Module gpu 

GPU-accelerated vector search and k-means.

Two independent GPU backends, both via runtime libloading — no compile-time GPU SDK required. Either backend returns None when its hardware/libraries are absent; callers fall back to CPU automatically.

• NVIDIA CUDA: cuBLAS SGEMM via dlopen of libcudart + libcublas.
• AMD ROCm: hipBLAS SGEMM via dlopen of libamdhip64 + libhipblas.

Functions

try_nvidia_kmeans

k-means on an NVIDIA GPU via cuBLAS SGEMM.

try_nvidia_search_batch

Batch top-k vector search on an NVIDIA GPU via cuBLAS SGEMM.

try_rocm_kmeans

k-means on an AMD ROCm GPU.

try_rocm_search_batch

Batch top-k vector search on an AMD ROCm GPU via hipBLAS SGEMM.