From a7fc352669cfca14365b698b5d25e4b6f54dc7a8 Mon Sep 17 00:00:00 2001
From: ZaneNi <zhihenn@clemson.edu>
Date: Tue, 17 Feb 2026 13:56:08 -0500
Subject: [PATCH 1/4] feat: image resize for qwen2/3 vl

---
 src/common/AutoModel/modeling_qwen3vl.cpp     |  6 ++-
 .../AutoModel/modeling_qwen3vl_image.cpp      | 37 +++++++++++++++++--
 src/include/AutoModel/automodel.hpp           |  3 +-
 src/include/AutoModel/modeling_qwen3vl.hpp    |  5 ++-
 src/include/utils/vm_args.hpp                 |  5 +++
 src/server/rest_handler.cpp                   |  5 ++-
 src/server/rest_handler.hpp                   |  3 +-
 src/server/server.cpp                         |  4 +-
 src/src/main.cpp                              |  5 ++-
 9 files changed, 59 insertions(+), 14 deletions(-)

diff --git a/src/common/AutoModel/modeling_qwen3vl.cpp b/src/common/AutoModel/modeling_qwen3vl.cpp
index e098d36f..57d2c554 100644
--- a/src/common/AutoModel/modeling_qwen3vl.cpp
+++ b/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);
     
@@ -122,7 +126,7 @@ bool Qwen3VL::insert(chat_meta_info_t& meta_info, lm_uniform_input_t& input) {
                         total_images++;
                     }
                     qwen3vl_image_t image = this->load_image_base64(img_str);
-                    preprocess_image(image, image_payload._data__processed);
+                    preprocess_image(image, image_payload._data__processed, resize);
                     image_payload.images.push_back(image);
                     image_payload.num_images++;
                 }
diff --git a/src/common/AutoModel/modeling_qwen3vl_image.cpp b/src/common/AutoModel/modeling_qwen3vl_image.cpp
index c7cdf86f..d3b4590a 100644
--- a/src/common/AutoModel/modeling_qwen3vl_image.cpp
+++ b/src/common/AutoModel/modeling_qwen3vl_image.cpp
@@ -508,21 +508,50 @@ void Qwen3VL::smart_resize(
 ///@note: Converts uint8 image to BF16 format, data is already in (3, H, W) CHW layout
 ///@param: image: the image to preprocess (already in CHW format)
 ///@return: the preprocessed image in BF16 format
-void Qwen3VL::preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_values) {
+void Qwen3VL::preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_values, int resize) {
     const int width = image.width;
     const int height = image.height;
     const int channels = 3; // RGB
     int resized_height; 
-    int resized_width;    
+    int resized_width;   
+    int aresized_height;
+    int aresized_width;
+
+    std::cout << "height "<< height << " width " << width <<std::endl;
+
+    // do same ratio resize here
+    if (resize == -1) {
+        // the longer side turn to 1080
+        // the other side the same radio decrese
+        const float target_longest_edge = 1080.0f;
+        float scale = target_longest_edge / std::max(height, width);
+
+        // Apply scale
+        int raw_new_h = static_cast<int>(height * scale);
+        int raw_new_w = static_cast<int>(width * scale);
+
+        // CRITICAL: Align to QWEN3_PATCH_SIZE. 
+        // If we don't align, (resized_h / patch_size) will truncate, losing data.
+        // We round to the nearest multiple of the patch size.
+        aresized_height = ((raw_new_h + QWEN3_PATCH_SIZE / 2) / QWEN3_PATCH_SIZE) * QWEN3_PATCH_SIZE;
+        aresized_width = ((raw_new_w + QWEN3_PATCH_SIZE / 2) / QWEN3_PATCH_SIZE) * QWEN3_PATCH_SIZE;
+
+        // Ensure we explicitly enforce the minimum size (1 patch) to avoid 0 dimension
+        aresized_height = std::max(resized_height, (int)QWEN3_PATCH_SIZE);
+        aresized_width = std::max(resized_width, (int)QWEN3_PATCH_SIZE);
+
+    }
+
     // do the automatically resizing in here 
     smart_resize(
-        height, width,
+        aresized_height, aresized_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;
diff --git a/src/include/AutoModel/automodel.hpp b/src/include/AutoModel/automodel.hpp
index e28c2745..d2565d3d 100644
--- a/src/include/AutoModel/automodel.hpp
+++ b/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() {
diff --git a/src/include/AutoModel/modeling_qwen3vl.hpp b/src/include/AutoModel/modeling_qwen3vl.hpp
index 22fc5049..4f0f33c8 100644
--- a/src/include/AutoModel/modeling_qwen3vl.hpp
+++ b/src/include/AutoModel/modeling_qwen3vl.hpp
@@ -48,6 +48,8 @@ class Qwen3VL : public AutoModel {
     
