Fix: ci.py crash on macOS from duplicate libomp load#520
Open
ChaoWao wants to merge 1 commit intohw-native-sys:mainfrom
Open
Fix: ci.py crash on macOS from duplicate libomp load#520ChaoWao wants to merge 1 commit intohw-native-sys:mainfrom
ChaoWao wants to merge 1 commit intohw-native-sys:mainfrom
Conversation
On macOS, `python ci.py -p a2a3sim` (or a5sim) aborts every task with "OMP: Error hw-native-sys#15: Initializing libomp.dylib, but found libomp.dylib already initialized" (SIGABRT) before any DeviceRunner code runs. Two distinct libomp.dylib copies get mapped into the single CI process: homebrew's /opt/homebrew/opt/libomp/lib/libomp.dylib (via numpy -> openblas) and pip torch's .venv/.../torch/lib/libomp.dylib. They have different install names, so dyld loads them both and Intel's libomp aborts on the second init. Surfaced after hw-native-sys#493 collapsed sim CI into one long-lived Python process; each golden's `import numpy`/`import torch` now accumulates conflicting libomps in the same address space. - Set KMP_DUPLICATE_LIB_OK=TRUE at the top of ci.py on darwin, before any import that can transitively pull in numpy or torch. This is Intel's documented escape hatch; safe for our workload where numpy and torch are only used for golden reference math, not parallel OMP regions. - Document the full root cause, debugging steps, and explicit "what not to do" list in docs/macos-libomp-collision.md so future contributors don't re-investigate. Link it from docs/ci.md. - Rewrite the two remaining numpy-based goldens (a2a3/{aicpu,host}_build_graph/bgemm) in torch for style consistency with the rest of examples/. Note this does not avoid the libomp collision on its own -- `import torch` transitively imports numpy. Verified: `python ci.py` passes 32/32 sim tests (20 a2a3sim + 12 a5sim) on macOS without KMP_DUPLICATE_LIB_OK needing to be set manually.
There was a problem hiding this comment.
Code Review
This pull request introduces a workaround for a libomp collision issue on macOS that causes SIGABRT when both numpy and torch are loaded in the same process. The fix involves setting KMP_DUPLICATE_LIB_OK=TRUE at the top of ci.py before other imports. Detailed documentation explaining the root cause and mitigation has been added in docs/macos-libomp-collision.md and referenced in docs/ci.md. Additionally, several golden reference scripts were updated to use torch instead of numpy for input generation and computation. I have no feedback to provide as there were no review comments to evaluate.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
On macOS,
python ci.py -p a2a3sim(ora5sim) aborts every task withOMP: Error #15: Initializing libomp.dylib, but found libomp.dylib already initialized(SIGABRT) before any DeviceRunner code runs.Root cause: Two distinct
libomp.dylibcopies get mapped into the single CI process:/opt/homebrew/opt/libomp/lib/libomp.dylibpulled in bynumpy → openblas.venv/.../torch/lib/libomp.dylibThey have different install names, so dyld loads both and Intel's libomp aborts on the second init. Surfaced after #493 collapsed sim CI into one long-lived Python process — now every golden's
import numpy/import torchaccumulates conflicting libomps in the same address space.Changes
ci.py: SetKMP_DUPLICATE_LIB_OK=TRUEat the top of the file on darwin, before any import that can transitively pull in numpy or torch. This is Intel's documented escape hatch; safe for our workload where numpy/torch are only used for golden reference math, not parallel OMP regions.docs/macos-libomp-collision.md(new): Full root cause analysis, debugging steps, reproducer, and explicit "what NOT to do" list so future contributors don't re-investigate the same rabbit hole. Linked fromdocs/ci.md.examples/a2a3/{aicpu,host}_build_graph/bgemm/golden.py: Rewrite the two remaining numpy-based goldens in torch for style consistency with the rest ofexamples/. Note this does not avoid the libomp collision on its own —import torchtransitively imports numpy.Also investigated: alternatives like
ctypes.CDLL(..., RTLD_GLOBAL)pre-loading andDYLD_INSERT_LIBRARIESdo not fix this, because the two dylibs have distinctLC_ID_DYLIBinstall names and dyld resolves dependencies by install name, not by symbol. See the doc for details.Test plan
python ci.py -p a2a3simon macOS — 20/20 pass (previously 20/20 fail with SIGABRT)python ci.py -p a5simon macOS — 12/12 pass (previously 12/12 fail with SIGABRT)python ci.py(both sims together) on macOS — 32/32 passKMP_DUPLICATE_LIB_OKis only set onsys.platform == "darwin")