Vladislav-Kalinkin/ullis

Ullis Engine — reasoning ternary KAN in Rust. Contribute to Vladislav-Kalinkin/ullis development by creating an account on GitHub.

0 stars · 0 forks

README

Ullis Standalone ternary Mixture-of-Bumps Kolmogorov–Arnold engine. Zero Python at runtime. One binary: train, chat, smoke. tokens → embed → L × (causal mixer + TernaryKanLinear) → RMSNorm → tied logits

Thinking is no longer silent. The REPL streams the hidden trajectory in dim italic gray, prints └──, then streams the output block in bold green. Ephemeral GC (ReasoningScratch::clear) wipes the think ring the instant the output stream ends.

Absolute 8 GB Mac M1 benchmarks

Surface Number Notes

Pre-training throughput ~7500 tok/s 4-phase ternary QAT, Metal, defaults d=32 L=3 G=4→12 T=96 B=4

Deep inference RSS < 35 MB xhigh resonance included; think scratch is ephemeral

Unified memory 8 GB M1 SGD (no Adam states); JSONL I/O independent of corpus size

Working set is designed to stay flat: dialogue cache never stores thinking tokens, the token ring is capped at 32 768 ids (~128 KB), and Gauss–Jordan grid projection stays on Metal (G ≤ 12).

Quick start cargo build --release ./target/release/ullis train --data data/thinking-train.jsonl --steps 200 --out checkpoints/ ./target/release/ullis chat --model checkpoints/packed.bin --thinking medium ./target/release/ullis chat --model checkpoints/packed.bin --thinking xhigh --prompt "fn add(" ./target/release/ullis smoke --thinking low|medium|high|xhigh sets the KAN eval budget. low masks routed (thinking) weights and emits output immediately. xhigh runs three residual KAN loops per block on the G=12 MoE stack, streamed live.

Data — strict 4-key JSONL Every training line is exactly: { "system": "...", "user": "...", "thinking": "...", "output": "..." } Packed as: <|system|> … <|user|> … <|thinking|> … <|/thinking|> <|output|> …

Loss is masked onto thinking + output so the KAN layer learns the logical chain (bracket matching, import tracking, lifetime ownership, pipeline quoting). A verified sample corpus lives at data/thinking-train.jsonl (Rust, Python, Bash). Legacy {"text","lang"} lines are still lifted in-stream.

Visual reasoning UI

Lane ANSI Marker

thinking \x1b[2;3m dim + italic [Ullis is thinking...]

close reset └──

output \x1b1;32m bold green code stream

Colors honor NO_COLOR and non-TTY stdout. Persistent dialogue keeps only system / user / output.

Crate map

Module Role

quant TWN threshold, STE, 2-bit pack/unpack

gauss G×G Gauss–Jordan (matmul / broadcast / cat, Metal-safe)

kan TernaryKanLinear + ReLU-bump basis + MoB router

mixers CausalShift (0 params) / tiny causal attention