Add python bindings to VLMPipeline for encrypted models

samples/cpp/visual_language_chat/CMakeLists.txt

@@ -28,6 +28,21 @@ install(TARGETS visual_language_chat
         COMPONENT samples_bin
         EXCLUDE_FROM_ALL)
 
+# create encrypted model sample executable
+
+add_executable(encrypted_model_vlm encrypted_model_vlm.cpp load_image.cpp)
+target_include_directories(encrypted_model_vlm PRIVATE "${CMAKE_CURRENT_SOUCE_DIR}" "${CMAKE_BINARY_DIR}")
+target_link_libraries(encrypted_model_vlm PRIVATE openvino::genai)
+
+set_target_properties(encrypted_model_vlm PROPERTIES
+    # Ensure out of box LC_RPATH on macOS with SIP
+    INSTALL_RPATH_USE_LINK_PATH ON)
+
+install(TARGETS encrypted_model_vlm
+        RUNTIME DESTINATION samples_bin/
+        COMPONENT samples_bin
+        EXCLUDE_FROM_ALL)
+
 # create benchmark executable
 
 add_executable(benchmark_vlm benchmark_vlm.cpp load_image.cpp)
@@ -40,4 +55,4 @@ set_target_properties(benchmark_vlm PROPERTIES
 install(TARGETS benchmark_vlm
         RUNTIME DESTINATION samples_bin/
         COMPONENT samples_bin
-        EXCLUDE_FROM_ALL)
+        EXCLUDE_FROM_ALL)

samples/cpp/visual_language_chat/encrypted_model_vlm.cpp

@@ -0,0 +1,104 @@
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+
+#include <filesystem>
+#include <fstream>
+
+#include "load_image.hpp"
+#include "openvino/genai/visual_language/pipeline.hpp"
+
+std::pair<std::filesystem::path, std::filesystem::path> create_full_model_path(const std::filesystem::path& path, const std::string model_name) {
+    std::filesystem::path model_path, weight_path;
+    model_path = weight_path = path / model_name;
+    model_path += ".xml";
+    weight_path += ".bin";
+    return {model_path, weight_path};
+}
+
+std::pair<std::string, ov::Tensor> decrypt_model(const std::filesystem::path& model_path, const std::filesystem::path& weights_path) {
+    std::ifstream model_file(model_path);
+    std::ifstream weights_file(weights_path, std::ios::binary);
+    if (!model_file.is_open() || !weights_file.is_open()) {
+        throw std::runtime_error("Cannot open model or weights file");
+    }
+
+    // User can add file decryption of model_file and weights_file in memory here.
+
+    std::string model_str((std::istreambuf_iterator<char>(model_file)), std::istreambuf_iterator<char>());
+
+    weights_file.seekg(0, std::ios::end);
+    auto weight_size = static_cast<unsigned>(weights_file.tellg());
+    weights_file.seekg(0, std::ios::beg);
+    auto weights_tensor = ov::Tensor(ov::element::u8, {weight_size});
+    if (!weights_file.read(static_cast<char*>(weights_tensor.data()), weight_size)) {
+        throw std::runtime_error("Cannot read weights file");
+    }
+
+    return {model_str, weights_tensor};
+}
+
+ov::genai::Tokenizer decrypt_tokenizer(const std::filesystem::path& models_path) {
+    auto [tok_model_str, tok_weights_tensor] = std::apply(decrypt_model, create_full_model_path(models_path, "openvino_tokenizer"));
+    auto [detok_model_str, detok_weights_tensor] = std::apply(decrypt_model, create_full_model_path(models_path, "openvino_detokenizer"));
+
+    return ov::genai::Tokenizer(tok_model_str, tok_weights_tensor, detok_model_str, detok_weights_tensor);
+}
+
+bool print_subword(std::string&& subword) {
+    return !(std::cout << subword << std::flush);
+}
+
+int main(int argc, char* argv[]) try {
+    if (4 != argc) {
+        throw std::runtime_error(std::string{"Usage "} + argv[0] + " <MODEL_DIR> <IMAGE_FILE OR DIR_WITH_IMAGES> <PROMPT>");
+    }
+
+    //read and encrypt models
+    std::filesystem::path models_path = argv[1];
+    ov::genai::ModelsMap models_map;
+
+    std::map<std::string, std::string> model_name_to_file_map = {
+        {"language", "openvino_language_model"},
+        {"resampler", "openvino_resampler_model"},
+        {"text_embeddings", "openvino_text_embeddings_model"},
+        {"vision_embeddings", "openvino_vision_embeddings_model"}};
+
+    for (const auto& [model_name, file_name] : model_name_to_file_map) {
+        auto model_pair = std::apply(decrypt_model, create_full_model_path(models_path, file_name));
+        models_map.emplace(model_name, std::move(model_pair));
+    }
+
+    ov::genai::Tokenizer tokenizer = decrypt_tokenizer(models_path);
+
+    // GPU can be used as well.
+    std::string device = "CPU";
+    ov::AnyMap enable_compile_cache;
+    if (device == "GPU") {
+        // Cache compiled models on disk for GPU to save time on the
+        // next run. It's not beneficial for CPU.
+        enable_compile_cache.insert({ov::cache_dir("vlm_cache")});
+    }
+    ov::genai::VLMPipeline pipe(models_map, tokenizer, models_path,  device, enable_compile_cache);
+
+    ov::genai::GenerationConfig generation_config;
+    generation_config.max_new_tokens = 100;
+
+    std::vector<ov::Tensor> rgbs = utils::load_images(argv[2]);
+
+    std::string prompt = argv[3];
+    pipe.generate(prompt,
+                  ov::genai::images(rgbs),
+                  ov::genai::generation_config(generation_config),
+                  ov::genai::streamer(print_subword));
+
+} catch (const std::exception& error) {
+    try {
+        std::cerr << error.what() << '\n';
+    } catch (const std::ios_base::failure&) {}
+    return EXIT_FAILURE;
+} catch (...) {
+    try {
+        std::cerr << "Non-exception object thrown\n";
+    } catch (const std::ios_base::failure&) {}
+    return EXIT_FAILURE;
+}

samples/python/visual_language_chat/encrypted_model_vlm.py

@@ -0,0 +1,111 @@
+#!/usr/bin/env python3
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import argparse
+
+import numpy as np
+import openvino_genai
+import openvino
+from PIL import Image
+from openvino import Tensor
+from pathlib import Path
+import typing
+
+def decrypt_model(model_dir, model_file_name, weights_file_name):
+    with open(model_dir + '/' + model_file_name, "r") as file:
+        model = file.read()
+    # decrypt model
+
+    with open(model_dir + '/' + weights_file_name, "rb") as file:
+        binary_data = file.read()
+    # decrypt weights
+    weights = np.frombuffer(binary_data, dtype=np.uint8).astype(np.uint8)
+
+    return model, Tensor(weights)
+
+def read_tokenizer(model_dir):
+    tokenizer_model_name = 'openvino_tokenizer.xml'
+    tokenizer_weights_name = 'openvino_tokenizer.bin'
+    tokenizer_model, tokenizer_weights = decrypt_model(model_dir, tokenizer_model_name, tokenizer_weights_name)
+
+    detokenizer_model_name = 'openvino_detokenizer.xml'
+    detokenizer_weights_name = 'openvino_detokenizer.bin'
+    detokenizer_model, detokenizer_weights = decrypt_model(model_dir, detokenizer_model_name, detokenizer_weights_name)
+
+    return openvino_genai.Tokenizer(tokenizer_model, tokenizer_weights, detokenizer_model, detokenizer_weights)
+
+def streamer(subword: str) -> bool:
+    '''
+
+    Args:
+        subword: sub-word of the generated text.
+
+    Returns: Return flag corresponds whether generation should be stopped.
+
+    '''
+    print(subword, end='', flush=True)
+
+    # No value is returned as in this example we don't want to stop the generation in this method.
+    # "return None" will be treated the same as "return openvino_genai.StreamingStatus.RUNNING".
+
+
+def read_image(path: str) -> Tensor:
+    '''
+
+    Args:
+        path: The path to the image.
+
+    Returns: the ov.Tensor containing the image.
+
+    '''
+    pic = Image.open(path).convert("RGB")
+    image_data = np.array(pic)
+    return Tensor(image_data)
+
+
+def read_images(path: str) -> list[Tensor]:
+    entry = Path(path)
+    if entry.is_dir():
+        return [read_image(str(file)) for file in sorted(entry.iterdir())]
+    return [read_image(path)]
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument('model_dir')
+    parser.add_argument('image_dir', help="Image file or dir with images")
+    parser.add_argument('prompt', help="Image file or dir with images")
+    args = parser.parse_args()
+
+    model_name_to_file_map = {
+        ('language', 'openvino_language_model'),
+        ('resampler', 'openvino_resampler_model'),
+        ('text_embeddings', 'openvino_text_embeddings_model'),
+        ('vision_embeddings', 'openvino_vision_embeddings_model')}
+
+    models_map = dict()
+    for model_name, file_name in model_name_to_file_map:
+        model, weights = decrypt_model(args.model_dir, file_name + '.xml', file_name + '.bin')
+        models_map[model_name] = (model, weights)
+
+    tokenizer = read_tokenizer(args.model_dir)
+
+    # GPU can be used as well.
+    device = 'CPU'
+    enable_compile_cache = dict()
+    if "GPU" == device:
+        # Cache compiled models on disk for GPU to save time on the
+        # next run. It's not beneficial for CPU.
+        enable_compile_cache["CACHE_DIR"] = "vlm_cache"
+
+    pipe = openvino_genai.VLMPipeline(models_map, tokenizer, args.model_dir, device, **enable_compile_cache)
+
+    config = openvino_genai.GenerationConfig()
+    config.max_new_tokens = 100
+
+    rgbs = read_images(args.image_dir)
+
+    pipe.generate(args.prompt, images=rgbs, generation_config=config, streamer=streamer)
+
+if '__main__' == __name__:
+    main()

