Add image pre‑resize for qwen3‑vl and qwen2.5‑vl

src/common/AutoModel/modeling_qwen2vl.cpp

@@ -13,6 +13,10 @@
 /************              Qwen2VL family            **************/
 Qwen2VL::Qwen2VL(xrt::device* npu_device_inst) : AutoModel(npu_device_inst, "Qwen2VL") {}
 
+void Qwen2VL::set_special_flags(int resize) {
+    this->resize = resize;
+}
+
 void Qwen2VL::load_model(std::string model_path, json model_info, int default_context_length, bool enable_preemption) {
     this->_shared_load_model(model_path, model_info, default_context_length, enable_preemption);
 

src/common/AutoModel/modeling_qwen2vl_image.cpp

@@ -512,17 +512,35 @@ void Qwen2VL::preprocess_image(qwen2vl_image_t& image, std::vector<bf16> &pixel_
     const int width = image.width;
     const int height = image.height;
     const int channels = 3; // RGB
+    int downscaled_height;
+    int downscaled_width;
     int resized_height; 
     int resized_width;    
+
+    //std::cout << "height " << height << " width " << width << std::endl;
+
+    int target_longest_edge = (resize == 0) ? 1080 : (resize == 1) ? 1920 : (resize == 2) ? 2560 : 0;
+
+    if (target_longest_edge > 0) {
+        float scale = target_longest_edge / static_cast<float>(std::max(height, width));
+        scale = std::min(scale, 1.0f);
+        downscaled_height = static_cast<int>(height * scale);
+        downscaled_width = static_cast<int>(width * scale);
+    }
+    else {
+        downscaled_height = height;
+        downscaled_width = width;
+    }
+
     // do the automatically resizing in here 
     smart_resize(
-        height, width,
+        downscaled_height, downscaled_width,
         resized_height, resized_width,
         QWEN2_PATCH_SIZE * QWEN2_IMAGE_MERGE_SIZE,
         QWEN2_SHORTEST_EDGE,
         QWEN2_LONGEST_EDGE
     );
-    // std::cout << "resized_height "<< resized_height << " resized_width " << resized_width <<std::endl;
+     //std::cout << "resized_height "<< resized_height << " resized_width " << resized_width <<std::endl;
 
     // Use non-optimized path for consistent results across platforms
     auto resize_image =  imgproc::avx512::resize_bicubic_antialias_rgb_planar_avx512(

src/common/AutoModel/modeling_qwen3vl.cpp

@@ -13,6 +13,10 @@
 /************              Qwen3VL family            **************/
 Qwen3VL::Qwen3VL(xrt::device* npu_device_inst) : AutoModel(npu_device_inst, "Qwen3VL") {}
 
+void Qwen3VL::set_special_flags(int resize) {
+    this->resize = resize;
+}
+
 void Qwen3VL::load_model(std::string model_path, json model_info, int default_context_length, bool enable_preemption) {
     this->_shared_load_model(model_path, model_info, default_context_length, enable_preemption);
 

src/common/AutoModel/modeling_qwen3vl_image.cpp

@@ -512,17 +512,36 @@ void Qwen3VL::preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_
     const int width = image.width;
     const int height = image.height;
     const int channels = 3; // RGB
+    int downscaled_height;
+    int downscaled_width;
     int resized_height; 
-    int resized_width;    
+    int resized_width;   
+
+    //std::cout << "height "<< height << " width " << width <<std::endl;
+
+    int target_longest_edge = (resize == 0) ? 1080 : (resize == 1) ? 1920 : (resize == 2) ? 2560 : 0;
+
+    if (target_longest_edge > 0) {
+        float scale = target_longest_edge / static_cast<float>(std::max(height, width));
+        scale = std::min(scale, 1.0f); 
+        downscaled_height = static_cast<int>(height * scale);
+        downscaled_width = static_cast<int>(width * scale);
+    }
+    else {
+        downscaled_height = height;
+        downscaled_width = width;
+    }
+
     // do the automatically resizing in here 
     smart_resize(
-        height, width,
+        downscaled_height, downscaled_width,
         resized_height, resized_width,
         QWEN3_PATCH_SIZE * QWEN3_IMAGE_MERGE_SIZE,
         QWEN3_SHORTEST_EDGE,
         QWEN3_LONGEST_EDGE
     );
-    // std::cout << "resized_height "<< resized_height << " resized_width " << resized_width <<std::endl;
+
+    //std::cout << "resized_height " << resized_height << " resized_width " << resized_width << std::endl;
 
