feat: add OpenCL backend support (#680)

Some SDXL LoRAs (eg. PCM) can exceed 12k nodes.

.github/workflows/build.yml

@@ -155,17 +155,17 @@ jobs:
       matrix:
         include:
           - build: "noavx"
-            defines: "-DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS=ON"
+            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD_BUILD_SHARED_LIBS=ON"
           - build: "avx2"
-            defines: "-DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
+            defines: "-DGGML_NATIVE=OFF -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
           - build: "avx"
-            defines: "-DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON"
+            defines: "-DGGML_NATIVE=OFF -DGGML_AVX=ON -DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS=ON"
           - build: "avx512"
-            defines: "-DGGML_AVX512=ON -DSD_BUILD_SHARED_LIBS=ON"
+            defines: "-DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON"
           - build: "cuda12"
-            defines: "-DSD_CUBLAS=ON -DSD_BUILD_SHARED_LIBS=ON"
-          - build: "rocm5.5"
-            defines: '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1100;gfx1102;gfx1030" -DSD_BUILD_SHARED_LIBS=ON'
+            defines: "-DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES=90;89;80;75"
+          # - build: "rocm5.5"
+          #   defines: '-G Ninja -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1100;gfx1102;gfx1030" -DSD_BUILD_SHARED_LIBS=ON'
           - build: 'vulkan'
             defines: "-DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON"
     steps:
@@ -178,9 +178,9 @@ jobs:
       - name: Install cuda-toolkit
         id: cuda-toolkit
         if: ${{ matrix.build == 'cuda12' }}
-        uses: Jimver/cuda-toolkit@v0.2.11
+        uses: Jimver/cuda-toolkit@v0.2.19
         with:
-          cuda: "12.2.0"
+          cuda: "12.6.2"
           method: "network"
           sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
 

CMakeLists.txt

@@ -24,20 +24,21 @@ endif()
 # general
 #option(SD_BUILD_TESTS                "sd: build tests"    ${SD_STANDALONE})
 option(SD_BUILD_EXAMPLES             "sd: build examples" ${SD_STANDALONE})
-option(SD_CUBLAS                     "sd: cuda backend" OFF)
+option(SD_CUDA                       "sd: cuda backend" OFF)
 option(SD_HIPBLAS                    "sd: rocm backend" OFF)
 option(SD_METAL                      "sd: metal backend" OFF)
 option(SD_VULKAN                     "sd: vulkan backend" OFF)
+option(SD_OPENCL                     "sd: opencl backend" OFF)
 option(SD_SYCL                       "sd: sycl backend" OFF)
-option(SD_FLASH_ATTN                 "sd: use flash attention for x4 less memory usage" OFF)
+option(SD_MUSA                       "sd: musa backend" OFF)
 option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
-if(SD_CUBLAS)
-    message("-- Use CUBLAS as backend stable-diffusion")
+if(SD_CUDA)
+    message("-- Use CUDA as backend stable-diffusion")
     set(GGML_CUDA ON)
-    add_definitions(-DSD_USE_CUBLAS)
+    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_METAL)
@@ -52,23 +53,33 @@ if (SD_VULKAN)
     add_definitions(-DSD_USE_VULKAN)
 endif ()
 
+if (SD_OPENCL)
+    message("-- Use OpenCL as backend stable-diffusion")
+    set(GGML_OPENCL ON)
+    add_definitions(-DSD_USE_OPENCL)
+endif ()
+
 if (SD_HIPBLAS)
     message("-- Use HIPBLAS as backend stable-diffusion")
-    set(GGML_HIPBLAS ON)
-    add_definitions(-DSD_USE_CUBLAS)
+    set(GGML_HIP ON)
+    add_definitions(-DSD_USE_CUDA)
     if(SD_FAST_SOFTMAX)
         set(GGML_CUDA_FAST_SOFTMAX ON)
     endif()
 endif ()
 
-if(SD_FLASH_ATTN)
-    message("-- Use Flash Attention for memory optimization")
-    add_definitions(-DSD_USE_FLASH_ATTENTION)
+if(SD_MUSA)
+    message("-- Use MUSA as backend stable-diffusion")
+    set(GGML_MUSA ON)
+    add_definitions(-DSD_USE_CUDA)
+    if(SD_FAST_SOFTMAX)
+        set(GGML_CUDA_FAST_SOFTMAX ON)
+    endif()
 endif()
 
 set(SD_LIB stable-diffusion)
 
-file(GLOB SD_LIB_SOURCES 
+file(GLOB SD_LIB_SOURCES
     "*.h"
     "*.cpp"
     "*.hpp"
@@ -92,6 +103,7 @@ endif()
 if(SD_SYCL)
     message("-- Use SYCL as backend stable-diffusion")
     set(GGML_SYCL ON)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
     add_definitions(-DSD_USE_SYCL)
     # disable fast-math on host, see:
     # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html

Dockerfile.musa

@@ -0,0 +1,22 @@
+ARG MUSA_VERSION=rc3.1.1
+
+FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu22.04 as build
+
+RUN apt-get update && apt-get install -y ccache cmake git
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ \
+        -DCMAKE_C_FLAGS="${CMAKE_C_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
+        -DCMAKE_CXX_FLAGS="${CMAKE_CXX_FLAGS} -fopenmp -I/usr/lib/llvm-14/lib/clang/14.0.0/include -L/usr/lib/llvm-14/lib" \
+        -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release
+
+FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu22.04 as runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

README.md

@@ -13,7 +13,8 @@ Inference of Stable Diffusion and Flux in pure C/C++
 - SD1.x, SD2.x, SDXL and [SD3/SD3.5](./docs/sd3.md) support
     - !!!The VAE in SDXL encounters NaN issues under FP16, but unfortunately, the ggml_conv_2d only operates under FP16. Hence, a parameter is needed to specify the VAE that has fixed the FP16 NaN issue. You can find it here: [SDXL VAE FP16 Fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix/blob/main/sdxl_vae.safetensors).
 - [Flux-dev/Flux-schnell Support](./docs/flux.md)
-
+- [FLUX.1-Kontext-dev](./docs/kontext.md)
+- [Chroma](./docs/chroma.md)
 - [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo) and [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo) support
 - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
 - 16-bit, 32-bit float support
@@ -21,10 +22,10 @@ Inference of Stable Diffusion and Flux in pure C/C++
 - Accelerated memory-efficient CPU inference
     - Only requires ~2.3GB when using txt2img with fp16 precision to generate a 512x512 image, enabling Flash Attention just requires ~1.8GB.
 - AVX, AVX2 and AVX512 support for x86 architectures
-- Full CUDA, Metal, Vulkan and SYCL backend for GPU acceleration.
+- Full CUDA, Metal, Vulkan, OpenCL and SYCL backend for GPU acceleration.
 - Can load ckpt, safetensors and diffusers models/checkpoints. Standalone VAEs models
     - No need to convert to `.ggml` or `.gguf` anymore!
-- Flash Attention for memory usage optimization (only cpu for now)
+- Flash Attention for memory usage optimization
 - Original `txt2img` and `img2img` mode
 - Negative prompt
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
@@ -113,12 +114,12 @@ cmake .. -DGGML_OPENBLAS=ON
 cmake --build . --config Release
 ```
 
-##### Using CUBLAS
+##### Using CUDA
 
 This provides BLAS acceleration using the CUDA cores of your Nvidia GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
 
 ```
-cmake .. -DSD_CUBLAS=ON
+cmake .. -DSD_CUDA=ON
 cmake --build . --config Release
 ```
 
@@ -132,6 +133,14 @@ cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_H
 cmake --build . --config Release
 ```
 
+##### Using MUSA
+
+This provides BLAS acceleration using the MUSA cores of your Moore Threads GPU. Make sure to have the MUSA toolkit installed.
+
+```bash
+cmake .. -DCMAKE_C_COMPILER=/usr/local/musa/bin/clang -DCMAKE_CXX_COMPILER=/usr/local/musa/bin/clang++ -DSD_MUSA=ON -DCMAKE_BUILD_TYPE=Release
+cmake --build . --config Release
+```
 
 ##### Using Metal
 
@@ -151,6 +160,73 @@ cmake .. -DSD_VULKAN=ON
 cmake --build . --config Release
 ```
 
+##### Using OpenCL (for Adreno GPU)
+
+Currently, it supports only Adreno GPUs and is primarily optimized for Q4_0 type
+
+To build for Windows ARM please refers to [Windows 11 Arm64
+](https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md#windows-11-arm64)
+
+Building for Android:
+
+  Android NDK:
+       Download and install the Android NDK from the [official Android developer site](https://developer.android.com/ndk/downloads).
+
+Setup OpenCL Dependencies for NDK:
+
+You need to provide OpenCL headers and the ICD loader library to your NDK sysroot.
+
+*   OpenCL Headers:
+    ```bash
+    # In a temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-Headers
+    cd OpenCL-Headers
+    # Replace <YOUR_NDK_PATH> with your actual NDK installation path
+    # e.g., cp -r CL /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    sudo cp -r CL <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include
+    cd ..
+    ```
+
+*   OpenCL ICD Loader:
+    ```bash
+    # In the same temporary working directory
+    git clone https://github.com/KhronosGroup/OpenCL-ICD-Loader
+    cd OpenCL-ICD-Loader
+    mkdir build_ndk && cd build_ndk
+
+    # Replace <YOUR_NDK_PATH> in the CMAKE_TOOLCHAIN_FILE and OPENCL_ICD_LOADER_HEADERS_DIR
+    cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+      -DOPENCL_ICD_LOADER_HEADERS_DIR=<YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/include \
+      -DANDROID_ABI=arm64-v8a \
+      -DANDROID_PLATFORM=24 \
+      -DANDROID_STL=c++_shared
+
+    ninja
+    # Replace <YOUR_NDK_PATH>
+    # e.g., cp libOpenCL.so /path/to/android-ndk-r26c/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    sudo cp libOpenCL.so <YOUR_NDK_PATH>/toolchains/llvm/prebuilt/linux-x86_64/sysroot/usr/lib/aarch64-linux-android
+    cd ../..
+    ```
+
+Build `stable-diffusion.cpp` for Android with OpenCL:
+
+```bash
+mkdir build-android && cd build-android
+
+# Replace <YOUR_NDK_PATH> with your actual NDK installation path
+# e.g., -DCMAKE_TOOLCHAIN_FILE=/path/to/android-ndk-r26c/build/cmake/android.toolchain.cmake
+cmake .. -G Ninja \
+  -DCMAKE_TOOLCHAIN_FILE=<YOUR_NDK_PATH>/build/cmake/android.toolchain.cmake \
+  -DANDROID_ABI=arm64-v8a \
+  -DANDROID_PLATFORM=android-28 \
+  -DGGML_OPENMP=OFF \
+  -DSD_OPENCL=ON
+
+ninja
+```
+*(Note: Don't forget to include `LD_LIBRARY_PATH=/vendor/lib64` in your command line before running the binary)*
+
 ##### Using SYCL
 
 Using SYCL makes the computation run on the Intel GPU. Please make sure you have installed the related driver and [Intel® oneAPI Base toolkit](https://www.intel.com/content/www/us/en/developer/tools/oneapi/base-toolkit.html) before start. More details and steps can refer to [llama.cpp SYCL backend](https://github.com/ggerganov/llama.cpp/blob/master/docs/backend/SYCL.md#linux).
@@ -182,11 +258,21 @@ Example of text2img by using SYCL backend:
 
 ##### Using Flash Attention
 
-Enabling flash attention reduces memory usage by at least 400 MB. At the moment, it is not supported when CUBLAS is enabled because the kernel implementation is missing.
+Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
+eg.:
+ - flux 768x768 ~600mb
+ - SD2 768x768 ~1400mb
 
+For most backends, it slows things down, but for cuda it generally speeds it up too.
+At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
+
+Run by adding `--diffusion-fa` to the arguments and watch for:
 ```
-cmake .. -DSD_FLASH_ATTN=ON
-cmake --build . --config Release
+[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
+```
+and the compute buffer shrink in the debug log:
+```
+[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
 ```
 
 ### Run
@@ -202,7 +288,7 @@ arguments:
   -m, --model [MODEL]                path to full model
   --diffusion-model                  path to the standalone diffusion model
   --clip_l                           path to the clip-l text encoder
-  --clip_g                           path to the clip-l text encoder
+  --clip_g                           path to the clip-g text encoder
   --t5xxl                            path to the the t5xxl text encoder
   --vae [VAE]                        path to vae
   --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
@@ -213,22 +299,32 @@ arguments:
   --normalize-input                  normalize PHOTOMAKER input id images
   --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now
   --upscale-repeats                  Run the ESRGAN upscaler this many times (default 1)
-  --type [TYPE]                      weight type (f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_k, q3_k, q4_k)
+  --type [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
   --lora-model-dir [DIR]             lora model directory
   -i, --init-img [IMAGE]             path to the input image, required by img2img
+  --mask [MASK]                      path to the mask image, required by img2img with mask
   --control-image [IMAGE]            path to image condition, control net
+  -r, --ref_image [PATH]             reference image for Flux Kontext models (can be used multiple times)
   -o, --output OUTPUT                path to write result image to (default: ./output.png)
   -p, --prompt [PROMPT]              the prompt to render
   -n, --negative-prompt PROMPT       the negative prompt (default: "")
   --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)
+  --guidance SCALE                   guidance scale for img2img (default: 3.5)
+  --slg-scale SCALE                  skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium
+  --eta SCALE                        eta in DDIM, only for DDIM and TCD: (default: 0)
+  --skip-layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])
+  --skip-layer-start START           SLG enabling point: (default: 0.01)
+  --skip-layer-end END               SLG disabling point: (default: 0.2)
+                                     SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])
   --strength STRENGTH                strength for noising/unnoising (default: 0.75)
   --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20%)
   --control-strength STRENGTH        strength to apply Control Net (default: 0.9)
                                      1.0 corresponds to full destruction of information in init image
   -H, --height H                     image height, in pixel space (default: 512)
   -W, --width W                      image width, in pixel space (default: 512)
-  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm}
+  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
                                      sampling method (default: "euler_a")
   --steps  STEPS                     number of sample steps (default: 20)
   --rng {std_default, cuda}          RNG (default: cuda)
@@ -240,9 +336,15 @@ arguments:
   --vae-tiling                       process vae in tiles to reduce memory usage
   --vae-on-cpu                       keep vae in cpu (for low vram)
   --clip-on-cpu                      keep clip in cpu (for low vram)
+  --diffusion-fa                     use flash attention in the diffusion model (for low vram)
+                                     Might lower quality, since it implies converting k and v to f16.
+                                     This might crash if it is not supported by the backend.
   --control-net-cpu                  keep controlnet in cpu (for low vram)
   --canny                            apply canny preprocessor (edge detection)
-  --color                            Colors the logging tags according to level
+  --color                            colors the logging tags according to level
+  --chroma-disable-dit-mask          disable dit mask for chroma
+  --chroma-enable-t5-mask            enable t5 mask for chroma
+  --chroma-t5-mask-pad  PAD_SIZE     t5 mask pad size of chroma
   -v, --verbose                      print extra info
 ```
 
@@ -292,12 +394,16 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra
 
 * Golang: [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
 * C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
+* Python: [william-murray1204/stable-diffusion-cpp-python](https://github.com/william-murray1204/stable-diffusion-cpp-python)
+* Rust: [newfla/diffusion-rs](https://github.com/newfla/diffusion-rs)
 
 ## UIs
 
 These projects use `stable-diffusion.cpp` as a backend for their image generation.
 
 - [Jellybox](https://jellybox.com)
+- [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
+- [Stable Diffusion CLI-GUI](https://github.com/piallai/stable-diffusion.cpp)
 
 ## Contributors
 

clip.hpp

@@ -343,6 +343,13 @@ public:
         }
     }
 
+    std::string clean_up_tokenization(std::string& text) {
+        std::regex pattern(R"( ,)");
+        // Replace " ," with ","
+        std::string result = std::regex_replace(text, pattern, ",");
+        return result;
+    }
+
     std::string decode(const std::vector<int>& tokens) {
         std::string text = "";
         for (int t : tokens) {
@@ -351,8 +358,12 @@ public:
             std::u32string ts = decoder[t];
             // printf("%d, %s \n", t,  utf32_to_utf8(ts).c_str());
             std::string s = utf32_to_utf8(ts);
-            if (s.length() >= 4 && ends_with(s, "</w>")) {
-                text += " " + s.replace(s.length() - 4, s.length() - 1, "");
+            if (s.length() >= 4) {
+                if (ends_with(s, "</w>")) {
+                    text += s.replace(s.length() - 4, s.length() - 1, "") + " ";
+                } else {
+                    text += s;
+                }
             } else {
                 text += " " + s;
             }
@@ -364,6 +375,7 @@ public:
 
         // std::string s((char *)bytes.data());
         // std::string s = "";
+        text = clean_up_tokenization(text);
         return trim(text);
     }
 
@@ -533,9 +545,12 @@ protected:
     int64_t vocab_size;
     int64_t num_positions;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, wtype, embed_dim, vocab_size);
-        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, embed_dim, num_positions);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type token_wtype    = GGML_TYPE_F32;  //(tensor_types.find(prefix + "token_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "token_embedding.weight"] : GGML_TYPE_F32;
+        enum ggml_type position_wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "position_embedding.weight"] : GGML_TYPE_F32;
+
+        params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
+        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
     }
 
 public:
@@ -579,11 +594,14 @@ protected:
     int64_t image_size;
     int64_t num_patches;
     int64_t num_positions;
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type patch_wtype    = GGML_TYPE_F16;  // tensor_types.find(prefix + "patch_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "patch_embedding.weight"] : GGML_TYPE_F16;
+        enum ggml_type class_wtype    = GGML_TYPE_F32;  // tensor_types.find(prefix + "class_embedding") != tensor_types.end() ? tensor_types[prefix + "class_embedding"] : GGML_TYPE_F32;
+        enum ggml_type position_wtype = GGML_TYPE_F32;  // tensor_types.find(prefix + "position_embedding.weight") != tensor_types.end() ? tensor_types[prefix + "position_embedding.weight"] : GGML_TYPE_F32;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["patch_embedding.weight"]    = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, patch_size, patch_size, num_channels, embed_dim);
-        params["class_embedding"]           = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, embed_dim);
-        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, embed_dim, num_positions);
+        params["patch_embedding.weight"]    = ggml_new_tensor_4d(ctx, patch_wtype, patch_size, patch_size, num_channels, embed_dim);
+        params["class_embedding"]           = ggml_new_tensor_1d(ctx, class_wtype, embed_dim);
+        params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
     }
 
 public:
@@ -639,9 +657,10 @@ enum CLIPVersion {
 
 class CLIPTextModel : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
         if (version == OPEN_CLIP_VIT_BIGG_14) {
-            params["text_projection"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, projection_dim, hidden_size);
+            enum ggml_type wtype      = GGML_TYPE_F32;  // tensor_types.find(prefix + "text_projection") != tensor_types.end() ? tensor_types[prefix + "text_projection"] : GGML_TYPE_F32;
+            params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
         }
     }
 
@@ -659,8 +678,8 @@ public:
     bool with_final_ln        = true;
 
     CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
-                  int clip_skip_value = -1,
-                  bool with_final_ln  = true)
+                  bool with_final_ln  = true,
+                  int clip_skip_value = -1)
         : version(version), with_final_ln(with_final_ln) {
         if (version == OPEN_CLIP_VIT_H_14) {
             hidden_size       = 1024;
@@ -682,7 +701,7 @@ public:
 
     void set_clip_skip(int skip) {
         if (skip <= 0) {
-            return;
+            skip = -1;
         }
         clip_skip = skip;
     }
@@ -765,14 +784,17 @@ public:
         auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
         x      = pre_layernorm->forward(ctx, x);
         x      = encoder->forward(ctx, x, -1, false);
-        x      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
+        // print_ggml_tensor(x, true, "ClipVisionModel x: ");
+        auto last_hidden_state = x;
+        x                      = post_layernorm->forward(ctx, x);  // [N, n_token, hidden_size]
 
         GGML_ASSERT(x->ne[3] == 1);
         if (return_pooled) {
             ggml_tensor* pooled = ggml_cont(ctx, ggml_view_2d(ctx, x, x->ne[0], x->ne[2], x->nb[2], 0));
             return pooled;  // [N, hidden_size]
         } else {
-            return x;  // [N, n_token, hidden_size]
+            // return x;  // [N, n_token, hidden_size]
+            return last_hidden_state;  // [N, n_token, hidden_size]
         }
     }
 };
@@ -783,9 +805,9 @@ protected:
     int64_t out_features;
     bool transpose_weight;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = tensor_types.find(prefix + "weight") != tensor_types.end() ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
         if (transpose_weight) {
-            LOG_ERROR("transpose_weight");
             params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
         } else {
             params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
@@ -846,12 +868,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
     CLIPTextModel model;
 
     CLIPTextModelRunner(ggml_backend_t backend,
-                        ggml_type wtype,
+                        std::map<std::string, enum ggml_type>& tensor_types,
+                        const std::string prefix,
                         CLIPVersion version = OPENAI_CLIP_VIT_L_14,
-                        int clip_skip_value = 1,
-                        bool with_final_ln  = true)
-        : GGMLRunner(backend, wtype), model(version, clip_skip_value, with_final_ln) {
-        model.init(params_ctx, wtype);
+                        bool with_final_ln  = true,
+                        int clip_skip_value = -1)
+        : GGMLRunner(backend), model(version, with_final_ln, clip_skip_value) {
+        model.init(params_ctx, tensor_types, prefix);
     }
 
     std::string get_desc() {
@@ -893,13 +916,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
         struct ggml_tensor* embeddings = NULL;
 
         if (num_custom_embeddings > 0 && custom_embeddings_data != NULL) {
-            auto custom_embeddings = ggml_new_tensor_2d(compute_ctx,
-                                                        wtype,
-                                                        model.hidden_size,
-                                                        num_custom_embeddings);
+            auto token_embed_weight = model.get_token_embed_weight();
+            auto custom_embeddings  = ggml_new_tensor_2d(compute_ctx,
+                                                         token_embed_weight->type,
+                                                         model.hidden_size,
+                                                         num_custom_embeddings);
             set_backend_tensor_data(custom_embeddings, custom_embeddings_data);
 
-            auto token_embed_weight = model.get_token_embed_weight();
             // concatenate custom embeddings
             embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
         }

common.hpp

@@ -56,7 +56,7 @@ public:
         // x: [N, channels, h, w]
         auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
 
-        x = ggml_upscale(ctx, x, 2);  // [N, channels, h*2, w*2]
+        x = ggml_upscale(ctx, x, 2, GGML_SCALE_MODE_NEAREST);  // [N, channels, h*2, w*2]
         x = conv->forward(ctx, x);    // [N, out_channels, h*2, w*2]
         return x;
     }
@@ -182,9 +182,11 @@ protected:
     int64_t dim_in;
     int64_t dim_out;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["proj.weight"] = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
-        params["proj.bias"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, dim_out * 2);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
+        enum ggml_type wtype      = (tensor_types.find(prefix + "proj.weight") != tensor_types.end()) ? tensor_types[prefix + "proj.weight"] : GGML_TYPE_F32;
+        enum ggml_type bias_wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "proj.bias") != tensor_types.end()) ? tensor_types[prefix + "proj.bias"] : GGML_TYPE_F32;
+        params["proj.weight"]     = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
+        params["proj.bias"]       = ggml_new_tensor_1d(ctx, bias_wtype, dim_out * 2);
     }
 
 public:
@@ -245,16 +247,19 @@ protected:
     int64_t context_dim;
     int64_t n_head;
     int64_t d_head;
+    bool flash_attn;
 
 public:
     CrossAttention(int64_t query_dim,
                    int64_t context_dim,
                    int64_t n_head,
-                   int64_t d_head)
+                   int64_t d_head,
+                   bool flash_attn = false)
         : n_head(n_head),
           d_head(d_head),
           query_dim(query_dim),
-          context_dim(context_dim) {
+          context_dim(context_dim),
+          flash_attn(flash_attn) {
         int64_t inner_dim = d_head * n_head;
 
         blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, false));
@@ -283,7 +288,7 @@ public:
         auto k = to_k->forward(ctx, context);  // [N, n_context, inner_dim]
         auto v = to_v->forward(ctx, context);  // [N, n_context, inner_dim]
 
-        x = ggml_nn_attention_ext(ctx, q, k, v, n_head, NULL, false);  // [N, n_token, inner_dim]
+        x = ggml_nn_attention_ext(ctx, q, k, v, n_head, NULL, false, false, flash_attn);  // [N, n_token, inner_dim]
 
         x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
         return x;
@@ -301,15 +306,16 @@ public:
                           int64_t n_head,
                           int64_t d_head,
                           int64_t context_dim,
-                          bool ff_in = false)
+                          bool ff_in      = false,
+                          bool flash_attn = false)
         : n_head(n_head), d_head(d_head), ff_in(ff_in) {
         // disable_self_attn is always False
         // disable_temporal_crossattention is always False
         // switch_temporal_ca_to_sa is always False
         // inner_dim is always None or equal to dim
         // gated_ff is always True
-        blocks["attn1"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, dim, n_head, d_head));
-        blocks["attn2"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, context_dim, n_head, d_head));
+        blocks["attn1"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, dim, n_head, d_head, flash_attn));
+        blocks["attn2"] = std::shared_ptr<GGMLBlock>(new CrossAttention(dim, context_dim, n_head, d_head, flash_attn));
         blocks["ff"]    = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim));
         blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
         blocks["norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
@@ -374,7 +380,8 @@ public:
                        int64_t n_head,
                        int64_t d_head,
                        int64_t depth,
-                       int64_t context_dim)
+                       int64_t context_dim,
+                       bool flash_attn = false)
         : in_channels(in_channels),
           n_head(n_head),
           d_head(d_head),
@@ -388,7 +395,7 @@ public:
 
         for (int i = 0; i < depth; i++) {
             std::string name = "transformer_blocks." + std::to_string(i);
-            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim));
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new BasicTransformerBlock(inner_dim, n_head, d_head, context_dim, false, flash_attn));
         }
 
         blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(inner_dim, in_channels, {1, 1}));
@@ -433,8 +440,10 @@ public:
 
 class AlphaBlender : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["mix_factor"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
+        // Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
+        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.ypes.find(prefix + "mix_factor") != tensor_types.end()) ? tensor_types[prefix + "mix_factor"] : GGML_TYPE_F32;
+        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
     }
 
     float get_alpha() {
@@ -511,4 +520,4 @@ public:
     }
 };
 
-#endif  // __COMMON_HPP__
+#endif  // __COMMON_HPP__

conditioner.hpp

@@ -43,71 +43,74 @@ struct Conditioner {
 // ldm.modules.encoders.modules.FrozenCLIPEmbedder
 // Ref: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/cad87bf4e3e0b0a759afa94e933527c3123d59bc/modules/sd_hijack_clip.py#L283
 struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
-    SDVersion version = VERSION_SD1;
+    SDVersion version    = VERSION_SD1;
+    PMVersion pm_version = PM_VERSION_1;
     CLIPTokenizer tokenizer;
-    ggml_type wtype;
     std::shared_ptr<CLIPTextModelRunner> text_model;
     std::shared_ptr<CLIPTextModelRunner> text_model2;
 
     std::string trigger_word = "img";  // should be user settable
     std::string embd_dir;
-    int32_t num_custom_embeddings = 0;
+    int32_t num_custom_embeddings   = 0;
+    int32_t num_custom_embeddings_2 = 0;
     std::vector<uint8_t> token_embed_custom;
     std::vector<std::string> readed_embeddings;
 
