Arguments API

Overview

Slime uses a comprehensive argument system built on top of Megatron-LM arguments with extensive RL-specific additions.

Parsing Arguments

parse_args()

Main entry point for parsing all arguments.

from slime.utils.arguments import parse_args

args = parse_args()

Returns: Namespace object containing all parsed arguments Source: slime/utils/arguments.py:1, defined via get_slime_extra_args_provider()

Argument Categories

Cluster Configuration

Arguments for distributed training setup.

--actor-num-nodes

int

default:"1"

Number of nodes for actor training

--actor-num-gpus-per-node

int

default:"8"

GPUs per node for actor training

--critic-num-nodes

int

Number of nodes for critic training (if using critic)

--critic-num-gpus-per-node

int

GPUs per node for critic training

--rollout-num-gpus

int

Total GPUs for rollout inference (ignored if --colocate)

--rollout-num-gpus-per-engine

int

default:"1"

GPUs per inference engine (tensor parallel size)

--num-gpus-per-node

int

default:"8"

GPUs per node for resource allocation

--colocate

bool

default:"False"

Colocate inference engines and actor on same GPUs (enables offloading)

--offload

bool

default:"False"

Enable both --offload-train and --offload-rollout

--offload-train

bool

Offload training model to CPU during rollout

--offload-rollout

bool

Offload rollout model to CPU during training

Source: slime/utils/arguments.py:37-108

Training Configuration

--qkv-format

str

default:"thd"

QKV layout for Megatron: “thd” or “bshd”

--true-on-policy-mode

bool

default:"False"

Enable strict on-policy training

--train-env-vars

dict

default:"{}"

Extra environment variables for training (JSON string)

--train-memory-margin-bytes

int

default:"1073741824"

Memory margin for training allocation (default: 1GB)

--enable-weights-backuper

bool

default:"True"

Enable weights backuper (disable to save host memory)

--megatron-to-hf-mode

str

default:"raw"

Megatron to HuggingFace conversion: “raw” or “bridge”

--custom-model-provider-path

str

Path to custom model provider function

--recompute-loss-function

bool

default:"False"

Recompute loss function to save memory

--log-probs-chunk-size

int

default:"-1"

Chunk size for log prob computation

--only-train-params-name-list

list[str]

Regex patterns of parameters to train (freeze all others)

--freeze-params-name-list

list[str]

Regex patterns of parameters to freeze

Source: slime/utils/arguments.py:111-214

Rollout Configuration

--hf-checkpoint

str

HuggingFace checkpoint path for rollout model

--model-name

str

Model name for weight conversion (auto-detected if not set)

--rollout-function-path

str

default:"slime.rollout.sglang_rollout.generate_rollout"

Path to rollout generation function

--rollout-temperature

float

default:"1.0"

Sampling temperature

--rollout-top-p

float

default:"1.0"

Top-p (nucleus) sampling

--rollout-top-k

int

default:"-1"

Top-k sampling (-1 for disabled)

--rollout-max-context-len

int

Maximum context length (must not exceed model’s max position embeddings)

--rollout-max-prompt-len

int

Maximum prompt length (filters long prompts at init)

--rollout-max-response-len

int

Maximum response length (max_tokens in SGLang)

--rollout-skip-special-tokens

bool

default:"False"

Skip special tokens in decoded response

--rollout-stop

list[str]

Stop strings for generation

--rollout-stop-token-ids

list[int]

Stop token IDs for generation

--rollout-shuffle

bool

default:"False"

Shuffle prompts during rollout

--rollout-seed

int

default:"42"

Random seed for rollout

Source: slime/utils/arguments.py:218-339

Sampling Configuration

--over-sampling-batch-size

int

Sampling batch granularity (defaults to rollout_batch_size)

--dynamic-sampling-filter-path

str

Path to dynamic sampling filter function

--partial-rollout

bool

default:"False"

Enable partial rollout with sample recycling

--mask-offpolicy-in-partial-rollout

bool

default:"False"

Mask off-policy tokens in partial rollout

--custom-generate-function-path

str

Path to custom generate function

--custom-rollout-log-function-path

str

Path to custom rollout logging function

--buffer-filter-path

str

Path to buffer filter function

Source: slime/utils/arguments.py:341-426

Data Configuration

--num-rollout

int

Total number of rollout steps (calculated from num_epoch if not set)

--num-epoch

int

Number of training epochs

--rollout-global-dataset

bool

default:"True"

Use global dataset for rollout (disable for custom data management)

--data-source-path

str

Data source class path

--prompt-data

str

Path to prompt dataset (JSONL format)

--apply-chat-template

bool

default:"False"

Apply chat template to prompts

--apply-chat-template-kwargs

dict

default:"{}"

Kwargs for chat template (JSON string)

--input-key

str

default:"input"

JSON key for input/prompt

--label-key

str

JSON key for labels/ground truth

--multimodal-keys

dict

JSON mapping of media types to data keys, e.g. {"image": "image_file"}

--metadata-key

str

default:"metadata"

JSON key for metadata

--tool-key

str

default:"tools"

JSON key for tools in function calling

Source: slime/utils/arguments.py:500-573

Batch Size Configuration

--rollout-batch-size

int

required

Number of prompts per rollout step

--n-samples-per-prompt

int

default:"1"

Number of responses per prompt

--global-batch-size

int

Global batch size for training (across all data parallel ranks)

--num-steps-per-rollout

int

Number of training steps per rollout (alternative to setting global-batch-size)

--micro-batch-size

int

default:"1"

Micro batch size per GPU (ignored if using dynamic batch size)

--balance-data

bool

default:"False"

Balance tokens across DP ranks using Karmarkar-Karp algorithm

--use-dynamic-batch-size

bool

default:"False"

Dynamically adjust micro batch size based on sequence lengths

--max-tokens-per-gpu

int

Maximum tokens per GPU for dynamic batching

Source: slime/utils/arguments.py:585-654

Evaluation Configuration

--eval-function-path

str

Path to eval function (defaults to rollout_function_path)

--eval-interval

int

Evaluation interval in rollout steps (None to disable)

--eval-prompt-data

list[str]

Eval dataset paths: dataset_name /path/to/data.jsonl ...

