auto detect z-image-omni

.github/workflows/build.yml

@@ -70,7 +70,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -123,7 +123,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -177,7 +177,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
@@ -240,7 +240,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -340,7 +340,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2
 
       - name: Pack artifacts
         id: pack_artifacts
@@ -463,7 +463,7 @@ jobs:
       - name: Get commit hash
         id: commit
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2
 
       - name: Pack artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -485,146 +485,6 @@ jobs:
           path: |
             sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip
 
-  ubuntu-latest-rocm:
-    runs-on: ubuntu-latest
-    container: rocm/dev-ubuntu-24.04:7.2
-
-    env:
-      ROCM_VERSION: "7.2"
-      UBUNTU_VERSION: "24.04"
-      GPU_TARGETS: "gfx1151;gfx1150;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
-
-    steps:
-      - run: apt-get update && apt-get install -y git
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-        with:
-          submodules: recursive
-
-      - name: Free disk space
-        run: |
-          # Remove preinstalled SDKs and caches not needed for this job
-          sudo rm -rf /usr/share/dotnet || true
-          sudo rm -rf /usr/local/lib/android || true
-          sudo rm -rf /opt/ghc || true
-          sudo rm -rf /usr/local/.ghcup || true
-          sudo rm -rf /opt/hostedtoolcache || true
-
-          # Remove old package lists and caches
-          sudo rm -rf /var/lib/apt/lists/* || true
-          sudo apt clean
-
-      - name: Dependencies
-        id: depends
-        run: |
-          sudo apt-get update
-          sudo apt install -y \
-            cmake \
-            hip-dev \
-            hipblas-dev \
-            ninja-build \
-            rocm-dev \
-            zip
-          # Clean apt caches to recover disk space
-          sudo apt clean
-          sudo rm -rf /var/lib/apt/lists/* || true
-
-      - name: Setup ROCm Environment
-        run: |
-          # Add ROCm to PATH for current session
-          echo "/opt/rocm/bin" >> $GITHUB_PATH
-
-          # Build regex pattern from ${{ env.GPU_TARGETS }} (match target as substring)
-          TARGET_REGEX="($(printf '%s' "${{ env.GPU_TARGETS }}" | sed 's/;/|/g'))"
-
-          # Remove library files for architectures we're not building for to save disk space
-          echo "Cleaning up unneeded architecture files..."
-          cd /opt/rocm/lib/rocblas/library
-          # Keep only our target architectures
-          for file in *; do
-            if printf '%s' "$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
-            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
-        id: cmake_build
-        run: |
-          mkdir build
-          cd build
-          cmake .. -G Ninja \
-            -DCMAKE_CXX_COMPILER=amdclang++ \
-            -DCMAKE_C_COMPILER=amdclang \
-            -DCMAKE_BUILD_TYPE=Release \
-            -DSD_HIPBLAS=ON \
-            -DGPU_TARGETS="${{ env.GPU_TARGETS }}" \
-            -DAMDGPU_TARGETS="${{ env.GPU_TARGETS }}" \
-            -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
-            -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
-            -DSD_BUILD_SHARED_LIBS=ON
-          cmake --build . --config Release
-
-      - name: Get commit hash
-        id: commit
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: prompt/actions-commit-hash@v2
-
-      - name: Prepare artifacts
-        id: prepare_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          # Copy licenses
-          cp ggml/LICENSE ./build/bin/ggml.txt
-          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
-
-          # Move ROCm runtime libraries (to avoid double space consumption)
-          sudo mv /opt/rocm/lib/librocsparse.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhsa-runtime64.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libamdhip64.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhipblas.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhipblaslt.so* ./build/bin/
-          sudo mv /opt/rocm/lib/librocblas.so* ./build/bin/
-          sudo mv /opt/rocm/lib/rocblas/ ./build/bin/
-          sudo mv /opt/rocm/lib/hipblaslt/ ./build/bin/
-
-      - name: Fetch system info
-        id: system-info
-        run: |
-          echo "CPU_ARCH=`uname -m`" >> "$GITHUB_OUTPUT"
-          echo "OS_NAME=`lsb_release -s -i`" >> "$GITHUB_OUTPUT"
-          echo "OS_VERSION=`lsb_release -s -r`" >> "$GITHUB_OUTPUT"
-          echo "OS_TYPE=`uname -s`" >> "$GITHUB_OUTPUT"
-
-      - name: Pack artifacts
-        id: pack_artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          cp ggml/LICENSE ./build/bin/ggml.txt
-          cp LICENSE ./build/bin/stable-diffusion.cpp.txt
-          zip -y -r sd-${{ 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: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        uses: actions/upload-artifact@v4
-        with:
-          name: sd-${{ 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
-          path: |
-            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
-
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
 
@@ -633,7 +493,6 @@ jobs:
     needs:
       - ubuntu-latest-cmake
       - ubuntu-latest-cmake-vulkan
-      - ubuntu-latest-rocm
       - build-and-push-docker-images
       - macOS-latest-cmake
       - windows-latest-cmake
@@ -660,7 +519,7 @@ jobs:
 
       - name: Get commit hash
         id: commit
-        uses: prompt/actions-commit-hash@v2
+        uses: pr-mpt/actions-commit-hash@v2
 
       - name: Create release
         id: create_release

CMakeLists.txt

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

README.md

@@ -53,7 +53,6 @@ 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)
@@ -140,7 +139,6 @@ 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)

clip.hpp

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

common.hpp

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

conditioner.hpp

@@ -39,7 +39,6 @@ struct Conditioner {
     virtual void free_params_buffer()                                                      = 0;
     virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors)    = 0;
     virtual size_t get_params_buffer_size()                                                = 0;
-    virtual void set_flash_attention_enabled(bool enabled)                                 = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
     virtual std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
                                                                                           int n_threads,
@@ -121,13 +120,6 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         return buffer_size;
     }
 
-    void set_flash_attention_enabled(bool enabled) override {
-        text_model->set_flash_attention_enabled(enabled);
-        if (sd_version_is_sdxl(version)) {
-            text_model2->set_flash_attention_enabled(enabled);
-        }
-    }
-
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         text_model->set_weight_adapter(adapter);
         if (sd_version_is_sdxl(version)) {
@@ -796,18 +788,6 @@ struct SD3CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
-    void set_flash_attention_enabled(bool enabled) override {
-        if (clip_l) {
-            clip_l->set_flash_attention_enabled(enabled);
-        }
-        if (clip_g) {
-            clip_g->set_flash_attention_enabled(enabled);
-        }
-        if (t5) {
-            t5->set_flash_attention_enabled(enabled);
-        }
-    }
-
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (clip_l) {
             clip_l->set_weight_adapter(adapter);
@@ -1216,15 +1196,6 @@ struct FluxCLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
-    void set_flash_attention_enabled(bool enabled) override {
-        if (clip_l) {
-            clip_l->set_flash_attention_enabled(enabled);
-        }
-        if (t5) {
-            t5->set_flash_attention_enabled(enabled);
-        }
-    }
-
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {
         if (clip_l) {
             clip_l->set_weight_adapter(adapter);
@@ -1474,12 +1445,6 @@ struct T5CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
-    void set_flash_attention_enabled(bool enabled) override {
-        if (t5) {
-            t5->set_flash_attention_enabled(enabled);
-        }
-    }
-
     void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
         if (t5) {
             t5->set_weight_adapter(adapter);
@@ -1641,142 +1606,6 @@ 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;
@@ -1826,10 +1655,6 @@ struct LLMEmbedder : public Conditioner {
         return 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 {
         if (llm) {
             llm->set_weight_adapter(adapter);

control.hpp

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

denoiser.hpp

@@ -1,8 +1,6 @@
 #ifndef __DENOISER_HPP__
 #define __DENOISER_HPP__
 
-#include <cmath>
-
 #include "ggml_extend.hpp"
 #include "gits_noise.inl"
 
@@ -353,95 +351,6 @@ struct SmoothStepScheduler : SigmaScheduler {
     }
 };
 
-struct BongTangentScheduler : SigmaScheduler {
-    static constexpr float kPi = 3.14159265358979323846f;
-
-    static std::vector<float> get_bong_tangent_sigmas(int steps, float slope, float pivot, float start, float end) {
-        std::vector<float> sigmas;
-        if (steps <= 0) {
-            return sigmas;
-        }
-
-        float smax   = ((2.0f / kPi) * atanf(-slope * (0.0f - pivot)) + 1.0f) * 0.5f;
-        float smin   = ((2.0f / kPi) * atanf(-slope * ((float)(steps - 1) - pivot)) + 1.0f) * 0.5f;
-        float srange = smax - smin;
-        float sscale = start - end;
-
-        sigmas.reserve(steps);
-
-        if (fabsf(srange) < 1e-8f) {
-            if (steps == 1) {
-                sigmas.push_back(start);
-                return sigmas;
-            }
-            for (int i = 0; i < steps; ++i) {
-                float t = (float)i / (float)(steps - 1);
-                sigmas.push_back(start + (end - start) * t);
-            }
-            return sigmas;
-        }
-
-        float inv_srange = 1.0f / srange;
-        for (int x = 0; x < steps; ++x) {
-            float v     = ((2.0f / kPi) * atanf(-slope * ((float)x - pivot)) + 1.0f) * 0.5f;
-            float sigma = ((v - smin) * inv_srange) * sscale + end;
-            sigmas.push_back(sigma);
-        }
-
-        return sigmas;
-    }
-
-    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t /*t_to_sigma*/) override {
-        std::vector<float> result;
-        if (n == 0) {
-            return result;
-        }
-
-        float start  = sigma_max;
-        float end    = sigma_min;
-        float middle = sigma_min + (sigma_max - sigma_min) * 0.5f;
-
-        float pivot_1 = 0.6f;
-        float pivot_2 = 0.6f;
-        float slope_1 = 0.2f;
-        float slope_2 = 0.2f;
-
-        int steps     = static_cast<int>(n) + 2;
-        int midpoint  = static_cast<int>(((float)steps * pivot_1 + (float)steps * pivot_2) * 0.5f);
-        int pivot_1_i = static_cast<int>((float)steps * pivot_1);
-        int pivot_2_i = static_cast<int>((float)steps * pivot_2);
-
-        float slope_scale = (float)steps / 40.0f;
-        slope_1           = slope_1 / slope_scale;
-        slope_2           = slope_2 / slope_scale;
-
-        int stage_2_len = steps - midpoint;
-        int stage_1_len = steps - stage_2_len;
-
-        std::vector<float> sigmas_1 = get_bong_tangent_sigmas(stage_1_len, slope_1, (float)pivot_1_i, start, middle);
-        std::vector<float> sigmas_2 = get_bong_tangent_sigmas(stage_2_len, slope_2, (float)(pivot_2_i - stage_1_len), middle, end);
-
-        if (!sigmas_1.empty()) {
-            sigmas_1.pop_back();
-        }
-
-        result.reserve(n + 1);
-        result.insert(result.end(), sigmas_1.begin(), sigmas_1.end());
-        result.insert(result.end(), sigmas_2.begin(), sigmas_2.end());
-
-        if (result.size() < n + 1) {
-            while (result.size() < n + 1) {
-                result.push_back(end);
-            }
-        } else if (result.size() > n + 1) {
-            result.resize(n + 1);
-        }
-
-        result[n] = 0.0f;
-        return result;
-    }
-};
-
 struct KLOptimalScheduler : SigmaScheduler {
     std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
         std::vector<float> sigmas;
@@ -522,10 +431,6 @@ struct Denoiser {
                 LOG_INFO("get_sigmas with SmoothStep scheduler");
                 scheduler = std::make_shared<SmoothStepScheduler>();
                 break;
-            case BONG_TANGENT_SCHEDULER:
-                LOG_INFO("get_sigmas with bong_tangent scheduler");
-                scheduler = std::make_shared<BongTangentScheduler>();
-                break;
             case KL_OPTIMAL_SCHEDULER:
                 LOG_INFO("get_sigmas with KL Optimal scheduler");
                 scheduler = std::make_shared<KLOptimalScheduler>();
@@ -657,8 +562,9 @@ struct DiscreteFlowDenoiser : public Denoiser {
 
     float sigma_data = 1.0f;
 
-    DiscreteFlowDenoiser(float shift = 3.0f) {
-        set_shift(shift);
+    DiscreteFlowDenoiser(float shift = 3.0f)
+        : shift(shift) {
+        set_parameters();
     }
 
     void set_parameters() {
@@ -667,11 +573,6 @@ struct DiscreteFlowDenoiser : public Denoiser {
         }
     }
 
-    void set_shift(float shift) {
-        this->shift = shift;
-        set_parameters();
-    }
-
     float sigma_min() override {
         return sigmas[0];
     }
@@ -714,8 +615,34 @@ 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 DiscreteFlowDenoiser {
-    FluxFlowDenoiser() = default;
+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];
+    }
 
     float sigma_to_t(float sigma) override {
         return sigma;
@@ -725,6 +652,26 @@ struct FluxFlowDenoiser : public DiscreteFlowDenoiser {
         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 {
@@ -1687,216 +1634,6 @@ static bool sample_k_diffusion(sample_method_t method,
                 }
             }
         } break;
-        case RES_MULTISTEP_SAMPLE_METHOD:  // Res Multistep sampler
-        {
-            struct ggml_tensor* noise        = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* old_denoised = ggml_dup_tensor(work_ctx, x);
-
-            bool have_old_sigma  = false;
-            float old_sigma_down = 0.0f;
-
-            auto t_fn     = [](float sigma) -> float { return -logf(sigma); };
-            auto sigma_fn = [](float t) -> float { return expf(-t); };
-            auto phi1_fn  = [](float t) -> float {
-                if (fabsf(t) < 1e-6f) {
-                    return 1.0f + t * 0.5f + (t * t) / 6.0f;
-                }
-                return (expf(t) - 1.0f) / t;
-            };
-            auto phi2_fn = [&](float t) -> float {
-                if (fabsf(t) < 1e-6f) {
-                    return 0.5f + t / 6.0f + (t * t) / 24.0f;
-                }
-                float phi1_val = phi1_fn(t);
-                return (phi1_val - 1.0f) / t;
-            };
-
-            for (int i = 0; i < steps; i++) {
-                ggml_tensor* denoised = model(x, sigmas[i], i + 1);
-                if (denoised == nullptr) {
-                    return false;
-                }
-
-                float sigma_from = sigmas[i];
-                float sigma_to   = sigmas[i + 1];
-                float sigma_up   = 0.0f;
-                float sigma_down = sigma_to;
-
-                if (eta > 0.0f) {
-                    float sigma_from_sq = sigma_from * sigma_from;
-                    float sigma_to_sq   = sigma_to * sigma_to;
-                    if (sigma_from_sq > 0.0f) {
-                        float term = sigma_to_sq * (sigma_from_sq - sigma_to_sq) / sigma_from_sq;
-                        if (term > 0.0f) {
-                            sigma_up = eta * std::sqrt(term);
-                        }
-                    }
-                    sigma_up            = std::min(sigma_up, sigma_to);
-                    float sigma_down_sq = sigma_to_sq - sigma_up * sigma_up;
-                    sigma_down          = sigma_down_sq > 0.0f ? std::sqrt(sigma_down_sq) : 0.0f;
-                }
-
-                if (sigma_down == 0.0f || !have_old_sigma) {
-                    float dt            = sigma_down - sigma_from;
-                    float* vec_x        = (float*)x->data;
-                    float* vec_denoised = (float*)denoised->data;
-
-                    for (int j = 0; j < ggml_nelements(x); j++) {
-                        float d  = (vec_x[j] - vec_denoised[j]) / sigma_from;
-                        vec_x[j] = vec_x[j] + d * dt;
-                    }
-                } else {
-                    float t      = t_fn(sigma_from);
-                    float t_old  = t_fn(old_sigma_down);
-                    float t_next = t_fn(sigma_down);
-                    float t_prev = t_fn(sigmas[i - 1]);
-                    float h      = t_next - t;
-                    float c2     = (t_prev - t_old) / h;
-
-                    float phi1_val = phi1_fn(-h);
-                    float phi2_val = phi2_fn(-h);
-                    float b1       = phi1_val - phi2_val / c2;
-                    float b2       = phi2_val / c2;
-
-                    if (!std::isfinite(b1)) {
-                        b1 = 0.0f;
-                    }
-                    if (!std::isfinite(b2)) {
-                        b2 = 0.0f;
-                    }
-
-                    float sigma_h           = sigma_fn(h);
-                    float* vec_x            = (float*)x->data;
-                    float* vec_denoised     = (float*)denoised->data;
-                    float* vec_old_denoised = (float*)old_denoised->data;
-
-                    for (int j = 0; j < ggml_nelements(x); j++) {
-                        vec_x[j] = sigma_h * vec_x[j] + h * (b1 * vec_denoised[j] + b2 * vec_old_denoised[j]);
-                    }
-                }
-
-                if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
-                    ggml_ext_im_set_randn_f32(noise, rng);
-                    float* vec_x     = (float*)x->data;
-                    float* vec_noise = (float*)noise->data;
-
-                    for (int j = 0; j < ggml_nelements(x); j++) {
-                        vec_x[j] = vec_x[j] + vec_noise[j] * sigma_up;
-                    }
-                }
-
-                float* vec_old_denoised = (float*)old_denoised->data;
-                float* vec_denoised     = (float*)denoised->data;
-                for (int j = 0; j < ggml_nelements(x); j++) {
-                    vec_old_denoised[j] = vec_denoised[j];
-                }
-
-                old_sigma_down = sigma_down;
-                have_old_sigma = true;
-            }
-        } break;
-        case RES_2S_SAMPLE_METHOD:  // Res 2s sampler
-        {
-            struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* x0    = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* x2    = ggml_dup_tensor(work_ctx, x);
-
-            const float c2 = 0.5f;
-            auto t_fn      = [](float sigma) -> float { return -logf(sigma); };
-            auto phi1_fn   = [](float t) -> float {
-                if (fabsf(t) < 1e-6f) {
-                    return 1.0f + t * 0.5f + (t * t) / 6.0f;
-                }
-                return (expf(t) - 1.0f) / t;
-            };
-            auto phi2_fn = [&](float t) -> float {
-                if (fabsf(t) < 1e-6f) {
-                    return 0.5f + t / 6.0f + (t * t) / 24.0f;
-                }
-                float phi1_val = phi1_fn(t);
-                return (phi1_val - 1.0f) / t;
-            };
-
-            for (int i = 0; i < steps; i++) {
-                float sigma_from = sigmas[i];
-                float sigma_to   = sigmas[i + 1];
-
-                ggml_tensor* denoised = model(x, sigma_from, -(i + 1));
-                if (denoised == nullptr) {
-                    return false;
-                }
-
-                float sigma_up   = 0.0f;
-                float sigma_down = sigma_to;
-                if (eta > 0.0f) {
-                    float sigma_from_sq = sigma_from * sigma_from;
-                    float sigma_to_sq   = sigma_to * sigma_to;
-                    if (sigma_from_sq > 0.0f) {
-                        float term = sigma_to_sq * (sigma_from_sq - sigma_to_sq) / sigma_from_sq;
-                        if (term > 0.0f) {
-                            sigma_up = eta * std::sqrt(term);
-                        }
-                    }
-                    sigma_up            = std::min(sigma_up, sigma_to);
-                    float sigma_down_sq = sigma_to_sq - sigma_up * sigma_up;
-                    sigma_down          = sigma_down_sq > 0.0f ? std::sqrt(sigma_down_sq) : 0.0f;
-                }
-
-                float* vec_x  = (float*)x->data;
-                float* vec_x0 = (float*)x0->data;
-                for (int j = 0; j < ggml_nelements(x); j++) {
-                    vec_x0[j] = vec_x[j];
-                }
-
-                if (sigma_down == 0.0f || sigma_from == 0.0f) {
-                    float* vec_denoised = (float*)denoised->data;
-                    for (int j = 0; j < ggml_nelements(x); j++) {
-                        vec_x[j] = vec_denoised[j];
-                    }
-                } else {
-                    float t      = t_fn(sigma_from);
-                    float t_next = t_fn(sigma_down);
-                    float h      = t_next - t;
-
-                    float a21      = c2 * phi1_fn(-h * c2);
-                    float phi1_val = phi1_fn(-h);
-                    float phi2_val = phi2_fn(-h);
-                    float b2       = phi2_val / c2;
-                    float b1       = phi1_val - b2;
-
-                    float sigma_c2 = expf(-(t + h * c2));
-
-                    float* vec_denoised = (float*)denoised->data;
-                    float* vec_x2       = (float*)x2->data;
-                    for (int j = 0; j < ggml_nelements(x); j++) {
-                        float eps1 = vec_denoised[j] - vec_x0[j];
-                        vec_x2[j]  = vec_x0[j] + h * a21 * eps1;
-                    }
-
-                    ggml_tensor* denoised2 = model(x2, sigma_c2, i + 1);
-                    if (denoised2 == nullptr) {
-                        return false;
-                    }
-                    float* vec_denoised2 = (float*)denoised2->data;
-
-                    for (int j = 0; j < ggml_nelements(x); j++) {
-                        float eps1 = vec_denoised[j] - vec_x0[j];
-                        float eps2 = vec_denoised2[j] - vec_x0[j];
-                        vec_x[j]   = vec_x0[j] + h * (b1 * eps1 + b2 * eps2);
-                    }
-                }
-
-                if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
-                    ggml_ext_im_set_randn_f32(noise, rng);
-                    float* vec_x     = (float*)x->data;
-                    float* vec_noise = (float*)noise->data;
-
-                    for (int j = 0; j < ggml_nelements(x); j++) {
-                        vec_x[j] = vec_x[j] + vec_noise[j] * sigma_up;
-                    }
-                }
-            }
-        } break;
 
