perf: improved flux attention qkv unpacking (#1306)

* feat: turn flow_shift into a generation parameter

* format code

* simplify set_shift/set_parameters

* fix sd_sample_params_to_str

* remove unused variable

* update docs

---------

Co-authored-by: leejet <leejet714@gmail.com>

.github/workflows/build.yml

@@ -535,31 +535,30 @@ jobs:
           # Add ROCm to PATH for current session
           echo "/opt/rocm/bin" >> $GITHUB_PATH
 
-          # Build case pattern from GPU_TARGETS
-          PATTERN=$(printf '%s' "$GPU_TARGETS" | sed 's/;/\*|\*/g')
-          PATTERN="*${PATTERN}*"
+          # Build regex pattern from ${{ env.GPU_TARGETS }} (match target as substring)
+          TARGET_REGEX="($(printf '%s' "${{ env.GPU_TARGETS }}" | sed 's/;/|/g'))"
 
           # Remove library files for architectures we're not building for to save disk space
           echo "Cleaning up unneeded architecture files..."
           cd /opt/rocm/lib/rocblas/library
           # Keep only our target architectures
           for file in *; do
-            case "$file" in
-            $PATTERN)
-              ;;
-            *)
-              sudo rm -f "$file" ;;
-            esac;
+            if printf '%s' "$file" | grep -q 'gfx'; then
+              if ! printf '%s' "$file" | grep -Eq "$TARGET_REGEX"; then
+                echo "Removing $file" &&
+                sudo rm -f "$file";
+              fi
+            fi
           done
 
           cd /opt/rocm/lib/hipblaslt/library
           for file in *; do
-            case "$file" in
-              $PATTERN)
-                ;;
-              *)
-                sudo rm -f "$file" ;;
-            esac;
+            if printf '%s' "$file" | grep -q 'gfx'; then
+              if ! printf '%s' "$file" | grep -Eq "$TARGET_REGEX"; then
+                echo "Removing $file" &&
+                sudo rm -f "$file";
+              fi
+            fi
           done
 
       - name: Build
@@ -592,21 +591,15 @@ jobs:
           cp ggml/LICENSE ./build/bin/ggml.txt
           cp LICENSE ./build/bin/stable-diffusion.cpp.txt
 
-          # Create directories for ROCm libraries
-          mkdir -p ./build/bin/rocblas/library
-          mkdir -p ./build/bin/hipblaslt/library
-
-          # Copy ROCm runtime libraries (use || true to continue if files don't exist)
-          cp /opt/rocm/lib/librocsparse.so* ./build/bin/ || true
-          cp /opt/rocm/lib/libhsa-runtime64.so* ./build/bin/ || true
-          cp /opt/rocm/lib/libamdhip64.so* ./build/bin/ || true
-          cp /opt/rocm/lib/libhipblas.so* ./build/bin/ || true
-          cp /opt/rocm/lib/libhipblaslt.so* ./build/bin/ || true
-          cp /opt/rocm/lib/librocblas.so* ./build/bin/ || true
-
-          # Copy library files (already filtered to target architectures)
-          cp /opt/rocm/lib/rocblas/library/* ./build/bin/rocblas/library/ || true
-          cp /opt/rocm/lib/hipblaslt/library/* ./build/bin/hipblaslt/library/ || true
+          # Move ROCm runtime libraries (to avoid double space consumption)
+          sudo mv /opt/rocm/lib/librocsparse.so* ./build/bin/
+          sudo mv /opt/rocm/lib/libhsa-runtime64.so* ./build/bin/
+          sudo mv /opt/rocm/lib/libamdhip64.so* ./build/bin/
+          sudo mv /opt/rocm/lib/libhipblas.so* ./build/bin/
+          sudo mv /opt/rocm/lib/libhipblaslt.so* ./build/bin/
+          sudo mv /opt/rocm/lib/librocblas.so* ./build/bin/
+          sudo mv /opt/rocm/lib/rocblas/ ./build/bin/
+          sudo mv /opt/rocm/lib/hipblaslt/ ./build/bin/
 
       - name: Fetch system info
         id: system-info
@@ -622,7 +615,7 @@ jobs:
         run: |
           cp ggml/LICENSE ./build/bin/ggml.txt
           cp LICENSE ./build/bin/stable-diffusion.cpp.txt
-          zip -j sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip ./build/bin/*
+          zip -y -r sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip ./build/bin
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}

CMakeLists.txt

@@ -87,9 +87,11 @@ endif()
 set(SD_LIB stable-diffusion)
 
 file(GLOB SD_LIB_SOURCES
-    "*.h"
-    "*.cpp"
-    "*.hpp"
+    "src/*.h"
+    "src/*.cpp"
+    "src/*.hpp"
+    "src/vocab/*.h"
+    "src/vocab/*.cpp"
 )
 
 find_program(GIT_EXE NAMES git git.exe NO_CMAKE_FIND_ROOT_PATH)
@@ -119,7 +121,7 @@ endif()
 message(STATUS "stable-diffusion.cpp commit ${SDCPP_BUILD_COMMIT}")
 
 set_property(
-  SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/version.cpp
+  SOURCE ${CMAKE_CURRENT_SOURCE_DIR}/src/version.cpp
   APPEND PROPERTY COMPILE_DEFINITIONS
   SDCPP_BUILD_COMMIT=${SDCPP_BUILD_COMMIT} SDCPP_BUILD_VERSION=${SDCPP_BUILD_VERSION}
 )
@@ -182,6 +184,7 @@ endif()
 add_subdirectory(thirdparty)
 
 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
+target_include_directories(${SD_LIB} PUBLIC . include)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
 target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)
 
@@ -190,7 +193,7 @@ if (SD_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
 
-set(SD_PUBLIC_HEADERS stable-diffusion.h)
+set(SD_PUBLIC_HEADERS include/stable-diffusion.h)
 set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
 
 install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)

README.md

@@ -53,6 +53,7 @@ API and command-line option may change frequently.***
     - [Qwen Image](./docs/qwen_image.md)
     - [Z-Image](./docs/z_image.md)
     - [Ovis-Image](./docs/ovis_image.md)
+    - [Anima](./docs/anima.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
@@ -139,6 +140,7 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [🔥Wan2.1/Wan2.2](./docs/wan.md)
 - [🔥Z-Image](./docs/z_image.md)
 - [Ovis-Image](./docs/ovis_image.md)
+- [Anima](./docs/anima.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)

docs/anima.md

@@ -0,0 +1,20 @@
+# How to Use
+
+## Download weights
+
+- Download Anima
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/Bedovyy/Anima-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
+- Download Qwen3-0.6B-Base
+    - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen3-0.6B-Base-GGUF/tree/main
+
+## Examples
+
+```sh
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\anima-preview.safetensors --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="anima image example" src="../assets/anima/example.png" />

examples/cli/README.md

@@ -44,7 +44,6 @@ Context Options:
                                            CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
   --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --vae-tiling                             process vae in tiles to reduce memory usage
   --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
   --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
@@ -109,6 +108,7 @@ Generation Options:
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
   --eta <float>                            eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
+  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
   --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
   --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)

examples/cli/main.cpp

@@ -394,12 +394,15 @@ bool save_results(const SDCliParams& cli_params,
 
     fs::path base_path = out_path;
     fs::path ext       = out_path.has_extension() ? out_path.extension() : fs::path{};
-    if (!ext.empty())
-        base_path.replace_extension();
 
     std::string ext_lower = ext.string();
     std::transform(ext_lower.begin(), ext_lower.end(), ext_lower.begin(), ::tolower);
     bool is_jpg = (ext_lower == ".jpg" || ext_lower == ".jpeg" || ext_lower == ".jpe");
+    if (!ext.empty()) {
+        if (is_jpg || ext_lower == ".png") {
+            base_path.replace_extension();
+        }
+    }
 
     int output_begin_idx = cli_params.output_begin_idx;
     if (output_begin_idx < 0) {
@@ -409,7 +412,7 @@ bool save_results(const SDCliParams& cli_params,
     auto write_image = [&](const fs::path& path, int idx) {
         const sd_image_t& img = results[idx];
         if (!img.data)
-            return;
+            return false;
 
         std::string params = get_image_params(cli_params, ctx_params, gen_params, gen_params.seed + idx);
         int ok             = 0;
@@ -419,8 +422,11 @@ bool save_results(const SDCliParams& cli_params,
             ok = stbi_write_png(path.string().c_str(), img.width, img.height, img.channel, img.data, 0, params.c_str());
         }
         LOG_INFO("save result image %d to '%s' (%s)", idx, path.string().c_str(), ok ? "success" : "failure");
+        return ok != 0;
     };
 
+    int sucessful_reults = 0;
+
     if (std::regex_search(cli_params.output_path, format_specifier_regex)) {
         if (!is_jpg && ext_lower != ".png")
             ext = ".png";
@@ -429,9 +435,12 @@ bool save_results(const SDCliParams& cli_params,
 
         for (int i = 0; i < num_results; ++i) {
             fs::path img_path = format_frame_idx(pattern.string(), output_begin_idx + i);
-            write_image(img_path, i);
+            if (write_image(img_path, i)) {
+                sucessful_reults++;
+            }
         }
-        return true;
+        LOG_INFO("%d/%d images saved", sucessful_reults, num_results);
+        return sucessful_reults != 0;
     }
 
     if (cli_params.mode == VID_GEN && num_results > 1) {
@@ -439,9 +448,13 @@ bool save_results(const SDCliParams& cli_params,
             ext = ".avi";
         fs::path video_path = base_path;
         video_path += ext;
-        create_mjpg_avi_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps);
-        LOG_INFO("save result MJPG AVI video to '%s'", video_path.string().c_str());
-        return true;
+        if (create_mjpg_avi_from_sd_images(video_path.string().c_str(), results, num_results, gen_params.fps) == 0) {
+            LOG_INFO("save result MJPG AVI video to '%s'", video_path.string().c_str());
+            return true;
+        } else {
+            LOG_ERROR("Failed to save result MPG AVI video to '%s'", video_path.string().c_str());
+            return false;
+        }
     }
 
     if (!is_jpg && ext_lower != ".png")
@@ -453,10 +466,12 @@ bool save_results(const SDCliParams& cli_params,
             img_path += "_" + std::to_string(output_begin_idx + i);
         }
         img_path += ext;
-        write_image(img_path, i);
+        if (write_image(img_path, i)) {
+            sucessful_reults++;
+        }
     }
-
-    return true;
+    LOG_INFO("%d/%d images saved", sucessful_reults, num_results);
+    return sucessful_reults != 0;
 }
 
 int main(int argc, const char* argv[]) {

examples/common/common.hpp

@@ -581,10 +581,6 @@ struct SDContextParams {
              "--vae-tile-overlap",
              "tile overlap for vae tiling, in fraction of tile size (default: 0.5)",
              &vae_tiling_params.target_overlap},
-            {"",
-             "--flow-shift",
-             "shift value for Flow models like SD3.x or WAN (default: auto)",
-             &flow_shift},
         };
 
         options.bool_options = {
@@ -903,7 +899,6 @@ struct SDContextParams {
             << "  photo_maker_path: \"" << photo_maker_path << "\",\n"
             << "  rng_type: " << sd_rng_type_name(rng_type) << ",\n"
             << "  sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
-            << "  flow_shift: " << (std::isinf(flow_shift) ? "INF" : std::to_string(flow_shift)) << "\n"
             << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
             << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
             << "  control_net_cpu: " << (control_net_cpu ? "true" : "false") << ",\n"
@@ -986,7 +981,6 @@ struct SDContextParams {
             chroma_use_t5_mask,
             chroma_t5_mask_pad,
             qwen_image_zero_cond_t,
-            flow_shift,
         };
         return sd_ctx_params;
     }
@@ -1206,6 +1200,10 @@ struct SDGenerationParams {
              "--eta",
              "eta in DDIM, only for DDIM and TCD (default: 0)",
              &sample_params.eta},
+            {"",
+             "--flow-shift",
+             "shift value for Flow models like SD3.x or WAN (default: auto)",
+             &sample_params.flow_shift},
             {"",
              "--high-noise-cfg-scale",
              "(high noise) unconditional guidance scale: (default: 7.0)",
@@ -1606,6 +1604,7 @@ struct SDGenerationParams {
         load_if_exists("cfg_scale", sample_params.guidance.txt_cfg);
         load_if_exists("img_cfg_scale", sample_params.guidance.img_cfg);
         load_if_exists("guidance", sample_params.guidance.distilled_guidance);
+        load_if_exists("flow_shift", sample_params.flow_shift);
 
         auto load_sampler_if_exists = [&](const char* key, enum sample_method_t& out) {
             if (j.contains(key) && j[key].is_string()) {

examples/server/README.md

@@ -36,7 +36,6 @@ Context Options:
                                            CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
   --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --vae-tiling                             process vae in tiles to reduce memory usage
   --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
   --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
@@ -101,6 +100,7 @@ Default Generation Options:
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
   --eta <float>                            eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
+  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
   --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
   --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)

examples/server/main.cpp

@@ -266,6 +266,7 @@ void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
 struct LoraEntry {
     std::string name;
     std::string path;
+    std::string fullpath;
 };
 
 int main(int argc, const char** argv) {
@@ -321,7 +322,8 @@ int main(int argc, const char** argv) {
 
                 LoraEntry e;
                 e.name          = p.stem().u8string();
-                std::string rel = fs::relative(p, lora_dir).u8string();
+                e.fullpath      = p.u8string();
+                std::string rel = p.lexically_relative(lora_dir).u8string();
                 std::replace(rel.begin(), rel.end(), '\\', '/');
                 e.path = rel;
 
@@ -340,10 +342,11 @@ int main(int argc, const char** argv) {
         }
     };
 
-    auto is_valid_lora_path = [&](const std::string& path) -> bool {
+    auto get_lora_full_path = [&](const std::string& path) -> std::string {
         std::lock_guard<std::mutex> lock(lora_mutex);
-        return std::any_of(lora_cache.begin(), lora_cache.end(),
+        auto it = std::find_if(lora_cache.begin(), lora_cache.end(),
                            [&](const LoraEntry& e) { return e.path == path; });
+        return (it != lora_cache.end()) ? it->fullpath : "";
     };
 
     httplib::Server svr;
@@ -404,8 +407,8 @@ int main(int argc, const char** argv) {
             std::string size          = j.value("size", "");
             std::string output_format = j.value("output_format", "png");
             int output_compression    = j.value("output_compression", 100);
-            int width                 = 512;
-            int height                = 512;
+            int width                 = default_gen_params.width > 0 ? default_gen_params.width : 512;
+            int height                = default_gen_params.width > 0 ? default_gen_params.height : 512;
             if (!size.empty()) {
                 auto pos = size.find('x');
                 if (pos != std::string::npos) {
@@ -565,7 +568,8 @@ int main(int argc, const char** argv) {
             std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(prompt);
 
             size_t image_count = req.form.get_file_count("image[]");
-            if (image_count == 0) {
+            bool has_legacy_image = req.form.has_file("image");
+            if (image_count == 0 && !has_legacy_image) {
                 res.status = 400;
                 res.set_content(R"({"error":"at least one image[] required"})", "application/json");
                 return;
@@ -576,6 +580,10 @@ int main(int argc, const char** argv) {
                 auto file = req.form.get_file("image[]", i);
                 images_bytes.emplace_back(file.content.begin(), file.content.end());
             }
+            if (image_count == 0 && has_legacy_image) {
+                auto file = req.form.get_file("image");
+                images_bytes.emplace_back(file.content.begin(), file.content.end());
+            }
 
             std::vector<uint8_t> mask_bytes;
             if (req.form.has_file("mask")) {
@@ -593,7 +601,7 @@ int main(int argc, const char** argv) {
             n = std::clamp(n, 1, 8);
 
             std::string size = req.form.get_field("size");
-            int width = 512, height = 512;
+            int width = -1, height = -1;
             if (!size.empty()) {
                 auto pos = size.find('x');
                 if (pos != std::string::npos) {
@@ -650,15 +658,31 @@ int main(int argc, const char** argv) {
 
             LOG_DEBUG("%s\n", gen_params.to_string().c_str());
 
-            sd_image_t init_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
-            sd_image_t control_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+            sd_image_t init_image    = {0, 0, 3, nullptr};
+            sd_image_t control_image = {0, 0, 3, nullptr};
             std::vector<sd_image_t> pmid_images;
 
+            auto get_resolved_width = [&gen_params, &default_gen_params]() -> int {
+                if (gen_params.width > 0)
+                    return gen_params.width;
+                if (default_gen_params.width > 0)
+                    return default_gen_params.width;
+                return 512;
+            };
+            auto get_resolved_height = [&gen_params, &default_gen_params]() -> int {
+                if (gen_params.height > 0)
+                    return gen_params.height;
+                if (default_gen_params.height > 0)
+                    return default_gen_params.height;
+                return 512;
+            };
+
             std::vector<sd_image_t> ref_images;
             ref_images.reserve(images_bytes.size());
             for (auto& bytes : images_bytes) {
-                int img_w           = width;
-                int img_h           = height;
+                int img_w;
+                int img_h;
+
                 uint8_t* raw_pixels = load_image_from_memory(
                     reinterpret_cast<const char*>(bytes.data()),
                     static_cast<int>(bytes.size()),
@@ -670,22 +694,31 @@ int main(int argc, const char** argv) {
                 }
 
                 sd_image_t img{(uint32_t)img_w, (uint32_t)img_h, 3, raw_pixels};
+                gen_params.set_width_and_height_if_unset(img.width, img.height);
                 ref_images.push_back(img);
             }
 
             sd_image_t mask_image = {0};
             if (!mask_bytes.empty()) {
-                int mask_w        = width;
-                int mask_h        = height;
+                int expected_width  = 0;
+                int expected_height = 0;
+                if (gen_params.width_and_height_are_set()) {
+                    expected_width  = gen_params.width;
+                    expected_height = gen_params.height;
+                }
+                int mask_w;
+                int mask_h;
+
                 uint8_t* mask_raw = load_image_from_memory(
                     reinterpret_cast<const char*>(mask_bytes.data()),
                     static_cast<int>(mask_bytes.size()),
                     mask_w, mask_h,
-                    width, height, 1);
+                    expected_width, expected_height, 1);
                 mask_image = {(uint32_t)mask_w, (uint32_t)mask_h, 1, mask_raw};
+                gen_params.set_width_and_height_if_unset(mask_image.width, mask_image.height);
             } else {
-                mask_image.width   = width;
-                mask_image.height  = height;
+                mask_image.width   = get_resolved_width();
+                mask_image.height  = get_resolved_height();
                 mask_image.channel = 1;
                 mask_image.data    = nullptr;
             }
@@ -702,8 +735,8 @@ int main(int argc, const char** argv) {
                 gen_params.auto_resize_ref_image,
                 gen_params.increase_ref_index,
                 mask_image,
-                gen_params.width,
-                gen_params.height,
+                get_resolved_width(),
+                get_resolved_height(),
                 gen_params.sample_params,
                 gen_params.strength,
                 gen_params.seed,
@@ -786,8 +819,8 @@ int main(int argc, const char** argv) {
             std::string negative_prompt = j.value("negative_prompt", "");
             int width                   = j.value("width", 512);
             int height                  = j.value("height", 512);
-            int steps                   = j.value("steps", -1);
-            float cfg_scale             = j.value("cfg_scale", 7.f);
+            int steps                   = j.value("steps", default_gen_params.sample_params.sample_steps);
+            float cfg_scale             = j.value("cfg_scale", default_gen_params.sample_params.guidance.txt_cfg);
             int64_t seed                = j.value("seed", -1);
             int batch_size              = j.value("batch_size", 1);
             int clip_skip               = j.value("clip_skip", -1);
@@ -837,11 +870,12 @@ int main(int argc, const char** argv) {
                         return bad("lora.path required");
                     }
 
-                    if (!is_valid_lora_path(path)) {
+                    std::string fullpath = get_lora_full_path(path);
+                    if (fullpath.empty()) {
                         return bad("invalid lora path: " + path);
                     }
 
-                    lora_path_storage.push_back(path);
+                    lora_path_storage.push_back(fullpath);
                     sd_lora_t l;
                     l.is_high_noise = is_high_noise;
                     l.multiplier    = multiplier;
@@ -883,16 +917,13 @@ int main(int argc, const char** argv) {
 
             enum scheduler_t scheduler = str_to_scheduler(scheduler_name.c_str());
 
-            // avoid excessive resource usage
-
-            SDGenerationParams gen_params         = default_gen_params;
-            gen_params.prompt                     = prompt;
-            gen_params.negative_prompt            = negative_prompt;
-            gen_params.width                      = width;
-            gen_params.height                     = height;
-            gen_params.seed                       = seed;
-            gen_params.sample_params.sample_steps = steps;
-            gen_params.batch_count                = batch_size;
+            SDGenerationParams gen_params             = default_gen_params;
+            gen_params.prompt                         = prompt;
+            gen_params.negative_prompt                = negative_prompt;
+            gen_params.seed                           = seed;
+            gen_params.sample_params.sample_steps     = steps;
+            gen_params.batch_count                    = batch_size;
+            gen_params.sample_params.guidance.txt_cfg = cfg_scale;
 
             if (clip_skip > 0) {
                 gen_params.clip_skip = clip_skip;
@@ -906,38 +937,66 @@ int main(int argc, const char** argv) {
                 gen_params.sample_params.scheduler = scheduler;
             }
 
+            // re-read to avoid applying 512 as default before the provided
+            // images and/or server command-line
+            gen_params.width  = j.value("width", -1);
+            gen_params.height = j.value("height", -1);
+
             LOG_DEBUG("%s\n", gen_params.to_string().c_str());
 
-            sd_image_t init_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
-            sd_image_t control_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
-            sd_image_t mask_image    = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 1, nullptr};
+            sd_image_t init_image    = {0, 0, 3, nullptr};
+            sd_image_t control_image = {0, 0, 3, nullptr};
+            sd_image_t mask_image    = {0, 0, 1, nullptr};
             std::vector<uint8_t> mask_data;
             std::vector<sd_image_t> pmid_images;
             std::vector<sd_image_t> ref_images;
 
-            if (img2img) {
-                auto decode_image = [](sd_image_t& image, std::string encoded) -> bool {
-                    // remove data URI prefix if present ("data:image/png;base64,")
-                    auto comma_pos = encoded.find(',');
-                    if (comma_pos != std::string::npos) {
-                        encoded = encoded.substr(comma_pos + 1);
-                    }
-                    std::vector<uint8_t> img_data = base64_decode(encoded);
-                    if (!img_data.empty()) {
-                        int img_w         = image.width;
-                        int img_h         = image.height;
-                        uint8_t* raw_data = load_image_from_memory(
-                            (const char*)img_data.data(), (int)img_data.size(),
-                            img_w, img_h,
-                            image.width, image.height, image.channel);
-                        if (raw_data) {
-                            image = {(uint32_t)img_w, (uint32_t)img_h, image.channel, raw_data};
-                            return true;
-                        }
-                    }
-                    return false;
-                };
+            auto get_resolved_width = [&gen_params, &default_gen_params]() -> int {
+                if (gen_params.width > 0)
+                    return gen_params.width;
+                if (default_gen_params.width > 0)
+                    return default_gen_params.width;
+                return 512;
+            };
+            auto get_resolved_height = [&gen_params, &default_gen_params]() -> int {
+                if (gen_params.height > 0)
+                    return gen_params.height;
+                if (default_gen_params.height > 0)
+                    return default_gen_params.height;
+                return 512;
+            };
 
