Pinned Loading

1. commvq-kv-cache-mlx commvq-kv-cache-mlx Public

   RoPE-commutative additive vector quantization of the KV cache in MLX, so quantized keys can be rotated without re-encoding.

   Python

2. cppo-mlx cppo-mlx Public

   Toy-scale MLX study of CPPO: pruning low-advantage completions in GRPO so each update needs far fewer sampled completions, plus dynamic completion allocation.

   Python

3. delta-attention-lm delta-attention-lm Public

   A small hybrid LM in MLX mixing gated delta-rule attention, NoPE latent MoE and attention residuals over depth, after Kimi K3.

   Python

4. japan-fire-calculator japan-fire-calculator Public

   TypeScript

5. kv-cache-fusion kv-cache-fusion Public

   Cache-to-cache communication: a trainable fuser maps one small LM's KV cache into another's so it can answer with knowledge it never learned.

   Python

6. warica warica Public

   Warica - bill splitting & financial planning (wari.ca)

   TypeScript