Publishing ipex-from-scratch and Building a Shim Runtime

A GitHub-Safe Snapshot

I prepared a publishable snapshot of ipex-from-scratch (90c9e933). The project had accumulated large binary artifacts, IGC dumps, and generated files that don’t belong in a repository. Curating what to publish forced me to separate source from artifacts, which is genuinely useful—the repo is now smaller and the build is clearer.

The Shim Runtime

I added a shim runtime with bootstrap docs and an iteration queue (eb0df504). The shim is the layer that interposes the IPEX implementations on top of upstream llama.cpp. A runtime shim, as opposed to a compile-time link, means I can enable or disable specific interposers and measure each one.

Advancing the Interposers

I tightened the benchmark loop and advanced the shim interposers (142b4685). The iteration loop matters here: change one interposed op, run the benchmark, compare throughput and correctness, record the result, repeat. Making that loop fast is what makes the whole project tractable.

Why Publish It

Most of my local-inference work lives in private repos. ipex-from-scratch is worth publishing because the reverse-engineering methodology is broadly useful—how to go from a proprietary kernel to a verifiable reimplementation. Publishing also imposes a quality bar that a private repo never does.