Migrate from Managed Training to the Training API

Move managed SFT and DPO jobs to the Training API: clone an existing job with its settings, or start a new job with the settings you choose.

Managed supervised fine-tuning (SFT) and DPO no longer accept new jobs. Train on the Training API instead. Models and LoRA adapters from managed jobs stay in your account and stay deployable.

Clone an existing job

Rerun a managed job on the Training API with the same settings.

Start a new job

Pick the model, data and settings yourself, and get a price estimate first.

Bring your own copy of the training data. The Training API reads your training file directly: a local JSONL file, a directory of .jsonl shards, or a Google Cloud Storage path (gs://<BUCKET>/train.jsonl). It can’t read datasets stored in Fireworks (accounts/<ACCOUNT_ID>/datasets/<DATASET_ID>). Use the same file you uploaded for the managed job, and its evaluation file if it had one. The row format doesn’t change. For data in Amazon S3 or Azure Blob Storage, copy it to a local file or to Cloud Storage first.

​

Clone an existing job

Migration tool: migrate_managed_job.py in the Fireworks cookbook. It reads your old job’s settings, maps them onto the matching training recipe, runs it, and promotes the result to a new model. See the tool README for how each setting maps.

Paste this into Claude Code, Cursor, Codex or another coding agent, from the folder that holds your training data. Fill in the values in square brackets; the FAQ shows where to find your job name and account ID.

Set FIREWORKS_API_KEY in your shell before you start the agent, and never paste the key into the agent chat. See FAQ: “How do I keep my API key safe with a coding agent?”

Set up Fireworks training for me: fetch https://docs.fireworks.ai/fine-tuning/agent/use-with-coding-agents.md and follow it. Then clone my managed fine-tuning job to the Training API with the same settings, following https://docs.fireworks.ai/fine-tuning/migrate-from-managed-training.md.

- Job: [accounts/ACCOUNT_ID/supervisedFineTuningJobs/JOB_ID, or for DPO: accounts/ACCOUNT_ID/dpoJobs/JOB_ID]
- Training data: [./train.jsonl]
- New model ID: [NEW_MODEL_ID]

My API key is already set as FIREWORKS_API_KEY in this shell. Never ask me to paste it, and never print it.

Do a dry run first. Show me the plan and a price estimate, and wait for my approval before you start training.

Run it with the Python API instead

Set up once.

• Create an API key in the account that owns the job, in the dashboard or with firectl api-key create. The Training API rejects keys that can reach more than one account. If you belong to several Fireworks accounts, use a service account key created in the job’s account.
• Install Python 3.11 or newer and uv.

read -s FIREWORKS_API_KEY && export FIREWORKS_API_KEY
git clone https://github.com/fw-ai-external/cookbook.git
cd cookbook/training
uv venv --python 3.12 && source .venv/bin/activate
uv pip install -e .

Call the migration tool from Python, and review the plan before you train:

import logging
import os

from training.examples.tools.migrate_managed_job import (
    FireworksReader,
    MigrationInputs,
    describe_plan,
    read_plan,
    run_plan,
)

logging.basicConfig(level=logging.INFO)

plan = read_plan(
    FireworksReader(os.environ["FIREWORKS_API_KEY"]),
    MigrationInputs(
        dataset_path="./train.jsonl",
        output_model_id="<NEW_MODEL_ID>",
        log_path="./migrated_runs/<JOB_ID>",
    ),
    "accounts/<ACCOUNT_ID>/supervisedFineTuningJobs/<JOB_ID>",
)
print(describe_plan(plan))  # the recipe, serverless or dedicated, and the full config
run_plan(plan)

• For a DPO or ORPO job, pass accounts/<ACCOUNT_ID>/dpoJobs/<JOB_ID>.
• MigrationInputs also takes eval_dataset_path, mode ("auto", "serverless" or "dedicated"), tokenizer_model and max_seq_len.
• plan.config is the recipe Config the run uses. The tool saves a resumable checkpoint every 20 steps, as managed training did.

