[intel-npu] Adding NPU_DYNAMIC_QUANTIZATION property

docs/articles_en/openvino-workflow/running-inference/inference-devices-and-modes/npu-device.rst

@@ -143,6 +143,7 @@ offer a limited set of supported OpenVINO features.
          ov::enable_profiling
          ov::workload_type
          ov::intel_npu::compilation_mode_params
+         ov::intel_npu::compiler_dynamic_quantization
          ov::intel_npu::turbo
          ov::intel_npu::tiles
          ov::intel_npu::max_tiles

src/inference/include/openvino/runtime/intel_npu/properties.hpp

@@ -61,6 +61,14 @@ static constexpr ov::Property<uint32_t, ov::PropertyMutability::RO> driver_versi
  */
 static constexpr ov::Property<std::string> compilation_mode_params{"NPU_COMPILATION_MODE_PARAMS"};
 
+/**
+ * @brief [Only for NPU compiler]
+ * Type: boolean
+ * Set or verify state of dynamic quantization in  the NPU compiler
+ * @ingroup ov_runtime_npu_prop_cpp_api
+ */
+static constexpr ov::Property<bool> compiler_dynamic_quantization{"NPU_COMPILER_DYNAMIC_QUANTIZATION"};
+
 /**
  * @brief [Only for NPU plugin]
  * Type: std::bool

src/plugins/intel_npu/README.md

@@ -172,6 +172,7 @@ The following properties are supported:
 | `ov::intel_npu::device_total_mem_size`/</br>`NPU_DEVICE_TOTAL_MEM_SIZE` | RO | Size of available NPU DDR memory | `N/A` | `N/A` |
 | `ov::intel_npu::driver_version`/</br>`NPU_DRIVER_VERSION` | RO | NPU driver version. | `N/A` | `N/A` |
 | `ov::intel_npu::compilation_mode_params`/</br>`NPU_COMPILATION_MODE_PARAMS` | RW | Set various parameters supported by the NPU compiler. (See bellow) | `<std::string>`| `N/A` |
+| `ov::intel_npu::compiler_dynamic_quantization`/</br>`NPU_COMPILER_DYNAMIC_QUANTIZATION` | RW | Enable/Disable dynamic quantization by NPU compiler | `YES` / `NO` | `N/A` |
 | `ov::intel_npu::turbo`/</br>`NPU_TURBO` | RW | Set Turbo mode on/off | `YES`/ `NO`| `NO` |
 | `ov::intel_npu::tiles`/</br>`NPU_TILES` | RW | Sets the number of npu tiles to compile the model for | `[0-]` | `-1` |
 | `ov::intel_npu::max_tiles`/</br>`NPU_MAX_TILES` | RW | Maximum number of tiles supported by the device we compile for. Can be set for offline compilation. If not set, it will be populated by driver.| `[0-]` | `[1-6] depends on npu platform` |

src/plugins/intel_npu/src/al/include/intel_npu/config/compiler.hpp

@@ -357,4 +357,26 @@ struct COMPILATION_NUM_THREADS final : OptionBase<COMPILATION_NUM_THREADS, int32
     }
 };
 
+//
+// NPU_COMPILER_DYNAMIC_QUANTIZATION
+//
+
+struct COMPILER_DYNAMIC_QUANTIZATION final : OptionBase<COMPILER_DYNAMIC_QUANTIZATION, bool> {
+    static std::string_view key() {
+        return ov::intel_npu::compiler_dynamic_quantization.name();
+    }
+
+    static bool defaultValue() {
+        return false;
+    }
+
+    static OptionMode mode() {
+        return OptionMode::CompileTime;
+    }
+
+    static bool isPublic() {
+        return true;
+    }
+};
+
 }  // namespace intel_npu

src/plugins/intel_npu/src/compiler_adapter/src/driver_compiler_adapter.cpp

@@ -538,6 +538,16 @@ std::string DriverCompilerAdapter::serializeConfig(const Config& config,
         content = std::regex_replace(content, std::regex(batchstr.str()), "");
     }
 
+    // COMPILER_DYNAMIC_QUANTIZATION is not supported in versions < 6.1 - need to remove it
+    if ((compilerVersion.major < 6) || (compilerVersion.major == 6 && compilerVersion.minor < 3)) {
+        std::ostringstream dqstr;
+        dqstr << ov::intel_npu::compiler_dynamic_quantization.name() << KEY_VALUE_SEPARATOR << VALUE_DELIMITER << "\\S+"
+              << VALUE_DELIMITER;
+        logger.warning("COMPILER_DYNAMIC_QUANTIZATION property is not suppored by this compiler version. Removing from "
+                       "parameters");
+        content = std::regex_replace(content, std::regex(dqstr.str()), "");
+    }
+
     // NPU_DEFER_WEIGHTS_LOAD is needed at runtime only
     {
         std::ostringstream batchstr;

src/plugins/intel_npu/src/plugin/include/metrics.hpp

@@ -58,6 +58,7 @@ class Metrics final {
     };
     const std::vector<ov::PropertyName> _cachingProperties = {ov::device::architecture.name(),
                                                               ov::intel_npu::compilation_mode_params.name(),
+                                                              ov::intel_npu::compiler_dynamic_quantization.name(),
                                                               ov::intel_npu::tiles.name(),
                                                               ov::intel_npu::dpu_groups.name(),
                                                               ov::intel_npu::dma_engines.name(),

src/plugins/intel_npu/src/plugin/src/compiled_model.cpp

@@ -260,6 +260,12 @@ void CompiledModel::initialize_properties() {
           [](const Config& config) {
               return config.get<COMPILATION_MODE_PARAMS>();
           }}},
+        {ov::intel_npu::compiler_dynamic_quantization.name(),
+         {true,
+          ov::PropertyMutability::RO,
+          [](const Config& config) {
+              return config.get<COMPILER_DYNAMIC_QUANTIZATION>();
+          }}},
         {ov::intel_npu::turbo.name(),
          {isPropertySupported(ov::intel_npu::turbo.name()),
           ov::PropertyMutability::RO,

src/plugins/intel_npu/src/plugin/src/plugin.cpp

@@ -460,6 +460,12 @@ Plugin::Plugin()
           [](const Config& config) {
               return config.get<COMPILATION_MODE_PARAMS>();
           }}},
+        {ov::intel_npu::compiler_dynamic_quantization.name(),
+         {true,
+          ov::PropertyMutability::RW,
+          [](const Config& config) {
+              return config.get<COMPILER_DYNAMIC_QUANTIZATION>();
+          }}},
         {ov::intel_npu::turbo.name(),
          {_backends->isCommandQueueExtSupported(),
           ov::PropertyMutability::RW,