Add comprehensive architectural analysis documentation for AReaLite vs Core

[Copilot]

This PR addresses the request to analyze the main differences between AReaLite and Core from a software architecture perspective, using the sync examples as a starting point.

Overview

Added a complete documentation set that provides in-depth architectural analysis of AReaL's two parallel systems:

• AReaLite (/arealite/, /examples/arealite/) - Lightweight, AI researcher-friendly framework
• Core (/realhf/, /training/) - Full-scale distributed system for production training

Key Findings

The analysis reveals fundamental architectural differences:

Aspect	AReaLite	Core
Design Philosophy	AI-centric, algorithm-focused	System-centric, production-focused
Abstraction Layers	3 layers	6 layers
Code Complexity	~233 lines entry point	70 + 300 lines configuration
Learning Curve	1-2 days	1-2 weeks
Target Scale	1-64 GPUs	64-1000+ GPUs

Documentation Structure

📋 docs/architecture/README.md

Navigation index with quick access paths for different user types (AI researchers, industrial teams, system engineers).

📊 docs/architecture/final_summary_report.md

Executive summary with strategic recommendations and clear selection criteria based on team capabilities and project scale.

🔧 docs/architecture/arealite_vs_core_analysis.md

Detailed technical comparison covering:

• Entry point architecture differences
• System design patterns (Component-based vs Master-Worker)
• Synchronous/asynchronous training implementations
• Code complexity analysis
• Customization approaches

💻 docs/architecture/code_examples_comparison.md

Real code snippets showing practical differences:

• AReaLite's direct PyTorch-style APIs vs Core's declarative configuration
• Single-file customization vs multi-module registration systems
• SPMD programming model vs distributed worker architecture

📈 docs/architecture/architecture_diagrams.md

Visual representations using Mermaid diagrams to illustrate system architectures, abstraction layers, and training workflows.

Architecture Insights

AReaLite Example (simple and direct):

def main(args):
    rollout = RemoteSGLangEngine(config.rollout)
    actor = FSDPPPOActor(config.actor)
    
    for step in range(max_steps):
        batch = rollout.rollout_batch(data, workflow)
        stats = actor.ppo_update(batch)
        actor.upload_weights(weight_update_meta)

Core Example (declarative and complex):

class PPOMATHConfig:
    def make_dfg(self):
        dfg.add_mfc_dataflow(prompt_mfc, gen_mfc, {"prompt"})
        dfg.add_mfc_dataflow(gen_mfc, ppo_mfc, {"response"})
        return dfg

Selection Guidelines

The documentation provides clear decision criteria:

• Choose AReaLite for: Research, rapid prototyping, teams with PyTorch background, <64 GPU scale
• Choose Core for: Production training, >64 GPU scale, teams with dedicated system engineers

Impact

This analysis helps users make informed architectural decisions and understand the "90% functionality, 10% complexity" philosophy that drives AReaLite's design as the foundation for future AReaL API evolution.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[zhshgmail]

approved