How to transpose a tensor and have the transpose reflected in contiguous memory with CUDA backend

[balisujohn]

I have a contiguous tensor of shape [18,64,16,4] If I print out the first and last 3 memory locations I get

0.00137997
0.00800999
0.0252284
0.18625
-0.075262
-0.00594251

I want to transpose this tensor and have the transpose reflected in contiguous memory. So I try the code:

KQV = ggml_transpose(ctx0, KQV);
KQV = ggml_cont_3d(ctx0,ggml_view_1d(ctx0, KQV, 64*18*16*4,0),64,18,64);

and I get the same results when printing out the first and last 3 values from memory:

0.00137997
0.00800999
0.0252284
0.18625
-0.075262
-0.00594251

How can I get a transpose that's actually reflected in contiguous memory for the first 2 dimensions of a 4d tensor? I should add that this is with a cuda backend, and ggml_cont_4d() doesn't work with 4d tensors for the cuda backend.

[balisujohn]

This strategy seems to work actually:

            KQV = ggml_reshape_3d(ctx0, KQV, 18,64,64);
            KQV = ggml_permute(ctx0, KQV, 1,0,2,3);
            KQV = ggml_cont_3d(ctx0, KQV, 64,18,64);
            KQV = ggml_reshape_4d(ctx0, KQV, 64,18,16,4);

[balisujohn]

(though im not getting exactly the transpose I want still even when I tried 1,0,2,3, but that's a separate issue)
updated the version above so it now works correctly!

[ggerganov]

KQV = ggml_cont(ggml_transpose(ctx0, KQV));

should also work