         default:
             LOG_ERROR("Attempting to sample with nonexisting sample method %i", method);

diffusion_model.hpp

@@ -1,7 +1,6 @@
 #ifndef __DIFFUSION_MODEL_H__
 #define __DIFFUSION_MODEL_H__
 
-#include "anima.hpp"
 #include "flux.hpp"
 #include "mmdit.hpp"
 #include "qwen_image.hpp"
@@ -24,6 +23,8 @@ struct DiffusionParams {
     struct ggml_tensor* vace_context          = nullptr;
     float vace_strength                       = 1.f;
     std::vector<int> skip_layers              = {};
+    std::vector<struct ggml_tensor*> extra_contexts;  // for z-image-omni
+    std::vector<struct ggml_tensor*> ref_clip_feats;  // for z-image-omni
 };
 
 struct DiffusionModel {
@@ -39,7 +40,7 @@ struct DiffusionModel {
     virtual size_t get_params_buffer_size()                                             = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
     virtual int64_t get_adm_in_channels()                            = 0;
-    virtual void set_flash_attention_enabled(bool enabled)           = 0;
+    virtual void set_flash_attn_enabled(bool enabled)                = 0;
     virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
 };
 
@@ -85,7 +86,7 @@ struct UNetModel : public DiffusionModel {
         return unet.unet.adm_in_channels;
     }
 
-    void set_flash_attention_enabled(bool enabled) {
+    void set_flash_attn_enabled(bool enabled) {
         unet.set_flash_attention_enabled(enabled);
     }
 
@@ -150,7 +151,7 @@ struct MMDiTModel : public DiffusionModel {
         return 768 + 1280;
     }
 
-    void set_flash_attention_enabled(bool enabled) {
+    void set_flash_attn_enabled(bool enabled) {
         mmdit.set_flash_attention_enabled(enabled);
     }
 
@@ -216,7 +217,7 @@ struct FluxModel : public DiffusionModel {
         return 768;
     }
 
-    void set_flash_attention_enabled(bool enabled) {
+    void set_flash_attn_enabled(bool enabled) {
         flux.set_flash_attention_enabled(enabled);
     }
 
@@ -243,72 +244,6 @@ 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;
@@ -353,7 +288,7 @@ struct WanModel : public DiffusionModel {
         return 768;
     }
 
-    void set_flash_attention_enabled(bool enabled) {
+    void set_flash_attn_enabled(bool enabled) {
         wan.set_flash_attention_enabled(enabled);
     }
 
@@ -424,7 +359,7 @@ struct QwenImageModel : public DiffusionModel {
         return 768;
     }
 
-    void set_flash_attention_enabled(bool enabled) {
+    void set_flash_attn_enabled(bool enabled) {
         qwen_image.set_flash_attention_enabled(enabled);
     }
 
@@ -491,7 +426,7 @@ struct ZImageModel : public DiffusionModel {
         return 768;
     }
 
-    void set_flash_attention_enabled(bool enabled) {
+    void set_flash_attn_enabled(bool enabled) {
         z_image.set_flash_attention_enabled(enabled);
     }
 
@@ -503,12 +438,14 @@ struct ZImageModel : public DiffusionModel {
                  DiffusionParams diffusion_params,
                  struct ggml_tensor** output     = nullptr,
                  struct ggml_context* output_ctx = nullptr) override {
+        std::vector<ggml_tensor*> contexts = {diffusion_params.context};
+        contexts.insert(contexts.end(), diffusion_params.extra_contexts.begin(), diffusion_params.extra_contexts.end());
         return z_image.compute(n_threads,
                                diffusion_params.x,
                                diffusion_params.timesteps,
-                               diffusion_params.context,
+                               contexts,
                                diffusion_params.ref_latents,
-                               true,  // increase_ref_index
+                               diffusion_params.ref_clip_feats,
                                output,
                                output_ctx);
     }

docs/anima.md

@@ -1,20 +0,0 @@
-# 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" />

docs/z_image.md

@@ -7,9 +7,6 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 - Download Z-Image-Turbo
     - safetensors: https://huggingface.co/Comfy-Org/z_image_turbo/tree/main/split_files/diffusion_models
     - gguf: https://huggingface.co/leejet/Z-Image-Turbo-GGUF/tree/main
-- Download Z-Image
-    - safetensors: https://huggingface.co/Comfy-Org/z_image/tree/main/split_files/diffusion_models
-    - gguf: https://huggingface.co/unsloth/Z-Image-GGUF/tree/main
 - Download vae
     - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-schnell/tree/main
 - Download Qwen3 4b
@@ -18,22 +15,12 @@ You can run Z-Image with stable-diffusion.cpp on GPUs with 4GB of VRAM — or ev
 
 ## Examples
 
-### Z-Image-Turbo
-
 ```
 .\bin\Release\sd-cli.exe --diffusion-model  z_image_turbo-Q3_K.gguf --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\Qwen3-4B-Instruct-2507-Q4_K_M.gguf -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 1.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
 ```
 
 <img width="256" alt="z-image example" src="../assets/z_image/q3_K.png" />
 
-### Z-Image-Base
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\z_image_bf16.safetensors --vae ..\..\ComfyUI\models\vae\ae.sft  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_4b.safetensors -p "A cinematic, melancholic photograph of a solitary hooded figure walking through a sprawling, rain-slicked metropolis at night. The city lights are a chaotic blur of neon orange and cool blue, reflecting on the wet asphalt. The scene evokes a sense of being a single component in a vast machine. Superimposed over the image in a sleek, modern, slightly glitched font is the philosophical quote: 'THE CITY IS A CIRCUIT BOARD, AND I AM A BROKEN TRANSISTOR.' -- moody, atmospheric, profound, dark academic" --cfg-scale 5.0 -v --offload-to-cpu --diffusion-fa -H 1024 -W 512
-```
-
-<img width="256" alt="z-image example" src="../assets/z_image/base_bf16.png" />
-
 ## Comparison of Different Quantization Types
 
 | bf16 | q8_0 | q6_K | q5_0 | q4_K | q4_0 | q3_K | q2_K|

examples/cli/README.md

@@ -4,12 +4,11 @@
 usage: ./bin/sd-cli  [options]
 
 CLI Options:
-  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default:
-                              ./output.png) (eg. output_%03d.png)
+  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default: ./output.png) (eg. output_%03d.png)
+  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
   --preview-path <string>     path to write preview image to (default: ./preview.png)
   --preview-interval <int>    interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at
                               every step)
-  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
   --canny                     apply canny preprocessor (edge detection)
   --convert-name              convert tensor name (for convert mode)
   -v, --verbose               print extra info
@@ -45,6 +44,7 @@ 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
@@ -52,15 +52,13 @@ Context Options:
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
   --vae-on-cpu                             keep vae in cpu (for low vram)
-  --fa                                     use flash attention
-  --diffusion-fa                           use flash attention in the diffusion model only
+  --diffusion-fa                           use flash attention in the diffusion model
   --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
   --vae-conv-direct                        use ggml_conv2d_direct in the vae model
   --circular                               enable circular padding for convolutions
   --circularx                              enable circular RoPE wrapping on x-axis (width) only
   --circulary                              enable circular RoPE wrapping on y-axis (height) only
   --chroma-disable-dit-mask                disable dit mask for chroma
-  --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
   --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
                                            type of the weight file
@@ -110,7 +108,6 @@ 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 and TCD (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)
@@ -127,13 +124,11 @@ Generation Options:
   --disable-auto-resize-ref-image          disable auto resize of ref images
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
   --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd, res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a
-                                           otherwise)
+                                           tcd] (default: euler for Flux/SD3/Wan, euler_a otherwise)
   --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           ddim_trailing, tcd, res_multistep, res_2s] default: euler for Flux/SD3/Wan,
-                                           euler_a otherwise
+                                           ddim_trailing, tcd] default: euler for Flux/SD3/Wan, euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           kl_optimal, lcm, bong_tangent], default: discrete
+                                           kl_optimal, lcm], default: discrete
   --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])

examples/cli/main.cpp

@@ -394,15 +394,12 @@ 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) {
@@ -412,7 +409,7 @@ 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 false;
+            return;
 
         std::string params = get_image_params(cli_params, ctx_params, gen_params, gen_params.seed + idx);
         int ok             = 0;
@@ -422,11 +419,8 @@ 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";
@@ -435,12 +429,9 @@ 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);
-            if (write_image(img_path, i)) {
-                sucessful_reults++;
-            }
+            write_image(img_path, i);
         }
-        LOG_INFO("%d/%d images saved", sucessful_reults, num_results);
-        return sucessful_reults != 0;
+        return true;
     }
 
     if (cli_params.mode == VID_GEN && num_results > 1) {
@@ -448,13 +439,9 @@ bool save_results(const SDCliParams& cli_params,
             ext = ".avi";
         fs::path video_path = base_path;
         video_path += ext;
-        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;
-        }
+        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 (!is_jpg && ext_lower != ".png")
@@ -466,12 +453,10 @@ bool save_results(const SDCliParams& cli_params,
             img_path += "_" + std::to_string(output_begin_idx + i);
         }
         img_path += ext;
-        if (write_image(img_path, i)) {
-            sucessful_reults++;
-        }
+        write_image(img_path, i);
     }
-    LOG_INFO("%d/%d images saved", sucessful_reults, num_results);
-    return sucessful_reults != 0;
+
+    return true;
 }
 
 int main(int argc, const char* argv[]) {

examples/common/common.hpp

@@ -457,7 +457,6 @@ struct SDContextParams {
     bool control_net_cpu        = false;
     bool clip_on_cpu            = false;
     bool vae_on_cpu             = false;
-    bool flash_attn             = false;
     bool diffusion_flash_attn   = false;
     bool diffusion_conv_direct  = false;
     bool vae_conv_direct        = false;
@@ -581,6 +580,10 @@ 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 = {
@@ -612,13 +615,9 @@ struct SDContextParams {
              "--vae-on-cpu",
              "keep vae in cpu (for low vram)",
              true, &vae_on_cpu},
-            {"",
-             "--fa",
-             "use flash attention",
-             true, &flash_attn},
             {"",
              "--diffusion-fa",
-             "use flash attention in the diffusion model only",
+             "use flash attention in the diffusion model",
              true, &diffusion_flash_attn},
             {"",
              "--diffusion-conv-direct",
@@ -899,12 +898,12 @@ 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"
             << "  clip_on_cpu: " << (clip_on_cpu ? "true" : "false") << ",\n"
             << "  vae_on_cpu: " << (vae_on_cpu ? "true" : "false") << ",\n"
-            << "  flash_attn: " << (flash_attn ? "true" : "false") << ",\n"
             << "  diffusion_flash_attn: " << (diffusion_flash_attn ? "true" : "false") << ",\n"
             << "  diffusion_conv_direct: " << (diffusion_conv_direct ? "true" : "false") << ",\n"
             << "  vae_conv_direct: " << (vae_conv_direct ? "true" : "false") << ",\n"
@@ -969,7 +968,6 @@ struct SDContextParams {
             clip_on_cpu,
             control_net_cpu,
             vae_on_cpu,
-            flash_attn,
             diffusion_flash_attn,
             taesd_preview,
             diffusion_conv_direct,
@@ -981,6 +979,7 @@ struct SDContextParams {
             chroma_use_t5_mask,
             chroma_t5_mask_pad,
             qwen_image_zero_cond_t,
+            flow_shift,
         };
         return sd_ctx_params;
     }
@@ -1200,10 +1199,6 @@ 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)",
@@ -1483,17 +1478,17 @@ struct SDGenerationParams {
              on_seed_arg},
             {"",
              "--sampling-method",
-             "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s] "
+             "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd] "
              "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
              on_sample_method_arg},
             {"",
              "--high-noise-sampling-method",
-             "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s]"
+             "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd]"
              " default: euler for Flux/SD3/Wan, euler_a otherwise",
              on_high_noise_sample_method_arg},
             {"",
              "--scheduler",
-             "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default: discrete",
+             "denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple, kl_optimal, lcm], default: discrete",
              on_scheduler_arg},
             {"",
              "--sigmas",
@@ -1604,7 +1599,6 @@ 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

@@ -4,12 +4,12 @@
 usage: ./bin/sd-server  [options]
 
 Svr Options:
-  -l, --listen-ip <string>      server listen ip (default: 127.0.0.1)        
-  --serve-html-path <string>    path to HTML file to serve at root (optional)
-  --listen-port <int>           server listen port (default: 1234)
-  -v, --verbose                 print extra info
-  --color                       colors the logging tags according to level   
-  -h, --help                    show this help message and exit
+  -l, --listen-ip <string>    server listen ip (default: 127.0.0.1)
+  --listen-port <int>         server listen port (default: 1234)
+  --serve-html-path <string>  path to HTML file to serve at root (optional)
+  -v, --verbose               print extra info
+  --color                     colors the logging tags according to level
+  -h, --help                  show this help message and exit
 
 Context Options:
   -m, --model <string>                     path to full model
@@ -36,22 +36,21 @@ 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
-  --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
   --vae-on-cpu                             keep vae in cpu (for low vram)
-  --fa                                     use flash attention
-  --diffusion-fa                           use flash attention in the diffusion model only
+  --mmap                                   whether to memory-map model
+  --diffusion-fa                           use flash attention in the diffusion model
   --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
   --vae-conv-direct                        use ggml_conv2d_direct in the vae model
   --circular                               enable circular padding for convolutions
   --circularx                              enable circular RoPE wrapping on x-axis (width) only
   --circulary                              enable circular RoPE wrapping on y-axis (height) only
   --chroma-disable-dit-mask                disable dit mask for chroma
-  --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
   --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
                                            type of the weight file
@@ -101,7 +100,6 @@ 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 and TCD (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)
@@ -118,13 +116,11 @@ Default Generation Options:
   --disable-auto-resize-ref-image          disable auto resize of ref images
   -s, --seed                               RNG seed (default: 42, use random seed for < 0)
   --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd, res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a
-                                           otherwise)
+                                           tcd] (default: euler for Flux/SD3/Wan, euler_a otherwise)
   --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           ddim_trailing, tcd, res_multistep, res_2s] default: euler for Flux/SD3/Wan,
-                                           euler_a otherwise
+                                           ddim_trailing, tcd] default: euler for Flux/SD3/Wan, euler_a otherwise
   --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           kl_optimal, lcm, bong_tangent], default: discrete
+                                           kl_optimal, lcm], default: discrete
   --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])

examples/server/main.cpp

@@ -263,24 +263,6 @@ void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
     log_print(level, log, svr_params->verbose, svr_params->color);
 }
 
-struct LoraEntry {
-    std::string name;
-    std::string path;
-    std::string fullpath;
-};
-
-void free_results(sd_image_t* result_images, int num_results) {
-    if (result_images) {
-        for (int i = 0; i < num_results; ++i) {
-            if (result_images[i].data) {
-                stbi_image_free(result_images[i].data);
-                result_images[i].data = nullptr;
-            }
-        }
-    }
-    free(result_images);
-}
-
 int main(int argc, const char** argv) {
     if (argc > 1 && std::string(argv[1]) == "--version") {
         std::cout << version_string() << "\n";
@@ -311,56 +293,6 @@ int main(int argc, const char** argv) {
 
     std::mutex sd_ctx_mutex;
 
-    std::vector<LoraEntry> lora_cache;
-    std::mutex lora_mutex;
-
-    auto refresh_lora_cache = [&]() {
-        std::vector<LoraEntry> new_cache;
-
-        fs::path lora_dir = ctx_params.lora_model_dir;
-        if (fs::exists(lora_dir) && fs::is_directory(lora_dir)) {
-            auto is_lora_ext = [](const fs::path& p) {
-                auto ext = p.extension().string();
-                std::transform(ext.begin(), ext.end(), ext.begin(), ::tolower);
-                return ext == ".gguf" || ext == ".pt" || ext == ".pth" || ext == ".safetensors";
-            };
-
-            for (auto& entry : fs::recursive_directory_iterator(lora_dir)) {
-                if (!entry.is_regular_file())
-                    continue;
-                const fs::path& p = entry.path();
-                if (!is_lora_ext(p))
-                    continue;
-
-                LoraEntry e;
-                e.name          = p.stem().u8string();
-                e.fullpath      = p.u8string();
-                std::string rel = p.lexically_relative(lora_dir).u8string();
-                std::replace(rel.begin(), rel.end(), '\\', '/');
-                e.path = rel;
-
-                new_cache.push_back(std::move(e));
-            }
-        }
-
-        std::sort(new_cache.begin(), new_cache.end(),
-                  [](const LoraEntry& a, const LoraEntry& b) {
-                      return a.path < b.path;
-                  });
-
-        {
-            std::lock_guard<std::mutex> lock(lora_mutex);
-            lora_cache = std::move(new_cache);
-        }
-    };
-
-    auto get_lora_full_path = [&](const std::string& path) -> std::string {
-        std::lock_guard<std::mutex> lock(lora_mutex);
-        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;
 
     svr.set_pre_routing_handler([](const httplib::Request& req, httplib::Response& res) {
@@ -380,7 +312,7 @@ int main(int argc, const char** argv) {
         return httplib::Server::HandlerResponse::Unhandled;
     });
 
