Logging API

Overview

Slime provides comprehensive logging support through wandb, TensorBoard, and structured console logging.

Logger Configuration

configure_logger()

Configure the Python logging system with slime formatting.

from slime.utils.logging_utils import configure_logger

configure_logger(prefix="")

prefix

str

default:"''"

Optional prefix for log messages

Log Format:

[2024-03-15 14:23:45] train.py:10 - Training started

Behavior:

Sets logging level to INFO
Configures timestamp and file location formatting
Only configures once (subsequent calls are no-ops)
Uses force=True to override existing configuration

Source: slime/utils/logging_utils.py:12

Tracking Initialization

init_tracking()

Initialize experiment tracking (wandb/TensorBoard).

from slime.utils.logging_utils import init_tracking

init_tracking(args, primary=True)

args

Namespace

required

Arguments containing tracking configuration:

use_wandb: Enable wandb
use_tensorboard: Enable TensorBoard
wandb_project, wandb_team, wandb_group, etc.

primary

bool

default:"True"

Whether this is the primary tracking process:

True: Main training process (initializes wandb run)
False: Secondary process (joins existing run)

**kwargs

dict

Additional kwargs passed to wandb.init()

Source: slime/utils/logging_utils.py:27

Logging Metrics

log()

Log metrics to configured tracking backends.

from slime.utils.logging_utils import log

log(
    args,
    metrics={
        "train/loss": 0.234,
        "train/grad_norm": 1.23,
        "rollout_step": 42
    },
    step_key="rollout_step"
)

args

Namespace

required

Arguments with use_wandb and use_tensorboard flags

metrics

dict

required

Dictionary of metric names to values

step_key

str

required

Key in metrics dict to use as step counter (e.g., “rollout_step”, “train_step”)

Behavior:

Wandb: Calls wandb.log(metrics) (includes step automatically)
TensorBoard: Extracts step from metrics, logs remaining metrics to TensorBoard

Source: slime/utils/logging_utils.py:35

Wandb Configuration

Primary Arguments

--use-wandb

bool

default:"False"

Enable Weights & Biases logging

--wandb-project

str

W&B project name

--wandb-team

str

W&B team/entity name

--wandb-group

str

W&B run group for organizing related runs

--wandb-mode

str

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

--wandb-dir

str

Directory to store wandb logs (default: ./wandb)

--wandb-key

str

W&B API key (alternative to WANDB_API_KEY env var)

--wandb-host

str

W&B host URL (for self-hosted instances)

--wandb-run-id

str

W&B run ID (for resuming runs)

--wandb-random-suffix

bool

default:"True"

Add random 6-character suffix to run name

Source: slime/utils/arguments.py:1044-1113

Advanced Wandb Arguments

--wandb-always-use-train-step

bool

default:"False"

Always use train step as x-axis (instead of rollout step for most metrics)

--log-multi-turn

bool

default:"False"

Log statistics for multi-turn rollout

--log-passrate

bool

default:"False"

Log pass@n statistics for responses

--log-reward-category

str

Log reward category statistics from specified metadata key

--log-correct-samples

bool

default:"False"

Log correct samples to wandb

Source: slime/utils/arguments.py:1076-1112

TensorBoard Configuration

--use-tensorboard

bool

default:"False"

Enable TensorBoard logging

--tb-project-name

str

TensorBoard log directory (default: $TENSORBOARD_DIR env var)

--tb-experiment-name

str

TensorBoard experiment name (subdirectory)

Source: slime/utils/arguments.py:1115-1127

Logged Metrics

Training Metrics

Logged during training steps:

{
    "train/loss": 0.234,
    "train/pg_loss": 0.123,
    "train/value_loss": 0.089,
    "train/entropy_loss": 0.022,
    "train/grad_norm": 1.23,
    "train/lr": 1e-6,
    "train/kl": 0.045,
    "train/clipfrac": 0.15,
    "train_step": 100,
    "rollout_step": 42
}

Rollout Metrics

Logged after rollout generation:

{
    "rollout/avg_reward": 0.67,
    "rollout/max_reward": 1.0,
    "rollout/min_reward": 0.0,
    "rollout/avg_response_length": 234.5,
    "rollout/truncation_rate": 0.05,
    "rollout/generation_time": 12.3,
    "rollout/samples_per_second": 45.6,
    "rollout_step": 42
}

Evaluation Metrics

Logged during evaluation:

{
    "eval/gsm8k_accuracy": 0.87,
    "eval/gsm8k_avg_reward": 0.87,
    "eval/math_accuracy": 0.52,
    "eval/math_avg_reward": 0.52,
    "rollout_step": 42
}

Performance Metrics

System performance metrics:

{
    "perf/train_time": 8.5,
    "perf/rollout_time": 12.3,
    "perf/total_time": 20.8,
    "perf/gpu_memory_allocated_gb": 45.2,
    "perf/gpu_memory_reserved_gb": 48.0,
    "rollout_step": 42
}

Custom Logging Functions

Custom Rollout Logging

Provide custom logging for rollout data:

# my_module.py
def log_rollout_data(rollout_id, args, samples, rollout_extra_metrics, rollout_time):
    """Custom rollout logging function.
    
