M1 MacBook下CocoaPod无法同时集成LibTorch-Lite与TensorFlowLiteSwift

问题背景 #

我在为自研CocoaPod同时添加以下两个依赖时遇到冲突：单独安装任意一个依赖均可正常工作，但无法同时集成：

s.dependency 'TensorFlowLiteSwift'
s.dependency 'LibTorch-Lite'

LibTorch-Lite需要配置特定的spec参数才能避免架构编译错误：

s.static_framework = true

s.pod_target_xcconfig = {
      'HEADER_SEARCH_PATHS' => '$(inherited) ${PODS_ROOT}/LibTorch-Lite/install/include',
      'EXCLUDED_ARCHS[sdk=iphonesimulator*]' => 'arm64'
  }
  s.user_target_xcconfig = {
      'EXCLUDED_ARCHS[sdk=iphonesimulator*]' => 'arm64'
  }

不配置上述参数会触发以下错误：

libtorch.a(empty.cpp.o), building for iOS Simulator, but linking in object file built for iOS, file '/var/folders/3x/x73_p70n2tg0gnwq9lct7rmr0000gn/T/CocoaPods-Lint-20221219-7300-1w8h64c-PyTorchTensorFlowCocoaPod/Pods/LibTorch-Lite/install/lib/libtorch.a' for architecture arm64

但配置EXCLUDED_ARCHS[sdk=iphonesimulator*]为arm64后，集成TensorFlowLiteSwift时又会出现模块找不到的错误：

main.swift:1:8: error: could not find module 'PyTorchTensorFlowCocoaPod' for target 'arm64-apple-ios-simulator'; found: x86_64-apple-ios-simulator, at: /Users/steve.ham/Library/Developer/Xcode/DerivedData/App-dnxlikhybfsytahdpnmquhblzyly/Build/Products/Release-iphonesimulator/PyTorchTensorFlowCocoaPod/PyTorchTensorFlowCocoaPod.framework/Modules/PyTorchTensorFlowCocoaPod.swiftmodule

解决方案 #

1. 仅为LibTorch-Lite单独配置架构排除规则 #

不要在全局的user_target_xcconfig中设置EXCLUDED_ARCHS，仅将该规则限定在LibTorch-Lite的Pod目标范围内，避免影响TensorFlowLiteSwift和主Pod的架构编译。修改后的spec配置如下：

s.static_framework = true

# 仅为LibTorch-Lite的Pod目标配置架构排除，不影响全局用户目标
s.pod_target_xcconfig = {
    'HEADER_SEARCH_PATHS' => '$(inherited) ${PODS_ROOT}/LibTorch-Lite/install/include',
    'EXCLUDED_ARCHS[sdk=iphonesimulator*]' => 'arm64'
}

# 移除user_target_xcconfig中的EXCLUDED_ARCHS配置

2. 为主Pod添加多架构编译支持 #

确保主Pod的spec明确支持模拟器的arm64和x86_64架构，让Xcode生成通用架构的产物：

# 在主Pod的spec中添加以下配置
s.pod_target_xcconfig = {
    'VALID_ARCHS[sdk=iphonesimulator*]' => 'arm64 x86_64',
    'ARCHS[sdk=iphonesimulator*]' => 'arm64 x86_64'
}

3. 清理缓存并重新安装依赖 #

执行以下命令清除CocoaPods和Xcode的编译缓存，避免旧缓存导致的冲突：

# 清理所有Pod缓存
pod cache clean --all
# 删除Pod锁定文件和Pods目录
rm -rf Podfile.lock Pods
# 重新安装依赖
pod install
# 清理Xcode派生数据
rm -rf ~/Library/Developer/Xcode/DerivedData

4. 升级TensorFlowLiteSwift版本 #

确认使用的TensorFlowLiteSwift版本支持arm64模拟器架构，较新版本已修复相关兼容性问题，建议升级到最新稳定版。

内容的提问来源于stack exchange，提问作者Steve Ham