fix: avoid writable mmap for read-only weights (#1698)

README.md

@@ -50,12 +50,14 @@ API and command-line option may change frequently.***
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
     - [ERNIE-Image](./docs/ernie_image.md)
+    - [Boogu Image](./docs/boogu_image.md)
     - [HiDream-O1-Image](./docs/hidream_o1_image.md)
     - [Ideogram4](./docs/ideogram4.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
     - [LongCat Image Edit](./docs/longcat_image.md)
+    - [Boogu Image Edit](./docs/boogu_image.md)
   - Video Models
     - [Wan2.1/Wan2.2](./docs/wan.md)
     - [LTX-2.3](./docs/ltx2.md)

docs/boogu_image.md

@@ -0,0 +1,31 @@
+# How to Use
+
+Boogu Image uses a Boogu diffusion transformer, the FLUX VAE, and Qwen3-VL as the LLM text and vision encoder.
+
+## Download weights
+
+- Download Boogu Image
+    - safetensors: https://huggingface.co/Comfy-Org/Boogu-Image/tree/main/diffusion_models
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download Qwen3-VL 8B
+    - gguf: https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/tree/main
+        - For image editing with GGUF text encoders, also download the matching mmproj file and pass it with `--llm_vision`.
+
+## Examples
+
+### Boogu Image Base
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\boogu_image_base_bf16.safetensors --llm ..\..\llm\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ae.sft -p "a lovely cat" --diffusion-fa -v --offload-to-cpu
+```
+
+<img width="256" alt="Boogu Image Base example" src="../assets/boogu/example.png" />
+
+### Boogu Image Edit
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\boogu_image_edit_bf16.safetensors --llm ..\..\llm\Qwen3VL-8B-Instruct-Q4_K_M.gguf --llm_vision ..\..\llm\mmproj-Qwen3VL-8B-Instruct-F16.gguf --vae ..\..\ComfyUI\models\vae\ae.sft --diffusion-fa -v --offload-to-cpu -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'boogu.cpp'"
+```
+
+<img width="256" alt="Boogu Image Edit example" src="../assets/boogu/edit_example.png" />

examples/cli/main.cpp

@@ -62,18 +62,22 @@ struct SDCliParams {
             {"-o",
              "--output",
              "path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs support .avi, .webm, and animated .webp",
+             0,
              &output_path},
             {"",
              "--image",
              "path to the image to inspect (for metadata mode)",
+             0,
              &image_path},
             {"",
              "--metadata-format",
              "metadata output format, one of [text, json] (default: text)",
+             0,
              &metadata_format},
             {"",
              "--preview-path",
              "path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp",
+             0,
              &preview_path},
         };
 

examples/common/common.cpp

@@ -6,6 +6,7 @@
 #include <cstdlib>
 #include <ctime>
 #include <filesystem>
+#include <fstream>
 #include <iomanip>
 #include <iostream>
 #include <regex>
@@ -260,8 +261,15 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
                         invalid_arg = true;
                         return;
                     }
-                    *option.target = argv_to_utf8(i, argv);
+                    if (option.concat && !option.target->empty()) {
-                    found_arg      = true;
+                        if (option.concat > 0 && option.concat <= 0xff) {
+                            *option.target += static_cast<char>(option.concat);
+                        }
+                        *option.target += argv_to_utf8(i, argv);
+                    } else {
+                        *option.target = argv_to_utf8(i, argv);
+                    }
+                    found_arg = true;
                 }))
                 break;
 
@@ -324,120 +332,151 @@ ArgOptions SDContextParams::get_options() {
         {"-m",
          "--model",
          "path to full model",
+         0,
          &model_path},
         {"",
          "--clip_l",
-         "path to the clip-l text encoder", &clip_l_path},
+         "path to the clip-l text encoder",
+         0,
+         &clip_l_path},
         {"", "--clip_g",
          "path to the clip-g text encoder",
+         0,
          &clip_g_path},
         {"",
          "--clip_vision",
          "path to the clip-vision encoder",
+         0,
          &clip_vision_path},
         {"",
          "--t5xxl",
          "path to the t5xxl text encoder",
+         0,
          &t5xxl_path},
         {"",
          "--llm",
          "path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)",
+         0,
          &llm_path},
         {"",
          "--llm_vision",
          "path to the llm vit",
+         0,
          &llm_vision_path},
         {"",
          "--qwen2vl",
          "alias of --llm. Deprecated.",
+         0,
          &llm_path},
         {"",
          "--qwen2vl_vision",
          "alias of --llm_vision. Deprecated.",
+         0,
          &llm_vision_path},
         {"",
          "--diffusion-model",
          "path to the standalone diffusion model",
+         0,
          &diffusion_model_path},
         {"",
          "--high-noise-diffusion-model",
          "path to the standalone high noise diffusion model",
+         0,
          &high_noise_diffusion_model_path},
         {"",
          "--uncond-diffusion-model",
          "path to the standalone unconditional diffusion model, currently used by Ideogram4 CFG",
+         0,
          &uncond_diffusion_model_path},
         {"",
          "--embeddings-connectors",
          "path to LTXAV embeddings connectors",
+         0,
          &embeddings_connectors_path},
         {"",
          "--vae",
          "path to standalone vae model",
+         0,
          &vae_path},
         {"",
          "--vae-format",
          "VAE latent format override: auto, flux, sd3, or flux2 (default: auto)",
+         0,
          &vae_format},
         {"",
          "--audio-vae",
          "path to standalone LTX audio vae model",
+         0,
          &audio_vae_path},
         {"",
          "--taesd",
          "path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)",
+         0,
          &taesd_path},
         {"",
          "--tae",
          "alias of --taesd",
+         0,
          &taesd_path},
         {"",
          "--control-net",
          "path to control net model",
+         0,
          &control_net_path},
         {"",
          "--embd-dir",
          "embeddings directory",
+         0,
          &embedding_dir},
         {"",
          "--lora-model-dir",
          "lora model directory",
+         0,
          &lora_model_dir},
         {"",
          "--hires-upscalers-dir",
          "highres fix upscaler model directory",
+         0,
          &hires_upscalers_dir},
         {"",
          "--tensor-type-rules",
          "weight type per tensor pattern (example: \"^vae\\.=f16,model\\.=q8_0\")",
+         (int)',',
          &tensor_type_rules},
         {"",
          "--photo-maker",
          "path to PHOTOMAKER model",
+         0,
          &photo_maker_path},
         {"",
          "--pulid-weights",
          "path to PuLID Flux weights",
+         0,
          &pulid_weights_path},
         {"",
          "--upscale-model",
          "path to esrgan model.",
+         0,
          &esrgan_path},
         {"",
          "--backend",
          "runtime backend assignment, e.g. cpu or clip=cpu,vae=cuda0,diffusion=vulkan0",
+         (int)',',
          &backend},
         {"",
          "--params-backend",
          "parameter backend assignment, e.g. disk, cpu, or diffusion=disk,clip=cpu",
+         (int)',',
          &params_backend},
         {"",
          "--rpc-servers",
          "comma-separated list of RPC servers to connect to for offloading, in the format host:port, e.g. localhost:50052,192.168.1.3:50052",
+         (int)',',
          &rpc_servers},
         {"",
          "--max-vram",
          "maximum VRAM budget in GiB for graph-cut segmented execution. Accepts a single value or assignments by backend/device, e.g. 6 or cuda0=6,vulkan0=4. 0 disables graph splitting; a negative value auto-detects free VRAM, sparing the specified value",
+         0,
          &max_vram},
     };
 
@@ -458,6 +497,10 @@ ArgOptions SDContextParams::get_options() {
          "--stream-layers",
          "enable residency+prefetch streaming on top of --max-vram (no effect without --max-vram; defaults to false)",
          true, &stream_layers},
+        {"",
+         "--eager-load",
+         "load all params into the params backend at model-load time instead of lazily on first use (defaults to false)",
+         true, &eager_load},
         {"",
          "--force-sdxl-vae-conv-scale",
          "force use of conv scale on sdxl vae",
@@ -761,6 +804,7 @@ std::string SDContextParams::to_string() const {
         << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
         << "  max_vram: \"" << max_vram << "\",\n"
         << "  stream_layers: " << (stream_layers ? "true" : "false") << ",\n"
+        << "  eager_load: " << (eager_load ? "true" : "false") << ",\n"
         << "  backend: \"" << backend << "\",\n"
         << "  params_backend: \"" << params_backend << "\",\n"
         << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
@@ -840,6 +884,7 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.vae_format                      = str_to_vae_format(vae_format);
     sd_ctx_params.max_vram                        = max_vram.c_str();
     sd_ctx_params.stream_layers                   = stream_layers;
+    sd_ctx_params.eager_load                      = eager_load;
     sd_ctx_params.backend                         = effective_backend.c_str();
     sd_ctx_params.params_backend                  = effective_params_backend.c_str();
     sd_ctx_params.rpc_servers                     = rpc_servers.c_str();
@@ -857,58 +902,71 @@ ArgOptions SDGenerationParams::get_options() {
         {"-p",
          "--prompt",
          "the prompt to render",
+         0,
          &prompt},
         {"-n",
          "--negative-prompt",
          "the negative prompt (default: \"\")",
+         0,
          &negative_prompt},
         {"-i",
          "--init-img",
          "path to the init image",
+         0,
          &init_image_path},
         {"",
          "--end-img",
          "path to the end image, required by flf2v",
+         0,
          &end_image_path},
         {"",
          "--mask",
          "path to the mask image",
+         0,
          &mask_image_path},
         {"",
          "--control-image",
          "path to control image, control net",
+         0,
          &control_image_path},
         {"",
          "--control-video",
          "path to control video frames, It must be a directory path. The video frames inside should be stored as images in "
          "lexicographical (character) order. For example, if the control video path is `frames`, the directory contain images "
          "such as 00.png, 01.png, ... etc.",
+         0,
          &control_video_path},
         {"",
          "--pm-id-images-dir",
          "path to PHOTOMAKER input id images dir",
+         0,
          &pm_id_images_dir},
         {"",
          "--pm-id-embed-path",
          "path to PHOTOMAKER v2 id embed",
+         0,
          &pm_id_embed_path},
         {"",
          "--pulid-id-embedding",
          "path to PuLID id embedding",
+         0,
          &pulid_id_embedding_path},
         {"",
          "--hires-upscaler",
          "highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent (nearest-exact), "
          "Latent (antialiased), Latent (bicubic), Latent (bicubic antialiased), or a model name "
          "under --hires-upscalers-dir (default: Latent)",
+         0,
          &hires_upscaler},
         {"",
          "--extra-sample-args",
-         "extra sampler/scheduler/guidance args, key=value list. APG supports apg_eta, apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma",
+         "extra sampler/scheduler/guidance args, key=value list. CFG supports guidance_schedule; APG supports apg_eta, apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma;",
+         (int)',',
          &extra_sample_args},
         {"",
          "--extra-tiling-args",
          "extra VAE tiling args, key=value list. LTX video VAE supports temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)",
+         (int)',',
          &extra_tiling_args},
     };
 
@@ -1364,6 +1422,42 @@ ArgOptions SDGenerationParams::get_options() {
         return 1;
     };
 
+    auto on_prompt_file_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        const char* arg = argv[index];
+        std::ifstream f(arg, std::ios::binary);
+        try {
+            prompt = std::string(std::istreambuf_iterator<char>{f}, {});
+        } catch (const std::ios_base::failure&) {
+            f.setstate(std::ios_base::failbit);
+        }
+        if (f.fail()) {
+            LOG_ERROR("error: failed to read prompt file '%s'\n", arg);
+            return -1;
+        }
+        return 1;
+    };
+
+    auto on_negative_prompt_file_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        const char* arg = argv[index];
+        std::ifstream f(arg, std::ios::binary);
+        try {
+            negative_prompt = std::string(std::istreambuf_iterator<char>{f}, {});
+        } catch (const std::ios_base::failure&) {
+            f.setstate(std::ios_base::failbit);
+        }
+        if (f.fail()) {
+            LOG_ERROR("error: failed to read negative prompt file '%s'\n", arg);
+            return -1;
+        }
+        return 1;
+    };
+
     options.manual_options = {
         {"-s",
          "--seed",
@@ -1427,6 +1521,14 @@ ArgOptions SDGenerationParams::get_options() {
          "--vae-relative-tile-size",
          "relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)",
          on_relative_tile_size_arg},
+        {"",
+         "--prompt-file",
+         "path to the file containing the prompt to render",
+         on_prompt_file_arg},
+        {"",
+         "--negative-prompt-file",
+         "path to the file containing the negative prompt",
+         on_negative_prompt_file_arg},
 