    Args:
        rollout_id: Current rollout step
        args: Training arguments
        samples: List of sample groups
        rollout_extra_metrics: Extra metrics from rollout function
        rollout_time: Time taken for rollout
    
    Returns:
        bool: Whether to skip default logging (True = skip, False = run default)
    """
    import wandb
    
    # Custom metrics
    avg_reward = sum(s[0].reward for s in samples) / len(samples)
    wandb.log({"custom/avg_group_reward": avg_reward})
    
    # Return False to also run default logging
    return False

Configuration:

python train.py \
  --custom-rollout-log-function-path my_module:log_rollout_data

Source: slime/utils/arguments.py:397-404

Custom Eval Logging

Provide custom logging for evaluation:

# my_module.py
def log_eval_rollout_data(rollout_id, args, data, extra_metrics):
    """Custom eval logging function.
    
    Args:
        rollout_id: Current rollout step
        args: Training arguments
        data: Dict mapping dataset names to eval results
        extra_metrics: Extra metrics from eval function
    
    Returns:
        bool: Whether to skip default logging
    """
    import wandb
    
    # Log per-dataset custom metrics
    for dataset_name, results in data.items():
        samples = results["samples"]
        # Custom analysis
        wandb.log({f"custom/{dataset_name}_metric": compute_metric(samples)})
    
    return False

Configuration:

python train.py \
  --custom-eval-rollout-log-function-path my_module:log_eval_rollout_data

Source: slime/utils/arguments.py:406-414

Example Usage

Basic Logging Setup

from slime.utils.logging_utils import configure_logger, init_tracking, log
from slime.utils.arguments import parse_args

# Parse arguments
args = parse_args()

# Configure logging
configure_logger()

# Initialize tracking
init_tracking(args, primary=True)

# Training loop
for rollout_id in range(args.num_rollout):
    # ... training code ...
    
    # Log metrics
    log(args, {
        "train/loss": loss,
        "train/grad_norm": grad_norm,
        "rollout_step": rollout_id
    }, step_key="rollout_step")

Wandb with Grouping

python train.py \
  --use-wandb \
  --wandb-project slime-experiments \
  --wandb-team my-team \
  --wandb-group qwen-32b-experiment \
  --wandb-mode online

Runs will be grouped under “qwen-32b-experiment” in the W&B UI.

Offline Logging

# Log offline (sync later)
python train.py \
  --use-wandb \
  --wandb-mode offline \
  --wandb-dir ./wandb_logs

# Later, sync to cloud
wandb sync ./wandb_logs

Combined Wandb + TensorBoard

python train.py \
  --use-wandb \
  --wandb-project my-project \
  --use-tensorboard \
  --tb-project-name ./tb_logs \
  --tb-experiment-name run-001

Implementation Details

Wandb Utils

Wandb initialization is handled by wandb_utils.py:

# Primary process (main training)
def init_wandb_primary(args, **kwargs):
    wandb.init(
        project=args.wandb_project,
        entity=args.wandb_team,
        group=args.wandb_group,
        config=vars(args),
        **kwargs
    )

# Secondary process (distributed workers)
def init_wandb_secondary(args, **kwargs):
    # Join existing run without creating new
    ...

Source: slime/utils/logging_utils.py:5

TensorBoard Adapter

TensorBoard logging is handled by _TensorboardAdapter:

class _TensorboardAdapter:
    def __init__(self, args):
        self.writer = SummaryWriter(log_dir=...)
    
    def log(self, data, step):
        for key, value in data.items():
            self.writer.add_scalar(key, value, step)

Source: slime/utils/logging_utils.py:6

Best Practices

Always configure logger first:

configure_logger()  # Before any logging calls

Use consistent step keys:
- Use "rollout_step" for rollout/eval metrics
- Use "train_step" for training metrics
Group related runs:
```
--wandb-group hyperparameter-sweep-001
```

Use offline mode for unstable networks:

--wandb-mode offline --wandb-dir ./persistent_storage/wandb

Log hierarchical metrics:

{
    "train/loss": ...,
    "train/pg_loss": ...,
    "eval/gsm8k_accuracy": ...,
    "rollout/avg_reward": ...
}

See Also:

Arguments API - Logging configuration
Data Structures API - Metrics from samples
Training API - Training metrics

Core Modules

Utilities

Overview

Logger Configuration

configure_logger()

Tracking Initialization

init_tracking()

Logging Metrics

log()

Wandb Configuration

Primary Arguments

Advanced Wandb Arguments

TensorBoard Configuration

Logged Metrics

Training Metrics

Rollout Metrics

Evaluation Metrics

Performance Metrics

Custom Logging Functions

Custom Rollout Logging

Custom Eval Logging

Example Usage

Basic Logging Setup

Wandb with Grouping

Offline Logging

Combined Wandb + TensorBoard

Implementation Details

Wandb Utils

TensorBoard Adapter

Best Practices

Build docs developers (and LLMs) love

Core Modules

Utilities

​Overview

​Logger Configuration

​configure_logger()

​Tracking Initialization

​init_tracking()

​Logging Metrics

​log()

​Wandb Configuration

​Primary Arguments

​Advanced Wandb Arguments

​TensorBoard Configuration

​Logged Metrics

​Training Metrics

​Rollout Metrics

​Evaluation Metrics

​Performance Metrics

​Custom Logging Functions

​Custom Rollout Logging

​Custom Eval Logging

​Example Usage

​Basic Logging Setup

​Wandb with Grouping

​Offline Logging

​Combined Wandb + TensorBoard

​Implementation Details

​Wandb Utils

​TensorBoard Adapter

​Best Practices

Build docs developers (and LLMs) love

Overview

Logger Configuration

configure_logger()

Tracking Initialization

init_tracking()

Logging Metrics

log()

Wandb Configuration

Primary Arguments

Advanced Wandb Arguments

TensorBoard Configuration

Logged Metrics

Training Metrics

Rollout Metrics

Evaluation Metrics

Performance Metrics

Custom Logging Functions

Custom Rollout Logging

Custom Eval Logging

Example Usage

Basic Logging Setup

Wandb with Grouping

Offline Logging

Combined Wandb + TensorBoard

Implementation Details

Wandb Utils

TensorBoard Adapter

Best Practices