AI 日报hiw3c.com

ROCmFix和InferBench - AMD Local-LLM设置和Vulkan与髋关节基准测试

原文标题 · ROCmFix and InferBench – AMD Local-LLM Setup and Vulkan vs. Hip Benchmarking
Hacker News Top github.com 网页快照
正文为英文,可一键机器翻译(仅首次需要等待)

Notifications You must be signed in to change notification settings

Latest commit

History

Folders and files

Repository files navigation

🔧 ROCmFix

Auto-detect your AMD GPU, fix ROCm overrides, and benchmark your AI backends in seconds.

One file. Zero installs. No pip. No venv. Just run it.

❓ What is this?

Running local AI (llama.cpp, Ollama, LM Studio, PyTorch) on AMD GPUs often leads to cryptic errors like HSA_STATUS_ERROR_INVALID_ISA or silent fallbacks to CPU. Usually, you just need to set the HSA_OVERRIDE_GFX_VERSION environment variable.

ROCmFix detects your exact AMD GPU, checks a live community database, automatically applies the correct shell environment override, and runs health diagnostics to ensure your setup is ready.

✨ Features

Auto-Apply & Undo: Automatically modifies CMD, PowerShell, Bash, Zsh, or Fish configs to apply overrides safely.

rocmfix doctor : Scans your hardware, Adrenalin drivers, HIP SDK, and Vulkan API to diagnose issues.

rocmfix bench : A 10-second backend race! Temporarily isolates Ollama or LM Studio, runs a generation test using Vulkan, then HIP, and tells you which is faster on your PC.

rocmfix install-hip : Missing the HIP SDK on Windows? This safely downloads and launches the official AMD installer for you.

Live Database & Self-Updater: Fetches new GPUs from the cloud daily, and updates itself via rocmfix update .

Export System Report: Generate a Markdown file of your setup for Reddit/Discord help posts.

🚀 Quick Start

Download (One Command)

Invoke-WebRequest - Uri " https://raw.githubusercontent.com/xanpavle/rocmfix/main/rocmfix.py " - OutFile " rocmfix.py "
curl -O https://raw.githubusercontent.com/xanpavle/rocmfix/main/rocmfix.py

Run

python rocmfix.py

📋 Commands

🎮 Supported GPUs

ROCmFix supports RDNA4 (RX 9070/9060 series), RDNA3 (RX 7000 series), RDNA2 (RX 6000 series), and Integrated Graphics.

New GPUs are synced automatically from our live database without needing to update the app!

Don't see your card? Run python rocmfix.py — it will generate a pre-filled GitHub issue link so you can contribute your GPU!

🔥 HIP vs Vulkan — Which should I use?

If you're running llama.cpp / LM Studio / Ollama on a consumer AMD card (especially on Windows):

TL;DR

Running LLMs on Windows? → Try Vulkan first. Run rocmfix bench to verify.

PyTorch / Stable Diffusion? → Run rocmfix install-hip , then apply the override ROCmFix gives you.

📊 Privacy & Telemetry

ROCmFix includes an opt-in telemetry system to help grow the community database. If enabled, ROCmFix securely sends your GPU model, driver version, OS, and benchmark results to a private collector. We never collect usernames, IPs, file paths, or AI prompts. Read the full details in PRIVACY.md .

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

About

Auto-detect AMD GPUs on Windows/Linux and instantly get the correct HSA_OVERRIDE_GFX_VERSION for local AI (ROCm/HIP).

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages