Error handling for indexes not supported on GPU?

[mlomeli1]

Summary

Currently, the indexes that are supported in GPU are clearly listed in the wiki: https://github.com/facebookresearch/faiss/wiki/Faiss-on-the-GPU#implemented-indexes, however, it would be great if we could throw an error when the user attempts to move an unsupported GPU index instead of failing silently and running in the CPU.

Platform

Faiss version: faiss-gpu 1.7.4

Installed from: conda

Running on:

• GPU

Interface:

• Python

Reproduction instructions

I have created an IndexPQ, attempted to move it to the GPU and conducted search. This results in the knn search taking a long time 236.77 seconds plus nvidia-smi pmon showcases we indeed never use the GPU:

# gpu         pid  type    sm    mem    enc    dec    command
# Idx           #   C/G     %      %      %      %    name
    0    2711586     C     0     0     -     -   python3.10
    1          -     -     -     -     -     -   -
    0    2711586     C     0     0     -     -   python3.10
    1          -     -     -     -     -     -   -
    0    2711586     C     0     0     -     -   python3.10

import faiss
import numpy as np
import time

dimension = 128                          
n = 1000000                
db_vectors = np.random.random((n, dimension)).astype('float32') 
k = 3 
dimension=128
query_vectors = np.random.random((n, dimension)).astype('float32')
code_size=16
res = faiss.StandardGpuResources()
start = time.time()
#define the pq index in cpu
index_pq = faiss.IndexPQ(dimension, code_size, 8) 
index_pq.train(db_vectors)  
index_pq.add(db_vectors)  
gpu_index_pq = faiss.index_cpu_to_gpu(res, 0, index_pq)
distances_pq, indices_pq = gpu_index_pq.search(query_vectors, k)
end = time.time()
total_time = end - start
print("\n"+ str(total_time))

In contrast, with an IVFPQ (which is GPU supported) the run time is 3.48 seconds utilising the same number of vectors and we use the GPU according to nvidia-smi pmon as expected.

gpu         pid  type    sm    mem    enc    dec    command
# Idx           #   C/G     %      %      %      %    name
    0    2711586     C     0     0     -     -   python3.10
    1          -     -     -     -     -     -   -
    0    2711586     C     0     0     -     -   python3.10
    1          -     -     -     -     -     -   -
    0    2711586     C     0     0     -     -   python3.10
    1          -     -     -     -     -     -   -
    0    2711586     C    25     6     -     -   python3.10
    1          -     -     -     -     -     -   -
    0    2711586     C    10     3     -     -   python3.10

dimension = 128                          
n = 1000000                
db_vectors = np.random.random((n, dimension)).astype('float32') 
k = 3 
dimension=128
query_vectors = np.random.random((n, dimension)).astype('float32')
code_size=16
nlist=math.floor(math.sqrt(n))
res = faiss.StandardGpuResources()
#define the ivfpq index in cpu
start = time.time()
quantizer = faiss.IndexFlatL2(dimension)  # coarse quantizer
#define the inverted index 
index_ivfpq = faiss.IndexIVFPQ(quantizer, dimension, nlist, code_size, 8)
index_ivfpq.train(db_vectors) 
index_ivfpq.add(db_vectors) 
gpu_index_ivfpq = faiss.index_cpu_to_gpu(res, 0, index_ivfpq)
distances_ivfpq, indices_ivfpq = gpu_index_ivfpq.search(query_vectors, k)
end = time.time()
total_time = end - start
print("\n"+ str(total_time))

[mdouze]

+1

[mdouze]

The issue here is that wrapped indexes (like IDMap and Pretransform) should be cloned on CPU, that's normal but indexes that contain actual data should be moved to GPU (and the cloning function should throw if this is not the case).

[ramilbakhshyiev]

Fixed in #3336