     FrozenCLIPEmbedderWithCustomWords(ggml_backend_t backend,
-                                      ggml_type wtype,
+                                      std::map<std::string, enum ggml_type>& tensor_types,
                                       const std::string& embd_dir,
                                       SDVersion version = VERSION_SD1,
+                                      PMVersion pv      = PM_VERSION_1,
                                       int clip_skip     = -1)
-        : version(version), tokenizer(version == VERSION_SD2 ? 0 : 49407), embd_dir(embd_dir), wtype(wtype) {
-        if (clip_skip <= 0) {
-            clip_skip = 1;
-            if (version == VERSION_SD2 || version == VERSION_SDXL) {
-                clip_skip = 2;
-            }
-        }
-        if (version == VERSION_SD1) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPENAI_CLIP_VIT_L_14, clip_skip);
-        } else if (version == VERSION_SD2) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPEN_CLIP_VIT_H_14, clip_skip);
-        } else if (version == VERSION_SDXL) {
-            text_model  = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPENAI_CLIP_VIT_L_14, clip_skip, false);
-            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
+        : version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407), embd_dir(embd_dir) {
+        if (sd_version_is_sd1(version)) {
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14);
+        } else if (sd_version_is_sd2(version)) {
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14);
+        } else if (sd_version_is_sdxl(version)) {
+            text_model  = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
+            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
         }
+        set_clip_skip(clip_skip);
     }
 
     void set_clip_skip(int clip_skip) {
+        if (clip_skip <= 0) {
+            clip_skip = 1;
+            if (sd_version_is_sd2(version) || sd_version_is_sdxl(version)) {
+                clip_skip = 2;
+            }
+        }
         text_model->set_clip_skip(clip_skip);
-        if (version == VERSION_SDXL) {
+        if (sd_version_is_sdxl(version)) {
             text_model2->set_clip_skip(clip_skip);
         }
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
         text_model->get_param_tensors(tensors, "cond_stage_model.transformer.text_model");
-        if (version == VERSION_SDXL) {
+        if (sd_version_is_sdxl(version)) {
             text_model2->get_param_tensors(tensors, "cond_stage_model.1.transformer.text_model");
         }
     }
 
     void alloc_params_buffer() {
         text_model->alloc_params_buffer();
-        if (version == VERSION_SDXL) {
+        if (sd_version_is_sdxl(version)) {
             text_model2->alloc_params_buffer();
         }
     }
 
     void free_params_buffer() {
         text_model->free_params_buffer();
-        if (version == VERSION_SDXL) {
+        if (sd_version_is_sdxl(version)) {
             text_model2->free_params_buffer();
         }
     }
 
     size_t get_params_buffer_size() {
         size_t buffer_size = text_model->get_params_buffer_size();
-        if (version == VERSION_SDXL) {
+        if (sd_version_is_sdxl(version)) {
             buffer_size += text_model2->get_params_buffer_size();
         }
         return buffer_size;
@@ -130,28 +133,55 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         params.no_alloc               = false;
         struct ggml_context* embd_ctx = ggml_init(params);
         struct ggml_tensor* embd      = NULL;
-        int64_t hidden_size           = text_model->model.hidden_size;
+        struct ggml_tensor* embd2     = NULL;
         auto on_load                  = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) {
-            if (tensor_storage.ne[0] != hidden_size) {
-                LOG_DEBUG("embedding wrong hidden size, got %i, expected %i", tensor_storage.ne[0], hidden_size);
-                return false;
+            if (tensor_storage.ne[0] != text_model->model.hidden_size) {
+                if (text_model2) {
+                    if (tensor_storage.ne[0] == text_model2->model.hidden_size) {
+                        embd2       = ggml_new_tensor_2d(embd_ctx, tensor_storage.type, text_model2->model.hidden_size, tensor_storage.n_dims > 1 ? tensor_storage.ne[1] : 1);
+                        *dst_tensor = embd2;
+                    } else {
+                        LOG_DEBUG("embedding wrong hidden size, got %i, expected %i or %i", tensor_storage.ne[0], text_model->model.hidden_size, text_model2->model.hidden_size);
+                        return false;
+                    }
+                } else {
+                    LOG_DEBUG("embedding wrong hidden size, got %i, expected %i", tensor_storage.ne[0], text_model->model.hidden_size);
+                    return false;
+                }
+            } else {
+                embd        = ggml_new_tensor_2d(embd_ctx, tensor_storage.type, text_model->model.hidden_size, tensor_storage.n_dims > 1 ? tensor_storage.ne[1] : 1);
+                *dst_tensor = embd;
             }
-            embd        = ggml_new_tensor_2d(embd_ctx, wtype, hidden_size, tensor_storage.n_dims > 1 ? tensor_storage.ne[1] : 1);
-            *dst_tensor = embd;
             return true;
         };
         model_loader.load_tensors(on_load, NULL);
         readed_embeddings.push_back(embd_name);
-        token_embed_custom.resize(token_embed_custom.size() + ggml_nbytes(embd));
-        memcpy((void*)(token_embed_custom.data() + num_custom_embeddings * hidden_size * ggml_type_size(wtype)),
-               embd->data,
-               ggml_nbytes(embd));
-        for (int i = 0; i < embd->ne[1]; i++) {
-            bpe_tokens.push_back(text_model->model.vocab_size + num_custom_embeddings);
-            // LOG_DEBUG("new custom token: %i", text_model.vocab_size + num_custom_embeddings);
-            num_custom_embeddings++;
+        if (embd) {
+            int64_t hidden_size = text_model->model.hidden_size;
+            token_embed_custom.resize(token_embed_custom.size() + ggml_nbytes(embd));
+            memcpy((void*)(token_embed_custom.data() + num_custom_embeddings * hidden_size * ggml_type_size(embd->type)),
+                   embd->data,
+                   ggml_nbytes(embd));
+            for (int i = 0; i < embd->ne[1]; i++) {
+                bpe_tokens.push_back(text_model->model.vocab_size + num_custom_embeddings);
+                // LOG_DEBUG("new custom token: %i", text_model.vocab_size + num_custom_embeddings);
+                num_custom_embeddings++;
+            }
+            LOG_DEBUG("embedding '%s' applied, custom embeddings: %i", embd_name.c_str(), num_custom_embeddings);
+        }
+        if (embd2) {
+            int64_t hidden_size = text_model2->model.hidden_size;
+            token_embed_custom.resize(token_embed_custom.size() + ggml_nbytes(embd2));
+            memcpy((void*)(token_embed_custom.data() + num_custom_embeddings_2 * hidden_size * ggml_type_size(embd2->type)),
+                   embd2->data,
+                   ggml_nbytes(embd2));
+            for (int i = 0; i < embd2->ne[1]; i++) {
+                bpe_tokens.push_back(text_model2->model.vocab_size + num_custom_embeddings_2);
+                // LOG_DEBUG("new custom token: %i", text_model.vocab_size + num_custom_embeddings);
+                num_custom_embeddings_2++;
+            }
+            LOG_DEBUG("embedding '%s' applied, custom embeddings: %i (text model 2)", embd_name.c_str(), num_custom_embeddings_2);
         }
-        LOG_DEBUG("embedding '%s' applied, custom embeddings: %i", embd_name.c_str(), num_custom_embeddings);
         return true;
     }
 
@@ -268,7 +298,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                 std::vector<int> clean_input_ids_tmp;
                 for (uint32_t i = 0; i < class_token_index[0]; i++)
                     clean_input_ids_tmp.push_back(clean_input_ids[i]);
-                for (uint32_t i = 0; i < num_input_imgs; i++)
+                for (uint32_t i = 0; i < (pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs); i++)
                     clean_input_ids_tmp.push_back(class_token);
                 for (uint32_t i = class_token_index[0] + 1; i < clean_input_ids.size(); i++)
                     clean_input_ids_tmp.push_back(clean_input_ids[i]);
@@ -279,13 +309,16 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             tokens.insert(tokens.end(), clean_input_ids.begin(), clean_input_ids.end());
             weights.insert(weights.end(), clean_input_ids.size(), curr_weight);
         }
-        tokens.insert(tokens.begin(), tokenizer.BOS_TOKEN_ID);
-        weights.insert(weights.begin(), 1.0);
+        // BUG!! double couting, pad_tokens will add BOS at the beginning
+        // tokens.insert(tokens.begin(), tokenizer.BOS_TOKEN_ID);
+        // weights.insert(weights.begin(), 1.0);
 
         tokenizer.pad_tokens(tokens, weights, max_length, padding);
-
+        int offset = pm_version == PM_VERSION_2 ? 2 * num_input_imgs : num_input_imgs;
         for (uint32_t i = 0; i < tokens.size(); i++) {
-            if (class_idx + 1 <= i && i < class_idx + 1 + num_input_imgs)
+            // if (class_idx + 1 <= i && i < class_idx + 1 + 2*num_input_imgs) // photomaker V2 has num_tokens(=2)*num_input_imgs
+            if (class_idx + 1 <= i && i < class_idx + 1 + offset)  // photomaker V2 has num_tokens(=2)*num_input_imgs
+                                                                   // hardcode for now
                 class_token_mask.push_back(true);
             else
                 class_token_mask.push_back(false);
@@ -398,7 +431,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             auto input_ids                 = vector_to_ggml_tensor_i32(work_ctx, chunk_tokens);
             struct ggml_tensor* input_ids2 = NULL;
             size_t max_token_idx           = 0;
-            if (version == VERSION_SDXL) {
+            if (sd_version_is_sdxl(version)) {
                 auto it = std::find(chunk_tokens.begin(), chunk_tokens.end(), tokenizer.EOS_TOKEN_ID);
                 if (it != chunk_tokens.end()) {
                     std::fill(std::next(it), chunk_tokens.end(), 0);
@@ -423,7 +456,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                     false,
                                     &chunk_hidden_states1,
                                     work_ctx);
-                if (version == VERSION_SDXL) {
+                if (sd_version_is_sdxl(version)) {
                     text_model2->compute(n_threads,
                                          input_ids2,
                                          0,
@@ -482,7 +515,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                         ggml_nelements(hidden_states) / chunk_hidden_states->ne[0]);
 
         ggml_tensor* vec = NULL;
-        if (version == VERSION_SDXL) {
+        if (sd_version_is_sdxl(version)) {
             int out_dim = 256;
             vec         = ggml_new_tensor_1d(work_ctx, GGML_TYPE_F32, adm_in_channels);
             // [0:1280]
@@ -585,9 +618,9 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
 struct FrozenCLIPVisionEmbedder : public GGMLRunner {
     CLIPVisionModelProjection vision_model;
 
-    FrozenCLIPVisionEmbedder(ggml_backend_t backend, ggml_type wtype)
-        : vision_model(OPEN_CLIP_VIT_H_14, true), GGMLRunner(backend, wtype) {
-        vision_model.init(params_ctx, wtype);
+    FrozenCLIPVisionEmbedder(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types)
+        : vision_model(OPEN_CLIP_VIT_H_14, true), GGMLRunner(backend) {
+        vision_model.init(params_ctx, tensor_types, "cond_stage_model.transformer");
     }
 
     std::string get_desc() {
@@ -622,7 +655,6 @@ struct FrozenCLIPVisionEmbedder : public GGMLRunner {
 };
 
 struct SD3CLIPEmbedder : public Conditioner {
-    ggml_type wtype;
     CLIPTokenizer clip_l_tokenizer;
     CLIPTokenizer clip_g_tokenizer;
     T5UniGramTokenizer t5_tokenizer;
@@ -631,18 +663,19 @@ struct SD3CLIPEmbedder : public Conditioner {
     std::shared_ptr<T5Runner> t5;
 
     SD3CLIPEmbedder(ggml_backend_t backend,
-                    ggml_type wtype,
+                    std::map<std::string, enum ggml_type>& tensor_types,
                     int clip_skip = -1)
-        : wtype(wtype), clip_g_tokenizer(0) {
-        if (clip_skip <= 0) {
-            clip_skip = 2;
-        }
-        clip_l = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPENAI_CLIP_VIT_L_14, clip_skip, false);
-        clip_g = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
-        t5     = std::make_shared<T5Runner>(backend, wtype);
+        : clip_g_tokenizer(0) {
+        clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
+        clip_g = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
+        t5     = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
+        set_clip_skip(clip_skip);
     }
 
     void set_clip_skip(int clip_skip) {
+        if (clip_skip <= 0) {
+            clip_skip = 2;
+        }
         clip_l->set_clip_skip(clip_skip);
         clip_g->set_clip_skip(clip_skip);
     }
@@ -716,7 +749,7 @@ struct SD3CLIPEmbedder : public Conditioner {
 
         clip_l_tokenizer.pad_tokens(clip_l_tokens, clip_l_weights, max_length, padding);
         clip_g_tokenizer.pad_tokens(clip_g_tokens, clip_g_weights, max_length, padding);
-        t5_tokenizer.pad_tokens(t5_tokens, t5_weights, max_length, padding);
+        t5_tokenizer.pad_tokens(t5_tokens, t5_weights, NULL, max_length, padding);
 
         // for (int i = 0; i < clip_l_tokens.size(); i++) {
         //     std::cout << clip_l_tokens[i] << ":" << clip_l_weights[i] << ", ";
@@ -871,6 +904,7 @@ struct SD3CLIPEmbedder : public Conditioner {
 
                 t5->compute(n_threads,
                             input_ids,
+                            NULL,
                             &chunk_hidden_states_t5,
                             work_ctx);
                 {
@@ -969,24 +1003,24 @@ struct SD3CLIPEmbedder : public Conditioner {
 };
 
 struct FluxCLIPEmbedder : public Conditioner {
-    ggml_type wtype;
     CLIPTokenizer clip_l_tokenizer;
     T5UniGramTokenizer t5_tokenizer;
     std::shared_ptr<CLIPTextModelRunner> clip_l;
     std::shared_ptr<T5Runner> t5;
+    size_t chunk_len = 256;
 
     FluxCLIPEmbedder(ggml_backend_t backend,
-                     ggml_type wtype,
-                     int clip_skip = -1)
-        : wtype(wtype) {
-        if (clip_skip <= 0) {
-            clip_skip = 2;
-        }
-        clip_l = std::make_shared<CLIPTextModelRunner>(backend, wtype, OPENAI_CLIP_VIT_L_14, clip_skip, true);
-        t5     = std::make_shared<T5Runner>(backend, wtype);
+                     std::map<std::string, enum ggml_type>& tensor_types,
+                     int clip_skip = -1) {
+        clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
+        t5     = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
+        set_clip_skip(clip_skip);
     }
 
     void set_clip_skip(int clip_skip) {
+        if (clip_skip <= 0) {
+            clip_skip = 2;
+        }
         clip_l->set_clip_skip(clip_skip);
     }
 
@@ -1048,7 +1082,7 @@ struct FluxCLIPEmbedder : public Conditioner {
         }
 
         clip_l_tokenizer.pad_tokens(clip_l_tokens, clip_l_weights, 77, padding);
-        t5_tokenizer.pad_tokens(t5_tokens, t5_weights, max_length, padding);
+        t5_tokenizer.pad_tokens(t5_tokens, t5_weights, NULL, max_length, padding);
 
         // for (int i = 0; i < clip_l_tokens.size(); i++) {
         //     std::cout << clip_l_tokens[i] << ":" << clip_l_weights[i] << ", ";
@@ -1080,7 +1114,6 @@ struct FluxCLIPEmbedder : public Conditioner {
         struct ggml_tensor* pooled              = NULL;  // [768,]
         std::vector<float> hidden_states_vec;
 
-        size_t chunk_len   = 256;
         size_t chunk_count = t5_tokens.size() / chunk_len;
         for (int chunk_idx = 0; chunk_idx < chunk_count; chunk_idx++) {
             // clip_l
@@ -1118,6 +1151,7 @@ struct FluxCLIPEmbedder : public Conditioner {
 
                 t5->compute(n_threads,
                             input_ids,
+                            NULL,
                             &chunk_hidden_states,
                             work_ctx);
                 {
@@ -1167,7 +1201,7 @@ struct FluxCLIPEmbedder : public Conditioner {
                                       int height,
                                       int adm_in_channels        = -1,
                                       bool force_zero_embeddings = false) {
-        auto tokens_and_weights = tokenize(text, 256, true);
+        auto tokens_and_weights = tokenize(text, chunk_len, true);
         return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, force_zero_embeddings);
     }
 
@@ -1189,4 +1223,206 @@ struct FluxCLIPEmbedder : public Conditioner {
     }
 };
 
-#endif
+struct PixArtCLIPEmbedder : public Conditioner {
+    T5UniGramTokenizer t5_tokenizer;
+    std::shared_ptr<T5Runner> t5;
+    size_t chunk_len = 512;
+    bool use_mask    = false;
+    int mask_pad     = 1;
+
+    PixArtCLIPEmbedder(ggml_backend_t backend,
+                       std::map<std::string, enum ggml_type>& tensor_types,
+                       int clip_skip = -1,
+                       bool use_mask = false,
+                       int mask_pad  = 1)
+        : use_mask(use_mask), mask_pad(mask_pad) {
+        t5 = std::make_shared<T5Runner>(backend, tensor_types, "text_encoders.t5xxl.transformer");
+    }
+
+    void set_clip_skip(int clip_skip) {
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+        t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
+    }
+
+    void alloc_params_buffer() {
+        t5->alloc_params_buffer();
+    }
+
+    void free_params_buffer() {
+        t5->free_params_buffer();
+    }
+
+    size_t get_params_buffer_size() {
+        size_t buffer_size = 0;
+
+        buffer_size += t5->get_params_buffer_size();
+
+        return buffer_size;
+    }
+
+    std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
+                                                                                  size_t max_length = 0,
+                                                                                  bool padding      = false) {
+        auto parsed_attention = parse_prompt_attention(text);
+
+        {
+            std::stringstream ss;
+            ss << "[";
+            for (const auto& item : parsed_attention) {
+                ss << "['" << item.first << "', " << item.second << "], ";
+            }
+            ss << "]";
+            LOG_DEBUG("parse '%s' to %s", text.c_str(), ss.str().c_str());
+        }
+
+        auto on_new_token_cb = [&](std::string& str, std::vector<int32_t>& bpe_tokens) -> bool {
+            return false;
+        };
+
+        std::vector<int> t5_tokens;
+        std::vector<float> t5_weights;
+        std::vector<float> t5_mask;
+        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);
+        }
+
+        t5_tokenizer.pad_tokens(t5_tokens, t5_weights, &t5_mask, max_length, padding);
+
+        return {t5_tokens, t5_weights, t5_mask};
+    }
+
+    void modify_mask_to_attend_padding(struct ggml_tensor* mask, int max_seq_length, int num_extra_padding = 8) {
+        float* mask_data = (float*)mask->data;
+        int num_pad      = 0;
+        for (int64_t i = 0; i < max_seq_length; i++) {
+            if (num_pad >= num_extra_padding) {
+                break;
+            }
+            if (std::isinf(mask_data[i])) {
+                mask_data[i] = 0;
+                ++num_pad;
+            }
+        }
+        // LOG_DEBUG("PAD: %d", num_pad);
+    }
+
+    SDCondition get_learned_condition_common(ggml_context* work_ctx,
+                                             int n_threads,
+                                             std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> token_and_weights,
+                                             int clip_skip,
+                                             bool force_zero_embeddings = false) {
+        auto& t5_tokens        = std::get<0>(token_and_weights);
+        auto& t5_weights       = std::get<1>(token_and_weights);
+        auto& t5_attn_mask_vec = std::get<2>(token_and_weights);
+
+        int64_t t0                              = ggml_time_ms();
+        struct ggml_tensor* hidden_states       = NULL;                                               // [N, n_token, 4096]
+        struct ggml_tensor* chunk_hidden_states = NULL;                                               // [n_token, 4096]
+        struct ggml_tensor* pooled              = NULL;                                               // [768,]
+        struct ggml_tensor* t5_attn_mask        = vector_to_ggml_tensor(work_ctx, t5_attn_mask_vec);  // [768,]
+
+        std::vector<float> hidden_states_vec;
+
+        size_t chunk_count = t5_tokens.size() / chunk_len;
+
+        for (int chunk_idx = 0; chunk_idx < chunk_count; chunk_idx++) {
+            // t5
+            std::vector<int> chunk_tokens(t5_tokens.begin() + chunk_idx * chunk_len,
+                                          t5_tokens.begin() + (chunk_idx + 1) * chunk_len);
+            std::vector<float> chunk_weights(t5_weights.begin() + chunk_idx * chunk_len,
+                                             t5_weights.begin() + (chunk_idx + 1) * chunk_len);
+            std::vector<float> chunk_mask(t5_attn_mask_vec.begin() + chunk_idx * chunk_len,
+                                          t5_attn_mask_vec.begin() + (chunk_idx + 1) * chunk_len);
+
+            auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, chunk_tokens);
+            auto t5_attn_mask_chunk = use_mask ? vector_to_ggml_tensor(work_ctx, chunk_mask) : NULL;
+
+            t5->compute(n_threads,
+                        input_ids,
+                        t5_attn_mask_chunk,
+                        &chunk_hidden_states,
+                        work_ctx);
+            {
+                auto tensor         = chunk_hidden_states;
+                float original_mean = ggml_tensor_mean(tensor);
+                for (int i2 = 0; i2 < tensor->ne[2]; i2++) {
+                    for (int i1 = 0; i1 < tensor->ne[1]; i1++) {
+                        for (int i0 = 0; i0 < tensor->ne[0]; i0++) {
+                            float value = ggml_tensor_get_f32(tensor, i0, i1, i2);
+                            value *= chunk_weights[i1];
+                            ggml_tensor_set_f32(tensor, value, i0, i1, i2);
+                        }
+                    }
+                }
+                float new_mean = ggml_tensor_mean(tensor);
+                ggml_tensor_scale(tensor, (original_mean / new_mean));
+            }
+
+            int64_t t1 = ggml_time_ms();
+            LOG_DEBUG("computing condition graph completed, taking %" PRId64 " ms", t1 - t0);
+            if (force_zero_embeddings) {
+                float* vec = (float*)chunk_hidden_states->data;
+                for (int i = 0; i < ggml_nelements(chunk_hidden_states); i++) {
+                    vec[i] = 0;
+                }
+            }
+
+            hidden_states_vec.insert(hidden_states_vec.end(),
+                                     (float*)chunk_hidden_states->data,
+                                     ((float*)chunk_hidden_states->data) + ggml_nelements(chunk_hidden_states));
+        }
+
+        if (hidden_states_vec.size() > 0) {
+            hidden_states = vector_to_ggml_tensor(work_ctx, hidden_states_vec);
+            hidden_states = ggml_reshape_2d(work_ctx,
+                                            hidden_states,
+                                            chunk_hidden_states->ne[0],
+                                            ggml_nelements(hidden_states) / chunk_hidden_states->ne[0]);
+        } else {
+            hidden_states = ggml_new_tensor_2d(work_ctx, GGML_TYPE_F32, 4096, 256);
+            ggml_set_f32(hidden_states, 0.f);
+        }
+
+        modify_mask_to_attend_padding(t5_attn_mask, ggml_nelements(t5_attn_mask), mask_pad);
+
+        return SDCondition(hidden_states, t5_attn_mask, NULL);
+    }
+
+    SDCondition get_learned_condition(ggml_context* work_ctx,
+                                      int n_threads,
+                                      const std::string& text,
+                                      int clip_skip,
+                                      int width,
+                                      int height,
+                                      int adm_in_channels        = -1,
+                                      bool force_zero_embeddings = false) {
+        auto tokens_and_weights = tokenize(text, chunk_len, true);
+        return get_learned_condition_common(work_ctx, n_threads, tokens_and_weights, clip_skip, force_zero_embeddings);
+    }
+
+    std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(ggml_context* work_ctx,
+                                                                                  int n_threads,
+                                                                                  const std::string& text,
+                                                                                  int clip_skip,
+                                                                                  int width,
+                                                                                  int height,
+                                                                                  int num_input_imgs,
+                                                                                  int adm_in_channels        = -1,
+                                                                                  bool force_zero_embeddings = false) {
+        GGML_ASSERT(0 && "Not implemented yet!");
+    }
+
+    std::string remove_trigger_from_prompt(ggml_context* work_ctx,
+                                           const std::string& prompt) {
+        GGML_ASSERT(0 && "Not implemented yet!");
+    }
+};
+
+#endif

control.hpp

@@ -34,11 +34,11 @@ public:
 
     ControlNetBlock(SDVersion version = VERSION_SD1)
         : version(version) {
-        if (version == VERSION_SD2) {
+        if (sd_version_is_sd2(version)) {
             context_dim       = 1024;
             num_head_channels = 64;
             num_heads         = -1;
-        } else if (version == VERSION_SDXL) {
+        } else if (sd_version_is_sdxl(version)) {
             context_dim           = 2048;
             attention_resolutions = {4, 2};
             channel_mult          = {1, 2, 4};
@@ -58,7 +58,7 @@ public:
         // time_embed_1 is nn.SiLU()
         blocks["time_embed.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim));
 
-        if (version == VERSION_SDXL || version == VERSION_SVD) {
+        if (sd_version_is_sdxl(version) || version == VERSION_SVD) {
             blocks["label_emb.0.0"] = std::shared_ptr<GGMLBlock>(new Linear(adm_in_channels, time_embed_dim));
             // label_emb_1 is nn.SiLU()
             blocks["label_emb.0.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim));
@@ -317,10 +317,10 @@ struct ControlNet : public GGMLRunner {
     bool guided_hint_cached         = false;
 
     ControlNet(ggml_backend_t backend,
-               ggml_type wtype,
+               std::map<std::string, enum ggml_type>& tensor_types,
                SDVersion version = VERSION_SD1)
-        : GGMLRunner(backend, wtype), control_net(version) {
-        control_net.init(params_ctx, wtype);
+        : GGMLRunner(backend), control_net(version) {
+        control_net.init(params_ctx, tensor_types, "");
     }
 
