Merge branch 'add-VLM-matching-test'

Author: Wovchena

.github/workflows/causal_lm_cpp.yml

@@ -702,19 +702,44 @@ jobs:
         run: |
           source ./ov/setupvars.sh
           python -m pip install ./thirdparty/openvino_tokenizers/[transformers] --pre --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
-          python -m pip install --upgrade-strategy eager -r ./samples/requirements.txt --pre --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
+          python -m pip install --upgrade-strategy eager -r ./samples/requirements.txt opencv-python --pre --extra-index-url https://storage.openvinotoolkit.org/simple/wheels/nightly
       - name: Download and convert MiniCPM-V-2_6 model and an image
         run: |
           python -m pip install git+https://github.com/eaidova/optimum-intel.git@ea/minicpmv
           python -m pip install -U "optimum<1.23" --no-dependencies
           source ./ov/setupvars.sh
           optimum-cli export openvino -m openbmb/MiniCPM-V-2_6 MiniCPM-V-2_6 --trust-remote-code
           wget https://github.com/openvinotoolkit/openvino_notebooks/assets/29454499/d5fbbd1a-d484-415c-88cb-9986625b7b11 --output-document cat.jpg
+      - name: Generate reference
+        shell: python
+        run: |
+          from optimum.intel.openvino import OVModelForVisualCausalLM
+          from transformers import AutoProcessor
+          from PIL import Image
+          import cv2
+          import numpy as np
+          res = 448, 448
+          lines = np.arange(res[0] * res[1] * 3, dtype=np.uint8) % 255
+          lines = lines.reshape([*res, 3])
+          cv2.imwrite("lines.png", lines)
+          lines = Image.open("lines.png").convert('RGB')
+          model_id = "openbmb/MiniCPM-V-2_6"
+          processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
+          prompt = processor.tokenizer.apply_chat_template([{"role": "user", "content": "(<image>./</image>)\nWhat is unusual on this image?"}], tokenize=False, add_generation_prompt=True)
+          inputs = processor([prompt], [lines], return_tensors="pt")
+          model = OVModelForVisualCausalLM.from_pretrained("MiniCPM-V-2_6", device="CPU", trust_remote_code=True)
+          result = model.generate(**inputs, max_new_tokens=200)
+          decoded = processor.tokenizer.batch_decode(result[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True)[0]
+          print(decoded)
+          with open("ref.txt", "w") as f:
+              f.write(f"question:\n{decoded}\n----------\nquestion:\n")
+
       - name: Run visual_language_chat C++ sample - MiniCPM-V-2_6
         run: >
           source ./ov/setupvars.sh
-          && timeout 120s ./build/samples/cpp/visual_language_chat/visual_language_chat ./MiniCPM-V-2_6/ cat.jpg
-          <<< $'What is on the image?\nWhat is special on the image?'
+          && timeout 120s ./build/samples/cpp/visual_language_chat/visual_language_chat ./MiniCPM-V-2_6/ lines.png
+          <<< $'What is unusual on this image?' | tee cpp.txt
+      - run: diff cpp.txt ref.txt
       - name: Download and convert LLaVa 1.5 model and an image
         run: |
           source ./ov/setupvars.sh

src/cpp/src/visual_language/clip.hpp

@@ -6,25 +6,9 @@
 #include <vector>
 #include <numeric>
 
-//#define CLIP_DEBUG_FUNCTIONS
-enum projector_type {
-    PROJECTOR_TYPE_RESAMPLER,
-    PROJECTOR_TYPE_UNKNOWN,
-};
-
 struct clip_ctx {
-    bool has_text_encoder = false;
-    bool has_vision_encoder = false;
-    bool has_minicpmv_projector = false;
-
-    float image_mean[3];
-    float image_std[3];
-    int32_t ftype = 1;
-
-    std::vector<uint8_t> buf_compute_meta;
-
-    projector_type proj_type = PROJECTOR_TYPE_RESAMPLER;
-    size_t patch_size = 0;
+    float image_mean[3] = {0.0f, 0.0f, 0.0f};
+    float image_std[3] = {1.0f, 1.0f, 1.0f};
     size_t image_size = 0;
 };
 

