Geohot has published the Tinygrad library↗ for machine learning. This library is small and simple, which makes it a potentially good target for jetting. Doing so would give basic machine learning functionality in Urbit.
Components
- Hoon, /lib/tinygrad as part of the machine learning suite in urbit/numerics; should utilize /sur/lagoon types along with /lib/lagoon and /lib/math for numerics.
- Jets, something of an open design question right now. C or Rust FFI function calls to the conventional Python tinygrad functions may be necessary.
Milestones
- (3 stars) Implement tinygrad opcodes and functions in Hoon. Produce a transformer, if necessary, from standard tinygrad input to a Nock-compatible layer (likely as a mark). (Some prior art exists for us on this step.) Fulfill basic testing requirements as outlined below.
- (0 stars) Spec the jet layer with ~lagrev-nocfep, ~mopfel-winrux, and ~rovnys-ricfer.
- (3–5 stars, depending on spec) Implement jet-accelerated code for tinygrad. Fulfill LLaMA/Stable Diffusion operating requirements.
Timeline
- MS1 should take up to two months.
- MS2 should take three to four months, depending on implementation details.
Testing
tinygrad can run LLaMA and Stable Diffusion. Without jets, these are likely to be prohibitively slow, so the Quick Start Guide should be the testing target for the Hoon-only version. The Showcase is a good next step.
Documentation
Compliant examples starting from tiny corp materials should be supplied with each milestone.