     ~ControlNet() {

denoiser.hpp

@@ -474,7 +474,8 @@ static void sample_k_diffusion(sample_method_t method,
                                ggml_context* work_ctx,
                                ggml_tensor* x,
                                std::vector<float> sigmas,
-                               std::shared_ptr<RNG> rng) {
+                               std::shared_ptr<RNG> rng,
+                               float eta) {
     size_t steps = sigmas.size() - 1;
     // sample_euler_ancestral
     switch (method) {
@@ -1005,6 +1006,370 @@ static void sample_k_diffusion(sample_method_t method,
                 }
             }
         } break;
+        case DDIM_TRAILING:  // Denoising Diffusion Implicit Models
+                             // with the "trailing" timestep spacing
+        {
+            // See J. Song et al., "Denoising Diffusion Implicit
+            // Models", arXiv:2010.02502 [cs.LG]
+            //
+            // DDIM itself needs alphas_cumprod (DDPM, J. Ho et al.,
+            // arXiv:2006.11239 [cs.LG] with k-diffusion's start and
+            // end beta) (which unfortunately k-diffusion's data
+            // structure hides from the denoiser), and the sigmas are
+            // also needed to invert the behavior of CompVisDenoiser
+            // (k-diffusion's LMSDiscreteScheduler)
+            float beta_start = 0.00085f;
+            float beta_end   = 0.0120f;
+            std::vector<double> alphas_cumprod;
+            std::vector<double> compvis_sigmas;
+
+            alphas_cumprod.reserve(TIMESTEPS);
+            compvis_sigmas.reserve(TIMESTEPS);
+            for (int i = 0; i < TIMESTEPS; i++) {
+                alphas_cumprod[i] =
+                    (i == 0 ? 1.0f : alphas_cumprod[i - 1]) *
+                    (1.0f -
+                     std::pow(sqrtf(beta_start) +
+                                  (sqrtf(beta_end) - sqrtf(beta_start)) *
+                                      ((float)i / (TIMESTEPS - 1)),
+                              2));
+                compvis_sigmas[i] =
+                    std::sqrt((1 - alphas_cumprod[i]) /
+                              alphas_cumprod[i]);
+            }
+
+            struct ggml_tensor* pred_original_sample =
+                ggml_dup_tensor(work_ctx, x);
+            struct ggml_tensor* variance_noise =
+                ggml_dup_tensor(work_ctx, x);
+
+            for (int i = 0; i < steps; i++) {
+                // The "trailing" DDIM timestep, see S. Lin et al.,
+                // "Common Diffusion Noise Schedules and Sample Steps
+                // are Flawed", arXiv:2305.08891 [cs], p. 4, Table
+                // 2. Most variables below follow Diffusers naming
+                //
+                // Diffuser naming vs. Song et al. (2010), p. 5, (12)
+                // and p. 16, (16) (<variable name> -> <name in
+                // paper>):
+                //
+                // - pred_noise_t -> epsilon_theta^(t)(x_t)
+                // - pred_original_sample -> f_theta^(t)(x_t) or x_0
+                // - std_dev_t -> sigma_t (not the LMS sigma)
+                // - eta -> eta (set to 0 at the moment)
+                // - pred_sample_direction -> "direction pointing to
+                //   x_t"
+                // - pred_prev_sample -> "x_t-1"
+                int timestep =
+                    roundf(TIMESTEPS -
+                           i * ((float)TIMESTEPS / steps)) -
+                    1;
+                // 1. get previous step value (=t-1)
+                int prev_timestep = timestep - TIMESTEPS / steps;
+                // The sigma here is chosen to cause the
+                // CompVisDenoiser to produce t = timestep
+                float sigma = compvis_sigmas[timestep];
+                if (i == 0) {
+                    // The function add_noise intializes x to
+                    // Diffusers' latents * sigma (as in Diffusers'
+                    // pipeline) or sample * sigma (Diffusers'
+                    // scheduler), where this sigma = init_noise_sigma
+                    // in Diffusers. For DDPM and DDIM however,
+                    // init_noise_sigma = 1. But the k-diffusion
+                    // model() also evaluates F_theta(c_in(sigma) x;
+                    // ...) instead of the bare U-net F_theta, with
+                    // c_in = 1 / sqrt(sigma^2 + 1), as defined in
+                    // T. Karras et al., "Elucidating the Design Space
+                    // of Diffusion-Based Generative Models",
+                    // arXiv:2206.00364 [cs.CV], p. 3, Table 1. Hence
+                    // the first call has to be prescaled as x <- x /
+                    // (c_in * sigma) with the k-diffusion pipeline
+                    // and CompVisDenoiser.
+                    float* vec_x = (float*)x->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] *= std::sqrt(sigma * sigma + 1) /
+                                    sigma;
+                    }
+                } else {
+                    // For the subsequent steps after the first one,
+                    // at this point x = latents or x = sample, and
+                    // needs to be prescaled with x <- sample / c_in
+                    // to compensate for model() applying the scale
+                    // c_in before the U-net F_theta
+                    float* vec_x = (float*)x->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] *= std::sqrt(sigma * sigma + 1);
+                    }
+                }
+                // Note (also noise_pred in Diffuser's pipeline)
+                // model_output = model() is the D(x, sigma) as
+                // defined in Karras et al. (2022), p. 3, Table 1 and
+                // p. 8 (7), compare also p. 38 (226) therein.
+                struct ggml_tensor* model_output =
+                    model(x, sigma, i + 1);
+                // Here model_output is still the k-diffusion denoiser
+                // output, not the U-net output F_theta(c_in(sigma) x;
+                // ...) in Karras et al. (2022), whereas Diffusers'
+                // model_output is F_theta(...). Recover the actual
+                // model_output, which is also referred to as the
+                // "Karras ODE derivative" d or d_cur in several
+                // samplers above.
+                {
+                    float* vec_x = (float*)x->data;
+                    float* vec_model_output =
+                        (float*)model_output->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_model_output[j] =
+                            (vec_x[j] - vec_model_output[j]) *
+                            (1 / sigma);
+                    }
+                }
+                // 2. compute alphas, betas
+                float alpha_prod_t = alphas_cumprod[timestep];
+                // Note final_alpha_cumprod = alphas_cumprod[0] due to
+                // trailing timestep spacing
+                float alpha_prod_t_prev = prev_timestep >= 0 ? alphas_cumprod[prev_timestep] : alphas_cumprod[0];
+                float beta_prod_t       = 1 - alpha_prod_t;
+                // 3. compute predicted original sample from predicted
+                // noise also called "predicted x_0" of formula (12)
+                // from https://arxiv.org/pdf/2010.02502.pdf
+                {
+                    float* vec_x = (float*)x->data;
+                    float* vec_model_output =
+                        (float*)model_output->data;
+                    float* vec_pred_original_sample =
+                        (float*)pred_original_sample->data;
+                    // Note the substitution of latents or sample = x
+                    // * c_in = x / sqrt(sigma^2 + 1)
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_pred_original_sample[j] =
+                            (vec_x[j] / std::sqrt(sigma * sigma + 1) -
+                             std::sqrt(beta_prod_t) *
+                                 vec_model_output[j]) *
+                            (1 / std::sqrt(alpha_prod_t));
+                    }
+                }
+                // Assuming the "epsilon" prediction type, where below
+                // pred_epsilon = model_output is inserted, and is not
+                // defined/copied explicitly.
+                //
+                // 5. compute variance: "sigma_t(eta)" -> see formula
+                // (16)
+                //
+                // sigma_t = sqrt((1 - alpha_t-1)/(1 - alpha_t)) *
+                // sqrt(1 - alpha_t/alpha_t-1)
+                float beta_prod_t_prev = 1 - alpha_prod_t_prev;
+                float variance         = (beta_prod_t_prev / beta_prod_t) *
+                                 (1 - alpha_prod_t / alpha_prod_t_prev);
+                float std_dev_t = eta * std::sqrt(variance);
+                // 6. compute "direction pointing to x_t" of formula
+                // (12) from https://arxiv.org/pdf/2010.02502.pdf
+                // 7. compute x_t without "random noise" of formula
+                // (12) from https://arxiv.org/pdf/2010.02502.pdf
+                {
+                    float* vec_model_output = (float*)model_output->data;
+                    float* vec_pred_original_sample =
+                        (float*)pred_original_sample->data;
+                    float* vec_x = (float*)x->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        // Two step inner loop without an explicit
+                        // tensor
+                        float pred_sample_direction =
+                            std::sqrt(1 - alpha_prod_t_prev -
+                                      std::pow(std_dev_t, 2)) *
+                            vec_model_output[j];
+                        vec_x[j] = std::sqrt(alpha_prod_t_prev) *
+                                       vec_pred_original_sample[j] +
+                                   pred_sample_direction;
+                    }
+                }
+                if (eta > 0) {
+                    ggml_tensor_set_f32_randn(variance_noise, rng);
+                    float* vec_variance_noise =
+                        (float*)variance_noise->data;
+                    float* vec_x = (float*)x->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] += std_dev_t * vec_variance_noise[j];
+                    }
+                }
+                // See the note above: x = latents or sample here, and
+                // is not scaled by the c_in. For the final output
+                // this is correct, but for subsequent iterations, x
+                // needs to be prescaled again, since k-diffusion's
+                // model() differes from the bare U-net F_theta by the
+                // factor c_in.
+            }
+        } break;
+        case TCD:  // Strategic Stochastic Sampling (Algorithm 4) in
+                   // Trajectory Consistency Distillation
+        {
+            // See J. Zheng et al., "Trajectory Consistency
+            // Distillation: Improved Latent Consistency Distillation
+            // by Semi-Linear Consistency Function with Trajectory
+            // Mapping", arXiv:2402.19159 [cs.CV]
+            float beta_start = 0.00085f;
+            float beta_end   = 0.0120f;
+            std::vector<double> alphas_cumprod;
+            std::vector<double> compvis_sigmas;
+
+            alphas_cumprod.reserve(TIMESTEPS);
+            compvis_sigmas.reserve(TIMESTEPS);
+            for (int i = 0; i < TIMESTEPS; i++) {
+                alphas_cumprod[i] =
+                    (i == 0 ? 1.0f : alphas_cumprod[i - 1]) *
+                    (1.0f -
+                     std::pow(sqrtf(beta_start) +
+                                  (sqrtf(beta_end) - sqrtf(beta_start)) *
+                                      ((float)i / (TIMESTEPS - 1)),
+                              2));
+                compvis_sigmas[i] =
+                    std::sqrt((1 - alphas_cumprod[i]) /
+                              alphas_cumprod[i]);
+            }
+            int original_steps = 50;
+
+            struct ggml_tensor* pred_original_sample =
+                ggml_dup_tensor(work_ctx, x);
+            struct ggml_tensor* noise =
+                ggml_dup_tensor(work_ctx, x);
+
+            for (int i = 0; i < steps; i++) {
+                // Analytic form for TCD timesteps
+                int timestep = TIMESTEPS - 1 -
+                               (TIMESTEPS / original_steps) *
+                                   (int)floor(i * ((float)original_steps / steps));
+                // 1. get previous step value
+                int prev_timestep = i >= steps - 1 ? 0 : TIMESTEPS - 1 - (TIMESTEPS / original_steps) * (int)floor((i + 1) * ((float)original_steps / steps));
+                // Here timestep_s is tau_n' in Algorithm 4. The _s
+                // notation appears to be that from C. Lu,
+                // "DPM-Solver: A Fast ODE Solver for Diffusion
+                // Probabilistic Model Sampling in Around 10 Steps",
+                // arXiv:2206.00927 [cs.LG], but this notation is not
+                // continued in Algorithm 4, where _n' is used.
+                int timestep_s =
+                    (int)floor((1 - eta) * prev_timestep);
+                // Begin k-diffusion specific workaround for
+                // evaluating F_theta(x; ...) from D(x, sigma), same
+                // as in DDIM (and see there for detailed comments)
+                float sigma = compvis_sigmas[timestep];
+                if (i == 0) {
+                    float* vec_x = (float*)x->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] *= std::sqrt(sigma * sigma + 1) /
+                                    sigma;
+                    }
+                } else {
+                    float* vec_x = (float*)x->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] *= std::sqrt(sigma * sigma + 1);
+                    }
+                }
+                struct ggml_tensor* model_output =
+                    model(x, sigma, i + 1);
+                {
+                    float* vec_x = (float*)x->data;
+                    float* vec_model_output =
+                        (float*)model_output->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_model_output[j] =
+                            (vec_x[j] - vec_model_output[j]) *
+                            (1 / sigma);
+                    }
+                }
+                // 2. compute alphas, betas
+                //
+                // When comparing TCD with DDPM/DDIM note that Zheng
+                // et al. (2024) follows the DPM-Solver notation for
+                // alpha. One can find the following comment in the
+                // original DPM-Solver code
+                // (https://github.com/LuChengTHU/dpm-solver/):
+                // "**Important**: Please pay special attention for
+                // the args for `alphas_cumprod`: The `alphas_cumprod`
+                // is the \hat{alpha_n} arrays in the notations of
+                // DDPM. [...] Therefore, the notation \hat{alpha_n}
+                // is different from the notation alpha_t in
+                // DPM-Solver. In fact, we have alpha_{t_n} =
+                // \sqrt{\hat{alpha_n}}, [...]"
+                float alpha_prod_t = alphas_cumprod[timestep];
+                float beta_prod_t  = 1 - alpha_prod_t;
+                // Note final_alpha_cumprod = alphas_cumprod[0] since
+                // TCD is always "trailing"
+                float alpha_prod_t_prev = prev_timestep >= 0 ? alphas_cumprod[prev_timestep] : alphas_cumprod[0];
+                // The subscript _s are the only portion in this
+                // section (2) unique to TCD
+                float alpha_prod_s = alphas_cumprod[timestep_s];
+                float beta_prod_s  = 1 - alpha_prod_s;
+                // 3. Compute the predicted noised sample x_s based on
+                // the model parameterization
+                //
+                // This section is also exactly the same as DDIM
+                {
+                    float* vec_x = (float*)x->data;
+                    float* vec_model_output =
+                        (float*)model_output->data;
+                    float* vec_pred_original_sample =
+                        (float*)pred_original_sample->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_pred_original_sample[j] =
+                            (vec_x[j] / std::sqrt(sigma * sigma + 1) -
+                             std::sqrt(beta_prod_t) *
+                                 vec_model_output[j]) *
+                            (1 / std::sqrt(alpha_prod_t));
+                    }
+                }
+                // This consistency function step can be difficult to
+                // decipher from Algorithm 4, as it is simply stated
+                // using a consistency function. This step is the
+                // modified DDIM, i.e. p. 8 (32) in Zheng et
+                // al. (2024), with eta set to 0 (see the paragraph
+                // immediately thereafter that states this somewhat
+                // obliquely).
+                {
+                    float* vec_pred_original_sample =
+                        (float*)pred_original_sample->data;
+                    float* vec_model_output =
+                        (float*)model_output->data;
+                    float* vec_x = (float*)x->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        // Substituting x = pred_noised_sample and
+                        // pred_epsilon = model_output
+                        vec_x[j] =
+                            std::sqrt(alpha_prod_s) *
+                                vec_pred_original_sample[j] +
+                            std::sqrt(beta_prod_s) *
+                                vec_model_output[j];
+                    }
+                }
+                // 4. Sample and inject noise z ~ N(0, I) for
+                // MultiStep Inference Noise is not used on the final
+                // timestep of the timestep schedule. This also means
+                // that noise is not used for one-step sampling. Eta
+                // (referred to as "gamma" in the paper) was
+                // introduced to control the stochasticity in every
+                // step. When eta = 0, it represents deterministic
+                // sampling, whereas eta = 1 indicates full stochastic
+                // sampling.
+                if (eta > 0 && i != steps - 1) {
+                    // In this case, x is still pred_noised_sample,
+                    // continue in-place
+                    ggml_tensor_set_f32_randn(noise, rng);
+                    float* vec_x     = (float*)x->data;
+                    float* vec_noise = (float*)noise->data;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        // Corresponding to (35) in Zheng et
+                        // al. (2024), substituting x =
+                        // pred_noised_sample
+                        vec_x[j] =
+                            std::sqrt(alpha_prod_t_prev /
+                                      alpha_prod_s) *
+                                vec_x[j] +
+                            std::sqrt(1 - alpha_prod_t_prev /
+                                              alpha_prod_s) *
+                                vec_noise[j];
+                    }
+                }
+            }
+        } break;
 
         default:
             LOG_ERROR("Attempting to sample with nonexisting sample method %i", method);

diffusion_model.hpp

@@ -13,6 +13,7 @@ struct DiffusionModel {
                          struct ggml_tensor* c_concat,
                          struct ggml_tensor* y,
                          struct ggml_tensor* guidance,
+                         std::vector<ggml_tensor*> ref_latents     = {},
                          int num_video_frames                      = -1,
                          std::vector<struct ggml_tensor*> controls = {},
                          float control_strength                    = 0.f,
@@ -31,9 +32,10 @@ struct UNetModel : public DiffusionModel {
     UNetModelRunner unet;
 
     UNetModel(ggml_backend_t backend,
-              ggml_type wtype,
-              SDVersion version = VERSION_SD1)
-        : unet(backend, wtype, version) {
+              std::map<std::string, enum ggml_type>& tensor_types,
+              SDVersion version = VERSION_SD1,
+              bool flash_attn   = false)
+        : unet(backend, tensor_types, "model.diffusion_model", version, flash_attn) {
     }
 
     void alloc_params_buffer() {
@@ -67,6 +69,7 @@ struct UNetModel : public DiffusionModel {
                  struct ggml_tensor* c_concat,
                  struct ggml_tensor* y,
                  struct ggml_tensor* guidance,
+                 std::vector<ggml_tensor*> ref_latents     = {},
                  int num_video_frames                      = -1,
                  std::vector<struct ggml_tensor*> controls = {},
                  float control_strength                    = 0.f,
@@ -82,9 +85,8 @@ struct MMDiTModel : public DiffusionModel {
     MMDiTRunner mmdit;
 
     MMDiTModel(ggml_backend_t backend,
-               ggml_type wtype,
-               SDVersion version = VERSION_SD3_2B)
-        : mmdit(backend, wtype, version) {
+               std::map<std::string, enum ggml_type>& tensor_types)
+        : mmdit(backend, tensor_types, "model.diffusion_model") {
     }
 
     void alloc_params_buffer() {
@@ -118,6 +120,7 @@ struct MMDiTModel : public DiffusionModel {
                  struct ggml_tensor* c_concat,
                  struct ggml_tensor* y,
                  struct ggml_tensor* guidance,
+                 std::vector<ggml_tensor*> ref_latents     = {},
                  int num_video_frames                      = -1,
                  std::vector<struct ggml_tensor*> controls = {},
                  float control_strength                    = 0.f,
@@ -132,9 +135,11 @@ struct FluxModel : public DiffusionModel {
     Flux::FluxRunner flux;
 
     FluxModel(ggml_backend_t backend,
-              ggml_type wtype,
-              SDVersion version = VERSION_FLUX_DEV)
-        : flux(backend, wtype, version) {
+              std::map<std::string, enum ggml_type>& tensor_types,
+              SDVersion version = VERSION_FLUX,
+              bool flash_attn   = false,
+              bool use_mask     = false)
+        : flux(backend, tensor_types, "model.diffusion_model", version, flash_attn, use_mask) {
     }
 
     void alloc_params_buffer() {
@@ -168,14 +173,15 @@ struct FluxModel : public DiffusionModel {
                  struct ggml_tensor* c_concat,
                  struct ggml_tensor* y,
                  struct ggml_tensor* guidance,
+                 std::vector<ggml_tensor*> ref_latents     = {},
                  int num_video_frames                      = -1,
                  std::vector<struct ggml_tensor*> controls = {},
                  float control_strength                    = 0.f,
                  struct ggml_tensor** output               = NULL,
                  struct ggml_context* output_ctx           = NULL,
                  std::vector<int> skip_layers              = std::vector<int>()) {
-        return flux.compute(n_threads, x, timesteps, context, y, guidance, output, output_ctx, skip_layers);
+        return flux.compute(n_threads, x, timesteps, context, c_concat, y, guidance, ref_latents, output, output_ctx, skip_layers);
     }
 };
 
-#endif
+#endif

docs/chroma.md

@@ -0,0 +1,33 @@
+# How to Use
+
+You can run Chroma using stable-diffusion.cpp with a GPU that has 6GB or even 4GB of VRAM, without needing to offload to RAM.
+
+## Download weights
+
+- Download Chroma
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF)
+    - Otherwise, download chroma's safetensors from [lodestones/Chroma](https://huggingface.co/lodestones/Chroma)
+- Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## Convert Chroma weights
+
+You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](https://huggingface.co/silveroxides/Chroma-GGUF), this way you don't have to do the conversion yourself.
+
+```
+.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\chroma-unlocked-v40.safetensors -o ..\models\chroma-unlocked-v40-q8_0.gguf -v --type q8_0
+```
+
+## Run
+
+### Example
+For example:
+
+```
+ .\bin\Release\sd.exe -diffusion-model  ..\models\chroma-unlocked-v40-q8_0.gguf --vae ..\models\ae.sft --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'chroma.cpp'" --cfg-scale 4.0 --sampling-method euler -v --chroma-disable-dit-mask
+```
+
+![](../assets/flux/chroma_v40.png)
+
+
+

docs/kontext.md

@@ -0,0 +1,39 @@
+# How to Use
+
+You can run Kontext using stable-diffusion.cpp with a GPU that has 6GB or even 4GB of VRAM, without needing to offload to RAM.
+
+## Download weights
+
+- Download Kontext
+    - If you don't want to do the conversion yourself, download the preconverted gguf model from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF)
+    - Otherwise, download FLUX.1-Kontext-dev from https://huggingface.co/black-forest-labs/FLUX.1-Kontext-dev/blob/main/flux1-kontext-dev.safetensors
+- Download vae from https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- Download clip_l from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/clip_l.safetensors
+- Download t5xxl from https://huggingface.co/comfyanonymous/flux_text_encoders/blob/main/t5xxl_fp16.safetensors
+
+## Convert Kontext weights
+
+You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](https://huggingface.co/QuantStack/FLUX.1-Kontext-dev-GGUF), this way you don't have to do the conversion yourself.
+
+```
+.\bin\Release\sd.exe -M convert -m ..\..\ComfyUI\models\unet\flux1-kontext-dev.safetensors -o ..\models\flux1-kontext-dev-q8_0.gguf -v --type q8_0
+```
+
+## Run
+
+- `--cfg-scale` is recommended to be set to 1. 
+
+### Example
+For example:
+
+```
+ .\bin\Release\sd.exe -M edit -r .\flux1-dev-q8_0.png --diffusion-model  ..\models\flux1-kontext-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors -p "change 'flux.cpp' to 'kontext.cpp'" --cfg-scale 1.0 --sampling-method euler -v
+```
+
+
+| ref_image | prompt  | output  |
+| ---- | ----  |----  |
+| ![](../assets/flux/flux1-dev-q8_0.png) | change 'flux.cpp' to 'kontext.cpp' |![](../assets/flux/kontext1_dev_output.png) |
+
+
+