     };
 

examples/common/common.h

@@ -31,6 +31,7 @@ struct StringOption {
     std::string short_name;
     std::string long_name;
     std::string desc;
+    int concat;
     std::string* target;
 };
 
@@ -147,6 +148,7 @@ struct SDContextParams {
     bool offload_params_to_cpu  = false;
     std::string max_vram        = "0";
     bool stream_layers          = false;
+    bool eager_load             = false;
     std::string backend;
     std::string params_backend;
     std::string rpc_servers;

examples/server/runtime.cpp

@@ -190,8 +190,8 @@ ArgOptions SDSvrParams::get_options() {
     ArgOptions options;
 
     options.string_options = {
-        {"-l", "--listen-ip", "server listen ip (default: 127.0.0.1)", &listen_ip},
+        {"-l", "--listen-ip", "server listen ip (default: 127.0.0.1)", 0, &listen_ip},
-        {"", "--serve-html-path", "path to HTML file to serve at root (optional)", &serve_html_path},
+        {"", "--serve-html-path", "path to HTML file to serve at root (optional)", 0, &serve_html_path},
     };
 
     options.int_options = {

include/stable-diffusion.h

@@ -219,6 +219,7 @@ typedef struct {
     enum sd_vae_format_t vae_format;
     const char* max_vram;  // GiB budget or backend assignment spec for graph-cut segmented param offload (0 = disabled, -1 = auto)
     bool stream_layers;  // Enable residency+prefetch streaming on top of --max-vram (no effect without --max-vram)
+    bool eager_load;  // Load all params into the params backend at model-load time instead of lazily on first use
     const char* backend;
     const char* params_backend;
     const char* rpc_servers;

src/conditioning/conditioner.hpp

@@ -1518,7 +1518,7 @@ struct LLMEmbedder : public Conditioner {
             arch = LLM::LLMArch::GPT_OSS_20B;
         } else if (sd_version_is_pid(version)) {
             arch = LLM::LLMArch::GEMMA2_2B;
-        } else if (sd_version_is_ideogram4(version)) {
+        } else if (sd_version_is_ideogram4(version) || sd_version_is_boogu_image(version)) {
             arch = LLM::LLMArch::QWEN3_VL;
         } else if (sd_version_is_z_image(version) || version == VERSION_OVIS_IMAGE || version == VERSION_FLUX2_KLEIN) {
             arch = LLM::LLMArch::QWEN3;
@@ -1778,6 +1778,65 @@ struct LLMEmbedder : public Conditioner {
 
                 prompt += "<|im_end|>\n<|im_start|>assistant\n";
             }
+        } else if (sd_version_is_boogu_image(version)) {
+            prompt_template_encode_start_idx = 0;
+
+            const std::string t2i_system_prompt =
+                "You are a helpful assistant that generates high-quality images based on user instructions. The instructions are as follows.";
+            const std::string edit_system_prompt =
+                "Describe 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.";
+            const bool has_ref_images = llm->enable_vision && conditioner_params.ref_images != nullptr && !conditioner_params.ref_images->empty();
+            const bool text_empty     = conditioner_params.text.find_first_not_of(" \t\r\n") == std::string::npos;
+
+            if (has_ref_images) {
+                LOG_INFO("BooguImageEditPipeline");
+                const std::string prompt_prefix = "<|im_start|>system\n" + edit_system_prompt + "<|im_end|>\n<|im_start|>user\n";
+                std::string img_prompt;
+                const std::string placeholder = "<|image_pad|>";
+
+                for (int i = 0; i < conditioner_params.ref_images->size(); i++) {
+                    const auto& image = (*conditioner_params.ref_images)[i];
+                    double factor     = llm->config.vision.patch_size * llm->config.vision.spatial_merge_size;
+                    int height        = static_cast<int>(image.shape()[1]);
+                    int width         = static_cast<int>(image.shape()[0]);
+                    double beta       = std::sqrt((384.0 * 384.0) / (static_cast<double>(height) * static_cast<double>(width)));
+                    int h_bar         = std::max(static_cast<int>(factor),
+                                                 static_cast<int>(std::round(height * beta / factor)) * static_cast<int>(factor));
+                    int w_bar         = std::max(static_cast<int>(factor),
+                                                 static_cast<int>(std::round(width * beta / factor)) * static_cast<int>(factor));
+
+                    LOG_DEBUG("resize conditioner ref image %d from %dx%d to %dx%d", i, height, width, h_bar, w_bar);
+
+                    auto resized_image = clip_preprocess(image, w_bar, h_bar);
+                    auto image_embed   = llm->encode_image(n_threads, resized_image, false, true, true);
+                    GGML_ASSERT(!image_embed.empty());
+
+                    std::string image_prefix = prompt_prefix + img_prompt + "<|vision_start|>";
+                    int image_embed_idx      = static_cast<int>(tokenizer->encode(image_prefix, nullptr).size());
+                    image_embeds.emplace_back(image_embed_idx, image_embed);
+
+                    img_prompt += "<|vision_start|>";
+                    int64_t num_image_tokens = image_embed.shape()[1];
+                    img_prompt.reserve(img_prompt.size() + static_cast<size_t>(num_image_tokens) * placeholder.size() + 32);
+                    for (int j = 0; j < num_image_tokens; j++) {
+                        img_prompt += placeholder;
+                    }
+                    img_prompt += "<|vision_end|>";
+                }
+
+                prompt                  = prompt_prefix + img_prompt;
+                prompt_attn_range.first = static_cast<int>(prompt.size());
+                prompt += conditioner_params.text;
+                prompt_attn_range.second = static_cast<int>(prompt.size());
+                prompt += "<|im_end|>\n";
+            } else {
+                const std::string& system_prompt = text_empty ? edit_system_prompt : t2i_system_prompt;
+                prompt                           = "<|im_start|>system\n" + system_prompt + "<|im_end|>\n<|im_start|>user\n";
+                prompt_attn_range.first          = static_cast<int>(prompt.size());
+                prompt += conditioner_params.text;
+                prompt_attn_range.second = static_cast<int>(prompt.size());
+                prompt += "<|im_end|>\n";
+            }
         } else if (sd_version_is_longcat(version)) {
             spell_quotes = true;
 

src/model.h

@@ -42,6 +42,7 @@ enum SDVersion {
     VERSION_LTXAV,
     VERSION_HIDREAM_O1,
     VERSION_Z_IMAGE,
+    VERSION_BOOGU_IMAGE,
     VERSION_OVIS_IMAGE,
     VERSION_ERNIE_IMAGE,
     VERSION_LENS,
@@ -143,6 +144,13 @@ static inline bool sd_version_is_z_image(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_boogu_image(SDVersion version) {
+    if (version == VERSION_BOOGU_IMAGE) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_longcat(SDVersion version) {
     if (version == VERSION_LONGCAT) {
         return true;
@@ -178,6 +186,13 @@ static inline bool sd_version_is_ideogram4(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_uses_flux_vae(SDVersion version) {
+    if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_boogu_image(version) || sd_version_is_longcat(version)) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_uses_flux2_vae(SDVersion version) {
     if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version) || sd_version_is_lens(version) || sd_version_is_ideogram4(version)) {
         return true;
@@ -206,6 +221,7 @@ static inline bool sd_version_is_dit(SDVersion version) {
         version == VERSION_HIDREAM_O1 ||
         sd_version_is_anima(version) ||
         sd_version_is_z_image(version) ||
+        sd_version_is_boogu_image(version) ||
         sd_version_is_ernie_image(version) ||
         sd_version_is_lens(version) ||
         sd_version_is_longcat(version) ||

src/model/common/rope.hpp

@@ -899,10 +899,12 @@ namespace Rope {
         // q,k,v: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2]
         // return: [N, L, n_head*d_head]
+        int64_t n_head = q->ne[1];
+
         q = apply_rope(ctx->ggml_ctx, q, pe, rope_interleaved);  // [N*n_head, L, d_head]
         k = apply_rope(ctx->ggml_ctx, k, pe, rope_interleaved);  // [N*n_head, L, d_head]
 
-        auto x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, v->ne[1], mask, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*d_head]
+        auto x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, true, ctx->flash_attn_enabled, kv_scale);  // [N, L, n_head*d_head]
         return x;
     }
 };  // namespace Rope

