spack/var/spack/repos/builtin/packages/py-tensorflow/package.py
Andrew W Elble fd087107ea
py-tensorflow: fix linking with ubuntu's gcc (#45437)
gcc on ubuntu has fix-cortex-a53-843419 set by default - this causes linking
issues (symbol relocation errors) for tf, even when compiling for different
cpus.
2024-09-25 15:19:58 +02:00

902 lines
40 KiB
Python

# Copyright 2013-2024 Lawrence Livermore National Security, LLC and other
# Spack Project Developers. See the top-level COPYRIGHT file for details.
#
# SPDX-License-Identifier: (Apache-2.0 OR MIT)
import glob
import os
import sys
import tempfile
from spack.build_environment import optimization_flags
from spack.package import *
rocm_dependencies = [
"hip",
"rocrand",
"rocblas",
"rocfft",
"hipfft",
"rccl",
"hipsparse",
"rocprim",
"llvm-amdgpu",
"hsa-rocr-dev",
"rocminfo",
"hipsolver",
"hiprand",
"rocsolver",
"hipsolver",
"hipblas",
"hipcub",
"rocm-core",
"roctracer-dev",
"miopen-hip",
]
class PyTensorflow(Package, CudaPackage, ROCmPackage, PythonExtension):
"""TensorFlow is an open source machine learning framework for everyone."""
homepage = "https://www.tensorflow.org"
url = "https://github.com/tensorflow/tensorflow/archive/v2.3.1.tar.gz"
git = "https://github.com/tensorflow/tensorflow.git"
maintainers("adamjstewart", "aweits")
import_modules = ["tensorflow"]
license("Apache-2.0")
version("2.17.0", sha256="9cc4d5773b8ee910079baaecb4086d0c28939f024dd74b33fc5e64779b6533dc")
version("2.16.2", sha256="023849bf253080cb1e4f09386f5eb900492da2288274086ed6cfecd6d99da9eb")
version("2.16.1", sha256="c729e56efc945c6df08efe5c9f5b8b89329c7c91b8f40ad2bb3e13900bd4876d")
version(
"2.16.1-rocm-enhanced",
sha256="e1b63b1b5d5b014194ed33113c7fa7f26ecb8d36333282b8c550e795e0eb31c6",
url="https://github.com/ROCm/tensorflow-upstream/archive/refs/tags/v2.16.1-rocm-enhanced.tar.gz",
)
version("2.15.1", sha256="f36416d831f06fe866e149c7cd752da410a11178b01ff5620e9f265511ed57cf")
version("2.15.0", sha256="9cec5acb0ecf2d47b16891f8bc5bc6fbfdffe1700bdadc0d9ebe27ea34f0c220")
version("2.14.1", sha256="6b31ed347ed7a03c45b906aa41628ac91c3db7c84cb816971400d470e58ba494")
version(
"2.14-rocm-enhanced",
git="https://github.com/ROCm/tensorflow-upstream.git",
branch="r2.14-rocm-enhanced-nohipblaslt-build",
)
version("2.14.0", sha256="ce357fd0728f0d1b0831d1653f475591662ec5bca736a94ff789e6b1944df19f")
version("2.13.1", sha256="89c07aebd4f41fbe0d08cc88aef00305542134f2f16d3b62918dc3c1182f33e2")
version("2.13.0", sha256="e58c939079588623e6fa1d054aec2f90f95018266e0a970fd353a5244f5173dc")
version("2.12.1", sha256="6bc4600cc0b88e9e40f1800096f5bddbbd3b6e5527a030dea631b87f2ae46b5b")
version("2.12.0", sha256="c030cb1905bff1d2446615992aad8d8d85cbe90c4fb625cee458c63bf466bc8e")
version("2.11.1", sha256="624ed1cc170cdcc19e8a15d8cdde989a9a1c6b0534c90b38a6b2f06fb2963e5f")
version(
"2.11.0-rocm-enhanced",
sha256="0c4ee8d83bc72215cbc1a5cd3e88cde1a9cf7304237d3e3d8d105ff09827d903",
url="https://github.com/ROCmSoftwarePlatform/tensorflow-upstream/archive/refs/tags/v2.11.0-rocm-enhanced.tar.gz",
)
version("2.11.0", sha256="99c732b92b1b37fc243a559e02f9aef5671771e272758aa4aec7f34dc92dac48")
version("2.10.1", sha256="622a92e22e6f3f4300ea43b3025a0b6122f1cc0e2d9233235e4c628c331a94a3")
version("2.10.0", sha256="b5a1bb04c84b6fe1538377e5a1f649bb5d5f0b2e3625a3c526ff3a8af88633e8")
version("2.9.3", sha256="59d09bd00eef6f07477eea2f50778582edd4b7b2850a396f1fd0c646b357a573")
version("2.9.2", sha256="8cd7ed82b096dc349764c3369331751e870d39c86e73bbb5374e1664a59dcdf7")
version("2.9.1", sha256="6eaf86ead73e23988fe192da1db68f4d3828bcdd0f3a9dc195935e339c95dbdc")
version("2.9.0", sha256="8087cb0c529f04a4bfe480e49925cd64a904ad16d8ec66b98e2aacdfd53c80ff")
version("2.8.4", sha256="c08a222792bdbff9da299c7885561ee27b95d414d1111c426efac4ccdce92cde")
version("2.8.3", sha256="4b7ecbe50b36887e1615bc2a582cb86df1250004d8bb540e18336d539803b5a7")
version("2.8.2", sha256="b3f860c02c22a30e9787e2548ca252ab289a76b7778af6e9fa763d4aafd904c7")
version("2.8.1", sha256="4b487a63d6f0c1ca46a2ac37ba4687eabdc3a260c222616fa414f6df73228cec")
version("2.8.0", sha256="66b953ae7fba61fd78969a2e24e350b26ec116cf2e6a7eb93d02c63939c6f9f7")
version(
"2.7.4-rocm-enhanced",
sha256="45b79c125edfdc008274f1b150d8b5a53b3ff4713fd1ad1ff4738f515aad8191",
url="https://github.com/ROCmSoftwarePlatform/tensorflow-upstream/archive/refs/tags/v2.7.4-rocm-enhanced.tar.gz",
)
version("2.7.4", sha256="75b2e40a9623df32da16d8e97528f5e02e4a958e23b1f2ee9637be8eec5d021b")
version("2.7.3", sha256="b576c2e124cd6d4d04cbfe985430a0d955614e882172b2258217f0ec9b61f39b")
version("2.7.2", sha256="b3c8577f3b7cc82368ff7f9315821d506abd2f716ea6692977d255b7d8bc54c0")
version("2.7.1", sha256="abebe2cf5ca379e18071693ca5f45b88ade941b16258a21cc1f12d77d5387a21")
version("2.7.0", sha256="bb124905c7fdacd81e7c842b287c169bbf377d29c74c9dacc04f96c9793747bb")
version("2.6.5", sha256="305da42845ac584a42494e521c92a88ce92ee47d93022d4c0bb45180b5c19a8c")
version("2.6.4", sha256="6a9e54f46039ef0a6f0a1adf19befa510044d3203d1e124dba8318ec4b1e0210")
version("2.6.3", sha256="7a71dde0987677b9512b202eb6ae119e0e308b1ea15b66dcfce001a44873997b")
version("2.6.2", sha256="e68c1d346fc3d529653530ca346b2c62f5b31bd4fcca7ffc9c65bb39ab2f6ed3")
version("2.6.1", sha256="8e457f617bc2eb43de2a51900e7922b60a8107e2524b2576438f1acccee1d043")
version("2.6.0", sha256="41b32eeaddcbc02b0583660bcf508469550e4cd0f86b22d2abe72dfebeacde0f")
version("2.5.3", sha256="58d69b7163f7624debc243750976d27fa7dddbc6fb7c5215aec94732bcc670e1")
version("2.5.2", sha256="bcccc6ba0b8ac1d10d3302f766eed71911acecc0bc43d0bd27d97a1e7ce275a8")
version("2.5.1", sha256="8d2728e155a3aa6befd9cb3d0980fabd25e2142d124f8f6b6c78cdf17ff79da5")
version("2.5.0", sha256="233875ea27fc357f6b714b2a0de5f6ff124b50c1ee9b3b41f9e726e9e677b86c")
version("2.4.4", sha256="f1abc3ed92c3ce955db2a7db5ec422a3a98f015331183194f97b99fe77a09bb4")
version("2.4.3", sha256="cafd520c753f8755a9eb1262932f685dc722d8658f08373f8ec88d8acd58d7d4")
version("2.4.2", sha256="edc88da97277906513d53eeee57997a2036fa32ac1f1937730301764fa06cdc0")
version("2.4.1", sha256="f681331f8fc0800883761c7709d13cda11942d4ad5ff9f44ad855e9dc78387e0")
version("2.4.0", sha256="26c833b7e1873936379e810a39d14700281125257ddda8cd822c89111db6f6ae")
version("2.3.4", sha256="195947838b0918c15d79bc6ed85ff714b24d6d564b4d07ba3de0b745a2f9b656")
version("2.3.3", sha256="b91e5bcd373b942c4a62c6bcb7ff6f968b1448152b82f54a95dfb0d8fb9c6093")
version("2.3.2", sha256="21a703d2e68cd0677f6f9ce329198c24fd8203125599d791af9f1de61aadf31f")
version("2.3.2", sha256="21a703d2e68cd0677f6f9ce329198c24fd8203125599d791af9f1de61aadf31f")
version("2.3.1", sha256="ee534dd31a811f7a759453567257d1e643f216d8d55a25c32d2fbfff8153a1ac")
version("2.3.0", sha256="2595a5c401521f20a2734c4e5d54120996f8391f00bb62a57267d930bce95350")
version("2.2.3", sha256="5e6c779ca8392864d436d88893461dcce783c3a8d46dcb2b2f2ee8ece3cc4538")
version("2.2.2", sha256="fb4b5d26c5b983350f7ce8297b71176a86a69e91faf66e6ebb1e58538ad3bb51")
version("2.2.1", sha256="e6a28e64236d729e598dbeaa02152219e67d0ac94d6ed22438606026a02e0f88")
version("2.2.0", sha256="69cd836f87b8c53506c4f706f655d423270f5a563b76dc1cfa60fbc3184185a3")
depends_on("c", type="build") # generated
depends_on("cxx", type="build") # generated
variant("mkl", default=False, description="Build with MKL support")
variant("jemalloc", default=False, description="Build with jemalloc as malloc support")
variant("gcp", default=False, description="Build with Google Cloud Platform support")
variant("hdfs", default=False, description="Build with Hadoop File System support")
variant("aws", default=False, description="Build with Amazon AWS Platform support")
variant("xla", default=sys.platform != "darwin", description="Build with XLA JIT support")
variant("gdr", default=False, description="Build with GDR support")
variant("verbs", default=False, description="Build with libverbs support")
variant("ngraph", default=False, description="Build with Intel nGraph support")
variant("opencl", default=False, description="Build with OpenCL SYCL support")
variant("computecpp", default=False, description="Build with ComputeCPP support")
variant(
"tensorrt", default=False, description="Build with TensorRT support"
) # TODO: enable when TensorRT in Spack
variant("cuda", default=sys.platform != "darwin", description="Build with CUDA support")
variant(
"nccl", default=sys.platform.startswith("linux"), description="Enable NVIDIA NCCL support"
)
variant("mpi", default=False, description="Build with MPI support")
variant("android", default=False, description="Configure for Android builds")
variant("ios", default=False, description="Build with iOS support (macOS only)")
