Repository navigation
Build the ggml CPU engine on Apple Silicon - #25359
Conversation
Summary: The ggml CPU backend now builds for Apple Silicon (previous diff), so `CPU_COMPAT` allows arm64 macOS as well as x86_64 Linux. The targets that link the ggml CPU engine without torch, its unit test, `ggml_ptn_runner`, `run_hf_llm_ggml_cpu` and `run_classification_ggml_cpu`, now build on a Mac. The ggml sample tests, `test_ggml_ops` and `test_cross_engine` also run torch, through the sample generator or the test itself, so they take the new `TORCH_COMPAT` and stay x86_64 Linux. The native runtime's package reader uses `nlohmann_json`, whose fbcode external dep does not build for macOS. fbcode now resolves `nlohmann_json` to `fbsource//third-party/nlohmann-json`, the header-only target xplat already uses, as it does for gtest and flatbuffers. Authored with Claude Code. Reviewed By: digantdesai Differential Revision: D133538765
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/25359
Note: Links to docs will display an error until the docs builds have been completed. ⏳ 1 Pending, 1 Unrelated FailureAs of commit e7eea11 with merge base ba0d826 ( BROKEN TRUNK - The following job failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
|
@JakeStevens has exported this pull request. If you are a Meta employee, you can view the originating Diff in D133538765. |
digantdesai
left a comment
There was a problem hiding this comment.
Review automatically exported from Phabricator review in Meta.
This PR needs a
|
Summary:
The ggml CPU backend now builds for Apple Silicon (previous diff), so
CPU_COMPATallows arm64 macOS as well as x86_64 Linux. The targets that link the ggml CPU engine without torch, its unit test,ggml_ptn_runner,run_hf_llm_ggml_cpuandrun_classification_ggml_cpu, now build on a Mac. The ggml sample tests,test_ggml_opsandtest_cross_enginealso run torch, through the sample generator or the test itself, so they take the newTORCH_COMPATand stay x86_64 Linux.The native runtime's package reader uses
nlohmann_json, whose fbcode external dep does not build for macOS. fbcode now resolvesnlohmann_jsontofbsource//third-party/nlohmann-json, the header-only target xplat already uses, as it does for gtest and flatbuffers.Authored with Claude Code.
Reviewed By: digantdesai
Differential Revision: D133538765