VLM: change infer to start_async/wait (#1947)

Author: dkalinowski

src/cpp/src/visual_language/embedding_model.cpp

@@ -78,7 +78,8 @@ ov::Tensor EmbeddingsModel::infer(const ov::Tensor& input_idx, bool return_remot
     } else {
         req.ireq.set_output_tensor(req.cpu_tensor);
     }
-    req.ireq.infer();
+    req.ireq.start_async();
+    req.ireq.wait();
     return req.ireq.get_output_tensor();
 }
 

src/cpp/src/visual_language/internvl_chat/classes.cpp

@@ -136,7 +136,8 @@ EncodedImage VisionEncoderInternVLChat::encode(const ov::Tensor& image, const ov
     ov::Tensor pixel_values = get_pixel_values_internvl(image, config);
 
     encoder.set_tensor("pixel_values", pixel_values);
-    encoder.infer();
+    encoder.start_async();
+    encoder.wait();
 
     const ov::Tensor& infer_output = encoder.get_output_tensor();
     ov::Tensor image_features(infer_output.get_element_type(), infer_output.get_shape());

src/cpp/src/visual_language/llava/classes.cpp

@@ -76,7 +76,8 @@ EncodedImage VisionEncoderLLaVA::encode( const ov::Tensor& image, const ov::AnyM
     ov::Tensor pixel_values = get_pixel_values_llava(image, config);
 
     encoder.set_tensor("pixel_values", pixel_values);
-    encoder.infer();
+    encoder.start_async();
+    encoder.wait();
 
     const ov::Tensor& infer_output = encoder.get_output_tensor();
     ov::Tensor image_features(infer_output.get_element_type(), infer_output.get_shape());

src/cpp/src/visual_language/llava_next/classes.cpp

@@ -57,7 +57,8 @@ EncodedImage VisionEncoderLLaVANext::encode(const ov::Tensor& image, const ov::A
     ov::Tensor pixel_values = get_pixel_values_llava_next(image, config);
 
     encoder.set_tensor("pixel_values", pixel_values);
-    encoder.infer();
+    encoder.start_async();
+    encoder.wait();
 
     const ov::Tensor& infer_output = encoder.get_output_tensor();
     ov::Tensor image_features(infer_output.get_element_type(), infer_output.get_shape());

src/cpp/src/visual_language/minicpm/classes.cpp

@@ -377,7 +377,8 @@ EncodedImage llava_image_embed_make_with_bytes_slice(clip_ctx& ctx_clip, const o
     }
     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();
+    encoder.start_async();
+    encoder.wait();
     const ov::Tensor& output_tensor = encoder.get_output_tensor();
 
     if (1 == preprocessed.size()) {
@@ -746,7 +747,8 @@ ov::Tensor InputsEmbedderMiniCPM::resample(const ov::Tensor& encoded_image, cons
     resampler.set_tensor("image_feature", encoded_image);  // [N, H*W, old_hidden_size]
     resampler.set_tensor("pos_embed", pos_embed);  // [H*W, N, new_hidden_size]
     resampler.set_tensor("key_padding_mask", key_padding_mask);  // [N, H*W]
-    resampler.infer();
+    resampler.start_async();
+    resampler.wait();
     return resampler.get_output_tensor();  // [N, query_num, new_hidden_size]
 }
 

src/cpp/src/visual_language/phi3_vision/classes.cpp

@@ -220,7 +220,8 @@ EncodedImage VisionEncoderPhi3V::encode(const ov::Tensor& image, const ov::AnyMa
     encoder.set_input_tensor(pixel_values);
     ov::Tensor res{ov::element::f32, encoder.get_output_tensor().get_shape()};
     encoder.set_output_tensor(res);
-    encoder.infer();
+    encoder.start_async();
+    encoder.wait();
     return {std::move(res), image_size};
 }
 
@@ -365,7 +366,8 @@ ov::Tensor reshape_hd_patches_2x2merge(const ov::Tensor& image_features, size_t
     hd_feature_transformer.set_input_tensor(1, height);
     ov::Tensor width{ov::element::i32, {}, &w_crop};
     hd_feature_transformer.set_input_tensor(2, width);
-    hd_feature_transformer.infer();
+    hd_feature_transformer.start_async();
+    hd_feature_transformer.wait();
     return hd_feature_transformer.get_output_tensor();
 }
 
@@ -432,7 +434,8 @@ ov::Tensor hd_feature_transform(const EncodedImage& image_features, InferRequest
     ov::Tensor sub_image_features_hd_newline = add_image_newline(sub_image_features_hd, sub_GN);  // [1,h_crop*12*(w_crop*12+1), 4096]
     ov::Tensor image_embeddings = concatenate_2d(sub_image_features_hd_newline, glb_GN, global_image_features_hd_newline);  // [1,l,4096]
     vision_projection.set_input_tensor(image_embeddings);
-    vision_projection.infer();
+    vision_projection.start_async();
+    vision_projection.wait();
     ov::Tensor out = vision_projection.get_output_tensor();
     ov::Tensor res{out.get_element_type(), out.get_shape()};
     out.copy_to(res);

src/cpp/src/visual_language/qwen2vl/classes.cpp

@@ -233,7 +233,8 @@ EncodedImage VisionEncoderQwen2VL::encode(const ov::Tensor& image, const ov::Any
     std::memcpy(flattened_patches.data(), transposed_patches.data(), transposed_patches.get_byte_size());
 
     encoder.set_tensor("hidden_states", flattened_patches);
-    encoder.infer();
+    encoder.start_async();
+    encoder.wait();
 
     const ov::Tensor& infer_output = encoder.get_output_tensor();
     ov::Tensor image_features(infer_output.get_element_type(), infer_output.get_shape());
@@ -435,7 +436,8 @@ ov::Tensor InputsEmbedderQwen2VL::merge_text_and_image_embeddings_qwen2vl(
     vision_embeddings_merger.set_tensor("hidden_states", concatenated_images);
     vision_embeddings_merger.set_tensor("attention_mask", attention_mask);
     vision_embeddings_merger.set_tensor("rotary_pos_emb", rotary_pos_emb);
-    vision_embeddings_merger.infer();
+    vision_embeddings_merger.start_async();
+    vision_embeddings_merger.wait();
     ov::Tensor processed_vision_embeds = vision_embeddings_merger.get_output_tensor();
 
     ov::Tensor merged_embeds(text_embeds.get_element_type(), text_embeds.get_shape());