amd-radeon

一位 Reddit 用户用 8 张 Radeon Pro V620（256GB VRAM）和自定义 vLLM fork，让 Qwen3.8-Flash-Next 达到 60-100 t/s 解码速度和 3000+ t/s 预填充，比 llama.cpp 快约 800%。

The author is building a dual GPU system for running large language models, evaluating AMD Radeon AI Pro R9700 versus NVIDIA RTX 3090 options while aiming to reduce AI subscription costs.

The Lucebox engine now supports running the Qwen3.8-27B model on a single AMD Radeon AI PRO R9700 GPU, achieving up to 227 tok/s on code tasks using the DFlash2 drafter with lossless verification.