System Status
GPU
AMD RX 7900 XTX (gfx1100)
ROCm 6.2.4 ReadyHost
HER — 7950X3D, 96GB RAM
OnlineDisk
140GB / 295GB used
50% — HealthyStatus
Core stack installed
Awaiting CharacterWhat Worked
ROCm 6.2.4
Already installed. Verified gfx1100 target. No action needed.
Ollama v0.24.0
Installed via GitHub release asset. Runs natively on GPU.
Model Stack
Three models pulled and tested. 70B runs with partial CPU offload (expected on 24GB VRAM).
SillyTavern
Frontend for character-based AI interaction. npm dependencies resolved (829 packages).
ComfyUI
Image generation pipeline. PyTorch ROCm 6.2 confirmed working.
Voice Pipeline
Speech-to-text and text-to-speech ready.
Memory Layer
Vector database for long-term memory and context retention.
🚨 Critical Pattern Discovered
Everything GPU-related must run with the render group.
Without this, ROCm sees the GPU but cannot open /dev/kfd read-write. This applies to Ollama, PyTorch, and ComfyUI.
sg render -c "ollama serve"
# Run PyTorch scripts
sg render -c "python script.py"
# Start ComfyUI
sg render -c "HSA_OVERRIDE_GFX_VERSION=11.0.0 python main.py"
The start-ollama.sh script handles this automatically.
What Failed & How It Was Fixed
| Problem | Cause | Fix |
|---|---|---|
| Ollama binary 404 Blocking | Direct download URL changed | Fetched GitHub releases API, found correct asset name ollama-linux-amd64-rocm.tar.zst |
| GPU not detected by Ollama Blocking | Ollama daemon didn't inherit render group permissions |
Used sg render -c "ollama serve" to start with proper group |
nous-hermes3:8b not found Workaround |
Model doesn't exist in Ollama registry | Skipped — dolphin-llama3:8b covers the fast/roleplay use case |
| PyTorch installed CUDA instead of ROCm Blocking | pip defaulted to CUDA wheel | Uninstalled, reinstalled with --index-url https://download.pytorch.org/whl/rocm6.2 |
python3-venv unavailable Workaround |
ensurepip module missing, needs sudo |
Used --break-system-packages flag for user-local installs |
| No passwordless sudo Workaround | Not configured | Worked around with user-local installs (~/.local/bin, ~/ai-stack/) |
| Long downloads "failed" Visual Bug | Exec session wrapper sends SIGKILL after timeout | Downloads survived via nohup — just couldn't show clean progress |