[TextRerankPipeline] Force left padding for qwen3 reranker tests (openvinotoolkit#3109)

<!-- Keep your pull requests (PRs) as atomic as possible. That increases
the likelihood that an individual PR won't be stuck because of adjacent
problems, merge conflicts, or code review.
Your merged PR is going to appear in the automatically generated release
notes on GitHub. So the clearer the title the better. -->
## Description
There is an issue with default tokenizer padding side for qwen3. Force
left padding so far, ticket for investigation: 177405

Author: as-suvorov

src/cpp/include/openvino/genai/rag/text_rerank_pipeline.hpp

@@ -21,6 +21,16 @@ class OPENVINO_GENAI_EXPORTS TextRerankPipeline {
          */
         std::optional<size_t> max_length;
 
+        /**
+         * @brief If 'true', model input tensors are padded to the maximum length
+         */
+        std::optional<bool> pad_to_max_length;
+
+        /**
+         * @brief Side to use for padding "left" or "right"
+         */
+        std::optional<std::string> padding_side;
+
         /**
          * @brief Constructs text rerank pipeline configuration
          */

src/cpp/src/rag/text_rerank_pipeline.cpp

@@ -23,6 +23,8 @@ ov::AnyMap remove_config_properties(const ov::AnyMap& properties) {
 
     properties_copy.erase(top_n.name());
     properties_copy.erase(max_length.name());
+    properties_copy.erase(pad_to_max_length.name());
+    properties_copy.erase(padding_side.name());
 
     return properties_copy;
 }
@@ -138,6 +140,8 @@ using utils::read_anymap_param;
 TextRerankPipeline::Config::Config(const ov::AnyMap& properties) {
     read_anymap_param(properties, ov::genai::top_n.name(), top_n);
     read_anymap_param(properties, ov::genai::max_length.name(), max_length);
+    read_anymap_param(properties, ov::genai::padding_side.name(), padding_side);
+    read_anymap_param(properties, ov::genai::pad_to_max_length.name(), pad_to_max_length);
 };
 
 class TextRerankPipeline::TextRerankPipelineImpl {
@@ -154,6 +158,14 @@ class TextRerankPipeline::TextRerankPipelineImpl {
             m_tokenization_params.insert({max_length.name(), *m_config.max_length});
         }
 
+        if (m_config.pad_to_max_length) {
+            m_tokenization_params.insert({pad_to_max_length.name(), *m_config.pad_to_max_length});
+        }
+
+        if (m_config.padding_side) {
+            m_tokenization_params.insert({padding_side.name(), *m_config.padding_side});
+        }
+
         // qwen3 tokenizer doesn't support add_second_input(true)
         m_tokenizer = Tokenizer(models_path, ov::genai::add_second_input(!is_qwen3));
 

src/python/openvino_genai/py_openvino_genai.pyi

@@ -3448,7 +3448,13 @@ class TextRerankPipeline:
                 Number of documents to return sorted by score.
             max_length (int, optional):
                 Maximum length of tokens passed to the embedding model.
+            pad_to_max_length (bool, optional):
+                If 'True', model input tensors are padded to the maximum length.
+            padding_side (str, optional):
+                Side to use for padding "left" or "right"
         """
+        pad_to_max_length: bool | None
+        padding_side: str | None
         @typing.overload
         def __init__(self) -> None:
             ...

src/python/py_rag.cpp

@@ -53,6 +53,10 @@ Structure to keep TextRerankPipeline configuration parameters.
         Number of documents to return sorted by score.
     max_length (int, optional):
         Maximum length of tokens passed to the embedding model.
+    pad_to_max_length (bool, optional):
+        If 'True', model input tensors are padded to the maximum length.
+    padding_side (str, optional):
+        Side to use for padding "left" or "right"
 )";
 
 }  // namespace
@@ -226,7 +230,9 @@ kwargs: Plugin and/or config properties
             return ov::genai::TextRerankPipeline::Config(pyutils::kwargs_to_any_map(kwargs));
         }))
         .def_readwrite("top_n", &ov::genai::TextRerankPipeline::Config::top_n)
-        .def_readwrite("max_length", &ov::genai::TextRerankPipeline::Config::max_length);
+        .def_readwrite("max_length", &ov::genai::TextRerankPipeline::Config::max_length)
+        .def_readwrite("pad_to_max_length", &ov::genai::TextRerankPipeline::Config::pad_to_max_length)
+        .def_readwrite("padding_side", &ov::genai::TextRerankPipeline::Config::padding_side);
 
     text_rerank_pipeline.def(
         py::init([](const std::filesystem::path& models_path,

tests/python_tests/test_rag.py

@@ -71,7 +71,6 @@ def rerank_model(request) -> OVConvertedModelSchema:
     return download_and_convert_model_class(model_id, OVModelForSequenceClassification)
 
 
-
 @pytest.fixture(scope="module")
 def emb_model(request) -> OVConvertedModelSchema:
     model_id = request.param
@@ -199,7 +198,6 @@ def validate_embedding_results(result_1: EmbeddingResult, result_2: EmbeddingRes
 
     max_error = np.abs(np_result_1 - np_result_2).max()
     assert max_error < MAX_EMBEDDING_ERROR, f"Max error: {max_error} is greater than allowed {MAX_EMBEDDING_ERROR}"
-    
 
 
 def run_text_embedding_pipeline_with_ref(
@@ -620,10 +618,17 @@ def test_qwen3_seq_cls_rerank_documents(rerank_model: OVConvertedModelSchema, qu
 @pytest.mark.parametrize(
     "config",
     [
-        TextRerankPipeline.Config(top_n=4),
+        pytest.param(
+            TextRerankPipeline.Config(top_n=4),
+            marks=pytest.mark.skip(
+                reason="Qwen3 Reranker different default tokenizer padding side on Win vs Linux: 177405"
+            ),
+        ),
+        TextRerankPipeline.Config(top_n=4, padding_side="left"),
     ],
     ids=[
         "top_n=4",
+        "top_n=4, padding_side=left",
     ],
 )
 @pytest.mark.xfail(condition=(sys.platform == "darwin"), reason="Ticket - 174635")