Lightgbm does not work on GPU?

[code1704]

Description

It's really hard to install a gpu version.

"Please recompile with CMake option -DUSE_CUDAP=1"

??? What is CUDAP?

cmake -DUSE_CUDAP=1 ..
but still got the same error.

Reproducible example

import lightgbm as lgb
import numpy as np

X = np.random.rand(500, 10) 
y = np.random.randint(2, size=500)
dataset = lgb.Dataset(X, label=y)
# params = {'device': 'gpu'}
params = {'device': 'cuda'}
gbm = lgb.train(params, dataset)

Environment info

centos 7 (in docker)

LightGBM version or commit hash:
master branch (4.1?)

Command(s) you used to install LightGBM

cmake -DUSE_CUDAP=1 ..
sh ./build-python.sh install --precompile
# the command above upgrades numpy and lightgbm will fails to import (numpy.float error). why do you upgrade that then?
pip3 install numpy==1.20

Additional Comments

[jameslamb]

LightGBM does not work on GPU

That is absolutely not true. There are 2 GPU-accelerated builds of LightGBM:

• "device": "gpu" (-DUSE_GPU=1) = OpenCL-based implementation targeting a wide range of GPUs from multiple manufacturers
• "device": "cuda" (-DUSE_CUDA=1) = CUDA kernel targeting NVIDIA GPUs

To install the Python package using these builds, see the guide in https://github.com/microsoft/LightGBM/blob/master/python-package/README.rst.

What is CUDAP?

Thanks for pointing this out! That's a typo.

I found that error message using GitHub search: https://github.com/search?q=repo%3Amicrosoft%2FLightGBM+%22Please+recompile+with+CMake+option+-DUSE_CUDAP%3D1%22&type=code

LightGBM/src/treelearner/cuda/cuda_single_gpu_tree_learner.hpp

Line 158 in 18dbd65

"Please recompile with CMake option -DUSE_CUDAP=1");

It should be -DUSE_CUDA (no P). We'd be grateful if you'd open a pull request fixing that. Are you interested in contributing?

[code1704]

Thanks I will try USE_CUDA later.

I just compiled with opencl. It tells "LightGBMError: No OpenCL device found", which clinfo founds nvidia cuda.

clinfo
Number of platforms                               1
  Platform Name                                   NVIDIA CUDA
  Platform Vendor                                 NVIDIA Corporation
  Platform Version                                OpenCL 3.0 CUDA 11.8.87
  Platform Profile                                FULL_PROFILE
  Platform Extensions                             cl_khr_global_int32_base_atomics cl_khr_global_int32_extended_atomics cl_khr_local_int32_base_atomics cl_khr_local_int32_extended_atomics cl_khr_fp64 cl_khr_3d_image_writes cl_khr_byte_addressable_store cl_khr_icd cl_khr_gl_sharing cl_nv_compiler_options cl_nv_device_attribute_query cl_nv_pragma_unroll cl_nv_copy_opts cl_nv_create_buffer cl_khr_int64_base_atomics cl_khr_int64_extended_atomics cl_khr_device_uuid cl_khr_pci_bus_info cl_khr_external_semaphore cl_khr_external_memory cl_khr_external_semaphore_opaque_fd cl_khr_external_memory_opaque_fd
  Platform Host timer resolution                  0ns
  Platform Extensions function suffix             NV

  Platform Name                                   NVIDIA CUDA
Number of devices                                 1
  Device Name                                     NVIDIA GeForce GTX 1080 Ti

[jameslamb]

No OpenCL device found

Please see the advice in #5914 (comment).

[code1704]

Thanks, this works. Will use USE_CUDA much faster than USE_GPU for nvidia GPU?

And this is another typo DUSE_GPU :)

[jameslamb]

another typo

ha sorry about that! Thanks, just edited

USE_CUDA much faster than USE_GPU for nvidia GPU?

Yes. The "device": "cuda" build should be faster, as:

1. it handles more operations on the GPU (i.e. not falling back to CPU) than the "device": "gpu" version
2. it was written much more recently, and takes advantage of newer versions of CUDA which have performance enhancements for newer NVIDIA GPUs

If you have an NVIDIA GPU, use the USE_CUDA version.

this works

great, we'll close this issue

[code1704]

great thanks

[code1704]

Can lightgbm use multi gpu cards? I have 4 cards, seems it uses only one.

[jmoralez]

Not yet, that's being tracked in #5076.