求助：基于32位ARM9架构用Bazel构建TensorFlow Wheel文件

一、准备ARM9交叉编译工具链 #

• 下载适配ARM9（ARMv5TE指令集）的32位交叉编译工具链，比如arm-linux-gnueabihf工具链，解压到本地路径（例：/opt/arm-linux-gnueabihf）。
• 设置环境变量绑定工具链：

  export CC=/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-gcc
  export CXX=/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-g++
  export AR=/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-ar
  export LD=/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-ld
  

二、配置TensorFlow源码与Bazel工具链 #

1. 克隆TensorFlow稳定版本源码（推荐2.15.x，新版本已移除32位ARM支持）：

   git clone https://github.com/tensorflow/tensorflow.git -b v2.15.0
   

2. 运行配置脚本，按以下选项配置：
   • 选择交叉编译模式，指定目标架构为arm，目标CPU为arm926ej-s
   • 禁用CUDA、XLA、MKL等非必需加速模块
   • 指定与ARM9设备匹配的Python版本路径
3. 编写Bazel工具链配置文件arm_toolchain.bzl：

   def _arm_toolchain_impl(ctx):
       toolchain = cc_common.create_cc_toolchain_config_info(
           ctx = ctx,
           toolchain_identifier = "arm-linux-gnueabihf",
           host_system_name = "x86_64-linux-gnu",
           target_system_name = "arm-linux-gnueabihf",
           target_cpu = "arm",
           target_libc = "gnu",
           compiler = "gcc",
           abi_version = "gnu",
           abi_libc_version = "gnu",
           tool_paths = {
               "gcc": "/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-gcc",
               "g++": "/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-g++",
               "ar": "/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-ar",
               "ld": "/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-ld",
               "nm": "/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-nm",
               "objcopy": "/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-objcopy",
               "objdump": "/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-objdump",
               "strip": "/opt/arm-linux-gnueabihf/bin/arm-linux-gnueabihf-strip",
           },
           cxx_builtin_include_directories = [
               "/opt/arm-linux-gnueabihf/include",
               "/opt/arm-linux-gnueabihf/lib/gcc/arm-linux-gnueabihf/9.3.0/include",
           ],
       )
       return [toolchain]
   
   arm_toolchain = rule(
       implementation = _arm_toolchain_impl,
       attrs = {},
       provides = [CcToolchainConfigInfo],
   )
   

4. 在WORKSPACE文件中注册工具链：

   load("//:arm_toolchain.bzl", "arm_toolchain")
   
   arm_toolchain(name = "arm_linux_toolchain")
   register_toolchains("//:arm_linux_toolchain")
   

三、编译并生成Wheel文件 #

• 执行Bazel编译命令，指定ARM9专属编译参数：

  bazel build --config=arm --copt=-march=armv5te --copt=-mtune=arm926ej-s --copt=-mfloat-abi=hard --copt=-mfpu=vfp2 //tensorflow/tools/pip_package:build_pip_package
  

• 生成Wheel包：

  ./bazel-bin/tensorflow/tools/pip_package/build_pip_package /tmp/tensorflow_wheel
  

四、适配ARM9环境 #

• 用auditwheel（需32位Python环境运行）修复Wheel的兼容性标签：

  auditwheel repair /tmp/tensorflow_wheel/tensorflow-*.whl --plat manylinux_2_17_armv5l -w /tmp/fixed_wheel
  

• 将修复后的Wheel拷贝到ARM9设备，使用对应版本的pip安装即可。

关键注意事项 #

• 必须选择TensorFlow 2.15.x及以下版本，新版本已彻底放弃32位ARM支持
• 编译过程需充足内存（建议8G以上，或配置swap分区）
• 禁用所有非必需功能可大幅降低编译复杂度与失败概率

内容的提问来源于stack exchange，提问作者sai manohar