Yushi Bai

icon indicating a website link https://bys0318.github.io/

icon company Tsinghua University icon location UTC+8 icon location Tsinghua University, China

Results 102 comments of Yushi Bai

Needel_test CUDA OOM 了应该怎么解决？

如果条件允许的话，可以用多gpu推理，只需要在load模型时传入`device_map="auto"`

train 成功了，分享一些细节。 trained successfully! Share some details with guys.

Thanks! Here is the English version: The training was successful! Here are the details of the environment: - Environment: - python==3.11.9 - transformers==4.33.0 - pytorch==2.2.0 - flash-attn==2.6.3 - ninja==1.11.1.1 -...

train 成功了，分享一些细节。 trained successfully! Share some details with guys.

> @LYCnight 大佬 训练的时候预计要多大的显存 GLM-4-9b 32k训练需要8卡80G。如果显存不够可以试试lora或者qlora。

train 成功了，分享一些细节。 trained successfully! Share some details with guys.

> @LYCnight 按照大佬的环境，成功的开启了训练，但是为什么训练完后的文件超级大呢？存储空间不够了…… -rw-r--r-- 1 root root 4984147224 Sep 5 10:49 model-00001-of-00004.safetensors -rw-r--r-- 1 root root 4895071360 Sep 5 10:49 model-00002-of-00004.safetensors -rw-r--r-- 1 root root 4895071384 Sep 5 10:49 model-00003-of-00004.safetensors...

Error making gguf: KeyError: '<|user|>'

Hi! You can get the token id by `tokenizer.get_command("")`.

Error making gguf: KeyError: '<|user|>'

Have you updated to our most recent model files? Also, please use `transformers>=4.43.0`.

Error making gguf: KeyError: '<|user|>'

This is correct. Thanks for sharing!

Token indices sequence length is longer than the specified maximum sequence length for this model (1113927 > 128000). Running this sequence through the model will result in indexing errors

This is just a warning, right? You can ignore it.

什么时候能上架智谱AI开放平台？

你好，目前我们还没有确定的上架时间，之后我们会release具有更强的长输出性能而且更大的模型，应该会在1-2个月内。

什么时候能上架智谱AI开放平台？

sorry～我们暂时还不能透露