SlimeRunner

module: agentcore_rl_toolkit.backends.slime.runner

SlimeRunner — one Python entry point for slime-backed training.

Users instantiate SlimeRunner with a handful of per-experiment fields and call .train(); the runner reproduces what train.sh does today (stop stale processes, start a Ray head, source the slime model script, submit the slime training job) via subprocess.

train.sh stays in the repo as the low-level escape hatch; this class is the primary entry point.

class SlimeRunner

One Python entry point for slime-backed training.

Constructor

SlimeRunner(
    # --- Required: per-experiment ---
    exp_id: str,
    agent_runtime_arn: str,
    s3_bucket: str,
    model_dir: str,
    data_path: str,
    model_type: str,

    # --- Optional: cluster ---
    num_gpus: int = 8,
    tp_size: int = 2,
    rollout_gpus_per_engine: int = 2,
    slime_dir: str = '/root/slime',
    megatron_dir: str = '/root/Megatron-LM',

    # --- Optional: ACR / toolkit (forwarded to slime via custom-config yaml) ---
    model_id: str = 'default',
    acr_timeout: int = 900,
    acr_tps_limit: int = 25,
    max_concurrent: int = 100,
    gateway_port: int = 9090,
    reward_postprocessing: str = 'grpo',

    # --- Optional: training hyperparameters ---
    rollout_batch_size: int = 32,
    n_samples_per_prompt: int = 8,
    rollout_max_response_len: int = 1024,
    rollout_temperature: float = 1.0,
    lr: float = 1e-06,
    eps_clip: float = 0.2,
    eps_clip_high: float = 0.28,
    weight_decay: float = 0.1,
    adam_beta2: float = 0.98,
    sglang_mem_fraction_static: float = 0.7,
    max_tokens_per_gpu: int = 9216,

    # --- Wandb (opt-in; no defaults injected if unset) ---
    wandb_project: str | None = None,
    wandb_group: str | None = None,

    # --- Escape hatch ---
    extra_flags: list[str] = list(),
)

Methods

from_yaml(path: str | os.PathLike) -> 'SlimeRunner'

Load kwargs from a YAML file (convenience for config-file workflows).

train(num_rollout: int = 1) -> None

Run the training job. Blocks until the slime job exits.

Mirrors train.sh step-by-step: stop stale sglang/ray, start a Ray head, source the slime model script, submit the slime training job via ray job submit. Streams stdout/stderr to the parent process.

Attributes

• acr_timeout (int)

• acr_tps_limit (int)

• adam_beta2 (float)

• agent_runtime_arn (str)

• data_path (str)

• eps_clip (float)

• eps_clip_high (float)

• exp_id (str)

• extra_flags (list[str])

• gateway_port (int)

• lr (float)

• max_concurrent (int)

• max_tokens_per_gpu (int)

• megatron_dir (str)

• model_dir (str)

• model_id (str)

• model_type (str)

• n_samples_per_prompt (int)

• num_gpus (int)

• reward_postprocessing (str)

• rollout_batch_size (int)

• rollout_gpus_per_engine (int)

• rollout_max_response_len (int)

• rollout_temperature (float)

• s3_bucket (str)

• sglang_mem_fraction_static (float)

• slime_dir (str)

• tp_size (int)

• wandb_group (str | None)

• wandb_project (str | None)

• weight_decay (float)