add --ramtorch_transformer_percent and --ramtorch_text_encoder_percent to treat it more like block swap

[bghira]

This pull request introduces support for partial offloading of Linear layers to RamTorch, allowing users to specify the percentage of eligible layers to offload for both transformer and text encoder components. This provides finer control over VRAM usage and training speed. The changes update documentation in multiple languages and modify the codebase to support and utilize the new percentage-based offloading.

Feature: Partial RamTorch Offloading

• Added new configuration options ramtorch_transformer_percent and ramtorch_text_encoder_percent to allow specifying the percentage (0-100) of eligible Linear layers to offload for transformers and text encoders, respectively. These are documented in English, Spanish, Portuguese, Japanese, Hindi, and Chinese option files. [1] [2] [3] [4] [5] [6]

Core Logic Updates

• Updated the _apply_ramtorch_layers method and its usage across all model helper files to accept and pass the percent argument, enabling partial offloading for both transformer models and text encoders. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12]

RamTorch Utility Enhancements

• Refactored replace_linear_layers_with_ramtorch in ramtorch.py to support percentage-based layer replacement, including logic to count eligible layers and select only the specified percentage for replacement. [1] [2] [3] [4]

These changes collectively provide users with more flexibility in managing memory and performance trade-offs during model training.