parallax

Why we included this project

Parallax lets teams turn the machines they already own into a single inference cluster, so a couple of spare GPUs, a Mac, and a box in another office can share one workload. It shards large open-weight models across nodes with pipeline parallelism, batches requests with a paged KV cache on Apple silicon, and routes traffic on demand, which is how scattered hardware ends up acting like one serving endpoint. The GPU side leans on SGLang and vLLM, while Macs run through MLX LM, so it slots into an existing serving stack instead of asking you to rebuild one. For anyone hosting DeepSeek, Qwen, GLM, or similar open-weight models who wants more throughput, reliability, or data locality than a single machine allows, this is a practical way to get it from hardware that is already sitting around.