src/model/diffusion/boogu.hpp

@@ -0,0 +1,835 @@
+#ifndef __SD_MODEL_DIFFUSION_BOOGU_HPP__
+#define __SD_MODEL_DIFFUSION_BOOGU_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <tuple>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model/diffusion/qwen_image.hpp"
+#include "model_loader.h"
+
+namespace Boogu {
+    constexpr int BOOGU_GRAPH_SIZE = 65536;
+
+    struct BooguConfig {
+        int patch_size                   = 2;
+        int64_t in_channels              = 16;
+        int64_t out_channels             = 16;
+        int64_t hidden_size              = 3360;
+        int64_t num_layers               = 32;
+        int64_t num_double_stream_layers = 8;
+        int64_t num_refiner_layers       = 2;
+        int64_t num_attention_heads      = 28;
+        int64_t num_kv_heads             = 7;
+        int64_t head_dim                 = 120;
+        int64_t multiple_of              = 256;
+        int64_t instruction_feat_dim     = 4096;
+        int64_t timestep_embed_dim       = 1024;
+        int theta                        = 10000;
+        float timestep_scale             = 1000.0f;
+        float norm_eps                   = 1e-5f;
+        std::vector<int> axes_dim        = {40, 40, 40};
+        int64_t axes_dim_sum             = 120;
+
+        static int64_t count_blocks(const String2TensorStorage& tensor_storage_map,
+                                    const std::string& prefix,
+                                    const std::string& block_prefix) {
+            int64_t count = 0;
+            for (const auto& [name, _] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                size_t pos = name.find(block_prefix);
+                if (pos == std::string::npos) {
+                    continue;
+                }
+                auto items = split_string(name.substr(pos), '.');
+                if (items.size() > 1) {
+                    count = std::max<int64_t>(count, atoi(items[1].c_str()) + 1);
+                }
+            }
+            return count;
+        }
+
+        static BooguConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            BooguConfig config;
+            int64_t detected_head_dim = 0;
+            int64_t detected_kv_dim   = 0;
+
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "x_embedder.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area = config.patch_size * config.patch_size;
+                    config.in_channels = tensor_storage.ne[0] / patch_area;
+                    config.hidden_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "time_caption_embed.caption_embedder.1.weight") && tensor_storage.n_dims == 2) {
+                    config.instruction_feat_dim = tensor_storage.ne[0];
+                    config.hidden_size          = tensor_storage.ne[1];
+                } else if (ends_with(name, "single_stream_layers.0.attn.norm_q.weight") && tensor_storage.n_dims == 1) {
+                    detected_head_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "double_stream_layers.0.img_self_attn.norm_q.weight") && tensor_storage.n_dims == 1) {
+                    detected_head_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "single_stream_layers.0.attn.to_k.weight") && tensor_storage.n_dims == 2) {
+                    detected_kv_dim = tensor_storage.ne[1];
+                } else if (ends_with(name, "double_stream_layers.0.img_instruct_attn.processor.img_to_k.weight") && tensor_storage.n_dims == 2) {
+                    detected_kv_dim = tensor_storage.ne[1];
+                } else if (ends_with(name, "norm_out.linear_2.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area  = config.patch_size * config.patch_size;
+                    config.out_channels = tensor_storage.ne[1] / patch_area;
+                }
+            }
+
+            config.num_layers               = std::max<int64_t>(1, count_blocks(tensor_storage_map, prefix, "single_stream_layers."));
+            config.num_double_stream_layers = std::max<int64_t>(0, count_blocks(tensor_storage_map, prefix, "double_stream_layers."));
+            int64_t noise_refiner_layers    = count_blocks(tensor_storage_map, prefix, "noise_refiner.");
+            int64_t ref_refiner_layers      = count_blocks(tensor_storage_map, prefix, "ref_image_refiner.");
+            int64_t context_refiner_layers  = count_blocks(tensor_storage_map, prefix, "context_refiner.");
+            config.num_refiner_layers       = std::max<int64_t>(1, std::max(noise_refiner_layers, std::max(ref_refiner_layers, context_refiner_layers)));
+
+            if (detected_head_dim > 0) {
+                config.head_dim            = detected_head_dim;
+                config.num_attention_heads = config.hidden_size / config.head_dim;
+                config.axes_dim_sum        = config.head_dim;
+                if (detected_kv_dim > 0) {
+                    config.num_kv_heads = detected_kv_dim / config.head_dim;
+                }
+                if (config.axes_dim_sum == 120) {
+                    config.axes_dim = {40, 40, 40};
+                } else if (config.axes_dim_sum % 3 == 0) {
+                    int axis        = static_cast<int>(config.axes_dim_sum / 3);
+                    config.axes_dim = {axis, axis, axis};
+                }
+            }
+            config.timestep_embed_dim = std::min<int64_t>(config.hidden_size, 1024);
+
+            LOG_DEBUG("boogu_image: layers=%" PRId64 ", double_stream_layers=%" PRId64 ", refiner_layers=%" PRId64 ", hidden=%" PRId64 ", heads=%" PRId64 ", kv_heads=%" PRId64 ", head_dim=%" PRId64 ", in_channels=%" PRId64 ", out_channels=%" PRId64,
+                      config.num_layers,
+                      config.num_double_stream_layers,
+                      config.num_refiner_layers,
+                      config.hidden_size,
+                      config.num_attention_heads,
+                      config.num_kv_heads,
+                      config.head_dim,
+                      config.in_channels,
+                      config.out_channels);
+            return config;
+        }
+    };
+
+    __STATIC_INLINE__ ggml_tensor* scale_modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* scale) {
+        scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);
+        return ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+    }
+
+    __STATIC_INLINE__ ggml_tensor* gate_residual(ggml_context* ctx, ggml_tensor* residual, ggml_tensor* x, ggml_tensor* gate) {
+        gate = ggml_tanh(ctx, gate);
+        gate = ggml_reshape_3d(ctx, gate, gate->ne[0], 1, gate->ne[1]);
+        x    = ggml_mul(ctx, x, gate);
+        return ggml_add(ctx, residual, x);
+    }
+
+    struct LuminaCombinedTimestepCaptionEmbedding : public GGMLBlock {
+        int64_t frequency_embedding_size;
+        float timestep_scale;
+
+        LuminaCombinedTimestepCaptionEmbedding(int64_t hidden_size,
+                                               int64_t instruction_feat_dim,
+                                               int64_t frequency_embedding_size,
+                                               float norm_eps,
+                                               float timestep_scale)
+            : frequency_embedding_size(frequency_embedding_size),
+              timestep_scale(timestep_scale) {
+            blocks["timestep_embedder"]  = std::make_shared<Qwen::TimestepEmbedding>(frequency_embedding_size, std::min<int64_t>(hidden_size, 1024));
+            blocks["caption_embedder.0"] = std::make_shared<RMSNorm>(instruction_feat_dim, norm_eps);
+            blocks["caption_embedder.1"] = std::make_shared<Linear>(instruction_feat_dim, hidden_size, true);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* timestep, ggml_tensor* text_hidden_states) {
+            auto timestep_embedder  = std::dynamic_pointer_cast<Qwen::TimestepEmbedding>(blocks["timestep_embedder"]);
+            auto caption_embedder_0 = std::dynamic_pointer_cast<RMSNorm>(blocks["caption_embedder.0"]);
+            auto caption_embedder_1 = std::dynamic_pointer_cast<Linear>(blocks["caption_embedder.1"]);
+
+            auto timestep_proj = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep, static_cast<int>(frequency_embedding_size), 10000, timestep_scale);
+            auto time_embed    = timestep_embedder->forward(ctx, timestep_proj);
+            auto caption_embed = caption_embedder_1->forward(ctx, caption_embedder_0->forward(ctx, text_hidden_states));
+            return {time_embed, caption_embed};
+        }
+    };
+
+    struct LuminaRMSNormZero : public GGMLBlock {
+        LuminaRMSNormZero(int64_t embedding_dim, int64_t conditioning_embedding_dim, float norm_eps) {
+            blocks["linear"] = std::make_shared<Linear>(conditioning_embedding_dim, 4 * embedding_dim, true);
+            blocks["norm"]   = std::make_shared<RMSNorm>(embedding_dim, norm_eps);
+        }
+
+        std::tuple<ggml_tensor*, ggml_tensor*, ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* emb) {
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            auto norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+
+            emb       = linear->forward(ctx, ggml_silu(ctx->ggml_ctx, emb));
+            auto mods = ggml_ext_chunk(ctx->ggml_ctx, emb, 4, 0);
+
+            auto scale_msa = mods[0];
+            auto gate_msa  = mods[1];
+            auto scale_mlp = mods[2];
+            auto gate_mlp  = mods[3];
+
+            x = scale_modulate(ctx->ggml_ctx, norm->forward(ctx, x), scale_msa);
+            return {x, gate_msa, scale_mlp, gate_mlp};
+        }
+    };
+
+    struct LuminaFeedForward : public GGMLBlock {
+        LuminaFeedForward(int64_t dim, int64_t inner_dim, int64_t multiple_of) {
+            inner_dim          = multiple_of * ((inner_dim + multiple_of - 1) / multiple_of);
+            blocks["linear_1"] = std::make_shared<Linear>(dim, inner_dim, false);
+            blocks["linear_2"] = std::make_shared<Linear>(inner_dim, dim, false);
+            blocks["linear_3"] = std::make_shared<Linear>(dim, inner_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["linear_1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
+            auto linear_3 = std::dynamic_pointer_cast<Linear>(blocks["linear_3"]);
+
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                linear_2->set_force_prec_f32(true);
+            }
+
+            auto h1 = linear_1->forward(ctx, x);
+            auto h2 = linear_3->forward(ctx, x);
+            x       = ggml_swiglu_split(ctx->ggml_ctx, h1, h2);
+            x       = linear_2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct LuminaLayerNormContinuous : public GGMLBlock {
+        LuminaLayerNormContinuous(int64_t embedding_dim,
+                                  int64_t conditioning_embedding_dim,
+                                  int64_t out_dim) {
+            blocks["linear_1"] = std::make_shared<Linear>(conditioning_embedding_dim, embedding_dim, true);
+            blocks["norm"]     = std::make_shared<LayerNorm>(embedding_dim, 1e-6f, false);
+            blocks["linear_2"] = std::make_shared<Linear>(embedding_dim, out_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* conditioning_embedding) {
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["linear_1"]);
+            auto norm     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
+
+            auto emb = linear_1->forward(ctx, ggml_silu(ctx->ggml_ctx, conditioning_embedding));
+            x        = scale_modulate(ctx->ggml_ctx, norm->forward(ctx, x), emb);
+            x        = linear_2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct Attention : public GGMLBlock {
+        int64_t dim_head;
+        int64_t heads;
+        int64_t kv_heads;
+
+        Attention(int64_t query_dim, int64_t dim_head, int64_t heads, int64_t kv_heads, float eps = 1e-5f)
+            : dim_head(dim_head), heads(heads), kv_heads(kv_heads) {
+            blocks["to_q"]     = std::make_shared<Linear>(query_dim, heads * dim_head, false);
+            blocks["to_k"]     = std::make_shared<Linear>(query_dim, kv_heads * dim_head, false);
+            blocks["to_v"]     = std::make_shared<Linear>(query_dim, kv_heads * dim_head, false);
+            blocks["norm_q"]   = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_k"]   = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["to_out.0"] = std::make_shared<Linear>(heads * dim_head, query_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* hidden_states,
+                             ggml_tensor* encoder_hidden_states,
+                             ggml_tensor* rotary_emb,
+                             ggml_tensor* attention_mask = nullptr) {
+            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 norm_q   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                to_out_0->set_force_prec_f32(true);
+            }
+
+            int64_t N  = hidden_states->ne[2];
+            int64_t Lq = hidden_states->ne[1];
+            int64_t Lk = encoder_hidden_states->ne[1];
+
+            auto q = to_q->forward(ctx, hidden_states);
+            q      = ggml_reshape_4d(ctx->ggml_ctx, q, dim_head, heads, Lq, N);
+            auto k = to_k->forward(ctx, encoder_hidden_states);
+            k      = ggml_reshape_4d(ctx->ggml_ctx, k, dim_head, kv_heads, Lk, N);
+            auto v = to_v->forward(ctx, encoder_hidden_states);
+            v      = ggml_reshape_4d(ctx->ggml_ctx, v, dim_head, kv_heads, Lk, N);
+
+            q = norm_q->forward(ctx, q);
+            k = norm_k->forward(ctx, k);
+
+            auto out = Rope::attention(ctx, q, k, v, rotary_emb, attention_mask);
+            out      = to_out_0->forward(ctx, out);
+            return out;
+        }
+    };
+
+    struct BooguImageTransformerBlock : public GGMLBlock {
+        bool modulation;
+
+        BooguImageTransformerBlock(int64_t dim,
+                                   int64_t num_attention_heads,
+                                   int64_t num_kv_heads,
+                                   int64_t multiple_of,
+                                   float norm_eps,
+                                   bool modulation)
+            : modulation(modulation) {
+            int64_t head_dim       = dim / num_attention_heads;
+            blocks["attn"]         = std::make_shared<Attention>(dim, head_dim, num_attention_heads, num_kv_heads, 1e-5f);
+            blocks["feed_forward"] = std::make_shared<LuminaFeedForward>(dim, 4 * dim, multiple_of);
+            if (modulation) {
+                blocks["norm1"] = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            } else {
+                blocks["norm1"] = std::make_shared<RMSNorm>(dim, norm_eps);
+            }
+            blocks["ffn_norm1"] = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["norm2"]     = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["ffn_norm2"] = std::make_shared<RMSNorm>(dim, norm_eps);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* hidden_states,
+                             ggml_tensor* rotary_emb,
+                             ggml_tensor* temb           = nullptr,
+                             ggml_tensor* attention_mask = nullptr) {
+            auto attn         = std::dynamic_pointer_cast<Attention>(blocks["attn"]);
+            auto feed_forward = std::dynamic_pointer_cast<LuminaFeedForward>(blocks["feed_forward"]);
+            auto ffn_norm1    = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm1"]);
+            auto norm2        = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"]);
+            auto ffn_norm2    = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm2"]);
+
+            if (modulation) {
+                auto norm1 = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["norm1"]);
+                auto mods  = norm1->forward(ctx, hidden_states, temb);
+
+                auto norm_hidden_states = std::get<0>(mods);
+                auto gate_msa           = std::get<1>(mods);
+                auto scale_mlp          = std::get<2>(mods);
