Adjust options

Author: mengweiguo

src/cpp/src/rag/text_embedding_pipeline.cpp

@@ -245,17 +245,16 @@ class TextEmbeddingPipeline::TextEmbeddingPipelineImpl {
             }
         }
 
-        bool should_reshape = (device != "NPU" && (m_config.batch_size.has_value() || m_config.max_length.has_value())) ||
-            (device == "NPU" && m_config.batch_size.has_value() && is_fixed_size);
-        if (should_reshape) {
+        bool should_reshape_non_npu = (device != "NPU" && (m_config.batch_size.has_value() || m_config.max_length.has_value()));
+        bool should_reshape_npu = (device == "NPU" && m_config.batch_size.has_value() && is_fixed_size);
+        if (should_reshape_non_npu || should_reshape_npu) {
             reshape_model(model);
         }
 
         ov::CompiledModel compiled_model;
         if (device == "NPU" && model->is_dynamic()) {
-            const auto is_padding = m_config.pad_to_max_length.has_value() && m_config.pad_to_max_length.value();
-            OPENVINO_ASSERT(!is_padding || config.pooling_type == TextEmbeddingPipeline::PoolingType::MEAN,
-                            "Padding is only supported for the mean post-processing type");
+            OPENVINO_ASSERT(!(is_padding_on_left && is_fixed_size) || config.pooling_type == TextEmbeddingPipeline::PoolingType::MEAN,
+                            "Padding on left is only supported for the mean post-processing type");
 
             auto kv_pos = ov::genai::utils::get_kv_axes_pos(model);
             utils::KVDesc kv_desc;

src/cpp/src/utils.cpp

@@ -151,6 +151,12 @@ namespace ov {
 namespace genai {
 namespace utils {
 
+enum class ModelType {
+    Standard,
+    Whisper,
+    TextEmbedding
+};
+
 Tensor init_attention_mask(const Tensor& input_ids) {
     auto shape = input_ids.get_shape();
     auto attention_mask = ov::Tensor{input_ids.get_element_type(), shape};
@@ -608,10 +614,7 @@ void import_npu_model(ov::CompiledModel& compiled,
 }
 
 void export_npu_model(ov::CompiledModel& compiled,
-                        std::string& blob_path) {
-    if (blob_path.empty()) {
-        blob_path = "openvino_model.blob";
-    }
+                        const std::string& blob_path) {
     // Check the path is full
     const int EXT_SIZE = 5; // ".blob"
     if (blob_path.size() < EXT_SIZE) {
@@ -657,8 +660,7 @@ std::pair<ov::CompiledModel, KVDesc>
 compile_decoder_for_npu_impl(const std::shared_ptr<ov::Model>& model,
                         const ov::AnyMap& config,
                         const KVAxesPosition& kv_pos,
-                        const bool is_whisper,
-                        const bool is_text_embedding,
+                        ModelType model_type,
                         const ov::AnyMap& text_embed_config = {}) {
     ov::CompiledModel compiled;
     ov::AnyMap properties = config;
@@ -671,15 +673,25 @@ compile_decoder_for_npu_impl(const std::shared_ptr<ov::Model>& model,
     if (do_import) {
         import_npu_model(compiled, kv_desc, properties, blob_path);
     } else {
-        if (is_text_embedding) {
-            get_npu_text_embedding_config(properties, kv_pos, kv_desc, text_embed_config);
-        } else {
-            get_npu_model_config(properties, kv_pos, kv_desc, is_whisper);
+        switch (model_type) {
+            case ModelType::TextEmbedding:
+                get_npu_text_embedding_config(properties, kv_pos, kv_desc, text_embed_config);
+                break;
+            case ModelType::Whisper:
+                get_npu_model_config(properties, kv_pos, kv_desc, true);
+                break;
+            case ModelType::Standard:
+            default:
+                get_npu_model_config(properties, kv_pos, kv_desc, false);
+                break;
         }
 
         compiled = ov::genai::utils::singleton_core().compile_model(model, "NPU", properties);
         // Also export compiled model if required
         if (export_blob) {
+            if (blob_path.empty()) {
+                blob_path = "openvino_model.blob";
+            }
             export_npu_model(compiled, blob_path);
         }
     }
@@ -692,15 +704,17 @@ compile_decoder_for_npu(const std::shared_ptr<ov::Model>& model,
                         const ov::AnyMap& config,
                         const KVAxesPosition& kv_pos,
                         const bool is_whisper) {
-    return compile_decoder_for_npu_impl(model, config, kv_pos, is_whisper, false);
+    return compile_decoder_for_npu_impl(model, config, kv_pos,
+        is_whisper ? ModelType::Whisper : ModelType::Standard
+    );
 }
 
 std::pair<ov::CompiledModel, KVDesc>
 compile_decoder_for_npu_text_embedding(const std::shared_ptr<ov::Model>& model,
                         const ov::AnyMap& config,
                         const KVAxesPosition& kv_pos,
                         const ov::AnyMap& text_embed_config) {
-    return compile_decoder_for_npu_impl(model, config, kv_pos, false, true, text_embed_config);
+    return compile_decoder_for_npu_impl(model, config, kv_pos, ModelType::TextEmbedding, text_embed_config);
 }
 
 std::optional<ov::Any> pop_option(ov::AnyMap& config, const std::string& option_name) {