search

THUDM / VisualGLM-6B

Chinese and English multimodal conversational language model | 多模态中英双语对话语言模型
Apache License 2.0
4.05k stars 414 forks source link

visualglm进行QLoRA微调时报错，RuntimeError: mat1 and mat2 shapes cannot be multiplied (320x4096 and 1x25165824) [2024-03-07 07:26:17,037] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 13210 #343

Open munDane117 opened 4 months ago

munDane117 commented 4 months ago

运行环境：colab v100 51G系统RAM 16G GPU RAM

python==3.10.13 bitsandbytes==0.42.0 已是最新版

运行 bash finetune/finetune_visualglm_qlora.sh 脚本时报错 RuntimeError: mat1 and mat2 shapes cannot be multiplied (320x4096 and 1x25165824) [2024-03-07 07:26:17,037] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 13210

issue2

2232141528 commented 3 months ago

同问，请问兄弟问题解决了吗？

2232141528 commented 3 months ago

运行环境：colab v100 51G系统RAM 16G GPU RAM

python==3.10.13 bitsandbytes==0.42.0 已是最新版

运行 bash finetune/finetune_visualglm_qlora.sh 脚本时报错 RuntimeError: mat1 and mat2 shapes cannot be multiplied (320x4096 and 1x25165824) [2024-03-07 07:26:17,037] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 13210

issue2

请问问题解决了吗兄弟

2232141528 commented 3 months ago

运行环境：colab v100 51G系统RAM 16G GPU RAM

python==3.10.13 bitsandbytes==0.42.0 已是最新版本

运行 bashfinetune/finetune_visualglm_qlora.sh 脚本表述错 RuntimeError: mat1 and mat2 Shapes can be multiplied (320x4096 and 1x25165824) [2024-03-07 07:26:17,037] [INFO] [launch.py​​:316:sigkill_handler] Killing subprocess 13210

问题2

请问是怎么解决这个问题的？

munDane117 commented 3 months ago

运行环境：colab v100 51G系统RAM 16G GPU RAM python==3.10.13 bitsandbytes==0.42.0 已是最新版本 运行 bashfinetune/finetune_visualglm_qlora.sh 脚本表述错 RuntimeError: mat1 and mat2 Shapes can be multiplied (320x4096 and 1x25165824) [2024-03-07 07:26:17,037] [INFO] [launch.py​​:316:sigkill_handler] Killing subprocess 13210 问题2

请问是怎么解决这个问题的？

我没找到解决办法，换了张4090本地部署就没有再遇到这个问题

2232141528 commented 3 months ago

运行环境：colab v100 51G系统RAM 16G GPU RAM python==3.10.13 bitsandbytes==0.42.0 已是最新版本 运行 bashfinetune/finetune_visualglm_qlora.sh 脚本淀粉错RuntimeError: mat1 and mat2 Shapes can be multiplied (320x4096 and 1x25165824) [2024-03-07 07:26:17,037] [INFO] [launch.py​​​​:316:sigkill_handler] 杀死子进程 13210 问题2

请问这个问题是怎么解决的？

我没有找到解决办法，换了张4090本地部署就没有再遇到这个问题

我感觉像是我用的哪个huggingface的镜像网站的问题，我仔细看了那个网站里边的visualglm模型是HF版本的，而我用的是sat版本的，之前为了解决AttributeError: 'FakeTokenizer' object has no attribute 'encode'，偷懒用了镜像，估计应该是两个的encode不一样的。