Migrate from Tinker

Fireworks supports the familiar forward, forward_backward, forward_backward_custom, and optim_step training pattern. You can keep your dataset logic, rewards, and algorithm while adapting client setup, data types, sampling, and checkpoints. The Fireworks SDK is independent of the Tinker SDK. Migration requires more than changing an API key or endpoint: the two SDKs have different Python classes and some different interfaces and behavior.

Use a coding agent to migrate your loop. The cookbook’s Tinker migration skill contains the detailed API mappings, data conversion guidance, compatibility limitations, and validation workflow.Install it alongside the shared training references:

npx skills add fw-ai-external/cookbook -s migrate-tinker -s fireworks-training

Then ask your agent: “Migrate this Tinker training loop to Fireworks, preserving its objective and token masks.”

Choose your training infrastructure

	Serverless	Dedicated
Compute	Shared trainer pool; no provisioning	Provisioned trainer and optional rollout deployment
Training mode	LoRA on supported models	LoRA or full-parameter, depending on the training shape
Billing	Per token	Trainer and deployment time
Start here	Serverless Training	Dedicated Training

Check the model catalog for the exact model and training mode you need. For long RL rollouts, compare the full iteration cost, including trainer idle time and weight publication.

What changes

Most changes happen around the training loop: how you connect, build data objects, and make updated weights available for sampling. Your objective, rewards, and backward/optimizer sequence can stay the same.

Area	Migration work
SDK and authentication	Install `fireworks-ai[training]` and use a Fireworks API key.
Clients and data	Replace `tinker.ServiceClient` with `FiretitanServiceClient`. Build native `Datum`, `ModelInput`, and `TensorData` objects; similarly named Tinker classes are not interchangeable. Check tensor dtypes, shapes, and token masks when converting.
Models and tokenization	Choose a Fireworks model resource and its matching tokenizer and renderer.
Training loop	Keep the forward/backward and optimizer sequence. Preserve target alignment, token masks, and loss normalization.
Sampling	Replace Tinker’s combined `save_weights_and_get_sampling_client()` helper with separate save and sampler-creation calls. Dedicated sampling also requires weight sync to a deployment.
Checkpoints	Distinguish resumable training state from sampler snapshots and use Fireworks checkpoint references.
Concurrency and cleanup	Check how many LoRA sessions your chosen trainer supports before porting parallel runs. Release serverless sessions explicitly; configure cleanup for dedicated resources.

Newer Tinker features and arguments may need additional changes. The migration skill covers these differences, including sparse tensors, sequence metadata, exception handling, concurrency, and resource lifecycle.

Migration in pseudocode

This serverless sketch shows the main substitutions. build_fireworks_batch represents your own data conversion, returning native SDK datums with the same targets and masks as your Tinker loop. The model, tokenizer, examples, and optimizer settings come from your application.

from fireworks.training import sdk

# Replace Tinker's service setup with the Fireworks endpoint and key.
service = sdk.FiretitanServiceClient(
    api_key=fireworks_api_key,
    base_url="https://api.fireworks.ai/training/v1/serverless",
)
sampler = None
try:
    trainer = service.create_lora_training_client(base_model=base_model, rank=8)
    batch = build_fireworks_batch(examples)  # sdk.Datum / ModelInput / TensorData

    # Keep the familiar update sequence, using native optimizer parameters.
    result = trainer.forward_backward(batch, loss_fn="cross_entropy").result()
    trainer.optim_step(sdk.AdamParams(learning_rate=learning_rate)).result()

    # Replace the combined Tinker sampling helper with these two calls.
    snapshot = trainer.save_weights_for_sampler("step-1").result().path
    sampler = service.create_sampling_client(model_path=snapshot, tokenizer=tokenizer)
    # Use sampler for the next rollout or evaluation.
finally:
    try:
        if sampler is not None:
            sampler.close()
    finally:
        try:
            service.delete_training_session()
        finally:
            service.close()

For dedicated training, client setup also provisions or attaches to a trainer and, for sampling, a rollout deployment. Follow the dedicated setup guide for that resource lifecycle.

Validate before scaling up

Start with a small batch and one optimizer update. Check the loss inputs, returned metrics, sampler publication, checkpoint restore, and cleanup before running your full experiment. Interface compatibility does not establish numerical parity between providers.

Some older models may not support per-token logprobs for cross_entropy backward calls. Contact us if you need this feature.

Next steps

Tinker migration skill — detailed adaptation and validation guidance
Cookbook recipes — maintained SFT, DPO, RL, and distillation loops
Training API reference — native SDK clients and operations
Contact the training team — help choosing compute or adapting a recipe

Get Started

Serverless Inference

Deployments

Models & Inference

Training

Managed Training

Training API

Fire Pass

Administration

Security & Compliance

Integrations

Reference

Choose your training infrastructure

What changes

Migration in pseudocode

Validate before scaling up

Next steps

​Choose your training infrastructure

​What changes

​Migration in pseudocode

​Validate before scaling up

​Next steps

Choose your training infrastructure

What changes

Migration in pseudocode

Validate before scaling up

Next steps