search

ArvinZhuang / DSI-QG

The official repository for "Bridging the Gap Between Indexing and Retrieval for Differentiable Search Index with Query Generation", Shengyao Zhuang, Houxing Ren, Linjun Shou, Jian Pei, Ming Gong, Guido Zuccon and Daxin Jiang.
MIT License
110 stars 16 forks source link

Single GPU feasibility #5

Open gusye1234 opened 1 year ago

gusye1234 commented 1 year ago

Hi. Note that the demo command will launch the training in 8 GPU. Have you tested running this task on a single GPU, and how long will it take?

I tried to follow up on this work but only got one GPU... So if the training speed is hard to bear, I might consider adding more GPUs.

ArvinZhuang commented 1 year ago

Hi @gusye1234! thanks for using our code! You definitely could try one GPU with --gradient_accumulation_steps first, DSI with QG actually converges pretty fast with 8 gpus, so I think speed on one GPU is acceptable.