-    // root
+    // health
     svr.Get("/", [&](const httplib::Request&, httplib::Response& res) {
         if (!svr_params.serve_html_path.empty()) {
             std::ifstream file(svr_params.serve_html_path);
@@ -419,8 +351,8 @@ 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                 = default_gen_params.width > 0 ? default_gen_params.width : 512;
-            int height                = default_gen_params.width > 0 ? default_gen_params.height : 512;
+            int width                 = 512;
+            int height                = 512;
             if (!size.empty()) {
                 auto pos = size.find('x');
                 if (pos != std::string::npos) {
@@ -549,7 +481,6 @@ int main(int argc, const char** argv) {
                 item["b64_json"] = b64;
                 out["data"].push_back(item);
             }
-            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;
@@ -580,9 +511,8 @@ 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[]");
-            bool has_legacy_image = req.form.has_file("image");
-            if (image_count == 0 && !has_legacy_image) {
+            size_t image_count = req.form.get_file_count("image[]");
+            if (image_count == 0) {
                 res.status = 400;
                 res.set_content(R"({"error":"at least one image[] required"})", "application/json");
                 return;
@@ -593,10 +523,6 @@ 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")) {
@@ -614,7 +540,7 @@ int main(int argc, const char** argv) {
             n = std::clamp(n, 1, 8);
 
             std::string size = req.form.get_field("size");
-            int width = -1, height = -1;
+            int width = 512, height = 512;
             if (!size.empty()) {
                 auto pos = size.find('x');
                 if (pos != std::string::npos) {
@@ -671,31 +597,15 @@ int main(int argc, const char** argv) {
 
             LOG_DEBUG("%s\n", gen_params.to_string().c_str());
 
-            sd_image_t init_image    = {0, 0, 3, nullptr};
-            sd_image_t control_image = {0, 0, 3, nullptr};
+            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};
             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;
-                int img_h;
-
+                int img_w           = width;
+                int img_h           = height;
                 uint8_t* raw_pixels = load_image_from_memory(
                     reinterpret_cast<const char*>(bytes.data()),
                     static_cast<int>(bytes.size()),
@@ -707,31 +617,22 @@ 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 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;
-
+                int mask_w        = width;
+                int mask_h        = height;
                 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,
-                    expected_width, expected_height, 1);
+                    width, 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   = get_resolved_width();
-                mask_image.height  = get_resolved_height();
+                mask_image.width   = width;
+                mask_image.height  = height;
                 mask_image.channel = 1;
                 mask_image.data    = nullptr;
             }
@@ -748,8 +649,8 @@ int main(int argc, const char** argv) {
                 gen_params.auto_resize_ref_image,
                 gen_params.increase_ref_index,
                 mask_image,
-                get_resolved_width(),
-                get_resolved_height(),
+                gen_params.width,
+                gen_params.height,
                 gen_params.sample_params,
                 gen_params.strength,
                 gen_params.seed,
@@ -794,7 +695,6 @@ int main(int argc, const char** argv) {
                 item["b64_json"] = b64;
                 out["data"].push_back(item);
             }
-            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;
@@ -833,8 +733,8 @@ 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", default_gen_params.sample_params.sample_steps);
-            float cfg_scale             = j.value("cfg_scale", default_gen_params.sample_params.guidance.txt_cfg);
+            int steps                   = j.value("steps", -1);
+            float cfg_scale             = j.value("cfg_scale", 7.f);
             int64_t seed                = j.value("seed", -1);
             int batch_size              = j.value("batch_size", 1);
             int clip_skip               = j.value("clip_skip", -1);
@@ -867,38 +767,6 @@ int main(int argc, const char** argv) {
                 return bad("prompt required");
             }
 
-            std::vector<sd_lora_t> sd_loras;
-            std::vector<std::string> lora_path_storage;
-
-            if (j.contains("lora") && j["lora"].is_array()) {
-                for (const auto& item : j["lora"]) {
-                    if (!item.is_object()) {
-                        continue;
-                    }
-
-                    std::string path   = item.value("path", "");
-                    float multiplier   = item.value("multiplier", 1.0f);
-                    bool is_high_noise = item.value("is_high_noise", false);
-
-                    if (path.empty()) {
-                        return bad("lora.path required");
-                    }
-
-                    std::string fullpath = get_lora_full_path(path);
-                    if (fullpath.empty()) {
-                        return bad("invalid lora path: " + path);
-                    }
-
-                    lora_path_storage.push_back(fullpath);
-                    sd_lora_t l;
-                    l.is_high_noise = is_high_noise;
-                    l.multiplier    = multiplier;
-                    l.path          = lora_path_storage.back().c_str();
-
-                    sd_loras.push_back(l);
-                }
-            }
-
             auto get_sample_method = [](std::string name) -> enum sample_method_t {
                 enum sample_method_t result = str_to_sample_method(name.c_str());
                 if (result != SAMPLE_METHOD_COUNT) return result;
@@ -917,11 +785,7 @@ int main(int argc, const char** argv) {
                     {"lcm", LCM_SAMPLE_METHOD},
                     {"ddim", DDIM_TRAILING_SAMPLE_METHOD},
                     {"dpm++ 2m", DPMPP2M_SAMPLE_METHOD},
-                    {"k_dpmpp_2m", DPMPP2M_SAMPLE_METHOD},
-                    {"res multistep", RES_MULTISTEP_SAMPLE_METHOD},
-                    {"k_res_multistep", RES_MULTISTEP_SAMPLE_METHOD},
-                    {"res 2s", RES_2S_SAMPLE_METHOD},
-                    {"k_res_2s", RES_2S_SAMPLE_METHOD}};
+                    {"k_dpmpp_2m", DPMPP2M_SAMPLE_METHOD}};
                 auto it            = hardcoded.find(name);
                 if (it != hardcoded.end()) return it->second;
                 return SAMPLE_METHOD_COUNT;
@@ -931,13 +795,16 @@ int main(int argc, const char** argv) {
 
             enum scheduler_t scheduler = str_to_scheduler(scheduler_name.c_str());
 
-            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;
+            // 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;
 
             if (clip_skip > 0) {
                 gen_params.clip_skip = clip_skip;
@@ -951,66 +818,38 @@ 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    = {0, 0, 3, nullptr};
-            sd_image_t control_image = {0, 0, 3, nullptr};
-            sd_image_t mask_image    = {0, 0, 1, nullptr};
+            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};
             std::vector<uint8_t> mask_data;
             std::vector<sd_image_t> pmid_images;
             std::vector<sd_image_t> ref_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;
-            };
-
-            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) {
+                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;
+                };
+
                 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);
@@ -1026,15 +865,23 @@ int main(int argc, const char** argv) {
                         }
                     }
                 } else {
-                    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_data          = std::vector<uint8_t>(width * height, 255);
+                    mask_image.width   = width;
+                    mask_image.height  = 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);
@@ -1042,19 +889,9 @@ 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()),
+                gen_params.lora_vec.data(),
+                static_cast<uint32_t>(gen_params.lora_vec.size()),
                 gen_params.prompt.c_str(),
                 gen_params.negative_prompt.c_str(),
                 gen_params.clip_skip,
@@ -1064,8 +901,8 @@ int main(int argc, const char** argv) {
                 gen_params.auto_resize_ref_image,
                 gen_params.increase_ref_index,
                 mask_image,
-                get_resolved_width(),
-                get_resolved_height(),
+                gen_params.width,
+                gen_params.height,
                 gen_params.sample_params,
                 gen_params.strength,
                 gen_params.seed,
@@ -1115,7 +952,6 @@ int main(int argc, const char** argv) {
                 std::string b64 = base64_encode(image_bytes);
                 out["images"].push_back(b64);
             }
-            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;
@@ -1147,23 +983,6 @@ int main(int argc, const char** argv) {
         sdapi_any2img(req, res, true);
     });
 
-    svr.Get("/sdapi/v1/loras", [&](const httplib::Request&, httplib::Response& res) {
-        refresh_lora_cache();
-
-        json result = json::array();
-        {
-            std::lock_guard<std::mutex> lock(lora_mutex);
-            for (const auto& e : lora_cache) {
-                json item;
-                item["name"] = e.name;
-                item["path"] = e.path;
-                result.push_back(item);
-            }
-        }
-
-        res.set_content(result.dump(), "application/json");
-    });
-
     svr.Get("/sdapi/v1/samplers", [&](const httplib::Request&, httplib::Response& res) {
         std::vector<std::string> sampler_names;
         sampler_names.push_back("default");

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)
+
     

flux.hpp

@@ -4,7 +4,7 @@
 #include <memory>
 #include <vector>
 
-#include "common_dit.hpp"
+#include "ggml_extend.hpp"
 #include "model.h"
 #include "rope.hpp"
 
@@ -103,13 +103,11 @@ namespace Flux {
             auto norm     = std::dynamic_pointer_cast<QKNorm>(blocks["norm"]);
 
             auto qkv         = qkv_proj->forward(ctx, x);
-            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);
+            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]);
             q                = norm->query_norm(ctx, q);
             k                = norm->key_norm(ctx, k);
             return {q, k, v};
@@ -493,14 +491,15 @@ 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;
 
-            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 = 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]
 
             q         = norm->query_norm(ctx, q);
             k         = norm->key_norm(ctx, k);
@@ -847,6 +846,70 @@ 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,
@@ -997,7 +1060,7 @@ namespace Flux {
             int pad_h      = (patch_size - H % patch_size) % patch_size;
             int pad_w      = (patch_size - W % patch_size) % patch_size;
 
-            auto img      = DiT::pad_to_patch_size(ctx, x, params.patch_size, params.patch_size);
+            auto img      = pad_to_patch_size(ctx, x);
             auto orig_img = img;
 
             if (params.chroma_radiance_params.fake_patch_size_x2) {
@@ -1019,7 +1082,7 @@ 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    = DiT::patchify(ctx->ggml_ctx, orig_img, patch_size, patch_size);  // [N, num_patches, C * patch_size * patch_size]
+            auto nerf_pixels    = patchify(ctx->ggml_ctx, orig_img);  // [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,
@@ -1039,7 +1102,7 @@ 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 = 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]
+            img_dct = unpatchify(ctx->ggml_ctx, img_dct, (H + pad_h) / patch_size, (W + pad_w) / patch_size);                               // [N, nerf_hidden_size, H, W]
 
             out = nerf_final_layer_conv->forward(ctx, img_dct);  // [N, C, H, W]
 
@@ -1071,7 +1134,7 @@ namespace Flux {
             int pad_h      = (patch_size - H % patch_size) % patch_size;
             int pad_w      = (patch_size - W % patch_size) % patch_size;
 
-            auto img           = DiT::pad_and_patchify(ctx, x, patch_size, patch_size);
+            auto img           = process_img(ctx, x);
             int64_t img_tokens = img->ne[1];
 
             if (params.version == VERSION_FLUX_FILL) {
@@ -1079,8 +1142,8 @@ 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 = DiT::pad_and_patchify(ctx, masked, patch_size, patch_size);
-                mask   = DiT::pad_and_patchify(ctx, mask, patch_size, patch_size);
+                masked = process_img(ctx, masked);
+                mask   = process_img(ctx, mask);
 
                 img = ggml_concat(ctx->ggml_ctx, img, ggml_concat(ctx->ggml_ctx, masked, mask, 0), 0);
             } else if (params.version == VERSION_FLEX_2) {
@@ -1089,21 +1152,21 @@ 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  = 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);
+                masked  = process_img(ctx, masked);
+                mask    = process_img(ctx, mask);
+                control = process_img(ctx, control);
 
                 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 = DiT::pad_and_patchify(ctx, c_concat, patch_size, patch_size);
+                auto control = process_img(ctx, c_concat);
                 img          = ggml_concat(ctx->ggml_ctx, img, control, 0);
             }
 
             if (ref_latents.size() > 0) {
                 for (ggml_tensor* ref : ref_latents) {
-                    ref = DiT::pad_and_patchify(ctx, ref, patch_size, patch_size);
+                    ref = process_img(ctx, ref);
                     img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                 }
             }
@@ -1115,7 +1178,8 @@ namespace Flux {
                 out = ggml_cont(ctx->ggml_ctx, out);
             }
 
-            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, patch_size, patch_size);  // [N, C, H, W]
+            // 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]
             return out;
         }
 

format-code.sh

@@ -1,4 +1,4 @@
-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
+for f in *.cpp *.h *.hpp 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

ggml_extend.hpp

@@ -491,16 +491,12 @@ __STATIC_INLINE__ void ggml_ext_tensor_split_2d(struct ggml_tensor* input,
     int64_t height   = output->ne[1];
     int64_t channels = output->ne[2];
     int64_t ne3      = output->ne[3];
-
-    int64_t input_width  = input->ne[0];
-    int64_t input_height = input->ne[1];
-
     GGML_ASSERT(input->type == GGML_TYPE_F32 && output->type == GGML_TYPE_F32);
     for (int iy = 0; iy < height; iy++) {
         for (int ix = 0; ix < width; ix++) {
             for (int k = 0; k < channels; k++) {
                 for (int l = 0; l < ne3; l++) {
-                    float value = ggml_ext_tensor_get_f32(input, (ix + x) % input_width, (iy + y) % input_height, k, l);
+                    float value = ggml_ext_tensor_get_f32(input, ix + x, iy + y, k, l);
                     ggml_ext_tensor_set_f32(output, value, ix, iy, k, l);
                 }
             }
@@ -520,8 +516,6 @@ __STATIC_INLINE__ void ggml_ext_tensor_merge_2d(struct ggml_tensor* input,
                                                 int y,
                                                 int overlap_x,
                                                 int overlap_y,
-                                                bool circular_x,
-                                                bool circular_y,
                                                 int x_skip = 0,
                                                 int y_skip = 0) {
     int64_t width    = input->ne[0];
@@ -539,12 +533,12 @@ __STATIC_INLINE__ void ggml_ext_tensor_merge_2d(struct ggml_tensor* input,
                 for (int l = 0; l < ne3; l++) {
                     float new_value = ggml_ext_tensor_get_f32(input, ix, iy, k, l);
                     if (overlap_x > 0 || overlap_y > 0) {  // blend colors in overlapped area
-                        float old_value = ggml_ext_tensor_get_f32(output, (x + ix) % img_width, (y + iy) % img_height, k, l);
+                        float old_value = ggml_ext_tensor_get_f32(output, x + ix, y + iy, k, l);
 
-                        const float x_f_0 = (circular_x || (overlap_x > 0 && x > 0)) ? (ix - x_skip) / float(overlap_x) : 1;
-                        const float x_f_1 = (circular_x || (overlap_x > 0 && x < (img_width - width))) ? (width - ix) / float(overlap_x) : 1;
-                        const float y_f_0 = (circular_y || (overlap_y > 0 && y > 0)) ? (iy - y_skip) / float(overlap_y) : 1;
-                        const float y_f_1 = (circular_y || (overlap_y > 0 && y < (img_height - height))) ? (height - iy) / float(overlap_y) : 1;
+                        const float x_f_0 = (overlap_x > 0 && x > 0) ? (ix - x_skip) / float(overlap_x) : 1;
+                        const float x_f_1 = (overlap_x > 0 && x < (img_width - width)) ? (width - ix) / float(overlap_x) : 1;
+                        const float y_f_0 = (overlap_y > 0 && y > 0) ? (iy - y_skip) / float(overlap_y) : 1;
+                        const float y_f_1 = (overlap_y > 0 && y < (img_height - height)) ? (height - iy) / float(overlap_y) : 1;
 
                         const float x_f = std::min(std::min(x_f_0, x_f_1), 1.f);
                         const float y_f = std::min(std::min(y_f_0, y_f_1), 1.f);
@@ -552,9 +546,9 @@ __STATIC_INLINE__ void ggml_ext_tensor_merge_2d(struct ggml_tensor* input,
                         ggml_ext_tensor_set_f32(
                             output,
                             old_value + new_value * smootherstep_f32(y_f) * smootherstep_f32(x_f),
-                            (x + ix) % img_width, (y + iy) % img_height, k, l);
+                            x + ix, y + iy, k, l);
                     } else {
-                        ggml_ext_tensor_set_f32(output, new_value, (x + ix) % img_width, (y + iy) % img_height, k, l);
+                        ggml_ext_tensor_set_f32(output, new_value, x + ix, y + iy, k, l);
                     }
                 }
             }
@@ -773,37 +767,16 @@ __STATIC_INLINE__ ggml_tensor* ggml_ext_silu_act(ggml_context* ctx, ggml_tensor*
     return x;
 }
 
-typedef std::function<bool(ggml_tensor*, ggml_tensor*, bool)> on_tile_process;
+typedef std::function<void(ggml_tensor*, ggml_tensor*, bool)> on_tile_process;
 
 __STATIC_INLINE__ void sd_tiling_calc_tiles(int& num_tiles_dim,
                                             float& tile_overlap_factor_dim,
                                             int small_dim,
                                             int tile_size,
-                                            const float tile_overlap_factor,
-                                            bool circular) {
+                                            const float tile_overlap_factor) {
     int tile_overlap     = static_cast<int>(tile_size * tile_overlap_factor);
     int non_tile_overlap = tile_size - tile_overlap;
 
-    if (circular) {
-        // circular means the last and first tile are overlapping (wraping around)
-        num_tiles_dim = small_dim / non_tile_overlap;
-
-        if (num_tiles_dim < 1) {
-            num_tiles_dim = 1;
-        }
-
-        tile_overlap_factor_dim = (tile_size - small_dim / num_tiles_dim) / (float)tile_size;
-
-        // if single tile and tile_overlap_factor is not 0, add one to ensure we have at least two overlapping tiles
-        if (num_tiles_dim == 1 && tile_overlap_factor_dim > 0) {
-            num_tiles_dim++;
-            tile_overlap_factor_dim = 0.5;
-        }
-
-        return;
-    }
-    // else, non-circular means the last and first tile are not overlapping
-
     num_tiles_dim     = (small_dim - tile_overlap) / non_tile_overlap;
     int overshoot_dim = ((num_tiles_dim + 1) * non_tile_overlap + tile_overlap) % small_dim;
 
@@ -832,8 +805,6 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
                                             const int p_tile_size_x,
                                             const int p_tile_size_y,
                                             const float tile_overlap_factor,
-                                            const bool circular_x,
-                                            const bool circular_y,
                                             on_tile_process on_processing) {
     output = ggml_set_f32(output, 0);
 
@@ -858,11 +829,11 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
 
     int num_tiles_x;
     float tile_overlap_factor_x;
-    sd_tiling_calc_tiles(num_tiles_x, tile_overlap_factor_x, small_width, p_tile_size_x, tile_overlap_factor, circular_x);
+    sd_tiling_calc_tiles(num_tiles_x, tile_overlap_factor_x, small_width, p_tile_size_x, tile_overlap_factor);
 
     int num_tiles_y;
     float tile_overlap_factor_y;
-    sd_tiling_calc_tiles(num_tiles_y, tile_overlap_factor_y, small_height, p_tile_size_y, tile_overlap_factor, circular_y);
+    sd_tiling_calc_tiles(num_tiles_y, tile_overlap_factor_y, small_height, p_tile_size_y, tile_overlap_factor);
 