+            auto decode_image = [&gen_params](sd_image_t& image, std::string encoded) -> bool {
+                // remove data URI prefix if present ("data:image/png;base64,")
+                auto comma_pos = encoded.find(',');
+                if (comma_pos != std::string::npos) {
+                    encoded = encoded.substr(comma_pos + 1);
+                }
+                std::vector<uint8_t> img_data = base64_decode(encoded);
+                if (!img_data.empty()) {
+                    int expected_width  = 0;
+                    int expected_height = 0;
+                    if (gen_params.width_and_height_are_set()) {
+                        expected_width  = gen_params.width;
+                        expected_height = gen_params.height;
+                    }
+                    int img_w;
+                    int img_h;
+
+                    uint8_t* raw_data = load_image_from_memory(
+                        (const char*)img_data.data(), (int)img_data.size(),
+                        img_w, img_h,
+                        expected_width, expected_height, image.channel);
+                    if (raw_data) {
+                        image = {(uint32_t)img_w, (uint32_t)img_h, image.channel, raw_data};
+                        gen_params.set_width_and_height_if_unset(image.width, image.height);
+                        return true;
+                    }
+                }
+                return false;
+            };
+
+            if (img2img) {
                 if (j.contains("init_images") && j["init_images"].is_array() && !j["init_images"].empty()) {
                     std::string encoded = j["init_images"][0].get<std::string>();
                     decode_image(init_image, encoded);
@@ -953,23 +1012,15 @@ int main(int argc, const char** argv) {
                         }
                     }
                 } else {
-                    mask_data          = std::vector<uint8_t>(width * height, 255);
-                    mask_image.width   = width;
-                    mask_image.height  = height;
+                    int m_width        = get_resolved_width();
+                    int m_height       = get_resolved_height();
+                    mask_data          = std::vector<uint8_t>(m_width * m_height, 255);
+                    mask_image.width   = m_width;
+                    mask_image.height  = m_height;
                     mask_image.channel = 1;
                     mask_image.data    = mask_data.data();
                 }
 
-                if (j.contains("extra_images") && j["extra_images"].is_array()) {
-                    for (auto extra_image : j["extra_images"]) {
-                        std::string encoded  = extra_image.get<std::string>();
-                        sd_image_t tmp_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
-                        if (decode_image(tmp_image, encoded)) {
-                            ref_images.push_back(tmp_image);
-                        }
-                    }
-                }
-
                 float denoising_strength = j.value("denoising_strength", -1.f);
                 if (denoising_strength >= 0.f) {
                     denoising_strength  = std::min(denoising_strength, 1.0f);
@@ -977,6 +1028,16 @@ int main(int argc, const char** argv) {
                 }
             }
 
+            if (j.contains("extra_images") && j["extra_images"].is_array()) {
+                for (auto extra_image : j["extra_images"]) {
+                    std::string encoded  = extra_image.get<std::string>();
+                    sd_image_t tmp_image = {(uint32_t)gen_params.width, (uint32_t)gen_params.height, 3, nullptr};
+                    if (decode_image(tmp_image, encoded)) {
+                        ref_images.push_back(tmp_image);
+                    }
+                }
+            }
+
             sd_img_gen_params_t img_gen_params = {
                 sd_loras.data(),
                 static_cast<uint32_t>(sd_loras.size()),
@@ -989,8 +1050,8 @@ int main(int argc, const char** argv) {
                 gen_params.auto_resize_ref_image,
                 gen_params.increase_ref_index,
                 mask_image,
-                gen_params.width,
-                gen_params.height,
+                get_resolved_width(),
+                get_resolved_height(),
                 gen_params.sample_params,
                 gen_params.strength,
                 gen_params.seed,

format-code.sh

@@ -1,4 +1,4 @@
-for f in *.cpp *.h *.hpp examples/cli/*.cpp examples/common/*.hpp examples/cli/*.h examples/server/*.cpp; do
+for f in src/*.cpp src/*.h src/*.hpp src/vocab/*.h src/vocab/*.cpp examples/cli/*.cpp examples/common/*.hpp examples/cli/*.h examples/server/*.cpp; do
   [[ "$f" == vocab* ]] && continue
   echo "formatting '$f'"
   # if [ "$f" != "stable-diffusion.h" ]; then

include/stable-diffusion.h

@@ -201,7 +201,6 @@ typedef struct {
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
     bool qwen_image_zero_cond_t;
-    float flow_shift;
 } sd_ctx_params_t;
 
 typedef struct {
@@ -235,6 +234,7 @@ typedef struct {
     int shifted_timestep;
     float* custom_sigmas;
     int custom_sigmas_count;
+    float flow_shift;
 } sd_sample_params_t;
 
 typedef struct {

script/face_detect.py

@@ -1,88 +1,88 @@
-import os
-import sys
-
-import numpy as np
-import torch
-from diffusers.utils import load_image
-# pip install insightface==0.7.3
-from insightface.app import FaceAnalysis
-from insightface.data import get_image as ins_get_image
-from safetensors.torch import save_file
-
-### 
-# https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/165#issue-2055829543
-###
-class FaceAnalysis2(FaceAnalysis):
-    # NOTE: allows setting det_size for each detection call.
-    # the model allows it but the wrapping code from insightface
-    # doesn't show it, and people end up loading duplicate models
-    # for different sizes where there is absolutely no need to
-    def get(self, img, max_num=0, det_size=(640, 640)):
-        if det_size is not None:
-            self.det_model.input_size = det_size
-
-        return super().get(img, max_num)
-
-def analyze_faces(face_analysis: FaceAnalysis, img_data: np.ndarray, det_size=(640, 640)):
-    # NOTE: try detect faces, if no faces detected, lower det_size until it does
-    detection_sizes = [None] + [(size, size) for size in range(640, 256, -64)] + [(256, 256)]
-
-    for size in detection_sizes:
-        faces = face_analysis.get(img_data, det_size=size)
-        if len(faces) > 0:
-            return faces
-
-    return []
-
-if __name__ == "__main__":
-    #face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
-    face_detector = FaceAnalysis2(providers=['CPUExecutionProvider'], allowed_modules=['detection', 'recognition'])
-    face_detector.prepare(ctx_id=0, det_size=(640, 640))
-    #input_folder_name = './scarletthead_woman'
-    input_folder_name = sys.argv[1]
-    image_basename_list = os.listdir(input_folder_name)
-    image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
-
-    input_id_images = []
-    for image_path in image_path_list:
-        input_id_images.append(load_image(image_path))
-    
-    id_embed_list = []
-    
-    for img in input_id_images:
-        img = np.array(img)
-        img = img[:, :, ::-1]
-        faces = analyze_faces(face_detector, img)
-        if len(faces) > 0:
-            id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
-    
-    if len(id_embed_list) == 0:
-        raise ValueError(f"No face detected in input image pool")
-    
-    id_embeds = torch.stack(id_embed_list)    
-    
-    # for r in id_embeds:
-    #     print(r)
-    # #torch.save(id_embeds, input_folder_name+'/id_embeds.pt');
-    # weights = dict()
-    # weights["id_embeds"] = id_embeds
-    # save_file(weights, input_folder_name+'/id_embeds.safetensors')
-
-    binary_data = id_embeds.numpy().tobytes()
-    two = 4
-    zero = 0
-    one = 1
-    tensor_name = "id_embeds"
-# Write binary data to a file
-    with open(input_folder_name+'/id_embeds.bin', "wb") as f:
-        f.write(two.to_bytes(4, byteorder='little'))
-        f.write((len(tensor_name)).to_bytes(4, byteorder='little'))
-        f.write(zero.to_bytes(4, byteorder='little'))
-        f.write((id_embeds.shape[1]).to_bytes(4, byteorder='little'))
-        f.write((id_embeds.shape[0]).to_bytes(4, byteorder='little'))
-        f.write(one.to_bytes(4, byteorder='little'))
-        f.write(one.to_bytes(4, byteorder='little'))
-        f.write(tensor_name.encode('ascii'))
-        f.write(binary_data)
-
+import os
+import sys
+
+import numpy as np
+import torch
+from diffusers.utils import load_image
+# pip install insightface==0.7.3
+from insightface.app import FaceAnalysis
+from insightface.data import get_image as ins_get_image
+from safetensors.torch import save_file
+
+### 
+# https://github.com/cubiq/ComfyUI_IPAdapter_plus/issues/165#issue-2055829543
+###
+class FaceAnalysis2(FaceAnalysis):
+    # NOTE: allows setting det_size for each detection call.
+    # the model allows it but the wrapping code from insightface
+    # doesn't show it, and people end up loading duplicate models
+    # for different sizes where there is absolutely no need to
+    def get(self, img, max_num=0, det_size=(640, 640)):
+        if det_size is not None:
+            self.det_model.input_size = det_size
+
+        return super().get(img, max_num)
+
+def analyze_faces(face_analysis: FaceAnalysis, img_data: np.ndarray, det_size=(640, 640)):
+    # NOTE: try detect faces, if no faces detected, lower det_size until it does
+    detection_sizes = [None] + [(size, size) for size in range(640, 256, -64)] + [(256, 256)]
+
+    for size in detection_sizes:
+        faces = face_analysis.get(img_data, det_size=size)
+        if len(faces) > 0:
+            return faces
+
+    return []
+
+if __name__ == "__main__":
+    #face_detector = FaceAnalysis2(providers=['CUDAExecutionProvider'], allowed_modules=['detection', 'recognition'])
+    face_detector = FaceAnalysis2(providers=['CPUExecutionProvider'], allowed_modules=['detection', 'recognition'])
+    face_detector.prepare(ctx_id=0, det_size=(640, 640))
+    #input_folder_name = './scarletthead_woman'
+    input_folder_name = sys.argv[1]
+    image_basename_list = os.listdir(input_folder_name)
+    image_path_list = sorted([os.path.join(input_folder_name, basename) for basename in image_basename_list])
+
+    input_id_images = []
+    for image_path in image_path_list:
+        input_id_images.append(load_image(image_path))
+    
+    id_embed_list = []
+    
+    for img in input_id_images:
+        img = np.array(img)
+        img = img[:, :, ::-1]
+        faces = analyze_faces(face_detector, img)
+        if len(faces) > 0:
+            id_embed_list.append(torch.from_numpy((faces[0]['embedding'])))
+    
+    if len(id_embed_list) == 0:
+        raise ValueError(f"No face detected in input image pool")
+    
+    id_embeds = torch.stack(id_embed_list)    
+    
+    # for r in id_embeds:
+    #     print(r)
+    # #torch.save(id_embeds, input_folder_name+'/id_embeds.pt');
+    # weights = dict()
+    # weights["id_embeds"] = id_embeds
+    # save_file(weights, input_folder_name+'/id_embeds.safetensors')
+
+    binary_data = id_embeds.numpy().tobytes()
+    two = 4
+    zero = 0
+    one = 1
+    tensor_name = "id_embeds"
+# Write binary data to a file
+    with open(input_folder_name+'/id_embeds.bin', "wb") as f:
+        f.write(two.to_bytes(4, byteorder='little'))
+        f.write((len(tensor_name)).to_bytes(4, byteorder='little'))
+        f.write(zero.to_bytes(4, byteorder='little'))
+        f.write((id_embeds.shape[1]).to_bytes(4, byteorder='little'))
+        f.write((id_embeds.shape[0]).to_bytes(4, byteorder='little'))
+        f.write(one.to_bytes(4, byteorder='little'))
+        f.write(one.to_bytes(4, byteorder='little'))
+        f.write(tensor_name.encode('ascii'))
+        f.write(binary_data)
+
     

