Reproduce “Expert Threshold Routing for Autoregressive Language Modeling with Dynamic Computation Allocation and Load Balancing” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Hanchi Sun, Yixin Liu, Yonghui Wu, Lichao Sun
Expert Threshold (ET) routing offers a causal mechanism for Mixture-of-Experts (MoE) models, enabling dynamic computation allocation and load balancing for autoregressive language generation. This approach yields a 0.067 lower cross-entropy loss compared to Token Choice routing on a 2.4B parameter model, performing similarly to large-batch Expert Choice without its causal limitations.
Key insights
- ET routing approximates the global, stable routing behavior of Expert Choice (EC) in infinitely large batches through historically estimated thresholds, making it causal and suitable for autoregressive models.
- The method inherently achieves load balancing in expectation across the data distribution without the need for auxiliary losses, simplifying MoE training.
- ET routing allows for dynamic computation allocation per token, where the 'fanout' (number of activated experts) can vary based on token complexity and position, similar to EC but in a causal context.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.