NPU fix q4 perf regression

ggml/src/ggml-openvino/openvino/pass/squeeze_matmul.cpp

@@ -0,0 +1,58 @@
+#include "squeeze_matmul.hpp"
+
+#include <openvino/core/graph_util.hpp>
+#include <openvino/core/rt_info.hpp>
+#include <openvino/op/constant.hpp>
+#include <openvino/op/matmul.hpp>
+#include <openvino/op/squeeze.hpp>
+#include <openvino/op/unsqueeze.hpp>
+#include <openvino/pass/pattern/op/label.hpp>
+#include <openvino/pass/pattern/op/pattern.hpp>
+#include <openvino/pass/pattern/op/wrap_type.hpp>
+
+namespace opp = ov::pass::pattern;
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+// For quantized models, NPUW expects the activation to be 3d in DQ(DynamicQuantization) opt, e.g. DQMatMulGQ2i
+SqueezeMatmul::SqueezeMatmul() {
+    auto m_act = opp::any_input();
+    auto m_wei = opp::any_input();
+    auto m_matmul = opp::wrap_type<ov::op::v0::MatMul>({m_act, m_wei});
+
+    const auto callback = [=](ov::pass::pattern::Matcher & m) {
+        const auto & pattern_map = m.get_pattern_value_map();
+        auto matmul_node =
+            std::dynamic_pointer_cast<ov::op::v0::MatMul>(pattern_map.at(m_matmul).get_node_shared_ptr());
+        auto act = pattern_map.at(m_act);
+        auto wei = pattern_map.at(m_wei);
+        auto act_shape = act.get_partial_shape();
+        auto wei_shape = wei.get_partial_shape();
+        if (act_shape.rank().is_dynamic() || wei_shape.rank().is_dynamic()) {
+            return false;
+        }
+        if (act_shape.rank().get_length() == 4 && wei_shape.rank().get_length() == 2) {
+            auto axis = ov::op::v0::Constant::create(ov::element::i64, ov::Shape{1}, {0});
+            auto squeezed_act = std::make_shared<ov::op::v0::Squeeze>(act, axis);
+            auto new_matmul = std::make_shared<ov::op::v0::MatMul>(squeezed_act, wei, matmul_node->get_transpose_a(),
+                                                                   matmul_node->get_transpose_b());
+            auto unsqueezed_output = std::make_shared<ov::op::v0::Unsqueeze>(new_matmul, axis);
+            unsqueezed_output->set_friendly_name(matmul_node->get_friendly_name());
+            ov::copy_runtime_info(matmul_node, {squeezed_act, new_matmul, unsqueezed_output});
+            ov::replace_node(matmul_node, unsqueezed_output);
+            return true;
+        }
+        return false;
+    };
+
+    register_matcher(std::make_shared<ov::pass::pattern::Matcher>(m_matmul, "ov::frontend::ggml::pass::SqueezeMatmul"),
+                     callback);
+}
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

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

@@ -0,0 +1,17 @@
+#include "openvino/pass/matcher_pass.hpp"
+
+namespace ov {
+namespace frontend {
+namespace ggml {
+namespace pass {
+
+class SqueezeMatmul : public ov::pass::MatcherPass {
+public:
+    OPENVINO_MATCHER_PASS_RTTI("ov::frontend::ggml::pass::SqueezeMatmul")
+    SqueezeMatmul();
+};
+
+}  // namespace pass
+}  // namespace ggml
+}  // namespace frontend
+}  // namespace ov

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

@@ -5,6 +5,7 @@
 #include "input_model.hpp"
 #include "pass/eliminate_zp.hpp"
 #include "pass/mark_decompression_convert_constant_folding.hpp"
+#include "pass/squeeze_matmul.hpp"
 
 #include <cstdint>
 #include <cstdlib>
@@ -231,9 +232,10 @@ std::shared_ptr<Model> TranslateSession::apply_transformations(std::shared_ptr<M
         //     manager.register_pass<ov::pass::MakeStateful>(kv_param_res_pairs);
         // }
 
-        // if (ggml_model_decoder->is_static()) {
-        manager.register_pass<pass::EliminateZeroPoints>();
-        // }
+        if (ggml_model_decoder->is_static()) {
+            manager.register_pass<pass::EliminateZeroPoints>();
+            manager.register_pass<pass::SqueezeMatmul>();
+        }
         manager.run_passes(model);
     }
     return model;