variant("monolithic", default=False, description="Static monolithic build")
variant("numa", default=False, description="Build with NUMA support")
variant(
"dynamic_kernels",
default=sys.platform.startswith("linux"),
description="Build kernels into separate shared objects",
)
extends("python")
with default_args(type="build"):
# See .bazelversion
depends_on("bazel@6.5.0", when="@2.16:")
depends_on("bazel@6.1.0", when="@2.14:2.15")
depends_on("bazel@5.3.0", when="@2.11:2.13")
depends_on("bazel@5.1.1", when="@2.10")
# See _TF_MIN_BAZEL_VERSION and _TF_MAX_BAZEL_VERSION in configure.py
depends_on("bazel@4.2.2:5.99.0", when="@2.9")
depends_on("bazel@4.2.1:4.99.0", when="@2.8")
depends_on("bazel@3.7.2:4.99.0", when="@2.7")
depends_on("bazel@3.7.2:3.99.0", when="@2.5:2.6")
depends_on("bazel@3.1.0:3.99.0", when="@2.3:2.4")
depends_on("bazel@2.0.0", when="@2.2")
# tensorflow/tools/pip_package/build_pip_package.sh
depends_on("patchelf", when="@2.13: platform=linux")
# https://github.com/tensorflow/tensorflow/issues/60179#issuecomment-1491238631
depends_on("coreutils", when="@2.13: platform=darwin")
depends_on("swig")
depends_on("py-pip")
depends_on("py-wheel")
with default_args(type=("build", "run")):
# Python support based on wheel availability
depends_on("python@3.9:3.12", when="@2.16:")
depends_on("python@3.9:3.11", when="@2.14:2.15")
depends_on("python@3.8:3.11", when="@2.12:2.13")
depends_on("python@:3.10", when="@2.8:2.11")
depends_on("python@:3.9", when="@2.5:2.7")
depends_on("python@:3.8", when="@2.2:2.4")
# Listed under REQUIRED_PACKAGES in tensorflow/tools/pip_package/setup.py
depends_on("py-absl-py@1:", when="@2.9:")
depends_on("py-absl-py@0.4:", when="@2.7:2.8")
depends_on("py-absl-py@0.10:0", when="@2.4:2.6")
depends_on("py-absl-py@0.7:", when="@:2.3")
depends_on("py-astunparse@1.6:", when="@2.7:")
depends_on("py-astunparse@1.6.3:1.6", when="@2.4:2.6")
depends_on("py-astunparse@1.6.3", when="@2.2:2.3")
depends_on("py-flatbuffers@24.3.25:", when="@2.17:")
depends_on("py-flatbuffers@23.5.26:", when="@2.14:")
depends_on("py-flatbuffers@23.1.21:", when="@2.13")
depends_on("py-flatbuffers@2:", when="@2.10:2.12")
depends_on("py-flatbuffers@1.12:1", when="@2.9")
depends_on("py-flatbuffers@1.12:", when="@2.8")
depends_on("py-flatbuffers@1.12:2", when="@2.7")
depends_on("py-flatbuffers@1.12", when="@2.4:2.6")
depends_on("py-gast@0.2.1:0.4,0.5.3:", when="@2.14:")
depends_on("py-gast@0.2.1:0.4.0", when="@2.9:2.13")
depends_on("py-gast@0.2.1:", when="@2.8")
depends_on("py-gast@0.2.1:0.4", when="@2.7")
depends_on("py-gast@0.4.0", when="@2.5:2.6")
depends_on("py-gast@0.3.3", when="@2.2:2.4")
depends_on("py-gast@0.2.2", when="@:2.1")
depends_on("py-google-pasta@0.1.1:", when="@2.7:")
depends_on("py-google-pasta@0.2:0", when="@2.4:2.6")
depends_on("py-google-pasta@0.1.8:", when="@2.2:2.3")
depends_on("py-google-pasta@0.1.6:", when="@:2.1")
depends_on("py-h5py@3.10:", when="@2.16:")
depends_on("py-h5py@2.9:", when="@2.7:2.15")
depends_on("py-h5py@3.1", when="@2.5:2.6")
depends_on("py-h5py@2.10", when="@2.2:2.4")
depends_on("py-h5py@:2.10.0", when="@2.1.3:2.1")
# propagate the mpi variant setting for h5py/hdf5 to avoid unexpected crashes
depends_on("py-h5py+mpi", when="@2.1.3:+mpi")
depends_on("py-h5py~mpi", when="@2.1.3:~mpi")
depends_on("hdf5+mpi", when="@2.1.3:+mpi")
depends_on("hdf5~mpi", when="@2.1.3:~mpi")
depends_on("py-libclang@13:", when="@2.9:")
depends_on("py-libclang@9.0.1:", when="@2.7:2.8")
depends_on("py-ml-dtypes@0.3.1:0.4", when="@2.17:")
depends_on("py-ml-dtypes@0.3.1:0.3", when="@2.15.1:2.16")
depends_on("py-ml-dtypes@0.2", when="@2.15.0")
depends_on("py-ml-dtypes@0.2.0", when="@2.14")
depends_on("py-numpy@1.23.5:", when="@2.14:")
depends_on("py-numpy@1.22:1.24.3", when="@2.13:")
depends_on("py-numpy@1.22:1.23", when="@2.12")
depends_on("py-numpy@1.20:", when="@2.8:2.11")
depends_on("py-numpy@1.14.5:", when="@2.7")
depends_on("py-numpy@1.19.2:1.19", when="@2.4:2.6")
# https://github.com/tensorflow/tensorflow/issues/40688
depends_on("py-numpy@1.16.0:1.18", when="@:2.3")
# https://github.com/tensorflow/tensorflow/issues/67291
depends_on("py-numpy@:1")
depends_on("py-opt-einsum@2.3.2:", when="@:2.3,2.7:")
depends_on("py-opt-einsum@3.3", when="@2.4:2.6")
depends_on("py-packaging", when="@2.9:")
depends_on("py-protobuf@3.20.3:4.20,4.21.6:4", when="@2.12:")
depends_on("py-protobuf@3.9.2:", when="@2.3:2.11")
depends_on("py-protobuf@3.8.0:", when="@:2.2")
# https://github.com/protocolbuffers/protobuf/issues/10051
# https://github.com/tensorflow/tensorflow/issues/56266
depends_on("py-protobuf@:3.19", when="@:2.11")
depends_on("py-requests@2.21:2", when="@2.16:")
depends_on("py-requests")
depends_on("py-setuptools")
depends_on("py-six@1.12:", when="@:2.3,2.7:")
depends_on("py-six@1.15", when="@2.4:2.6")
depends_on("py-termcolor@1.1:", when="@:2.3,2.7:")
depends_on("py-termcolor@1.1", when="@2.4:2.6")
depends_on("py-typing-extensions@3.6.6:", when="@2.7:2.12,2.14:")
depends_on("py-typing-extensions@3.6.6:4.5", when="@2.13")
depends_on("py-typing-extensions@3.7.4:3.7", when="@2.4:2.6")
depends_on("py-wrapt@1.11:", when="@2.7:2.11,2.13,2.16:")
depends_on("py-wrapt@1.11:1.14", when="@2.12,2.14:2.15")
depends_on("py-wrapt@1.12.1:1.12", when="@2.4:2.6")
depends_on("py-wrapt@1.11.1:", when="@:2.3")
# TODO: add packages for these dependencies
# depends_on('py-tensorflow-io-gcs-filesystem@0.23.1:', when='@2.8:')
# depends_on('py-tensorflow-io-gcs-filesystem@0.21:', when='@2.7')
if sys.byteorder == "little":
# Only builds correctly on little-endian machines
depends_on("py-grpcio@1.24.3:1", when="@2.7:")
depends_on("py-grpcio@1.37.0:1", when="@2.6")
depends_on("py-grpcio@1.34", when="@2.5")
depends_on("py-grpcio@1.32", when="@2.4")
depends_on("py-grpcio@1.8.6:", when="@:2.3")
for minor_ver in range(2, 18):
depends_on("py-tensorboard@2.{}".format(minor_ver), when="@2.{}".format(minor_ver))
# TODO: support circular run-time dependencies
# depends_on('py-tensorflow-estimator')
# depends_on('py-keras')
# Historical dependencies
depends_on("py-jax@0.3.15:", when="@2.12")
depends_on("py-keras-preprocessing@1.1.1:", when="@2.7:2.10")
depends_on("py-keras-preprocessing@1.1.2:1.1", when="@2.4:2.6")
depends_on("py-keras-preprocessing@1.1.1:1.1", when="@2.3")
depends_on("py-keras-preprocessing@1.1:", when="@2.2")
depends_on("py-scipy@1.4.1", when="@2.2.0,2.3.0")
depends_on("py-wheel@0.32:0", when="@2.7")
depends_on("py-wheel@0.35:0", when="@2.4:2.6")
depends_on("py-wheel@0.26:", when="@:2.3")
# TODO: add packages for some of these dependencies
depends_on("mkl", when="+mkl")
depends_on("curl", when="+gcp")
# depends_on('computecpp', when='+opencl+computecpp')
# depends_on('trisycl', when='+opencl~computepp')
with when("+cuda"):
# https://www.tensorflow.org/install/source#gpu
depends_on("cuda@12.3:", when="@2.16:")
depends_on("cuda@12.2:", when="@2.15:")
depends_on("cuda@11.8:", when="@2.12:")
depends_on("cuda@11.2:", when="@2.5:")
depends_on("cuda@11.0:", when="@2.4:")
depends_on("cuda@10.1:", when="@2.1:")
depends_on("cuda@:11.7.0", when="@:2.9")
depends_on("cuda@:11.4", when="@2.4:2.7")
depends_on("cuda@:10.2", when="@:2.3")
depends_on("cudnn@8.9:8", when="@2.15:")
depends_on("cudnn@8.7:8", when="@2.14:")
depends_on("cudnn@8.6:8", when="@2.12:")
depends_on("cudnn@8.1:8", when="@2.5:")
depends_on("cudnn@8.0:8", when="@2.4:")
depends_on("cudnn@7.6:8", when="@2.1:")
depends_on("cudnn@:7", when="@:2.2")
# depends_on('tensorrt', when='+tensorrt')
depends_on("nccl", when="+nccl+cuda")
depends_on("mpi", when="+mpi")
# depends_on('android-ndk@10:18', when='+android')
# depends_on('android-sdk', when='+android')
with when("+rocm"):
for pkg_dep in rocm_dependencies:
depends_on(f"{pkg_dep}@6.0:", when="@2.14:")
depends_on(pkg_dep)
# Check configure and configure.py to see when these variants are supported
conflicts("+mkl", when="platform=darwin", msg="Darwin is not yet supported")
conflicts(
"+jemalloc",
when="platform=darwin",
msg="Currently jemalloc is only support on Linux platform",
)
conflicts("+opencl", when="platform=windows")
conflicts("+computecpp", when="~opencl")
conflicts(
"+cuda",
when="+rocm",
msg="CUDA / ROCm are mututally exclusive. At most 1 GPU platform can be configured",
)
conflicts("+cuda", when="platform=darwin", msg="There is no GPU support for macOS")
conflicts(
"cuda_arch=none",
when="+cuda",
msg="Must specify CUDA compute capabilities of your GPU, see https://developer.nvidia.com/cuda-gpus",
)
conflicts("cuda_arch=20", msg="TensorFlow only supports compute capabilities >= 3.5")
conflicts("cuda_arch=30", msg="TensorFlow only supports compute capabilities >= 3.5")
