rx-9070-xt

Hardware-focused llama.cpp fork for Windows and dual AMD RDNA4 RX 9070 XT GPUs: ROCm/HIP + Vulkan backends, PyQt6 GUI (RDNA LLM Studio), MTP speculative decoding, FP8 attention, long-context Qwen3.8 benchmarks

Qwen3.8-27B at 256k context on a 16 GB AMD RX 9070 XT (llama.cpp Vulkan): measured, quality-gated, reproducible. 56.6-58.5 t/s at working depth, ~42-46 t/s at 200k resident with a selective-attention tile-skip.

Ollama with ROCm 7 GPU acceleration for AMD RDNA 4 (RX 9070, RX 9070 XT, RX 9060 XT - gfx1201) on Windows 11

Dynamic GPU voltage control and a full Radeon tuning suite for Windows, built on AMD's official ADLX SDK. Live clock-based voltage switching plus manual tuning, fan curves, 3D & display settings, per-app boost, profiles, and a crash flight-recorder — one persistent bridge, GUI + CLI.

Reproducible Windows ROCm deployment kit + validated ComfyUI workflows for AMD RX 9070 XT. No GPU required to verify.

Running capable LLMs locally on an AMD RX 9070 XT (RDNA4) with Vulkan, local-Qwen3.6 + Claude workflow.