feat: add Qwen Image support (#851)

* add qwen tokenizer

* add qwen2.5 vl support

* mv qwen.hpp -> qwenvl.hpp

* add qwen image model

* add qwen image t2i pipeline

* fix qwen image flash attn

* add qwen image i2i pipline

* change encoding of vocab_qwen.hpp to utf8

* fix get_first_stage_encoding

* apply jeffbolz f32 patch

https://github.com/leejet/stable-diffusion.cpp/pull/851#issuecomment-3335515302

* fix the issue that occurs when using CUDA with k-quants weights

* optimize the handling of the FeedForward precision fix

* to_add_out precision fix

* update docs

README.md

@@ -21,6 +21,7 @@ API and command-line option may change frequently.***
     - [SD3/SD3.5](./docs/sd3.md)
     - [Flux-dev/Flux-schnell](./docs/flux.md)
     - [Chroma](./docs/chroma.md)
+    - [Qwen Image](./docs/qwen_image.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
   - Video Models
@@ -296,6 +297,7 @@ arguments:
   --clip_g                           path to the clip-g text encoder
   --clip_vision                      path to the clip-vision encoder
   --t5xxl                            path to the t5xxl text encoder
+  --qwen2vl                          path to the qwen2vl text encoder
   --vae [VAE]                        path to vae
   --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
   --control-net [CONTROL_PATH]       path to control net model
@@ -464,6 +466,7 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 ## References
 
 - [ggml](https://github.com/ggerganov/ggml)
+- [diffusers](https://github.com/huggingface/diffusers)
 - [stable-diffusion](https://github.com/CompVis/stable-diffusion)
 - [sd3-ref](https://github.com/Stability-AI/sd3-ref)
 - [stable-diffusion-stability-ai](https://github.com/Stability-AI/stablediffusion)

clip.hpp

@@ -6,7 +6,7 @@
 
 /*================================================== CLIPTokenizer ===================================================*/
 
-std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remove_lora(std::string text) {
+__STATIC_INLINE__ std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remove_lora(std::string text) {
     std::regex re("<lora:([^:]+):([^>]+)>");
     std::smatch matches;
     std::unordered_map<std::string, float> filename2multiplier;
@@ -31,7 +31,7 @@ std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remov
     return std::make_pair(filename2multiplier, text);
 }
 
-std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
+__STATIC_INLINE__ std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
     std::vector<std::pair<int, std::u32string>> byte_unicode_pairs;
     std::set<int> byte_set;
     for (int b = static_cast<int>('!'); b <= static_cast<int>('~'); ++b) {

common.hpp

@@ -177,7 +177,7 @@ public:
     }
 };
 
-class GEGLU : public GGMLBlock {
+class GEGLU : public UnaryBlock {
 protected:
     int64_t dim_in;
     int64_t dim_out;
@@ -216,23 +216,57 @@ public:
     }
 };
 
+class GELU : public UnaryBlock {
+public:
+    GELU(int64_t dim_in, int64_t dim_out, bool bias = true) {
+        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out, bias));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        // x: [ne3, ne2, ne1, dim_in]
+        // return: [ne3, ne2, ne1, dim_out]
+        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+
+        x = proj->forward(ctx, x);
+        x = ggml_gelu_inplace(ctx, x);
+        return x;
+    }
+};
+
 class FeedForward : public GGMLBlock {
 public:
+    enum class Activation {
+        GEGLU,
+        GELU
+    };
     FeedForward(int64_t dim,
                 int64_t dim_out,
-                int64_t mult = 4) {
+                int64_t mult          = 4,
+                Activation activation = Activation::GEGLU,
+                bool precision_fix    = false) {
         int64_t inner_dim = dim * mult;
+        if (activation == Activation::GELU) {
+            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GELU(dim, inner_dim));
+        } else {
+            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
+        }
 
-        blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
         // net_1 is nn.Dropout(), skip for inference
-        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out));
+        float scale = 1.f;
+        if (precision_fix) {
+            scale = 1.f / 128.f;
+        }
+        // The purpose of the scale here is to prevent NaN issues in certain situations.
+        // For example, when using Vulkan without enabling force_prec_f32,
+        // or when using CUDA but the weights are k-quants.
+        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, false, scale));
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
         // x: [ne3, ne2, ne1, dim]
         // return: [ne3, ne2, ne1, dim_out]
 
-        auto net_0 = std::dynamic_pointer_cast<GEGLU>(blocks["net.0"]);
+        auto net_0 = std::dynamic_pointer_cast<UnaryBlock>(blocks["net.0"]);
         auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
 
         x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]

conditioner.hpp

@@ -2,6 +2,7 @@
 #define __CONDITIONER_HPP__
 
 #include "clip.hpp"
+#include "qwenvl.hpp"
 #include "t5.hpp"
 
 struct SDCondition {
@@ -22,11 +23,11 @@ struct Conditioner {
                                               int width,
                                               int height,
                                               int adm_in_channels  = -1,
-                                              bool zero_out_masked = false)                                             = 0;
+                                              bool zero_out_masked = false)             = 0;
-    virtual void alloc_params_buffer()                                                                                  = 0;
+    virtual void alloc_params_buffer()                                                  = 0;
-    virtual void free_params_buffer()                                                                                   = 0;
+    virtual void free_params_buffer()                                                   = 0;
-    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors)                                 = 0;
+    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) = 0;
-    virtual size_t get_params_buffer_size()                                                                             = 0;
+    virtual size_t get_params_buffer_size()                                             = 0;
     virtual std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
                                                                                           int n_threads,
                                                                                           const std::string& text,
@@ -35,9 +36,13 @@ struct Conditioner {
                                                                                           int height,
                                                                                           int num_input_imgs,
                                                                                           int adm_in_channels  = -1,
-                                                                                          bool zero_out_masked = false) = 0;
+                                                                                          bool zero_out_masked = false) {
+        GGML_ABORT("Not implemented yet!");
+    }
     virtual std::string remove_trigger_from_prompt(ggml_context* work_ctx,
-                                                   const std::string& prompt)                                           = 0;
+                                                   const std::string& prompt) {
+        GGML_ABORT("Not implemented yet!");
+    }
 };
 
 // ldm.modules.encoders.modules.FrozenCLIPEmbedder
@@ -978,23 +983,6 @@ struct SD3CLIPEmbedder : public Conditioner {
         auto tokens_and_weights = tokenize(text, 77, true);
         return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, zero_out_masked);
     }
-
-    std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
-                                                                                  int n_threads,
-                                                                                  const std::string& text,
-                                                                                  int clip_skip,
-                                                                                  int width,
-                                                                                  int height,
-                                                                                  int num_input_imgs,
-                                                                                  int adm_in_channels  = -1,
-                                                                                  bool zero_out_masked = false) {
-        GGML_ASSERT(0 && "Not implemented yet!");
-    }
-
-    std::string remove_trigger_from_prompt(ggml_context* work_ctx,
-                                           const std::string& prompt) {
-        GGML_ASSERT(0 && "Not implemented yet!");
-    }
 };
 
 struct FluxCLIPEmbedder : public Conditioner {
@@ -1195,23 +1183,6 @@ struct FluxCLIPEmbedder : public Conditioner {
         auto tokens_and_weights = tokenize(text, chunk_len, true);
         return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, zero_out_masked);
     }
-
-    std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
-                                                                                  int n_threads,
-                                                                                  const std::string& text,
-                                                                                  int clip_skip,
-                                                                                  int width,
-                                                                                  int height,
-                                                                                  int num_input_imgs,
-                                                                                  int adm_in_channels  = -1,
-                                                                                  bool zero_out_masked = false) {
-        GGML_ASSERT(0 && "Not implemented yet!");
-    }
-
-    std::string remove_trigger_from_prompt(ggml_context* work_ctx,
-                                           const std::string& prompt) {
-        GGML_ASSERT(0 && "Not implemented yet!");
-    }
 };
 
 struct T5CLIPEmbedder : public Conditioner {
@@ -1398,22 +1369,135 @@ struct T5CLIPEmbedder : public Conditioner {
         auto tokens_and_weights = tokenize(text, chunk_len, true);
         return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, zero_out_masked);
     }
+};
 
-    std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
+struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
-                                                                                  int n_threads,
+    Qwen::Qwen2Tokenizer tokenizer;
-                                                                                  const std::string& text,
+    std::shared_ptr<Qwen::Qwen2_5_VLRunner> qwenvl;
-                                                                                  int clip_skip,
+    int prompt_template_encode_start_idx = 34;
-                                                                                  int width,
+
-                                                                                  int height,
+    Qwen2_5_VLCLIPEmbedder(ggml_backend_t backend,
-                                                                                  int num_input_imgs,
+                           bool offload_params_to_cpu,
-                                                                                  int adm_in_channels  = -1,
+                           const String2GGMLType& tensor_types = {},
-                                                                                  bool zero_out_masked = false) {
+                           const std::string prefix            = "") {
-        GGML_ASSERT(0 && "Not implemented yet!");
+        qwenvl = std::make_shared<Qwen::Qwen2_5_VLRunner>(backend, offload_params_to_cpu, tensor_types, "text_encoders.qwen2vl");
     }
 
-    std::string remove_trigger_from_prompt(ggml_context* work_ctx,
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
-                                           const std::string& prompt) {
+        qwenvl->get_param_tensors(tensors, "text_encoders.qwen2vl");
-        GGML_ASSERT(0 && "Not implemented yet!");
+    }
+
+    void alloc_params_buffer() {
+        qwenvl->alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        qwenvl->free_params_buffer();
+    }
+
+    size_t get_params_buffer_size() {
+        size_t buffer_size = 0;
+        buffer_size += qwenvl->get_params_buffer_size();
+        return buffer_size;
+    }
+
+    std::tuple<std::vector<int>, std::vector<float>> tokenize(std::string text,
+                                                              size_t max_length = 0,
+                                                              bool padding      = false) {
+        auto parsed_attention = parse_prompt_attention(text);
+
+        {
+            std::stringstream ss;
+            ss << "[";
+            for (const auto& item : parsed_attention) {
+                ss << "['" << item.first << "', " << item.second << "], ";
+            }
+            ss << "]";
+            LOG_DEBUG("parse '%s' to %s", text.c_str(), ss.str().c_str());
+        }
+
+        std::vector<int> tokens;
+        std::vector<float> weights;
+        for (const auto& item : parsed_attention) {
+            const std::string& curr_text = item.first;
+            float curr_weight            = item.second;
+            std::vector<int> curr_tokens = tokenizer.tokenize(curr_text, nullptr);
+            tokens.insert(tokens.end(), curr_tokens.begin(), curr_tokens.end());
+            weights.insert(weights.end(), curr_tokens.size(), curr_weight);
+        }
+
+        tokenizer.pad_tokens(tokens, weights, max_length, padding);
+
+        // for (int i = 0; i < tokens.size(); i++) {
+        //     std::cout << tokens[i] << ":" << weights[i] << ", ";
+        // }
+        // std::cout << std::endl;
+
+        return {tokens, weights};
+    }
+
+    SDCondition get_learned_condition_common(ggml_context* work_ctx,
+                                             int n_threads,
+                                             std::tuple<std::vector<int>, std::vector<float>> token_and_weights,
+                                             int clip_skip,
+                                             bool zero_out_masked = false) {
+        auto& tokens  = std::get<0>(token_and_weights);
+        auto& weights = std::get<1>(token_and_weights);
+
+        int64_t t0                        = ggml_time_ms();
+        struct ggml_tensor* hidden_states = NULL;  // [N, n_token, 3584]
+
+        auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
+
+        qwenvl->compute(n_threads,
+                        input_ids,
+                        &hidden_states,
+                        work_ctx);
+        {
+            auto tensor         = hidden_states;
+            float original_mean = ggml_tensor_mean(tensor);
+            for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                    for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                        float value = ggml_tensor_get_f32(tensor, i0, i1, i2);
+                        value *= weights[i1];
+                        ggml_tensor_set_f32(tensor, value, i0, i1, i2);
+                    }
+                }
+            }
+            float new_mean = ggml_tensor_mean(tensor);
+            ggml_tensor_scale(tensor, (original_mean / new_mean));
+        }
+
+        GGML_ASSERT(hidden_states->ne[1] > prompt_template_encode_start_idx);
+
+        ggml_tensor* new_hidden_states = ggml_new_tensor_3d(work_ctx,
+                                                            GGML_TYPE_F32,
+                                                            hidden_states->ne[0],
+                                                            hidden_states->ne[1] - prompt_template_encode_start_idx,
+                                                            hidden_states->ne[2]);
+
+        ggml_tensor_iter(new_hidden_states, [&](ggml_tensor* new_hidden_states, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+            float value = ggml_tensor_get_f32(hidden_states, i0, i1 + prompt_template_encode_start_idx, i2, i3);
+            ggml_tensor_set_f32(new_hidden_states, value, i0, i1, i2, i3);
+        });
+
+        int64_t t1 = ggml_time_ms();
+        LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
+        return SDCondition(new_hidden_states, nullptr, nullptr);
+    }
+
+    SDCondition get_learned_condition(ggml_context* work_ctx,
+                                      int n_threads,
+                                      const std::string& text,
+                                      int clip_skip,
+                                      int width,
+                                      int height,
+                                      int adm_in_channels  = -1,
+                                      bool zero_out_masked = false) {
+        std::string prompt      = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n" + text + "<|im_end|>\n<|im_start|>assistant\n";
+        auto tokens_and_weights = tokenize(prompt, 0, false);
+        return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, zero_out_masked);
     }
 };
 

diffusion_model.hpp

@@ -3,6 +3,7 @@
 
 #include "flux.hpp"
 #include "mmdit.hpp"
+#include "qwen_image.hpp"
 #include "unet.hpp"
 #include "wan.hpp"
 
@@ -263,4 +264,58 @@ struct WanModel : public DiffusionModel {
     }
 };
 
+struct QwenImageModel : public DiffusionModel {
+    std::string prefix;
+    Qwen::QwenImageRunner qwen_image;
+
+    QwenImageModel(ggml_backend_t backend,
+                   bool offload_params_to_cpu,
+                   const String2GGMLType& tensor_types = {},
+                   const std::string prefix            = "model.diffusion_model",
+                   SDVersion version                   = VERSION_QWEN_IMAGE,
+                   bool flash_attn                     = false)
+        : prefix(prefix), qwen_image(backend, offload_params_to_cpu, tensor_types, prefix, version, flash_attn) {
+    }
+
+    std::string get_desc() {
+        return qwen_image.get_desc();
+    }
+
+    void alloc_params_buffer() {
+        qwen_image.alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        qwen_image.free_params_buffer();
+    }
+
+    void free_compute_buffer() {
+        qwen_image.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        qwen_image.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() {
+        return qwen_image.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() {
+        return 768;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = NULL,
+                 struct ggml_context* output_ctx = NULL) {
+        return qwen_image.compute(n_threads,
+                                  diffusion_params.x,
+                                  diffusion_params.timesteps,
+                                  diffusion_params.context,
+                                  output,
+                                  output_ctx);
+    }
+};
+
 #endif

docs/qwen_image.md

@@ -0,0 +1,23 @@
+# How to Use
+
+## Download weights
+
+- Download Qwen Image
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/QuantStack/Qwen-Image-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
+- Download qwen_2.5_vl 7b
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-GGUF/tree/main
+
+## Examples
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --qwen2vl ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
+```
+
+<img alt="qwen example" src="../assets/qwen/example.png" />
+
+
+

examples/cli/main.cpp

@@ -61,6 +61,7 @@ struct SDParams {
     std::string clip_g_path;
     std::string clip_vision_path;
     std::string t5xxl_path;
+    std::string qwen2vl_path;
     std::string diffusion_model_path;
     std::string high_noise_diffusion_model_path;
     std::string vae_path;
@@ -146,6 +147,7 @@ void print_params(SDParams params) {
     printf("    clip_g_path:                       %s\n", params.clip_g_path.c_str());
     printf("    clip_vision_path:                  %s\n", params.clip_vision_path.c_str());
     printf("    t5xxl_path:                        %s\n", params.t5xxl_path.c_str());
+    printf("    qwen2vl_path:                      %s\n", params.qwen2vl_path.c_str());
     printf("    diffusion_model_path:              %s\n", params.diffusion_model_path.c_str());
     printf("    high_noise_diffusion_model_path:   %s\n", params.high_noise_diffusion_model_path.c_str());
     printf("    vae_path:                          %s\n", params.vae_path.c_str());
@@ -217,6 +219,7 @@ void print_usage(int argc, const char* argv[]) {
     printf("  --clip_g                           path to the clip-g text encoder\n");
     printf("  --clip_vision                      path to the clip-vision encoder\n");
     printf("  --t5xxl                            path to the t5xxl text encoder\n");
+    printf("  --qwen2vl                          path to the qwen2vl text encoder\n");
     printf("  --vae [VAE]                        path to vae\n");
     printf("  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)\n");
     printf("  --control-net [CONTROL_PATH]       path to control net model\n");
@@ -486,6 +489,7 @@ void parse_args(int argc, const char** argv, SDParams& params) {
         {"", "--clip_g", "", &params.clip_g_path},
         {"", "--clip_vision", "", &params.clip_vision_path},
         {"", "--t5xxl", "", &params.t5xxl_path},
+        {"", "--qwen2vl", "", &params.qwen2vl_path},
         {"", "--diffusion-model", "", &params.diffusion_model_path},
         {"", "--high-noise-diffusion-model", "", &params.high_noise_diffusion_model_path},
         {"", "--vae", "", &params.vae_path},
@@ -948,7 +952,7 @@ std::string get_image_params(SDParams params, int64_t seed) {
         parameter_string += " " + std::string(sd_schedule_name(params.sample_params.scheduler));
     }
     parameter_string += ", ";
-    for (const auto& te : {params.clip_l_path, params.clip_g_path, params.t5xxl_path}) {
+    for (const auto& te : {params.clip_l_path, params.clip_g_path, params.t5xxl_path, params.qwen2vl_path}) {
         if (!te.empty()) {
             parameter_string += "TE: " + sd_basename(te) + ", ";
         }
@@ -1331,6 +1335,7 @@ int main(int argc, const char* argv[]) {
         params.clip_g_path.c_str(),
         params.clip_vision_path.c_str(),
         params.t5xxl_path.c_str(),
+        params.qwen2vl_path.c_str(),
         params.diffusion_model_path.c_str(),
         params.high_noise_diffusion_model_path.c_str(),
         params.vae_path.c_str(),

flux.hpp

@@ -120,14 +120,15 @@ namespace Flux {
                                                     struct ggml_tensor* v,
                                                     struct ggml_tensor* pe,
                                                     struct ggml_tensor* mask,
-                                                    bool flash_attn) {
+                                                    bool flash_attn,
+                                                    float kv_scale = 1.0f) {
         // q,k,v: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2]
         // return: [N, L, n_head*d_head]
         q = apply_rope(ctx, q, pe);  // [N*n_head, L, d_head]
         k = apply_rope(ctx, k, pe);  // [N*n_head, L, d_head]
 
