Facing no kernel image is available for execution on the device" returned from 'cudaGetLastError()

[Nayan JHA]

Hi all,

I am facing the error " no kernel image is available for execution on the device" returned from 'cudaGetLastError()" while trying to execute the TDNN training recipe .

I am using RTX 3080 GPU which has cuda capability : 8.6  , along with the latest version of kaldi .

I have searched for solutions in the forum and have implemented the following two solutions but they didn't work for me .

Solution 1 Implemented : https://groups.google.com/g/kaldi-help/c/yKwZGRBF6cg/m/zgxVLiY4CQAJ

Solution  2 Implemented : https://groups.google.com/g/kaldi-help/c/Fs6s3v3404w/m/8sThqr_xAQAJ

I am attaching the screenshot for the configs as well as the files and logs associated.

Crux of solution Implemented by me :

cd kaldi/src

Add  " -gencode arch=compute_86,code=sm_86 " in Kaldi.mk file for the CUDA_ARCH option

make clean

make depend -j 8
make -j 8

Kindly help me resolve this error.

Looking forward to hearing from you all !

Thanks & Regards,

Nayan

[Jan Yenda Trmal]

I think you need cuda 11 for the 30xx cards but I'm not super sure. Try to google around if those cards are supported in cuda 10

--
Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/e23c84a0-b7a8-46c9-8836-3ff94af9ade3n%40googlegroups.com.

[Nayan JHA]

Hi Yenda,

It finally worked for me , I had to change the cuda version to 11.0 from 11.4 . Turns out 30XX cards don't support  cuda version below 11.0 .

I had to  force mount the cuda during the configure stage ( ./configure  --cuda-dir= path to cuda dir ).

Thanks for all the help and support !

Regards,

Nayan

[Kirill 'kkm' Katsnelson]

Always compile Kaldi for the arch that you have. configure is very generic. Your solution in almost correct! :-) A fully correct one is to run ./configure --help, read it, and set your compute and code arch using a configure switch. I more often than not do that. Can't remember the switch name, sorry. CUDA code will also comple much faster if you select a single arch.

BTW, modifying kaldi.mk by hand is OK. We can't configure some complex scenarios automatically.

 -kkm

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/4422093f-b995-41a8-adb0-9094ae065b4en%40googlegroups.com.