_t, "result", 0};
|
||||||
|
|
||||||
// TODO: Remove me when MacOS-14 is no longer supported
|
|
||||||
std::optional<Tensor> output_c;
|
|
||||||
if (!is_macos_15_plus && is_channels_last) {
|
|
||||||
output_c = at::empty_like(output_t, output_t.options().memory_format(kContiguous));
|
|
||||||
}
|
|
||||||
|
|
||||||
if (!is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_1_PLUS)) {
|
if (!is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_1_PLUS)) {
|
||||||
// On macOS < 15.1, MPS convolution kernel does not support output channels > 2^16
|
// On macOS < 15.1, MPS convolution kernel does not support output channels > 2^16
|
||||||
for (auto elem : output_t.sizes()) {
|
for (auto elem : output_t.sizes()) {
|
||||||
@ -184,22 +186,32 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
getArrayRefString(dilation),
|
getArrayRefString(dilation),
|
||||||
getArrayRefString(padding),
|
getArrayRefString(padding),
|
||||||
groups,
|
groups,
|
||||||
input_suggested_layout == kChannelsLast,
|
is_channels_last,
|
||||||
mps::getTensorsStringKey({input_t, weight_t}),
|
mps::getTensorsStringKey({input_t, weight_t}),
|
||||||
bias_defined,
|
bias_defined,
|
||||||
bias_shape_key);
|
bias_shape_key);
|
||||||
|
|
||||||
auto inputShape = mps::getMPSShape(input_t, input_suggested_layout);
|
MPSShape* inputShape = mps::getMPSShape(input_t, memory_format);
|
||||||
auto outputShape = mps::getMPSShape(output_t, input_suggested_layout);
|
MPSShape* outputShape = mps::getMPSShape(output_t, memory_format);
|
||||||
auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
|
MPSNDArray* inputNDArray = nil;
|
||||||
bool isDepthwiseConv =
|
MPSNDArray* outputNDArray = nil;
|
||||||
(groups > 1 && weight_t.size(1) == 1) && input_t.dim() >= 4 && weight_t.dim() >= 4 && !is_channels_last;
|
|
||||||
|
|
||||||
auto inputTensor = mpsGraphRankedPlaceHolder(mpsGraph, getMPSScalarType(input_t), inputShape);
|
if (input_t.is_contiguous(memory_format) && output_t.is_contiguous(memory_format) && is_macOS_15_0_or_newer) {
|
||||||
auto weightTensor = mpsGraphRankedPlaceHolder(mpsGraph, weight_t);
|
inputNDArray = getMPSNDArray(input_t, inputShape);
|
||||||
MPSGraphTensor* outputTensor = nil;
|
outputNDArray = getMPSNDArray(*output, outputShape);
|
||||||
|
}
|
||||||
|
|
||||||
|
auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
|
||||||
|
MPSShape* weightShape = mps::getMPSShape(weight_t);
|
||||||
|
bool isDepthwiseConv = ((groups > 1 && (weightShape[1].intValue == 1)) && inputShape.count >= 4 &&
|
||||||
|
weightShape.count >= 4 && !is_channels_last);
|
||||||
|
|
||||||
|
MPSGraphTensor* inputTensor =
|
||||||
|
mpsGraphRankedPlaceHolder(mpsGraph, getMPSScalarType(input_t.scalar_type()), inputShape);
|
||||||
|
MPSGraphTensor* weightTensor = mpsGraphRankedPlaceHolder(mpsGraph, weight_t);
|
||||||
|
MPSGraphTensor* outputTensor;
|
||||||
if (is3DConv) {
|
if (is3DConv) {
|
||||||
auto conv3dDescriptor_ = [[MPSGraphConvolution3DOpDescriptor new] autorelease];
|
MPSGraphConvolution3DOpDescriptor* conv3dDescriptor_ = [[MPSGraphConvolution3DOpDescriptor new] autorelease];
|
||||||
fill_conv3d_desc(conv3dDescriptor_,
|
fill_conv3d_desc(conv3dDescriptor_,
|
||||||
stride[2],
|
stride[2],
|
||||||
stride[1],
|
stride[1],
|
||||||
@ -217,9 +229,17 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
descriptor:conv3dDescriptor_
|
descriptor:conv3dDescriptor_
|
||||||
name:nil];
|
name:nil];
|
||||||
} else if (isDepthwiseConv) {
|
} else if (isDepthwiseConv) {
|
||||||
auto depthWiseConv3dDescriptor_ = [[MPSGraphDepthwiseConvolution3DOpDescriptor new] autorelease];
|
MPSGraphDepthwiseConvolution3DOpDescriptor* depthWiseConv3dDescriptor_ =
|
||||||
fill_depthwise_conv_desc(
|
[[MPSGraphDepthwiseConvolution3DOpDescriptor new] autorelease];
|
||||||
depthWiseConv3dDescriptor_, stride[1], stride[0], dilation[1], dilation[0], padding[1], padding[0]);
|
fill_depthwise_conv_desc(depthWiseConv3dDescriptor_,
|
||||||
|
stride[1],
|
||||||
|
stride[0],
|
||||||
|
dilation[1],
|
||||||
|
dilation[0],
|
||||||
|
padding[1],
|
||||||
|
padding[0],
|
||||||
|
memory_format,
|
||||||
|
groups);
|
||||||
|
|
||||||
MPSGraphTensor* weightTransposeTensor = [mpsGraph transposeTensor:weightTensor
|
MPSGraphTensor* weightTransposeTensor = [mpsGraph transposeTensor:weightTensor
|
||||||
dimension:-3
|
dimension:-3
|
||||||
@ -238,7 +258,7 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
dilation[0],
|
dilation[0],
|
||||||
padding[1],
|
padding[1],
|
||||||
padding[0],
|
padding[0],
|
||||||
input_suggested_layout,
|
memory_format,
|
||||||
groups);
|
groups);
|
||||||
|
|
||||||
outputTensor = [mpsGraph convolution2DWithSourceTensor:inputTensor
|
outputTensor = [mpsGraph convolution2DWithSourceTensor:inputTensor
|
||||||
@ -250,6 +270,13 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
MPSGraphTensor* biasTensor = nil;
|
MPSGraphTensor* biasTensor = nil;
|
||||||
if (bias_defined) {
|
if (bias_defined) {
|
||||||
biasTensor = mpsGraphUnrankedPlaceHolder(mpsGraph, getMPSDataType(bias_opt.value()));
|
biasTensor = mpsGraphUnrankedPlaceHolder(mpsGraph, getMPSDataType(bias_opt.value()));
|
||||||
|
}
|
||||||
|
|
||||||
|
if (is_channels_last && !is_macOS_15_0_or_newer) {
|
||||||
|
outputTensor = mps::convertNHWCtoNCHW(mpsGraph, outputTensor);
|
||||||
|
}
|
||||||
|
|
||||||
|
if (bias_defined) {
|
||||||
outputTensor = [mpsGraph additionWithPrimaryTensor:outputTensor secondaryTensor:biasTensor name:nil];
|
outputTensor = [mpsGraph additionWithPrimaryTensor:outputTensor secondaryTensor:biasTensor name:nil];
|
||||||
}
|
}
|
||||||
newCachedGraph->inputTensor_ = inputTensor;
|
newCachedGraph->inputTensor_ = inputTensor;
|
||||||
@ -258,26 +285,27 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
newCachedGraph->outputTensor_ = outputTensor;
|
newCachedGraph->outputTensor_ = outputTensor;
|
||||||
});
|
});
|
||||||
|
|
||||||
auto inputPlaceholder = input_suggested_layout == kContiguous
|
auto inputPlaceholder = inputNDArray ? Placeholder(cachedGraph->inputTensor_, inputNDArray)
|
||||||
? Placeholder(cachedGraph->inputTensor_, output_c || is3DConv ? input_t.contiguous() : input_t)
|
: Placeholder(cachedGraph->inputTensor_, input_t, inputShape);
|
||||||
: Placeholder(cachedGraph->inputTensor_, getMPSNDArray(input_t, inputShape));
|
auto weightsPlaceholder = Placeholder(cachedGraph->weightTensor_, weight_t);
|
||||||
auto outputPlaceholder = input_suggested_layout == kContiguous
|
|
||||||
? Placeholder(cachedGraph->outputTensor_, output_c ? *output_c : output_t)
|
|
||||||
: Placeholder(cachedGraph->outputTensor_, getMPSNDArray(output_t, outputShape));
|
|
||||||
auto weightsPlaceholder = Placeholder(cachedGraph->weightTensor_, output_c ? weight_t.contiguous() : weight_t);
