Method build_pos_bias

subprojects/llama.cpp/src/llama-graph.cpp:1575–1588

}

ggml_tensor * llm_graph_context::build_pos_bias(ggml_tensor * pos_bucket, ggml_tensor * attn_rel_b) const {
    ggml_tensor * pos_bucket_1d = ggml_reshape_1d(ctx0, pos_bucket, pos_bucket->ne[0] * pos_bucket->ne[1]);
    cb(pos_bucket_1d, "pos_bucket_1d", -1);

    ggml_tensor * pos_bias = ggml_get_rows(ctx0, attn_rel_b, pos_bucket_1d);

    pos_bias = ggml_reshape_3d(ctx0, pos_bias, pos_bias->ne[0], pos_bucket->ne[0], pos_bucket->ne[1]);
    pos_bias = ggml_permute   (ctx0, pos_bias, 2, 0, 1, 3);
    pos_bias = ggml_cont      (ctx0, pos_bias);

    cb(pos_bias, "pos_bias", -1);

    return pos_bias;
}

ggml_tensor * llm_graph_context::build_attn_mha(
         ggml_tensor * q,