PLeX

Build and run portable, self-contained software artifacts. Compose compiled graphs, WebAssembly orchestration, native providers, weights, and assets into a versioned .mln.

Learn PLeX

Follow this path for a first artifact:

  1. Install the tools.
  2. Complete the Quickstart to create, inspect, build, and load an artifact.
  3. Read Core concepts and Architecture to understand the boundaries.
  4. Use the project guide when composing your own packages and deployment policy.

How it fits together

flowchart LR
    src["Sources + synps.toml"] --> syn["SYNaPSe: resolve, validate, compile"]
    syn --> art["Versioned .mln artifact"]
    art --> mln["MyeLiN Runtime"]
    app["Host application"] --> mln

SYNaPSe resolves packages and toolchains, validates the manifest, and builds the artifact. MyeLiN provides the authoring contracts, artifact format, and runtime APIs. A deployment machine needs the runtime and compatible drivers, not the SYNaPSe compiler.

The PLeX core and authoring SDK are implemented in C++. Build components can use additional language toolchains where their execution domain supports them.

Execution domains

Domain Role Typical output
Graph Tensor computation compiled ahead of time through IREE. IREE bytecode and parameter data
Orchestration Imperative control flow around graph calls. WebAssembly module
Native Platform integration and providers. ABI-compatible shared library
Driver Runtime backend integration. Versioned runtime driver library

The manifest keeps domain and language separate: for example, a C++ toolchain can produce graph, orchestration, or native components where that combination is supported. The build-domain guide explains the contracts and compatibility rules.

Choose a task

Current support boundary

The repository currently targets Linux. The bundled Compiler enables the IREE LLVM CPU target by default; CUDA and Vulkan targets and runtime drivers are optional build features. A deployment requires a compiler target and runtime driver for the same backend. See Installation before building and Devices for backend selection details.

PLeX is not an eager tensor framework and does not provide autograd, optimizers, or model training. It focuses on composing and deploying compiled workloads.