感谢大佬们开源本项目。记录并分享下自己遇到的问题以及如何解决的 #251
wangyongshuai88
started this conversation in
General
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
-------跑通代码阶段-----
1 遇到一些版本错误问题,最终使用的是以下版本:
py-cpuinfo==8.0.0
pathtools-0.1.1
bitsandbytes== 0.37.2
scipy== 1.10.1
zipp==3.14.0
ffmpy==0.2.0
pydub==0.25.0
cudatoolkit==11.7.1
2 遇到的问题:
RuntimeError: Internal: unk is not defined.
原因:下载模型以后可能会遇到tokenizer.model 为0B的情况,需要重新下载。
3 /home/sher/anaconda3/envs/chinese_vicuna_llama_lora/lib/python3.8/site-packages/bitsandbytes/libbitsandbytes_cpu.so: undefined symbol: cget_col_row_stats
nvidia-smi
cp libbitsandbytes_cuda自己的版本.so libbitsandbytes_cpu.so
4 torch没有自动安装cuda, 需要在conda 里面直接安装一下cudatoolkit==11.7.1
5 使用wsl的时候bitandbytes 找不到,由于wsl 里没有安装驱动,使用的是nvidia特殊定制的cuda调用的是windows里的驱动。
export CUDA_HOME=/usr/local/cuda-11.8
export LD_LIBRARY_PATH=/usr/local/cuda-11.8/lib64:/usr/lib/wsl/lib
export PATH=$CUDA_HOME/bin:$PATH
6 AttributeError: 'NoneType' object has no attribute 'device'
需要调整:
----------------微调阶段----------------
下载下来的guanaco_belle_merge_v1.0 是json 格式,需要转成jsonl 格式的
可以使用以下代码。:
另外8bit 微调13B模型的时候显存占用,这么少的显存是正常的吗?可以交流下。
(MICRO_BATCH_SIZE=4 BATCH_SIZE=128 ) 显存占用是16GB,CUDA使用率70%
(MICRO_BATCH_SIZE=16 BATCH_SIZE=128 ) 显存占用是20GB,CUDA使用率95%以上
(MICRO_BATCH_SIZE=26 BATCH_SIZE=128 ) 显存占用是22GB
(MICRO_BATCH_SIZE=32 BATCH_SIZE=128 ) 显存占用是22.9GB, CUDA 使用率98%
Beta Was this translation helpful? Give feedback.
All reactions