TensorTune 1.2.0 - Fixed GPU layers override and improved progress bar handling

TensorTune 1.2.0 Changelog and Update Instructions

Changelog: Core Improvements and Bug Fixes

1. Enhanced Model Layer Detection and GPU Layer Allocation

   • Completely redesigned analyze_filename() function with more accurate model family recognition
   • Added special handling for MoE model series
   • Improved layer count detection for large models (20B+)
   • Added fallback mechanisms for models with missing layer information

2. Rebuilt GPU Layer Allocation Logic

   • Completely rewrote get_gpu_layers_for_level() to provide smoother transitions between CPU/GPU bias levels
   • Eliminated the "big jump" issue where models would suddenly jump from a few layers to 999
   • Created proper percentage-based scaling that respects the actual layer count of different model architectures
   • Enhanced support for MoE models with specialized allocation strategies

3. More Informative Offload Descriptions

   • Updated get_offload_description() function to provide clearer insights into tensor offloading strategy
   • Added better display of GPU layer counts and total model layers
   • Improved qualitative descriptions of offloading levels

4. CLI Version Improvements

   • Fixed issue with Rich library progress bar display during monitoring
   • Improved progress bar feedback with automatic completion when API is detected
   • Added small delay after successful API detection to show completed progress
   • Fixed multiple bugs in monitoring code

5. Tuning and Parameter Controls

   • Added the ability to manually specify GPU layers rather than relying solely on automatic calculation
   • Custom GPU layers settings now properly override the automatic OT level-based calculations
   • Fixed issue with preferred configurations not being respected when starting new tuning sessions
   • Added last-run history display when restarting tuning sessions

6. Model-Specific Improvements

   • Improved model size detection and parameter estimation
   • Added refinement process for model analysis to better handle special cases

7. Better Historical Configuration Handling

   • Added proper persistence of successful configurations marked as preferred
   • Fixed issue where tuning would not retain the exact level from successful previous runs
   • Improved handling of VRAM usage history

8. UX and UI Improvements

   • Better error handling and user feedback
   • Improved handling of model-specific configurations
   • Fixed TclError when closing dialogs in GUI version

Technical Changes Summary

This release contains significant improvements to both GUI and CLI versions:

1. Manual GPU Layer Control: Both GUI and CLI now support manually specifying GPU layers which will override the automatic calculation based on OT level.

2. GPU Layer / Tensor Offload Balance: Refined the relationship between OT levels and GPU layer counts for more predictable behavior with the "More GPU" and "More CPU" controls.

3. CLI Monitoring Improvements: Fixed issues with the progress display and added visual feedback for completed loading.

4. Bug Fixes:

   • Fixed TclError on dialog close in GUI version
   • Fixed progress bar display issues in CLI
   • Corrected the handling of model-specific arguments and settings

5. Core Enhancements:

   • Improved model analysis with better refinement for special cases
   • Enhanced parameter estimation for large and MoE models
   • Better historical configuration selection

These updates provide a more robust and user-friendly experience while maintaining the core functionality that makes TensorTune an effective tool for managing KoboldCpp tensor offloading.