src/cpp/src/visual_language/pipeline.cpp

@@ -557,6 +557,13 @@ ov::Tensor pack_image_features_llava_next(
         return result;
     }
 }
+
+// It's not possible to pass a GPU tensor from one model to another GPU
+// model on a different ov::Core instance.
+ov::Core singleton_core() {
+    static ov::Core core;
+    return core;
+}
 }
 
 class ov::genai::VLMPipeline::VLMPipelineImpl {
@@ -604,21 +611,22 @@ class ov::genai::VLMPipeline::VLMPipelineImpl {
             )
         },
         m_tokenizer{Tokenizer(model_dir.string(), device_config)},
-        m_vision_encoder(model_dir, m_vlm_config.model_type, device, device_config, ov::Core{}),
+        m_vision_encoder(model_dir, m_vlm_config.model_type, device, device_config, singleton_core()),
         m_is_chat_conversation{false},
         m_image_id{0} {
+            ov::Core core = singleton_core();
             if (m_vlm_config.model_type == VLMModelType::MINICPM) {
-                m_resampler = ov::Core{}.compile_model(
+                m_resampler = core.compile_model(
                     model_dir / "openvino_resampler_model.xml", device, device_config
                 ).create_infer_request();
 
                 m_pos_embed_cache = get_2d_sincos_pos_embed(m_vlm_config.hidden_size, {70, 70});
             }
-            m_embedding = ov::Core{}.compile_model(
+            m_embedding = core.compile_model(
                 model_dir / "openvino_text_embeddings_model.xml", device, device_config
             ).create_infer_request();
 
-            m_language = ov::Core{}.compile_model(
+            m_language = core.compile_model(
                 model_dir / "openvino_language_model.xml", device, device_config
             ).create_infer_request();
 

src/cpp/src/visual_language/vision_encoder.cpp

@@ -242,7 +242,6 @@ ov::Tensor prepare_vis_position_ids(
     });
     size_t position_ids_batch_elem = max_nb_patches_h * max_nb_patches_w;
     ov::Tensor position_ids{ov::element::i64, {batch_size, position_ids_batch_elem}};
-    // throw std::runtime_error("");
     int64_t* res_data = position_ids.data<int64_t>();
     std::fill_n(res_data, position_ids.get_size(), 0);
 
@@ -285,66 +284,84 @@ EncodedImage llava_image_embed_make_with_bytes_slice(clip_ctx& ctx_clip, const o
     std::vector<std::vector<ov::Tensor>> results;
     std::vector<std::vector<ImageSize>> sizes;
 
-    // std::vector<clip_image_f32*> img_res_v; // format N x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
     std::vector<std::vector<clip_image_f32>> preprocessed{imgs.size()};
-    std::transform(imgs.begin(), imgs.end(), preprocessed.begin(), [&ctx_clip](const std::vector<clip_image_u8>& row) {
+    size_t max_h = 0, max_w = 0, n_images = 0;
+    std::transform(imgs.begin(), imgs.end(), preprocessed.begin(), [&ctx_clip, &max_h, &max_w, &n_images](const std::vector<clip_image_u8>& row) {
         std::vector<clip_image_f32> processed_row{row.size()};
-        std::transform(row.begin(), row.end(), processed_row.begin(), [&ctx_clip](const clip_image_u8& raw) {
-            return clip_image_preprocess(ctx_clip, raw);
+        std::transform(row.begin(), row.end(), processed_row.begin(), [&ctx_clip, &max_h, &max_w, &n_images](const clip_image_u8& raw) {
+            clip_image_f32 im = clip_image_preprocess(ctx_clip, raw);
+            max_h = std::max(size_t(im.ny), max_h);
+            max_w = std::max(size_t(im.nx), max_w);
+            ++n_images;
+            return im;
         });
         return processed_row;
     });
 
