Weighted Adaptive Variance-Aware Expert Selection for Speech Enhancement
An LLM inference engine written in pure Rust, designed to run large models on hardware that would normally refuse them.
A TensorFlow Mixture-of-Experts Transformer built from scratch, with rotary attention, sparse routing, trained checkpoints, and dense baselines.
Research code for ProbeRoute, a probe-initialized sparse routing method for frozen-backbone multi-token prediction
TensorFlow mixture-of-experts from scratch: sparse token dispatch, load-balancing ablation, routing diagnostics, trained checkpoints, and CPU benchmarks.
To associate your repository with the sparse-routing topic, visit your repo's landing page and select "manage topics."