     // Cache size calculations for efficiency
     const uint32_t single_frame_size = resized_height * resized_width * channels;

src/include/AutoModel/automodel.hpp

@@ -311,7 +311,8 @@ class AutoModel {
 	}
 
 	//************ Unique for each model *************/
-
+	virtual void set_special_flags(int) {}
+
 	virtual void load_model(std::string model_path, json model_info, int default_context_length = -1, bool enable_preemption = false) {}
 	virtual std::string generate(chat_meta_info_t& meta_info, int length_limit, std::ostream& os, std::function<bool()> is_cancelled = [] { return false; }) = 0;
 	virtual chat_template_type_t get_chat_template_type() {

src/include/AutoModel/modeling_qwen2vl.hpp

@@ -46,7 +46,8 @@ class Qwen2VL : public AutoModel {
     qwen2vl_image_t load_image_base64(const std::string& base64_string);
 
 
-    int debug_count= 0;
+    int debug_count = 0;
+    int resize = 0;
     void smart_resize(
     int height, int width,
     int& h_bar,int& w_bar,
@@ -59,6 +60,7 @@ class Qwen2VL : public AutoModel {
 public:
     Qwen2VL(xrt::device* npu_device_inst);
 
+    void set_special_flags(int resize) override;
     void load_model(std::string model_path, json model_inf, int default_context_length = -1, bool enable_preemption = false) override;
     //void toggle_enable_think() override;
     bool insert(chat_meta_info_t& meta_info, lm_uniform_input_t& input) override;

src/include/AutoModel/modeling_qwen3vl.hpp

@@ -47,19 +47,22 @@ class Qwen3VL : public AutoModel {
     qwen3vl_image_t load_image_base64(const std::string& base64_string);
 
 
-    int debug_count= 0;
+    int debug_count = 0;
+    int resize = 0;
+
     void smart_resize(
     int height, int width,
     int& h_bar,int& w_bar,
     int factor,
     int min_pixels,
     int max_pixels);
 
-    void preprocess_image(qwen3vl_image_t& image,  std::vector<bf16> &pixel_values);
+    void preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_values);
 
 public:
     Qwen3VL(xrt::device* npu_device_inst);
 
+    void set_special_flags(int resize) override;
     void load_model(std::string model_path, json model_inf, int default_context_length = -1, bool enable_preemption = false) override;
     //void toggle_enable_think() override;
     bool insert(chat_meta_info_t& meta_info, lm_uniform_input_t& input) override;

src/include/utils/vm_args.hpp

@@ -28,6 +28,7 @@ struct ParsedArgs {
     bool quiet_list;
     bool preemption;
     int ctx_length;
+    int resize;
     size_t max_socket_connections;
     size_t max_npu_queue;
     int port;
@@ -70,6 +71,8 @@ bool parse_options(int argc, char *argv[], ParsedArgs& parsed_args) {
             ("quiet", "Hide emojis in the model list (can be used with --filter)")
             ("ctx-len,c", po::value<int>(&parsed_args.ctx_length)->default_value(-1),
              "Set context length")
+            ("resize,r", po::value<int>(&parsed_args.resize)->default_value(0),
+             "Pre-resize the image")
             ("socket,s", po::value<size_t>(&parsed_args.max_socket_connections)->default_value(10),
             "Set the maximum number of socket connections allowed (for serve command)")
             ("q-len,q", po::value<size_t>(&parsed_args.max_npu_queue)->default_value(10),
@@ -132,6 +135,7 @@ bool parse_options(int argc, char *argv[], ParsedArgs& parsed_args) {
             std::cout << "  " << argv[0] << " serve llama3.2:1b --cors 0" << std::endl;
             std::cout << "  " << argv[0] << " serve llama3.2:1b --asr 1" << std::endl;
             std::cout << "  " << argv[0] << " serve llama3.2:1b --embed 1" << std::endl;
+            std::cout << "  " << argv[0] << " serve qwen3vl-it:4b --resize 0" << std::endl;
             std::cout << "  " << argv[0] << " list" << std::endl;
             std::cout << "  " << argv[0] << " list --quiet" << std::endl;
             std::cout << "  " << argv[0] << " list --filter installed" << std::endl;
@@ -176,6 +180,7 @@ bool parse_options(int argc, char *argv[], ParsedArgs& parsed_args) {
                 std::cout << "  " << argv[0] << " serve llama3.2:1b --cors 0" << std::endl;
                 std::cout << "  " << argv[0] << " serve llama3.2:1b --asr 1" << std::endl;
                 std::cout << "  " << argv[0] << " serve llama3.2:1b --embed 1" << std::endl;
+                std::cout << "  " << argv[0] << " serve qwen3vl-it:4b --resize 0" << std::endl;
                 std::cout << "  " << argv[0] << " list" << std::endl;
                 std::cout << "  " << argv[0] << " list --quiet" << std::endl;
                 std::cout << "  " << argv[0] << " list --filter installed" << std::endl;