docs/photo_maker.md

@@ -29,4 +29,26 @@ Example:
 
 ```bash
 bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --stacked-id-embd-dir ../models/photomaker-v1.safetensors --input-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --style-ratio 10 --vae-on-cpu -o output.png
-```
+```
+
+## PhotoMaker Version 2
+
+[PhotoMaker Version 2 (PMV2)](https://github.com/TencentARC/PhotoMaker/blob/main/README_pmv2.md) has some key improvements. Unfortunately it has a very heavy dependency which makes running it a bit involved in ```SD.cpp```. 
+
+Running PMV2 is now a two-step process:
+
+- Run a python script ```face_detect.py``` to obtain **id_embeds** for the given input images
+```
+python face_detect.py input_image_dir
+```
+An ```id_embeds.safetensors``` file will be generated in ```input_images_dir```
+
+**Note: this step is only needed to run once; the same ```id_embeds``` can be reused**
+
+- Run the same command as in version 1 but replacing ```photomaker-v1.safetensors``` with ```photomaker-v2.safetensors```.
+
+  You can download ```photomaker-v2.safetensors``` from [here](https://huggingface.co/bssrdf/PhotoMakerV2)
+
+- All the command line parameters from Version 1 remain the same for Version 2
+
+

esrgan.hpp

@@ -130,8 +130,8 @@ public:
         body_feat = conv_body->forward(ctx, body_feat);
         feat      = ggml_add(ctx, feat, body_feat);
         // upsample
-        feat     = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2)));
-        feat     = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2)));
+        feat     = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+        feat     = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
         auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
         return out;
     }
@@ -142,10 +142,9 @@ struct ESRGAN : public GGMLRunner {
     int scale     = 4;
     int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
 
-    ESRGAN(ggml_backend_t backend,
-           ggml_type wtype)
-        : GGMLRunner(backend, wtype) {
-        rrdb_net.init(params_ctx, wtype);
+    ESRGAN(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types)
+        : GGMLRunner(backend) {
+        rrdb_net.init(params_ctx, tensor_types, "");
     }
 
     std::string get_desc() {

examples/cli/main.cpp

@@ -39,6 +39,8 @@ const char* sample_method_str[] = {
     "ipndm",
     "ipndm_v",
     "lcm",
+    "ddim_trailing",
+    "tcd",
 };
 
 // Names of the sigma schedule overrides, same order as sample_schedule in stable-diffusion.h
@@ -55,6 +57,7 @@ const char* modes_str[] = {
     "txt2img",
     "img2img",
     "img2vid",
+    "edit",
     "convert",
 };
 
@@ -62,6 +65,7 @@ enum SDMode {
     TXT2IMG,
     IMG2IMG,
     IMG2VID,
+    EDIT,
     CONVERT,
     MODE_COUNT
 };
@@ -85,13 +89,16 @@ struct SDParams {
     std::string lora_model_dir;
     std::string output_path = "output.png";
     std::string input_path;
+    std::string mask_path;
     std::string control_image_path;
+    std::vector<std::string> ref_image_paths;
 
     std::string prompt;
     std::string negative_prompt;
     float min_cfg     = 1.0f;
     float cfg_scale   = 7.0f;
     float guidance    = 3.5f;
+    float eta         = 0.f;
     float style_ratio = 20.f;
     int clip_skip     = -1;  // <= 0 represents unspecified
     int width         = 512;
@@ -116,14 +123,19 @@ struct SDParams {
     bool normalize_input          = false;
     bool clip_on_cpu              = false;
     bool vae_on_cpu               = false;
+    bool diffusion_flash_attn     = false;
     bool canny_preprocess         = false;
     bool color                    = false;
     int upscale_repeats           = 1;
 
     std::vector<int> skip_layers = {7, 8, 9};
-    float slg_scale              = 0.;
-    float skip_layer_start       = 0.01;
-    float skip_layer_end         = 0.2;
+    float slg_scale              = 0.f;
+    float skip_layer_start       = 0.01f;
+    float skip_layer_end         = 0.2f;
+
+    bool chroma_use_dit_mask = true;
+    bool chroma_use_t5_mask  = false;
+    int chroma_t5_mask_pad   = 1;
 };
 
 void print_params(SDParams params) {
@@ -147,10 +159,16 @@ void print_params(SDParams params) {
     printf("    normalize input image :  %s\n", params.normalize_input ? "true" : "false");
     printf("    output_path:       %s\n", params.output_path.c_str());
     printf("    init_img:          %s\n", params.input_path.c_str());
+    printf("    mask_img:          %s\n", params.mask_path.c_str());
     printf("    control_image:     %s\n", params.control_image_path.c_str());
+    printf("    ref_images_paths:\n");
+    for (auto& path : params.ref_image_paths) {
+        printf("        %s\n", path.c_str());
+    };
     printf("    clip on cpu:       %s\n", params.clip_on_cpu ? "true" : "false");
     printf("    controlnet cpu:    %s\n", params.control_net_cpu ? "true" : "false");
     printf("    vae decoder on cpu:%s\n", params.vae_on_cpu ? "true" : "false");
+    printf("    diffusion flash attention:%s\n", params.diffusion_flash_attn ? "true" : "false");
     printf("    strength(control): %.2f\n", params.control_strength);
     printf("    prompt:            %s\n", params.prompt.c_str());
     printf("    negative_prompt:   %s\n", params.negative_prompt.c_str());
@@ -158,6 +176,7 @@ void print_params(SDParams params) {
     printf("    cfg_scale:         %.2f\n", params.cfg_scale);
     printf("    slg_scale:         %.2f\n", params.slg_scale);
     printf("    guidance:          %.2f\n", params.guidance);
+    printf("    eta:               %.2f\n", params.eta);
     printf("    clip_skip:         %d\n", params.clip_skip);
     printf("    width:             %d\n", params.width);
     printf("    height:            %d\n", params.height);
@@ -170,6 +189,9 @@ void print_params(SDParams params) {
     printf("    batch_count:       %d\n", params.batch_count);
     printf("    vae_tiling:        %s\n", params.vae_tiling ? "true" : "false");
     printf("    upscale_repeats:   %d\n", params.upscale_repeats);
+    printf("    chroma_use_dit_mask:   %s\n", params.chroma_use_dit_mask ? "true" : "false");
+    printf("    chroma_use_t5_mask:    %s\n", params.chroma_use_t5_mask ? "true" : "false");
+    printf("    chroma_t5_mask_pad:    %d\n", params.chroma_t5_mask_pad);
 }
 
 void print_usage(int argc, const char* argv[]) {
@@ -183,7 +205,7 @@ void print_usage(int argc, const char* argv[]) {
     printf("  -m, --model [MODEL]                path to full model\n");
     printf("  --diffusion-model                  path to the standalone diffusion model\n");
     printf("  --clip_l                           path to the clip-l text encoder\n");
-    printf("  --clip_g                           path to the clip-l text encoder\n");
+    printf("  --clip_g                           path to the clip-g text encoder\n");
     printf("  --t5xxl                            path to the the t5xxl text encoder\n");
     printf("  --vae [VAE]                        path to vae\n");
     printf("  --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)\n");
@@ -194,20 +216,24 @@ void print_usage(int argc, const char* argv[]) {
     printf("  --normalize-input                  normalize PHOTOMAKER input id images\n");
     printf("  --upscale-model [ESRGAN_PATH]      path to esrgan model. Upscale images after generate, just RealESRGAN_x4plus_anime_6B supported by now\n");
     printf("  --upscale-repeats                  Run the ESRGAN upscaler this many times (default 1)\n");
-    printf("  --type [TYPE]                      weight type (f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_k, q3_k, q4_k)\n");
+    printf("  --type [TYPE]                      weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K)\n");
     printf("                                     If not specified, the default is the type of the weight file\n");
     printf("  --lora-model-dir [DIR]             lora model directory\n");
     printf("  -i, --init-img [IMAGE]             path to the input image, required by img2img\n");
+    printf("  --mask [MASK]                      path to the mask image, required by img2img with mask\n");
     printf("  --control-image [IMAGE]            path to image condition, control net\n");
+    printf("  -r, --ref_image [PATH]             reference image for Flux Kontext models (can be used multiple times) \n");
     printf("  -o, --output OUTPUT                path to write result image to (default: ./output.png)\n");
     printf("  -p, --prompt [PROMPT]              the prompt to render\n");
     printf("  -n, --negative-prompt PROMPT       the negative prompt (default: \"\")\n");
     printf("  --cfg-scale SCALE                  unconditional guidance scale: (default: 7.0)\n");
+    printf("  --guidance SCALE                   guidance scale for img2img (default: 3.5)\n");
     printf("  --slg-scale SCALE                  skip layer guidance (SLG) scale, only for DiT models: (default: 0)\n");
     printf("                                     0 means disabled, a value of 2.5 is nice for sd3.5 medium\n");
-    printf("  --skip_layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])\n");
-    printf("  --skip_layer_start START           SLG enabling point: (default: 0.01)\n");
-    printf("  --skip_layer_end END               SLG disabling point: (default: 0.2)\n");
+    printf("  --eta SCALE                        eta in DDIM, only for DDIM and TCD: (default: 0)\n");
+    printf("  --skip-layers LAYERS               Layers to skip for SLG steps: (default: [7,8,9])\n");
+    printf("  --skip-layer-start START           SLG enabling point: (default: 0.01)\n");
+    printf("  --skip-layer-end END               SLG disabling point: (default: 0.2)\n");
     printf("                                     SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])\n");
     printf("  --strength STRENGTH                strength for noising/unnoising (default: 0.75)\n");
     printf("  --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20%%)\n");
@@ -215,7 +241,7 @@ void print_usage(int argc, const char* argv[]) {
     printf("                                     1.0 corresponds to full destruction of information in init image\n");
     printf("  -H, --height H                     image height, in pixel space (default: 512)\n");
     printf("  -W, --width W                      image width, in pixel space (default: 512)\n");
-    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm}\n");
+    printf("  --sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}\n");
     printf("                                     sampling method (default: \"euler_a\")\n");
     printf("  --steps  STEPS                     number of sample steps (default: 20)\n");
     printf("  --rng {std_default, cuda}          RNG (default: cuda)\n");
@@ -227,9 +253,15 @@ void print_usage(int argc, const char* argv[]) {
     printf("  --vae-tiling                       process vae in tiles to reduce memory usage\n");
     printf("  --vae-on-cpu                       keep vae in cpu (for low vram)\n");
     printf("  --clip-on-cpu                      keep clip in cpu (for low vram)\n");
+    printf("  --diffusion-fa                     use flash attention in the diffusion model (for low vram)\n");
+    printf("                                     Might lower quality, since it implies converting k and v to f16.\n");
+    printf("                                     This might crash if it is not supported by the backend.\n");
     printf("  --control-net-cpu                  keep controlnet in cpu (for low vram)\n");
     printf("  --canny                            apply canny preprocessor (edge detection)\n");
-    printf("  --color                            Colors the logging tags according to level\n");
+    printf("  --color                            colors the logging tags according to level\n");
+    printf("  --chroma-disable-dit-mask          disable dit mask for chroma\n");
+    printf("  --chroma-enable-t5-mask            enable t5 mask for chroma\n");
+    printf("  --chroma-t5-mask-pad  PAD_SIZE     t5 mask pad size of chroma\n");
     printf("  -v, --verbose                      print extra info\n");
 }
 
@@ -341,30 +373,30 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 invalid_arg = true;
                 break;
             }
-            std::string type = argv[i];
-            if (type == "f32") {
-                params.wtype = SD_TYPE_F32;
-            } else if (type == "f16") {
-                params.wtype = SD_TYPE_F16;
-            } else if (type == "q4_0") {
-                params.wtype = SD_TYPE_Q4_0;
-            } else if (type == "q4_1") {
-                params.wtype = SD_TYPE_Q4_1;
-            } else if (type == "q5_0") {
-                params.wtype = SD_TYPE_Q5_0;
-            } else if (type == "q5_1") {
-                params.wtype = SD_TYPE_Q5_1;
-            } else if (type == "q8_0") {
-                params.wtype = SD_TYPE_Q8_0;
-            } else if (type == "q2_k") {
-                params.wtype = SD_TYPE_Q2_K;
-            } else if (type == "q3_k") {
-                params.wtype = SD_TYPE_Q3_K;
-            } else if (type == "q4_k") {
-                params.wtype = SD_TYPE_Q4_K;
-            } else {
-                fprintf(stderr, "error: invalid weight format %s, must be one of [f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_k, q3_k, q4_k]\n",
-                        type.c_str());
+            std::string type        = argv[i];
+            bool found              = false;
+            std::string valid_types = "";
+            for (size_t i = 0; i < SD_TYPE_COUNT; i++) {
+                auto trait = ggml_get_type_traits((ggml_type)i);
+                std::string name(trait->type_name);
+                if (name == "f32" || trait->to_float && trait->type_size) {
+                    if (i)
+                        valid_types += ", ";
+                    valid_types += name;
+                    if (type == name) {
+                        if (ggml_quantize_requires_imatrix((ggml_type)i)) {
+                            printf("\033[35;1m[WARNING]\033[0m: type %s requires imatrix to work properly. A dummy imatrix will be used, expect poor quality.\n", trait->type_name);
+                        }
+                        params.wtype = (enum sd_type_t)i;
+                        found        = true;
+                        break;
+                    }
+                }
+            }
+            if (!found) {
+                fprintf(stderr, "error: invalid weight format %s, must be one of [%s]\n",
+                        type.c_str(),
+                        valid_types.c_str());
                 exit(1);
             }
         } else if (arg == "--lora-model-dir") {
@@ -379,6 +411,12 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.input_path = argv[i];
+        } else if (arg == "--mask") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.mask_path = argv[i];
         } else if (arg == "--control-image") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -425,6 +463,12 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.guidance = std::stof(argv[i]);
+        } else if (arg == "--eta") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.eta = std::stof(argv[i]);
         } else if (arg == "--strength") {
             if (++i >= argc) {
                 invalid_arg = true;
@@ -477,6 +521,8 @@ void parse_args(int argc, const char** argv, SDParams& params) {
             params.clip_on_cpu = true;  // will slow down get_learned_condiotion but necessary for low MEM GPUs
         } else if (arg == "--vae-on-cpu") {
             params.vae_on_cpu = true;  // will slow down latent decoding but necessary for low MEM GPUs
+        } else if (arg == "--diffusion-fa") {
+            params.diffusion_flash_attn = true;  // can reduce MEM significantly
         } else if (arg == "--canny") {
             params.canny_preprocess = true;
         } else if (arg == "-b" || arg == "--batch-count") {
@@ -601,6 +647,22 @@ void parse_args(int argc, const char** argv, SDParams& params) {
                 break;
             }
             params.skip_layer_end = std::stof(argv[i]);
+        } else if (arg == "-r" || arg == "--ref-image") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.ref_image_paths.push_back(argv[i]);
+        } else if (arg == "--chroma-disable-dit-mask") {
+            params.chroma_use_dit_mask = false;
+        } else if (arg == "--chroma-use-t5-mask") {
+            params.chroma_use_t5_mask = true;
+        } else if (arg == "--chroma-t5-mask-pad") {
+            if (++i >= argc) {
+                invalid_arg = true;
+                break;
+            }
+            params.chroma_t5_mask_pad = std::stoi(argv[i]);
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             print_usage(argc, argv);
@@ -629,7 +691,13 @@ void parse_args(int argc, const char** argv, SDParams& params) {
     }
 
     if ((params.mode == IMG2IMG || params.mode == IMG2VID) && params.input_path.length() == 0) {
-        fprintf(stderr, "error: when using the img2img mode, the following arguments are required: init-img\n");
+        fprintf(stderr, "error: when using the img2img/img2vid mode, the following arguments are required: init-img\n");
+        print_usage(argc, argv);
+        exit(1);
+    }
+
+    if (params.mode == EDIT && params.ref_image_paths.size() == 0) {
+        fprintf(stderr, "error: when using the edit mode, the following arguments are required: ref-image\n");
         print_usage(argc, argv);
         exit(1);
     }
@@ -702,6 +770,7 @@ std::string get_image_params(SDParams params, int64_t seed) {
         parameter_string += "Skip layer end: " + std::to_string(params.skip_layer_end) + ", ";
     }
     parameter_string += "Guidance: " + std::to_string(params.guidance) + ", ";
+    parameter_string += "Eta: " + std::to_string(params.eta) + ", ";
     parameter_string += "Seed: " + std::to_string(seed) + ", ";
     parameter_string += "Size: " + std::to_string(params.width) + "x" + std::to_string(params.height) + ", ";
     parameter_string += "Model: " + sd_basename(params.model_path) + ", ";
@@ -796,6 +865,9 @@ int main(int argc, const char* argv[]) {
     bool vae_decode_only          = true;
     uint8_t* input_image_buffer   = NULL;
     uint8_t* control_image_buffer = NULL;
+    uint8_t* mask_image_buffer    = NULL;
+    std::vector<sd_image_t> ref_images;
+
     if (params.mode == IMG2IMG || params.mode == IMG2VID) {
         vae_decode_only = false;
 
@@ -846,6 +918,37 @@ int main(int argc, const char* argv[]) {
             free(input_image_buffer);
             input_image_buffer = resized_image_buffer;
         }
+    } else if (params.mode == EDIT) {
+        vae_decode_only = false;
+        for (auto& path : params.ref_image_paths) {
+            int c                 = 0;
+            int width             = 0;
+            int height            = 0;
+            uint8_t* image_buffer = stbi_load(path.c_str(), &width, &height, &c, 3);
+            if (image_buffer == NULL) {
+                fprintf(stderr, "load image from '%s' failed\n", path.c_str());
+                return 1;
+            }
+            if (c < 3) {
+                fprintf(stderr, "the number of channels for the input image must be >= 3, but got %d channels\n", c);
+                free(image_buffer);
+                return 1;
+            }
+            if (width <= 0) {
+                fprintf(stderr, "error: the width of image must be greater than 0\n");
+                free(image_buffer);
+                return 1;
+            }
+            if (height <= 0) {
+                fprintf(stderr, "error: the height of image must be greater than 0\n");
+                free(image_buffer);
+                return 1;
+            }
+            ref_images.push_back({(uint32_t)width,
+                                  (uint32_t)height,
+                                  3,
+                                  image_buffer});
+        }
     }
 
     sd_ctx_t* sd_ctx = new_sd_ctx(params.model_path.c_str(),
@@ -868,7 +971,11 @@ int main(int argc, const char* argv[]) {
                                   params.schedule,
                                   params.clip_on_cpu,
                                   params.control_net_cpu,
-                                  params.vae_on_cpu);
+                                  params.vae_on_cpu,
+                                  params.diffusion_flash_attn,
+                                  params.chroma_use_dit_mask,
+                                  params.chroma_use_t5_mask,
+                                  params.chroma_t5_mask_pad);
 
     if (sd_ctx == NULL) {
         printf("new_sd_ctx_t failed\n");
@@ -899,6 +1006,18 @@ int main(int argc, const char* argv[]) {
         }
     }
 
+    std::vector<uint8_t> default_mask_image_vec(params.width * params.height, 255);
+    if (params.mask_path != "") {
+        int c             = 0;
+        mask_image_buffer = stbi_load(params.mask_path.c_str(), &params.width, &params.height, &c, 1);
+    } else {
+        mask_image_buffer = default_mask_image_vec.data();
+    }
+    sd_image_t mask_image = {(uint32_t)params.width,
+                             (uint32_t)params.height,
+                             1,
+                             mask_image_buffer};
+
     sd_image_t* results;
     if (params.mode == TXT2IMG) {
         results = txt2img(sd_ctx,
@@ -907,6 +1026,7 @@ int main(int argc, const char* argv[]) {
                           params.clip_skip,
                           params.cfg_scale,
                           params.guidance,
+                          params.eta,
                           params.width,
                           params.height,
                           params.sample_method,
@@ -918,11 +1038,12 @@ int main(int argc, const char* argv[]) {
                           params.style_ratio,
                           params.normalize_input,
                           params.input_id_images_path.c_str(),
-                          params.skip_layers,
+                          params.skip_layers.data(),
+                          params.skip_layers.size(),
                           params.slg_scale,
                           params.skip_layer_start,
                           params.skip_layer_end);
-    } else {
+    } else if (params.mode == IMG2IMG || params.mode == IMG2VID) {
         sd_image_t input_image = {(uint32_t)params.width,
                                   (uint32_t)params.height,
                                   3,
@@ -967,11 +1088,13 @@ int main(int argc, const char* argv[]) {
         } else {
             results = img2img(sd_ctx,
                               input_image,
+                              mask_image,
                               params.prompt.c_str(),
                               params.negative_prompt.c_str(),
                               params.clip_skip,
                               params.cfg_scale,
                               params.guidance,
+                              params.eta,
                               params.width,
                               params.height,
                               params.sample_method,
@@ -983,8 +1106,39 @@ int main(int argc, const char* argv[]) {
                               params.control_strength,
                               params.style_ratio,
                               params.normalize_input,
-                              params.input_id_images_path.c_str());
+                              params.input_id_images_path.c_str(),
+                              params.skip_layers.data(),
+                              params.skip_layers.size(),
+                              params.slg_scale,
+                              params.skip_layer_start,
+                              params.skip_layer_end);
         }
+    } else {  // EDIT
+        results = edit(sd_ctx,
+                       ref_images.data(),
+                       ref_images.size(),
+                       params.prompt.c_str(),
+                       params.negative_prompt.c_str(),
+                       params.clip_skip,
+                       params.cfg_scale,
+                       params.guidance,
+                       params.eta,
+                       params.width,
+                       params.height,
+                       params.sample_method,
+                       params.sample_steps,
+                       params.strength,
+                       params.seed,
+                       params.batch_count,
+                       control_image,
+                       params.control_strength,
+                       params.style_ratio,
+                       params.normalize_input,
+                       params.skip_layers.data(),
+                       params.skip_layers.size(),
+                       params.slg_scale,
+                       params.skip_layer_start,
+                       params.skip_layer_end);
     }
 
     if (results == NULL) {
@@ -996,8 +1150,7 @@ int main(int argc, const char* argv[]) {
     int upscale_factor = 4;  // unused for RealESRGAN_x4plus_anime_6B.pth
     if (params.esrgan_path.size() > 0 && params.upscale_repeats > 0) {
         upscaler_ctx_t* upscaler_ctx = new_upscaler_ctx(params.esrgan_path.c_str(),
-                                                        params.n_threads,
-                                                        params.wtype);
+                                                        params.n_threads);
 
         if (upscaler_ctx == NULL) {
             printf("new_upscaler_ctx failed\n");
@@ -1021,16 +1174,41 @@ int main(int argc, const char* argv[]) {
         }
     }
 
-    size_t last            = params.output_path.find_last_of(".");
-    std::string dummy_name = last != std::string::npos ? params.output_path.substr(0, last) : params.output_path;
+    std::string dummy_name, ext, lc_ext;
+    bool is_jpg;
+    size_t last      = params.output_path.find_last_of(".");
+    size_t last_path = std::min(params.output_path.find_last_of("/"),
+                                params.output_path.find_last_of("\\"));
+    if (last != std::string::npos  // filename has extension
+        && (last_path == std::string::npos || last > last_path)) {
+        dummy_name = params.output_path.substr(0, last);
+        ext = lc_ext = params.output_path.substr(last);
+        std::transform(ext.begin(), ext.end(), lc_ext.begin(), ::tolower);
+        is_jpg = lc_ext == ".jpg" || lc_ext == ".jpeg" || lc_ext == ".jpe";
+    } else {
+        dummy_name = params.output_path;
+        ext = lc_ext = "";
+        is_jpg       = false;
+    }
+    // appending ".png" to absent or unknown extension
+    if (!is_jpg && lc_ext != ".png") {
+        dummy_name += ext;
+        ext = ".png";
+    }
     for (int i = 0; i < params.batch_count; i++) {
         if (results[i].data == NULL) {
             continue;
         }
-        std::string final_image_path = i > 0 ? dummy_name + "_" + std::to_string(i + 1) + ".png" : dummy_name + ".png";
-        stbi_write_png(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
-                       results[i].data, 0, get_image_params(params, params.seed + i).c_str());
-        printf("save result image to '%s'\n", final_image_path.c_str());
+        std::string final_image_path = i > 0 ? dummy_name + "_" + std::to_string(i + 1) + ext : dummy_name + ext;
+        if (is_jpg) {
+            stbi_write_jpg(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
+                           results[i].data, 90, get_image_params(params, params.seed + i).c_str());
+            printf("save result JPEG image to '%s'\n", final_image_path.c_str());
+        } else {
+            stbi_write_png(final_image_path.c_str(), results[i].width, results[i].height, results[i].channel,
+                           results[i].data, 0, get_image_params(params, params.seed + i).c_str());
+            printf("save result PNG image to '%s'\n", final_image_path.c_str());
+        }
         free(results[i].data);
         results[i].data = NULL;
     }
@@ -1040,4 +1218,4 @@ int main(int argc, const char* argv[]) {
     free(input_image_buffer);
 
     return 0;
-}
+}

face_detect.py

@@ -0,0 +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)
+
+    

flux.hpp

@@ -35,8 +35,9 @@ namespace Flux {
         int64_t hidden_size;
         float eps;
 
-        void init_params(struct ggml_context* ctx, ggml_type wtype) {
-            params["scale"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
+        void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+            ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "scale") != tensor_types.end()) ? tensor_types[prefix + "scale"] : GGML_TYPE_F32;
+            params["scale"] = ggml_new_tensor_1d(ctx, wtype, hidden_size);
         }
 
     public:
@@ -115,25 +116,29 @@ namespace Flux {
                                                     struct ggml_tensor* q,
                                                     struct ggml_tensor* k,
                                                     struct ggml_tensor* v,
-                                                    struct ggml_tensor* pe) {
+                                                    struct ggml_tensor* pe,
+                                                    struct ggml_tensor* mask,
+                                                    bool flash_attn) {
         // q,k,v: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2]
         // return: [N, L, n_head*d_head]
         q = apply_rope(ctx, q, pe);  // [N*n_head, L, d_head]
         k = apply_rope(ctx, k, pe);  // [N*n_head, L, d_head]
 
-        auto x = ggml_nn_attention_ext(ctx, q, k, v, v->ne[1], NULL, false, true);  // [N, L, n_head*d_head]
+        auto x = ggml_nn_attention_ext(ctx, q, k, v, v->ne[1], mask, false, true, flash_attn);  // [N, L, n_head*d_head]
         return x;
     }
 
     struct SelfAttention : public GGMLBlock {
     public:
         int64_t num_heads;
+        bool flash_attn;
 
     public:
         SelfAttention(int64_t dim,
                       int64_t num_heads = 8,
-                      bool qkv_bias     = false)
+                      bool qkv_bias     = false,
+                      bool flash_attn   = false)
             : num_heads(num_heads) {
             int64_t head_dim = dim / num_heads;
             blocks["qkv"]    = std::shared_ptr<GGMLBlock>(new Linear(dim, dim * 3, qkv_bias));
@@ -163,13 +168,13 @@ namespace Flux {
             return x;
         }
 
-        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* pe) {
+        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* pe, struct ggml_tensor* mask) {
             // x: [N, n_token, dim]
             // pe: [n_token, d_head/2, 2, 2]
             // return [N, n_token, dim]
-            auto qkv = pre_attention(ctx, x);                       // q,k,v: [N, n_token, n_head, d_head]
-            x        = attention(ctx, qkv[0], qkv[1], qkv[2], pe);  // [N, n_token, dim]
-            x        = post_attention(ctx, x);                      // [N, n_token, dim]
