Skip to main content

Documentation Index

Fetch the complete documentation index at: https://fireworks.ai/docs/llms.txt

Use this file to discover all available pages before exploring further.

​
What this is

RLOR trainer jobs and hotload-enabled deployments hold GPU resources. Always clean up after experiments — especially if jobs terminate unexpectedly.

​
Cleaning up RLOR trainer jobs

import os
from fireworks.training.sdk import TrainerJobManager, DeploymentManager

api_key = os.environ["FIREWORKS_API_KEY"]
base_url = os.environ.get("FIREWORKS_BASE_URL", "https://api.fireworks.ai")

rlor_mgr = TrainerJobManager(api_key=api_key, base_url=base_url)
deploy_mgr = DeploymentManager(api_key=api_key, base_url=base_url)

# Delete known trainer jobs from this run
for job_id in ["<policy-job-id>", "<reference-job-id>"]:
    rlor_mgr.delete(job_id=job_id)

​
Cleaning up deployments

deploy_mgr.delete(deployment_id="<deployment-id>")
If you want to keep the deployment resource but release GPUs (lighter alternative to delete): 
deploy_mgr.scale_to_zero(deployment_id="<deployment-id>")
This sets both minReplicaCount and maxReplicaCount to 0, releasing all accelerators while keeping the deployment available for future scale-up.

​
Automatic cleanup with ResourceCleanup

The cookbook provides ResourceCleanup, a context manager that automatically deletes registered trainers and deployments on scope exit — including on exceptions and Ctrl+C: 
from training.utils.infra import ResourceCleanup

with ResourceCleanup(rlor_mgr, deploy_mgr) as cleanup:
    # Create trainer first (trainer owns the hot-load bucket)
    endpoint = rlor_mgr.create_and_wait(config)
    cleanup.trainer(endpoint.job_id)

    # Create deployment linked to the trainer's bucket
    deploy_config.hot_load_trainer_job = endpoint.job_name
    deploy_mgr.create_or_get(deploy_config)
    cleanup.deployment("research-loop-serving")

    run_training_loop()
Resources are deleted in reverse creation order. Pre-existing resources that should survive are simply not registered. Deployments can be scaled to zero instead of deleted: 
cleanup.deployment("research-loop-serving", action="scale_to_zero")

​
Manual cleanup with try/finally

If you’re not using the cookbook, use try/finally directly: 
policy_job_id = "<policy-job-id>"
reference_job_id = "<reference-job-id>"
deployment_id = "research-loop-serving"

try:
    run_training_loop()
finally:
    rlor_mgr.delete(policy_job_id)
    rlor_mgr.delete(reference_job_id)
    deploy_mgr.delete(deployment_id)

​
Checking for leaked resources

Track the IDs you create (trainer job IDs + deployment ID) and clean those explicitly. For broad account-wide discovery, use the Fireworks console or the managed fw.*.list() APIs.

​
Operational guidance

  • Delete both policy and reference trainers when running GRPO (which uses 2 RLOR jobs).
  • Register cleanup on atexit in your training scripts for automatic cleanup on Ctrl+C or exceptions.
  • Don’t delete a trainer while a save_weights_for_sampler_ext operation is in progress — wait for it to complete first.