The fastest decode and prefill for 4-bit Qwen3.8-27B on one RTX 3090 that preserves the model's intelligence

This repository hasn't been mirrored yet. Its content will appear here once it syncs.