context : fix reorder logic (#13267)

ggml-ci

diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index a88b9a5ff90da607fc52dcc509f8ca4cb736e9ab..45591be992d8788303fc2f164984040ce103fff3 100644 (file)
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -1050,6 +1050,9 @@ int llama_context::decode(llama_batch & inp_batch) {
     // finalize the batch processing
     kv_guard.commit();
 
+    // set to total number of outputs in the batch, for use in llama_get_logits_ith
+    n_outputs = n_outputs_all;
+
     // set output mappings
     {
         bool sorted_output = true;
@@ -1103,9 +1106,6 @@ int llama_context::decode(llama_batch & inp_batch) {
         }
     }
 
-    // set to total number of outputs in the batch, for use in llama_get_logits_ith
-    n_outputs = n_outputs_all;
-
     // wait for the computation to finish (automatically done when obtaining the model output)
     //synchronize();