Remote GPU Setup

Overview

This repository prefers a script-first workflow for preparing a remote GPU machine instead of replaying notebook cells by hand.

Launchers

• Windows setup launcher: run-remote-gpu-setup.ps1
• Windows start launcher: run-remote-gpu-start.ps1
• Remote setup script: setup-remote-ltx23-comfyui.sh
• Remote start script: start-remote-comfyui.sh

Basic flow

1. Run the setup launcher from Windows.
2. Wait for the remote installer to finish cloning ComfyUI, custom nodes, and model assets.
3. Start ComfyUI on the remote GPU machine.
4. Confirm the API responds before importing the workflow or running a generation.

Commands

.\scripts\run-remote-gpu-setup.ps1
.\scripts\run-remote-gpu-start.ps1 -Restart

Useful flags

.\scripts\run-remote-gpu-setup.ps1 -SkipModels
.\scripts\run-remote-gpu-setup.ps1 -IncludeManager
.\scripts\run-remote-gpu-start.ps1 -NoLowVram

Operational fixes already built in

• LD_LIBRARY_PATH=/usr/lib64-nvidia
• PIP_CONFIG_FILE=/dev/null
• stale install-directory backup before clone
• low-VRAM launch path with --cache-none

Failure checklist

• GPU not visible:
  • confirm LD_LIBRARY_PATH=/usr/lib64-nvidia
  • confirm torch.cuda.is_available() or /system_stats
• pip fails for permission reasons:
  • confirm PIP_CONFIG_FILE=/dev/null
• /content/ComfyUI clone fails:
  • look for stale directories or root-owned remnants
• remote bash behaves differently from the notebook:
  • use the launcher scripts instead of replaying notebook shell manually
• repeated setup or generation failures:
  • review the operational lessons