feat: add flux2 support (#1016)

* add flux2 support

* rename qwenvl to llm

* add Flux2FlowDenoiser

* update docs

Author: leejet

README.md

@@ -37,7 +37,8 @@ API and command-line option may change frequently.***
     - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
     - [Some SD1.x and SDXL distilled models](./docs/distilled_sd.md)
     - [SD3/SD3.5](./docs/sd3.md)
-    - [Flux-dev/Flux-schnell](./docs/flux.md)
+    - [FlUX.1-dev/FlUX.1-schnell](./docs/flux.md)
+    - [FLUX.2-dev](./docs/flux2.md)
     - [Chroma](./docs/chroma.md)
     - [Chroma1-Radiance](./docs/chroma_radiance.md)
     - [Qwen Image](./docs/qwen_image.md)
@@ -118,7 +119,8 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 
 - [SD1.x/SD2.x/SDXL](./docs/sd.md)
 - [SD3/SD3.5](./docs/sd3.md)
-- [Flux-dev/Flux-schnell](./docs/flux.md)
+- [FlUX.1-dev/FlUX.1-schnell](./docs/flux.md)
+- [FLUX.2-dev](./docs/flux2.md)
 - [FLUX.1-Kontext-dev](./docs/kontext.md)
 - [Chroma](./docs/chroma.md)
 - [🔥Qwen Image](./docs/qwen_image.md)

assets/flux2/example.png

@@ -2,7 +2,7 @@
 #define __CONDITIONER_HPP__
 
 #include "clip.hpp"
-#include "qwenvl.hpp"
+#include "llm.hpp"
 #include "t5.hpp"
 
 struct SDCondition {
@@ -1623,61 +1623,72 @@ struct T5CLIPEmbedder : public Conditioner {
     }
 };
 
-struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
-    Qwen::Qwen2Tokenizer tokenizer;
-    std::shared_ptr<Qwen::Qwen2_5_VLRunner> qwenvl;
-
-    Qwen2_5_VLCLIPEmbedder(ggml_backend_t backend,
-                           bool offload_params_to_cpu,
-                           const String2TensorStorage& tensor_storage_map = {},
-                           const std::string prefix                       = "",
-                           bool enable_vision                             = false) {
-        qwenvl = std::make_shared<Qwen::Qwen2_5_VLRunner>(backend,
-                                                          offload_params_to_cpu,
-                                                          tensor_storage_map,
-                                                          "text_encoders.qwen2vl",
-                                                          enable_vision);
+struct LLMEmbedder : public Conditioner {
+    SDVersion version;
+    std::shared_ptr<LLM::BPETokenizer> tokenizer;
+    std::shared_ptr<LLM::LLMRunner> llm;
+
+    LLMEmbedder(ggml_backend_t backend,
+                bool offload_params_to_cpu,
+                const String2TensorStorage& tensor_storage_map = {},
+                SDVersion version                              = VERSION_QWEN_IMAGE,
+                const std::string prefix                       = "",
+                bool enable_vision                             = false)
+        : version(version) {
+        LLM::LLMArch arch = LLM::LLMArch::QWEN2_5_VL;
+        if (sd_version_is_flux2(version)) {
+            arch = LLM::LLMArch::MISTRAL_SMALL_3_2;
+        }
+        if (arch == LLM::LLMArch::MISTRAL_SMALL_3_2) {
+            tokenizer = std::make_shared<LLM::MistralTokenizer>();
+        } else {
+            tokenizer = std::make_shared<LLM::Qwen2Tokenizer>();
+        }
+        llm = std::make_shared<LLM::LLMRunner>(arch,
+                                               backend,
+                                               offload_params_to_cpu,
+                                               tensor_storage_map,
+                                               "text_encoders.llm",
+                                               enable_vision);
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
-        qwenvl->get_param_tensors(tensors, "text_encoders.qwen2vl");
+        llm->get_param_tensors(tensors, "text_encoders.llm");
     }
 
     void alloc_params_buffer() override {
-        qwenvl->alloc_params_buffer();
+        llm->alloc_params_buffer();
     }
 
     void free_params_buffer() override {
-        qwenvl->free_params_buffer();
+        llm->free_params_buffer();
     }
 
     size_t get_params_buffer_size() override {
         size_t buffer_size = 0;
-        buffer_size += qwenvl->get_params_buffer_size();
+        buffer_size += llm->get_params_buffer_size();
         return buffer_size;
     }
 
