numpy version conflicts

[huangjundashuaige]

install from setup.py (numpy version 1.21.6)

~/alpa$ python3 tests/test_install.py
RuntimeError: module compiled against API version 0xf but this version of numpy is 0xe

after install numpy=1.22.0

python3 tests/test_install.py
Traceback (most recent call last):
  File "tests/test_install.py", line 13, in <module>
    from alpa import (init, parallelize, grad, ShardParallel, automatic_layer_construction,
  File "/home/ubuntu/alpa/alpa/__init__.py", line 1, in <module>
    from alpa.api import (init, shutdown, parallelize, grad, value_and_grad,
  File "/home/ubuntu/alpa/alpa/api.py", line 15, in <module>
    from alpa.parallel_method import ParallelMethod, ShardParallel
  File "/home/ubuntu/alpa/alpa/parallel_method.py", line 26, in <module>
    from alpa.pipeline_parallel.compile_executable import compile_pipeshard_executable
  File "/home/ubuntu/alpa/alpa/pipeline_parallel/compile_executable.py", line 27, in <module>
    from alpa.pipeline_parallel.stage_construction import (
  File "/home/ubuntu/alpa/alpa/pipeline_parallel/stage_construction.py", line 10, in <module>
    import numba
  File "/home/ubuntu/.local/lib/python3.8/site-packages/numba/__init__.py", line 200, in <module>
    _ensure_critical_deps()
  File "/home/ubuntu/.local/lib/python3.8/site-packages/numba/__init__.py", line 140, in _ensure_critical_deps
    raise ImportError("Numba needs NumPy 1.21 or less")
ImportError: Numba needs NumPy 1.21 or less

I disable numba importerror in the end and passed this test, was wondering what would be the potential problems.

[zhisbug]

Thank you. Could you try numpy==1.19.5 first. We'll fix this shortly

[huangjundashuaige]

Successfully installed numpy-1.19.5
:~/alpa$ python3 tests/test_install.py
RuntimeError: module compiled against API version 0xf but this version of numpy is 0xd
ImportError: numpy.core._multiarray_umath failed to import
ImportError: numpy.core.umath failed to import
2022-05-20 21:57:35.710782: F external/org_tensorflow/tensorflow/compiler/xla/python/xla.cc:83] Check failed: tensorflow::RegisterNumpyBfloat16() 
Aborted (core dumped)

this is the error after install numpy==1.19.5

not sure which module "compiled against API version 0xf " of numpy, I guess "0xf" means numpy==1.22.

[zhisbug]

try numpy==1.20?

[merrymercy]

I use numpy==1.20.3 and tensorflow==2.8.0

After installing numpy, please recompile jaxlib or tensorflow again

[marwage]

Hi, I had the same issue and I fixed it by removing the version requirement from numpy in setup.py

alpa/setup.py

Line 88 in 3995d72

"numpy<1.22",

My current version is numpy==1.22.4. I hope that helps others too.

[zhisbug]

@huangjundashuaige Hello, has your issue been addressed?

It is because, during the compilation of jaxlib, a higher version of numpy is used, but when running Alpa, a lower version of numpy is used.

This is possible to happen because when running pip install cupy-cuda11x, pip will resolve the dependency and install a newer version of numpy (>1.22)

But alpa's setup.py declares numpy<1.22.

We have addressed this issue via PR #492 and PR #499