[NPUW] Model builder: name projection MatMuls optimum-style for dynamic LoRA

make_linear named the projection MatMul with the bare layer name
("model.layers.N.self_attn.q_proj"). optimum-exported models name it
"<layer>/ov_ext::linear/MatMul". GenAI's dynamic-LoRA path builds its
lora_state_* variable ids from the MatMul friendly name, and NPUW's
ReshapeToStatic only pins the LoRA rank static for ids matching
".*MatMul\.(A|B|alpha)". With a bare name the rank stays dynamic and NPU
compilation throws "to_shape was called on a dynamic shape".

Append "/MatMul" to the MatMul friendly name so GenAI --lora --lora_mode
dynamic compiles for builder-generated models. Add a unit test asserting
projection MatMuls carry the /MatMul suffix.

Author: dylanneve1

src/plugins/intel_npu/tests/functional/behavior/npuw/test_engine/models/model_builder.cpp

@@ -72,7 +72,12 @@ ov::Output<ov::Node> make_linear(const ov::Output<ov::Node>& input,
     auto weight_output = weight_fn(name + ".weight", ov::Shape{out_features, in_features}, precision);
 
     auto matmul = std::make_shared<ov::opset11::MatMul>(input, weight_output, false, true);
-    matmul->set_friendly_name(name);
+    // Mirror optimum-exported naming ("<layer>/ov_ext::linear/MatMul"). GenAI's
+    // dynamic-LoRA path derives its lora_state_* variable ids from this MatMul's
+    // friendly name, and NPUW's ReshapeToStatic only pins the LoRA rank static for
+    // ids matching ".*MatMul\\.(A|B|alpha)". A bare name (no "MatMul") leaves the
+    // rank dynamic -> to_shape throws at compile. Keep the "MatMul" suffix.
+    matmul->set_friendly_name(name + "/MatMul");
 
     ov::Output<ov::Node> result = matmul->output(0);
 

src/plugins/intel_npu/tests/unit/npuw/model_builder_lora_test.cpp

@@ -53,4 +53,27 @@ TEST(ModelBuilderLoraTest, BuildLlmWithoutLoraHasNoLoraParameters) {
     EXPECT_FALSE(has_param(model, "lora_state_"));
 }
 
+// Projection MatMuls must be named optimum-style ("<layer>/MatMul"). GenAI's
+// dynamic-LoRA path derives lora_state_* variable ids from the MatMul friendly
+// name, and NPUW's ReshapeToStatic only pins the LoRA rank static for ids
+// matching ".*MatMul\\.(A|B|alpha)". A bare name leaves the rank dynamic and
+// compilation throws "to_shape was called on a dynamic shape".
+TEST(ModelBuilderLoraTest, ProjectionMatMulsAreNamedOptimumStyle) {
+    auto model = ov::test::npuw::build_llm_test_model();
+    ASSERT_TRUE(model);
+    bool found_qproj_matmul = false;
+    for (const auto& op : model->get_ordered_ops()) {
+        if (std::string(op->get_type_name()) != "MatMul") {
+            continue;
+        }
+        const auto name = op->get_friendly_name();
+        if (name.find("q_proj") != std::string::npos) {
+            EXPECT_NE(name.find("/MatMul"), std::string::npos)
+                << "projection MatMul '" << name << "' must end in /MatMul";
+            found_qproj_matmul = true;
+        }
+    }
+    EXPECT_TRUE(found_qproj_matmul);
+}
+
 }  // namespace