Skip to content

CUDA context pollution #5371

Description

@oseanblue

When using faiss.index_cpu_to_all_gpus() and calling search(), FAISS appears to permanently modify the CUDA context in a way that prevents subsequent PyTorch kernels from setting their required shared memory size. The error persists even after deleting the index object and calling torch.cuda.empty_cache(). This is a process-global state mutation with no cleanup path — any ML system running FAISS GPU alongside PyTorch in a single process is affected.

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions