mulmat type conversion update

ggml/src/ggml-openvino/openvino/op/mulmat.cpp

@@ -12,6 +12,7 @@
 #include <openvino/op/slice.hpp>
 #include <openvino/op/transpose.hpp>
 #include <openvino/op/unsqueeze.hpp>
+#include <openvino/op/util/op_types.hpp>
 #include <vector>
 
 #include "../node_context.hpp"
@@ -29,8 +30,13 @@ OutputVector translate_mulmat(const NodeContext& context) {
     ov::Output<Node> res;
     ov::Output<ov::Node> B = context.get_input(0);
     ov::Output<ov::Node> A = context.get_input(1);
-    if (context.get_input_type(0) != context.get_input_type(1)) {
+
+    bool convert_out_type = false;
+    if (ov::op::util::is_constant(B.get_node()) && context.get_input_type(0) != context.get_input_type(1)) {
         B = std::make_shared<ov::op::v0::Convert>(context.get_input(0), context.get_input_type(1));
+    } else if (context.get_input_type(0) != context.get_input_type(1)) {
+        A = std::make_shared<ov::op::v0::Convert>(context.get_input(1), context.get_input_type(0));
+        convert_out_type = true;
     }
 
     auto B_shape = context.get_input_shape(0).to_shape();
@@ -65,7 +71,12 @@ OutputVector translate_mulmat(const NodeContext& context) {
             A = Z;
         }
 
-        res = std::make_shared<ov::op::v0::MatMul>(A, B, false, true);
+        if (convert_out_type) {
+            auto result_lp = std::make_shared<ov::op::v0::MatMul>(A, B, false, true);
+            res = std::make_shared<ov::op::v0::Convert>(result_lp, context.get_output_type(0));
+        } else {
+            res = std::make_shared<ov::op::v0::MatMul>(A, B, false, true);
+        }
 
         return rename_outputs_with_suffix({res}, context.get_name());
 }

ggml/src/ggml-openvino/openvino/pass/mark_decompression_convert_constant_folding.hpp

@@ -0,0 +1,29 @@
+#pragma once
+
+#include "mark_decompression_convert_constant_folding.hpp"
+#include "openvino/pass/matcher_pass.hpp"
+#include "openvino/core/visibility.hpp"
+
+#ifdef OPENVINO_STATIC_LIBRARY
+#    define TRANSFORMATIONS_API
+#else
+#    ifdef IMPLEMENT_OPENVINO_API
+#        define TRANSFORMATIONS_API OPENVINO_CORE_EXPORTS
+#    else
+#        define TRANSFORMATIONS_API OPENVINO_CORE_IMPORTS
+#    endif  // IMPLEMENT_OPENVINO_API
+#endif      // OPENVINO_STATIC_LIBRARY
+
+namespace ov {
+namespace pass {
+
+class TRANSFORMATIONS_API MarkCompressedFloatConstants;
+
+}  // namespace pass
+}  // namespace ov
+
+class ov::pass::MarkCompressedFloatConstants : public MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("MarkCompressedFloatConstants");
+    MarkCompressedFloatConstants();
+};

ggml/src/ggml-openvino/openvino/translate_session.cpp

@@ -28,6 +28,7 @@
 #include "ggml-openvino/openvino/utils.hpp"
 #include "input_model.hpp"
 #include "pass/fuse_to_sdpa.hpp"
+#include "pass/mark_decompression_convert_constant_folding.hpp"
 
 namespace ov {
 namespace frontend {
@@ -259,6 +260,8 @@ std::shared_ptr<Model> TranslateSession::apply_transformations(std::shared_ptr<M
     {
         ov::pass::Manager manager;
         manager.set_per_pass_validation(true);
+        manager.register_pass<ov::pass::MarkCompressedFloatConstants>();
+        manager.register_pass<ov::pass::ConstantFolding>();
 
         if (!ggml_model_decoder->is_static()) {
             const auto kv_param_res_names = ggml_model_decoder->get_kv_param_res_names();
@@ -267,7 +270,7 @@ std::shared_ptr<Model> TranslateSession::apply_transformations(std::shared_ptr<M
         }
 
         // SDPA is even worse on performance
-        // manager.register_pass<pass::FuseToSDPA>();
+        manager.register_pass<pass::FuseToSDPA>();
         manager.run_passes(model);
     }
     auto preprocessor = ov::preprocess::PrePostProcessor(model);