StarSpace GPU working?

[loretoparisi]

StarSpace GPU working?

[loretoparisi]

@bluebyte9001 I will check it out, thanks.

[loretoparisi]

@bluebyte9001 I have checked your issue.
When using gpu/build.sh you came out with a cuda:8.0 supported ubuntu, once you have a gpu. At that point running nvdia-smi you should see some output like:

cd gpu/
$ ./run.sh 
# nvidia-smi
Tue Mar 28 22:25:50 2017       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 367.57                 Driver Version: 367.57                    |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GRID K520           Off  | 0000:00:03.0     Off |                  N/A |
| N/A   41C    P8    17W / 125W |      0MiB /  4036MiB |      0%      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID  Type  Process name                               Usage      |
|=============================================================================|
|  No running processes found                                                 |
+-

If so the gpu is connected via nvidia-docker. Hope this helps. Feel free to re-open if need more help!

[geirberge]

Thank you.

I run 384.111 version of nvndia driver, with cuda 8. I get the nvidia-smi to work as outlined by you,
but only after editing run.sh as follows (edits: added the --runtime=nvidia parameter, updated nvidia version):
docker run --runtime=nvidia --rm -it --device=/dev/nvidiactl --device=/dev/nvidia-uvm
--device=/dev/nvidia0 -v nvidia_driver_384.111:/usr/local/nvidia:ro -v ml:/home/user/ml:rw --name $IMAGE $IMAGE $CMD

The main issue is connected to Starspace's lack of utilization of the GPU. It only uses the CPU.
The GPU-util is not changing when running nvidia-smi while running Starspace.

[geirberge]

I thought Starspace was modified in your package to run on GPU. But I guess not? I cannot find so much information on Starspace, but fasText is a Facebook relative:
"Can we run fastText program on a GPU?
FastText only works on CPU for accessibility. That being said, fastText has been implemented in the caffe2 library which can be run on GPU."

[loretoparisi]

@bluebyte9001 yes I can confirm that fastText and Starspace only run on CPU in the default C++ implementation. I'm aware there were some attempts to run on GPU, but unofficial. Also consider that, while FastText does not use any BLAS library, Starspace is using the Boost library - see facebookresearch/StarSpace#49 about this. By the way, I have asked the FastText Facebook community here about the GPU and Caffe2, since I cannot see any reference to the FastText API in the current Caffe2 C++/Python API. Also I will re-open this, let's collect more info.

[geirberge]

Thanks a lot for checking out. Would have been nice to have either or both run on the gpu.