toml-config

โดย 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

ดาวน์โหลด ZIP GitHub

TOML Config

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

Running with configs

# Load a config file with @ syntax
uv run inference @ configs/debug/infer.toml
uv run sft @ configs/debug/sft/train.toml
uv run rl @ configs/debug/rl/train.toml

# CLI overrides (take precedence over TOML)
uv run inference @ config.toml --model.name Qwen/Qwen3-0.6B --server.port 8001

# Boolean flags: no value needed
uv run inference --model.enforce-eager          # sets to true
uv run inference --no-model.enforce-eager       # sets to false

# CLI-only (no TOML file)
uv run inference --model.name Qwen/Qwen3-0.6B --model.max-model-len 2048

# Compose multiple config files (later files override earlier ones)
uv run rl @ examples/reverse_text/rl.toml @ examples/reverse_text/slurm_rl.toml

# Nested config files: load a config for a specific section
uv run rl --model @ model.toml --data @ data.toml

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)