-        auto x = ggml_nn_attention_ext(ctx, backend, q, k, v, v->ne[1], mask, false, true, flash_attn);  // [N, L, n_head*d_head]
+        auto x = ggml_nn_attention_ext(ctx, backend, q, k, v, v->ne[1], mask, false, true, flash_attn, kv_scale);  // [N, L, n_head*d_head]
         return x;
     }
 

ggml_extend.hpp

@@ -56,6 +56,10 @@
 #define __STATIC_INLINE__ static inline
 #endif
 
+#ifndef SD_UNUSED
+#define SD_UNUSED(x) (void)(x)
+#endif
+
 __STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void*) {
     switch (level) {
         case GGML_LOG_LEVEL_DEBUG:
@@ -939,8 +943,19 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_group_norm_32(struct ggml_context* ct
 __STATIC_INLINE__ struct ggml_tensor* ggml_nn_linear(struct ggml_context* ctx,
                                                      struct ggml_tensor* x,
                                                      struct ggml_tensor* w,
-                                                     struct ggml_tensor* b) {
+                                                     struct ggml_tensor* b,
+                                                     bool force_prec_f32 = false,
+                                                     float scale         = 1.f) {
+    if (scale != 1.f) {
+        x = ggml_scale(ctx, x, scale);
+    }
     x = ggml_mul_mat(ctx, w, x);
+    if (force_prec_f32) {
+        ggml_mul_mat_set_prec(x, GGML_PREC_F32);
+    }
+    if (scale != 1.f) {
+        x = ggml_scale(ctx, x, 1.f / scale);
+    }
     if (b != NULL) {
         x = ggml_add_inplace(ctx, x, b);
     }
@@ -1125,9 +1140,9 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention(struct ggml_context* ctx
     return kqv;
 }
 
-// q: [N, L_q, C] or [N*n_head, L_q, d_head]
+// q: [N, L_q, C(n_head*d_head)] or [N*n_head, L_q, d_head]
-// k: [N, L_k, C] or [N*n_head, L_k, d_head]
+// k: [N, L_k, n_kv_head*d_head] or [N*n_kv_head, L_k, d_head]
-// v: [N, L_k, C] or [N, L_k, n_head, d_head]
+// v: [N, L_k, n_kv_head*d_head] or [N, L_k, n_kv_head, d_head]
 // mask: [N, L_q, L_k]
 // return: [N, L_q, C]
 __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context* ctx,
@@ -1139,33 +1154,38 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
                                                             struct ggml_tensor* mask = NULL,
                                                             bool diag_mask_inf       = false,
                                                             bool skip_reshape        = false,
-                                                            bool flash_attn          = false) {
+                                                            bool flash_attn          = false,  // avoid overflow
+                                                            float kv_scale           = 1.0f) {
     int64_t L_q;
     int64_t L_k;
     int64_t C;
     int64_t N;
     int64_t d_head;
+    int64_t n_kv_head;
     if (!skip_reshape) {
-        L_q    = q->ne[1];
+        L_q       = q->ne[1];
-        L_k    = k->ne[1];
+        L_k       = k->ne[1];
-        C      = q->ne[0];
+        C         = q->ne[0];
-        N      = q->ne[2];
+        N         = q->ne[2];
-        d_head = C / n_head;
+        d_head    = C / n_head;
-        q      = ggml_reshape_4d(ctx, q, d_head, n_head, L_q, N);      // [N, L_q, n_head, d_head]
+        n_kv_head = k->ne[0] / d_head;
-        q      = ggml_nn_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3));  // [N, n_head, L_q, d_head]
-        q      = ggml_reshape_3d(ctx, q, d_head, L_q, n_head * N);     // [N * n_head, L_q, d_head]
 
-        k = ggml_reshape_4d(ctx, k, d_head, n_head, L_k, N);      // [N, L_k, n_head, d_head]
+        q = ggml_reshape_4d(ctx, q, d_head, n_head, L_q, N);      // [N, L_q, n_head, d_head]
-        k = ggml_nn_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3));  // [N, n_head, L_k, d_head]
+        q = ggml_nn_cont(ctx, ggml_permute(ctx, q, 0, 2, 1, 3));  // [N, n_head, L_q, d_head]
-        k = ggml_reshape_3d(ctx, k, d_head, L_k, n_head * N);     // [N * n_head, L_k, d_head]
+        q = ggml_reshape_3d(ctx, q, d_head, L_q, n_head * N);     // [N * n_head, L_q, d_head]
 
-        v = ggml_reshape_4d(ctx, v, d_head, n_head, L_k, N);  // [N, L_k, n_head, d_head]
+        k = ggml_reshape_4d(ctx, k, d_head, n_kv_head, L_k, N);   // [N, L_k, n_kv_head, d_head]
+        k = ggml_nn_cont(ctx, ggml_permute(ctx, k, 0, 2, 1, 3));  // [N, n_kv_head, L_k, d_head]
+        k = ggml_reshape_3d(ctx, k, d_head, L_k, n_kv_head * N);  // [N * n_kv_head, L_k, d_head]
+
+        v = ggml_reshape_4d(ctx, v, d_head, n_kv_head, L_k, N);  // [N, L_k, n_kv_head, d_head]
     } else {
-        L_q    = q->ne[1];
+        L_q       = q->ne[1];
-        L_k    = k->ne[1];
+        L_k       = k->ne[1];
-        d_head = v->ne[0];
+        d_head    = v->ne[0];
-        N      = v->ne[3];
+        N         = v->ne[3];
-        C      = d_head * n_head;
+        n_kv_head = k->ne[2] / N;
+        C         = d_head * n_head;
     }
 
     float scale = (1.0f / sqrt((float)d_head));
@@ -1177,13 +1197,19 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
         if (kv_pad != 0) {
             k_in = ggml_pad(ctx, k_in, 0, kv_pad, 0, 0);
         }
+        if (kv_scale != 1.0f) {
+            k_in = ggml_scale(ctx, k_in, kv_scale);
+        }
         k_in = ggml_cast(ctx, k_in, GGML_TYPE_F16);
 
         v_in = ggml_nn_cont(ctx, ggml_permute(ctx, v_in, 0, 2, 1, 3));
-        v_in = ggml_reshape_3d(ctx, v_in, d_head, L_k, n_head * N);
+        v_in = ggml_reshape_3d(ctx, v_in, d_head, L_k, n_kv_head * N);
         if (kv_pad != 0) {
             v_in = ggml_pad(ctx, v_in, 0, kv_pad, 0, 0);
         }
+        if (kv_scale != 1.0f) {
+            v_in = ggml_scale(ctx, v_in, kv_scale);
+        }
         v_in = ggml_cast(ctx, v_in, GGML_TYPE_F16);
 
         if (mask_in != nullptr) {
@@ -1207,8 +1233,11 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
             mask_in = ggml_cast(ctx, mask_in, GGML_TYPE_F16);
         }
 
-        auto out = ggml_flash_attn_ext(ctx, q_in, k_in, v_in, mask_in, scale, 0, 0);
+        auto out = ggml_flash_attn_ext(ctx, q_in, k_in, v_in, mask_in, scale / kv_scale, 0, 0);
         ggml_flash_attn_ext_set_prec(out, GGML_PREC_F32);
+        if (kv_scale != 1.0f) {
+            out = ggml_scale(ctx, out, 1.0f / kv_scale);
+        }
         return out;
     };
 
@@ -1238,8 +1267,8 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
         // if (flash_attn) {
         //     LOG_DEBUG("fallback to default attention, L_q:%d L_k:%d n_head:%d C:%d d_head:%d N:%d", L_q, L_k, n_head, C, d_head, N);
         // }
-        v = ggml_nn_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));  // [N, n_head, d_head, L_k]
+        v = ggml_nn_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));  // [N, n_kv_head, d_head, L_k]
-        v = ggml_reshape_3d(ctx, v, L_k, d_head, n_head * N);     // [N * n_head, d_head, L_k]
+        v = ggml_reshape_3d(ctx, v, L_k, d_head, n_kv_head * N);  // [N * n_kv_head, d_head, L_k]
 
         auto kq = ggml_mul_mat(ctx, k, q);  // [N * n_head, L_q, L_k]
         kq      = ggml_scale_inplace(ctx, kq, scale);
@@ -1355,15 +1384,13 @@ __STATIC_INLINE__ std::vector<float> arange(float start, float end, float step =
 // Ref: https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/util.py#L151
 __STATIC_INLINE__ std::vector<float> timestep_embedding(std::vector<float> timesteps,
                                                         int dim,
-                                                        int max_period = 10000) {
+                                                        int max_period       = 10000,
+                                                        bool flip_sin_to_cos = true,
+                                                        float scale          = 1.f) {
     // timesteps: [N,]
     // embedding: [N, dim]
-    size_t N        = timesteps.size();
+    size_t N = timesteps.size();
-    int acutual_dim = dim;
+    std::vector<float> embedding(N * dim, 0.f);
-    if (dim % 2 != 0) {
-        acutual_dim = dim + 1;
-    }
-    std::vector<float> embedding(N * acutual_dim, 0.f);
     int half = dim / 2;
     std::vector<float> freqs(half);
     for (int i = 0; i < half; ++i) {
@@ -1371,9 +1398,14 @@ __STATIC_INLINE__ std::vector<float> timestep_embedding(std::vector<float> times
     }
     for (int i = 0; i < N; ++i) {
         for (int j = 0; j < half; ++j) {
-            float arg                             = timesteps[i] * freqs[j];
+            float arg = timesteps[i] * freqs[j] * scale;
-            embedding[i * acutual_dim + j]        = std::cos(arg);
+            if (flip_sin_to_cos) {
-            embedding[i * acutual_dim + j + half] = std::sin(arg);
+                embedding[i * dim + j]        = std::cos(arg);
+                embedding[i * dim + j + half] = std::sin(arg);
+            } else {
+                embedding[i * dim + j]        = std::sin(arg);
+                embedding[i * dim + j + half] = std::cos(arg);
+            }
         }
     }
     return embedding;
@@ -1394,11 +1426,7 @@ __STATIC_INLINE__ struct ggml_tensor* new_timestep_embedding(struct ggml_context
     // timesteps: [N,]
     // embedding: [N, dim]
     std::vector<float> embedding_vec = timestep_embedding(timesteps, dim, max_period);
-    int acutual_dim                  = dim;
+    struct ggml_tensor* embedding    = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, dim, timesteps.size());
-    if (dim % 2 != 0) {
-        acutual_dim = dim + 1;
-    }
-    struct ggml_tensor* embedding = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, acutual_dim, timesteps.size());
     if (embedding->data != NULL) {
         memcpy(((char*)embedding->data), ((char*)embedding_vec.data()), ggml_nbytes(embedding));
     } else {
@@ -1940,6 +1968,8 @@ protected:
     int64_t out_features;
     bool bias;
     bool force_f32;
+    bool force_prec_f32;
+    float scale;
 
     void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
         enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
@@ -1956,12 +1986,16 @@ protected:
 public:
     Linear(int64_t in_features,
            int64_t out_features,
-           bool bias      = true,
+           bool bias           = true,
-           bool force_f32 = false)
+           bool force_f32      = false,
+           bool force_prec_f32 = false,
+           float scale         = 1.f)
         : in_features(in_features),
           out_features(out_features),
           bias(bias),
-          force_f32(force_f32) {}
+          force_f32(force_f32),
+          force_prec_f32(force_prec_f32),
+          scale(scale) {}
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
         struct ggml_tensor* w = params["weight"];
@@ -1969,7 +2003,7 @@ public:
         if (bias) {
             b = params["bias"];
         }
-        return ggml_nn_linear(ctx, x, w, b);
+        return ggml_nn_linear(ctx, x, w, b, force_prec_f32, scale);
     }
 };
 

model.cpp

@@ -17,6 +17,7 @@
 #include "stable-diffusion.h"
 #include "util.h"
 #include "vocab.hpp"
+#include "vocab_qwen.hpp"
 #include "vocab_umt5.hpp"
 
 #include "ggml-alloc.h"
@@ -110,6 +111,9 @@ const char* unused_tensors[] = {
     "embedding_manager",
     "denoiser.sigmas",
     "text_encoders.t5xxl.transformer.encoder.embed_tokens.weight",  // only used during training
+    "text_encoders.qwen2vl.output.weight",
+    "text_encoders.qwen2vl.lm_head.",
+    "text_encoders.qwen2vl.visual.",
 };
 
 bool is_unused_tensor(std::string name) {
@@ -193,6 +197,21 @@ std::unordered_map<std::string, std::string> pmid_v2_name_map = {
      "pmid.qformer_perceiver.token_proj.fc2.weight"},
 };
 
+std::unordered_map<std::string, std::string> qwenvl_name_map{
+    {"token_embd.", "model.embed_tokens."},
+    {"blk.", "model.layers."},
+    {"attn_q.", "self_attn.q_proj."},
+    {"attn_k.", "self_attn.k_proj."},
+    {"attn_v.", "self_attn.v_proj."},
+    {"attn_output.", "self_attn.o_proj."},
+    {"attn_norm.", "input_layernorm."},
+    {"ffn_down.", "mlp.down_proj."},
+    {"ffn_gate.", "mlp.gate_proj."},
+    {"ffn_up.", "mlp.up_proj."},
+    {"ffn_norm.", "post_attention_layernorm."},
+    {"output_norm.", "model.norm."},
+};
+
 std::string convert_cond_model_name(const std::string& name) {
     std::string new_name = name;
     std::string prefix;
@@ -250,6 +269,13 @@ std::string convert_cond_model_name(const std::string& name) {
         if (pos != std::string::npos) {
             new_name.replace(pos, 11, "layer.0.SelfAttention.relative_attention_bias.");
         }
+    } else if (contains(name, "qwen2vl")) {
+        for (auto kv : qwenvl_name_map) {
+            size_t pos = new_name.find(kv.first);
+            if (pos != std::string::npos) {
+                new_name.replace(pos, kv.first.size(), kv.second);
+            }
+        }
     } else if (name == "text_encoders.t5xxl.transformer.token_embd.weight") {
         new_name = "text_encoders.t5xxl.transformer.shared.weight";
     }
@@ -580,7 +606,11 @@ std::string convert_tensor_name(std::string name) {
     //     name.replace(pos, strlen("lora_B"), "lora_down");
     // }
     std::string new_name = name;
-    if (starts_with(name, "cond_stage_model.") || starts_with(name, "conditioner.embedders.") || starts_with(name, "text_encoders.") || ends_with(name, ".vision_model.visual_projection.weight")) {
+    if (starts_with(name, "cond_stage_model.") ||
+        starts_with(name, "conditioner.embedders.") ||
+        starts_with(name, "text_encoders.") ||
+        ends_with(name, ".vision_model.visual_projection.weight") ||
+        starts_with(name, "qwen2vl")) {
         new_name = convert_cond_model_name(name);
     } else if (starts_with(name, "first_stage_model.decoder")) {
         new_name = convert_vae_decoder_name(name);
@@ -699,6 +729,7 @@ void preprocess_tensor(TensorStorage tensor_storage,
 
     // convert unet transformer linear to conv2d 1x1
     if (starts_with(new_name, "model.diffusion_model.") &&
+        !starts_with(new_name, "model.diffusion_model.proj_out.") &&
         (ends_with(new_name, "proj_in.weight") || ends_with(new_name, "proj_out.weight"))) {
         tensor_storage.unsqueeze();
     }
@@ -1732,6 +1763,9 @@ SDVersion ModelLoader::get_sd_version() {
             if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
                 return VERSION_SD3;
             }
+            if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
+                return VERSION_QWEN_IMAGE;
+            }
             if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
                 is_wan = true;
             }
@@ -1945,6 +1979,11 @@ std::string ModelLoader::load_merges() {
     return merges_utf8_str;
 }
 
+std::string ModelLoader::load_qwen2_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(qwen2_merges_utf8_c_str), sizeof(qwen2_merges_utf8_c_str));
+    return merges_utf8_str;
+}
+
 std::string ModelLoader::load_t5_tokenizer_json() {
     std::string json_str(reinterpret_cast<const char*>(t5_tokenizer_json_str), sizeof(t5_tokenizer_json_str));
     return json_str;

model.h

@@ -36,6 +36,7 @@ enum SDVersion {
     VERSION_WAN2,
     VERSION_WAN2_2_I2V,
     VERSION_WAN2_2_TI2V,
+    VERSION_QWEN_IMAGE,
     VERSION_COUNT,
 };
 
@@ -81,6 +82,13 @@ static inline bool sd_version_is_wan(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_qwen_image(SDVersion version) {
+    if (version == VERSION_QWEN_IMAGE) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_inpaint(SDVersion version) {
     if (version == VERSION_SD1_INPAINT || version == VERSION_SD2_INPAINT || version == VERSION_SDXL_INPAINT || version == VERSION_FLUX_FILL || version == VERSION_FLEX_2) {
         return true;
@@ -89,7 +97,10 @@ static inline bool sd_version_is_inpaint(SDVersion version) {
 }
 
 static inline bool sd_version_is_dit(SDVersion version) {
-    if (sd_version_is_flux(version) || sd_version_is_sd3(version) || sd_version_is_wan(version)) {
+    if (sd_version_is_flux(version) ||
+        sd_version_is_sd3(version) ||
+        sd_version_is_wan(version) ||
+        sd_version_is_qwen_image(version)) {
         return true;
     }
     return false;
@@ -272,6 +283,7 @@ public:
     ~ModelLoader() = default;
 
     static std::string load_merges();
+    static std::string load_qwen2_merges();
     static std::string load_t5_tokenizer_json();
     static std::string load_umt5_tokenizer_json();
 };