 
     int debug_count= 0;
+    int resize = -1;
+
     void smart_resize(
     int height, int width,
     int& h_bar,int& w_bar,
@@ -55,11 +57,12 @@ class Qwen3VL : public AutoModel {
     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, int resize = -1);
 
 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;
diff --git a/src/include/utils/vm_args.hpp b/src/include/utils/vm_args.hpp
index 22ff4e78..05ba20f9 100644
--- a/src/include/utils/vm_args.hpp
+++ b/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(-1),
+             "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;
diff --git a/src/server/rest_handler.cpp b/src/server/rest_handler.cpp
index adc37ea5..49d63b7c 100644
--- a/src/server/rest_handler.cpp
+++ b/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;
     }
diff --git a/src/server/rest_handler.hpp b/src/server/rest_handler.hpp
index 0b212d5a..3fffdef6 100644
--- a/src/server/rest_handler.hpp
+++ b/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 = -1, 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;
diff --git a/src/server/server.cpp b/src/server/server.cpp
index 8b3f5e07..6cda9601 100644
--- a/src/server/server.cpp
+++ b/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",
diff --git a/src/src/main.cpp b/src/src/main.cpp
index f061a5a5..91e7f949 100644
--- a/src/src/main.cpp
+++ b/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;
@@ -339,7 +340,7 @@ int main(int argc, char* argv[]) {
             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

From 687758741dcb5253415918449c262b8a9037129a Mon Sep 17 00:00:00 2001
From: ZaneNi <zhihenn@clemson.edu>
Date: Tue, 17 Feb 2026 15:44:41 -0500
Subject: [PATCH 2/4] feat: image resize for qwen2/3 vl

---
 src/common/AutoModel/modeling_qwen2vl.cpp     |  4 ++
 .../AutoModel/modeling_qwen2vl_image.cpp      | 23 ++++++++++-
 src/common/AutoModel/modeling_qwen3vl.cpp     |  2 +-
 .../AutoModel/modeling_qwen3vl_image.cpp      | 38 +++++++------------
 src/include/AutoModel/modeling_qwen2vl.hpp    |  2 +
 src/include/AutoModel/modeling_qwen3vl.hpp    |  2 +-
 src/include/utils/vm_args.hpp                 |  2 +-
 src/server/rest_handler.hpp                   |  2 +-
 8 files changed, 45 insertions(+), 30 deletions(-)

diff --git a/src/common/AutoModel/modeling_qwen2vl.cpp b/src/common/AutoModel/modeling_qwen2vl.cpp
index 13181608..c4ae9c84 100644
--- a/src/common/AutoModel/modeling_qwen2vl.cpp
+++ b/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);
     
