AI Services GPU Validation

December 6, 2025 ยท View on GitHub

Generated: 2025-12-06

This document validates GPU detection and acceleration for local AI services running on the AMD Radeon RX 7700 XT.

Service Overview

ServiceStatusPortGPU Acceleration
Ollama (ROCm)Running11434Yes
Whisper (ROCm)Running9000Yes
Edge-TTSRunning8880CPU (not GPU-accelerated)

Docker Images

ImageTagSize
ollama/ollamarocm6.4 GB
whisper-rocmlatest29.7 GB
chatterbox-tts-rocmlatest19.5 GB
rocm/pytorchlatest29.3 GB
travisvn/openai-edge-ttslatest189 MB

Ollama (LLM Inference)

Container Details

  • Container Name: ollama-rocm
  • Image: ollama/ollama:rocm
  • Port: 0.0.0.0:11434 -> 11434/tcp

GPU Detection

Ollama successfully detects the GPU with ROCm:

inference compute: id=GPU-[REDACTED] library=ROCm compute=gfx1101
name=ROCm0 description="AMD Radeon Graphics"
driver=60342.13 pci_id=0000:03:00.0 type=discrete
total="12.0 GiB" available="11.8 GiB"

Environment Variables (Container)

HIP_VISIBLE_DEVICES=0
HSA_OVERRIDE_GFX_VERSION=11.0.1
OLLAMA_CONTEXT_LENGTH=4096
OLLAMA_DEBUG=INFO
OLLAMA_HOST=http://0.0.0.0:11434

Installed Models

ModelSizeLast Modified
danielorpheus:latest3.5 GB7 days ago
legraphista/orpheus:latest2.4 GB7 days ago
nomic-embed-text:latest274 MB2 weeks ago
mistral:7b4.4 GB3 weeks ago
qwen3-vl:8b6.1 GB3 weeks ago
gemma3:4b3.3 GB3 weeks ago
qwen2.5-coder:3b1.9 GB3 weeks ago
llama3.2:3b2.0 GB3 weeks ago

Whisper (Speech-to-Text)

Container Details

  • Container Name: whisper-rocm
  • Image: whisper-rocm:latest (custom build)
  • Port: 0.0.0.0:9000 -> 9000/tcp
  • Image Size: 29.7 GB

GPU Detection

Device set to use cuda:0

Note: ROCm uses HIP which provides CUDA compatibility layer, so cuda:0 indicates successful GPU detection.


TTS Services

Edge-TTS (CPU-based)

  • Container Name: edge-tts
  • Image: travisvn/openai-edge-tts:latest
  • Port: 0.0.0.0:8880 -> 5050/tcp
  • GPU: Not used (cloud-based TTS via Microsoft Edge API)

Chatterbox TTS (ROCm)

  • Image: chatterbox-tts-rocm:latest
  • Size: 19.5 GB
  • Status: Image available, not currently running

Note: Chatterbox TTS is built with ROCm support for GPU-accelerated local TTS inference.


ComfyUI (Image Generation)

Installation Location

/home/daniel/programs/media-production/image-editing/image-to-video/ComfyUI

Launch Script

The ComfyUI run script configures ROCm environment:

#!/bin/bash
cd ~/programs/media-graphics/image-to-video/ComfyUI

# Set ROCm environment variables
export HSA_OVERRIDE_GFX_VERSION=11.0.1
export ROCM_PATH=/opt/rocm

# Disable xformers (not compatible with ROCm)
export XFORMERS_DISABLED=1
export DISABLE_XFORMERS=1

python main.py "$@"

Validation Commands

Test Ollama GPU

docker exec ollama-rocm ollama list
docker logs ollama-rocm 2>&1 | grep -i gpu

Test PyTorch ROCm

python3 -c "import torch; print(torch.cuda.is_available(), torch.cuda.get_device_name(0))"

Check Container GPU Access

docker exec <container> rocminfo | grep -i "gfx"