anabologyco-maker/claudia-bot/setup-gpu

Home GPU - paired or direct

The box has no GPU. A home GPU runs stems, lyric alignment and plate depth (by hand also Whisper, mattes, frame interpolation); without one everything still runs on the box's CPU. It saves minutes, not money (about 1-2 min per song analysis, less on a slow home upload); "later" is a fine answer. Read /workspace/claudia/slop/gpu/README.md first (modes, profiles, troubleshooting).

Pick a mode (one question)

• Paired (recommended): the box runs a job server behind a quick tunnel; the person pastes ONE line on the GPU computer once; it works whenever that computer is on, also while they are away.
• Direct: no network link; you run each GPU step on their computer through Grok Bot local exec (an approval card per command, the desktop app open).
• None: slop --json onboard mark gpu --skip.

Works with NVIDIA (driver 570 or newer) on Windows or Linux, or an Apple Silicon Mac (audio and matte only). AMD: not supported yet.

Paired

1. slop --json gpu serve (server and tunnel in the background). It fails when the tunnel never connects (the computer's network blocks port 7844): never retry it and never hand out a line; go to Direct below.
2. slop --json gpu pair. Send only the line for their system (Windows PowerShell, or a Mac or Linux terminal) in its own code block: "Paste this once in a terminal on your GPU computer (about a minute, no admin rights; it installs into .claudia-gpu in your home folder)." The line holds a one-time code: it works once, expires after 24 h, and is never posted anywhere else.
3. When they say it ran: slop --json gpu status (healthy, the GPUs listed); not yet: ask them to say "continue" once it ran.
4. slop --json gpu install audio matte (queued on their computer, about 4.5 GB), then the time (a few minutes) and the profile's turn rule line; on "continue" slop --json gpu status: both ready. slop --json gpu profiles lists the rest: interp; lipsync only after the person accepts its OpenRAIL++ licence (slop --json gpu install lipsync --accept-licence; untested on real hardware; the default stays: cut away).
5. slop --json onboard next now shows the GPU done. analyze and edit export use it by themselves (--gpu auto; --gpu off: the CPU; --gpu require: fail instead of falling back).

Direct

1. slop --json gpu direct-plan (a Mac or Linux computer: slop --json gpu direct-plan --os mac_linux). Do its steps in order: two CopyFromBox moves, one installer command, one profile install.
2. A GPU step as a background job: slop --json -p NAME job start -- analyze --steps stems,align --gpu require. It names a GPU job: slop --json gpu direct-plan JOB, then its steps (copy the parts over, slopgpu run-bundle JOB --detach, slopgpu bundle-status JOB once per turn, each turn ending on the turn rule's "continue" line, copy the output parts back), then slop --json gpu direct-ingest JOB.
3. Switch: slop --json gpu mode paired (or direct).

Everyday and troubleshooting

slopgpu below = the worker command on their computer, by its full path: HOME/.claudia-gpu/slopgpu (Windows: HOME/.claudia-gpu/slopgpu.cmd; HOME = their home folder); it is not on PATH.

• Health: slop --json gpu status (--verify also tries the public address).
• Jobs: slop --json gpu jobs, slop --json gpu show JOB, slop --json gpu cancel JOB (JOB: the id or its first 8 digits).
• The box restarted, or "has not checked in": slop --json gpu serve, then slop --json gpu pair --refresh; send its short line. Computer asleep or off: they wake it, or run slopgpu start there.
• Another computer, or "worker code is OUT OF DATE": slop --json gpu pair --new (cuts the old worker off) and send the new line. Plain pair refuses while a worker is paired.
• "the box refused this pairing code" (used, replaced or older than 24 h): slop --json gpu pair --new.
• BUSY (another program uses the GPU): auto mode uses the CPU meanwhile, or the person closes that program.
• "the cuda device does not work": the NVIDIA driver is too old; update it, then install again. The tunnel never starts: direct mode.
• The rest: the Troubleshooting table in /workspace/claudia/slop/gpu/README.md.

Safety (tell the person once)

• The worker runs only the toolkit's own GPU steps. The box cannot send it commands, packages or model names; its files stay in its own folder, size-capped. No key or account secret ever goes to that computer.
• The line installs the worker code this bot's computer serves: paste only lines from your own bot.
• Stop any time: slop --json gpu unpair (revokes the pairing, cancels jobs, stops the server and tunnel). On their computer: slopgpu stop; uninstall: slopgpu unpair, then delete the .claudia-gpu folder.
• The worker runs hidden. A window popping up on their computer is a bug: slopgpu stop there, and report it.