How to use an intermediate buffer during graph building ?

[phanichinimilli]

I'm trying to implement a state space model using GGML, I'm finding difficulty to induce usage of an intermediate buffer during graph building state. I'm depending on backend tensor allocation for both weights and graph compute.

My code logic looks like

build_graph () {
-
-
struct ggml_tensor *h_reals_arr = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 2048, 24);
-
struct ggml_tensor *h_reals_cur = ggml_view_1d(ctx, h_reals_arr, 2048, 0 * 2048 * sizeof(float));
ggml_cpy(ctx, data, h_reals_cur)
-
return ggml_mul_mat(ctx, h_reals_arr, C);
}
ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend));
ggml_gallocr_alloc_graph(allocr, gf);
ggml_backend_graph_compute(model.backend, gf);

After the compute,
-I'm seeing all zeros in tensor that include h_reals_arr/ h_reals_cur
-rest of the tensors I can see asper my intended operations applied.
Any help on this I would appreciate it.