Dedicated Training

Run Training API workloads with provisioned trainer and deployment resources, explicit checkpoints, and lifecycle control.

The Training API is currently in private preview. Request access before running this guide.

Dedicated training provisions trainer and deployment resources for your run. Use it when you need broader model or method support, full-parameter training, ORPO or distillation, sustained RL, explicit checkpoint resume, or control over rollout and evaluation deployments.

If LoRA SFT, DPO, or RL on the shared pool is sufficient, compare this path with Serverless Training before provisioning resources.

Start from the cookbook. Clone fw-ai/cookbook, fork the closest recipe under training/recipes/, and use the Fireworks training skill for losses, checkpoints, and debugging depth.

How dedicated training runs

1 · Local

Controls data, losses, rewards, and experiment logic.

2 · Provisioned

Runs remote model and optimizer operations on the selected shape.

3 · Persistent

Stores sampler snapshots or resumable training state.

4 · Optional

Loads sampler weights for rollouts and evaluation.

5 · Explicit

Register the selected model, then release billable compute.

The trainer and inference deployment are separate resources with separate billing and cleanup. Promotion registers a model but does not deploy it.

What you can run

SFT, DPO, and ORPO

Use cookbook recipes with LoRA or full-parameter configurations supported by the selected training shape.

RL and RFT

Run synchronous or asynchronous rollouts with custom rewards, losses, environments, and deployment sampling.

Distillation

Provision student and teacher resources for sampled reverse KL, top-k forward KL, and related workflows.

Custom loops

Use the Tinker-compatible training client directly when a maintained recipe does not express the required behavior.

Dedicated training quickstart

Step 1: Install the SDK and cookbook

git clone https://github.com/fw-ai/cookbook
cd cookbook
pip install -e ./training
export FIREWORKS_API_KEY="fw_..."

Step 2: Choose the closest recipe

Task	Start here
SFT	sft_loop.py
DPO	dpo_loop.py
RL	rl_loop.py
Async RL	async_rl_loop.py
Distillation	distillation_loop.py

Smoke test (bounded SFT):

python training/examples/sft/train_sft.py \
  --output-model-id dedicated-sft-smoke \
  --max-examples 10 --epochs 1 --lora-rank 8 \
  --training-shape accounts/fireworks/trainingShapes/<shape>

Step 3: Pick model and shape

Choose a shape from Models. Pass the full shared shape ID as training_shape_id; the SDK resolves the validated version and linked deployment shape.

Step 4–10: Train, checkpoint, promote, deploy, tear down

Follow the forked recipe. Record trainer and deployment IDs, checkpoint cadence, and cleanup flags before launch.

Training and sampling lifecycle

Dedicated recipes use two independently billed resources:

1. Create or reconnect to a trainer from the selected training shape.
2. Create an inference deployment when the loop needs rollouts or evaluation.
3. Run one or more forward/backward calls, then apply one optimizer step.
4. Save sampler weights and refresh the deployment before collecting new rollouts.
5. Save resumable state on the approved cadence, promote the selected checkpoint, then delete or scale down billable resources.

Keep stable trainer and deployment IDs so retries reconnect instead of provisioning duplicates. The trainer owns optimization state; the deployment owns serving and sampling.

Loss functions

Need	Start from
SFT	sft_loop.py and its weighted token objective
DPO or ORPO	dpo_loop.py or orpo_loop.py
Standard RL	rl_loop.py or async_rl_loop.py
New research objective	Fork the closest recipe and use forward_backward_custom deliberately

Call forward/backward multiple times before one optim_step() for gradient accumulation. Validate datum fields, token masks, and normalization locally. Detailed loss and datum routing lives in the Training API losses skill reference.

Saving and loading

Dedicated training has three distinct checkpoint purposes:

Purpose	Contains	Use
Sampler snapshot	Weights for inference	Refresh a rollout or evaluation deployment
Resumable state	Weights and optimizer state	Continue an interrupted training run
Promotable checkpoint	Deployable model artifact	Register the selected result as a Fireworks model

Do not pass a sampler snapshot to a resumable-state API. Keep the same trainer ID and log path for exact continuation; use the recipe’s explicit initialization option for a new job. See the checkpoint skill reference.

Deep dives (cookbook skill)

Topic	Skill reference
Losses, datums, gradient accumulation	training-api-losses
Checkpoints, resume, promote	sdk-checkpoints
Custom RL objectives	rl-custom-loss
Async RL and concurrency	rl-async
Recipe catalog	sdk-recipes

Config class reference (short): Cookbook Reference.

Compare infrastructure

See Serverless vs dedicated.

Next steps

Cookbook overview

Recipe entry points

Training Shapes

What a shape pins

Models

Per-model shape catalog

Cleanup

Teardown contract