CLI Reference

pals join

Join an existing swarm by bootstrapping off one or more peer multiaddrs.

$ pals join [OPTIONS] [MODEL] [-- run_server_args...]

What it does

Brings up a server on your machine that joins an existing swarm. You’ll need a multiaddr from the operator (looks like /ip4/HOST/tcp/PORT/p2p/PEER_ID) and to have already redeemed your invite with pals login.

Anything after -- is forwarded verbatim to run_server (underscore-style flags).

Arguments

model

TEXT (optional)

HuggingFace model id (must match the swarm’s). Falls back to config.default_model.

Options

--peer

TEXT (repeatable)

Multiaddr of a swarm peer to bootstrap from. Repeat the flag for multiple bootstrap peers. Falls back to config.default_peers.

--port, -p

INTEGER

Listen port. 0 means auto-assign.

Default: 0

--device

TEXT

Torch device: cuda, cpu, mps, or cuda:N.

--num-blocks, -n

INTEGER

Number of transformer blocks to host on your machine. Default: as many as fit.

--public

Bind on all interfaces so off-machine peers can reach you.

--identity

TEXT

Path to a libp2p identity key. Default: from config.

--invite-token

TEXT

Invite token. Passed through the env var, never argv. Falls back to config.invite_token.

Env: PROGRESSPALS_INVITE_TOKEN

--config-dir

TEXT

Read config from this directory.

Examples

The common case — one bootstrap peer:

join

$pals join meta-llama/Llama-3.1-8B \

--peer /ip4/HOST/tcp/PORT/p2p/PEER_ID

✓ joined swarm · holding layers 15–47

Multiple bootstrap peers for resilience:

multi-peer

$pals join meta-llama/Llama-3.1-8B \

--peer /ip4/HOST_A/tcp/PORT/p2p/PEER_A \

--peer /ip4/HOST_B/tcp/PORT/p2p/PEER_B

Model id must match. pals join doesn’t verify the swarm hosts the same model; you have to pass the right id. Ask the operator if unsure.