Fix build

src/cpp/src/continuous_batching_impl.cpp

@@ -105,7 +105,7 @@ void ContinuousBatchingPipeline::ContinuousBatchingImpl::initialize_pipeline(
             }
 
             size_t max_sequence_cache_occupation_length_in_blocks = scheduler_config.max_num_batched_tokens / m_scheduler->get_block_size()  + 1;
-            size_t embedding_size = device_config.get_head_size();
+            size_t embedding_size = device_config.get_k_head_size(0);
             m_cache_rotation_calculator = std::make_shared<CacheRotationCalculator>(
                 m_scheduler->get_block_size(),
                 max_sequence_cache_occupation_length_in_blocks,

src/cpp/src/device_config.hpp

@@ -158,8 +158,8 @@ class DeviceConfig {
         return m_key_cache_shape[id];
     }
 
-    size_t get_head_size() const {
-        return m_head_size;
+    size_t get_k_head_size(size_t layer_id) const {
+        return m_kv_heads_config[layer_id].k_head_size;
     }
 
     ov::PartialShape get_value_cache_shape(size_t id) const {

src/python/openvino_genai/py_openvino_genai.pyi

@@ -326,7 +326,9 @@ class CacheEvictionConfig:
         :param aggregation_mode: The mode used to compute the importance of tokens for eviction
         :type aggregation_mode: openvino_genai.AggregationMode
     
-        :param apply_rotation: Whether to apply cache rotation (RoPE-based) after each eviction
+        :param apply_rotation: Whether to apply cache rotation (RoPE-based) after each eviction.
+          Set this to false if your model has different RoPE scheme from the one used in the
+          original llama model and you experience accuracy issues with cache eviction enabled.
         :type apply_rotation: bool
     """
     aggregation_mode: AggregationMode