Coordinator-driven runtime for exact synchronous edge foundation-model training.
Wasp decomposes GEMM-dominated training into row/column sub-GEMM shards and elevates each shard to a first-class runtime object with explicit placement, transport-commit, cache, and recovery state. It preserves cloud-style synchronous update semantics while running on heterogeneous, volatile edge devices.
Wasp consumes the Morphling
runtime (EdgeSys '26) through its public morphling.api surface and runs its
examples on Morphling's measurement-calibrated emulator. The canonical
environment is a Docker image layered on Morphling's image.
See examples/opt125m_emulated/ for a runnable
OPT-125M end-to-end run on emulated devices. The default placement backend is
the license-free heuristic solver; the exact Gurobi MIP backend is optional
(pip install "wasp[solver]").
If you use Wasp, please cite the paper (see CITATION.cff and
docs/paper.md).
Apache-2.0. See LICENSE, NOTICE, and
THIRD_PARTY_LICENSES.md.