Tags: OriPekelman/toy
Tags
toy v0.7.0-pre-alpha — CLI-driven framework: 5-layer refactor, train(… …from-scratch/lora/warm-start)+cuda/metal, infer/eval/serve, eval lmc, vit-tiny; first RubyGems release (prerelease)
v0.4.0-pre-alpha — modern MoE works, V Q8 + flash perf realized
Real OLMoE-1B-7B-Instruct produces coherent factual answers
("The capital of France is called Paris"). Gemma 2 extras land
(soft-cap, embed scale, pre+post norms, alternating SWA). Flash
attention finally beats baseline on Qwen3-1.7B (12% faster) after
the V cache layout flip that was eating its win. SSM op primitives
bound (Mamba enable, speculative). Upstream ggml issue filed for a
real mul_mat_id × K-quant bug we found.
See CHANGELOG.md for the full inventory.
v0.3.0-pre-alpha — Metal backend (Apple Silicon). SmolLM2-135M F32 ru… …ns end-to-end on Metal with bit-identical CPU parity. Closes #2.
v0.2.0-pre-alpha — Qwen3 dense, SentencePiece, RoPE scaling, bench ha… …rness Three new model families work end-to-end as text → text: - Qwen3-0.6B (dense, with QK-norm + explicit head_dim) - Mistral-7B-Instruct-v0.2 (SentencePiece tokenizer) - TinyLlama-1.1B-Chat-v1.0 (SentencePiece tokenizer) Plus: - RoPE scaling (YaRN / llama3-style / linear) - Bench harness with regression gate (make bench) - Card-drift detector (make check-cards) - Chrome Trace Format observability primitive (TRACE=path.json) - Spinel master bump (example_serve unblocked, RBS warnings down) Still pre-alpha. See CHANGELOG.md for the full inventory.
v0.1.0-pre-alpha — first tagged cut. See CHANGELOG.md.