If I understand the interface correctly, the input x must be a tensor fully loaded on CPU. There are use cases (like the one described in this work) where both x and d are so large the inputs woudn't fit even in CPU memory. Would you consider adding support for inputs using an np.memmap() style interface to load chunks of the array dynamically?
Thank you for this great work!
If I understand the interface correctly, the input
xmust be a tensor fully loaded on CPU. There are use cases (like the one described in this work) where bothxanddare so large the inputs woudn't fit even in CPU memory. Would you consider adding support for inputs using annp.memmap()style interface to load chunks of the array dynamically?Thank you for this great work!