RCreddit.com
17
·开发者社区 · RSS
GitHub - kallewoof/tftf: Transforming Transformers -- ultra light-weight pipeline for enormous transformer model manipulation with minimal overhead
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
tftf（transforming transformers）项目旨在解决处理大型模型时，因模型大小超出显存和内存限制而导致的LoRA合并或格式转换等操作困难的问题。该工具的核心思想是对每个张量进行操作，避免将整个模型加载到（显存）内存中，从而以最小的开销实现对大型Transformer模型的处理。tftf主要面向需要微调模型，特别是大型模型的用户。
Working with large models that don't fit in your VRAM+RAM is extremely annoying when you want to do things like LoRA merging or converting between formats, so I started the tftf (transforming transformers) project.
The idea is simple: do all operations on a per tensor level. Never load "everything" into (V)RAM. This tool is mostly aimed at people who finetune models, especially big models.
Vibe-coded for the most part, but figured I'd post it, since it is probably useful to finetuners.
主题标签GitHub模型发布开源代码
原始关键词#manipulation#transformers#transforming#transformer#kallewoof#enormous