src/anima.hpp

@@ -0,0 +1,686 @@
+#ifndef __ANIMA_HPP__
+#define __ANIMA_HPP__
+
+#include <cmath>
+#include <memory>
+#include <utility>
+#include <vector>
+
+#include "common_block.hpp"
+#include "flux.hpp"
+#include "rope.hpp"
+
+namespace Anima {
+    constexpr int ANIMA_GRAPH_SIZE = 65536;
+
+    __STATIC_INLINE__ struct ggml_tensor* apply_gate(struct ggml_context* ctx,
+                                                     struct ggml_tensor* x,
+                                                     struct ggml_tensor* gate) {
+        gate = ggml_reshape_3d(ctx, gate, gate->ne[0], 1, gate->ne[1]);  // [N, 1, C]
+        return ggml_mul(ctx, x, gate);
+    }
+
+    struct XEmbedder : public GGMLBlock {
+    public:
+        XEmbedder(int64_t in_dim, int64_t out_dim) {
+            blocks["proj.1"] = std::make_shared<Linear>(in_dim, out_dim, false);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+            auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj.1"]);
+            return proj->forward(ctx, x);
+        }
+    };
+
+    struct TimestepEmbedder : public GGMLBlock {
+    public:
+        TimestepEmbedder(int64_t in_dim, int64_t out_dim) {
+            blocks["1.linear_1"] = std::make_shared<Linear>(in_dim, in_dim, false);
+            blocks["1.linear_2"] = std::make_shared<Linear>(in_dim, out_dim, false);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1.linear_1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["1.linear_2"]);
+
+            x = linear_1->forward(ctx, x);
+            x = ggml_silu_inplace(ctx->ggml_ctx, x);
+            x = linear_2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct AdaLayerNormZero : public GGMLBlock {
+    protected:
+        int64_t in_features;
+
+    public:
+        AdaLayerNormZero(int64_t in_features, int64_t hidden_features = 256)
+            : in_features(in_features) {
+            blocks["norm"] = std::make_shared<LayerNorm>(in_features, 1e-6f, false, false);
+            blocks["1"]    = std::make_shared<Linear>(in_features, hidden_features, false);
+            blocks["2"]    = std::make_shared<Linear>(hidden_features, 3 * in_features, false);
+        }
+
+        std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                                    struct ggml_tensor* hidden_states,
+                                                                    struct ggml_tensor* embedded_timestep,
+                                                                    struct ggml_tensor* temb = nullptr) {
+            auto norm     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
+
+            auto emb = ggml_silu(ctx->ggml_ctx, embedded_timestep);
+            emb      = linear_1->forward(ctx, emb);
+            emb      = linear_2->forward(ctx, emb);  // [N, 3*C]
+
+            if (temb != nullptr) {
+                emb = ggml_add(ctx->ggml_ctx, emb, temb);
+            }
+
+            auto emb_chunks = ggml_ext_chunk(ctx->ggml_ctx, emb, 3, 0);
+            auto shift      = emb_chunks[0];
+            auto scale      = emb_chunks[1];
+            auto gate       = emb_chunks[2];
+
+            auto x = norm->forward(ctx, hidden_states);
+            x      = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+
+            return {x, gate};
+        }
+    };
+
+    struct AdaLayerNorm : public GGMLBlock {
+    protected:
+        int64_t embedding_dim;
+
+    public:
+        AdaLayerNorm(int64_t in_features, int64_t hidden_features = 256)
+            : embedding_dim(in_features) {
+            blocks["norm"] = std::make_shared<LayerNorm>(in_features, 1e-6f, false, false);
+            blocks["1"]    = std::make_shared<Linear>(in_features, hidden_features, false);
+            blocks["2"]    = std::make_shared<Linear>(hidden_features, 2 * in_features, false);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                    struct ggml_tensor* hidden_states,
+                                    struct ggml_tensor* embedded_timestep,
+                                    struct ggml_tensor* temb = nullptr) {
+            auto norm     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
+
+            auto emb = ggml_silu(ctx->ggml_ctx, embedded_timestep);
+            emb      = linear_1->forward(ctx, emb);
+            emb      = linear_2->forward(ctx, emb);  // [N, 2*C]
+
+            if (temb != nullptr) {
+                auto temb_2c = ggml_view_2d(ctx->ggml_ctx, temb, 2 * embedding_dim, temb->ne[1], temb->nb[1], 0);
+                emb          = ggml_add(ctx->ggml_ctx, emb, temb_2c);
+            }
+
+            auto emb_chunks = ggml_ext_chunk(ctx->ggml_ctx, emb, 2, 0);
+            auto shift      = emb_chunks[0];
+            auto scale      = emb_chunks[1];
+
+            auto x = norm->forward(ctx, hidden_states);
+            x      = Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+            return x;
+        }
+    };
+
+    struct AnimaAttention : public GGMLBlock {
+    protected:
+        int64_t num_heads;
+        int64_t head_dim;
+        std::string out_proj_name;
+
+    public:
+        AnimaAttention(int64_t query_dim,
+                       int64_t context_dim,
+                       int64_t num_heads,
+                       int64_t head_dim,
+                       const std::string& out_proj_name = "output_proj")
+            : num_heads(num_heads), head_dim(head_dim), out_proj_name(out_proj_name) {
+            int64_t inner_dim = num_heads * head_dim;
+
+            blocks["q_proj"]            = std::make_shared<Linear>(query_dim, inner_dim, false);
+            blocks["k_proj"]            = std::make_shared<Linear>(context_dim, inner_dim, false);
+            blocks["v_proj"]            = std::make_shared<Linear>(context_dim, inner_dim, false);
+            blocks["q_norm"]            = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm"]            = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks[this->out_proj_name] = std::make_shared<Linear>(inner_dim, query_dim, false);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                    struct ggml_tensor* hidden_states,
+                                    struct ggml_tensor* encoder_hidden_states = nullptr,
+                                    struct ggml_tensor* pe_q                  = nullptr,
+                                    struct ggml_tensor* pe_k                  = nullptr) {
+            if (encoder_hidden_states == nullptr) {
+                encoder_hidden_states = hidden_states;
+            }
+
+            auto q_proj   = std::dynamic_pointer_cast<Linear>(blocks["q_proj"]);
+            auto k_proj   = std::dynamic_pointer_cast<Linear>(blocks["k_proj"]);
+            auto v_proj   = std::dynamic_pointer_cast<Linear>(blocks["v_proj"]);
+            auto q_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+            auto out_proj = std::dynamic_pointer_cast<Linear>(blocks[out_proj_name]);
+
+            auto q = q_proj->forward(ctx, hidden_states);
+            auto k = k_proj->forward(ctx, encoder_hidden_states);
+            auto v = v_proj->forward(ctx, encoder_hidden_states);
+
+            int64_t N   = q->ne[2];
+            int64_t L_q = q->ne[1];
+            int64_t L_k = k->ne[1];
+
+            auto q4 = ggml_reshape_4d(ctx->ggml_ctx, q, head_dim, num_heads, L_q, N);  // [N, L_q, H, D]
+            auto k4 = ggml_reshape_4d(ctx->ggml_ctx, k, head_dim, num_heads, L_k, N);  // [N, L_k, H, D]
+            auto v4 = ggml_reshape_4d(ctx->ggml_ctx, v, head_dim, num_heads, L_k, N);  // [N, L_k, H, D]
+
+            q4 = q_norm->forward(ctx, q4);
+            k4 = k_norm->forward(ctx, k4);
+
+            struct ggml_tensor* attn_out = nullptr;
+            if (pe_q != nullptr || pe_k != nullptr) {
+                if (pe_q == nullptr) {
+                    pe_q = pe_k;
+                }
+                if (pe_k == nullptr) {
+                    pe_k = pe_q;
+                }
+                auto q_rope = Rope::apply_rope(ctx->ggml_ctx, q4, pe_q, false);
+                auto k_rope = Rope::apply_rope(ctx->ggml_ctx, k4, pe_k, false);
+                attn_out    = ggml_ext_attention_ext(ctx->ggml_ctx,
+                                                     ctx->backend,
+                                                     q_rope,
+                                                     k_rope,
+                                                     v4,
+                                                     num_heads,
+                                                     nullptr,
+                                                     true,
+                                                     ctx->flash_attn_enabled);
+            } else {
+                auto q_flat = ggml_reshape_3d(ctx->ggml_ctx, q4, head_dim * num_heads, L_q, N);
+                auto k_flat = ggml_reshape_3d(ctx->ggml_ctx, k4, head_dim * num_heads, L_k, N);
+                attn_out    = ggml_ext_attention_ext(ctx->ggml_ctx,
+                                                     ctx->backend,
+                                                     q_flat,
+                                                     k_flat,
+                                                     v,
+                                                     num_heads,
+                                                     nullptr,
+                                                     false,
+                                                     ctx->flash_attn_enabled);
+            }
+
+            return out_proj->forward(ctx, attn_out);
+        }
+    };
+
+    struct AnimaMLP : public GGMLBlock {
+    public:
+        AnimaMLP(int64_t dim, int64_t hidden_dim) {
+            blocks["layer1"] = std::make_shared<Linear>(dim, hidden_dim, false);
+            blocks["layer2"] = std::make_shared<Linear>(hidden_dim, dim, false);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+            auto layer1 = std::dynamic_pointer_cast<Linear>(blocks["layer1"]);
+            auto layer2 = std::dynamic_pointer_cast<Linear>(blocks["layer2"]);
+
+            x = layer1->forward(ctx, x);
+            x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
+            x = layer2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct AdapterMLP : public GGMLBlock {
+    public:
+        AdapterMLP(int64_t dim, int64_t hidden_dim) {
+            blocks["0"] = std::make_shared<Linear>(dim, hidden_dim, true);
+            blocks["2"] = std::make_shared<Linear>(hidden_dim, dim, true);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+            auto layer0 = std::dynamic_pointer_cast<Linear>(blocks["0"]);
+            auto layer2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
+
+            x = layer0->forward(ctx, x);
+            x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
+            x = layer2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct LLMAdapterBlock : public GGMLBlock {
+    public:
+        LLMAdapterBlock(int64_t model_dim = 1024, int64_t source_dim = 1024, int64_t num_heads = 16, int64_t head_dim = 64) {
+            blocks["norm_self_attn"]  = std::make_shared<RMSNorm>(model_dim, 1e-6f);
+            blocks["self_attn"]       = std::make_shared<AnimaAttention>(model_dim, model_dim, num_heads, head_dim, "o_proj");
+            blocks["norm_cross_attn"] = std::make_shared<RMSNorm>(model_dim, 1e-6f);
+            blocks["cross_attn"]      = std::make_shared<AnimaAttention>(model_dim, source_dim, num_heads, head_dim, "o_proj");
+            blocks["norm_mlp"]        = std::make_shared<RMSNorm>(model_dim, 1e-6f);
+            blocks["mlp"]             = std::make_shared<AdapterMLP>(model_dim, model_dim * 4);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* context,
+                                    struct ggml_tensor* target_pe,
+                                    struct ggml_tensor* context_pe) {
+            auto norm_self_attn  = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_self_attn"]);
+            auto self_attn       = std::dynamic_pointer_cast<AnimaAttention>(blocks["self_attn"]);
+            auto norm_cross_attn = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_cross_attn"]);
+            auto cross_attn      = std::dynamic_pointer_cast<AnimaAttention>(blocks["cross_attn"]);
+            auto norm_mlp        = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_mlp"]);
+            auto mlp             = std::dynamic_pointer_cast<AdapterMLP>(blocks["mlp"]);
+
+            auto h = norm_self_attn->forward(ctx, x);
+            h      = self_attn->forward(ctx, h, nullptr, target_pe, target_pe);
+            x      = ggml_add(ctx->ggml_ctx, x, h);
+
+            h = norm_cross_attn->forward(ctx, x);
+            h = cross_attn->forward(ctx, h, context, target_pe, context_pe);
+            x = ggml_add(ctx->ggml_ctx, x, h);
+
+            h = norm_mlp->forward(ctx, x);
+            h = mlp->forward(ctx, h);
+            x = ggml_add(ctx->ggml_ctx, x, h);
+
+            return x;
+        }
+    };
+
+    struct LLMAdapter : public GGMLBlock {
+    protected:
+        int num_layers;
+
+    public:
+        LLMAdapter(int64_t source_dim = 1024,
+                   int64_t target_dim = 1024,
+                   int64_t model_dim  = 1024,
+                   int num_layers     = 6,
+                   int num_heads      = 16)
+            : num_layers(num_layers) {
+            int64_t head_dim = model_dim / num_heads;
+
+            blocks["embed"] = std::make_shared<Embedding>(32128, target_dim);
+            for (int i = 0; i < num_layers; i++) {
+                blocks["blocks." + std::to_string(i)] =
+                    std::make_shared<LLMAdapterBlock>(model_dim, source_dim, num_heads, head_dim);
+            }
+            blocks["out_proj"] = std::make_shared<Linear>(model_dim, target_dim, true);
+            blocks["norm"]     = std::make_shared<RMSNorm>(target_dim, 1e-6f);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                    struct ggml_tensor* source_hidden_states,
+                                    struct ggml_tensor* target_input_ids,
+                                    struct ggml_tensor* target_pe,
+                                    struct ggml_tensor* source_pe) {
+            GGML_ASSERT(target_input_ids != nullptr);
+            if (ggml_n_dims(target_input_ids) == 1) {
+                target_input_ids = ggml_reshape_2d(ctx->ggml_ctx, target_input_ids, target_input_ids->ne[0], 1);
+            }
+
+            auto embed    = std::dynamic_pointer_cast<Embedding>(blocks["embed"]);
+            auto out_proj = std::dynamic_pointer_cast<Linear>(blocks["out_proj"]);
+            auto norm     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+
+            auto x = embed->forward(ctx, target_input_ids);  // [N, target_len, target_dim]
+
+            for (int i = 0; i < num_layers; i++) {
+                auto block = std::dynamic_pointer_cast<LLMAdapterBlock>(blocks["blocks." + std::to_string(i)]);
+                x          = block->forward(ctx, x, source_hidden_states, target_pe, source_pe);
+            }
+
+            x = out_proj->forward(ctx, x);
+            x = norm->forward(ctx, x);
+            return x;
+        }
+    };
+
+    struct TransformerBlock : public GGMLBlock {
+    public:
+        TransformerBlock(int64_t hidden_size,
+                         int64_t text_embed_dim,
+                         int64_t num_heads,
+                         int64_t head_dim,
+                         int64_t mlp_ratio      = 4,
+                         int64_t adaln_lora_dim = 256) {
+            blocks["adaln_modulation_self_attn"]  = std::make_shared<AdaLayerNormZero>(hidden_size, adaln_lora_dim);
+            blocks["self_attn"]                   = std::make_shared<AnimaAttention>(hidden_size, hidden_size, num_heads, head_dim);
+            blocks["adaln_modulation_cross_attn"] = std::make_shared<AdaLayerNormZero>(hidden_size, adaln_lora_dim);
+            blocks["cross_attn"]                  = std::make_shared<AnimaAttention>(hidden_size, text_embed_dim, num_heads, head_dim);
+            blocks["adaln_modulation_mlp"]        = std::make_shared<AdaLayerNormZero>(hidden_size, adaln_lora_dim);
+            blocks["mlp"]                         = std::make_shared<AnimaMLP>(hidden_size, hidden_size * mlp_ratio);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                    struct ggml_tensor* hidden_states,
+                                    struct ggml_tensor* encoder_hidden_states,
+                                    struct ggml_tensor* embedded_timestep,
+                                    struct ggml_tensor* temb,
+                                    struct ggml_tensor* image_pe) {
+            auto norm1 = std::dynamic_pointer_cast<AdaLayerNormZero>(blocks["adaln_modulation_self_attn"]);
+            auto attn1 = std::dynamic_pointer_cast<AnimaAttention>(blocks["self_attn"]);
+            auto norm2 = std::dynamic_pointer_cast<AdaLayerNormZero>(blocks["adaln_modulation_cross_attn"]);
+            auto attn2 = std::dynamic_pointer_cast<AnimaAttention>(blocks["cross_attn"]);
+            auto norm3 = std::dynamic_pointer_cast<AdaLayerNormZero>(blocks["adaln_modulation_mlp"]);
+            auto mlp   = std::dynamic_pointer_cast<AnimaMLP>(blocks["mlp"]);
+
+            auto [normed1, gate1] = norm1->forward(ctx, hidden_states, embedded_timestep, temb);
+            auto h                = attn1->forward(ctx, normed1, nullptr, image_pe, image_pe);
+            hidden_states         = ggml_add(ctx->ggml_ctx, hidden_states, apply_gate(ctx->ggml_ctx, h, gate1));
+
+            auto [normed2, gate2] = norm2->forward(ctx, hidden_states, embedded_timestep, temb);
+            h                     = attn2->forward(ctx, normed2, encoder_hidden_states, nullptr, nullptr);
+            hidden_states         = ggml_add(ctx->ggml_ctx, hidden_states, apply_gate(ctx->ggml_ctx, h, gate2));
+
+            auto [normed3, gate3] = norm3->forward(ctx, hidden_states, embedded_timestep, temb);
+            h                     = mlp->forward(ctx, normed3);
+            hidden_states         = ggml_add(ctx->ggml_ctx, hidden_states, apply_gate(ctx->ggml_ctx, h, gate3));
+
+            return hidden_states;
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+    protected:
+        int64_t hidden_size;
+        int64_t patch_size;
+        int64_t out_channels;
+
+    public:
+        FinalLayer(int64_t hidden_size, int64_t patch_size, int64_t out_channels)
+            : hidden_size(hidden_size), patch_size(patch_size), out_channels(out_channels) {
+            blocks["adaln_modulation"] = std::make_shared<AdaLayerNorm>(hidden_size, 256);
+            blocks["linear"]           = std::make_shared<Linear>(hidden_size, patch_size * patch_size * out_channels, false);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                    struct ggml_tensor* hidden_states,
+                                    struct ggml_tensor* embedded_timestep,
+                                    struct ggml_tensor* temb) {
+            auto adaln  = std::dynamic_pointer_cast<AdaLayerNorm>(blocks["adaln_modulation"]);
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+
+            hidden_states = adaln->forward(ctx, hidden_states, embedded_timestep, temb);
+            hidden_states = linear->forward(ctx, hidden_states);
+            return hidden_states;
+        }
+    };
+
+    struct AnimaNet : public GGMLBlock {
+    public:
+        int64_t in_channels       = 16;
+        int64_t out_channels      = 16;
+        int64_t hidden_size       = 2048;
+        int64_t text_embed_dim    = 1024;
+        int64_t num_heads         = 16;
+        int64_t head_dim          = 128;
+        int patch_size            = 2;
+        int64_t num_layers        = 28;
+        std::vector<int> axes_dim = {44, 42, 42};
+        int theta                 = 10000;
+
+    public:
+        AnimaNet() = default;
+        explicit AnimaNet(int64_t num_layers)
+            : num_layers(num_layers) {
+            blocks["x_embedder"]       = std::make_shared<XEmbedder>((in_channels + 1) * patch_size * patch_size, hidden_size);
+            blocks["t_embedder"]       = std::make_shared<TimestepEmbedder>(hidden_size, hidden_size * 3);
+            blocks["t_embedding_norm"] = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            for (int i = 0; i < num_layers; i++) {
+                blocks["blocks." + std::to_string(i)] = std::make_shared<TransformerBlock>(hidden_size,
+                                                                                           text_embed_dim,
+                                                                                           num_heads,
+                                                                                           head_dim);
+            }
+            blocks["final_layer"] = std::make_shared<FinalLayer>(hidden_size, patch_size, out_channels);
+            blocks["llm_adapter"] = std::make_shared<LLMAdapter>(1024, 1024, 1024, 6, 16);
+        }
+
+        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* timestep,
+                                    struct ggml_tensor* encoder_hidden_states,
+                                    struct ggml_tensor* image_pe,
+                                    struct ggml_tensor* t5_ids       = nullptr,
+                                    struct ggml_tensor* t5_weights   = nullptr,
+                                    struct ggml_tensor* adapter_q_pe = nullptr,
+                                    struct ggml_tensor* adapter_k_pe = nullptr) {
+            GGML_ASSERT(x->ne[3] == 1);
+
+            auto x_embedder       = std::dynamic_pointer_cast<XEmbedder>(blocks["x_embedder"]);
+            auto t_embedder       = std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder"]);
+            auto t_embedding_norm = std::dynamic_pointer_cast<RMSNorm>(blocks["t_embedding_norm"]);
+            auto final_layer      = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
+            auto llm_adapter      = std::dynamic_pointer_cast<LLMAdapter>(blocks["llm_adapter"]);
+
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+
+            auto padding_mask = ggml_ext_zeros(ctx->ggml_ctx, x->ne[0], x->ne[1], 1, x->ne[3]);
+            x                 = ggml_concat(ctx->ggml_ctx, x, padding_mask, 2);  // [N, C + 1, H, W]
+
+            x = DiT::pad_and_patchify(ctx, x, patch_size, patch_size);  // [N, h*w, (C+1)*ph*pw]
+
+            x = x_embedder->forward(ctx, x);
+
+            auto timestep_proj     = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep, static_cast<int>(hidden_size));
+            auto temb              = t_embedder->forward(ctx, timestep_proj);
+            auto embedded_timestep = t_embedding_norm->forward(ctx, timestep_proj);
+
+            if (t5_ids != nullptr) {
+                auto adapted_context = llm_adapter->forward(ctx, encoder_hidden_states, t5_ids, adapter_q_pe, adapter_k_pe);
+                if (t5_weights != nullptr) {
+                    auto w = t5_weights;
+                    if (ggml_n_dims(w) == 1) {
+                        w = ggml_reshape_3d(ctx->ggml_ctx, w, 1, w->ne[0], 1);
+                    }
+                    w               = ggml_repeat_4d(ctx->ggml_ctx, w, adapted_context->ne[0], adapted_context->ne[1], adapted_context->ne[2], 1);
+                    adapted_context = ggml_mul(ctx->ggml_ctx, adapted_context, w);
+                }
+                if (adapted_context->ne[1] < 512) {
+                    auto pad_ctx    = ggml_ext_zeros(ctx->ggml_ctx,
+                                                     adapted_context->ne[0],
+                                                     512 - adapted_context->ne[1],
+                                                     adapted_context->ne[2],
+                                                     1);
+                    adapted_context = ggml_concat(ctx->ggml_ctx, adapted_context, pad_ctx, 1);
+                } else if (adapted_context->ne[1] > 512) {
+                    adapted_context = ggml_ext_slice(ctx->ggml_ctx, adapted_context, 1, 0, 512);
+                }
+                encoder_hidden_states = adapted_context;
+            }
+
+            for (int i = 0; i < num_layers; i++) {
+                auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["blocks." + std::to_string(i)]);
+                x          = block->forward(ctx, x, encoder_hidden_states, embedded_timestep, temb, image_pe);
+            }
+
+            x = final_layer->forward(ctx, x, embedded_timestep, temb);  // [N, h*w, ph*pw*C]
+
+            x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, patch_size, patch_size, false);  // [N, C, H, W]
+
+            return x;
+        }
+    };
+
+    struct AnimaRunner : public GGMLRunner {
+    public:
+        std::vector<float> image_pe_vec;
+        std::vector<float> adapter_q_pe_vec;
+        std::vector<float> adapter_k_pe_vec;
+        AnimaNet net;
+
+        AnimaRunner(ggml_backend_t backend,
+                    bool offload_params_to_cpu,
+                    const String2TensorStorage& tensor_storage_map = {},
+                    const std::string prefix                       = "model.diffusion_model")
+            : GGMLRunner(backend, offload_params_to_cpu) {
+            int64_t num_layers    = 0;
+            std::string layer_tag = prefix + ".net.blocks.";
+            for (const auto& kv : tensor_storage_map) {
+                const std::string& tensor_name = kv.first;
+                size_t pos                     = tensor_name.find(layer_tag);
+                if (pos == std::string::npos) {
+                    continue;
+                }
+                size_t start = pos + layer_tag.size();
+                size_t end   = tensor_name.find('.', start);
+                if (end == std::string::npos) {
+                    continue;
+                }
+                int64_t layer_id = atoll(tensor_name.substr(start, end - start).c_str());
+                num_layers       = std::max(num_layers, layer_id + 1);
+            }
+            if (num_layers <= 0) {
+                num_layers = 28;
+            }
+            LOG_INFO("anima net layers: %" PRId64, num_layers);
+
+            net = AnimaNet(num_layers);
+            net.init(params_ctx, tensor_storage_map, prefix + ".net");
+        }
+
+        std::string get_desc() override {
+            return "anima";
+        }
+
+        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+            net.get_param_tensors(tensors, prefix + ".net");
+        }
+
+        static std::vector<float> gen_1d_rope_pe_vec(int64_t seq_len, int dim, float theta = 10000.f) {
+            std::vector<float> pos(seq_len);
+            for (int64_t i = 0; i < seq_len; i++) {
+                pos[i] = static_cast<float>(i);
+            }
+            auto rope_emb = Rope::rope(pos, dim, theta);
+            return Rope::flatten(rope_emb);
+        }
+
+        static float calc_ntk_factor(float extrapolation_ratio, int axis_dim) {
+            if (extrapolation_ratio == 1.0f || axis_dim <= 2) {
+                return 1.0f;
+            }
+            return std::pow(extrapolation_ratio, static_cast<float>(axis_dim) / static_cast<float>(axis_dim - 2));
+        }
+
+        static std::vector<float> gen_anima_image_pe_vec(int bs,
+                                                         int h,
+                                                         int w,
+                                                         int patch_size,
+                                                         int theta,
+                                                         const std::vector<int>& axes_dim,
+                                                         float h_extrapolation_ratio,
+                                                         float w_extrapolation_ratio,
+                                                         float t_extrapolation_ratio) {
+            static const std::vector<ggml_tensor*> empty_ref_latents;
+            auto ids = Rope::gen_flux_ids(h,
+                                          w,
+                                          patch_size,
+                                          bs,
+                                          static_cast<int>(axes_dim.size()),
+                                          0,
+                                          {},
+                                          empty_ref_latents,
+                                          false,
+                                          1.0f);
+
+            std::vector<float> axis_thetas = {
+                static_cast<float>(theta) * calc_ntk_factor(t_extrapolation_ratio, axes_dim[0]),
+                static_cast<float>(theta) * calc_ntk_factor(h_extrapolation_ratio, axes_dim[1]),
+                static_cast<float>(theta) * calc_ntk_factor(w_extrapolation_ratio, axes_dim[2]),
+            };
+            return Rope::embed_nd(ids, bs, axis_thetas, axes_dim);
+        }
+
+        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+                                        struct ggml_tensor* timesteps,
+                                        struct ggml_tensor* context,
+                                        struct ggml_tensor* t5_ids     = nullptr,
+                                        struct ggml_tensor* t5_weights = nullptr) {
+            GGML_ASSERT(x->ne[3] == 1);
+            struct ggml_cgraph* gf = new_graph_custom(ANIMA_GRAPH_SIZE);
+
+            x          = to_backend(x);
+            timesteps  = to_backend(timesteps);
+            context    = to_backend(context);
+            t5_ids     = to_backend(t5_ids);
+            t5_weights = to_backend(t5_weights);
+
+            int64_t pad_h = (net.patch_size - x->ne[1] % net.patch_size) % net.patch_size;
+            int64_t pad_w = (net.patch_size - x->ne[0] % net.patch_size) % net.patch_size;
+            int64_t h_pad = x->ne[1] + pad_h;
+            int64_t w_pad = x->ne[0] + pad_w;
+
+            image_pe_vec          = gen_anima_image_pe_vec(1,
+                                                           static_cast<int>(h_pad),
+                                                           static_cast<int>(w_pad),
+                                                           static_cast<int>(net.patch_size),
+                                                           net.theta,
+                                                           net.axes_dim,
+                                                           4.0f,
+                                                           4.0f,
+                                                           1.0f);
+            int64_t image_pos_len = static_cast<int64_t>(image_pe_vec.size()) / (2 * 2 * (net.head_dim / 2));
+            auto image_pe         = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, net.head_dim / 2, image_pos_len);
+            set_backend_tensor_data(image_pe, image_pe_vec.data());
+
+            ggml_tensor* adapter_q_pe = nullptr;
+            ggml_tensor* adapter_k_pe = nullptr;
+            if (t5_ids != nullptr) {
+                int64_t target_len = t5_ids->ne[0];
+                int64_t source_len = context->ne[1];
+
+                adapter_q_pe_vec = gen_1d_rope_pe_vec(target_len, 64, 10000.f);
+                adapter_k_pe_vec = gen_1d_rope_pe_vec(source_len, 64, 10000.f);
+
+                int64_t target_pos_len = static_cast<int64_t>(adapter_q_pe_vec.size()) / (2 * 2 * 32);
+                int64_t source_pos_len = static_cast<int64_t>(adapter_k_pe_vec.size()) / (2 * 2 * 32);
+
+                adapter_q_pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, 32, target_pos_len);
+                adapter_k_pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, 32, source_pos_len);
+                set_backend_tensor_data(adapter_q_pe, adapter_q_pe_vec.data());
+                set_backend_tensor_data(adapter_k_pe, adapter_k_pe_vec.data());
+            }
+
+            auto runner_ctx = get_context();
+            auto out        = net.forward(&runner_ctx,
+                                          x,
+                                          timesteps,
+                                          context,
+                                          image_pe,
+                                          t5_ids,
+                                          t5_weights,
+                                          adapter_q_pe,
+                                          adapter_k_pe);
+
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        bool compute(int n_threads,
+                     struct ggml_tensor* x,
+                     struct ggml_tensor* timesteps,
+                     struct ggml_tensor* context,
+                     struct ggml_tensor* t5_ids      = nullptr,
+                     struct ggml_tensor* t5_weights  = nullptr,
+                     struct ggml_tensor** output     = nullptr,
+                     struct ggml_context* output_ctx = nullptr) {
+            auto get_graph = [&]() -> struct ggml_cgraph* {
+                return build_graph(x, timesteps, context, t5_ids, t5_weights);
+            };
+            return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        }
+    };
+}  // namespace Anima
+
+#endif  // __ANIMA_HPP__

src/clip.hpp

@@ -4,6 +4,7 @@
 #include "ggml_extend.hpp"
 #include "model.h"
 #include "tokenize_util.h"
+#include "vocab/vocab.h"
 
 /*================================================== CLIPTokenizer ===================================================*/
 
@@ -110,7 +111,7 @@ public:
         if (merges_utf8_str.size() > 0) {
             load_from_merges(merges_utf8_str);
         } else {
-            load_from_merges(ModelLoader::load_merges());
+            load_from_merges(load_clip_merges());
         }
         add_special_token("<|startoftext|>");
         add_special_token("<|endoftext|>");

src/common_block.hpp

@@ -1,5 +1,5 @@
-#ifndef __COMMON_HPP__
-#define __COMMON_HPP__
+#ifndef __COMMON_BLOCK_HPP__
+#define __COMMON_BLOCK_HPP__
 
 #include "ggml_extend.hpp"
 
@@ -590,4 +590,4 @@ public:
     }
 };
 
-#endif  // __COMMON_HPP__
+#endif  // __COMMON_BLOCK_HPP__

src/common_dit.hpp

@@ -0,0 +1,108 @@
+#ifndef __COMMON_DIT_HPP__
+#define __COMMON_DIT_HPP__
+
+#include "ggml_extend.hpp"
+
+namespace DiT {
+    ggml_tensor* patchify(ggml_context* ctx,
+                          ggml_tensor* x,
+                          int pw,
+                          int ph,
+                          bool patch_last = true) {
+        // x: [N, C, H, W]
+        // return: [N, h*w, C*ph*pw] if patch_last else [N, h*w, ph*pw*C]
+        int64_t N = x->ne[3];
+        int64_t C = x->ne[2];
+        int64_t H = x->ne[1];
+        int64_t W = x->ne[0];
+        int64_t h = H / ph;
+        int64_t w = W / pw;
+
+        GGML_ASSERT(h * ph == H && w * pw == W);
+
+        x = ggml_reshape_4d(ctx, x, pw, w, ph, h * C * N);     // [N*C*h, ph, w, pw]
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, w, ph, pw]
+        x = ggml_reshape_4d(ctx, x, pw * ph, w * h, C, N);     // [N, C, h*w, ph*pw]
+        if (patch_last) {
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, h*w, C, ph*pw]
+            x = ggml_reshape_3d(ctx, x, pw * ph * C, w * h, N);    // [N, h*w, C*ph*pw]
+        } else {
+            x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3));  // [N, h*w, C, ph*pw]
+            x = ggml_reshape_3d(ctx, x, C * pw * ph, w * h, N);              // [N, h*w, ph*pw*C]
+        }
+        return x;
+    }
+
+    ggml_tensor* unpatchify(ggml_context* ctx,
+                            ggml_tensor* x,
+                            int64_t h,
+                            int64_t w,
+                            int ph,
+                            int pw,
+                            bool patch_last = true) {
+        // x: [N, h*w, C*ph*pw] if patch_last else [N, h*w, ph*pw*C]
+        // return: [N, C, H, W]
+        int64_t N = x->ne[2];
+        int64_t C = x->ne[0] / ph / pw;
+        int64_t H = h * ph;
+        int64_t W = w * pw;
+
+        GGML_ASSERT(C * ph * pw == x->ne[0]);
+
+        if (patch_last) {
+            x = ggml_reshape_4d(ctx, x, pw * ph, C, w * h, N);     // [N, h*w, C, ph*pw]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, C, h*w, ph*pw]
+        } else {
+            x = ggml_reshape_4d(ctx, x, C, pw * ph, w * h, N);     // [N, h*w, ph*pw, C]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 2, 0, 1, 3));  // [N, C, h*w, ph*pw]
+        }
+
+        x = ggml_reshape_4d(ctx, x, pw, ph, w, h * C * N);     // [N*C*h, w, ph, pw]
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, ph, w, pw]
+        x = ggml_reshape_4d(ctx, x, W, H, C, N);               // [N, C, h*ph, w*pw]
+
+        return x;
+    }
+
+    ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
+                                   ggml_tensor* x,
+                                   int ph,
+                                   int pw) {
+        int64_t W = x->ne[0];
+        int64_t H = x->ne[1];
+
+        int pad_h = (ph - H % ph) % ph;
+        int pad_w = (pw - W % pw) % pw;
+        x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
+        return x;
+    }
+
+    ggml_tensor* pad_and_patchify(GGMLRunnerContext* ctx,
+                                  ggml_tensor* x,
+                                  int ph,
+                                  int pw,
+                                  bool patch_last = true) {
+        x = pad_to_patch_size(ctx, x, ph, pw);
+        x = patchify(ctx->ggml_ctx, x, ph, pw, patch_last);
+        return x;
+    }
+
+    ggml_tensor* unpatchify_and_crop(ggml_context* ctx,
+                                     ggml_tensor* x,
+                                     int64_t H,
+                                     int64_t W,
+                                     int ph,
+                                     int pw,
+                                     bool patch_last = true) {
+        int pad_h = (ph - H % ph) % ph;
+        int pad_w = (pw - W % pw) % pw;
+        int64_t h = ((H + pad_h) / ph);
+        int64_t w = ((W + pad_w) / pw);
+        x         = unpatchify(ctx, x, h, w, ph, pw, patch_last);  // [N, C, H + pad_h, W + pad_w]
+        x         = ggml_ext_slice(ctx, x, 1, 0, H);               // [N, C, H, W + pad_w]
+        x         = ggml_ext_slice(ctx, x, 0, 0, W);               // [N, C, H, W]
+        return x;
+    }
+}  // namespace DiT
+
+#endif  // __COMMON_DIT_HPP__

src/conditioner.hpp

@@ -10,9 +10,14 @@ struct SDCondition {
     struct ggml_tensor* c_vector    = nullptr;  // aka y
     struct ggml_tensor* c_concat    = nullptr;
 