diff --git a/src/common/AutoModel/modeling_qwen2vl_image.cpp b/src/common/AutoModel/modeling_qwen2vl_image.cpp
index 1beb88c2..8470cece 100644
--- a/src/common/AutoModel/modeling_qwen2vl_image.cpp
+++ b/src/common/AutoModel/modeling_qwen2vl_image.cpp
@@ -512,17 +512,36 @@ 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 == 1) ? 1080 : (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(
diff --git a/src/common/AutoModel/modeling_qwen3vl.cpp b/src/common/AutoModel/modeling_qwen3vl.cpp
index 57d2c554..5b25e778 100644
--- a/src/common/AutoModel/modeling_qwen3vl.cpp
+++ b/src/common/AutoModel/modeling_qwen3vl.cpp
@@ -126,7 +126,7 @@ bool Qwen3VL::insert(chat_meta_info_t& meta_info, lm_uniform_input_t& input) {
                         total_images++;
                     }
                     qwen3vl_image_t image = this->load_image_base64(img_str);
-                    preprocess_image(image, image_payload._data__processed, resize);
+                    preprocess_image(image, image_payload._data__processed);
                     image_payload.images.push_back(image);
                     image_payload.num_images++;
                 }
diff --git a/src/common/AutoModel/modeling_qwen3vl_image.cpp b/src/common/AutoModel/modeling_qwen3vl_image.cpp
index d3b4590a..5adc28bf 100644
--- a/src/common/AutoModel/modeling_qwen3vl_image.cpp
+++ b/src/common/AutoModel/modeling_qwen3vl_image.cpp
@@ -508,43 +508,33 @@ void Qwen3VL::smart_resize(
 ///@note: Converts uint8 image to BF16 format, data is already in (3, H, W) CHW layout
 ///@param: image: the image to preprocess (already in CHW format)
 ///@return: the preprocessed image in BF16 format
-void Qwen3VL::preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_values, int resize) {
+void Qwen3VL::preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_values) {
     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 aresized_height;
-    int aresized_width;
 
     std::cout << "height "<< height << " width " << width <<std::endl;
 
-    // do same ratio resize here
-    if (resize == -1) {
-        // the longer side turn to 1080
-        // the other side the same radio decrese
-        const float target_longest_edge = 1080.0f;
-        float scale = target_longest_edge / std::max(height, width);
-
-        // Apply scale
-        int raw_new_h = static_cast<int>(height * scale);
-        int raw_new_w = static_cast<int>(width * scale);
-
-        // CRITICAL: Align to QWEN3_PATCH_SIZE. 
-        // If we don't align, (resized_h / patch_size) will truncate, losing data.
-        // We round to the nearest multiple of the patch size.
-        aresized_height = ((raw_new_h + QWEN3_PATCH_SIZE / 2) / QWEN3_PATCH_SIZE) * QWEN3_PATCH_SIZE;
-        aresized_width = ((raw_new_w + QWEN3_PATCH_SIZE / 2) / QWEN3_PATCH_SIZE) * QWEN3_PATCH_SIZE;
-
-        // Ensure we explicitly enforce the minimum size (1 patch) to avoid 0 dimension
-        aresized_height = std::max(resized_height, (int)QWEN3_PATCH_SIZE);
-        aresized_width = std::max(resized_width, (int)QWEN3_PATCH_SIZE);
+    int target_longest_edge = (resize == 1) ? 1080 : (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(
-        aresized_height, aresized_width,
+        downscaled_height, downscaled_width,
         resized_height, resized_width,
         QWEN3_PATCH_SIZE * QWEN3_IMAGE_MERGE_SIZE,
         QWEN3_SHORTEST_EDGE,
diff --git a/src/include/AutoModel/modeling_qwen2vl.hpp b/src/include/AutoModel/modeling_qwen2vl.hpp
index d567e4ba..d4589fa6 100644
--- a/src/include/AutoModel/modeling_qwen2vl.hpp
+++ b/src/include/AutoModel/modeling_qwen2vl.hpp
@@ -47,6 +47,7 @@ class Qwen2VL : public AutoModel {
     
 
     int debug_count= 0;
+    int resize;
     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;
diff --git a/src/include/AutoModel/modeling_qwen3vl.hpp b/src/include/AutoModel/modeling_qwen3vl.hpp
index 4f0f33c8..9f6e7ade 100644
--- a/src/include/AutoModel/modeling_qwen3vl.hpp
+++ b/src/include/AutoModel/modeling_qwen3vl.hpp
@@ -57,7 +57,7 @@ class Qwen3VL : public AutoModel {
     int min_pixels,
     int max_pixels);
     