     LOG_DEBUG("num tiles : %d, %d ", num_tiles_x, num_tiles_y);
     LOG_DEBUG("optimal overlap : %f, %f (targeting %f)", tile_overlap_factor_x, tile_overlap_factor_y, tile_overlap_factor);
@@ -916,7 +887,7 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
     float last_time = 0.0f;
     for (int y = 0; y < small_height && !last_y; y += non_tile_overlap_y) {
         int dy = 0;
-        if (!circular_y && y + tile_size_y >= small_height) {
+        if (y + tile_size_y >= small_height) {
             int _y = y;
             y      = small_height - tile_size_y;
             dy     = _y - y;
@@ -927,7 +898,7 @@ __STATIC_INLINE__ void sd_tiling_non_square(ggml_tensor* input,
         }
         for (int x = 0; x < small_width && !last_x; x += non_tile_overlap_x) {
             int dx = 0;
-            if (!circular_x && x + tile_size_x >= small_width) {
+            if (x + tile_size_x >= small_width) {
                 int _x = x;
                 x      = small_width - tile_size_x;
                 dx     = _x - x;
@@ -947,15 +918,12 @@ __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);
-            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, circular_x, circular_y, dx, dy);
+            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);
-            } else {
-                LOG_ERROR("Failed to process patch %d at (%d, %d)", tile_count, x, y);
-            }
+            int64_t t2 = ggml_time_ms();
+            last_time  = (t2 - t1) / 1000.0f;
+            pretty_progress(tile_count, num_tiles, last_time);
             tile_count++;
         }
         last_x = false;
@@ -971,10 +939,8 @@ __STATIC_INLINE__ void sd_tiling(ggml_tensor* input,
                                  const int scale,
                                  const int tile_size,
                                  const float tile_overlap_factor,
-                                 const bool circular_x,
-                                 const bool circular_y,
                                  on_tile_process on_processing) {
-    sd_tiling_non_square(input, output, scale, tile_size, tile_size, tile_overlap_factor, circular_x, circular_y, on_processing);
+    sd_tiling_non_square(input, output, scale, tile_size, tile_size, tile_overlap_factor, on_processing);
 }
 
 __STATIC_INLINE__ struct ggml_tensor* ggml_ext_group_norm_32(struct ggml_context* ctx,
@@ -1250,11 +1216,6 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_zeros(struct ggml_context* ctx,
     return ggml_ext_full(ctx, 0.f, ne0, ne1, ne2, ne3);
 }
 
-__STATIC_INLINE__ struct ggml_tensor* ggml_ext_zeros_like(struct ggml_context* ctx,
-                                                          struct ggml_tensor* x) {
-    return ggml_ext_zeros(ctx, x->ne[0], x->ne[1], x->ne[2], x->ne[3]);
-}
-
 __STATIC_INLINE__ struct ggml_tensor* ggml_ext_ones(struct ggml_context* ctx,
                                                     int64_t ne0,
                                                     int64_t ne1,
@@ -1263,11 +1224,6 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_ext_ones(struct ggml_context* ctx,
     return ggml_ext_full(ctx, 1.f, ne0, ne1, ne2, ne3);
 }
 
-__STATIC_INLINE__ struct ggml_tensor* ggml_ext_ones_like(struct ggml_context* ctx,
-                                                         struct ggml_tensor* x) {
-    return ggml_ext_ones(ctx, x->ne[0], x->ne[1], x->ne[2], x->ne[3]);
-}
-
 __STATIC_INLINE__ ggml_tensor* ggml_ext_cast_f32(ggml_context* ctx, ggml_tensor* a) {
 #ifdef SD_USE_VULKAN
     auto zero_index = ggml_get_tensor(ctx, "ggml_runner_build_in_tensor:zero_int");
@@ -1621,7 +1577,7 @@ struct WeightAdapter {
             bool force_prec_f32 = false;
             float scale         = 1.f;
         } linear;
-        struct conv2d_params_t {
+        struct {
             int s0          = 1;
             int s1          = 1;
             int p0          = 0;
@@ -2667,180 +2623,11 @@ public:
             v = v_proj->forward(ctx, x);
         }
 
-        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, false);  // [N, n_token, embed_dim]
+        x = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask);  // [N, n_token, embed_dim]
 
         x = out_proj->forward(ctx, x);  // [N, n_token, embed_dim]
         return x;
     }
 };
 
-__STATIC_INLINE__ struct ggml_tensor* ggml_ext_lokr_forward(
-    struct ggml_context* ctx,
-    struct ggml_tensor* h,    // Input: [q, batch] or [W, H, q, batch]
-    struct ggml_tensor* w1,   // Outer C (Full rank)
-    struct ggml_tensor* w1a,  // Outer A (Low rank part 1)
-    struct ggml_tensor* w1b,  // Outer B (Low rank part 2)
-    struct ggml_tensor* w2,   // Inner BA (Full rank)
-    struct ggml_tensor* w2a,  // Inner A (Low rank part 1)
-    struct ggml_tensor* w2b,  // Inner B (Low rank part 2)
-    bool is_conv,
-    WeightAdapter::ForwardParams::conv2d_params_t conv_params,
-    float scale) {
-    GGML_ASSERT((w1 != NULL || (w1a != NULL && w1b != NULL)));
-    GGML_ASSERT((w2 != NULL || (w2a != NULL && w2b != NULL)));
-
-    int uq = (w1 != NULL) ? (int)w1->ne[0] : (int)w1a->ne[0];
-    int up = (w1 != NULL) ? (int)w1->ne[1] : (int)w1b->ne[1];
-
-    int q_actual = is_conv ? (int)h->ne[2] : (int)h->ne[0];
-    int vq       = q_actual / uq;
-
-    int vp = (w2 != NULL) ? (is_conv ? (int)w2->ne[3] : (int)w2->ne[1])
-                          : (int)w2a->ne[1];
-    GGML_ASSERT(q_actual == (uq * vq) && "Input dimension mismatch for LoKR split");
-
-    struct ggml_tensor* hb;
-
-    if (!is_conv) {
-        int batch          = (int)h->ne[1];
-        int merge_batch_uq = batch;
-        int merge_batch_vp = batch;
-
-#if SD_USE_VULKAN
-        if (batch > 1) {
-            // no access to backend here, worst case is slightly worse perfs for other backends when built alongside Vulkan backend
-            int max_batch    = 65535;
-            int max_batch_uq = max_batch / uq;
-            merge_batch_uq   = 1;
-            for (int i = max_batch_uq; i > 0; i--) {
-                if (batch % i == 0) {
-                    merge_batch_uq = i;
-                    break;
-                }
-            }
-
-            int max_batch_vp = max_batch / vp;
-            merge_batch_vp   = 1;
-            for (int i = max_batch_vp; i > 0; i--) {
-                if (batch % i == 0) {
-                    merge_batch_vp = i;
-                    break;
-                }
-            }
-        }
-#endif
-
-        struct ggml_tensor* h_split = ggml_reshape_3d(ctx, h, vq, uq * merge_batch_uq, batch / merge_batch_uq);
-        if (w2 != NULL) {
-            hb = ggml_mul_mat(ctx, w2, h_split);
-        } else {
-            hb = ggml_mul_mat(ctx, w2b, ggml_mul_mat(ctx, w2a, h_split));
-        }
-
-        if (batch > 1) {
-            hb = ggml_reshape_3d(ctx, hb, vp, uq, batch);
-        }
-        struct ggml_tensor* hb_t = ggml_cont(ctx, ggml_transpose(ctx, hb));
-        hb_t                     = ggml_reshape_3d(ctx, hb_t, uq, vp * merge_batch_vp, batch / merge_batch_vp);
-
-        struct ggml_tensor* hc_t;
-        if (w1 != NULL) {
-            hc_t = ggml_mul_mat(ctx, w1, hb_t);
-        } else {
-            hc_t = ggml_mul_mat(ctx, w1b, ggml_mul_mat(ctx, w1a, hb_t));
-        }
-
-        if (batch > 1) {
-            hc_t = ggml_reshape_3d(ctx, hc_t, up, vp, batch);
-        }
-
-        struct ggml_tensor* hc  = ggml_transpose(ctx, hc_t);
-        struct ggml_tensor* out = ggml_reshape_2d(ctx, ggml_cont(ctx, hc), up * vp, batch);
-        return ggml_scale(ctx, out, scale);
-    } else {
-        int batch = (int)h->ne[3];
-        // 1. Reshape input: [W, H, vq*uq, batch] -> [W, H, vq, uq * batch]
-        struct ggml_tensor* h_split = ggml_reshape_4d(ctx, h, h->ne[0], h->ne[1], vq, uq * batch);
-
-        if (w2 != NULL) {
-            hb = ggml_ext_conv_2d(ctx, h_split, w2, nullptr,
-                                  conv_params.s0,
-                                  conv_params.s1,
-                                  conv_params.p0,
-                                  conv_params.p1,
-                                  conv_params.d0,
-                                  conv_params.d1,
-                                  conv_params.direct,
-                                  conv_params.circular_x,
-                                  conv_params.circular_y,
-                                  conv_params.scale);
-        } else {
-            // swap a and b order for conv lora
-            struct ggml_tensor* a = w2b;
-            struct ggml_tensor* b = w2a;
-
-            // unpack conv2d weights if needed
-            if (ggml_n_dims(a) < 4) {
-                int k = (int)sqrt(a->ne[0] / h_split->ne[2]);
-                GGML_ASSERT(k * k * h_split->ne[2] == a->ne[0]);
-                a = ggml_reshape_4d(ctx, a, k, k, a->ne[0] / (k * k), a->ne[1]);
-            } else if (a->ne[2] != h_split->ne[2]) {
-                int k = (int)sqrt(a->ne[2] / h_split->ne[2]);
-                GGML_ASSERT(k * k * h_split->ne[2] == a->ne[2]);
-                a = ggml_reshape_4d(ctx, a, a->ne[0] * k, a->ne[1] * k, a->ne[2] / (k * k), a->ne[3]);
-            }
-            struct ggml_tensor* ha = ggml_ext_conv_2d(ctx, h_split, a, nullptr,
-                                                      conv_params.s0,
-                                                      conv_params.s1,
-                                                      conv_params.p0,
-                                                      conv_params.p1,
-                                                      conv_params.d0,
-                                                      conv_params.d1,
-                                                      conv_params.direct,
-                                                      conv_params.circular_x,
-                                                      conv_params.circular_y,
-                                                      conv_params.scale);
-
-            // not supporting lora_mid here
-            hb = ggml_ext_conv_2d(ctx,
-                                  ha,
-                                  b,
-                                  nullptr,
-                                  1,
-                                  1,
-                                  0,
-                                  0,
-                                  1,
-                                  1,
-                                  conv_params.direct,
-                                  conv_params.circular_x,
-                                  conv_params.circular_y,
-                                  conv_params.scale);
-        }
-
-        // Current hb shape: [W_out, H_out, vp, uq * batch]
-        int w_out = (int)hb->ne[0];
-        int h_out = (int)hb->ne[1];
-
-        // struct ggml_tensor* hb_cat = ggml_reshape_4d(ctx, hb, w_out , h_out , vp * uq, batch);
-        // [W_out, H_out, vp * uq,  batch]
-        // Now left to compute (W1 kr Id) * hb_cat == (W1 kr W2) cv h
-
-        // merge the uq groups of size vp*w_out*h_out
-        struct ggml_tensor* hb_merged = ggml_reshape_2d(ctx, hb, w_out * h_out * vp, uq * batch);
-        struct ggml_tensor* hc_t;
-        struct ggml_tensor* hb_merged_t = ggml_cont(ctx, ggml_transpose(ctx, hb_merged));
-        if (w1 != NULL) {
-            // Would be great to be able to transpose w1 instead to avoid transposing both hb and hc
-            hc_t = ggml_mul_mat(ctx, w1, hb_merged_t);
-        } else {
-            hc_t = ggml_mul_mat(ctx, w1b, ggml_mul_mat(ctx, w1a, hb_merged_t));
-        }
-        struct ggml_tensor* hc = ggml_transpose(ctx, hc_t);
-        // ungroup
-        struct ggml_tensor* out = ggml_reshape_4d(ctx, ggml_cont(ctx, hc), w_out, h_out, up * vp, batch);
-        return ggml_scale(ctx, out, scale);
-    }
-}
-
 #endif  // __GGML_EXTEND__HPP__

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);
+            }
+        }
+    }
+}

llm.hpp

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

lora.hpp

@@ -468,10 +468,10 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
+    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora = true) {
         // lora
         ggml_tensor* diff = nullptr;
-        if (with_lora_and_lokr) {
+        if (with_lora) {
             diff = get_lora_weight_diff(model_tensor_name, ctx);
         }
         // diff
@@ -483,7 +483,7 @@ struct LoraModel : public GGMLRunner {
             diff = get_loha_weight_diff(model_tensor_name, ctx);
         }
         // lokr
-        if (diff == nullptr && with_lora_and_lokr) {
+        if (diff == nullptr) {
             diff = get_lokr_weight_diff(model_tensor_name, ctx);
         }
         if (diff != nullptr) {
@@ -514,108 +514,6 @@ struct LoraModel : public GGMLRunner {
             } else {
                 key = model_tensor_name + "." + std::to_string(index);
             }
-            bool is_conv2d = forward_params.op_type == WeightAdapter::ForwardParams::op_type_t::OP_CONV2D;
-
-            std::string lokr_w1_name   = "lora." + key + ".lokr_w1";
-            std::string lokr_w1_a_name = "lora." + key + ".lokr_w1_a";
-            // if either of these is found, then we have a lokr lora
-            auto iter   = lora_tensors.find(lokr_w1_name);
-            auto iter_a = lora_tensors.find(lokr_w1_a_name);
-            if (iter != lora_tensors.end() || iter_a != lora_tensors.end()) {
-                std::string lokr_w1_b_name = "lora." + key + ".lokr_w1_b";
-                std::string lokr_w2_name   = "lora." + key + ".lokr_w2";
-                std::string lokr_w2_a_name = "lora." + key + ".lokr_w2_a";
-                std::string lokr_w2_b_name = "lora." + key + ".lokr_w2_b";
-                std::string alpha_name     = "lora." + key + ".alpha";
-
-                ggml_tensor* lokr_w1   = nullptr;
-                ggml_tensor* lokr_w1_a = nullptr;
-                ggml_tensor* lokr_w1_b = nullptr;
-                ggml_tensor* lokr_w2   = nullptr;
-                ggml_tensor* lokr_w2_a = nullptr;
-                ggml_tensor* lokr_w2_b = nullptr;
-
-                if (iter != lora_tensors.end()) {
-                    lokr_w1 = iter->second;
-                }
-                iter = iter_a;
-                if (iter != lora_tensors.end()) {
-                    lokr_w1_a = iter->second;
-                }
-                iter = lora_tensors.find(lokr_w1_b_name);
-                if (iter != lora_tensors.end()) {
-                    lokr_w1_b = iter->second;
-                }
-
-                iter = lora_tensors.find(lokr_w2_name);
-                if (iter != lora_tensors.end()) {
-                    lokr_w2 = iter->second;
-                    if (is_conv2d && lokr_w2->type != GGML_TYPE_F16) {
-                        lokr_w2 = ggml_cast(ctx, lokr_w2, GGML_TYPE_F16);
-                    }
-                }
-                iter = lora_tensors.find(lokr_w2_a_name);
-                if (iter != lora_tensors.end()) {
-                    lokr_w2_a = iter->second;
-                    if (is_conv2d && lokr_w2_a->type != GGML_TYPE_F16) {
-                        lokr_w2_a = ggml_cast(ctx, lokr_w2_a, GGML_TYPE_F16);
-                    }
-                }
-                iter = lora_tensors.find(lokr_w2_b_name);
-                if (iter != lora_tensors.end()) {
-                    lokr_w2_b = iter->second;
-                    if (is_conv2d && lokr_w2_b->type != GGML_TYPE_F16) {
-                        lokr_w2_b = ggml_cast(ctx, lokr_w2_b, GGML_TYPE_F16);
-                    }
-                }
-
-                int rank = 1;
-                if (lokr_w1_b) {
-                    rank = (int)lokr_w1_b->ne[ggml_n_dims(lokr_w1_b) - 1];
-                }
-                if (lokr_w2_b) {
-                    rank = (int)lokr_w2_b->ne[ggml_n_dims(lokr_w2_b) - 1];
-                }
-
-                float scale_value = 1.0f;
-                iter              = lora_tensors.find(alpha_name);
-                if (iter != lora_tensors.end()) {
-                    float alpha = ggml_ext_backend_tensor_get_f32(iter->second);
-                    scale_value = alpha / rank;
-                    applied_lora_tensors.insert(alpha_name);
-                }
-
-                if (rank == 1) {
-                    scale_value = 1.0f;
-                }
-                scale_value *= multiplier;
-
-                auto curr_out_diff = ggml_ext_lokr_forward(ctx, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
-                if (out_diff == nullptr) {
-                    out_diff = curr_out_diff;
-                } else {
-                    out_diff = ggml_concat(ctx, out_diff, curr_out_diff, 0);
-                }
-
-                if (lokr_w1)
-                    applied_lora_tensors.insert(lokr_w1_name);
-                if (lokr_w1_a)
-                    applied_lora_tensors.insert(lokr_w1_a_name);
-                if (lokr_w1_b)
-                    applied_lora_tensors.insert(lokr_w1_b_name);
-                if (lokr_w2)
-                    applied_lora_tensors.insert(lokr_w2_name);
-                if (lokr_w2_a)
-                    applied_lora_tensors.insert(lokr_w2_name);
-                if (lokr_w2_b)
-                    applied_lora_tensors.insert(lokr_w2_b_name);
-                applied_lora_tensors.insert(alpha_name);
-
-                index++;
-                continue;
-            }
-
-            // not a lokr, normal lora path
 
             std::string lora_down_name = "lora." + key + ".lora_down";
             std::string lora_up_name   = "lora." + key + ".lora_up";
@@ -627,7 +525,9 @@ struct LoraModel : public GGMLRunner {
             ggml_tensor* lora_mid  = nullptr;
             ggml_tensor* lora_down = nullptr;
 
-            iter = lora_tensors.find(lora_up_name);
+            bool is_conv2d = forward_params.op_type == WeightAdapter::ForwardParams::op_type_t::OP_CONV2D;
+
+            auto iter = lora_tensors.find(lora_up_name);
             if (iter != lora_tensors.end()) {
                 lora_up = iter->second;
                 if (is_conv2d && lora_up->type != GGML_TYPE_F16) {
@@ -841,9 +741,9 @@ public:
         : lora_models(lora_models) {
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name, bool with_lora) {
         for (auto& lora_model : lora_models) {
-            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, ctx, weight, with_lora_and_lokr);
+            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, ctx, weight, with_lora);
             if (diff == nullptr) {
                 continue;
             }

ltxv.hpp

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

mmdit.hpp

@@ -745,6 +745,28 @@ 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,
@@ -789,11 +811,11 @@ 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]
@@ -812,7 +834,7 @@ public:
 
         x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 
-        x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, patch_size, patch_size, /*patch_last*/ false);  // [N, C, H, W]
+        x = unpatchify(ctx->ggml_ctx, x, h, w);  // [N, C, H, W]
 
         return x;
     }

model.cpp

@@ -16,6 +16,10 @@
 #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"
@@ -1035,6 +1039,8 @@ SDVersion ModelLoader::get_sd_version() {
     bool is_xl                       = false;
     bool is_flux                     = false;
     bool is_flux2                    = false;
+    bool is_z_image                  = false;
+    bool is_z_image_omni             = false;
     bool has_single_block_47         = false;
     bool is_wan                      = false;
     int64_t patch_embedding_channels = 0;
@@ -1057,9 +1063,6 @@ 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;
             }
@@ -1070,7 +1073,10 @@ SDVersion ModelLoader::get_sd_version() {
                 return VERSION_OVIS_IMAGE;
             }
             if (tensor_storage.name.find("model.diffusion_model.cap_embedder.0.weight") != std::string::npos) {
-                return VERSION_Z_IMAGE;
+                is_z_image = true;
+            }
+            if (tensor_storage.name.find("model.diffusion_model.siglip_embedder.0.weight") != std::string::npos) {
+                is_z_image_omni = true;
             }
             if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
                 is_wan = true;
@@ -1173,6 +1179,13 @@ SDVersion ModelLoader::get_sd_version() {
         return VERSION_FLUX2_KLEIN;
     }
 
+    if (is_z_image) {
+        if (is_z_image_omni) {
+            return VERSION_Z_IMAGE_OMNI;
+        }
+        return VERSION_Z_IMAGE;
+    }
+
     if (token_embedding_weight.ne[0] == 768) {
         if (is_inpaint) {
             return VERSION_SD1_INPAINT;
@@ -1339,6 +1352,36 @@ 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);

model.h

@@ -45,10 +45,10 @@ enum SDVersion {
     VERSION_WAN2_2_I2V,
     VERSION_WAN2_2_TI2V,
     VERSION_QWEN_IMAGE,
-    VERSION_ANIMA,
     VERSION_FLUX2,
     VERSION_FLUX2_KLEIN,
     VERSION_Z_IMAGE,
+    VERSION_Z_IMAGE_OMNI,
     VERSION_OVIS_IMAGE,
     VERSION_COUNT,
 };
@@ -123,15 +123,8 @@ 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) {
+    if (version == VERSION_Z_IMAGE || version == VERSION_Z_IMAGE_OMNI) {
         return true;
     }
     return false;
@@ -154,7 +147,6 @@ 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;
     }
@@ -340,6 +332,13 @@ 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__

name_conversion.cpp

@@ -653,14 +653,6 @@ 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);
@@ -672,8 +664,6 @@ 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;
 }
@@ -852,7 +842,6 @@ std::string convert_sep_to_dot(std::string name) {
         "conv_in",
         "conv_out",
         "lora_down",
-        "lora_mid",
         "lora_up",
         "diff_b",
         "hada_w1_a",
@@ -1008,13 +997,10 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
     if (is_lora) {
         std::map<std::string, std::string> lora_suffix_map = {
             {".lora_down.weight", ".weight.lora_down"},
-            {".lora_mid.weight", ".weight.lora_mid"},
             {".lora_up.weight", ".weight.lora_up"},
             {".lora.down.weight", ".weight.lora_down"},
-            {".lora.mid.weight", ".weight.lora_mid"},
             {".lora.up.weight", ".weight.lora_up"},
             {"_lora.down.weight", ".weight.lora_down"},
-            {"_lora.mid.weight", ".weight.lora_mid"},
             {"_lora.up.weight", ".weight.lora_up"},
             {".lora_A.weight", ".weight.lora_down"},
             {".lora_B.weight", ".weight.lora_up"},

qwen_image.hpp

@@ -3,8 +3,9 @@
 
 #include <memory>
 
-#include "common_block.hpp"
+#include "common.hpp"
 #include "flux.hpp"
+#include "ggml_extend.hpp"
 
 namespace Qwen {
     constexpr int QWEN_IMAGE_GRAPH_SIZE = 20480;
@@ -211,7 +212,7 @@ namespace Qwen {
 
             blocks["txt_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
             blocks["txt_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
-            blocks["txt_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU, true));
+            blocks["txt_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU));
 
             blocks["attn"] = std::shared_ptr<GGMLBlock>(new QwenImageAttention(dim,
                                                                                attention_head_dim,
@@ -389,6 +390,69 @@ 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,
@@ -404,7 +468,7 @@ namespace Qwen {
 
             auto t_emb = time_text_embed->forward(ctx, timestep);
             if (params.zero_cond_t) {
-                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep));
+                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros(ctx->ggml_ctx, timestep->ne[0], timestep->ne[1], timestep->ne[2], timestep->ne[3]));
                 t_emb        = ggml_concat(ctx->ggml_ctx, t_emb, t_emb_0, 1);
             }
             auto img = img_in->forward(ctx, x);
@@ -448,16 +512,19 @@ namespace Qwen {
             int64_t C = x->ne[2];
             int64_t N = x->ne[3];
 
-            auto img           = DiT::pad_and_patchify(ctx, x, params.patch_size, params.patch_size);
+            auto img           = process_img(ctx, x);
             int64_t img_tokens = img->ne[1];
 
             if (ref_latents.size() > 0) {
                 for (ggml_tensor* ref : ref_latents) {
-                    ref = DiT::pad_and_patchify(ctx, ref, params.patch_size, params.patch_size);
+                    ref = process_img(ctx, ref);
                     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) {
@@ -466,7 +533,11 @@ 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 = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, params.patch_size, params.patch_size);  // [N, C, H, W]
+            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]
 
             return out;
         }

rope.hpp

@@ -43,7 +43,7 @@ namespace Rope {
 
     __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos,
                                                            int dim,
-                                                           float theta,
+                                                           int 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,
-                                                  const std::vector<float>& axis_thetas,
+                                                  int theta,
                                                   const std::vector<int>& axes_dim,
                                                   const std::vector<std::vector<int>>& wrap_dims = {}) {
         std::vector<std::vector<float>> trans_ids = transpose(ids);
@@ -188,12 +188,8 @@ 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], axis_theta, axis_wrap_dims);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
+                rope(trans_ids[i], axes_dim[i], 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) {
@@ -207,15 +203,6 @@ 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,
@@ -345,7 +332,7 @@ namespace Rope {
                 }
             }
         }
-        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
+        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int h,
@@ -434,7 +421,7 @@ namespace Rope {
                 }
             }
         }
-        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
+        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
@@ -488,7 +475,7 @@ 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, static_cast<float>(theta), axes_dim);
+        return embed_nd(ids, bs, theta, axes_dim);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen2vl_ids(int grid_h,
@@ -524,67 +511,124 @@ 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, static_cast<float>(theta), axes_dim);
+        return embed_nd(ids, 1, theta, axes_dim);
     }
 
     __STATIC_INLINE__ int bound_mod(int a, int m) {
         return (m - (a % m)) % m;
     }
 
-    __STATIC_INLINE__ std::vector<std::vector<float>> gen_z_image_ids(int h,
-                                                                      int w,
-                                                                      int patch_size,
-                                                                      int bs,
-                                                                      int context_len,
-                                                                      int seq_multi_of,
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_z_image_ids(ggml_tensor* x,
+                                                                      const std::vector<ggml_tensor*>& contexts,
                                                                       const std::vector<ggml_tensor*>& ref_latents,
-                                                                      bool increase_ref_index) {
-        int padded_context_len = context_len + bound_mod(context_len, seq_multi_of);
-        auto txt_ids           = std::vector<std::vector<float>>(bs * padded_context_len, std::vector<float>(3, 0.0f));
-        for (int i = 0; i < bs * padded_context_len; i++) {
-            txt_ids[i][0] = (i % padded_context_len) + 1.f;
+                                                                      const std::vector<ggml_tensor*>& siglip_feats,
+                                                                      int patch_size,
+                                                                      int seq_multi_of,
+                                                                      int bs) {
+        GGML_ASSERT(contexts.size() > ref_latents.size());
+        GGML_ASSERT(contexts.size() >= siglip_feats.size());
+        int context_cu_len = 1;
+        std::vector<int> context_end_pos;
+        std::vector<std::vector<float>> txt_ids;
+        for (auto context : contexts) {
+            int padded_context_len = static_cast<int>(context->ne[1]) + bound_mod(static_cast<int>(context->ne[1]), seq_multi_of);
+            auto curr_txt_ids      = std::vector<std::vector<float>>(bs * padded_context_len, std::vector<float>(3, 0.0f));
+            for (int i = 0; i < bs * padded_context_len; i++) {
+                curr_txt_ids[i][0] = static_cast<float>((i % padded_context_len) + context_cu_len);
+            }
+            context_cu_len += padded_context_len;
+            context_end_pos.push_back(context_cu_len);
+            context_cu_len += 2;  // for image and siglip tokens
+            txt_ids = concat_ids(txt_ids, curr_txt_ids, bs);
         }
 
-        int axes_dim_num = 3;
-        int index        = padded_context_len + 1;
-        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, index);
+        std::vector<std::vector<float>> img_ids;
+        std::vector<ggml_tensor*> all_img = ref_latents;
+        all_img.push_back(x);
+        for (int i = 0; i < all_img.size(); i++) {
+            int axes_dim_num  = 3;
+            int index         = context_end_pos[i];
+            auto curr_img_ids = gen_flux_img_ids(static_cast<int>(all_img[i]->ne[1]), static_cast<int>(all_img[i]->ne[0]), patch_size, bs, axes_dim_num, index);
 
-        int img_pad_len = bound_mod(static_cast<int>(img_ids.size() / bs), seq_multi_of);
-        if (img_pad_len > 0) {
-            std::vector<std::vector<float>> img_pad_ids(bs * img_pad_len, std::vector<float>(3, 0.f));
-            img_ids = concat_ids(img_ids, img_pad_ids, bs);
+            int img_pad_len = bound_mod(static_cast<int>(curr_img_ids.size() / bs), seq_multi_of);
+            if (img_pad_len > 0) {
+                std::vector<std::vector<float>> img_pad_ids(bs * img_pad_len, std::vector<float>(3, 0.f));
+                curr_img_ids = concat_ids(curr_img_ids, img_pad_ids, bs);
+            }
+            img_ids = concat_ids(img_ids, curr_img_ids, bs);
+        }
+
+        std::vector<std::vector<float>> sig_ids;
+        for (int i = 0; i < siglip_feats.size(); i++) {
+            int axes_dim_num = 3;
+            int index        = context_end_pos[i] + 1;
+            int h_len        = static_cast<int>(siglip_feats[i]->ne[1]);
+            int w_len        = static_cast<int>(siglip_feats[i]->ne[0]);
+
+            std::vector<std::vector<float>> curr_sig_ids(bs * h_len * w_len, std::vector<float>(axes_dim_num, 0.0));
+
+            // scale position IDs to match img resolution
+            std::vector<float> row_ids = linspace<float>(0, static_cast<float>(all_img[i]->ne[1]) - 1.f, h_len);
+            std::vector<float> col_ids = linspace<float>(0, static_cast<float>(all_img[i]->ne[0]) - 1.f, w_len);
+
+            for (int ib = 0; ib < bs; ++ib) {
+                for (int ih = 0; ih < h_len; ++ih) {
+                    for (int iw = 0; iw < w_len; ++iw) {
+                        curr_sig_ids[ib * h_len * w_len + ih * w_len + iw][0] = static_cast<float>(index);
+                        curr_sig_ids[ib * h_len * w_len + ih * w_len + iw][1] = row_ids[ih];
+                        curr_sig_ids[ib * h_len * w_len + ih * w_len + iw][2] = col_ids[iw];
+                    }
+                }
+            }
+
+            int sig_pad_len = bound_mod(static_cast<int>(curr_sig_ids.size() / bs), seq_multi_of);
+            if (sig_pad_len > 0) {
+                std::vector<std::vector<float>> sig_pad_ids(bs * sig_pad_len, std::vector<float>(3, 0.f));
+                curr_sig_ids = concat_ids(curr_sig_ids, sig_pad_ids, bs);
+            }
+            sig_ids = concat_ids(sig_ids, curr_sig_ids, bs);
         }
 
         auto ids = concat_ids(txt_ids, img_ids, bs);
 
-        // ignore ref_latents for now
+        if (!sig_ids.empty()) {
+            ids = concat_ids(ids, sig_ids, bs);
+        }
+
         return ids;
     }
 
     // Generate z_image positional embeddings
-    __STATIC_INLINE__ std::vector<float> gen_z_image_pe(int h,
-                                                        int w,
-                                                        int patch_size,
-                                                        int bs,
-                                                        int context_len,
-                                                        int seq_multi_of,
+    __STATIC_INLINE__ std::vector<float> gen_z_image_pe(ggml_tensor* x,
+                                                        const std::vector<ggml_tensor*>& contexts,
                                                         const std::vector<ggml_tensor*>& ref_latents,
-                                                        bool increase_ref_index,
+                                                        const std::vector<ggml_tensor*>& siglip_feats,
+                                                        int patch_size,
+                                                        int seq_multi_of,
                                                         int theta,
+                                                        const std::vector<int>& axes_dim,
                                                         bool circular_h,
                                                         bool circular_w,
-                                                        const std::vector<int>& axes_dim) {
-        std::vector<std::vector<float>> ids = gen_z_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, ref_latents, increase_ref_index);
+                                                        int bs) {
+        std::vector<std::vector<float>> ids = gen_z_image_ids(x, contexts, ref_latents, siglip_feats, patch_size, seq_multi_of, bs);
         std::vector<std::vector<int>> wrap_dims;
         if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int context_len = 0;
+            for (auto context : contexts) {
+                int padded_context_len = static_cast<int>(context->ne[1]) + bound_mod(static_cast<int>(context->ne[1]), seq_multi_of);
+                context_len += padded_context_len;
+            }
+            int h     = static_cast<int>(x->ne[1]);
+            int w     = static_cast<int>(x->ne[0]);
             int pad_h = (patch_size - (h % patch_size)) % patch_size;
             int pad_w = (patch_size - (w % patch_size)) % patch_size;
             int h_len = (h + pad_h) / patch_size;
             int w_len = (w + pad_w) / patch_size;
+
             if (h_len > 0 && w_len > 0) {
                 size_t pos_len = ids.size() / bs;
                 wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
-                size_t cursor     = context_len + bound_mod(context_len, seq_multi_of);  // skip text (and its padding)
+                size_t cursor     = context_len;  // skip text (and its padding)
                 size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
                 for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
                     if (circular_h) {
@@ -597,7 +641,7 @@ namespace Rope {
             }
         }
 
-        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
+        return embed_nd(ids, bs, theta, axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ struct ggml_tensor* apply_rope(struct ggml_context* ctx,

src/anima.hpp

@@ -1,686 +0,0 @@
-#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/common_dit.hpp

@@ -1,108 +0,0 @@
-#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/vocab/vocab.cpp

@@ -1,35 +0,0 @@
-#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

@@ -1,13 +0,0 @@
-#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__

stable-diffusion.cpp

@@ -48,10 +48,10 @@ const char* model_version_to_str[] = {
     "Wan 2.2 I2V",
     "Wan 2.2 TI2V",
     "Qwen Image",
-    "Anima",
     "Flux.2",
     "Flux.2 klein",
     "Z-Image",
+    "Z-Image-Omni",
     "Ovis Image",
 };
 
@@ -68,8 +68,6 @@ const char* sampling_methods_str[] = {
     "LCM",
     "DDIM \"trailing\"",
     "TCD",
-    "Res Multistep",
-    "Res 2s",
 };
 
 /*================================================== Helper Functions ================================================*/
@@ -108,18 +106,13 @@ public:
 
     SDVersion version;
     bool vae_decode_only         = false;
-    bool external_vae_is_invalid = false;
     bool free_params_immediately = false;
 
-    bool circular_x = false;
-    bool circular_y = false;
-
     std::shared_ptr<RNG> rng         = std::make_shared<PhiloxRNG>();
     std::shared_ptr<RNG> sampler_rng = nullptr;
     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
@@ -327,7 +320,6 @@ 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;
             }
         }
 
@@ -409,7 +401,6 @@ 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;
@@ -453,7 +444,7 @@ public:
                     }
                 }
                 if (is_chroma) {
-                    if ((sd_ctx_params->flash_attn || sd_ctx_params->diffusion_flash_attn) && sd_ctx_params->chroma_use_dit_mask) {
+                    if (sd_ctx_params->diffusion_flash_attn && sd_ctx_params->chroma_use_dit_mask) {
                         LOG_WARN(
                             "!!!It looks like you are using Chroma with flash attention. "
                             "This is currently unsupported. "
@@ -540,14 +531,6 @@ 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,
@@ -587,6 +570,14 @@ public:
                 }
             }
 
+            if (sd_ctx_params->diffusion_flash_attn) {
+                LOG_INFO("Using flash attention in the diffusion model");
+                diffusion_model->set_flash_attn_enabled(true);
+                if (high_noise_diffusion_model) {
+                    high_noise_diffusion_model->set_flash_attn_enabled(true);
+                }
+            }
+
             cond_stage_model->alloc_params_buffer();
             cond_stage_model->get_param_tensors(tensors);
 
@@ -610,7 +601,7 @@ public:
             }
 
             if (!(use_tiny_autoencoder || version == VERSION_SDXS) || tae_preview_only) {
-                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
+                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
                     first_stage_model = std::make_shared<WAN::WanVAERunner>(vae_backend,
                                                                             offload_params_to_cpu,
                                                                             tensor_storage_map,
@@ -635,10 +626,10 @@ 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 || external_vae_is_invalid)) {
+                        (strlen(SAFE_STR(sd_ctx_params->vae_path)) == 0 || sd_ctx_params->force_sdxl_vae_conv_scale)) {
                         float vae_conv_2d_scale = 1.f / 32.f;
                         LOG_WARN(
-                            "No valid VAE specified with --vae or --force-sdxl-vae-conv-scale flag set, "
+                            "No 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);
@@ -648,7 +639,7 @@ public:
                 }
             }
             if (use_tiny_autoencoder || version == VERSION_SDXS) {
-                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
+                if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
                     tae_first_stage = std::make_shared<TinyVideoAutoEncoder>(vae_backend,
                                                                              offload_params_to_cpu,
                                                                              tensor_storage_map,
@@ -733,28 +724,6 @@ public:
                 pmid_model->get_param_tensors(tensors, "pmid");
             }
 
-            if (sd_ctx_params->flash_attn) {
-                LOG_INFO("Using flash attention");
-                cond_stage_model->set_flash_attention_enabled(true);
-                if (clip_vision) {
-                    clip_vision->set_flash_attention_enabled(true);
-                }
-                if (first_stage_model) {
-                    first_stage_model->set_flash_attention_enabled(true);
-                }
-                if (tae_first_stage) {
-                    tae_first_stage->set_flash_attention_enabled(true);
-                }
-            }
-
-            if (sd_ctx_params->flash_attn || sd_ctx_params->diffusion_flash_attn) {
-                LOG_INFO("Using flash attention in the diffusion model");
-                diffusion_model->set_flash_attention_enabled(true);
-                if (high_noise_diffusion_model) {
-                    high_noise_diffusion_model->set_flash_attention_enabled(true);
-                }
-            }
-
             diffusion_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
             if (high_noise_diffusion_model) {
                 high_noise_diffusion_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
@@ -762,8 +731,12 @@ public:
             if (control_net) {
                 control_net->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
             }
-            circular_x = sd_ctx_params->circular_x;
-            circular_y = sd_ctx_params->circular_y;
+            if (first_stage_model) {
+                first_stage_model->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
+            }
+            if (tae_first_stage) {
+                tae_first_stage->set_circular_axes(sd_ctx_params->circular_x, sd_ctx_params->circular_y);
+            }
         }
 
         struct ggml_init_params params;
@@ -891,6 +864,7 @@ 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)) {
@@ -913,22 +887,24 @@ 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 (sd_version_is_wan(version)) {
-                        default_flow_shift = 5.f;
-                    } else {
-                        default_flow_shift = 3.f;
+                    if (flow_shift == INFINITY) {
+                        if (sd_version_is_wan(version)) {
+                            flow_shift = 5.f;
+                        } else {
+                            flow_shift = 3.f;
+                        }
                     }
                 } else if (sd_version_is_flux(version)) {
                     pred_type = FLUX_FLOW_PRED;
 
-                    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;
+                    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;
+                            }
                         }
                     }
                 } else if (sd_version_is_flux2(version)) {
@@ -952,12 +928,12 @@ public:
                     break;
                 case FLOW_PRED: {
                     LOG_INFO("running in FLOW mode");
-                    denoiser = std::make_shared<DiscreteFlowDenoiser>();
+                    denoiser = std::make_shared<DiscreteFlowDenoiser>(flow_shift);
                     break;
                 }
                 case FLUX_FLOW_PRED: {
                     LOG_INFO("running in Flux FLOW mode");
-                    denoiser = std::make_shared<FluxFlowDenoiser>();
+                    denoiser = std::make_shared<FluxFlowDenoiser>(flow_shift);
                     break;
                 }
                 case FLUX2_FLOW_PRED: {
@@ -1097,18 +1073,6 @@ 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;
         }
@@ -1478,7 +1442,7 @@ public:
         sd_progress_cb_t cb = sd_get_progress_callback();
         void* cbd           = sd_get_progress_callback_data();
         sd_set_progress_callback((sd_progress_cb_t)suppress_pp, nullptr);
-        sd_tiling(input, output, scale, tile_size, tile_overlap_factor, circular_x, circular_y, on_processing);
+        sd_tiling(input, output, scale, tile_size, tile_overlap_factor, on_processing);
         sd_set_progress_callback(cb, cbd);
     }
 
@@ -1525,7 +1489,7 @@ 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) || sd_version_is_anima(version)) {
+                } else if (sd_version_is_wan(version) || sd_version_is_qwen_image(version)) {
                     latent_rgb_proj = wan_21_latent_rgb_proj;
                     latent_rgb_bias = wan_21_latent_rgb_bias;
                 } else {
@@ -1579,7 +1543,7 @@ 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) {
-                        return first_stage_model->compute(n_threads, in, true, &out, nullptr);
+                        first_stage_model->compute(n_threads, in, true, &out, nullptr);
                     };
                     silent_tiling(latents, result, get_vae_scale_factor(), 32, 0.5f, on_tiling);
 
@@ -1598,7 +1562,7 @@ 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) {
-                        return tae_first_stage->compute(n_threads, in, true, &out, nullptr);
+                        tae_first_stage->compute(n_threads, in, true, &out, nullptr);
                     };
                     silent_tiling(latents, result, get_vae_scale_factor(), 64, 0.5f, on_tiling);
                 } else {
@@ -1652,12 +1616,13 @@ public:
                         const std::vector<float>& sigmas,
                         int start_merge_step,
                         SDCondition id_cond,
-                        std::vector<ggml_tensor*> ref_latents = {},
-                        bool increase_ref_index               = false,
-                        ggml_tensor* denoise_mask             = nullptr,
-                        ggml_tensor* vace_context             = nullptr,
-                        float vace_strength                   = 1.f,
-                        const sd_cache_params_t* cache_params = nullptr) {
+                        std::vector<ggml_tensor*> ref_latents    = {},
+                        std::vector<ggml_tensor*> ref_clip_feats = {},
+                        bool increase_ref_index                  = false,
+                        ggml_tensor* denoise_mask                = nullptr,
+                        ggml_tensor* vace_context                = nullptr,
+                        float vace_strength                      = 1.f,
+                        const sd_cache_params_t* cache_params    = nullptr) {
         if (shifted_timestep > 0 && !sd_version_is_sdxl(version)) {
             LOG_WARN("timestep shifting is only supported for SDXL models!");
             shifted_timestep = 0;
@@ -2006,9 +1971,6 @@ 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 {
@@ -2049,6 +2011,7 @@ public:
             diffusion_params.timesteps          = timesteps;
             diffusion_params.guidance           = guidance_tensor;
             diffusion_params.ref_latents        = ref_latents;
+            diffusion_params.ref_clip_feats     = ref_clip_feats;
             diffusion_params.increase_ref_index = increase_ref_index;
             diffusion_params.controls           = controls;
             diffusion_params.control_strength   = control_strength;
@@ -2059,10 +2022,11 @@ public:
             struct ggml_tensor** active_output  = &out_cond;
             if (start_merge_step == -1 || step <= start_merge_step) {
                 // cond
-                diffusion_params.context  = cond.c_crossattn;
-                diffusion_params.c_concat = cond.c_concat;
-                diffusion_params.y        = cond.c_vector;
-                active_condition          = &cond;
+                diffusion_params.context        = cond.c_crossattn;
+                diffusion_params.extra_contexts = cond.extra_c_crossattns;
+                diffusion_params.c_concat       = cond.c_concat;
+                diffusion_params.y              = cond.c_vector;
+                active_condition                = &cond;
             } else {
                 diffusion_params.context  = id_cond.c_crossattn;
                 diffusion_params.c_concat = cond.c_concat;
@@ -2093,12 +2057,13 @@ public:
                         LOG_ERROR("controlnet compute failed");
                     }
                 }
-                current_step_skipped      = cache_step_is_skipped();
-                diffusion_params.controls = controls;
-                diffusion_params.context  = uncond.c_crossattn;
-                diffusion_params.c_concat = uncond.c_concat;
-                diffusion_params.y        = uncond.c_vector;
-                bool skip_uncond          = cache_before_condition(&uncond, out_uncond);
+                current_step_skipped            = cache_step_is_skipped();
+                diffusion_params.controls       = controls;
+                diffusion_params.context        = uncond.c_crossattn;
+                diffusion_params.extra_contexts = uncond.extra_c_crossattns;
+                diffusion_params.c_concat       = uncond.c_concat;
+                diffusion_params.y              = uncond.c_vector;
+                bool skip_uncond                = cache_before_condition(&uncond, out_uncond);
                 if (!skip_uncond) {
                     if (!work_diffusion_model->compute(n_threads,
                                                        diffusion_params,
@@ -2113,10 +2078,11 @@ public:
 
             float* img_cond_data = nullptr;
             if (has_img_cond) {
-                diffusion_params.context  = img_cond.c_crossattn;
-                diffusion_params.c_concat = img_cond.c_concat;
-                diffusion_params.y        = img_cond.c_vector;
-                bool skip_img_cond        = cache_before_condition(&img_cond, out_img_cond);
+                diffusion_params.context        = img_cond.c_crossattn;
+                diffusion_params.extra_contexts = img_cond.extra_c_crossattns;
+                diffusion_params.c_concat       = img_cond.c_concat;
+                diffusion_params.y              = img_cond.c_vector;
+                bool skip_img_cond              = cache_before_condition(&img_cond, out_img_cond);
                 if (!skip_img_cond) {
                     if (!work_diffusion_model->compute(n_threads,
                                                        diffusion_params,
@@ -2420,7 +2386,7 @@ public:
     }
 
     void process_latent_in(ggml_tensor* latent) {
-        if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version) || sd_version_is_flux2(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_qwen_image(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;
@@ -2459,7 +2425,7 @@ public:
     }
 
     void process_latent_out(ggml_tensor* latent) {
-        if (sd_version_is_wan(version) || sd_version_is_qwen_image(version) || sd_version_is_anima(version) || sd_version_is_flux2(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_qwen_image(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;
@@ -2537,7 +2503,7 @@ public:
             // TODO wan2.2 vae support?
             int64_t ne2;
             int64_t ne3;
-            if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
+            if (sd_version_is_qwen_image(version)) {
                 ne2 = 1;
                 ne3 = C * x->ne[3];
             } else {
@@ -2555,7 +2521,7 @@ public:
             result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, ne2, ne3);
         }
 
-        if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
+        if (sd_version_is_qwen_image(version)) {
             x = ggml_reshape_4d(work_ctx, x, x->ne[0], x->ne[1], 1, x->ne[2] * x->ne[3]);
         }
 
@@ -2570,9 +2536,9 @@ 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) {
-                    return first_stage_model->compute(n_threads, in, false, &out, work_ctx);
+                    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, circular_x, circular_y, on_tiling);
+                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, false, &result, work_ctx);
             }
@@ -2581,9 +2547,9 @@ 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) {
-                    return tae_first_stage->compute(n_threads, in, false, &out, nullptr);
+                    tae_first_stage->compute(n_threads, in, false, &out, nullptr);
                 };
-                sd_tiling(x, result, vae_scale_factor, 64, 0.5f, circular_x, circular_y, on_tiling);
+                sd_tiling(x, result, vae_scale_factor, 64, 0.5f, on_tiling);
             } else {
                 tae_first_stage->compute(n_threads, x, false, &result, work_ctx);
             }
@@ -2628,7 +2594,6 @@ 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) {
@@ -2648,7 +2613,7 @@ public:
         if (!use_tiny_autoencoder) {
             process_latent_in(latent);
         }
-        if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
+        if (sd_version_is_qwen_image(version)) {
             latent = ggml_reshape_4d(work_ctx, latent, latent->ne[0], latent->ne[1], latent->ne[3], 1);
         }
         return latent;
@@ -2686,7 +2651,7 @@ public:
         }
         int64_t t0 = ggml_time_ms();
         if (!use_tiny_autoencoder) {
-            if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
+            if (sd_version_is_qwen_image(version)) {
                 x = ggml_reshape_4d(work_ctx, x, x->ne[0], x->ne[1], 1, x->ne[2] * x->ne[3]);
             }
             process_latent_out(x);
@@ -2700,15 +2665,11 @@ public:
 
                 // split latent in 32x32 tiles and compute in several steps
                 auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
-                    return first_stage_model->compute(n_threads, in, true, &out, nullptr);
+                    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, circular_x, circular_y, on_tiling);
+                sd_tiling_non_square(x, result, vae_scale_factor, tile_size_x, tile_size_y, tile_overlap, on_tiling);
             } else {
-                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->compute(n_threads, x, true, &result, work_ctx);
             }
             first_stage_model->free_compute_buffer();
             process_vae_output_tensor(result);
@@ -2716,15 +2677,11 @@ 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) {
-                    return tae_first_stage->compute(n_threads, in, true, &out);
+                    tae_first_stage->compute(n_threads, in, true, &out);
                 };
-                sd_tiling(x, result, vae_scale_factor, 64, 0.5f, circular_x, circular_y, on_tiling);
+                sd_tiling(x, result, vae_scale_factor, 64, 0.5f, on_tiling);
             } else {
-                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->compute(n_threads, x, true, &result);
             }
             tae_first_stage->free_compute_buffer();
         }
@@ -2734,16 +2691,6 @@ 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 ==================================================*/
@@ -2802,8 +2749,6 @@ const char* sample_method_to_str[] = {
     "lcm",
     "ddim_trailing",
     "tcd",
-    "res_multistep",
-    "res_2s",
 };
 
 const char* sd_sample_method_name(enum sample_method_t sample_method) {
@@ -2833,7 +2778,6 @@ const char* scheduler_to_str[] = {
     "smoothstep",
     "kl_optimal",
     "lcm",
-    "bong_tangent",
 };
 
 const char* sd_scheduler_name(enum scheduler_t scheduler) {
@@ -2964,6 +2908,7 @@ 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) {
@@ -2998,7 +2943,6 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              "keep_clip_on_cpu: %s\n"
              "keep_control_net_on_cpu: %s\n"
              "keep_vae_on_cpu: %s\n"
-             "flash_attn: %s\n"
              "diffusion_flash_attn: %s\n"
              "circular_x: %s\n"
              "circular_y: %s\n"
@@ -3030,7 +2974,6 @@ char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params) {
              BOOL_STR(sd_ctx_params->keep_clip_on_cpu),
              BOOL_STR(sd_ctx_params->keep_control_net_on_cpu),
              BOOL_STR(sd_ctx_params->keep_vae_on_cpu),
-             BOOL_STR(sd_ctx_params->flash_attn),
              BOOL_STR(sd_ctx_params->diffusion_flash_attn),
              BOOL_STR(sd_ctx_params->circular_x),
              BOOL_STR(sd_ctx_params->circular_y),
@@ -3055,7 +2998,6 @@ 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) {
@@ -3076,8 +3018,7 @@ char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
              "sample_method: %s, "
              "sample_steps: %d, "
              "eta: %.2f, "
-             "shifted_timestep: %d, "
-             "flow_shift: %.2f)",
+             "shifted_timestep: %d)",
              sample_params->guidance.txt_cfg,
              std::isfinite(sample_params->guidance.img_cfg)
                  ? sample_params->guidance.img_cfg
@@ -3091,8 +3032,7 @@ 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->flow_shift);
+             sample_params->shifted_timestep);
 
     return buf;
 }
@@ -3262,6 +3202,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
                                     sd_pm_params_t pm_params,
                                     std::vector<sd_image_t*> ref_images,
                                     std::vector<ggml_tensor*> ref_latents,
+                                    std::vector<ggml_tensor*> ref_clip_feats,
                                     bool increase_ref_index,
                                     ggml_tensor* concat_latent            = nullptr,
                                     ggml_tensor* denoise_mask             = nullptr,
@@ -3458,6 +3399,7 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
                                                      start_merge_step,
                                                      id_cond,
                                                      ref_latents,
+                                                     ref_clip_feats,
                                                      increase_ref_index,
                                                      denoise_mask,
                                                      nullptr,
@@ -3506,7 +3448,6 @@ 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;
@@ -3521,9 +3462,8 @@ sd_image_t* generate_image_internal(sd_ctx_t* sd_ctx,
 
 sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_gen_params) {
     sd_ctx->sd->vae_tiling_params = sd_img_gen_params->vae_tiling_params;
-
-    int width  = sd_img_gen_params->width;
-    int height = sd_img_gen_params->height;
+    int width                     = sd_img_gen_params->width;
+    int height                    = sd_img_gen_params->height;
 
     int vae_scale_factor            = sd_ctx->sd->get_vae_scale_factor();
     int diffusion_model_down_factor = sd_ctx->sd->get_diffusion_model_down_factor();
@@ -3537,40 +3477,6 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
         LOG_WARN("align up %dx%d to %dx%d (multiple=%d)", sd_img_gen_params->width, sd_img_gen_params->height, width, height, spatial_multiple);
     }
 
-    bool circular_x = sd_ctx->sd->circular_x;
-    bool circular_y = sd_ctx->sd->circular_y;
-
-    if (!sd_img_gen_params->vae_tiling_params.enabled) {
-        if (sd_ctx->sd->first_stage_model) {
-            sd_ctx->sd->first_stage_model->set_circular_axes(sd_ctx->sd->circular_x, sd_ctx->sd->circular_y);
-        }
-        if (sd_ctx->sd->tae_first_stage) {
-            sd_ctx->sd->tae_first_stage->set_circular_axes(sd_ctx->sd->circular_x, sd_ctx->sd->circular_y);
-        }
-    } else {
-        int tile_size_x, tile_size_y;
-        float _overlap;
-        int latent_size_x = width / sd_ctx->sd->get_vae_scale_factor();
-        int latent_size_y = height / sd_ctx->sd->get_vae_scale_factor();
-        sd_ctx->sd->get_tile_sizes(tile_size_x, tile_size_y, _overlap, sd_img_gen_params->vae_tiling_params, latent_size_x, latent_size_y);
-
-        // force disable circular padding for vae if tiling is enabled unless latent is smaller than tile size
-        // otherwise it will cause artifacts at the edges of the tiles
-        sd_ctx->sd->circular_x = sd_ctx->sd->circular_x && (tile_size_x >= latent_size_x);
-        sd_ctx->sd->circular_y = sd_ctx->sd->circular_y && (tile_size_y >= latent_size_y);
-
-        if (sd_ctx->sd->first_stage_model) {
-            sd_ctx->sd->first_stage_model->set_circular_axes(sd_ctx->sd->circular_x, sd_ctx->sd->circular_y);
-        }
-        if (sd_ctx->sd->tae_first_stage) {
-            sd_ctx->sd->tae_first_stage->set_circular_axes(sd_ctx->sd->circular_x, sd_ctx->sd->circular_y);
-        }
-
-        // disable circular tiling if it's enabled for the VAE
-        sd_ctx->sd->circular_x = circular_x && (tile_size_x < latent_size_x);
-        sd_ctx->sd->circular_y = circular_y && (tile_size_y < latent_size_y);
-    }
-
     LOG_DEBUG("generate_image %dx%d", width, height);
     if (sd_ctx == nullptr || sd_img_gen_params == nullptr) {
         return nullptr;
@@ -3598,8 +3504,6 @@ 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);
 
@@ -3762,6 +3666,7 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
     }
 
     std::vector<ggml_tensor*> ref_latents;
+    std::vector<ggml_tensor*> ref_clip_feats;
     for (int i = 0; i < ref_images.size(); i++) {
         ggml_tensor* img;
         if (sd_img_gen_params->auto_resize_ref_image) {
@@ -3808,6 +3713,11 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
 
         ggml_tensor* latent = sd_ctx->sd->encode_first_stage(work_ctx, img);
         ref_latents.push_back(latent);
+
+        if (sd_ctx->sd->version == VERSION_Z_IMAGE_OMNI) {
+            auto clip_vision_output = sd_ctx->sd->get_clip_vision_output(work_ctx, *ref_images[i], false, -2);
+            ref_clip_feats.push_back(clip_vision_output);
+        }
     }
 
     if (sd_img_gen_params->init_image.data != nullptr || sd_img_gen_params->ref_images_count > 0) {
@@ -3835,15 +3745,12 @@ sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_g
                                                         sd_img_gen_params->pm_params,
                                                         ref_images,
                                                         ref_latents,
+                                                        ref_clip_feats,
                                                         sd_img_gen_params->increase_ref_index,
                                                         concat_latent,
                                                         denoise_mask,
                                                         &sd_img_gen_params->cache);
 
-    // restore circular params
-    sd_ctx->sd->circular_x = circular_x;
-    sd_ctx->sd->circular_y = circular_y;
-
     size_t t2 = ggml_time_ms();
 
