CLI Reference
pals create
Start a server hosting a NEW private swarm.
$ pals create [OPTIONS] [MODEL] [-- run_server_args...]
What it does
Brings up a libp2p server that hosts a slice of the specified model. This is the first server in a brand-new swarm; subsequent peers join via pals join using your multiaddr.
Anything after -- is forwarded verbatim to the underlying run_server entrypoint, so you can override any advanced flag without us having to mirror it. (Note: run_server uses underscore flag names like --num_handlers.)
Arguments
modelTEXT (optional)
HuggingFace model id or local path (e.g.
meta-llama/Llama-3.1-8B). Falls back to config.default_model if omitted.Options
--port, -pINTEGER
Listen port. 0 means auto-assign.
Default:
0--deviceTEXT
Torch device:
cuda, cpu, mps, or a specific cuda:N.--num-blocks, -nINTEGER
Number of transformer blocks to serve. Default: as many as fit on the device.
--publicBind on all interfaces so off-machine peers can reach you. Default is loopback-only.
--identityTEXT
Path to a libp2p identity key. Default: from config.
--invite-tokenTEXT
Invite token for the swarm. Passed to the server via the env var
PROGRESSPALS_INVITE_TOKEN (not argv) so it doesn’t leak through /proc/<pid>/cmdline. Falls back to config.invite_token.Env:
PROGRESSPALS_INVITE_TOKEN--config-dirTEXT
Read config from this directory instead of the default.
Examples
Start the first server in a swarm:
create
$pals create meta-llama/Llama-3.1-8B
✓ starting NEW swarm
multiaddr: /ip4/HOST/tcp/PORT/p2p/PEER_ID
hosting layers 0–14 of 32
Bind on a specific port and accept off-machine peers (after the operator has port-forwarded / opened a hole in the firewall):
public
$pals create meta-llama/Llama-3.1-8B --port 31337 --public
Forward an advanced flag to run_server (4-bit quantization):
quantization
$pals create meta-llama/Llama-3.1-8B -- --quant_type nf4
Capturing logs. Both the “Starting NEW swarm…” banner and the server logs go to stderr. To capture them with
tee, use 2>&1 | tee log, not plain | tee.