-    void preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_values, int resize = -1);
+    void preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_values);
 
 public:
     Qwen3VL(xrt::device* npu_device_inst);
diff --git a/src/include/utils/vm_args.hpp b/src/include/utils/vm_args.hpp
index 05ba20f9..4c2a71a3 100644
--- a/src/include/utils/vm_args.hpp
+++ b/src/include/utils/vm_args.hpp
@@ -71,7 +71,7 @@ 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(-1),
+            ("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)")
diff --git a/src/server/rest_handler.hpp b/src/server/rest_handler.hpp
index 3fffdef6..47c16b9e 100644
--- a/src/server/rest_handler.hpp
+++ b/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, int resize = -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,

From 748b0a963636762d9721a25e51dfe64ed6faaac3 Mon Sep 17 00:00:00 2001
From: ZaneNi <zhihenn@clemson.edu>
Date: Tue, 17 Feb 2026 15:46:48 -0500
Subject: [PATCH 3/4] hide debug info

---
 src/common/AutoModel/modeling_qwen2vl_image.cpp | 5 ++---
 src/common/AutoModel/modeling_qwen3vl_image.cpp | 4 ++--
 2 files changed, 4 insertions(+), 5 deletions(-)

diff --git a/src/common/AutoModel/modeling_qwen2vl_image.cpp b/src/common/AutoModel/modeling_qwen2vl_image.cpp
index 8470cece..5f1b0844 100644
--- a/src/common/AutoModel/modeling_qwen2vl_image.cpp
+++ b/src/common/AutoModel/modeling_qwen2vl_image.cpp
@@ -517,8 +517,7 @@ void Qwen2VL::preprocess_image(qwen2vl_image_t& image, std::vector<bf16> &pixel_
     int resized_height; 
     int resized_width;    
 
-    std::cout << "height " << height << " width " << width << std::endl;
-
+    //std::cout << "height " << height << " width " << width << std::endl;
 
     int target_longest_edge = (resize == 1) ? 1080 : (resize == 2) ? 2560 : 0;
 
@@ -541,7 +540,7 @@ void Qwen2VL::preprocess_image(qwen2vl_image_t& image, std::vector<bf16> &pixel_
         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(
diff --git a/src/common/AutoModel/modeling_qwen3vl_image.cpp b/src/common/AutoModel/modeling_qwen3vl_image.cpp
index 5adc28bf..1e7980e5 100644
--- a/src/common/AutoModel/modeling_qwen3vl_image.cpp
+++ b/src/common/AutoModel/modeling_qwen3vl_image.cpp
@@ -517,7 +517,7 @@ void Qwen3VL::preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_
     int resized_height; 
     int resized_width;   
 
-    std::cout << "height "<< height << " width " << width <<std::endl;
+    //std::cout << "height "<< height << " width " << width <<std::endl;
 
     int target_longest_edge = (resize == 1) ? 1080 : (resize == 2) ? 2560 : 0;
 
@@ -541,7 +541,7 @@ void Qwen3VL::preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_
         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;

From fbee4d70bd60ee0fac06bf86f97d6d20d2527ea5 Mon Sep 17 00:00:00 2001
From: ZaneNi <zhihenn@clemson.edu>
Date: Wed, 18 Feb 2026 10:48:52 -0500
Subject: [PATCH 4/4] feat: improve resize