--eval-config

str

Path to YAML/JSON eval config (overrides --eval-prompt-data)

--skip-eval-before-train

bool

default:"False"

Skip initial evaluation before training

--eval-input-key

str

Override input key for eval

--eval-label-key

str

Override label key for eval

--n-samples-per-eval-prompt

int

default:"1"

Responses per eval prompt

--eval-temperature

float

Override temperature for eval

--eval-max-response-len

int

Override max response length for eval

Source: slime/utils/arguments.py:657-714

Algorithm Configuration

--ref-load

str

Reference model checkpoint path

--load

str

Actor checkpoint to load

--save

str

Checkpoint save directory

--save-interval

int

Checkpoint save interval

--save-hf

str

Save HuggingFace format checkpoints (Megatron backend)

--lr

float

default:"1e-6"

Learning rate

--clip-grad

float

default:"1.0"

Gradient clipping value

--seed

int

default:"1234"

Random seed

Critic Configuration:

--num-critic-only-steps

int

default:"0"

Number of critic-only training steps before actor training

--critic-load

str

Critic checkpoint path

--critic-save

str

Critic save directory

--critic-lr

float

Critic learning rate (defaults to --lr)

PPO Configuration:

--eps-clip

float

default:"0.2"

PPO clip range

--eps-clip-high

float

PPO upper clip range

--eps-clip-c

float

Dual-clip PPO lower bound

--value-clip

float

default:"0.2"

Value function clip range

--kl-coef

float

default:"0.0"

KL penalty coefficient for reward shaping

--loss-type

str

default:"policy_loss"

Loss type: “policy_loss”, “sft_loss”, or “custom_loss”

--advantage-estimator

str

default:"grpo"

Advantage estimator: “grpo”, “gspo”, “reinforce_plus_plus”, “ppo”

--gamma

float

default:"1.0"

PPO GAE gamma (discount factor)

--lambd

float

default:"1.0"

PPO GAE lambda

--normalize-advantages

bool

default:"False"

Normalize advantages

Source: slime/utils/arguments.py:718-963

Router Configuration

--use-slime-router

bool

default:"False"

Use SlimeRouter instead of SGLang router

--slime-router-middleware-paths

list[str]

default:"[]"

Middleware function paths

--slime-router-timeout

float

HTTP request timeout in seconds

--slime-router-max-connections

int

Maximum concurrent connections

--slime-router-health-check-failure-threshold

int

default:"3"

Consecutive failures before marking worker dead

Source: slime/utils/arguments.py:1009-1041

Logging Configuration

--use-wandb

bool

default:"False"

Enable Weights & Biases logging

--wandb-mode

str

W&B mode: “online”, “offline”, or “disabled”

--wandb-project

str

W&B project name

--wandb-team

str

W&B team/entity name

--wandb-group

str

W&B run group

--use-tensorboard

bool

default:"False"

Enable TensorBoard logging

--tb-project-name

str

TensorBoard log directory

Source: slime/utils/arguments.py:1043-1127

Example Configuration

python train.py \
  --actor-num-nodes 2 \
  --actor-num-gpus-per-node 8 \
  --rollout-num-gpus 8 \
  --rollout-batch-size 256 \
  --n-samples-per-prompt 4 \
  --global-batch-size 1024 \
  --hf-checkpoint Qwen/Qwen2.5-32B-Instruct \
  --prompt-data data/prompts.jsonl \
  --lr 1e-6 \
  --num-epoch 3 \
  --save checkpoints/my_run \
  --use-wandb \
  --wandb-project slime-training

See Also:

Training API - Using parsed arguments
Rollout API - Rollout configuration
Logging API - Tracking configuration

Core Modules

Utilities

Arguments API

Overview

Parsing Arguments

parse_args()

Argument Categories

Cluster Configuration

Training Configuration

Rollout Configuration

Sampling Configuration

Data Configuration

Batch Size Configuration

Evaluation Configuration

Algorithm Configuration

Router Configuration

Logging Configuration

Example Configuration

Build docs developers (and LLMs) love

Core Modules

Utilities

​Overview

​Parsing Arguments

​parse_args()

​Argument Categories

​Cluster Configuration

​Training Configuration

​Rollout Configuration

​Sampling Configuration

​Data Configuration

​Batch Size Configuration

​Evaluation Configuration

​Algorithm Configuration

​Router Configuration

​Logging Configuration

​Example Configuration

Build docs developers (and LLMs) love

Overview

Parsing Arguments

parse_args()

Argument Categories

Cluster Configuration

Training Configuration

Rollout Configuration

Sampling Configuration

Data Configuration

Batch Size Configuration

Evaluation Configuration

Algorithm Configuration

Router Configuration

Logging Configuration

Example Configuration