|
|
||||||
auto biasPlaceholder = Placeholder();
|
auto biasPlaceholder = Placeholder();
|
||||||
// Reshape the bias to be broadcastable with output of conv2d or conv3d
|
// Reshape the bias to be broadcastable with output of conv2d or conv3d
|
||||||
if (bias_defined) {
|
if (bias_defined) {
|
||||||
if (is3DConv) {
|
if (is3DConv) {
|
||||||
biasPlaceholder = Placeholder(cachedGraph->biasTensor_, bias_opt->view({1, bias_shape[0], 1, 1, 1}));
|
biasPlaceholder = Placeholder(cachedGraph->biasTensor_, (bias_opt.value()).view({1, bias_shape[0], 1, 1, 1}));
|
||||||
} else if (input_suggested_layout == kChannelsLast) {
|
|
||||||
biasPlaceholder = Placeholder(cachedGraph->biasTensor_, bias_opt->view({1, 1, 1, bias_shape[0]}));
|
|
||||||
} else {
|
} else {
|
||||||
biasPlaceholder = Placeholder(cachedGraph->biasTensor_, bias_opt->view({1, bias_shape[0], 1, 1}));
|
if (is_channels_last && is_macOS_15_0_or_newer) {
|
||||||
|
biasPlaceholder = Placeholder(cachedGraph->biasTensor_, (bias_opt.value()).view({1, 1, 1, bias_shape[0]}));
|
||||||
|
} else {
|
||||||
|
biasPlaceholder = Placeholder(cachedGraph->biasTensor_, (bias_opt.value()).view({1, bias_shape[0], 1, 1}));
|
||||||
|
}
|
||||||
}
|
}
|
||||||
}
|
}
|
||||||
|
auto outputPlaceholder = outputNDArray ? Placeholder(cachedGraph->outputTensor_, outputNDArray)
|
||||||
|
: Placeholder(cachedGraph->outputTensor_, *output);
|
||||||
|
|
||||||
auto feeds = [[[NSMutableDictionary alloc] initWithCapacity:3] autorelease];
|
NSMutableDictionary<MPSGraphTensor*, MPSGraphTensorData*>* feeds =
|
||||||
|
[[[NSMutableDictionary alloc] initWithCapacity:3] autorelease];
|
||||||
feeds[inputPlaceholder.getMPSGraphTensor()] = inputPlaceholder.getMPSGraphTensorData();
|
feeds[inputPlaceholder.getMPSGraphTensor()] = inputPlaceholder.getMPSGraphTensorData();
|
||||||
feeds[weightsPlaceholder.getMPSGraphTensor()] = weightsPlaceholder.getMPSGraphTensorData();
|
feeds[weightsPlaceholder.getMPSGraphTensor()] = weightsPlaceholder.getMPSGraphTensorData();
|
||||||
if (bias_defined) {
|
if (bias_defined) {
|
||||||
@ -287,11 +315,7 @@ static Tensor _mps_convolution_impl(const Tensor& input_t,
|
|||||||
runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
|
runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
|
||||||
}
|
}
|
||||||
|
|
||||||
if (output_c) {
|
return *output;
|
||||||
output_t.copy_(*output_c);
|
|
||||||
}
|
|
||||||
|
|
||||||
return output_t;
|
|
||||||
}
|
}
|
||||||
|
|
||||||
Tensor _mps_convolution(const Tensor& input_t,
|
Tensor _mps_convolution(const Tensor& input_t,
|
||||||
@ -327,21 +351,14 @@ static Tensor mps_convolution_backward_input(IntArrayRef input_size,
|
|||||||
TensorArg grad_output{grad_output_t, "grad_output", 1}, weight{weight_t, "weight", 2};
|
TensorArg grad_output{grad_output_t, "grad_output", 1}, weight{weight_t, "weight", 2};
|
||||||
checkAllSameType(c, {grad_output, weight});
|
checkAllSameType(c, {grad_output, weight});
|
||||||
checkAllSameGPU(c, {grad_output, weight});
|
checkAllSameGPU(c, {grad_output, weight});
|
||||||
constexpr auto kChannelsLast = at::MemoryFormat::ChannelsLast;
|
auto memory_format = grad_output_t.suggest_memory_format();
|
||||||
bool is_channels_last = mps_conv_use_channels_last(grad_output_t, weight_t) && !is3DConv;
|
bool is_channels_last = (memory_format == at::MemoryFormat::ChannelsLast) && !is3DConv;
|
||||||
auto grad_input_t =
|
auto grad_input_t = at::empty(input_size, grad_output_t.options(), std::nullopt);
|
||||||
at::empty(input_size, grad_output_t.options(), is_channels_last ? std::optional(kChannelsLast) : std::nullopt);
|
|
||||||
|
|
||||||
// Avoid "grad_input" when this is being used as transposed convolution
|
// Avoid "grad_input" when this is being used as transposed convolution
|
||||||
TensorArg grad_input{grad_input_t, "result", 0};
|
TensorArg grad_input{grad_input_t, "result", 0};
|
||||||
convolution_shape_check(c, grad_input, weight, grad_output, padding, stride, dilation, groups);
|
convolution_shape_check(c, grad_input, weight, grad_output, padding, stride, dilation, groups);
|
||||||
|
|
||||||
// TODO: Remove me when MacOS-14 is no longer supported
|
|
||||||
std::optional<Tensor> grad_input_c;
|
|
||||||
if (!is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_0_PLUS) && is_channels_last) {
|
|
||||||
grad_input_c = at::empty_like(grad_input_t, grad_input_t.options().memory_format(MemoryFormat::Contiguous));
|
|
||||||
}
|
|
||||||
|
|
||||||
// Derive from MPSCachedGraph
|
// Derive from MPSCachedGraph
|
||||||
struct CachedGraph : public MPSCachedGraph {
|
struct CachedGraph : public MPSCachedGraph {
|
||||||
CachedGraph(MPSGraph* graph) : MPSCachedGraph(graph) {}
|
CachedGraph(MPSGraph* graph) : MPSCachedGraph(graph) {}
|
||||||
@ -353,6 +370,7 @@ static Tensor mps_convolution_backward_input(IntArrayRef input_size,
|
|||||||
// Add backward with input
|
// Add backward with input
|
||||||
@autoreleasepool {
|
@autoreleasepool {
|
||||||
MPSStream* stream = getCurrentMPSStream();
|
MPSStream* stream = getCurrentMPSStream();
|
||||||
|
|
||||||
MPSShape* mps_input_shape = getMPSShape(input_size);
|
MPSShape* mps_input_shape = getMPSShape(input_size);
|
||||||
std::string key = fmt::format("mps_{}_convolution_backward_input:{}:{}:{}:{}:{}:{}",
|
std::string key = fmt::format("mps_{}_convolution_backward_input:{}:{}:{}:{}:{}:{}",
|
||||||
is3DConv ? "3d_" : "",
|
is3DConv ? "3d_" : "",
|
||||||
@ -393,8 +411,15 @@ static Tensor mps_convolution_backward_input(IntArrayRef input_size,
|
|||||||
} else if (isDepthwiseConv) {
|
} else if (isDepthwiseConv) {
|
||||||
MPSGraphDepthwiseConvolution3DOpDescriptor* depthWiseConv3dDescriptor_ =
|
MPSGraphDepthwiseConvolution3DOpDescriptor* depthWiseConv3dDescriptor_ =
|
||||||
[[MPSGraphDepthwiseConvolution3DOpDescriptor new] autorelease];
|
[[MPSGraphDepthwiseConvolution3DOpDescriptor new] autorelease];
|
||||||
fill_depthwise_conv_desc(
|
fill_depthwise_conv_desc(depthWiseConv3dDescriptor_,
|
||||||
depthWiseConv3dDescriptor_, stride[1], stride[0], dilation[1], dilation[0], padding[1], padding[0]);
|
stride[1],
|
||||||
|
stride[0],
|
||||||
|
dilation[1],
|
||||||
|
dilation[0],
|
||||||
|
padding[1],
|
||||||
|
padding[0],
|
||||||
|
at::MemoryFormat::Contiguous,
|
||||||
|
groups);
|
||||||
MPSGraphTensor* weightTransposeTensor = [mpsGraph transposeTensor:weightTensor
|
MPSGraphTensor* weightTransposeTensor = [mpsGraph transposeTensor:weightTensor
|
||||||
dimension:-3
|
dimension:-3
|
||||||
withDimension:-4
|
withDimension:-4
|
||||||
@ -429,18 +454,14 @@ static Tensor mps_convolution_backward_input(IntArrayRef input_size,
|
|||||||
newCachedGraph->gradInputTensor_ = gradInputTensor;
