A general-purpose NN inference stack running on WebGPU (Deno + browser, pure TypeScript + WGSL,
zero runtime dependencies). Composed of three JSR packages — @karume/runtime (IR execution) /
@karume/hub (model resolution, fetch, and caching) / @karume/models (pipelines and tokenizer)
— plus the PyPI package karume (exporter CLI) that lowers PyTorch models to IR.
Status: WIP (this description gets written during release preparation. The current design's source of truth is docs/decisions/)
License: MIT (LICENSE)