Should there be a matrix check in ggml_transpose?

[crnsh]

I apologize in advance for any mistakes in understanding I may have! Still very new to tensors and ML frameworks.

Based on my understanding, ggml_tranpose only applies to 2D tensors (matrices). For tensors of higher dimension, ggml_permute should be used.

• Is my understanding correct?
• If it is correct, shouldn't there be a check in ggml_tranpose for whether the input tensor is 2D or not? It seems to me that if it isn't 2D, then the ggml_tranpose function doesn't make any sense right?

Thanks!

[slaren]

ggml_tranpose swaps the first two dimensions of the tensor. That is usually fine even if the tensor has a higher dimension than 2, since the other dimensions are usually the batch size. So what you end with is a batch of transposed matrices, which is generally what you want. As you say, ggml_permute can be used for more advanced transposes that apply to any dimensions.