graph: backend: compiler: fix updated llama mlp pattern

oneapi-src · Oct 17, 2023 · 595543d · 595543d
1 parent 4207105
commit 595543d
Showing 1 changed file with 5 additions and 1 deletion.
diff --git a/src/graph/backend/graph_compiler/patterns/mlp_pattern.hpp b/src/graph/backend/graph_compiler/patterns/mlp_pattern.hpp
@@ -353,8 +353,12 @@ void create_llama_mlp(const std::shared_ptr<pb_graph_t> &pgraph,
         auto quant1 = pgraph->append_op(graph::op_kind::Quantize,
                 {in_edge(0, extra_cast_after_mul, 0)});
         if (split_smooth_quant) {
+            auto extra_cast_before_mul_rhs
+                    = append_single_op_repetition_subgraph(
+                            pgraph, graph::op_kind::TypeCast, norm1);
             auto smooth_quant_mul1_rhs = append_single_op_repetition_subgraph(
-                    pgraph, graph::op_kind::Multiply, extra_cast_before_mul);
+                    pgraph, graph::op_kind::Multiply,
+                    extra_cast_before_mul_rhs);
             auto extra_cast_after_mul_rhs
                     = append_single_op_repetition_subgraph(pgraph,
                             graph::op_kind::TypeCast, smooth_quant_mul1_rhs, 0,