编译过程碰到的问题

[lileilai]

按照编译的文档：https://github.com/bytedance/lightseq/blob/master/docs/inference/build.md 碰到了下面的问题，请问可以怎么解决？

[godweiyang]

There are no libcudadevrt.so and libcudart_static.so in your lib64 dir.

[lileilai]

There are no libcudadevrt.so and libcudart_static.so in your lib64 dir.

I am confused with the cuda version, according to the build.md，cuda version should be <11 and >=10.1。but，at the CmakeList.txt，it's required cuda11

[hexisyztem]

This document is out of date

There are no libcudadevrt.so and libcudart_static.so in your lib64 dir.

I am confused with the cuda version, according to the build.md，cuda version should be <11 and >=10.1。but，at the CmakeList.txt，it's required cuda11

[lileilai]

document

Thanks for your reply, i have build success and got a whl package as follows:

but, where i try to import T5/Bert/Bart in python， indicating there are no such modules:

[hexisyztem]

document

Thanks for your reply, i have build success and got a whl package as follows:

but, where i try to import T5/Bert/Bart in python， indicating there are no such modules:

How did you compile the whl package? I recommend you to use this command: " ENABLE_FP32=0 ENABLE_DEBUG=0 ENABLE_NEW_ARCH=0 python3 setup.py install"

[hexisyztem]

In addition, you can get the Bert model through lsi.Bert(...)

[hexisyztem]

You can find interface code in this file: https://github.com/bytedance/lightseq/blob/master/lightseq/inference/pywrapper/wrapper.cc#L919

[hexisyztem]

for example: import lightseq.inference as lsi model = lsi.Bert(model_file, max_batch_size)

In addition, you can get the Bert model through lsi.Bert(...)

[lileilai]

document

Thanks for your reply, i have build success and got a whl package as follows: but, where i try to import T5/Bert/Bart in python， indicating there are no such modules:

How did you compile the whl package? I recommend you to use this command: " ENABLE_FP32=0 ENABLE_DEBUG=0 ENABLE_NEW_ARCH=0 python3 setup.py install"

i got the wheel file by following:

and pip install **

[hexisyztem]

It looks like you did compile correctly, it's just the wrong way to call it

document

Thanks for your reply, i have build success and got a whl package as follows: but, where i try to import T5/Bert/Bart in python， indicating there are no such modules:

How did you compile the whl package? I recommend you to use this command: " ENABLE_FP32=0 ENABLE_DEBUG=0 ENABLE_NEW_ARCH=0 python3 setup.py install"

i got the wheel file by following:

and pip install **

[lileilai]

for example: import lightseq.inference as lsi model = lsi.Bert(model_file, max_batch_size)

In addition, you can get the Bert model through lsi.Bert(...)

i have try this way, but it said "module not found"

[lileilai]

i try to check the compile process, my cuda version is 11.6, and i can not build lightseq program with the script "sh build.sh", and the error as follows:

Then , i switch the cuda version to cuda-10.2, and error as following:

[Taka152]

You could check if you install cudatoolkit-dev correctly with nvcc, cublas, etc.

[Fizzmy]

I meet the same problem. And I wonder have you solve it now？