jerome
05eccd5b53
- Moved scattered scripts, templates, and documentation into organized directories (applications/, scripts/, assets/). - Updated .gitignore to strictly exclude secrets, state files, and IDE configs. - Added comprehensive README.md outlining repository structure and best practices. - Preserved all existing documentation and technical architecture files. - Prepared infrastructure/ for AWS Org and Proxmox Terraform management.
56 lines
1.5 KiB
Markdown
56 lines
1.5 KiB
Markdown
# 03 - AMD GPU Passthrough (Strix Halo) — Home
|
|
**Purpose:** Detailed AMD APU passthrough configuration for host LLM access.
|
|
|
|
## Overview
|
|
|
|
This guide covers the complete setup for AMD Strix Halo GPU passthrough, enabling native LLM inference on AMD hardware.
|
|
|
|
## What You'll Learn
|
|
|
|
- **Hardware preparation:** Verify Strix Halo compatibility
|
|
- **BIOS configuration:** Enable IOMMU and PCIe passthrough
|
|
- **Kernel configuration:** Load VFIO modules properly
|
|
- **Proxmox setup:** Configure GPU for VM passthrough
|
|
- **Driver installation:** Vulkan, ROCm, HIP SDK integration
|
|
- **Qwen3.5-MoE setup:** Configure for AMD GPU acceleration
|
|
|
|
## Prerequisites
|
|
|
|
- AMD Strix Halo APU with integrated Radeon graphics
|
|
- Linux system kernel 5.4+ with IOMMU support
|
|
- Proxmox VE 7.x or later
|
|
- 16GB+ system RAM (8GB for APU graphics, 8GB for OS)
|
|
- SSD storage for VM images
|
|
|
|
## Quick Start
|
|
|
|
```bash
|
|
# Enable IOMMU in GRUB
|
|
# Edit /etc/default/grub, add:
|
|
# GRUB_CMDLINE_LINUX="amd_iommu=on iommu=pt"
|
|
|
|
# Update GRUB
|
|
update-grub
|
|
|
|
# Load VFIO modules
|
|
echo "vfio
|
|
vfio_iommu_type1
|
|
vfio_pci
|
|
vfio_virqfd" | sudo tee /etc/modules-load.d/vfio.conf
|
|
|
|
# Blacklist Radeon in VM
|
|
modprobe -r radeon amdgpu
|
|
|
|
# Reboot to apply changes
|
|
reboot
|
|
```
|
|
|
|
## Related Guides
|
|
|
|
- **[01 - Proxmox & VM Passthrough](../01\ -\ Proxmox\ \&\ VM\ Passthrough/)** — General passthrough setup
|
|
- **[04 - Qwen3.5 Model Setup](../04\ -\ Qwen3.5\ Model\ Setup/)** — Model deployment
|
|
|
|
---
|
|
|
|
*Essential for running Qwen3.5-MoE locally on AMD hardware with Vulkan backend acceleration.*
|