add missing config features from CV

src/cpp/continuous_batching/src/continuous_batching_pipeline.cpp

@@ -116,7 +116,7 @@ class ContinuousBatchingPipeline::Impl {
     }
 
     GenerationHandle add_request(uint64_t request_id, std::string prompt, ov::genai::GenerationConfig sampling_params) {
-        sampling_params.eos_token_id = m_tokenizer->get_eos_token_id();
+        sampling_params.set_eos_token_id(m_tokenizer->get_eos_token_id());
         sampling_params.validate();
 
         ov::Tensor input_ids;

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

@@ -94,6 +94,10 @@ class OPENVINO_GENAI_EXPORTS GenerationConfig {
     // EOS special token
     int64_t eos_token_id = -1;
 
+    /** @brief sets eos_token_id to tokenizer_eos_token_id if eos_token_id is less than 0.
+     * Otherwise verifies eos_token_id == tokenizer_eos_token_id.
+     */
+    void set_eos_token_id(size_t tokenizer_eos_token_id);
     size_t get_max_new_tokens(size_t prompt_length = 0) const;
     bool is_greedy_decoding() const;
     bool is_beam_search() const;

src/cpp/src/generation_config.cpp

@@ -49,6 +49,16 @@ GenerationConfig::GenerationConfig(const std::string& json_path) {
     }
 }
 
+void GenerationConfig::set_eos_token_id(size_t tokenizer_eos_token_id) {
+    if (eos_token_id < 0) {
+        eos_token_id = tokenizer_eos_token_id;
+    } else {
+        OPENVINO_ASSERT(eos_token_id == tokenizer_eos_token_id,
+            "EOS token ID is different in generation config (", eos_token_id, ") and tokenizer (",
+            tokenizer_eos_token_id, ")");
+    }
+}
+
 void GenerationConfig::update_generation_config(const ov::AnyMap& config_map) {
     using ov::genai::utils::read_anymap_param;
 
@@ -98,6 +108,7 @@ void GenerationConfig::validate() const {
                     "or set num_beams=1 if you with to use multinomial sampling.");
     OPENVINO_ASSERT(num_return_sequences > 0, "num_return_sequences must be greater than 0");
     OPENVINO_ASSERT(max_new_tokens > 0, "'max_new_tokens' must be greater than 0");
+    OPENVINO_ASSERT(min_new_tokens <= max_new_tokens, "min_new_tokens must be less or equal max_new_tokens");
 
     // max_new_tokens has priority over max_length
     // if max_new_tokens is defined no need to check max_length
@@ -123,6 +134,12 @@ void GenerationConfig::validate() const {
 
     OPENVINO_ASSERT(eos_token_id != -1 || max_new_tokens != SIZE_MAX || max_length != SIZE_MAX,
                     "Either 'eos_token_id', or 'max_new_tokens', or 'max_length' should be defined.");
+    if (is_beam_search()) {
+        OPENVINO_ASSERT(no_repeat_ngram_size > 0, "no_repeat_ngram_size must be positive");
+    } else {
+        OPENVINO_ASSERT(frequency_penalty >= -2.0f && frequency_penalty <= 2.0f, "frequence_penalty penalty must be a [-2; +2]");
+        OPENVINO_ASSERT(presence_penalty >= -2.0f && presence_penalty <= 2.0f, "presence_penalty penalty must be a [-2; +2]");
+    }
 }
 
 GenerationConfig beam_search() {

src/cpp/src/llm_pipeline.cpp

@@ -75,7 +75,7 @@ class StatefulLLMPipeline final : public LLMPipelineImplBase {
 
         // If eos_token_id was not provided, take value
         if (m_generation_config.eos_token_id == -1)
-            m_generation_config.eos_token_id = m_tokenizer.get_eos_token_id();
+            m_generation_config.set_eos_token_id(m_tokenizer.get_eos_token_id());
     }
 
     StatefulLLMPipeline(

src/cpp/src/llm_pipeline_static.cpp

@@ -203,7 +203,7 @@ EncodedResults StaticLLMPipeline::generate(
     GenerationConfig config = (generation_config.has_value()) ? *generation_config : m_generation_config;
     // If eos_token_id was not provided, take value from default m_generation_config
     if (config.eos_token_id == -1)
-        config.eos_token_id = m_generation_config.eos_token_id;
+        config.set_eos_token_id(m_generation_config.eos_token_id);
     config.validate();
 
     std::shared_ptr<StreamerBase> streamer_ptr;

src/python/py_generate_pipeline.cpp

@@ -137,7 +137,7 @@ OptionalGenerationConfig update_config_from_kwargs(const OptionalGenerationConfi
         } else if (key == "repetition_penalty") {
             res_config.repetition_penalty = py::cast<float>(item.second);
         } else if (key == "eos_token_id") {
-            res_config.eos_token_id = py::cast<int>(item.second);
+            res_config.set_eos_token_id(py::cast<int>(item.second));
         } else {
             throw(std::invalid_argument("'" + key + "' is incorrect GenerationConfig parameter name. "
                                         "Use help(openvino_genai.GenerationConfig) to get list of acceptable parameters."));
@@ -512,6 +512,7 @@ PYBIND11_MODULE(py_generate_pipeline, m) {
         .def_readwrite("presence_penalty", &GenerationConfig::presence_penalty)
         .def_readwrite("frequency_penalty", &GenerationConfig::frequency_penalty)
         .def_readwrite("rng_seed", &GenerationConfig::rng_seed)
+        .def("set_eos_token_id", &GenerationConfig::set_eos_token_id)
         .def("is_beam_search", &GenerationConfig::is_beam_search);
 
     py::class_<DecodedResults>(m, "DecodedResults")

tests/python_tests/continuous_batching/test_sampling.py

@@ -3,6 +3,7 @@
 import os
 import pytest
 import shutil
+import sys
 from dataclasses import dataclass
 from pathlib import Path
 from openvino_genai.py_continuous_batching import ContinuousBatchingPipeline
@@ -23,7 +24,7 @@
 
 @pytest.mark.precommit
 @pytest.mark.parametrize("model_id", get_models_list(os.path.join(os.path.dirname(os.path.realpath(__file__)), "models", "precommit")))
-@pytest.mark.xfail(reason='CPU: head size must be multiple of 16, current: 8', raises=RuntimeError, strict=True)
+@pytest.mark.xfail(reason='CPU: head size must be multiple of 16, current: 8. Ticket 145986.', raises=RuntimeError, strict=True)
 def test_sampling_precommit(tmp_path, model_id):
     run_test_pipeline(tmp_path, model_id)
 
@@ -165,6 +166,13 @@ class RandomSamplingTestStruct:
              "greedy_with_penalties",
              "multinomial_max_and_min_token"])
 def test_individual_generation_configs_random(tmp_path, test_struct: RandomSamplingTestStruct):
+    if test_struct in (
+        RANDOM_SAMPLING_TEST_CASES[1],
+        RANDOM_SAMPLING_TEST_CASES[3],
+        RANDOM_SAMPLING_TEST_CASES[6],
+        RANDOM_SAMPLING_TEST_CASES[10],
+    ) and sys.platform.startswith("win"):
+        pytest.xfail("assert ref_text == ov_text fails")
     generation_config = test_struct.generation_config
 
     prompts = test_struct.prompts