+                auto gate_mlp           = std::get<3>(mods);
+
+                auto attn_output = attn->forward(ctx, norm_hidden_states, norm_hidden_states, rotary_emb, attention_mask);
+                hidden_states    = gate_residual(ctx->ggml_ctx, hidden_states, norm2->forward(ctx, attn_output), gate_msa);
+
+                auto mlp_input  = scale_modulate(ctx->ggml_ctx, ffn_norm1->forward(ctx, hidden_states), scale_mlp);
+                auto mlp_output = feed_forward->forward(ctx, mlp_input);
+                hidden_states   = gate_residual(ctx->ggml_ctx, hidden_states, ffn_norm2->forward(ctx, mlp_output), gate_mlp);
+            } else {
+                auto norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"]);
+
+                auto norm_hidden_states = norm1->forward(ctx, hidden_states);
+                auto attn_output        = attn->forward(ctx, norm_hidden_states, norm_hidden_states, rotary_emb, attention_mask);
+                hidden_states           = ggml_add(ctx->ggml_ctx, hidden_states, norm2->forward(ctx, attn_output));
+
+                auto mlp_output = feed_forward->forward(ctx, ffn_norm1->forward(ctx, hidden_states));
+                hidden_states   = ggml_add(ctx->ggml_ctx, hidden_states, ffn_norm2->forward(ctx, mlp_output));
+            }
+            return hidden_states;
+        }
+    };
+
+    struct BooguImageJointAttention : public GGMLBlock {
+        int64_t dim_head;
+        int64_t heads;
+        int64_t kv_heads;
+
+        BooguImageJointAttention(int64_t dim, int64_t dim_head, int64_t heads, int64_t kv_heads)
+            : dim_head(dim_head), heads(heads), kv_heads(kv_heads) {
+            blocks["norm_q"]                  = std::make_shared<RMSNorm>(dim_head, 1e-5f);
+            blocks["norm_k"]                  = std::make_shared<RMSNorm>(dim_head, 1e-5f);
+            blocks["to_out.0"]                = std::make_shared<Linear>(heads * dim_head, dim, false);
+            blocks["processor.img_to_q"]      = std::make_shared<Linear>(dim, heads * dim_head, false);
+            blocks["processor.img_to_k"]      = std::make_shared<Linear>(dim, kv_heads * dim_head, false);
+            blocks["processor.img_to_v"]      = std::make_shared<Linear>(dim, kv_heads * dim_head, false);
+            blocks["processor.instruct_to_q"] = std::make_shared<Linear>(dim, heads * dim_head, false);
+            blocks["processor.instruct_to_k"] = std::make_shared<Linear>(dim, kv_heads * dim_head, false);
+            blocks["processor.instruct_to_v"] = std::make_shared<Linear>(dim, kv_heads * dim_head, false);
+            blocks["processor.instruct_out"]  = std::make_shared<Linear>(heads * dim_head, dim, false);
+            blocks["processor.img_out"]       = std::make_shared<Linear>(heads * dim_head, dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* img_hidden_states,
+                             ggml_tensor* instruct_hidden_states,
+                             ggml_tensor* rotary_emb,
+                             ggml_tensor* attention_mask = nullptr) {
+            auto norm_q        = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k        = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto to_out_0      = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+            auto img_to_q      = std::dynamic_pointer_cast<Linear>(blocks["processor.img_to_q"]);
+            auto img_to_k      = std::dynamic_pointer_cast<Linear>(blocks["processor.img_to_k"]);
+            auto img_to_v      = std::dynamic_pointer_cast<Linear>(blocks["processor.img_to_v"]);
+            auto instruct_to_q = std::dynamic_pointer_cast<Linear>(blocks["processor.instruct_to_q"]);
+            auto instruct_to_k = std::dynamic_pointer_cast<Linear>(blocks["processor.instruct_to_k"]);
+            auto instruct_to_v = std::dynamic_pointer_cast<Linear>(blocks["processor.instruct_to_v"]);
+            auto instruct_out  = std::dynamic_pointer_cast<Linear>(blocks["processor.instruct_out"]);
+            auto img_out       = std::dynamic_pointer_cast<Linear>(blocks["processor.img_out"]);
+
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                to_out_0->set_force_prec_f32(true);
+            }
+
+            int64_t N          = img_hidden_states->ne[2];
+            int64_t L_img      = img_hidden_states->ne[1];
+            int64_t L_instruct = instruct_hidden_states->ne[1];
+
+            auto img_q = img_to_q->forward(ctx, img_hidden_states);
+            img_q      = ggml_reshape_4d(ctx->ggml_ctx, img_q, dim_head, heads, L_img, N);
+            auto img_k = img_to_k->forward(ctx, img_hidden_states);
+            img_k      = ggml_reshape_4d(ctx->ggml_ctx, img_k, dim_head, kv_heads, L_img, N);
+            auto img_v = img_to_v->forward(ctx, img_hidden_states);
+            img_v      = ggml_reshape_4d(ctx->ggml_ctx, img_v, dim_head, kv_heads, L_img, N);
+
+            auto instruct_q = instruct_to_q->forward(ctx, instruct_hidden_states);
+            instruct_q      = ggml_reshape_4d(ctx->ggml_ctx, instruct_q, dim_head, heads, L_instruct, N);
+            auto instruct_k = instruct_to_k->forward(ctx, instruct_hidden_states);
+            instruct_k      = ggml_reshape_4d(ctx->ggml_ctx, instruct_k, dim_head, kv_heads, L_instruct, N);
+            auto instruct_v = instruct_to_v->forward(ctx, instruct_hidden_states);
+            instruct_v      = ggml_reshape_4d(ctx->ggml_ctx, instruct_v, dim_head, kv_heads, L_instruct, N);
+
+            auto q = ggml_concat(ctx->ggml_ctx, instruct_q, img_q, 2);
+            auto k = ggml_concat(ctx->ggml_ctx, instruct_k, img_k, 2);
+            auto v = ggml_concat(ctx->ggml_ctx, instruct_v, img_v, 2);
+            q      = norm_q->forward(ctx, q);
+            k      = norm_k->forward(ctx, k);
+
+            auto hidden_states = Rope::attention(ctx, q, k, v, rotary_emb, attention_mask);
+            auto instruct_attn = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, 0, L_instruct);
+            auto img_attn      = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, L_instruct, L_instruct + L_img);
+
+            instruct_attn = instruct_out->forward(ctx, instruct_attn);
+            img_attn      = img_out->forward(ctx, img_attn);
+            hidden_states = ggml_concat(ctx->ggml_ctx, instruct_attn, img_attn, 1);
+            hidden_states = to_out_0->forward(ctx, hidden_states);
+            return hidden_states;
+        }
+    };
+
+    struct BooguImageDoubleStreamBlock : public GGMLBlock {
+        BooguImageDoubleStreamBlock(int64_t dim,
+                                    int64_t num_attention_heads,
+                                    int64_t num_kv_heads,
+                                    int64_t multiple_of,
+                                    float norm_eps) {
+            int64_t head_dim                = dim / num_attention_heads;
+            blocks["img_instruct_attn"]     = std::make_shared<BooguImageJointAttention>(dim, head_dim, num_attention_heads, num_kv_heads);
+            blocks["img_self_attn"]         = std::make_shared<Attention>(dim, head_dim, num_attention_heads, num_kv_heads, 1e-5f);
+            blocks["img_feed_forward"]      = std::make_shared<LuminaFeedForward>(dim, 4 * dim, multiple_of);
+            blocks["instruct_feed_forward"] = std::make_shared<LuminaFeedForward>(dim, 4 * dim, multiple_of);
+            blocks["img_norm1"]             = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["img_norm2"]             = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["img_norm3"]             = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["instruct_norm1"]        = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["instruct_norm2"]        = std::make_shared<LuminaRMSNormZero>(dim, std::min<int64_t>(dim, 1024), norm_eps);
+            blocks["img_attn_norm"]         = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["img_self_attn_norm"]    = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["img_ffn_norm1"]         = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["img_ffn_norm2"]         = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["instruct_attn_norm"]    = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["instruct_ffn_norm1"]    = std::make_shared<RMSNorm>(dim, norm_eps);
+            blocks["instruct_ffn_norm2"]    = std::make_shared<RMSNorm>(dim, norm_eps);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* img_hidden_states,
+                                                      ggml_tensor* instruct_hidden_states,
+                                                      ggml_tensor* joint_rotary_emb,
+                                                      ggml_tensor* img_rotary_emb,
+                                                      ggml_tensor* temb) {
+            auto img_instruct_attn     = std::dynamic_pointer_cast<BooguImageJointAttention>(blocks["img_instruct_attn"]);
+            auto img_self_attn         = std::dynamic_pointer_cast<Attention>(blocks["img_self_attn"]);
+            auto img_feed_forward      = std::dynamic_pointer_cast<LuminaFeedForward>(blocks["img_feed_forward"]);
+            auto instruct_feed_forward = std::dynamic_pointer_cast<LuminaFeedForward>(blocks["instruct_feed_forward"]);
+            auto img_norm1             = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["img_norm1"]);
+            auto img_norm2             = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["img_norm2"]);
+            auto img_norm3             = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["img_norm3"]);
+            auto instruct_norm1        = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["instruct_norm1"]);
+            auto instruct_norm2        = std::dynamic_pointer_cast<LuminaRMSNormZero>(blocks["instruct_norm2"]);
+            auto img_attn_norm         = std::dynamic_pointer_cast<RMSNorm>(blocks["img_attn_norm"]);
+            auto img_self_attn_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["img_self_attn_norm"]);
+            auto img_ffn_norm1         = std::dynamic_pointer_cast<RMSNorm>(blocks["img_ffn_norm1"]);
+            auto img_ffn_norm2         = std::dynamic_pointer_cast<RMSNorm>(blocks["img_ffn_norm2"]);
+            auto instruct_attn_norm    = std::dynamic_pointer_cast<RMSNorm>(blocks["instruct_attn_norm"]);
+            auto instruct_ffn_norm1    = std::dynamic_pointer_cast<RMSNorm>(blocks["instruct_ffn_norm1"]);
+            auto instruct_ffn_norm2    = std::dynamic_pointer_cast<RMSNorm>(blocks["instruct_ffn_norm2"]);
+
+            int64_t L_instruct = instruct_hidden_states->ne[1];
+
+            auto img_norm1_out_vec      = img_norm1->forward(ctx, img_hidden_states, temb);
+            auto img_norm2_out_vec      = img_norm2->forward(ctx, img_hidden_states, temb);
+            auto img_norm3_out_vec      = img_norm3->forward(ctx, img_hidden_states, temb);
+            auto instruct_norm1_out_vec = instruct_norm1->forward(ctx, instruct_hidden_states, temb);
+            auto instruct_norm2_out_vec = instruct_norm2->forward(ctx, instruct_hidden_states, temb);
+
+            auto img_norm1_out = std::get<0>(img_norm1_out_vec);
+            auto img_gate_msa  = std::get<1>(img_norm1_out_vec);
+            auto img_scale_mlp = std::get<2>(img_norm1_out_vec);
+            auto img_gate_mlp  = std::get<3>(img_norm1_out_vec);
+
+            auto img_norm2_out = std::get<0>(img_norm2_out_vec);
+            auto img_shift_mlp = std::get<1>(img_norm2_out_vec);
+
+            auto img_norm3_out = std::get<0>(img_norm3_out_vec);
+            auto img_gate_self = std::get<1>(img_norm3_out_vec);
+
+            auto instruct_norm1_out = std::get<0>(instruct_norm1_out_vec);
+            auto instruct_gate_msa  = std::get<1>(instruct_norm1_out_vec);
+            auto instruct_scale_mlp = std::get<2>(instruct_norm1_out_vec);
+            auto instruct_gate_mlp  = std::get<3>(instruct_norm1_out_vec);
+
+            auto instruct_norm2_out = std::get<0>(instruct_norm2_out_vec);
+            auto instruct_shift_mlp = std::get<1>(instruct_norm2_out_vec);
+
+            auto joint_attn_out    = img_instruct_attn->forward(ctx, img_norm1_out, instruct_norm1_out, joint_rotary_emb);
+            auto instruct_attn_out = ggml_ext_slice(ctx->ggml_ctx, joint_attn_out, 1, 0, L_instruct);
+            auto img_attn_out      = ggml_ext_slice(ctx->ggml_ctx, joint_attn_out, 1, L_instruct, joint_attn_out->ne[1]);
+
+            auto img_self_attn_out = img_self_attn->forward(ctx, img_norm3_out, img_norm3_out, img_rotary_emb);
+
+            img_hidden_states = gate_residual(ctx->ggml_ctx, img_hidden_states, img_attn_norm->forward(ctx, img_attn_out), img_gate_msa);
+            img_hidden_states = gate_residual(ctx->ggml_ctx, img_hidden_states, img_self_attn_norm->forward(ctx, img_self_attn_out), img_gate_self);
+
+            auto img_mlp_input = scale_modulate(ctx->ggml_ctx, img_norm2_out, img_scale_mlp);
+            img_shift_mlp      = ggml_reshape_3d(ctx->ggml_ctx, img_shift_mlp, img_shift_mlp->ne[0], 1, img_shift_mlp->ne[1]);
+            img_mlp_input      = ggml_add(ctx->ggml_ctx, img_mlp_input, img_shift_mlp);
+            auto img_mlp_out   = img_feed_forward->forward(ctx, img_ffn_norm1->forward(ctx, img_mlp_input));
+            img_hidden_states  = gate_residual(ctx->ggml_ctx, img_hidden_states, img_ffn_norm2->forward(ctx, img_mlp_out), img_gate_mlp);
+
+            instruct_hidden_states  = gate_residual(ctx->ggml_ctx, instruct_hidden_states, instruct_attn_norm->forward(ctx, instruct_attn_out), instruct_gate_msa);
+            auto instruct_mlp_input = scale_modulate(ctx->ggml_ctx, instruct_norm2_out, instruct_scale_mlp);
+            instruct_shift_mlp      = ggml_reshape_3d(ctx->ggml_ctx, instruct_shift_mlp, instruct_shift_mlp->ne[0], 1, instruct_shift_mlp->ne[1]);
+            instruct_mlp_input      = ggml_add(ctx->ggml_ctx, instruct_mlp_input, instruct_shift_mlp);
+            auto instruct_mlp_out   = instruct_feed_forward->forward(ctx, instruct_ffn_norm1->forward(ctx, instruct_mlp_input));
+            instruct_hidden_states  = gate_residual(ctx->ggml_ctx, instruct_hidden_states, instruct_ffn_norm2->forward(ctx, instruct_mlp_out), instruct_gate_mlp);
+
+            return {img_hidden_states, instruct_hidden_states};
+        }
+    };
+
+    struct BooguImageModel : public GGMLBlock {
+        BooguConfig config;
+
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+            GGML_UNUSED(tensor_storage_map);
+            GGML_UNUSED(prefix);
+            params["image_index_embedding"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, config.hidden_size, 5);
+        }
+
+        BooguImageModel() = default;
+        BooguImageModel(BooguConfig config)
+            : config(std::move(config)) {
+            blocks["x_embedder"]               = std::make_shared<Linear>(this->config.patch_size * this->config.patch_size * this->config.in_channels, this->config.hidden_size, true);
+            blocks["ref_image_patch_embedder"] = std::make_shared<Linear>(this->config.patch_size * this->config.patch_size * this->config.in_channels, this->config.hidden_size, true);
+            blocks["time_caption_embed"]       = std::make_shared<LuminaCombinedTimestepCaptionEmbedding>(this->config.hidden_size,
+                                                                                                    this->config.instruction_feat_dim,
+                                                                                                    256,
+                                                                                                    this->config.norm_eps,
+                                                                                                    this->config.timestep_scale);
+
+            for (int i = 0; i < this->config.num_refiner_layers; i++) {
+                blocks["noise_refiner." + std::to_string(i)]     = std::make_shared<BooguImageTransformerBlock>(this->config.hidden_size,
