graphs

Author: pwilkin

src/llama-graph.h

@@ -628,6 +628,10 @@ struct llm_graph_params {
         return
             cparams.embeddings  == other.cparams.embeddings  &&
             cparams.causal_attn == other.cparams.causal_attn &&
+            // mtp_step selects which MTP block the graph references; reusing a
+            // graph built for step k with step k' != k would silently feed the
+            // wrong block's weights.
+            cparams.mtp_step    == other.cparams.mtp_step    &&
             arch  == other.arch  &&
             gtype == other.gtype &&
             cvec  == other.cvec  &&