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        if (qwenvl) {
-            qwenvl->set_weight_adapter(adapter);
+        if (llm) {
+            llm->set_weight_adapter(adapter);
         }
     }
 
     std::tuple<std::vector<int>, std::vector<float>> tokenize(std::string text,
-                                                              size_t max_length           = 0,
-                                                              size_t system_prompt_length = 0,
-                                                              bool padding                = false) {
+                                                              std::pair<int, int> attn_range,
+                                                              size_t max_length = 0,
+                                                              bool padding      = false) {
         std::vector<std::pair<std::string, float>> parsed_attention;
-        if (system_prompt_length > 0) {
-            parsed_attention.emplace_back(text.substr(0, system_prompt_length), 1.f);
-            auto new_parsed_attention = parse_prompt_attention(text.substr(system_prompt_length, text.size() - system_prompt_length));
+        parsed_attention.emplace_back(text.substr(0, attn_range.first), 1.f);
+        if (attn_range.second - attn_range.first > 0) {
+            auto new_parsed_attention = parse_prompt_attention(text.substr(attn_range.first, attn_range.second - attn_range.first));
             parsed_attention.insert(parsed_attention.end(),
                                     new_parsed_attention.begin(),
                                     new_parsed_attention.end());
-        } else {
-            parsed_attention = parse_prompt_attention(text);
         }
-
+        parsed_attention.emplace_back(text.substr(attn_range.second), 1.f);
         {
             std::stringstream ss;
             ss << "[";
@@ -1693,12 +1704,12 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
         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);
+            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);
+        tokenizer->pad_tokens(tokens, weights, max_length, padding);
 
         // for (int i = 0; i < tokens.size(); i++) {
         //     std::cout << tokens[i] << ":" << weights[i] << ", " << i << std::endl;
@@ -1713,9 +1724,10 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
                                       const ConditionerParams& conditioner_params) override {
         std::string prompt;
         std::vector<std::pair<int, ggml_tensor*>> image_embeds;
-        size_t system_prompt_length          = 0;
+        std::pair<int, int> prompt_attn_range;
         int prompt_template_encode_start_idx = 34;
-        if (qwenvl->enable_vision && conditioner_params.ref_images.size() > 0) {
+        std::set<int> out_layers;
+        if (llm->enable_vision && conditioner_params.ref_images.size() > 0) {
             LOG_INFO("QwenImageEditPlusPipeline");
             prompt_template_encode_start_idx = 64;
             int image_embed_idx              = 64 + 6;
@@ -1727,7 +1739,7 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
 
             for (int i = 0; i < conditioner_params.ref_images.size(); i++) {
                 sd_image_f32_t image = sd_image_t_to_sd_image_f32_t(*conditioner_params.ref_images[i]);
-                double factor        = qwenvl->params.vision.patch_size * qwenvl->params.vision.spatial_merge_size;
+                double factor        = llm->params.vision.patch_size * llm->params.vision.spatial_merge_size;
                 int height           = image.height;
                 int width            = image.width;
                 int h_bar            = static_cast<int>(std::round(height / factor)) * factor;
@@ -1757,7 +1769,7 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
                 resized_image.data = nullptr;
 
                 ggml_tensor* image_embed = nullptr;
-                qwenvl->encode_image(n_threads, image_tensor, &image_embed, work_ctx);
+                llm->encode_image(n_threads, image_tensor, &image_embed, work_ctx);
                 image_embeds.emplace_back(image_embed_idx, image_embed);
                 image_embed_idx += 1 + image_embed->ne[1] + 6;
 
@@ -1771,17 +1783,37 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
             }
 
             prompt = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n";
-
-            system_prompt_length = prompt.size();
-
             prompt += img_prompt;
+
+            prompt_attn_range.first = prompt.size();
             prompt += conditioner_params.text;
+            prompt_attn_range.second = prompt.size();
+
             prompt += "<|im_end|>\n<|im_start|>assistant\n";
+        } else if (sd_version_is_flux2(version)) {
+            prompt_template_encode_start_idx = 0;
+            out_layers                       = {10, 20, 30};
+
+            prompt = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]";
+
+            prompt_attn_range.first = prompt.size();
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = prompt.size();
+
+            prompt += "[/INST]";
         } else {
-            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" + conditioner_params.text + "<|im_end|>\n<|im_start|>assistant\n";
+            prompt_template_encode_start_idx = 34;
+
+            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";
+
+            prompt_attn_range.first = prompt.size();
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = prompt.size();
+
+            prompt += "<|im_end|>\n<|im_start|>assistant\n";
         }
 
-        auto tokens_and_weights = tokenize(prompt, 0, system_prompt_length, false);
+        auto tokens_and_weights = tokenize(prompt, prompt_attn_range, 0, false);
         auto& tokens            = std::get<0>(tokens_and_weights);
         auto& weights           = std::get<1>(tokens_and_weights);
 
@@ -1790,11 +1822,12 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
 
         auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
 
-        qwenvl->compute(n_threads,
-                        input_ids,
-                        image_embeds,
-                        &hidden_states,
-                        work_ctx);
+        llm->compute(n_threads,
+                     input_ids,
+                     image_embeds,
+                     out_layers,
+                     &hidden_states,
+                     work_ctx);
         {
             auto tensor         = hidden_states;
             float original_mean = ggml_ext_tensor_mean(tensor);
@@ -1813,14 +1846,25 @@ struct Qwen2_5_VLCLIPEmbedder : public Conditioner {
 
