Error when I used pytorch to run

[michelleqyhqyh]

python pytorch/decoder_sample.py 8 6 32 8 64 --time

[byshiue]

Please describe the hardward, environment, docker, branch and building scripts you use, thanks.

[michelleqyhqyh]

Please describe the hardward, environment, docker, branch and building scripts you use, thanks.

I compile it in my local sever. So I didn't use docker. My system is CENTOS7. My GPU is V100.

1. I downloaded the git, cd FasterTransformer mkdir -p build cd build
2. cmake -DSM=70 -DCMAKE_BUILD_TYPE=Release -DBUILD_PYT=ON ..
3. make
4. pip install transformers==2.5.1
5. ./bin/encoder_gemm 32 32 12 64 0 0
6. python pytorch/encoder_sample.py 32 12 32 12 64 --time

And the error is from step 6.

[byshiue]

It seems that there are some issues for the fused mha initialization. Please try to use the v5.0_beta branch.

[michelleqyhqyh]

It seems that there are some issues for the fused mha initialization. Please try to use the v5.0_beta bran

I compiled dev_v5.0_beta, there is no the docment pytorch/ and file ./bin/encoder_gemm to generate. How can I use with the dev_v5.0_beta branch?

[michelleqyhqyh]

Besides, what does 'sm' mean?

[byshiue]

SM is the compute capacity of GPU. The SM of V100 is 70. You can find the SM of other GPUs on NVIDIA official website.

[JeremyZhao1998]

@michelleqyhqyh @byshiue Hi, I'm also running this file without using docker. I'm facing another error: It seems like error occurred in loading op. Have you met such error?

[byshiue]

This bug is fixed in latest release. It is caused by CUDA_API version.

[byshiue]

Close this bug because it is inactivated. Feel free to re-open this issue if you still have any problem.