+            auto qkv = pre_attention(ctx, x);                                         // q,k,v: [N, n_token, n_head, d_head]
+            x        = attention(ctx, qkv[0], qkv[1], qkv[2], pe, mask, flash_attn);  // [N, n_token, dim]
+            x        = post_attention(ctx, x);                                        // [N, n_token, dim]
             return x;
         }
     };
@@ -181,6 +186,13 @@ namespace Flux {
 
         ModulationOut(ggml_tensor* shift = NULL, ggml_tensor* scale = NULL, ggml_tensor* gate = NULL)
             : shift(shift), scale(scale), gate(gate) {}
+
+        ModulationOut(struct ggml_context* ctx, ggml_tensor* vec, int64_t offset) {
+            int64_t stride = vec->nb[1] * vec->ne[1];
+            shift          = ggml_view_2d(ctx, vec, vec->ne[0], vec->ne[1], vec->nb[1], stride * (offset + 0));  // [N, dim]
+            scale          = ggml_view_2d(ctx, vec, vec->ne[0], vec->ne[1], vec->nb[1], stride * (offset + 1));  // [N, dim]
+            gate           = ggml_view_2d(ctx, vec, vec->ne[0], vec->ne[1], vec->nb[1], stride * (offset + 2));  // [N, dim]
+        }
     };
 
     struct Modulation : public GGMLBlock {
@@ -206,19 +218,12 @@ namespace Flux {
             auto m = ggml_reshape_3d(ctx, out, vec->ne[0], multiplier, vec->ne[1]);  // [N, multiplier, dim]
             m      = ggml_cont(ctx, ggml_permute(ctx, m, 0, 2, 1, 3));               // [multiplier, N, dim]
 
-            int64_t offset = m->nb[1] * m->ne[1];
-            auto shift_0   = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 0);  // [N, dim]
-            auto scale_0   = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 1);  // [N, dim]
-            auto gate_0    = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 2);  // [N, dim]
-
+            ModulationOut m_0 = ModulationOut(ctx, m, 0);
             if (is_double) {
-                auto shift_1 = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 3);  // [N, dim]
-                auto scale_1 = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 4);  // [N, dim]
-                auto gate_1  = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 5);  // [N, dim]
-                return {ModulationOut(shift_0, scale_0, gate_0), ModulationOut(shift_1, scale_1, gate_1)};
+                return {m_0, ModulationOut(ctx, m, 3)};
             }
 
-            return {ModulationOut(shift_0, scale_0, gate_0), ModulationOut()};
+            return {m_0, ModulationOut()};
         }
     };
 
@@ -237,24 +242,36 @@ namespace Flux {
     }
 
     struct DoubleStreamBlock : public GGMLBlock {
+        bool flash_attn;
+        bool prune_mod;
+        int idx = 0;
+
     public:
         DoubleStreamBlock(int64_t hidden_size,
                           int64_t num_heads,
                           float mlp_ratio,
-                          bool qkv_bias = false) {
+                          int idx         = 0,
+                          bool qkv_bias   = false,
+                          bool flash_attn = false,
+                          bool prune_mod  = false)
+            : idx(idx), flash_attn(flash_attn), prune_mod(prune_mod) {
             int64_t mlp_hidden_dim = hidden_size * mlp_ratio;
-            blocks["img_mod"]      = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
-            blocks["img_norm1"]    = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
-            blocks["img_attn"]     = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias));
+            if (!prune_mod) {
+                blocks["img_mod"] = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
+            }
+            blocks["img_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
+            blocks["img_attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias, flash_attn));
 
             blocks["img_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
             blocks["img_mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, mlp_hidden_dim));
             // img_mlp.1 is nn.GELU(approximate="tanh")
             blocks["img_mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(mlp_hidden_dim, hidden_size));
 
-            blocks["txt_mod"]   = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
+            if (!prune_mod) {
+                blocks["txt_mod"] = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, true));
+            }
             blocks["txt_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
-            blocks["txt_attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias));
+            blocks["txt_attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qkv_bias, flash_attn));
 
             blocks["txt_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
             blocks["txt_mlp.0"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, mlp_hidden_dim));
@@ -262,17 +279,34 @@ namespace Flux {
             blocks["txt_mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(mlp_hidden_dim, hidden_size));
         }
 
+        std::vector<ModulationOut> get_distil_img_mod(struct ggml_context* ctx, struct ggml_tensor* vec) {
+            // TODO: not hardcoded?
+            const int single_blocks_count = 38;
+            const int double_blocks_count = 19;
+
+            int64_t offset = 6 * idx + 3 * single_blocks_count;
+            return {ModulationOut(ctx, vec, offset), ModulationOut(ctx, vec, offset + 3)};
+        }
+
+        std::vector<ModulationOut> get_distil_txt_mod(struct ggml_context* ctx, struct ggml_tensor* vec) {
+            // TODO: not hardcoded?
+            const int single_blocks_count = 38;
+            const int double_blocks_count = 19;
+
+            int64_t offset = 6 * idx + 6 * double_blocks_count + 3 * single_blocks_count;
+            return {ModulationOut(ctx, vec, offset), ModulationOut(ctx, vec, offset + 3)};
+        }
+
         std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
                                                                     struct ggml_tensor* img,
                                                                     struct ggml_tensor* txt,
                                                                     struct ggml_tensor* vec,
-                                                                    struct ggml_tensor* pe) {
+                                                                    struct ggml_tensor* pe,
+                                                                    struct ggml_tensor* mask = NULL) {
             // img: [N, n_img_token, hidden_size]
             // txt: [N, n_txt_token, hidden_size]
             // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
             // return: ([N, n_img_token, hidden_size], [N, n_txt_token, hidden_size])
-
-            auto img_mod   = std::dynamic_pointer_cast<Modulation>(blocks["img_mod"]);
             auto img_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm1"]);
             auto img_attn  = std::dynamic_pointer_cast<SelfAttention>(blocks["img_attn"]);
 
@@ -280,7 +314,6 @@ namespace Flux {
             auto img_mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.0"]);
             auto img_mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.2"]);
 
-            auto txt_mod   = std::dynamic_pointer_cast<Modulation>(blocks["txt_mod"]);
             auto txt_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm1"]);
             auto txt_attn  = std::dynamic_pointer_cast<SelfAttention>(blocks["txt_attn"]);
 
@@ -288,10 +321,22 @@ namespace Flux {
             auto txt_mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["txt_mlp.0"]);
             auto txt_mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["txt_mlp.2"]);
 
-            auto img_mods          = img_mod->forward(ctx, vec);
+            std::vector<ModulationOut> img_mods;
+            if (prune_mod) {
+                img_mods = get_distil_img_mod(ctx, vec);
+            } else {
+                auto img_mod = std::dynamic_pointer_cast<Modulation>(blocks["img_mod"]);
+                img_mods     = img_mod->forward(ctx, vec);
+            }
             ModulationOut img_mod1 = img_mods[0];
             ModulationOut img_mod2 = img_mods[1];
-            auto txt_mods          = txt_mod->forward(ctx, vec);
+            std::vector<ModulationOut> txt_mods;
+            if (prune_mod) {
+                txt_mods = get_distil_txt_mod(ctx, vec);
+            } else {
+                auto txt_mod = std::dynamic_pointer_cast<Modulation>(blocks["txt_mod"]);
+                txt_mods     = txt_mod->forward(ctx, vec);
+            }
             ModulationOut txt_mod1 = txt_mods[0];
             ModulationOut txt_mod2 = txt_mods[1];
 
@@ -316,7 +361,7 @@ namespace Flux {
             auto k = ggml_concat(ctx, txt_k, img_k, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
             auto v = ggml_concat(ctx, txt_v, img_v, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
 
-            auto attn         = attention(ctx, q, k, v, pe);                          // [N, n_txt_token + n_img_token, n_head*d_head]
+            auto attn         = attention(ctx, q, k, v, pe, mask, flash_attn);        // [N, n_txt_token + n_img_token, n_head*d_head]
             attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));  // [n_txt_token + n_img_token, N, hidden_size]
             auto txt_attn_out = ggml_view_3d(ctx,
                                              attn,
@@ -364,13 +409,19 @@ namespace Flux {
         int64_t num_heads;
         int64_t hidden_size;
         int64_t mlp_hidden_dim;
+        bool flash_attn;
+        bool prune_mod;
+        int idx = 0;
 
     public:
         SingleStreamBlock(int64_t hidden_size,
                           int64_t num_heads,
                           float mlp_ratio = 4.0f,
-                          float qk_scale  = 0.f)
-            : hidden_size(hidden_size), num_heads(num_heads) {
+                          int idx         = 0,
+                          float qk_scale  = 0.f,
+                          bool flash_attn = false,
+                          bool prune_mod  = false)
+            : hidden_size(hidden_size), num_heads(num_heads), idx(idx), flash_attn(flash_attn), prune_mod(prune_mod) {
             int64_t head_dim = hidden_size / num_heads;
             float scale      = qk_scale;
             if (scale <= 0.f) {
@@ -383,26 +434,37 @@ namespace Flux {
             blocks["norm"]     = std::shared_ptr<GGMLBlock>(new QKNorm(head_dim));
             blocks["pre_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-6f, false));
             // mlp_act is nn.GELU(approximate="tanh")
-            blocks["modulation"] = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, false));
+            if (!prune_mod) {
+                blocks["modulation"] = std::shared_ptr<GGMLBlock>(new Modulation(hidden_size, false));
+            }
+        }
+
+        ModulationOut get_distil_mod(struct ggml_context* ctx, struct ggml_tensor* vec) {
+            int64_t offset = 3 * idx;
+            return ModulationOut(ctx, vec, offset);
         }
 
         struct ggml_tensor* forward(struct ggml_context* ctx,
                                     struct ggml_tensor* x,
                                     struct ggml_tensor* vec,
-                                    struct ggml_tensor* pe) {
+                                    struct ggml_tensor* pe,
+                                    struct ggml_tensor* mask = NULL) {
             // x: [N, n_token, hidden_size]
             // pe: [n_token, d_head/2, 2, 2]
             // return: [N, n_token, hidden_size]
 
-            auto linear1    = std::dynamic_pointer_cast<Linear>(blocks["linear1"]);
-            auto linear2    = std::dynamic_pointer_cast<Linear>(blocks["linear2"]);
-            auto norm       = std::dynamic_pointer_cast<QKNorm>(blocks["norm"]);
-            auto pre_norm   = std::dynamic_pointer_cast<LayerNorm>(blocks["pre_norm"]);
-            auto modulation = std::dynamic_pointer_cast<Modulation>(blocks["modulation"]);
-
-            auto mods         = modulation->forward(ctx, vec);
-            ModulationOut mod = mods[0];
+            auto linear1  = std::dynamic_pointer_cast<Linear>(blocks["linear1"]);
+            auto linear2  = std::dynamic_pointer_cast<Linear>(blocks["linear2"]);
+            auto norm     = std::dynamic_pointer_cast<QKNorm>(blocks["norm"]);
+            auto pre_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["pre_norm"]);
+            ModulationOut mod;
+            if (prune_mod) {
+                mod = get_distil_mod(ctx, vec);
+            } else {
+                auto modulation = std::dynamic_pointer_cast<Modulation>(blocks["modulation"]);
 
+                mod = modulation->forward(ctx, vec)[0];
+            }
             auto x_mod   = Flux::modulate(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]
             qkv_mlp      = ggml_cont(ctx, ggml_permute(ctx, qkv_mlp, 2, 0, 1, 3));  // [hidden_size * 3 + mlp_hidden_dim, N, n_token]
@@ -433,7 +495,7 @@ namespace Flux {
             auto v           = ggml_reshape_4d(ctx, qkv_vec[2], head_dim, num_heads, qkv_vec[2]->ne[1], qkv_vec[2]->ne[2]);  // [N, n_token, n_head, d_head]
             q                = norm->query_norm(ctx, q);
             k                = norm->key_norm(ctx, k);
-            auto attn        = attention(ctx, q, k, v, pe);  // [N, n_token, hidden_size]
+            auto attn        = attention(ctx, q, k, v, pe, mask, flash_attn);  // [N, n_token, hidden_size]
 
             auto attn_mlp = ggml_concat(ctx, attn, ggml_gelu_inplace(ctx, mlp), 0);  // [N, n_token, hidden_size + mlp_hidden_dim]
             auto output   = linear2->forward(ctx, attn_mlp);                         // [N, n_token, hidden_size]
@@ -444,13 +506,28 @@ namespace Flux {
     };
 
     struct LastLayer : public GGMLBlock {
+        bool prune_mod;
+
     public:
         LastLayer(int64_t hidden_size,
                   int64_t patch_size,
-                  int64_t out_channels) {
-            blocks["norm_final"]         = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
-            blocks["linear"]             = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, patch_size * patch_size * out_channels));
-            blocks["adaLN_modulation.1"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, 2 * hidden_size));
+                  int64_t out_channels,
+                  bool prune_mod = false)
+            : prune_mod(prune_mod) {
+            blocks["norm_final"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
+            blocks["linear"]     = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, patch_size * patch_size * out_channels));
+            if (!prune_mod) {
+                blocks["adaLN_modulation.1"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, 2 * hidden_size));
+            }
+        }
+
+        ModulationOut get_distil_mod(struct ggml_context* ctx, struct ggml_tensor* vec) {
+            int64_t offset = vec->ne[2] - 2;
+            int64_t stride = vec->nb[1] * vec->ne[1];
+            auto shift     = ggml_view_2d(ctx, vec, vec->ne[0], vec->ne[1], vec->nb[1], stride * (offset + 0));  // [N, dim]
+            auto scale     = ggml_view_2d(ctx, vec, vec->ne[0], vec->ne[1], vec->nb[1], stride * (offset + 1));  // [N, dim]
+            // No gate
+            return ModulationOut(shift, scale, NULL);
         }
 
         struct ggml_tensor* forward(struct ggml_context* ctx,
@@ -459,17 +536,24 @@ namespace Flux {
             // x: [N, n_token, hidden_size]
             // c: [N, hidden_size]
             // return: [N, n_token, patch_size * patch_size * out_channels]
-            auto norm_final         = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_final"]);
-            auto linear             = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
-            auto adaLN_modulation_1 = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.1"]);
+            auto norm_final = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_final"]);
+            auto linear     = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            struct ggml_tensor *shift, *scale;
+            if (prune_mod) {
+                auto mod = get_distil_mod(ctx, c);
+                shift    = mod.shift;
+                scale    = mod.scale;
+            } else {
+                auto adaLN_modulation_1 = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.1"]);
 
-            auto m = adaLN_modulation_1->forward(ctx, ggml_silu(ctx, c));  // [N, 2 * hidden_size]
-            m      = ggml_reshape_3d(ctx, m, c->ne[0], 2, c->ne[1]);       // [N, 2, hidden_size]
-            m      = ggml_cont(ctx, ggml_permute(ctx, m, 0, 2, 1, 3));     // [2, N, hidden_size]
+                auto m = adaLN_modulation_1->forward(ctx, ggml_silu(ctx, c));  // [N, 2 * hidden_size]
+                m      = ggml_reshape_3d(ctx, m, c->ne[0], 2, c->ne[1]);       // [N, 2, hidden_size]
+                m      = ggml_cont(ctx, ggml_permute(ctx, m, 0, 2, 1, 3));     // [2, N, hidden_size]
 
-            int64_t offset = m->nb[1] * m->ne[1];
-            auto shift     = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 0);  // [N, hidden_size]
-            auto scale     = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 1);  // [N, hidden_size]
+                int64_t offset = m->nb[1] * m->ne[1];
+                shift          = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 0);  // [N, hidden_size]
+                scale          = ggml_view_2d(ctx, m, m->ne[0], m->ne[1], m->nb[1], offset * 1);  // [N, hidden_size]
+            }
 
             x = Flux::modulate(ctx, norm_final->forward(ctx, x), shift, scale);
             x = linear->forward(ctx, x);
@@ -478,8 +562,37 @@ namespace Flux {
         }
     };
 
+    struct ChromaApproximator : public GGMLBlock {
+        int64_t inner_size = 5120;
+        int64_t n_layers   = 5;
+        ChromaApproximator(int64_t in_channels = 64, int64_t hidden_size = 3072) {
+            blocks["in_proj"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, inner_size, true));
+            for (int i = 0; i < n_layers; i++) {
+                blocks["norms." + std::to_string(i)]  = std::shared_ptr<GGMLBlock>(new RMSNorm(inner_size));
+                blocks["layers." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(inner_size, inner_size));
+            }
+            blocks["out_proj"] = std::shared_ptr<GGMLBlock>(new Linear(inner_size, hidden_size, true));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+            auto in_proj  = std::dynamic_pointer_cast<Linear>(blocks["in_proj"]);
+            auto out_proj = std::dynamic_pointer_cast<Linear>(blocks["out_proj"]);
+
+            x = in_proj->forward(ctx, x);
+            for (int i = 0; i < n_layers; i++) {
+                auto norm  = std::dynamic_pointer_cast<RMSNorm>(blocks["norms." + std::to_string(i)]);
+                auto embed = std::dynamic_pointer_cast<MLPEmbedder>(blocks["layers." + std::to_string(i)]);
+                x          = ggml_add_inplace(ctx, x, embed->forward(ctx, norm->forward(ctx, x)));
+            }
+            x = out_proj->forward(ctx, x);
+
+            return x;
+        }
+    };
+
     struct FluxParams {
         int64_t in_channels         = 64;
+        int64_t out_channels        = 64;
         int64_t vec_in_dim          = 768;
         int64_t context_in_dim      = 4096;
         int64_t hidden_size         = 3072;
@@ -492,6 +605,8 @@ namespace Flux {
         int theta                   = 10000;
         bool qkv_bias               = true;
         bool guidance_embed         = true;
+        bool flash_attn             = true;
+        bool is_chroma              = false;
     };
 
     struct Flux : public GGMLBlock {
@@ -558,17 +673,22 @@ namespace Flux {
         }
 
         // Generate IDs for image patches and text
-        std::vector<std::vector<float>> gen_ids(int h, int w, int patch_size, int bs, int context_len) {
+        std::vector<std::vector<float>> gen_txt_ids(int bs, int context_len) {
+            return std::vector<std::vector<float>>(bs * context_len, std::vector<float>(3, 0.0));
+        }
+
+        std::vector<std::vector<float>> gen_img_ids(int h, int w, int patch_size, int bs, int index = 0, int h_offset = 0, int w_offset = 0) {
             int h_len = (h + (patch_size / 2)) / patch_size;
             int w_len = (w + (patch_size / 2)) / patch_size;
 
             std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(3, 0.0));
 
-            std::vector<float> row_ids = linspace(0, h_len - 1, h_len);
-            std::vector<float> col_ids = linspace(0, w_len - 1, w_len);
+            std::vector<float> row_ids = linspace(h_offset, h_len - 1 + h_offset, h_len);
+            std::vector<float> col_ids = linspace(w_offset, w_len - 1 + w_offset, w_len);
 
             for (int i = 0; i < h_len; ++i) {
                 for (int j = 0; j < w_len; ++j) {
+                    img_ids[i * w_len + j][0] = index;
                     img_ids[i * w_len + j][1] = row_ids[i];
                     img_ids[i * w_len + j][2] = col_ids[j];
                 }
@@ -580,24 +700,54 @@ namespace Flux {
                     img_ids_repeated[i * img_ids.size() + j] = img_ids[j];
                 }
             }
+            return img_ids_repeated;
+        }
 
-            std::vector<std::vector<float>> txt_ids(bs * context_len, std::vector<float>(3, 0.0));
-            std::vector<std::vector<float>> ids(bs * (context_len + img_ids.size()), std::vector<float>(3));
+        std::vector<std::vector<float>> concat_ids(const std::vector<std::vector<float>>& a,
+                                                   const std::vector<std::vector<float>>& b,
+                                                   int bs) {
+            size_t a_len = a.size() / bs;
+            size_t b_len = b.size() / bs;
+            std::vector<std::vector<float>> ids(a.size() + b.size(), std::vector<float>(3));
             for (int i = 0; i < bs; ++i) {
-                for (int j = 0; j < context_len; ++j) {
-                    ids[i * (context_len + img_ids.size()) + j] = txt_ids[j];
+                for (int j = 0; j < a_len; ++j) {
+                    ids[i * (a_len + b_len) + j] = a[i * a_len + j];
                 }
-                for (int j = 0; j < img_ids.size(); ++j) {
-                    ids[i * (context_len + img_ids.size()) + context_len + j] = img_ids_repeated[i * img_ids.size() + j];
+                for (int j = 0; j < b_len; ++j) {
+                    ids[i * (a_len + b_len) + a_len + j] = b[i * b_len + j];
                 }
             }
+            return ids;
+        }
 
+        std::vector<std::vector<float>> gen_ids(int h, int w, int patch_size, int bs, int context_len, std::vector<ggml_tensor*> ref_latents) {
+            auto txt_ids = gen_txt_ids(bs, context_len);
+            auto img_ids = gen_img_ids(h, w, patch_size, bs);
+
+            auto ids               = concat_ids(txt_ids, img_ids, bs);
+            uint64_t curr_h_offset = 0;
+            uint64_t curr_w_offset = 0;
+            for (ggml_tensor* ref : ref_latents) {
+                uint64_t h_offset = 0;
+                uint64_t w_offset = 0;
+                if (ref->ne[1] + curr_h_offset > ref->ne[0] + curr_w_offset) {
+                    w_offset = curr_w_offset;
+                } else {
+                    h_offset = curr_h_offset;
+                }
+
+                auto ref_ids = gen_img_ids(ref->ne[1], ref->ne[0], patch_size, bs, 1, h_offset, w_offset);
+                ids          = concat_ids(ids, ref_ids, bs);
+
+                curr_h_offset = std::max(curr_h_offset, ref->ne[1] + h_offset);
+                curr_w_offset = std::max(curr_w_offset, ref->ne[0] + w_offset);
+            }
             return ids;
         }
 
         // Generate positional embeddings
-        std::vector<float> gen_pe(int h, int w, int patch_size, int bs, int context_len, int theta, const std::vector<int>& axes_dim) {
-            std::vector<std::vector<float>> ids       = gen_ids(h, w, patch_size, bs, context_len);
+        std::vector<float> gen_pe(int h, int w, int patch_size, int bs, int context_len, std::vector<ggml_tensor*> ref_latents, int theta, const std::vector<int>& axes_dim) {
+            std::vector<std::vector<float>> ids       = gen_ids(h, w, patch_size, bs, context_len, ref_latents);
             std::vector<std::vector<float>> trans_ids = transpose(ids);
             size_t pos_len                            = ids.size();
             int num_axes                              = axes_dim.size();
@@ -631,14 +781,17 @@ namespace Flux {
         Flux() {}
         Flux(FluxParams params)
             : params(params) {
-            int64_t out_channels = params.in_channels;
-            int64_t pe_dim       = params.hidden_size / params.num_heads;
+            int64_t pe_dim = params.hidden_size / params.num_heads;
 
-            blocks["img_in"]    = std::shared_ptr<GGMLBlock>(new Linear(params.in_channels, params.hidden_size, true));
-            blocks["time_in"]   = std::shared_ptr<GGMLBlock>(new MLPEmbedder(256, params.hidden_size));
-            blocks["vector_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(params.vec_in_dim, params.hidden_size));
-            if (params.guidance_embed) {
-                blocks["guidance_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(256, params.hidden_size));
+            blocks["img_in"] = std::shared_ptr<GGMLBlock>(new Linear(params.in_channels, params.hidden_size, true));
+            if (params.is_chroma) {
+                blocks["distilled_guidance_layer"] = std::shared_ptr<GGMLBlock>(new ChromaApproximator(params.in_channels, params.hidden_size));
+            } else {
+                blocks["time_in"]   = std::shared_ptr<GGMLBlock>(new MLPEmbedder(256, params.hidden_size));
+                blocks["vector_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(params.vec_in_dim, params.hidden_size));
+                if (params.guidance_embed) {
+                    blocks["guidance_in"] = std::shared_ptr<GGMLBlock>(new MLPEmbedder(256, params.hidden_size));
+                }
             }
             blocks["txt_in"] = std::shared_ptr<GGMLBlock>(new Linear(params.context_in_dim, params.hidden_size, true));
 
@@ -646,16 +799,23 @@ namespace Flux {
                 blocks["double_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new DoubleStreamBlock(params.hidden_size,
                                                                                                                 params.num_heads,
                                                                                                                 params.mlp_ratio,
-                                                                                                                params.qkv_bias));
+                                                                                                                i,
+                                                                                                                params.qkv_bias,
+                                                                                                                params.flash_attn,
+                                                                                                                params.is_chroma));
             }
 
             for (int i = 0; i < params.depth_single_blocks; i++) {
                 blocks["single_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new SingleStreamBlock(params.hidden_size,
                                                                                                                 params.num_heads,
-                                                                                                                params.mlp_ratio));
+                                                                                                                params.mlp_ratio,
+                                                                                                                i,
+                                                                                                                0.f,
+                                                                                                                params.flash_attn,
+                                                                                                                params.is_chroma));
             }
 
-            blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new LastLayer(params.hidden_size, 1, out_channels));
+            blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new LastLayer(params.hidden_size, 1, params.out_channels, params.is_chroma));
         }
 
         struct ggml_tensor* patchify(struct ggml_context* ctx,
@@ -712,25 +872,55 @@ namespace Flux {
                                          struct ggml_tensor* y,
                                          struct ggml_tensor* guidance,
                                          struct ggml_tensor* pe,
-                                         std::vector<int> skip_layers = std::vector<int>()) {
+                                         struct ggml_tensor* mod_index_arange = NULL,
+                                         std::vector<int> skip_layers         = {}) {
             auto img_in      = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
-            auto time_in     = std::dynamic_pointer_cast<MLPEmbedder>(blocks["time_in"]);
-            auto vector_in   = std::dynamic_pointer_cast<MLPEmbedder>(blocks["vector_in"]);
             auto txt_in      = std::dynamic_pointer_cast<Linear>(blocks["txt_in"]);
             auto final_layer = std::dynamic_pointer_cast<LastLayer>(blocks["final_layer"]);
 
-            img      = img_in->forward(ctx, img);
-            auto vec = time_in->forward(ctx, ggml_nn_timestep_embedding(ctx, timesteps, 256, 10000, 1000.f));
+            img = img_in->forward(ctx, img);
+            struct ggml_tensor* vec;
+            struct ggml_tensor* txt_img_mask = NULL;
+            if (params.is_chroma) {
+                int64_t mod_index_length = 344;
+                auto approx              = std::dynamic_pointer_cast<ChromaApproximator>(blocks["distilled_guidance_layer"]);
+                auto distill_timestep    = ggml_nn_timestep_embedding(ctx, timesteps, 16, 10000, 1000.f);
+                auto distill_guidance    = ggml_nn_timestep_embedding(ctx, guidance, 16, 10000, 1000.f);
 
-            if (params.guidance_embed) {
-                GGML_ASSERT(guidance != NULL);
-                auto guidance_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["guidance_in"]);
-                // bf16 and fp16 result is different
-                auto g_in = ggml_nn_timestep_embedding(ctx, guidance, 256, 10000, 1000.f);
-                vec       = ggml_add(ctx, vec, guidance_in->forward(ctx, g_in));
+                // auto mod_index_arange  = ggml_arange(ctx, 0, (float)mod_index_length, 1);
+                // ggml_arange tot working on a lot of backends, precomputing it on CPU instead
+                GGML_ASSERT(arange != NULL);
