Pytorch training slow

1

I have this project file using RoBERTa to create a model about multiclass text classification. Before I tested on my own laptop, I’m using google colab. It only took about 20-30 minutes for training. Then you know, it’s free version so I’ve reach my usage limit.
After that I’m installing jupyterlab & pytorch cuda version 12.1 on my laptop. I didn’t install cuda toolkit from nvidia website. Then I run it, it works. But when training my model, it says approximately 6-7 hours. Is that normal? Or I have some settings/configs wrong. Tbh my gpu laptop isn’t that good too. You know google colab has Nvidia Tesla T4 and my laptop was RTX 2050. But the difference is so big, I tried to use my cpu instead gpu then it says approximately 12 hours.

Please let me know if this normal or there’s something isn’t right. I have knowledge in machine learning before, but for NLP this is my first time

image
image1110×557 114 KB

2

You could profile your code with a visual profiler, e.g. Nsight Systems, to see where the bottleneck in your code is. E.g. it could be the data loading, the CPU workload (i.e. scheduling the actual kernels), or the GPU itself.