nvcc fatal : Unsupported gpu architecture 'compute_20'
dllen opened this issue · 10 comments
dllen commented
when install throws this error
nvcc fatal : Unsupported gpu architecture 'compute_20' make[2]: *** [cu-kernels.o] Error 1 make[2]: *** Waiting for unfinished jobs.... ar -cru kaldi-tree.a event-map.o context-dep.o clusterable-classes.o cluster-utils.o build-tree-utils.o build-tree.o build-tree-questions.o tree-renderer.o
galv commented
Are you using CUDA 9.0?
Fermi (aka 2.x) was deprecated in CUDA 8.0 and removed in CUDA 9.0. Kaldi
hasn't been fully updated for CUDA 9.0 quite yet. It should be a simple
fix, though.
…On Thu, Sep 28, 2017 at 12:03 AM, dllen ***@***.***> wrote:
when install throws this error
nvcc fatal : Unsupported gpu architecture 'compute_20' make[2]: ***
[cu-kernels.o] Error 1 make[2]: *** Waiting for unfinished jobs.... ar -cru
kaldi-tree.a event-map.o context-dep.o clusterable-classes.o
cluster-utils.o build-tree-utils.o build-tree.o build-tree-questions.o
tree-renderer.o
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#1918>, or mute the thread
<https://github.com/notifications/unsubscribe-auth/AEi_UKIkDkGMXc1yTZ8ukGJHJ_bqAtWqks5sm0StgaJpZM4Pmz7D>
.
--
Daniel Galvez
dllen commented
Sorry, would you provide more information how can I fix it in detail? Thanks so much.
galv commented
Oh, sorry, there is definitely an issue in master right now.
First get the latest master branch. I assume you know how to do this.
Delete "-gencode arch=compute_20,code=sm_20 "
in this line:
https://github.com/kaldi-asr/kaldi/blob/master/src/configure#L431
The ./configure && make depend && make && make test
Let us know if that fixes it. I could check this myself, but I've been
crunched for time lately and don't want to get CUDA 9 onto one of my dev
machines right this instant.
…On Thu, Sep 28, 2017 at 12:26 AM, dllen ***@***.***> wrote:
Sorry, would you provide more information how can I fix it in detail?
Thanks so much.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#1918 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AEi_UEN0Ui4mxMcoIBFanwMvVYr6HrANks5sm0oPgaJpZM4Pmz7D>
.
--
Daniel Galvez
jtrmal commented
Give me a couple of minutes, I'll fix it in master
y.
On Thu, Sep 28, 2017 at 10:23 AM, Daniel Galvez <notifications@github.com>
wrote:
… Oh, sorry, there is definitely an issue in master right now.
First get the latest master branch. I assume you know how to do this.
Delete "-gencode arch=compute_20,code=sm_20 "
in this line:
https://github.com/kaldi-asr/kaldi/blob/master/src/configure#L431
The ./configure && make depend && make && make test
Let us know if that fixes it. I could check this myself, but I've been
crunched for time lately and don't want to get CUDA 9 onto one of my dev
machines right this instant.
On Thu, Sep 28, 2017 at 12:26 AM, dllen ***@***.***> wrote:
> Sorry, would you provide more information how can I fix it in detail?
> Thanks so much.
>
> —
> You are receiving this because you commented.
> Reply to this email directly, view it on GitHub
> <#1918 (comment)>,
> or mute the thread
> <https://github.com/notifications/unsubscribe-auth/AEi_
UEN0Ui4mxMcoIBFanwMvVYr6HrANks5sm0oPgaJpZM4Pmz7D>
> .
>
--
Daniel Galvez
—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
<#1918 (comment)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/AKisX5d0tXvULHGP0KnmMtImhKdxm5Poks5sm6vGgaJpZM4Pmz7D>
.
jtrmal commented
It should be fixed in the master.
y.
…On Thu, Sep 28, 2017 at 10:24 AM, Jan Trmal ***@***.***> wrote:
Give me a couple of minutes, I'll fix it in master
y.
On Thu, Sep 28, 2017 at 10:23 AM, Daniel Galvez ***@***.***>
wrote:
> Oh, sorry, there is definitely an issue in master right now.
>
> First get the latest master branch. I assume you know how to do this.
>
> Delete "-gencode arch=compute_20,code=sm_20 "
>
> in this line:
>
> https://github.com/kaldi-asr/kaldi/blob/master/src/configure#L431
>
> The ./configure && make depend && make && make test
>
> Let us know if that fixes it. I could check this myself, but I've been
> crunched for time lately and don't want to get CUDA 9 onto one of my dev
> machines right this instant.
>
> On Thu, Sep 28, 2017 at 12:26 AM, dllen ***@***.***> wrote:
>
> > Sorry, would you provide more information how can I fix it in detail?
> > Thanks so much.
> >
> > —
> > You are receiving this because you commented.
> > Reply to this email directly, view it on GitHub
> > <#1918 (comment)
> >,
> > or mute the thread
> > <https://github.com/notifications/unsubscribe-auth/AEi_UEN0U
> i4mxMcoIBFanwMvVYr6HrANks5sm0oPgaJpZM4Pmz7D>
> > .
> >
>
>
>
> --
> Daniel Galvez
>
> —
> You are receiving this because you are subscribed to this thread.
> Reply to this email directly, view it on GitHub
> <#1918 (comment)>,
> or mute the thread
> <https://github.com/notifications/unsubscribe-auth/AKisX5d0tXvULHGP0KnmMtImhKdxm5Poks5sm6vGgaJpZM4Pmz7D>
> .
>
silvacarloss commented
If the problem persists after removing the -gencode arch=compute_20,code=sm_20 line, try to erase the following file: /caffe/build/CMakeCache.txt and run cmake .. again. It worked for me.
CodesFarmer commented
U can remove the line -gencode arch=compute_20,code=sm_20 and -gencode arch=compute_20,code=sm_21 in Makefile.config.
danpovey commented
I just pushed a change to remove compute_20 from the nvcc flags.
serggehu commented
BorisGudnov commented
Deleting "-gencode arch=compute_20,code=sm_20 " doesn't work if you're using cmake to compile, as Makefile.config is not used (it's only used if you compile using make). The solution is to edit caffe/Cuda.cmake. On line 9 simply delete 20 21(20) from the list of GPU architectures.