+    std::vector<struct ggml_tensor*> extra_c_crossattns;
+
     SDCondition() = default;
-    SDCondition(struct ggml_tensor* c_crossattn, struct ggml_tensor* c_vector, struct ggml_tensor* c_concat)
-        : c_crossattn(c_crossattn), c_vector(c_vector), c_concat(c_concat) {}
+    SDCondition(struct ggml_tensor* c_crossattn,
+                struct ggml_tensor* c_vector,
+                struct ggml_tensor* c_concat,
+                const std::vector<struct ggml_tensor*>& extra_c_crossattns = {})
+        : c_crossattn(c_crossattn), c_vector(c_vector), c_concat(c_concat), extra_c_crossattns(extra_c_crossattns) {}
 };
 
 struct ConditionerParams {
@@ -1636,6 +1641,142 @@ struct T5CLIPEmbedder : public Conditioner {
     }
 };
 
+struct AnimaConditioner : public Conditioner {
+    std::shared_ptr<LLM::BPETokenizer> qwen_tokenizer;
+    T5UniGramTokenizer t5_tokenizer;
+    std::shared_ptr<LLM::LLMRunner> llm;
+
+    AnimaConditioner(ggml_backend_t backend,
+                     bool offload_params_to_cpu,
+                     const String2TensorStorage& tensor_storage_map = {}) {
+        qwen_tokenizer = std::make_shared<LLM::Qwen2Tokenizer>();
+        llm            = std::make_shared<LLM::LLMRunner>(LLM::LLMArch::QWEN3,
+                                               backend,
+                                               offload_params_to_cpu,
+                                               tensor_storage_map,
+                                               "text_encoders.llm",
+                                               false);
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        llm->get_param_tensors(tensors, "text_encoders.llm");
+    }
+
+    void alloc_params_buffer() override {
+        llm->alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        llm->free_params_buffer();
+    }
+
+    size_t get_params_buffer_size() override {
+        return llm->get_params_buffer_size();
+    }
+
+    void set_flash_attention_enabled(bool enabled) override {
+        llm->set_flash_attention_enabled(enabled);
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        llm->set_weight_adapter(adapter);
+    }
+
+    std::tuple<std::vector<int>, std::vector<float>, std::vector<int>, std::vector<float>> tokenize(std::string text) {
+        auto parsed_attention = parse_prompt_attention(text);
+
+        {
+            std::stringstream ss;
+            ss << "[";
+            for (const auto& item : parsed_attention) {
+                ss << "['" << item.first << "', " << item.second << "], ";
+            }
+            ss << "]";
+            LOG_DEBUG("parse '%s' to %s", text.c_str(), ss.str().c_str());
+        }
+
+        std::vector<int> qwen_tokens;
+        std::vector<float> qwen_weights;
+        std::vector<int> t5_tokens;
+        std::vector<float> t5_weights;
+
+        for (const auto& item : parsed_attention) {
+            const std::string& curr_text = item.first;
+            std::vector<int> curr_tokens = qwen_tokenizer->tokenize(curr_text, nullptr);
+            qwen_tokens.insert(qwen_tokens.end(), curr_tokens.begin(), curr_tokens.end());
+            // Anima uses uniform Qwen token weights.
+            qwen_weights.insert(qwen_weights.end(), curr_tokens.size(), 1.f);
+        }
+        if (qwen_tokens.empty()) {
+            qwen_tokens.push_back(151643);  // qwen3 pad token
+            qwen_weights.push_back(1.f);
+        }
+
+        for (const auto& item : parsed_attention) {
+            const std::string& curr_text = item.first;
+            float curr_weight            = item.second;
+            std::vector<int> curr_tokens = t5_tokenizer.Encode(curr_text, true);
+            t5_tokens.insert(t5_tokens.end(), curr_tokens.begin(), curr_tokens.end());
+            t5_weights.insert(t5_weights.end(), curr_tokens.size(), curr_weight);
+        }
+
+        return {qwen_tokens, qwen_weights, t5_tokens, t5_weights};
+    }
+
+    SDCondition get_learned_condition(ggml_context* work_ctx,
+                                      int n_threads,
+                                      const ConditionerParams& conditioner_params) override {
+        int64_t t0 = ggml_time_ms();
+
+        auto tokenized     = tokenize(conditioner_params.text);
+        auto& qwen_tokens  = std::get<0>(tokenized);
+        auto& qwen_weights = std::get<1>(tokenized);
+        auto& t5_tokens    = std::get<2>(tokenized);
+        auto& t5_weights   = std::get<3>(tokenized);
+
+        auto input_ids = vector_to_ggml_tensor_i32(work_ctx, qwen_tokens);
+
+        struct ggml_tensor* hidden_states = nullptr;  // [N, n_token, 1024]
+        llm->compute(n_threads,
+                     input_ids,
+                     nullptr,
+                     {},
+                     {},
+                     &hidden_states,
+                     work_ctx);
+
+        {
+            auto tensor         = hidden_states;
+            float original_mean = ggml_ext_tensor_mean(tensor);
+            for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                    for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                        float value = ggml_ext_tensor_get_f32(tensor, i0, i1, i2);
+                        value *= qwen_weights[i1];
+                        ggml_ext_tensor_set_f32(tensor, value, i0, i1, i2);
+                    }
+                }
+            }
+            float new_mean = ggml_ext_tensor_mean(tensor);
+            if (new_mean != 0.f) {
+                ggml_ext_tensor_scale_inplace(tensor, (original_mean / new_mean));
+            }
+        }
+
+        struct ggml_tensor* t5_ids_tensor    = nullptr;
+        struct ggml_tensor* t5_weight_tensor = nullptr;
+        if (!t5_tokens.empty()) {
+            t5_ids_tensor    = vector_to_ggml_tensor_i32(work_ctx, t5_tokens);
+            t5_weight_tensor = vector_to_ggml_tensor(work_ctx, t5_weights);
+        }
+
+        int64_t t1 = ggml_time_ms();
+        LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
+
+        return {hidden_states, t5_weight_tensor, t5_ids_tensor};
+    }
+};
+
 struct LLMEmbedder : public Conditioner {
     SDVersion version;
     std::shared_ptr<LLM::BPETokenizer> tokenizer;
@@ -1696,18 +1837,23 @@ struct LLMEmbedder : public Conditioner {
     }
 
     std::tuple<std::vector<int>, std::vector<float>> tokenize(std::string text,
-                                                              std::pair<int, int> attn_range,
+                                                              const std::pair<int, int>& attn_range,
                                                               size_t max_length = 0,
                                                               bool padding      = false) {
         std::vector<std::pair<std::string, float>> parsed_attention;
-        parsed_attention.emplace_back(text.substr(0, attn_range.first), 1.f);
-        if (attn_range.second - attn_range.first > 0) {
-            auto new_parsed_attention = parse_prompt_attention(text.substr(attn_range.first, attn_range.second - attn_range.first));
-            parsed_attention.insert(parsed_attention.end(),
-                                    new_parsed_attention.begin(),
-                                    new_parsed_attention.end());
+        if (attn_range.first >= 0 && attn_range.second > 0) {
+            parsed_attention.emplace_back(text.substr(0, attn_range.first), 1.f);
+            if (attn_range.second - attn_range.first > 0) {
+                auto new_parsed_attention = parse_prompt_attention(text.substr(attn_range.first, attn_range.second - attn_range.first));
+                parsed_attention.insert(parsed_attention.end(),
+                                        new_parsed_attention.begin(),
+                                        new_parsed_attention.end());
+            }
+            parsed_attention.emplace_back(text.substr(attn_range.second), 1.f);
+        } else {
+            parsed_attention.emplace_back(text, 1.f);
         }
-        parsed_attention.emplace_back(text.substr(attn_range.second), 1.f);
+
         {
             std::stringstream ss;
             ss << "[";
@@ -1738,156 +1884,27 @@ struct LLMEmbedder : public Conditioner {
         return {tokens, weights};
     }
 
-    SDCondition get_learned_condition(ggml_context* work_ctx,
-                                      int n_threads,
-                                      const ConditionerParams& conditioner_params) override {
-        std::string prompt;
-        std::vector<std::pair<int, ggml_tensor*>> image_embeds;
-        std::pair<int, int> prompt_attn_range;
-        int prompt_template_encode_start_idx = 34;
-        int max_length                       = 0;
-        std::set<int> out_layers;
-        std::vector<int> tokens;
-        std::vector<float> weights;
+    ggml_tensor* encode_prompt(ggml_context* work_ctx,
+                               int n_threads,
+                               const std::string prompt,
+                               const std::pair<int, int>& prompt_attn_range,
+                               int max_length,
+                               int min_length,
+                               std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+                               const std::set<int>& out_layers,
+                               int prompt_template_encode_start_idx) {
+        auto tokens_and_weights = tokenize(prompt, prompt_attn_range);
+        auto& tokens            = std::get<0>(tokens_and_weights);
+        auto& weights           = std::get<1>(tokens_and_weights);
         std::vector<float> mask;
-        if (llm->enable_vision && conditioner_params.ref_images.size() > 0) {
-            LOG_INFO("QwenImageEditPlusPipeline");
-            prompt_template_encode_start_idx = 64;
-            int image_embed_idx              = 64 + 6;
-
-            int min_pixels          = 384 * 384;
-            int max_pixels          = 560 * 560;
-            std::string placeholder = "<|image_pad|>";
-            std::string img_prompt;
-
-            for (int i = 0; i < conditioner_params.ref_images.size(); i++) {
-                sd_image_f32_t image = sd_image_t_to_sd_image_f32_t(*conditioner_params.ref_images[i]);
-                double factor        = llm->params.vision.patch_size * llm->params.vision.spatial_merge_size;
-                int height           = image.height;
-                int width            = image.width;
-                int h_bar            = static_cast<int>(std::round(height / factor) * factor);
-                int w_bar            = static_cast<int>(std::round(width / factor) * factor);
-
-                if (static_cast<double>(h_bar) * w_bar > max_pixels) {
-                    double beta = std::sqrt((height * width) / static_cast<double>(max_pixels));
-                    h_bar       = std::max(static_cast<int>(factor),
-                                           static_cast<int>(std::floor(height / beta / factor)) * static_cast<int>(factor));
-                    w_bar       = std::max(static_cast<int>(factor),
-                                           static_cast<int>(std::floor(width / beta / factor)) * static_cast<int>(factor));
-                } else if (static_cast<double>(h_bar) * w_bar < min_pixels) {
-                    double beta = std::sqrt(static_cast<double>(min_pixels) / (height * width));
-                    h_bar       = static_cast<int>(std::ceil(height * beta / factor)) * static_cast<int>(factor);
-                    w_bar       = static_cast<int>(std::ceil(width * beta / factor)) * static_cast<int>(factor);
-                }
-
-                LOG_DEBUG("resize conditioner ref image %d from %dx%d to %dx%d", i, image.height, image.width, h_bar, w_bar);
-
-                sd_image_f32_t resized_image = clip_preprocess(image, w_bar, h_bar);
-                free(image.data);
-                image.data = nullptr;
-
-                ggml_tensor* image_tensor = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, resized_image.width, resized_image.height, 3, 1);
-                sd_image_f32_to_ggml_tensor(resized_image, image_tensor, false);
-                free(resized_image.data);
-                resized_image.data = nullptr;
-
-                ggml_tensor* image_embed = nullptr;
-                llm->encode_image(n_threads, image_tensor, &image_embed, work_ctx);
-                image_embeds.emplace_back(image_embed_idx, image_embed);
-                image_embed_idx += 1 + static_cast<int>(image_embed->ne[1]) + 6;
-
-                img_prompt += "Picture " + std::to_string(i + 1) + ": <|vision_start|>";  // [24669, 220, index, 25, 220, 151652]
-                int64_t num_image_tokens = image_embed->ne[1];
-                img_prompt.reserve(num_image_tokens * placeholder.size());
-                for (int j = 0; j < num_image_tokens; j++) {
-                    img_prompt += placeholder;
-                }
-                img_prompt += "<|vision_end|>";
-            }
-
-            prompt = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n";
-            prompt += 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<|im_start|>assistant\n";
-        } else if (version == VERSION_FLUX2) {
-            prompt_template_encode_start_idx = 0;
-            out_layers                       = {10, 20, 30};
-
-            prompt = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]";
-
-            prompt_attn_range.first = static_cast<int>(prompt.size());
-            prompt += conditioner_params.text;
-            prompt_attn_range.second = static_cast<int>(prompt.size());
-
-            prompt += "[/INST]";
-        } else if (sd_version_is_z_image(version)) {
-            prompt_template_encode_start_idx = 0;
-            out_layers                       = {35};  // -2
-
-            prompt = "<|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<|im_start|>assistant\n";
-        } else if (version == VERSION_FLUX2_KLEIN) {
-            prompt_template_encode_start_idx = 0;
-            max_length                       = 512;
-            out_layers                       = {9, 18, 27};
-
-            prompt = "<|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<|im_start|>assistant\n<think>\n\n</think>\n\n";
-
-            auto tokens_and_weights = tokenize(prompt, prompt_attn_range, 0, false);
-            tokens                  = std::get<0>(tokens_and_weights);
-            weights                 = std::get<1>(tokens_and_weights);
 
+        if (max_length > 0 && tokens.size() < max_length) {
             mask.insert(mask.end(), tokens.size(), 1.f);
-            if (tokens.size() < max_length) {
-                mask.insert(mask.end(), max_length - tokens.size(), 0.f);
-                tokenizer->pad_tokens(tokens, weights, max_length, true);
-            }
-        } else if (version == VERSION_OVIS_IMAGE) {
-            prompt_template_encode_start_idx = 28;
-            max_length                       = prompt_template_encode_start_idx + 256;
-
-            prompt = "<|im_start|>user\nDescribe the image by detailing the color, quantity, text, shape, size, texture, spatial relationships of the objects and background:";
-
-            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<|im_start|>assistant\n<think>\n\n</think>\n\n";
-        } else {
-            prompt_template_encode_start_idx = 34;
-
-            prompt = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n";
-
-            prompt_attn_range.first = static_cast<int>(prompt.size());
-            prompt += conditioner_params.text;
-            prompt_attn_range.second = static_cast<int>(prompt.size());
-
-            prompt += "<|im_end|>\n<|im_start|>assistant\n";
+            mask.insert(mask.end(), max_length - tokens.size(), 0.f);
+            tokenizer->pad_tokens(tokens, weights, max_length, true);
         }
 
-        if (tokens.empty()) {
-            auto tokens_and_weights = tokenize(prompt, prompt_attn_range, max_length, max_length > 0);
-            tokens                  = std::get<0>(tokens_and_weights);
-            weights                 = std::get<1>(tokens_and_weights);
-        }
-
-        int64_t t0                        = ggml_time_ms();
-        struct ggml_tensor* hidden_states = nullptr;  // [N, n_token, 3584]
+        struct ggml_tensor* hidden_states = nullptr;  // [N, n_token, hidden_size]
 
         auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
 
@@ -1930,11 +1947,6 @@ struct LLMEmbedder : public Conditioner {
 
         GGML_ASSERT(hidden_states->ne[1] > prompt_template_encode_start_idx);
 
-        int64_t min_length = 0;
-        if (version == VERSION_FLUX2) {
-            min_length = 512;
-        }
-
         int64_t zero_pad_len = 0;
         if (min_length > 0) {
             if (hidden_states->ne[1] - prompt_template_encode_start_idx < min_length) {
@@ -1956,11 +1968,186 @@ struct LLMEmbedder : public Conditioner {
             ggml_ext_tensor_set_f32(new_hidden_states, value, i0, i1, i2, i3);
         });
 
-        // print_ggml_tensor(new_hidden_states);
+        return new_hidden_states;
+    }
+
+    SDCondition get_learned_condition(ggml_context* work_ctx,
+                                      int n_threads,
+                                      const ConditionerParams& conditioner_params) override {
+        std::string prompt;
+        std::pair<int, int> prompt_attn_range;
+        std::vector<std::string> extra_prompts;
+        std::vector<std::pair<int, int>> extra_prompts_attn_range;
+        std::vector<std::pair<int, ggml_tensor*>> image_embeds;
+        int prompt_template_encode_start_idx = 34;
+        int max_length                       = 0;  // pad tokens
+        int min_length                       = 0;  // zero pad hidden_states
+        std::set<int> out_layers;
+
+        int64_t t0 = ggml_time_ms();
+
+        if (sd_version_is_qwen_image(version)) {
+            if (llm->enable_vision && !conditioner_params.ref_images.empty()) {
+                LOG_INFO("QwenImageEditPlusPipeline");
+                prompt_template_encode_start_idx = 64;
+                int image_embed_idx              = 64 + 6;
+
+                int min_pixels          = 384 * 384;
+                int max_pixels          = 560 * 560;
+                std::string placeholder = "<|image_pad|>";
+                std::string img_prompt;
+
+                for (int i = 0; i < conditioner_params.ref_images.size(); i++) {
+                    sd_image_f32_t image = sd_image_t_to_sd_image_f32_t(*conditioner_params.ref_images[i]);
+                    double factor        = llm->params.vision.patch_size * llm->params.vision.spatial_merge_size;
+                    int height           = image.height;
+                    int width            = image.width;
+                    int h_bar            = static_cast<int>(std::round(height / factor) * factor);
+                    int w_bar            = static_cast<int>(std::round(width / factor) * factor);
+
+                    if (static_cast<double>(h_bar) * w_bar > max_pixels) {
+                        double beta = std::sqrt((height * width) / static_cast<double>(max_pixels));
+                        h_bar       = std::max(static_cast<int>(factor),
+                                               static_cast<int>(std::floor(height / beta / factor)) * static_cast<int>(factor));
+                        w_bar       = std::max(static_cast<int>(factor),
+                                               static_cast<int>(std::floor(width / beta / factor)) * static_cast<int>(factor));
+                    } else if (static_cast<double>(h_bar) * w_bar < min_pixels) {
+                        double beta = std::sqrt(static_cast<double>(min_pixels) / (height * width));
+                        h_bar       = static_cast<int>(std::ceil(height * beta / factor)) * static_cast<int>(factor);
+                        w_bar       = static_cast<int>(std::ceil(width * beta / factor)) * static_cast<int>(factor);
+                    }
+
+                    LOG_DEBUG("resize conditioner ref image %d from %dx%d to %dx%d", i, image.height, image.width, h_bar, w_bar);
+
+                    sd_image_f32_t resized_image = clip_preprocess(image, w_bar, h_bar);
+                    free(image.data);
+                    image.data = nullptr;
+
+                    ggml_tensor* image_tensor = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, resized_image.width, resized_image.height, 3, 1);
+                    sd_image_f32_to_ggml_tensor(resized_image, image_tensor, false);
+                    free(resized_image.data);
+                    resized_image.data = nullptr;
+
+                    ggml_tensor* image_embed = nullptr;
+                    llm->encode_image(n_threads, image_tensor, &image_embed, work_ctx);
+                    image_embeds.emplace_back(image_embed_idx, image_embed);
+                    image_embed_idx += 1 + static_cast<int>(image_embed->ne[1]) + 6;
+
+                    img_prompt += "Picture " + std::to_string(i + 1) + ": <|vision_start|>";  // [24669, 220, index, 25, 220, 151652]
+                    int64_t num_image_tokens = image_embed->ne[1];
+                    img_prompt.reserve(num_image_tokens * placeholder.size());
+                    for (int j = 0; j < num_image_tokens; j++) {
+                        img_prompt += placeholder;
+                    }
+                    img_prompt += "<|vision_end|>";
+                }
+
+                prompt = "<|im_start|>system\nDescribe the key features of the input image (color, shape, size, texture, objects, background), then explain how the user's text instruction should alter or modify the image. Generate a new image that meets the user's requirements while maintaining consistency with the original input where appropriate.<|im_end|>\n<|im_start|>user\n";
+                prompt += 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<|im_start|>assistant\n";
+            } else {
+                prompt_template_encode_start_idx = 34;
+
+                prompt = "<|im_start|>system\nDescribe the image by detailing the color, shape, size, texture, quantity, text, spatial relationships of the objects and background:<|im_end|>\n<|im_start|>user\n";
+
+                prompt_attn_range.first = static_cast<int>(prompt.size());
+                prompt += conditioner_params.text;
+                prompt_attn_range.second = static_cast<int>(prompt.size());
+
+                prompt += "<|im_end|>\n<|im_start|>assistant\n";
+            }
+        } else if (version == VERSION_FLUX2) {
+            prompt_template_encode_start_idx = 0;
+            min_length                       = 512;
+            out_layers                       = {10, 20, 30};
+
+            prompt = "[SYSTEM_PROMPT]You are an AI that reasons about image descriptions. You give structured responses focusing on object relationships, object\nattribution and actions without speculation.[/SYSTEM_PROMPT][INST]";
+
+            prompt_attn_range.first = static_cast<int>(prompt.size());
+            prompt += conditioner_params.text;
+            prompt_attn_range.second = static_cast<int>(prompt.size());
+
+            prompt += "[/INST]";
+        } else if (sd_version_is_z_image(version)) {
+            prompt_template_encode_start_idx = 0;
+            out_layers                       = {35};  // -2
+
+            if (!conditioner_params.ref_images.empty()) {
+                LOG_INFO("ZImageOmniPipeline");
+                prompt = "<|im_start|>user\n<|vision_start|>";
+                for (int i = 0; i < conditioner_params.ref_images.size() - 1; i++) {
+                    extra_prompts.push_back("<|vision_end|><|vision_start|>");
+                }
+                extra_prompts.push_back("<|vision_end|>" + conditioner_params.text + "<|im_end|>\n<|im_start|>assistant\n<|vision_start|>");
+                extra_prompts.push_back("<|vision_end|><|im_end|>");
+            } else {
+                prompt = "<|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<|im_start|>assistant\n";
+            }
+        } else if (version == VERSION_FLUX2_KLEIN) {
+            prompt_template_encode_start_idx = 0;
+            max_length                       = 512;
+            out_layers                       = {9, 18, 27};
+
+            prompt = "<|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<|im_start|>assistant\n<think>\n\n</think>\n\n";
+        } else if (version == VERSION_OVIS_IMAGE) {
+            prompt_template_encode_start_idx = 28;
+            max_length                       = prompt_template_encode_start_idx + 256;
+
+            prompt = "<|im_start|>user\nDescribe the image by detailing the color, quantity, text, shape, size, texture, spatial relationships of the objects and background:";
+
+            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<|im_start|>assistant\n<think>\n\n</think>\n\n";
+        } else {
+            GGML_ABORT("unknown version %d", version);
+        }
+
+        auto hidden_states = encode_prompt(work_ctx,
+                                           n_threads,
+                                           prompt,
+                                           prompt_attn_range,
+                                           max_length,
+                                           min_length,
+                                           image_embeds,
+                                           out_layers,
+                                           prompt_template_encode_start_idx);
+
+        std::vector<ggml_tensor*> extra_hidden_states_vec;
+        for (int i = 0; i < extra_prompts.size(); i++) {
+            auto extra_hidden_states = encode_prompt(work_ctx,
+                                                     n_threads,
+                                                     extra_prompts[i],
+                                                     extra_prompts_attn_range[i],
+                                                     max_length,
+                                                     min_length,
+                                                     image_embeds,
+                                                     out_layers,
+                                                     prompt_template_encode_start_idx);
+            extra_hidden_states_vec.push_back(extra_hidden_states);
+        }
 
         int64_t t1 = ggml_time_ms();
         LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
-        return {new_hidden_states, nullptr, nullptr};
+        return {hidden_states, nullptr, nullptr, extra_hidden_states_vec};
     }
 };
 

src/control.hpp

@@ -1,8 +1,7 @@
 #ifndef __CONTROL_HPP__
 #define __CONTROL_HPP__
 
-#include "common.hpp"
-#include "ggml_extend.hpp"
+#include "common_block.hpp"
 #include "model.h"
 