         GGML_ASSERT(hidden_states->ne[1] > prompt_template_encode_start_idx);
 
+        int64_t zero_pad_len = 0;
+        if (sd_version_is_flux2(version)) {
+            int64_t min_length = 512;
+            if (hidden_states->ne[1] - prompt_template_encode_start_idx < min_length) {
+                zero_pad_len = min_length - 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[1] - prompt_template_encode_start_idx + zero_pad_len,
                                                             hidden_states->ne[2]);
 
         ggml_ext_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_ext_tensor_get_f32(hidden_states, i0, i1 + prompt_template_encode_start_idx, i2, i3);
+            float value = 0.f;
+            if (i1 + prompt_template_encode_start_idx < hidden_states->ne[1]) {
+                value = ggml_ext_tensor_get_f32(hidden_states, i0, i1 + prompt_template_encode_start_idx, i2, i3);
+            }
             ggml_ext_tensor_set_f32(new_hidden_states, value, i0, i1, i2, i3);
         });
 

conditioner.hpp

@@ -356,7 +356,7 @@ struct Denoiser {
     virtual ggml_tensor* noise_scaling(float sigma, ggml_tensor* noise, ggml_tensor* latent) = 0;
     virtual ggml_tensor* inverse_noise_scaling(float sigma, ggml_tensor* latent)             = 0;
 
-    virtual std::vector<float> get_sigmas(uint32_t n, scheduler_t scheduler_type, SDVersion version) {
+    virtual std::vector<float> get_sigmas(uint32_t n, int /*image_seq_len*/, scheduler_t scheduler_type, SDVersion version) {
         auto bound_t_to_sigma = std::bind(&Denoiser::t_to_sigma, this, std::placeholders::_1);
         std::shared_ptr<SigmaScheduler> scheduler;
         switch (scheduler_type) {
@@ -582,10 +582,14 @@ struct FluxFlowDenoiser : public Denoiser {
         set_parameters(shift);
     }
 
-    void set_parameters(float shift = 1.15f) {
+    void set_shift(float shift) {
         this->shift = shift;
-        for (int i = 1; i < TIMESTEPS + 1; i++) {
-            sigmas[i - 1] = t_to_sigma(i / TIMESTEPS * TIMESTEPS);
+    }
+
+    void set_parameters(float shift) {
+        set_shift(shift);
+        for (int i = 0; i < TIMESTEPS; i++) {
+            sigmas[i] = t_to_sigma(i);
         }
     }
 
@@ -627,6 +631,38 @@ struct FluxFlowDenoiser : public Denoiser {
     }
 };
 
+struct Flux2FlowDenoiser : public FluxFlowDenoiser {
+    Flux2FlowDenoiser() = default;
+
+    float compute_empirical_mu(uint32_t n, int image_seq_len) {
+        const float a1 = 8.73809524e-05f;
+        const float b1 = 1.89833333f;
+        const float a2 = 0.00016927f;
+        const float b2 = 0.45666666f;
+
+        if (image_seq_len > 4300) {
+            float mu = a2 * image_seq_len + b2;
+            return mu;
+        }
+
+        float m_200 = a2 * image_seq_len + b2;
+        float m_10  = a1 * image_seq_len + b1;
+
+        float a  = (m_200 - m_10) / 190.0f;
+        float b  = m_200 - 200.0f * a;
+        float mu = a * n + b;
+
+        return mu;
+    }
+
+    std::vector<float> get_sigmas(uint32_t n, int image_seq_len, scheduler_t scheduler_type, SDVersion version) override {
+        float mu = compute_empirical_mu(n, image_seq_len);
+        LOG_DEBUG("Flux2FlowDenoiser: set shift to %.3f", mu);
+        set_shift(mu);
+        return Denoiser::get_sigmas(n, image_seq_len, scheduler_type, version);
+    }
+};
+
 typedef std::function<ggml_tensor*(ggml_tensor*, float, int)> denoise_cb_t;
 
 // k diffusion reverse ODE: dx = (x - D(x;\sigma)) / \sigma dt; \sigma(t) = t

denoiser.hpp

@@ -0,0 +1,21 @@
+# How to Use
+
+## Download weights
+
+- Download FLUX.2-dev
+    - gguf: https://huggingface.co/city96/FLUX.2-dev-gguf/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download Mistral-Small-3.2-24B-Instruct-2506-GGUF
+    - gguf: https://huggingface.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF/tree/main
+
+## Examples
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\flux2-dev-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf -r .\kontext_input.png -p "change 'flux.cpp' to 'flux2-dev.cpp'" --cfg-scale 1.0 --sampling-method euler -v --diffusion-fa --offload-to-cpu
+```
+
+<img alt="flux2 example" src="../assets/flux2/example.png" />
+
+
+

docs/flux2.md

@@ -14,7 +14,7 @@
 ## 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
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\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" />