+                                                                                                            this->config.num_attention_heads,
+                                                                                                            this->config.num_kv_heads,
+                                                                                                            this->config.multiple_of,
+                                                                                                            this->config.norm_eps,
+                                                                                                            true);
+                blocks["ref_image_refiner." + std::to_string(i)] = std::make_shared<BooguImageTransformerBlock>(this->config.hidden_size,
+                                                                                                                this->config.num_attention_heads,
+                                                                                                                this->config.num_kv_heads,
+                                                                                                                this->config.multiple_of,
+                                                                                                                this->config.norm_eps,
+                                                                                                                true);
+                blocks["context_refiner." + std::to_string(i)]   = std::make_shared<BooguImageTransformerBlock>(this->config.hidden_size,
+                                                                                                              this->config.num_attention_heads,
+                                                                                                              this->config.num_kv_heads,
+                                                                                                              this->config.multiple_of,
+                                                                                                              this->config.norm_eps,
+                                                                                                              false);
+            }
+
+            for (int i = 0; i < this->config.num_double_stream_layers; i++) {
+                blocks["double_stream_layers." + std::to_string(i)] = std::make_shared<BooguImageDoubleStreamBlock>(this->config.hidden_size,
+                                                                                                                    this->config.num_attention_heads,
+                                                                                                                    this->config.num_kv_heads,
+                                                                                                                    this->config.multiple_of,
+                                                                                                                    this->config.norm_eps);
+            }
+
+            for (int i = 0; i < this->config.num_layers; i++) {
+                blocks["single_stream_layers." + std::to_string(i)] = std::make_shared<BooguImageTransformerBlock>(this->config.hidden_size,
+                                                                                                                   this->config.num_attention_heads,
+                                                                                                                   this->config.num_kv_heads,
+                                                                                                                   this->config.multiple_of,
+                                                                                                                   this->config.norm_eps,
+                                                                                                                   true);
+            }
+
+            blocks["norm_out"] = std::make_shared<LuminaLayerNormContinuous>(this->config.hidden_size,
+                                                                             this->config.timestep_embed_dim,
+                                                                             this->config.patch_size * this->config.patch_size * this->config.out_channels);
+        }
+
+        ggml_tensor* image_index_embedding(GGMLRunnerContext* ctx, int index) {
+            GGML_ASSERT(index >= 0 && index < 5);
+            auto embedding = params["image_index_embedding"];
+            auto out       = ggml_view_1d(ctx->ggml_ctx,
+                                          embedding,
+                                          config.hidden_size,
+                                          index * config.hidden_size * ggml_element_size(embedding));
+            out            = ggml_reshape_3d(ctx->ggml_ctx, out, config.hidden_size, 1, 1);
+            return out;
+        }
+
+        ggml_tensor* embed_refs(GGMLRunnerContext* ctx, const std::vector<ggml_tensor*>& ref_latents) {
+            if (ref_latents.empty()) {
+                return nullptr;
+            }
+            auto ref_image_patch_embedder = std::dynamic_pointer_cast<Linear>(blocks["ref_image_patch_embedder"]);
+
+            ggml_tensor* ref_img = nullptr;
+            for (int i = 0; i < static_cast<int>(ref_latents.size()); i++) {
+                auto ref = DiT::pad_and_patchify(ctx, ref_latents[i], config.patch_size, config.patch_size, false);
+                ref      = ref_image_patch_embedder->forward(ctx, ref);
+                ref      = ggml_add(ctx->ggml_ctx, ref, image_index_embedding(ctx, std::min(i, 4)));
+                ref_img  = ref_img == nullptr ? ref : ggml_concat(ctx->ggml_ctx, ref_img, ref, 1);
+            }
+            return ref_img;
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timesteps,
+                             ggml_tensor* context,
+                             ggml_tensor* pe,
+                             std::vector<ggml_tensor*> ref_latents = {}) {
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t N = x->ne[3];
+            GGML_ASSERT(N == 1);
+
+            auto x_embedder         = std::dynamic_pointer_cast<Linear>(blocks["x_embedder"]);
+            auto time_caption_embed = std::dynamic_pointer_cast<LuminaCombinedTimestepCaptionEmbedding>(blocks["time_caption_embed"]);
+            auto norm_out           = std::dynamic_pointer_cast<LuminaLayerNormContinuous>(blocks["norm_out"]);
+
+            auto timestep = ggml_sub(ctx->ggml_ctx, ggml_ext_ones_like(ctx->ggml_ctx, timesteps), timesteps);
+            auto embeds   = time_caption_embed->forward(ctx, timestep, context);
+            auto temb     = embeds.first;
+            auto txt      = embeds.second;
+
+            auto img        = DiT::pad_and_patchify(ctx, x, config.patch_size, config.patch_size, false);
+            int64_t img_len = img->ne[1];
+            img             = x_embedder->forward(ctx, img);
+            auto ref_img    = embed_refs(ctx, ref_latents);
+            int64_t ref_len = ref_img != nullptr ? ref_img->ne[1] : 0;
+            int64_t txt_len = txt->ne[1];
+
+            GGML_ASSERT(pe->ne[3] == txt_len + ref_len + img_len);
+            auto txt_pe   = ggml_ext_slice(ctx->ggml_ctx, pe, 3, 0, txt_len);
+            auto noise_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt_len + ref_len, txt_len + ref_len + img_len);
+
+            for (int i = 0; i < config.num_refiner_layers; i++) {
+                auto block = std::dynamic_pointer_cast<BooguImageTransformerBlock>(blocks["context_refiner." + std::to_string(i)]);
+                txt        = block->forward(ctx, txt, txt_pe);
+                sd::ggml_graph_cut::mark_graph_cut(txt, "boogu.context_refiner." + std::to_string(i), "txt");
+            }
+
+            for (int i = 0; i < config.num_refiner_layers; i++) {
+                auto block = std::dynamic_pointer_cast<BooguImageTransformerBlock>(blocks["noise_refiner." + std::to_string(i)]);
+                img        = block->forward(ctx, img, noise_pe, temb);
+                sd::ggml_graph_cut::mark_graph_cut(img, "boogu.noise_refiner." + std::to_string(i), "img");
+            }
+
+            ggml_tensor* combined_img = img;
+            if (ref_img != nullptr) {
+                auto ref_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt_len, txt_len + ref_len);
+                for (int i = 0; i < config.num_refiner_layers; i++) {
+                    auto block = std::dynamic_pointer_cast<BooguImageTransformerBlock>(blocks["ref_image_refiner." + std::to_string(i)]);
+                    ref_img    = block->forward(ctx, ref_img, ref_pe, temb);
+                    sd::ggml_graph_cut::mark_graph_cut(ref_img, "boogu.ref_image_refiner." + std::to_string(i), "ref_img");
+                }
+                combined_img = ggml_concat(ctx->ggml_ctx, ref_img, img, 1);
+            }
+
+            auto img_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt_len, txt_len + combined_img->ne[1]);
+            for (int i = 0; i < config.num_double_stream_layers; i++) {
+                auto block   = std::dynamic_pointer_cast<BooguImageDoubleStreamBlock>(blocks["double_stream_layers." + std::to_string(i)]);
+                auto result  = block->forward(ctx, combined_img, txt, pe, img_pe, temb);
+                combined_img = result.first;
+                txt          = result.second;
+                sd::ggml_graph_cut::mark_graph_cut(combined_img, "boogu.double_stream_layers." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "boogu.double_stream_layers." + std::to_string(i), "txt");
+            }
+
+            auto hidden_states = ggml_concat(ctx->ggml_ctx, txt, combined_img, 1);
+            for (int i = 0; i < config.num_layers; i++) {
+                auto block    = std::dynamic_pointer_cast<BooguImageTransformerBlock>(blocks["single_stream_layers." + std::to_string(i)]);
+                hidden_states = block->forward(ctx, hidden_states, pe, temb);
+                sd::ggml_graph_cut::mark_graph_cut(hidden_states, "boogu.single_stream_layers." + std::to_string(i), "hidden_states");
+            }
+
+            hidden_states = norm_out->forward(ctx, hidden_states, temb);
+            hidden_states = ggml_ext_slice(ctx->ggml_ctx, hidden_states, 1, hidden_states->ne[1] - img_len, hidden_states->ne[1]);
+            hidden_states = DiT::unpatchify_and_crop(ctx->ggml_ctx, hidden_states, H, W, config.patch_size, config.patch_size, false);
+            hidden_states = ggml_ext_scale(ctx->ggml_ctx, hidden_states, -1.f);
+            return hidden_states;
+        }
+    };
+
+    __STATIC_INLINE__ int patched_token_count(int64_t size, int patch_size) {
+        int pad = (patch_size - (static_cast<int>(size) % patch_size)) % patch_size;
+        return (static_cast<int>(size) + pad) / patch_size;
+    }
+
+    __STATIC_INLINE__ void append_spatial_ids(std::vector<std::vector<float>>& ids,
+                                              int bs,
+                                              int pe_shift,
+                                              int h_tokens,
+                                              int w_tokens) {
+        std::vector<std::vector<float>> image_ids(h_tokens * w_tokens, std::vector<float>(3, 0.0f));
+        for (int h = 0; h < h_tokens; h++) {
+            for (int w = 0; w < w_tokens; w++) {
+                image_ids[h * w_tokens + w][0] = static_cast<float>(pe_shift);
+                image_ids[h * w_tokens + w][1] = static_cast<float>(h);
+                image_ids[h * w_tokens + w][2] = static_cast<float>(w);
+            }
+        }
+        for (int b = 0; b < bs; b++) {
+            ids.insert(ids.end(), image_ids.begin(), image_ids.end());
+        }
+    }
+
+    __STATIC_INLINE__ std::vector<float> gen_boogu_pe(int h,
+                                                      int w,
+                                                      int patch_size,
+                                                      int bs,
+                                                      int context_len,
+                                                      const std::vector<ggml_tensor*>& ref_latents,
+                                                      int theta,
+                                                      const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids;
+        ids.reserve(static_cast<size_t>(bs) * context_len);
+        for (int b = 0; b < bs; b++) {
+            for (int i = 0; i < context_len; i++) {
+                float pos = static_cast<float>(i);
+                ids.push_back({pos, pos, pos});
+            }
+        }
+
+        int pe_shift = context_len;
+        for (ggml_tensor* ref : ref_latents) {
+            int ref_h_tokens = patched_token_count(ref->ne[1], patch_size);
+            int ref_w_tokens = patched_token_count(ref->ne[0], patch_size);
+            append_spatial_ids(ids, bs, pe_shift, ref_h_tokens, ref_w_tokens);
+            pe_shift += std::max(ref_h_tokens, ref_w_tokens);
+        }
+
+        int h_tokens = patched_token_count(h, patch_size);
+        int w_tokens = patched_token_count(w, patch_size);
+        append_spatial_ids(ids, bs, pe_shift, h_tokens, w_tokens);
+
+        return Rope::embed_nd(ids, bs, static_cast<float>(theta), axes_dim);
+    }
+
+    struct BooguImageRunner : public DiffusionModelRunner {
+        BooguConfig config;
+        BooguImageModel boogu;
+        std::vector<float> pe_vec;
+
+        BooguImageRunner(ggml_backend_t backend,
+                         const String2TensorStorage& tensor_storage_map      = {},
+                         const std::string prefix                            = "",
+                         SDVersion version                                   = VERSION_BOOGU_IMAGE,
+                         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(BooguConfig::detect_from_weights(tensor_storage_map, prefix)) {
+            boogu = BooguImageModel(config);
+            boogu.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "boogu_image";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            boogu.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 const std::vector<sd::Tensor<float>>& ref_latents_tensor = {}) {
+            ggml_cgraph* gf        = new_graph_custom(BOOGU_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(!context_tensor.empty());
+            ggml_tensor* context = make_input(context_tensor);
+
+            std::vector<ggml_tensor*> ref_latents;
+            ref_latents.reserve(ref_latents_tensor.size());
+            for (const auto& ref_latent_tensor : ref_latents_tensor) {
+                ref_latents.push_back(make_input(ref_latent_tensor));
+            }
+
+            pe_vec      = gen_boogu_pe(static_cast<int>(x->ne[1]),
+                                       static_cast<int>(x->ne[0]),
+                                       config.patch_size,
+                                       static_cast<int>(x->ne[3]),
+                                       static_cast<int>(context->ne[1]),
+                                       ref_latents,
+                                       config.theta,
+                                       config.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
+            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = boogu.forward(&runner_ctx, x, timesteps, context, pe, ref_latents);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  const std::vector<sd::Tensor<float>>& ref_latents = {}) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, ref_latents);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            static const std::vector<sd::Tensor<float>> empty_ref_latents;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents);
+        }
+    };
+}  // namespace Boogu
+
+#endif  // __SD_MODEL_DIFFUSION_BOOGU_HPP__

