Implementation:Danijar Dreamerv3 Parallel Logger

Knowledge Sources	DreamerV3
Domains	Reinforcement_Learning, Distributed_Systems, Monitoring
Last Updated	2026-02-15 09:00 GMT

Overview

Concrete tool for running the centralized logging process that aggregates metrics from all distributed DreamerV3 processes via RPC.

Description

The parallel_logger() function in embodied/run/parallel.py deserializes the logger factory, creates the elements.Logger, and starts a portal.Server with two RPC endpoints: add (for scalar metrics) and tran (for episode transitions). It maintains per-environment episode accumulators, computes episode statistics on completion, drops stale episodes after episode_timeout, and periodically flushes metrics to all configured backends.

Usage

Spawned as a separate process by combined(). All other processes connect as clients.

Code Reference

Source Location

• Repository: dreamerv3
• File: embodied/run/parallel.py
• Lines: L317-413

Signature

def parallel_logger(make_logger, args):
    """
    Run the centralized logging process.

    Args:
        make_logger: bytes or Callable -> Logger
        args: elements.Config with logger_addr, log_every, save_every, episode_timeout.
    """

Import

from embodied.run.parallel import parallel_logger

I/O Contract

Inputs

Name	Type	Required	Description
make_logger	bytes or Callable	Yes	Factory for Logger (cloudpickle-serialized)
args	elements.Config	Yes	logger_addr, log_every, save_every, episode_timeout

Outputs

Name	Type	Description
RPC Server	portal.Server	Running at args.logger_addr with add and tran endpoints
Metrics	Files	Written to JSONL, TensorBoard, WandB at log_every intervals
Checkpoints	Files	Logger step counter saved at save_every intervals

Usage Examples

# Spawned by combined() as a separate process
import portal
import cloudpickle

portal.Process(parallel_logger, cloudpickle.dumps(make_logger), args)

Related Pages

Implements Principle

• Principle:Danijar_Dreamerv3_Distributed_Logger_Aggregation