Single AMD R9700 Hits 153 tok/s on Qwen3.8 27B with NVFP4

A developer doubled inference performance on a single AMD R9700 running Qwen3.8 27B NVFP4, reaching 153 tok/s peak decode. Online NVFP4-to-MXFP4 conversion in a vLLM build exploits the card's MXFP4 fast path, hitting 5809 tok/s prefill.

2026-09-16 ~ 2026-09-17 · 2 related posts

• NVFP4-to-MXFP4 online conversion hits 5,809 tok/s prefill on AMD R9700 — whodoneit1 · 2026-09-16
• Single AMD R9700 Hits 153 tok/s Running Qwen3.8 27B NVFP4 After Optimization — whodoneit1 · 2026-09-17