OVEP Stateful: Improve accuracy on NPU for sequence lengths >= 2048

Author: RyanMetcalfeInt8

onnxruntime/core/providers/openvino/ov_interface.cc

@@ -131,6 +131,13 @@ OVExeNetwork OVCore::StatefulCompileModel(std::shared_ptr<OVNetwork>& model,
     }
 
     UpdateNPUConfig(config, kv_pos, kv_desc);
+
+    if ((kv_desc.max_prompt_len + kv_desc.min_response_len) >= 2048) {
+      // This improves accuracy for generation sequences that exceed 2k tokens.
+      config["++NPUW_LLM_PREFILL_CONFIG"] = ov::AnyMap{{"NPUW_DEVICES", "NPU,CPU"}, {"NPUW_ONLINE_AVOID", "P:SinCos/NPU"}};
+      config["++NPUW_LLM_GENERATE_CONFIG"] = ov::AnyMap{{"NPUW_DEVICES", "NPU,CPU"}, {"NPUW_ONLINE_AVOID", "P:SinCos/NPU"}};
+    }
+
   } else {
     // This patches the OV IR model so that it only produces the logits required for sampling.
     // Actually either way that happens within NPUW::LLMCompiledModel creation for NPU device,