M* Documentation#

mstar is the package behind M* (“M-star”), a disaggregated, any-to-any multimodal inference engine. It serves models built from structurally distinct components — vision encoders, transformer backbones, diffusion / flow heads, audio codecs, action generators, world-model predictors — whose execution path changes with the input and the task.

The core abstraction is the graph walk: each model declares its computation as a dataflow graph of components, and every request is a walk over that graph. A request flows HTTP / SDK → API server → conductor → workers → streamed results; the conductor schedules the model’s graph walks to coordinate multi-engine pipelines across one or more GPUs. Logical graph structure is decoupled from physical placement, so the same model runs single-GPU or fully disaggregated by changing only a YAML config.

One runtime serves unified multimodal models, omni models, speech LMs, vision-language-action policies, and world models — through a Python SDK, an OpenAI-compatible API, and a native streaming endpoint.

Reference

Contributing

Adding a New Model