With the recent progresses in Generative Artificial Intelligence (GenAI) and especially with Large Language Models (LLM), many use cases (e.g. chat interfaces on serverless, chaining models) motivates the need for bringing the existing libraries and tools over to WASM/WASI. Rebuilding all the ecosystems would be a titanic work and duplicate the maintenance surface, making it unsustainable. So far, there are only a few libraries that have been manually adapted (see 1 and 2) and there is no known initiative towards a systematic and automated bridge.

Tackling this challenge would make immediately available more than 550'000 packages (with a total download of 1.5 billion a day) to the WASM/WASI ecosystem, and offer an alternative runtime for Python based on all the advantages that WASM brings. With this scoping, we have four specific research questions to be answered:

• What are the technological limitations with native libraries and extensions when it comes to WASM/WASI compatibility?
• To what extent can a library with native libraries and extensions be adapted automatically into a compatible WASM/WASI module?
• Is the approach suggested by Pyodide the only viable one or are they alternative to would enable to build a WASM compatible registry on the top of PyPi (the de facto Python registry), e.g. WASM host interface (sidecar process)?
• If successful, can this approach be combined with the WebGPU standard to facilitate the deployment of machine learning inference directly in applications?