llama : fix command-r inference when omitting outputs (ggerganov#6367)

hodlen · Apr 1, 2024 · 3b1c814 · 3b1c814
1 parent 8dc38b1
commit 3b1c814
Showing 1 changed file with 3 additions and 2 deletions.
diff --git a/llama.cpp b/llama.cpp
@@ -9152,8 +9152,9 @@ struct llm_build_context {
             if (il == n_layer - 1) {
                 // skip computing output for unused tokens
                 struct ggml_tensor * inp_out_ids = build_inp_out_ids();
-                cur  = ggml_get_rows(ctx0,  cur, inp_out_ids);
-                inpL = ggml_get_rows(ctx0, inpL, inp_out_ids);
+                cur     = ggml_get_rows(ctx0,     cur, inp_out_ids);
+                inpL    = ggml_get_rows(ctx0,    inpL, inp_out_ids);
+                ffn_inp = ggml_get_rows(ctx0, ffn_inp, inp_out_ids);
             }
 
             struct ggml_tensor * attn_out = cur;