[MRG] Load triton before TensorFlow in ot.backend (closes #816) - #839
Open
deeb01 wants to merge 1 commit into
Open
[MRG] Load triton before TensorFlow in ot.backend (closes #816)#839deeb01 wants to merge 1 commit into
deeb01 wants to merge 1 commit into
Conversation
triton and TensorFlow both ship a statically linked LLVM. Loading triton's libtriton.so into a process that has already imported TensorFlow segfaults inside dlopen, during the static initializers of a 461 MB shared object. torch imports triton lazily rather than at import time: torch.optim's Optimizer.__init__ calls add_param_group, which is decorated with @torch._disable_dynamo, whose wrapper imports torch._dynamo on first call, which calls torch.utils._triton.has_triton_package(), which imports triton. So the first construction of any torch optimizer loads libtriton, and in a process that imported ot.backend that happens after TensorFlow is resident. This is why the crash only appeared in the jobs that install every backend into one process, and why the individual tests all passed on their own. It is not specific to torch 2.12; that was simply when torch started shipping a triton version that collides. ot.backend already imports torch before TensorFlow, so probing for triton at the torch import site loads it while the process is still clean. The probe is guarded on TensorFlow actually being importable, so installations without TensorFlow do not pay the cost of loading libtriton. Reproduced on Linux with torch 2.13.0, triton 3.7.1 and TensorFlow 2.21.0: the crash is a SIGSEGV at triton/knobs.py:15 inside create_module, matching the traceback in the issue. Verified that importing triton after TensorFlow still crashes, so it is the ordering and not the import itself that matters. Adds a non-regression test that builds a torch optimizer after importing ot.backend in a subprocess, so the segfault is reported as a test failure instead of killing the test runner. Removes the torch<2.12 pin from .github/requirements_doctests.txt and docs/requirements.txt.
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
1 participant
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.
Types of changes
Bug fix (non-breaking change which fixes an issue).
Motivation and context / Related issue
Closes #816.
triton and TensorFlow both ship a statically linked LLVM; loading
libtriton.soafter TensorFlow segfaults insidedlopen. torch imports triton lazily, on the firsttorch.optimconstructor — which in a POT process is after TensorFlow is resident. TensorFlow is the trigger, not JAX.ot/backend.pyimports torch ~60 lines before TensorFlow, so the probe goes there, guarded on TensorFlow being importable. Not inconftest.py: the doc build never loads it. Bothtorch<2.12pins removed.Upstream reproducer:
import tensorflow; import triton.How has this been tested (if it applies)
On Linux with torch 2.13.0+cu130, triton 3.7.1, TensorFlow 2.21.0,
test/conftest.pyuntouched:test_free_support_barycenter_generic_costs_auto_ground_baryreturncode=-11test_ot.py+test_backend.pyThe new test runs in a subprocess so the segfault is reported as a failure rather than killing the runner.
PR checklist