src/cpp/src/visual_language/pipeline.cpp

@@ -122,7 +122,7 @@ class VLMPipeline::VLMPipelineImpl : public VLMPipelineBase{
     ) :
         m_generation_config{generation_config} {
         m_is_npu = device.find("NPU") != std::string::npos;
-        OPENVINO_ASSERT(m_is_npu,
+        OPENVINO_ASSERT(!m_is_npu,
             "VLMPipeline initialization from string isn't supported for NPU device");
 
         m_inputs_embedder = std::make_shared<InputsEmbedder>(models_map, tokenizer, config_dir_path, device, properties);

src/python/openvino_genai/py_openvino_genai.pyi

@@ -2092,14 +2092,25 @@ class VLMPipeline:
     """
     This class is used for generation with VLMs
     """
+    @typing.overload
     def __init__(self, models_path: os.PathLike, device: str, **kwargs) -> None:
         """
-        device on which inference will be done
                     VLMPipeline class constructor.
                     models_path (os.PathLike): Path to the folder with exported model files.
                     device (str): Device to run the model on (e.g., CPU, GPU). Default is 'CPU'.
                     kwargs: Device properties
         """
+    @typing.overload
+    def __init__(self, models: dict[str, tuple[str, openvino._pyopenvino.Tensor]], tokenizer: Tokenizer, config_dir_path: os.PathLike, device: str, generation_config: GenerationConfig | None = None, **kwargs) -> None:
+        """
+                    VLMPipeline class constructor.
+                    models (typing.Dict[str, typing.Tuple[str, openvino.Tensor]]): Map with decrypted models. It should contains next models: language, resampler, text_embeddings, vision_embeddings.
+                    tokenizer (Tokenizer): Genai Tokenizers.
+                    config_dir_path (os.PathLike): Path to folder with model configs.
+                    device (str): Device to run the model on (e.g., CPU, GPU). Default is 'CPU'.
+                    generation_config (GenerationConfig | None): Device properties.
+                    kwargs: Device properties
+        """
     def finish_chat(self) -> None:
         ...
     @typing.overload

src/python/py_vlm_pipeline.cpp

@@ -165,14 +165,39 @@ void init_vlm_pipeline(py::module_& m) {
             return std::make_unique<ov::genai::VLMPipeline>(models_path, device, pyutils::kwargs_to_any_map(kwargs));
         }),
         py::arg("models_path"), "folder with exported model files",
-        py::arg("device"), "device on which inference will be done"
+        py::arg("device"), "device on which inference will be done",
         R"(
             VLMPipeline class constructor.
             models_path (os.PathLike): Path to the folder with exported model files.
             device (str): Device to run the model on (e.g., CPU, GPU). Default is 'CPU'.
             kwargs: Device properties
         )")
 
+        .def(py::init([](
+            const ov::genai::ModelsMap& models,
+            const ov::genai::Tokenizer& tokenizer,
+            const std::filesystem::path& config_dir_path,
+            const std::string& device,
+            const ov::genai::OptionalGenerationConfig& generation_config,
+            const py::kwargs& kwargs
+        ) {
+            return std::make_unique<ov::genai::VLMPipeline>(models, tokenizer, config_dir_path, device, pyutils::kwargs_to_any_map(kwargs), generation_config.value_or(ov::genai::GenerationConfig()));
+        }),
+        py::arg("models"), "map with decrypted models",
+        py::arg("tokenizer"), "genai Tokenizers",
+        py::arg("config_dir_path"), "Path to folder with model configs",
+        py::arg("device"), "device on which inference will be done",
+        py::arg("generation_config")  = std::nullopt, "generation config",
+        R"(
+            VLMPipeline class constructor.
+            models (typing.Dict[str, typing.Tuple[str, openvino.Tensor]]): Map with decrypted models. It should contains next models: language, resampler, text_embeddings, vision_embeddings.
+            tokenizer (Tokenizer): Genai Tokenizers.
+            config_dir_path (os.PathLike): Path to folder with model configs.
+            device (str): Device to run the model on (e.g., CPU, GPU). Default is 'CPU'.
+            generation_config (GenerationConfig | None): Device properties.
+            kwargs: Device properties
+        )")
+
         .def("start_chat", &ov::genai::VLMPipeline::start_chat, py::arg("system_message") = "")
         .def("finish_chat", &ov::genai::VLMPipeline::finish_chat)
         .def("set_chat_template", &ov::genai::VLMPipeline::set_chat_template, py::arg("chat_template"))

tests/python_tests/samples/test_encrypted_model_vlm.py

@@ -0,0 +1,39 @@
+# Copyright (C) 2025 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import os
+import pytest
+import sys
+
+from conftest import SAMPLES_PY_DIR, SAMPLES_CPP_DIR
+from test_utils import run_sample
+
+class TestEncryptedVLM:
+    @pytest.mark.llm
+    @pytest.mark.samples
+    @pytest.mark.parametrize("convert_model", ["tiny-random-minicpmv-2_6"], indirect=True)
+    @pytest.mark.parametrize("sample_args", ["Describe the images."])
+    @pytest.mark.parametrize("download_test_content", ["images/image.png"], indirect=True)
+    @pytest.mark.parametrize("generate_test_content", ["images/lines.png"], indirect=True)
+
+    def test_sample_encrypted_lm(self, convert_model, download_test_content, generate_test_content, sample_args):
+        # Test Python sample
+        py_script = os.path.join(SAMPLES_PY_DIR, "visual_language_chat/encrypted_model_vlm.py")
+        py_command = [sys.executable, py_script, convert_model, os.path.dirname(generate_test_content), sample_args]
+        py_result = run_sample(py_command)
+
+        # Test CPP sample
+        cpp_sample = os.path.join(SAMPLES_CPP_DIR, 'encrypted_model_vlm')
+        cpp_command =[cpp_sample, convert_model, os.path.dirname(generate_test_content), sample_args]
+        cpp_result = run_sample(cpp_command)
+
+        # Test common sample
+        py_common_script = os.path.join(SAMPLES_PY_DIR, "visual_language_chat/visual_language_chat.py")
+        py_common_command = [sys.executable, py_common_script, convert_model, os.path.dirname(generate_test_content)]
+        py_common_result = run_sample(py_common_command, sample_args)
+
+        # Compare results
+        assert py_result.stdout == cpp_result.stdout, f"Results should match"
+        # results from visual_language_chat sample contains also additional outputs like "question:".
+        # So just check if results of encrypted_model_vlm sample is a substring of it.
+        assert py_result.stdout in py_common_result.stdout, f"Results should match"