gnh1201/Open-Llama
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

update server

Browse Source
This commit is contained in:
LiangSong 2023-04-02 21:51:32 +08:00
parent 3d0d4e8af7
commit 9857ad03f1
1 changed files with 3 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

11
server.py
View File

@ -32,10 +32,7 @@ raw_model = LlamaForCausalLM(
shared_input_output_embedding=True, shared_input_output_embedding=True,
) )
) )
ckpt = torch.load( ckpt = torch.load("data/saved_ckpt/instruction_tuning/14001.pt", map_location="cpu")
"data/saved_ckpt/instruction_tuning/14001.pt",
map_location="cpu",
)
raw_model.load_state_dict(ckpt) raw_model.load_state_dict(ckpt)
raw_model.eval() raw_model.eval()
model = raw_model.cuda() model = raw_model.cuda()
@ -68,10 +65,8 @@ demo = gr.Interface(
"Can you explain to me what quantum mechanics is and how it relates to quantum computing?", "Can you explain to me what quantum mechanics is and how it relates to quantum computing?",
"I'm feeling a bit unwell in my stomach today. Do you have any suggestions for dinner?", "I'm feeling a bit unwell in my stomach today. Do you have any suggestions for dinner?",
], ],
).queue(
concurrency_count=1,
title="Open-Llama", title="Open-Llama",
description="不基于其他预训练模型，完全使用Open-Llama项目从0开始训练的Instruct-GPT模型，总共训练花费在2w美元以内，训练时间80h。", description="不基于其他预训练模型，完全使用[Open-Llama](https://github.com/Bayes-Song/Open-Llama)项目从0开始训练的Instruct-GPT模型，总训练成本不超过2w美元。由于请求需要经Gradio进行转发，可能出现请求丢失的现象，当长时间无响应（如20s以上）可刷新重试。当前体验服务生成的所有内容都是由人工智能模型生成，我们对其生成内容的准确性、完整性和功能性不做任何保证，并且其生成的内容不代表我们的态度或观点。",
article="[关于作者](http://home.ustc.edu.cn/~sl9292/resume.html)", article="[关于作者](http://home.ustc.edu.cn/~sl9292/resume.html)",
) ).queue(concurrency_count=1)
demo.launch(share=True) demo.launch(share=True)