cutorch fails on getDeviceCount()

[sergei-sh]

Installed torch, cuda.

test.py

import lutorpy as lua
lua.require('cutorch')

Output:

THCudaCheck FAIL file=/home/serj/torch/extra/cutorch/lib/THC/THCGeneral.c line=66 error=30 : unknown error
Traceback (most recent call last):
  File "test.py", line 4, in <module>
    lua.require('cutorch')
  File "/home/serj/work/beautytorch/venv/local/lib/python2.7/site-packages/lutorpy/__init__.py", line 112, in require
    ret = luaRuntime.require(module_name)
  File "lutorpy/_lupa.pyx", line 318, in lutorpy._lupa.LuaRuntime.require (lutorpy/_lupa.c:6189)
  File "lutorpy/_lupa.pyx", line 1658, in lutorpy._lupa.call_lua (lutorpy/_lupa.c:25877)
  File "lutorpy/_lupa.pyx", line 1667, in lutorpy._lupa.execute_lua_call (lutorpy/_lupa.c:25987)
  File "lutorpy/_lupa.pyx", line 1620, in lutorpy._lupa.raise_lua_error (lutorpy/_lupa.c:25297)
lutorpy._lupa.LuaError: cuda runtime error (30) : unknown error at /home/serj/torch/extra/cutorch/lib/THC/THCGeneral.c:66

Related code:

{
  if (!state->cudaDeviceAllocator) {
    state->cudaDeviceAllocator = &defaultDeviceAllocator;
  }
  if (!state->cudaHostAllocator) {
    state->cudaHostAllocator = &THCudaHostAllocator;
  }
  if (!state->cudaUVAAllocator) {
    state->cudaUVAAllocator = &THCUVAAllocator;
  }

  int numDevices = 0;
  THCudaCheck(cudaGetDeviceCount(&numDevices)); // FAILS HERE!!!

Maybe the problem relates to the fact that my GPU is not NVidia, but cutorch is needed as a part of 3-rd party project which I'm trying to use

[albanD]

Please do not post the same question at multiple places. It is useless and creates a lot of noise.

cutorch/cunn only supports nvidia CUDA devices as explained here

[sergei-sh]

Thanks for your comment.

cutorch/cunn only supports nvidia CUDA devices as explained here

Sure. Can I avoid crash when importing as a part of a bigger project, sometimes running with non-supported devices?

[albanD]

Do not require it? You won't be able to use anything from it anyway.

[sergei-sh]

I would use other parts of the enclosing project. Do you mean it should fail in this situation by design? I suggest some other control path.

[sergei-sh]

Any workaround ?...

[albanD]

Make sure the library does not require cutorch, cunn or cudnn when gpu_mode is set to false.

[sergei-sh]

The problem is that I need those requirements. If I can do without them finally, there will be no issue (I don't like to spend your time).

This is the example code. It converts GPU model to CPU model so that I could run it without NVidia GPU.

Convert GPU model to CPU compatible version
Author: Eren Golge - erengolge@gmail.com

require 'torch'
require 'cudnn'
require 'nn'
require 'cunn'
require "models/dropresnet"

cmd = torch.CmdLine()
cmd:text()
cmd:text()
cmd:text('convert GPU model to CPU version')
cmd:text()
cmd:text('Options')
cmd:option('-loadPath','','Path to load GPU model.')
cmd:option('-savePath','', 'Path to save CPU model')
cmd:text()

params = cmd:parse(arg)

print(params.loadPath)
local model = torch.load(params.loadPath)
model_cpu = cudnn.convert(model, nn):float()
torch.save(params.savePath, model_cpu)

As you can see the task doesn't require the hardware (does it?). At the same time it will fail on torch.load() without cudnn and cunn.

/home/serj/torch/install/share/lua/5.1/torch/File.lua:343: unknown Torch class <cudnn.SpatialConvolution>

[albanD]

The thing is that if you have a GPU model (with modules built to run on GPU and parameters in CUDA tensors), you need cutorch and cunn to load them. Unfortunately there is no workaround: you need a CUDA device to load a torch GPU model.
The simplest thing to do is to ask whoever gave you the GPU model (or someone with a GPU) to run the conversion script locally and give you the CPU model directly.

[sergei-sh]

Thanks a lot, this is much helpful.