v0.11.0

PyFlowGraph v0.11.0

Release Date: September 1, 2025
Build: 0d60178

Major Changes

Single Process Execution Architecture

Replaced subprocess isolation with direct Python function calls for significant performance improvements.

Implementation:

• Added SingleProcessExecutor class for direct function execution
• Modified GraphExecutor to use single process instead of subprocess.run()
• Implemented persistent namespace for imports and variables
• Added direct object reference passing without JSON serialization

Performance Results:

• Execution speed: 96,061 operations/second (0.01ms per operation)
• Eliminated subprocess overhead (50-200ms → <1ms per node)
• Removed JSON serialization overhead completely
• Import persistence eliminates re-import costs

Native Object Passing System

Implemented zero-copy object passing for ML frameworks.

Supported Types:

• PyTorch tensors with GPU memory sharing
• NumPy arrays with direct references
• Pandas DataFrames without copying
• TensorFlow tensors and variables
• JAX arrays with hardware acceleration

Measured Performance Improvements:

• PyTorch operations: 4,000x-7,500x faster
• NumPy array processing: 95x-100x faster
• Pandas DataFrame operations: 20x-50x faster
• Large object transfers: eliminated (now uses references)

NVIDIA GPU Computer Vision Pipeline

Added complete computer vision workflow demonstrating GPU acceleration.

Components:

• Image preprocessing with CUDA tensors
• ResNet50 classification with ImageNet labels
• GPU device detection and reporting
• Automatic GPU memory management

Requirements:

• NVIDIA GPU with CUDA support
• CUDA-enabled PyTorch (2.0.0+cu118)
• CUDA Toolkit installation

Technical Improvements

Virtual Environment Management

• Fixed environment switching between different graphs
• Added proper package loading from graph-specific virtual environments
• Implemented dynamic sys.path injection for package availability
• Added Dependencies section support to FlowSpec format

GUI Fixes

• Fixed undo history dialog text visibility in dark theme
• Fixed Clear Results button functionality in computer vision pipeline
• Added proper device information display (cuda:0 vs cpu)
• Implemented absolute path resolution for relative image paths
• Added meaningful ImageNet class names instead of numeric IDs

Testing

Added 36 new tests covering:

• Native object passing validation
• ML framework integration testing
• Memory leak detection
• Performance benchmarking
• Edge cases (circular references, concurrent access)

Bug Fixes

Virtual Environment Issues

• Fixed environment switching when loading graphs with different requirements
• Resolved package loading issues with proper sys.path management
• Fixed PyTorch deprecation warnings (updated torchvision API usage)
• Corrected import path consistency for isinstance() compatibility

GUI State Management

• Fixed Clear Results button scope issues in GUI State Handler
• Resolved node reference access in GUI functions
• Fixed device information connection routing
• Improved connection source accuracy for device reporting

Breaking Changes

Architecture Change: PyFlowGraph now uses single-process execution instead of subprocess isolation.

Impact:

• Performance: Significant speed improvements for all workflows
• Security: Reduced process isolation (traded for performance)
• Compatibility: Existing graphs work unchanged
• Environment: Virtual environment isolation maintained

Installation

Requirements:

• Windows OS with Python 3.8+
• PySide6
• Optional: NVIDIA GPU + CUDA for computer vision features
• Optional: PyTorch, NumPy, Pandas for ML workflows

Installation:

1. Download and extract PyFlowGraph-Windows-v0.11.0.zip
2. Run PyFlowGraph.exe

Documentation Updates

• Updated Epic 3 specifications for single-process architecture
• Added performance benchmarks and metrics
• Enhanced FlowSpec format with Dependencies section
• Added LLM documentation for AI-assisted graph creation
• Updated memory management and GPU cleanup documentation