InternVL icon indicating copy to clipboard operation
InternVL copied to clipboard

InternVL2_5-4B-MPO lora微调

Open ChenJian7578 opened this issue 11 months ago • 4 comments

在目录下没有看到MPO lora的微调,请问是目前不支持吗,还是说MPO的lora微调用的是2_5的lora微调脚本?

ChenJian7578 avatar Jan 09 '25 06:01 ChenJian7578

你好,我尝试用2.5的lora微调脚本来进行MPO的lora微调,但我遇到了warning: shape mismatch: value tensor of shape [4608, 4096] cannot be broadcast to indexing result of shape [1098, 4096], i nput_embeds[selected].shape=torch.Size([1098, 4096]), vit_embeds.shape=torch.Size([4608, 4096])这样的问题,请问你遇到了吗,或者你成功使用lora微调了吗,感谢

JackeyHRan avatar Jan 10 '25 12:01 JackeyHRan

没有试过,应该是还没支持吧 ---- 回复的原邮件 ---- 发件人haoran @.>发送日期2025年01月10日 20:34 @.> @.>, @.>主题Re: [OpenGVLab/InternVL] InternVL2_5-4B-MPO lora微调 (Issue #839) 你好,我尝试用2.5的lora微调脚本来进行MPO的lora微调,但我遇到了warning: shape mismatch: value tensor of shape [4608, 4096] cannot be broadcast to indexing result of shape [1098, 4096], i nput_embeds[selected].shape=torch.Size([1098, 4096]), vit_embeds.shape=torch.Size([4608, 4096])这样的问题,请问你遇到了吗,或者你成功使用lora微调了吗,感谢 — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: @.***>

ChenJian7578 avatar Jan 11 '25 12:01 ChenJian7578

同问,MPO训练是否支持lora方式

zwang-datascience avatar Jan 21 '25 03:01 zwang-datascience

同问

cooperong avatar Mar 06 '25 01:03 cooperong