spcl / substation Goto Github PK

please delete

Hi,

I am trying to execute the Substation encoder layer provided in the pytorch_module folder. According to the readme I am trying to execute the following command

python run_encoder.py --batch-size 8 --max-seq-len 512 --num-heads 16 --emb-size 1024 --layout layouts-bert-b8-l512-h16-e1024.pickle

but this results in an error message:

OSError: ./compilation_dir/26e5dc71cb611d368b14bab45256feed/./encoder.so: cannot open shared object file: No such file or directory

I have then tried to run make encoder to hopefully build the encoder.so library. This unfortunately fails as well:

encoder.cu:30:10: fatal error: encoder_parameters.cuh: No such file or directory

even when I set CPATH to include the compilation_dir folder:

nvcc --compiler-options -Wfatal-errors -O3 -o encoder encoder.cu -gencode arch=compute_61,code=sm_61 -gencode arch=compute_70,code=sm_70 -lcublas
/usr/bin/ld: /usr/lib/gcc/x86_64-linux-gnu/9/../../../x86_64-linux-gnu/Scrt1.o: in function _start':
(.text+0x24): undefined reference to main'

Could you help me get this running or point me to the corresponding part of the readme on how to fix this?
Many thanks in advance!

I run the 'tc_profiling/compile.sh', show error:

cublas-gemm.cpp: In function 'int main(int argc, char **argv)':
cublas-gemm.cpp: 26:10: error: 'err' does not name a type
               auto err = (expr);  
cublas-gemm.cpp: 37:3: note: in expansion of macro ‘CUDA_CHECK’
               CUDA_CHECK(cublasCreate(&handle));x
....
 

GPU type: V100
cuda version:10.1

What is your compilation environment?

Hi, Thanks for your code.

When I using

sh compile.sh

I fonud bug follows:

compile.sh: line 4: syntax error near unexpected token `('

compile.sh: line 4: for i (seq 0 23`) { echo $i; nvcc -O3 -lcudadevrt -lcuda -arch sm_70 -lcublas cublas-gemm.cpp -

DCUBLAS_ALGO=CUBLAS_GEMM_ALGO$i -o gemm$i }'

Can you give me some advices?

Thanks:)