You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I am using the sparse variational GPyTorch framework to perform 7500 tasks. I have 4800 data points, and I am using batch sizes (so both the inout and output matrices have dimension (4800,7500). Even with a btach size of 1 I get an memory allocation error. Where as, as I am using the variational framework, this should not be an issue. Also I was not having this issue with the RBF kernel, even when I was not using batches and I had the full dataset on the GPU.
CUDA out of memory. Tried to allocate 172.00 MiB (GPU 5; 47.54 GiB total capacity; 5.74 GiB already allocated; 105.56 MiB free; 5.76 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
// Paste the bad output here!
I do not have this memory allocation issue with other kernels, only with the periodic kernel. I think there is a bug somewhere, since with just a small batch size and with the variational framework, there should a priori not be any memory issue.
- <!-- GPyTorch Version (run `print(gpytorch.__version__)` -->1.11
- <!-- PyTorch Version (run `print(torch.__version__)` -->1.13.1
- <!-- Computer OS -->Linux
The text was updated successfully, but these errors were encountered:
The periodic kernel makes an ... x d x m x n tensor (the ... are for batch dimensions, d is for dimensions of the input, m and n are for the number of points) so that might be causing the memory error. For the RBF kernel you get ... x m x n tensor only. Maybe try with less data and check what size of the covariance matrix you are getting?
馃悰 Bug
I am using the sparse variational GPyTorch framework to perform 7500 tasks. I have 4800 data points, and I am using batch sizes (so both the inout and output matrices have dimension (4800,7500). Even with a btach size of 1 I get an memory allocation error. Where as, as I am using the variational framework, this should not be an issue. Also I was not having this issue with the RBF kernel, even when I was not using batches and I had the full dataset on the GPU.CUDA out of memory. Tried to allocate 172.00 MiB (GPU 5; 47.54 GiB total capacity; 5.74 GiB already allocated; 105.56 MiB free; 5.76 GiB reserved in total by PyTorch) If reserved memory is >> allocated memory try setting max_split_size_mb to avoid fragmentation. See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
// Paste the bad output here!
The text was updated successfully, but these errors were encountered: