Port from 24.6 release to master

.github/workflows/causal_lm_cpp.yml

@@ -63,13 +63,13 @@ jobs:
           PYTHONPATH: "./build"
       - run: >
           . ./ov/setupvars.sh
-          && timeout 25s ./build/samples/cpp/greedy_causal_lm/greedy_causal_lm ./open_llama_3b_v2/ "return 0"
-          | diff <(timeout 25s samples/python/greedy_causal_lm/greedy_causal_lm.py ./open_llama_3b_v2/ "return 0") -
+          && timeout 25s ./build/samples/cpp/text_generation/greedy_causal_lm ./open_llama_3b_v2/ "return 0"
+          | diff <(timeout 25s samples/python/text_generation/greedy_causal_lm.py ./open_llama_3b_v2/ "return 0") -
         env:
           PYTHONPATH: "./build"
       - run: >
           . ./ov/setupvars.sh
-          && samples/python/greedy_causal_lm/lora.py ./TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T/ adapter_model.safetensors "How to create a table with two columns, one of them has type float, another one has type int?"
+          && samples/python/text_generation/lora.py ./TinyLlama/TinyLlama-1.1B-intermediate-step-1431k-3T/ adapter_model.safetensors "How to create a table with two columns, one of them has type float, another one has type int?"
         env:
           PYTHONPATH: "./build"
 
@@ -249,7 +249,7 @@ jobs:
       - run: >
           set PATH=.\build\openvino_genai\;%PATH%
           && call .\ov\setupvars.bat
-          && .\build\samples\cpp\greedy_causal_lm\Release\greedy_causal_lm.exe .\TinyLlama-1.1B-Chat-v1.0\ 69 > .\cpp.txt
+          && .\build\samples\cpp\text_generation\Release\greedy_causal_lm.exe .\TinyLlama-1.1B-Chat-v1.0\ 69 > .\cpp.txt
       - run: |
           echo import transformers > ref.py
           echo predictions = open('cpp.txt', 'r').read() >> ref.py
@@ -266,13 +266,13 @@ jobs:
           set PATH=.\build\openvino_genai\;%PATH%
           && set "PYTHONPATH=./build/"
           && call .\ov\setupvars.bat
-          && python samples\python\greedy_causal_lm\greedy_causal_lm.py .\TinyLlama-1.1B-Chat-v1.0\ 69 > .\py.txt
+          && python samples\python\text_generation\greedy_causal_lm.py .\TinyLlama-1.1B-Chat-v1.0\ 69 > .\py.txt
       - run: fc .\cpp.txt .\py.txt
       - run: >
           set PATH=.\build\openvino_genai\;%PATH%
           && set "PYTHONPATH=./build/"
           && call .\ov\setupvars.bat
-          && python samples\python\greedy_causal_lm\lora.py .\TinyLlama\TinyLlama-1.1B-intermediate-step-1431k-3T\ adapter_model.safetensors "How to create a table with two columns, one of them has type float, another one has type int?"
+          && python samples\python\text_generation\lora.py .\TinyLlama\TinyLlama-1.1B-intermediate-step-1431k-3T\ adapter_model.safetensors "How to create a table with two columns, one of them has type float, another one has type int?"
 
   cpp-greedy_causal_lm-Qwen-7B-Chat:
     runs-on: ubuntu-20.04-16-cores
@@ -304,7 +304,7 @@ jobs:
           optimum-cli export openvino --trust-remote-code --weight-format fp16 --model Qwen/Qwen-7B-Chat Qwen-7B-Chat
       - run: >
           . ./ov/setupvars.sh
-          && timeout 2m ./build/samples/cpp/greedy_causal_lm/greedy_causal_lm ./Qwen-7B-Chat/ 69 | diff <(timeout 2m samples/python/greedy_causal_lm/greedy_causal_lm.py ./Qwen-7B-Chat/ 69) -
+          && timeout 2m ./build/samples/cpp/text_generation/greedy_causal_lm ./Qwen-7B-Chat/ 69 | diff <(timeout 2m samples/python/text_generation/greedy_causal_lm.py ./Qwen-7B-Chat/ 69) -
         env:
           PYTHONPATH: "./build"
 
@@ -446,7 +446,7 @@ jobs:
         run: |
           source ./ov/setupvars.sh
           ./build/samples/cpp/speculative_decoding_lm/speculative_decoding_lm ./dolly-v2-7b/ ./dolly-v2-3b/ "Alan Turing was a" > predictions_speculative.txt
-          ./build/samples/cpp/greedy_causal_lm/greedy_causal_lm ./dolly-v2-7b/ "Alan Turing was a" > predictions_greedy.txt
+          ./build/samples/cpp/text_generation/greedy_causal_lm ./dolly-v2-7b/ "Alan Turing was a" > predictions_greedy.txt
           python ./samples/python/speculative_decoding_lm/speculative_decoding_lm.py ./dolly-v2-7b/ ./dolly-v2-3b/ "Alan Turing was a" > predictions_py.txt
           python -c "
           with open('predictions_greedy.txt', 'r') as f:
@@ -504,7 +504,7 @@ jobs:
           A:' > ./prompt.txt
 
           ./build/samples/cpp/prompt_lookup_decoding_lm/prompt_lookup_decoding_lm ./TinyLlama-1.1B-Chat-v1.0/ "$(<prompt.txt)" > predictions_prompt_lookup.txt
-          ./build/samples/cpp/greedy_causal_lm/greedy_causal_lm ./TinyLlama-1.1B-Chat-v1.0/ "$(<prompt.txt)" > predictions_greedy.txt
+          ./build/samples/cpp/text_generation/greedy_causal_lm ./TinyLlama-1.1B-Chat-v1.0/ "$(<prompt.txt)" > predictions_greedy.txt
           python -c "
           with open('predictions_greedy.txt', 'r') as f:
               predicted_greedy = f.readline()
@@ -525,7 +525,7 @@ jobs:
           A:' > ./prompt.txt
 
           ./build/samples/cpp/prompt_lookup_decoding_lm/prompt_lookup_decoding_lm ./Qwen-7B-Chat/ "$(<prompt.txt)" > predictions_prompt_lookup.txt
-          ./build/samples/cpp/greedy_causal_lm/greedy_causal_lm ./Qwen-7B-Chat/ "$(<prompt.txt)" > predictions_greedy.txt
+          ./build/samples/cpp/text_generation/greedy_causal_lm ./Qwen-7B-Chat/ "$(<prompt.txt)" > predictions_greedy.txt
           python -c "
           with open('predictions_greedy.txt', 'r') as f:
               predicted_greedy = f.readline()
@@ -566,7 +566,7 @@ jobs:
       - name: Run Generation
         run: |
           source ./ov/setupvars.sh
-          timeout 50s ./build/samples/cpp/greedy_causal_lm/greedy_causal_lm ./phi-1_5/ "Alan Turing was a" > ./pred_greedy.txt
+          timeout 50s ./build/samples/cpp/text_generation/greedy_causal_lm ./phi-1_5/ "Alan Turing was a" > ./pred_greedy.txt
       - name: Compare
         run: |
           python -c "
@@ -585,7 +585,7 @@ jobs:
           echo Phi-1_5 passed
       - run: >
           . ./ov/setupvars.sh
-          && timeout 50s samples/python/greedy_causal_lm/greedy_causal_lm.py ./phi-1_5/ "Alan Turing was a"
+          && timeout 50s samples/python/text_generation/greedy_causal_lm.py ./phi-1_5/ "Alan Turing was a"
           | diff ./pred_greedy.txt -
         env:
           PYTHONPATH: "./build"
@@ -621,7 +621,7 @@ jobs:
       - name: Run Generation
         run: |
           source ./ov/setupvars.sh
-          timeout 50s ./build/samples/cpp/greedy_causal_lm/greedy_causal_lm ./redpajama-3b-chat/ "Alan Turing was a" > ./pred_greedy.txt
+          timeout 50s ./build/samples/cpp/text_generation/greedy_causal_lm ./redpajama-3b-chat/ "Alan Turing was a" > ./pred_greedy.txt
       - name: Compare
         run: |
           python -c "
@@ -640,7 +640,7 @@ jobs:
           echo "Alan Turing was a" passed
       - run: >
           . ./ov/setupvars.sh
-          && timeout 50s samples/python/greedy_causal_lm/greedy_causal_lm.py ./redpajama-3b-chat/ "Alan Turing was a"
+          && timeout 50s samples/python/text_generation/greedy_causal_lm.py ./redpajama-3b-chat/ "Alan Turing was a"
           | diff ./pred_greedy.txt -
         env:
           PYTHONPATH: "./build"

samples/CMakeLists.txt

@@ -5,7 +5,7 @@
 add_subdirectory(cpp/beam_search_causal_lm)
 add_subdirectory(cpp/benchmark_genai)
 add_subdirectory(cpp/chat_sample)
-add_subdirectory(cpp/greedy_causal_lm)
+add_subdirectory(cpp/text_generation)
 add_subdirectory(cpp/lora_greedy_causal_lm)
 add_subdirectory(cpp/multinomial_causal_lm)
 add_subdirectory(cpp/prompt_lookup_decoding_lm)
@@ -25,7 +25,7 @@ install(DIRECTORY
             cpp/beam_search_causal_lm
             cpp/benchmark_genai
             cpp/chat_sample
-            cpp/greedy_causal_lm
+            cpp/text_generation
             cpp/lora_greedy_causal_lm
             cpp/multinomial_causal_lm
             # Don't install prompt_lookup_decoding_lm because it doesn't use openvino_genai library and is not verified yet.
@@ -39,7 +39,7 @@ install(DIRECTORY
             python/beam_search_causal_lm
             python/benchmark_genai
             python/chat_sample
-            python/greedy_causal_lm
+            python/text_generation
             python/multinomial_causal_lm
             python/speculative_decoding_lm
             python/text2image

samples/cpp/greedy_causal_lm/CMakeLists.txt → samples/cpp/text_generation/CMakeLists.txt

@@ -20,3 +20,16 @@ install(TARGETS greedy_causal_lm
         RUNTIME DESTINATION samples_bin/
         COMPONENT samples_bin
         EXCLUDE_FROM_ALL)
+
+add_executable(encrypted_model_causal_lm encrypted_model_causal_lm.cpp)
+target_link_libraries(encrypted_model_causal_lm PRIVATE openvino::genai)
+set_target_properties(encrypted_model_causal_lm PROPERTIES
+    COMPILE_PDB_NAME encrypted_model_causal_lm
+    # Ensure out of box LC_RPATH on macOS with SIP
+    INSTALL_RPATH_USE_LINK_PATH ON)
+target_compile_features(encrypted_model_causal_lm PRIVATE cxx_std_11)
+
+install(TARGETS encrypted_model_causal_lm
+        RUNTIME DESTINATION samples_bin/
+        COMPONENT samples_bin
+        EXCLUDE_FROM_ALL)

samples/cpp/greedy_causal_lm/README.md → samples/cpp/text_generation/README.md

@@ -24,6 +24,18 @@ Discrete GPUs (dGPUs) usually provide better performance compared to CPUs. It is
 
 See https://github.com/openvinotoolkit/openvino.genai/blob/master/src/README.md#supported-models for the list of supported models.
 
+## Using encrypted models
+
+LLMPipeline and Tokenizer objects can be initialized directly from the memory buffer, e.g. when user stores only encrypted files and decrypts them on-the-fly. 
+The following code snippet demonstrates how to load the model from the memory buffer:
+
+```cpp
+auto [model_str, weights_tensor] = decrypt_model(models_path + "/openvino_model.xml", models_path + "/openvino_model.bin");
+ov::genai::Tokenizer tokenizer(models_path);
+ov::genai::LLMPipeline pipe(model_str, weights_tensor, tokenizer, device);
+```
+For the sake of brevity the code above does not include Tokenizer decryption. For more details look to encrypted_model_causal_lm sample.
+
 ### Troubleshooting
 
 #### Unicode characters encoding error on Windows

samples/cpp/text_generation/encrypted_model_causal_lm.cpp

@@ -0,0 +1,59 @@
+// Copyright (C) 2023-2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+
+#include "openvino/genai/llm_pipeline.hpp"
+#include <fstream>
+
+std::pair<std::string, ov::Tensor> decrypt_model(const std::string& model_path, const std::string& 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>());
+    std::vector<char> weights_buffer((std::istreambuf_iterator<char>(weights_file)), std::istreambuf_iterator<char>());
+    auto weights_tensor = ov::Tensor(ov::element::u8, {weights_buffer.size()}, weights_buffer.data());
+    return {model_str, weights_tensor};
+}
+
+ov::genai::Tokenizer decrypt_tokenizer(const std::string& models_path) {
+    std::string tok_model_path = models_path + "/openvino_tokenizer.xml";
+    std::string tok_weights_path = models_path + "/openvino_tokenizer.bin";
+    auto [tok_model_str, tok_weights_tensor] = decrypt_model(tok_model_path, tok_weights_path);
+
+    std::string detok_model_path = models_path + "/openvino_detokenizer.xml";
+    std::string detok_weights_path = models_path + "/openvino_detokenizer.bin";
+    auto [detok_model_str, detok_weights_tensor] = decrypt_model(tok_model_path, tok_weights_path);
+
+    return ov::genai::Tokenizer(tok_model_str, tok_weights_tensor, detok_model_str, detok_weights_tensor);
+}
+
+int main(int argc, char* argv[]) try {
+    if (3 > argc)
+        throw std::runtime_error(std::string{"Usage: "} + argv[0] + " <MODEL_DIR> \"<PROMPT>\"");
+
+    std::string device = "CPU";  // GPU, NPU can be used as well
+    std::string models_path = argv[1];
+    std::string prompt = argv[2];
+
+    auto [model_str, model_weights] = decrypt_model(models_path + "/openvino_model.xml", models_path + "/openvino_model.bin");
+    ov::genai::Tokenizer tokenizer = decrypt_tokenizer(models_path);
+
+    ov::genai::LLMPipeline pipe(model_str, model_weights, tokenizer, device);
+
+    std::string result = pipe.generate(prompt, ov::genai::max_new_tokens(100));
+    std::cout << result << std::endl;
+} 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/cpp/visual_language_chat/visual_language_chat.cpp

@@ -18,7 +18,7 @@ int main(int argc, char* argv[]) try {
 
     std::string device = "CPU";  // GPU can be used as well
     ov::AnyMap enable_compile_cache;
-    if ("GPU" == device) {
+    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")});

samples/python/multinomial_causal_lm/multinomial_causal_lm.py

@@ -85,6 +85,7 @@ def put(self, token_id: int) -> bool:
         text = self.tokenizer.decode(self.tokens_cache)
 
         word = ''
+        delay_n_chars = 4
         if len(text) > self.print_len and '\n' == text[-1]:
             # Flush the cache after the new line symbol.
             word = text[self.print_len:]            
@@ -93,11 +94,13 @@ def put(self, token_id: int) -> bool:
         elif len(text) >= 3 and text[-3:] == chr(65533):
             # Don't print incomplete text.
             pass
-        elif len(text) > self.print_len:
+        elif len(text) > self.print_len + delay_n_chars:
             # It is possible to have a shorter text after adding new token.
             # Print to output only if text length is increaesed.
-            word = text[self.print_len:]
-            self.print_len = len(text)
+            # E.g. when apostrophe removing regex had worked after adding new tokens.
+            # Several last characters are delayed before flushed to output.
+            word = text[self.print_len:-delay_n_chars]
+            self.print_len = len(text) - delay_n_chars
         self.put_word(word)        
 
         if self.get_stop_flag():

src/cpp/include/openvino/genai/continuous_batching_pipeline.hpp

@@ -88,6 +88,32 @@ class OPENVINO_GENAI_EXPORTS ContinuousBatchingPipeline {
         const ov::AnyMap& properties = {}
     );
 
+    /**
+     * @brief Constructs a ContinuousBatchingPipeline from already existing model and tokenizer.
+     * 
+     * This constructor allows for the creation of a ContinuousBatchingPipeline using an existing model
+     * represented as a string and a weights tensor, along with a manually initialized tokenizer.
+     * This is useful when the model and tokenizer are already loaded or created in memory and do not
+     * need to be loaded from files.
+     *
+     * @param model_str A string representation of the model.
+     * @param weights_tensor A tensor containing the weights of the model.
+     * @param tokenizer A manually initialized ov::genai::Tokenizer.
+     * @param scheduler_config Configuration for the scheduler.
+     * @param device The device to run the pipeline on (e.g., CPU, GPU).
+     * @param properties Optional properties for the pipeline.
+     * @param generation_config Optional generation configuration for the pipeline.
+     */
+    ContinuousBatchingPipeline(
+        const std::string& model_str,
+        const ov::Tensor& weights_tensor,
+        const ov::genai::Tokenizer& tokenizer,
+        const SchedulerConfig& scheduler_config,
+        const std::string& device,
+        const ov::AnyMap& properties = {},
+        const ov::genai::GenerationConfig& generation_config = {}
+    );
+
     ov::genai::Tokenizer get_tokenizer();
 
     ov::genai::GenerationConfig get_config() const;

src/cpp/include/openvino/genai/llm_pipeline.hpp

@@ -112,6 +112,15 @@ class OPENVINO_GENAI_EXPORTS LLMPipeline {
         const ov::AnyMap& properties = {}
     );
 
+    LLMPipeline(
+        const std::string& model_str,
+        const ov::Tensor& weights_tensor,
+        const ov::genai::Tokenizer& tokenizer,
+        const std::string& device,
+        const ov::AnyMap& properties = {},
+        const ov::genai::GenerationConfig& generation_config = {}
+    );
+
     OPENVINO_DEPRECATED("Please, specify device explicitly when create LLMPipeline. This overload will be removed in 2025.0.0 release")
     explicit LLMPipeline(const std::filesystem::path& path) :
         LLMPipeline(path, "CPU") { }
@@ -274,6 +283,14 @@ class OPENVINO_GENAI_EXPORTS LLMPipeline {
 OPENVINO_GENAI_EXPORTS std::pair<std::string, Any> streamer(StreamerVariant func);
 OPENVINO_GENAI_EXPORTS std::pair<std::string, Any> generation_config(const GenerationConfig& config);
 
+OPENVINO_GENAI_EXPORTS std::pair<std::string, Any> draft_model(
+    std::string& model_str,
+    ov::Tensor& weights_tensor,
+    const ov::genai::Tokenizer& tokenizer,
+    const std::string& device = {},
+    const ov::AnyMap& properties = {},
+    const ov::genai::GenerationConfig& generation_config = {});
+
 OPENVINO_GENAI_EXPORTS std::pair<std::string, Any> draft_model(
     const std::filesystem::path& models_path,
     const std::string& device = {},