SmolFactory / src /trackio.py
Tonic's picture
adds monkey patch for trackio monitoring in torch and readme creator improvements
39db0ca verified
raw
history blame
5.74 kB
"""
Trackio Module Interface for TRL Library
Provides the interface expected by TRL library while integrating with our custom monitoring system
"""
import os
import logging
from typing import Dict, Any, Optional
from datetime import datetime
# Import our custom monitoring
from monitoring import SmolLM3Monitor
logger = logging.getLogger(__name__)
# Global monitor instance
_monitor = None
def init(
project_name: str,
experiment_name: Optional[str] = None,
**kwargs
) -> str:
"""
Initialize trackio experiment (TRL interface)
Args:
project_name: Name of the project
experiment_name: Name of the experiment (optional)
**kwargs: Additional configuration parameters
Returns:
Experiment ID
"""
global _monitor
try:
# Extract configuration from kwargs
trackio_url = kwargs.get('trackio_url') or os.environ.get('TRACKIO_URL')
trackio_token = kwargs.get('trackio_token') or os.environ.get('TRACKIO_TOKEN')
hf_token = kwargs.get('hf_token') or os.environ.get('HF_TOKEN')
dataset_repo = kwargs.get('dataset_repo') or os.environ.get('TRACKIO_DATASET_REPO', 'tonic/trackio-experiments')
# Use experiment_name if provided, otherwise use project_name
exp_name = experiment_name or project_name
# Create monitor instance
_monitor = SmolLM3Monitor(
experiment_name=exp_name,
trackio_url=trackio_url,
trackio_token=trackio_token,
enable_tracking=True,
log_artifacts=True,
log_metrics=True,
log_config=True,
hf_token=hf_token,
dataset_repo=dataset_repo
)
# Generate experiment ID
experiment_id = f"trl_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
_monitor.experiment_id = experiment_id
logger.info(f"Trackio initialized for experiment: {exp_name}")
logger.info(f"Experiment ID: {experiment_id}")
return experiment_id
except Exception as e:
logger.error(f"Failed to initialize trackio: {e}")
# Return a fallback experiment ID
return f"trl_fallback_{datetime.now().strftime('%Y%m%d_%H%M%S')}"
def log(
metrics: Dict[str, Any],
step: Optional[int] = None,
**kwargs
):
"""
Log metrics to trackio (TRL interface)
Args:
metrics: Dictionary of metrics to log
step: Current training step
**kwargs: Additional parameters
"""
global _monitor
try:
if _monitor is None:
logger.warning("Trackio not initialized, skipping log")
return
# Log metrics using our custom monitor
_monitor.log_metrics(metrics, step)
# Also log system metrics if available
_monitor.log_system_metrics(step)
except Exception as e:
logger.error(f"Failed to log metrics: {e}")
def finish():
"""
Finish trackio experiment (TRL interface)
"""
global _monitor
try:
if _monitor is None:
logger.warning("Trackio not initialized, skipping finish")
return
# Close the monitoring session
_monitor.close()
logger.info("Trackio experiment finished")
except Exception as e:
logger.error(f"Failed to finish trackio experiment: {e}")
def log_config(config: Dict[str, Any]):
"""
Log configuration to trackio (TRL interface)
Args:
config: Configuration dictionary to log
"""
global _monitor
try:
if _monitor is None:
logger.warning("Trackio not initialized, skipping config log")
return
# Log configuration using our custom monitor
_monitor.log_configuration(config)
except Exception as e:
logger.error(f"Failed to log config: {e}")
def log_checkpoint(checkpoint_path: str, step: Optional[int] = None):
"""
Log checkpoint to trackio (TRL interface)
Args:
checkpoint_path: Path to the checkpoint file
step: Current training step
"""
global _monitor
try:
if _monitor is None:
logger.warning("Trackio not initialized, skipping checkpoint log")
return
# Log checkpoint using our custom monitor
_monitor.log_model_checkpoint(checkpoint_path, step)
except Exception as e:
logger.error(f"Failed to log checkpoint: {e}")
def log_evaluation_results(results: Dict[str, Any], step: Optional[int] = None):
"""
Log evaluation results to trackio (TRL interface)
Args:
results: Evaluation results dictionary
step: Current training step
"""
global _monitor
try:
if _monitor is None:
logger.warning("Trackio not initialized, skipping evaluation log")
return
# Log evaluation results using our custom monitor
_monitor.log_evaluation_results(results, step)
except Exception as e:
logger.error(f"Failed to log evaluation results: {e}")
# Additional utility functions for TRL compatibility
def get_experiment_url() -> Optional[str]:
"""Get the URL to view the experiment"""
global _monitor
if _monitor is not None:
return _monitor.get_experiment_url()
return None
def is_available() -> bool:
"""Check if trackio is available and initialized"""
return _monitor is not None and _monitor.enable_tracking
def get_monitor():
"""Get the current monitor instance (for advanced usage)"""
return _monitor