Skip to content

PROJECT_STATUS.md

Latest commit

 

History

History
96 lines (74 loc) · 3.25 KB

PROJECT_STATUS.md

File metadata and controls

96 lines (74 loc) · 3.25 KB

CortexGPT Project Status

✅ Completed Tasks

1. Fixed OOM Issues for RTX 3090

  • Created train_cortexgpt_consumer_gpu.py with GPU-specific profiles
  • Implemented gradient accumulation support in unified trainer
  • Reduced default configurations for consumer GPUs

2. Neuroscience Features Support

  • Created train_neuroscience_3090.py for selective feature enabling
  • Added memory-optimized configurations for Phase 2 features
  • Updated READMEs with detailed neuroscience training instructions

3. Documentation Updates

  • Updated README.md and README_KR.md with:
    • Consumer GPU configurations
    • Neuroscience feature usage
    • Memory optimization techniques
    • Troubleshooting guides

4. Code Cleanup

  • Removed old phase-specific implementations (phase2, phase3, stable)
  • Moved documentation to docs/ directory
  • Cleaned up test and temporary files
  • Removed wandb logs

📁 Current Project Structure

Core Files (Unified Implementation)

  • cortexgpt/models/cortex_gpt.py - Main model interface
  • cortexgpt/models/cortex_gpt_unified.py - Unified implementation with all phases
  • cortexgpt/training/train_cortex_gpt.py - Unified trainer with gradient accumulation

Training Scripts

  • scripts/train_cortexgpt.py - Main training script
  • scripts/train_cortexgpt_consumer_gpu.py - Consumer GPU optimized training
  • scripts/train_neuroscience_3090.py - Neuroscience features for RTX 3090
  • scripts/quick_start_unified.py - Quick start guide with auto-detection

Documentation

  • README.md - Main documentation (updated)
  • README_KR.md - Korean documentation (updated)
  • PHASE1_SUMMARY.md - Phase 1 development history
  • FIXES_SUMMARY.md - Recent fixes documentation
  • docs/ - Additional technical documentation

🚀 Ready-to-Use Commands

For RTX 3090 Users

Basic Training (Minimal Features)

uv run scripts/train_cortexgpt_consumer_gpu.py --auto-detect --epochs 10

Neuroscience Features

# Both homeostasis and sleep-wake (default)
uv run scripts/train_neuroscience_3090.py --epochs 20

# Only homeostasis (lower memory)
uv run scripts/train_neuroscience_3090.py --homeostasis-only --epochs 20

Quick Start

uv run scripts/quick_start_unified.py
# Choose option 2 for consumer GPU optimized

Memory Usage Guidelines

Configuration Memory Usage Features
Minimal 8-10GB Base model only
Phase 1 10-12GB + Stability features
+ Homeostasis 12-15GB + Homeostatic plasticity
+ Sleep-Wake 15-18GB + Sleep-wake cycles
Full (Default) >20GB All features enabled

🔧 Key Improvements

  1. Gradient Accumulation: Enables larger effective batch sizes on limited memory
  2. Selective Feature Enabling: Turn on only needed features to save memory
  3. Auto GPU Detection: Automatically configures for your GPU
  4. Memory Monitoring: Built-in warnings and recommendations

📝 Notes

  • Default configuration requires >20GB memory (not suitable for consumer GPUs)
  • Use provided scripts for consumer GPU training
  • Monitor GPU memory with watch -n 1 nvidia-smi
  • Start with minimal features and gradually enable more

The project is now fully functional on consumer GPUs with intelligent memory management!