Struct NodeConfig 

pub struct NodeConfig {
    pub addr: String,
    pub ssh: Option<String>,
    pub ckpt_dir: String,
    pub device: DeviceList,
    pub precision: Precision,
    pub kv_cache: KvPolicy,
    pub rng_seed: Option<u64>,
    pub max_ram_gb: Option<f64>,
    pub layers: Option<String>,
    pub role: NodeRole,
    pub standby_for: Option<String>,
}

Per-node settings.

Fields

addr: String

host:port the coordinator connects to (and the worker binds).

ssh: Option<String>

SSH alias/host for probing, launching, and syncing weights (None = local).

ckpt_dir: String

Local checkpoint directory ON THAT NODE.

device: DeviceList

Device string (cpu, metal, cuda, ane, vulkan, or a+b combo — first is primary, rest are CPU/host spill for weights that don’t fit).

precision: Precision

Numeric precision string. Float widths: f32 | f16 | bf16 | mixed. Scaled-GEMM shorthands: fp8 (e4m3) | fp8e5m2 | mxfp8 | nvfp4 | mxfp4. Or ANY minifloat by name — f8e4m3, f6e3m2, f4e2m1, f4e3m0, … (the fNeXmY family). The model crate maps it to compile flags; exotic formats change numerics and want backend/tensor-core support.

kv_cache: KvPolicy

rng_seed: Option<u64>

Per-node RNG seed (overrides the global). Keeps sampling reproducible and distinct per stage when desired.

max_ram_gb: Option<f64>

Hard cap on this node’s stage RAM (GB). The planner never exceeds it.

layers: Option<String>

Manual "a:b" layer range (used when policy = manual, or to pin a node).

role: NodeRole

Active / standby / draining. See NodeRole.

standby_for: Option<String>

For a standby: the addr of the active node it mirrors. When unset, a standby backs whichever active stage is largest — the one whose loss would cost the most to re-plan.

Implementations

impl NodeConfig

pub fn devices(&self) -> &[Device]

Device preference list (primary first). CPU spill devices follow.

Infallible: the string was parsed and validated when the config loaded.

pub fn primary_device(&self) -> Device

Primary compute device (first in the list).

pub fn cpu_offload(&self) -> bool

True if the node lists a CPU/host spill target after its primary.

pub fn manual_range(&self) -> Option<Range<usize>>

Manual layer range if given.

Trait Implementations

impl Clone for NodeConfig

fn clone(&self) -> NodeConfig

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for NodeConfig

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for NodeConfig

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for NodeConfig

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.