RDNA3-tuned llama.cpp fork doubles Qwen3.8-27B speed on RX 7900 XTX to ~100 tok/s
Jebbyk1 · reddit · 2026-11-12
A Reddit user ran Qwen3.8-27B (Q4KM) on an RX 7900 XTX with a community llama.cpp fork specifically optimized for RDNA3 GPUs, roughly doubling throughput to 97-110 tok/s decode (spiking to 130-150 on code) with no visible quality loss.
Key details:
- prefill 890-1015 tok/s, VRAM 21-22GB of 24GB
- Uses DFlash draft-model speculative decoding, q80 KV cache, flash-attention, Resizable BAR enabled
- The iGPU runs the desktop UI so all XTX VRAM goes to the LLM
The full working llama-server launch command is included verbatim, so RDNA3 owners can reproduce the setup directly.
More from Infra
- Local LLM on 3090 + 64GB RAM hits 70-80 t/s with QFN quants; eyeing a $3-3.5k Strix Halo box — makkalot · 2026-11-12
- Running ComfyUI on NVIDIA DGX Spark with Docker Compose — setup shared — Ok_Joke5408 · 2026-11-12
- Beff Jezos: Even the toilets in Japan are running 3B-parameter models — beffjezos · 2026-11-12
- Curated 9-paper reading list for LLM quantization beginners, from QLoRA to BitNet — East-Muffin-6472 · 2026-11-12
- llama.cpp PR #31915 fixes Gemma4 tool calling, user reports it "completely fixed" agentic coding — jacek2023 · 2026-11-12
- Nvidia board reportedly working with outside firm to find Jensen Huang successor — kimmonismus · 2026-11-12