Install

kern is one static-looking binary: it links no CUDA library and loads the NVIDIA driver and cuBLAS at first use. Install a release, or build it from source; either way there is no toolkit and no Python to set up.

Requirements


OS / arch	Linux x86_64 or aarch64, glibc 2.28 or newer
Driver	supports CUDA 13 (r580 or newer)
Runtime libraries	`libcublas.so.13` and `libcublasLt.so.13` on the loader path: a CUDA 13 toolkit, an `nvidia/cuda:13.*-runtime` image, or `pip install nvidia-cublas-cu13` with `LD_LIBRARY_PATH` pointing into it
GPU	whatever the kernels you load were built for (the published Qwen3.8 registry is `sm_103a`)

From a release

curl -fsSL https://kern-baa.pages.dev/install.sh | sh
kern --version

The script downloads kern-<arch>-unknown-linux-gnu.tar.gz and SHA256SUMS from the latest GitHub release, verifies the checksum and installs one file to ~/.local/bin/kern. It does not use sudo or edit your shell profile; if the directory is not on PATH it prints the export to add. Afterwards it looks at the machine and warns, without failing, about a missing driver or cuBLAS.

KERN_VERSION=v0.2.3 pins a release
KERN_INSTALL_DIR=/opt/bin picks the directory
KERN_BASE_URL=… points at a mirror or an offline copy of the assets

Upgrading is the same line again; uninstalling is deleting the file.

From source

git clone https://github.com/pegainfer-project/kern.git
cd kern
cargo build --release
./target/release/kern --version

Needs Rust and a C++ compiler. The CUDA API the runtime binds is fixed by the pinned cudarc feature, so no CUDA toolkit is involved in the build. The repository's kern.toml names fixture targets; with their kernels and checkpoints in place, ./target/release/kern run qwen3-4b works as-is.

kern-serve

The HTTP server is a separate workspace (it carries the OpenAI front end and its dependencies) and is not in the release archive yet. It builds from the same checkout and needs pkg-config, libssl-dev and protobuf-compiler on Debian/Ubuntu:

cd crates/kern-serve && cargo build --release
target/release/kern-serve --help

Verify a manifest without a GPU

The manifest verifier is a pure Rust crate. It is what CI and editors can run on a machine with no CUDA at all:

cargo run -p kern-manifest --example verify -- examples/qwen3-4b.json
# examples/qwen3-4b.json: ok, 3 forwards, 6 fills

./target/release/kern verify path/to/manifest.json   # the CLI's version: also prints the serving protocol

The published JSON Schema gives editor completion and early validation; kern verify remains authoritative, since it runs the verifier the runtime itself uses.

Quick start: run Qwen3.8-27B end to end
The model artifact: what a manifest declares

Install ​

Requirements ​

From a release ​

From source ​

kern-serve ​

Verify a manifest without a GPU ​

Next ​