SSD-LLaMA Runs Large MoE Models on Consumer GPUs

The SSD-LLaMA pipeline enables large Mixture-of-Experts models to run on consumer-grade hardware by utilizing a three-tier SSD-RAM-VRAM memory hierarchy. This method fetches specific experts from SSD storage to bypass VRAM capacity limits.