conflicts("cuda_arch=32", msg="TensorFlow only supports compute capabilities >= 3.5")
conflicts("+tensorrt", when="~cuda")
conflicts(
"+tensorrt",
when="platform=darwin",
msg="Currently TensorRT is only supported on Linux platform",
)
conflicts("+nccl", when="~cuda~rocm")
conflicts(
"+nccl", when="platform=darwin", msg="Currently NCCL is only supported on Linux platform"
)
conflicts("+mpi", when="platform=windows")
conflicts("+ios", when="platform=linux", msg="iOS support only available on macOS")
# https://github.com/tensorflow/tensorflow/pull/45404
conflicts("platform=darwin target=aarch64:", when="@:2.4")
# https://github.com/tensorflow/tensorflow/pull/39225
conflicts("target=aarch64:", when="@:2.2")
conflicts(
"~rocm",
when="@2.7.4-rocm-enhanced,2.11.0-rocm-enhanced,2.14-rocm-enhanced,2.16.1-rocm-enhanced",
)
conflicts("+rocm", when="@:2.7.4-a,2.7.4.0:2.11.0-a,2.11.0.0:2.14-a,2.14-z:2.16.1-a,2.16.1-z:")
# wheel 0.40 upgrades vendored packaging, trips over tensorflow-io-gcs-filesystem identifier
conflicts("^py-wheel@0.40:", when="@2.11:2.13")
# Must be matching versions of py-protobuf and protobuf
conflicts("^py-protobuf~cpp")
# https://www.tensorflow.org/install/source#tested_build_configurations
# https://github.com/tensorflow/tensorflow/issues/70199
# (-mavx512fp16 exists in gcc@12:)
conflicts("%gcc@13:", when="@:2.14")
conflicts("%gcc@:11", when="@2.17:")
conflicts("%gcc@:9.3.0", when="@2.9:")
conflicts("%gcc@:7.3.0")
# zlib is vendored and downloaded directly from zlib.org (or mirrors), but
# old downloads are removed from that site immediately after a new release.
# If the tf mirrors don't work, make sure the fallback is to something existing.
patch(
"https://github.com/tensorflow/tensorflow/commit/76b9fa22857148a562f3d9b5af6843402a93c15b.patch?full_index=1",
sha256="f9e26c544da729cfd376dbd3b096030e3777d3592459add1f3c78b1b9828d493",
when="@2.9:2.10.0",
)
# Version 2.10 produces an error related to cuBLAS:
# E tensorflow/stream_executor/cuda/cuda_blas.cc:2981] Unable to register
# cuBLAS factory: Attempting to register factory for plugin cuBLAS when one
# has already been registered
# See https://github.com/tensorflow/tensorflow/issues/57663
# This is fixed for 2.11 but 2.10 needs the following patch.
patch(
"https://github.com/tensorflow/tensorflow/pull/56691.patch?full_index=1",
sha256="d635ea6d6c1571505871d0caba3e2cd939ea0f4aff972095d552913a8109def3",
when="@2.10",
)
# needed for protobuf 3.16+
patch("example_parsing.patch", when="@:2.7 ^protobuf@3.16:")
# allow linker to be found in PATH
# https://github.com/tensorflow/tensorflow/issues/39263
patch("null_linker_bin_path.patch", when="@2.5:")
# Reset import order to that of 2.4. Part of
# https://bugs.gentoo.org/800824#c3 From the patch:
# When tensorflow and python protobuf use the same instance of libprotobuf,
# pywrap_tensorflow must be imported before anything else that would import
# protobuf definitions.
patch("0008-Fix-protobuf-errors-when-using-system-protobuf.patch", when="@2.5:2.6")
# see https://github.com/tensorflow/tensorflow/issues/62490
# and https://github.com/abseil/abseil-cpp/issues/1665
patch("absl_neon.patch", when="@2.16.1: target=aarch64:")
# reverting change otherwise the c467913 commit patch won't apply
patch(
"https://github.com/ROCm/tensorflow-upstream/commit/fd6b0a4356c66f5f30cedbc62b24f18d9e32806f.patch?full_index=1",
sha256="43f1519dfc618b4fb568f760d559c063234248fa12c47a35c1cf3b7114756424",
when="@2.16.1-rocm-enhanced +rocm",
reverse=True,
)
patch(
"https://github.com/ROCm/tensorflow-upstream/commit/c467913bf4411ce2681391f37a9adf6031d23c2c.patch?full_index=1",
sha256="82554a84d19d99180a6bec274c6106dd217361e809b446e2e4bc4b6b979bdf7a",
when="@2.16.1-rocm-enhanced +rocm",
)
patch(
"https://github.com/ROCm/tensorflow-upstream/commit/f4f4e8698b90755b0b5ea2d9da1933b0b988b111.patch?full_index=1",
sha256="a4c0fd62a0af3ba113c8933fa531dd17fa6667e507202a144715cd87fbdaf476",
when="@2.16.1-rocm-enhanced: +rocm",
)
patch(
"https://github.com/ROCm/tensorflow-upstream/commit/8b7fcccb2914078737689347540cb79ace579bbb.patch?full_index=1",
sha256="75a61a79ce3aae51fda920f677f4dc045374b20e25628626eb37ca19c3a3b4c4",
when="@2.16.1-rocm-enhanced +rocm",
)
phases = ["configure", "build", "install"]
def flag_handler(self, name, flags):
