package cutorch, cunn not found about char-rnn HOT 16 CLOSED

sorry about this, I was just hacking on this part and may have broken the build somehow. I'll fix this asap, I don't have access to a GPU machine to figure out what's going on at this moment. Hopefully within few hours.

from char-rnn.

No worries, I can take a peek at it myself. Do you have a file name I can look at or a commit hash I could possibly revert?

from char-rnn.

I was trying to be clever with a few of the most recent commits. The issue is where you see them imported, local ok,cunn = pcall..., maybe they don't work as I thought they did. I'll look into this in a bit

from char-rnn.

I think I got it. There shouldn't be local for the pcall lines. Taking it out works I think.
EDIT: nvm, it's all confusing. I don't think this should be a fix.

from char-rnn.

I just took them out, doesn't appear to be working. I'll tinker around with it a bit to see if it's specific to my implementation.

from char-rnn.

hey, I finally had a chance to check this and current code works without issues on my GPU machine. It could be something with your configuration.

from char-rnn.

Roger that. I'll mark it closed and start checking paths and global/env variables. Thanks for getting to this so quickly!

from char-rnn.

Got it...was in fact implementation specific. Forcing the require without pcall gave me a stacktrace with a CUDA runtime error.

from char-rnn.

Interesting. Is the stacktrace helpful? Let me know if you resolve this.

from char-rnn.

Yeah, it was an error about insufficient versions. I forced installing Cuda 5.5 as I saw it in one of the libraries (Torch, maybe?) but I'm going to see if Cuda 7 will work, unless you've got a better suggestion.

from char-rnn.

Can you post that stacktrace here? Cuda 5.5 to 7.0 should work comfortably.

Your gpu needs to be atleast CUDA 2.0 capable.

On Sunday, June 14, 2015, Bret Mattingly II [email protected]
wrote:

Yeah, it was an error about insufficient versions. I forced installing
Cuda 5.5 as I saw it in one of the libraries (Torch, maybe?) but I'm going
to see if Cuda 7 will work, unless you've got a better suggestion.

—
Reply to this email directly or view it on GitHub
#37 (comment).

from char-rnn.

Yeah, it'll be an hour or so, right now I'm getting an Ubuntu distro on another partition to see if I can get this running sooner.

from char-rnn.

For me, removing the pcall for the require changes the error to 'no CUDA-capable device', so what I thought was a package error for the past several hours actually wasn't.

So now I'm back to figuring out what's up with my CUDA installation...
Edit:
So CUDA hadn't actually installed that time. But it looks like no matter how I install it, it breaks my ability to actually drive my display :(

from char-rnn.

@eternaldensity What OS and GPU are you using?

I was able to get this working on Ubuntu 12.04 LTS without issue. A lack of full headers (version.h) in Kali makes CUDA installation difficult. I've submitted a PR which generates a little more feedback, since pcall will return false on ANY require failure, not just missing packages.

from char-rnn.

@bomattin
Ubuntu 17.04 LTS
I've tried Cuda 7, and then 6.5 when I found that was recommended for my GPU (GTS 250M). I installed via the runfile method, first by installing everything, and a second attempt extracting the driver installer so I could run it with --no-opengl-files which didn't help.

Actually the attempt with Cuda 7 was the most successful because that time I could still get to the Ubuntu login screen (it had the wrong resolution, and attempts to login resulted in a screen flicker and threw me back to the login screen) while with Cuda 6.5 it never succeeded in reaching full graphics mode at all.

Since the Cuda 7 attempt I've upgraded my Nvidia drivers from 331 to 340 (recommended for my GPU on the Nvidia site) so maybe another attempt will be more successful. Tomorrow.

from char-rnn.

Hi, I'm using Ubuntu 14.04 with CUDA 7.0.27 and solved the same problem by upgrading cmake.
Additional details below.

First, I successfully installed the following with no errors:

$ luarocks install cutorch
$ luarocks install cunn

However, Torch complained about not being able to find them, and fell back to CPU:

package cunn not found!
package cutorch not found!

So I opened a Torch REPL and just tried loading require 'cutorch':

th> require 'cutorch'
/home/epark/torch/install/share/lua/5.1/trepl/init.lua:363: cuda runtime error (30) : unknown error at /tmp/user/1000/luarocks_cutorch-scm-1-1963/cutorch/lib/THC/THCGeneral.c:16
stack traceback:
        [C]: in function 'error'
        /home/epark/torch/install/share/lua/5.1/trepl/init.lua:363: in function 'require'
        [string "_RESULT={require 'cutorch'}"]:1: in main chunk
        [C]: in function 'xpcall'
        /home/epark/torch/install/share/lua/5.1/trepl/init.lua:630: in function 'repl'
        ...park/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:185: in main chunk
        [C]: at 0x00406670

Googled the traceback and found an issue at the cutorch repo saying that I'd have to upgrade my cmake (torch/cutorch#126):

$ cmake --version
cmake version 2.8.12.2

I had to use a PPA to upgrade my cmake to version 3.2:

$ sudo apt-get install software-properties-common
$ sudo add-apt-repository ppa:george-edison55/cmake-3.x
$ sudo apt-get update
$ sudo apt-get install cmake
$ cmake --version
cmake version 3.2.2

CMake suite maintained and supported by Kitware (kitware.com/cmake).

Then I reinstalled cutorch and cunn according to torch/cutorch#210 (In case this doesn't work on the first try, check if you have the permissions for ~/.cache/luarocks):

$ luarocks install cutorch
$ luarocks install cunn

Hope that helps anyone else who encounters the same problem.

p.s. I also saw someone say that the cuda runtime error (30) I got in the first place may be due to sudoing. (nagadomi/waifu2x#22 (comment))

from char-rnn.