LatentPort native runtime
Pinned native binaries for the shared LatentPort Python library. Source code and build instructions live in the LatentPort repository.
These ABI-2 binaries implement LatentPort state operations, including compaction. LPCompact is the library's compaction feature; it does not have a separate product release.
| Platform | CUDA runtime | GPU architecture | Native ABI |
|---|---|---|---|
| Windows x64 | 13.0 | NVIDIA Blackwell SM120 | 2 |
| Linux x64 | 12.8 | NVIDIA Blackwell SM120 | 2 |
The archives are byte-identical to the previously qualified native builds.
native-runtime.json records their platform, compiler/runtime compatibility,
source revisions, byte counts, and SHA-256 values. Each archive also includes
native-build.json with hashes of its libraries. SHA256SUMS.txt verifies the ZIP files.
Base models, bridge weights, and NVIDIA drivers are distributed separately.
Installation
Use LatentPort 0.2.1 or later, whose download manifest pins this repository to an immutable commit:
python -m pip install --upgrade latentport
latentport install-native
LatentPort 0.2.0 used GitHub release downloads; upgrade before a fresh native installation. See the installation guide for backend dependencies, CUDA libraries, and supported hardware.
Licenses
LatentPort source is Apache-2.0 licensed (LICENSE). The patched llama.cpp/ggml
components retain their MIT license (LICENSE.llama.cpp), also included in each archive.
Preserve these notices when redistributing the binaries.