munmap_chunk(): invalid pointer \n segmentation fault (core dump), when feature dim is big in index.train

[cywuuuu]

Summary

I am trying to run a baseline model VQREC pq.py I met munmap_chunk(): invalid pointer \n segmentation fault (core dump), when feat_size is bigger than 1024, but when feat_size = 768 its ok. I minimized the code below, so it can help with Reproduction the issue. (I am new to faiss, but I would like know how to fix it, thanks)

# -*- coding: utf-8 -*-
import numpy as np
import faiss
res = faiss.StandardGpuResources()
feat_size = 1024 # segment fault core dump
feat_size = 768 # ok 

index_cpu = faiss.index_factory(feat_size, "OPQ32,IVF1,PQ8x8", faiss.METRIC_INNER_PRODUCT)
index_cpu.verbose = True

filtered_feat = np.random.randn(1000, feat_size).astype('float32')

index_cpu.train(filtered_feat)

print('Training completed.')

Platform

OS: Ubuntu 20.04 Faiss version: faiss 1.9.0 Installed from: anaconda Faiss compilation options:

Running on:

• [yes] CPU
• [ ] GPU

Interface:

• [ ] C++
• [yes] Python

Reproduction instructions

Just run the code, with different feat_size.

[cywuuuu]

My RAM is 500GB, and i have not observe the surge of ram usage when running the code (or maybe everything happened too fast?), If anyone could help me understand the exact cause, I would greatly appreciate it.

[junjieqi]

hi @cywuuuu, I just tried to re-run the code you provided and use feat_size = 1024. The code is running fine. Probably you want to isolate the issue from your code to see where the seg_fault happened

[cywuuuu]

Well, it is quite strange because I do have that issue when rerunning the code, I give my detail as below(The segment fault core dump is described in Chinese 已放弃 (核心已转储), sorry for that because i am using Chinese Ver of Ubuntu)

And the following is the package I am using:

I am curious why. Hopefully you can help me out, I would greatly appreciate it,thanks @junjieqi

[junjieqi]

@cywuuuu thanks for providing additional information. I'm wondering why you are using faiss-gpu. Since I saw you initialize res = faiss.StandardGpuResources(), but that parameter never got used. Do you want to remove that one first and re-run the code?