qwen_image.hpp

@@ -0,0 +1,643 @@
+#ifndef __QWEN_IMAGE_HPP__
+#define __QWEN_IMAGE_HPP__
+
+#include "common.hpp"
+#include "flux.hpp"
+#include "ggml_extend.hpp"
+
+namespace Qwen {
+    constexpr int QWEN_IMAGE_GRAPH_SIZE = 20480;
+
+    struct TimestepEmbedding : public GGMLBlock {
+    public:
+        TimestepEmbedding(int64_t in_channels,
+                          int64_t time_embed_dim,
+                          int64_t out_dim       = 0,
+                          int64_t cond_proj_dim = 0,
+                          bool sample_proj_bias = true) {
+            blocks["linear_1"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, time_embed_dim, sample_proj_bias));
+            if (cond_proj_dim > 0) {
+                blocks["cond_proj"] = std::shared_ptr<GGMLBlock>(new Linear(cond_proj_dim, in_channels, false));
+            }
+            if (out_dim <= 0) {
+                out_dim = time_embed_dim;
+            }
+            blocks["linear_2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, out_dim, sample_proj_bias));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* sample,
+                                    struct ggml_tensor* condition = nullptr) {
+            if (condition != nullptr) {
+                auto cond_proj = std::dynamic_pointer_cast<Linear>(blocks["cond_proj"]);
+                sample         = ggml_add(ctx, sample, cond_proj->forward(ctx, condition));
+            }
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["linear_1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
+
+            sample = linear_1->forward(ctx, sample);
+            sample = ggml_silu_inplace(ctx, sample);
+            sample = linear_2->forward(ctx, sample);
+            return sample;
+        }
+    };
+
+    struct QwenTimestepProjEmbeddings : public GGMLBlock {
+    public:
+        QwenTimestepProjEmbeddings(int64_t embedding_dim) {
+            blocks["timestep_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedding(256, embedding_dim));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* timesteps) {
+            // timesteps: [N,]
+            // return: [N, embedding_dim]
+            auto timestep_embedder = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["timestep_embedder"]);
+
+            auto timesteps_proj = ggml_nn_timestep_embedding(ctx, timesteps, 256, 10000, 1.f);
+            auto timesteps_emb  = timestep_embedder->forward(ctx, timesteps_proj);
+            return timesteps_emb;
+        }
+    };
+
+    struct QwenImageAttention : public GGMLBlock {
+    protected:
+        int64_t dim_head;
+        bool flash_attn;
+
+    public:
+        QwenImageAttention(int64_t query_dim,
+                           int64_t dim_head,
+                           int64_t num_heads,
+                           int64_t out_dim         = 0,
+                           int64_t out_context_dim = 0,
+                           bool bias               = true,
+                           bool out_bias           = true,
+                           float eps               = 1e-6,
+                           bool flash_attn         = false)
+            : dim_head(dim_head), flash_attn(flash_attn) {
+            int64_t inner_dim = out_dim > 0 ? out_dim : dim_head * num_heads;
+            out_dim           = out_dim > 0 ? out_dim : query_dim;
+            out_context_dim   = out_context_dim > 0 ? out_context_dim : query_dim;
+
+            blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["to_k"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["to_v"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+
+            blocks["norm_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+            blocks["norm_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+
+            blocks["add_q_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["add_k_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["add_v_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+
+            blocks["norm_added_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+            blocks["norm_added_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+
+            blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_dim, out_bias));
+            // to_out.1 is nn.Dropout
+
+            float scale = 1.f / 32.f;
+            // The purpose of the scale here is to prevent NaN issues in certain situations.
+            // For example when using CUDA but the weights are k-quants (not all prompts).
+            blocks["to_add_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_context_dim, out_bias, false, false, scale));
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,
+                                                      ggml_backend_t backend,
+                                                      struct ggml_tensor* img,
+                                                      struct ggml_tensor* txt,
+                                                      struct ggml_tensor* pe,
+                                                      struct ggml_tensor* mask = nullptr) {
+            // img: [N, n_img_token, hidden_size]
+            // txt: [N, n_txt_token, hidden_size]
+            // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
+            // return: ([N, n_img_token, hidden_size], [N, n_txt_token, hidden_size])
+
+            auto norm_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_q"]);
+            auto norm_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_k"]);
+
+            auto to_q     = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+            auto to_k     = std::dynamic_pointer_cast<Linear>(blocks["to_k"]);
+            auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
+            auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+
+            auto norm_added_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_q"]);
+            auto norm_added_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_k"]);
+
+            auto add_q_proj = std::dynamic_pointer_cast<Linear>(blocks["add_q_proj"]);
+            auto add_k_proj = std::dynamic_pointer_cast<Linear>(blocks["add_k_proj"]);
+            auto add_v_proj = std::dynamic_pointer_cast<Linear>(blocks["add_v_proj"]);
+            auto to_add_out = std::dynamic_pointer_cast<Linear>(blocks["to_add_out"]);
+
+            int64_t N           = img->ne[2];
+            int64_t n_img_token = img->ne[1];
+            int64_t n_txt_token = txt->ne[1];
+
+            auto img_q        = to_q->forward(ctx, img);
+            int64_t num_heads = img_q->ne[0] / dim_head;
+            img_q             = ggml_reshape_4d(ctx, img_q, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+            auto img_k        = to_k->forward(ctx, img);
+            img_k             = ggml_reshape_4d(ctx, img_k, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+            auto img_v        = to_v->forward(ctx, img);
+            img_v             = ggml_reshape_4d(ctx, img_v, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+
+            img_q = norm_q->forward(ctx, img_q);
+            img_k = norm_k->forward(ctx, img_k);
+
+            auto txt_q = add_q_proj->forward(ctx, txt);
+            txt_q      = ggml_reshape_4d(ctx, txt_q, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+            auto txt_k = add_k_proj->forward(ctx, txt);
+            txt_k      = ggml_reshape_4d(ctx, txt_k, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+            auto txt_v = add_v_proj->forward(ctx, txt);
+            txt_v      = ggml_reshape_4d(ctx, txt_v, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+
+            txt_q = norm_added_q->forward(ctx, txt_q);
+            txt_k = norm_added_k->forward(ctx, txt_k);
+
+            auto q = ggml_concat(ctx, txt_q, img_q, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
+            auto k = ggml_concat(ctx, txt_k, img_k, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
+            auto v = ggml_concat(ctx, txt_v, img_v, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
+
+            auto attn         = Flux::attention(ctx, backend, q, k, v, pe, mask, flash_attn, (1.0f / 128.f));  // [N, n_txt_token + n_img_token, n_head*d_head]
+            attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));                           // [n_txt_token + n_img_token, N, hidden_size]
+            auto txt_attn_out = ggml_view_3d(ctx,
+                                             attn,
+                                             attn->ne[0],
+                                             attn->ne[1],
+                                             txt->ne[1],
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             0);                                              // [n_txt_token, N, hidden_size]
+            txt_attn_out      = ggml_cont(ctx, ggml_permute(ctx, txt_attn_out, 0, 2, 1, 3));  // [N, n_txt_token, hidden_size]
+            auto img_attn_out = ggml_view_3d(ctx,
+                                             attn,
+                                             attn->ne[0],
+                                             attn->ne[1],
+                                             img->ne[1],
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             attn->nb[2] * txt->ne[1]);                       // [n_img_token, N, hidden_size]
+            img_attn_out      = ggml_cont(ctx, ggml_permute(ctx, img_attn_out, 0, 2, 1, 3));  // [N, n_img_token, hidden_size]
+
+            img_attn_out = to_out_0->forward(ctx, img_attn_out);
+            txt_attn_out = to_add_out->forward(ctx, txt_attn_out);
+
+            return {img_attn_out, txt_attn_out};
+        }
+    };
+
+    class QwenImageTransformerBlock : public GGMLBlock {
+    public:
+        QwenImageTransformerBlock(int64_t dim,
+                                  int64_t num_attention_heads,
+                                  int64_t attention_head_dim,
+                                  float eps       = 1e-6,
+                                  bool flash_attn = false) {
+            // img_mod.0 is nn.SiLU()
+            blocks["img_mod.1"] = std::shared_ptr<GGMLBlock>(new Linear(dim, 6 * dim, true));
+
+            blocks["img_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["img_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["img_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU, true));
+
+            // txt_mod.0 is nn.SiLU()
+            blocks["txt_mod.1"] = std::shared_ptr<GGMLBlock>(new Linear(dim, 6 * dim, true));
+
+            blocks["txt_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["txt_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["txt_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU));
+
+            blocks["attn"] = std::shared_ptr<GGMLBlock>(new QwenImageAttention(dim,
+                                                                               attention_head_dim,
+                                                                               num_attention_heads,
+                                                                               0,     // out_dim
+                                                                               0,     // out_context-dim
+                                                                               true,  // bias
+                                                                               true,  // out_bias
+                                                                               eps,
+                                                                               flash_attn));
+        }
+
+        virtual std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,
+                                                              ggml_backend_t backend,
+                                                              struct ggml_tensor* img,
+                                                              struct ggml_tensor* txt,
+                                                              struct ggml_tensor* t_emb,
+                                                              struct ggml_tensor* pe) {
+            // img: [N, n_img_token, hidden_size]
+            // txt: [N, n_txt_token, hidden_size]
+            // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
+            // return: ([N, n_img_token, hidden_size], [N, n_txt_token, hidden_size])
+
+            auto img_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["img_mod.1"]);
+            auto img_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm1"]);
+            auto img_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm2"]);
+            auto img_mlp   = std::dynamic_pointer_cast<FeedForward>(blocks["img_mlp"]);
+
+            auto txt_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["txt_mod.1"]);
+            auto txt_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm1"]);
+            auto txt_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm2"]);
+            auto txt_mlp   = std::dynamic_pointer_cast<FeedForward>(blocks["txt_mlp"]);
+
+            auto attn = std::dynamic_pointer_cast<QwenImageAttention>(blocks["attn"]);
+
+            auto img_mod_params    = ggml_silu(ctx, t_emb);
+            img_mod_params         = img_mod_1->forward(ctx, img_mod_params);
+            auto img_mod_param_vec = ggml_chunk(ctx, img_mod_params, 6, 0);
+
+            auto txt_mod_params    = ggml_silu(ctx, t_emb);
+            txt_mod_params         = txt_mod_1->forward(ctx, txt_mod_params);
+            auto txt_mod_param_vec = ggml_chunk(ctx, txt_mod_params, 6, 0);
+
+            auto img_normed    = img_norm1->forward(ctx, img);
+            auto img_modulated = Flux::modulate(ctx, img_normed, img_mod_param_vec[0], img_mod_param_vec[1]);
+            auto img_gate1     = img_mod_param_vec[2];
+
+            auto txt_normed    = txt_norm1->forward(ctx, txt);
+            auto txt_modulated = Flux::modulate(ctx, txt_normed, txt_mod_param_vec[0], txt_mod_param_vec[1]);
+            auto txt_gate1     = txt_mod_param_vec[2];
+
+            auto [img_attn_output, txt_attn_output] = attn->forward(ctx, backend, img_modulated, txt_modulated, pe);
+
+            img = ggml_add(ctx, img, ggml_mul(ctx, img_attn_output, img_gate1));
+            txt = ggml_add(ctx, txt, ggml_mul(ctx, txt_attn_output, txt_gate1));
+
+            auto img_normed2    = img_norm2->forward(ctx, img);
+            auto img_modulated2 = Flux::modulate(ctx, img_normed2, img_mod_param_vec[3], img_mod_param_vec[4]);
+            auto img_gate2      = img_mod_param_vec[5];
+
+            auto txt_normed2    = txt_norm2->forward(ctx, txt);
+            auto txt_modulated2 = Flux::modulate(ctx, txt_normed2, txt_mod_param_vec[3], txt_mod_param_vec[4]);
+            auto txt_gate2      = txt_mod_param_vec[5];
+
+            auto img_mlp_out = img_mlp->forward(ctx, img_modulated2);
+            auto txt_mlp_out = txt_mlp->forward(ctx, txt_modulated2);
+
+            img = ggml_add(ctx, img, ggml_mul(ctx, img_mlp_out, img_gate2));
+            txt = ggml_add(ctx, txt, ggml_mul(ctx, txt_mlp_out, txt_gate2));
+
+            return {img, txt};
+        }
+    };
+
+    struct AdaLayerNormContinuous : public GGMLBlock {
+    public:
+        AdaLayerNormContinuous(int64_t embedding_dim,
+                               int64_t conditioning_embedding_dim,
+                               bool elementwise_affine = true,
+                               float eps               = 1e-5f,
+                               bool bias               = true) {
+            blocks["norm"]   = std::shared_ptr<GGMLBlock>(new LayerNorm(conditioning_embedding_dim, eps, elementwise_affine, bias));
+            blocks["linear"] = std::shared_ptr<GGMLBlock>(new Linear(conditioning_embedding_dim, embedding_dim * 2, bias));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* c) {
+            // x: [N, n_token, hidden_size]
+            // c: [N, hidden_size]
+            // return: [N, n_token, patch_size * patch_size * out_channels]
+
+            auto norm   = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+
+            auto emb   = linear->forward(ctx, ggml_silu(ctx, c));
+            auto mods  = ggml_chunk(ctx, emb, 2, 0);
+            auto scale = mods[0];
+            auto shift = mods[1];
+
+            x = norm->forward(ctx, x);
+            x = Flux::modulate(ctx, x, shift, scale);
+
+            return x;
+        }
+    };
+
+    struct QwenImageParams {
+        int64_t patch_size          = 2;
+        int64_t in_channels         = 64;
+        int64_t out_channels        = 16;
+        int64_t num_layers          = 60;
+        int64_t attention_head_dim  = 128;
+        int64_t num_attention_heads = 24;
+        int64_t joint_attention_dim = 3584;
+        float theta                 = 10000;
+        std::vector<int> axes_dim   = {16, 56, 56};
+        int64_t axes_dim_sum        = 128;
+        bool flash_attn             = false;
+    };
+
+    class QwenImageModel : public GGMLBlock {
+    protected:
+        QwenImageParams params;
+
+    public:
+        QwenImageModel() {}
+        QwenImageModel(QwenImageParams params)
+            : params(params) {
+            int64_t inner_dim         = params.num_attention_heads * params.attention_head_dim;
+            blocks["time_text_embed"] = std::shared_ptr<GGMLBlock>(new QwenTimestepProjEmbeddings(inner_dim));
+            blocks["txt_norm"]        = std::shared_ptr<GGMLBlock>(new RMSNorm(params.joint_attention_dim, 1e-6f));
+            blocks["img_in"]          = std::shared_ptr<GGMLBlock>(new Linear(params.in_channels, inner_dim));
+            blocks["txt_in"]          = std::shared_ptr<GGMLBlock>(new Linear(params.joint_attention_dim, inner_dim));
+
+            // blocks
+            for (int i = 0; i < params.num_layers; i++) {
+                auto block                                        = std::shared_ptr<GGMLBlock>(new QwenImageTransformerBlock(inner_dim,
+                                                                                                                             params.num_attention_heads,
+                                                                                                                             params.attention_head_dim,
+                                                                                                                             1e-6f,
+                                                                                                                             params.flash_attn));
+                blocks["transformer_blocks." + std::to_string(i)] = block;
+            }
+
+            blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new AdaLayerNormContinuous(inner_dim, inner_dim, false, 1e-6f));
+            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, params.patch_size * params.patch_size * params.out_channels));
+        }
+
+        struct ggml_tensor* pad_to_patch_size(struct ggml_context* ctx,
+                                              struct ggml_tensor* x) {
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+
+            int pad_h = (params.patch_size - H % params.patch_size) % params.patch_size;
+            int pad_w = (params.patch_size - W % params.patch_size) % params.patch_size;
+            x         = ggml_pad(ctx, x, pad_w, pad_h, 0, 0);  // [N, C, H + pad_h, W + pad_w]
+            return x;
+        }
+
+        struct ggml_tensor* patchify(struct ggml_context* ctx,
+                                     struct ggml_tensor* x) {
+            // x: [N, C, H, W]
+            // return: [N, h*w, C * patch_size * patch_size]
+            int64_t N = x->ne[3];
+            int64_t C = x->ne[2];
+            int64_t H = x->ne[1];
+            int64_t W = x->ne[0];
+            int64_t p = params.patch_size;
+            int64_t h = H / params.patch_size;
+            int64_t w = W / params.patch_size;
+
+            GGML_ASSERT(h * p == H && w * p == W);
+
+            x = ggml_reshape_4d(ctx, x, p, w, p, h * C * N);       // [N*C*h, p, w, p]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, w, p, p]
+            x = ggml_reshape_4d(ctx, x, p * p, w * h, C, N);       // [N, C, h*w, p*p]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, h*w, C, p*p]
+            x = ggml_reshape_3d(ctx, x, p * p * C, w * h, N);      // [N, h*w, C*p*p]
+            return x;
+        }
+
+        struct ggml_tensor* unpatchify(struct ggml_context* ctx,
+                                       struct ggml_tensor* x,
+                                       int64_t h,
+                                       int64_t w) {
+            // x: [N, h*w, C*patch_size*patch_size]
+            // return: [N, C, H, W]
+            int64_t N = x->ne[2];
+            int64_t C = x->ne[0] / params.patch_size / params.patch_size;
+            int64_t H = h * params.patch_size;
+            int64_t W = w * params.patch_size;
+            int64_t p = params.patch_size;
+
+            GGML_ASSERT(C * p * p == x->ne[0]);
+
+            x = ggml_reshape_4d(ctx, x, p * p, C, w * h, N);       // [N, h*w, C, p*p]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, C, h*w, p*p]
+            x = ggml_reshape_4d(ctx, x, p, p, w, h * C * N);       // [N*C*h, w, p, p]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, p, w, p]
+            x = ggml_reshape_4d(ctx, x, W, H, C, N);               // [N, C, h*p, w*p]
+
+            return x;
+        }
+
+        struct ggml_tensor* forward_orig(struct ggml_context* ctx,
+                                         ggml_backend_t backend,
+                                         struct ggml_tensor* x,
+                                         struct ggml_tensor* timestep,
+                                         struct ggml_tensor* context,
+                                         struct ggml_tensor* pe) {
+            auto time_text_embed = std::dynamic_pointer_cast<QwenTimestepProjEmbeddings>(blocks["time_text_embed"]);
+            auto txt_norm        = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm"]);
+            auto img_in          = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
+            auto txt_in          = std::dynamic_pointer_cast<Linear>(blocks["txt_in"]);
+            auto norm_out        = std::dynamic_pointer_cast<AdaLayerNormContinuous>(blocks["norm_out"]);
+            auto proj_out        = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
+
+            auto t_emb = time_text_embed->forward(ctx, timestep);
+            auto img   = img_in->forward(ctx, x);
+            auto txt   = txt_norm->forward(ctx, context);
+            txt        = txt_in->forward(ctx, txt);
+
+            for (int i = 0; i < params.num_layers; i++) {
+                auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
+
+                auto result = block->forward(ctx, backend, img, txt, t_emb, pe);
+                img         = result.first;
+                txt         = result.second;
+            }
+
+            img = norm_out->forward(ctx, img, t_emb);
+            img = proj_out->forward(ctx, img);
+
+            return img;
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* timestep,
+                                    struct ggml_tensor* context,
+                                    struct ggml_tensor* pe) {
+            // Forward pass of DiT.
+            // x: [N, C, H, W]
+            // timestep: [N,]
+            // context: [N, L, D]
+            // pe: [L, d_head/2, 2, 2]
+            // return: [N, C, H, W]
+
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t C = x->ne[2];
+            int64_t N = x->ne[3];
+
+            x = pad_to_patch_size(ctx, x);
+            x = patchify(ctx, x);
+
+            int64_t h_len = ((H + (params.patch_size / 2)) / params.patch_size);
+            int64_t w_len = ((W + (params.patch_size / 2)) / params.patch_size);
+
+            auto out = forward_orig(ctx, backend, x, timestep, context, pe);  // [N, h_len*w_len, ph*pw*C]
+
+            out = unpatchify(ctx, out, h_len, w_len);  // [N, C, H + pad_h, W + pad_w]
+
+            // slice
+            out = ggml_slice(ctx, out, 1, 0, H);  // [N, C, H, W + pad_w]
+            out = ggml_slice(ctx, out, 0, 0, W);  // [N, C, H, W]
+
+            return out;
+        }
+    };
+
+    struct QwenImageRunner : public GGMLRunner {
+    public:
+        QwenImageParams qwen_image_params;
+        QwenImageModel qwen_image;
+        std::vector<float> pe_vec;
+        SDVersion version;
+
+        QwenImageRunner(ggml_backend_t backend,
+                        bool offload_params_to_cpu,
+                        const String2GGMLType& tensor_types = {},
+                        const std::string prefix            = "",
+                        SDVersion version                   = VERSION_QWEN_IMAGE,
+                        bool flash_attn                     = false)
+            : GGMLRunner(backend, offload_params_to_cpu) {
+            qwen_image_params.flash_attn = flash_attn;
+            qwen_image                   = QwenImageModel(qwen_image_params);
+            qwen_image.init(params_ctx, tensor_types, prefix);
+        }
+
+        std::string get_desc() {
+            return "qwen_image";
+        }
+
+        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+            qwen_image.get_param_tensors(tensors, prefix);
+        }
+
+        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+                                        struct ggml_tensor* timesteps,
+                                        struct ggml_tensor* context) {
+            GGML_ASSERT(x->ne[3] == 1);
+            struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, QWEN_IMAGE_GRAPH_SIZE, false);
+
+            x         = to_backend(x);
+            context   = to_backend(context);
+            timesteps = to_backend(timesteps);
+
+            pe_vec      = Rope::gen_qwen_image_pe(x->ne[1],
+                                                  x->ne[0],
+                                                  qwen_image_params.patch_size,
+                                                  x->ne[3],
+                                                  context->ne[1],
+                                                  qwen_image_params.theta,
+                                                  qwen_image_params.axes_dim);
+            int pos_len = pe_vec.size() / qwen_image_params.axes_dim_sum / 2;
+            // LOG_DEBUG("pos_len %d", pos_len);
+            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, qwen_image_params.axes_dim_sum / 2, pos_len);
+            // pe->data = pe_vec.data();
+            // print_ggml_tensor(pe);
+            // pe->data = NULL;
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            struct ggml_tensor* out = qwen_image.forward(compute_ctx,
+                                                         runtime_backend,
+                                                         x,
+                                                         timesteps,
+                                                         context,
+                                                         pe);
+
+            ggml_build_forward_expand(gf, out);
+
+            return gf;
+        }
+
+        void compute(int n_threads,
+                     struct ggml_tensor* x,
+                     struct ggml_tensor* timesteps,
+                     struct ggml_tensor* context,
+                     struct ggml_tensor** output     = NULL,
+                     struct ggml_context* output_ctx = NULL) {
+            // x: [N, in_channels, h, w]
+            // timesteps: [N, ]
+            // context: [N, max_position, hidden_size]
+            auto get_graph = [&]() -> struct ggml_cgraph* {
+                return build_graph(x, timesteps, context);
+            };
+
+            GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        }
+
+        void test() {
+            struct ggml_init_params params;
+            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
+            params.mem_buffer = NULL;
+            params.no_alloc   = false;
+
+            struct ggml_context* work_ctx = ggml_init(params);
+            GGML_ASSERT(work_ctx != NULL);
+
+            {
+                // auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 16, 16, 16, 1);
+                // ggml_set_f32(x, 0.01f);
+                auto x = load_tensor_from_file(work_ctx, "./qwen_image_x.bin");
+                print_ggml_tensor(x);
+
+                std::vector<float> timesteps_vec(1, 1000.f);
+                auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
+
+                // auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 3584, 256, 1);
+                // ggml_set_f32(context, 0.01f);
+                auto context = load_tensor_from_file(work_ctx, "./qwen_image_context.bin");
+                print_ggml_tensor(context);
+
+                struct ggml_tensor* out = NULL;
+
+                int t0 = ggml_time_ms();
+                compute(8, x, timesteps, context, &out, work_ctx);
+                int t1 = ggml_time_ms();
+
+                print_ggml_tensor(out);
+                LOG_DEBUG("qwen_image test done in %dms", t1 - t0);
+            }
+        }
+
+        static void load_from_file_and_test(const std::string& file_path) {
+            // cuda q8: pass
+            // cuda q8 fa: nan
+            // ggml_backend_t backend    = ggml_backend_cuda_init(0);
+            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_type model_data_type = GGML_TYPE_Q8_0;
+
+            ModelLoader model_loader;
+            if (!model_loader.init_from_file(file_path, "model.diffusion_model.")) {
+                LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
+                return;
+            }
+
+            auto tensor_types = model_loader.tensor_storages_types;
+            for (auto& item : tensor_types) {
+                // LOG_DEBUG("%s %u", item.first.c_str(), item.second);
+                if (ends_with(item.first, "weight")) {
+                    item.second = model_data_type;
+                }
+            }
+
+            std::shared_ptr<QwenImageRunner> qwen_image = std::shared_ptr<QwenImageRunner>(new QwenImageRunner(backend,
+                                                                                                               false,
+                                                                                                               tensor_types,
+                                                                                                               "model.diffusion_model",
+                                                                                                               VERSION_QWEN_IMAGE,
+                                                                                                               true));
+
+            qwen_image->alloc_params_buffer();
+            std::map<std::string, ggml_tensor*> tensors;
+            qwen_image->get_param_tensors(tensors, "model.diffusion_model");
+
+            bool success = model_loader.load_tensors(tensors);
+
+            if (!success) {
+                LOG_ERROR("load tensors from model loader failed");
+                return;
+            }
+
+            LOG_INFO("qwen_image model loaded");
+            qwen_image->test();
+        }
+    };
+
+}  // namespace name
+
+#endif  // __QWEN_IMAGE_HPP__

