0.2.0 Interim dev release - very beta

TorchDevice 0.2.0 - Neural Network Operations and Device Handling Overhaul

Major Changes

• Complete Neural Network Operations Refactoring

  • Centralized all neural network operations in dedicated device/nn.py module
  • Added comprehensive type safety and device compatibility checks
  • Implemented proper tensor dtype handling across operations
  • Added support for embedding, linear, and layer normalization operations
  • Fixed critical issues with embedding operations on MPS devices

• Enhanced Device Management

  • Improved device redirection logic for better compatibility
  • Added robust type conversion handling for tensor operations
  • Fixed device-specific normalization issues
  • Enhanced memory management for tensor operations

• Modular Architecture

  • Reorganized codebase into logical modules for better maintainability
  • Separated device-specific operations into dedicated modules
  • Implemented helper utilities for common tensor operations
  • Improved code reusability and reduced duplication

• Testing Infrastructure

  • Added comprehensive tests for neural network operations
  • Enhanced test coverage for device handling
  • Improved test reliability and reproducibility
  • Added transformer model integration tests

Breaking Changes

• Neural network operations now enforce stricter type checking
• Device handling may require explicit dtype specifications in some cases
• Embedding operations now handle normalization differently

Known Issues

• Some CUDA-specific operations may not have full MPS equivalents
• Performance implications when falling back to CPU for unsupported operations

Next Steps

• Implementation of attention mechanisms
• Support for more neural network operations
• Enhanced error handling and diagnostics
• Performance optimizations for device-specific operations

Testing Notes

We need testers to validate the following scenarios:

1. Transformer model inference on MPS devices
2. Large-scale embedding operations
3. Mixed-precision training workflows
4. Multi-device tensor operations

Please report any issues or unexpected behavior through the issue tracker.