"RuntimeError: Expected all tensors to be on the same device" faced when running tutorial

[gkstmdwn]

I am using python 3.10.18. I installed pytorch with cuda 12.8, then installed Botorch.
For some reason, I faced the following error when running the first tutorial, "q-Noisy Constrained EI.ipynb".
when calling generate_initial_data(10) , the following error occured:

RuntimeError                              Traceback (most recent call last)
Cell In[7], [line 37](vscode-notebook-cell:?execution_count=7&line=37)
     [29](vscode-notebook-cell:?execution_count=7&line=29) best_observed_ei, best_observed_nei, best_random = [], [], []
     [31](vscode-notebook-cell:?execution_count=7&line=31) # call helper functions to generate initial training data and initialize model
     [32](vscode-notebook-cell:?execution_count=7&line=32) (
     [33](vscode-notebook-cell:?execution_count=7&line=33)     train_x_ei,
     [34](vscode-notebook-cell:?execution_count=7&line=34)     train_obj_ei,
     [35](vscode-notebook-cell:?execution_count=7&line=35)     train_con_ei,
     [36](vscode-notebook-cell:?execution_count=7&line=36)     best_observed_value_ei,
---> [37](vscode-notebook-cell:?execution_count=7&line=37) ) = generate_initial_data(n=10)
     [38](vscode-notebook-cell:?execution_count=7&line=38) mll_ei, model_ei = initialize_model(train_x_ei, train_obj_ei, train_con_ei)
     [40](vscode-notebook-cell:?execution_count=7&line=40) train_x_nei, train_obj_nei, train_con_nei = train_x_ei, train_obj_ei, train_con_ei

Cell In[4], [line 12](vscode-notebook-cell:?execution_count=4&line=12)
      [9](vscode-notebook-cell:?execution_count=4&line=9) def generate_initial_data(n=10):
     [10](vscode-notebook-cell:?execution_count=4&line=10)     # generate training data
     [11](vscode-notebook-cell:?execution_count=4&line=11)     train_x = torch.rand(10, 6, device=device, dtype=dtype)
---> [12](vscode-notebook-cell:?execution_count=4&line=12)     exact_obj = neg_hartmann6(train_x).unsqueeze(-1)  # add output dimension
     [13](vscode-notebook-cell:?execution_count=4&line=13)     exact_con = outcome_constraint(train_x).unsqueeze(-1)  # add output dimension
     [14](vscode-notebook-cell:?execution_count=4&line=14)     train_obj = exact_obj + NOISE_SE * torch.randn_like(exact_obj)

File c:\Users\sugar\AppData\Local\anaconda3\envs\HornDesignBO\lib\site-packages\torch\nn\modules\module.py:1751, in Module._wrapped_call_impl(self, *args, **kwargs)
   [1749](file:///C:/Users/sugar/AppData/Local/anaconda3/envs/HornDesignBO/lib/site-packages/torch/nn/modules/module.py:1749)     return self._compiled_call_impl(*args, **kwargs)  # type: ignore[misc]
   [1750](file:///C:/Users/sugar/AppData/Local/anaconda3/envs/HornDesignBO/lib/site-packages/torch/nn/modules/module.py:1750) else:
...
--> [107](file:///C:/Users/sugar/AppData/Local/anaconda3/envs/HornDesignBO/lib/site-packages/botorch/test_functions/base.py:107) if not ((X >= bounds[0]).all() and (X <= bounds[1]).all()):
    [108](file:///C:/Users/sugar/AppData/Local/anaconda3/envs/HornDesignBO/lib/site-packages/botorch/test_functions/base.py:108)     raise ValueError("Expected `X` to be within the bounds of the test problem.")
    [109](file:///C:/Users/sugar/AppData/Local/anaconda3/envs/HornDesignBO/lib/site-packages/botorch/test_functions/base.py:109) for inds in [discrete_inds, categorical_inds]:

RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu!

I fixed it by moving some tensors(local variable in function) to cuda device.
I find this was weird, and wanted to contribute by making a pull request, so i cloned the repository and set up a new environment, following the instructions in CONTRIBUTING.md

But then... in this new environment no error occurs and the code runs perfectly.

I was comparing both environment's pip list and found out that following CONTRIBUTING.md, pytorch will not install with cuda, and the device varaible will be set to 'cpu', therefore no errror will occur.
However, by installing pytorch with cuda, the tutorial code won't run, raising runtime error above.

I thought maybe this is a bug, and wanted to ask if this is my personal problem or it really is a bug.

[Balandat]

Hi @gkstmdwn thanks for flagging - yes, there is an issue here. For one, we should put the problem on the correct device:

neg_hartmann6 = Hartmann(negate=True).to(device=device)

I also found another issue where the random points were not generated on the GPU:

def update_random_observations(best_random):
    """Simulates a random policy by taking a the current list of best values observed randomly,
    drawing a new random point, observing its value, and updating the list.
    """
    rand_x = torch.rand(BATCH_SIZE, 6, device=device)   <-- note the `device` here
    next_random_best = weighted_obj(rand_x).max().item()
    best_random.append(max(best_random[-1], next_random_best))
    return best_random

I'll put up a PR with a fix.

[gkstmdwn]

Hi Max,

Thanks for such a quick accurate fix!
I think you can close this discussion now.

Have a nice day.

[Balandat]

For future reference, this was fixed in #2906