qwenvl.hpp

@@ -0,0 +1,731 @@
+#ifndef __QWENVL_HPP__
+#define __QWENVL_HPP__
+
+#include <algorithm>
+#include <fstream>
+#include <iostream>
+#include <map>
+#include <optional>
+#include <regex>
+#include <set>
+#include <sstream>
+#include <string>
+#include <vector>
+
+#include "clip.hpp"
+#include "ggml_extend.hpp"
+#include "json.hpp"
+#include "tokenize_util.h"
+
+namespace Qwen {
+
+    class Qwen2Tokenizer {
+    private:
+        std::map<int, std::u32string> byte_encoder;
+        std::map<std::u32string, int> byte_decoder;
+        std::map<std::u32string, int> encoder;
+        std::map<int, std::u32string> decoder;
+        std::map<std::pair<std::u32string, std::u32string>, int> bpe_ranks;
+        std::regex pat;
+        int encoder_len;
+        int bpe_len;
+
+    public:
+        const std::string UNK_TOKEN = "<|endoftext|>";
+        const std::string EOS_TOKEN = "<|endoftext|>";
+        const std::string PAD_TOKEN = "<|endoftext|>";
+
+        const int UNK_TOKEN_ID = 151643;
+        const int EOS_TOKEN_ID = 151643;
+        const int PAD_TOKEN_ID = 151643;
+
+        std::vector<std::string> special_tokens = {
+            "<|endoftext|>",
+            "<|im_start|>",
+            "<|im_end|>",
+            "<|object_ref_start|>",
+            "<|object_ref_end|>",
+            "<|box_start|>",
+            "<|box_end|>",
+            "<|quad_start|>",
+            "<|quad_end|>",
+            "<|vision_start|>",
+            "<|vision_end|>",
+            "<|vision_pad|>",
+            "<|image_pad|>",
+            "<|video_pad|>",
+            "<tool_call>",
+            "</tool_call>",
+            "<|fim_prefix|>",
+            "<|fim_middle|>",
+            "<|fim_suffix|>",
+            "<|fim_pad|>",
+            "<|repo_name|>",
+            "<|file_sep|>",
+        };
+
+    private:
+        static std::string strip(const std::string& str) {
+            std::string::size_type start = str.find_first_not_of(" \t\n\r\v\f");
+            std::string::size_type end   = str.find_last_not_of(" \t\n\r\v\f");
+
+            if (start == std::string::npos) {
+                // String contains only whitespace characters
+                return "";
+            }
+
+            return str.substr(start, end - start + 1);
+        }
+
+        static std::string whitespace_clean(std::string text) {
+            text = std::regex_replace(text, std::regex(R"(\s+)"), " ");
+            text = strip(text);
+            return text;
+        }
+
+        static std::set<std::pair<std::u32string, std::u32string>> get_pairs(const std::vector<std::u32string>& subwords) {
+            std::set<std::pair<std::u32string, std::u32string>> pairs;
+            if (subwords.size() == 0) {
+                return pairs;
+            }
+            std::u32string prev_subword = subwords[0];
+            for (int i = 1; i < subwords.size(); i++) {
+                std::u32string subword = subwords[i];
+                std::pair<std::u32string, std::u32string> pair(prev_subword, subword);
+                pairs.insert(pair);
+                prev_subword = subword;
+            }
+            return pairs;
+        }
+
+        bool is_special_token(const std::string& token) {
+            for (auto& special_token : special_tokens) {
+                if (special_token == token) {
+                    return true;
+                }
+            }
+            return false;
+        }
+
+    public:
+        explicit Qwen2Tokenizer(const std::string& merges_utf8_str = "") {
+            if (merges_utf8_str.size() > 0) {
+                load_from_merges(merges_utf8_str);
+            } else {
+                load_from_merges(ModelLoader::load_qwen2_merges());
+            }
+        }
+
+        void load_from_merges(const std::string& merges_utf8_str) {
+            auto byte_unicode_pairs = bytes_to_unicode();
+            // printf("byte_unicode_pairs have %lu pairs \n", byte_unicode_pairs.size());
+            byte_encoder = std::map<int, std::u32string>(byte_unicode_pairs.begin(), byte_unicode_pairs.end());
+            for (auto& pair : byte_unicode_pairs) {
+                byte_decoder[pair.second] = pair.first;
+            }
+            // for (auto & pair: byte_unicode_pairs) {
+            //     std::cout << pair.first << ": " << pair.second << std::endl;
+            // }
+            std::vector<std::u32string> merges;
+            size_t start = 0;
+            size_t pos;
+            std::u32string merges_utf32_str = utf8_to_utf32(merges_utf8_str);
+            while ((pos = merges_utf32_str.find('\n', start)) != std::string::npos) {
+                merges.push_back(merges_utf32_str.substr(start, pos - start));
+                start = pos + 1;
+            }
+            LOG_DEBUG("merges size %llu", merges.size());
+            merges = std::vector<std::u32string>(merges.begin(), merges.end());
+            std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
+            for (const auto& merge : merges) {
+                size_t space_pos = merge.find(' ');
+                merge_pairs.emplace_back(merge.substr(0, space_pos), merge.substr(space_pos + 1));
+                // LOG_DEBUG("%s", utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
+                // printf("%s :: %s | %s \n", utf32_to_utf8(merge).c_str(), utf32_to_utf8(merge.substr(0, space_pos)).c_str(),
+                //                     utf32_to_utf8(merge.substr(space_pos + 1)).c_str());
+            }
+
+            std::vector<std::u32string> vocab;
+            for (const auto& pair : byte_unicode_pairs) {
+                vocab.push_back(pair.second);
+            }
+            for (const auto& merge : merge_pairs) {
+                vocab.push_back(merge.first + merge.second);
+            }
+            for (auto& special_token : special_tokens) {
+                vocab.push_back(utf8_to_utf32(special_token));
+            }
+
+            LOG_DEBUG("vocab size: %llu", vocab.size());
+            int i = 0;
+            for (const auto& token : vocab) {
+                encoder[token] = i;
+                decoder[i]     = token;
+                i++;
+            }
+            encoder_len = i;
+
+            int rank = 0;
+            for (const auto& merge : merge_pairs) {
+                bpe_ranks[merge] = rank++;
+            }
+            bpe_len = rank;
+        };
+
+        std::u32string bpe(const std::u32string& token) {
+            std::vector<std::u32string> word;
+
+            for (int i = 0; i < token.size(); i++) {
+                word.emplace_back(1, token[i]);
+            }
+
+            std::set<std::pair<std::u32string, std::u32string>> pairs = get_pairs(word);
+
+            if (pairs.empty()) {
+                return token;
+            }
+
+            while (true) {
+                auto min_pair_iter = std::min_element(pairs.begin(),
+                                                      pairs.end(),
+                                                      [&](const std::pair<std::u32string, std::u32string>& a,
+                                                          const std::pair<std::u32string, std::u32string>& b) {
+                                                          if (bpe_ranks.find(a) == bpe_ranks.end()) {
+                                                              return false;
+                                                          } else if (bpe_ranks.find(b) == bpe_ranks.end()) {
+                                                              return true;
+                                                          }
+                                                          return bpe_ranks.at(a) < bpe_ranks.at(b);
+                                                      });
+
+                const std::pair<std::u32string, std::u32string>& bigram = *min_pair_iter;
+
+                if (bpe_ranks.find(bigram) == bpe_ranks.end()) {
+                    break;
+                }
+
+                std::u32string first  = bigram.first;
+                std::u32string second = bigram.second;
+                std::vector<std::u32string> new_word;
+                int32_t i = 0;
+
+                while (i < word.size()) {
+                    auto it = std::find(word.begin() + i, word.end(), first);
+                    if (it == word.end()) {
+                        new_word.insert(new_word.end(), word.begin() + i, word.end());
+                        break;
+                    }
+                    new_word.insert(new_word.end(), word.begin() + i, it);
+                    i = static_cast<int32_t>(std::distance(word.begin(), it));
+
+                    if (word[i] == first && i < static_cast<int32_t>(word.size()) - 1 && word[i + 1] == second) {
+                        new_word.push_back(first + second);
+                        i += 2;
+                    } else {
+                        new_word.push_back(word[i]);
+                        i += 1;
+                    }
+                }
+
+                word = new_word;
+
+                if (word.size() == 1) {
+                    break;
+                }
+                pairs = get_pairs(word);
+            }
+
+            std::u32string result;
+            for (int i = 0; i < word.size(); i++) {
+                result += word[i];
+                if (i != word.size() - 1) {
+                    result += utf8_to_utf32(" ");
+                }
+            }
+
+            return result;
+        }
+
+        std::vector<int> tokenize(std::string text,
+                                  on_new_token_cb_t on_new_token_cb = nullptr,
+                                  size_t max_length                 = 0,
+                                  bool padding                      = false) {
+            std::vector<int32_t> tokens = encode(text, on_new_token_cb);
+
+            if (max_length > 0) {
+                if (tokens.size() < max_length) {
+                    tokens.resize(max_length);
+                } else {
+                    if (padding) {
+                        tokens.insert(tokens.end(), max_length - tokens.size(), PAD_TOKEN_ID);
+                    }
+                }
+            }
+
+            return tokens;
+        }
+
+        void pad_tokens(std::vector<int>& tokens,
+                        std::vector<float>& weights,
+                        size_t max_length = 0,
+                        bool padding      = false) {
+            if (max_length > 0 && padding) {
+                size_t n = std::ceil(tokens.size() * 1.0 / max_length);
+                if (n == 0) {
+                    n = 1;
+                }
+                size_t length = max_length * n;
+                LOG_DEBUG("token length: %llu", length);
+                tokens.insert(tokens.end(), length - tokens.size(), PAD_TOKEN_ID);
+                weights.insert(weights.end(), length - weights.size(), 1.0);
+            }
+        }
+
+        std::vector<int> encode(std::string text, on_new_token_cb_t on_new_token_cb = nullptr) {
+            std::string original_text = text;
+            std::vector<int32_t> bpe_tokens;
+            std::vector<std::string> token_strs;
+
+            auto splited_texts = split_with_special_tokens(text, special_tokens);
+
+            for (auto& splited_text : splited_texts) {
+                if (is_special_token(splited_text)) {
+                    bpe_tokens.push_back(encoder[utf8_to_utf32(splited_text)]);
+                    token_strs.push_back(splited_text);
+                    continue;
+                }
+                auto tokens = token_split(splited_text);
+                for (auto& token : tokens) {
+                    if (on_new_token_cb != nullptr) {
+                        bool skip = on_new_token_cb(token, bpe_tokens);
+                        if (skip) {
+                            continue;
+                        }
+                    }
+
+                    std::string token_str = token;
+                    std::u32string utf32_token;
+                    for (int i = 0; i < token_str.length(); i++) {
+                        unsigned char b = token_str[i];
+                        utf32_token += byte_encoder[b];
+                    }
+                    auto bpe_strs = bpe(utf32_token);
+                    size_t start  = 0;
+                    size_t pos;
+                    while ((pos = bpe_strs.find(' ', start)) != std::u32string::npos) {
+                        auto bpe_str = bpe_strs.substr(start, pos - start);
+                        bpe_tokens.push_back(encoder[bpe_str]);
+                        token_strs.push_back(utf32_to_utf8(bpe_str));
+
+                        start = pos + 1;
+                    }
+                    auto bpe_str = bpe_strs.substr(start, bpe_strs.size() - start);
+                    bpe_tokens.push_back(encoder[bpe_str]);
+                    token_strs.push_back(utf32_to_utf8(bpe_str));
+                }
+            }
+
+            std::stringstream ss;
+            ss << "[";
+            for (auto token : token_strs) {
+                ss << "\"" << token << "\", ";
+            }
+            ss << "]";
+            // LOG_DEBUG("split prompt \"%s\" to tokens %s", original_text.c_str(), ss.str().c_str());
+            // printf("split prompt \"%s\" to tokens %s \n", original_text.c_str(), ss.str().c_str());
+            return bpe_tokens;
+        }
+    };
+
+    struct Qwen2_5_VLMLP : public GGMLBlock {
+    public:
+        Qwen2_5_VLMLP(int64_t hidden_size, int64_t intermediate_size, bool bias = false) {
+            blocks["gate_proj"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, intermediate_size, false));
+            blocks["up_proj"]   = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, intermediate_size, false));
+            blocks["down_proj"] = std::shared_ptr<GGMLBlock>(new Linear(intermediate_size, hidden_size, false));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+            // x: [N, n_token, hidden_size]
+            auto gate_proj = std::dynamic_pointer_cast<Linear>(blocks["gate_proj"]);
+            auto up_proj   = std::dynamic_pointer_cast<Linear>(blocks["up_proj"]);
+            auto down_proj = std::dynamic_pointer_cast<Linear>(blocks["down_proj"]);
+
+            auto h = gate_proj->forward(ctx, x);
+            h      = ggml_silu_inplace(ctx, h);
+            h      = ggml_mul_inplace(ctx, h, up_proj->forward(ctx, x));
+            h      = down_proj->forward(ctx, h);
+            return h;
+        }
+    };
+
+    struct Qwen2_5_VLAttention : public GGMLBlock {
+    protected:
+        int64_t head_dim;
+        int64_t num_heads;
+        int64_t num_kv_heads;
+
+    public:
+        Qwen2_5_VLAttention(int64_t hidden_size,
+                            int64_t num_heads,
+                            int64_t num_kv_heads)
+            : num_heads(num_heads), num_kv_heads(num_kv_heads) {
+            head_dim = hidden_size / num_heads;
+            GGML_ASSERT(num_heads * head_dim == hidden_size);
+            blocks["q_proj"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, num_heads * head_dim));
+            blocks["k_proj"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, num_kv_heads * head_dim));
+            blocks["v_proj"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, num_kv_heads * head_dim));
+            blocks["o_proj"] = std::shared_ptr<GGMLBlock>(new Linear(num_heads * head_dim, hidden_size, false));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* input_pos) {
+            // x: [N, n_token, hidden_size]
+            int64_t n_token = x->ne[1];
+            int64_t N       = x->ne[2];
+            auto q_proj     = std::dynamic_pointer_cast<Linear>(blocks["q_proj"]);
+            auto k_proj     = std::dynamic_pointer_cast<Linear>(blocks["k_proj"]);
+            auto v_proj     = std::dynamic_pointer_cast<Linear>(blocks["v_proj"]);
+            auto out_proj   = std::dynamic_pointer_cast<Linear>(blocks["o_proj"]);
+
+            auto q = q_proj->forward(ctx, x);  // [N, n_token, num_heads*head_dim]
+            auto k = k_proj->forward(ctx, x);  // [N, n_token, num_kv_heads*head_dim]
+            auto v = v_proj->forward(ctx, x);  // [N, n_token, num_kv_heads*head_dim]
+
+            q = ggml_reshape_4d(ctx, q, head_dim, num_heads, n_token, N);     // [N, n_token, num_heads, head_dim]
+            k = ggml_reshape_4d(ctx, k, head_dim, num_kv_heads, n_token, N);  // [N, n_token, num_kv_heads, head_dim]
+            v = ggml_reshape_4d(ctx, v, head_dim, num_kv_heads, n_token, N);  // [N, n_token, num_kv_heads, head_dim]
+
+            int sections[4] = {16, 24, 24, 0};
+            q               = ggml_rope_multi(ctx, q, input_pos, nullptr, head_dim, sections, GGML_ROPE_TYPE_MROPE, 128000, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
+            k               = ggml_rope_multi(ctx, k, input_pos, nullptr, head_dim, sections, GGML_ROPE_TYPE_MROPE, 128000, 1000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
+
+            q = ggml_cont(ctx, ggml_torch_permute(ctx, q, 0, 2, 1, 3));            // [N, num_heads, n_token, head_dim]
+            q = ggml_reshape_3d(ctx, q, q->ne[0], q->ne[1], q->ne[2] * q->ne[3]);  // [N*num_heads, n_token, head_dim]
+
+            k = ggml_cont(ctx, ggml_torch_permute(ctx, k, 0, 2, 1, 3));            // [N, num_kv_heads, n_token, head_dim]
+            k = ggml_reshape_3d(ctx, k, k->ne[0], k->ne[1], k->ne[2] * k->ne[3]);  // [N*num_kv_heads, n_token, head_dim]
+
+            x = ggml_nn_attention_ext(ctx, backend, q, k, v, num_heads, nullptr, true, true, false);  // [N, n_token, hidden_size]
+
+            x = out_proj->forward(ctx, x);  // [N, n_token, hidden_size]
+            return x;
+        }
+    };
+
+    struct Qwen2_5_VLBlock : public GGMLBlock {
+    public:
+        Qwen2_5_VLBlock(int64_t hidden_size,
+                        int64_t intermediate_size,
+                        int64_t num_heads,
+                        int64_t num_kv_heads,
+                        float eps = 1e-6f) {
+            blocks["self_attn"]                = std::shared_ptr<GGMLBlock>(new Qwen2_5_VLAttention(hidden_size, num_heads, num_kv_heads));
+            blocks["mlp"]                      = std::shared_ptr<GGMLBlock>(new Qwen2_5_VLMLP(hidden_size, intermediate_size));
+            blocks["input_layernorm"]          = std::shared_ptr<GGMLBlock>(new RMSNorm(hidden_size, eps));
+            blocks["post_attention_layernorm"] = std::shared_ptr<GGMLBlock>(new RMSNorm(hidden_size, eps));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* input_pos) {
+            // x: [N, n_token, hidden_size]
+            auto self_attn                = std::dynamic_pointer_cast<Qwen2_5_VLAttention>(blocks["self_attn"]);
+            auto mlp                      = std::dynamic_pointer_cast<Qwen2_5_VLMLP>(blocks["mlp"]);
+            auto input_layernorm          = std::dynamic_pointer_cast<RMSNorm>(blocks["input_layernorm"]);
+            auto post_attention_layernorm = std::dynamic_pointer_cast<RMSNorm>(blocks["post_attention_layernorm"]);
+
+            auto residual = x;
+            x             = input_layernorm->forward(ctx, x);
+            x             = self_attn->forward(ctx, backend, x, input_pos);
+            x             = ggml_add_inplace(ctx, x, residual);
+
+            residual = x;
+            x        = post_attention_layernorm->forward(ctx, x);
+            x        = mlp->forward(ctx, x);
+            x        = ggml_add_inplace(ctx, x, residual);
+
+            return x;
+        }
+    };
+
+    struct Qwen2_5_VLTextModel : public GGMLBlock {
+    protected:
+        int64_t num_layers;
+
+    public:
+        Qwen2_5_VLTextModel(int64_t num_layers,
+                            int64_t vocab_size,
+                            int64_t hidden_size,
+                            int64_t intermediate_size,
+                            int64_t num_heads,
+                            int64_t num_kv_heads,
+                            float eps = 1e-6f)
+            : num_layers(num_layers) {
+            blocks["embed_tokens"] = std::shared_ptr<GGMLBlock>(new Embedding(vocab_size, hidden_size));
+            for (int i = 0; i < num_layers; i++) {
+                blocks["layers." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new Qwen2_5_VLBlock(hidden_size,
+                                                                                                       intermediate_size,
+                                                                                                       num_heads,
+                                                                                                       num_kv_heads));
+            }
+            blocks["norm"] = std::shared_ptr<GGMLBlock>(new RMSNorm(hidden_size, eps));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* input_ids,
+                                    struct ggml_tensor* input_pos) {
+            // input_ids: [N, n_token]
+            // return: [N, n_token, hidden_size]
+
+            auto embed_tokens = std::dynamic_pointer_cast<Embedding>(blocks["embed_tokens"]);
+            auto norm         = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+
+            auto x = embed_tokens->forward(ctx, input_ids);
+
+            for (int i = 0; i < num_layers; i++) {
+                auto block = std::dynamic_pointer_cast<Qwen2_5_VLBlock>(blocks["layers." + std::to_string(i)]);
+
+                x = block->forward(ctx, backend, x, input_pos);
+            }
+
+            x = norm->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct Qwen2_5_VLParams {
+        int64_t num_layers        = 28;
+        int64_t hidden_size       = 3584;
+        int64_t intermediate_size = 18944;
+        int64_t num_heads         = 28;
+        int64_t num_kv_heads      = 4;
+        int64_t vocab_size        = 152064;
+        float rms_norm_eps        = 1e-06f;
+    };
+
+    struct Qwen2_5_VL : public GGMLBlock {
+        Qwen2_5_VLParams params;
+
+    public:
+        Qwen2_5_VL() {}
+        Qwen2_5_VL(Qwen2_5_VLParams params)
+            : params(params) {
+            blocks["model"] = std::shared_ptr<GGMLBlock>(new Qwen2_5_VLTextModel(params.num_layers,
+                                                                                 params.vocab_size,
+                                                                                 params.hidden_size,
+                                                                                 params.intermediate_size,
+                                                                                 params.num_heads,
+                                                                                 params.num_kv_heads,
+                                                                                 params.rms_norm_eps));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* input_ids,
+                                    struct ggml_tensor* input_pos) {
+            // input_ids: [N, n_token]
+            auto model = std::dynamic_pointer_cast<Qwen2_5_VLTextModel>(blocks["model"]);
+
+            auto x = model->forward(ctx, backend, input_ids, input_pos);
+            return x;
+        }
+    };
+
+    struct Qwen2_5_VLRunner : public GGMLRunner {
+        Qwen2_5_VLParams params;
+        Qwen2_5_VL model;
+
+        std::vector<int> input_pos_vec;
+
+        Qwen2_5_VLRunner(ggml_backend_t backend,
+                         bool offload_params_to_cpu,
+                         const String2GGMLType& tensor_types,
+                         const std::string prefix)
+            : GGMLRunner(backend, offload_params_to_cpu) {
+            model = Qwen2_5_VL(params);
+            model.init(params_ctx, tensor_types, prefix);
+        }
+
+        std::string get_desc() {
+            return "qwenvl2.5";
+        }
+
+        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* input_ids,
+                                    struct ggml_tensor* input_pos) {
+            auto hidden_states = model.forward(ctx, backend, input_ids, input_pos);  // [N, n_token, hidden_size]
+            return hidden_states;
+        }
+
+        struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids) {
+            struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+            input_ids = to_backend(input_ids);
+
+            int64_t n_tokens = input_ids->ne[0];
+            input_pos_vec.resize(n_tokens * 4);
+            for (int i = 0; i < n_tokens; ++i) {
+                input_pos_vec[i]                = i;
+                input_pos_vec[n_tokens + i]     = i;
+                input_pos_vec[2 * n_tokens + i] = i;
+                input_pos_vec[3 * n_tokens + i] = 0;
+            }
+
+            auto input_pos = ggml_new_tensor_1d(compute_ctx,
+                                                GGML_TYPE_I32,
+                                                n_tokens * 4);
+            set_backend_tensor_data(input_pos, input_pos_vec.data());
+
+            struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, input_pos);
+
+            ggml_build_forward_expand(gf, hidden_states);
+
+            return gf;
+        }
+
+        void compute(const int n_threads,
+                     struct ggml_tensor* input_ids,
+                     ggml_tensor** output,
+                     ggml_context* output_ctx = NULL) {
+            auto get_graph = [&]() -> struct ggml_cgraph* {
+                return build_graph(input_ids);
+            };
+            GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+        }
+    };
+
+    struct Qwen2_5_VLEmbedder {
+        Qwen2Tokenizer tokenizer;
+        Qwen2_5_VLRunner model;
+
+        Qwen2_5_VLEmbedder(ggml_backend_t backend,
+                           bool offload_params_to_cpu,
+                           const String2GGMLType& tensor_types = {},
+                           const std::string prefix            = "")
+            : model(backend, offload_params_to_cpu, tensor_types, prefix) {
+        }
+
+        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        void alloc_params_buffer() {
+            model.alloc_params_buffer();
+        }
+
+        std::tuple<std::vector<int>, std::vector<float>> tokenize(std::string text,
+                                                                  size_t max_length = 0,
+                                                                  bool padding      = false) {
+            auto parsed_attention = parse_prompt_attention(text);
+
+            {
+                std::stringstream ss;
+                ss << "[";
+                for (const auto& item : parsed_attention) {
+                    ss << "['" << item.first << "', " << item.second << "], ";
+                }
+                ss << "]";
+                LOG_DEBUG("parse '%s' to %s", text.c_str(), ss.str().c_str());
+            }
+
+            std::vector<int> tokens;
+            std::vector<float> weights;
+            for (const auto& item : parsed_attention) {
+                const std::string& curr_text = item.first;
+                float curr_weight            = item.second;
+                std::vector<int> curr_tokens = tokenizer.tokenize(curr_text, nullptr);
+                tokens.insert(tokens.end(), curr_tokens.begin(), curr_tokens.end());
+                weights.insert(weights.end(), curr_tokens.size(), curr_weight);
+            }
+
+            tokenizer.pad_tokens(tokens, weights, max_length, padding);
+
+            // for (int i = 0; i < tokens.size(); i++) {
+            //     std::cout << tokens[i] << ":" << weights[i] << ", ";
+            // }
+            // std::cout << std::endl;
+
+            return {tokens, weights};
+        }
+
+        void test() {
+            struct ggml_init_params params;
+            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
+            params.mem_buffer = NULL;
+            params.no_alloc   = false;
+
+            struct ggml_context* work_ctx = ggml_init(params);
+            GGML_ASSERT(work_ctx != NULL);
+
+            {
+                std::string text("<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\na lovely cat<|im_end|>\n<|im_start|>assistant\n");
+                auto tokens_and_weights     = tokenize(text, 0, false);
+                std::vector<int>& tokens    = std::get<0>(tokens_and_weights);
+                std::vector<float>& weights = std::get<1>(tokens_and_weights);
+                for (auto token : tokens) {
+                    printf("%d ", token);
+                }
+                printf("\n");
+                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
+                struct ggml_tensor* out = NULL;
+
+                int t0 = ggml_time_ms();
+                model.compute(8, input_ids, &out, work_ctx);
+                int t1 = ggml_time_ms();
+
+                print_ggml_tensor(out);
+                LOG_DEBUG("qwen2vl test done in %dms", t1 - t0);
+            }
+        }
+
+        static void load_from_file_and_test(const std::string& file_path) {
+            // cpu f16: pass
+            // ggml_backend_t backend = ggml_backend_cuda_init(0);
+            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_type model_data_type = GGML_TYPE_Q8_0;
+
+            ModelLoader model_loader;
+            if (!model_loader.init_from_file(file_path, "qwen2vl.")) {
+                LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
+                return;
+            }
+
+            auto tensor_types = model_loader.tensor_storages_types;
+            for (auto& item : tensor_types) {
+                // LOG_DEBUG("%s %u", item.first.c_str(), item.second);
+                if (ends_with(item.first, "weight")) {
+                    item.second = model_data_type;
+                }
+            }
+
+            std::shared_ptr<Qwen2_5_VLEmbedder> qwenvl = std::shared_ptr<Qwen2_5_VLEmbedder>(new Qwen2_5_VLEmbedder(backend, false, tensor_types, "qwen2vl"));
+
+            qwenvl->alloc_params_buffer();
+            std::map<std::string, ggml_tensor*> tensors;
+            qwenvl->get_param_tensors(tensors, "qwen2vl");
+
+            bool success = model_loader.load_tensors(tensors);
+
+            if (!success) {
+                LOG_ERROR("load tensors from model loader failed");
+                return;
+            }
+
+            LOG_INFO("qwenvl model loaded");
+            qwenvl->test();
+        }
+    };
+
+};  // Qwen
+
+#endif  // __QWENVL_HPP__