|
newCachedGraph->gradInputTensor_ = gradInputTensor;
|
||||||
});
|
});
|
||||||
|
|
||||||
auto gradOutputPlaceholder =
|
auto gradOutputPlaceholder = Placeholder(cachedGraph->gradOutputTensor_, grad_output_t);
|
||||||
Placeholder(cachedGraph->gradOutputTensor_, grad_input_c ? grad_output_t.contiguous() : grad_output_t);
|
auto weightsPlaceholder = Placeholder(cachedGraph->weightTensor_, weight_t);
|
||||||
auto weightsPlaceholder = Placeholder(cachedGraph->weightTensor_, grad_input_c ? weight_t.contiguous() : weight_t);
|
auto outputPlaceholder = Placeholder(cachedGraph->gradInputTensor_, *grad_input);
|
||||||
auto outputPlaceholder = Placeholder(cachedGraph->gradInputTensor_, grad_input_c ? *grad_input_c : grad_input_t);
|
|
||||||
|
|
||||||
auto feeds = dictionaryFromPlaceholders(gradOutputPlaceholder, weightsPlaceholder);
|
auto feeds = dictionaryFromPlaceholders(gradOutputPlaceholder, weightsPlaceholder);
|
||||||
runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
|
runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
|
||||||
}
|
}
|
||||||
if (grad_input_c) {
|
return *grad_input;
|
||||||
grad_input_t.copy_(*grad_input_c);
|
|
||||||
}
|
|
||||||
return grad_input_t;
|
|
||||||
}
|
}
|
||||||
|
|
||||||
static Tensor mps_convolution_backward_weights(IntArrayRef weight_size,
|
static Tensor mps_convolution_backward_weights(IntArrayRef weight_size,
|
||||||
@ -453,11 +474,9 @@ static Tensor mps_convolution_backward_weights(IntArrayRef weight_size,
|
|||||||
bool bias_defined) {
|
bool bias_defined) {
|
||||||
using namespace at::native::mps;
|
using namespace at::native::mps;
|
||||||
using namespace mps;
|
using namespace mps;
|
||||||
const bool is3DConv = input_t.dim() == 5;
|
bool is3DConv = input_t.dim() == 5;
|
||||||
TORCH_CHECK(isFloatingType(grad_output_t.scalar_type()), "Convolution is supported only for Floating types");
|
TORCH_CHECK(isFloatingType(grad_output_t.scalar_type()), "Convolution is supported only for Floating types");
|
||||||
CheckedFrom c = "mps_convolution_backward_weights";
|
CheckedFrom c = "mps_convolution_backward_weights";
|
||||||
constexpr auto kChannelsLast = at::MemoryFormat::ChannelsLast;
|
|
||||||
bool is_channels_last = mps_conv_use_channels_last(input_t, grad_output_t) && !is3DConv;
|
|
||||||
|
|
||||||
// For uniformity with everything else, although it seems grad_weight
|
// For uniformity with everything else, although it seems grad_weight
|
||||||
// would be unambiguous too.
|
// would be unambiguous too.
|
||||||
@ -468,8 +487,7 @@ static Tensor mps_convolution_backward_weights(IntArrayRef weight_size,
|
|||||||
checkAllSameGPU(c, {grad_output, input});
|
checkAllSameGPU(c, {grad_output, input});
|
||||||
|
|
||||||
auto grad_weight_t =
|
auto grad_weight_t =
|
||||||
at::empty(weight_size, grad_output_t.options(), is_channels_last ? std::optional(kChannelsLast) : std::nullopt);
|
at::empty(weight_size, grad_output_t.scalar_type(), std::nullopt, kMPS, std::nullopt, std::nullopt);
|
||||||
|
|
||||||
TensorArg grad_weight{grad_weight_t, "result", 0};
|
TensorArg grad_weight{grad_weight_t, "result", 0};
|
||||||
|
|
||||||
convolution_shape_check(c, input, grad_weight, grad_output, padding, stride, dilation, groups);
|
convolution_shape_check(c, input, grad_weight, grad_output, padding, stride, dilation, groups);
|
||||||
@ -482,23 +500,16 @@ static Tensor mps_convolution_backward_weights(IntArrayRef weight_size,
|
|||||||
MPSGraphTensor* gradWeightTensor_ = nil;
|
MPSGraphTensor* gradWeightTensor_ = nil;
|
||||||
};
|
};
|
||||||
|
|
||||||
// TODO: Remove me when MacOS-14 is no longer supported
|
|
||||||
std::optional<Tensor> grad_weight_c;
|
|
||||||
if (!is_macos_13_or_newer(MacOSVersion::MACOS_VER_15_0_PLUS) && is_channels_last) {
|
|
||||||
grad_weight_c = at::empty_like(grad_weight_t, grad_weight_t.options().memory_format(MemoryFormat::Contiguous));
|
|
||||||
}
|
|
||||||
|
|
||||||
@autoreleasepool {
|
@autoreleasepool {
|
||||||
MPSStream* stream = getCurrentMPSStream();
|
MPSStream* stream = getCurrentMPSStream();
|
||||||
|
|
||||||
MPSShape* mps_weight_shape = getMPSShape(weight_size);
|
MPSShape* mps_weight_shape = getMPSShape(weight_size);
|
||||||
std::string key = fmt::format("mps_{}convolution_backward_weights:{}:{}:{}:{}:{}:{}",
|
std::string key = fmt::format("mps_{}convolution_backward_weights:{}:{}:{}:{}:{}",
|
||||||
is3DConv ? "3d_" : "",
|
is3DConv ? "3d_" : "",
|
||||||
getArrayRefString(stride),
|
getArrayRefString(stride),
|
||||||
getArrayRefString(dilation),
|
getArrayRefString(dilation),
|
||||||
getArrayRefString(padding),
|
getArrayRefString(padding),
|
||||||
groups,
|
groups,
|
||||||
is_channels_last,
|
|
||||||
getTensorsStringKey({grad_output_t, input_t, grad_weight_t}));
|
getTensorsStringKey({grad_output_t, input_t, grad_weight_t}));
|
||||||
auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
|
auto cachedGraph = LookUpOrCreateCachedGraph<CachedGraph>(key, [&](auto mpsGraph, auto newCachedGraph) {
|
||||||
MPSShape* inputShape = getMPSShape(input_t);
|
MPSShape* inputShape = getMPSShape(input_t);
|
||||||
@ -530,8 +541,15 @@ static Tensor mps_convolution_backward_weights(IntArrayRef weight_size,
|
|||||||
} else if (isDepthwiseConv) {
|
} else if (isDepthwiseConv) {
|
||||||
MPSGraphDepthwiseConvolution3DOpDescriptor* depthWiseConv3dDescriptor_ =
|
MPSGraphDepthwiseConvolution3DOpDescriptor* depthWiseConv3dDescriptor_ =
|
||||||
[[MPSGraphDepthwiseConvolution3DOpDescriptor new] autorelease];
|
[[MPSGraphDepthwiseConvolution3DOpDescriptor new] autorelease];
|
||||||
fill_depthwise_conv_desc(
|
fill_depthwise_conv_desc(depthWiseConv3dDescriptor_,
|
||||||
depthWiseConv3dDescriptor_, stride[1], stride[0], dilation[1], dilation[0], padding[1], padding[0]);
|
stride[1],
|
||||||
|
stride[0],
|
||||||
|
dilation[1],
|
||||||
|
dilation[0],
|
||||||
|
padding[1],
|
||||||
|
padding[0],
|
||||||
|
at::MemoryFormat::Contiguous,
|
||||||
|
groups);
|
||||||
NSNumber* outputFeatChannelDim = mps_weight_shape[0];
|
NSNumber* outputFeatChannelDim = mps_weight_shape[0];
|
||||||
MPSShape* weightShapeTranspose = @[ @1, outputFeatChannelDim, mps_weight_shape[2], mps_weight_shape[3] ];
|
MPSShape* weightShapeTranspose = @[ @1, outputFeatChannelDim, mps_weight_shape[2], mps_weight_shape[3] ];
|
||||||
MPSGraphTensor* gradWeightTensorTranspose =
|
MPSGraphTensor* gradWeightTensorTranspose =
|
||||||
@ -565,19 +583,14 @@ static Tensor mps_convolution_backward_weights(IntArrayRef weight_size,
|
|||||||
newCachedGraph->gradWeightTensor_ = gradWeightTensor;
|
newCachedGraph->gradWeightTensor_ = gradWeightTensor;
|
||||||
});
|
});
|
||||||
|
|
||||||
auto gradOutputPlaceholder =
|
