Merge branch 'master' into an/fix-android-jit-eltwise

Author: allnes

.github/dockerfiles/docker_tag

@@ -1 +1 @@
-pr-32459
+pr-33047

.github/dockerfiles/ov_build/debian_10_arm/Dockerfile

@@ -4,10 +4,16 @@ FROM ${REGISTRY}/library/debian:10.13
 USER root
 
 # APT configuration
+# WARNING: Debian 10 "Buster" is no longer officially supported,
+# so repositories were moved to "archive.debian.org"
+# See: https://www.debian.org/releases/
 RUN echo 'Acquire::Retries "10";' > /etc/apt/apt.conf && \
     echo 'APT::Get::Assume-Yes "true";' >> /etc/apt/apt.conf && \
     echo 'APT::Get::Fix-Broken "true";' >> /etc/apt/apt.conf && \
-    echo 'APT::Get::no-install-recommends "true";' >> /etc/apt/apt.conf
+    echo 'APT::Get::no-install-recommends "true";' >> /etc/apt/apt.conf && \
+    sed -i 's/deb.debian.org/archive.debian.org/g' /etc/apt/sources.list && \
+    sed -i 's/security.debian.org/archive.debian.org/g' /etc/apt/sources.list && \
+    sed -i '/stretch-updates/d' /etc/apt/sources.list
 
 ENV DEBIAN_FRONTEND="noninteractive" \
     TZ="Europe/London"

.github/dockerfiles/ov_test/debian_10_arm/Dockerfile

@@ -4,10 +4,16 @@ FROM ${REGISTRY}/library/debian:10.13
 USER root
 
 # APT configuration
+# WARNING: Debian 10 "Buster" is no longer officially supported,
+# so repositories were moved to "archive.debian.org"
+# See: https://www.debian.org/releases/
 RUN echo 'Acquire::Retries "10";' > /etc/apt/apt.conf && \
     echo 'APT::Get::Assume-Yes "true";' >> /etc/apt/apt.conf && \
     echo 'APT::Get::Fix-Broken "true";' >> /etc/apt/apt.conf && \
-    echo 'APT::Get::no-install-recommends "true";' >> /etc/apt/apt.conf
+    echo 'APT::Get::no-install-recommends "true";' >> /etc/apt/apt.conf && \
+    sed -i 's/deb.debian.org/archive.debian.org/g' /etc/apt/sources.list && \
+    sed -i 's/security.debian.org/archive.debian.org/g' /etc/apt/sources.list && \
+    sed -i '/stretch-updates/d' /etc/apt/sources.list
 
 ENV DEBIAN_FRONTEND="noninteractive" \
     TZ="Europe/London"

.github/dockerfiles/ov_test/debian_10_py310/Dockerfile

@@ -4,10 +4,16 @@ FROM ${REGISTRY}/library/debian:10.13
 USER root
 
 # APT configuration
+# WARNING: Debian 10 "Buster" is no longer officially supported,
+# so repositories were moved to "archive.debian.org"
+# See: https://www.debian.org/releases/
 RUN echo 'Acquire::Retries "10";' > /etc/apt/apt.conf && \
     echo 'APT::Get::Assume-Yes "true";' >> /etc/apt/apt.conf && \
     echo 'APT::Get::Fix-Broken "true";' >> /etc/apt/apt.conf && \
-    echo 'APT::Get::no-install-recommends "true";' >> /etc/apt/apt.conf
+    echo 'APT::Get::no-install-recommends "true";' >> /etc/apt/apt.conf && \
+    sed -i 's/deb.debian.org/archive.debian.org/g' /etc/apt/sources.list && \
+    sed -i 's/security.debian.org/archive.debian.org/g' /etc/apt/sources.list && \
+    sed -i '/stretch-updates/d' /etc/apt/sources.list
 
 ENV DEBIAN_FRONTEND="noninteractive" \
     TZ="Europe/London"

docs/articles_en/about-openvino/performance-benchmarks.rst

@@ -158,7 +158,7 @@ For a listing of all platforms and configurations used for testing, refer to the
 **Disclaimers**
 
 * System configurations used for Intel® Distribution of OpenVINO™ toolkit performance results
-  are based on release 2025.3, as of September 3rd, 2025.
+  are based on release 2025.4, as of December 1st, 2025.
 
