[python-package] Using sparse features with CUDA is currently not supported.

[OrangeAoo]

Description

I got the following warnings when setting device_type='cuda'

 [Warning] Using sparse features with CUDA is currently not supported.
 [Fatal] CUDA Tree Learner was not enabled in this build.
Please recompile with CMake option -DUSE_CUDA=1

Reproducible example

## Train-test split 
# X_train.shape==(18840083, 3), y_train.shape==(18840083,1)
# X_val.shape==(2379462, 3), y_val.shape==(2379462, 1)

## parameters 
params = {
    'learning_rate': 0.01,
    'metric':'mse',
    'max_bin':63,
    'max_depth':15,
    'num_leaves':10000,
    'gpu_use_dp':True,
    'n_estimators':750,
    'tree_type':'data_parallel',
    'device_type':'cuda',
    'n_jobs':-1,
    'verbose':1
}

## Train 
start=time.time()
LGBmodel = lgb.LGBMRegressor(**params)
LGBmodel.fit(X_train,y_train,eval_set=[(X_val,y_val)])  
print(f"Training LGBoost completed in {time.time()-start:.4f} seconds!")

[jameslamb]

Thanks for using LightGBM.

That warning is expected, but the error means you haven't installed a CUDA-enabled build of LightGBM. To install a CUDA-enabled build of LightGBM, follow the instructions in the docs:

• pip: https://github.com/microsoft/LightGBM/blob/master/python-package/README.rst#build-cuda-version
• conda (preferred): https://github.com/microsoft/LightGBM/blob/master/python-package/README.rst#install-from-conda-forge-channel

[github-actions]

This issue has been automatically closed because it has been awaiting a response for too long. When you have time to to work with the maintainers to resolve this issue, please post a new comment and it will be re-opened. If the issue has been locked for editing by the time you return to it, please open a new issue and reference this one. Thank you for taking the time to improve LightGBM!

[Zaffer]

using pip install lightgbm --config-settings=cmake.define.USE_CUDA=ON does not fix this issue. It simply returns the same error. I tried purging the pip cache and using -no-cache-dir, but still does nothing.

[jameslamb]

@Zaffer if you can provide the complete logs from running that command, we can help you. I strongly suspect there's an error in those logs somewhere, or maybe a warning where pip is refusing to update your already-installed lightgbm.

[StrikerRUS]

@Zaffer Could you please try the following command?

pip install lightgbm --no-binary lightgbm --config-settings=cmake.define.USE_CUDA=ON

I guess instructions in the README document are incorrect. Without --no-binary flag pip doesn't try to build from sources but simply installs precompiled wheel file regardless specified --config-settings flags.

[jameslamb]

Oh wow, thanks @StrikerRUS ! I could swear that it used to be the case that if you passed --config-settings, pip would automatically try to download a source distribution and build from source.

But I tested that tonight with a couple of different versions of pip and couldn't reproduce it, so maybe I'm misremembering.

If there are any places in the docs that have pip install --config-settings ... without --no-binary lightgbm, I'd support a PR adding that --no-binary lightgbm.

[StrikerRUS]

@jameslamb
Already on the way: #6767.

[jameslamb]

Oh great, thanks!