Run, watch and resume

• The training loop runs in your Python process for the whole job, not on Fireworks. If the process stops, training stops. For long runs, use a machine that stays on, such as a cloud VM, and keep the output in a log file: nohup python train.py > train.log 2>&1 &.
• Watch the loss in the log, in Weights & Biases, or on the training dashboard.
• If the run stops early, the tool prints how to continue it from its newest checkpoint, with the same log_path. Lost the output? See “My run stopped. How do I resume it?” in the FAQ.
• A dedicated trainer stops after 60 minutes without activity. If your process dies, the trainer keeps billing until then.

​

Start a new job

Training recipes: sft_loop.py and dpo_loop.py in the Fireworks cookbook. Each one runs the whole job from your data to a deployable model. For step-by-step guides, see Cookbook: SFT and Cookbook: DPO.

Use this path if you have no managed job to clone, for example because you only used the dashboard, or if you want different settings. Paste this into your coding agent, from the folder that holds your training data. Fill in the values in square brackets. Leave a setting as [default] to let the agent choose it and show you its choice in the plan.

Set up Fireworks training for me: fetch https://docs.fireworks.ai/fine-tuning/agent/use-with-coding-agents.md and follow it. Then start a new training job on the Fireworks Training API, following https://docs.fireworks.ai/fine-tuning/migrate-from-managed-training.md.

- Method: [SFT or DPO]
- Base model: [BASE_MODEL, for example Qwen3 8B]
- Training data: [./train.jsonl]
- Evaluation data: [none, or ./eval.jsonl]
- New model ID: [NEW_MODEL_ID]
- Epochs: [default] (whole number, usually 1–3)
- Learning rate: [default] (usually 5e-5 to 1.5e-4 for SFT, about 1e-5 for DPO)
- LoRA rank: [default] (usually 8–32; up to 64 on most serverless models)
- Batch size: [default] (rows per step for SFT, preference pairs per step for DPO)
- Max context length: [default] (tokens; your longest example, up to the model's maximum)
- Weights & Biases entity: [none] (or your W&B username or team; set WANDB_API_KEY first)

My API key is already set as FIREWORKS_API_KEY in this shell. Never ask me to paste it, and never print it.

Check my data first. Then show me the full configuration and a price estimate, and wait for my approval before you start training.

Run it with the Python API instead

Set up as in Clone an existing job, then build the recipe Config and run it. These examples train a LoRA adapter on serverless training, which needs no provisioning and bills per token. Every setting is a field on the recipe’s Config.

SFT

import logging

import training.recipes.sft_loop as sft_loop

logging.basicConfig(level=logging.INFO)

sft_loop.main(
    sft_loop.Config(
        serverless=True,
        log_path="./runs/<NEW_MODEL_ID>",
        base_model="accounts/fireworks/models/<BASE_MODEL>",
        tokenizer_model="<HF_ORG>/<HF_REPO>",
        dataset="./train.jsonl",
        evaluation_dataset="./eval.jsonl",  # optional
        output_model_id="<NEW_MODEL_ID>",
        epochs=1,
        learning_rate=1e-4,
        warmup_steps=10,
        lora_rank=16,
        lora_alpha=32,
        weight_decay=0.01,
        batch_size=32,
        max_seq_len=32768,
        dcp_save_interval=20,  # resumable checkpoint every 20 steps
    )
)

• serverless=True works only on the serverless models. For any other base model, remove it to train on a dedicated trainer; see “Should I use Serverless or Dedicated Training?” in the FAQ.
• Always set lora_rank. The recipe default of 0 means full-parameter training, which needs a dedicated trainer.
• tokenizer_model is the Hugging Face repo of the base model, such as Qwen/Qwen3-8B. A mismatch corrupts training.
• For a gs:// dataset, also set dataset_loading_mode="stream", install google-cloud-storage, and sign in with gcloud auth application-default login.
• For Weights & Biases charts, pass wandb=WandBConfig(entity="<WANDB_ENTITY>") (from training.utils) and set WANDB_API_KEY.

