commit a30854e8c89f02c88bca34fe5fae6b37188aae60
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Sep 11 00:52:48 2026 +0200

dev-python/instanttensor-bin: trim 0.1.9 comments

commit cdf1380170f8c434c3bc3378e5a3c23199b2608a
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Sep 9 19:40:56 2026 +0200

dev-python/instanttensor-bin: new package, add 0.1.9

vllm 0.29.0's requirements/cuda.txt declares instanttensor >= 0.1.9
unconditionally. It backs the new "instanttensor" --load-format, which
streams safetensors straight into GPU memory over io_uring, libaio or
cuFile and pipelines the prefetch across the ranks of a torch process
group.

Packaged from the upstream wheels rather than the sdist: the sdist
vendors and statically links libaio, liburing, a Boost header subset,
pybind11 and dlpack, and drives their in-tree Makefiles from a custom
build_ext, so unbundling those to the system copies is a packaging job of
its own. The extension links no libtorch or libc10, so unlike a torch C++
extension it is not ABI-locked to a torch minor and upstream's
torch>=2.8.0 floor is the real bound.