Add prompt lookup decoding (#379)

Ticket: 138549

.github/workflows/causal_lm_cpp.yml

@@ -353,6 +353,52 @@ jobs:
           "
           echo "Alan Turing was a" passed
 
+
+  cpp-prompt_lookup_decoding_lm-ubuntu:
+    runs-on: ubuntu-20.04-16-cores
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          submodules: recursive
+      - uses: actions/setup-python@v4
+        with:
+          python-version: 3.8
+      - name: Install OpenVINO
+        run: |
+          mkdir ./ov/
+          curl https://storage.openvinotoolkit.org/repositories/openvino/packages/nightly/2024.1.0-14645-e6dc0865128/l_openvino_toolkit_ubuntu20_2024.1.0.dev20240304_x86_64.tgz | tar --directory ./ov/ --strip-components 1 -xz
+          sudo ./ov/install_dependencies/install_openvino_dependencies.sh
+      - name: Download, convert and build
+        run: |
+          source ./ov/setupvars.sh
+          python -m pip install --upgrade-strategy eager "optimum>=1.14" -r ./llm_bench/python/requirements.txt "transformers<4.38" ./thirdparty/openvino_tokenizers/[transformers] --extra-index-url https://download.pytorch.org/whl/cpu
+          python ./llm_bench/python/convert.py --model_id TinyLlama/TinyLlama-1.1B-Chat-v1.0 --output_dir ./TinyLlama-1.1B-Chat-v1.0/ --precision FP16
+          convert_tokenizer ./TinyLlama-1.1B-Chat-v1.0/pytorch/dldt/FP16/ --output ./TinyLlama-1.1B-Chat-v1.0/pytorch/dldt/FP16/ --with-detokenizer
+          cmake -DCMAKE_BUILD_TYPE=Release -S ./text_generation/causal_lm/cpp/ -B ./build/
+          cmake --build ./build/ --config Release -j
+          wait
+      - name: run and compare
+        run: |
+          source ./ov/setupvars.sh
+
+          echo 'Code:```python
+          def add(a, b):
+              return a + b
+          ```
+          Question: Can you please add 2 and 3
+          A:' > ./prompt.txt
+
+          ./build/prompt_lookup_decoding_lm ./TinyLlama-1.1B-Chat-v1.0/pytorch/dldt/FP16/ "$(<prompt.txt)" > predictions_prompt_lookup.txt
+          ./build/greedy_causal_lm ./TinyLlama-1.1B-Chat-v1.0/pytorch/dldt/FP16/ "$(<prompt.txt)" > predictions_greedy.txt
+          python -c "
+          with open('predictions_greedy.txt', 'r') as f:
+              predicted_greedy = f.readline()
+          with open('predictions_prompt_lookup.txt', 'r') as f:
+              predicted_prompt_lookup = f.readline()
+          assert predicted_greedy == predicted_prompt_lookup
+          "
+          echo "Prompt lookup" passed
+
   cpp-Phi-1_5:
     runs-on: ubuntu-20.04-16-cores
     steps:

text_generation/causal_lm/cpp/CMakeLists.txt

@@ -28,3 +28,11 @@ find_package(OpenVINO REQUIRED COMPONENTS Runtime)
 target_link_libraries(speculative_decoding_lm PRIVATE openvino::runtime)
 set_target_properties(speculative_decoding_lm PROPERTIES CXX_STANDARD 17)
 set_target_properties(speculative_decoding_lm PROPERTIES CXX_STANDARD_REQUIRED ON)
+
+add_executable(prompt_lookup_decoding_lm prompt_lookup_decoding_lm.cpp)
+target_compile_definitions(prompt_lookup_decoding_lm PRIVATE OPENVINO_TOKENIZERS_PATH=\"$<TARGET_FILE:openvino_tokenizers>\")
+target_include_directories(prompt_lookup_decoding_lm PRIVATE ./)
+find_package(OpenVINO REQUIRED COMPONENTS Runtime)
+target_link_libraries(prompt_lookup_decoding_lm PRIVATE openvino::runtime)
+set_target_properties(prompt_lookup_decoding_lm PROPERTIES CXX_STANDARD 17)
+set_target_properties(prompt_lookup_decoding_lm PROPERTIES CXX_STANDARD_REQUIRED ON)

text_generation/causal_lm/cpp/README.md

@@ -36,14 +36,18 @@ The program loads a tokenizer, a detokenizer and a model (`.xml` and `.bin`) to
 
 The program loads a tokenizer, a detokenizer and a model (`.xml` and `.bin`) to OpenVINO. A prompt is tokenized and passed to the model. The model predicts a distribution over the next tokens and group beam search samples from that distribution to explore possible sequesnses. The result is converted to chars and printed.
 
-### speculative_sampling_lm
