Labsco
huggingface logo

toml-config

✓ Official

by huggingface · part of huggingface/prime-rl

How to write and use TOML configs in prime-rl. Use when creating config files, running commands with configs, or overriding config values via CLI.

🔥🔥✓ VerifiedFreeNeeds API keys
#devops#git#maps
🔒 Repo-maintenance skill. It exists to help maintain huggingface/prime-rl itself — it's only useful if you contribute code to that project.

This is the playbook your agent receives when the skill activates — you don't need to read it to use the skill, but it's here to audit before installing.

by huggingface

How to write and use TOML configs in prime-rl. Use when creating config files, running commands with configs, or overriding config values via CLI. npx skills add https://github.com/huggingface/prime-rl --skill toml-config Download ZIPGitHub

TOML Config

All prime-rl commands use pydantic_config (tyro-backed) with TOML configs and CLI overrides.

TOML structure

Top-level fields must come before any [section] header — this is a TOML rule.

# Top-level fields first
gpu_memory_utilization = 0.5
seed = 42

# Then sections
[model]
name = "Qwen/Qwen3-0.6B"
max_model_len = 4096

[server]
port = 8000

Putting a top-level field after a section header nests it inside that section, which causes validation errors.

Setting None

Use the string "None" in TOML to set a field to None:

max_model_len = "None"

SLURM mode

Both rl and sft commands support SLURM execution via an optional [slurm] section. When present, the run is submitted as a SLURM job instead of running locally.

SLURM configs are composed with the base config via CLI:

uv run rl @ examples/reverse_text/rl.toml @ examples/reverse_text/slurm_rl.toml

RL SLURM

output_dir = "/shared/experiments/my-run"

[deployment]
type = "multi_node"
num_train_nodes = 2
num_infer_nodes = 1
gpus_per_node = 8
# nodes_per_fsdp_group = 1

[slurm]
job_name = "my-rl-job"
# dry_run = true # generate script without submitting
# template_path = "path/to/custom.sh.j2"
# project_dir = "/path/to/project"

When [slurm] is set for RL:

  • output_dir must be explicitly set (the default outputs is rejected)

  • Teacher inference is not supported in multi-node deployment

SFT SLURM

output_dir = "/shared/experiments/my-sft-run"

[deployment]
type = "multi_node"
num_nodes = 2
gpus_per_node = 8
# nodes_per_fsdp_group = 1

[slurm]
job_name = "my-sft-job"
# dry_run = true
# template_path = "path/to/custom.sh.j2"
# project_dir = "/path/to/project"

SFT deployment follows the same pattern as RL:

  • [deployment] configures node/GPU allocation (single_node default or multi_node)

  • [slurm] configures SLURM submission (job name, partition, template)

  • output_dir must be explicitly set when using SLURM

  • Multi-node deployment requires [slurm] to be set

Available commands

All accept @ config.toml and CLI overrides:

Command Config class Description uv run rl full RL pipeline Orchestrator + inference + trainer (local or SLURM) uv run inference InferenceConfig vLLM inference server uv run trainer trainer config RL trainer uv run orchestrator orchestrator config Rollout orchestrator uv run env-server env server config Environment server uv run sft SFT config Supervised fine-tuning (local or SLURM)

Key files

  • src/prime_rl/utils/config.py — BaseConfig, cli, get_all_fields

  • src/prime_rl/entrypoints/rl.py — unified RL entrypoint (local + SLURM)

  • src/prime_rl/configs/rl.py — RLConfig, SlurmConfig, DeploymentConfig

  • src/prime_rl/entrypoints/sft.py — unified SFT entrypoint (local + SLURM)

  • src/prime_rl/configs/sft.py — SFTConfig

  • configs/ — all config files, organized by task