cuda runtime error (10)

[lixin4ever]

Dear all,

I have a issue when using the cutorch package

I install Torch7 with self-contained Lua on the redhat machine. No error occurs during the installation. Then I installed cutorch, and the output in the shell is below:

-- Installing: /home/liuxuebo/distro/install/include/THC/generic/THCTensorSort.cu
cd build
Updating manifest for /home/liuxuebo/distro/install/lib/luarocks/rocks
cutorch scm-1 is now built and installed in /home/liuxuebo/distro/install/ (license: BSD)

which seems normal. Then I try to import cutorch in th shell and the output is ok. However, when I ran a torch program using GPU, I met an error:

THCudaCheck FAIL file=/tmp/luarocks_cutorch-scm-1-8352/cutorch/init.c line=720 error=10 : invalid device ordinal
/home/liuxuebo/distro/install/bin/luajit: main.lua:282: cuda runtime error (10) : invalid device ordinal at /tmp/luarocks_cutorch-scm-1-8352/cutorch/init.c:720
stack traceback:
[C]: in function 'setDevice'
main.lua:282: in function 'main'
main.lua:354: in main chunk
[C]: in function 'dofile'
...ebo/distro/install/lib/luarocks/rocks/trepl/scm-1/bin/th:145: in main chunk
[C]: at 0x00405810

I wonder that do the settings of cuda or the way I install torch cause this problem?

(P.S., version of cuda is 7.0)

[soumith]

invalid device ordinal
means you are accessing a GPU device index that is not accessible to the program

[allanj]

How to solve this problem?

[liuxiaojun56]

I get the same problem. Any help on how to solve this problem ?

[shubhvachher]

I'm new to torch but the way I got this working is

$th
th> require 'cutorch'
th> cutorch.getDevice()

Then plugging the output from above into cutorch.setDevice in the code that isn't working. More experienced torch devs could point out if there is a better way