rope.hpp

@@ -203,6 +203,38 @@ struct Rope {
         return embed_nd(ids, bs, theta, axes_dim);
     }
 
+    static std::vector<std::vector<float>> gen_qwen_image_ids(int h,
+                                                              int w,
+                                                              int patch_size,
+                                                              int bs,
+                                                              int context_len) {
+        int h_len        = (h + (patch_size / 2)) / patch_size;
+        int w_len        = (w + (patch_size / 2)) / patch_size;
+        int txt_id_start = std::max(h_len, w_len);
+        auto txt_ids     = linspace<float>(txt_id_start, context_len + txt_id_start, context_len);
+        std::vector<std::vector<float>> txt_ids_repeated(bs * context_len, std::vector<float>(3));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < txt_ids.size(); ++j) {
+                txt_ids_repeated[i * txt_ids.size() + j] = {txt_ids[j], txt_ids[j], txt_ids[j]};
+            }
+        }
+        auto img_ids = gen_img_ids(h, w, patch_size, bs);
+        auto ids     = concat_ids(txt_ids_repeated, img_ids, bs);
+        return ids;
+    }
+
+    // Generate qwen_image positional embeddings
+    static std::vector<float> gen_qwen_image_pe(int h,
+                                                int w,
+                                                int patch_size,
+                                                int bs,
+                                                int context_len,
+                                                int theta,
+                                                const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_qwen_image_ids(h, w, patch_size, bs, context_len);
+        return embed_nd(ids, bs, theta, axes_dim);
+    }
+
     static std::vector<std::vector<float>> gen_vid_ids(int t,
                                                        int h,
                                                        int w,

stable-diffusion.cpp

@@ -42,6 +42,7 @@ const char* model_version_to_str[] = {
     "Wan 2.x",
     "Wan 2.2 I2V",
     "Wan 2.2 TI2V",
+    "Qwen Image",
 };
 
 const char* sampling_methods_str[] = {
@@ -253,6 +254,13 @@ public:
             }
         }
 
+        if (strlen(SAFE_STR(sd_ctx_params->qwen2vl_path)) > 0) {
+            LOG_INFO("loading qwen2vl from '%s'", sd_ctx_params->qwen2vl_path);
+            if (!model_loader.init_from_file(sd_ctx_params->qwen2vl_path, "text_encoders.qwen2vl.")) {
+                LOG_WARN("loading qwen2vl from '%s' failed", sd_ctx_params->qwen2vl_path);
+            }
+        }
+
         if (strlen(SAFE_STR(sd_ctx_params->vae_path)) > 0) {
             LOG_INFO("loading vae from '%s'", sd_ctx_params->vae_path);
             if (!model_loader.init_from_file(sd_ctx_params->vae_path, "vae.")) {
@@ -318,7 +326,7 @@ public:
         } else if (sd_version_is_flux(version)) {
             scale_factor = 0.3611f;
             // TODO: shift_factor
-        } else if (sd_version_is_wan(version)) {
+        } else if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
             scale_factor = 1.0f;
         }
 
@@ -332,7 +340,7 @@ public:
         {
             clip_backend   = backend;
             bool use_t5xxl = false;
-            if (sd_version_is_dit(version)) {
+            if (sd_version_is_dit(version) && !sd_version_is_qwen_image(version)) {
                 use_t5xxl = true;
             }
             if (!clip_on_cpu && !ggml_backend_is_cpu(backend) && use_t5xxl) {
@@ -418,6 +426,16 @@ public:
                     clip_vision->alloc_params_buffer();
                     clip_vision->get_param_tensors(tensors);
                 }
+            } else if (sd_version_is_qwen_image(version)) {
+                cond_stage_model = std::make_shared<Qwen2_5_VLCLIPEmbedder>(clip_backend,
+                                                                            offload_params_to_cpu,
+                                                                            model_loader.tensor_storages_types);
+                diffusion_model  = std::make_shared<QwenImageModel>(backend,
+                                                                   offload_params_to_cpu,
+                                                                   model_loader.tensor_storages_types,
+                                                                   "model.diffusion_model",
+                                                                   version,
+                                                                   sd_ctx_params->diffusion_flash_attn);
             } else {  // SD1.x SD2.x SDXL
                 if (strstr(SAFE_STR(sd_ctx_params->photo_maker_path), "v2")) {
                     cond_stage_model = std::make_shared<FrozenCLIPEmbedderWithCustomWords>(clip_backend,
@@ -466,7 +484,7 @@ public:
                 vae_backend = backend;
             }
 
-            if (sd_version_is_wan(version)) {
+            if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
                 first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
                                                                         offload_params_to_cpu,
                                                                         model_loader.tensor_storages_types,
@@ -711,6 +729,13 @@ public:
                 shift = 5.0;
             }
             denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
+        } else if (sd_version_is_qwen_image(version)) {
+            LOG_INFO("running in FLOW mode");
+            float shift = sd_ctx_params->flow_shift;
+            if (shift == INFINITY) {
+                shift = 3.0;
+            }
+            denoiser = std::make_shared<DiscreteFlowDenoiser>(shift);
         } else if (is_using_v_parameterization) {
             LOG_INFO("running in v-prediction mode");
             denoiser = std::make_shared<CompVisVDenoiser>();
@@ -989,7 +1014,7 @@ public:
                     ggml_tensor_scale(noise, augmentation_level);
                     ggml_tensor_add(init_img, noise);
                 }
-                ggml_tensor* moments = encode_first_stage(work_ctx, init_img);
+                ggml_tensor* moments = vae_encode(work_ctx, init_img);
                 c_concat             = get_first_stage_encoding(work_ctx, moments);
             }
         }
@@ -1298,118 +1323,8 @@ public:
         return x;
     }
 
-    // ldm.models.diffusion.ddpm.LatentDiffusion.get_first_stage_encoding
-    ggml_tensor* get_first_stage_encoding(ggml_context* work_ctx, ggml_tensor* moments) {
-        // ldm.modules.distributions.distributions.DiagonalGaussianDistribution.sample
-        ggml_tensor* latent       = ggml_new_tensor_4d(work_ctx, moments->type, moments->ne[0], moments->ne[1], moments->ne[2] / 2, moments->ne[3]);
-        struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, latent);
-        ggml_tensor_set_f32_randn(noise, rng);
-        {
-            float mean   = 0;
-            float logvar = 0;
-            float value  = 0;
-            float std_   = 0;
-            for (int i = 0; i < latent->ne[3]; i++) {
-                for (int j = 0; j < latent->ne[2]; j++) {
-                    for (int k = 0; k < latent->ne[1]; k++) {
-                        for (int l = 0; l < latent->ne[0]; l++) {
-                            mean   = ggml_tensor_get_f32(moments, l, k, j, i);
-                            logvar = ggml_tensor_get_f32(moments, l, k, j + (int)latent->ne[2], i);
-                            logvar = std::max(-30.0f, std::min(logvar, 20.0f));
-                            std_   = std::exp(0.5f * logvar);
-                            value  = mean + std_ * ggml_tensor_get_f32(noise, l, k, j, i);
-                            value  = value * scale_factor;
-                            // printf("%d %d %d %d -> %f\n", i, j, k, l, value);
-                            ggml_tensor_set_f32(latent, value, l, k, j, i);
-                        }
-                    }
-                }
-            }
-        }
-        return latent;
-    }
-
-    void get_tile_sizes(int& tile_size_x,
-                        int& tile_size_y,
-                        float& tile_overlap,
-                        const sd_tiling_params_t& params,
-                        int latent_x,
-                        int latent_y,
-                        float encoding_factor = 1.0f) {
-        tile_overlap       = std::max(std::min(params.target_overlap, 0.5f), 0.0f);
-        auto get_tile_size = [&](int requested_size, float factor, int latent_size) {
-            const int default_tile_size  = 32;
-            const int min_tile_dimension = 4;
-            int tile_size                = default_tile_size;
-            // factor <= 1 means simple fraction of the latent dimension
-            // factor > 1 means number of tiles across that dimension
-            if (factor > 0.f) {
-                if (factor > 1.0)
-                    factor = 1 / (factor - factor * tile_overlap + tile_overlap);
-                tile_size = std::round(latent_size * factor);
-            } else if (requested_size >= min_tile_dimension) {
-                tile_size = requested_size;
-            }
-            tile_size *= encoding_factor;
-            return std::max(std::min(tile_size, latent_size), min_tile_dimension);
-        };
-
-        tile_size_x = get_tile_size(params.tile_size_x, params.rel_size_x, latent_x);
-        tile_size_y = get_tile_size(params.tile_size_y, params.rel_size_y, latent_y);
-    }
-
-    ggml_tensor* encode_first_stage(ggml_context* work_ctx, ggml_tensor* x, bool encode_video = false) {
-        int64_t t0          = ggml_time_ms();
-        ggml_tensor* result = NULL;
-        int W               = x->ne[0] / 8;
-        int H               = x->ne[1] / 8;
-        if (vae_tiling_params.enabled && !encode_video) {
-            // TODO wan2.2 vae support?
-            int C = sd_version_is_dit(version) ? 16 : 4;
-            if (!use_tiny_autoencoder) {
-                C *= 2;
-            }
-            result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, C, x->ne[3]);
-        }
-
-        if (!use_tiny_autoencoder) {
-            process_vae_input_tensor(x);
-            if (vae_tiling_params.enabled && !encode_video) {
-                float tile_overlap;
-                int tile_size_x, tile_size_y;
-                // multiply tile size for encode to keep the compute buffer size consistent
-                get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, vae_tiling_params, W, H, 1.30539f);
-
-                LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
-
-                auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                    first_stage_model->compute(n_threads, in, false, &out, work_ctx);
-                };
-                sd_tiling_non_square(x, result, 8, tile_size_x, tile_size_y, tile_overlap, on_tiling);
-            } else {
-                first_stage_model->compute(n_threads, x, false, &result, work_ctx);
-            }
-            first_stage_model->free_compute_buffer();
-        } else {
-            if (vae_tiling_params.enabled && !encode_video) {
-                // split latent in 32x32 tiles and compute in several steps
-                auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                    tae_first_stage->compute(n_threads, in, false, &out, NULL);
-                };
-                sd_tiling(x, result, 8, 64, 0.5f, on_tiling);
-            } else {
-                tae_first_stage->compute(n_threads, x, false, &result, work_ctx);
-            }
-            tae_first_stage->free_compute_buffer();
-        }
-
-        int64_t t1 = ggml_time_ms();
-        LOG_DEBUG("computing vae encode graph completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
-        return result;
-    }
-
     void process_latent_in(ggml_tensor* latent) {
-        if (sd_version_is_wan(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
             GGML_ASSERT(latent->ne[3] == 16 || latent->ne[3] == 48);
             std::vector<float> latents_mean_vec = {-0.7571f, -0.7089f, -0.9113f, 0.1075f, -0.1745f, 0.9653f, -0.1517f, 1.5508f,
                                                    0.4134f, -0.0715f, 0.5517f, -0.3632f, -0.1922f, -0.9497f, 0.2503f, -0.2921f};
@@ -1449,7 +1364,7 @@ public:
     }
 
     void process_latent_out(ggml_tensor* latent) {
-        if (sd_version_is_wan(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
             GGML_ASSERT(latent->ne[3] == 16 || latent->ne[3] == 48);
             std::vector<float> latents_mean_vec = {-0.7571f, -0.7089f, -0.9113f, 0.1075f, -0.1745f, 0.9653f, -0.1517f, 1.5508f,
                                                    0.4134f, -0.0715f, 0.5517f, -0.3632f, -0.1922f, -0.9497f, 0.2503f, -0.2921f};
@@ -1488,6 +1403,146 @@ public:
         }
     }
 
+    void get_tile_sizes(int& tile_size_x,
+                        int& tile_size_y,
+                        float& tile_overlap,
+                        const sd_tiling_params_t& params,
+                        int latent_x,
+                        int latent_y,
+                        float encoding_factor = 1.0f) {
+        tile_overlap       = std::max(std::min(params.target_overlap, 0.5f), 0.0f);
+        auto get_tile_size = [&](int requested_size, float factor, int latent_size) {
+            const int default_tile_size  = 32;
+            const int min_tile_dimension = 4;
+            int tile_size                = default_tile_size;
+            // factor <= 1 means simple fraction of the latent dimension
+            // factor > 1 means number of tiles across that dimension
+            if (factor > 0.f) {
+                if (factor > 1.0)
+                    factor = 1 / (factor - factor * tile_overlap + tile_overlap);
+                tile_size = std::round(latent_size * factor);
+            } else if (requested_size >= min_tile_dimension) {
+                tile_size = requested_size;
+            }
+            tile_size *= encoding_factor;
+            return std::max(std::min(tile_size, latent_size), min_tile_dimension);
+        };
+
+        tile_size_x = get_tile_size(params.tile_size_x, params.rel_size_x, latent_x);
+        tile_size_y = get_tile_size(params.tile_size_y, params.rel_size_y, latent_y);
+    }
+
+    ggml_tensor* vae_encode(ggml_context* work_ctx, ggml_tensor* x, bool encode_video = false) {
+        int64_t t0          = ggml_time_ms();
+        ggml_tensor* result = NULL;
+        int W               = x->ne[0] / 8;
+        int H               = x->ne[1] / 8;
+        if (vae_tiling_params.enabled && !encode_video) {
+            // TODO wan2.2 vae support?
+            int C = sd_version_is_dit(version) ? 16 : 4;
+            if (!use_tiny_autoencoder) {
+                C *= 2;
+            }
+            result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, C, x->ne[3]);
+        }
+
+        if (sd_version_is_qwen_image(version)) {
+            x = ggml_reshape_4d(work_ctx, x, x->ne[0], x->ne[1], 1, x->ne[2] * x->ne[3]);
+        }
+
+        if (!use_tiny_autoencoder) {
+            process_vae_input_tensor(x);
+            if (vae_tiling_params.enabled && !encode_video) {
+                float tile_overlap;
+                int tile_size_x, tile_size_y;
+                // multiply tile size for encode to keep the compute buffer size consistent
+                get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, vae_tiling_params, W, H, 1.30539f);
+
+                LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
+
+                auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
+                    first_stage_model->compute(n_threads, in, false, &out, work_ctx);
+                };
+                sd_tiling_non_square(x, result, 8, tile_size_x, tile_size_y, tile_overlap, on_tiling);
+            } else {
+                first_stage_model->compute(n_threads, x, false, &result, work_ctx);
+            }
+            first_stage_model->free_compute_buffer();
+        } else {
+            if (vae_tiling_params.enabled && !encode_video) {
+                // split latent in 32x32 tiles and compute in several steps
+                auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
+                    tae_first_stage->compute(n_threads, in, false, &out, NULL);
+                };
+                sd_tiling(x, result, 8, 64, 0.5f, on_tiling);
+            } else {
+                tae_first_stage->compute(n_threads, x, false, &result, work_ctx);
+            }
+            tae_first_stage->free_compute_buffer();
+        }
+
+        int64_t t1 = ggml_time_ms();
+        LOG_DEBUG("computing vae encode graph completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+        return result;
+    }
+
+    ggml_tensor* gaussian_latent_sample(ggml_context* work_ctx, ggml_tensor* moments) {
+        // ldm.modules.distributions.distributions.DiagonalGaussianDistribution.sample
+        ggml_tensor* latent       = ggml_new_tensor_4d(work_ctx, moments->type, moments->ne[0], moments->ne[1], moments->ne[2] / 2, moments->ne[3]);
+        struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, latent);
+        ggml_tensor_set_f32_randn(noise, rng);
+        {
+            float mean   = 0;
+            float logvar = 0;
+            float value  = 0;
+            float std_   = 0;
+            for (int i = 0; i < latent->ne[3]; i++) {
+                for (int j = 0; j < latent->ne[2]; j++) {
+                    for (int k = 0; k < latent->ne[1]; k++) {
+                        for (int l = 0; l < latent->ne[0]; l++) {
+                            mean   = ggml_tensor_get_f32(moments, l, k, j, i);
+                            logvar = ggml_tensor_get_f32(moments, l, k, j + (int)latent->ne[2], i);
+                            logvar = std::max(-30.0f, std::min(logvar, 20.0f));
+                            std_   = std::exp(0.5f * logvar);
+                            value  = mean + std_ * ggml_tensor_get_f32(noise, l, k, j, i);
+                            // printf("%d %d %d %d -> %f\n", i, j, k, l, value);
+                            ggml_tensor_set_f32(latent, value, l, k, j, i);
+                        }
+                    }
+                }
+            }
+        }
+        return latent;
+    }
+
+    ggml_tensor* get_first_stage_encoding(ggml_context* work_ctx, ggml_tensor* vae_output) {
+        ggml_tensor* latent;
+        if (use_tiny_autoencoder || sd_version_is_qwen_image(version) || sd_version_is_wan(version)) {
+            latent = vae_output;
+        } else if (version == VERSION_SD1_PIX2PIX) {
+            latent = ggml_view_3d(work_ctx,
+                                  vae_output,
+                                  vae_output->ne[0],
+                                  vae_output->ne[1],
+                                  vae_output->ne[2] / 2,
+                                  vae_output->nb[1],
+                                  vae_output->nb[2],
+                                  0);
+        } else {
+            latent = gaussian_latent_sample(work_ctx, vae_output);
+        }
+        process_latent_in(latent);
+        if (sd_version_is_qwen_image(version)) {
+            latent = ggml_reshape_4d(work_ctx, latent, latent->ne[0], latent->ne[1], latent->ne[3], 1);
+        }
+        return latent;
+    }
+
+    ggml_tensor* encode_first_stage(ggml_context* work_ctx, ggml_tensor* x, bool encode_video = false) {
+        ggml_tensor* vae_output = vae_encode(work_ctx, x, encode_video);
+        return get_first_stage_encoding(work_ctx, vae_output);
+    }
+
     ggml_tensor* decode_first_stage(ggml_context* work_ctx, ggml_tensor* x, bool decode_video = false) {
         int64_t W           = x->ne[0] * 8;
         int64_t H           = x->ne[1] * 8;
@@ -1518,6 +1573,9 @@ public:
         }
         int64_t t0 = ggml_time_ms();
         if (!use_tiny_autoencoder) {
+            if (sd_version_is_qwen_image(version)) {
+                x = ggml_reshape_4d(work_ctx, x, x->ne[0], x->ne[1], 1, x->ne[2] * x->ne[3]);
+            }
             process_latent_out(x);
             // x = load_tensor_from_file(work_ctx, "wan_vae_z.bin");
             if (vae_tiling_params.enabled && !decode_video) {
@@ -1689,6 +1747,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              "clip_g_path: %s\n"
              "clip_vision_path: %s\n"
              "t5xxl_path: %s\n"
+             "qwen2vl_path: %s\n"
              "diffusion_model_path: %s\n"
              "high_noise_diffusion_model_path: %s\n"
              "vae_path: %s\n"
@@ -1716,6 +1775,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              SAFE_STR(sd_ctx_params->clip_g_path),
              SAFE_STR(sd_ctx_params->clip_vision_path),
              SAFE_STR(sd_ctx_params->t5xxl_path),
+             SAFE_STR(sd_ctx_params->qwen2vl_path),
              SAFE_STR(sd_ctx_params->diffusion_model_path),
              SAFE_STR(sd_ctx_params->high_noise_diffusion_model_path),
              SAFE_STR(sd_ctx_params->vae_path),
@@ -2079,6 +2139,8 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
         C = 16;
     } else if (sd_version_is_flux(sd_ctx->sd->version)) {
         C = 16;
+    } else if (sd_version_is_qwen_image(sd_ctx->sd->version)) {
+        C = 16;
     }
     int W = width / 8;
     int H = height / 8;
@@ -2086,12 +2148,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
 
     struct ggml_tensor* control_latent = NULL;
     if (sd_version_is_control(sd_ctx->sd->version) && image_hint != NULL) {
-        if (!sd_ctx->sd->use_tiny_autoencoder) {
+        control_latent = sd_ctx->sd->encode_first_stage(work_ctx, image_hint);
-            struct ggml_tensor* control_moments = sd_ctx->sd->encode_first_stage(work_ctx, image_hint);
-            control_latent                      = sd_ctx->sd->get_first_stage_encoding(work_ctx, control_moments);
-        } else {
-            control_latent = sd_ctx->sd->encode_first_stage(work_ctx, image_hint);
-        }
         ggml_tensor_scale(control_latent, control_strength);
     }
 
@@ -2278,6 +2335,8 @@ ggml_tensor* generate_init_latent(sd_ctx_t* sd_ctx,
         C = 16;
     } else if (sd_version_is_flux(sd_ctx->sd->version)) {
         C = 16;
+    } else if (sd_version_is_qwen_image(sd_ctx->sd->version)) {
+        C = 16;
     } else if (sd_version_is_wan(sd_ctx->sd->version)) {
         C = 16;
         T = ((T - 1) / 4) + 1;
@@ -2354,7 +2413,6 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
     std::vector<float> sigmas = sd_ctx->sd->denoiser->get_sigmas(sample_steps);
 
     ggml_tensor* init_latent   = NULL;
-    ggml_tensor* init_moments  = NULL;
     ggml_tensor* concat_latent = NULL;
     ggml_tensor* denoise_mask  = NULL;
     if (sd_img_gen_params->init_image.data) {
@@ -2374,12 +2432,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
         sd_image_to_tensor(sd_img_gen_params->mask_image, mask_img);
         sd_image_to_tensor(sd_img_gen_params->init_image, init_img);
 
-        if (!sd_ctx->sd->use_tiny_autoencoder) {
+        init_latent = sd_ctx->sd->encode_first_stage(work_ctx, init_img);
-            init_moments = sd_ctx->sd->encode_first_stage(work_ctx, init_img);
-            init_latent  = sd_ctx->sd->get_first_stage_encoding(work_ctx, init_moments);
-        } else {
-            init_latent = sd_ctx->sd->encode_first_stage(work_ctx, init_img);
-        }
 
         if (sd_version_is_inpaint(sd_ctx->sd->version)) {
             int64_t mask_channels = 1;
@@ -2389,16 +2442,12 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
                 mask_channels = 1 + init_latent->ne[2];
             }
             ggml_tensor* masked_latent = NULL;
+
             if (sd_ctx->sd->version != VERSION_FLEX_2) {
                 // most inpaint models mask before vae
                 ggml_tensor* masked_img = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, height, 3, 1);
                 sd_apply_mask(init_img, mask_img, masked_img);
-                if (!sd_ctx->sd->use_tiny_autoencoder) {
+                masked_latent = sd_ctx->sd->encode_first_stage(work_ctx, masked_img);
-                    ggml_tensor* moments = sd_ctx->sd->encode_first_stage(work_ctx, masked_img);
-                    masked_latent        = sd_ctx->sd->get_first_stage_encoding(work_ctx, moments);
-                } else {
-                    masked_latent = sd_ctx->sd->encode_first_stage(work_ctx, masked_img);
-                }
             } else {
                 // mask after vae
                 masked_latent = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, init_latent->ne[0], init_latent->ne[1], init_latent->ne[2], 1);
@@ -2458,7 +2507,6 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
                 }
             }
         }
-
     } else {
         LOG_INFO("TXT2IMG");
         if (sd_version_is_inpaint(sd_ctx->sd->version)) {
@@ -2497,23 +2545,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
                                               1);
         sd_image_to_tensor(*ref_images[i], img);
 
-        ggml_tensor* latent = NULL;
+        ggml_tensor* latent = sd_ctx->sd->encode_first_stage(work_ctx, img);
-        if (sd_ctx->sd->use_tiny_autoencoder) {
-            latent = sd_ctx->sd->encode_first_stage(work_ctx, img);
-        } else if (sd_ctx->sd->version == VERSION_SD1_PIX2PIX) {
-            latent = sd_ctx->sd->encode_first_stage(work_ctx, img);
-            latent = ggml_view_3d(work_ctx,
-                                  latent,
-                                  latent->ne[0],
-                                  latent->ne[1],
-                                  latent->ne[2] / 2,
-                                  latent->nb[1],
-                                  latent->nb[2],
-                                  0);
-        } else {
-            ggml_tensor* moments = sd_ctx->sd->encode_first_stage(work_ctx, img);
-            latent               = sd_ctx->sd->get_first_stage_encoding(work_ctx, moments);
-        }
         ref_latents.push_back(latent);
     }
 
@@ -2675,8 +2707,6 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
         int64_t t2 = ggml_time_ms();
         LOG_INFO("encode_first_stage completed, taking %" PRId64 " ms", t2 - t1);
 
-        sd_ctx->sd->process_latent_in(concat_latent);
-
         ggml_tensor* concat_mask = ggml_new_tensor_4d(work_ctx,
                                                       GGML_TYPE_F32,
                                                       concat_latent->ne[0],
@@ -2702,7 +2732,7 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
         sd_image_to_tensor(sd_vid_gen_params->init_image, init_img);
         init_img = ggml_reshape_4d(work_ctx, init_img, width, height, 1, 3);
 
-        auto init_image_latent = sd_ctx->sd->encode_first_stage(work_ctx, init_img);  // [b*c, 1, h/16, w/16]
+        auto init_image_latent = sd_ctx->sd->vae_encode(work_ctx, init_img);  // [b*c, 1, h/16, w/16]
 
         init_latent  = generate_init_latent(sd_ctx, work_ctx, width, height, frames, true);
         denoise_mask = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, init_latent->ne[0], init_latent->ne[1], init_latent->ne[2], 1);
@@ -2733,7 +2763,6 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
             ref_img = ggml_reshape_4d(work_ctx, ref_img, width, height, 1, 3);
 
             ref_image_latent = sd_ctx->sd->encode_first_stage(work_ctx, ref_img);  // [b*c, 1, h/16, w/16]
-            sd_ctx->sd->process_latent_in(ref_image_latent);
             auto zero_latent = ggml_dup_tensor(work_ctx, ref_image_latent);
             ggml_set_f32(zero_latent, 0.f);
             ref_image_latent = ggml_tensor_concat(work_ctx, ref_image_latent, zero_latent, 3);  // [b*2*c, 1, h/16, w/16]
@@ -2765,9 +2794,6 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
         inactive = sd_ctx->sd->encode_first_stage(work_ctx, inactive);  // [b*c, t, h/8, w/8]
         reactive = sd_ctx->sd->encode_first_stage(work_ctx, reactive);  // [b*c, t, h/8, w/8]
 