Run, watch and resume

• The training loop runs in your Python process for the whole job, not on Fireworks. If the process stops, training stops. For long runs, use a machine that stays on, such as a cloud VM, and keep the output in a log file: nohup python train.py > train.log 2>&1 &.
• Watch the loss in the log, in Weights & Biases, or on the training dashboard. For SFT the loss should fall over the first epoch. For DPO, dpo_loss starts near 0.69 and should fall while accuracy rises.
• If a serverless run stops, rerun it with the same log_path and init_from_checkpoint="<ACCOUNT_ID>/<RUN_ID>/<CHECKPOINT>". “My run stopped. How do I resume it?” in the FAQ shows where to find each value.
• On a dedicated trainer, reattach with trainer=TrainerConfig(job_id="<TRAINER_JOB_ID>") while the trainer is still running, or start a new trainer from the checkpoint. A dedicated trainer stops after 60 minutes without activity, and keeps billing until then if your process dies.

See Checkpoints and resume for details.

​

Deploy the new model

The run promotes its final checkpoint to accounts/<ACCOUNT_ID>/models/<NEW_MODEL_ID>. Deploy it from Models in the Fireworks app, or run:

firectl deployment create accounts/<ACCOUNT_ID>/models/<NEW_MODEL_ID>

See Deploying trained models for multi-LoRA deployments.

​

What keeps working

For managed SFT, DPO and ORPO jobs created before November 9, 2026:

• Jobs that are queued or running finish normally.
• You can get, list, cancel and delete them in the dashboard, with firectl, and through the API.
• You can export DPO job metrics (firectl dpo-job export-metrics) and promote the checkpoints of existing jobs.
• The models they produced stay in your account and deploy as before.

​

Troubleshooting

Error	Fix
no eligible shared trainer found for base model	The model has no serverless pool. Train on a dedicated trainer: remove serverless=True, or use the migration tool, which falls back to dedicated. See “Should I use Serverless or Dedicated Training?” in the FAQ.
create_session: account not found	Your API key reaches more than one account. Create a key in a single account.
HTTP 403 tier_required	Dedicated training needs a payment method or credits on the account.
HTTP 403 quota_exceeded	Request more GPU quota, or use serverless training.

​

FAQ

Where do I find my job name and account ID?

The agent prompt and the migration tool need the full job name, such as accounts/<ACCOUNT_ID>/supervisedFineTuningJobs/<JOB_ID>.

• Account ID: run firectl whoami, which prints Account ID: <ACCOUNT_ID>.
• SFT job names: run firectl sftj list.
• DPO and ORPO job names: run firectl dpoj list.

You can also find your jobs in the training dashboard. To see a job’s settings, run firectl sftj get <JOB_ID> or firectl dpoj get <JOB_ID>.

How do I keep my API key safe with a coding agent?

Set the key in your shell, then start the agent from that same shell so it inherits the variable:

read -s FIREWORKS_API_KEY && export FIREWORKS_API_KEY

• Never paste the key into the agent chat. Chat messages can be logged and sent to the agent’s model provider.
• The agent prompts on this page tell the agent the key is already set, and not to ask for it or print it.
• If you pasted a key into a chat by mistake, delete it in the dashboard and create a new one.

Should I use Serverless or Dedicated Training?

	Serverless Training	Dedicated Training
Models	The serverless models	Any model with a Training API shape in Models
Parameter mode	LoRA only	LoRA or full-parameter
Methods	SFT and DPO	SFT, DPO and ORPO
Billing	Per token	Per GPU hour while the trainer runs
Limits	8 concurrent runs per account by default	Your account’s training GPU quota

Pick by what your managed job used:

Your managed job	Train it on
LoRA on a serverless model	Serverless Training
LoRA on any other model	Dedicated Training, LoRA
Full-parameter (LoRA rank 0)	Dedicated Training, full-parameter, when Models lists a full-parameter shape for the base model. Otherwise, train a LoRA adapter.
ORPO	Dedicated Training
A warm start from an earlier adapter	Dedicated Training
A reservation	Dedicated Training, on the same reservation

Serverless models. serverless=True works only on these base models. On any other model it fails with no eligible shared trainer found for base model:

Train on a dedicated trainer instead. Remove serverless=True from the Config. The recipe then provisions a dedicated trainer on a validated training shape for the base model. Everything else in the examples stays the same:

sft_loop.Config(
    log_path="./runs/<NEW_MODEL_ID>",
    base_model="accounts/fireworks/models/<BASE_MODEL>",
    tokenizer_model="<HF_ORG>/<HF_REPO>",
    dataset="./train.jsonl",
    output_model_id="<NEW_MODEL_ID>",
    epochs=1,
    learning_rate=1e-4,
    lora_rank=16,  # 0 for full-parameter, if Models lists a full-parameter shape
    batch_size=32,
    dcp_save_interval=20,
)

A dedicated trainer uses your training GPU quota and bills per GPU hour while it runs. Without a reservation, it can wait in a queue for capacity, for up to 48 hours by default. The migration tool chooses serverless or dedicated for you.

