Xingchen Song(宋星辰)

me either

also confused about this Q

@Colanim Hi, i think the answer is : For the conventional language model, we try to predict the next word based on the previous state, but in XLNet ,the model...

> @kimiyoung Hi, I also have the problem here. I tried the source code **_local_perm** in **data_utils** with a given input. > inputs: 1,2,3,4,5,6,7,8,9,10 > is_masked:False,False,False,False,False,True, True, False,False, False >...

also confused about this Q.

and also see this: https://github.com/wenet-e2e/wenet/issues/583#issuecomment-907994058 A simple solution is setting `language_type = kIndoEuropean ;`

cc @pengzhendong

your libtorch version dose not match system arch (x86/aarch64/... etc)

Maybe it is just a `warmup` issue, in https://github.com/wenet-e2e/wenet/pull/942, we can clearly see that warmup round(round-1) is significantly slower than the forllowing rounds. We can draw a conclusion that both...

1. yes, and it is mostly caused by different versions of libtorch 2. I think the time spend for warmup is not a big issue in production env since we...