src/model/diffusion/ernie_image.hpp

@@ -162,6 +162,8 @@ namespace ErnieImage {
             int64_t S = x->ne[1];
             int64_t N = x->ne[2];
 
+            float scale = (sd_backend_is(ctx->backend, "Vulkan") && ctx->flash_attn_enabled) ? 1.0f / 32.0f : 1.0f;
+
             auto q = to_q->forward(ctx, x);
             auto k = to_k->forward(ctx, x);
             auto v = to_v->forward(ctx, x);
@@ -182,7 +184,7 @@ namespace ErnieImage {
             k = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, k, 0, 2, 1, 3));  // [N, heads, S, head_dim]
             k = ggml_reshape_3d(ctx->ggml_ctx, k, k->ne[0], k->ne[1], k->ne[2] * k->ne[3]);
 
-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, attention_mask, true, ctx->flash_attn_enabled);  // [N, S, hidden_size]
+            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, num_heads, attention_mask, true, ctx->flash_attn_enabled, scale);  // [N, S, hidden_size]
             x = to_out_0->forward(ctx, x);
             return x;
         }

src/model/te/llm.hpp

@@ -79,6 +79,7 @@ namespace LLM {
         int window_size                     = 112;
         int num_position_embeddings         = 0;
         std::set<int> fullatt_block_indexes = {7, 15, 23, 31};
+        bool split_patch_embed              = false;
     };
 
     struct LLMConfig {
@@ -179,7 +180,8 @@ namespace LLM {
                 config.num_experts_per_tok     = 4;
             }
 
-            config.num_layers = 0;
+            config.num_layers          = 0;
+            int detected_vision_layers = 0;
             for (const auto& [name, tensor_storage] : tensor_storage_map) {
                 if (!starts_with(name, prefix)) {
                     continue;
@@ -190,6 +192,38 @@ namespace LLM {
                     if (contains(name, "attn.q_proj")) {
                         config.llama_cpp_style = true;
                     }
+                    if (contains(name, "visual.patch_embed.proj.1.weight")) {
+                        config.vision.split_patch_embed = true;
+                    }
+                    if (contains(name, "visual.patch_embed.proj.0.weight")) {
+                        config.vision.patch_size  = static_cast<int>(tensor_storage.ne[0]);
+                        config.vision.in_channels = tensor_storage.ne[2];
+                        config.vision.hidden_size = tensor_storage.ne[3];
+                    }
+                    if (contains(name, "visual.patch_embed.bias")) {
+                        config.vision.hidden_size = tensor_storage.ne[0];
+                    }
+                    if (contains(name, "visual.pos_embed.weight")) {
+                        config.vision.hidden_size             = tensor_storage.ne[0];
+                        config.vision.num_position_embeddings = static_cast<int>(tensor_storage.ne[1]);
+                    }
+                    if (contains(name, "visual.blocks.")) {
+                        auto items = split_string(name.substr(pos), '.');
+                        if (items.size() > 2) {
+                            int block_index = atoi(items[2].c_str());
+                            if (block_index + 1 > detected_vision_layers) {
+                                detected_vision_layers = block_index + 1;
+                            }
+                        }
+                    }
+                    if (contains(name, "visual.blocks.0.mlp.linear_fc1.weight") ||
+                        contains(name, "visual.blocks.0.mlp.gate_proj.weight")) {
+                        config.vision.intermediate_size = tensor_storage.ne[1];
+                    }
+                    if (contains(name, "visual.merger.linear_fc2.weight") ||
+                        contains(name, "visual.merger.mlp.2.weight")) {
+                        config.vision.out_hidden_size = tensor_storage.ne[1];
+                    }
                     continue;
                 }
                 pos = name.find("layers.");
@@ -219,6 +253,9 @@ namespace LLM {
             if (arch == LLMArch::QWEN3 && config.num_layers == 28) {
                 config.num_heads = 16;
             }
+            if (detected_vision_layers > 0) {
+                config.vision.num_layers = detected_vision_layers;
+            }
             LOG_DEBUG("llm: num_layers = %" PRId64 ", vocab_size = %" PRId64 ", hidden_size = %" PRId64 ", intermediate_size = %" PRId64,
                       config.num_layers,
                       config.vocab_size,
@@ -539,40 +576,51 @@ namespace LLM {
 
     struct VisionPatchEmbed : public GGMLBlock {
     protected:
-        bool llama_cpp_style;
+        bool split_patch_embed;
+        bool bias;
         int patch_size;
         int temporal_patch_size;
         int64_t in_channels;
         int64_t embed_dim;
 
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            GGML_UNUSED(tensor_storage_map);
+            GGML_UNUSED(prefix);
+            if (split_patch_embed && bias) {
+                params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, embed_dim);
+            }
+        }
+
     public:
-        VisionPatchEmbed(bool llama_cpp_style,
+        VisionPatchEmbed(bool split_patch_embed,
                          LLMVisionArch arch,
                          int patch_size          = 14,
                          int temporal_patch_size = 2,
                          int64_t in_channels     = 3,
                          int64_t embed_dim       = 1152)
-            : llama_cpp_style(llama_cpp_style),
+            : split_patch_embed(split_patch_embed),
+              bias(arch == LLMVisionArch::QWEN3_VL),
               patch_size(patch_size),
               temporal_patch_size(temporal_patch_size),
               in_channels(in_channels),
               embed_dim(embed_dim) {
-            bool bias = arch == LLMVisionArch::QWEN3_VL;
+            if (split_patch_embed) {
-            if (llama_cpp_style) {
                 blocks["proj.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels,
                                                                          embed_dim,
                                                                          {patch_size, patch_size},
                                                                          {patch_size, patch_size},
                                                                          {0, 0},
                                                                          {1, 1},
-                                                                         bias));
+                                                                         false));
                 blocks["proj.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels,
                                                                          embed_dim,
                                                                          {patch_size, patch_size},
                                                                          {patch_size, patch_size},
                                                                          {0, 0},
                                                                          {1, 1},
-                                                                         bias));
+                                                                         false));
             } else {
                 std::tuple<int, int, int> kernel_size = {(int)temporal_patch_size, (int)patch_size, (int)patch_size};
                 blocks["proj"]                        = std::shared_ptr<GGMLBlock>(new Conv3d(in_channels,
@@ -593,7 +641,7 @@ namespace LLM {
                                 temporal_patch_size,
                                 ggml_nelements(x) / (temporal_patch_size * patch_size * patch_size));
 
-            if (llama_cpp_style) {
+            if (split_patch_embed) {
                 auto proj_0 = std::dynamic_pointer_cast<Conv2d>(blocks["proj.0"]);
                 auto proj_1 = std::dynamic_pointer_cast<Conv2d>(blocks["proj.1"]);
 
@@ -606,6 +654,10 @@ namespace LLM {
                 x1      = proj_1->forward(ctx, x1);
 
                 x = ggml_add(ctx->ggml_ctx, x0, x1);
+                if (bias) {
+                    auto b = ggml_reshape_4d(ctx->ggml_ctx, params["bias"], 1, 1, embed_dim, 1);
+                    x      = ggml_add_inplace(ctx->ggml_ctx, x, b);
+                }
             } else {
                 auto proj = std::dynamic_pointer_cast<Conv3d>(blocks["proj"]);
 
