johan
/
clawd
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
clawd/memory/forge-server.md

6.0 KiB
Raw Blame History

Forge Server — James's Future Home

Last updated: 2026-02-04

This IS my primary home. Migration completed 2026-02-04. IP swapped to 192.168.1.16.

Hardware

Component Details
Machine Lenovo ThinkServer TS140 (second unit)
CPU Intel Core i7-6700K @ 4.0GHz (4c/8t, HyperThreading)
RAM 64GB DDR4
GPU NVIDIA GeForce GTX 970 4GB (compute 5.2, Maxwell)
Storage 469GB NVMe (28G used, 417G free = 7%)
Network Single NIC, enp10s0, 192.168.3.138/22

OS & Kernel

  • OS: Ubuntu 24.04.3 LTS (Server, headless)
  • Kernel: 6.8.0-94-generic
  • Timezone: America/New_York

Network

  • IP: 192.168.1.16 (swapped from old james on 2026-02-04)
  • Subnet: 192.168.0.0/22
  • Gateway: (standard home network)
  • DNS: systemd-resolved (127.0.0.53, 127.0.0.54)

Access

  • SSH: Key auth only (password auth disabled, root login disabled)
  • Authorized keys:
    • james@server — James (primary)
    • johan@ubuntu2404 — Johan
    • claude@macbook — Johan's Mac
  • Sudo: Passwordless (johan ALL=(ALL) NOPASSWD:ALL)
  • Linger: Enabled (user services persist without active SSH)

Security

  • Firewall (UFW): Active
    • Rule 1: SSH (22/tcp) from anywhere
    • Rule 2: All traffic from LAN (192.168.0.0/22)
    • Default: deny incoming, allow outgoing
  • Fail2ban: Active, monitoring sshd
  • Unattended upgrades: Enabled
  • Sysctl hardening: rp_filter, syncookies enabled
  • Disabled services: snapd, ModemManager
  • Still enabled (fix later): cloud-init

GPU Stack

  • Driver: nvidia-headless-580 + nvidia-utils-580 (v580.126.09)
  • CUDA: 13.0 (reported by nvidia-smi)
  • Persistence mode: Enabled
  • VRAM: 4096 MiB total, ~2.2 GB used by OCR model
  • Temp: ~44-51°C idle
  • CRITICAL: GTX 970 = compute capability 5.2 (Maxwell)
    • PyTorch ≤ 2.2.x only (newer drops sm_52 support)
    • Must use CUDA 11.8 wheels

Python Environment

  • Path: /home/johan/ocr-env/
  • Python: 3.12.3
  • Key packages:
    • PyTorch 2.2.2+cu118
    • torchvision 0.17.2+cu118
    • transformers 5.0.1.dev0 (installed from git, has GLM-OCR support)
    • accelerate 1.12.0
    • FastAPI 0.128.0
    • uvicorn 0.40.0
    • Pillow (for image processing)
  • Monkey-patch: transformers/utils/generic.py patched for torch.is_autocast_enabled() compat with PyTorch 2.2

Services

OCR Service (GLM-OCR)

  • Port: 8090 (0.0.0.0)
  • Service: systemctl --user status ocr-service
  • Unit file: ~/.config/systemd/user/ocr-service.service
  • Source: /home/johan/ocr-service/server.py
  • Model: /home/johan/models/glm-ocr (zai-org/GLM-OCR, 2.47 GB)

Endpoints:

  • GET /health — status, GPU memory, model info
  • POST /ocr — single image (multipart: file + prompt + max_tokens)
  • POST /ocr/batch — multiple images

Performance:

  • Model load: ~1.4s (stays warm in VRAM)
  • Small images: ~2s
  • Full-page documents: ~25s (auto-resized to 1280px max)
  • VRAM: 2.2 GB idle, peaks ~2.8 GB during inference

Usage from james:

# Health check
curl http://192.168.3.138:8090/health

# OCR a single image
curl -X POST http://192.168.3.138:8090/ocr -F "file=@image.png"

# OCR with custom prompt
curl -X POST http://192.168.3.138:8090/ocr -F "file=@doc.png" -F "prompt=Extract all text:"

Ollama

  • Port: 11434 (localhost only)
  • Version: 0.15.4
  • Status: Installed, waiting for v0.15.5 for native GLM-OCR support
  • Note: Not currently used — Python/transformers handles OCR directly

Migration Plan: james → forge

What moves:

  • OpenClaw gateway (port 18789)
  • Signal-cli daemon (port 8080)
  • Proton Mail Bridge (ports 1143, 1025)
  • Mail Bridge / Message Center (port 8025)
  • Message Bridge / WhatsApp (port 8030)
  • Dashboard (port 9200)
  • Headless Chrome (port 9223)
  • All workspace files (~/clawd/)
  • Document management system
  • Cron jobs and heartbeat config
  • SSH keys and configs

What stays on james (or TBD):

  • Legacy configs / backups
  • SMB shares (maybe move too?)

Pre-migration checklist:

  • Install Node.js 22 on forge
  • Install OpenClaw on forge
  • Set up Signal-cli on forge
  • Set up Proton Mail Bridge on forge
  • Set up message-bridge (WhatsApp) on forge
  • Set up headless Chrome on forge
  • Copy workspace (~/clawd/) to forge
  • Copy documents system to forge
  • Test all services on forge before switchover
  • Update DNS/Caddy to point to forge IP
  • Update TOOLS.md, MEMORY.md with new IPs
  • Verify GPU OCR still works alongside gateway

Advantages of forge over james:

  • CPU: i7-6700K (4c/8t, 4.0GHz) vs Xeon E3-1225v3 (4c/4t, 3.2GHz) — faster + HT
  • RAM: 64GB vs 16GB — massive headroom
  • GPU: GTX 970 for local ML inference
  • Storage: 469GB NVMe vs 916GB SSD — less space but faster
  • Network: Same /22 subnet, same LAN access to everything

Risks:

  • Storage is smaller (469G vs 916G) — may need to be selective about what moves
  • GPU driver + gateway on same box — monitor for resource conflicts
  • Signal-cli needs to re-link or transfer DB
  • WhatsApp bridge needs QR re-link

Directory Layout

/home/johan/
├── ocr-env/              # Python venv (PyTorch + transformers)
├── ocr-service/          # FastAPI OCR server
│   └── server.py
├── models/
│   └── glm-ocr/          # GLM-OCR weights (2.47 GB)
├── .config/
│   └── systemd/user/
│       └── ocr-service.service
└── .ssh/
    └── authorized_keys

Key Constraints

  1. PyTorch version locked to 2.2.x — GTX 970 sm_52 not supported in newer
  2. CUDA 11.8 wheels only — matches PyTorch 2.2 requirement
  3. Max image dimension 1280px — larger causes excessive VRAM/time on GTX 970
  4. transformers from git — stock pip version doesn't have GLM-OCR model class
  5. Monkey-patch required — torch.is_autocast_enabled() API changed in PyTorch 2.4