+    ov::Tensor batched_images{ov::element::f32, {n_images, 3, max_h, max_w}};
+    float* batched_data = batched_images.data<float>();
     const clip_image_f32& resized_preprocessed = preprocessed.at(0).at(0);
-    ImageSize resized_source_size{resized_preprocessed.ny / patch_size, resized_preprocessed.nx / patch_size};
-    ov::Tensor input_tensor{ov::element::f32, {1, 3, size_t(resized_preprocessed.ny), size_t(resized_preprocessed.nx)}, (void*)(resized_preprocessed.buf.data())};
-    ov::Tensor pixel_values = preprocess_for_encoder(input_tensor, patch_size);
+    std::copy(resized_preprocessed.buf.begin(), resized_preprocessed.buf.end(), batched_data);
+    if (1 < preprocessed.size()) {
+        for (size_t row = 1; row < preprocessed.size(); ++row) {
+            size_t n_slices = preprocessed.at(row).size();
+            for (size_t col = 0; col < n_slices; ++col) {
+                const clip_image_f32& elem = preprocessed.at(row).at(col);
+                std::copy(elem.buf.begin(), elem.buf.end(), batched_data + ((row - 1) * n_slices + col + 1) * 3 * max_h * max_w);
+            }
+        }
+    }
+    ov::Tensor pixel_values = preprocess_for_encoder(batched_images, patch_size);
     encoder.set_tensor("pixel_values", pixel_values);
-    ov::Tensor patch_attention_mask{ov::element::f32, {pixel_values.get_shape().at(0), 1, resized_source_size.height * resized_source_size.width}};
-    std::fill_n(patch_attention_mask.data<float>(), patch_attention_mask.get_size(), 1.0f);
+
+    ov::Tensor patch_attention_mask{ov::element::f32, {pixel_values.get_shape().at(0), 1, max_h / patch_size * max_w / patch_size}};
+    float* attention_data = patch_attention_mask.data<float>();
+    std::fill_n(attention_data, patch_attention_mask.get_size(), 0.0f);
+    std::fill_n(attention_data, resized_preprocessed.ny / patch_size * resized_preprocessed.nx / patch_size, 1.0f);
+    if (1 < preprocessed.size()) {
+        for (size_t row = 1; row < preprocessed.size(); ++row) {
+            size_t n_slices = preprocessed.at(row).size();
+            for (size_t col = 0; col < n_slices; ++col) {
+                const clip_image_f32& elem = preprocessed.at(row).at(col);
+                std::fill_n(attention_data + ((row - 1) * n_slices + col + 1) * max_h / patch_size * max_w / patch_size, elem.ny / patch_size * elem.nx / patch_size, 1.0f);
+            }
+        }
+    }
     encoder.set_tensor("patch_attention_mask", patch_attention_mask);
-    ov::Tensor position_ids = prepare_vis_position_ids(pixel_values, patch_attention_mask, {resized_source_size}, ctx_clip.patch_size, ctx_clip.image_size / ctx_clip.patch_size);
+
+    ImageSize resized_source_size{resized_preprocessed.ny / patch_size, resized_preprocessed.nx / patch_size};
+    std::vector<ImageSize> tgt_sizes{resized_source_size};
+    if (1 < preprocessed.size()) {
+        for (const std::vector<clip_image_f32>& row : preprocessed) {
+            for (const clip_image_f32& elem : row) {
+                tgt_sizes.push_back({elem.ny / patch_size, elem.nx / patch_size});
+            }
+        }
+    }
+    ov::Tensor position_ids = prepare_vis_position_ids(pixel_values, patch_attention_mask, tgt_sizes, patch_size, ctx_clip.image_size / patch_size);
     encoder.set_tensor("position_ids", position_ids);
     encoder.infer();
     const ov::Tensor& output_tensor = encoder.get_output_tensor();
-    ov::Tensor resized_source{ov::element::f32, output_tensor.get_shape()};
-    output_tensor.copy_to(resized_source);
 
     if (1 == preprocessed.size()) {
+        ov::Tensor resized_source{ov::element::f32, output_tensor.get_shape()};
+        output_tensor.copy_to(resized_source);
         return {std::move(resized_source), resized_source_size};
     }
 
