CASIA-IVA-Lab / AnomalyGPT

[AAAI 2024 Oral] AnomalyGPT: Detecting Industrial Anomalies Using Large Vision-Language Models
https://anomalygpt.github.io
Other
672 stars 81 forks source link

vicuna7B的delta权重模型一定要用V0版本的么?高于V0版本,1.1或者1.5行不行? #63

Closed plshuaishuai2 closed 5 months ago

plshuaishuai2 commented 5 months ago

vicuna7B的delta权重模型一定要用V0版本的么?高于V0版本,1.1或者1.5行不行?V0版本训练词表的维度和llama词表的维度不一致,V0版本词表是32001,而llama词表是32000,因此对输入embedding编码的参数维度不一致

zhtstar commented 5 months ago

Maybe the bug is caused by the versions of fschat and transformers you are using. This page introduces the corresponding relationship between model weights and the versions of fschat and transformers. I hope it can help you. https://github.com/lm-sys/FastChat/blob/main/docs/vicuna_weights_version.md#how-to-apply-delta-weights-for-weights-v11-and-v0

Here are my settings, no errors reported: vicuna-7b-delta-v1.1 llama-7b-hf fschat version is 0.2.1 transformers version is 4.37.0.dev0

the flowing also is ok: vicuna-7b-delta-v0 llama-7b-hf fschat version is 0.1.10 transformers version is 4.37.0.dev0

FantasticGNU commented 5 months ago

@zhtstar Thank you for your answer!