Metal backend: exported `.pte` does not load outside a process that already has PyTorch's libomp (embedded AOTI object links `/opt/llvm-openmp/lib/libomp.dylib`)
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21日。
评估
- 难度
- 4/5
- 预计耗时
- 3-5 天
- 新手友好度
- 48/100
- Issue 类型
- 缺陷
- 描述清晰度
- 基本清楚
- 活跃度
- 活跃
调研方向
Start with metal_backend.cpp around line 352 and the MetalBackend.codesign_so path, then inspect run_metal_test.sh's existing library-reference rewrite. Reproduce export and loading outside executor_runner, and determine whether the accepted fix should remove unused dependencies or reject unsupported ones at export time. Done means the chosen behavior is verified for an app without PyTorch's libomp.
由索引模型根据 Issue 内容生成。
描述
Summary
A Metal .pte loads under executor_runner built by run_metal_test.sh, but fails in an application that links the ExecuTorch runtime on its own:
E executorch:metal_backend.cpp:352] Failed to load shared library: dlopen(.../<hash>_so_blob<pid>.so, 0x0005):
Library not loaded: /opt/llvm-openmp/lib/libomp.dylib
E executorch:method.cpp:132] Init failed for backend MetalBackend: 0x22
Cause
On macOS inductor's _get_openmp_args always adds -lomp. The libomp it resolves is the one bundled in the PyTorch wheel (torch/lib/libomp.dylib), whose LC_ID_DYLIB is the absolute path /opt/llvm-openmp/lib/libomp.dylib, and that path is recorded as a dependency of the AOTI object embedded in the .pte. The path exists on no end-user machine.
It goes unnoticed in-tree because run_metal_test.sh already rewrites the same reference in executor_runner to @rpath/libomp.dylib with an rpath into torch/lib. The runner therefore loads PyTorch's libomp first, and dyld then satisfies the model's dependency by install name. An app that ships its own libomp (or none) has no image with that install name loaded, so dlopen fails.
The dependency is not used: nm -u on the embedded object shows no omp / kmp symbol for a Metal-delegated MobileNet, as expected for a graph that runs on the GPU.
Possible directions
- Link the AOTI object with
-Wl,-dead_strip_dylibsfor the Metal backend, so dylibs the object does not reference are dropped and libomp stays only if a model really needs it. Inductor has no option for extra link flags today; we do it from outside by pointingtorch._inductor.config.cpp.cxxat a wrapper compiler that appends the flag on link steps. The resulting object links onlylibc++andlibSystemand loads in a process with no libomp at all. - Independently, a check next to
MetalBackend.codesign_sothat the compiled object has no dependency outside/usr/liband/Systemwould turn this from a load-time surprise on someone else's machine into an export-time error.
Happy to send a PR for the check if that direction sounds right; the link-flag part probably belongs in inductor.
- 主要语言
- Python
- 星标
- 5k
- 派生
- 1.2k
- 平均合并
- 2 天 9 小时
- 30 天内合并 PR
- 555
环境准备
从这里开始
- 先读完整个 Issue，再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库，在一个分支上完成修改。
- 提交 Pull Request，并在描述里引用这个 Issue 编号。
pytorch/executorch 的其他 Issue
-
enhancement triaged
难度 2/5 半天 新手友好度 68/100
pytorch/executorch#21640 ·
维护者通常 1 天内回复
-
enhancement module: examples
难度 5/5 一周以上 新手友好度 20/100
pytorch/executorch#23164 · 7 条评论 · 1 个 reaction ·
维护者通常 1 天内回复
-
Qualcomm: 8-bit per-channel weight scales are floored at the 16-bit eps, and the HTP miscomputes near-zero channels可能已有人在做 @psiddh 于 2 天前认领。 未关闭module: qnn partner: qualcomm
pytorch/executorch#23160 · 1 条评论 · 已指派 1 人 ·
维护者通常 1 天内回复
-
[cpu kernels] native_layer_norm: layer_norm_scalar returns NaN on large-mean rows; Half/BF16 at N>=256 slow after #23153可能已有人在做 @JakeStevens 于 2 天前认领。 未关闭module: kernels
pytorch/executorch#23159 · 2 条评论 · 已指派 1 人 ·
维护者通常 1 天内回复
-
module: vulkan
难度 3/5 1-2 天 新手友好度 66/100
pytorch/executorch#23158 ·
维护者通常 1 天内回复
查看 pytorch/executorch 的全部 Issue
相似的 Issue
-
docs pydanty:is-working
难度 2/5 1-3 小时 新手友好度 75/100
pydantic/pydantic-ai#8863 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 68/100
run-llama/llama_index#23278 ·
维护者通常 2 天内回复
-
documentation from-review-extraction github-actions priority: low severity:nit
难度 1/5 1 小时以内 新手友好度 92/100
LearningCircuit/local-deep-research#6946 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 82/100
oracle/langchain-oracle#323 ·
维护者通常 1 天内回复
-
难度 1/5 1 小时以内 新手友好度 88/100
tenstorrent/tt-metal#58057 · 1 条评论 ·
维护者通常 1 天内回复