 * OpenVINO Model Server performance results are based on release 2025.3, as of September 3rd, 2025.
 

docs/articles_en/about-openvino/performance-benchmarks/model-accuracy-int8-fp32.rst

@@ -41,7 +41,7 @@ the table for more information.
    * - mobilenet-v2
      - ImageNet2012
      - accuracy @ top1
-     - -0.93%
+     - -0.91%
      - -0.93%
      - -0.91%
      - -1.03%
@@ -96,28 +96,28 @@ the table for more information.
      - 0.00%
      - 0.00%
      - 0.02%
-     - 0.01%
+     - 0.02%
    * - resnet-50
      - ImageNet2012
      - accuracy @ top1
      - 0.00%
      - 0.00%
      - 0.00%
-     - -0.04%
+     - -0.01%
    * - ssd-resnet34-1200
      - COCO2017_detection_80cl_bkgr
      - map
      - 0.02%
      - 0.02%
      - 0.02%
-     - 0.06%
+     - -0.23%
    * - yolo_v11
      - COCO2017_detection_80cl
      - [email protected]:0.05:0.95
-     - 0.00%
-     - 0.00%
-     - 0.00%
-     - 
+     - 0.03%
+     - -2.21%
+     - -2.21%
+     - -2.21%
 .. list-table:: Model Accuracy for AMX-FP16, AMX-INT4, Arc-FP16 and Arc-INT4 (Arc™ B-series)
    :header-rows: 1
 
@@ -134,69 +134,62 @@ the table for more information.
      - 98.1%
      - 94.4%
      - 99.5%
-     - 92.6%
+     - 94.0%
    * - DeepSeek-R1-Distill-Qwen-1.5B
      - Data Default WWB
      - Similarity
      - 96.5%
      - 92.4%
      - 99.7%
-     - 92.1%
-   * - Gemma-3-1B-it
+     - 92.3%
+   * - Gemma-3-4B-it
      - Data Default WWB
      - Similarity
-     - 97.3%
      - 92.0%
-     - 99.2%
-     - 91.5%
-   * - GLM4-9B-Chat
-     - Data Default WWB
-     - Similarity
-     - 98.8%
-     - 93.3%
-     - %
-     - 95.0%
+     - 83.9%
+     - 
+     - 84.9%
    * - Llama-2-7B-chat
      - Data Default WWB
      - Similarity
      - 99.3%
      - 93.4%
      - 99.8%
-     - 91.9%
+     - 93.4%
    * - Llama-3-8B
      - Data Default WWB
      - Similarity
      - 98.8%
      - 94.3%
-     - %
+     - 99.7%
      - 94.5%
    * - Llama-3.2-3b-instruct
      - Data Default WWB
      - Similarity
-     - 98.2%
-     - 93.2%
-     - 98.4%
-     - 94.0%
-   * - Mistral-7b-instruct-V0.3
-     - Data Default WWB
-     - Similarity
-     - 98.3%
-     - 92.8%
-     - 99.9%
-     - 93.6%
+     - 97.9%
+     - 94.2%
+     - 99.7%
+     - 94.1%
    * - Phi4-mini-instruct
      - Data Default WWB
      - Similarity
-     - 96.4%
-     - 92.0%
-     - 99.3%
-     - 91.7%
+     - 89.1%
+     - 92.1%
+     - 99.5%
+     - 92.4%
    * - Qwen2-VL-7B
      - Data Default WWB
      - Similarity
-     - 97.8%
-     - 92.4%
+     - 97.5%
+     - 88.1%
      - 99.8%
+     - 91.4%
+   * - Qwen3-8B
+     - Data Default WWB
+     - Similarity
+     - 97.8%
+     - 92.3%
+     - 
      - 93.0%
    * - Flux.1-schnell
      - Data Default WWB
@@ -208,10 +201,10 @@ the table for more information.
    * - Stable-Diffusion-V1-5
      - Data Default WWB
      - Similarity
-     - 97.3%
-     - 95.1%
+     - 96.3%
+     - 93.3%
      - 99.5%
-     - 91.5%
+     - 93.7%
 
 Notes: For all accuracy metrics a "-", (minus sign), indicates an accuracy drop.
 The Similarity metric is the distance from "perfect" and as such always positive. 

docs/articles_en/about-openvino/performance-benchmarks/performance-benchmarks-faq.rst

@@ -55,11 +55,7 @@ Performance Information F.A.Q.
         - DeepSeek, HF
         - Auto regressive language
         - 128K
-      * - `GLM4-9B-chat <https://huggingface.co/THUDM/glm-4-9b-chat/tree/main>`__
-        - THUDM
-        - Transformer
-        - 128K
-      * - `Gemma-3-1B-it <https://huggingface.co/google/gemma-3-1b-it>`__
+      * - `Gemma-3-4B-it <https://huggingface.co/google/gemma-3-4b-it>`__
         - Hugginface
         - Text-To-Text Decoder-only
         - 128K
@@ -75,14 +71,6 @@ Performance Information F.A.Q.
         - Meta AI
         - Auto regressive language
         - 128K
-      * - `Mistral-7b-Instruct-V0.3 <https://huggingface.co/mistralai/Mistral-7B-v0.3>`__
-        - Mistral AI
-        - Auto regressive language
-        - 32K
-      * - `Phi3-4k-mini-Instruct <https://huggingface.co/microsoft/Phi-3-mini-4k-instruct>`__
-        - Huggingface
-        - Auto regressive language
-        - 4096
       * - `Phi4-mini-Instruct <https://huggingface.co/microsoft/Phi-4-mini-instruct>`__
         - Huggingface
         - Auto regressive language

docs/articles_en/assets/snippets/npu_remote_objects_creation.cpp

@@ -44,14 +44,24 @@ int main() {
     {
         //! [wrap_nt_handle]
         void* shared_buffer = nullptr;
-        auto remote_tensor = npu_context.create_tensor(in_element_type, in_shape, shared_buffer);
+        ov::intel_npu::MemType memory_type = ov::intel_npu::MemType::SHARED_BUF;
+        auto remote_tensor = npu_context.create_tensor(in_element_type, in_shape, shared_buffer, memory_type);
         //! [wrap_nt_handle]
     }
 
