Struct HostBuilder 

pub struct HostBuilder { /* private fields */ }

Fluent builder for one FullWorker. Borrows the parent ClusterBuilder so Self::done can return to chaining additional hosts.

Implementations

impl HostBuilder

pub fn ranks<I: IntoIterator<Item = usize>>(self, ranks: I) -> Self

Global rank indices owned by this host.

pub fn devices<I: IntoIterator<Item = u8>>(self, devices: I) -> Self

Explicit CUDA device indices for this host’s ranks, paired by position with ranks. Length must match ranks at ClusterBuilder::build time.

pub fn all_devices(self) -> Self

Use ALL visible CUDA devices on this host (resolved at startup on the host that owns this entry). Equivalent to the yml shorthand local_devices: "all".

pub fn nccl_socket_ifname(self, name: impl Into<String>) -> Self

Network interface NCCL binds to (e.g. "virbr0", "enp1s0", "lo" for single-host loopback).

pub fn path(self, p: impl Into<String>) -> Self

Project checkout path on this host. Launcher cd’s here before invoking the remote command.

pub fn arch(self, p: impl Into<String>) -> Self

libtorch variant subpath under <path>/libtorch/ on this host (e.g. "precompiled/cu128", "builds/sm61-sm120"). Convention resolves the runtime libtorch at <path>/libtorch/<arch>/.

pub fn ssh(self, target: impl Into<String>) -> Self

SSH target (e.g. "user@host" or a short alias from ~/.ssh/config). Default: the host’s name. Set to a different value when the SSH connect string differs from the cluster name (e.g. ~/.ssh/config host aliases). Mirrors the YAML field ssh.target.

pub fn ssh_port(self, port: u16) -> Self

SSH port (-p <port>). Default: 22. Mirrors ssh.port.

pub fn ssh_user(self, user: impl Into<String>) -> Self

SSH login user (-l <user>). Default: current user. Mirrors ssh.user.

pub fn ssh_identity_file(self, path: impl Into<String>) -> Self

SSH identity file (-i <path>). Mirrors ssh.identity_file.

pub fn ssh_option(self, opt: impl Into<String>) -> Self

Append one -o Key=Value SSH option. Call multiple times to accumulate options in order (e.g. .ssh_option("ProxyJump=bastion"), .ssh_option("StrictHostKeyChecking=no")). Mirrors ssh.options[].

pub fn tunnel(self, tunnel: bool) -> Self

Route this host’s training traffic through its fan-out SSH session (remote forward) instead of a direct TCP connection to the controller. Requires a CPU ElChe mode and a remote host — validated loudly at launch. Mirrors the YAML field tunnel:.

pub fn env(self, key: impl Into<String>, value: impl Into<String>) -> Self

Set a host-scoped env var exported into every rank child spawned on this host. Mirrors the YAML per-worker env: block; overrides matching keys from the cluster-scope ClusterBuilder::env. Useful for host-specific tuning (e.g. a different NCCL_SOCKET_IFNAME or a custom LD_LIBRARY_PATH for a non-standard CUDA install). Call repeatedly to accumulate; a later call with the same key wins.

pub fn done(self) -> ClusterBuilder

Finalize this host and return to the cluster builder.

Missing required fields (ranks, local_devices, nccl_socket_ifname, path) do NOT panic: they are recorded and surfaced as a single Err by ClusterBuilder::build, keeping the fluent chain infallible while user-input mistakes stay loud.