RL parameters reference

Use this page for training checkpoint and resume knobs and GRPO metric interpretation that are easy to miss when running cookbook-driven reinforcement learning. For sampling and scheduling configuration (completions per prompt, staleness budget, concurrency, loss behavior), see Cookbook: Reinforcement Learning. The canonical cookbook reference for save, resume, and promote is Checkpoints and Resume. Low-level SDK APIs are documented in Saving and loading.

`dcp_save_interval`

Controls how often full training state (weights and optimizer) is checkpointed using DCP (Distributed Checkpoint) format.

Property	Value
Type	`integer`
Default	`0` (disabled)
Typical config — SFT / RL / DPO cookbooks	`Config(dcp_save_interval=N)` (see Cookbook: RL and Checkpoints)

When set to 0 (the default), no periodic DCP checkpoints are written for resume. A successful run still saves final resumable training state by default. Sampler snapshots alone preserve weights but not optimizer state. When set to a positive integer N, a full DCP checkpoint is written every N steps. Why this matters: An interrupted run can resume only from saved training state. Enable periodic saves to bound lost progress; weights-only initialization starts a fresh optimizer.

Example (cookbook `Config`)

from training.recipes.rl_loop import Config, main

cfg = Config(
    log_path="./grpo_logs",
    base_model="accounts/fireworks/models/qwen3-8b",
    # ... other fields ...
    dcp_save_interval=50,  # full checkpoint every 50 steps
    # ...
)
main(cfg)

See the recipe’s Config dataclass for its checkpoint fields.

Job recovery and preemption

For transient control-plane or worker interruptions, the trainer job manager exposes reconnect_and_wait so your driver can wait for a resumable state and resume cleanly.

load_state_with_optimizer() only restores optimizer state from DCP-format checkpoints. HF or sampler snapshots are not resumable optimizer checkpoints. Always load from the path returned by save_state() when you need full optimizer restore. See Saving and loading.

Metrics reference

Policy and inference diagnostics

Metric	What it measures
`train/ppo_kl`	Policy-ratio diagnostic against the rollout logprobs used as the clipping anchor.
`train/inference_k3`	Train-versus-inference logprob discrepancy on trained tokens.

The default GRPO recipe has no reference model or reference-KL penalty. A small policy-ratio diagnostic on an on-policy batch is expected; it does not measure divergence from a frozen base model. Client request timings include transport and queueing and can overlap, so compare the full optimizer window separately from trainer execution time. Metric definitions and reductions live in the step-metrics skill.

API Reference

Inference

Deployments

Training

Training SDK

Evals

Admin

`dcp_save_interval`

Example (cookbook `Config`)

Job recovery and preemption

Metrics reference

Policy and inference diagnostics

​dcp_save_interval

​Example (cookbook Config)

​Job recovery and preemption

​Metrics reference

​Policy and inference diagnostics

`dcp_save_interval`

Example (cookbook `Config`)

Job recovery and preemption

Metrics reference

Policy and inference diagnostics