 #define CONTROL_NET_GRAPH_SIZE 1536

src/denoiser.hpp

@@ -657,9 +657,8 @@ struct DiscreteFlowDenoiser : public Denoiser {
 
     float sigma_data = 1.0f;
 
-    DiscreteFlowDenoiser(float shift = 3.0f)
-        : shift(shift) {
-        set_parameters();
+    DiscreteFlowDenoiser(float shift = 3.0f) {
+        set_shift(shift);
     }
 
     void set_parameters() {
@@ -668,6 +667,11 @@ struct DiscreteFlowDenoiser : public Denoiser {
         }
     }
 
+    void set_shift(float shift) {
+        this->shift = shift;
+        set_parameters();
+    }
+
     float sigma_min() override {
         return sigmas[0];
     }
@@ -710,34 +714,8 @@ float flux_time_shift(float mu, float sigma, float t) {
     return ::expf(mu) / (::expf(mu) + ::powf((1.0f / t - 1.0f), sigma));
 }
 
-struct FluxFlowDenoiser : public Denoiser {
-    float sigmas[TIMESTEPS];
-    float shift = 1.15f;
-
-    float sigma_data = 1.0f;
-
-    FluxFlowDenoiser(float shift = 1.15f) {
-        set_parameters(shift);
-    }
-
-    void set_shift(float shift) {
-        this->shift = shift;
-    }
-
-    void set_parameters(float shift) {
-        set_shift(shift);
-        for (int i = 0; i < TIMESTEPS; i++) {
-            sigmas[i] = t_to_sigma(static_cast<float>(i));
-        }
-    }
-
-    float sigma_min() override {
-        return sigmas[0];
-    }
-
-    float sigma_max() override {
-        return sigmas[TIMESTEPS - 1];
-    }
+struct FluxFlowDenoiser : public DiscreteFlowDenoiser {
+    FluxFlowDenoiser() = default;
 
     float sigma_to_t(float sigma) override {
         return sigma;
@@ -747,26 +725,6 @@ struct FluxFlowDenoiser : public Denoiser {
         t = t + 1;
         return flux_time_shift(shift, 1.0f, t / TIMESTEPS);
     }
-
-    std::vector<float> get_scalings(float sigma) override {
-        float c_skip = 1.0f;
-        float c_out  = -sigma;
-        float c_in   = 1.0f;
-        return {c_skip, c_out, c_in};
-    }
-
-    // this function will modify noise/latent
-    ggml_tensor* noise_scaling(float sigma, ggml_tensor* noise, ggml_tensor* latent) override {
-        ggml_ext_tensor_scale_inplace(noise, sigma);
-        ggml_ext_tensor_scale_inplace(latent, 1.0f - sigma);
-        ggml_ext_tensor_add_inplace(latent, noise);
-        return latent;
-    }
-
-    ggml_tensor* inverse_noise_scaling(float sigma, ggml_tensor* latent) override {
-        ggml_ext_tensor_scale_inplace(latent, 1.0f / (1.0f - sigma));
-        return latent;
-    }
 };
 
 struct Flux2FlowDenoiser : public FluxFlowDenoiser {

src/diffusion_model.hpp

@@ -1,6 +1,7 @@
 #ifndef __DIFFUSION_MODEL_H__
 #define __DIFFUSION_MODEL_H__
 
+#include "anima.hpp"
 #include "flux.hpp"
 #include "mmdit.hpp"
 #include "qwen_image.hpp"
@@ -242,6 +243,72 @@ struct FluxModel : public DiffusionModel {
     }
 };
 
+struct AnimaModel : public DiffusionModel {
+    std::string prefix;
+    Anima::AnimaRunner anima;
+
+    AnimaModel(ggml_backend_t backend,
+               bool offload_params_to_cpu,
+               const String2TensorStorage& tensor_storage_map = {},
+               const std::string prefix                       = "model.diffusion_model")
+        : prefix(prefix), anima(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
+    }
+
+    std::string get_desc() override {
+        return anima.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        anima.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        anima.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        anima.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        anima.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() override {
+        return anima.get_params_buffer_size();
+    }
+
+    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+        anima.set_weight_adapter(adapter);
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768;
+    }
+
+    void set_flash_attention_enabled(bool enabled) {
+        anima.set_flash_attention_enabled(enabled);
+    }
+
+    void set_circular_axes(bool circular_x, bool circular_y) override {
+        anima.set_circular_axes(circular_x, circular_y);
+    }
+
+    bool compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return anima.compute(n_threads,
+                             diffusion_params.x,
+                             diffusion_params.timesteps,
+                             diffusion_params.context,
+                             diffusion_params.c_concat,
+                             diffusion_params.y,
+                             output,
+                             output_ctx);
+    }
+};
+
 struct WanModel : public DiffusionModel {
     std::string prefix;
     WAN::WanRunner wan;

src/flux.hpp

@@ -4,7 +4,7 @@
 #include <memory>
 #include <vector>
 
-#include "ggml_extend.hpp"
+#include "common_dit.hpp"
 #include "model.h"
 #include "rope.hpp"
 
@@ -103,11 +103,13 @@ namespace Flux {
             auto norm     = std::dynamic_pointer_cast<QKNorm>(blocks["norm"]);
 
             auto qkv         = qkv_proj->forward(ctx, x);
-            auto qkv_vec     = ggml_ext_chunk(ctx->ggml_ctx, qkv, 3, 0, true);
-            int64_t head_dim = qkv_vec[0]->ne[0] / num_heads;
-            auto q           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[0], head_dim, num_heads, qkv_vec[0]->ne[1], qkv_vec[0]->ne[2]);
-            auto k           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[1], head_dim, num_heads, qkv_vec[1]->ne[1], qkv_vec[1]->ne[2]);
-            auto v           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[2], head_dim, num_heads, qkv_vec[2]->ne[1], qkv_vec[2]->ne[2]);
+            int64_t head_dim = qkv->ne[0] / 3 / num_heads;
+            auto q           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], 0);
+            auto k           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], (qkv->nb[0]) * qkv->ne[0] / 3);
+            auto v           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], (qkv->nb[0]) * 2 * qkv->ne[0] / 3);
             q                = norm->query_norm(ctx, q);
             k                = norm->key_norm(ctx, k);
             return {q, k, v};
@@ -491,15 +493,14 @@ namespace Flux {
             auto x_mod   = Flux::modulate(ctx->ggml_ctx, pre_norm->forward(ctx, x), mod.shift, mod.scale);
             auto qkv_mlp = linear1->forward(ctx, x_mod);  // [N, n_token, hidden_size * 3 + mlp_hidden_dim*mlp_mult_factor]
 
-            auto q = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], 0);
-            auto k = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], hidden_size * qkv_mlp->nb[0]);
-            auto v = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], hidden_size * 2 * qkv_mlp->nb[0]);
-
             int64_t head_dim = hidden_size / num_heads;
 
-            q = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, q), head_dim, num_heads, q->ne[1], q->ne[2]);  // [N, n_token, n_head, d_head]
-            k = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, k), head_dim, num_heads, k->ne[1], k->ne[2]);  // [N, n_token, n_head, d_head]
-            v = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, v), head_dim, num_heads, v->ne[1], v->ne[2]);  // [N, n_token, n_head, d_head]
+            auto q = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], 0);
+            auto k = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], (qkv_mlp->nb[0]) * hidden_size);
+            auto v = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], (qkv_mlp->nb[0]) * 2 * hidden_size);
 
             q         = norm->query_norm(ctx, q);
             k         = norm->key_norm(ctx, k);
@@ -846,70 +847,6 @@ namespace Flux {
             }
         }
 
-        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
-                                              struct ggml_tensor* x) {
-            int64_t W = x->ne[0];
-            int64_t H = x->ne[1];
-
-            int pad_h = (params.patch_size - H % params.patch_size) % params.patch_size;
-            int pad_w = (params.patch_size - W % params.patch_size) % params.patch_size;
-            x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
-            return x;
-        }
-
-        struct ggml_tensor* patchify(struct ggml_context* ctx,
-                                     struct ggml_tensor* x) {
-            // x: [N, C, H, W]
-            // return: [N, h*w, C * patch_size * patch_size]
-            int64_t N = x->ne[3];
-            int64_t C = x->ne[2];
-            int64_t H = x->ne[1];
-            int64_t W = x->ne[0];
-            int64_t p = params.patch_size;
-            int64_t h = H / params.patch_size;
-            int64_t w = W / params.patch_size;
-
-            GGML_ASSERT(h * p == H && w * p == W);
-
-            x = ggml_reshape_4d(ctx, x, p, w, p, h * C * N);       // [N*C*h, p, w, p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, w, p, p]
-            x = ggml_reshape_4d(ctx, x, p * p, w * h, C, N);       // [N, C, h*w, p*p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, h*w, C, p*p]
-            x = ggml_reshape_3d(ctx, x, p * p * C, w * h, N);      // [N, h*w, C*p*p]
-            return x;
-        }
-
-        struct ggml_tensor* process_img(GGMLRunnerContext* ctx,
-                                        struct ggml_tensor* x) {
-            // img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
-            x = pad_to_patch_size(ctx, x);
-            x = patchify(ctx->ggml_ctx, x);
-            return x;
-        }
-
-        struct ggml_tensor* unpatchify(struct ggml_context* ctx,
-                                       struct ggml_tensor* x,
-                                       int64_t h,
-                                       int64_t w) {
-            // x: [N, h*w, C*patch_size*patch_size]
-            // return: [N, C, H, W]
-            int64_t N = x->ne[2];
-            int64_t C = x->ne[0] / params.patch_size / params.patch_size;
-            int64_t H = h * params.patch_size;
-            int64_t W = w * params.patch_size;
-            int64_t p = params.patch_size;
-
-            GGML_ASSERT(C * p * p == x->ne[0]);
-
-            x = ggml_reshape_4d(ctx, x, p * p, C, w * h, N);       // [N, h*w, C, p*p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, C, h*w, p*p]
-            x = ggml_reshape_4d(ctx, x, p, p, w, h * C * N);       // [N*C*h, w, p, p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, p, w, p]
-            x = ggml_reshape_4d(ctx, x, W, H, C, N);               // [N, C, h*p, w*p]
-
-            return x;
-        }
-
         struct ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
                                          struct ggml_tensor* img,
                                          struct ggml_tensor* txt,
@@ -1060,7 +997,7 @@ namespace Flux {
             int pad_h      = (patch_size - H % patch_size) % patch_size;
             int pad_w      = (patch_size - W % patch_size) % patch_size;
 
-            auto img      = pad_to_patch_size(ctx, x);
+            auto img      = DiT::pad_to_patch_size(ctx, x, params.patch_size, params.patch_size);
             auto orig_img = img;
 
             if (params.chroma_radiance_params.fake_patch_size_x2) {
@@ -1082,7 +1019,7 @@ namespace Flux {
             auto nerf_image_embedder   = std::dynamic_pointer_cast<NerfEmbedder>(blocks["nerf_image_embedder"]);
             auto nerf_final_layer_conv = std::dynamic_pointer_cast<NerfFinalLayerConv>(blocks["nerf_final_layer_conv"]);
 
-            auto nerf_pixels    = patchify(ctx->ggml_ctx, orig_img);  // [N, num_patches, C * patch_size * patch_size]
+            auto nerf_pixels    = DiT::patchify(ctx->ggml_ctx, orig_img, patch_size, patch_size);  // [N, num_patches, C * patch_size * patch_size]
             int64_t num_patches = nerf_pixels->ne[1];
             nerf_pixels         = ggml_reshape_3d(ctx->ggml_ctx,
                                                   nerf_pixels,
@@ -1102,7 +1039,7 @@ namespace Flux {
 
             img_dct = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, img_dct, 1, 0, 2, 3));                                 // [N*num_patches, nerf_hidden_size, patch_size*patch_size]
             img_dct = ggml_reshape_3d(ctx->ggml_ctx, img_dct, img_dct->ne[0] * img_dct->ne[1], num_patches, img_dct->ne[2] / num_patches);  // [N, num_patches, nerf_hidden_size*patch_size*patch_size]
-            img_dct = unpatchify(ctx->ggml_ctx, img_dct, (H + pad_h) / patch_size, (W + pad_w) / patch_size);                               // [N, nerf_hidden_size, H, W]
+            img_dct = DiT::unpatchify(ctx->ggml_ctx, img_dct, (H + pad_h) / patch_size, (W + pad_w) / patch_size, patch_size, patch_size);  // [N, nerf_hidden_size, H, W]
 
             out = nerf_final_layer_conv->forward(ctx, img_dct);  // [N, C, H, W]
 
@@ -1134,7 +1071,7 @@ namespace Flux {
             int pad_h      = (patch_size - H % patch_size) % patch_size;
             int pad_w      = (patch_size - W % patch_size) % patch_size;
 
-            auto img           = process_img(ctx, x);
+            auto img           = DiT::pad_and_patchify(ctx, x, patch_size, patch_size);
             int64_t img_tokens = img->ne[1];
 
             if (params.version == VERSION_FLUX_FILL) {
@@ -1142,8 +1079,8 @@ namespace Flux {
                 ggml_tensor* masked = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], C, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], 0);
                 ggml_tensor* mask   = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], 8 * 8, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C);
 
-                masked = process_img(ctx, masked);
-                mask   = process_img(ctx, mask);
+                masked = DiT::pad_and_patchify(ctx, masked, patch_size, patch_size);
+                mask   = DiT::pad_and_patchify(ctx, mask, patch_size, patch_size);
 
                 img = ggml_concat(ctx->ggml_ctx, img, ggml_concat(ctx->ggml_ctx, masked, mask, 0), 0);
             } else if (params.version == VERSION_FLEX_2) {
@@ -1152,21 +1089,21 @@ namespace Flux {
                 ggml_tensor* mask    = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], 1, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C);
                 ggml_tensor* control = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], C, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * (C + 1));
 
-                masked  = process_img(ctx, masked);
-                mask    = process_img(ctx, mask);
-                control = process_img(ctx, control);
+                masked  = DiT::pad_and_patchify(ctx, masked, patch_size, patch_size);
+                mask    = DiT::pad_and_patchify(ctx, mask, patch_size, patch_size);
+                control = DiT::pad_and_patchify(ctx, control, patch_size, patch_size);
 
                 img = ggml_concat(ctx->ggml_ctx, img, ggml_concat(ctx->ggml_ctx, ggml_concat(ctx->ggml_ctx, masked, mask, 0), control, 0), 0);
             } else if (params.version == VERSION_FLUX_CONTROLS) {
                 GGML_ASSERT(c_concat != nullptr);
 
-                auto control = process_img(ctx, c_concat);
+                auto control = DiT::pad_and_patchify(ctx, c_concat, patch_size, patch_size);
                 img          = ggml_concat(ctx->ggml_ctx, img, control, 0);
             }
 
             if (ref_latents.size() > 0) {
                 for (ggml_tensor* ref : ref_latents) {
-                    ref = process_img(ctx, ref);
+                    ref = DiT::pad_and_patchify(ctx, ref, patch_size, patch_size);
                     img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                 }
             }
@@ -1178,8 +1115,7 @@ namespace Flux {
                 out = ggml_cont(ctx->ggml_ctx, out);
             }
 
-            // rearrange(out, "b (h w) (c ph pw) -> b c (h ph) (w pw)", h=h_len, w=w_len, ph=2, pw=2)
-            out = unpatchify(ctx->ggml_ctx, out, (H + pad_h) / patch_size, (W + pad_w) / patch_size);  // [N, C, H + pad_h, W + pad_w]
+            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, patch_size, patch_size);  // [N, C, H, W]
             return out;
         }
 

src/ggml_extend.hpp

@@ -767,7 +767,7 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_silu_act(ggml_context* ctx, ggml_tensor*
     return x;
 }
 
-typedef std::function<void(ggml_tensor*, ggml_tensor*, bool)> on_tile_process;
+typedef std::function<bool(ggml_tensor*, ggml_tensor*, bool)> on_tile_process;
 
 __STATIC_INLINE__ void sd_tiling_calc_tiles(int& num_tiles_dim,
                                             float& tile_overlap_factor_dim,
@@ -918,12 +918,15 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
 
             int64_t t1 = ggml_time_ms();
             ggml_ext_tensor_split_2d(input, input_tile, x_in, y_in);
-            on_processing(input_tile, output_tile, false);
-            ggml_ext_tensor_merge_2d(output_tile, output, x_out, y_out, overlap_x_out, overlap_y_out, dx, dy);
+            if (on_processing(input_tile, output_tile, false)) {
+                ggml_ext_tensor_merge_2d(output_tile, output, x_out, y_out, overlap_x_out, overlap_y_out, dx, dy);
 
-            int64_t t2 = ggml_time_ms();
-            last_time  = (t2 - t1) / 1000.0f;
-            pretty_progress(tile_count, num_tiles, last_time);
+                int64_t t2 = ggml_time_ms();
+                last_time  = (t2 - t1) / 1000.0f;
+                pretty_progress(tile_count, num_tiles, last_time);
+            } else {
+                LOG_ERROR("Failed to process patch %d at (%d, %d)", tile_count, x, y);
+            }
             tile_count++;
         }
         last_x = false;

