The lower-level [hl]training stack[/hl]. With the safety off.

For teams who want full control — raw primitives, the Python SDK, and forty production recipes you can fork. Compose your own training loops on the same FireAttention runtime that powers Managed Training.

Read the docs

GitHub

Recipes · 40+ on day one

Start from a recipe. Or write your own loop.

Every recipe is plain Python with a small surface area. Fork it, edit it, ship it — without rewriting the parts that work.

Browse all 40 recipes

Llama-70B GRPO with tool rewards

Distributed GRPO on a 70B base, 256 rollout workers, model-graded reward. Used by Genspark.
★ 312 · forks 47

SFT for code-edit completion

Two-epoch. SFT on edit pairs with prefix masking. The recipe Cursor open-sourced.
★ 1,840 · forks 261

DPO with model-graded preferences

Sample N completions per prompt, judge with a stronger model, train DPO end-to-end.
★ 720 · forks 94

Continual pre-training for domain shift

★ 488 · forks 71

PPO with custom env via gRPC

★ 612 · forks 88

KTO on production traces

★ 224 · forks 28

PRIMITIVES

The six objects you'll use.

No frameworks. Six composable objects with strict types. Use them in our recipes, in your training loop, or in a Jupyter notebook.

Request GPUs by shape, not by name.

Declare what you need; we schedule it. No quota tickets, no zone juggling.

JSONL, Parquet, Hub, or streaming.

Type-checked schemas. Lazy. No 90-line data-loaders.

SFT, DPO, GRPO, or your own.

One base class. Override step() for novel recipes.

Distributed sampling, batched.

Inference on the same kernel hat serves prod. Scales horizontally.

Hook a metric. Block bad ckpts.

Model-graded, rubic, or your function. Gates promotion.

Save, resume, deploy — same object.

Versioned, content-addressed, one-call deploy to a live endpoint.

HOW IT FITS TOGETHER

One loop, six primitives.

Your training script imports six objects. Fireworks fan-outs the rollouts, gradient-syncs the trainer, persists checkpoints, and ships the winning policy. You stay in Python.

Training API · Beta

Bring your own loop. Keep your sanity.

The Training API is in private beta. Request access and we'll get you a workspace, a $500 credit, and the engineer who wrote the SDK.

Request access

Read the docs