-        sd_ctx->sd->process_latent_in(inactive);
-        sd_ctx->sd->process_latent_in(reactive);
-
         int64_t length = inactive->ne[2];
         if (ref_image_latent) {
             length += 1;

stable-diffusion.h

@@ -131,6 +131,7 @@ typedef struct {
     const char* clip_g_path;
     const char* clip_vision_path;
     const char* t5xxl_path;
+    const char* qwen2vl_path;
     const char* diffusion_model_path;
     const char* high_noise_diffusion_model_path;
     const char* vae_path;

tokenize_util.cpp

@@ -0,0 +1,985 @@
+#include <algorithm>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "tokenize_util.h"
+
+bool is_number(char32_t ch) {
+    return (ch >= U'0' && ch <= U'9');
+}
+
+bool is_letter(char32_t ch) {
+    static const struct { char32_t start, end; } ranges[] = {
+        {0x41, 0x5A},
+        {0x61, 0x7A},
+        {0xAA, 0xAA},
+        {0xB5, 0xB5},
+        {0xBA, 0xBA},
+        {0xC0, 0xD6},
+        {0xD8, 0xF6},
+        {0xF8, 0x2C1},
+        {0x2C6, 0x2D1},
+        {0x2E0, 0x2E4},
+        {0x2EC, 0x2EC},
+        {0x2EE, 0x2EE},
+        {0x370, 0x374},
+        {0x376, 0x377},
+        {0x37A, 0x37D},
+        {0x37F, 0x37F},
+        {0x386, 0x386},
+        {0x388, 0x38A},
+        {0x38C, 0x38C},
+        {0x38E, 0x3A1},
+        {0x3A3, 0x3F5},
+        {0x3F7, 0x481},
+        {0x48A, 0x52F},
+        {0x531, 0x556},
+        {0x559, 0x559},
+        {0x560, 0x588},
+        {0x5D0, 0x5EA},
+        {0x5EF, 0x5F2},
+        {0x620, 0x64A},
+        {0x66E, 0x66F},
+        {0x671, 0x6D3},
+        {0x6D5, 0x6D5},
+        {0x6E5, 0x6E6},
+        {0x6EE, 0x6EF},
+        {0x6FA, 0x6FC},
+        {0x6FF, 0x6FF},
+        {0x710, 0x710},
+        {0x712, 0x72F},
+        {0x74D, 0x7A5},
+        {0x7B1, 0x7B1},
+        {0x7CA, 0x7EA},
+        {0x7F4, 0x7F5},
+        {0x7FA, 0x7FA},
+        {0x800, 0x815},
+        {0x81A, 0x81A},
+        {0x824, 0x824},
+        {0x828, 0x828},
+        {0x840, 0x858},
+        {0x860, 0x86A},
+        {0x870, 0x887},
+        {0x889, 0x88F},
+        {0x8A0, 0x8C9},
+        {0x904, 0x939},
+        {0x93D, 0x93D},
+        {0x950, 0x950},
+        {0x958, 0x961},
+        {0x971, 0x980},
+        {0x985, 0x98C},
+        {0x98F, 0x990},
+        {0x993, 0x9A8},
+        {0x9AA, 0x9B0},
+        {0x9B2, 0x9B2},
+        {0x9B6, 0x9B9},
+        {0x9BD, 0x9BD},
+        {0x9CE, 0x9CE},
+        {0x9DC, 0x9DD},
+        {0x9DF, 0x9E1},
+        {0x9F0, 0x9F1},
+        {0x9FC, 0x9FC},
+        {0xA05, 0xA0A},
+        {0xA0F, 0xA10},
+        {0xA13, 0xA28},
+        {0xA2A, 0xA30},
+        {0xA32, 0xA33},
+        {0xA35, 0xA36},
+        {0xA38, 0xA39},
+        {0xA59, 0xA5C},
+        {0xA5E, 0xA5E},
+        {0xA72, 0xA74},
+        {0xA85, 0xA8D},
+        {0xA8F, 0xA91},
+        {0xA93, 0xAA8},
+        {0xAAA, 0xAB0},
+        {0xAB2, 0xAB3},
+        {0xAB5, 0xAB9},
+        {0xABD, 0xABD},
+        {0xAD0, 0xAD0},
+        {0xAE0, 0xAE1},
+        {0xAF9, 0xAF9},
+        {0xB05, 0xB0C},
+        {0xB0F, 0xB10},
+        {0xB13, 0xB28},
+        {0xB2A, 0xB30},
+        {0xB32, 0xB33},
+        {0xB35, 0xB39},
+        {0xB3D, 0xB3D},
+        {0xB5C, 0xB5D},
+        {0xB5F, 0xB61},
+        {0xB71, 0xB71},
+        {0xB83, 0xB83},
+        {0xB85, 0xB8A},
+        {0xB8E, 0xB90},
+        {0xB92, 0xB95},
+        {0xB99, 0xB9A},
+        {0xB9C, 0xB9C},
+        {0xB9E, 0xB9F},
+        {0xBA3, 0xBA4},
+        {0xBA8, 0xBAA},
+        {0xBAE, 0xBB9},
+        {0xBD0, 0xBD0},
+        {0xC05, 0xC0C},
+        {0xC0E, 0xC10},
+        {0xC12, 0xC28},
+        {0xC2A, 0xC39},
+        {0xC3D, 0xC3D},
+        {0xC58, 0xC5A},
+        {0xC5C, 0xC5D},
+        {0xC60, 0xC61},
+        {0xC80, 0xC80},
+        {0xC85, 0xC8C},
+        {0xC8E, 0xC90},
+        {0xC92, 0xCA8},
+        {0xCAA, 0xCB3},
+        {0xCB5, 0xCB9},
+        {0xCBD, 0xCBD},
+        {0xCDC, 0xCDE},
+        {0xCE0, 0xCE1},
+        {0xCF1, 0xCF2},
+        {0xD04, 0xD0C},
+        {0xD0E, 0xD10},
+        {0xD12, 0xD3A},
+        {0xD3D, 0xD3D},
+        {0xD4E, 0xD4E},
+        {0xD54, 0xD56},
+        {0xD5F, 0xD61},
+        {0xD7A, 0xD7F},
+        {0xD85, 0xD96},
+        {0xD9A, 0xDB1},
+        {0xDB3, 0xDBB},
+        {0xDBD, 0xDBD},
+        {0xDC0, 0xDC6},
+        {0xE01, 0xE30},
+        {0xE32, 0xE33},
+        {0xE40, 0xE46},
+        {0xE81, 0xE82},
+        {0xE84, 0xE84},
+        {0xE86, 0xE8A},
+        {0xE8C, 0xEA3},
+        {0xEA5, 0xEA5},
+        {0xEA7, 0xEB0},
+        {0xEB2, 0xEB3},
+        {0xEBD, 0xEBD},
+        {0xEC0, 0xEC4},
+        {0xEC6, 0xEC6},
+        {0xEDC, 0xEDF},
+        {0xF00, 0xF00},
+        {0xF40, 0xF47},
+        {0xF49, 0xF6C},
+        {0xF88, 0xF8C},
+        {0x1000, 0x102A},
+        {0x103F, 0x103F},
+        {0x1050, 0x1055},
+        {0x105A, 0x105D},
+        {0x1061, 0x1061},
+        {0x1065, 0x1066},
+        {0x106E, 0x1070},
+        {0x1075, 0x1081},
+        {0x108E, 0x108E},
+        {0x10A0, 0x10C5},
+        {0x10C7, 0x10C7},
+        {0x10CD, 0x10CD},
+        {0x10D0, 0x10FA},
+        {0x10FC, 0x1248},
+        {0x124A, 0x124D},
+        {0x1250, 0x1256},
+        {0x1258, 0x1258},
+        {0x125A, 0x125D},
+        {0x1260, 0x1288},
+        {0x128A, 0x128D},
+        {0x1290, 0x12B0},
+        {0x12B2, 0x12B5},
+        {0x12B8, 0x12BE},
+        {0x12C0, 0x12C0},
+        {0x12C2, 0x12C5},
+        {0x12C8, 0x12D6},
+        {0x12D8, 0x1310},
+        {0x1312, 0x1315},
+        {0x1318, 0x135A},
+        {0x1380, 0x138F},
+        {0x13A0, 0x13F5},
+        {0x13F8, 0x13FD},
+        {0x1401, 0x166C},
+        {0x166F, 0x167F},
+        {0x1681, 0x169A},
+        {0x16A0, 0x16EA},
+        {0x16F1, 0x16F8},
+        {0x1700, 0x1711},
+        {0x171F, 0x1731},
+        {0x1740, 0x1751},
+        {0x1760, 0x176C},
+        {0x176E, 0x1770},
+        {0x1780, 0x17B3},
+        {0x17D7, 0x17D7},
+        {0x17DC, 0x17DC},
+        {0x1820, 0x1878},
+        {0x1880, 0x1884},
+        {0x1887, 0x18A8},
+        {0x18AA, 0x18AA},
+        {0x18B0, 0x18F5},
+        {0x1900, 0x191E},
+        {0x1950, 0x196D},
+        {0x1970, 0x1974},
+        {0x1980, 0x19AB},
+        {0x19B0, 0x19C9},
+        {0x1A00, 0x1A16},
+        {0x1A20, 0x1A54},
+        {0x1AA7, 0x1AA7},
+        {0x1B05, 0x1B33},
+        {0x1B45, 0x1B4C},
+        {0x1B83, 0x1BA0},
+        {0x1BAE, 0x1BAF},
+        {0x1BBA, 0x1BE5},
+        {0x1C00, 0x1C23},
+        {0x1C4D, 0x1C4F},
+        {0x1C5A, 0x1C7D},
+        {0x1C80, 0x1C8A},
+        {0x1C90, 0x1CBA},
+        {0x1CBD, 0x1CBF},
+        {0x1CE9, 0x1CEC},
+        {0x1CEE, 0x1CF3},
+        {0x1CF5, 0x1CF6},
+        {0x1CFA, 0x1CFA},
+        {0x1D00, 0x1DBF},
+        {0x1E00, 0x1F15},
+        {0x1F18, 0x1F1D},
+        {0x1F20, 0x1F45},
+        {0x1F48, 0x1F4D},
+        {0x1F50, 0x1F57},
+        {0x1F59, 0x1F59},
+        {0x1F5B, 0x1F5B},
+        {0x1F5D, 0x1F5D},
+        {0x1F5F, 0x1F7D},
+        {0x1F80, 0x1FB4},
+        {0x1FB6, 0x1FBC},
+        {0x1FBE, 0x1FBE},
+        {0x1FC2, 0x1FC4},
+        {0x1FC6, 0x1FCC},
+        {0x1FD0, 0x1FD3},
+        {0x1FD6, 0x1FDB},
+        {0x1FE0, 0x1FEC},
+        {0x1FF2, 0x1FF4},
+        {0x1FF6, 0x1FFC},
+        {0x2071, 0x2071},
+        {0x207F, 0x207F},
+        {0x2090, 0x209C},
+        {0x2102, 0x2102},
+        {0x2107, 0x2107},
+        {0x210A, 0x2113},
+        {0x2115, 0x2115},
+        {0x2119, 0x211D},
+        {0x2124, 0x2124},
+        {0x2126, 0x2126},
+        {0x2128, 0x2128},
+        {0x212A, 0x212D},
+        {0x212F, 0x2139},
+        {0x213C, 0x213F},
+        {0x2145, 0x2149},
+        {0x214E, 0x214E},
+        {0x2183, 0x2184},
+        {0x2C00, 0x2CE4},
+        {0x2CEB, 0x2CEE},
+        {0x2CF2, 0x2CF3},
+        {0x2D00, 0x2D25},
+        {0x2D27, 0x2D27},
+        {0x2D2D, 0x2D2D},
+        {0x2D30, 0x2D67},
+        {0x2D6F, 0x2D6F},
+        {0x2D80, 0x2D96},
+        {0x2DA0, 0x2DA6},
+        {0x2DA8, 0x2DAE},
+        {0x2DB0, 0x2DB6},
+        {0x2DB8, 0x2DBE},
+        {0x2DC0, 0x2DC6},
+        {0x2DC8, 0x2DCE},
+        {0x2DD0, 0x2DD6},
+        {0x2DD8, 0x2DDE},
+        {0x2E2F, 0x2E2F},
+        {0x3005, 0x3006},
+        {0x3031, 0x3035},
+        {0x303B, 0x303C},
+        {0x3041, 0x3096},
+        {0x309D, 0x309F},
+        {0x30A1, 0x30FA},
+        {0x30FC, 0x30FF},
+        {0x3105, 0x312F},
+        {0x3131, 0x318E},
+        {0x31A0, 0x31BF},
+        {0x31F0, 0x31FF},
+        {0x3400, 0x4DBF},
+        {0x4E00, 0xA48C},
+        {0xA4D0, 0xA4FD},
+        {0xA500, 0xA60C},
+        {0xA610, 0xA61F},
+        {0xA62A, 0xA62B},
+        {0xA640, 0xA66E},
+        {0xA67F, 0xA69D},
+        {0xA6A0, 0xA6E5},
+        {0xA717, 0xA71F},
+        {0xA722, 0xA788},
+        {0xA78B, 0xA7DC},
+        {0xA7F1, 0xA801},
+        {0xA803, 0xA805},
+        {0xA807, 0xA80A},
+        {0xA80C, 0xA822},
+        {0xA840, 0xA873},
+        {0xA882, 0xA8B3},
+        {0xA8F2, 0xA8F7},
+        {0xA8FB, 0xA8FB},
+        {0xA8FD, 0xA8FE},
+        {0xA90A, 0xA925},
+        {0xA930, 0xA946},
+        {0xA960, 0xA97C},
+        {0xA984, 0xA9B2},
+        {0xA9CF, 0xA9CF},
+        {0xA9E0, 0xA9E4},
+        {0xA9E6, 0xA9EF},
+        {0xA9FA, 0xA9FE},
+        {0xAA00, 0xAA28},
+        {0xAA40, 0xAA42},
+        {0xAA44, 0xAA4B},
+        {0xAA60, 0xAA76},
+        {0xAA7A, 0xAA7A},
+        {0xAA7E, 0xAAAF},
+        {0xAAB1, 0xAAB1},
+        {0xAAB5, 0xAAB6},
+        {0xAAB9, 0xAABD},
+        {0xAAC0, 0xAAC0},
+        {0xAAC2, 0xAAC2},
+        {0xAADB, 0xAADD},
+        {0xAAE0, 0xAAEA},
+        {0xAAF2, 0xAAF4},
+        {0xAB01, 0xAB06},
+        {0xAB09, 0xAB0E},
+        {0xAB11, 0xAB16},
+        {0xAB20, 0xAB26},
+        {0xAB28, 0xAB2E},
+        {0xAB30, 0xAB5A},
+        {0xAB5C, 0xAB69},
+        {0xAB70, 0xABE2},
+        {0xAC00, 0xD7A3},
+        {0xD7B0, 0xD7C6},
+        {0xD7CB, 0xD7FB},
+        {0xF900, 0xFA6D},
+        {0xFA70, 0xFAD9},
+        {0xFB00, 0xFB06},
+        {0xFB13, 0xFB17},
+        {0xFB1D, 0xFB1D},
+        {0xFB1F, 0xFB28},
+        {0xFB2A, 0xFB36},
+        {0xFB38, 0xFB3C},
+        {0xFB3E, 0xFB3E},
+        {0xFB40, 0xFB41},
+        {0xFB43, 0xFB44},
+        {0xFB46, 0xFBB1},
+        {0xFBD3, 0xFD3D},
+        {0xFD50, 0xFD8F},
+        {0xFD92, 0xFDC7},
+        {0xFDF0, 0xFDFB},
+        {0xFE70, 0xFE74},
+        {0xFE76, 0xFEFC},
+        {0xFF21, 0xFF3A},
+        {0xFF41, 0xFF5A},
+        {0xFF66, 0xFFBE},
+        {0xFFC2, 0xFFC7},
+        {0xFFCA, 0xFFCF},
+        {0xFFD2, 0xFFD7},
+        {0xFFDA, 0xFFDC},
+        {0x10000, 0x1000B},
+        {0x1000D, 0x10026},
+        {0x10028, 0x1003A},
+        {0x1003C, 0x1003D},
+        {0x1003F, 0x1004D},
+        {0x10050, 0x1005D},
+        {0x10080, 0x100FA},
+        {0x10280, 0x1029C},
+        {0x102A0, 0x102D0},
+        {0x10300, 0x1031F},
+        {0x1032D, 0x10340},
+        {0x10342, 0x10349},
+        {0x10350, 0x10375},
+        {0x10380, 0x1039D},
+        {0x103A0, 0x103C3},
+        {0x103C8, 0x103CF},
+        {0x10400, 0x1049D},
+        {0x104B0, 0x104D3},
+        {0x104D8, 0x104FB},
+        {0x10500, 0x10527},
+        {0x10530, 0x10563},
+        {0x10570, 0x1057A},
+        {0x1057C, 0x1058A},
+        {0x1058C, 0x10592},
+        {0x10594, 0x10595},
+        {0x10597, 0x105A1},
+        {0x105A3, 0x105B1},
+        {0x105B3, 0x105B9},
+        {0x105BB, 0x105BC},
+        {0x105C0, 0x105F3},
+        {0x10600, 0x10736},
+        {0x10740, 0x10755},
+        {0x10760, 0x10767},
+        {0x10780, 0x10785},
+        {0x10787, 0x107B0},
+        {0x107B2, 0x107BA},
+        {0x10800, 0x10805},
+        {0x10808, 0x10808},
+        {0x1080A, 0x10835},
+        {0x10837, 0x10838},
+        {0x1083C, 0x1083C},
+        {0x1083F, 0x10855},
+        {0x10860, 0x10876},
+        {0x10880, 0x1089E},
+        {0x108E0, 0x108F2},
+        {0x108F4, 0x108F5},
+        {0x10900, 0x10915},
+        {0x10920, 0x10939},
+        {0x10940, 0x10959},
+        {0x10980, 0x109B7},
+        {0x109BE, 0x109BF},
+        {0x10A00, 0x10A00},
+        {0x10A10, 0x10A13},
+        {0x10A15, 0x10A17},
+        {0x10A19, 0x10A35},
+        {0x10A60, 0x10A7C},
+        {0x10A80, 0x10A9C},
+        {0x10AC0, 0x10AC7},
+        {0x10AC9, 0x10AE4},
+        {0x10B00, 0x10B35},
+        {0x10B40, 0x10B55},
+        {0x10B60, 0x10B72},
+        {0x10B80, 0x10B91},
+        {0x10C00, 0x10C48},
+        {0x10C80, 0x10CB2},
+        {0x10CC0, 0x10CF2},
+        {0x10D00, 0x10D23},
+        {0x10D4A, 0x10D65},
+        {0x10D6F, 0x10D85},
+        {0x10E80, 0x10EA9},
+        {0x10EB0, 0x10EB1},
+        {0x10EC2, 0x10EC7},
+        {0x10F00, 0x10F1C},
+        {0x10F27, 0x10F27},
+        {0x10F30, 0x10F45},
+        {0x10F70, 0x10F81},
+        {0x10FB0, 0x10FC4},
+        {0x10FE0, 0x10FF6},
+        {0x11003, 0x11037},
+        {0x11071, 0x11072},
+        {0x11075, 0x11075},
+        {0x11083, 0x110AF},
+        {0x110D0, 0x110E8},
+        {0x11103, 0x11126},
+        {0x11144, 0x11144},
+        {0x11147, 0x11147},
+        {0x11150, 0x11172},
+        {0x11176, 0x11176},
+        {0x11183, 0x111B2},
+        {0x111C1, 0x111C4},
+        {0x111DA, 0x111DA},
+        {0x111DC, 0x111DC},
+        {0x11200, 0x11211},
+        {0x11213, 0x1122B},
+        {0x1123F, 0x11240},
+        {0x11280, 0x11286},
+        {0x11288, 0x11288},
+        {0x1128A, 0x1128D},
+        {0x1128F, 0x1129D},
+        {0x1129F, 0x112A8},
+        {0x112B0, 0x112DE},
+        {0x11305, 0x1130C},
+        {0x1130F, 0x11310},
+        {0x11313, 0x11328},
+        {0x1132A, 0x11330},
+        {0x11332, 0x11333},
+        {0x11335, 0x11339},
+        {0x1133D, 0x1133D},
+        {0x11350, 0x11350},
+        {0x1135D, 0x11361},
+        {0x11380, 0x11389},
+        {0x1138B, 0x1138B},
+        {0x1138E, 0x1138E},
+        {0x11390, 0x113B5},
+        {0x113B7, 0x113B7},
+        {0x113D1, 0x113D1},
+        {0x113D3, 0x113D3},
+        {0x11400, 0x11434},
+        {0x11447, 0x1144A},
+        {0x1145F, 0x11461},
+        {0x11480, 0x114AF},
+        {0x114C4, 0x114C5},
+        {0x114C7, 0x114C7},
+        {0x11580, 0x115AE},
+        {0x115D8, 0x115DB},
+        {0x11600, 0x1162F},
+        {0x11644, 0x11644},
+        {0x11680, 0x116AA},
+        {0x116B8, 0x116B8},
+        {0x11700, 0x1171A},
+        {0x11740, 0x11746},
+        {0x11800, 0x1182B},
+        {0x118A0, 0x118DF},
+        {0x118FF, 0x11906},
+        {0x11909, 0x11909},
