[CUDA] Defaulting to malloc in CHAllocator

[pr1nce-harry]

Hi all,

New to fitting with gpu, I'm running with device: "cuda" with "max_bin": 63

I see heaps of these warnings on initialization:
[LightGBM] [Warning] Defaulting to malloc in CHAllocator!!!

Is this expected, or is something wrong? Is there a dataset param that I'm not setting correctly?

Luke

[jameslamb]

Thanks for using LightGBM.

It's very difficult to help you based only on an error message. Could you give us some more information please?

• what version of LightGBM?
• operating system?
• how you installed LightGBM (e.g. conda vs. pip vs. building from source)?
• how you are using LightGBM (CLI? Python? something else?)?
• CUDA version and GPU type (output of nvidia-smi command)?

[pr1nce-harry]

Sorry here's more info:

LightGBM tag v4.6.0 built from source (sh ./build-python.sh install --cuda), running in python
Docker image based on nvidia/cuda:12.6.0-cudnn-devel-ubuntu24.04
nvidia-smi screenshot

[pr1nce-harry]

After a lot of time, and "[LightGBM] [Warning] Defaulting to malloc in CHAllocator!!!" messages
I get the following error:
lightgbm.basic.LightGBMError: Check failed: (split_indices_block_size_data_partition) > (0) at /src/LightGBM/lightgbm-python/src/treelearner/cuda/cuda_data_partition.cpp, line 280