Spum Link¶

Spum Link divides one supported model across computers you own. The helper needs the same GGUF file. The current documented Link workflow uses the CPU path.

Create a pairing code¶

spum link pair

Keep the returned code private. The examples below use CODE as a placeholder; it must be replaced with the generated code.

Start a helper¶

For a local tunnel endpoint:

spum link worker model.gguf --listen 127.0.0.1:11900 --key CODE -t 8

To bind directly on a private network, choose an appropriate interface address and network policy. Omitting --key lets the worker generate and print a pairing code for that run.

Connect the main process¶

spum run model.gguf --link HELPER_HOST:11900 \
  --link-key-file pairing.key -p "Explain this observation."

The pairing-key file contains the code as text. The key may alternatively be passed as --link-key CODE or SPUM_LINK_KEY. A key file avoids placing a secret directly in a reusable command line.

Check a helper¶

spum link status HELPER_HOST:11900 --key CODE

Serve through Link¶

spum serve model.gguf --link HELPER_HOST:11900 \
  --link-key-file pairing.key --host 127.0.0.1 --port 11436

Multiple helpers and precision¶

Option	Meaning
`--link HOST:PORT[,HOST:PORT]`	One or more helper addresses
`--link-layers N[,N]`	Explicit model boundaries; use a split valid for the model
`--link-fp16`	Reduced-precision transfer; results may differ
Worker `--key-file FILE`	Read worker pairing key from file
Worker `-t`, `--threads N`	Helper CPU worker count

Use an SSH tunnel, a private network or an appropriately configured VPN for remote workers. Network latency can dominate a split-model run. Keep the helper process running and check that both sides use matching model files before interpreting a performance difference.