+                auto modulation_index = ggml_nn_timestep_embedding(ctx, mod_index_arange, 32, 10000, 1000.f);  // [1, 344, 32]
+
+                // Batch broadcast (will it ever be useful)
+                modulation_index = ggml_repeat(ctx, modulation_index, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, modulation_index->ne[0], modulation_index->ne[1], img->ne[2]));  // [N, 344, 32]
+
+                auto timestep_guidance = ggml_concat(ctx, distill_timestep, distill_guidance, 0);  // [N, 1, 32]
+                timestep_guidance      = ggml_repeat(ctx, timestep_guidance, modulation_index);    // [N, 344, 32]
+
+                vec = ggml_concat(ctx, timestep_guidance, modulation_index, 0);  // [N, 344, 64]
+                // Permute for consistency with non-distilled modulation implementation
+                vec = ggml_cont(ctx, ggml_permute(ctx, vec, 0, 2, 1, 3));  // [344, N, 64]
+                vec = approx->forward(ctx, vec);                           // [344, N, hidden_size]
+
+                if (y != NULL) {
+                    txt_img_mask = ggml_pad(ctx, y, img->ne[1], 0, 0, 0);
+                }
+            } else {
+                auto time_in   = std::dynamic_pointer_cast<MLPEmbedder>(blocks["time_in"]);
+                auto vector_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["vector_in"]);
+                vec            = time_in->forward(ctx, ggml_nn_timestep_embedding(ctx, timesteps, 256, 10000, 1000.f));
+                if (params.guidance_embed) {
+                    GGML_ASSERT(guidance != NULL);
+                    auto guidance_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["guidance_in"]);
+                    // bf16 and fp16 result is different
+                    auto g_in = ggml_nn_timestep_embedding(ctx, guidance, 256, 10000, 1000.f);
+                    vec       = ggml_add(ctx, vec, guidance_in->forward(ctx, g_in));
+                }
+
+                vec = ggml_add(ctx, vec, vector_in->forward(ctx, y));
             }
 
-            vec = ggml_add(ctx, vec, vector_in->forward(ctx, y));
             txt = txt_in->forward(ctx, txt);
 
             for (int i = 0; i < params.depth; i++) {
@@ -740,7 +930,7 @@ namespace Flux {
 
                 auto block = std::dynamic_pointer_cast<DoubleStreamBlock>(blocks["double_blocks." + std::to_string(i)]);
 
-                auto img_txt = block->forward(ctx, img, txt, vec, pe);
+                auto img_txt = block->forward(ctx, img, txt, vec, pe, txt_img_mask);
                 img          = img_txt.first;   // [N, n_img_token, hidden_size]
                 txt          = img_txt.second;  // [N, n_txt_token, hidden_size]
             }
@@ -752,7 +942,7 @@ namespace Flux {
                 }
                 auto block = std::dynamic_pointer_cast<SingleStreamBlock>(blocks["single_blocks." + std::to_string(i)]);
 
-                txt_img = block->forward(ctx, txt_img, vec, pe);
+                txt_img = block->forward(ctx, txt_img, vec, pe, txt_img_mask);
             }
 
             txt_img = ggml_cont(ctx, ggml_permute(ctx, txt_img, 0, 2, 1, 3));  // [n_txt_token + n_img_token, N, hidden_size]
@@ -767,29 +957,11 @@ namespace Flux {
             img     = ggml_cont(ctx, ggml_permute(ctx, img, 0, 2, 1, 3));  // [N, n_img_token, hidden_size]
 
             img = final_layer->forward(ctx, img, vec);  // (N, T, patch_size ** 2 * out_channels)
-
             return img;
         }
 
-        struct ggml_tensor* forward(struct ggml_context* ctx,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* timestep,
-                                    struct ggml_tensor* context,
-                                    struct ggml_tensor* y,
-                                    struct ggml_tensor* guidance,
-                                    struct ggml_tensor* pe,
-                                    std::vector<int> skip_layers = std::vector<int>()) {
-            // Forward pass of DiT.
-            // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
-            // timestep: (N,) tensor of diffusion timesteps
-            // context: (N, L, D)
-            // y: (N, adm_in_channels) tensor of class labels
-            // guidance: (N,)
-            // pe: (L, d_head/2, 2, 2)
-            // return: (N, C, H, W)
-
-            GGML_ASSERT(x->ne[3] == 1);
-
+        struct ggml_tensor* process_img(struct ggml_context* ctx,
+                                        struct ggml_tensor* x) {
             int64_t W          = x->ne[0];
             int64_t H          = x->ne[1];
             int64_t patch_size = 2;
@@ -799,8 +971,65 @@ namespace Flux {
 
             // img = rearrange(x, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=patch_size, pw=patch_size)
             auto img = patchify(ctx, x, patch_size);  // [N, h*w, C * patch_size * patch_size]
+            return img;
+        }
 
-            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, skip_layers);  // [N, h*w, C * patch_size * patch_size]
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* timestep,
+                                    struct ggml_tensor* context,
+                                    struct ggml_tensor* c_concat,
+                                    struct ggml_tensor* y,
+                                    struct ggml_tensor* guidance,
+                                    struct ggml_tensor* pe,
+                                    struct ggml_tensor* mod_index_arange  = NULL,
+                                    std::vector<ggml_tensor*> ref_latents = {},
+                                    std::vector<int> skip_layers          = {}) {
+            // Forward pass of DiT.
+            // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
+            // timestep: (N,) tensor of diffusion timesteps
+            // context: (N, L, D)
+            // c_concat: NULL, or for (N,C+M, H, W) for Fill
+            // y: (N, adm_in_channels) tensor of class labels
+            // guidance: (N,)
+            // pe: (L, d_head/2, 2, 2)
+            // return: (N, C, H, W)
+
+            GGML_ASSERT(x->ne[3] == 1);
+
+            int64_t W          = x->ne[0];
+            int64_t H          = x->ne[1];
+            int64_t C          = x->ne[2];
+            int64_t patch_size = 2;
+            int pad_h          = (patch_size - H % patch_size) % patch_size;
+            int pad_w          = (patch_size - W % patch_size) % patch_size;
+
+            auto img            = process_img(ctx, x);
+            uint64_t img_tokens = img->ne[1];
+
+            if (c_concat != NULL) {
+                ggml_tensor* masked = ggml_view_4d(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, c_concat, c_concat->ne[0], c_concat->ne[1], 8 * 8, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C);
+
+                masked = process_img(ctx, masked);
+                mask   = process_img(ctx, mask);
+
+                img = ggml_concat(ctx, img, ggml_concat(ctx, masked, mask, 0), 0);
+            }
+
+            if (ref_latents.size() > 0) {
+                for (ggml_tensor* ref : ref_latents) {
+                    ref = process_img(ctx, ref);
+                    img = ggml_concat(ctx, img, ref, 1);
+                }
+            }
+
+            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, mod_index_arange, skip_layers);  // [N, num_tokens, C * patch_size * patch_size]
+            if (out->ne[1] > img_tokens) {
+                out = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));  // [num_tokens, N, C * patch_size * patch_size]
+                out = ggml_view_3d(ctx, out, out->ne[0], out->ne[1], img_tokens, out->nb[1], out->nb[2], 0);
+                out = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));  // [N, h*w, C * patch_size * patch_size]
+            }
 
             // 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, out, (H + pad_h) / patch_size, (W + pad_w) / patch_size, patch_size);  // [N, C, H + pad_h, W + pad_w]
@@ -810,23 +1039,69 @@ namespace Flux {
     };
 
     struct FluxRunner : public GGMLRunner {
+        static std::map<std::string, enum ggml_type> empty_tensor_types;
+
     public:
         FluxParams flux_params;
         Flux flux;
-        std::vector<float> pe_vec;  // for cache
+        std::vector<float> pe_vec;
+        std::vector<float> mod_index_arange_vec;  // for cache
+        SDVersion version;
+        bool use_mask = false;
 
         FluxRunner(ggml_backend_t backend,
-                   ggml_type wtype,
-                   SDVersion version = VERSION_FLUX_DEV)
-            : GGMLRunner(backend, wtype) {
-            if (version == VERSION_FLUX_SCHNELL) {
-                flux_params.guidance_embed = false;
+                   std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
+                   const std::string prefix                            = "",
+                   SDVersion version                                   = VERSION_FLUX,
+                   bool flash_attn                                     = false,
+                   bool use_mask                                       = false)
+            : GGMLRunner(backend), use_mask(use_mask) {
+            flux_params.flash_attn          = flash_attn;
+            flux_params.guidance_embed      = false;
+            flux_params.depth               = 0;
+            flux_params.depth_single_blocks = 0;
+            if (version == VERSION_FLUX_FILL) {
+                flux_params.in_channels = 384;
             }
-            if (version == VERSION_FLUX_LITE) {
-                flux_params.depth = 8;
+            for (auto pair : tensor_types) {
+                std::string tensor_name = pair.first;
+                if (tensor_name.find("model.diffusion_model.") == std::string::npos)
+                    continue;
+                if (tensor_name.find("guidance_in.in_layer.weight") != std::string::npos) {
+                    // not schnell
+                    flux_params.guidance_embed = true;
+                }
+                if (tensor_name.find("distilled_guidance_layer.in_proj.weight") != std::string::npos) {
+                    // Chroma
+                    flux_params.is_chroma = true;
+                }
+                size_t db = tensor_name.find("double_blocks.");
+                if (db != std::string::npos) {
+                    tensor_name     = tensor_name.substr(db);  // remove prefix
+                    int block_depth = atoi(tensor_name.substr(14, tensor_name.find(".", 14)).c_str());
+                    if (block_depth + 1 > flux_params.depth) {
+                        flux_params.depth = block_depth + 1;
+                    }
+                }
+                size_t sb = tensor_name.find("single_blocks.");
+                if (sb != std::string::npos) {
+                    tensor_name     = tensor_name.substr(sb);  // remove prefix
+                    int block_depth = atoi(tensor_name.substr(14, tensor_name.find(".", 14)).c_str());
+                    if (block_depth + 1 > flux_params.depth_single_blocks) {
+                        flux_params.depth_single_blocks = block_depth + 1;
+                    }
+                }
             }
+
+            LOG_INFO("Flux blocks: %d double, %d single", flux_params.depth, flux_params.depth_single_blocks);
+            if (flux_params.is_chroma) {
+                LOG_INFO("Using pruned modulation (Chroma)");
+            } else if (!flux_params.guidance_embed) {
+                LOG_INFO("Flux guidance is disabled (Schnell mode)");
+            }
+
             flux = Flux(flux_params);
-            flux.init(params_ctx, wtype);
+            flux.init(params_ctx, tensor_types, prefix);
         }
 
         std::string get_desc() {
@@ -840,21 +1115,44 @@ namespace Flux {
         struct ggml_cgraph* build_graph(struct ggml_tensor* x,
                                         struct ggml_tensor* timesteps,
                                         struct ggml_tensor* context,
+                                        struct ggml_tensor* c_concat,
                                         struct ggml_tensor* y,
                                         struct ggml_tensor* guidance,
-                                        std::vector<int> skip_layers = std::vector<int>()) {
+                                        std::vector<ggml_tensor*> ref_latents = {},
+                                        std::vector<int> skip_layers          = {}) {
             GGML_ASSERT(x->ne[3] == 1);
             struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, FLUX_GRAPH_SIZE, false);
 
-            x         = to_backend(x);
-            context   = to_backend(context);
-            y         = to_backend(y);
+            struct ggml_tensor* mod_index_arange = NULL;
+
+            x       = to_backend(x);
+            context = to_backend(context);
+            if (c_concat != NULL) {
+                c_concat = to_backend(c_concat);
+            }
+            if (flux_params.is_chroma) {
+                guidance = ggml_set_f32(guidance, 0);
+
+                if (!use_mask) {
+                    y = NULL;
+                }
+
+                // ggml_arange is not working on some backends, precompute it
+                mod_index_arange_vec = arange(0, 344);
+                mod_index_arange     = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_F32, mod_index_arange_vec.size());
+                set_backend_tensor_data(mod_index_arange, mod_index_arange_vec.data());
+            }
+            y = to_backend(y);
+
             timesteps = to_backend(timesteps);
-            if (flux_params.guidance_embed) {
+            if (flux_params.guidance_embed || flux_params.is_chroma) {
                 guidance = to_backend(guidance);
             }
+            for (int i = 0; i < ref_latents.size(); i++) {
+                ref_latents[i] = to_backend(ref_latents[i]);
+            }
 
-            pe_vec      = flux.gen_pe(x->ne[1], x->ne[0], 2, x->ne[3], context->ne[1], flux_params.theta, flux_params.axes_dim);
+            pe_vec      = flux.gen_pe(x->ne[1], x->ne[0], 2, x->ne[3], context->ne[1], ref_latents, flux_params.theta, flux_params.axes_dim);
             int pos_len = pe_vec.size() / flux_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, flux_params.axes_dim_sum / 2, pos_len);
@@ -867,9 +1165,12 @@ namespace Flux {
                                                    x,
                                                    timesteps,
                                                    context,
+                                                   c_concat,
                                                    y,
                                                    guidance,
                                                    pe,
+                                                   mod_index_arange,
+                                                   ref_latents,
                                                    skip_layers);
 
             ggml_build_forward_expand(gf, out);
@@ -881,18 +1182,20 @@ namespace Flux {
                      struct ggml_tensor* x,
                      struct ggml_tensor* timesteps,
                      struct ggml_tensor* context,
+                     struct ggml_tensor* c_concat,
                      struct ggml_tensor* y,
                      struct ggml_tensor* guidance,
-                     struct ggml_tensor** output     = NULL,
-                     struct ggml_context* output_ctx = NULL,
-                     std::vector<int> skip_layers    = std::vector<int>()) {
+                     std::vector<ggml_tensor*> ref_latents = {},
+                     struct ggml_tensor** output           = NULL,
+                     struct ggml_context* output_ctx       = NULL,
+                     std::vector<int> skip_layers          = std::vector<int>()) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             // y: [N, adm_in_channels] or [1, adm_in_channels]
             // guidance: [N, ]
             auto get_graph = [&]() -> struct ggml_cgraph* {
-                return build_graph(x, timesteps, context, y, guidance, skip_layers);
+                return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, skip_layers);
             };
 
             GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
@@ -932,7 +1235,7 @@ namespace Flux {
                 struct ggml_tensor* out = NULL;
 
                 int t0 = ggml_time_ms();
-                compute(8, x, timesteps, context, y, guidance, &out, work_ctx);
+                compute(8, x, timesteps, context, NULL, y, guidance, {}, &out, work_ctx);
                 int t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);
@@ -944,7 +1247,7 @@ namespace Flux {
             // ggml_backend_t backend    = ggml_backend_cuda_init(0);
             ggml_backend_t backend           = ggml_backend_cpu_init();
             ggml_type model_data_type        = GGML_TYPE_Q8_0;
-            std::shared_ptr<FluxRunner> flux = std::shared_ptr<FluxRunner>(new FluxRunner(backend, model_data_type));
+            std::shared_ptr<FluxRunner> flux = std::shared_ptr<FluxRunner>(new FluxRunner(backend));
             {
                 LOG_INFO("loading from '%s'", file_path.c_str());
 
@@ -973,4 +1276,4 @@ namespace Flux {
 
 }  // namespace Flux
 
-#endif  // __FLUX_HPP__
+#endif  // __FLUX_HPP__

ggml_extend.hpp

@@ -22,9 +22,12 @@
 
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
+#include "ggml-cpu.h"
 #include "ggml.h"
 
-#ifdef SD_USE_CUBLAS
+#include "model.h"
+
+#ifdef SD_USE_CUDA
 #include "ggml-cuda.h"
 #endif
 
@@ -36,6 +39,10 @@
 #include "ggml-vulkan.h"
 #endif
 
+#ifdef SD_USE_OPENCL
+#include "ggml-opencl.h"
+#endif
+
 #ifdef SD_USE_SYCL
 #include "ggml-sycl.h"
 #endif
@@ -49,6 +56,72 @@
 #define __STATIC_INLINE__ static inline
 #endif
 
+// n-mode trensor-matrix product
+// example: 2-mode product
+// A: [ne03, k, ne01, ne00]
+// B: k rows, m columns => [k, m]
+// result is [ne03, m, ne01, ne00]
+__STATIC_INLINE__ struct ggml_tensor* ggml_mul_n_mode(struct ggml_context* ctx, struct ggml_tensor* a, struct ggml_tensor* b, int mode = 0) {
+    // reshape A
+    // swap 0th and nth axis
+    a       = ggml_cont(ctx, ggml_permute(ctx, a, mode, mode != 1 ? 1 : 0, mode != 2 ? 2 : 0, mode != 3 ? 3 : 0));
+    int ne1 = a->ne[1];
+    int ne2 = a->ne[2];
+    int ne3 = a->ne[3];
+    // make 2D
+    a = ggml_cont(ctx, ggml_reshape_2d(ctx, a, a->ne[0], (ne3 * ne2 * ne1)));
+
+    struct ggml_tensor* result = ggml_cont(ctx, ggml_transpose(ctx, ggml_mul_mat(ctx, a, b)));
+
+    // reshape output (same shape as a after permutation except first dim)
+    result = ggml_reshape_4d(ctx, result, result->ne[0], ne1, ne2, ne3);
+    // swap back 0th and nth axis
+    result = ggml_permute(ctx, result, mode, mode != 1 ? 1 : 0, mode != 2 ? 2 : 0, mode != 3 ? 3 : 0);
+    return result;
+}
+
+__STATIC_INLINE__ struct ggml_tensor* ggml_merge_lora(ggml_context* ctx, struct ggml_tensor* lora_down, struct ggml_tensor* lora_up, struct ggml_tensor* lora_mid = NULL) {
+    struct ggml_tensor* updown;
+    // flat lora tensors to multiply it
+    int64_t lora_up_rows  = lora_up->ne[ggml_n_dims(lora_up) - 1];
+    lora_up               = ggml_reshape_2d(ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
+    auto lora_down_n_dims = ggml_n_dims(lora_down);
+    // assume n_dims should always be a multiple of 2 (otherwise rank 1 doesn't work)
+    lora_down_n_dims       = (lora_down_n_dims + lora_down_n_dims % 2);
+    int64_t lora_down_rows = lora_down->ne[lora_down_n_dims - 1];
+    lora_down              = ggml_reshape_2d(ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
+
+    // ggml_mul_mat requires tensor b transposed
+    lora_down = ggml_cont(ctx, ggml_transpose(ctx, lora_down));
+    if (lora_mid == NULL) {
+        updown = ggml_mul_mat(ctx, lora_up, lora_down);
+        updown = ggml_cont(ctx, ggml_transpose(ctx, updown));
+    } else {
+        // undoing tucker decomposition for conv layers.
+        // lora_mid  has shape (3,    3,   Rank, Rank)
+        // lora_down has shape (Rank, In,  1,    1)
+        // lora_up   has shape (Rank, Out, 1,    1)
+        // conv layer shape is (3,    3,   Out,  In)
+        updown = ggml_mul_n_mode(ctx, ggml_mul_n_mode(ctx, lora_mid, lora_down, 3), lora_up, 2);
+        updown = ggml_cont(ctx, updown);
+    }
+    return updown;
+}
+
+// Kronecker product
+// [ne03,ne02,ne01,ne00] x [ne13,ne12,ne11,ne10] => [ne03*ne13,ne02*ne12,ne01*ne11,ne00*ne10]
+__STATIC_INLINE__ struct ggml_tensor* ggml_kronecker(ggml_context* ctx, struct ggml_tensor* a, struct ggml_tensor* b) {
+    return ggml_mul(ctx,
+                    ggml_upscale_ext(ctx,
+                                     a,
+                                     a->ne[0] * b->ne[0],
+                                     a->ne[1] * b->ne[1],
+                                     a->ne[2] * b->ne[2],
+                                     a->ne[3] * b->ne[3],
+                                     GGML_SCALE_MODE_NEAREST), 
+                    b);
+}
+
 __STATIC_INLINE__ void ggml_log_callback_default(ggml_log_level level, const char* text, void* user_data) {
     (void)level;
     (void)user_data;
@@ -100,17 +173,11 @@ __STATIC_INLINE__ ggml_fp16_t ggml_tensor_get_f16(const ggml_tensor* tensor, int
 
 static struct ggml_tensor* get_tensor_from_graph(struct ggml_cgraph* gf, const char* name) {
     struct ggml_tensor* res = NULL;
-    for (int i = 0; i < gf->n_nodes; i++) {
-        // printf("%d, %s \n", i, gf->nodes[i]->name);
-        if (strcmp(ggml_get_name(gf->nodes[i]), name) == 0) {
-            res = gf->nodes[i];
-            break;
-        }
-    }
-    for (int i = 0; i < gf->n_leafs; i++) {
-        // printf("%d, %s \n", i, gf->leafs[i]->name);
-        if (strcmp(ggml_get_name(gf->leafs[i]), name) == 0) {
-            res = gf->leafs[i];
+    for (int i = 0; i < ggml_graph_n_nodes(gf); i++) {
+        struct ggml_tensor* node = ggml_graph_node(gf, i);
+        // printf("%d, %s \n", i, ggml_get_name(node));
+        if (strcmp(ggml_get_name(node), name) == 0) {
+            res = node;
             break;
         }
     }
@@ -293,6 +360,44 @@ __STATIC_INLINE__ void sd_image_to_tensor(const uint8_t* image_data,
     }
 }
 
+__STATIC_INLINE__ void sd_mask_to_tensor(const uint8_t* image_data,
+                                         struct ggml_tensor* output,
+                                         bool scale = true) {
+    int64_t width    = output->ne[0];
+    int64_t height   = output->ne[1];
+    int64_t channels = output->ne[2];
+    GGML_ASSERT(channels == 1 && output->type == GGML_TYPE_F32);
+    for (int iy = 0; iy < height; iy++) {
+        for (int ix = 0; ix < width; ix++) {
+            float value = *(image_data + iy * width * channels + ix);
+            if (scale) {
+                value /= 255.f;
+            }
+            ggml_tensor_set_f32(output, value, ix, iy);
+        }
+    }
+}
+
+__STATIC_INLINE__ void sd_apply_mask(struct ggml_tensor* image_data,
+                                     struct ggml_tensor* mask,
+                                     struct ggml_tensor* output) {
+    int64_t width    = output->ne[0];
+    int64_t height   = output->ne[1];
+    int64_t channels = output->ne[2];
+    GGML_ASSERT(output->type == GGML_TYPE_F32);
+    for (int ix = 0; ix < width; ix++) {
+        for (int iy = 0; iy < height; iy++) {
+            float m = ggml_tensor_get_f32(mask, ix, iy);
+            m       = round(m);  // inpaint models need binary masks
+            ggml_tensor_set_f32(mask, m, ix, iy);
+            for (int k = 0; k < channels; k++) {
+                float value = (1 - m) * (ggml_tensor_get_f32(image_data, ix, iy, k) - .5) + .5;
+                ggml_tensor_set_f32(output, value, ix, iy, k);
+            }
+        }
+    }
+}
+
 __STATIC_INLINE__ void sd_mul_images_to_tensor(const uint8_t* image_data,
                                                struct ggml_tensor* output,
                                                int idx,
@@ -675,18 +780,16 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention(struct ggml_context* ctx
                                                         struct ggml_tensor* k,
                                                         struct ggml_tensor* v,
                                                         bool mask = false) {
-#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUBLAS) && !defined(SD_USE_METAL) && !defined(SD_USE_VULKAN) && !defined(SD_USE_SYCL)
+#if defined(SD_USE_FLASH_ATTENTION) && !defined(SD_USE_CUDA) && !defined(SD_USE_METAL) && !defined(SD_USE_VULKAN) && !defined(SD_USE_SYCL)
     struct ggml_tensor* kqv = ggml_flash_attn(ctx, q, k, v, false);  // [N * n_head, n_token, d_head]
 #else
-    float d_head = (float)q->ne[0];
-
+    float d_head           = (float)q->ne[0];
     struct ggml_tensor* kq = ggml_mul_mat(ctx, k, q);  // [N * n_head, n_token, n_k]
     kq                     = ggml_scale_inplace(ctx, kq, 1.0f / sqrt(d_head));
     if (mask) {
         kq = ggml_diag_mask_inf_inplace(ctx, kq, 0);
     }
-    kq = ggml_soft_max_inplace(ctx, kq);
-
+    kq                      = ggml_soft_max_inplace(ctx, kq);
     struct ggml_tensor* kqv = ggml_mul_mat(ctx, v, kq);  // [N * n_head, n_token, d_head]
 #endif
     return kqv;
@@ -703,7 +806,8 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
                                                             int64_t n_head,
                                                             struct ggml_tensor* mask = NULL,
                                                             bool diag_mask_inf       = false,
-                                                            bool skip_reshape        = false) {
+                                                            bool skip_reshape        = false,
+                                                            bool flash_attn          = false) {
     int64_t L_q;
     int64_t L_k;
     int64_t C;
@@ -734,13 +838,54 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
 
     float scale = (1.0f / sqrt((float)d_head));
 
-    bool use_flash_attn = false;
-    ggml_tensor* kqv    = NULL;
-    if (use_flash_attn) {
+    // if (flash_attn) {
+    //     LOG_DEBUG("attention_ext L_q:%d L_k:%d n_head:%d C:%d d_head:%d N:%d", L_q, L_k, n_head, C, d_head, N);
+    // }
+    //  is there anything oddly shaped?? ping Green-Sky if you can trip this assert
+    GGML_ASSERT(((L_k % 256 == 0) && L_q == L_k) || !(L_k % 256 == 0));
+
+    bool can_use_flash_attn = true;
+    can_use_flash_attn      = can_use_flash_attn && L_k % 256 == 0;
+    can_use_flash_attn      = can_use_flash_attn && d_head % 64 == 0;  // double check
+
+    // cuda max d_head seems to be 256, cpu does seem to work with 512
+    can_use_flash_attn = can_use_flash_attn && d_head <= 256;  // double check
+
+    if (mask != nullptr) {
+        // TODO(Green-Sky): figure out if we can bend t5 to work too
+        can_use_flash_attn = can_use_flash_attn && mask->ne[2] == 1;
+        can_use_flash_attn = can_use_flash_attn && mask->ne[3] == 1;
+    }
+
+    // TODO(Green-Sky): more pad or disable for funny tensor shapes
+
+    ggml_tensor* kqv = nullptr;
+    // GGML_ASSERT((flash_attn && can_use_flash_attn) || !flash_attn);
+    if (can_use_flash_attn && flash_attn) {
+        // LOG_DEBUG("using flash attention");
+        k = ggml_cast(ctx, k, GGML_TYPE_F16);
+
         v = ggml_cont(ctx, ggml_permute(ctx, v, 0, 2, 1, 3));  // [N, n_head, L_k, d_head]
         v = ggml_reshape_3d(ctx, v, d_head, L_k, n_head * N);  // [N * n_head, L_k, d_head]
-        LOG_DEBUG("k->ne[1] == %d", k->ne[1]);
+        v = ggml_cast(ctx, v, GGML_TYPE_F16);
+
+        if (mask != nullptr) {
+            mask = ggml_transpose(ctx, mask);
+
+            if (mask->ne[1] < GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD)) {
+                LOG_DEBUG("mask dims %ld, %ld, %ld, %ld\n", mask->ne[0], mask->ne[1], mask->ne[2], mask->ne[3]);
+                LOG_DEBUG("needs padding, padding from %ld to %ld\n", mask->ne[1], GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD));
+                mask = ggml_pad(ctx, mask, 0, GGML_PAD(q->ne[1], GGML_KQ_MASK_PAD) - mask->ne[1], 0, 0);
+            }
+
+            mask = ggml_cast(ctx, mask, GGML_TYPE_F16);
+        }
+
         kqv = ggml_flash_attn_ext(ctx, q, k, v, mask, scale, 0, 0);
+        ggml_flash_attn_ext_set_prec(kqv, GGML_PREC_F32);
+
+        // kqv = ggml_view_3d(ctx, kqv, d_head, n_head, L_k, kqv->nb[1], kqv->nb[2], 0);
+        kqv = ggml_view_3d(ctx, kqv, d_head, n_head, L_q, kqv->nb[1], kqv->nb[2], 0);
     } else {
         v = ggml_cont(ctx, ggml_permute(ctx, v, 1, 2, 0, 3));  // [N, n_head, d_head, L_k]
         v = ggml_reshape_3d(ctx, v, L_k, d_head, n_head * N);  // [N * n_head, d_head, L_k]
@@ -748,7 +893,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
         auto kq = ggml_mul_mat(ctx, k, q);  // [N * n_head, L_q, L_k]
         kq      = ggml_scale_inplace(ctx, kq, scale);
         if (mask) {
-            kq = ggml_add(ctx, kq, mask);
+            kq = ggml_add_inplace(ctx, kq, mask);
         }
         if (diag_mask_inf) {
             kq = ggml_diag_mask_inf_inplace(ctx, kq, 0);
@@ -756,10 +901,12 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
         kq = ggml_soft_max_inplace(ctx, kq);
 
         kqv = ggml_mul_mat(ctx, v, kq);  // [N * n_head, L_q, d_head]
+
+        kqv = ggml_reshape_4d(ctx, kqv, d_head, L_q, n_head, N);  // [N, n_head, L_q, d_head]
+        kqv = ggml_permute(ctx, kqv, 0, 2, 1, 3);                 // [N, L_q, n_head, d_head]
     }
 
-    kqv = ggml_reshape_4d(ctx, kqv, d_head, L_q, n_head, N);   // [N, n_head, L_q, d_head]
-    kqv = ggml_cont(ctx, ggml_permute(ctx, kqv, 0, 2, 1, 3));  // [N, L_q, n_head, d_head]
+    kqv = ggml_cont(ctx, kqv);
     kqv = ggml_reshape_3d(ctx, kqv, d_head * n_head, L_q, N);  // [N, L_q, C]
 
     return kqv;
@@ -801,7 +948,7 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_group_norm(struct ggml_context* ct
 }
 
 __STATIC_INLINE__ void ggml_backend_tensor_get_and_sync(ggml_backend_t backend, const struct ggml_tensor* tensor, void* data, size_t offset, size_t size) {
-#if defined(SD_USE_CUBLAS) || defined(SD_USE_SYCL)
+#if defined(SD_USE_CUDA) || defined(SD_USE_SYCL)
     if (!ggml_backend_is_cpu(backend)) {
         ggml_backend_tensor_get_async(backend, tensor, data, offset, size);
         ggml_backend_synchronize(backend);
@@ -924,8 +1071,8 @@ __STATIC_INLINE__ size_t ggml_tensor_num(ggml_context* ctx) {
 }
 
 /* SDXL with LoRA requires more space */
-#define MAX_PARAMS_TENSOR_NUM 15360
-#define MAX_GRAPH_SIZE 15360
+#define MAX_PARAMS_TENSOR_NUM 32768
+#define MAX_GRAPH_SIZE 32768
 
 struct GGMLRunner {
 protected:
@@ -939,7 +1086,6 @@ protected:
 
     std::map<struct ggml_tensor*, const void*> backend_tensor_data_map;
 
-    ggml_type wtype        = GGML_TYPE_F32;
     ggml_backend_t backend = NULL;
 
     void alloc_params_ctx() {
@@ -1015,8 +1161,8 @@ protected:
 public:
     virtual std::string get_desc() = 0;
 
-    GGMLRunner(ggml_backend_t backend, ggml_type wtype = GGML_TYPE_F32)
