libai
libai copied to clipboard
add gpt2 based llama2
基于libai gpt2实现的llama2
初步测试
在推理过程,基于gpt2实现llama2结果与原始llama相同,均为
[{'generated_text': 'Give three tips for staying healthy.\nWhat is the best way to stay healthy?\nWhat are the 5 ways to stay healthy?\nWhat are the 5 ways to stay healthy?\nWhat are the '}]
训练过程对比原始libai llama实现初步时间 基于gpt2实现llama训练总时间约为11小时[4卡],其余参数按照原仓库配置
[01/22 13:07:19 lb.utils.events]: eta: 11:00:05 iteration: 9/37320 consumed_samples: 320 total_loss: 1.49 time: 1.0820 s/iter data_time: 0.0027 s/iter total_throughput: 3.70 samples/s lr: 3.62e-08
[01/22 13:07:30 lb.utils.events]: eta: 11:06:09 iteration: 19/37320 consumed_samples: 640 total_loss: 1.415 time: 1.0827 s/iter data_time: 0.0026 s/iter total_throughput: 3.69 samples/s lr: 7.64e-08
原始llama实现训练总时间也约为11小时[4卡]
[01/22 13:03:18 lb.utils.events]: eta: 10:55:21 iteration: 9/37320 consumed_samples: 320 total_loss: 1.49 time: 1.0764 s/iter data_time: 0.0026 s/iter total_throughput: 3.72 samples/s lr: 3.62e-08
[01/22 13:03:29 lb.utils.events]: eta: 11:03:12 iteration: 19/37320 consumed_samples: 640 total_loss: 1.415 time: 1.0771 s/iter data_time: 0.0026 s/iter total_throughput: 3.71 samples/s lr: 7.64e-08
似乎应该需要再调整一下实现的gpt-based模型