src/latent-preview.h

@@ -1,234 +1,234 @@
-#include <cstddef>
-#include <cstdint>
-#include "ggml.h"
-
-const float wan_21_latent_rgb_proj[16][3] = {
-    {0.015123f, -0.148418f, 0.479828f},
-    {0.003652f, -0.010680f, -0.037142f},
-    {0.212264f, 0.063033f, 0.016779f},
-    {0.232999f, 0.406476f, 0.220125f},
-    {-0.051864f, -0.082384f, -0.069396f},
-    {0.085005f, -0.161492f, 0.010689f},
-    {-0.245369f, -0.506846f, -0.117010f},
-    {-0.151145f, 0.017721f, 0.007207f},
-    {-0.293239f, -0.207936f, -0.421135f},
-    {-0.187721f, 0.050783f, 0.177649f},
-    {-0.013067f, 0.265964f, 0.166578f},
-    {0.028327f, 0.109329f, 0.108642f},
-    {-0.205343f, 0.043991f, 0.148914f},
-    {0.014307f, -0.048647f, -0.007219f},
-    {0.217150f, 0.053074f, 0.319923f},
-    {0.155357f, 0.083156f, 0.064780f}};
-float wan_21_latent_rgb_bias[3] = {-0.270270f, -0.234976f, -0.456853f};
-
-const float wan_22_latent_rgb_proj[48][3] = {
-    {0.017126f, -0.027230f, -0.019257f},
-    {-0.113739f, -0.028715f, -0.022885f},
-    {-0.000106f, 0.021494f, 0.004629f},
-    {-0.013273f, -0.107137f, -0.033638f},
-    {-0.000381f, 0.000279f, 0.025877f},
-    {-0.014216f, -0.003975f, 0.040528f},
-    {0.001638f, -0.000748f, 0.011022f},
-    {0.029238f, -0.006697f, 0.035933f},
-    {0.021641f, -0.015874f, 0.040531f},
-    {-0.101984f, -0.070160f, -0.028855f},
-    {0.033207f, -0.021068f, 0.002663f},
-    {-0.104711f, 0.121673f, 0.102981f},
-    {0.082647f, -0.004991f, 0.057237f},
-    {-0.027375f, 0.031581f, 0.006868f},
-    {-0.045434f, 0.029444f, 0.019287f},
-    {-0.046572f, -0.012537f, 0.006675f},
-    {0.074709f, 0.033690f, 0.025289f},
-    {-0.008251f, -0.002745f, -0.006999f},
-    {0.012685f, -0.061856f, -0.048658f},
-    {0.042304f, -0.007039f, 0.000295f},
-    {-0.007644f, -0.060843f, -0.033142f},
-    {0.159909f, 0.045628f, 0.367541f},
-    {0.095171f, 0.086438f, 0.010271f},
-    {0.006812f, 0.019643f, 0.029637f},
-    {0.003467f, -0.010705f, 0.014252f},
-    {-0.099681f, -0.066272f, -0.006243f},
-    {0.047357f, 0.037040f, 0.000185f},
-    {-0.041797f, -0.089225f, -0.032257f},
-    {0.008928f, 0.017028f, 0.018684f},
-    {-0.042255f, 0.016045f, 0.006849f},
-    {0.011268f, 0.036462f, 0.037387f},
-    {0.011553f, -0.016375f, -0.048589f},
-    {0.046266f, -0.027189f, 0.056979f},
-    {0.009640f, -0.017576f, 0.030324f},
-    {-0.045794f, -0.036083f, -0.010616f},
-    {0.022418f, 0.039783f, -0.032939f},
-    {-0.052714f, -0.015525f, 0.007438f},
-    {0.193004f, 0.223541f, 0.264175f},
-    {-0.059406f, -0.008188f, 0.022867f},
-    {-0.156742f, -0.263791f, -0.007385f},
-    {-0.015717f, 0.016570f, 0.033969f},
-    {0.037969f, 0.109835f, 0.200449f},
-    {-0.000782f, -0.009566f, -0.008058f},
-    {0.010709f, 0.052960f, -0.044195f},
-    {0.017271f, 0.045839f, 0.034569f},
-    {0.009424f, 0.013088f, -0.001714f},
-    {-0.024805f, -0.059378f, -0.033756f},
-    {-0.078293f, 0.029070f, 0.026129f}};
-float wan_22_latent_rgb_bias[3] = {0.013160f, -0.096492f, -0.071323f};
-
-const float flux_latent_rgb_proj[16][3] = {
-    {-0.041168f, 0.019917f, 0.097253f},
-    {0.028096f, 0.026730f, 0.129576f},
-    {0.065618f, -0.067950f, -0.014651f},
-    {-0.012998f, -0.014762f, 0.081251f},
-    {0.078567f, 0.059296f, -0.024687f},
-    {-0.015987f, -0.003697f, 0.005012f},
-    {0.033605f, 0.138999f, 0.068517f},
-    {-0.024450f, -0.063567f, -0.030101f},
-    {-0.040194f, -0.016710f, 0.127185f},
-    {0.112681f, 0.088764f, -0.041940f},
-    {-0.023498f, 0.093664f, 0.025543f},
-    {0.082899f, 0.048320f, 0.007491f},
-    {0.075712f, 0.074139f, 0.081965f},
-    {-0.143501f, 0.018263f, -0.136138f},
-    {-0.025767f, -0.082035f, -0.040023f},
-    {-0.111849f, -0.055589f, -0.032361f}};
-float flux_latent_rgb_bias[3] = {0.024600f, -0.006937f, -0.008089f};
-
-const float flux2_latent_rgb_proj[32][3] = {
-    {0.000736f, -0.008385f, -0.019710f},
-    {-0.001352f, -0.016392f, 0.020693f},
-    {-0.006376f, 0.002428f, 0.036736f},
-    {0.039384f, 0.074167f, 0.119789f},
-    {0.007464f, -0.005705f, -0.004734f},
-    {-0.004086f, 0.005287f, -0.000409f},
-    {-0.032835f, 0.050802f, -0.028120f},
-    {-0.003158f, -0.000835f, 0.000406f},
-    {-0.112840f, -0.084337f, -0.023083f},
-    {0.001462f, -0.006656f, 0.000549f},
-    {-0.009980f, -0.007480f, 0.009702f},
-    {0.032540f, 0.000214f, -0.061388f},
-    {0.011023f, 0.000694f, 0.007143f},
-    {-0.001468f, -0.006723f, -0.001678f},
-    {-0.005921f, -0.010320f, -0.003907f},
-    {-0.028434f, 0.027584f, 0.018457f},
-    {0.014349f, 0.011523f, 0.000441f},
-    {0.009874f, 0.003081f, 0.001507f},
-    {0.002218f, 0.005712f, 0.001563f},
-    {0.053010f, -0.019844f, 0.008683f},
-    {-0.002507f, 0.005384f, 0.000938f},
-    {-0.002177f, -0.011366f, 0.003559f},
-    {-0.000261f, 0.015121f, -0.003240f},
-    {-0.003944f, -0.002083f, 0.005043f},
-    {-0.009138f, 0.011336f, 0.003781f},
-    {0.011429f, 0.003985f, -0.003855f},
-    {0.010518f, -0.005586f, 0.010131f},
-    {0.007883f, 0.002912f, -0.001473f},
-    {-0.003318f, -0.003160f, 0.003684f},
-    {-0.034560f, -0.008740f, 0.012996f},
-    {0.000166f, 0.001079f, -0.012153f},
-    {0.017772f, 0.000937f, -0.011953f}};
-float flux2_latent_rgb_bias[3] = {-0.028738f, -0.098463f, -0.107619f};
-
-// This one was taken straight from
-// https://github.com/Stability-AI/sd3.5/blob/8565799a3b41eb0c7ba976d18375f0f753f56402/sd3_impls.py#L288-L303
-// (MiT Licence)
-const float sd3_latent_rgb_proj[16][3] = {
-    {-0.0645f, 0.0177f, 0.1052f},
-    {0.0028f, 0.0312f, 0.0650f},
-    {0.1848f, 0.0762f, 0.0360f},
-    {0.0944f, 0.0360f, 0.0889f},
-    {0.0897f, 0.0506f, -0.0364f},
-    {-0.0020f, 0.1203f, 0.0284f},
-    {0.0855f, 0.0118f, 0.0283f},
-    {-0.0539f, 0.0658f, 0.1047f},
-    {-0.0057f, 0.0116f, 0.0700f},
-    {-0.0412f, 0.0281f, -0.0039f},
-    {0.1106f, 0.1171f, 0.1220f},
-    {-0.0248f, 0.0682f, -0.0481f},
-    {0.0815f, 0.0846f, 0.1207f},
-    {-0.0120f, -0.0055f, -0.0867f},
-    {-0.0749f, -0.0634f, -0.0456f},
-    {-0.1418f, -0.1457f, -0.1259f},
-};
-float sd3_latent_rgb_bias[3] = {0, 0, 0};
-
-const float sdxl_latent_rgb_proj[4][3] = {
-    {0.258303f, 0.277640f, 0.329699f},
-    {-0.299701f, 0.105446f, 0.014194f},
-    {0.050522f, 0.186163f, -0.143257f},
-    {-0.211938f, -0.149892f, -0.080036f}};
-float sdxl_latent_rgb_bias[3] = {0.144381f, -0.033313f, 0.007061f};
-
-const float sd_latent_rgb_proj[4][3] = {
-    {0.337366f, 0.216344f, 0.257386f},
-    {0.165636f, 0.386828f, 0.046994f},
-    {-0.267803f, 0.237036f, 0.223517f},
-    {-0.178022f, -0.200862f, -0.678514f}};
-float sd_latent_rgb_bias[3] = {-0.017478f, -0.055834f, -0.105825f};
-
-void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
-    size_t buffer_head = 0;
-
-    uint32_t latent_width  = static_cast<uint32_t>(latents->ne[0]);
-    uint32_t latent_height = static_cast<uint32_t>(latents->ne[1]);
-    uint32_t dim           = static_cast<uint32_t>(latents->ne[ggml_n_dims(latents) - 1]);
-    uint32_t frames        = 1;
-    if (ggml_n_dims(latents) == 4) {
-        frames = static_cast<uint32_t>(latents->ne[2]);
-    }
-
-    uint32_t rgb_width  = latent_width * patch_size;
-    uint32_t rgb_height = latent_height * patch_size;
-
-    uint32_t unpatched_dim = dim / (patch_size * patch_size);
-
-    for (uint32_t k = 0; k < frames; k++) {
-        for (uint32_t rgb_x = 0; rgb_x < rgb_width; rgb_x++) {
-            for (uint32_t rgb_y = 0; rgb_y < rgb_height; rgb_y++) {
-                int latent_x = rgb_x / patch_size;
-                int latent_y = rgb_y / patch_size;
-
-                int channel_offset = 0;
-                if (patch_size > 1) {
-                    channel_offset = ((rgb_y % patch_size) * patch_size + (rgb_x % patch_size));
-                }
-
-                size_t latent_id = (latent_x * latents->nb[0] + latent_y * latents->nb[1] + k * latents->nb[2]);
-
-                // should be incremented by 1 for each pixel
-                size_t pixel_id = k * rgb_width * rgb_height + rgb_y * rgb_width + rgb_x;
-
-                float r = 0, g = 0, b = 0;
-                if (latent_rgb_proj != nullptr) {
-                    for (uint32_t d = 0; d < unpatched_dim; d++) {
-                        float value = *(float*)((char*)latents->data + latent_id + (d * patch_size * patch_size + channel_offset) * latents->nb[ggml_n_dims(latents) - 1]);
-                        r += value * latent_rgb_proj[d][0];
-                        g += value * latent_rgb_proj[d][1];
-                        b += value * latent_rgb_proj[d][2];
-                    }
-                } else {
-                    // interpret first 3 channels as RGB
-                    r = *(float*)((char*)latents->data + latent_id + 0 * latents->nb[ggml_n_dims(latents) - 1]);
-                    g = *(float*)((char*)latents->data + latent_id + 1 * latents->nb[ggml_n_dims(latents) - 1]);
-                    b = *(float*)((char*)latents->data + latent_id + 2 * latents->nb[ggml_n_dims(latents) - 1]);
-                }
-                if (latent_rgb_bias != nullptr) {
-                    // bias
-                    r += latent_rgb_bias[0];
-                    g += latent_rgb_bias[1];
-                    b += latent_rgb_bias[2];
-                }
-                // change range
-                r = r * .5f + .5f;
-                g = g * .5f + .5f;
-                b = b * .5f + .5f;
-
-                // clamp rgb values to [0,1] range
-                r = r >= 0 ? r <= 1 ? r : 1 : 0;
-                g = g >= 0 ? g <= 1 ? g : 1 : 0;
-                b = b >= 0 ? b <= 1 ? b : 1 : 0;
-
-                buffer[pixel_id * 3 + 0] = (uint8_t)(r * 255);
-                buffer[pixel_id * 3 + 1] = (uint8_t)(g * 255);
-                buffer[pixel_id * 3 + 2] = (uint8_t)(b * 255);
-            }
-        }
-    }
-}
+#include <cstddef>
+#include <cstdint>
+#include "ggml.h"
+
+const float wan_21_latent_rgb_proj[16][3] = {
+    {0.015123f, -0.148418f, 0.479828f},
+    {0.003652f, -0.010680f, -0.037142f},
+    {0.212264f, 0.063033f, 0.016779f},
+    {0.232999f, 0.406476f, 0.220125f},
+    {-0.051864f, -0.082384f, -0.069396f},
+    {0.085005f, -0.161492f, 0.010689f},
+    {-0.245369f, -0.506846f, -0.117010f},
+    {-0.151145f, 0.017721f, 0.007207f},
+    {-0.293239f, -0.207936f, -0.421135f},
+    {-0.187721f, 0.050783f, 0.177649f},
+    {-0.013067f, 0.265964f, 0.166578f},
+    {0.028327f, 0.109329f, 0.108642f},
+    {-0.205343f, 0.043991f, 0.148914f},
+    {0.014307f, -0.048647f, -0.007219f},
+    {0.217150f, 0.053074f, 0.319923f},
+    {0.155357f, 0.083156f, 0.064780f}};
+float wan_21_latent_rgb_bias[3] = {-0.270270f, -0.234976f, -0.456853f};
+
+const float wan_22_latent_rgb_proj[48][3] = {
+    {0.017126f, -0.027230f, -0.019257f},
+    {-0.113739f, -0.028715f, -0.022885f},
+    {-0.000106f, 0.021494f, 0.004629f},
+    {-0.013273f, -0.107137f, -0.033638f},
+    {-0.000381f, 0.000279f, 0.025877f},
+    {-0.014216f, -0.003975f, 0.040528f},
+    {0.001638f, -0.000748f, 0.011022f},
+    {0.029238f, -0.006697f, 0.035933f},
+    {0.021641f, -0.015874f, 0.040531f},
+    {-0.101984f, -0.070160f, -0.028855f},
+    {0.033207f, -0.021068f, 0.002663f},
+    {-0.104711f, 0.121673f, 0.102981f},
+    {0.082647f, -0.004991f, 0.057237f},
+    {-0.027375f, 0.031581f, 0.006868f},
+    {-0.045434f, 0.029444f, 0.019287f},
+    {-0.046572f, -0.012537f, 0.006675f},
+    {0.074709f, 0.033690f, 0.025289f},
+    {-0.008251f, -0.002745f, -0.006999f},
+    {0.012685f, -0.061856f, -0.048658f},
+    {0.042304f, -0.007039f, 0.000295f},
+    {-0.007644f, -0.060843f, -0.033142f},
+    {0.159909f, 0.045628f, 0.367541f},
+    {0.095171f, 0.086438f, 0.010271f},
+    {0.006812f, 0.019643f, 0.029637f},
+    {0.003467f, -0.010705f, 0.014252f},
+    {-0.099681f, -0.066272f, -0.006243f},
+    {0.047357f, 0.037040f, 0.000185f},
+    {-0.041797f, -0.089225f, -0.032257f},
+    {0.008928f, 0.017028f, 0.018684f},
+    {-0.042255f, 0.016045f, 0.006849f},
+    {0.011268f, 0.036462f, 0.037387f},
+    {0.011553f, -0.016375f, -0.048589f},
+    {0.046266f, -0.027189f, 0.056979f},
+    {0.009640f, -0.017576f, 0.030324f},
+    {-0.045794f, -0.036083f, -0.010616f},
+    {0.022418f, 0.039783f, -0.032939f},
+    {-0.052714f, -0.015525f, 0.007438f},
+    {0.193004f, 0.223541f, 0.264175f},
+    {-0.059406f, -0.008188f, 0.022867f},
+    {-0.156742f, -0.263791f, -0.007385f},
+    {-0.015717f, 0.016570f, 0.033969f},
+    {0.037969f, 0.109835f, 0.200449f},
+    {-0.000782f, -0.009566f, -0.008058f},
+    {0.010709f, 0.052960f, -0.044195f},
+    {0.017271f, 0.045839f, 0.034569f},
+    {0.009424f, 0.013088f, -0.001714f},
+    {-0.024805f, -0.059378f, -0.033756f},
+    {-0.078293f, 0.029070f, 0.026129f}};
+float wan_22_latent_rgb_bias[3] = {0.013160f, -0.096492f, -0.071323f};
+
+const float flux_latent_rgb_proj[16][3] = {
+    {-0.041168f, 0.019917f, 0.097253f},
+    {0.028096f, 0.026730f, 0.129576f},
+    {0.065618f, -0.067950f, -0.014651f},
+    {-0.012998f, -0.014762f, 0.081251f},
+    {0.078567f, 0.059296f, -0.024687f},
+    {-0.015987f, -0.003697f, 0.005012f},
+    {0.033605f, 0.138999f, 0.068517f},
+    {-0.024450f, -0.063567f, -0.030101f},
+    {-0.040194f, -0.016710f, 0.127185f},
+    {0.112681f, 0.088764f, -0.041940f},
+    {-0.023498f, 0.093664f, 0.025543f},
+    {0.082899f, 0.048320f, 0.007491f},
+    {0.075712f, 0.074139f, 0.081965f},
+    {-0.143501f, 0.018263f, -0.136138f},
+    {-0.025767f, -0.082035f, -0.040023f},
+    {-0.111849f, -0.055589f, -0.032361f}};
+float flux_latent_rgb_bias[3] = {0.024600f, -0.006937f, -0.008089f};
+
+const float flux2_latent_rgb_proj[32][3] = {
+    {0.000736f, -0.008385f, -0.019710f},
+    {-0.001352f, -0.016392f, 0.020693f},
+    {-0.006376f, 0.002428f, 0.036736f},
+    {0.039384f, 0.074167f, 0.119789f},
+    {0.007464f, -0.005705f, -0.004734f},
+    {-0.004086f, 0.005287f, -0.000409f},
+    {-0.032835f, 0.050802f, -0.028120f},
+    {-0.003158f, -0.000835f, 0.000406f},
+    {-0.112840f, -0.084337f, -0.023083f},
+    {0.001462f, -0.006656f, 0.000549f},
+    {-0.009980f, -0.007480f, 0.009702f},
+    {0.032540f, 0.000214f, -0.061388f},
+    {0.011023f, 0.000694f, 0.007143f},
+    {-0.001468f, -0.006723f, -0.001678f},
+    {-0.005921f, -0.010320f, -0.003907f},
+    {-0.028434f, 0.027584f, 0.018457f},
+    {0.014349f, 0.011523f, 0.000441f},
+    {0.009874f, 0.003081f, 0.001507f},
+    {0.002218f, 0.005712f, 0.001563f},
+    {0.053010f, -0.019844f, 0.008683f},
+    {-0.002507f, 0.005384f, 0.000938f},
+    {-0.002177f, -0.011366f, 0.003559f},
+    {-0.000261f, 0.015121f, -0.003240f},
+    {-0.003944f, -0.002083f, 0.005043f},
+    {-0.009138f, 0.011336f, 0.003781f},
+    {0.011429f, 0.003985f, -0.003855f},
+    {0.010518f, -0.005586f, 0.010131f},
+    {0.007883f, 0.002912f, -0.001473f},
+    {-0.003318f, -0.003160f, 0.003684f},
+    {-0.034560f, -0.008740f, 0.012996f},
+    {0.000166f, 0.001079f, -0.012153f},
+    {0.017772f, 0.000937f, -0.011953f}};
+float flux2_latent_rgb_bias[3] = {-0.028738f, -0.098463f, -0.107619f};
+
+// This one was taken straight from
+// https://github.com/Stability-AI/sd3.5/blob/8565799a3b41eb0c7ba976d18375f0f753f56402/sd3_impls.py#L288-L303
+// (MiT Licence)
+const float sd3_latent_rgb_proj[16][3] = {
+    {-0.0645f, 0.0177f, 0.1052f},
+    {0.0028f, 0.0312f, 0.0650f},
+    {0.1848f, 0.0762f, 0.0360f},
+    {0.0944f, 0.0360f, 0.0889f},
+    {0.0897f, 0.0506f, -0.0364f},
+    {-0.0020f, 0.1203f, 0.0284f},
+    {0.0855f, 0.0118f, 0.0283f},
+    {-0.0539f, 0.0658f, 0.1047f},
+    {-0.0057f, 0.0116f, 0.0700f},
+    {-0.0412f, 0.0281f, -0.0039f},
+    {0.1106f, 0.1171f, 0.1220f},
+    {-0.0248f, 0.0682f, -0.0481f},
+    {0.0815f, 0.0846f, 0.1207f},
+    {-0.0120f, -0.0055f, -0.0867f},
+    {-0.0749f, -0.0634f, -0.0456f},
+    {-0.1418f, -0.1457f, -0.1259f},
+};
+float sd3_latent_rgb_bias[3] = {0, 0, 0};
+
+const float sdxl_latent_rgb_proj[4][3] = {
+    {0.258303f, 0.277640f, 0.329699f},
+    {-0.299701f, 0.105446f, 0.014194f},
+    {0.050522f, 0.186163f, -0.143257f},
+    {-0.211938f, -0.149892f, -0.080036f}};
+float sdxl_latent_rgb_bias[3] = {0.144381f, -0.033313f, 0.007061f};
+
+const float sd_latent_rgb_proj[4][3] = {
+    {0.337366f, 0.216344f, 0.257386f},
+    {0.165636f, 0.386828f, 0.046994f},
+    {-0.267803f, 0.237036f, 0.223517f},
+    {-0.178022f, -0.200862f, -0.678514f}};
+float sd_latent_rgb_bias[3] = {-0.017478f, -0.055834f, -0.105825f};
+
+void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
+    size_t buffer_head = 0;
+
+    uint32_t latent_width  = static_cast<uint32_t>(latents->ne[0]);
+    uint32_t latent_height = static_cast<uint32_t>(latents->ne[1]);
+    uint32_t dim           = static_cast<uint32_t>(latents->ne[ggml_n_dims(latents) - 1]);
+    uint32_t frames        = 1;
+    if (ggml_n_dims(latents) == 4) {
+        frames = static_cast<uint32_t>(latents->ne[2]);
+    }
+
+    uint32_t rgb_width  = latent_width * patch_size;
+    uint32_t rgb_height = latent_height * patch_size;
+
+    uint32_t unpatched_dim = dim / (patch_size * patch_size);
+
+    for (uint32_t k = 0; k < frames; k++) {
+        for (uint32_t rgb_x = 0; rgb_x < rgb_width; rgb_x++) {
+            for (uint32_t rgb_y = 0; rgb_y < rgb_height; rgb_y++) {
+                int latent_x = rgb_x / patch_size;
+                int latent_y = rgb_y / patch_size;
+
+                int channel_offset = 0;
+                if (patch_size > 1) {
+                    channel_offset = ((rgb_y % patch_size) * patch_size + (rgb_x % patch_size));
+                }
+
+                size_t latent_id = (latent_x * latents->nb[0] + latent_y * latents->nb[1] + k * latents->nb[2]);
+
+                // should be incremented by 1 for each pixel
+                size_t pixel_id = k * rgb_width * rgb_height + rgb_y * rgb_width + rgb_x;
+
+                float r = 0, g = 0, b = 0;
+                if (latent_rgb_proj != nullptr) {
+                    for (uint32_t d = 0; d < unpatched_dim; d++) {
+                        float value = *(float*)((char*)latents->data + latent_id + (d * patch_size * patch_size + channel_offset) * latents->nb[ggml_n_dims(latents) - 1]);
+                        r += value * latent_rgb_proj[d][0];
+                        g += value * latent_rgb_proj[d][1];
+                        b += value * latent_rgb_proj[d][2];
+                    }
+                } else {
+                    // interpret first 3 channels as RGB
+                    r = *(float*)((char*)latents->data + latent_id + 0 * latents->nb[ggml_n_dims(latents) - 1]);
+                    g = *(float*)((char*)latents->data + latent_id + 1 * latents->nb[ggml_n_dims(latents) - 1]);
+                    b = *(float*)((char*)latents->data + latent_id + 2 * latents->nb[ggml_n_dims(latents) - 1]);
+                }
+                if (latent_rgb_bias != nullptr) {
+                    // bias
+                    r += latent_rgb_bias[0];
+                    g += latent_rgb_bias[1];
+                    b += latent_rgb_bias[2];
+                }
+                // change range
+                r = r * .5f + .5f;
+                g = g * .5f + .5f;
+                b = b * .5f + .5f;
+
+                // clamp rgb values to [0,1] range
+                r = r >= 0 ? r <= 1 ? r : 1 : 0;
+                g = g >= 0 ? g <= 1 ? g : 1 : 0;
+                b = b >= 0 ? b <= 1 ? b : 1 : 0;
+
+                buffer[pixel_id * 3 + 0] = (uint8_t)(r * 255);
+                buffer[pixel_id * 3 + 1] = (uint8_t)(g * 255);
+                buffer[pixel_id * 3 + 2] = (uint8_t)(b * 255);
+            }
+        }
+    }
+}

src/llm.hpp

@@ -19,6 +19,7 @@
 #include "json.hpp"
 #include "rope.hpp"
 #include "tokenize_util.h"
+#include "vocab/vocab.h"
 
 namespace LLM {
     constexpr int LLM_GRAPH_SIZE = 10240;
@@ -365,7 +366,7 @@ namespace LLM {
             if (merges_utf8_str.size() > 0) {
                 load_from_merges(merges_utf8_str);
             } else {
-                load_from_merges(ModelLoader::load_qwen2_merges());
+                load_from_merges(load_qwen2_merges());
             }
         }
     };
@@ -466,7 +467,7 @@ namespace LLM {
             if (merges_utf8_str.size() > 0 && vocab_utf8_str.size() > 0) {
                 load_from_merges(merges_utf8_str, vocab_utf8_str);
             } else {
-                load_from_merges(ModelLoader::load_mistral_merges(), ModelLoader::load_mistral_vocab_json());
+                load_from_merges(load_mistral_merges(), load_mistral_vocab_json());
             }
         }
     };

src/ltxv.hpp

@@ -1,8 +1,7 @@
 #ifndef __LTXV_HPP__
 #define __LTXV_HPP__
 
-#include "common.hpp"
-#include "ggml_extend.hpp"
+#include "common_block.hpp"
 
 namespace LTXV {
 

src/mmdit.hpp

@@ -745,28 +745,6 @@ public:
         return spatial_pos_embed;
     }
 
-    struct ggml_tensor* unpatchify(struct ggml_context* ctx,
-                                   struct ggml_tensor* x,
-                                   int64_t h,
-                                   int64_t w) {
-        // x: [N, H*W, patch_size * patch_size * C]
-        // return: [N, C, H, W]
-        int64_t n = x->ne[2];
-        int64_t c = out_channels;
-        int64_t p = patch_size;
-        h         = (h + 1) / p;
-        w         = (w + 1) / p;
-
-        GGML_ASSERT(h * w == x->ne[1]);
-
-        x = ggml_reshape_4d(ctx, x, c, p * p, w * h, n);       // [N, H*W, P*P, C]
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 2, 0, 1, 3));  // [N, C, H*W, P*P]
-        x = ggml_reshape_4d(ctx, x, p, p, w, h * c * n);       // [N*C*H, W, P, P]
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*H, P, W, P]
-        x = ggml_reshape_4d(ctx, x, p * w, p * h, c, n);       // [N, C, H*P, W*P]
-        return x;
-    }
-
     struct ggml_tensor* forward_core_with_concat(GGMLRunnerContext* ctx,
                                                  struct ggml_tensor* x,
                                                  struct ggml_tensor* c_mod,
@@ -811,11 +789,11 @@ public:
         auto x_embedder = std::dynamic_pointer_cast<PatchEmbed>(blocks["x_embedder"]);
         auto t_embedder = std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder"]);
 
