Added notes about what is actually stored in batch's pos

Author: rujialiu

src/llama-graph.cpp

@@ -55,6 +55,10 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
             ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
         } else {
             llama_pos * pos_ptr = ubatch->pos;
+            // Normally, ubatch->pos stores linearly increasing position
+            // However, some multi-modal models requires special position embedding (e.g. M-Rope in qwen2vl and qwen2.5vl)
+            // But linearly increasing position is still needed for proper causal attention masking
+            // So we store both of them: the first n_tokens elements are not changed, while model-specific positions are appended after that.
             if (ubatch->embd && n_pos_per_embd > 1) pos_ptr += n_tokens; // use mrope positions
             ggml_backend_tensor_set(pos, pos_ptr, 0, n_tokens * n_pos_per_embd * ggml_element_size(pos));
         }

tools/mtmd/mtmd-helper.cpp

@@ -55,6 +55,11 @@ llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks) {
 
 // helper struct to make working with embd batch easier
 // note: this will be removed after llama_batch_ext refactoring
+// notes2: Normally, batch's `pos` stores linearly increasing position
+// However, some multi-modal models requires special position embedding (e.g. M-Rope in qwen2vl and qwen2.5vl)
+// But linearly increasing position is still needed for proper causal attention masking
+// So we store both of them: the first n_tokens elements are not changed, while model-specific positions are appended after that.
+// So `pos` has `n_tokens * (n_pos_per_embd + 1)` elements
 struct decode_embd_batch {
     int n_pos_per_embd;
     int n_mmproj_embd;