spec = self.spec
# ubuntu gcc has this workaround turned on by default in aarch64
# and it causes issues with symbol relocation during link
# note, archspec doesn't currently ever report cortex_a53!
if (
name == "ldflags"
and spec.target.family == "aarch64"
and "ubuntu" in spec.os
and spec.compiler.name == "gcc"
and "cortex_a53" not in spec.target.name
):
flags.append("-mno-fix-cortex-a53-843419")
return (flags, None, None)
# https://www.tensorflow.org/install/source
def setup_build_environment(self, env):
spec = self.spec
# Please specify the location of python
env.set("PYTHON_BIN_PATH", python.path)
# Please input the desired Python library path to use
env.set("PYTHON_LIB_PATH", python_platlib)
env.set("TF_PYTHON_VERSION", spec["python"].version.up_to(2))
# Ensure swig is in PATH or set SWIG_PATH
env.set("SWIG_PATH", spec["swig"].prefix.bin.swig)
# Do you wish to build TensorFlow with MKL support?
if "+mkl" in spec:
env.set("TF_NEED_MKL", "1")
# Do you wish to download MKL LIB from the web?
env.set("TF_DOWNLOAD_MKL", "0")
# Please specify the location where MKL is installed
env.set("MKL_INSTALL_PATH", spec["mkl"].prefix)
else:
env.set("TF_NEED_MKL", "0")
# Do you wish to build TensorFlow with jemalloc as malloc support?
if "+jemalloc" in spec:
env.set("TF_NEED_JEMALLOC", "1")
else:
env.set("TF_NEED_JEMALLOC", "0")
# Do you wish to build TensorFlow with Google Cloud Platform support?
if "+gcp" in spec:
env.set("TF_NEED_GCP", "1")
else:
env.set("TF_NEED_GCP", "0")
# Do you wish to build TensorFlow with Hadoop File System support?
if "+hdfs" in spec:
env.set("TF_NEED_HDFS", "1")
else:
env.set("TF_NEED_HDFS", "0")
# Do you wish to build TensorFlow with Amazon AWS Platform support?
if "+aws" in spec:
env.set("TF_NEED_AWS", "1")
env.set("TF_NEED_S3", "1")
else:
env.set("TF_NEED_AWS", "0")
env.set("TF_NEED_S3", "0")
# Do you wish to build TensorFlow with XLA JIT support?
if "+xla" in spec:
env.set("TF_ENABLE_XLA", "1")
else:
env.set("TF_ENABLE_XLA", "0")
# Do you wish to build TensorFlow with GDR support?
if "+gdr" in spec:
env.set("TF_NEED_GDR", "1")
else:
env.set("TF_NEED_GDR", "0")
# Do you wish to build TensorFlow with VERBS support?
if "+verbs" in spec:
env.set("TF_NEED_VERBS", "1")
else:
env.set("TF_NEED_VERBS", "0")
# Do you wish to build TensorFlow with nGraph support?
if "+ngraph" in spec:
env.set("TF_NEED_NGRAPH", "1")
else:
env.set("TF_NEED_NGRAPH", "0")
# Do you wish to build TensorFlow with OpenCL SYCL support?
if "+opencl" in spec:
env.set("TF_NEED_OPENCL_SYCL", "1")
env.set("TF_NEED_OPENCL", "1")
# Please specify which C++ compiler should be used as the host
# C++ compiler
env.set("HOST_CXX_COMPILER", spack_cxx)
# Please specify which C compiler should be used as the host
# C compiler
env.set("HOST_C_COMPILER", spack_cc)
# Do you wish to build TensorFlow with ComputeCPP support?
if "+computecpp" in spec:
env.set("TF_NEED_COMPUTECPP", "1")
# Please specify the location where ComputeCpp is installed
env.set("COMPUTECPP_TOOLKIT_PATH", spec["computecpp"].prefix)
else:
env.set("TF_NEED_COMPUTECPP", "0")
# Please specify the location of the triSYCL include directory
env.set("TRISYCL_INCLUDE_DIR", spec["trisycl"].prefix.include)
else:
env.set("TF_NEED_OPENCL_SYCL", "0")
env.set("TF_NEED_OPENCL", "0")
# Do you wish to build TensorFlow with ROCm support?
if "+rocm" in spec:
env.set("TF_NEED_ROCM", "1")
env.set("TF_HIPBLASLT", "0")
env.set("MIOPEN_PATH", spec["miopen-hip"].prefix)
env.set("ROCTRACER_PATH", spec["roctracer-dev"].prefix)
env.set("LLVM_PATH", spec["llvm-amdgpu"].prefix)
for pkg_dep in rocm_dependencies:
pkg_dep_cap = pkg_dep.upper().replace("-", "_")
env.set(f"{pkg_dep_cap}_PATH", spec[pkg_dep].prefix)
env.set("TF_ROCM_AMDGPU_TARGETS", ",".join(self.spec.variants["amdgpu_target"].value))
else:
env.set("TF_NEED_ROCM", "0")