src/runner/runner.cpp

@@ -36,7 +36,7 @@ std::map<std::string, runner_cmd_t> cmd_map = {
 /// \param supported_models - the list of supported models
 /// \param downloader - the downloader for the models
 /// \param tag - the tag of the model to load
-Runner::Runner(model_list& supported_models, ModelDownloader& downloader, std::string& tag, bool asr, bool embed, int ctx_length, bool preemption)
+Runner::Runner(model_list& supported_models, ModelDownloader& downloader, std::string& tag, bool asr, bool embed, int ctx_length, int resize, bool preemption)
     : supported_models(supported_models), downloader(downloader), tag(tag), asr(asr), embed(embed) {
 
     this->npu_device_inst = xrt::device(0);
@@ -83,6 +83,7 @@ Runner::Runner(model_list& supported_models, ModelDownloader& downloader, std::s
         this->downloader.pull_model(this->tag);
     }
     auto [new_tag, model_info] = this->supported_models.get_model_info(this->tag);
+    this->auto_chat_engine->set_special_flags(resize);
     this->auto_chat_engine->load_model(this->supported_models.get_model_path(new_tag), model_info, this->ctx_length, this->preemption);
 
     this->generate_limit = -1;

src/runner/runner.hpp

@@ -43,7 +43,7 @@ typedef enum {
 /// \brief Runner class
 class Runner {
     public: 
-        Runner(model_list& supported_models, ModelDownloader& downloader, std::string& tag, bool asr, bool embed, int ctx_length, bool preemption);
+        Runner(model_list& supported_models, ModelDownloader& downloader, std::string& tag, bool asr, bool embed, int ctx_length, int resize, bool preemption);
         void run();
     private:
         std::string tag;

src/server/rest_handler.cpp

@@ -146,8 +146,8 @@ static json normalize_template(json messages) {
 ///@param downloader the downloader
 ///@param default_tag the default tag
 ///@return the rest handler
-RestHandler::RestHandler(model_list& models, ModelDownloader& downloader, const std::string& default_tag, bool asr, bool embed, int ctx_length, bool preemption)
-    : supported_models(models), downloader(downloader), default_model_tag(default_tag), current_model_tag(""), asr(asr), embed(embed), preemption(preemption){
+RestHandler::RestHandler(model_list& models, ModelDownloader& downloader, const std::string& default_tag, bool asr, bool embed, int ctx_length, int resize, bool preemption)
+    : supported_models(models), downloader(downloader), default_model_tag(default_tag), current_model_tag(""), asr(asr), embed(embed), resize(resize), preemption(preemption){
     this->npu_device_inst = xrt::device(0);
 
     if (ctx_length != -1) {
@@ -207,6 +207,7 @@ void RestHandler::ensure_model_loaded(const std::string& model_tag) {
             downloader.pull_model(ensure_tag);
         }
         auto [new_ensure_tag, model_info] = supported_models.get_model_info(ensure_tag);
+        auto_chat_engine->set_special_flags(resize);
         auto_chat_engine->load_model(supported_models.get_model_path(new_ensure_tag), model_info, ctx_length, preemption);
         current_model_tag = ensure_tag;
     }

