Locallab

Fixed
- Fixed critical server startup error related to missing 'lifespan' attribute in ServerWithCallback class
- Fixed KeyError in 'locallab info' command by properly handling RAM information
- Significantly improved CLI startup speed through lazy loading of imports
- Enhanced error handling in system information display
- Fixed environment variable conflicts between CLI configuration and OS environment variables
- Improved configuration system to properly handle both CLI and environment variable settings
- Optimized server startup process for faster response time

Changed
- Reduced unnecessary operations during CLI startup for better performance
- Improved memory usage reporting with proper unit conversion (GB instead of MB)
- Enhanced ServerWithCallback class with proper lifespan initialization
- Updated configuration system to use a unified approach for all settings

Added
- Enhanced CLI with interactive configuration wizard
- Added persistent configuration storage
- Implemented environment detection for smart defaults
- Added command groups: start, config, info
- Added support for configuring optimizations through CLI
- Improved Google Colab integration with context-aware prompts
- Added system information command

Fixed
- Improved streaming generation quality to match non-streaming responses
- Added proper stopping conditions for streaming to prevent endless generation
- Implemented repetition detection to stop low-quality streaming responses
- Reduced token chunk size for better quality control in streaming mode
- Ensured consistent generation parameters between streaming and non-streaming modes

Added
- Added memory monitoring to prevent CUDA out of memory errors
- Implemented adaptive token generation for streaming responses
- Added CUDA memory configuration with expandable segments

Fixed
- Fixed torch.compile() errors by adding proper error handling and fallback to eager mode
- Fixed early stopping warning by correctly setting num_beams parameter
- Improved streaming generation with smaller token chunks for more responsive output
- Added memory-aware generation that adapts to available GPU resources
- Implemented error recovery for out-of-memory situations during generation

Fixed
- Fixed issue with banners (running banner, system instructions, model configuration, API documentation) repeating in the console at regular intervals
- Added flag to ensure startup information is only displayed once during server initialization
- Improved server callback handling to prevent duplicate banner displays

Fixed
- Fixed Env Configuration by removing the duplicated Env Configuration.