# Do you wish to build TensorFlow with CUDA support?
if "+cuda" in spec:
env.set("TF_NEED_CUDA", "1")
# Do you want to use clang as CUDA compiler?
env.set("TF_CUDA_CLANG", "0")
# Please specify which gcc nvcc should use as the host compiler
env.set("GCC_HOST_COMPILER_PATH", spack_cc)
cuda_paths = [spec["cuda"].prefix, spec["cudnn"].prefix]
# Do you wish to build TensorFlow with TensorRT support?
if "+tensorrt" in spec:
env.set("TF_NEED_TENSORRT", "1")
cuda_paths.append(spec["tensorrt"].prefix)
# Please specify the TensorRT version you want to use
env.set("TF_TENSORRT_VERSION", spec["tensorrt"].version.up_to(1))
# Please specify the location where TensorRT is installed
env.set("TENSORRT_INSTALL_PATH", spec["tensorrt"].prefix)
else:
env.set("TF_NEED_TENSORRT", "0")
env.unset("TF_TENSORRT_VERSION")
# Please specify the CUDA SDK version you want to use
env.set("TF_CUDA_VERSION", spec["cuda"].version.up_to(2))
# Please specify the cuDNN version you want to use
env.set("TF_CUDNN_VERSION", spec["cudnn"].version.up_to(1))
if "+nccl" in spec:
cuda_paths.append(spec["nccl"].prefix)
# Please specify the locally installed NCCL version to use
env.set("TF_NCCL_VERSION", spec["nccl"].version.up_to(1))
# Please specify the location where NCCL is installed
env.set("NCCL_INSTALL_PATH", spec["nccl"].prefix)
env.set("NCCL_HDR_PATH", spec["nccl"].prefix.include)
else:
env.unset("TF_NCCL_VERSION")
# Please specify the comma-separated list of base paths to
# look for CUDA libraries and headers
env.set("TF_CUDA_PATHS", ",".join(cuda_paths))
# Please specify the location where CUDA toolkit is installed
env.set("CUDA_TOOLKIT_PATH", spec["cuda"].prefix)
# Please specify the location where CUDNN library is installed
env.set("CUDNN_INSTALL_PATH", spec["cudnn"].prefix)
# Please specify a list of comma-separated CUDA compute
# capabilities you want to build with. You can find the compute
# capability of your device at:
# https://developer.nvidia.com/cuda-gpus.
# Please note that each additional compute capability significantly
# increases your build time and binary size, and that TensorFlow
# only supports compute capabilities >= 3.5
capabilities = CudaPackage.compute_capabilities(spec.variants["cuda_arch"].value)
env.set("TF_CUDA_COMPUTE_CAPABILITIES", ",".join(capabilities))
else:
env.set("TF_NEED_CUDA", "0")
# Do you want to use Clang to build TensorFlow?
if "%clang" in spec:
env.set("TF_NEED_CLANG", "1")
else:
env.set("TF_NEED_CLANG", "0")
# Do you wish to download a fresh release of clang? (Experimental)
env.set("TF_DOWNLOAD_CLANG", "0")
# Do you wish to build TensorFlow with MPI support?
if "+mpi" in spec:
env.set("TF_NEED_MPI", "1")
# Please specify the MPI toolkit folder
env.set("MPI_HOME", spec["mpi"].prefix)
else:
env.set("TF_NEED_MPI", "0")
env.unset("MPI_HOME")
# Please specify optimization flags to use during compilation when
# bazel option '--config=opt' is specified
env.set("CC_OPT_FLAGS", optimization_flags(self.compiler, spec.target))
# Would you like to interactively configure ./WORKSPACE for
# Android builds?
if "+android" in spec:
env.set("TF_SET_ANDROID_WORKSPACE", "1")
# Please specify the home path of the Android NDK to use
env.set("ANDROID_NDK_HOME", spec["android-ndk"].prefix)
env.set("ANDROID_NDK_API_LEVEL", spec["android-ndk"].version)
# Please specify the home path of the Android SDK to use
env.set("ANDROID_SDK_HOME", spec["android-sdk"].prefix)
env.set("ANDROID_SDK_API_LEVEL", spec["android-sdk"].version)
# Please specify the Android SDK API level to use
env.set("ANDROID_API_LEVEL", spec["android-sdk"].version)
# Please specify an Android build tools version to use
env.set("ANDROID_BUILD_TOOLS_VERSION", spec["android-sdk"].version)
else:
env.set("TF_SET_ANDROID_WORKSPACE", "0")
# Do you wish to build TensorFlow with iOS support?
if "+ios" in spec:
env.set("TF_CONFIGURE_IOS", "1")
else:
env.set("TF_CONFIGURE_IOS", "0")
# set tmpdir to a non-NFS filesystem
# (because bazel uses ~/.cache/bazel)
# TODO: This should be checked for non-nfsy filesystem, but the current
# best idea for it is to check
# subprocess.call([
# 'stat', '--file-system', '--format=%T', tmp_path
# ])
# to not be nfs. This is only valid for Linux and we'd like to
# stay at least also OSX compatible
tmp_path = tempfile.mkdtemp(prefix="spack")
env.set("TEST_TMPDIR", tmp_path)
def configure(self, spec, prefix):
