segfault when doing index.search using faiss-cpu about faiss-wheels

[kyamagu]

Not sure, likely the ABI compatibility issue in the runtime. Do you have any stack trace?

[rom1504]

#0  0x00007fffe814e28d in ?? () from /media/donnees/non_sauvegarde/image_embeddings_env/lib/python3.7/site-packages/faiss/_swigfaiss.cpython-37m-x86_64-linux-gnu.so
#1  0x00007fffe8174dc0 in ?? () from /media/donnees/non_sauvegarde/image_embeddings_env/lib/python3.7/site-packages/faiss/_swigfaiss.cpython-37m-x86_64-linux-gnu.so
#2  0x00000000005c8663 in _PyMethodDef_RawFastCallKeywords ()
#3  0x0000000000535990 in ?? ()
#4  0x000000000053c5a1 in _PyEval_EvalFrameDefault ()
#5  0x00000000005c916b in _PyFunction_FastCallKeywords ()
#6  0x0000000000535880 in ?? ()
#7  0x0000000000538713 in _PyEval_EvalFrameDefault ()
#8  0x00000000005365e7 in _PyEval_EvalCodeWithName ()
#9  0x00000000005c9468 in _PyFunction_FastCallKeywords ()
#10 0x0000000000535880 in ?? ()
#11 0x0000000000538713 in _PyEval_EvalFrameDefault ()
#12 0x00000000005365e7 in _PyEval_EvalCodeWithName ()
#13 0x000000000064cbb3 in PyEval_EvalCode ()
#14 0x00000000006402a3 in ?? ()
#15 0x0000000000640357 in PyRun_FileExFlags ()
#16 0x000000000064110a in PyRun_SimpleFileExFlags ()

Not sure if that really helps

[rom1504]

Ah interesting...
It seems I'm using the same issue as #1
In my environment I had tensorflow 2.2.0
If I make an environment with only faiss-cpu, it works

[rom1504]

Do you think a similar fix as #1 (comment) could be possible ?

[rom1504]

Ok no I identified the actual issue.
In order to install tensorflow 2, it's needed to update pip.
So I updated to pip pip-20.1.1.
Run pip install -U pip in a venv for this

These commands reproduce my issue :

python3 -m venv env
source env/bin/activate
pip install -U pip
pip install faiss-cpu
python test.py

With test.py content being the thing in the first post

I'm using python 3.7.5 if that can make a difference

Can you reproduce ? any idea why that would happen ?

[kyamagu]

There could be many possibilities, but I could not reproduce this on docker ubuntu 20.04 image. It could be bad interaction with multiple package managers (pip and conda), it could be lack of cpu feature (avx2), etc.

[kyamagu]

Also, if you import tensorflow as tf, always do that before importing other pacakges.

[rom1504]

rom1504@rom1504-W35-37ET:~$ cat /proc/cpuinfo
processor       : 0
vendor_id       : GenuineIntel
cpu family      : 6
model           : 58
model name      : Intel(R) Core(TM) i5-3230M CPU @ 2.60GHz
stepping        : 9
microcode       : 0x21
cpu MHz         : 1197.274
cache size      : 3072 KB
physical id     : 0
siblings        : 4
core id         : 0
cpu cores       : 2
apicid          : 0
initial apicid  : 0
fpu             : yes
fpu_exception   : yes
cpuid level     : 13
wp              : yes
flags           : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx rdtscp lm constant_tsc arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm cpuid_fault epb pti ssbd ibrs ibpb stibp tpr_shadow vnmi flexpriority ept vpid fsgsbase smep erms xsaveopt dtherm ida arat pln pts md_clear flush_l1d
bugs            : cpu_meltdown spectre_v1 spectre_v2 spec_store_bypass l1tf mds swapgs itlb_multihit srbds
bogomips        : 5188.15
clflush size    : 64
cache_alignment : 64
address sizes   : 36 bits physical, 48 bits virtual
power management:

Indeed looks like my CPU does not support avx2

What I'm quite confused about is if I do not update pip (so keeping pip 18.1 with python 3.7) then everything works.
If I do update pip to 20.1.1, I'm getting that segfault, even without installing tensorflow.
I don't use conda

[rom1504]

Ah I think I found something.
If I use pip 1.18.1 (working) it's using faiss_cpu-1.6.3-cp37-cp37m-manylinux1_x86_64.whl
whereas if I use pip 20.1.1 (not working) it's using faiss_cpu-1.6.3-cp37-cp37m-manylinux2010_x86_64.whl

Are there some differences in this second file that might explain my result ?
Maybe this second file requires avx2 ?

If that's the case, would it be possible to provide a non avx2 build or do you advise I build it myself ?

(and do you know a way to identify if the issue is missing avx2 ?)

[rom1504]

I tried using python 3.8 instead of 3.7 as per #8 (comment) but this makes no difference in my case

[rom1504]

Tried to build myself but getting some other errors when trying to follow https://github.com/kyamagu/faiss-wheels#linux , I opened a separate issue for this (#17)
I think it may make sense for me to just not support hardware without AVX2 for the lib I'm building, and advising to use google colab in that case.

[kyamagu]

manylinux1 and manylinux2010 have different glibc version dependency, and they are usually forward compatible: manylinux1 works on centos 5 environment and above, manylinux2010 works on centos 6 and above. ubuntu 20.04 should be fine with both.

Hmm, it makes sense to support option to disable / enable avx2.

[kyamagu]

If you build from source, you can also try the official approach:

cd faiss
./configure
make
make -C python

[rom1504]

So as discussed in the github action PR, the new version will not use avx2 instructions. That's great. I understand it's not possible to change an existing release and you follow faiss versioning which makes sense.
In the mean time, I forked your repo (https://github.com/rom1504/faiss-wheels), and thanks to the use of github action, I was able by just changing NAME in setup.py to release this package https://pypi.org/project/faiss-cpu-noavx2 which completely solves my problem

I will use it until faiss release a new version, at this point I'll be able to go back to faiss-cpu

Again thanks for the support and for building this simple and useful repo, I think at some point you could consider simply doing a PR to faiss main repo, they might just accept it and maintain it in the future.

[kyamagu]

@rom1504 Great.

As a side note, faiss maintainers seem not to be interested in supporting PyPI: facebookresearch/faiss#170 (comment)
This is why this unofficial repo exists.