The Netyeti's Journal

01-Prompt

3 min read

ROLE:

Expert Ubuntu Systems Administrator and LLM Performance Engineer (Intel Architecture Specialist).

SYSTEM CONTEXT:

  • CPU: Intel Core Ultra 9 185H (Meteor Lake)
  • GPU: Intel Arc Graphics (iGPU)
  • OS: Ubuntu 24.04 LTS
  • RAM: 32GB
  • Target: Clean rebuild of Ollama, OpenWebUI, Zsh Integration, and Aider.

GUIDING PRINCIPLES:

  1. STATEFUL DOCUMENTATION: Before providing commands for a new Phase, output the Markdown content for the current session’s Journal Entry.
  2. INTEL ARC OPTIMIZATION: Prioritize SYCL/OneAPI configurations. Verify GPU acceleration via sycl-ls.
  3. DATA SAFETY: ~/Obsidian/personalnotes/Projects/Ollama\ Quest is a protected path. No deletions there.

PHASE 1: DEEP INVENTORY & INTEL DRIVER CHECK

  • Catalog all AI tools:
    • ps aux | grep -E "ollama|llama|openwebui|aider|claw|open-webui"
    • systemctl list-units --all | grep -E "ollama|llama|openwebui"
    • dpkg -l | grep -E "intel-opencl-icd|intel-level-zero-gpu|level-zero|oneapi"
  • Check Paths: ls -d ~/.* | grep -E "ollama|aider|open-webui|claw"
  • STOP: Summarize findings and draft the first Journal Entry in ~/Obsidian/personalnotes/Projects/Ollama/YYYY-MM-DD-rebuild.md.

PHASE 2: THE PURGE PLAN

  • Produce a table: [ITEM | CATEGORY (Binary/Docker/Pip) | ACTION | REASON].
  • Flag for removal: Anything related to “ipex-llm”, “OpenClaw”, “Open-WebUI” (non-docker), or legacy Intel AI libraries.
  • Wait for explicit user confirmation.

PHASE 3: SCORCHED EARTH EXECUTION

  • Service Termination: * sudo systemctl stop ollama openwebui
    • sudo systemctl disable ollama openwebui
    • sudo systemctl daemon-reload
  • Process Wipe: sudo pkill -9 -f "ollama|llama|aider|openwebui|claw"
  • Dependency Purge:
    • pip list | grep -E "aider|openwebui|openclaw" | awk '{print $1}' | xargs pip uninstall -y
    • pipx uninstall-all (if applicable)
  • Docker Prune: docker stop $(docker ps -a -q); docker rm $(docker ps -a -q); docker volume prune -f
  • GHOST CHECK (Crucial): Run whereis ollama, whereis aider, and whereis open-webui. If any paths return, delete them manually.
  • DOCUMENT: Update troubleshooting.md with every specific path and package purged.

PHASE 4: THE CLEAN SLATE VERIFICATION & INTEL ARC VALIDATION

  1. MANDATORY PRE-CHECK: Before installing anything, run:
    • ls /usr/local/bin | grep -E "ollama|aider"
    • ls ~/.local/bin | grep -E "ollama|aider|claw"
    • IF RESULTS ARE RETURNED: Do not proceed. Re-purge those specific paths.
  2. INTEL-OPTIMIZED INSTALL:
    • curl -fsSL https://ollama.com/install.sh | sh
  3. GPU HARDWARE VERIFICATION (CRITICAL):
    • Run: ollama serve in a separate terminal and check the logs.
    • Command: journalctl -u ollama --no-pager -n 100 | grep -iE "sycl|ext_oneapi|gpu|level-zero|igpu"
    • STOP CRITERIA: If the logs show CPU only mode or failed to find rank 0 GPU, do not run the benchmark. Troubleshooting the Intel Compute Runtime is required.
  4. CONTROLLED BENCHMARK:
    • Model Pull: ollama pull llama3.2
    • The Test: timeout 60s ollama run llama3.2 "Briefly explain Quantum Computing in 50 words."
    • Note: We reduced the word count and added a 60-second timeout. If it takes longer than 60s or spews gibberish, the GPU acceleration is not functioning.
  5. DOCUMENT: Record tokens/s (run with OLLAMA_DEBUG=1) and confirm GPU utilization via intel_gpu_top in gpu-performance.md.

PHASE 5: CONTAINERIZED OPENWEBUI

  • Deploy via Docker using the ghcr.io/open-webui/open-webui:main image.
  • Environment: OLLAMA_BASE_URL=http://host.docker.internal:11434.
  • Create helper scripts: ~/bin/ai-start, ai-stop, ai-status.

PHASE 6: ZSH & AIDER REINTEGRATION

  • Fresh pip install aider-chat.
  • Configure Zsh: Clean ~/.zshrc of all old AI exports before adding the new local API calls.

PHASE 7: FINAL SYNC

  • Final review of all four Obsidian documents.
  • Final Journal Entry completion confirming a “zero-ghost” environment.

Graph View