EvoESAP: Non-Uniform Expert Pruning for Sparse MoE

EvoESAP is a specialized research project addressing a critical bottleneck in Sparse Mixture-of-Experts (SMoE) models: the massive VRAM footprint required to store all experts. While the method of using evolutionary search for non-uniform layer-wise pruning is a sound 'novel combination' of existing techniques, the project lacks any defensibility. With 0 stars and 5 forks, it is currently a static research artifact accompanying an arXiv paper rather than a living software project. From a competitive standpoint, this is high-risk for frontier lab absorption. Labs like OpenAI, Anthropic, and DeepSeek (the primary purveyors of MoE) are already deeply invested in post-training optimization. If this technique proves superior to uniform pruning or standard quantization (like GGUF/EXL2), it will be integrated into inference engines like vLLM, TensorRT-LLM, or TGI within months. There is no 'moat' here; the value lies entirely in the mathematical approach, which is trivially reproducible by any senior ML engineer once the paper is read. The displacement horizon is short (6 months) because the field of MoE compression is moving at breakneck speed, and more integrated solutions (like expert merging or dynamic routing) are likely to emerge from larger research teams.