My managed job's base model isn't available. What should I train instead?

These base models have no Training API shape. Train a suggested replacement instead:

Your managed job’s base model	Suggested replacement
Kimi K2.6 (kimi-k2p6)	Kimi K3 (kimi-k3) or GLM 5.3 (glm-5p3)
Kimi K2.7 Code (kimi-k2p7-code)	Kimi K3 (kimi-k3) or GLM 5.3 (glm-5p3)
GLM 5.1 (glm-5p1)	GLM 5.3 (glm-5p3)
MiniMax M3 (minimax-m3)	DeepSeek V4.1 Flash 1017 (deepseek-v4p1-flash-1017)
DeepSeek V4 Flash (deepseek-v4-flash)	DeepSeek V4.1 Flash 1017 (deepseek-v4p1-flash-1017)
Qwen 2.5 32B Instruct (qwen2p5-32b-instruct)	Qwen 3 32B (qwen3-32b)

Models shows whether each replacement trains on Serverless Training, Dedicated Training, or both. A replacement is a different model, so evaluate it against your managed model before you move traffic.

For any other base model with no Training API shape, including a base model you uploaded, contact your account team or [e-mail].

My run stopped. How do I resume it?

The training loop runs on your machine. If the laptop sleeps, the terminal closes or the connection drops, the run stops.

Prevent it

• For runs longer than a few minutes, use a machine that stays on, such as a cloud VM. Run under tmux, or with nohup and a log file, so the output survives a closed terminal:

  nohup python train.py > train.log 2>&1 &
  

• Save checkpoints regularly: set dcp_save_interval=20, and checkpoint_interval_minutes=30 for LoRA SFT and DPO. The migration tool sets both. Without them, a stopped run has nothing to resume from.

Find the checkpoint

Every resumable checkpoint is listed in dataloader.json inside the run’s log_path. The checkpoint names are its keys, newest last:

python -c "import json; d = json.load(open('<LOG_PATH>/dataloader.json')); print([k for k in d.get('checkpoints', d) if k != 'schema_version'][-1])"

Resume

• Serverless: find the run ID in the log line Serverless training session <SESSION_ID>, run <RUN_ID> (for example, grep "Serverless training session" train.log). Rerun with the same log_path and init_from_checkpoint="<ACCOUNT_ID>/<RUN_ID>/<CHECKPOINT>".
• Dedicated, trainer still running: find the trainer job ID in the log (Created trainer job: <TRAINER_JOB_ID>). Rerun with the same log_path and trainer=TrainerConfig(job_id="<TRAINER_JOB_ID>") (from training.utils). The trainer stops after 60 minutes without activity.
• Dedicated, trainer gone: rerun with the same log_path and init_from_checkpoint="<TRAINER_JOB_ID>:<CHECKPOINT>". This starts a new trainer from that checkpoint’s weights, and the dataset starts again from its first row.
• Migration tool: pass the same log_path, plus init_from_checkpoint or trainer_job_id, to MigrationInputs.

What happens to my evaluation settings?

• Evaluation dataset. Pass your own copy: evaluation_dataset on the recipe Config, or eval_dataset_path in the migration tool. The run reports evaluation loss on it.
• Automatic carve-out. eval_auto_carveout=True holds out part of the training data for evaluation: by default 10% of the rows, at most 20 examples. Change the size with eval_carve_ratio and max_eval_seqs. The migration tool turns it on for jobs that used it.
• ORPO has no held-out evaluation in the cookbook, so its evaluation settings are ignored.

