expert-kit/expert-kit

Dist-MoE implements Speed-of-Light distributed Mixture-of-Experts forward and backward kernels. They are highly optimized and fused comm-compute kernels for training and inference on SM100+ GPUs

MoonEP: A Perfectly Balanced Expert Parallelism Library via Dynamic Redundant Experts

Tutel MoE: Optimized Mixture-of-Experts Library, Support GLM/Kimi/Gpt/DeepSeek/Qwen using FP8/NVFP4/MXFP4

A Flexible Framework for Experiencing Heterogeneous LLM Inference/Fine-tune Optimizations

PyTorch library for cost-effective, fast and easy serving of MoE models.

MoE expert offload for low-VRAM GPUs — run 100B+ MoE models (DeepSeek, Qwen, Mixtral) on 8 GB cards. Expert cache with LRU hot experts, OpenAI-compatible proxy, GGUF multi-shard.

Compact and Agent-Native MoE Training System

DeepEP: an efficient expert-parallel communication library

Run MoE models bigger than your RAM. Frontier-size MoE on a phone, laptop, edge. Lossless, on stock llama.cpp. An adaptive streaming optimizer. Any hardware, any OS, any model, any quantization, behind a local OpenAI-compatible server.