Visit Apps Ved →

openlaked

OpenLake storage daemon for LLM inference and GPU workloads

TLDR

Start the store with a config file
$ openlaked --config [path/to/config.toml]
Start after building from source
$ ./target/release/openlaked --config crates/openlake_server/configs/storage-tcp-local.toml
Install the vLLM connector helper package
$ pip install openlake-vllm
$ openlaked

SYNOPSIS

openlaked [--config path] [options]

DESCRIPTION

openlaked is the OpenLake server binary: a high-performance storage engine for GPU-centric workloads, including KV-cache offload for LLM inference, checkpoint storage, and S3-compatible object storage. It is implemented in Rust and optimized for Linux (`io_uring`, optional RDMA/GPUDirect paths).Typical use: run `openlaked` on GPU hosts, then point vLLM (via `openlake-vllm` / OpenLakeConnector) at the node list so long-context prefixes can be reused from host RAM/disk. In object-store mode it exposes an S3-compatible endpoint for tools like the AWS CLI.

PARAMETERS

--config path

Path to OpenLake server configuration (TCP local storage, RDMA KV, etc.)

CAVEATS

Production RDMA and GPUDirect features target Linux GPU clusters; macOS is for development only. Single-node defaults use local data directories you must create. S3 demo credentials in docs are for local testing only—change for any shared deployment.

HISTORY

OpenLake is an Apache-2.0 project for AI infrastructure storage (KV pool + object store).

SEE ALSO

vllm(1), aws(1)

RESOURCES

Source code · Homepage · Documentation