[intel-npu] Publishing NPU_DEFER_WEIGHTS_LOAD property

csoka · Nov 28, 2024 · 960c435 · 960c435
1 parent a1920c4
commit 960c435
Show file tree

Hide file tree

Showing 5 changed files with 16 additions and 14 deletions.
diff --git a/.../openvino-workflow/running-inference/inference-devices-and-modes/npu-device.rst b/.../openvino-workflow/running-inference/inference-devices-and-modes/npu-device.rst
@@ -146,6 +146,8 @@ offer a limited set of supported OpenVINO features.
          ov::intel_npu::turbo
          ov::intel_npu::tiles
          ov::intel_npu::max_tiles
+         ov::intel_npu::bypass_umd_caching
+         ov::intel_npu::defer_weights_load
 
    .. tab-item:: Read-only properties
 
@@ -168,7 +170,6 @@ offer a limited set of supported OpenVINO features.
          ov::intel_npu::device_alloc_mem_size
          ov::intel_npu::device_total_mem_size
          ov::intel_npu::driver_version
-         ov::intel_npu::bypass_umd_caching
 
 
 .. note::

diff --git a/src/inference/include/openvino/runtime/intel_npu/properties.hpp b/src/inference/include/openvino/runtime/intel_npu/properties.hpp
@@ -95,5 +95,12 @@ static constexpr ov::Property<int64_t> max_tiles{"NPU_MAX_TILES"};
  */
 static constexpr ov::Property<bool> bypass_umd_caching{"NPU_BYPASS_UMD_CACHING"};
 
+/**
+ * @brief [Only for NPU Plugin]
+ * Type: boolean, default is false
+ * This option allows to delay loading the weights until inference is created
+ */
+static constexpr ov::Property<bool> defer_weights_load{"NPU_DEFER_WEIGHTS_LOAD"};
+
 }  // namespace intel_npu
 }  // namespace ov
diff --git a/src/plugins/intel_npu/README.md b/src/plugins/intel_npu/README.md
@@ -176,6 +176,7 @@ The following properties are supported:
 | `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` |
 | `ov::intel_npu::bypass_umd_caching`/</br>`NPU_BYPASS_UMD_CACHING` | RW | Bypass the caching of compiled models in UMD. | `YES`/ `NO`| `NO` |
+| `ov::intel_npu::defer_weights_load`/</br>`NPU_DEFER_WEIGHTS_LOAD` | RW | Delay loading the weights until inference is created. | `YES`/ `NO`| `NO` |
 
 &nbsp;
 ### Performance Hint: Default Number of DPU Groups / DMA Engines

diff --git a/src/plugins/intel_npu/src/al/include/intel_npu/npu_private_properties.hpp b/src/plugins/intel_npu/src/al/include/intel_npu/npu_private_properties.hpp
@@ -305,13 +305,6 @@ static constexpr ov::Property<BatchMode> batch_mode{"NPU_BATCH_MODE"};
  */
 static constexpr ov::Property<int64_t> create_executor{"NPU_CREATE_EXECUTOR"};
 
-/**
- * @brief [Only for NPU Plugin]
- * Type: boolean, default is false
- * This option allows to omit loading the weights until inference is created
- */
-static constexpr ov::Property<bool> defer_weights_load{"NPU_DEFER_WEIGHTS_LOAD"};
-
 /**
  * @brief Read-only property to get the name of used backend
  */

diff --git a/src/plugins/intel_npu/src/plugin/src/plugin.cpp b/src/plugins/intel_npu/src/plugin/src/plugin.cpp
@@ -489,6 +489,12 @@ Plugin::Plugin()
           [](const Config& config) {
               return config.get<BYPASS_UMD_CACHING>();
           }}},
+        {ov::intel_npu::defer_weights_load.name(),
+         {true,
+          ov::PropertyMutability::RW,
+          [](const Config& config) {
+              return config.get<DEFER_WEIGHTS_LOAD>();
+          }}},
         // NPU Private
         // =========
         {ov::intel_npu::dma_engines.name(),
@@ -544,12 +550,6 @@ Plugin::Plugin()
           [](const Config& config) {
               return config.get<CREATE_EXECUTOR>();
           }}},
-        {ov::intel_npu::defer_weights_load.name(),
-         {false,
-          ov::PropertyMutability::RW,
-          [](const Config& config) {
-              return config.get<DEFER_WEIGHTS_LOAD>();
-          }}},
         {ov::intel_npu::dynamic_shape_to_static.name(),
          {false,
           ov::PropertyMutability::RW,