Support multiple images for InternVL models (openvinotoolkit#1099)

Ticket: CVS-155384

Author: yatarkan

src/cpp/src/visual_language/input_embedder.cpp src/cpp/src/visual_language/inputs_embedder.cpp

@@ -1,7 +1,7 @@
 // Copyright (C) 2023-2024 Intel Corporation
 // SPDX-License-Identifier: Apache-2.0
 
-#include "visual_language/image_embedder.hpp"
+#include "visual_language/inputs_embedder.hpp"
 
 #include "visual_language/clip.hpp"
 #include "visual_language/vision_encoder.hpp"
@@ -905,56 +905,60 @@ class InputsEmbedderInternVLChat : public InputsEmbedder::IInputsEmbedder {
         IInputsEmbedder(vlm_config, model_dir, device, device_config) { }
 
     virtual ov::Tensor get_inputs_embeds(const std::string& prompt, const std::vector<ov::Tensor>& images) override {
-        if (images.empty()) {
-            ov::Tensor input_ids = get_encoded_input_ids(prompt);
-            return m_embedding.infer(input_ids);
-        } else {
-            OPENVINO_ASSERT(1 == images.size(), "Only a single image allowed");
-            EncodedImage encoded_image = m_vision_encoder.encode(images.at(0));
-            ov::Tensor image_embeds = encoded_image.resized_source;
-            
-            std::string image_start_token = m_vlm_config.image_start_token;
-            std::string image_context_token = m_vlm_config.image_context_token;
-            std::string image_end_token = m_vlm_config.image_end_token;
+        std::string image_start_token = m_vlm_config.image_start_token;
+        std::string image_context_token = m_vlm_config.image_context_token;
+        std::string image_end_token = m_vlm_config.image_end_token;
+        
+        std::vector<ov::Tensor> single_images = to_single_image_tensors(images);
+
+        std::string formatted_prompt;
+        std::vector<ov::Tensor> image_embeds;
+        image_embeds.reserve(single_images.size());
+        
+        for (const auto& image : single_images) {
+            EncodedImage encoded_image = m_vision_encoder.encode(image);
+            ov::Tensor single_image_embeds = encoded_image.resized_source;
 
-            const size_t num_patches = image_embeds.get_shape().at(0);
-            const size_t num_image_tokens = image_embeds.get_shape().at(1);
+            const size_t num_patches = single_image_embeds.get_shape().at(0);
+            const size_t num_image_tokens = single_image_embeds.get_shape().at(1);
 
-            std::string concated_image_tokens;
-            concated_image_tokens += image_start_token;
+            formatted_prompt += image_start_token;
             for (int i = 0; i < num_patches * num_image_tokens; ++i) {
-                concated_image_tokens += image_context_token;
+                formatted_prompt += image_context_token;
             }
-            concated_image_tokens += image_end_token;
+            formatted_prompt += image_end_token + "\n";
 
-            std::string formatted_prompt = concated_image_tokens + "\n" + prompt;
-            
-            ov::Tensor input_ids = get_encoded_input_ids(formatted_prompt);
-            ov::Tensor text_embeds = m_embedding.infer(input_ids);
+            image_embeds.push_back(std::move(single_image_embeds));
+        }
+        formatted_prompt += prompt;
 
-            ov::Tensor encoded_image_context_token = m_tokenizer.encode(image_context_token, ov::genai::add_special_tokens(false)).input_ids;
-            int64_t image_context_token_id = encoded_image_context_token.data<int64_t>()[encoded_image_context_token.get_size() - 1];
+        ov::Tensor input_ids = get_encoded_input_ids(formatted_prompt);
+        ov::Tensor text_embeds = m_embedding.infer(input_ids);
 
-            return merge_text_and_image_embeddings_internvl(input_ids, text_embeds, image_embeds, image_context_token_id);
+        if (images.empty()) {
+            return text_embeds;
         }
+
+        ov::Tensor encoded_image_context_token = m_tokenizer.encode(image_context_token, ov::genai::add_special_tokens(false)).input_ids;
+        int64_t image_context_token_id = encoded_image_context_token.data<int64_t>()[encoded_image_context_token.get_size() - 1];
+
+        return merge_text_and_image_embeddings_internvl(input_ids, text_embeds, image_embeds, image_context_token_id);
     }
 
 protected:
     ov::Tensor merge_text_and_image_embeddings_internvl(
         const ov::Tensor& input_ids,
         const ov::Tensor& text_embeds,
-        const ov::Tensor& image_embeds,
+        const std::vector<ov::Tensor>& image_embeds,
         int64_t image_context_token_id
     ) {
         auto text_embeds_shape = text_embeds.get_shape();
-        auto image_embeds_shape = image_embeds.get_shape();
         size_t batch_size = text_embeds_shape.at(0);
         size_t seq_len = text_embeds_shape.at(1);
         size_t embed_dim = text_embeds_shape.at(2);
 
         ov::Tensor merged_embeds(text_embeds.get_element_type(), text_embeds_shape);
 
-        const float* image_embeds_data = image_embeds.data<float>();
         const float* text_embeds_data = text_embeds.data<float>();
         const int64_t* input_ids_data = input_ids.data<int64_t>();
         float* merged_embeds_data = merged_embeds.data<float>();
@@ -972,15 +976,27 @@ class InputsEmbedderInternVLChat : public InputsEmbedder::IInputsEmbedder {
 
         OPENVINO_ASSERT(image_context_tokens_count > 0, "input_ids does not contain image context token ids");
 
-        size_t vision_idx = 0;
+        size_t image_idx = 0;
+        size_t image_context_token_idx = 0;
         for (size_t i = 0; i < batch_size; ++i) {
             for (size_t j = 0; j < seq_len; ++j) {
                 size_t flat_idx = i * seq_len + j;
                 size_t offset = flat_idx * embed_dim;
 
                 if (image_context_tokens_mask[flat_idx]) {
-                    std::copy_n(image_embeds_data + vision_idx * embed_dim, embed_dim, merged_embeds_data + offset);
-                    ++vision_idx;
+                    const ov::Tensor& single_image_embeds = image_embeds[image_idx];
+                    const size_t num_all_image_tokens = single_image_embeds.get_shape().at(0) * single_image_embeds.get_shape().at(1); // num_patches * num_image_tokens
+                    const float* image_embeds_data = single_image_embeds.data<float>();
+                    std::copy_n(image_embeds_data + image_context_token_idx * embed_dim,
+                                embed_dim,
+                                merged_embeds_data + offset);
+                    
+                    ++image_context_token_idx;
+
+                    if (image_context_token_idx == num_all_image_tokens) {
+                        ++image_idx;
+                        image_context_token_idx = 0;
+                    }
                 } else {
                     std::copy_n(text_embeds_data + offset, embed_dim, merged_embeds_data + offset);
                 }

src/cpp/src/visual_language/image_embedder.hpp src/cpp/src/visual_language/inputs_embedder.hpp

@@ -8,7 +8,7 @@
 #include "openvino/genai/tokenizer.hpp"
 
 #include "visual_language/vlm_config.hpp"
-#include "visual_language/image_embedder.hpp"
+#include "visual_language/inputs_embedder.hpp"
 #include "visual_language/embedding_model.hpp"
 
 #include "sampler.hpp"