CUDA error when running local/chain/run_tdnn.sh (recipe for AMI corpus)

981 views
Skip to first unread message

Stefan-Adrian Toma

unread,
Mar 22, 2017, 12:18:21 PM3/22/17
to kaldi-help
I get this error when running run.sh for the AMI corpus. 

local/chain/run_tdnn.sh --mic ihm
This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
If you want to use GPUs (and have them), go to src/, and configure and make on a machine
where "nvcc" is installed.

In understand that Kaldi can also run on CPU only. I don't have access to GPUs but I have access to VMs with multiple cores.
How do i run the scripts using only the CPU?

I found the lines that check for CUDA in the beginning if the run_tdnn.sh script (see below). Can Kaldi use CPU automatically if no GPU is present? Will it work if I comment the lines checking for CUDA at the beginning of local/chain/run_tdnn.sh? 
(local/chain/run_tdnn.sh is a link to /local/chain/tuning/run_tdnn_1b.sh)

if ! cuda-compiled; then
  cat
<<EOF && exit 1
This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
If you want to use GPUs (and have them), go to src/, and configure and make on a machine
where "nvcc" is installed.
EOF
fi

Regards,
Stefan  

Daniel Povey

unread,
Mar 22, 2017, 1:26:06 PM3/22/17
to kaldi-help
You should comment those lines, and you may also have to add the option --use-gpu false to certain scripts that it calls (mostly the neural net training)- just see what scripts fail.

However, it will be extremely slow- like, a hundred times slower than using the GPU.
In nnet2 we used to have support for each job to use multiple threads on the CPU.  But the performance was so much worse than the GPUs that we mostly stopped using it, and in nnet3 that functionality was removed in order to keep the code simple.

Da

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
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+unsubscribe@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

Giorgos Giannios

unread,
Sep 7, 2021, 8:05:00 AM9/7/21
to kaldi-help
Hello everyone! I m new to kaldi so i want to ask a question about this issue. When i run a .sh file, i see the same message:

This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
If you want to use GPUs (and have them), go to src/, and configure and make on a machine
where "nvcc" is installed.


I have GPU on my laptop. More specifically, with this command i get the details about GPU:

lspci | grep VGA
, i get the following output:
00:02.0 VGA compatible controller: Intel Corporation Device 9b41 (rev 02)

while with this command: sudo lshw -C video
, i get the following output:
  *-display                 
       description: VGA compatible controller
       product: Intel Corporation
       vendor: Intel Corporation
       physical id: 2
       bus info: pci@0000:00:02.0
       version: 02
       width: 64 bits
       clock: 33MHz
       capabilities: pciexpress msi pm vga_controller bus_master cap_list rom
       configuration: driver=i915 latency=0
       resources: irq:147 memory:70000000-70ffffff memory:60000000-6fffffff ioport:3000(size=64) memory:c0000-dffff

My question is, how can i configure and make on a machine where "nvcc" is installed (using this GPU), according to terminal's suggestion

Thanks !
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.

Daniel Povey

unread,
Sep 7, 2021, 8:18:29 AM9/7/21
to kaldi-help
By GPUs, we mean NVidia GPU cards.

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.
Reply all
Reply to author
Forward
0 new messages