SlimeConfig

SlimeConfig

from modal_training_gym.frameworks.slime.config import SlimeConfig

slime GRPO training configuration.

Launcher Instructions

Field	Type	Default	Description
environment	dict	{'PYTHONPATH': '/root/Megatron-LM/', 'CUDA_DEVICE_MAX_CONNECTIONS': '1', 'NCCL_NVLS_ENABLE': '1'}	Injected into the Ray job runtime env.
async_mode	bool	False	When True, uses train_async.py. Default False.
slime_model_script	str	""	Shell script path relative to /root/slime. Default "".
model	`ModelConfiguration	None`	None

Composed Configs

Field	Type	Default	Description
dataset	`DatasetConfig	None`	None
wandb	`WandbConfig	None`	None
modal	`ModalConfig	None`	None
app_tags	dict	{}	Extra Modal app tags. Default {}.

Cluster and Parallelism

Field	Type	Default	Description
actor_num_nodes	`int	None`	None
actor_num_gpus_per_node	`int	None`	None
colocate	`bool	None`	None
rollout_num_gpus	`int	None`	None
rollout_num_gpus_per_engine	int	1	GPUs per SGLang rollout engine. Default 1.
tensor_model_parallel_size	`int	None`	None
use_critic	bool	False	Use a separate critic network. Default False.
critic_num_nodes	`int	None`	None
critic_num_gpus_per_node	`int	None`	None

RL Algorithm

Field	Type	Default	Description
advantage_estimator	str	"grpo"	Advantage estimation method. Default "grpo".
n_samples_per_prompt	int	2	Rollout samples per prompt. Default 2.
eps_clip	float	0.2	PPO clipping epsilon. Default 0.2.
eps_clip_high	float	0.28	Asymmetric high-side PPO clip. Default 0.28.
use_kl_loss	bool	True	Enable KL divergence loss. Default False.
kl_loss_type	str	"low_var_kl"	KL loss variant. Default "low_var_kl".
kl_loss_coef	float	0.0	KL loss coefficient. Default 0.0.
entropy_coef	float	0.0	Entropy bonus coefficient. Default 0.0.
ref_load	str	""	Reference model checkpoint path. Default "".

Rollout

Field	Type	Default	Description
num_rollout	int	1	Number of rollout episodes per step. Default 1.
rollout_batch_size	int	8	Batch size for rollout generation. Default 8.
rollout_max_response_len	int	8192	Maximum response length during rollout. Default 8192.
rollout_temperature	float	1.0	Sampling temperature for rollouts. Default 1.0.
sglang_mem_fraction_static	float	0.7	SGLang static memory fraction. Default 0.7.

Training

Field	Type	Default	Description
global_batch_size	int	16	Global batch size. Default 16.
lr	float	1e-06	Learning rate. Default 1e-6.
lr_decay_style	str	"constant"	LR decay schedule. Default "constant".
weight_decay	float	0.1	Weight decay. Default 0.0.
adam_beta1	float	0.9	Adam beta1. Default 0.9.
adam_beta2	float	0.98	Adam beta2. Default 0.95.
optimizer	str	"adam"	Optimizer name. Default "adam".
name	str	""

Memory and Precision

Field	Type	Default	Description
attention_dropout	float	0.0	Attention dropout. Default 0.0.
hidden_dropout	float	0.0	Hidden layer dropout. Default 0.0.
attention_softmax_in_fp32	bool	True	Compute softmax in FP32. Default True.
accumulate_allreduce_grads_in_fp32	bool	True	Accumulate allreduce grads in FP32. Default True.
recompute_granularity	str	"full"	Activation recomputation granularity. Default "".
recompute_method	str	"uniform"	Activation recomputation method. Default "".
recompute_num_layers	int	1	Number of layers to recompute. Default None.

Dynamic Batching

Field	Type	Default	Description
use_dynamic_batch_size	bool	True	Use dynamic batch sizing. Default False.
max_tokens_per_gpu	int	9216	Max tokens per GPU for dynamic batching. Default None.

Eval

Field	Type	Default	Description
eval_interval	int	20	Evaluation interval in training steps. Default 20.
n_samples_per_eval_prompt	int	4	Eval samples per prompt. Default 4.
eval_max_response_len	int	16384	Max response length for eval. Default 16384.
eval_top_p	float	1.0	Top-p sampling for eval. Default 1.0.
eval_config	`dict	None`	None

Checkpointing

Field	Type	Default	Description
save	str	"/checkpoints"	Enable checkpoint saving. Default True.
save_interval	int	1000	Checkpoint save interval. Default 1000.
megatron_to_hf_mode	str	"bridge"	Checkpoint conversion mode ("bridge" or "raw"). Default "bridge".
use_fault_tolerance	bool	True	Enable fault tolerance. Default False.

Custom Reward

Field	Type	Default	Description
custom_rm_path	str	""	Python import path for custom reward function. Default "".

SGLang

Field	Type	Default	Description
sglang_config	`dict	None`	None
custom_config_path	`dict	None`	None
apply_chat_template_kwargs	str	""	Extra kwargs for chat template. Default None.

Other Fields

Field	Type	Default	Description
image_run_commands	list[str]	[]
local_python_sources	list[str]	[]
gpu_type	`str	None`	None
sequence_parallel	`bool	None`	None
rm_type	str	"math"

Methods

build_app(self, *, name: str | None = None, modal: modal_training_gym.frameworks.slime.config.ModalConfig | None = None) -> 'App'

cli_args(self) -> list[str]

slime CLI arguments derived from this config.

prepare_data(self) -> None

Materialize the training data.

to_dataset_config(self) -> 'DatasetConfig'

Extract dataset-related slime flags back into a DatasetConfig.

to_model(self) -> 'ModelConfiguration'

Extract model-related slime flags back into a ModelConfiguration.

to_wandb_config(self) -> 'WandbConfig'

Extract wandb-related slime flags back into a WandbConfig.

total_nodes(self) -> int

Total Modal cluster nodes required by this config.

Related Tutorials

• Shared concepts: config containers, framework factories, volume layout, running the pipeline
• Qwen3-4B GRPO on GSM8K (colocated)
• Customizing your slime run — scaling nodes, parallelism, and throughput
• Qwen3-4B GRPO on haiku poems — structure score + LLM judge

Source: modal_training_gym/frameworks/slime/config.py