Repository navigation
Conversation
meta-cla
Bot
added
the
CLA Signed
Contributor
|
@yutingye has exported this pull request. If you are a Meta employee, you can view the originating Diff in D110929696. |
meta-codesync
Bot
changed the title
meta-codesync Bot
pushed a commit
that referenced
this pull request
Jul 7, 2026
Summary: When using pymomentum on linux, importing its submodules before importing torch will crash. ```python #import pymomentum.geometry #import torch # crashes ``` It's because the wheel packaging step re-packages and renames stdlibc++, so importing torch next creates a duplicated c runtime. The reason we did it is because different runtime libs require different gcc versions... This diff adds pymomentum/__init__.py to handle this conflict for linux. Differential Revision: D110929696
meta-codesync
Bot
force-pushed
the
export-D110929696
branch
from
July 7, 2026 22:02
36a651c to
20bb32c
Compare
meta-codesync Bot
pushed a commit
that referenced
this pull request
Jul 8, 2026
Summary: Linux PyPI wheels bundle a renamed private libstdc++ because conda-forge runtime dependencies require newer GLIBCXX symbols than manylinux_2_28's system libstdc++ provides. The original import failure was caused by loading that private C++ runtime before PyTorch under Python's default RTLD_LOCAL mode: import pymomentum.geometry import torch That lets a second libstdc++ instance exist in the process before torch._C is loaded, and torch can segfault during import. A scoped RTLD_GLOBAL loader makes that order pass, but it is not a correct fix: it exposes PyMomentum's private libstdc++ symbols to the whole process. In the opposite order, that can poison later native library loads. In particular, Triton's compiler fails in ConvertTritonGPUToLLVM after this order: import torch import pymomentum.geometry import pymomentum.quaternion # backend=\"triton\" Fix this by changing the Linux wheel repair shim to import torch, when torch is available, immediately before loading any PyMomentum native extension. The extension still loads with Python's normal RTLD_LOCAL behavior, so PyMomentum's private libstdc++ remains hidden from later native extensions such as Triton, while torch owns the process-global C++ runtime state before PyMomentum's private runtime is loaded. The generated pymomentum/__init__.py is added only during Linux wheel repair. This keeps macOS and Windows behavior unchanged for now because the known bug is in Linux ELF/dlopen/libstdc++ interaction. If we later want one package shape on all platforms, we can move a no-op-on-non-Linux __init__.py into the source tree, but that should be a separate deliberate package-shape change. Also expand wheel smoke tests so every wheel variant installs torch and checks the pymomentum.geometry-before-torch order. The strict GPU wheel test continues to require CUDA and Triton. Differential Revision: D110929696
meta-codesync
Bot
force-pushed
the
export-D110929696
branch
from
July 8, 2026 01:39
20bb32c to
f999266
Compare
Summary: Linux PyPI wheels bundle a renamed private libstdc++ because conda-forge runtime dependencies require newer GLIBCXX symbols than manylinux_2_28's system libstdc++ provides. The original import failure was caused by loading that private C++ runtime before PyTorch under Python's default RTLD_LOCAL mode: import pymomentum.geometry import torch That lets a second libstdc++ instance exist in the process before torch._C is loaded, and torch can segfault during import. A scoped RTLD_GLOBAL loader makes that order pass, but it is not a correct fix: it exposes PyMomentum's private libstdc++ symbols to the whole process. In the opposite order, that can poison later native library loads. In particular, Triton's compiler fails in ConvertTritonGPUToLLVM after this order: import torch import pymomentum.geometry import pymomentum.quaternion # backend=\"triton\" Fix this by changing the Linux wheel repair shim to import torch, when torch is available, immediately before loading any PyMomentum native extension. The extension still loads with Python's normal RTLD_LOCAL behavior, so PyMomentum's private libstdc++ remains hidden from later native extensions such as Triton, while torch owns the process-global C++ runtime state before PyMomentum's private runtime is loaded. The generated pymomentum/__init__.py is added only during Linux wheel repair. This keeps macOS and Windows behavior unchanged for now because the known bug is in Linux ELF/dlopen/libstdc++ interaction. If we later want one package shape on all platforms, we can move a no-op-on-non-Linux __init__.py into the source tree, but that should be a separate deliberate package-shape change. Also expand wheel smoke tests so every wheel variant installs torch and checks the pymomentum.geometry-before-torch order. The strict GPU wheel test continues to require CUDA and Triton. Reviewed By: cstollmeta Differential Revision: D110929696
meta-codesync
Bot
force-pushed
the
export-D110929696
branch
from
July 14, 2026 00:10
f999266 to
edf33e4
Compare
Contributor
|
This pull request has been merged in 3b45e02. |
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:
Linux PyPI wheels bundle a renamed private libstdc++ because conda-forge
runtime dependencies require newer GLIBCXX symbols than manylinux_2_28's
system libstdc++ provides.
The original import failure was caused by loading that private C++ runtime before
PyTorch under Python's default RTLD_LOCAL mode:
import pymomentum.geometry
import torch
That lets a second libstdc++ instance exist in the process before torch._C is
loaded, and torch can segfault during import.
A scoped RTLD_GLOBAL loader makes that order pass, but it is not a correct fix:
it exposes PyMomentum's private libstdc++ symbols to the whole process. In the
opposite order, that can poison later native library loads. In particular,
Triton's compiler fails in ConvertTritonGPUToLLVM after this order:
import torch
import pymomentum.geometry
import pymomentum.quaternion # backend="triton"
Fix this by changing the Linux wheel repair shim to import torch, when torch is
available, immediately before loading any PyMomentum native extension. The
extension still loads with Python's normal RTLD_LOCAL behavior, so PyMomentum's
private libstdc++ remains hidden from later native extensions such as Triton,
while torch owns the process-global C++ runtime state before PyMomentum's private
runtime is loaded.
The generated pymomentum/init.py is added only during Linux wheel repair.
This keeps macOS and Windows behavior unchanged for now because the known bug is
in Linux ELF/dlopen/libstdc++ interaction. If we later want one package shape on
all platforms, we can move a no-op-on-non-Linux init.py into the source tree,
but that should be a separate deliberate package-shape change.
Also expand wheel smoke tests so every wheel variant installs torch and checks
the pymomentum.geometry-before-torch order. The strict GPU wheel test continues
to require CUDA and Triton.
Reviewed By: cstollmeta
Differential Revision: D110929696