auto gradOutputPlaceholder = Placeholder(cachedGraph->gradOutputTensor_, grad_output_t);
|
||||||
Placeholder(cachedGraph->gradOutputTensor_, grad_weight_c ? grad_output_t.contiguous() : grad_output_t);
|
auto inputPlaceholder = Placeholder(cachedGraph->inputTensor_, input_t);
|
||||||
auto inputPlaceholder = Placeholder(cachedGraph->inputTensor_, grad_weight_c ? input_t.contiguous() : input_t);
|
auto outputPlaceholder = Placeholder(cachedGraph->gradWeightTensor_, grad_weight_t);
|
||||||
auto outputPlaceholder =
|
|
||||||
Placeholder(cachedGraph->gradWeightTensor_, grad_weight_c ? *grad_weight_c : grad_weight_t);
|
|
||||||
|
|
||||||
auto feeds = dictionaryFromPlaceholders(gradOutputPlaceholder, inputPlaceholder);
|
auto feeds = dictionaryFromPlaceholders(gradOutputPlaceholder, inputPlaceholder);
|
||||||
runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
|
runMPSGraph(stream, cachedGraph->graph(), feeds, outputPlaceholder);
|
||||||
}
|
}
|
||||||
|
|
||||||
if (grad_weight_c) {
|
|
||||||
grad_weight_t.copy_(*grad_weight_c);
|
|
||||||
}
|
|
||||||
return grad_weight_t;
|
return grad_weight_t;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
|||||||
@ -66,12 +66,11 @@ static std::tuple<Tensor, Tensor, Tensor, Tensor> _embedding_bag_mps_impl(
|
|||||||
int64_t num_indices = indices.size(0);
|
int64_t num_indices = indices.size(0);
|
||||||
int64_t num_bags = offsets.size(0);
|
int64_t num_bags = offsets.size(0);
|
||||||
if (include_last_offset) {
|
if (include_last_offset) {
|
||||||
TORCH_CHECK(num_bags >= 1, "include_last_offset: number of offsets should be at least 1");
|
|
||||||
num_bags -= 1;
|
num_bags -= 1;
|
||||||
}
|
}
|
||||||
int64_t feature_size = weight.size(1);
|
int64_t feature_size = weight.size(1);
|
||||||
|
|
||||||
auto bag_size = at::empty({num_bags}, indices.options());
|
auto bag_size = at::empty(offsets.sizes(), indices.options());
|
||||||
auto offset2bag = at::empty({indices.size(0)}, indices.options());
|
auto offset2bag = at::empty({indices.size(0)}, indices.options());
|
||||||
auto output = at::empty({num_bags, feature_size}, weight.options());
|
auto output = at::empty({num_bags, feature_size}, weight.options());
|
||||||
|
|
||||||
@ -95,7 +94,7 @@ static std::tuple<Tensor, Tensor, Tensor, Tensor> _embedding_bag_mps_impl(
|
|||||||
}
|
}
|
||||||
|
|
||||||
bool use_per_sample_weights = per_sample_weights_opt.has_value() && per_sample_weights_opt->defined();
|
bool use_per_sample_weights = per_sample_weights_opt.has_value() && per_sample_weights_opt->defined();
|
||||||
params.per_sample_weights_stride = use_per_sample_weights ? per_sample_weights_opt->stride(0) : 0;
|
params.per_sample_weights_strides = use_per_sample_weights ? per_sample_weights_opt->stride(0) : 0;
|
||||||
|
|
||||||
params.num_indices = num_indices;
|
params.num_indices = num_indices;
|
||||||
params.num_bags = num_bags;
|
params.num_bags = num_bags;
|
||||||
|
|||||||
@ -20,7 +20,6 @@
|
|||||||
#include <ATen/ops/baddbmm_native.h>
|
#include <ATen/ops/baddbmm_native.h>
|
||||||
#include <ATen/ops/bmm_native.h>
|
#include <ATen/ops/bmm_native.h>
|
||||||
#include <ATen/ops/cholesky_native.h>
|
#include <ATen/ops/cholesky_native.h>
|
||||||
#include <ATen/ops/eye_native.h>
|
|
||||||
#include <ATen/ops/linalg_cholesky_ex_native.h>
|
#include <ATen/ops/linalg_cholesky_ex_native.h>
|
||||||
#include <ATen/ops/linalg_inv_ex_native.h>
|
#include <ATen/ops/linalg_inv_ex_native.h>
|
||||||
#include <ATen/ops/linalg_lu_factor_ex_native.h>
|
#include <ATen/ops/linalg_lu_factor_ex_native.h>
|
||||||
@ -497,24 +496,26 @@ static void linalg_inv_ex_out_mps_impl(const Tensor& A, bool check_errors, const
|
|||||||
using namespace mps;
|
using namespace mps;
|
||||||
TORCH_CHECK(result.is_mps(), "Output tensor is not MPS");
|
TORCH_CHECK(result.is_mps(), "Output tensor is not MPS");
|
||||||
TORCH_CHECK(!A.is_complex(), "linalg_inv: not supported for complex types yet!");
|
TORCH_CHECK(!A.is_complex(), "linalg_inv: not supported for complex types yet!");
|
||||||
|
using CachedGraph = MPSUnaryCachedGraph;
|
||||||
|
|
||||||
|
MPSStream* stream = getCurrentMPSStream();
|
||||||
info.zero_();
|
info.zero_();
|
||||||
|
|
||||||
if (A.numel() == 0) {
|
if (A.numel() == 0) {
|
||||||
return;
|
return;
|
||||||
}
|
}
|
||||||
|
|
||||||
|
if (!result.is_contiguous()) {
|
||||||
|
result.unsafeGetTensorImpl()->empty_tensor_restride(MemoryFormat::Contiguous);
|
||||||
|
}
|
||||||
auto A_sizes = A.sizes();
|
auto A_sizes = A.sizes();
|
||||||
int ndim = A.dim();
|
int ndim = A.dim();
|
||||||
|
|
||||||
Tensor LU = empty_like(A, MemoryFormat::Contiguous);
|
Tensor LU = empty_like(A);
|
||||||
Tensor identity = eye(A.size(-2), A.size(-1), A.scalar_type(), A.options().layout(), A.device()).expand_as(A);
|
Tensor identity = zeros_like(A);
|
||||||
Tensor pivots = empty({A_sizes.begin(), A_sizes.end() - 1}, A.options().dtype(kInt));
|
Tensor pivots = empty({A_sizes.begin(), A_sizes.end() - 1}, A.options().dtype(kInt));
|
||||||
// need to do this to keep the strides of the result tensor
|
(ndim == 2 ? identity.diagonal() : identity.diagonal(0, -2, -1)).fill_(1);
|
||||||
// mps's solve expects row major layout, while inductor
|
linalg_solve_out_mps_impl(A, identity, true, check_errors, result, LU, pivots, info);
|
||||||
// expects result to be column major
|
|
||||||
Tensor tmp = empty_like(A, MemoryFormat::Contiguous);
|
|
||||||
linalg_solve_out_mps_impl(A, identity, true, check_errors, tmp, LU, pivots, info);
|
|
||||||
result.copy_(tmp);
|
|
||||||
}
|
}
|
||||||
|
|
||||||
static Tensor& mm_out_mps_impl(const Tensor& self, const Tensor& other, Tensor& output) {
|
static Tensor& mm_out_mps_impl(const Tensor& self, const Tensor& other, Tensor& output) {
|
||||||
|
|||||||
@ -519,13 +519,6 @@ static void max_unpool_out_mps_template(const Tensor& input,
|
|||||||
Tensor& output,
|
Tensor& output,
|
||||||
const int32_t pooling_dims,
|
const int32_t pooling_dims,
|
||||||
const std::string& op_name) {
|
const std::string& op_name) {
|
||||||
TORCH_CHECK(output_size_.size() == static_cast<size_t>(pooling_dims),
|
|
||||||
op_name,
|
|
||||||
"There should be exactly ",
|
|
||||||
pooling_dims,
|
|
||||||
" elements but got ",
|
|
||||||
output_size_.size());
|
|
||||||
|
|
||||||
auto dims = input.dim();
|
auto dims = input.dim();
|
||||||
auto leading_dims = input.dim() - pooling_dims;
|
auto leading_dims = input.dim() - pooling_dims;
|
||||||
|
|
||||||
|
|||||||
@ -9,22 +9,11 @@
|
|||||||
#else
|
#else
|
||||||
#include <ATen/ops/_unique2.h>
|
#include <ATen/ops/_unique2.h>
|
||||||
#include <ATen/ops/_unique2_native.h>
|
#include <ATen/ops/_unique2_native.h>
|
||||||
#include <ATen/ops/arange.h>
|
|
||||||
#include <ATen/ops/argsort.h>
