Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Revert "【Hackathon 5th No.83】PaddleMIX ppdiffusers models模块功能升级同步HF" -part #354

Merged
merged 1 commit into from
Dec 19, 2023

Conversation

JunnYu
Copy link
Member

@JunnYu JunnYu commented Dec 19, 2023

Reverts #322
回滚该PR,等之后修复后再合入,下面是对应的问题。

  1. fastdeploy的sd和sdxl 推理造成了错误。无法顺利加载vae的decoder模型,可能当前升级有部分修改,导致这个发生错误。
  2. sd的预训练和lora微调存在性能下降10%,可能怀疑是recompute粒度发生了修改,部分模块之前没有加recompute现在加上了recompute导致的。

@JunnYu JunnYu requested a review from westfish December 19, 2023 05:12
Copy link

paddle-bot bot commented Dec 19, 2023

Thanks for your contribution!

@JunnYu JunnYu merged commit 449bb34 into develop Dec 19, 2023
1 of 3 checks passed
@co63oc
Copy link
Contributor

co63oc commented Dec 19, 2023

1 测试vae_decoder目录已生成,再测试问题

2 这是使用 recompute的文件和类
ppdiffusers/ppdiffusers/models/transformer_2d.py Transformer2DModel
ppdiffusers/ppdiffusers/models/unet_2d_blocks.py UNetMidBlock2DCrossAttn
ppdiffusers/ppdiffusers/models/unet_3d_blocks.py DownBlockMotion CrossAttnDownBlockMotion CrossAttnUpBlockMotion UpBlockMotion UNetMidBlockCrossAttnMotion
ppdiffusers/ppdiffusers/models/vae.py EncoderTiny DecoderTiny
recompute是需要开启,如果不开启recompute对性能有影响那是其他问题,修改主要是增加lora scale参数,增加scale的计算,应该也没有这样的影响

@JunnYu
Copy link
Member Author

JunnYu commented Dec 19, 2023

@co63oc
问题1中,跑下面的代码会报错,不是导出时候报错,而是加载的时候报错了

python infer.py --model_dir static_model/stable-diffusion-v1-5 --scheduler "preconfig-euler-ancestral" --backend paddle_tensorrt --device gpu --task_name text2img
python infer.py --model_dir static_model/stable-diffusion-v1-5 --scheduler "preconfig-euler-ancestral" --backend paddle_tensorrt --device gpu --task_name img2img
python infer.py --model_dir static_model/stable-diffusion-v1-5 --scheduler "preconfig-euler-ancestral" --backend paddle_tensorrt --device gpu --task_name inpaint_legacy

image

@Ligoml Ligoml changed the title Revert "【Hackathon 5th No.83】PaddleMIX ppdiffusers models模块功能升级同步HF" Revert "【Hackathon 5th No.83】PaddleMIX ppdiffusers models模块功能升级同步HF" -part Dec 20, 2023
@jerrywgz jerrywgz deleted the revert-322-attention branch August 1, 2024 03:11
westfish pushed a commit to westfish/PaddleMIX that referenced this pull request Sep 25, 2024
…addlePaddle#354)

Reverts PaddlePaddle#322
回滚该PR,等之后修复后再合入,下面是对应的问题。
1. fastdeploy的sd和sdxl 推理造成了错误。无法顺利加载vae的decoder模型,可能当前升级有部分修改,导致这个发生错误。
2.
sd的预训练和lora微调存在性能下降10%,可能怀疑是recompute粒度发生了修改,部分模块之前没有加recompute现在加上了recompute导致的。
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants