ep command-line interface can inspect evaluation runs locally, upload evaluators, and create reinforcement fine-tuning jobs on Fireworks.
Global Options
These options can be used with any command:Enable verbose logging (Aliases:
-v)--server
Fireworks API server hostname or URL (e.g., dev.api.fireworks.ai or https://dev.api.fireworks.ai)
Commands
ep logs
Serve logs with file watching and real-time updates
Port to bind to (default: 8000)
Enable debug mode
Disable Elasticsearch setup
Use env vars for Elasticsearch config (requires ELASTICSEARCH_URL, ELASTICSEARCH_API_KEY, ELASTICSEARCH_INDEX_NAME)
Force Fireworks tracing backend for logs UI (overrides env auto-detection)
Force Elasticsearch backend for logs UI (overrides env auto-detection)
ep upload
Scan for evaluation tests, select, and upload as Fireworks evaluators
Path to search for evaluation tests (default: current directory)
--entry
Entrypoint of evaluation test to upload (module:function or path::function). For multiple, separate by commas.
Non-interactive: upload all discovered evaluation tests (Aliases:
-y)--env-file
Path to .env file containing secrets to upload (default: .env in current directory)
Overwrite existing evaluator with the same ID
Default dataset to use with this evaluator (Aliases:
--default-dataset)Description for evaluator (Aliases:
--description)Display name for evaluator (defaults to ID) (Aliases:
--name, --display-name)Pytest-style entrypoint (e.g., test_file.py::test_func). Auto-detected if not provided. (Aliases:
--entry-point)Requirements for evaluator (auto-detected from requirements.txt if not provided) (Aliases:
--requirements)Evaluator ID to use (if multiple selections, a numeric suffix is appended) (Aliases:
--id)ep create rft
Create a Reinforcement Fine-tuning Job on Fireworks
Non-interactive mode (Aliases:
-y)Print planned SDK call without sending
Overwrite existing evaluator with the same ID
Skip local dataset/evaluator validation
Ignore Dockerfile even if present; run pytest on host during evaluator validation
Extra flags to pass to ‘docker build’ when validating evaluator (quoted string, e.g. “—no-cache —pull —progress=plain”)
Extra flags to pass to ‘docker run’ when validating evaluator (quoted string, e.g. “—env-file .env —memory=8g”)
--env-file
Path to .env file containing secrets to upload to Fireworks (default: .env in project root)
--source-job
The source reinforcement fine-tuning job to copy configuration from. If other flags are set, they will override the source job’s configuration.
If set, only errors will be printed.
The name of the dataset used for training.
The evaluator resource name to use for RLOR fine-tuning job.
ID of the reinforcement fine-tuning job, a random UUID will be generated if not specified. (Aliases:
--job-id)Data chunking for rollout, default size 200, enabled when dataset > 300. Valid range is 1-10,000.
Whether to auto-carve the dataset for eval.
The name of a separate dataset to use for evaluation.
Additional parameters for the inference request as a JSON string. For example:
”{“stop”: [“\n”]}”. (Aliases:
--extra-body)Maximum number of tokens to generate per response. (Aliases:
--max-output-tokens)Number of response candidates to generate per input. (Aliases:
--response-candidates-count)Sampling temperature, typically between 0 and 2. (Aliases:
--temperature)Top-k sampling parameter, limits the token selection to the top k tokens. (Aliases:
--top-k)Top-p sampling parameter, typically between 0 and 1. (Aliases:
--top-p)KL coefficient (beta) override for GRPO-like methods. If unset, the trainer
default is used. (Aliases:
--rl-kl-beta, --kl-beta)RL loss method for underlying trainers. One of {grpo,dapo}. (Aliases:
--rl-loss-method, --method)The MCP server resource name to use for the reinforcement fine-tuning job. (Optional)
The number of nodes to use for the fine-tuning job. If not specified, the default is 1. (Aliases:
--nodes)The name of the base model to be fine-tuned Only one of ‘base_model’ or
‘warm_start_from’ should be specified. (Aliases:
--base-model)The maximum packed number of tokens per batch for training in sequence packing. (Aliases:
--batch-size)The number of epochs to train for. (Aliases:
--epochs)The number of batches to accumulate gradients before updating the model parameters. The effective batch size will be batch-size multiplied by this value. (Aliases:
--gradient-accumulation-steps)The learning rate used for training. (Aliases:
--learning-rate)The number of learning rate warmup steps for the reinforcement fine-tuning job. (Aliases:
--learning-rate-warmup-steps)The rank of the LoRA layers. (Aliases:
--lora-rank)The maximum context length to use with the model. (Aliases:
--max-context-length)The model ID to be assigned to the resulting fine-tuned model.If not specified, the job ID will be used. (Aliases:
--output-model)The PEFT addon model in Fireworks format to be fine-tuned from Only one of
‘base_model’ or ‘warm_start_from’ should be specified. (Aliases:
--warm-start-from)The API key for the wandb service. (Aliases:
--wandb-api-key, --api-key)Whether to enable wandb logging. (Aliases:
--wandb, --enabled)The entity name for the wandb service. (Aliases:
--wandb-entity, --entity)The project name for the wandb service. (Aliases:
--wandb-project, --project)ep local-test
Select an evaluation test and run it locally. If a Dockerfile exists, build and run via Docker; otherwise run on host.
--entry
Entrypoint to run (path::function or path). If not provided, a selector will be shown (unless —yes).
Ignore Dockerfile even if present; run pytest on host
Non-interactive: if multiple tests exist and no —entry, fails with guidance (Aliases:
-y)Extra flags to pass to ‘docker build’ (quoted string, e.g. “—no-cache —pull —progress=plain”)
Extra flags to pass to ‘docker run’ (quoted string, e.g. “—env-file .env —memory=8g”)