-        : backend(backend), wtype(wtype) {
+    GGMLRunner(ggml_backend_t backend)
+        : backend(backend) {
         alloc_params_ctx();
     }
 
@@ -1047,6 +1193,11 @@ public:
                   params_buffer_size / (1024.0 * 1024.0),
                   ggml_backend_is_cpu(backend) ? "RAM" : "VRAM",
                   num_tensors);
+        // printf("%s params backend buffer size = % 6.2f MB(%s) (%i tensors)\n",
+        //           get_desc().c_str(),
+        //           params_buffer_size / (1024.0 * 1024.0),
+        //           ggml_backend_is_cpu(backend) ? "RAM" : "VRAM",
+        //           num_tensors);
         return true;
     }
 
@@ -1107,18 +1258,12 @@ public:
             ggml_backend_cpu_set_n_threads(backend, n_threads);
         }
 
-#ifdef SD_USE_METAL
-        if (ggml_backend_is_metal(backend)) {
-            ggml_backend_metal_set_n_cb(backend, n_threads);
-        }
-#endif
         ggml_backend_graph_compute(backend, gf);
-
 #ifdef GGML_PERF
         ggml_graph_print(gf);
 #endif
         if (output != NULL) {
-            auto result = gf->nodes[gf->n_nodes - 1];
+            auto result = ggml_graph_node(gf, -1);
             if (*output == NULL && output_ctx != NULL) {
                 *output = ggml_dup_tensor(output_ctx, result);
             }
@@ -1140,20 +1285,22 @@ protected:
     GGMLBlockMap blocks;
     ParameterMap params;
 
-    void init_blocks(struct ggml_context* ctx, ggml_type wtype) {
+    void init_blocks(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
         for (auto& pair : blocks) {
             auto& block = pair.second;
-
-            block->init(ctx, wtype);
+            block->init(ctx, tensor_types, prefix + pair.first);
         }
     }
 
-    virtual void init_params(struct ggml_context* ctx, ggml_type wtype) {}
+    virtual void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {}
 
 public:
-    void init(struct ggml_context* ctx, ggml_type wtype) {
-        init_blocks(ctx, wtype);
-        init_params(ctx, wtype);
+    void init(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
+        if (prefix.size() > 0) {
+            prefix = prefix + ".";
+        }
+        init_blocks(ctx, tensor_types, prefix);
+        init_params(ctx, tensor_types, prefix);
     }
 
     size_t get_params_num() {
@@ -1209,13 +1356,15 @@ protected:
     bool bias;
     bool force_f32;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = (tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
         if (in_features % ggml_blck_size(wtype) != 0 || force_f32) {
             wtype = GGML_TYPE_F32;
         }
         params["weight"] = ggml_new_tensor_2d(ctx, wtype, in_features, out_features);
         if (bias) {
-            params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_features);
+            enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.ypes.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
+            params["bias"]       = ggml_new_tensor_1d(ctx, wtype, out_features);
         }
     }
 
@@ -1243,9 +1392,9 @@ class Embedding : public UnaryBlock {
 protected:
     int64_t embedding_dim;
     int64_t num_embeddings;
-
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["weight"] = ggml_new_tensor_2d(ctx, wtype, embedding_dim, num_embeddings);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = (tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
+        params["weight"]     = ggml_new_tensor_2d(ctx, wtype, embedding_dim, num_embeddings);
     }
 
 public:
@@ -1283,10 +1432,12 @@ protected:
     std::pair<int, int> dilation;
     bool bias;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["weight"] = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, kernel_size.second, kernel_size.first, in_channels, out_channels);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = GGML_TYPE_F16;  //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F16;
+        params["weight"]     = ggml_new_tensor_4d(ctx, wtype, kernel_size.second, kernel_size.first, in_channels, out_channels);
         if (bias) {
-            params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
+            enum ggml_type wtype = GGML_TYPE_F32;  // (tensor_types.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
+            params["bias"]       = ggml_new_tensor_1d(ctx, wtype, out_channels);
         }
     }
 
@@ -1326,10 +1477,12 @@ protected:
     int64_t dilation;
     bool bias;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["weight"] = ggml_new_tensor_4d(ctx, GGML_TYPE_F16, 1, kernel_size, in_channels, out_channels);  // 5d => 4d
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = GGML_TYPE_F16;                                                              //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F16;
+        params["weight"]     = ggml_new_tensor_4d(ctx, wtype, 1, kernel_size, in_channels, out_channels);  // 5d => 4d
         if (bias) {
-            params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, out_channels);
+            enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
+            params["bias"]       = ggml_new_tensor_1d(ctx, wtype, out_channels);
         }
     }
 
@@ -1368,11 +1521,13 @@ protected:
     bool elementwise_affine;
     bool bias;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
         if (elementwise_affine) {
-            params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, normalized_shape);
+            enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.ypes.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
+            params["weight"]     = ggml_new_tensor_1d(ctx, wtype, normalized_shape);
             if (bias) {
-                params["bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, normalized_shape);
+                enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.ypes.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
+                params["bias"]       = ggml_new_tensor_1d(ctx, wtype, normalized_shape);
             }
         }
     }
@@ -1408,10 +1563,12 @@ protected:
     float eps;
     bool affine;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
         if (affine) {
-            params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_channels);
-            params["bias"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_channels);
+            enum ggml_type wtype      = GGML_TYPE_F32;  //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
+            enum ggml_type bias_wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "bias") != tensor_types.end()) ? tensor_types[prefix + "bias"] : GGML_TYPE_F32;
+            params["weight"]          = ggml_new_tensor_1d(ctx, wtype, num_channels);
+            params["bias"]            = ggml_new_tensor_1d(ctx, bias_wtype, num_channels);
         }
     }
 

gits_noise.inl

@@ -329,21 +329,21 @@ const std::vector<std::vector<float>> GITS_NOISE_1_50 = {
 };
 
 const std::vector<const std::vector<std::vector<float>>*> GITS_NOISE = {
-    { &GITS_NOISE_0_80 },
-    { &GITS_NOISE_0_85 },
-    { &GITS_NOISE_0_90 },
-    { &GITS_NOISE_0_95 },
-    { &GITS_NOISE_1_00 },
-    { &GITS_NOISE_1_05 },
-    { &GITS_NOISE_1_10 },
-    { &GITS_NOISE_1_15 },
-    { &GITS_NOISE_1_20 },
-    { &GITS_NOISE_1_25 },
-    { &GITS_NOISE_1_30 },
-    { &GITS_NOISE_1_35 },
-    { &GITS_NOISE_1_40 },
-    { &GITS_NOISE_1_45 },
-    { &GITS_NOISE_1_50 }
+    &GITS_NOISE_0_80,
+    &GITS_NOISE_0_85,
+    &GITS_NOISE_0_90,
+    &GITS_NOISE_0_95,
+    &GITS_NOISE_1_00,
+    &GITS_NOISE_1_05,
+    &GITS_NOISE_1_10,
+    &GITS_NOISE_1_15,
+    &GITS_NOISE_1_20,
+    &GITS_NOISE_1_25,
+    &GITS_NOISE_1_30,
+    &GITS_NOISE_1_35,
+    &GITS_NOISE_1_40,
+    &GITS_NOISE_1_45,
+    &GITS_NOISE_1_50
 };
 
 #endif // GITS_NOISE_INL

lora.hpp

@@ -3,9 +3,93 @@
 
 #include "ggml_extend.hpp"
 
-#define LORA_GRAPH_SIZE 10240
+#define LORA_GRAPH_SIZE 15360
 
 struct LoraModel : public GGMLRunner {
+    enum lora_t {
+        REGULAR      = 0,
+        DIFFUSERS    = 1,
+        DIFFUSERS_2  = 2,
+        DIFFUSERS_3  = 3,
+        TRANSFORMERS = 4,
+        LORA_TYPE_COUNT
+    };
+
+    const std::string lora_ups[LORA_TYPE_COUNT] = {
+        ".lora_up",
+        "_lora.up",
+        ".lora_B",
+        ".lora.up",
+        ".lora_linear_layer.up",
+    };
+
+    const std::string lora_downs[LORA_TYPE_COUNT] = {
+        ".lora_down",
+        "_lora.down",
+        ".lora_A",
+        ".lora.down",
+        ".lora_linear_layer.down",
+    };
+
+    const std::string lora_pre[LORA_TYPE_COUNT] = {
+        "lora.",
+        "",
+        "",
+        "",
+        "",
+    };
+
+    const std::map<std::string, std::string> alt_names = {
+        // mmdit
+        {"final_layer.adaLN_modulation.1", "norm_out.linear"},
+        {"pos_embed", "pos_embed.proj"},
+        {"final_layer.linear", "proj_out"},
+        {"y_embedder.mlp.0", "time_text_embed.text_embedder.linear_1"},
+        {"y_embedder.mlp.2", "time_text_embed.text_embedder.linear_2"},
+        {"t_embedder.mlp.0", "time_text_embed.timestep_embedder.linear_1"},
+        {"t_embedder.mlp.2", "time_text_embed.timestep_embedder.linear_2"},
+        {"x_block.mlp.fc1", "ff.net.0.proj"},
+        {"x_block.mlp.fc2", "ff.net.2"},
+        {"context_block.mlp.fc1", "ff_context.net.0.proj"},
+        {"context_block.mlp.fc2", "ff_context.net.2"},
+        {"x_block.adaLN_modulation.1", "norm1.linear"},
+        {"context_block.adaLN_modulation.1", "norm1_context.linear"},
+        {"context_block.attn.proj", "attn.to_add_out"},
+        {"x_block.attn.proj", "attn.to_out.0"},
+        {"x_block.attn2.proj", "attn2.to_out.0"},
+        // flux
+        // singlestream
+        {"linear2", "proj_out"},
+        {"modulation.lin", "norm.linear"},
+        // doublestream
+        {"txt_attn.proj", "attn.to_add_out"},
+        {"img_attn.proj", "attn.to_out.0"},
+        {"txt_mlp.0", "ff_context.net.0.proj"},
+        {"txt_mlp.2", "ff_context.net.2"},
+        {"img_mlp.0", "ff.net.0.proj"},
+        {"img_mlp.2", "ff.net.2"},
+        {"txt_mod.lin", "norm1_context.linear"},
+        {"img_mod.lin", "norm1.linear"},
+    };
+
+    const std::map<std::string, std::string> qkv_prefixes = {
+        // mmdit
+        {"context_block.attn.qkv", "attn.add_"},  // suffix "_proj"
+        {"x_block.attn.qkv", "attn.to_"},
+        {"x_block.attn2.qkv", "attn2.to_"},
+        // flux
+        // doublestream
+        {"txt_attn.qkv", "attn.add_"},  // suffix "_proj"
+        {"img_attn.qkv", "attn.to_"},
+    };
+    const std::map<std::string, std::string> qkvm_prefixes = {
+        // flux
+        // singlestream
+        {"linear1", ""},
+    };
+
+    const std::string* type_fingerprints = lora_ups;
+
     float multiplier = 1.0f;
     std::map<std::string, struct ggml_tensor*> lora_tensors;
     std::string file_path;
@@ -14,12 +98,12 @@ struct LoraModel : public GGMLRunner {
     bool applied                    = false;
     std::vector<int> zero_index_vec = {0};
     ggml_tensor* zero_index         = NULL;
+    enum lora_t type                = REGULAR;
 
     LoraModel(ggml_backend_t backend,
-              ggml_type wtype,
               const std::string& file_path = "",
-              const std::string& prefix    = "")
-        : file_path(file_path), GGMLRunner(backend, wtype) {
+              const std::string prefix     = "")
+        : file_path(file_path), GGMLRunner(backend) {
         if (!model_loader.init_from_file(file_path, prefix)) {
             load_failed = true;
         }
@@ -45,6 +129,13 @@ struct LoraModel : public GGMLRunner {
                 // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
                 return true;
             }
+            // LOG_INFO("%s", name.c_str());
+            for (int i = 0; i < LORA_TYPE_COUNT; i++) {
+                if (name.find(type_fingerprints[i]) != std::string::npos) {
+                    type = (lora_t)i;
+                    break;
+                }
+            }
 
             if (dry_run) {
                 struct ggml_tensor* real = ggml_new_tensor(params_ctx,
@@ -62,10 +153,12 @@ struct LoraModel : public GGMLRunner {
 
         model_loader.load_tensors(on_new_tensor_cb, backend);
         alloc_params_buffer();
-
+        // exit(0);
         dry_run = false;
         model_loader.load_tensors(on_new_tensor_cb, backend);
 
+        LOG_DEBUG("lora type: \"%s\"/\"%s\"", lora_downs[type].c_str(), lora_ups[type].c_str());
+
         LOG_DEBUG("finished loaded lora");
         return true;
     }
@@ -77,7 +170,74 @@ struct LoraModel : public GGMLRunner {
         return out;
     }
 
-    struct ggml_cgraph* build_lora_graph(std::map<std::string, struct ggml_tensor*> model_tensors) {
+    std::vector<std::string> to_lora_keys(std::string blk_name, SDVersion version) {
+        std::vector<std::string> keys;
+        // if (!sd_version_is_sd3(version) || blk_name != "model.diffusion_model.pos_embed") {
+        size_t k_pos = blk_name.find(".weight");
+        if (k_pos == std::string::npos) {
+            return keys;
+        }
+        blk_name = blk_name.substr(0, k_pos);
+        // }
+        keys.push_back(blk_name);
+        keys.push_back("lora." + blk_name);
+        if (sd_version_is_dit(version)) {
+            if (blk_name.find("model.diffusion_model") != std::string::npos) {
+                blk_name.replace(blk_name.find("model.diffusion_model"), sizeof("model.diffusion_model") - 1, "transformer");
+            }
+
+            if (blk_name.find(".single_blocks") != std::string::npos) {
+                blk_name.replace(blk_name.find(".single_blocks"), sizeof(".single_blocks") - 1, ".single_transformer_blocks");
+            }
+            if (blk_name.find(".double_blocks") != std::string::npos) {
+                blk_name.replace(blk_name.find(".double_blocks"), sizeof(".double_blocks") - 1, ".transformer_blocks");
+            }
+
+            if (blk_name.find(".joint_blocks") != std::string::npos) {
+                blk_name.replace(blk_name.find(".joint_blocks"), sizeof(".joint_blocks") - 1, ".transformer_blocks");
+            }
+
+            if (blk_name.find("text_encoders.clip_l") != std::string::npos) {
+                blk_name.replace(blk_name.find("text_encoders.clip_l"), sizeof("text_encoders.clip_l") - 1, "cond_stage_model");
+            }
+
+            for (const auto& item : alt_names) {
+                size_t match = blk_name.find(item.first);
+                if (match != std::string::npos) {
+                    blk_name = blk_name.substr(0, match) + item.second;
+                }
+            }
+            for (const auto& prefix : qkv_prefixes) {
+                size_t match = blk_name.find(prefix.first);
+                if (match != std::string::npos) {
+                    std::string split_blk = "SPLIT|" + blk_name.substr(0, match) + prefix.second;
+                    keys.push_back(split_blk);
+                }
+            }
+            for (const auto& prefix : qkvm_prefixes) {
+                size_t match = blk_name.find(prefix.first);
+                if (match != std::string::npos) {
+                    std::string split_blk = "SPLIT_L|" + blk_name.substr(0, match) + prefix.second;
+                    keys.push_back(split_blk);
+                }
+            }
+            keys.push_back(blk_name);
+        }
+
+        std::vector<std::string> ret;
+        for (std::string& key : keys) {
+            ret.push_back(key);
+            replace_all_chars(key, '.', '_');
+            // fix for some sdxl lora, like lcm-lora-xl
+            if (key == "model_diffusion_model_output_blocks_2_2_conv") {
+                ret.push_back("model_diffusion_model_output_blocks_2_1_conv");
+            }
+            ret.push_back(key);
+        }
+        return ret;
+    }
+
+    struct ggml_cgraph* build_lora_graph(std::map<std::string, struct ggml_tensor*> model_tensors, SDVersion version) {
         struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, LORA_GRAPH_SIZE, false);
 
         zero_index = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_I32, 1);
@@ -89,91 +249,573 @@ struct LoraModel : public GGMLRunner {
             std::string k_tensor       = it.first;
             struct ggml_tensor* weight = model_tensors[it.first];
 
-            size_t k_pos = k_tensor.find(".weight");
-            if (k_pos == std::string::npos) {
+            std::vector<std::string> keys = to_lora_keys(k_tensor, version);
+            if (keys.size() == 0)
                 continue;
-            }
-            k_tensor = k_tensor.substr(0, k_pos);
-            replace_all_chars(k_tensor, '.', '_');
-            // LOG_DEBUG("k_tensor %s", k_tensor.c_str());
-            std::string lora_up_name = "lora." + k_tensor + ".lora_up.weight";
-            if (lora_tensors.find(lora_up_name) == lora_tensors.end()) {
-                if (k_tensor == "model_diffusion_model_output_blocks_2_2_conv") {
-                    // fix for some sdxl lora, like lcm-lora-xl
-                    k_tensor     = "model_diffusion_model_output_blocks_2_1_conv";
-                    lora_up_name = "lora." + k_tensor + ".lora_up.weight";
+
+            for (auto& key : keys) {
+                bool is_qkv_split = starts_with(key, "SPLIT|");
+                if (is_qkv_split) {
+                    key = key.substr(sizeof("SPLIT|") - 1);
                 }
+                bool is_qkvm_split = starts_with(key, "SPLIT_L|");
+                if (is_qkvm_split) {
+                    key = key.substr(sizeof("SPLIT_L|") - 1);
+                }
+                struct ggml_tensor* updown = NULL;
+                float scale_value          = 1.0f;
+                std::string fk             = lora_pre[type] + key;
+                if (lora_tensors.find(fk + ".hada_w1_a") != lora_tensors.end()) {
+                    // LoHa mode
+
+                    // TODO: split qkv convention for LoHas (is it ever used?)
+                    if (is_qkv_split || is_qkvm_split) {
+                        LOG_ERROR("Split qkv isn't supported for LoHa models.");
+                        break;
+                    }
+                    std::string alpha_name = "";
+
+                    ggml_tensor* hada_1_mid  = NULL;  // tau for tucker decomposition
+                    ggml_tensor* hada_1_up   = NULL;
+                    ggml_tensor* hada_1_down = NULL;
+
+                    ggml_tensor* hada_2_mid  = NULL;  // tau for tucker decomposition
+                    ggml_tensor* hada_2_up   = NULL;
+                    ggml_tensor* hada_2_down = NULL;
+
+                    std::string hada_1_mid_name  = "";
+                    std::string hada_1_down_name = "";
+                    std::string hada_1_up_name   = "";
+
+                    std::string hada_2_mid_name  = "";
+                    std::string hada_2_down_name = "";
+                    std::string hada_2_up_name   = "";
+
+                    hada_1_down_name = fk + ".hada_w1_b";
+                    hada_1_up_name   = fk + ".hada_w1_a";
+                    hada_1_mid_name  = fk + ".hada_t1";
+                    if (lora_tensors.find(hada_1_down_name) != lora_tensors.end()) {
+                        hada_1_down = to_f32(compute_ctx, lora_tensors[hada_1_down_name]);
+                    }
+                    if (lora_tensors.find(hada_1_up_name) != lora_tensors.end()) {
+                        hada_1_up = to_f32(compute_ctx, lora_tensors[hada_1_up_name]);
+                    }
+                    if (lora_tensors.find(hada_1_mid_name) != lora_tensors.end()) {
+                        hada_1_mid = to_f32(compute_ctx, lora_tensors[hada_1_mid_name]);
+                        applied_lora_tensors.insert(hada_1_mid_name);
+                        hada_1_up = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, hada_1_up));
+                    }
+
+                    hada_2_down_name = fk + ".hada_w2_b";
+                    hada_2_up_name   = fk + ".hada_w2_a";
+                    hada_2_mid_name  = fk + ".hada_t2";
+                    if (lora_tensors.find(hada_2_down_name) != lora_tensors.end()) {
+                        hada_2_down = to_f32(compute_ctx, lora_tensors[hada_2_down_name]);
+                    }
+                    if (lora_tensors.find(hada_2_up_name) != lora_tensors.end()) {
+                        hada_2_up = to_f32(compute_ctx, lora_tensors[hada_2_up_name]);
+                    }
+                    if (lora_tensors.find(hada_2_mid_name) != lora_tensors.end()) {
+                        hada_2_mid = to_f32(compute_ctx, lora_tensors[hada_2_mid_name]);
+                        applied_lora_tensors.insert(hada_2_mid_name);
+                        hada_2_up = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, hada_2_up));
+                    }
+
+                    alpha_name = fk + ".alpha";
+
+                    applied_lora_tensors.insert(hada_1_down_name);
+                    applied_lora_tensors.insert(hada_1_up_name);
+                    applied_lora_tensors.insert(hada_2_down_name);
+                    applied_lora_tensors.insert(hada_2_up_name);
+
+                    applied_lora_tensors.insert(alpha_name);
+                    if (hada_1_up == NULL || hada_1_down == NULL || hada_2_up == NULL || hada_2_down == NULL) {
+                        continue;
+                    }
+
+                    struct ggml_tensor* updown_1 = ggml_merge_lora(compute_ctx, hada_1_down, hada_1_up, hada_1_mid);
+                    struct ggml_tensor* updown_2 = ggml_merge_lora(compute_ctx, hada_2_down, hada_2_up, hada_2_mid);
+                    updown                       = ggml_mul_inplace(compute_ctx, updown_1, updown_2);
+
+                    // calc_scale
+                    // TODO: .dora_scale?
+                    int64_t rank = hada_1_down->ne[ggml_n_dims(hada_1_down) - 1];
+                    if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                        float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                        scale_value = alpha / rank;
+                    }
+                } else if (lora_tensors.find(fk + ".lokr_w1") != lora_tensors.end() || lora_tensors.find(fk + ".lokr_w1_a") != lora_tensors.end()) {
+                    // LoKr mode
+
+                    // TODO: split qkv convention for LoKrs (is it ever used?)
+                    if (is_qkv_split || is_qkvm_split) {
+                        LOG_ERROR("Split qkv isn't supported for LoKr models.");
+                        break;
+                    }
+
+                    std::string alpha_name = fk + ".alpha";
+
+                    ggml_tensor* lokr_w1 = NULL;
+                    ggml_tensor* lokr_w2 = NULL;
+
+                    std::string lokr_w1_name = "";
+                    std::string lokr_w2_name = "";
+
+                    lokr_w1_name = fk + ".lokr_w1";
+                    lokr_w2_name = fk + ".lokr_w2";
+
+                    if (lora_tensors.find(lokr_w1_name) != lora_tensors.end()) {
+                        lokr_w1 = to_f32(compute_ctx, lora_tensors[lokr_w1_name]);
+                        applied_lora_tensors.insert(lokr_w1_name);
+                    } else {
+                        ggml_tensor* down     = NULL;
+                        ggml_tensor* up       = NULL;
+                        std::string down_name = lokr_w1_name + "_b";
+                        std::string up_name   = lokr_w1_name + "_a";
+                        if (lora_tensors.find(down_name) != lora_tensors.end()) {
+                            // w1 should not be low rank normally, sometimes w1 and w2 are swapped
+                            down = to_f32(compute_ctx, lora_tensors[down_name]);
+                            applied_lora_tensors.insert(down_name);
+
+                            int64_t rank = down->ne[ggml_n_dims(down) - 1];
+                            if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                                float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                                scale_value = alpha / rank;
+                            }
+                        }
+                        if (lora_tensors.find(up_name) != lora_tensors.end()) {
+                            up = to_f32(compute_ctx, lora_tensors[up_name]);
+                            applied_lora_tensors.insert(up_name);
+                        }
+                        lokr_w1 = ggml_merge_lora(compute_ctx, down, up);
+                    }
+                    if (lora_tensors.find(lokr_w2_name) != lora_tensors.end()) {
+                        lokr_w2 = to_f32(compute_ctx, lora_tensors[lokr_w2_name]);
+                        applied_lora_tensors.insert(lokr_w2_name);
+                    } else {
+                        ggml_tensor* down     = NULL;
+                        ggml_tensor* up       = NULL;
+                        std::string down_name = lokr_w2_name + "_b";
+                        std::string up_name   = lokr_w2_name + "_a";
+                        if (lora_tensors.find(down_name) != lora_tensors.end()) {
+                            down = to_f32(compute_ctx, lora_tensors[down_name]);
+                            applied_lora_tensors.insert(down_name);
+
+                            int64_t rank = down->ne[ggml_n_dims(down) - 1];
+                            if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                                float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                                scale_value = alpha / rank;
+                            }
+                        }
+                        if (lora_tensors.find(up_name) != lora_tensors.end()) {
+                            up = to_f32(compute_ctx, lora_tensors[up_name]);
+                            applied_lora_tensors.insert(up_name);
+                        }
+                        lokr_w2 = ggml_merge_lora(compute_ctx, down, up);
+                    }
+
+                    // Technically it might be unused, but I believe it's the expected behavior
+                    applied_lora_tensors.insert(alpha_name);
+
+                    updown = ggml_kronecker(compute_ctx, lokr_w1, lokr_w2);
+
+                } else {
+                    // LoRA mode
+                    ggml_tensor* lora_mid  = NULL;  // tau for tucker decomposition
+                    ggml_tensor* lora_up   = NULL;
+                    ggml_tensor* lora_down = NULL;
+
+                    std::string alpha_name         = "";
+                    std::string scale_name         = "";
+                    std::string split_q_scale_name = "";
+                    std::string lora_mid_name      = "";
+                    std::string lora_down_name     = "";
+                    std::string lora_up_name       = "";
+
+                    if (is_qkv_split) {
+                        std::string suffix  = "";
+                        auto split_q_d_name = fk + "q" + suffix + lora_downs[type] + ".weight";
+
+                        if (lora_tensors.find(split_q_d_name) == lora_tensors.end()) {
+                            suffix         = "_proj";
+                            split_q_d_name = fk + "q" + suffix + lora_downs[type] + ".weight";
+                        }
+                        if (lora_tensors.find(split_q_d_name) != lora_tensors.end()) {
+                            // print_ggml_tensor(it.second, true);  //[3072, 21504, 1, 1]
+                            // find qkv and mlp up parts in LoRA model
+                            auto split_k_d_name = fk + "k" + suffix + lora_downs[type] + ".weight";
+                            auto split_v_d_name = fk + "v" + suffix + lora_downs[type] + ".weight";
+
+                            auto split_q_u_name = fk + "q" + suffix + lora_ups[type] + ".weight";
+                            auto split_k_u_name = fk + "k" + suffix + lora_ups[type] + ".weight";
+                            auto split_v_u_name = fk + "v" + suffix + lora_ups[type] + ".weight";
+
+                            auto split_q_scale_name = fk + "q" + suffix + ".scale";
+                            auto split_k_scale_name = fk + "k" + suffix + ".scale";
+                            auto split_v_scale_name = fk + "v" + suffix + ".scale";
+
+                            auto split_q_alpha_name = fk + "q" + suffix + ".alpha";
+                            auto split_k_alpha_name = fk + "k" + suffix + ".alpha";
+                            auto split_v_alpha_name = fk + "v" + suffix + ".alpha";
+
+                            ggml_tensor* lora_q_down = NULL;
+                            ggml_tensor* lora_q_up   = NULL;
+                            ggml_tensor* lora_k_down = NULL;
+                            ggml_tensor* lora_k_up   = NULL;
+                            ggml_tensor* lora_v_down = NULL;
+                            ggml_tensor* lora_v_up   = NULL;
+
+                            lora_q_down = to_f32(compute_ctx, lora_tensors[split_q_d_name]);
+
+                            if (lora_tensors.find(split_q_u_name) != lora_tensors.end()) {
+                                lora_q_up = to_f32(compute_ctx, lora_tensors[split_q_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_k_d_name) != lora_tensors.end()) {
+                                lora_k_down = to_f32(compute_ctx, lora_tensors[split_k_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_k_u_name) != lora_tensors.end()) {
+                                lora_k_up = to_f32(compute_ctx, lora_tensors[split_k_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_v_d_name) != lora_tensors.end()) {
+                                lora_v_down = to_f32(compute_ctx, lora_tensors[split_v_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_v_u_name) != lora_tensors.end()) {
+                                lora_v_up = to_f32(compute_ctx, lora_tensors[split_v_u_name]);
+                            }
+
+                            float q_rank = lora_q_up->ne[0];
+                            float k_rank = lora_k_up->ne[0];
+                            float v_rank = lora_v_up->ne[0];
+
+                            float lora_q_scale = 1;
+                            float lora_k_scale = 1;
+                            float lora_v_scale = 1;
+
+                            if (lora_tensors.find(split_q_scale_name) != lora_tensors.end()) {
+                                lora_q_scale = ggml_backend_tensor_get_f32(lora_tensors[split_q_scale_name]);
+                                applied_lora_tensors.insert(split_q_scale_name);
+                            }
+                            if (lora_tensors.find(split_k_scale_name) != lora_tensors.end()) {
+                                lora_k_scale = ggml_backend_tensor_get_f32(lora_tensors[split_k_scale_name]);
+                                applied_lora_tensors.insert(split_k_scale_name);
+                            }
+                            if (lora_tensors.find(split_v_scale_name) != lora_tensors.end()) {
+                                lora_v_scale = ggml_backend_tensor_get_f32(lora_tensors[split_v_scale_name]);
+                                applied_lora_tensors.insert(split_v_scale_name);
+                            }