+    {
+        //! [import_cpu_va]
+        void* standard_allocation = nullptr;
+        ov::intel_npu::MemType memory_type = ov::intel_npu::MemType::CPU_VA;
+        auto remote_tensor = npu_context.create_tensor(in_element_type, in_shape, standard_allocation, memory_type);
+        //! [import_cpu_va]
+    }
+
     {
         //! [wrap_dmabuf_fd]
         int32_t fd_heap = 0;  // create the DMA-BUF System Heap file descriptor
-        auto remote_tensor = npu_context.create_tensor(in_element_type, in_shape, fd_heap);
+        ov::intel_npu::MemType memory_type = ov::intel_npu::MemType::SHARED_BUF;
+        auto remote_tensor = npu_context.create_tensor(in_element_type, in_shape, fd_heap, memory_type);
         //! [wrap_dmabuf_fd]
     }
 

docs/articles_en/documentation/compatibility-and-support/supported-operations.rst

@@ -218,6 +218,7 @@ Data as of OpenVINO 2024.4, 18 Oct. 2024.
        ScatterElements
        ScatterND
        Selu
+       SequenceAt
        Shape
        Shrink
        Sigmoid
@@ -231,6 +232,7 @@ Data as of OpenVINO 2024.4, 18 Oct. 2024.
        Softsign
        SpaceToDepth
        Split
+       SplitToSequence
        Sqrt
        Squeeze
        STFT

docs/articles_en/openvino-workflow/running-inference/inference-devices-and-modes/npu-device/remote-tensor-api-npu-plugin.rst

@@ -12,8 +12,8 @@ The NPU plugin supports memory sharing between OpenVINO and native APIs such as
 It implements the ``ov::RemoteContext`` and ``ov::RemoteTensor`` interfaces, providing mechanisms for efficient memory sharing.
 On Windows, the plugin exports an NT handle; on Linux, it uses a DMA-BUF System Heap. You can share this memory by
 passing the pointer as the ``shared_buffer`` member to the ``remote_tensor(..., shared_buffer)`` create function.
-Another option is to share memory by mapping a file into memory. These methods help avoid memory copy overhead when
-plugging OpenVINO inference into an existing NPU pipeline.
+Another option is to import memory by mapping a file into memory or by using a CPU virtual address allocation. These methods
+help avoid memory copy overhead when plugging OpenVINO inference into an existing NPU pipeline.
 
 Supported scenario by the Remote Tensor API:
 
@@ -81,8 +81,15 @@ For more details, see the code snippets below:
                :language: cpp
                :fragment: [file_mapping]
 
+         .. tab-item:: Import CPU virtual address allocation
+            :sync: import-cpu-va
+
+            .. doxygensnippet:: docs/articles_en/assets/snippets/npu_remote_objects_creation.cpp
+               :language: cpp
+               :fragment: [import_cpu_va]
+
          .. tab-item:: NT handle
-            :sync: nthandle
+            :sync: nt-handle
 
             .. doxygensnippet:: docs/articles_en/assets/snippets/npu_remote_objects_creation.cpp
                :language: cpp
@@ -118,7 +125,23 @@ For more details, see the code snippets below:
 Limitations
 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
 
-* Allocation of the NT handle or DMA-BUF System Heap file descriptor is done manually.
+The NPU plugin does not support methods for direct allocation of native handles.
+
+.. warning::
+
+   **CPU Virtual Address Allocation Requirements**
+   When using CPU virtual address allocations, you **must** comply with the following requirements to prevent memory corruption and crashes:
+
+   **1. Memory Alignment (Mandatory)**
+   Both the allocation pointer and its size must be aligned to the standard page size (4KB). Non-aligned allocations will be rejected.
+
+   **2. Allocation Lifetime (Critical)**
+   The allocation must remain valid **until ALL** of the following have occurred:
+      * All inference requests using this remote tensor have completed execution, **AND**
+      * All inference requests using this remote tensor have been destroyed, **AND**
+      * The remote tensor has been destroyed
+
+   Failure to maintain the allocation for the entire lifecycle will result in undefined behavior and potential crashes.
 
 Low-Level Methods for RemoteContext and RemoteTensor Creation
 #############################################################