commit 071260b9634f526939edd76b2da6388dfa588b90
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu May 7 14:48:43 2026 +0200

dev-python/fastsafetensors: new package, 0.2.2

Tier 0 leaf for the vllm CUDA target packaging cycle. High-performance
safetensors model loader using NVIDIA GPUDirect Storage (CUFile) for
direct disk-to-GPU paths.

The C++ extension is built with pybind11 and links only against stdc++
— libcuda / libcudart / libcufile are dlopen'd at runtime through the
self-contained cuda_compat.h shim, so no CUDA headers are needed at
build time and no link-time CUDA dep. nvidia-cuda-toolkit is only
required to actually exercise the GDS path; the CPU fallback works
without it, so it is omitted from RDEPEND — the consuming vllm CUDA
target ebuild already pulls the toolkit.