Platform

OS: Debian 9

Faiss version: 1.6.3

Running on:

• CPU
• GPU

Interface:

• C++
• Python

Reproduction instructions

I’m trying to train faiss IVF index on Nvidia A100 instances. Currently A100 instances support cuda 11.0 I installed faiss with cuda toolkit 10.0

Code

>>> import numpy as np
>>> import faiss
>>> 
>>> d = 256
>>> quantizer = faiss.IndexBinaryFlat(d)
>>> index = faiss.IndexBinaryIVF(quantizer, d, 4096)
>>> xt = faiss.randint((100000, 256 // 8)).astype('uint8')
>>> 
>>> index.train(xt)
WARNING clustering 100000 points to 4096 centroids: please provide at least 159744 training points
>>> quantizer2 = faiss.IndexBinaryFlat(d)
>>> index2 = faiss.IndexBinaryIVF(quantizer2, d, 4096)
>>> index2 = faiss.IndexBinaryIVF(quantizer2, d, 4096)
>>> clustering_index = faiss.index_cpu_to_all_gpus(faiss.IndexFlatL2(d))

>>> 
>>> index2.clustering_index = clustering_index
>>> index2.train(xt)

Exception log

Faiss assertion 'err == CUBLAS_STATUS_SUCCESS' failed in void faiss::gpu::runMatrixMult(faiss::gpu::Tensor<float, 2, true>&, bool, faiss::gpu::Tensor<T, 2, true>&, bool, faiss::gpu::Tensor<IndexType, 2, true>&, bool, float, float, cublasHandle_t, cudaStream_t) [with AT = float; BT = float; cublasHandle_t = cublasContext*; cudaStream_t = CUstream_st*] at ./faiss/gpu/utils/MatrixMult-inl.cuh:133; details: cublas failed (13): (512, 256) x (4096, 256)' = (512, 4096)

I guess this is due to cuda version mismatch. Is there a plan to support cuda 11.0?