Crashes when running on GPU (cuda11.3)

[Sanaxen]

In the sample code it occurs in the following places

result <- luz::lr_finder(
  model, 
  transform(spec, train), 
  end_lr = 1,
  dataloader_options = list(
    batch_size = 64
  ),
  verbose = FALSE
)

and

fitted <- model %>% 
  fit(
    transform(spec),
    valid_data = transform(spec, new_data = valid),
    epochs = 100,
    callbacks = list(
      luz::luz_callback_keep_best_model(monitor = "valid_loss"),
      luz::luz_callback_early_stopping(
        monitor = "valid_loss", 
        patience = 5, 
        min_delta = 0.001
      )
    ),
    verbose = FALSE,
    dataloader_options = list(batch_size = 64, num_workers = 4)
  )

However, the CPU version works correctly. I have been able to install torch and tft without any trouble.

[Sanaxen]

A similar problem already existed. Rstudio crashes when fitting using luz #861 Rstudio crashes when fitting using luz #102

One thing I did find out.

On a PC with two GPUs it crashes, but on a PC with one GPU it works fine and exits with the correct result.

Does luz or torch support multi-GPU at all? If it is mentioned anywhere, I'm sure I'm missing it. Sorry if that is the case. Is there any way to specify which GPU to run on?

[Sanaxen]

fitted <- model %>% 
  fit(
    transform(spec),
    epochs = n_epochs,
    verbose = T,
    valid_data = transform(spec, new_data = valid),
    callbacks = list(
      luz::luz_callback_keep_best_model(monitor = "valid_loss"),
      luz::luz_callback_early_stopping(
        monitor = "valid_loss", 
        patience = 5, 
        min_delta = 0.001
      ),
    ),
    accelerator = tluz::accelerator( 
        device_placement = TRUE,
        cpu = TRUE,
        cuda_index = 0
    ) ,
    dataloader_options = list( batch_size = n_batch_size, num_workers = n_num_workers)
  )

We have found that explicitly specifying accelerator = tluz::accelerator allows for normal execution.