GPSampler crashes when torch default device is cuda #6113
Description
Expected behavior
Successful run to completion.
Environment
- Optuna version:4.3.0
- Python version:3.11.4
- OS:Windows-10-10.0.26100-SP0
- Torch version:2.6.0+cu124
Error messages, stack traces, or logs
...
File "C:\Users\Argusdusty\AppData\Local\Programs\Python\Python311\Lib\site-packages\optuna\samplers\_gp\sampler.py", line 230, in sample_relative
kernel_params = gp.fit_kernel_params(
^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Argusdusty\AppData\Local\Programs\Python\Python311\Lib\site-packages\optuna\_gp\gp.py", line 264, in fit_kernel_params
return _fit_kernel_params(
^^^^^^^^^^^^^^^^^^^
File "C:\Users\Argusdusty\AppData\Local\Programs\Python\Python311\Lib\site-packages\optuna\_gp\gp.py", line 183, in _fit_kernel_params
np.log(initial_kernel_params.inverse_squared_lengthscales.detach().numpy()),
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\Users\Argusdusty\AppData\Local\Programs\Python\Python311\Lib\site-packages\torch\utils\_device.py", line 104, in __torch_function__
return func(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^
TypeError: can't convert cuda:0 device type tensor to numpy. Use Tensor.cpu() to copy the tensor to host memory first.Steps to reproduce
import torch; torch.set_default_device("cuda")with a CUDA-enabled PyTorch.- Create study with
sampler=optuna.samplers.GPSampler() study.optimize(...)
import optuna
import torch
torch.set_default_device("cuda")
def objective(trial):
x = trial.suggest_float("x", -100, 100)
return x ** 2
study = optuna.create_study(sampler=optuna.samplers.GPSampler())
study.optimize(objective, n_trials=100)Additional context (optional)
A successful patch in my local workspace has been to modify all calls to torch.tensor, torch.ones, torch.empty, torch.eye, etc. (calls that create new tensors), under optuna/_gp/, to add device='cpu'