@@ -798,7 +850,7 @@ namespace LLM {
               spatial_merge_size(vision_params.spatial_merge_size),
               num_grid_per_side(vision_params.num_position_embeddings > 0 ? static_cast<int>(std::sqrt(vision_params.num_position_embeddings)) : 0),
               fullatt_block_indexes(vision_params.fullatt_block_indexes) {
-            blocks["patch_embed"] = std::shared_ptr<GGMLBlock>(new VisionPatchEmbed(llama_cpp_style,
+            blocks["patch_embed"] = std::shared_ptr<GGMLBlock>(new VisionPatchEmbed(vision_params.split_patch_embed,
                                                                                     arch_,
                                                                                     vision_params.patch_size,
                                                                                     vision_params.temporal_patch_size,

src/model/vae/auto_encoder_kl.hpp

@@ -682,7 +682,7 @@ struct AutoEncoderKL : public VAE {
         } else if (sd_version_is_sd3(version)) {
             scale_factor = 1.5305f;
             shift_factor = 0.0609f;
-        } else if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_longcat(version)) {
+        } else if (sd_version_uses_flux_vae(version)) {
             scale_factor = 0.3611f;
             shift_factor = 0.1159f;
         } else if (sd_version_uses_flux2_vae(version)) {

src/model_loader.cpp

@@ -485,6 +485,9 @@ SDVersion ModelLoader::get_sd_version() {
         if (tensor_storage.name.find("model.diffusion_model.cap_embedder.0.weight") != std::string::npos) {
             return VERSION_Z_IMAGE;
         }
+        if (tensor_storage.name.find("double_stream_layers.0.img_instruct_attn.processor.img_to_q.weight") != std::string::npos) {
+            return VERSION_BOOGU_IMAGE;
+        }
         if (tensor_storage.name.find("model.diffusion_model.layers.0.adaLN_sa_ln.weight") != std::string::npos) {
             return VERSION_ERNIE_IMAGE;
         }

src/model_manager.cpp

@@ -147,6 +147,17 @@ bool ModelManager::register_param_tensors(const std::string& desc,
     return true;
 }
 
+bool ModelManager::load_all_params_eagerly() {
+    std::vector<TensorState*> all_states;
+    all_states.reserve(tensor_states_.size());
+    for (const auto& s : tensor_states_) {
+        if (s != nullptr) {
+            all_states.push_back(s.get());
+        }
+    }
+    return load_tensors_to_params_backend(all_states);
+}
+
 bool ModelManager::validate_registered_tensors() {
     bool ok = true;
     for (const auto& state : tensor_states_) {
@@ -469,7 +480,7 @@ bool ModelManager::mmap_params(const std::vector<TensorState*>& states,
         return true;
     }
 
-    auto mmap_store = model_loader_.mmap_tensors(mmap_candidates, {}, true);
+    auto mmap_store = model_loader_.mmap_tensors(mmap_candidates, {}, writable_mmap_);
     if (mmap_store.empty()) {
         return true;
     }

src/model_manager.h

@@ -69,6 +69,7 @@ private:
     uint64_t current_lora_epoch_ = 0;
     int n_threads_               = 0;
     bool enable_mmap_            = false;
+    bool writable_mmap_          = false;
 
     void finish_compute_backend_usage(const std::vector<TensorState*>& states);
     void release_all();
@@ -110,6 +111,7 @@ public:
         model_loader_.set_n_threads(n_threads);
     }
     void set_enable_mmap(bool enable_mmap) { enable_mmap_ = enable_mmap; }
+    void set_writable_mmap(bool writable_mmap) { writable_mmap_ = writable_mmap; }
     void set_common_ignore_tensors(std::set<std::string> ignore_tensors);
     void set_loras(std::vector<LoraSpec> loras, SDVersion version);
 
@@ -158,6 +160,7 @@ public:
     }
 
     bool validate_registered_tensors();
+    bool load_all_params_eagerly();
 
     bool prepare_params(const std::vector<ggml_tensor*>& tensors) override;
     void release_compute_backend_params(const std::vector<ggml_tensor*>& tensors) override;

src/name_conversion.cpp

@@ -184,6 +184,27 @@ std::string convert_cond_stage_model_name(std::string name, std::string prefix)
     return name;
 }
 
+std::string convert_qwen3_vl_vision_name(std::string name) {
+    static const std::vector<std::pair<std::string, std::string>> qwen3_vl_vision_name_map{
+        {"mm.0.", "merger.linear_fc1."},
+        {"mm.2.", "merger.linear_fc2."},
+        {"v.post_ln.", "merger.norm."},
+        {"v.position_embd.weight", "pos_embed.weight"},
+        {"v.patch_embd.weight.1", "patch_embed.proj.1.weight"},
+        {"v.patch_embd.weight", "patch_embed.proj.0.weight"},
+        {"v.patch_embd.bias", "patch_embed.bias"},
+        {"v.blk.", "blocks."},
+        {"attn_qkv.", "attn.qkv."},
+        {"attn_out.", "attn.proj."},
+        {"ffn_up.", "mlp.linear_fc1."},
+        {"ffn_down.", "mlp.linear_fc2."},
+        {"ln1.", "norm1."},
+        {"ln2.", "norm2."},
+    };
+    replace_with_name_map(name, qwen3_vl_vision_name_map);
+    return name;
+}
+
 // ref: https://github.com/huggingface/diffusers/blob/main/scripts/convert_diffusers_to_original_stable_diffusion.py
 std::string convert_diffusers_unet_to_original_sd1(std::string name) {
     // (stable-diffusion, HF Diffusers)
@@ -1154,6 +1175,10 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
 
     replace_with_prefix_map(name, prefix_map);
 
+    if (sd_version_is_boogu_image(version) && starts_with(name, "text_encoders.llm.visual.")) {
+        name = convert_qwen3_vl_vision_name(std::move(name));
+    }
+
     // diffusion model
     {
         for (const auto& prefix : diffuison_model_prefix_vec) {

src/runtime/guidance.cpp

@@ -3,6 +3,7 @@
 #include <algorithm>
 #include <cmath>
 #include <cstdlib>
+#include <optional>
 #include <string>
 #include <utility>
 
@@ -63,6 +64,82 @@ namespace sd::guidance {
         return uncond;
     }
 
+    std::vector<float> parse_guidance_schedule_from_spec(std::string spec) {
+        std::vector<float> schedule;
+
+        while (!spec.empty()) {
+            auto sep     = spec.find('+');
+            auto segment = spec.substr(0, sep);
+
+            auto x = segment.find('x');
+            if (x == std::string::npos) {
+                LOG_ERROR("Invalid guidance schedule segment: '%s' (expected <guidance>x<count>)", segment.c_str());
+                return {};
+            }
+
+            float guidance;
+            int count;
+
+            auto guidance_str = segment.substr(0, x);
+            auto count_str    = segment.substr(x + 1);
+
+            try {
+                size_t idx = 0;
+                guidance   = std::stof(guidance_str, &idx);
+                if (idx != guidance_str.size()) {
+                    LOG_ERROR("Invalid guidance value in guidance schedule: '%s'", guidance_str.c_str());
+                    return {};
+                }
+            } catch (const std::exception&) {
+                LOG_ERROR("Invalid guidance value in guidance schedule: '%s'", guidance_str.c_str());
+                return {};
+            }
+
+            try {
+                size_t idx = 0;
+                count      = std::stoi(count_str, &idx);
+                if (idx != count_str.size()) {
+                    LOG_ERROR("Invalid count in guidance schedule: '%s'", count_str.c_str());
+                    return {};
+                }
+            } catch (const std::exception&) {
+                LOG_ERROR("Invalid count in guidance schedule: '%s'", count_str.c_str());
+                return {};
+            }
+
+            if (count <= 0) {
+                LOG_ERROR("Guidance schedule count must be positive");
+                return {};
+            }
+
+            schedule.insert(schedule.end(), count, guidance);
+
+            if (sep == std::string::npos) {
+                break;
+            }
+
+            spec = spec.substr(sep + 1);
+        }
+
+        return schedule;
+    }
+
+    std::vector<float> parse_guidance_schedule(const char* extra_sample_args) {
+        std::vector<float> guidance_schedule;
+        std::string guidance_schedule_str = "";
+        for (const auto& [key, value] : parse_key_value_args(extra_sample_args, "extra sample arg")) {
+            float parsed = 0.0f;
+            if (key == "guidance_schedule") {
+                guidance_schedule_str = value;
+            }
+        }
+
+        if (!guidance_schedule_str.empty()) {
+            guidance_schedule = parse_guidance_schedule_from_spec(guidance_schedule_str);
+        }
+        return guidance_schedule;
+    }
+
     ClassifierFreeGuidance::ClassifierFreeGuidance(float guidance_scale,
                                                    float image_guidance_scale)
         : guidance_scale_(guidance_scale),
@@ -70,8 +147,10 @@ namespace sd::guidance {
     }
 
     GuiderOutput ClassifierFreeGuidance::forward(const GuidanceInput& input,
-                                                 GuiderOutput previous) const {
+                                                 GuiderOutput previous,
+                                                 std::optional<float> scale_override) const {
         (void)previous;
+        float guidance_scale = scale_override.value_or(guidance_scale_);
 
         GuiderOutput output;
         if (!has_tensor(input.pred_cond)) {
@@ -86,14 +165,14 @@ namespace sd::guidance {
                 const sd::Tensor<float>& pred_img_uncond = *input.pred_img_uncond;
                 output.pred                              = pred_img_uncond +
                               image_guidance_scale_ * (pred_uncond - pred_img_uncond) +
-                              guidance_scale_ * (pred_cond - pred_uncond);
+                              guidance_scale * (pred_cond - pred_uncond);
 
             } else {
-                output.pred = pred_uncond + guidance_scale_ * (pred_cond - pred_uncond);
+                output.pred = pred_uncond + guidance_scale * (pred_cond - pred_uncond);
             }
         } else if (has_tensor(input.pred_img_uncond)) {
             const sd::Tensor<float>& pred_img_uncond = *input.pred_img_uncond;
-            output.pred                              = pred_img_uncond + guidance_scale_ * (pred_cond - pred_img_uncond);
+            output.pred                              = pred_img_uncond + guidance_scale * (pred_cond - pred_img_uncond);
         }
 
         return output;
@@ -128,8 +207,10 @@ namespace sd::guidance {
     }
 
     GuiderOutput AdaptiveProjectedGuidance::forward(const GuidanceInput& input,
-                                                    GuiderOutput previous) const {
+                                                    GuiderOutput previous,
+                                                    std::optional<float> scale_override) const {
         (void)previous;
+        float guidance_scale = scale_override.value_or(guidance_scale_);
 
         GuiderOutput output;
         if (!has_tensor(input.pred_cond)) {
@@ -144,13 +225,13 @@ namespace sd::guidance {
                 const sd::Tensor<float>& pred_img_uncond = *input.pred_img_uncond;
                 output.pred                              = pred_img_uncond +
                               image_guidance_scale_ * (pred_uncond - pred_img_uncond) +
-                              guidance_scale_ * (pred_cond - pred_uncond);
+                              guidance_scale * (pred_cond - pred_uncond);
             } else {
-                output.pred = pred_uncond + guidance_scale_ * (pred_cond - pred_uncond);
+                output.pred = pred_uncond + guidance_scale * (pred_cond - pred_uncond);
             }
         } else if (has_tensor(input.pred_img_uncond)) {
             const sd::Tensor<float>& pred_img_uncond = *input.pred_img_uncond;
-            output.pred                              = pred_img_uncond + guidance_scale_ * (pred_cond - pred_img_uncond);
+            output.pred                              = pred_img_uncond + guidance_scale * (pred_cond - pred_img_uncond);
         }
         if (!has_tensor(input.pred_uncond) && !has_tensor(input.pred_img_uncond)) {
             return output;
@@ -162,7 +243,7 @@ namespace sd::guidance {
         sd::Tensor<float> deltas = calculate_guidance_delta(pred_cond,
                                                             pred_uncond,
                                                             pred_img_uncond,
-                                                            guidance_scale_,
+                                                            guidance_scale,
                                                             image_guidance_scale_);
         if (params_.momentum != 0.0f) {
             if (momentum_buffer_.shape() != deltas.shape()) {
@@ -239,7 +320,8 @@ namespace sd::guidance {
     }
 
