Jacobew

> 这个错是openai没有额度了，现在新注册的号没有免费额度了 @zhayujie 怎么查看确认是否有免费额度呢？ clash的端口一定是7890么？我看控制台有多个端口，每个节点的配置也有对应的端口

> 这个错是openai没有额度了，现在新注册的号没有免费额度了 @zhayujie 你好，我看了下确实是没有免费额度了，那大陆地区有什么渠道可以申请付费usage呢？除了depay信用卡以外有什么方便的途径不？

> 新买个账号就可以了 请问现在还有可靠的买账号的途径吗？

> This isn't the library's fault, there is a problem with installation. Here are some possible fixes: Try running `pip install --upgrade revChatGPT` This should update the version, there should...

@Haijunlv Hi, any update here?

Another question: why restoring the order in the decoder can eliminate the need for the assignment-based loss? In this way, the decoder would output the elements in the same arbitrary...

I am also a little frustrated about figure1 (the color and the dashed box), for which I think a concrete example including input coordinate number and simple network transformation is...

Hi, have you solve this problem? I'm having the same issue.

+1, 请问下attention操作是否有被计算在内呢？看上去一个block中的mac等于qkv & proj的和，并没有考虑mha部分 @MrYxJ 

 hi， @JJJYmmm 我尝试加载qwen visual encoder的权重（Qwen2.5-VL-7B-Instruct，不包含merger，https://github.com/huggingface/transformers/blob/40a493c7ed4f19f08eadb0639cf26d49bfa5e180/src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py#L530） 用于下游任务的finetune，但是发现loss效果比较差，甚至比不过不带pretrain的vit-huge和vit-large 我对单个样本的forward进行了对齐，训练时visual encoder的输出能够和model.generate api的visual feature output完全对齐（检查过了均值归一化/dtype细节），用的sgd optimizer，输入encoder前的前处理参照了https://github.com/huggingface/transformers/blob/40a493c7ed4f19f08eadb0639cf26d49bfa5e180/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py#L168 进行了reshape & repeat，请问有什么debug思路么？ @JJJYmmm 非常感谢回复！麻烦再帮忙瞅瞅