-        int64_t w = x->ne[0];
-        int64_t h = x->ne[1];
+        int64_t W = x->ne[0];
+        int64_t H = x->ne[1];
 
         auto patch_embed = x_embedder->forward(ctx, x);                      // [N, H*W, hidden_size]
-        auto pos_embed   = cropped_pos_embed(ctx->ggml_ctx, h, w);           // [1, H*W, hidden_size]
+        auto pos_embed   = cropped_pos_embed(ctx->ggml_ctx, H, W);           // [1, H*W, hidden_size]
         x                = ggml_add(ctx->ggml_ctx, patch_embed, pos_embed);  // [N, H*W, hidden_size]
 
         auto c = t_embedder->forward(ctx, t);  // [N, hidden_size]
@@ -834,7 +812,7 @@ public:
 
         x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 
-        x = unpatchify(ctx->ggml_ctx, x, h, w);  // [N, C, H, W]
+        x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, patch_size, patch_size, /*patch_last*/ false);  // [N, C, H, W]
 
         return x;
     }

src/model.cpp

@@ -16,10 +16,6 @@
 #include "model.h"
 #include "stable-diffusion.h"
 #include "util.h"
-#include "vocab.hpp"
-#include "vocab_mistral.hpp"
-#include "vocab_qwen.hpp"
-#include "vocab_umt5.hpp"
 
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
@@ -1061,6 +1057,9 @@ SDVersion ModelLoader::get_sd_version() {
             if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
                 return VERSION_QWEN_IMAGE;
             }
+            if (tensor_storage.name.find("llm_adapter.blocks.0.cross_attn.q_proj.weight") != std::string::npos) {
+                return VERSION_ANIMA;
+            }
             if (tensor_storage.name.find("model.diffusion_model.double_stream_modulation_img.lin.weight") != std::string::npos) {
                 is_flux2 = true;
             }
@@ -1340,36 +1339,6 @@ void ModelLoader::set_wtype_override(ggml_type wtype, std::string tensor_type_ru
     }
 }
 
-std::string ModelLoader::load_merges() {
-    std::string merges_utf8_str(reinterpret_cast<const char*>(merges_utf8_c_str), sizeof(merges_utf8_c_str));
-    return merges_utf8_str;
-}
-
-std::string ModelLoader::load_qwen2_merges() {
-    std::string merges_utf8_str(reinterpret_cast<const char*>(qwen2_merges_utf8_c_str), sizeof(qwen2_merges_utf8_c_str));
-    return merges_utf8_str;
-}
-
-std::string ModelLoader::load_mistral_merges() {
-    std::string merges_utf8_str(reinterpret_cast<const char*>(mistral_merges_utf8_c_str), sizeof(mistral_merges_utf8_c_str));
-    return merges_utf8_str;
-}
-
-std::string ModelLoader::load_mistral_vocab_json() {
-    std::string json_str(reinterpret_cast<const char*>(mistral_vocab_json_utf8_c_str), sizeof(mistral_vocab_json_utf8_c_str));
-    return json_str;
-}
-
-std::string ModelLoader::load_t5_tokenizer_json() {
-    std::string json_str(reinterpret_cast<const char*>(t5_tokenizer_json_str), sizeof(t5_tokenizer_json_str));
-    return json_str;
-}
-
-std::string ModelLoader::load_umt5_tokenizer_json() {
-    std::string json_str(reinterpret_cast<const char*>(umt5_tokenizer_json_str), sizeof(umt5_tokenizer_json_str));
-    return json_str;
-}
-
 bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p, bool enable_mmap) {
     int64_t process_time_ms = 0;
     std::atomic<int64_t> read_time_ms(0);

src/model.h

@@ -45,6 +45,7 @@ enum SDVersion {
     VERSION_WAN2_2_I2V,
     VERSION_WAN2_2_TI2V,
     VERSION_QWEN_IMAGE,
+    VERSION_ANIMA,
     VERSION_FLUX2,
     VERSION_FLUX2_KLEIN,
     VERSION_Z_IMAGE,
@@ -122,6 +123,13 @@ static inline bool sd_version_is_qwen_image(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_anima(SDVersion version) {
+    if (version == VERSION_ANIMA) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_z_image(SDVersion version) {
     if (version == VERSION_Z_IMAGE) {
         return true;
@@ -146,6 +154,7 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_sd3(version) ||
         sd_version_is_wan(version) ||
         sd_version_is_qwen_image(version) ||
+        sd_version_is_anima(version) ||
         sd_version_is_z_image(version)) {
         return true;
     }
@@ -331,13 +340,6 @@ public:
     bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
     int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
     ~ModelLoader() = default;
-
-    static std::string load_merges();
-    static std::string load_qwen2_merges();
-    static std::string load_mistral_merges();
-    static std::string load_mistral_vocab_json();
-    static std::string load_t5_tokenizer_json();
-    static std::string load_umt5_tokenizer_json();
 };
 
 #endif  // __MODEL_H__

src/name_conversion.cpp

@@ -653,6 +653,14 @@ std::string convert_diffusers_dit_to_original_lumina2(std::string name) {
     return name;
 }
 
+std::string convert_other_dit_to_original_anima(std::string name) {
+    static const std::string anima_net_prefix = "net.";
+    if (!starts_with(name, anima_net_prefix)) {
+        name = anima_net_prefix + name;
+    }
+    return name;
+}
+
 std::string convert_diffusion_model_name(std::string name, std::string prefix, SDVersion version) {
     if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
         name = convert_diffusers_unet_to_original_sd1(name);
@@ -664,6 +672,8 @@ std::string convert_diffusion_model_name(std::string name, std::string prefix, S
         name = convert_diffusers_dit_to_original_flux(name);
     } else if (sd_version_is_z_image(version)) {
         name = convert_diffusers_dit_to_original_lumina2(name);
+    } else if (sd_version_is_anima(version)) {
+        name = convert_other_dit_to_original_anima(name);
     }
     return name;
 }

src/qwen_image.hpp

@@ -3,9 +3,8 @@
 
 #include <memory>
 
-#include "common.hpp"
+#include "common_block.hpp"
 #include "flux.hpp"
-#include "ggml_extend.hpp"
 
 namespace Qwen {
     constexpr int QWEN_IMAGE_GRAPH_SIZE = 20480;
@@ -390,69 +389,6 @@ namespace Qwen {
             blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, params.patch_size * params.patch_size * params.out_channels));
         }
 
-        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
-                                              struct ggml_tensor* x) {
-            int64_t W = x->ne[0];
-            int64_t H = x->ne[1];
-
-            int pad_h = (params.patch_size - H % params.patch_size) % params.patch_size;
-            int pad_w = (params.patch_size - W % params.patch_size) % params.patch_size;
-            x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
-            return x;
-        }
-
-        struct ggml_tensor* patchify(struct ggml_context* ctx,
-                                     struct ggml_tensor* x) {
-            // x: [N, C, H, W]
-            // return: [N, h*w, C * patch_size * patch_size]
-            int64_t N = x->ne[3];
-            int64_t C = x->ne[2];
-            int64_t H = x->ne[1];
-            int64_t W = x->ne[0];
-            int64_t p = params.patch_size;
-            int64_t h = H / params.patch_size;
-            int64_t w = W / params.patch_size;
-
-            GGML_ASSERT(h * p == H && w * p == W);
-
-            x = ggml_reshape_4d(ctx, x, p, w, p, h * C * N);       // [N*C*h, p, w, p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, w, p, p]
-            x = ggml_reshape_4d(ctx, x, p * p, w * h, C, N);       // [N, C, h*w, p*p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, h*w, C, p*p]
-            x = ggml_reshape_3d(ctx, x, p * p * C, w * h, N);      // [N, h*w, C*p*p]
-            return x;
-        }
-
-        struct ggml_tensor* process_img(GGMLRunnerContext* ctx,
-                                        struct ggml_tensor* x) {
-            x = pad_to_patch_size(ctx, x);
-            x = patchify(ctx->ggml_ctx, x);
-            return x;
-        }
-
-        struct ggml_tensor* unpatchify(struct ggml_context* ctx,
-                                       struct ggml_tensor* x,
-                                       int64_t h,
-                                       int64_t w) {
-            // x: [N, h*w, C*patch_size*patch_size]
-            // return: [N, C, H, W]
-            int64_t N = x->ne[2];
-            int64_t C = x->ne[0] / params.patch_size / params.patch_size;
-            int64_t H = h * params.patch_size;
-            int64_t W = w * params.patch_size;
-            int64_t p = params.patch_size;
-
-            GGML_ASSERT(C * p * p == x->ne[0]);
-
-            x = ggml_reshape_4d(ctx, x, p * p, C, w * h, N);       // [N, h*w, C, p*p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, C, h*w, p*p]
-            x = ggml_reshape_4d(ctx, x, p, p, w, h * C * N);       // [N*C*h, w, p, p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, p, w, p]
-            x = ggml_reshape_4d(ctx, x, W, H, C, N);               // [N, C, h*p, w*p]
-
-            return x;
-        }
-
         struct ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
                                          struct ggml_tensor* x,
                                          struct ggml_tensor* timestep,
@@ -512,19 +448,16 @@ namespace Qwen {
             int64_t C = x->ne[2];
             int64_t N = x->ne[3];
 
-            auto img           = process_img(ctx, x);
+            auto img           = DiT::pad_and_patchify(ctx, x, params.patch_size, params.patch_size);
             int64_t img_tokens = img->ne[1];
 
             if (ref_latents.size() > 0) {
                 for (ggml_tensor* ref : ref_latents) {
-                    ref = process_img(ctx, ref);
+                    ref = DiT::pad_and_patchify(ctx, ref, params.patch_size, params.patch_size);
                     img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                 }
             }
 
-            int64_t h_len = ((H + (params.patch_size / 2)) / params.patch_size);
-            int64_t w_len = ((W + (params.patch_size / 2)) / params.patch_size);
-
             auto out = forward_orig(ctx, img, timestep, context, pe, modulate_index);  // [N, h_len*w_len, ph*pw*C]
 
             if (out->ne[1] > img_tokens) {
@@ -533,11 +466,7 @@ namespace Qwen {
                 out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, out, 0, 2, 1, 3));  // [N, h*w, C * patch_size * patch_size]
             }
 
-            out = unpatchify(ctx->ggml_ctx, out, h_len, w_len);  // [N, C, H + pad_h, W + pad_w]
-
-            // slice
-            out = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, H);  // [N, C, H, W + pad_w]
-            out = ggml_ext_slice(ctx->ggml_ctx, out, 0, 0, W);  // [N, C, H, W]
+            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, params.patch_size, params.patch_size);  // [N, C, H, W]
 
             return out;
         }

src/rope.hpp

@@ -43,7 +43,7 @@ namespace Rope {
 
     __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos,
                                                            int dim,
-                                                           int theta,
+                                                           float theta,
                                                            const std::vector<int>& axis_wrap_dims = {}) {
         assert(dim % 2 == 0);
         int half_dim = dim / 2;
@@ -167,7 +167,7 @@ namespace Rope {
 
     __STATIC_INLINE__ std::vector<float> embed_nd(const std::vector<std::vector<float>>& ids,
                                                   int bs,
-                                                  int theta,
+                                                  const std::vector<float>& axis_thetas,
                                                   const std::vector<int>& axes_dim,
                                                   const std::vector<std::vector<int>>& wrap_dims = {}) {
         std::vector<std::vector<float>> trans_ids = transpose(ids);
@@ -188,8 +188,12 @@ namespace Rope {
             if (!wrap_dims.empty() && i < (int)wrap_dims.size()) {
                 axis_wrap_dims = wrap_dims[i];
             }
+            float axis_theta = 10000.0f;
+            if (!axis_thetas.empty()) {
+                axis_theta = axis_thetas[std::min(i, axis_thetas.size() - 1)];
+            }
             std::vector<std::vector<float>> rope_emb =
-                rope(trans_ids[i], axes_dim[i], theta, axis_wrap_dims);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
+                rope(trans_ids[i], axes_dim[i], axis_theta, axis_wrap_dims);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
             for (int b = 0; b < bs; ++b) {
                 for (int j = 0; j < pos_len; ++j) {
                     for (int k = 0; k < rope_emb[0].size(); ++k) {
@@ -203,6 +207,15 @@ namespace Rope {
         return flatten(emb);
     }
 
+    __STATIC_INLINE__ std::vector<float> embed_nd(const std::vector<std::vector<float>>& ids,
+                                                  int bs,
+                                                  float theta,
+                                                  const std::vector<int>& axes_dim,
+                                                  const std::vector<std::vector<int>>& wrap_dims = {}) {
+        std::vector<float> axis_thetas(axes_dim.size(), theta);
+        return embed_nd(ids, bs, axis_thetas, axes_dim, wrap_dims);
+    }
+
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_refs_ids(int patch_size,
                                                                    int bs,
                                                                    int axes_dim_num,
@@ -332,7 +345,7 @@ namespace Rope {
                 }
             }
         }
-        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int h,
@@ -421,7 +434,7 @@ namespace Rope {
                 }
             }
         }
-        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
@@ -475,7 +488,7 @@ namespace Rope {
                                                     int theta,
                                                     const std::vector<int>& axes_dim) {
         std::vector<std::vector<float>> ids = gen_vid_ids(t, h, w, pt, ph, pw, bs);
-        return embed_nd(ids, bs, theta, axes_dim);
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen2vl_ids(int grid_h,
@@ -511,7 +524,7 @@ namespace Rope {
                                                         int theta,
                                                         const std::vector<int>& axes_dim) {
         std::vector<std::vector<float>> ids = gen_qwen2vl_ids(grid_h, grid_w, merge_size, window_index);
-        return embed_nd(ids, 1, theta, axes_dim);
+        return embed_nd(ids, 1, static_cast<float>(theta), axes_dim);
     }
 
     __STATIC_INLINE__ int bound_mod(int a, int m) {
@@ -584,7 +597,7 @@ namespace Rope {
             }
         }
 
-        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ struct ggml_tensor* apply_rope(struct ggml_context* ctx,

src/stable-diffusion.cpp

@@ -48,6 +48,7 @@ const char* model_version_to_str[] = {
     "Wan 2.2 I2V",
     "Wan 2.2 TI2V",
     "Qwen Image",
+    "Anima",
     "Flux.2",
     "Flux.2 klein",
     "Z-Image",
@@ -107,6 +108,7 @@ public:
 
     SDVersion version;
     bool vae_decode_only         = false;
+    bool external_vae_is_invalid = false;
     bool free_params_immediately = false;
 
     std::shared_ptr<RNG> rng         = std::make_shared<PhiloxRNG>();
@@ -114,6 +116,7 @@ public:
     int n_threads                    = -1;
     float scale_factor               = 0.18215f;
     float shift_factor               = 0.f;
+    float default_flow_shift         = INFINITY;
 
     std::shared_ptr<Conditioner> cond_stage_model;
     std::shared_ptr<FrozenCLIPVisionEmbedder> clip_vision;  // for svd or wan2.1 i2v
@@ -321,6 +324,7 @@ public:
             LOG_INFO("loading vae from '%s'", sd_ctx_params->vae_path);
             if (!model_loader.init_from_file(sd_ctx_params->vae_path, "vae.")) {
                 LOG_WARN("loading vae from '%s' failed", sd_ctx_params->vae_path);
+                external_vae_is_invalid = true;
             }
         }
 
@@ -402,6 +406,7 @@ public:
             shift_factor = 0.1159f;
         } else if (sd_version_is_wan(version) ||
                    sd_version_is_qwen_image(version) ||
+                   sd_version_is_anima(version) ||
                    sd_version_is_flux2(version)) {
             scale_factor = 1.0f;
             shift_factor = 0.f;
@@ -532,6 +537,14 @@ public:
                                                                    "model.diffusion_model",
                                                                    version,
                                                                    sd_ctx_params->qwen_image_zero_cond_t);
+            } else if (sd_version_is_anima(version)) {
+                cond_stage_model = std::make_shared<AnimaConditioner>(clip_backend,
+                                                                      offload_params_to_cpu,
+                                                                      tensor_storage_map);
+                diffusion_model  = std::make_shared<AnimaModel>(backend,
+                                                               offload_params_to_cpu,
+                                                               tensor_storage_map,
+                                                               "model.diffusion_model");
             } else if (sd_version_is_z_image(version)) {
                 cond_stage_model = std::make_shared<LLMEmbedder>(clip_backend,
                                                                  offload_params_to_cpu,
@@ -594,7 +607,7 @@ public:
             }
 
             if (!(use_tiny_autoencoder || version == VERSION_SDXS) || tae_preview_only) {
-                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
+                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
                     first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
                                                                             offload_params_to_cpu,
                                                                             tensor_storage_map,
@@ -619,10 +632,10 @@ public:
                         first_stage_model->set_conv2d_direct_enabled(true);
                     }
                     if (sd_version_is_sdxl(version) &&
-                        (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 || sd_ctx_params->force_sdxl_vae_conv_scale)) {
+                        (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 || sd_ctx_params->force_sdxl_vae_conv_scale || external_vae_is_invalid)) {
                         float vae_conv_2d_scale = 1.f / 32.f;
                         LOG_WARN(
-                            "No VAE specified with --vae or --force-sdxl-vae-conv-scale flag set, "
+                            "No valid VAE specified with --vae or --force-sdxl-vae-conv-scale flag set, "
                             "using Conv2D scale %.3f",
                             vae_conv_2d_scale);
                         first_stage_model->set_conv2d_scale(vae_conv_2d_scale);
@@ -632,7 +645,7 @@ public:
                 }
             }
             if (use_tiny_autoencoder || version == VERSION_SDXS) {
-                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
+                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
                     tae_first_stage = std::make_shared<TinyVideoAutoEncoder>(vae_backend,
                                                                              offload_params_to_cpu,
                                                                              tensor_storage_map,
@@ -879,7 +892,6 @@ public:
         // init denoiser
         {
             prediction_t pred_type = sd_ctx_params->prediction;
-            float flow_shift       = sd_ctx_params->flow_shift;
 
             if (pred_type == PREDICTION_COUNT) {
                 if (sd_version_is_sd2(version)) {
@@ -902,24 +914,22 @@ public:
                 } else if (sd_version_is_sd3(version) ||
                            sd_version_is_wan(version) ||
                            sd_version_is_qwen_image(version) ||
+                           sd_version_is_anima(version) ||
                            sd_version_is_z_image(version)) {
                     pred_type = FLOW_PRED;
-                    if (flow_shift == INFINITY) {
-                        if (sd_version_is_wan(version)) {
-                            flow_shift = 5.f;
-                        } else {
-                            flow_shift = 3.f;
-                        }
+                    if (sd_version_is_wan(version)) {
+                        default_flow_shift = 5.f;
+                    } else {
+                        default_flow_shift = 3.f;
                     }
                 } else if (sd_version_is_flux(version)) {
                     pred_type = FLUX_FLOW_PRED;
 
-                    if (flow_shift == INFINITY) {
-                        flow_shift = 1.0f;  // TODO: validate
-                        for (const auto& [name, tensor_storage] : tensor_storage_map) {
-                            if (starts_with(name, "model.diffusion_model.guidance_in.in_layer.weight")) {
-                                flow_shift = 1.15f;
-                            }
+                    default_flow_shift = 1.0f;  // TODO: validate
+                    for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                        if (starts_with(name, "model.diffusion_model.guidance_in.in_layer.weight")) {
+                            default_flow_shift = 1.15f;
+                            break;
                         }
                     }
                 } else if (sd_version_is_flux2(version)) {
@@ -943,12 +953,12 @@ public:
                     break;
                 case FLOW_PRED: {
                     LOG_INFO("running in FLOW mode");
-                    denoiser = std::make_shared<DiscreteFlowDenoiser>(flow_shift);
+                    denoiser = std::make_shared<DiscreteFlowDenoiser>();
                     break;
                 }
                 case FLUX_FLOW_PRED: {
                     LOG_INFO("running in Flux FLOW mode");
-                    denoiser = std::make_shared<FluxFlowDenoiser>(flow_shift);
+                    denoiser = std::make_shared<FluxFlowDenoiser>();
                     break;
                 }
                 case FLUX2_FLOW_PRED: {
@@ -1088,6 +1098,18 @@ public:
         cond_stage_lora_models.clear();
         diffusion_lora_models.clear();
         first_stage_lora_models.clear();
+        if (cond_stage_model) {
+            cond_stage_model->set_weight_adapter(nullptr);
+        }
+        if (diffusion_model) {
+            diffusion_model->set_weight_adapter(nullptr);
+        }
+        if (high_noise_diffusion_model) {
+            high_noise_diffusion_model->set_weight_adapter(nullptr);
+        }
+        if (first_stage_model) {
+            first_stage_model->set_weight_adapter(nullptr);
+        }
         if (lora_state.empty()) {
             return;
         }
@@ -1504,7 +1526,7 @@ public:
                 } else if (sd_version_is_flux(version) || sd_version_is_z_image(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)) {
+                } else if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
                     latent_rgb_proj = wan_21_latent_rgb_proj;
                     latent_rgb_bias = wan_21_latent_rgb_bias;
                 } else {
@@ -1558,7 +1580,7 @@ public:
                 if (vae_tiling_params.enabled) {
                     // split latent in 32x32 tiles and compute in several steps
                     auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                        first_stage_model->compute(n_threads, in, true, &out, nullptr);
+                        return first_stage_model->compute(n_threads, in, true, &out, nullptr);
                     };
                     silent_tiling(latents, result, get_vae_scale_factor(), 32, 0.5f, on_tiling);
 
