strix-halo-setup

What problem does it solve?

AMD Strix Halo users face significant hurdles setting up PyTorch for AI workloads: official PyTorch builds don't work (causing "HIP error: invalid device function"), and default memory configurations limit GPU access to ~33GB, preventing the use of larger models. This skill automates the complex setup, fixes these critical software and hardware limitations, and unlocks your system's full potential for advanced AI.

Core Features & Use Cases

• Optimized PyTorch Installation: Automatically installs community-built PyTorch versions specifically compatible with AMD Strix Halo (gfx1151), ensuring your GPU works correctly for compute operations.
• Expanded GPU Memory Access: Configures GTT (Graphics Translation Table) to allow your GPU to access up to 113GB of system RAM, enabling you to run 30B+ parameter models in FP16.
• Comprehensive System Setup: Verifies ROCm installation, user group permissions, and sets up a ready-to-use Conda environment with all necessary environment variables and test scripts.
• Use Case: A machine learning engineer acquires a new AMD Strix Halo laptop and needs to quickly set up an environment to fine-tune a 30B parameter LLM. Instead of spending days debugging driver issues and memory limits, they use this skill to get a fully optimized, working environment in minutes, ready to deploy large models.

Quick Start

Configure my AMD Strix Halo system for PyTorch and large AI models.