Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

[iOS] Split the image processor into its own SwiftPM product instead of bundling it in executorch

未关闭
  • 哪里不对？
#22,722 0 条评论 0 个 reaction 已指派 1 人 在 GitHub 查看

维护者通常 1 天内回复

@metascroy 已经在做这个了。

开始于 2026年9月11日。

评估

这个 Issue 还没有评估数据。

描述

module: build/install module: ios
Problem

The image processor is bundled into the executorch xcframework, so the core runtime
product ends up carrying that helper's Apple framework dependencies.

scripts/build_apple_frameworks.sh merges nine static libraries into a single
executorch.xcframework, and libextension_image.a is one of them. Because it is merged,
a SwiftPM manifest cannot attach anything to that part alone: the smallest thing it can
describe is the whole product. So the frameworks the image processor needs have to be
declared on executorch itself.

That is what #22711 does, and it is the only place the current packaging allows. It reads,
though, as if the ExecuTorch runtime depends on Accelerate and Core Image, which it does
not. Every one of those symbols comes from the image processor:

image_processor_apple.o          8 undefined symbols   (vImage, from Accelerate)
image_processor_apple_gpu.o      8 undefined symbols   (Core Image)

No other object file in the shipped library references either framework.

There is also a size argument. The image processor is about 93 KB of a 967 KB library, so
roughly 10 percent, and every app that only wants Module and Tensor links it anyway.

Proposal

Give the image processor its own product, the way executorch_llm and executorch_dump
already work. Both build a separate xcframework and layer on the runtime rather than being
merged into it:

"executorch_llm": [
  "targets": ["executorch"],
],

An executorch_image product would follow that shape and own its dependencies:

"executorch_image": [
  "frameworks": ["Accelerate", "CoreGraphics", "CoreImage", "CoreVideo", "Foundation"],
  "targets": ["executorch"],
],

executorch then goes back to declaring only c++, and apps that do not process images
stop linking the helper at all.

What it involves
  1. Build a separate executorch_image.xcframework in scripts/build_apple_frameworks.sh,
    holding libextension_image.a.
  2. Add the product above to Package.swift, and remove the framework block from
    executorch. The same change is needed in Package.swift.template on the swiftpm
    branch, which is what prebuilt users actually get.
  3. Move the bindings, ExecuTorchImageProcessor.h/.mm and
    ExecuTorch+ImageProcessor.swift, out of extension/apple into the new product, and
    drop #import "ExecuTorchImageProcessor.h" from the ExecuTorch.h umbrella header.

Step 3 is a breaking change. Today import ExecuTorch gives you the image processor. After
the split it needs a second import, so anyone using it since the bindings landed has to add
a line. Worth calling out in release notes.

Nothing in the runtime depends on extension_image; the only reference is the extension
list in the top-level CMakeLists.txt. So there is no dependency to untangle beyond the
bindings.

Test plan
  • swift package --disable-sandbox dump-package shows executorch with only c++, and
    executorch_image with the five frameworks.
  • Build a consumer that depends on executorch alone and does not use the image processor,
    and confirm it links with no Apple frameworks declared.
  • Build a consumer that depends on executorch_image and uses the image processor, and
    confirm it links and runs on macOS, iOS device and the iOS simulator.
  • Run the iOS framework and demo app CI.

cc @shoumikhin @larryliu0820 @GregoryComer @cbilgin

主要语言
Python
星标
5k
派生
1.2k
平均合并
2 天 5 小时
30 天内合并 PR
491

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

pytorch/executorch 的其他 Issue

查看 pytorch/executorch 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。