-    ImageSize raw_size{
-        size_t(preprocessed.at(1).at(0).ny),
-        size_t(preprocessed.at(1).at(0).nx)
-    };
-    ImageSize slices_size{
-        raw_size.height / patch_size,
-        raw_size.width / patch_size
-    };
-    size_t n_patches = slices_size.height * slices_size.width,
-        old_hidden_size = resized_source.get_shape().at(2);
+    size_t old_hidden_size = output_tensor.get_shape().at(2);
+    const float* out = output_tensor.data<float>();
+    ov::Tensor resized_source{ov::element::f32, {1, resized_source_size.height * resized_source_size.width, old_hidden_size}};
+    std::copy_n(out, resized_source.get_size(), resized_source.data<float>());
+
+    size_t n_patches = tgt_sizes.at(1).height * tgt_sizes.at(1).width;
     ov::Tensor encoded_slices{ov::element::f32, {preprocessed.size() - 1, preprocessed.at(1).size(), n_patches, old_hidden_size}};
-    for (size_t row = 1; row < preprocessed.size(); ++row) {
-        for (size_t col = 0; col < preprocessed.at(row).size(); ++col) {
-            clip_image_f32& elem = preprocessed.at(row).at(col);
-            ov::Tensor pixel_values = preprocess_for_encoder(
-                {ov::element::f32, {1, 3, size_t(elem.ny), size_t(elem.nx)}, elem.buf.data()},
-                patch_size
-            );
-            encoder.set_tensor("pixel_values", pixel_values);
-            ov::Tensor patch_attention_mask{ov::element::f32, {1, 1, slices_size.height * slices_size.width}};
-            std::fill_n(patch_attention_mask.data<float>(), patch_attention_mask.get_size(), 1.0f);
-            encoder.set_tensor("patch_attention_mask", patch_attention_mask);
-            ov::Tensor position_ids = prepare_vis_position_ids(pixel_values, patch_attention_mask, {slices_size}, ctx_clip.patch_size, ctx_clip.image_size / ctx_clip.patch_size);
-            encoder.set_tensor("position_ids", position_ids);
-            const ov::Tensor& old = encoder.get_output_tensor();
-            encoder.set_output_tensor({ov::element::f32, {1, n_patches, old_hidden_size}, encoded_slices.data<float>() + ((row - 1) * preprocessed.at(row).size() + col) * n_patches * old_hidden_size});
-            encoder.infer();
-            encoder.set_output_tensor(old);
+    for (size_t col = 0; col < preprocessed.size() - 1; ++col) {
+        for (size_t row = 0; row < preprocessed.at(1).size(); ++row) {
+            std::copy_n(out + (col * preprocessed.at(1).size() + row + 1) * n_patches * old_hidden_size, n_patches * old_hidden_size, encoded_slices.data<float>() + (col * preprocessed.at(1).size() + row) * n_patches * old_hidden_size);
         }
     }
-    return {resized_source, resized_source_size, encoded_slices, slices_size};
+    return {resized_source, resized_source_size, encoded_slices, tgt_sizes.at(1)};
 }
 
 ProcessorConfig from_any_map(
@@ -504,7 +521,6 @@ EncodedImage VisionEncoder::encode(const ov::Tensor& image, const ov::AnyMap& co
 
 EncodedImage VisionEncoder::encode_minicpm(const ov::Tensor& image, const ProcessorConfig& config) {
     clip_ctx ctx_clip;
-    ctx_clip.patch_size = m_processor_config.patch_size;
     ctx_clip.image_size = m_processor_config.image_size;
     std::copy(config.norm_mean.begin(), config.norm_mean.end(), ctx_clip.image_mean);
     std::copy(config.norm_std.begin(), config.norm_std.end(), ctx_clip.image_std);