|
|
||||||
#include <ATen/ops/cat.h>
|
|
||||||
#include <ATen/ops/cumsum.h>
|
|
||||||
#include <ATen/ops/full.h>
|
|
||||||
#include <ATen/ops/masked_select.h>
|
|
||||||
#include <ATen/ops/nonzero.h>
|
|
||||||
#include <ATen/ops/ones.h>
|
|
||||||
#include <ATen/ops/ones_like.h>
|
|
||||||
#include <ATen/ops/slice.h>
|
#include <ATen/ops/slice.h>
|
||||||
#include <ATen/ops/unique_consecutive.h>
|
#include <ATen/ops/unique_consecutive.h>
|
||||||
#include <ATen/ops/unique_consecutive_native.h>
|
#include <ATen/ops/unique_consecutive_native.h>
|
||||||
#include <ATen/ops/unique_dim_consecutive.h>
|
#include <ATen/ops/unique_dim_consecutive.h>
|
||||||
#include <ATen/ops/unique_dim_consecutive_native.h>
|
#include <ATen/ops/unique_dim_consecutive_native.h>
|
||||||
#include <ATen/ops/unique_dim_native.h>
|
|
||||||
#include <ATen/ops/zeros.h>
|
|
||||||
#endif
|
#endif
|
||||||
|
|
||||||
namespace at::native {
|
namespace at::native {
|
||||||
@ -316,85 +305,4 @@ std::tuple<Tensor, Tensor, Tensor> _unique2_mps(const Tensor& self,
|
|||||||
return _unique_impl_mps(self, return_inverse, return_counts, false, std::nullopt);
|
return _unique_impl_mps(self, return_inverse, return_counts, false, std::nullopt);
|
||||||
}
|
}
|
||||||
|
|
||||||
static Tensor lexsort_rows_perm_mps(const Tensor& mat_2d) {
|
|
||||||
const auto rows = mat_2d.size(0), cols = mat_2d.size(1);
|
|
||||||
if (rows <= 1 || cols == 0) {
|
|
||||||
return arange(rows, mat_2d.options().dtype(kLong));
|
|
||||||
}
|
|
||||||
|
|
||||||
auto perm = arange(rows, mat_2d.options().dtype(kLong));
|
|
||||||
for (auto c = cols - 1; c >= 0; --c) {
|
|
||||||
auto keys = mat_2d.select(1, c).index_select(0, perm);
|
|
||||||
const auto idx = argsort(keys, /*dim=*/0, /*descending=*/false);
|
|
||||||
perm = perm.index_select(0, idx);
|
|
||||||
}
|
|
||||||
return perm;
|
|
||||||
}
|
|
||||||
|
|
||||||
static std::tuple<Tensor, Tensor, Tensor> unique_dim_sorted_mps_impl(const Tensor& self,
|
|
||||||
int64_t dim,
|
|
||||||
bool return_inverse,
|
|
||||||
bool return_counts) {
|
|
||||||
dim = maybe_wrap_dim(dim, self.dim());
|
|
||||||
|
|
||||||
auto sizes = self.sizes().vec();
|
|
||||||
auto num_zero_dims = std::count(sizes.begin(), sizes.end(), (int64_t)0);
|
|
||||||
if (self.size(dim) == 0) {
|
|
||||||
auto output = at::empty(sizes, self.options());
|
|
||||||
auto inverse_indices = at::empty({0}, self.options().dtype(kLong));
|
|
||||||
auto counts = at::empty({0}, self.options().dtype(kLong));
|
|
||||||
return {output, inverse_indices, counts};
|
|
||||||
}
|
|
||||||
|
|
||||||
auto transposed = self.moveaxis(dim, 0);
|
|
||||||
auto orig_sizes = transposed.sizes().vec();
|
|
||||||
auto rows = transposed.size(0);
|
|
||||||
auto input_flat = transposed.contiguous().view({rows, -1});
|
|
||||||
|
|
||||||
auto perm = lexsort_rows_perm_mps(input_flat);
|
|
||||||
auto input_sorted = input_flat.index_select(0, perm);
|
|
||||||
|
|
||||||
Tensor is_unique = at::zeros({rows}, self.options().dtype(kBool));
|
|
||||||
if (rows > 0) {
|
|
||||||
is_unique.narrow(0, 0, 1).fill_(true);
|
|
||||||
}
|
|
||||||
if (rows > 1) {
|
|
||||||
auto a = input_sorted.narrow(0, 1, rows - 1);
|
|
||||||
auto b = input_sorted.narrow(0, 0, rows - 1);
|
|
||||||
auto row_changed = a.ne(b).any(1);
|
|
||||||
is_unique.narrow(0, 1, rows - 1).copy_(row_changed);
|
|
||||||
}
|
|
||||||
|
|
||||||
auto unique_pos = nonzero(is_unique).squeeze(1);
|
|
||||||
auto group_id = cumsum(is_unique.to(kLong), 0).sub(1);
|
|
||||||
|
|
||||||
auto unique_rows_2d = input_sorted.index_select(0, unique_pos);
|
|
||||||
|
|
||||||
Tensor inverse_indices = empty({0}, self.options().dtype(kLong));
|
|
||||||
if (return_inverse) {
|
|
||||||
inverse_indices = empty({rows}, self.options().dtype(kLong));
|
|
||||||
inverse_indices.index_copy_(0, perm, group_id);
|
|
||||||
}
|
|
||||||
|
|
||||||
Tensor counts = empty({0}, self.options().dtype(kLong));
|
|
||||||
if (return_counts) {
|
|
||||||
const auto num_unique = unique_pos.size(0);
|
|
||||||
counts = zeros({num_unique}, self.options().dtype(kLong));
|
|
||||||
counts.scatter_add_(0, group_id, ones_like(group_id, group_id.options().dtype(kLong)));
|
|
||||||
}
|
|
||||||
|
|
||||||
orig_sizes[0] = unique_rows_2d.size(0);
|
|
||||||
auto output = unique_rows_2d.view(orig_sizes).moveaxis(0, dim);
|
|
||||||
|
|
||||||
return std::make_tuple(std::move(output), std::move(inverse_indices), std::move(counts));
|
|
||||||
}
|
|
||||||
|
|
||||||
std::tuple<Tensor, Tensor, Tensor> unique_dim_mps(const Tensor& self,
|
|
||||||
int64_t dim,
|
|
||||||
const bool /*sorted*/,
|
|
||||||
const bool return_inverse,
|
|
||||||
const bool return_counts) {
|
|
||||||
return unique_dim_sorted_mps_impl(self, dim, return_inverse, return_counts);
|
|
||||||
}
|
|
||||||
|
|
||||||
} // namespace at::native
|
} // namespace at::native
|
||||||
|
|||||||
@ -1409,7 +1409,7 @@
|
|||||||
- func: _sparse_broadcast_to(Tensor(a) self, int[] size) -> Tensor(a)
|
- func: _sparse_broadcast_to(Tensor(a) self, int[] size) -> Tensor(a)
|
||||||
variants: function
|
variants: function
|
||||||
dispatch:
|
dispatch:
|
||||||
SparseCPU, SparseCUDA, SparseMPS: sparse_broadcast_to
|
SparseCPU, SparseCUDA: sparse_broadcast_to
|
||||||
|
|
||||||
- func: cat(Tensor[] tensors, int dim=0) -> Tensor
|
- func: cat(Tensor[] tensors, int dim=0) -> Tensor
|
||||||
structured_delegate: cat.out
|
structured_delegate: cat.out
|
||||||
@ -3858,7 +3858,7 @@
|
|||||||
device_check: NoCheck # TensorIterator
|
device_check: NoCheck # TensorIterator
|
||||||
structured: True
|
structured: True
|
||||||
dispatch:
|
dispatch:
|
||||||
CPU, CUDA, MTIA: aminmax_out
|
CPU, CUDA: aminmax_out
|
||||||
MPS: aminmax_out_mps
|
MPS: aminmax_out_mps
|
||||||
|
|
||||||
- func: _compute_linear_combination(Tensor input, Tensor coefficients) -> Tensor
|
- func: _compute_linear_combination(Tensor input, Tensor coefficients) -> Tensor
|
||||||
@ -3909,7 +3909,7 @@
|
|||||||
- func: amax.out(Tensor self, int[1] dim=[], bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
|
- func: amax.out(Tensor self, int[1] dim=[], bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
|
||||||
structured: True
|
structured: True
|
||||||
dispatch:
|
dispatch:
|
||||||
CPU, CUDA, MTIA: amax_out
|
CPU, CUDA: amax_out
|
||||||
MPS: amax_out_mps
|
MPS: amax_out_mps
|
||||||
|
|
||||||
# Return: (Tensor output, Tensor indices)
|
# Return: (Tensor output, Tensor indices)
|
||||||
@ -4090,7 +4090,7 @@
|
|||||||
- func: amin.out(Tensor self, int[1] dim=[], bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
|
- func: amin.out(Tensor self, int[1] dim=[], bool keepdim=False, *, Tensor(a!) out) -> Tensor(a!)
