-
-
![image](https://github.com/THUDM/VisualGLM-6B/assets/45227019/f724a5af-51b5-402f-91b4-cee9119dfd65)
CUDA_VISIBLE_DEVICES=1 python cli_demo.py --from_pretrained ./checkpoints/finetune-visualglm-6b…
-
我想从本地加载已经下载好的模型来微调,在finetune_visualglm.sh中 的model_args里添加了一个参数 --from_pretrained 后面加上了自己的模型本地路径,但是启动后报错,显示不支持这个方式 ,那么我应该怎么从本地加载模型然后微调啊?
-
微调脚本运行成功了,可以使用cli_demo.py的方式运行。
未进行修改前的web_demo.py可以正常运行,但是修改inter_util.py中的模型路径后,再运行则报错。
![image](https://github.com/THUDM/VisualGLM-6B/assets/16079587/d0d42732-6fe3-4274-acd2-ceb9b38ca825)
![im…
-
` from transformers import AutoTokenizer, AutoModel
tokenizer = AutoTokenizer.from_pretrained("THUDM/visualglm-6b", trust_remote_code=True)
model = AutoModel.from_pretrained("THUDM/visualglm-6b", …
-
你好,目前使用3个节点对visualGLM进行微调(Lora),每个节点4张16GB的卡,一共是12张16GB的卡,依旧报OOM,请问我还需要增加多少才够用?
-
因为不想部署2个模型,请问VisualGLM-6B能同时支持普通的gpt问答、和图文问答吗?就像gpt4那样
-
加载微调模型的部分为:
```
model, model_args = AutoModel.from_pretrained(
‘/data/xxx/workspace/VisualGLM-6B-main/checkpoints/finetune-visualglm-6b-07-10-10-16’,
args=argparse.Namespace(
…
-
微调时报错如下:
![image](https://github.com/THUDM/VisualGLM-6B/assets/38753856/d45adf2a-69ef-4656-a015-65e3bebfc2a7)
应该是调用sat库的get_tokenizer出错了,有没有办法可以调用本地的tokenizer?
-
如题,微调所需要哦的sat的模型下载路径打不开了,visualglm-6b.zip文件可以去哪里获得呢