     LOG_INFO("generate_image completed in %.2fs", (t2 - t0) * 1.0f / 1000);
@@ -3879,8 +3786,6 @@ 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);
@@ -4210,8 +4115,9 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
                                  high_noise_sample_method,
                                  high_noise_sigmas,
                                  -1,
-                                 {},
-                                 {},
+                                 {},  // id_cond
+                                 {},  // ref_latents
+                                 {},  // ref_clip_feats
                                  false,
                                  denoise_mask,
                                  vace_context,
@@ -4247,8 +4153,9 @@ SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* s
                                           sample_method,
                                           sigmas,
                                           -1,
-                                          {},
-                                          {},
+                                          {},  // id_cond
+                                          {},  // ref_latents
+                                          {},  // ref_clip_feats
                                           false,
                                           denoise_mask,
                                           vace_context,

stable-diffusion.h

@@ -48,8 +48,6 @@ enum sample_method_t {
     LCM_SAMPLE_METHOD,
     DDIM_TRAILING_SAMPLE_METHOD,
     TCD_SAMPLE_METHOD,
-    RES_MULTISTEP_SAMPLE_METHOD,
-    RES_2S_SAMPLE_METHOD,
     SAMPLE_METHOD_COUNT
 };
 
@@ -64,7 +62,6 @@ enum scheduler_t {
     SMOOTHSTEP_SCHEDULER,
     KL_OPTIMAL_SCHEDULER,
     LCM_SCHEDULER,
-    BONG_TANGENT_SCHEDULER,
     SCHEDULER_COUNT
 };
 
@@ -189,7 +186,6 @@ typedef struct {
     bool keep_clip_on_cpu;
     bool keep_control_net_on_cpu;
     bool keep_vae_on_cpu;
-    bool flash_attn;
     bool diffusion_flash_attn;
     bool tae_preview_only;
     bool diffusion_conv_direct;
@@ -201,6 +197,7 @@ 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 {
@@ -234,7 +231,6 @@ typedef struct {
     int shifted_timestep;
     float* custom_sigmas;
     int custom_sigmas_count;
-    float flow_shift;
 } sd_sample_params_t;
 
 typedef struct {

t5.hpp

@@ -14,7 +14,6 @@
 #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.
@@ -342,9 +341,9 @@ protected:
 public:
     explicit T5UniGramTokenizer(bool is_umt5 = false) {
         if (is_umt5) {
-            InitializePieces(load_umt5_tokenizer_json());
+            InitializePieces(ModelLoader::load_umt5_tokenizer_json());
         } else {
-            InitializePieces(load_t5_tokenizer_json());
+            InitializePieces(ModelLoader::load_t5_tokenizer_json());
         }
 
         min_score_ = FLT_MAX;

tokenize_util.cpp

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

ucache.hpp

@@ -19,7 +19,6 @@ struct UCacheConfig {
     bool adaptive_threshold     = true;
     float early_step_multiplier = 0.5f;
     float late_step_multiplier  = 1.5f;
-    float relative_norm_gain    = 1.6f;
     bool reset_error_on_compute = true;
 };
 
@@ -46,16 +45,14 @@ struct UCacheState {
     bool has_output_prev_norm             = false;
     bool has_relative_transformation_rate = false;
     float relative_transformation_rate    = 0.0f;
+    float cumulative_change_rate          = 0.0f;
     float last_input_change               = 0.0f;
     bool has_last_input_change            = false;
-    float output_change_ema               = 0.0f;
-    bool has_output_change_ema            = false;
     int total_steps_skipped               = 0;
     int current_step_index                = -1;
     int steps_computed_since_active       = 0;
-    int expected_total_steps              = 0;
-    int consecutive_skipped_steps         = 0;
     float accumulated_error               = 0.0f;
+    float reference_output_norm           = 0.0f;
 
     struct BlockMetrics {
         float sum_transformation_rate = 0.0f;
@@ -109,16 +106,14 @@ struct UCacheState {
         has_output_prev_norm             = false;
         has_relative_transformation_rate = false;
         relative_transformation_rate     = 0.0f;
+        cumulative_change_rate           = 0.0f;
         last_input_change                = 0.0f;
         has_last_input_change            = false;
-        output_change_ema                = 0.0f;
-        has_output_change_ema            = false;
         total_steps_skipped              = 0;
         current_step_index               = -1;
         steps_computed_since_active      = 0;
-        expected_total_steps             = 0;
-        consecutive_skipped_steps        = 0;
         accumulated_error                = 0.0f;
+        reference_output_norm            = 0.0f;
         block_metrics.reset();
         total_active_steps = 0;
     }
@@ -138,8 +133,7 @@ struct UCacheState {
         if (!initialized || sigmas.size() < 2) {
             return;
         }
-        size_t n_steps       = sigmas.size() - 1;
-        expected_total_steps = static_cast<int>(n_steps);
+        size_t n_steps = sigmas.size() - 1;
 
         size_t start_step = static_cast<size_t>(config.start_percent * n_steps);
         size_t end_step   = static_cast<size_t>(config.end_percent * n_steps);
@@ -213,15 +207,11 @@ struct UCacheState {
         }
 
         int effective_total = estimated_total_steps;
-        if (effective_total <= 0) {
-            effective_total = expected_total_steps;
-        }
         if (effective_total <= 0) {
             effective_total = std::max(20, steps_computed_since_active * 2);
         }
 
         float progress = (effective_total > 0) ? (static_cast<float>(steps_computed_since_active) / effective_total) : 0.0f;
-        progress       = std::max(0.0f, std::min(1.0f, progress));
 
         float multiplier = 1.0f;
         if (progress < 0.2f) {
@@ -319,31 +309,17 @@ struct UCacheState {
 
         if (has_output_prev_norm && has_relative_transformation_rate &&
             last_input_change > 0.0f && output_prev_norm > 0.0f) {
-            float approx_output_change = relative_transformation_rate * last_input_change;
-            float approx_output_change_rate;
-            if (config.use_relative_threshold) {
-                float base_scale          = std::max(output_prev_norm, 1e-6f);
-                float dyn_scale           = has_output_change_ema
-                                                ? std::max(output_change_ema * std::max(1.0f, config.relative_norm_gain), 1e-6f)
-                                                : base_scale;
-                float scale               = std::sqrt(base_scale * dyn_scale);
-                approx_output_change_rate = approx_output_change / scale;
-            } else {
-                approx_output_change_rate = approx_output_change;
-            }
-            // Increase estimated error with skip horizon to avoid long extrapolation streaks
-            approx_output_change_rate *= (1.0f + 0.50f * consecutive_skipped_steps);
-            accumulated_error = accumulated_error * config.error_decay_rate + approx_output_change_rate;
+            float approx_output_change_rate = (relative_transformation_rate * last_input_change) / output_prev_norm;
+            accumulated_error               = accumulated_error * config.error_decay_rate + approx_output_change_rate;
 
             float effective_threshold = get_adaptive_threshold();
-            if (!config.use_relative_threshold && output_prev_norm > 0.0f) {
-                effective_threshold = effective_threshold * output_prev_norm;
+            if (config.use_relative_threshold && reference_output_norm > 0.0f) {
+                effective_threshold = effective_threshold * reference_output_norm;
             }
 
             if (accumulated_error < effective_threshold) {
                 skip_current_step = true;
                 total_steps_skipped++;
-                consecutive_skipped_steps++;
                 apply_cache(cond, input, output);
                 return true;
             } else if (config.reset_error_on_compute) {
@@ -364,8 +340,6 @@ struct UCacheState {
         if (cond != anchor_condition) {
             return;
         }
-        steps_computed_since_active++;
-        consecutive_skipped_steps = 0;
 
         size_t ne      = static_cast<size_t>(ggml_nelements(input));
         float* in_data = (float*)input->data;
@@ -385,14 +359,6 @@ struct UCacheState {
                 output_change /= static_cast<float>(ne);
             }
         }
-        if (std::isfinite(output_change) && output_change > 0.0f) {
-            if (!has_output_change_ema) {
-                output_change_ema     = output_change;
-                has_output_change_ema = true;
-            } else {
-                output_change_ema = 0.8f * output_change_ema + 0.2f * output_change;
-            }
-        }
 
         prev_output.resize(ne);
         for (size_t i = 0; i < ne; ++i) {
@@ -407,6 +373,10 @@ struct UCacheState {
         output_prev_norm     = (ne > 0) ? (mean_abs / static_cast<float>(ne)) : 0.0f;
         has_output_prev_norm = output_prev_norm > 0.0f;
 
+        if (reference_output_norm == 0.0f) {
+            reference_output_norm = output_prev_norm;
+        }
+
         if (has_last_input_change && last_input_change > 0.0f && output_change > 0.0f) {
             float rate = output_change / last_input_change;
             if (std::isfinite(rate)) {

unet.hpp

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

upscaler.cpp

@@ -89,11 +89,10 @@ 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) {
-            return esrgan_upscaler->compute(n_threads, in, &out);
+            esrgan_upscaler->compute(n_threads, in, &out);
         };
         int64_t t0 = ggml_time_ms();
-        // TODO: circular upscaling?
-        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, false, false, on_tiling);
+        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, on_tiling);
         esrgan_upscaler->free_compute_buffer();
         ggml_ext_tensor_clamp_inplace(upscaled, 0.f, 1.f);
         uint8_t* upscaled_data = ggml_tensor_to_sd_image(upscaled);

vae.hpp

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

vocab.hpp

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

vocab_mistral.hpp

@@ -1,4 +1,4 @@
-static const unsigned char mistral_merges_utf8_c_str[] = {
+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 @@ static const 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,
 };
-static const unsigned char mistral_vocab_json_utf8_c_str[] = {
+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,

vocab_qwen.hpp

@@ -1,4 +1,4 @@
-static const unsigned char qwen2_merges_utf8_c_str[] = {
+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,

vocab_umt5.hpp

@@ -1,4 +1,4 @@
-static const unsigned char umt5_tokenizer_json_str[] = {
+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,

wan.hpp

@@ -5,8 +5,9 @@
 #include <memory>
 #include <utility>
 
-#include "common_block.hpp"
+#include "common.hpp"
 #include "flux.hpp"
+#include "ggml_extend.hpp"
 #include "rope.hpp"
 #include "vae.hpp"
 
@@ -571,8 +572,8 @@ 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, false, 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, true, false);  // [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]

z_image.hpp

@@ -140,14 +140,37 @@ namespace ZImage {
 
     __STATIC_INLINE__ struct ggml_tensor* modulate(struct ggml_context* ctx,
                                                    struct ggml_tensor* x,
-                                                   struct ggml_tensor* scale) {
+                                                   struct ggml_tensor* scale,
+                                                   bool skip_reshape = false) {
         // x: [N, L, C]
-        // scale: [N, C]
-        scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
-        x     = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+        // scale: [N, C] or [N, L, C]
+        if (!skip_reshape) {
+            scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
+        }
+        x = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
         return x;
     }
 
+    __STATIC_INLINE__ struct ggml_tensor* select_per_token(struct ggml_context* ctx,
+                                                           struct ggml_tensor* index,
+                                                           struct ggml_tensor* mod_0,
+                                                           struct ggml_tensor* mod_1) {
+        // index: [N, L]
+        // mod_0/mod_1: [N, C]
+        // return: [N, L, C]
+        // mod_result = torch.where(index == 0, mod_0, mod_1)
+        // mod_result = (1 - index)*mod_0 + index*mod_1
+        index = ggml_reshape_3d(ctx, index, 1, index->ne[0], index->ne[1]);
+        index = ggml_repeat_4d(ctx, index, mod_0->ne[0], index->ne[1], index->ne[2], 1);  // [N, L, C]
+        mod_0 = ggml_reshape_3d(ctx, mod_0, mod_0->ne[0], 1, mod_0->ne[1]);               // [N, 1, C]
+        mod_1 = ggml_reshape_3d(ctx, mod_1, mod_1->ne[0], 1, mod_1->ne[1]);               // [N, 1, C]
+
+        mod_0           = ggml_sub(ctx, ggml_repeat(ctx, mod_0, index), ggml_mul(ctx, index, mod_0));  // [N, L, C]
+        mod_1           = ggml_mul(ctx, index, mod_1);                                                 // [N, L, C]
+        auto mod_result = ggml_add(ctx, mod_0, mod_1);
+        return mod_result;
+    }
+
     struct JointTransformerBlock : public GGMLBlock {
     protected:
         bool modulation;
@@ -179,7 +202,10 @@ namespace ZImage {
                                     struct ggml_tensor* x,
                                     struct ggml_tensor* pe,
                                     struct ggml_tensor* mask        = nullptr,
-                                    struct ggml_tensor* adaln_input = nullptr) {
+                                    struct ggml_tensor* adaln_input = nullptr,
+                                    struct ggml_tensor* noise_mask  = nullptr,
+                                    struct ggml_tensor* adaln_noisy = nullptr,
+                                    struct ggml_tensor* adaln_clean = nullptr) {
             auto attention       = std::dynamic_pointer_cast<JointAttention>(blocks["attention"]);
             auto feed_forward    = std::dynamic_pointer_cast<FeedForward>(blocks["feed_forward"]);
             auto attention_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["attention_norm1"]);
@@ -188,32 +214,55 @@ namespace ZImage {
             auto ffn_norm2       = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm2"]);
 
             if (modulation) {
-                GGML_ASSERT(adaln_input != nullptr);
                 auto adaLN_modulation_0 = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.0"]);
 
-                auto m         = adaLN_modulation_0->forward(ctx, adaln_input);  // [N, 4 * hidden_size]
-                auto mods      = ggml_ext_chunk(ctx->ggml_ctx, m, 4, 0);
-                auto scale_msa = mods[0];
-                auto gate_msa  = mods[1];
-                auto scale_mlp = mods[2];
-                auto gate_mlp  = mods[3];
+                struct ggml_tensor* scale_msa = nullptr;
+                struct ggml_tensor* gate_msa  = nullptr;
+                struct ggml_tensor* scale_mlp = nullptr;
+                struct ggml_tensor* gate_mlp  = nullptr;
+                bool skip_reshape             = false;
+
+                if (noise_mask != nullptr) {
+                    GGML_ASSERT(adaln_noisy != nullptr);
+                    GGML_ASSERT(adaln_clean != nullptr);
+
+                    auto mod_noisy = adaLN_modulation_0->forward(ctx, adaln_noisy);  // [N, 4 * hidden_size]
+                    auto mod_clean = adaLN_modulation_0->forward(ctx, adaln_clean);  // [N, 4 * hidden_size]
+
+                    auto mod_noisy_vec = ggml_ext_chunk(ctx->ggml_ctx, mod_noisy, 4, 0);
+                    auto mod_clean_vec = ggml_ext_chunk(ctx->ggml_ctx, mod_clean, 4, 0);
+
+                    scale_msa = select_per_token(ctx->ggml_ctx, noise_mask, mod_clean_vec[0], mod_noisy_vec[0]);
+                    gate_msa  = select_per_token(ctx->ggml_ctx, noise_mask, mod_clean_vec[1], mod_noisy_vec[1]);
+                    scale_mlp = select_per_token(ctx->ggml_ctx, noise_mask, mod_clean_vec[2], mod_noisy_vec[2]);
+                    gate_mlp  = select_per_token(ctx->ggml_ctx, noise_mask, mod_clean_vec[3], mod_noisy_vec[3]);
+
+                    skip_reshape = true;
+                } else {
+                    GGML_ASSERT(adaln_input != nullptr);
+
+                    auto mod     = adaLN_modulation_0->forward(ctx, adaln_input);  // [N, 4 * hidden_size]
+                    auto mod_vec = ggml_ext_chunk(ctx->ggml_ctx, mod, 4, 0);
+                    scale_msa    = mod_vec[0];
+                    gate_msa     = mod_vec[1];
+                    scale_mlp    = mod_vec[2];
+                    gate_mlp     = mod_vec[3];
+                }
 
                 auto residual = x;
-                x             = modulate(ctx->ggml_ctx, attention_norm1->forward(ctx, x), scale_msa);
+                x             = modulate(ctx->ggml_ctx, attention_norm1->forward(ctx, x), scale_msa, skip_reshape);
                 x             = attention->forward(ctx, x, pe, mask);
                 x             = attention_norm2->forward(ctx, x);
                 x             = ggml_mul(ctx->ggml_ctx, x, ggml_tanh(ctx->ggml_ctx, gate_msa));
                 x             = ggml_add(ctx->ggml_ctx, x, residual);
 
                 residual = x;
-                x        = modulate(ctx->ggml_ctx, ffn_norm1->forward(ctx, x), scale_mlp);
+                x        = modulate(ctx->ggml_ctx, ffn_norm1->forward(ctx, x), scale_mlp, skip_reshape);
                 x        = feed_forward->forward(ctx, x);
                 x        = ffn_norm2->forward(ctx, x);
                 x        = ggml_mul(ctx->ggml_ctx, x, ggml_tanh(ctx->ggml_ctx, gate_mlp));
                 x        = ggml_add(ctx->ggml_ctx, x, residual);
             } else {
-                GGML_ASSERT(adaln_input == nullptr);
-
                 auto residual = x;
                 x             = attention_norm1->forward(ctx, x);
                 x             = attention->forward(ctx, x, pe, mask);
@@ -243,7 +292,10 @@ namespace ZImage {
 
         struct ggml_tensor* forward(GGMLRunnerContext* ctx,
                                     struct ggml_tensor* x,
-                                    struct ggml_tensor* c) {
+                                    struct ggml_tensor* c,
+                                    struct ggml_tensor* noise_mask = nullptr,
+                                    struct ggml_tensor* c_noisy    = nullptr,
+                                    struct ggml_tensor* c_clean    = nullptr) {
             // x: [N, n_token, hidden_size]
             // c: [N, hidden_size]
             // return: [N, n_token, patch_size * patch_size * out_channels]
@@ -251,10 +303,28 @@ namespace ZImage {
             auto linear             = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
             auto adaLN_modulation_1 = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.1"]);
 
-            auto scale = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, hidden_size]
-            x          = norm_final->forward(ctx, x);
-            x          = modulate(ctx->ggml_ctx, x, scale);
-            x          = linear->forward(ctx, x);
+            struct ggml_tensor* scale = nullptr;
+            bool skip_reshape         = false;
+
+            if (noise_mask != nullptr) {
+                GGML_ASSERT(c_noisy != nullptr);
+                GGML_ASSERT(c_clean != nullptr);
+
+                auto scale_noisy = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c_noisy));  // [N, hidden_size]
+                auto scale_clean = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c_clean));  // [N, hidden_size]
+
+                scale = select_per_token(ctx->ggml_ctx, noise_mask, scale_clean, scale_noisy);
+
+                skip_reshape = true;
+            } else {
+                GGML_ASSERT(c != nullptr);
+
+                scale = adaLN_modulation_1->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, hidden_size]
+            }
+
+            x = norm_final->forward(ctx, x);
+            x = modulate(ctx->ggml_ctx, x, scale, skip_reshape);
+            x = linear->forward(ctx, x);
 
             return x;
         }
@@ -275,6 +345,7 @@ namespace ZImage {
         float norm_eps             = 1e-5f;
         bool qk_norm               = true;
         int64_t cap_feat_dim       = 2560;
+        int64_t siglip_feat_dim    = 0;
         int theta                  = 256;
         std::vector<int> axes_dim  = {32, 48, 48};
         int64_t axes_dim_sum       = 128;
@@ -287,6 +358,10 @@ namespace ZImage {
         void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             params["cap_pad_token"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, z_image_params.hidden_size);
             params["x_pad_token"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, z_image_params.hidden_size);
+
+            if (z_image_params.siglip_feat_dim > 0) {
+                params["siglip_pad_token"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, z_image_params.hidden_size);
+            }
         }
 
     public:
@@ -328,6 +403,26 @@ namespace ZImage {
                 blocks["context_refiner." + std::to_string(i)] = block;
             }
 