|
||||||
structured: True
|
structured: True
|
||||||
dispatch:
|
dispatch:
|
||||||
CPU, CUDA, MTIA: amin_out
|
CPU, CUDA: amin_out
|
||||||
MPS: amin_out_mps
|
MPS: amin_out_mps
|
||||||
|
|
||||||
# TODO: Add this function to MPS dispatch key so that we avoid declaring it in
|
# TODO: Add this function to MPS dispatch key so that we avoid declaring it in
|
||||||
@ -6450,7 +6450,6 @@
|
|||||||
dispatch:
|
dispatch:
|
||||||
CPU: unique_dim_cpu
|
CPU: unique_dim_cpu
|
||||||
CUDA: unique_dim_cuda
|
CUDA: unique_dim_cuda
|
||||||
MPS: unique_dim_mps
|
|
||||||
tags: dynamic_output_shape
|
tags: dynamic_output_shape
|
||||||
autogen: unique_dim.out
|
autogen: unique_dim.out
|
||||||
|
|
||||||
|
|||||||
@ -158,46 +158,12 @@ c10::intrusive_ptr<EmbeddingPackedParamsBase> PackedEmbeddingBagWeight::prepack(
|
|||||||
return packed_ptr;
|
return packed_ptr;
|
||||||
}
|
}
|
||||||
|
|
||||||
#ifdef USE_FBGEMM
|
|
||||||
namespace {
|
|
||||||
/// Number of columns in the rowwise min/max buffer passed to the quantization function(s)
|
|
||||||
constexpr int kRowwiseMinMaxNumCols = 2;
|
|
||||||
|
|
||||||
bool _validate_rowwise_min_max(
|
|
||||||
const at::Tensor& weight,
|
|
||||||
const std::optional<at::Tensor>& rowwise_min_max_opt) {
|
|
||||||
const auto is_valid_rowwise_min_max = rowwise_min_max_opt.has_value();
|
|
||||||
|
|
||||||
if (is_valid_rowwise_min_max) {
|
|
||||||
TORCH_CHECK(
|
|
||||||
(rowwise_min_max_opt->dim() == 2 &&
|
|
||||||
rowwise_min_max_opt->size(0) == weight.size(0) &&
|
|
||||||
rowwise_min_max_opt->size(1) == kRowwiseMinMaxNumCols),
|
|
||||||
"'rowwise_min_max' must be a 2D tensor with shape [num_rows(weight), 2].");
|
|
||||||
}
|
|
||||||
|
|
||||||
return is_valid_rowwise_min_max;
|
|
||||||
}
|
|
||||||
|
|
||||||
auto _get_rowwise_min_max_contig(
|
|
||||||
const std::optional<at::Tensor>& rowwise_min_max_opt) {
|
|
||||||
return rowwise_min_max_opt.has_value()
|
|
||||||
? rowwise_min_max_opt->expect_contiguous(rowwise_min_max_opt->suggest_memory_format())
|
|
||||||
: at::borrow_from_optional_tensor(rowwise_min_max_opt);
|
|
||||||
}
|
|
||||||
}
|
|
||||||
#endif // USE_FBGEMM
|
|
||||||
|
|
||||||
namespace at::native {
|
namespace at::native {
|
||||||
|
|
||||||
// Note - This is a temporary pack function for embedding bag which quantizes
|
// Note - This is a temporary pack function for embedding bag which quantizes
|
||||||
// and packs the float weight tensor. In the next step it will be replaced by a
|
// and packs the float weight tensor. In the next step it will be replaced by a
|
||||||
// quantize and pack function once we support FP scale and FP zero_point
|
// quantize and pack function once we support FP scale and FP zero_point
|
||||||
//
|
//
|
||||||
// The optional rowwise_min_max argument is to support callers to pass in the min/max
|
|
||||||
// values of the weight tensor. If the rowwise_min_max is not provided, the min/max
|
|
||||||
// values will be computed from the weight tensor.
|
|
||||||
//
|
|
||||||
// Python example examining a packed 8bit zero_point and scale:
|
// Python example examining a packed 8bit zero_point and scale:
|
||||||
//
|
//
|
||||||
// >> x = torch.from_numpy(np.array([[[10, 20], [30, 40]],[[50, 60], [70, 80]]],
|
// >> x = torch.from_numpy(np.array([[[10, 20], [30, 40]],[[50, 60], [70, 80]]],
|
||||||
@ -255,10 +221,7 @@ namespace at::native {
|
|||||||
//
|
//
|
||||||
// [[50. , 60.00000035],
|
// [[50. , 60.00000035],
|
||||||
// [70. , 80.00000035]]])
|
// [70. , 80.00000035]]])
|
||||||
Tensor& qembeddingbag_byte_prepack_out(
|
Tensor& qembeddingbag_byte_prepack_out(Tensor& output, const Tensor& weight) {
|
||||||
Tensor& output,
|
|
||||||
const Tensor& weight,
|
|
||||||
const std::optional<Tensor>& rowwise_min_max_opt) {
|
|
||||||
// The "last" dimension of an N-Dimensioned batch of embedding bags is
|
// The "last" dimension of an N-Dimensioned batch of embedding bags is
|
||||||
// quantization channel. E.g. for a 2D embedding bag, this has
|
// quantization channel. E.g. for a 2D embedding bag, this has
|
||||||
// [ row, col ] dimensions, for batched of embedding bags, dimensions might be
|
// [ row, col ] dimensions, for batched of embedding bags, dimensions might be
|
||||||
@ -293,16 +256,9 @@ Tensor& qembeddingbag_byte_prepack_out(
|
|||||||
auto* output_data = output.data_ptr<uint8_t>();
|
auto* output_data = output.data_ptr<uint8_t>();
|
||||||
|
|
||||||
#ifdef USE_FBGEMM
|
#ifdef USE_FBGEMM
|
||||||
// Move these outside of the ifdef when we support non-FBGEMM flow.
