radv

Qwen 3.8 27B ROCmFP4 on AMD Strix Halo (Ryzen AI Max+ 395). Up to 36 tok/s via MTP Speculation, TurboQuant & Mesa RADV Wave64.

Performance-tuned llama.cpp for AMD Strix Halo (gfx1151): FA + MoE-prefill fixes with a bundled current Mesa driver. Vulkan and HIP; portable dir, Docker, and distrobox.

High-performance Lemonade alternative for AMD Strix Halo & Radeon — latest MoE models (Ling-3.0-Flash, Qwen 3.8 Flash Next, Ornith 1.5), bleeding-edge RDNA 3.5 Wave64/ROCmFP4 kernels, and silicon-tuned profiles.

Kernel + Mesa/RADV patches that fix the ASRock BC-250's broken GPU compute queues: ~20% FPS from async compute

Open-source AMD Fluid Motion Frames (AFMF) for Linux: a Vulkan implicit layer that doubles the frame rate of any game, Proton or native, with FidelityFX Optical Flow. Tuned on RDNA4/RDNA3 with RADV; no game integration, no kernel module, no driver patch.

Run LLMs on AMD Vega8 / Vega10 APUs and GPUs (Ryzen 5700G / gfx90c and Vega 56/64): ROCm 7 gfx900 backport + Vulkan/RADV launchers, Docker images, benchmarks for llama.cpp and LM Studio

Async compute for the AMD BC-250 (GFX1013) on Bazzite / Fedora atomic. Patched RADV as a separate Vulkan driver.

Open Windows (WDDM) driver stack for AMD GPUs, developed first on the ASRock BC-250 (gfx1013). Not affiliated with AMD, ASRock or Microsoft.

Monorepo of 4 healthcare AI agent skills: chart review, HEDIS NLP, HCC NLP, HIPAA compliance. Install: npx skills add Yar177/medical-chart-review-skill --skill '*'

Self-hosted AI image and video generation for AMD Ryzen AI APUs on Linux: Vulkan (Mesa RADV) via stable-diffusion.cpp, no ROCm. Web UI with queue, model manager, multi-user auth and system diagnostics.

Controlled ten-boot A/B of amd_iommu=off on AMD Strix Halo (gfx1151): most of the ROCm-beats-Vulkan prefill gap appears to be the IOMMU. Five models, raw data.

🚀 Automated nightly builds & multi-backend releases (ROCm 7 TheRock, Vulkan RADV, CUDA, Metal) for CachyLLama & llama-ai on AMD Strix Halo (8060S), Strix Point (890M), Phoenix (780M), and Steam Deck with Homebrew support.

Local speech-to-text toolkit for Arch Linux. Transcribes audio/video to clean Markdown via whisper.cpp (Vulkan/RADV). CLI-first: local files, YouTube, Telegram, dictation, LLM cleanup. Clean text by default, timestamps optional. GPU: AMD RX 580 via RADV

Mirror: https://gitlab.freedesktop.org/mesa/mesa

Agentic auditor that flags unsubstantiated CMS-HCC risk-adjustment codes in clinical documentation, grounded against a deterministic Rust scoring engine as a verifiable oracle. Built on synthetic data.

Qwen3.8 Flash Next (UD-Q2_K_XL + 512K + MTP + Vision) on AMD Strix Halo — 2026-09-25 Dashboard 功率限制(90W 甜点) + 调用来源 + 模型来源 + llama.cpp 9-24 重编译

One buffer, two devices, no host copy: cross-GPU data handoff via Vulkan external memory \u2014 with the honest benchmark that beats its own thesis.