+            if (z_image_params.siglip_feat_dim > 0) {
+                blocks["siglip_embedder.0"] = std::make_shared<RMSNorm>(z_image_params.siglip_feat_dim, z_image_params.norm_eps);
+                blocks["siglip_embedder.1"] = std::make_shared<Linear>(z_image_params.siglip_feat_dim, z_image_params.hidden_size);
+
+                for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
+                    auto block = std::make_shared<JointTransformerBlock>(2000 + i,
+                                                                         z_image_params.hidden_size,
+                                                                         z_image_params.head_dim,
+                                                                         z_image_params.num_heads,
+                                                                         z_image_params.num_kv_heads,
+                                                                         z_image_params.multiple_of,
+                                                                         z_image_params.ffn_dim_multiplier,
+                                                                         z_image_params.norm_eps,
+                                                                         z_image_params.qk_norm,
+                                                                         false);
+
+                    blocks["siglip_refiner." + std::to_string(i)] = block;
+                }
+            }
+
             for (int i = 0; i < z_image_params.num_layers; i++) {
                 auto block = std::make_shared<JointTransformerBlock>(i,
                                                                      z_image_params.hidden_size,
@@ -346,11 +441,95 @@ 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* forward_core(GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
-                                         struct ggml_tensor* timestep,
-                                         struct ggml_tensor* context,
-                                         struct ggml_tensor* pe) {
+        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;
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> _pad_and_gen_noise_mask(GGMLRunnerContext* ctx,
+                                                                      ggml_tensor* x,
+                                                                      ggml_tensor* pad_token,
+                                                                      int N,
+                                                                      float noise_mask_value = 1.f) {
+            int64_t n_pad_token = Rope::bound_mod(static_cast<int>(x->ne[1]), SEQ_MULTI_OF);
+            if (n_pad_token > 0) {
+                auto pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, pad_token, pad_token->ne[0], n_pad_token, N, 1);
+                x               = ggml_concat(ctx->ggml_ctx, x, pad_tokens, 1);  // [N, n_token + n_pad_token, hidden_size]
+            }
+            ggml_tensor* noise_mask = nullptr;
+            if (noise_mask_value == 0.f) {
+                noise_mask = ggml_ext_zeros(ctx->ggml_ctx, x->ne[1], 1, 1, 1);
+            } else if (noise_mask_value == 1.f) {
+                noise_mask = ggml_ext_ones(ctx->ggml_ctx, x->ne[1], 1, 1, 1);
+            }
+            return {x, noise_mask};
+        }
+
+        struct ggml_tensor* forward_omni(GGMLRunnerContext* ctx,
+                                         ggml_tensor* x,
+                                         ggml_tensor* timestep,
+                                         std::vector<ggml_tensor*> contexts,
+                                         ggml_tensor* pe,
+                                         std::vector<ggml_tensor*> ref_latents,
+                                         std::vector<ggml_tensor*> siglip_feats) {
             auto x_embedder     = std::dynamic_pointer_cast<Linear>(blocks["x_embedder"]);
             auto t_embedder     = std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder"]);
             auto cap_embedder_0 = std::dynamic_pointer_cast<RMSNorm>(blocks["cap_embedder.0"]);
@@ -361,31 +540,145 @@ namespace ZImage {
             auto txt_pad_token = params["cap_pad_token"];
             auto img_pad_token = params["x_pad_token"];
 
-            int64_t N           = x->ne[2];
-            int64_t n_img_token = x->ne[1];
-            int64_t n_txt_token = context->ne[1];
+            bool omni_mode = ref_latents.size() > 0;
 
-            auto t_emb = t_embedder->forward(ctx, timestep);
+            int64_t N = x->ne[2];
 
-            auto txt = cap_embedder_1->forward(ctx, cap_embedder_0->forward(ctx, context));  // [N, n_txt_token, hidden_size]
-            auto img = x_embedder->forward(ctx, x);                                          // [N, n_img_token, hidden_size]
+            // noise mask of img: 0 for condition images (clean), 1 for target image (noisy)
+            // noise mask of txg/sig: same as the corresponding img. If there is no corresponding img, set to 1
 
-            int64_t n_txt_pad_token = Rope::bound_mod(static_cast<int>(n_txt_token), SEQ_MULTI_OF);
-            if (n_txt_pad_token > 0) {
-                auto txt_pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, txt_pad_token, txt_pad_token->ne[0], n_txt_pad_token, N, 1);
-                txt                 = ggml_concat(ctx->ggml_ctx, txt, txt_pad_tokens, 1);  // [N, n_txt_token + n_txt_pad_token, hidden_size]
+            ggml_tensor* txt            = nullptr;
+            ggml_tensor* txt_noise_mask = nullptr;
+            for (int i = 0; i < contexts.size(); i++) {
+                auto curr_txt_raw = cap_embedder_1->forward(ctx, cap_embedder_0->forward(ctx, contexts[i]));  // [N, n_txt_token, hidden_size]
+
+                float noise_mask_value = -1.f;  // empty noise mask
+                if (omni_mode) {
+                    noise_mask_value = (i < ref_latents.size() ? 0.f : 1.f);
+                }
+
+                auto [curr_txt, curr_txt_noise_mask] = _pad_and_gen_noise_mask(ctx, curr_txt_raw, txt_pad_token, static_cast<int>(N), noise_mask_value);
+                if (txt == nullptr) {
+                    txt = curr_txt;
+                } else {
+                    txt = ggml_concat(ctx->ggml_ctx, txt, curr_txt, 1);
+                }
+
+                if (omni_mode) {
+                    if (txt_noise_mask == nullptr) {
+                        txt_noise_mask = curr_txt_noise_mask;
+                    } else {
+                        txt_noise_mask = ggml_concat(ctx->ggml_ctx, txt_noise_mask, curr_txt_noise_mask, 0);
+                    }
+                }
             }
 
-            int64_t n_img_pad_token = Rope::bound_mod(static_cast<int>(n_img_token), SEQ_MULTI_OF);
-            if (n_img_pad_token > 0) {
-                auto img_pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, img_pad_token, img_pad_token->ne[0], n_img_pad_token, N, 1);
-                img                 = ggml_concat(ctx->ggml_ctx, img, img_pad_tokens, 1);  // [N, n_img_token + n_img_pad_token, hidden_size]
+            ggml_tensor* img            = nullptr;
+            ggml_tensor* img_noise_mask = nullptr;
+            for (ggml_tensor* ref : ref_latents) {
+                auto curr_img_raw = x_embedder->forward(ctx, ref);  // [N, n_img_token, hidden_size]
+
+                float noise_mask_value = -1.f;  // empty noise mask
+                if (omni_mode) {
+                    noise_mask_value = 0.f;
+                }
+
+                auto [curr_img, curr_img_noise_mask] = _pad_and_gen_noise_mask(ctx, curr_img_raw, img_pad_token, static_cast<int>(N), noise_mask_value);
+                if (img == nullptr) {
+                    img = curr_img;
+                } else {
+                    img = ggml_concat(ctx->ggml_ctx, img, curr_img, 1);
+                }
+
+                if (omni_mode) {
+                    if (img_noise_mask == nullptr) {
+                        img_noise_mask = curr_img_noise_mask;
+                    } else {
+                        img_noise_mask = ggml_concat(ctx->ggml_ctx, img_noise_mask, curr_img_noise_mask, 0);
+                    }
+                }
             }
 
-            GGML_ASSERT(txt->ne[1] + img->ne[1] == pe->ne[3]);
+            int64_t final_img_offset  = (img ? img->ne[1] : 0);
+            int64_t final_img_pad_len = 0;
+
+            {
+                auto curr_img_raw = x_embedder->forward(ctx, x);  // [N, n_img_token, hidden_size]
+
+                float noise_mask_value = -1.f;  // empty noise mask
+                if (omni_mode) {
+                    noise_mask_value = 0.f;
+                }
+
+                auto [curr_img, curr_img_noise_mask] = _pad_and_gen_noise_mask(ctx, curr_img_raw, img_pad_token, static_cast<int>(N), noise_mask_value);
+                if (img == nullptr) {
+                    img = curr_img;
+                } else {
+                    img = ggml_concat(ctx->ggml_ctx, img, curr_img, 1);
+                }
+
+                if (omni_mode) {
+                    if (img_noise_mask == nullptr) {
+                        img_noise_mask = curr_img_noise_mask;
+                    } else {
+                        img_noise_mask = ggml_concat(ctx->ggml_ctx, img_noise_mask, curr_img_noise_mask, 0);
+                    }
+                }
+
+                final_img_pad_len = Rope::bound_mod(static_cast<int>(curr_img_raw->ne[1]), SEQ_MULTI_OF);
+            }
+
+            ggml_tensor* sig            = nullptr;
+            ggml_tensor* sig_noise_mask = nullptr;
+            for (int i = 0; i < siglip_feats.size(); i++) {
+                auto sig_pad_token     = params["siglip_pad_token"];
+                auto siglip_embedder_0 = std::dynamic_pointer_cast<RMSNorm>(blocks["siglip_embedder.0"]);
+                auto siglip_embedder_1 = std::dynamic_pointer_cast<Linear>(blocks["siglip_embedder.1"]);
+
+                auto curr_sig_raw = siglip_embedder_1->forward(ctx, siglip_embedder_0->forward(ctx, siglip_feats[i]));  // [N, n_sig_token, hidden_size]
+
+                float noise_mask_value = -1.f;  // empty noise mask
+                if (omni_mode) {
+                    noise_mask_value = (i < ref_latents.size() ? 0.f : 1.f);
+                }
+
+                auto [curr_sig, curr_sig_noise_mask] = _pad_and_gen_noise_mask(ctx, curr_sig_raw, sig_pad_token, static_cast<int>(N), noise_mask_value);
+                if (sig == nullptr) {
+                    sig = curr_sig;
+                } else {
+                    sig = ggml_concat(ctx->ggml_ctx, sig, curr_sig, 1);
+                }
+
+                if (omni_mode) {
+                    if (sig_noise_mask == nullptr) {
+                        sig_noise_mask = curr_sig_noise_mask;
+                    } else {
+                        sig_noise_mask = ggml_concat(ctx->ggml_ctx, sig_noise_mask, curr_sig_noise_mask, 0);
+                    }
+                }
+            }
+
+            ggml_tensor* t_emb   = nullptr;
+            ggml_tensor* t_noisy = nullptr;
+            ggml_tensor* t_clean = nullptr;
+            if (omni_mode) {
+                t_noisy = t_embedder->forward(ctx, timestep);
+                t_clean = t_embedder->forward(ctx,
+                                              ggml_scale(ctx->ggml_ctx,
+                                                         ggml_ext_ones(ctx->ggml_ctx, timestep->ne[0], timestep->ne[1], timestep->ne[2], timestep->ne[3]),
+                                                         1000.f));
+            } else {
+                t_emb = t_embedder->forward(ctx, timestep);
+            }
+
+            if (sig) {
+                GGML_ASSERT(txt->ne[1] + img->ne[1] + sig->ne[1] == pe->ne[3]);
+            } else {
+                GGML_ASSERT(txt->ne[1] + img->ne[1] == pe->ne[3]);
+            }
 
             auto txt_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, 0, txt->ne[1]);
-            auto img_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt->ne[1], pe->ne[3]);
+            auto img_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt->ne[1], txt->ne[1] + img->ne[1]);
 
             for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["context_refiner." + std::to_string(i)]);
@@ -396,30 +689,50 @@ namespace ZImage {
             for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["noise_refiner." + std::to_string(i)]);
 
-                img = block->forward(ctx, img, img_pe, nullptr, t_emb);
+                img = block->forward(ctx, img, img_pe, nullptr, t_emb, img_noise_mask, t_noisy, t_clean);
             }
 
-            auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_txt_pad_token + n_img_token + n_img_pad_token, hidden_size]
+            auto unified = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
+
+            ggml_tensor* noise_mask = nullptr;
+            if (omni_mode) {
+                noise_mask = ggml_concat(ctx->ggml_ctx, txt_noise_mask, img_noise_mask, 0);  // [N, n_txt_token + n_img_token]
+            }
+
+            ggml_tensor* sig_pe = nullptr;
+            if (sig) {
+                sig_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt->ne[1] + img->ne[1], pe->ne[3]);
+
+                for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
+                    auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["siglip_refiner." + std::to_string(i)]);
+
+                    sig = block->forward(ctx, sig, sig_pe, nullptr, nullptr);
+                }
+
+                unified    = ggml_concat(ctx->ggml_ctx, unified, sig, 1);                // [N, n_txt_token + n_img_token + n_sig_token, hidden_size]
+                noise_mask = ggml_concat(ctx->ggml_ctx, noise_mask, sig_noise_mask, 0);  // [N, n_txt_token + n_img_token + n_sig_token]
+            }
 
             for (int i = 0; i < z_image_params.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["layers." + std::to_string(i)]);
 
-                txt_img = block->forward(ctx, txt_img, pe, nullptr, t_emb);
+                unified = block->forward(ctx, unified, pe, nullptr, t_emb, noise_mask, t_noisy, t_clean);
             }
 
-            txt_img = final_layer->forward(ctx, txt_img, t_emb);  // [N, n_txt_token + n_txt_pad_token + n_img_token + n_img_pad_token, ph*pw*C]
+            unified = final_layer->forward(ctx, unified, t_emb, noise_mask, t_noisy, t_clean);  // [N, n_txt_token + n_img_token + n_sig_token, ph*pw*C]
 
-            img = ggml_ext_slice(ctx->ggml_ctx, txt_img, 1, n_txt_token + n_txt_pad_token, n_txt_token + n_txt_pad_token + n_img_token);  // [N, n_img_token, ph*pw*C]
+            img = ggml_ext_slice(ctx->ggml_ctx, unified, 1, txt->ne[1] + final_img_offset, txt->ne[1] + img->ne[1] - final_img_pad_len);  // [N, n_final_img_token, ph*pw*C]
 
             return img;
         }
 
         struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* timestep,
-                                    struct ggml_tensor* context,
-                                    struct ggml_tensor* pe,
-                                    std::vector<ggml_tensor*> ref_latents = {}) {
+                                    ggml_tensor* x,
+                                    ggml_tensor* timestep,
+                                    std::vector<ggml_tensor*> contexts,
+                                    ggml_tensor* pe,
+                                    std::vector<ggml_tensor*> ref_latents  = {},
+                                    std::vector<ggml_tensor*> siglip_feats = {}) {
             // Forward pass of DiT.
             // x: [N, C, H, W]
             // timestep: [N,]
@@ -432,22 +745,24 @@ namespace ZImage {
             int64_t C = x->ne[2];
             int64_t N = x->ne[3];
 
-            int patch_size = z_image_params.patch_size;
+            auto img = process_img(ctx, x);
 
-            auto img             = DiT::pad_and_patchify(ctx, x, patch_size, patch_size, false);
-            uint64_t n_img_token = img->ne[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);
 
-            if (ref_latents.size() > 0) {
-                for (ggml_tensor* ref : ref_latents) {
-                    ref = DiT::pad_and_patchify(ctx, ref, patch_size, patch_size, false);
-                    img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
-                }
+            for (int i = 0; i < ref_latents.size(); i++) {
+                ref_latents[i] = process_img(ctx, ref_latents[i]);
             }
 
-            auto out = forward_core(ctx, img, timestep, context, pe);
+            auto out = forward_omni(ctx, img, timestep, contexts, pe, ref_latents, siglip_feats);  // [N, n_img_token, ph*pw*C]
 
-            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]
+            // auto out = forward_basic(ctx, img, timestep, contexts[0], pe);  // [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_scale(ctx->ggml_ctx, out, -1.f);
 
@@ -481,34 +796,37 @@ namespace ZImage {
             z_image.get_param_tensors(tensors, prefix);
         }
 
-        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
-                                        struct ggml_tensor* timesteps,
-                                        struct ggml_tensor* context,
-                                        std::vector<ggml_tensor*> ref_latents = {},
-                                        bool increase_ref_index               = false) {
+        struct ggml_cgraph* build_graph(ggml_tensor* x,
+                                        ggml_tensor* timesteps,
+                                        std::vector<ggml_tensor*> contexts,
+                                        std::vector<ggml_tensor*> ref_latents  = {},
+                                        std::vector<ggml_tensor*> siglip_feats = {}) {
             GGML_ASSERT(x->ne[3] == 1);
             struct ggml_cgraph* gf = new_graph_custom(Z_IMAGE_GRAPH_SIZE);
 
-            x         = to_backend(x);
-            context   = to_backend(context);
+            x = to_backend(x);
+
+            for (int i = 0; i < contexts.size(); i++) {
+                contexts[i] = to_backend(contexts[i]);
+            }
+
             timesteps = to_backend(timesteps);
 
             for (int i = 0; i < ref_latents.size(); i++) {
                 ref_latents[i] = to_backend(ref_latents[i]);
             }
 
-            pe_vec      = Rope::gen_z_image_pe(static_cast<int>(x->ne[1]),
-                                               static_cast<int>(x->ne[0]),
-                                               z_image_params.patch_size,
-                                               static_cast<int>(x->ne[3]),
-                                               static_cast<int>(context->ne[1]),
-                                               SEQ_MULTI_OF,
+            pe_vec      = Rope::gen_z_image_pe(x,
+                                               contexts,
                                                ref_latents,
-                                               increase_ref_index,
+                                               siglip_feats,
+                                               z_image_params.patch_size,
+                                               SEQ_MULTI_OF,
                                                z_image_params.theta,
+                                               z_image_params.axes_dim,
                                                circular_y_enabled,
                                                circular_x_enabled,
-                                               z_image_params.axes_dim);
+                                               static_cast<int>(x->ne[3]));
             int pos_len = static_cast<int>(pe_vec.size() / z_image_params.axes_dim_sum / 2);
             // LOG_DEBUG("pos_len %d", pos_len);
             auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, z_image_params.axes_dim_sum / 2, pos_len);
@@ -521,7 +839,7 @@ namespace ZImage {
             struct ggml_tensor* out = z_image.forward(&runner_ctx,
                                                       x,
                                                       timesteps,
-                                                      context,
+                                                      contexts,
                                                       pe,
                                                       ref_latents);
 
@@ -533,16 +851,16 @@ namespace ZImage {
         bool compute(int n_threads,
                      struct ggml_tensor* x,
                      struct ggml_tensor* timesteps,
-                     struct ggml_tensor* context,
-                     std::vector<ggml_tensor*> ref_latents = {},
-                     bool increase_ref_index               = false,
-                     struct ggml_tensor** output           = nullptr,
-                     struct ggml_context* output_ctx       = nullptr) {
+                     std::vector<ggml_tensor*> contexts,
+                     std::vector<ggml_tensor*> ref_latents  = {},
+                     std::vector<ggml_tensor*> siglip_feats = {},
+                     struct ggml_tensor** output            = nullptr,
+                     struct ggml_context* output_ctx        = nullptr) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             auto get_graph = [&]() -> struct ggml_cgraph* {
-                return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
+                return build_graph(x, timesteps, contexts, ref_latents, siglip_feats);
             };
 
             return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
@@ -574,7 +892,7 @@ namespace ZImage {
                 struct ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
-                compute(8, x, timesteps, context, {}, false, &out, work_ctx);
+                compute(8, x, timesteps, {context}, {}, {}, &out, work_ctx);
                 int64_t t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);