|
|
||||||
const auto is_valid_rowwise_min_max = _validate_rowwise_min_max(weight, rowwise_min_max_opt);
|
|
||||||
const auto rowwise_min_max_contig = _get_rowwise_min_max_contig(rowwise_min_max_opt);
|
|
||||||
|
|
||||||
if (weight_contig->scalar_type() == at::ScalarType::Half) {
|
if (weight_contig->scalar_type() == at::ScalarType::Half) {
|
||||||
const auto weight_data =
|
const auto weight_data =
|
||||||
static_cast<fbgemm::float16*>(weight_contig->data_ptr());
|
static_cast<fbgemm::float16*>(weight_contig->data_ptr());
|
||||||
const auto rowwise_min_max_data = is_valid_rowwise_min_max
|
|
||||||
? static_cast<fbgemm::float16*>(rowwise_min_max_contig->data_ptr())
|
|
||||||
: nullptr;
|
|
||||||
at::parallel_for(
|
at::parallel_for(
|
||||||
0, embedding_rows, 1, [&](int64_t start_idx, int64_t end_idx) {
|
0, embedding_rows, 1, [&](int64_t start_idx, int64_t end_idx) {
|
||||||
fbgemm::FloatOrHalfToFused8BitRowwiseQuantizedSBFloat<
|
fbgemm::FloatOrHalfToFused8BitRowwiseQuantizedSBFloat<
|
||||||
@ -310,21 +266,17 @@ Tensor& qembeddingbag_byte_prepack_out(
|
|||||||
weight_data + start_idx * embedding_cols,
|
weight_data + start_idx * embedding_cols,
|
||||||
end_idx - start_idx,
|
end_idx - start_idx,
|
||||||
embedding_cols,
|
embedding_cols,
|
||||||
output_data + start_idx * output_columns,
|
output_data + start_idx * output_columns);
|
||||||
(is_valid_rowwise_min_max ? (rowwise_min_max_data + start_idx * kRowwiseMinMaxNumCols) : nullptr));
|
|
||||||
});
|
});
|
||||||
} else {
|
} else {
|
||||||
const auto weight_data = weight_contig->data_ptr<float>();
|
const auto weight_data = weight_contig->data_ptr<float>();
|
||||||
const auto rowwise_min_max_data =
|
|
||||||
is_valid_rowwise_min_max ? rowwise_min_max_contig->data_ptr<float>() : nullptr;
|
|
||||||
at::parallel_for(
|
at::parallel_for(
|
||||||
0, embedding_rows, 1, [&](int64_t start_idx, int64_t end_idx) {
|
0, embedding_rows, 1, [&](int64_t start_idx, int64_t end_idx) {
|
||||||
fbgemm::FloatOrHalfToFused8BitRowwiseQuantizedSBFloat<float>(
|
fbgemm::FloatOrHalfToFused8BitRowwiseQuantizedSBFloat<float>(
|
||||||
weight_data + start_idx * embedding_cols,
|
weight_data + start_idx * embedding_cols,
|
||||||
end_idx - start_idx,
|
end_idx - start_idx,
|
||||||
embedding_cols,
|
embedding_cols,
|
||||||
output_data + start_idx * output_columns,
|
output_data + start_idx * output_columns);
|
||||||
(is_valid_rowwise_min_max ? (rowwise_min_max_data + start_idx * kRowwiseMinMaxNumCols) : nullptr));
|
|
||||||
});
|
});
|
||||||
}
|
}
|
||||||
|
|
||||||
@ -374,22 +326,6 @@ Tensor qembeddingbag_byte_prepack(const Tensor& weight) {
|
|||||||
return output;
|
return output;
|
||||||
}
|
}
|
||||||
|
|
||||||
static Tensor qembeddingbag_byte_prepack_with_rowwise_min_max(
|
|
||||||
const Tensor& weight,
|
|
||||||
const Tensor& rowwise_min_max) {
|
|
||||||
const auto weight_contig =
|
|
||||||
weight.expect_contiguous(weight.suggest_memory_format());
|
|
||||||
Tensor output = at::detail::empty_cpu(
|
|
||||||
{0},
|
|
||||||
at::kByte,
|
|
||||||
weight_contig->layout(),
|
|
||||||
weight_contig->device(),
|
|
||||||
std::nullopt,
|
|
||||||
std::nullopt);
|
|
||||||
qembeddingbag_byte_prepack_out(output, weight, rowwise_min_max);
|
|
||||||
return output;
|
|
||||||
}
|
|
||||||
|
|
||||||
Tensor qembeddingbag_byte_prepack_meta(const Tensor& weight) {
|
Tensor qembeddingbag_byte_prepack_meta(const Tensor& weight) {
|
||||||
const auto weight_contig =
|
const auto weight_contig =
|
||||||
weight.expect_contiguous(weight.suggest_memory_format());
|
weight.expect_contiguous(weight.suggest_memory_format());
|
||||||
@ -399,7 +335,7 @@ Tensor qembeddingbag_byte_prepack_meta(const Tensor& weight) {
|
|||||||
"'embedding_bag_byte_prepack' only support float32 or float16.");
|
"'embedding_bag_byte_prepack' only support float32 or float16.");
|
||||||
const auto weight_sizes = weight.sym_sizes();
|
const auto weight_sizes = weight.sym_sizes();
|
||||||
const auto cols_dim = weight.ndimension() - 1;
|
const auto cols_dim = weight.ndimension() - 1;
|
||||||
const auto& embedding_cols = weight_sizes[cols_dim];
|
const auto embedding_cols = weight_sizes[cols_dim];
|
||||||
// Add 8 bytes per column to store FP32 scale and zero_point per row.
|
// Add 8 bytes per column to store FP32 scale and zero_point per row.
|
||||||
const auto output_columns = embedding_cols + 2 * sizeof(float);
|
const auto output_columns = embedding_cols + 2 * sizeof(float);
|
||||||
|
|
||||||
@ -423,8 +359,7 @@ Tensor _qembeddingbag_nbit_prepack_helper(
|
|||||||
int bit_width,
|
int bit_width,
|
||||||
const bool optimized_qparams,
|
const bool optimized_qparams,
|
||||||
const int64_t nbins,
|
const int64_t nbins,
|
||||||
const double ratio,
|
const double ratio) {
|
||||||
const std::optional<Tensor>& rowwise_min_max_opt = std::nullopt) {
|
|
||||||
TORCH_CHECK(
|
TORCH_CHECK(
|
||||||
weight.scalar_type() == at::ScalarType::Float ||
|
weight.scalar_type() == at::ScalarType::Float ||
|
||||||
weight.scalar_type() == at::ScalarType::Half,
|
weight.scalar_type() == at::ScalarType::Half,
|
||||||
@ -466,17 +401,10 @@ Tensor _qembeddingbag_nbit_prepack_helper(
|
|||||||
auto* output_data = output.data_ptr<uint8_t>();
|
auto* output_data = output.data_ptr<uint8_t>();
|
||||||
|
|
||||||
#ifdef USE_FBGEMM
|
#ifdef USE_FBGEMM
|
||||||
// Move these outside of the ifdef when we support non-FBGEMM flow.
|
|
||||||
const auto is_valid_rowwise_min_max = _validate_rowwise_min_max(weight, rowwise_min_max_opt);
|
|
||||||
const auto rowwise_min_max_contig = _get_rowwise_min_max_contig(rowwise_min_max_opt);
|
|
||||||
|
|
||||||
if (!optimized_qparams) {
|
if (!optimized_qparams) {
|
||||||
if (weight_contig.scalar_type() == at::ScalarType::Half) {
|
if (weight_contig.scalar_type() == at::ScalarType::Half) {
|
||||||
const auto weight_data =
|
const auto weight_data =
|
||||||
static_cast<fbgemm::float16*>(weight_contig.data_ptr());
|
static_cast<fbgemm::float16*>(weight_contig.data_ptr());
|
||||||
const auto rowwise_min_max_data = is_valid_rowwise_min_max
|
|
||||||
? static_cast<fbgemm::float16*>(rowwise_min_max_contig->data_ptr())
|
|
||||||
: nullptr;
|
|
||||||
at::parallel_for(
|
at::parallel_for(
|
||||||
0, embedding_rows, 1, [&](int64_t start_idx, int64_t end_idx) {
|
0, embedding_rows, 1, [&](int64_t start_idx, int64_t end_idx) {
|
||||||
fbgemm::FloatOrHalfToFusedNBitRowwiseQuantizedSBHalf<
|
fbgemm::FloatOrHalfToFusedNBitRowwiseQuantizedSBHalf<
|
||||||
@ -485,13 +413,10 @@ Tensor _qembeddingbag_nbit_prepack_helper(
|
|||||||
weight_data + start_idx * embedding_cols,
|
weight_data + start_idx * embedding_cols,
|
||||||
end_idx - start_idx,
|
end_idx - start_idx,
|
||||||
static_cast<int>(embedding_cols),
|
static_cast<int>(embedding_cols),
|
||||||
output_data + start_idx * output_shape[1],
|
output_data + start_idx * output_shape[1]);
|
||||||
(is_valid_rowwise_min_max ? (rowwise_min_max_data + start_idx * kRowwiseMinMaxNumCols) : nullptr));
|
|
||||||
});
|
});
|
||||||
} else {
|
} else {
|
||||||
const auto weight_data = weight_contig.data_ptr<float>();
|
const auto weight_data = weight_contig.data_ptr<float>();
|
||||||
const auto rowwise_min_max_data =
|
|
||||||
is_valid_rowwise_min_max ? rowwise_min_max_contig->data_ptr<float>() : nullptr;
|
|
||||||
at::parallel_for(
|
at::parallel_for(
|
||||||
0, embedding_rows, 1, [&](int64_t start_idx, int64_t end_idx) {
|
0, embedding_rows, 1, [&](int64_t start_idx, int64_t end_idx) {
|
||||||
fbgemm::FloatOrHalfToFusedNBitRowwiseQuantizedSBHalf<float>(
|
fbgemm::FloatOrHalfToFusedNBitRowwiseQuantizedSBHalf<float>(
|
||||||
@ -499,8 +424,7 @@ Tensor _qembeddingbag_nbit_prepack_helper(
|
|||||||
weight_data + start_idx * embedding_cols,
|
weight_data + start_idx * embedding_cols,
|
||||||
end_idx - start_idx,
|
end_idx - start_idx,
|
||||||
static_cast<int>(embedding_cols),
|
static_cast<int>(embedding_cols),
|
||||||
