Not a name of any known instruction: "tanh"

[oreo-lp]

Description

Branch:main
GPU: Tesla T4
CUDA: 10.2
TensorRT: 8.4.0.6
Docker: No use

Reproduced Steps

When I build the BERT(C++) project, there has a error.
1. mkdir build && cd build
2. cmake -DSM=75 -DCMAKE_BUILD_TYPE=Release ..
3. make

The error is below:
"[19%] Building CUDA object src/fastertransformer/kernels/CmakeFiles/activation_kernels.dir/activation_kernels.cu.o
ptxas /tmp/tmpxft_0001d8f9_00000000-5_activation_kernels.ptx, line 95; error:Not a name of any known instruction: 'tanh' ..."

[byshiue]

Thank you for the feedback. This ptx instruction is only supported after CUDA 11. We have add CUDA version checking to fix this bug. Please try again.

[oreo-lp]

Thank you for the feedback. This ptx instruction is only supported after CUDA 11. We have add CUDA version checking to fix this bug. Please try again.

Thanks! Now I am back to use the fastertransformer-4.0 version, and build the project, but there has another error: "Could Not find MPI_CXX (missing: MPI_CXX_LIB_NAMES MPI_CXX_HEADER_DIR MPI_CXX_WORKS)" how can I install the mpi?

[byshiue]

We suggest to use latest main branch. We have support the compiling without mpi now.

I don't know how to install mpi. We assume this is installed in your environment.

[oreo-lp]

We suggest to use latest main branch. We have support the compiling without mpi now.

I don't know how to install mpi. We assume this is installed in your environment.

Thanks!

[oreo-lp]

Now, I have change cuda10.2 to cuda11.3-cudnn 7.6.5.However, during build the project, there also have a error: "FasterTransformer/src/fastertransformer/utils/conv2d.h:50:20: error: 'CUDNN_DATA_BFLOAT16' was not declared in this scops dataType = CUDNN_DATA_BFLOAT16".

[byshiue]

Your CUDNN is too old and it does not support bfloat16. You can remove https://github.com/NVIDIA/FasterTransformer/blob/main/CMakeLists.txt#L20 to disable bfloat16.

[oreo-lp]

Thanks, I have change cudnn7.6.5 to 8.3.2, but there also have a error: "./FasterTransformer/examples/cpp/xlnet/cpy.h:19:10: fatal error:zlib.h: No such file or directory #include <zlib.h>"

[byshiue]

You can install the package by

sudo apt-get install zlib1g-dev

We suggest using the docker image we use in document to prevent such environment setting issues.

Besides, you can also use

cmake --build . --target bert_example

to replace

make -j

to build the object you need, but not whole project.

[oreo-lp]

Thanks. I wanna to use Docker. However, for some reason, this server cannot install docker.

[oreo-lp]

You can install the package by

sudo apt-get install zlib1g-dev

We suggest using the docker image we use in document to prevent such environment setting issues.

Besides, you can also use

cmake --build . --target bert_example

to replace

make -j

to build the object you need, but not whole project.

Hi, I have build this project successfully. But there have a error during running bert_gemm: ./bin/bert_gemm 1 32 12 64 0 0 and the error is below: /lib64/libstdc++.so.6: version 'CXXABI_1.3.8' not found (required by ./bin/bert_gemm). My gcc version is gcc7.3.0. What gcc version does this project suggest? By the way, when I run strings /lib64/libstdc++.so.6 | grep CXXABI, it shows: "CXXABI_1.3 CXXABI_1.3.1 .... CXXABI_1.3.7"

[byshiue]

I am not familiar to these environment settings. So, I cannot provide many helps.

The gcc in our docker is 9.4. You can try it.

[oreo-lp]

I am not familiar to these environment settings. So, I cannot provide many helps.

The gcc in our docker is 9.4. You can try it.

Thanks, I changed the libstdc++.so.6.0.19 to libstdc++.so.6.0.26, it's OK. However, there have a error when I run : ./bin/bert_gemm 1 32 32 12 64 0 0, the error is below: terminate called agter throwing an instance of 'std::runtime_error' what(): [FT][ERROR] CUDA runtime error: initialization error /FasterTransformer/src/fastertransformer/models/bert/bert_gemm.cc: 54

[byshiue]

The error happens at

cudaMemGetInfo

It should be a error for cuda installation or driver.

[oreo-lp]

Thanks, I'll check it later.

[byshiue]

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