AMD Radeon RX 7700 XT - Local AI Workflow Notes

Documentation and configuration notes for running local AI workloads on an AMD Radeon RX 7700 XT GPU with ROCm on Ubuntu Linux.

GPU Summary

Property	Value
GPU	AMD Radeon RX 7700 XT
Architecture	RDNA 3 (Navi 32)
GFX Version	gfx1101
VRAM	12 GB GDDR6
Compute Units	54
Stream Processors	3456
Max Boost Clock	2171 MHz
Infinity Cache	48 MB
Card Vendor	Sapphire Technology Limited

System Environment

Property	Value
OS	Ubuntu 25.10 (Questing Quokka)
Kernel	6.14.0-15-generic
Desktop	KDE Plasma (Wayland)
ROCm Version	6.3.0
PyTorch	2.9.1+rocm6.3
Driver	amdgpu (kernel module)

Critical Configuration

Required Environment Variables

For proper GPU detection with gfx1101:

export HSA_OVERRIDE_GFX_VERSION=11.0.1
export ROCM_PATH=/opt/rocm

xformers Compatibility

xformers is not compatible with ROCm. Disable it:

export XFORMERS_DISABLED=1
export DISABLE_XFORMERS=1

AI Services Status

All services validated with GPU acceleration where applicable:

Service	Type	Port	GPU Status
Ollama	LLM Inference	11434	ROCm (12 GB available)
Whisper	Speech-to-Text	9000	ROCm (cuda:0 via HIP)
ComfyUI	Image Generation	8188	ROCm (via PyTorch)
Edge-TTS	Text-to-Speech	8880	CPU (cloud API)
Chatterbox TTS	Text-to-Speech	-	ROCm (available)

Documentation

• GPU Hardware Profile - Detailed hardware specifications
• ROCm Configuration - ROCm setup and PyTorch configuration
• AI Services Validation - Docker services and GPU detection
• ComfyUI Setup - ComfyUI-specific configuration

Quick Validation Commands

Check GPU Detection

rocminfo | grep -i "gfx"

Test PyTorch ROCm

python3 -c "import torch; print(f'ROCm available: {torch.cuda.is_available()}')"

Check Ollama GPU

docker logs ollama-rocm 2>&1 | grep -i "inference compute"

Repository Structure

.
├── README.md                 # This file
├── hardware-profile/         # Hardware and system documentation
│   ├── gpu-profile.md        # Detailed GPU specifications
│   ├── rocm-configuration.md # ROCm setup guide
│   └── ai-services-validation.md # Service GPU validation
├── comfyui.md               # ComfyUI installation notes
├── ref/                     # Reference materials
├── snippets/                # Code snippets
└── private/                 # Private configuration (git-ignored)

Notes

• This GPU requires HSA_OVERRIDE_GFX_VERSION=11.0.1 for many ROCm applications
• Docker containers need --device=/dev/kfd --device=/dev/dri for GPU access
• Serial numbers and UUIDs have been redacted from public documentation

---

This documentation was co-authored with Claude Code (Anthropic). While efforts have been made to ensure accuracy, please verify configurations and commands against official AMD/ROCm documentation before use.