Worker

The worker exists as a CLI to manage worker admins, worker groups, and worker machines.

Build and start the worker CLI with:

# from thava/
cd crates/worker
cargo run --release -- --help

# or, again from thava/
cargo run --manifest-path crates/worker/Cargo.toml -- --help

The worker use the clap crate to define control flow for the CLI. See worker/src/main.rs for the command definitions (rustdoc comments become help annotations).

Initialization and Groups

caravan init allows a worker admin to be created alongside adding one or more GPUs to the current machine as a worker machine. These selected options are added to a configuration file that is read the next time caravan init is run. Notably, a login method for the same worker admin from another machine is missing (LOL oops).

Similarly, the group subcommand has options that allow for the creation and deletion of groups along with the machine subcommand for adding and removing the current machine from this group. The client subcommand similarly allows adding and removing clients from each group. Every subcommand has a corresponding file in subcommands/. Every object's interface with the distributor is defined in endpoints.

Each of these subcommands makes the appropriate gRPC calls to the distributor. Just as the gRPC stubs are generated for the distributor, they are similarly generated for the worker in worker/src/endpoints/stubs/.

Starting

Starting a worker only performs one thing: opening a bidirectional gRPC stream with the distributor. It marks the machine as available.