@@ -1577,7 +1599,7 @@ public:
                 if (vae_tiling_params.enabled) {
                     // split latent in 64x64 tiles and compute in several steps
                     auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                        tae_first_stage->compute(n_threads, in, true, &out, nullptr);
+                        return tae_first_stage->compute(n_threads, in, true, &out, nullptr);
                     };
                     silent_tiling(latents, result, get_vae_scale_factor(), 64, 0.5f, on_tiling);
                 } else {
@@ -1985,6 +2007,9 @@ public:
                 shifted_t             = std::max((int64_t)0, std::min((int64_t)(TIMESTEPS - 1), shifted_t));
                 LOG_DEBUG("shifting timestep from %.2f to %" PRId64 " (sigma: %.4f)", t, shifted_t, sigma);
                 timesteps_vec.assign(1, (float)shifted_t);
+            } else if (sd_version_is_anima(version)) {
+                // Anima uses normalized flow timesteps.
+                timesteps_vec.assign(1, t / static_cast<float>(TIMESTEPS));
             } else if (sd_version_is_z_image(version)) {
                 timesteps_vec.assign(1, 1000.f - t);
             } else {
@@ -2396,7 +2421,7 @@ public:
     }
 
     void process_latent_in(ggml_tensor* latent) {
-        if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_flux2(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version) || sd_version_is_flux2(version)) {
             int channel_dim = sd_version_is_flux2(version) ? 2 : 3;
             std::vector<float> latents_mean_vec;
             std::vector<float> latents_std_vec;
@@ -2435,7 +2460,7 @@ public:
     }
 
     void process_latent_out(ggml_tensor* latent) {
-        if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_flux2(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version) || sd_version_is_flux2(version)) {
             int channel_dim = sd_version_is_flux2(version) ? 2 : 3;
             std::vector<float> latents_mean_vec;
             std::vector<float> latents_std_vec;
@@ -2513,7 +2538,7 @@ public:
             // TODO wan2.2 vae support?
             int64_t ne2;
             int64_t ne3;
-            if (sd_version_is_qwen_image(version)) {
+            if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
                 ne2 = 1;
                 ne3 = C * x->ne[3];
             } else {
@@ -2531,7 +2556,7 @@ public:
             result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, ne2, ne3);
         }
 
-        if (sd_version_is_qwen_image(version)) {
+        if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
             x = ggml_reshape_4d(work_ctx, x, x->ne[0], x->ne[1], 1, x->ne[2] * x->ne[3]);
         }
 
@@ -2546,7 +2571,7 @@ public:
                 LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
 
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                    first_stage_model->compute(n_threads, in, false, &out, work_ctx);
+                    return first_stage_model->compute(n_threads, in, false, &out, work_ctx);
                 };
                 sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, on_tiling);
             } else {
@@ -2557,7 +2582,7 @@ public:
             if (vae_tiling_params.enabled && !encode_video) {
                 // split latent in 32x32 tiles and compute in several steps
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                    tae_first_stage->compute(n_threads, in, false, &out, nullptr);
+                    return tae_first_stage->compute(n_threads, in, false, &out, nullptr);
                 };
                 sd_tiling(x, result, vae_scale_factor, 64, 0.5f, on_tiling);
             } else {
@@ -2604,6 +2629,7 @@ public:
         ggml_tensor* latent;
         if (use_tiny_autoencoder ||
             sd_version_is_qwen_image(version) ||
+            sd_version_is_anima(version) ||
             sd_version_is_wan(version) ||
             sd_version_is_flux2(version) ||
             version == VERSION_CHROMA_RADIANCE) {
@@ -2623,7 +2649,7 @@ public:
         if (!use_tiny_autoencoder) {
             process_latent_in(latent);
         }
-        if (sd_version_is_qwen_image(version)) {
+        if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
             latent = ggml_reshape_4d(work_ctx, latent, latent->ne[0], latent->ne[1], latent->ne[3], 1);
         }
         return latent;
@@ -2661,7 +2687,7 @@ public:
         }
         int64_t t0 = ggml_time_ms();
         if (!use_tiny_autoencoder) {
-            if (sd_version_is_qwen_image(version)) {
+            if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
                 x = ggml_reshape_4d(work_ctx, x, x->ne[0], x->ne[1], 1, x->ne[2] * x->ne[3]);
             }
             process_latent_out(x);
@@ -2675,11 +2701,15 @@ public:
 
                 // split latent in 32x32 tiles and compute in several steps
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                    first_stage_model->compute(n_threads, in, true, &out, nullptr);
+                    return first_stage_model->compute(n_threads, in, true, &out, nullptr);
                 };
                 sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, on_tiling);
             } else {
-                first_stage_model->compute(n_threads, x, true, &result, work_ctx);
+                if (!first_stage_model->compute(n_threads, x, true, &result, work_ctx)) {
+                    LOG_ERROR("Failed to decode latetnts");
+                    first_stage_model->free_compute_buffer();
+                    return nullptr;
+                }
             }
             first_stage_model->free_compute_buffer();
             process_vae_output_tensor(result);
@@ -2687,11 +2717,15 @@ public:
             if (vae_tiling_params.enabled) {
                 // split latent in 64x64 tiles and compute in several steps
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                    tae_first_stage->compute(n_threads, in, true, &out);
+                    return tae_first_stage->compute(n_threads, in, true, &out);
                 };
                 sd_tiling(x, result, vae_scale_factor, 64, 0.5f, on_tiling);
             } else {
-                tae_first_stage->compute(n_threads, x, true, &result);
+                if (!tae_first_stage->compute(n_threads, x, true, &result)) {
+                    LOG_ERROR("Failed to decode latetnts");
+                    tae_first_stage->free_compute_buffer();
+                    return nullptr;
+                }
             }
             tae_first_stage->free_compute_buffer();
         }
@@ -2701,6 +2735,16 @@ public:
         ggml_ext_tensor_clamp_inplace(result, 0.0f, 1.0f);
         return result;
     }
+
+    void set_flow_shift(float flow_shift = INFINITY) {
+        auto flow_denoiser = std::dynamic_pointer_cast<DiscreteFlowDenoiser>(denoiser);
+        if (flow_denoiser) {
+            if (flow_shift == INFINITY) {
+                flow_shift = default_flow_shift;
+            }
+            flow_denoiser->set_shift(flow_shift);
+        }
+    }
 };
 
 /*================================================= SD API ==================================================*/
@@ -2921,7 +2965,6 @@ void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params) {
     sd_ctx_params->chroma_use_dit_mask     = true;
     sd_ctx_params->chroma_use_t5_mask      = false;
     sd_ctx_params->chroma_t5_mask_pad      = 1;
-    sd_ctx_params->flow_shift              = INFINITY;
 }
 
 char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
@@ -3013,6 +3056,7 @@ void sd_sample_params_init(sd_sample_params_t* sample_params) {
     sample_params->sample_steps                = 20;
     sample_params->custom_sigmas               = nullptr;
     sample_params->custom_sigmas_count         = 0;
+    sample_params->flow_shift                  = INFINITY;
 }
 
 char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
@@ -3033,7 +3077,8 @@ char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
              "sample_method: %s, "
              "sample_steps: %d, "
              "eta: %.2f, "
-             "shifted_timestep: %d)",
+             "shifted_timestep: %d, "
+             "flow_shift: %.2f)",
              sample_params->guidance.txt_cfg,
              std::isfinite(sample_params->guidance.img_cfg)
                  ? sample_params->guidance.img_cfg
@@ -3047,7 +3092,8 @@ char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
              sd_sample_method_name(sample_params->sample_method),
              sample_params->sample_steps,
              sample_params->eta,
-             sample_params->shifted_timestep);
+             sample_params->shifted_timestep,
+             sample_params->flow_shift);
 
     return buf;
 }
@@ -3461,6 +3507,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
         ggml_free(work_ctx);
         return nullptr;
     }
+    memset(result_images, 0, batch_count * sizeof(sd_image_t));
 
     for (size_t i = 0; i < decoded_images.size(); i++) {
         result_images[i].width   = width;
@@ -3517,6 +3564,8 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
 
     size_t t0 = ggml_time_ms();
 
+    sd_ctx->sd->set_flow_shift(sd_img_gen_params->sample_params.flow_shift);
+
     // Apply lora
     sd_ctx->sd->apply_loras(sd_img_gen_params->loras, sd_img_gen_params->lora_count);
 
@@ -3792,6 +3841,8 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
     }
     LOG_INFO("generate_video %dx%dx%d", width, height, frames);
 
+    sd_ctx->sd->set_flow_shift(sd_vid_gen_params->sample_params.flow_shift);
+
     enum sample_method_t sample_method = sd_vid_gen_params->sample_params.sample_method;
     if (sample_method == SAMPLE_METHOD_COUNT) {
         sample_method = sd_get_default_sample_method(sd_ctx);

src/t5.hpp

@@ -14,6 +14,7 @@
 #include "ggml_extend.hpp"
 #include "json.hpp"
 #include "model.h"
+#include "vocab/vocab.h"
 
 // Port from: https://github.com/google/sentencepiece/blob/master/src/unigram_model.h
 // and https://github.com/google/sentencepiece/blob/master/src/unigram_model.h.
@@ -341,9 +342,9 @@ protected:
 public:
     explicit T5UniGramTokenizer(bool is_umt5 = false) {
         if (is_umt5) {
-            InitializePieces(ModelLoader::load_umt5_tokenizer_json());
+            InitializePieces(load_umt5_tokenizer_json());
         } else {
-            InitializePieces(ModelLoader::load_t5_tokenizer_json());
+            InitializePieces(load_t5_tokenizer_json());
         }
 
         min_score_ = FLT_MAX;

src/tokenize_util.cpp

@@ -919,15 +919,21 @@ std::vector<std::string> token_split(const std::string& text) {
 
         // `\s*[\r\n]+|\s+(?!\S)|\s+`
         if (is_space(cp)) {
-            std::string token = codepoint_to_utf8(cp);
-            ++i;
+            std::string token;
+            bool saw_new_line = false;
 
             while (i < cps.size() && is_space(cps[i])) {
                 token += codepoint_to_utf8(cps[i]);
-                ++i;
+
                 if (cps[i] == U'\r' || cps[i] == U'\n') {
-                    break;
+                    saw_new_line = true;
+                } else {
+                    if (saw_new_line) {
+                        break;
+                    }
                 }
+
+                ++i;
             }
 
             tokens.push_back(token);

src/unet.hpp

@@ -1,8 +1,7 @@
 #ifndef __UNET_HPP__
 #define __UNET_HPP__
 
-#include "common.hpp"
-#include "ggml_extend.hpp"
+#include "common_block.hpp"
 #include "model.h"
 
 /*==================================================== UnetModel =====================================================*/

src/upscaler.cpp

@@ -89,7 +89,7 @@ struct UpscalerGGML {
 
         ggml_tensor* upscaled = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, output_width, output_height, 3, 1);
         auto on_tiling        = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-            esrgan_upscaler->compute(n_threads, in, &out);
+            return esrgan_upscaler->compute(n_threads, in, &out);
         };
         int64_t t0 = ggml_time_ms();
         sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, on_tiling);

src/vae.hpp

@@ -1,8 +1,7 @@
 #ifndef __VAE_HPP__
 #define __VAE_HPP__
 
-#include "common.hpp"
-#include "ggml_extend.hpp"
+#include "common_block.hpp"
 
 /*================================================== AutoEncoderKL ===================================================*/
 
@@ -141,7 +140,7 @@ public:
             v = ggml_reshape_3d(ctx->ggml_ctx, v, c, h * w, n);                        // [N, h * w, in_channels]
         }
 
-        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, true, ctx->flash_attn_enabled);
+        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);
 
         if (use_linear) {
             h_ = proj_out->forward(ctx, h_);  // [N, h * w, in_channels]

src/vocab/clip_t5.hpp

@@ -1,4 +1,4 @@
-static unsigned char merges_utf8_c_str[] = {
+static const unsigned char clip_merges_utf8_c_str[] = {
     0x23,
     0x76,
     0x65,
@@ -524620,7 +524620,7 @@ static unsigned char merges_utf8_c_str[] = {
     0x0a,
 };
 
-static unsigned char t5_tokenizer_json_str[] = {
+static const unsigned char t5_tokenizer_json_str[] = {
     0x7b,
     0x0a,
     0x20,

src/vocab/mistral.hpp

@@ -1,4 +1,4 @@
-unsigned char mistral_merges_utf8_c_str[] = {
+static const unsigned char mistral_merges_utf8_c_str[] = {
   0xc4, 0xa0, 0x20, 0xc4, 0xa0, 0x0a, 0xc4, 0xa0, 0x20, 0x74, 0x0a, 0x65,
   0x20, 0x72, 0x0a, 0x69, 0x20, 0x6e, 0x0a, 0xc4, 0xa0, 0x20, 0xc4, 0xa0,
   0xc4, 0xa0, 0xc4, 0xa0, 0x0a, 0xc4, 0xa0, 0xc4, 0xa0, 0x20, 0xc4, 0xa0,
@@ -260614,7 +260614,7 @@ unsigned char mistral_merges_utf8_c_str[] = {
   0xc3, 0xa5, 0xc4, 0xb2, 0xc4, 0xb0, 0x20, 0xc3, 0xa6, 0xc2, 0xb1, 0xc4,
   0xab, 0xc3, 0xa4, 0xc2, 0xb9, 0xc2, 0xa6, 0x0a,
 };
-unsigned char mistral_vocab_json_utf8_c_str[] = {
+static const unsigned char mistral_vocab_json_utf8_c_str[] = {
   0x7b, 0x22, 0x3c, 0x75, 0x6e, 0x6b, 0x3e, 0x22, 0x3a, 0x20, 0x30, 0x2c,
   0x20, 0x22, 0x3c, 0x73, 0x3e, 0x22, 0x3a, 0x20, 0x31, 0x2c, 0x20, 0x22,
   0x3c, 0x2f, 0x73, 0x3e, 0x22, 0x3a, 0x20, 0x32, 0x2c, 0x20, 0x22, 0x5b,

src/vocab/qwen.hpp

@@ -1,4 +1,4 @@
-unsigned char qwen2_merges_utf8_c_str[] = {
+static const unsigned char qwen2_merges_utf8_c_str[] = {
   0xc4, 0xa0, 0x20, 0xc4, 0xa0, 0x0a, 0xc4, 0xa0, 0xc4, 0xa0, 0x20, 0xc4,
   0xa0, 0xc4, 0xa0, 0x0a, 0x69, 0x20, 0x6e, 0x0a, 0xc4, 0xa0, 0x20, 0x74,
   0x0a, 0xc4, 0xa0, 0xc4, 0xa0, 0xc4, 0xa0, 0xc4, 0xa0, 0x20, 0xc4, 0xa0,

src/vocab/umt5.hpp

@@ -1,4 +1,4 @@
-unsigned char umt5_tokenizer_json_str[] = {
+static const unsigned char umt5_tokenizer_json_str[] = {
   0x7b, 0x22, 0x76, 0x65, 0x72, 0x73, 0x69, 0x6f, 0x6e, 0x22, 0x3a, 0x20,
   0x22, 0x31, 0x2e, 0x30, 0x22, 0x2c, 0x20, 0x22, 0x74, 0x72, 0x75, 0x6e,
   0x63, 0x61, 0x74, 0x69, 0x6f, 0x6e, 0x22, 0x3a, 0x20, 0x6e, 0x75, 0x6c,

src/vocab/vocab.cpp

@@ -0,0 +1,35 @@
+#include "vocab.h"
+#include "clip_t5.hpp"
+#include "mistral.hpp"
+#include "qwen.hpp"
+#include "umt5.hpp"
+
+std::string load_clip_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(clip_merges_utf8_c_str), sizeof(clip_merges_utf8_c_str));
+    return merges_utf8_str;
+}
+
+std::string load_qwen2_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(qwen2_merges_utf8_c_str), sizeof(qwen2_merges_utf8_c_str));
+    return merges_utf8_str;
+}
+
+std::string load_mistral_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(mistral_merges_utf8_c_str), sizeof(mistral_merges_utf8_c_str));
+    return merges_utf8_str;
+}
+
+std::string load_mistral_vocab_json() {
+    std::string json_str(reinterpret_cast<const char*>(mistral_vocab_json_utf8_c_str), sizeof(mistral_vocab_json_utf8_c_str));
+    return json_str;
+}
+
+std::string load_t5_tokenizer_json() {
+    std::string json_str(reinterpret_cast<const char*>(t5_tokenizer_json_str), sizeof(t5_tokenizer_json_str));
+    return json_str;
+}
+
+std::string load_umt5_tokenizer_json() {
+    std::string json_str(reinterpret_cast<const char*>(umt5_tokenizer_json_str), sizeof(umt5_tokenizer_json_str));
+    return json_str;
+}

src/vocab/vocab.h

@@ -0,0 +1,13 @@
+#ifndef __VOCAB_H__
+#define __VOCAB_H__
+
+#include <string>
+
+std::string load_clip_merges();
+std::string load_qwen2_merges();
+std::string load_mistral_merges();
+std::string load_mistral_vocab_json();
+std::string load_t5_tokenizer_json();
+std::string load_umt5_tokenizer_json();
+
+#endif  // __VOCAB_H__

src/wan.hpp

@@ -5,9 +5,8 @@
 #include <memory>
 #include <utility>
 
-#include "common.hpp"
+#include "common_block.hpp"
 #include "flux.hpp"
-#include "ggml_extend.hpp"
 #include "rope.hpp"
 #include "vae.hpp"
 
@@ -572,8 +571,8 @@ namespace WAN {
             auto v = qkv_vec[2];
             v      = ggml_reshape_3d(ctx->ggml_ctx, v, h * w, c, n);  // [t, c, h * w]
 
-            v = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, v, 1, 0, 2, 3));                           // [t, h * w, c]
-            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, true, ctx->flash_attn_enabled);  // [t, h * w, c]
+            v = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, v, 1, 0, 2, 3));                            // [t, h * w, c]
+            x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);  // [t, h * w, c]
 
             x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 1, 0, 2, 3));  // [t, c, h * w]
             x = ggml_reshape_4d(ctx->ggml_ctx, x, w, h, c, n);                             // [t, c, h, w]

src/z_image.hpp

@@ -346,69 +346,6 @@ namespace ZImage {
             blocks["final_layer"] = std::make_shared<FinalLayer>(z_image_params.hidden_size, z_image_params.patch_size, z_image_params.out_channels);
         }
 
-        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
-                                              struct ggml_tensor* x) {
-            int64_t W = x->ne[0];
-            int64_t H = x->ne[1];
-
-            int pad_h = (z_image_params.patch_size - H % z_image_params.patch_size) % z_image_params.patch_size;
-            int pad_w = (z_image_params.patch_size - W % z_image_params.patch_size) % z_image_params.patch_size;
-            x         = ggml_ext_pad(ctx->ggml_ctx, x, pad_w, pad_h, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
-            return x;
-        }
-
-        struct ggml_tensor* patchify(struct ggml_context* ctx,
-                                     struct ggml_tensor* x) {
-            // x: [N, C, H, W]
-            // return: [N, h*w, patch_size*patch_size*C]
-            int64_t N = x->ne[3];
-            int64_t C = x->ne[2];
-            int64_t H = x->ne[1];
-            int64_t W = x->ne[0];
-            int64_t p = z_image_params.patch_size;
-            int64_t h = H / z_image_params.patch_size;
-            int64_t w = W / z_image_params.patch_size;
-
-            GGML_ASSERT(h * p == H && w * p == W);
-
-            x = ggml_reshape_4d(ctx, x, p, w, p, h * C * N);                 // [N*C*h, p, w, p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));            // [N*C*h, w, p, p]
-            x = ggml_reshape_4d(ctx, x, p * p, w * h, C, N);                 // [N, C, h*w, p*p]
-            x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 2, 0, 1, 3));  // [N, h*w, C, p*p]
-            x = ggml_reshape_3d(ctx, x, C * p * p, w * h, N);                // [N, h*w, p*p*C]
-            return x;
-        }
-
-        struct ggml_tensor* process_img(GGMLRunnerContext* ctx,
-                                        struct ggml_tensor* x) {
-            x = pad_to_patch_size(ctx, x);
-            x = patchify(ctx->ggml_ctx, x);
-            return x;
-        }
-
-        struct ggml_tensor* unpatchify(struct ggml_context* ctx,
-                                       struct ggml_tensor* x,
-                                       int64_t h,
-                                       int64_t w) {
-            // x: [N, h*w, patch_size*patch_size*C]
-            // return: [N, C, H, W]
-            int64_t N = x->ne[2];
-            int64_t C = x->ne[0] / z_image_params.patch_size / z_image_params.patch_size;
-            int64_t H = h * z_image_params.patch_size;
-            int64_t W = w * z_image_params.patch_size;
-            int64_t p = z_image_params.patch_size;
-
-            GGML_ASSERT(C * p * p == x->ne[0]);
-
-            x = ggml_reshape_4d(ctx, x, C, p * p, w * h, N);                 // [N, h*w, p*p, C]
-            x = ggml_cont(ctx, ggml_ext_torch_permute(ctx, x, 1, 2, 0, 3));  // [N, C, h*w, p*p]
-            x = ggml_reshape_4d(ctx, x, p, p, w, h * C * N);                 // [N*C*h, w, p, p]
-            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));            // [N*C*h, p, w, p]
-            x = ggml_reshape_4d(ctx, x, W, H, C, N);                         // [N, C, h*p, w*p]
-
-            return x;
-        }
-
         struct ggml_tensor* forward_core(GGMLRunnerContext* ctx,
                                          struct ggml_tensor* x,
                                          struct ggml_tensor* timestep,
@@ -495,27 +432,22 @@ namespace ZImage {
             int64_t C = x->ne[2];
             int64_t N = x->ne[3];
 
-            auto img             = process_img(ctx, x);
+            int patch_size = z_image_params.patch_size;
+
+            auto img             = DiT::pad_and_patchify(ctx, x, patch_size, patch_size, false);
             uint64_t n_img_token = img->ne[1];
 
             if (ref_latents.size() > 0) {
                 for (ggml_tensor* ref : ref_latents) {
-                    ref = process_img(ctx, ref);
+                    ref = DiT::pad_and_patchify(ctx, ref, patch_size, patch_size, false);
                     img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                 }
             }
 
-            int64_t h_len = ((H + (z_image_params.patch_size / 2)) / z_image_params.patch_size);
-            int64_t w_len = ((W + (z_image_params.patch_size / 2)) / z_image_params.patch_size);
-
             auto out = forward_core(ctx, img, timestep, context, pe);
 
-            out = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, n_img_token);  // [N, n_img_token, ph*pw*C]
-            out = unpatchify(ctx->ggml_ctx, out, h_len, w_len);           // [N, C, H + pad_h, W + pad_w]
-
-            // slice
-            out = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, H);  // [N, C, H, W + pad_w]
-            out = ggml_ext_slice(ctx->ggml_ctx, out, 0, 0, W);  // [N, C, H, W]
+            out = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, n_img_token);                              // [N, n_img_token, ph*pw*C]
+            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, patch_size, patch_size, false);  // [N, C, H, W]
 
             out = ggml_ext_scale(ctx->ggml_ctx, out, -1.f);