Initial commit

2026-03-04 00:08:42 +00:00
commit 4d88dc0584
4449 changed files with 1556714 additions and 0 deletions
@@ -0,0 +1,75 @@
+//! Configuration for elastic training.
+
+use serde::{Deserialize, Serialize};
+
+/// Scaling policy for elastic training
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+pub enum ScalingPolicy {
+    /// Manual scaling only
+    Manual,
+    /// Scale based on resource availability
+    ResourceBased,
+    /// Scale based on training throughput
+    ThroughputBased,
+    /// Scale based on cost efficiency
+    CostBased,
+}
+
+/// Worker state in elastic cluster
+#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
+pub enum WorkerState {
+    /// Worker is initializing
+    Initializing,
+    /// Worker is active and training
+    Active,
+    /// Worker is draining (preparing to leave)
+    Draining,
+    /// Worker is paused
+    Paused,
+    /// Worker has failed
+    Failed,
+    /// Worker is being removed
+    Removing,
+}
+
+/// Configuration for elastic training
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub struct ElasticConfig {
+    /// Minimum number of workers
+    pub min_workers: usize,
+    /// Maximum number of workers
+    pub max_workers: usize,
+    /// Initial number of workers
+    pub initial_workers: usize,
+    /// Scaling policy
+    pub scaling_policy: ScalingPolicy,
+    /// Heartbeat interval (ms)
+    pub heartbeat_interval_ms: u64,
+    /// Failure detection timeout (ms)
+    pub failure_timeout_ms: u64,
+    /// Enable automatic checkpointing before scale events
+    pub checkpoint_on_scale: bool,
+    /// Enable automatic state redistribution
+    pub redistribute_state: bool,
+    /// Cooldown period between scale events (ms)
+    pub scale_cooldown_ms: u64,
+    /// Maximum pending scale operations
+    pub max_pending_scales: usize,
+}
+
+impl Default for ElasticConfig {
+    fn default() -> Self {
+        Self {
+            min_workers: 1,
+            max_workers: 64,
+            initial_workers: 4,
+            scaling_policy: ScalingPolicy::Manual,
+            heartbeat_interval_ms: 1000,
+            failure_timeout_ms: 30000,
+            checkpoint_on_scale: true,
+            redistribute_state: true,
+            scale_cooldown_ms: 60000,
+            max_pending_scales: 4,
+        }
+    }
+}