Ecosyste.ms: Timeline

Browse the timeline of events for every public repo on GitHub. Data updated hourly from GH Archive.

zRzRzRzRzRzRzR

zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
哦我说的这个版本是FP8，BF16能正常的，fp8会出现你说的这个问题，我们还没处理过这个问题，目前人手有限，这个地方的尝试可能暂时没办法支持

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
Please note that for version 1.0, you need to go to the release version of the code. The sat in the main repository code is version 1.5; you can find the inference code for version 1.0 in the relea...

View on GitHub

zRzRzRzRzRzRzR closed an issue on THUDM/CogVideo
Any chance to have a gradio app for ease of use?
### Feature request / 功能建议 A gradio app will be useful to try the new model. ### Motivation / 动机 Ease of use and trial. ### Your contribution / 您的贡献 I saw the gradio composite demo on the rep...
zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
I didn't understand, do you want to use multiple images instead of a single image to complete this task? I haven't tried this before, and I believe the current structure is unable to accomplish suc...

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
可以，但是难点在准备数据和控制性上。 这个完全得看你的数据是否质量够高，此外，关于控制性，该模型的微调效果也同样存在跟原始模型差不多的控制性。需要自己测试

View on GitHub

zRzRzRzRzRzRzR closed an issue on THUDM/CogVideo
cogvideo能通过训练lora实现这种换发型的效果吗
### Feature request / 功能建议 通过训练lora实现这种换发型的效果 ### Motivation / 动机 通过训练lora实现这种换发型的效果 ### Your contribution / 您的贡献 ![GI132131232151511F](https://github.com/user-attachments/assets/aa6d1b85-7c5...
zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
Can you provide detailed explanations, thank you.

View on GitHub

zRzRzRzRzRzRzR created a review on a pull request on THUDM/CogVideo

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on THUDM/GLM-Edge
没有，这是一个基础的VQA模型

View on GitHub

zRzRzRzRzRzRzR closed an issue on THUDM/GLM-Edge
请问视觉模型是否有grounding能力？
### Feature request / 功能建议 请问视觉模型是否有grounding能力？ ### Motivation / 动机 请问视觉模型是否有grounding能力？ ### Your contribution / 您的贡献 请问视觉模型是否有grounding能力？
zRzRzRzRzRzRzR created a comment on an issue on THUDM/GLM-Edge
这个模型没有function call能力哦

View on GitHub

zRzRzRzRzRzRzR closed an issue on THUDM/GLM-Edge
4b-chat有tools插件调用能力吗，如果具备能提供代码例子吗，谢谢
### Feature request / 功能建议 4b-chat有tools插件调用能力吗，如果具备能提供代码例子吗，谢谢 ### Motivation / 动机 4b-chat有tools插件调用能力吗，如果具备能提供代码例子吗，谢谢 ### Your contribution / 您的贡献 4b-chat有tools插件调用能力吗，如果具备能提供代码例子吗，谢谢
zRzRzRzRzRzRzR pushed 1 commit to main THUDM/GLM-Edge

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
You can check the readme on the GitHub homepage. The main differences are changes in resolution and frame rate, with the resolution being improved and the effect also being better, supporting 16 fr...

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on a-r-r-o-w/cogvideox-factory
Therefore, during training, an image condition is needed, and during reasoning, when using the decoder, it is necessary to divide by the scale factor

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on a-r-r-o-w/cogvideox-factory
Thank you for your help @a-r-r-o-w. When CogVideoX 1.5 was trained, the image condition was not multiplied, so during inference, the image condition cannot be multiplied, but before the final deco...

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
这个版本的模型我的理解是无法挂lora的，目前fp8的显存消耗和BF16差不多，建议使用BF16 + cpu offload的方式

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on THUDM/GLM-Edge
> 感谢回复。目前看到端侧是decode的结构了，而glm应该是encoder的结构，这两者在部署态他们的结构差异在哪里呢? 另外请问高通部署使用genie sdk了吗？ @zRzRzRzRzRzRzR 开源的 GLM-Edge 以及 GLM-4-9B-Chat 模型是 decoder-only结构哈，可以自行查看模型文件中的modeling_chatglm.py

View on GitHub

zRzRzRzRzRzRzR pushed 1 commit to main THUDM/GLM-Edge

View on GitHub

zRzRzRzRzRzRzR pushed 1 commit to main zRzRzRzRzRzRzR/zRzRzRzRzRzRzR
  • Update research direction c8c43ae

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
全亮微调的实现和lora有一定区别，但是目前全量微调一定OOM，目前所有仓库都提供lora而非全量

View on GitHub

zRzRzRzRzRzRzR closed an issue on THUDM/CogVideo
CogVideoX1.5-5B-I2V how to finetune
can finetune CogVideoX1.5-5B-I2V？
zRzRzRzRzRzRzR closed an issue on THUDM/CogVideo
pad_image of rife_model.py is inaccurate can you fix it
This part causing error ``` def pad_image(img, scale): _, _, h, w = img.shape tmp = max(32, int(32 / scale)) ph = ((h - 1) // tmp + 1) * tmp pw = ((w - 1) // tmp + 1) * tmp...
zRzRzRzRzRzRzR closed an issue on THUDM/CogVideo
Generated video with third-party watermarks
### System Info / 系統信息 As tile, The generated videos appears as a "pond5" watermark. Is this normal? ### Information / 问题信息 - [ ] The official example scripts / 官方的示例脚本 - [ ] My own modi...
zRzRzRzRzRzRzR closed an issue on THUDM/CogVideo
想请教一下论文中的3D full attention的实现具体在哪里呢？
看到paper中提出了3D full attention和2D+1D attention的对比，并实验证明说生成视频的效果更好。论文中也提到可以用多种并行策略优化，想请教一下这部分有开源的代码吗
zRzRzRzRzRzRzR closed an issue on THUDM/CogVideo
mask策略咨询
### System Info / 系統信息 ubuntu20.04/A800 python3.10 cuda12.0 ### Information / 问题信息 - [X] The official example scripts / 官方的示例脚本 - [ ] My own modified scripts / 我自己修改的脚本和任务 ### Reproduction / 复现...
zRzRzRzRzRzRzR closed an issue on THUDM/CogVideo
video to video fine tune
### Feature request / 功能建议 Hi authors, thanks for open-sourcing such an amazing project! I've carefully read through your paper and the GitHub repository. I noticed there's a notebook called `Co...
zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
After adding pipe.enable_sequential_cpu_offload(), the speed is significantly slower, which is normal, but I don't understand why removing it causes your A100 to run out of memory, with the normal ...

View on GitHub

zRzRzRzRzRzRzR created a comment on an issue on THUDM/CogVideo
这个大量噪声估计是版本没对吧，源代码安装的diffusers并把生成的参数从1.0版本修改成1.5试试，更多的问题，请重新提一个带有具体报错或者复线脚本的issue，感谢理解

View on GitHub

Load more