+### speculative_decoding_lm
 
 Speculative decoding (or [assisted-generation](https://huggingface.co/blog/assisted-generation#understanding-text-generation-latency) in HF terminology) is a recent technique, that allows to speed up token generation when an additional smaller draft model is used alonside with the main model.
 
 Speculative decoding works the following way. The draft model predicts the next K tokens one by one in an autoregressive manner, while the main model validates these predictions and corrects them if necessary. We go through each predicted token, and if a difference is detected between the draft and main model, we stop and keep the last token predicted by the main model. Then the draft model gets the latest main prediction and again tries to predict the next K tokens, repeating the cycle.
 
 This approach reduces the need for multiple infer requests to the main model, enhancing performance. For instance, in more predictable parts of text generation, the draft model can, in best-case scenarios, generate the next K tokens that exactly match the target. In tha caste the are validated in a single inference request to the main model (which is bigger, more accurate but slower) instead of running K subsequent requests. More details can be found in the original paper https://arxiv.org/pdf/2211.17192.pdf, https://arxiv.org/pdf/2302.01318.pdf
 
+### prompt_lookup_decoding_lm
+
+[Prompt Lookup decoding](https://github.com/apoorvumang/prompt-lookup-decoding) is [assested-generation](https://huggingface.co/blog/assisted-generation#understanding-text-generation-latency) technique where the draft model is replaced with simple string matching the prompt to generate candidate token sequences. This method highly effective for input grounded generation (summarization, document QA, multi-turn chat, code editing), where there is high n-gram overlap between LLM input (prompt) and LLM output. This could be entity names, phrases, or code chunks that the LLM directly copies from the input while generating the output. Prompt lookup exploits this pattern to speed up autoregressive decoding in LLMs. This results in significant speedups with no effect on output quality.
+
 > [!NOTE]
 >Models should belong to the same family and have same tokenizers.
 
@@ -96,19 +100,22 @@ convert_tokenizer .\TinyLlama-1.1B-Chat-v1.0\pytorch\dldt\FP16\ --output .\TinyL
 ### Usage:
 1. `greedy_causal_lm <MODEL_DIR> "<PROMPT>"`
 2. `beam_search_causal_lm <MODEL_DIR> "<PROMPT>"`
-2. `speculative_decoding_lm <DRAFT_MODEL_DIR> <MAIN_MODEL_DIR> "<PROMPT>"`
+3. `speculative_decoding_lm <DRAFT_MODEL_DIR> <MAIN_MODEL_DIR> "<PROMPT>"`
+4. `prompt_lookup_decoding_lm <MODEL_DIR> "<PROMPT>"`
 
 ### Examples:
 
 #### Linux/MacOS:
 1. `./build/greedy_causal_lm ./TinyLlama-1.1B-Chat-v1.0/pytorch/dldt/FP16/ "Why is the Sun yellow?"`
 2. `./build/beam_search_causal_lm ./TinyLlama-1.1B-Chat-v1.0/pytorch/dldt/FP16/ "Why is the Sun yellow?"`
 3. `./build/speculative_decoding_lm ./TinyLlama-1.1B-Chat-v1.0/pytorch/dldt/FP16/ ./Llama-2-7b-chat-hf/pytorch/dldt/FP16/ "Why is the Sun yellow?"`
+4. `./build/prompt_lookup_decoding_lm ./TinyLlama-1.1B-Chat-v1.0/pytorch/dldt/FP16/ "Why is the Sun yellow?"`
 
 #### Windows:
 1. `.\build\Release\greedy_causal_lm .\TinyLlama-1.1B-Chat-v1.0\pytorch\dldt\FP16\ "Why is the Sun yellow?"`
 2. `.\build\Release\beam_search_causal_lm .\TinyLlama-1.1B-Chat-v1.0\pytorch\dldt\FP16\ "Why is the Sun yellow?"`
 3. `.\build\Release\speculative_decoding_lm .\TinyLlama-1.1B-Chat-v1.0\pytorch\dldt\FP16\ .\Llama-2-7b-chat-hf\pytorch\dldt\FP16\ "Why is the Sun yellow?"`
+4. `.\build\Release\prompt_lookup_decoding_lm .\TinyLlama-1.1B-Chat-v1.0\pytorch\dldt\FP16\ "Why is the Sun yellow?"`
 
 To enable Unicode characters for Windows cmd open `Region` settings from `Control panel`. `Administrative`->`Change system locale`->`Beta: Use Unicode UTF-8 for worldwide language support`->`OK`. Reboot.