output_data + start_idx * output_shape[1],
|
output_data + start_idx * output_shape[1]);
|
||||||
(is_valid_rowwise_min_max ? (rowwise_min_max_data + start_idx * kRowwiseMinMaxNumCols) : nullptr));
|
|
||||||
});
|
});
|
||||||
}
|
}
|
||||||
} else {
|
} else {
|
||||||
@ -590,16 +514,6 @@ Tensor qembeddingbag_4bit_prepack(
|
|||||||
weight, 4 /*bit_width*/, optimized_qparams, nbins, ratio);
|
weight, 4 /*bit_width*/, optimized_qparams, nbins, ratio);
|
||||||
}
|
}
|
||||||
|
|
||||||
Tensor qembeddingbag_4bit_prepack_with_rowwise_min_max(
|
|
||||||
const Tensor& weight,
|
|
||||||
const Tensor& rowwise_min_max,
|
|
||||||
const bool optimized_qparams,
|
|
||||||
const int64_t nbins,
|
|
||||||
const double ratio) {
|
|
||||||
return _qembeddingbag_nbit_prepack_helper(
|
|
||||||
weight, 4 /*bit_width*/, optimized_qparams, nbins, ratio, rowwise_min_max);
|
|
||||||
}
|
|
||||||
|
|
||||||
// Applies 2-bit row-wise quantization by determining the range
|
// Applies 2-bit row-wise quantization by determining the range
|
||||||
// (maximum - minimum) and bias (minimum value) of each row in the input
|
// (maximum - minimum) and bias (minimum value) of each row in the input
|
||||||
// matrix, and then scaling each element to an 2-bit number between 0 and
|
// matrix, and then scaling each element to an 2-bit number between 0 and
|
||||||
@ -617,16 +531,6 @@ Tensor qembeddingbag_2bit_prepack(
|
|||||||
weight, 2 /*bit_width*/, optimized_qparams, nbins, ratio);
|
weight, 2 /*bit_width*/, optimized_qparams, nbins, ratio);
|
||||||
}
|
}
|
||||||
|
|
||||||
Tensor qembeddingbag_2bit_prepack_with_rowwise_min_max(
|
|
||||||
const Tensor& weight,
|
|
||||||
const Tensor& rowwise_min_max,
|
|
||||||
const bool optimized_qparams,
|
|
||||||
const int64_t nbins,
|
|
||||||
const double ratio) {
|
|
||||||
return _qembeddingbag_nbit_prepack_helper(
|
|
||||||
weight, 2 /*bit_width*/, optimized_qparams, nbins, ratio, rowwise_min_max);
|
|
||||||
}
|
|
||||||
|
|
||||||
class QEmbeddingPackWeights final {
|
class QEmbeddingPackWeights final {
|
||||||
public:
|
public:
|
||||||
static c10::intrusive_ptr<EmbeddingPackedParamsBase> run(const at::Tensor& weight) {
|
static c10::intrusive_ptr<EmbeddingPackedParamsBase> run(const at::Tensor& weight) {
|
||||||
@ -638,21 +542,12 @@ TORCH_LIBRARY_IMPL(quantized, CPU, m) {
|
|||||||
m.impl(
|
m.impl(
|
||||||
TORCH_SELECTIVE_NAME("quantized::embedding_bag_byte_prepack"),
|
TORCH_SELECTIVE_NAME("quantized::embedding_bag_byte_prepack"),
|
||||||
TORCH_FN(qembeddingbag_byte_prepack));
|
TORCH_FN(qembeddingbag_byte_prepack));
|
||||||
m.impl(
|
|
||||||
TORCH_SELECTIVE_NAME("quantized::embedding_bag_byte_prepack_with_rowwise_min_max"),
|
|
||||||
TORCH_FN(qembeddingbag_byte_prepack_with_rowwise_min_max));
|
|
||||||
m.impl(
|
m.impl(
|
||||||
TORCH_SELECTIVE_NAME("quantized::embedding_bag_4bit_prepack"),
|
TORCH_SELECTIVE_NAME("quantized::embedding_bag_4bit_prepack"),
|
||||||
TORCH_FN(qembeddingbag_4bit_prepack));
|
TORCH_FN(qembeddingbag_4bit_prepack));
|
||||||
m.impl(
|
|
||||||
TORCH_SELECTIVE_NAME("quantized::embedding_bag_4bit_prepack_with_rowwise_min_max"),
|
|
||||||
TORCH_FN(qembeddingbag_4bit_prepack_with_rowwise_min_max));
|
|
||||||
m.impl(
|
m.impl(
|
||||||
TORCH_SELECTIVE_NAME("quantized::embedding_bag_2bit_prepack"),
|
TORCH_SELECTIVE_NAME("quantized::embedding_bag_2bit_prepack"),
|
||||||
TORCH_FN(qembeddingbag_2bit_prepack));
|
TORCH_FN(qembeddingbag_2bit_prepack));
|
||||||
m.impl(
|
|
||||||
TORCH_SELECTIVE_NAME("quantized::embedding_bag_2bit_prepack_with_rowwise_min_max"),
|
|
||||||
TORCH_FN(qembeddingbag_2bit_prepack_with_rowwise_min_max));
|
|
||||||
}
|
}
|
||||||
|
|
||||||
TORCH_LIBRARY_IMPL(quantized, QuantizedCPU, m) {
|
TORCH_LIBRARY_IMPL(quantized, QuantizedCPU, m) {
|
||||||
|
|||||||
@ -3,10 +3,7 @@
|
|||||||
|
|
||||||
namespace at::native {
|
namespace at::native {
|
||||||
|
|
||||||
Tensor& qembeddingbag_byte_prepack_out(
|
Tensor& qembeddingbag_byte_prepack_out(Tensor& output, const Tensor& weight);
|
||||||
Tensor& output,
|
|
||||||
const Tensor& weight,
|
|
||||||
const std::optional<Tensor>& rowwise_min_max_opt = std::nullopt);
|
|
||||||
|
|
||||||
Tensor qembeddingbag_byte_prepack(const Tensor& weight);
|
Tensor qembeddingbag_byte_prepack(const Tensor& weight);
|
||||||
|
|
||||||
|
|||||||
@ -121,12 +121,9 @@ TORCH_LIBRARY(quantized, m) {
|
|||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_unpack(__torch__.torch.classes.quantized.EmbeddingPackedParamsBase W_prepack) -> Tensor W_origin"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_unpack(__torch__.torch.classes.quantized.EmbeddingPackedParamsBase W_prepack) -> Tensor W_origin"), {at::Tag::pt2_compliant_tag});
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_byte_prepack(Tensor weight) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_byte_prepack(Tensor weight) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_byte_unpack(Tensor weight) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_byte_unpack(Tensor weight) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_byte_prepack_with_rowwise_min_max(Tensor weight, Tensor rowwise_min_max) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_4bit_prepack(Tensor weight, bool optimized_qparams=False, int nbins=200, float ratio=0.16) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_4bit_prepack(Tensor weight, bool optimized_qparams=False, int nbins=200, float ratio=0.16) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_4bit_prepack_with_rowwise_min_max(Tensor weight, Tensor rowwise_min_max, bool optimized_qparams=False, int nbins=200, float ratio=0.16) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_4bit_unpack(Tensor weight) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_4bit_unpack(Tensor weight) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_2bit_prepack(Tensor weight, bool optimized_qparams=False, int nbins=200, float ratio=0.16) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_2bit_prepack(Tensor weight, bool optimized_qparams=False, int nbins=200, float ratio=0.16) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_2bit_prepack_with_rowwise_min_max(Tensor weight, Tensor rowwise_min_max, bool optimized_qparams=False, int nbins=200, float ratio=0.16) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_2bit_unpack(Tensor weight) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_2bit_unpack(Tensor weight) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_byte_rowwise_offsets(Tensor weight, Tensor indices, Tensor? offsets=None, bool scale_grad_by_freq=False, int mode=0, bool pruned_weights=False, Tensor? per_sample_weights=None, Tensor? compressed_indices_mapping=None, bool include_last_offset=False) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_byte_rowwise_offsets(Tensor weight, Tensor indices, Tensor? offsets=None, bool scale_grad_by_freq=False, int mode=0, bool pruned_weights=False, Tensor? per_sample_weights=None, Tensor? compressed_indices_mapping=None, bool include_last_offset=False) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
||||||
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_4bit_rowwise_offsets(Tensor weight, Tensor indices, Tensor? offsets=None, bool scale_grad_by_freq=False, int mode=0, bool pruned_weights=False, Tensor? per_sample_weights=None, Tensor? compressed_indices_mapping=None, bool include_last_offset=False) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
m.def(TORCH_SELECTIVE_SCHEMA("quantized::embedding_bag_4bit_rowwise_offsets(Tensor weight, Tensor indices, Tensor? offsets=None, bool scale_grad_by_freq=False, int mode=0, bool pruned_weights=False, Tensor? per_sample_weights=None, Tensor? compressed_indices_mapping=None, bool include_last_offset=False) -> Tensor"), {at::Tag::pt2_compliant_tag});
|
||||||
|
|||||||
Some files were not shown because too many files have changed in this diff Show More
Reference in New Issue
Block a user