Circular imports when importing einops and torch._dynamo

[befelix]

Describe the bug
Importing einops before torch._dynamo currently leads to warnings. I'm not sure if this needs a fix on the pytorch or einops side. This is annoying for CI pipelines, where warnings are typically treated as errors. Note that, with sorted imports, einops will typically import before any torch namespaces.

$ python -W "error" -c 'import einops; import torch._dynamo'

Traceback (most recent call last):
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/einops/_torch_specific.py", line 106, in allow_ops_in_compiled_graph
    from torch._dynamo import allow_in_graph
ImportError: cannot import name 'allow_in_graph' from partially initialized module 'torch._dynamo' (most likely due to a circular import) (/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/__init__.py)

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "<string>", line 1, in <module>
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/__init__.py", line 2, in <module>
    from . import allowed_functions, convert_frame, eval_frame, resume_execution
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/convert_frame.py", line 62, in <module>
    from .output_graph import OutputGraph
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/output_graph.py", line 89, in <module>
    from .variables.builder import GraphArg, TrackedFake, VariableBuilder, wrap_fx_proxy
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/variables/builder.py", line 143, in <module>
    from .optimizer import OptimizerVariable
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/variables/optimizer.py", line 5, in <module>
    from ..decorators import mark_static_address
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/decorators.py", line 284, in <module>
    allowed_functions.add_module_init_func("einops", _allow_in_graph_einops)
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/allowed_functions.py", line 489, in add_module_init_func
    init_func()
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/torch/_dynamo/decorators.py", line 264, in _allow_in_graph_einops
    from einops._torch_specific import (  # noqa: F401
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/einops/_torch_specific.py", line 127, in <module>
    allow_ops_in_compiled_graph()
  File "/home/bfe2rng/Code/rlcore/venv/lib/python3.11/site-packages/einops/_torch_specific.py", line 108, in allow_ops_in_compiled_graph
    warnings.warn("allow_ops_in_compiled_graph failed to import torch: ensure pytorch >=2.0", ImportWarning)
ImportWarning: allow_ops_in_compiled_graph failed to import torch: ensure pytorch >=2.0

Reproduction steps
Steps to reproduce the behavior:

python -W "error" -c 'import einops; import torch._dynamo'

Expected behavior

No output, import order does not matter.

Your platform

Ubuntu 22.04, torch==2.2.2, einops==0.7.0

[arogozhnikov]

Hi @befelix , thank you for reporting, there is a circular import problem that we need to address.

step-by-step what happens in python -W "error" -c 'import einops; import torch._dynamo'

1. einops is imported
2. torch._dynamo is imported, it detects that einops is imported, and calls https://github.com/pytorch/pytorch/blob/88a71594933b2464d9d8b6b3533c5a945a4ac2ff/torch/_dynamo/decorators.py#L322
3. ... which imports from einops._torch_specific ...
4. einops._torch_specific tries to register ops by using allow_in_graph, that is needs to import first from torch._dynamo. That's where we run in circular imports

cc @wconstab @jansel

[arogozhnikov]

Update: this was fixed in pytorch 2.4, I'll leave the issue open until 2.5 comes out.

For pytorch < 2.4 use these instructions