Clustering on a selected GPU

[dzenilee]

Summary

Currently, this is what it says in the README:

Clustering on one or several GPUs can be done via the gpu=True (use all gpus) or gpu=3 (use 3 gpus) constructor option in the KMeans object.

So we can do something like:

kmeans = faiss.Kmeans(d, k, niter=niter, verbose=verbose, gpu=True)

But can we have the option to choose WHICH gpu(s) to use instead of specifying the number of gpus to use (or to use all gpus)?

faiss.index_cpu_to_gpu is only for a Faiss index, right?

Platform

OS:

Faiss version:

Installed from:

Faiss compilation options:

Running on:

• [ ] CPU
• [x] GPU

Interface:

• [ ] C++
• [x] Python

Reproduction instructions

[mdouze]

This precise configuration is not supported. You can say how many gpus you want with gpu=3 though. Refer to https://github.com/facebookresearch/faiss/blob/main/faiss/python/extra_wrappers.py#L456

[matrixji]

It seems could configure the CUDA_VISIBLE_DEVICES environment variable at CUDA level, this will configure CUDA only to see those devices.