src/server/rest_handler.hpp

@@ -31,7 +31,7 @@ using StreamResponseCallback = std::function<void(const json&, bool)>; // data,
 
 class RestHandler {
 public:
-    RestHandler(model_list& models, ModelDownloader& downloader, const std::string& default_tag, bool asr, bool embed, int ctx_length = -1, bool preemption = false);
+    RestHandler(model_list& models, ModelDownloader& downloader, const std::string& default_tag, bool asr, bool embed, int ctx_length = -1, int resize = 0, bool preemption = false);
     ~RestHandler();
 
     void handle_show(const json& request,
@@ -127,6 +127,7 @@ class RestHandler {
     int generate_context_id;
     int chat_context_id;
     int ctx_length;
+    int resize;
     std::string last_question;
     bool preemption;
     PromptCache prompt_cache;

src/server/server.cpp

@@ -741,9 +741,9 @@ bool WebServer::handle_request(http::request<http::string_body>& req,
 ///@param default_tag the default tag
 ///@param port the port
 ///@return the server
-std::unique_ptr<WebServer> create_lm_server(model_list& models, ModelDownloader& downloader, const std::string& default_tag, bool asr, bool embed, std::string host, int port, int ctx_length, bool cors, bool preemption) {
+std::unique_ptr<WebServer> create_lm_server(model_list& models, ModelDownloader& downloader, const std::string& default_tag, bool asr, bool embed, std::string host, int port, int ctx_length, int resize, bool cors, bool preemption) {
     auto server = std::make_unique<WebServer>(host, port, cors);
-    auto rest_handler = std::make_shared<RestHandler>(models, downloader, default_tag, asr, embed, ctx_length);
+    auto rest_handler = std::make_shared<RestHandler>(models, downloader, default_tag, asr, embed, ctx_length, resize);
 
     // Register Ollama-compatible routes
     server->register_handler("POST", "/api/show",

src/src/main.cpp

@@ -146,7 +146,7 @@ void handle_user_input() {
 ///@param default_tag the default tag
 ///@param port the port to listen on, default is 52625, same with the ollama server
 ///@return the server
-std::unique_ptr<WebServer> create_lm_server(model_list& models, ModelDownloader& downloader, const std::string& default_tag, bool asr, bool embed, std::string host, int port, int ctx_length, bool cors, bool preemption);
+std::unique_ptr<WebServer> create_lm_server(model_list& models, ModelDownloader& downloader, const std::string& default_tag, bool asr, bool embed, std::string host, int port, int ctx_length, int resize, bool cors, bool preemption);
 
 
 ///@brief get_server_port gets the server port from environment variable FLM_SERVE_PORT
@@ -264,6 +264,7 @@ int main(int argc, char* argv[]) {
     std::string power_mode = parsed_args.power_mode;
     bool got_power_mode = (power_mode != "performance"); // Check if user explicitly set power mode
     int ctx_length = parsed_args.ctx_length;
+    int resize = parsed_args.resize;
     bool preemption = parsed_args.preemption;
     size_t max_socket_connections = parsed_args.max_socket_connections;
     size_t max_npu_queue = parsed_args.max_npu_queue;
@@ -332,14 +333,14 @@ int main(int argc, char* argv[]) {
         }
         else if (command == "run") {
             check_and_notify_new_version();
-            Runner runner(availble_models, downloader, tag, asr, embed, ctx_length, preemption);
+            Runner runner(availble_models, downloader, tag, asr, embed, ctx_length, resize, preemption);
             runner.run();
 
         } else if (command == "serve") {
             check_and_notify_new_version();
             // Create the server
             int port = utils::get_server_port(user_port);
-            auto server = create_lm_server(availble_models, downloader, tag, asr, embed, user_host, port, ctx_length, cors, preemption);
+            auto server = create_lm_server(availble_models, downloader, tag, asr, embed, user_host, port, ctx_length, resize, cors, preemption);
             server->set_max_connections(max_socket_connections);           // Allow up to 10 concurrent connections
             server->set_io_threads(10);          // Allow up to 5 io threads
             server->set_npu_queue_length(max_npu_queue);           // Allow up to 10 concurrent queue