# NOTE: configure script is interactive. If you set the appropriate
# environment variables, this interactivity is skipped. If you don't,
# Spack hangs during the configure phase. Use `spack build-env` to
# determine which environment variables must be set for a particular
# version.
configure()
@run_after("configure")
def post_configure_fixes(self):
spec = self.spec
if spec.satisfies("@2.17:"):
filter_file(
"patchelf",
spec["patchelf"].prefix.bin.patchelf,
"tensorflow/tools/pip_package/build_pip_package.py",
string=True,
)
# make sure xla is actually turned off
if spec.satisfies("~xla"):
filter_file(
r"--define with_xla_support=true",
r"--define with_xla_support=false",
".tf_configure.bazelrc",
)
if spec.satisfies("~android"):
# env variable is somehow ignored -> brute force
# TODO: find a better solution
filter_file(r"if workspace_has_any_android_rule\(\)", r"if True", "configure.py")
if spec.satisfies("~gcp"):
# google cloud support seems to be installed on default, leading
# to boringssl error manually set the flag to false to avoid
# installing gcp support
# https://github.com/tensorflow/tensorflow/issues/20677#issuecomment-404634519
filter_file(
r"--define with_gcp_support=true",
r"--define with_gcp_support=false",
".tf_configure.bazelrc",
)
if spec.satisfies("~opencl"):
# 1.8.0 and 1.9.0 aborts with numpy import error during python_api
# generation somehow the wrong PYTHONPATH is used...
# set --distinct_host_configuration=false as a workaround
# https://github.com/tensorflow/tensorflow/issues/22395#issuecomment-431229451
with open(".tf_configure.bazelrc", mode="a") as f:
f.write("build --distinct_host_configuration=false\n")
f.write('build --action_env PYTHONPATH="{0}"\n'.format(env["PYTHONPATH"]))
if spec.satisfies("+cuda"):
libs = spec["cuda"].libs.directories
libs.extend(spec["cudnn"].libs.directories)
if "+nccl" in spec:
libs.extend(spec["nccl"].libs.directories)
if "+tensorrt" in spec:
libs.extend(spec["tensorrt"].libs.directories)
slibs = ":".join(libs)
with open(".tf_configure.bazelrc", mode="a") as f:
f.write('build --action_env LD_LIBRARY_PATH="' + slibs + '"')
if spec.satisfies("@2.16.1-rocm-enhanced +rocm"):
if os.path.exists(spec["llvm-amdgpu"].prefix.bin.clang):
filter_file(
"/usr/lib/llvm-17/bin/clang", spec["llvm-amdgpu"].prefix.bin.clang, ".bazelrc"
)
else:
filter_file(
"/usr/lib/llvm-17/bin/clang",
spec["llvm-amdgpu"].prefix.llvm.bin.clang,
".bazelrc",
)
filter_file("build:opt --copt=-march=native", "", ".tf_configure.bazelrc")
filter_file("build:opt --host_copt=-march=native", "", ".tf_configure.bazelrc")
def build(self, spec, prefix):
# Bazel needs the directory to exist on install
mkdirp(python_platlib)
tmp_path = env["TEST_TMPDIR"]
# https://docs.bazel.build/versions/master/command-line-reference.html
args = [
# Don't allow user or system .bazelrc to override build settings
"--nohome_rc",
"--nosystem_rc",
# Bazel does not work properly on NFS, switch to /tmp
"--output_user_root=" + tmp_path,
"build",
# Spack logs don't handle colored output well
"--color=no",
"--jobs={0}".format(make_jobs),
"--config=opt",
# Enable verbose output for failures
"--verbose_failures",
]
if spec.satisfies("^bazel@:3.5"):
# removed in bazel 3.6
args.append("--incompatible_no_support_tools_in_action_inputs=false")
# See .bazelrc for when each config flag is supported
if "+mkl" in spec:
args.append("--config=mkl")
if "+monolithic" in spec:
args.append("--config=monolithic")
if "+gdr" in spec:
args.append("--config=gdr")
if "+verbs" in spec:
args.append("--config=verbs")
if "+ngraph" in spec:
args.append("--config=ngraph")
if "+dynamic_kernels" in spec:
args.append("--config=dynamic_kernels")
if "+cuda" in spec:
args.append("--config=cuda")
if "+rocm" in spec:
args.append("--config=rocm")
if "~aws" in spec:
args.append("--config=noaws")
if "~gcp" in spec:
args.append("--config=nogcp")
if "~hdfs" in spec:
args.append("--config=nohdfs")
if "~nccl" in spec:
args.append("--config=nonccl")
# https://github.com/tensorflow/tensorflow/issues/63080
if self.spec.satisfies("@2.14:"):
args.append(f"--define=with_numa_support={'+numa' in spec}")
else:
if "+numa" in spec:
args.append("--config=numa")
args.append("--config=v2")
# https://github.com/tensorflow/tensorflow/issues/63298
if self.spec.satisfies("@2.17:"):
args.append("//tensorflow/tools/pip_package:wheel")
else:
args.append("//tensorflow/tools/pip_package:build_pip_package")
bazel(*args)
if self.spec.satisfies("@:2.16"):
build_pip_package = Executable(
"bazel-bin/tensorflow/tools/pip_package/build_pip_package"
)
buildpath = join_path(self.stage.source_path, "spack-build")
build_pip_package("--src", buildpath)
def install(self, spec, prefix):
tmp_path = env["TEST_TMPDIR"]
if self.spec.satisfies("@2.17:"):
buildpath = join_path(
self.stage.source_path, "bazel-bin/tensorflow/tools/pip_package/wheel_house/"
)
with working_dir(buildpath):
wheel = glob.glob("*.whl")[0]
args = std_pip_args + ["--prefix=" + prefix, wheel]
pip(*args)
else:
buildpath = join_path(self.stage.source_path, "spack-build")
with working_dir(buildpath):
args = std_pip_args + ["--prefix=" + prefix, "."]
pip(*args)
remove_linked_tree(tmp_path)