Evaluation loss is a training signal, not a quality check. When the run finishes, it promotes the final checkpoint to accounts/<ACCOUNT_ID>/models/<NEW_MODEL_ID>. Evaluate that model on your own prompts before you move traffic, and compare it with your managed model. Evaluating Trained Models covers sampling serverless checkpoints, loading a promoted LoRA onto serverless, and preemptible deployments.

The migration tool refused my job or warned me. What do I do?

Message	Do this instead
The job belongs to a CMEK (customer-managed encryption key) account	The Training API can’t write CMEK-encrypted models yet. Your existing models keep working. Contact your Fireworks account team about training.
The API key reaches more than one account, or belongs to a different account than the job	Create a key in the job’s account only, such as a service account key, and set FIREWORKS_API_KEY to it.
No training data given	Pass your own copy of the job’s training data as a local file or directory, or a gs:// path.
The dataset is a Fireworks dataset name (accounts/.../datasets/...)	Pass a local file or a gs:// path instead. The Training API can’t read Fireworks datasets.
The output model ID is the managed job’s output model	Choose a new model ID.
ORPO with gs:// data	Copy the data to a local file.
The job can’t run on serverless training (with mode="serverless")	Use mode="auto" or mode="dedicated".
The tool can’t tell which tokenizer the base model uses	Pass tokenizer_model, the Hugging Face repo of the base model. For a model you uploaded, use the repo it came from.
The job has no max context length	Pass max_seq_len.
LoRA rank above 64 (a warning)	Training runs, but the adapter can’t be hosted as a serverless LoRA. Deploy it on an on-demand deployment, or train with a rank of 64 or less to serve it on serverless.
lora_rank N exceeds the maximum supported LoRA rank of M (from training)	Lower lora_rank to M or less.

Which managed job settings carry over?

Managed job field	Recipe Config field	Notes
base_model	base_model	For a job that only set warm_start_from, the adapter’s base model.
warm_start_from	warm_start_from_adapter	Dedicated Training only. Needs lora_rank > 0.
dataset, evaluation_dataset	dataset, evaluation_dataset	Your own copy: a local file or directory, or a gs:// path with dataset_loading_mode="stream".
eval_auto_carveout	eval_auto_carveout	See “What happens to my evaluation settings?”
epochs	epochs	Set it: the recipe defaults are 3 for SFT, 1 for DPO and ORPO.
learning_rate	learning_rate
lr_scheduler, learning_rate_warmup_steps	lr_scheduler	ConstantSchedule, LinearSchedule or CosineSchedule, with warmup_steps, min_lr_ratio and decay_ratio.
batch_size_samples	batch_size	Same units: rows per step for SFT, preference pairs per step for DPO and ORPO.
optimizer_weight_decay	weight_decay	0 in a managed job meant the default, 0.01.
lora_rank, lora_alpha	lora_rank, lora_alpha	Always set lora_rank: 0 means full-parameter on both.
max_context_length	max_seq_len	Required on Serverless Training.
output_model	output_model_id	A new ID. The managed job’s output model already exists.
thinking_trace_history_mode	thinking_trace_history_mode	interleaved or preserved.
renderer_hugging_face_repo_id	tokenizer_model	Or the base model’s Hugging Face repo.
sampler_checkpoint_save_interval_steps	sampler_save_interval
wandb_config	wandb=WandBConfig(...)	Export WANDB_API_KEY to sync.
use_reservation, reservation_target	trainer.use_reservation, trainer.reservation_target	Set use_reservation explicitly: the recipe default is True.
DPO beta, ref_cache_concurrency, ref_cache_batch_size	Same names
ORPO orpo.lambda	orpo_lambda	ORPO runs on Dedicated Training only.

Managed jobs saved a resumable checkpoint every 20 steps, and LoRA SFT and DPO jobs also saved one every 30 minutes. To keep that, set dcp_save_interval=20, and checkpoint_interval_minutes=30 for LoRA SFT and DPO.

Which managed job settings have no equivalent?

• A token-based batch_size from older jobs. Recipes batch by rows; use batch_size_samples → batch_size.
• lora_target_modules, lora_dropout and tune_vision_encoder.
• Datasets in Amazon S3 or Azure Blob Storage. Copy the data to a local file or to Cloud Storage first.
• Region, accelerator and training shape. You don’t set these; the recipe picks a validated training shape.