+        {0x1190C, 0x11913},
+        {0x11915, 0x11916},
+        {0x11918, 0x1192F},
+        {0x1193F, 0x1193F},
+        {0x11941, 0x11941},
+        {0x119A0, 0x119A7},
+        {0x119AA, 0x119D0},
+        {0x119E1, 0x119E1},
+        {0x119E3, 0x119E3},
+        {0x11A00, 0x11A00},
+        {0x11A0B, 0x11A32},
+        {0x11A3A, 0x11A3A},
+        {0x11A50, 0x11A50},
+        {0x11A5C, 0x11A89},
+        {0x11A9D, 0x11A9D},
+        {0x11AB0, 0x11AF8},
+        {0x11BC0, 0x11BE0},
+        {0x11C00, 0x11C08},
+        {0x11C0A, 0x11C2E},
+        {0x11C40, 0x11C40},
+        {0x11C72, 0x11C8F},
+        {0x11D00, 0x11D06},
+        {0x11D08, 0x11D09},
+        {0x11D0B, 0x11D30},
+        {0x11D46, 0x11D46},
+        {0x11D60, 0x11D65},
+        {0x11D67, 0x11D68},
+        {0x11D6A, 0x11D89},
+        {0x11D98, 0x11D98},
+        {0x11DB0, 0x11DDB},
+        {0x11EE0, 0x11EF2},
+        {0x11F02, 0x11F02},
+        {0x11F04, 0x11F10},
+        {0x11F12, 0x11F33},
+        {0x11FB0, 0x11FB0},
+        {0x12000, 0x12399},
+        {0x12480, 0x12543},
+        {0x12F90, 0x12FF0},
+        {0x13000, 0x1342F},
+        {0x13441, 0x13446},
+        {0x13460, 0x143FA},
+        {0x14400, 0x14646},
+        {0x16100, 0x1611D},
+        {0x16800, 0x16A38},
+        {0x16A40, 0x16A5E},
+        {0x16A70, 0x16ABE},
+        {0x16AD0, 0x16AED},
+        {0x16B00, 0x16B2F},
+        {0x16B40, 0x16B43},
+        {0x16B63, 0x16B77},
+        {0x16B7D, 0x16B8F},
+        {0x16D40, 0x16D6C},
+        {0x16E40, 0x16E7F},
+        {0x16EA0, 0x16EB8},
+        {0x16EBB, 0x16ED3},
+        {0x16F00, 0x16F4A},
+        {0x16F50, 0x16F50},
+        {0x16F93, 0x16F9F},
+        {0x16FE0, 0x16FE1},
+        {0x16FE3, 0x16FE3},
+        {0x16FF2, 0x16FF3},
+        {0x17000, 0x18CD5},
+        {0x18CFF, 0x18D1E},
+        {0x18D80, 0x18DF2},
+        {0x1AFF0, 0x1AFF3},
+        {0x1AFF5, 0x1AFFB},
+        {0x1AFFD, 0x1AFFE},
+        {0x1B000, 0x1B122},
+        {0x1B132, 0x1B132},
+        {0x1B150, 0x1B152},
+        {0x1B155, 0x1B155},
+        {0x1B164, 0x1B167},
+        {0x1B170, 0x1B2FB},
+        {0x1BC00, 0x1BC6A},
+        {0x1BC70, 0x1BC7C},
+        {0x1BC80, 0x1BC88},
+        {0x1BC90, 0x1BC99},
+        {0x1D400, 0x1D454},
+        {0x1D456, 0x1D49C},
+        {0x1D49E, 0x1D49F},
+        {0x1D4A2, 0x1D4A2},
+        {0x1D4A5, 0x1D4A6},
+        {0x1D4A9, 0x1D4AC},
+        {0x1D4AE, 0x1D4B9},
+        {0x1D4BB, 0x1D4BB},
+        {0x1D4BD, 0x1D4C3},
+        {0x1D4C5, 0x1D505},
+        {0x1D507, 0x1D50A},
+        {0x1D50D, 0x1D514},
+        {0x1D516, 0x1D51C},
+        {0x1D51E, 0x1D539},
+        {0x1D53B, 0x1D53E},
+        {0x1D540, 0x1D544},
+        {0x1D546, 0x1D546},
+        {0x1D54A, 0x1D550},
+        {0x1D552, 0x1D6A5},
+        {0x1D6A8, 0x1D6C0},
+        {0x1D6C2, 0x1D6DA},
+        {0x1D6DC, 0x1D6FA},
+        {0x1D6FC, 0x1D714},
+        {0x1D716, 0x1D734},
+        {0x1D736, 0x1D74E},
+        {0x1D750, 0x1D76E},
+        {0x1D770, 0x1D788},
+        {0x1D78A, 0x1D7A8},
+        {0x1D7AA, 0x1D7C2},
+        {0x1D7C4, 0x1D7CB},
+        {0x1DF00, 0x1DF1E},
+        {0x1DF25, 0x1DF2A},
+        {0x1E030, 0x1E06D},
+        {0x1E100, 0x1E12C},
+        {0x1E137, 0x1E13D},
+        {0x1E14E, 0x1E14E},
+        {0x1E290, 0x1E2AD},
+        {0x1E2C0, 0x1E2EB},
+        {0x1E4D0, 0x1E4EB},
+        {0x1E5D0, 0x1E5ED},
+        {0x1E5F0, 0x1E5F0},
+        {0x1E6C0, 0x1E6DE},
+        {0x1E6E0, 0x1E6E2},
+        {0x1E6E4, 0x1E6E5},
+        {0x1E6E7, 0x1E6ED},
+        {0x1E6F0, 0x1E6F4},
+        {0x1E6FE, 0x1E6FF},
+        {0x1E7E0, 0x1E7E6},
+        {0x1E7E8, 0x1E7EB},
+        {0x1E7ED, 0x1E7EE},
+        {0x1E7F0, 0x1E7FE},
+        {0x1E800, 0x1E8C4},
+        {0x1E900, 0x1E943},
+        {0x1E94B, 0x1E94B},
+        {0x1EE00, 0x1EE03},
+        {0x1EE05, 0x1EE1F},
+        {0x1EE21, 0x1EE22},
+        {0x1EE24, 0x1EE24},
+        {0x1EE27, 0x1EE27},
+        {0x1EE29, 0x1EE32},
+        {0x1EE34, 0x1EE37},
+        {0x1EE39, 0x1EE39},
+        {0x1EE3B, 0x1EE3B},
+        {0x1EE42, 0x1EE42},
+        {0x1EE47, 0x1EE47},
+        {0x1EE49, 0x1EE49},
+        {0x1EE4B, 0x1EE4B},
+        {0x1EE4D, 0x1EE4F},
+        {0x1EE51, 0x1EE52},
+        {0x1EE54, 0x1EE54},
+        {0x1EE57, 0x1EE57},
+        {0x1EE59, 0x1EE59},
+        {0x1EE5B, 0x1EE5B},
+        {0x1EE5D, 0x1EE5D},
+        {0x1EE5F, 0x1EE5F},
+        {0x1EE61, 0x1EE62},
+        {0x1EE64, 0x1EE64},
+        {0x1EE67, 0x1EE6A},
+        {0x1EE6C, 0x1EE72},
+        {0x1EE74, 0x1EE77},
+        {0x1EE79, 0x1EE7C},
+        {0x1EE7E, 0x1EE7E},
+        {0x1EE80, 0x1EE89},
+        {0x1EE8B, 0x1EE9B},
+        {0x1EEA1, 0x1EEA3},
+        {0x1EEA5, 0x1EEA9},
+        {0x1EEAB, 0x1EEBB},
+        {0x20000, 0x2A6DF},
+        {0x2A700, 0x2B81D},
+        {0x2B820, 0x2CEAD},
+        {0x2CEB0, 0x2EBE0},
+        {0x2EBF0, 0x2EE5D},
+        {0x2F800, 0x2FA1D},
+        {0x30000, 0x3134A},
+        {0x31350, 0x33479},
+    };
+
+    for (const auto& r : ranges) {
+        if (ch >= r.start && ch <= r.end)
+            return true;
+    }
+    return false;
+}
+
+bool is_space(char32_t cp) {
+    switch (cp) {
+        case 0x0009:  // TAB \t
+        case 0x000A:  // LF \n
+        case 0x000B:  // VT
+        case 0x000C:  // FF
+        case 0x000D:  // CR \r
+        case 0x0020:  // Space
+        case 0x00A0:  // No-Break Space
+        case 0x1680:  // Ogham Space Mark
+        case 0x2000:  // En Quad
+        case 0x2001:  // Em Quad
+        case 0x2002:  // En Space
+        case 0x2003:  // Em Space
+        case 0x2004:  // Three-Per-Em Space
+        case 0x2005:  // Four-Per-Em Space
+        case 0x2006:  // Six-Per-Em Space
+        case 0x2007:  // Figure Space
+        case 0x2008:  // Punctuation Space
+        case 0x2009:  // Thin Space
+        case 0x200A:  // Hair Space
+        case 0x202F:  // Narrow No-Break Space
+        case 0x205F:  // Medium Mathematical Space
+        case 0x3000:  // Ideographic Space
+            return true;
+        default:
+            return false;
+    }
+}
+
+std::string str_to_lower(const std::string& input) {
+    std::string result = input;
+    std::transform(result.begin(), result.end(), result.begin(),
+                   [](unsigned char c) { return std::tolower(c); });
+    return result;
+}
+
+// UTF-8 -> Unicode code points
+std::vector<char32_t> utf8_to_codepoints(const std::string& str) {
+    std::vector<char32_t> codepoints;
+    size_t i = 0;
+    while (i < str.size()) {
+        unsigned char c    = str[i];
+        char32_t cp        = 0;
+        size_t extra_bytes = 0;
+
+        if ((c & 0x80) == 0)
+            cp = c;
+        else if ((c & 0xE0) == 0xC0) {
+            cp          = c & 0x1F;
+            extra_bytes = 1;
+        } else if ((c & 0xF0) == 0xE0) {
+            cp          = c & 0x0F;
+            extra_bytes = 2;
+        } else if ((c & 0xF8) == 0xF0) {
+            cp          = c & 0x07;
+            extra_bytes = 3;
+        } else {
+            ++i;
+            continue;
+        }  // Invalid UTF-8
+
+        if (i + extra_bytes >= str.size())
+            break;
+
+        for (size_t j = 1; j <= extra_bytes; ++j)
+            cp = (cp << 6) | (str[i + j] & 0x3F);
+
+        codepoints.push_back(cp);
+        i += 1 + extra_bytes;
+    }
+    return codepoints;
+}
+
+// Unicode code point -> UTF-8
+std::string codepoint_to_utf8(char32_t cp) {
+    std::string out;
+    if (cp <= 0x7F)
+        out.push_back(static_cast<char>(cp));
+    else if (cp <= 0x7FF) {
+        out.push_back(static_cast<char>(0xC0 | (cp >> 6)));
+        out.push_back(static_cast<char>(0x80 | (cp & 0x3F)));
+    } else if (cp <= 0xFFFF) {
+        out.push_back(static_cast<char>(0xE0 | (cp >> 12)));
+        out.push_back(static_cast<char>(0x80 | ((cp >> 6) & 0x3F)));
+        out.push_back(static_cast<char>(0x80 | (cp & 0x3F)));
+    } else {
+        out.push_back(static_cast<char>(0xF0 | (cp >> 18)));
+        out.push_back(static_cast<char>(0x80 | ((cp >> 12) & 0x3F)));
+        out.push_back(static_cast<char>(0x80 | ((cp >> 6) & 0x3F)));
+        out.push_back(static_cast<char>(0x80 | (cp & 0x3F)));
+    }
+    return out;
+}
+
+bool starts_with(const std::vector<char32_t>& text,
+                 const std::vector<char32_t>& prefix,
+                 std::size_t index) {
+    if (index > text.size()) {
+        return false;
+    }
+    if (prefix.size() > text.size() - index) {
+        return false;
+    }
+    return std::equal(prefix.begin(), prefix.end(), text.begin() + index);
+}
+
+std::vector<std::string> token_split(const std::string& text) {
+    std::vector<std::string> tokens;
+    auto cps = utf8_to_codepoints(text);
+    size_t i = 0;
+
+    while (i < cps.size()) {
+        char32_t cp = cps[i];
+
+        // `(?i:'s|'t|'re|'ve|'m|'ll|'d)`
+        if (cp == U'\'' && i + 1 < cps.size()) {
+            std::string next = str_to_lower(codepoint_to_utf8(cps[i + 1]));
+            if (next == "s" || next == "t" || next == "m") {
+                tokens.push_back("'" + next);
+                i += 2;
+                continue;
+            }
+            if (i + 2 < cps.size()) {
+                next += str_to_lower(codepoint_to_utf8(cps[i + 2]));
+                if (next == "re" || next == "ve" || next == "ll" || next == "d") {
+                    tokens.push_back("'" + next);
+                    i += 3;
+                    continue;
+                }
+            }
+        }
+
+        // `\p{N}`
+        if (is_number(cp)) {
+            tokens.push_back(codepoint_to_utf8(cp));
+            ++i;
+            continue;
+        }
+
+        // `[^\r\n\p{L}\p{N}]?\p{L}+`
+        {
+            // `[^\r\n\p{L}\p{N}]\p{L}+`
+            if (!is_letter(cp) && cp != U'\r' && cp != U'\n' && i + 1 < cps.size() && is_letter(cps[i + 1])) {
+                std::string token = codepoint_to_utf8(cp);
+                ++i;
+
+                while (i < cps.size() && is_letter(cps[i])) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+                tokens.push_back(token);
+                continue;
+            }
+
+            // `\p{L}+`
+            if (is_letter(cp)) {
+                std::string token = codepoint_to_utf8(cp);
+                ++i;
+                while (i < cps.size() && is_letter(cps[i])) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+                tokens.push_back(token);
+                continue;
+            }
+        }
+
+        // ` ?[^\s\p{L}\p{N}]+[\r\n]*`
+        {
+            // ` [^\s\p{L}\p{N}]+[\r\n]*`
+            if (cp == U' ' && i + 1 < cps.size() && !isspace(cps[i + 1]) && !is_letter(cps[i + 1]) && !is_number(cps[i + 1])) {
+                std::string token = codepoint_to_utf8(cp);
+                token += codepoint_to_utf8(cps[i + 1]);
+                i += 2;
+
+                while (i < cps.size() && !is_letter(cps[i]) && !is_number(cps[i]) && !isspace(cps[i])) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+
+                while (i < cps.size() && (cps[i] == U'\r' || cps[i] == U'\n')) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+
+                tokens.push_back(token);
+                continue;
+            }
+
+            // `[^\s\p{L}\p{N}]+[\r\n]*`
+            std::string token;
+            if (!is_letter(cps[i]) && !is_number(cps[i]) && !isspace(cps[i])) {
+                std::string token = codepoint_to_utf8(cp);
+                ++i;
+
+                while (i < cps.size() && !is_letter(cps[i]) && !is_number(cps[i]) && !isspace(cps[i])) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+
+                while (i < cps.size() && (cps[i] == U'\r' || cps[i] == U'\n')) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+
+                tokens.push_back(token);
+                continue;
+            }
+        }
+
+        // `\s*[\r\n]+|\s+(?!\S)|\s+`
+        if (is_space(cp)) {
+            std::string token = codepoint_to_utf8(cp);
+            ++i;
+
+            while (i < cps.size() && is_space(cps[i])) {
+                token += codepoint_to_utf8(cps[i]);
+                ++i;
+                if (cps[i] == U'\r' || cps[i] == U'\n') {
+                    break;
+                }
+            }
+
+            tokens.push_back(token);
+            continue;
+        }
+
+        // skip
+        ++i;
+    }
+
+    return tokens;
+}
+
+std::vector<std::string> split_with_special_tokens(
+    const std::string& text,
+    const std::vector<std::string>& special_tokens) {
+    std::vector<std::string> result;
+    size_t pos      = 0;
+    size_t text_len = text.size();
+
+    while (pos < text_len) {
+        size_t next_pos = text_len;
+        std::string matched_token;
+
+        for (const auto& token : special_tokens) {
+            size_t token_pos = text.find(token, pos);
+            if (token_pos != std::string::npos && token_pos < next_pos) {
+                next_pos      = token_pos;
+                matched_token = token;
+            }
+        }
+
+        if (next_pos > pos) {
+            result.push_back(text.substr(pos, next_pos - pos));
+        }
+
+        if (!matched_token.empty()) {
+            result.push_back(matched_token);
+            pos = next_pos + matched_token.size();
+        } else {
+            break;
+        }
+    }
+
+    return result;
+}
+
+// int main() {
+//     std::string text = "I'm testing C++ token_split function. 你好，世界! 123";
+//     auto tokens = token_split(text);
+
+//     for (const auto& t : tokens) {
+//         std::cout << "[" << t << "] ";
+//     }
+//     std::cout << "\n";
+//     return 0;
+// }

tokenize_util.h

@@ -0,0 +1,10 @@
+#ifndef __TOKENIZE_UTIL__
+#define __TOKENIZE_UTIL__
+
+#include <string>
+#include <vector>
+
+std::vector<std::string> token_split(const std::string& text);
+std::vector<std::string> split_with_special_tokens(const std::string& text, const std::vector<std::string>& special_tokens);
+
+#endif  // __TOKENIZE_UTIL__

wan.hpp

@@ -1833,7 +1833,7 @@ namespace WAN {
                                               struct ggml_tensor* x) {
             int64_t W = x->ne[0];
             int64_t H = x->ne[1];
-            int64_t T = x->ne[1];
+            int64_t T = x->ne[2];
 
             int pad_t = (std::get<0>(params.patch_size) - T % std::get<0>(params.patch_size)) % std::get<0>(params.patch_size);
             int pad_h = (std::get<1>(params.patch_size) - H % std::get<1>(params.patch_size)) % std::get<1>(params.patch_size);