lllyasviel/ControlNet

Training still OOM on 8GB gpu.

Open

#21 创建于 2023年2月12日

在 GitHub 查看
 (2 评论) (3 反应) (0 负责人)Python (27,526 star) (2,500 fork)batch import
help wanted

描述

It seems that I already used many tricks but training still OOM for 8GB gpu. But inference is good now.

This is strange because I know some textural inversion or dreambooth can be trained on 8GB.

What is the secrect of Automatic1111's optimization? Although xformers may help a bit, the currect sliced attention should require even smaller mem than xformers.

Does it make sence to move text encoder and vae outside gpu when training?

贡献者指南

技术栈
pythonpytorch
领域
machine learning
议题类型
bug
难度面向新贡献者的预计实现难度，1 表示很小改动，5 表示专家级工作。
3
预计时间有经验贡献者完成调查、实现、测试并准备 pull request 的粗略时间范围。
1-2 days
活动状态议题当前的可参与程度：新鲜、活跃、陈旧、阻塞或等待维护者输入。
stale
清晰度议题是否清楚说明期望改动、验收标准和下一步。
needs investigation
前置要求
PyTorch basicsGPU memory managementdiffusion modelsattention mechanisms
新手友好度1-100 的估计分数，表示该议题对首次贡献者的友好程度。
20
研究方向
Investigate the memory usage during training in ControlNet. Focus on the attention mechanism and compare with xformers. Profile memory using PyTorch's memory tools. Consider moving text encoder and VAE to CPU as suggested. Check if there are any known optimizations from Automatic1111's repository.