     GuiderOutput SkipLayerGuidance::forward(const GuidanceInput& input,
-                                            GuiderOutput output) const {
+                                            GuiderOutput output,
+                                            std::optional<float> /*scale_override*/) const {
         if (scale_ == 0.0f || !is_enabled_for_step(input) || !input.predict_skip_layer) {
             return output;
         }

src/runtime/guidance.h

@@ -3,6 +3,7 @@
 
 #include <cstddef>
 #include <functional>
+#include <optional>
 #include <vector>
 
 #include "core/tensor.hpp"
@@ -27,6 +28,7 @@ namespace sd::guidance {
     AdaptiveProjectedGuidanceParams parse_adaptive_projected_guidance_args(const char* extra_sample_args);
     bool is_adaptive_projected_guidance_enabled(const AdaptiveProjectedGuidanceParams& params);
     bool parse_skip_layer_guidance_uncond_arg(const char* extra_sample_args);
+    std::vector<float> parse_guidance_schedule(const char* extra_sample_args);
 
     struct GuidanceInput {
         int step                                 = 0;
@@ -40,9 +42,10 @@ namespace sd::guidance {
 
     class BaseGuidance {
     public:
-        virtual ~BaseGuidance()                                   = default;
+        virtual ~BaseGuidance()                                                                = default;
         virtual GuiderOutput forward(const GuidanceInput& input,
-                                     GuiderOutput previous) const = 0;
+                                     GuiderOutput previous,
+                                     std::optional<float> scale_override = std::nullopt) const = 0;
     };
 
     class ClassifierFreeGuidance : public BaseGuidance {
@@ -54,7 +57,8 @@ namespace sd::guidance {
                                float image_guidance_scale);
 
         GuiderOutput forward(const GuidanceInput& input,
-                             GuiderOutput previous) const override;
+                             GuiderOutput previous,
+                             std::optional<float> scale_override = std::nullopt) const override;
     };
 
     class AdaptiveProjectedGuidance : public BaseGuidance {
@@ -69,7 +73,8 @@ namespace sd::guidance {
                                   AdaptiveProjectedGuidanceParams params);
 
         GuiderOutput forward(const GuidanceInput& input,
-                             GuiderOutput previous) const override;
+                             GuiderOutput previous,
+                             std::optional<float> scale_override = std::nullopt) const override;
     };
 
     class SkipLayerGuidance : public BaseGuidance {
@@ -88,7 +93,8 @@ namespace sd::guidance {
         const std::vector<int>& layers() const;
 
         GuiderOutput forward(const GuidanceInput& input,
-                             GuiderOutput previous) const override;
+                             GuiderOutput previous,
+                             std::optional<float> scale_override = std::nullopt) const override;
     };
 
 }  // namespace sd::guidance

src/stable-diffusion.cpp

@@ -20,6 +20,7 @@
 #include "extensions/generation_extension.h"
 #include "model/adapter/lora.hpp"
 #include "model/diffusion/anima.hpp"
+#include "model/diffusion/boogu.hpp"
 #include "model/diffusion/control.hpp"
 #include "model/diffusion/ernie_image.hpp"
 #include "model/diffusion/flux.hpp"
@@ -87,6 +88,7 @@ const char* model_version_to_str[] = {
     "LTXAV",
     "HiDream O1",
     "Z-Image",
+    "Boogu Image",
     "Ovis Image",
     "Ernie Image",
     "Lens",
@@ -124,7 +126,8 @@ static bool sd_version_supports_ref_latent_img_cfg(SDVersion version) {
            sd_version_is_flux2(version) ||
            sd_version_is_qwen_image(version) ||
            sd_version_is_longcat(version) ||
-           sd_version_is_z_image(version);
+           sd_version_is_z_image(version) ||
+           sd_version_is_boogu_image(version);
 }
 
 static bool sd_version_supports_img_cfg(SDVersion version, bool has_ref_images) {
@@ -196,6 +199,7 @@ public:
     bool enable_mmap                     = false;
     sd::ggml_graph_cut::MaxVramAssignment max_vram_assignment;
     bool stream_layers = false;
+    bool eager_load    = false;
     std::string backend_spec;
     std::string params_backend_spec;
 
@@ -339,6 +343,7 @@ public:
         n_threads           = sd_ctx_params->n_threads;
         enable_mmap         = sd_ctx_params->enable_mmap;
         stream_layers       = sd_ctx_params->stream_layers;
+        eager_load          = sd_ctx_params->eager_load;
         backend_spec        = SAFE_STR(sd_ctx_params->backend);
         params_backend_spec = SAFE_STR(sd_ctx_params->params_backend);
         max_vram_assignment.reset(0.f);
@@ -527,7 +532,6 @@ public:
         if (wtype != GGML_TYPE_COUNT || tensor_type_rules.size() > 0) {
             model_loader.set_wtype_override(wtype, tensor_type_rules);
         }
-        model_loader.process_model_files(enable_mmap, true);
 
         std::map<ggml_type, uint32_t> wtype_stat                 = model_loader.get_wtype_stat();
         std::map<ggml_type, uint32_t> conditioner_wtype_stat     = model_loader.get_conditioner_wtype_stat();
@@ -581,9 +585,12 @@ public:
             apply_lora_immediately = false;
         }
 
+        bool needs_writable_mmap = enable_mmap && apply_lora_immediately;
+        model_manager->set_writable_mmap(needs_writable_mmap);
         if (enable_mmap && apply_lora_immediately) {
             LOG_WARN("in mode 'immediately', LoRAs will cause extra memory usage with mmap");
         }
+        model_loader.process_model_files(enable_mmap, needs_writable_mmap);
         load_alphas_cumprod(model_loader);
 
         size_t text_encoder_params_mem_size = 0;
@@ -784,6 +791,18 @@ public:
                                                                          "model.diffusion_model",
                                                                          version,
                                                                          model_manager);
+            } else if (sd_version_is_boogu_image(version)) {
+                cond_stage_model = std::make_shared<LLMEmbedder>(backend_for(SDBackendModule::TE),
+                                                                 tensor_storage_map,
+                                                                 version,
+                                                                 "",
+                                                                 true,
+                                                                 model_manager);
+                diffusion_model  = std::make_shared<Boogu::BooguImageRunner>(backend_for(SDBackendModule::DIFFUSION),
+                                                                            tensor_storage_map,
+                                                                            "model.diffusion_model",
+                                                                            version,
+                                                                            model_manager);
             } else if (sd_version_is_ernie_image(version)) {
                 cond_stage_model = std::make_shared<LLMEmbedder>(backend_for(SDBackendModule::TE),
                                                                  tensor_storage_map,
@@ -1138,7 +1157,15 @@ public:
             return false;
         }
 
-        LOG_DEBUG("model metadata validated; weights will be prepared lazily");
+        if (eager_load) {
+            if (!model_manager->load_all_params_eagerly()) {
+                LOG_ERROR("model params eager load failed");
+                return false;
+            }
+            LOG_DEBUG("model metadata validated; weights pre-loaded to params backend");
+        } else {
+            LOG_DEBUG("model metadata validated; weights will be prepared lazily");
+        }
 
         {
             size_t total_params_ram_size  = 0;
@@ -1220,6 +1247,7 @@ public:
                            sd_version_is_anima(version) ||
                            sd_version_is_ernie_image(version) ||
                            sd_version_is_z_image(version) ||
+                           sd_version_is_boogu_image(version) ||
                            sd_version_is_pid(version) ||
                            sd_version_is_ideogram4(version)) {
                     pred_type = FLOW_PRED;
@@ -1231,6 +1259,8 @@ public:
                         default_flow_shift = 1.5f;
                     } else if (sd_version_is_ideogram4(version)) {
                         default_flow_shift = 1.0f;
+                    } else if (sd_version_is_boogu_image(version)) {
+                        default_flow_shift = 3.16f;
                     } else {
                         default_flow_shift = 3.f;
                     }
@@ -1691,7 +1721,7 @@ public:
                 if (sd_version_is_sd3(version)) {
                     latent_rgb_proj = sd3_latent_rgb_proj;
                     latent_rgb_bias = sd3_latent_rgb_bias;
-                } else if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_longcat(version)) {
+                } else if (sd_version_uses_flux_vae(version)) {
                     latent_rgb_proj = flux_latent_rgb_proj;
                     latent_rgb_bias = flux_latent_rgb_bias;
                 } else if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
@@ -1786,6 +1816,9 @@ public:
         if (sd_version_is_anima(version)) {
             return std::vector<float>{t / static_cast<float>(TIMESTEPS)};
         }
+        if (sd_version_is_boogu_image(version)) {
+            return std::vector<float>{t / static_cast<float>(TIMESTEPS)};
+        }
         if (version == VERSION_HIDREAM_O1) {
             return std::vector<float>{1.0f - (t / static_cast<float>(TIMESTEPS))};
         }
@@ -1911,6 +1944,32 @@ public:
         float slg_scale     = guidance.slg.scale;
         bool slg_uncond     = sd::guidance::parse_skip_layer_guidance_uncond_arg(extra_sample_args);
 
+        std::vector<float> guidance_schedule = sd::guidance::parse_guidance_schedule(extra_sample_args);
+        if (!guidance_schedule.empty() && guidance_schedule.size() != sigmas.size() - 1) {
+            if (guidance_schedule.size() > sigmas.size()) {
+                LOG_WARN("guidance_schedule length (%zu) is greater than number of steps (%zu)", guidance_schedule.size(), sigmas.size() - 1);
+                LOG_WARN("truncating guidance_schedule to match step count");
+                guidance_schedule.resize(sigmas.size() - 1);
+            } else {
+                LOG_INFO("padding guidance_schedule with cfg_scale");
+                while (guidance_schedule.size() < sigmas.size() - 1) {
+                    guidance_schedule.push_back(cfg_scale);
+                }
+            }
+        }
+
+        if (!guidance_schedule.empty()) {
+            std::string schedule_str = "[";
+            for (size_t i = 0; i < guidance_schedule.size(); ++i) {
+                schedule_str += std::to_string(guidance_schedule[i]);
+                if (i < guidance_schedule.size() - 1) {
+                    schedule_str += ", ";
+                }
+            }
+            schedule_str += "]";
+            LOG_DEBUG("using guidance schedule: %s", schedule_str.c_str());
+        }
+
         sd_sample::SampleCacheRuntime cache_runtime = sd_sample::init_sample_cache_runtime(version,
                                                                                            cache_params,
                                                                                            denoiser.get(),
@@ -2151,7 +2210,7 @@ public:
             guidance_input.pred_uncond     = uncond_out.empty() ? nullptr : &uncond_out;
             guidance_input.pred_img_uncond = img_uncond_out.empty() ? nullptr : &img_uncond_out;
 
-            sd::guidance::GuiderOutput guided = primary_guidance.forward(guidance_input, {});
+            sd::guidance::GuiderOutput guided = guidance_schedule.empty() ? primary_guidance.forward(guidance_input, {}) : primary_guidance.forward(guidance_input, {}, guidance_schedule[guidance_schedule.size() - 1 - step]);
             if (guided.pred.empty()) {
                 return {};
             }
@@ -2675,6 +2734,7 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     sd_ctx_params->lora_apply_mode      = LORA_APPLY_AUTO;
     sd_ctx_params->max_vram             = nullptr;
     sd_ctx_params->stream_layers        = false;
+    sd_ctx_params->eager_load           = false;
     sd_ctx_params->enable_mmap          = false;
     sd_ctx_params->diffusion_flash_attn = false;
     sd_ctx_params->circular_x           = false;
@@ -2721,6 +2781,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              "prediction: %s\n"
              "max_vram: %s\n"
              "stream_layers: %s\n"
+             "eager_load: %s\n"
              "backend: %s\n"
              "params_backend: %s\n"
              "flash_attn: %s\n"
@@ -2756,6 +2817,7 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              sd_prediction_name(sd_ctx_params->prediction),
              SAFE_STR(sd_ctx_params->max_vram),
              BOOL_STR(sd_ctx_params->stream_layers),
+             BOOL_STR(sd_ctx_params->eager_load),
              SAFE_STR(sd_ctx_params->backend),
              SAFE_STR(sd_ctx_params->params_backend),
              BOOL_STR(sd_ctx_params->flash_attn),