---
 src/common/AutoModel/modeling_qwen2vl_image.cpp | 2 +-
 src/common/AutoModel/modeling_qwen3vl_image.cpp | 2 +-
 src/include/AutoModel/modeling_qwen2vl.hpp      | 4 ++--
 src/include/AutoModel/modeling_qwen3vl.hpp      | 4 ++--
 src/runner/runner.cpp                           | 3 ++-
 src/runner/runner.hpp                           | 2 +-
 src/src/main.cpp                                | 2 +-
 7 files changed, 10 insertions(+), 9 deletions(-)

diff --git a/src/common/AutoModel/modeling_qwen2vl_image.cpp b/src/common/AutoModel/modeling_qwen2vl_image.cpp
index 5f1b0844..bd34eb84 100644
--- a/src/common/AutoModel/modeling_qwen2vl_image.cpp
+++ b/src/common/AutoModel/modeling_qwen2vl_image.cpp
@@ -519,7 +519,7 @@ void Qwen2VL::preprocess_image(qwen2vl_image_t& image, std::vector<bf16> &pixel_
 
     //std::cout << "height " << height << " width " << width << std::endl;
 
-    int target_longest_edge = (resize == 1) ? 1080 : (resize == 2) ? 2560 : 0;
+    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));
diff --git a/src/common/AutoModel/modeling_qwen3vl_image.cpp b/src/common/AutoModel/modeling_qwen3vl_image.cpp
index 1e7980e5..5e0943e0 100644
--- a/src/common/AutoModel/modeling_qwen3vl_image.cpp
+++ b/src/common/AutoModel/modeling_qwen3vl_image.cpp
@@ -519,7 +519,7 @@ void Qwen3VL::preprocess_image(qwen3vl_image_t& image, std::vector<bf16> &pixel_
 
     //std::cout << "height "<< height << " width " << width <<std::endl;
 
-    int target_longest_edge = (resize == 1) ? 1080 : (resize == 2) ? 2560 : 0;
+    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));
diff --git a/src/include/AutoModel/modeling_qwen2vl.hpp b/src/include/AutoModel/modeling_qwen2vl.hpp
index d4589fa6..5e0bc649 100644
--- a/src/include/AutoModel/modeling_qwen2vl.hpp
+++ b/src/include/AutoModel/modeling_qwen2vl.hpp
@@ -46,8 +46,8 @@ class Qwen2VL : public AutoModel {
     qwen2vl_image_t load_image_base64(const std::string& base64_string);
     
 
-    int debug_count= 0;
-    int resize;
+    int debug_count = 0;
+    int resize = 0;
     void smart_resize(
     int height, int width,
     int& h_bar,int& w_bar,
diff --git a/src/include/AutoModel/modeling_qwen3vl.hpp b/src/include/AutoModel/modeling_qwen3vl.hpp
index 9f6e7ade..ab88caca 100644
--- a/src/include/AutoModel/modeling_qwen3vl.hpp
+++ b/src/include/AutoModel/modeling_qwen3vl.hpp
@@ -47,8 +47,8 @@ class Qwen3VL : public AutoModel {
     qwen3vl_image_t load_image_base64(const std::string& base64_string);
     
 
-    int debug_count= 0;
-    int resize = -1;
+    int debug_count = 0;
+    int resize = 0;
 
     void smart_resize(
     int height, int width,
diff --git a/src/runner/runner.cpp b/src/runner/runner.cpp
index 4532de1c..4e6cc9ef 100644
--- a/src/runner/runner.cpp
+++ b/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;
diff --git a/src/runner/runner.hpp b/src/runner/runner.hpp
index 93813166..b61b6199 100644
--- a/src/runner/runner.hpp
+++ b/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;
diff --git a/src/src/main.cpp b/src/src/main.cpp
index 91e7f949..e658699a 100644
--- a/src/src/main.cpp
+++ b/src/src/main.cpp
@@ -333,7 +333,7 @@ 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") {