jason law

Error with fast_jl

1

CUDA error: the provided PTX was compiled with an unsupported toolchain. What might be the cause for this mistake?

I use 8 A100 with 80G memorys, it always get out out memory in the second step. In my previous attempt, It will not meet such problems when fine-tuning LLaVA-Vicuna-13B,...

Questions about the influence value

1

I'm wondering what's the appropriate influence value in LESS setting. I'm reproducing it and the max influence value across all sub-tasks are about 0.1-0.4 for some samples. Is this value...

loss curve of llava-next-llama3

4

Thanks for your great work! I'm wondering if u can share the loss curve for training llava-next-llama3? I've observed some different behaviour compared to training llava-next-vicuna-7b. I'm wondering if it's...

Is there a seed for fine-tuning?

2

### Question I'm wondering if there is a seed for fine-tuning. Currently the reproduced results for second stage fine-tuning are all different from the results of official checkpoint. I'm certain...

Always output chinese in thought

1

"response": "\u4ece\u622a\u56fe\u53ef\u4ee5\u770b\u5230\uff0c\u8fd9\u662fLinux\u7cfb\u7edf\u684c\u9762\uff0cChrome\u6d4f\u89c8\u5668\u5df2\u7ecf\u6253\u5f00\uff0c\u5e76\u4e14\u6709\u4e00\u4e2a\u7f51\u9875\u5728\u663e\u793a\u4e2d\u3002\u6839\u636e\u4efb\u52a1\u8981\u6c42\uff0c\u6211\u8981\u5c06\u5f53\u524d\u7f51\u9875\u6dfb\u52a0\u5230\u4e66\u7b7e\u680f\u4ee5\u4fbf\u7a0d\u540e\u8fd4\u56de\u3002\n\n \u6211\u9700\u8981\u4f7f\u7528Chrome\u6d4f\u89c8\u5668\u7684\u5feb\u6377\u952e\"Ctrl+D\"\u6765\u5feb\u901f\u5c06\u5f53\u524d\u9875\u9762\u6dfb\u52a0\u5230\u4e66\u7b7e\u4e2d\u3002\u8fd9\u4e2a\u5feb\u6377\u952e\u662fChrome\u7684\u6807\u51c6\u64cd\u4f5c\u65b9\u5f0f\u3002\nAction: hotkey(key='ctrl d') Why does UI-TARS always output chinese in its thought? Is it designed to do so?

Problem about evaluation on OSWorld

6

Thanks for your great work! I'm trying to replicate the evaluation results of UI-TARS model on OSWorld. I'm using a remote server that does not support visible desktop(only supporting shell)....

Issues while trying to reproduce the results on LLaVA-v1.5

13

Thanks for your excellent work! I'm trying to reproduce this method on LLaVA-v1.5 model. But I've encounted one problem: File ~/anaconda3/envs/llava/lib/python3.10/site-packages/torch/autograd/__init__.py:200, in backward(tensors, grad_tensors, retain_graph, create_graph, grad_variables, inputs) 195 retain_graph...