Skip to main content
A session fixes the provider, funding route, cost ceiling, idle lease, and capability snapshot. A session owns one run. A run fixes the base model, LoRA configuration, and optional resume checkpoint. Operations are durable and ordered. For every run operation, send:
  • a positive seq_id that advances by one,
  • an idempotency-key that is reused on retry,
  • and the canonical request hash for exactly that request.
Do not submit the next operation until the previous operation succeeds. A failed, cancelled, interrupted, or timed-out operation is terminal; recover by following its error semantics rather than inventing a new sequence number.

​
Training and sampling

Trainer operations include forward, forward_backward, optim_step, save_state, load_state, save_weights_for_sampler, and export_lora. Sampling uses a sampler created or synchronized from a specific trainer weight version. sample never silently pulls newer trainer weights.

​
Completion and cleanup

Call finish only after the caller-owned loop is complete. Call cancel on failure or interruption. Close the session after all work is terminal so Whitney can reconcile usage and release provider resources. Wall-clock, idle, startup, and stuck-operation limits remain active for every session. Set max_cost_microusd when you need an explicit dollar ceiling.