From: Xuan-Son Nguyen <redacted>
Date: Tue, 29 Apr 2025 06:45:49 +0000 (+0200)
Subject: llama-graph : fix text position for mrope (#13159)
X-Git-Tag: upstream/0.0.5318~102
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=b6ce7430b7eb51f032152316880204e0a9c0470e;p=pkg%2Fggml%2Fsources%2Fllama.cpp

llama-graph : fix text position for mrope (#13159)

* llama-graph : fix text position for mrope

* fix typo

* explicitly set 4th dim in the loop
---

diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index 2706ea26..fabb9ca2 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -55,13 +55,16 @@ void llm_graph_input_pos::set_input(const llama_ubatch * ubatch) {
     if (ubatch->pos && pos) {
         const int64_t n_tokens = ubatch->n_tokens;
 
-        if (ubatch->token && n_pos_per_embd > 1) {
+        if (ubatch->token && n_pos_per_embd == 4) {
             // in case we're using M-RoPE with text tokens, convert the 1D positions to 4D
-            // the other dimensions are all 0, they are unused for text tokens
-            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd, 0);
+            // the 3 first dims are the same, and 4th dim is all 0
+            std::vector<llama_pos> pos_data(n_tokens*n_pos_per_embd);
             // copy the first dimension
             for (int i = 0; i < n_tokens; ++i) {
-                pos_data[i] = ubatch->pos[i];
+                pos_data[               i] = ubatch->pos[i];
+                pos_data[    n_tokens + i] = ubatch->pos[i];
+                pos_data[2 * n_tokens + i] = ubatch->pos[i];
+                pos_data[3 * n_tokens + i] = 0; // 4th dim is 0
             }
             ggml_backend_tensor_set(pos, pos_data.data(), 0, pos_data.size()*ggml_element_size(pos));
         } else {