New Rust-Based Inference Engine for DeepSeek v4.1-Flash Reaches 2060 TPS

A custom inference tool for the DeepSeek v4.1-flash model was built using Rust and CUDA. This engine reaches a decode speed of about 2060 tokens per second on 8 B300 GPUs. The project also records a 1M token prefill in approximately 12 seconds on 4 B300 GPUs.