how can i do equivalent of torch.cat in ggml tensor ?

[manmay-nakhashi]

i want to cocatenate two tensor with dim,1024 how can i do that ?

[slaren]

I don't think there is an op to do this directly, but you could get the same effect by creating two views with different offsets and copying each part separately:

// concat a,b into dst
dst = ggml_new_tensor_1d(ctx, a->type, a->ne[0] + b->ne[0]);
dst_lo = ggml_view_1d(ctx, dst, a->ne[0], 0);
dst_hi = ggml_view_1d(ctx, dst, b->ne[0], ggml_element_size(dst) * a->ne[0]);
ggml_cpy(ctx, a, dst_lo);
ggml_cpy(ctx, b, dst_hi);

[manmay-nakhashi]

and how can i do the same in 2d ?

[manmay-nakhashi]

        struct ggml_tensor * condL = ggml_new_tensor_2d(ctx0, cond->type, 1024, cond->ne[1] + textL->ne[1]);
        struct ggml_tensor * dst_lo = ggml_view_2d(ctx0, condL, 1024, cond->ne[1], cond->nb[1],0);
        struct ggml_tensor * dst_hi = ggml_view_2d(ctx0, condL, 1024, textL->ne[1], textL->nb[1], ggml_element_size(condL) * cond->ne[1]);
        ggml_cpy(ctx0, cond, dst_lo);
        ggml_cpy(ctx0, textL, dst_hi);

[manmay-nakhashi]

@slaren is this a right implementation ?

[slaren]

For dst_hi I think you want to set the offset to skip cond->ne[1] rows, but currently you are only skipping cond->ne[1] elements. This should work generally:

struct ggml_tensor * dst_hi = ggml_view_2d(ctx0, condL, 1024, textL->ne[1], textL->nb[1], condL->nb[1] * cond->ne[1]);