+
+                            if (lora_tensors.find(split_q_alpha_name) != lora_tensors.end()) {
+                                float lora_q_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_q_alpha_name]);
+                                applied_lora_tensors.insert(split_q_alpha_name);
+                                lora_q_scale = lora_q_alpha / q_rank;
+                            }
+                            if (lora_tensors.find(split_k_alpha_name) != lora_tensors.end()) {
+                                float lora_k_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_k_alpha_name]);
+                                applied_lora_tensors.insert(split_k_alpha_name);
+                                lora_k_scale = lora_k_alpha / k_rank;
+                            }
+                            if (lora_tensors.find(split_v_alpha_name) != lora_tensors.end()) {
+                                float lora_v_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_v_alpha_name]);
+                                applied_lora_tensors.insert(split_v_alpha_name);
+                                lora_v_scale = lora_v_alpha / v_rank;
+                            }
+
+                            ggml_scale_inplace(compute_ctx, lora_q_down, lora_q_scale);
+                            ggml_scale_inplace(compute_ctx, lora_k_down, lora_k_scale);
+                            ggml_scale_inplace(compute_ctx, lora_v_down, lora_v_scale);
+
+                            // print_ggml_tensor(lora_q_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_k_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_v_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_q_up, true);    //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_k_up, true);    //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_v_up, true);    //[R, 3072, 1, 1]
+
+                            // these need to be stitched together this way:
+                            //                          |q_up,0   ,0   |
+                            //                          |0   ,k_up,0   |
+                            //                          |0   ,0   ,v_up|
+                            // (q_down,k_down,v_down) . (q   ,k   ,v)
+
+                            // up_concat will be [9216, R*3, 1, 1]
+                            // down_concat will be [R*3, 3072, 1, 1]
+                            ggml_tensor* lora_down_concat = ggml_concat(compute_ctx, ggml_concat(compute_ctx, lora_q_down, lora_k_down, 1), lora_v_down, 1);
+
+                            ggml_tensor* z = ggml_dup_tensor(compute_ctx, lora_q_up);
+                            ggml_scale(compute_ctx, z, 0);
+                            ggml_tensor* zz = ggml_concat(compute_ctx, z, z, 1);
+
+                            ggml_tensor* q_up = ggml_concat(compute_ctx, lora_q_up, zz, 1);
+                            ggml_tensor* k_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, z, lora_k_up, 1), z, 1);
+                            ggml_tensor* v_up = ggml_concat(compute_ctx, zz, lora_v_up, 1);
+                            // print_ggml_tensor(q_up, true);  //[R, 9216, 1, 1]
+                            // print_ggml_tensor(k_up, true);  //[R, 9216, 1, 1]
+                            // print_ggml_tensor(v_up, true);  //[R, 9216, 1, 1]
+                            ggml_tensor* lora_up_concat = ggml_concat(compute_ctx, ggml_concat(compute_ctx, q_up, k_up, 0), v_up, 0);
+                            // print_ggml_tensor(lora_up_concat, true);  //[R*3, 9216, 1, 1]
+
+                            lora_down = ggml_cont(compute_ctx, lora_down_concat);
+                            lora_up   = ggml_cont(compute_ctx, lora_up_concat);
+
+                            applied_lora_tensors.insert(split_q_u_name);
+                            applied_lora_tensors.insert(split_k_u_name);
+                            applied_lora_tensors.insert(split_v_u_name);
+
+                            applied_lora_tensors.insert(split_q_d_name);
+                            applied_lora_tensors.insert(split_k_d_name);
+                            applied_lora_tensors.insert(split_v_d_name);
+                        }
+                    } else if (is_qkvm_split) {
+                        auto split_q_d_name = fk + "attn.to_q" + lora_downs[type] + ".weight";
+                        if (lora_tensors.find(split_q_d_name) != lora_tensors.end()) {
+                            // print_ggml_tensor(it.second, true);  //[3072, 21504, 1, 1]
+                            // find qkv and mlp up parts in LoRA model
+                            auto split_k_d_name = fk + "attn.to_k" + lora_downs[type] + ".weight";
+                            auto split_v_d_name = fk + "attn.to_v" + lora_downs[type] + ".weight";
+
+                            auto split_q_u_name = fk + "attn.to_q" + lora_ups[type] + ".weight";
+                            auto split_k_u_name = fk + "attn.to_k" + lora_ups[type] + ".weight";
+                            auto split_v_u_name = fk + "attn.to_v" + lora_ups[type] + ".weight";
+
+                            auto split_m_d_name = fk + "proj_mlp" + lora_downs[type] + ".weight";
+                            auto split_m_u_name = fk + "proj_mlp" + lora_ups[type] + ".weight";
+
+                            auto split_q_scale_name = fk + "attn.to_q" + ".scale";
+                            auto split_k_scale_name = fk + "attn.to_k" + ".scale";
+                            auto split_v_scale_name = fk + "attn.to_v" + ".scale";
+                            auto split_m_scale_name = fk + "proj_mlp" + ".scale";
+
+                            auto split_q_alpha_name = fk + "attn.to_q" + ".alpha";
+                            auto split_k_alpha_name = fk + "attn.to_k" + ".alpha";
+                            auto split_v_alpha_name = fk + "attn.to_v" + ".alpha";
+                            auto split_m_alpha_name = fk + "proj_mlp" + ".alpha";
+
+                            ggml_tensor* lora_q_down = NULL;
+                            ggml_tensor* lora_q_up   = NULL;
+                            ggml_tensor* lora_k_down = NULL;
+                            ggml_tensor* lora_k_up   = NULL;
+                            ggml_tensor* lora_v_down = NULL;
+                            ggml_tensor* lora_v_up   = NULL;
+
+                            ggml_tensor* lora_m_down = NULL;
+                            ggml_tensor* lora_m_up   = NULL;
+
+                            lora_q_up = to_f32(compute_ctx, lora_tensors[split_q_u_name]);
+
+                            if (lora_tensors.find(split_q_d_name) != lora_tensors.end()) {
+                                lora_q_down = to_f32(compute_ctx, lora_tensors[split_q_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_q_u_name) != lora_tensors.end()) {
+                                lora_q_up = to_f32(compute_ctx, lora_tensors[split_q_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_k_d_name) != lora_tensors.end()) {
+                                lora_k_down = to_f32(compute_ctx, lora_tensors[split_k_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_k_u_name) != lora_tensors.end()) {
+                                lora_k_up = to_f32(compute_ctx, lora_tensors[split_k_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_v_d_name) != lora_tensors.end()) {
+                                lora_v_down = to_f32(compute_ctx, lora_tensors[split_v_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_v_u_name) != lora_tensors.end()) {
+                                lora_v_up = to_f32(compute_ctx, lora_tensors[split_v_u_name]);
+                            }
+
+                            if (lora_tensors.find(split_m_d_name) != lora_tensors.end()) {
+                                lora_m_down = to_f32(compute_ctx, lora_tensors[split_m_d_name]);
+                            }
+
+                            if (lora_tensors.find(split_m_u_name) != lora_tensors.end()) {
+                                lora_m_up = to_f32(compute_ctx, lora_tensors[split_m_u_name]);
+                            }
+
+                            float q_rank = lora_q_up->ne[0];
+                            float k_rank = lora_k_up->ne[0];
+                            float v_rank = lora_v_up->ne[0];
+                            float m_rank = lora_v_up->ne[0];
+
+                            float lora_q_scale = 1;
+                            float lora_k_scale = 1;
+                            float lora_v_scale = 1;
+                            float lora_m_scale = 1;
+
+                            if (lora_tensors.find(split_q_scale_name) != lora_tensors.end()) {
+                                lora_q_scale = ggml_backend_tensor_get_f32(lora_tensors[split_q_scale_name]);
+                                applied_lora_tensors.insert(split_q_scale_name);
+                            }
+                            if (lora_tensors.find(split_k_scale_name) != lora_tensors.end()) {
+                                lora_k_scale = ggml_backend_tensor_get_f32(lora_tensors[split_k_scale_name]);
+                                applied_lora_tensors.insert(split_k_scale_name);
+                            }
+                            if (lora_tensors.find(split_v_scale_name) != lora_tensors.end()) {
+                                lora_v_scale = ggml_backend_tensor_get_f32(lora_tensors[split_v_scale_name]);
+                                applied_lora_tensors.insert(split_v_scale_name);
+                            }
+                            if (lora_tensors.find(split_m_scale_name) != lora_tensors.end()) {
+                                lora_m_scale = ggml_backend_tensor_get_f32(lora_tensors[split_m_scale_name]);
+                                applied_lora_tensors.insert(split_m_scale_name);
+                            }
+
+                            if (lora_tensors.find(split_q_alpha_name) != lora_tensors.end()) {
+                                float lora_q_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_q_alpha_name]);
+                                applied_lora_tensors.insert(split_q_alpha_name);
+                                lora_q_scale = lora_q_alpha / q_rank;
+                            }
+                            if (lora_tensors.find(split_k_alpha_name) != lora_tensors.end()) {
+                                float lora_k_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_k_alpha_name]);
+                                applied_lora_tensors.insert(split_k_alpha_name);
+                                lora_k_scale = lora_k_alpha / k_rank;
+                            }
+                            if (lora_tensors.find(split_v_alpha_name) != lora_tensors.end()) {
+                                float lora_v_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_v_alpha_name]);
+                                applied_lora_tensors.insert(split_v_alpha_name);
+                                lora_v_scale = lora_v_alpha / v_rank;
+                            }
+                            if (lora_tensors.find(split_m_alpha_name) != lora_tensors.end()) {
+                                float lora_m_alpha = ggml_backend_tensor_get_f32(lora_tensors[split_m_alpha_name]);
+                                applied_lora_tensors.insert(split_m_alpha_name);
+                                lora_m_scale = lora_m_alpha / m_rank;
+                            }
+
+                            ggml_scale_inplace(compute_ctx, lora_q_down, lora_q_scale);
+                            ggml_scale_inplace(compute_ctx, lora_k_down, lora_k_scale);
+                            ggml_scale_inplace(compute_ctx, lora_v_down, lora_v_scale);
+                            ggml_scale_inplace(compute_ctx, lora_m_down, lora_m_scale);
+
+                            // print_ggml_tensor(lora_q_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_k_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_v_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_m_down, true);  //[3072, R, 1, 1]
+                            // print_ggml_tensor(lora_q_up, true);  //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_k_up, true);  //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_v_up, true);  //[R, 3072, 1, 1]
+                            // print_ggml_tensor(lora_m_up, true);  //[R, 12288, 1, 1]
+
+                            // these need to be stitched together this way:
+                            //                                 |q_up,0   ,0   ,0   |
+                            //                                 |0   ,k_up,0   ,0   |
+                            //                                 |0   ,0   ,v_up,0   |
+                            //                                 |0   ,0   ,0   ,m_up|
+                            // (q_down,k_down,v_down,m_down) . (q   ,k   ,v   ,m)
+
+                            // up_concat will be [21504, R*4, 1, 1]
+                            // down_concat will be [R*4, 3072, 1, 1]
+
+                            ggml_tensor* lora_down_concat = ggml_concat(compute_ctx, ggml_concat(compute_ctx, lora_q_down, lora_k_down, 1), ggml_concat(compute_ctx, lora_v_down, lora_m_down, 1), 1);
+                            // print_ggml_tensor(lora_down_concat, true);  //[3072, R*4, 1, 1]
+
+                            // this also means that if rank is bigger than 672, it is less memory efficient to do it this way (should be fine)
+                            // print_ggml_tensor(lora_q_up, true);  //[3072, R, 1, 1]
+                            ggml_tensor* z     = ggml_dup_tensor(compute_ctx, lora_q_up);
+                            ggml_tensor* mlp_z = ggml_dup_tensor(compute_ctx, lora_m_up);
+                            ggml_scale(compute_ctx, z, 0);
+                            ggml_scale(compute_ctx, mlp_z, 0);
+                            ggml_tensor* zz = ggml_concat(compute_ctx, z, z, 1);
+
+                            ggml_tensor* q_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, lora_q_up, zz, 1), mlp_z, 1);
+                            ggml_tensor* k_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, z, lora_k_up, 1), ggml_concat(compute_ctx, z, mlp_z, 1), 1);
+                            ggml_tensor* v_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, zz, lora_v_up, 1), mlp_z, 1);
+                            ggml_tensor* m_up = ggml_concat(compute_ctx, ggml_concat(compute_ctx, zz, z, 1), lora_m_up, 1);
+                            // print_ggml_tensor(q_up, true);  //[R, 21504, 1, 1]
+                            // print_ggml_tensor(k_up, true);  //[R, 21504, 1, 1]
+                            // print_ggml_tensor(v_up, true);  //[R, 21504, 1, 1]
+                            // print_ggml_tensor(m_up, true);  //[R, 21504, 1, 1]
+
+                            ggml_tensor* lora_up_concat = ggml_concat(compute_ctx, ggml_concat(compute_ctx, q_up, k_up, 0), ggml_concat(compute_ctx, v_up, m_up, 0), 0);
+                            // print_ggml_tensor(lora_up_concat, true);  //[R*4, 21504, 1, 1]
+
+                            lora_down = ggml_cont(compute_ctx, lora_down_concat);
+                            lora_up   = ggml_cont(compute_ctx, lora_up_concat);
+
+                            applied_lora_tensors.insert(split_q_u_name);
+                            applied_lora_tensors.insert(split_k_u_name);
+                            applied_lora_tensors.insert(split_v_u_name);
+                            applied_lora_tensors.insert(split_m_u_name);
+
+                            applied_lora_tensors.insert(split_q_d_name);
+                            applied_lora_tensors.insert(split_k_d_name);
+                            applied_lora_tensors.insert(split_v_d_name);
+                            applied_lora_tensors.insert(split_m_d_name);
+                        }
+                    } else {
+                        lora_up_name   = fk + lora_ups[type] + ".weight";
+                        lora_down_name = fk + lora_downs[type] + ".weight";
+                        lora_mid_name  = fk + ".lora_mid.weight";
+
+                        alpha_name = fk + ".alpha";
+                        scale_name = fk + ".scale";
+
+                        if (lora_tensors.find(lora_up_name) != lora_tensors.end()) {
+                            lora_up = to_f32(compute_ctx, lora_tensors[lora_up_name]);
+                        }
+
+                        if (lora_tensors.find(lora_down_name) != lora_tensors.end()) {
+                            lora_down = to_f32(compute_ctx, lora_tensors[lora_down_name]);
+                        }
+
+                        if (lora_tensors.find(lora_mid_name) != lora_tensors.end()) {
+                            lora_mid = to_f32(compute_ctx, lora_tensors[lora_mid_name]);
+                            applied_lora_tensors.insert(lora_mid_name);
+                        }
+
+                        applied_lora_tensors.insert(lora_up_name);
+                        applied_lora_tensors.insert(lora_down_name);
+                        applied_lora_tensors.insert(alpha_name);
+                        applied_lora_tensors.insert(scale_name);
+                    }
+
+                    if (lora_up == NULL || lora_down == NULL) {
+                        continue;
+                    }
+                    // calc_scale
+                    // TODO: .dora_scale?
+                    int64_t rank = lora_down->ne[ggml_n_dims(lora_down) - 1];
+                    if (lora_tensors.find(scale_name) != lora_tensors.end()) {
+                        scale_value = ggml_backend_tensor_get_f32(lora_tensors[scale_name]);
+                    } else if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
+                        float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
+                        scale_value = alpha / rank;
+                    }
+
+                    updown = ggml_merge_lora(compute_ctx, lora_down, lora_up, lora_mid);
+                }
+                scale_value *= multiplier;
+                updown = ggml_reshape(compute_ctx, updown, weight);
+                GGML_ASSERT(ggml_nelements(updown) == ggml_nelements(weight));
+                updown = ggml_scale_inplace(compute_ctx, updown, scale_value);
+                ggml_tensor* final_weight;
+                if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
+                    // final_weight = ggml_new_tensor(compute_ctx, GGML_TYPE_F32, ggml_n_dims(weight), weight->ne);
+                    // final_weight = ggml_cpy(compute_ctx, weight, final_weight);
+                    final_weight = to_f32(compute_ctx, weight);
+                    final_weight = ggml_add_inplace(compute_ctx, final_weight, updown);
+                    final_weight = ggml_cpy(compute_ctx, final_weight, weight);
+                } else {
+                    final_weight = ggml_add_inplace(compute_ctx, weight, updown);
+                }
+                // final_weight = ggml_add_inplace(compute_ctx, weight, updown);  // apply directly
+                ggml_build_forward_expand(gf, final_weight);
+                break;
             }
-
-            std::string lora_down_name = "lora." + k_tensor + ".lora_down.weight";
-            std::string alpha_name     = "lora." + k_tensor + ".alpha";
-            std::string scale_name     = "lora." + k_tensor + ".scale";
-
-            ggml_tensor* lora_up   = NULL;
-            ggml_tensor* lora_down = NULL;
-
-            if (lora_tensors.find(lora_up_name) != lora_tensors.end()) {
-                lora_up = lora_tensors[lora_up_name];
-            }
-
-            if (lora_tensors.find(lora_down_name) != lora_tensors.end()) {
-                lora_down = lora_tensors[lora_down_name];
-            }
-
-            if (lora_up == NULL || lora_down == NULL) {
-                continue;
-            }
-
-            applied_lora_tensors.insert(lora_up_name);
-            applied_lora_tensors.insert(lora_down_name);
-            applied_lora_tensors.insert(alpha_name);
-            applied_lora_tensors.insert(scale_name);
-
-            // calc_cale
-            int64_t dim       = lora_down->ne[ggml_n_dims(lora_down) - 1];
-            float scale_value = 1.0f;
-            if (lora_tensors.find(scale_name) != lora_tensors.end()) {
-                scale_value = ggml_backend_tensor_get_f32(lora_tensors[scale_name]);
-            } else if (lora_tensors.find(alpha_name) != lora_tensors.end()) {
-                float alpha = ggml_backend_tensor_get_f32(lora_tensors[alpha_name]);
-                scale_value = alpha / dim;
-            }
-            scale_value *= multiplier;
-
-            // flat lora tensors to multiply it
-            int64_t lora_up_rows   = lora_up->ne[ggml_n_dims(lora_up) - 1];
-            lora_up                = ggml_reshape_2d(compute_ctx, lora_up, ggml_nelements(lora_up) / lora_up_rows, lora_up_rows);
-            int64_t lora_down_rows = lora_down->ne[ggml_n_dims(lora_down) - 1];
-            lora_down              = ggml_reshape_2d(compute_ctx, lora_down, ggml_nelements(lora_down) / lora_down_rows, lora_down_rows);
-
-            // ggml_mul_mat requires tensor b transposed
-            lora_down                  = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, lora_down));
-            struct ggml_tensor* updown = ggml_mul_mat(compute_ctx, lora_up, lora_down);
-            updown                     = ggml_cont(compute_ctx, ggml_transpose(compute_ctx, updown));
-            updown                     = ggml_reshape(compute_ctx, updown, weight);
-            GGML_ASSERT(ggml_nelements(updown) == ggml_nelements(weight));
-            updown = ggml_scale_inplace(compute_ctx, updown, scale_value);
-            ggml_tensor* final_weight;
-            if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
-                // final_weight = ggml_new_tensor(compute_ctx, GGML_TYPE_F32, ggml_n_dims(weight), weight->ne);
-                // final_weight = ggml_cpy(compute_ctx, weight, final_weight);
-                final_weight = to_f32(compute_ctx, weight);
-                final_weight = ggml_add_inplace(compute_ctx, final_weight, updown);
-                final_weight = ggml_cpy(compute_ctx, final_weight, weight);
-            } else {
-                final_weight = ggml_add_inplace(compute_ctx, weight, updown);
-            }
-            // final_weight = ggml_add_inplace(compute_ctx, weight, updown);  // apply directly
-            ggml_build_forward_expand(gf, final_weight);
         }
-
         size_t total_lora_tensors_count   = 0;
         size_t applied_lora_tensors_count = 0;
 
         for (auto& kv : lora_tensors) {
             total_lora_tensors_count++;
             if (applied_lora_tensors.find(kv.first) == applied_lora_tensors.end()) {
-                LOG_WARN("unused lora tensor %s", kv.first.c_str());
+                LOG_WARN("unused lora tensor |%s|", kv.first.c_str());
+                print_ggml_tensor(kv.second, true);
+                // exit(0);
             } else {
                 applied_lora_tensors_count++;
             }
@@ -192,9 +834,9 @@ struct LoraModel : public GGMLRunner {
         return gf;
     }
 
-    void apply(std::map<std::string, struct ggml_tensor*> model_tensors, int n_threads) {
+    void apply(std::map<std::string, struct ggml_tensor*> model_tensors, SDVersion version, int n_threads) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_lora_graph(model_tensors);
+            return build_lora_graph(model_tensors, version);
         };
         GGMLRunner::compute(get_graph, n_threads, true);
     }

mmdit.hpp

@@ -147,8 +147,9 @@ protected:
     int64_t hidden_size;
     float eps;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
+        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
+        params["weight"]     = ggml_new_tensor_1d(ctx, wtype, hidden_size);
     }
 
 public:
@@ -636,7 +637,6 @@ public:
 struct MMDiT : public GGMLBlock {
     // Diffusion model with a Transformer backbone.
 protected:
-    SDVersion version                = VERSION_SD3_2B;
     int64_t input_size               = -1;
     int64_t patch_size               = 2;
     int64_t in_channels              = 16;
@@ -652,13 +652,13 @@ protected:
     int64_t hidden_size;
     std::string qk_norm;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["pos_embed"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, hidden_size, num_patchs, 1);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, std::string prefix = "") {
+        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "pos_embed") != tensor_types.end()) ? tensor_types[prefix + "pos_embed"] : GGML_TYPE_F32;
+        params["pos_embed"]  = ggml_new_tensor_3d(ctx, wtype, hidden_size, num_patchs, 1);
     }
 
 public:
-    MMDiT(SDVersion version = VERSION_SD3_2B)
-        : version(version) {
+    MMDiT(std::map<std::string, enum ggml_type>& tensor_types) {
         // input_size is always None
         // learn_sigma is always False
         // register_length is alwalys 0
@@ -670,48 +670,44 @@ public:
         // pos_embed_scaling_factor is not used
         // pos_embed_offset is not used
         // context_embedder_config is always {'target': 'torch.nn.Linear', 'params': {'in_features': 4096, 'out_features': 1536}}
-        if (version == VERSION_SD3_2B) {
-            input_size               = -1;
-            patch_size               = 2;
-            in_channels              = 16;
-            depth                    = 24;
-            mlp_ratio                = 4.0f;
-            adm_in_channels          = 2048;
-            out_channels             = 16;
-            pos_embed_max_size       = 192;
-            num_patchs               = 36864;  // 192 * 192
-            context_size             = 4096;
-            context_embedder_out_dim = 1536;
-        } else if (version == VERSION_SD3_5_8B) {
-            input_size               = -1;
-            patch_size               = 2;
-            in_channels              = 16;
-            depth                    = 38;
-            mlp_ratio                = 4.0f;
-            adm_in_channels          = 2048;
-            out_channels             = 16;
-            pos_embed_max_size       = 192;
-            num_patchs               = 36864;  // 192 * 192
-            context_size             = 4096;
-            context_embedder_out_dim = 2432;
-            qk_norm                  = "rms";
-        } else if (version == VERSION_SD3_5_2B) {
-            input_size               = -1;
-            patch_size               = 2;
-            in_channels              = 16;
-            depth                    = 24;
-            d_self                   = 12;
-            mlp_ratio                = 4.0f;
-            adm_in_channels          = 2048;
-            out_channels             = 16;
-            pos_embed_max_size       = 384;
-            num_patchs               = 147456;
-            context_size             = 4096;
-            context_embedder_out_dim = 1536;
-            qk_norm                  = "rms";
+
+        // read tensors from tensor_types
+        for (auto pair : tensor_types) {
+            std::string tensor_name = pair.first;
+            if (tensor_name.find("model.diffusion_model.") == std::string::npos)
+                continue;
+            size_t jb = tensor_name.find("joint_blocks.");
+            if (jb != std::string::npos) {
+                tensor_name     = tensor_name.substr(jb);  // remove prefix
+                int block_depth = atoi(tensor_name.substr(13, tensor_name.find(".", 13)).c_str());
+                if (block_depth + 1 > depth) {
+                    depth = block_depth + 1;
+                }
+                if (tensor_name.find("attn.ln") != std::string::npos) {
+                    if (tensor_name.find(".bias") != std::string::npos) {
+                        qk_norm = "ln";
+                    } else {
+                        qk_norm = "rms";
+                    }
+                }
+                if (tensor_name.find("attn2") != std::string::npos) {
+                    if (block_depth > d_self) {
+                        d_self = block_depth;
+                    }
+                }
+            }
         }
+
+        if (d_self >= 0) {
+            pos_embed_max_size *= 2;
+            num_patchs *= 4;
+        }
+
+        LOG_INFO("MMDiT layers: %d (including %d MMDiT-x layers)", depth, d_self + 1);
+
         int64_t default_out_channels = in_channels;
         hidden_size                  = 64 * depth;
+        context_embedder_out_dim     = 64 * depth;
         int64_t num_heads            = depth;
 
         blocks["x_embedder"] = std::shared_ptr<GGMLBlock>(new PatchEmbed(input_size, patch_size, in_channels, hidden_size, true));
@@ -870,15 +866,16 @@ public:
         return x;
     }
 };
-
 struct MMDiTRunner : public GGMLRunner {
     MMDiT mmdit;
 
+    static std::map<std::string, enum ggml_type> empty_tensor_types;
+
     MMDiTRunner(ggml_backend_t backend,
-                ggml_type wtype,
-                SDVersion version = VERSION_SD3_2B)
-        : GGMLRunner(backend, wtype), mmdit(version) {
-        mmdit.init(params_ctx, wtype);
+                std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
+                const std::string prefix                            = "")
+        : GGMLRunner(backend), mmdit(tensor_types) {
+        mmdit.init(params_ctx, tensor_types, prefix);
     }
 
     std::string get_desc() {
@@ -975,7 +972,7 @@ struct MMDiTRunner : public GGMLRunner {
         // ggml_backend_t backend    = ggml_backend_cuda_init(0);
         ggml_backend_t backend             = ggml_backend_cpu_init();
         ggml_type model_data_type          = GGML_TYPE_F16;
-        std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend, model_data_type));
+        std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend));
         {
             LOG_INFO("loading from '%s'", file_path.c_str());
 

model.cpp

@@ -13,6 +13,7 @@
 
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
+#include "ggml-cpu.h"
 #include "ggml.h"
 
 #include "stable-diffusion.h"
@@ -25,6 +26,10 @@
 #include "ggml-vulkan.h"
 #endif
 
+#ifdef SD_USE_OPENCL
+#include "ggml-opencl.h"
+#endif
+
 #define ST_HEADER_SIZE_LEN 8
 
 uint64_t read_u64(uint8_t* buffer) {
@@ -146,6 +151,33 @@ std::unordered_map<std::string, std::string> vae_decoder_name_map = {
     {"first_stage_model.decoder.mid.attn_1.to_v.weight", "first_stage_model.decoder.mid.attn_1.v.weight"},
 };
 
+std::unordered_map<std::string, std::string> pmid_v2_name_map = {
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.0.1.1.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.0.1.1.fc1.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.0.1.3.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.0.1.1.fc2.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.1.1.1.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.1.1.1.fc1.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.1.1.3.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.1.1.1.fc2.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.2.1.1.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.2.1.1.fc1.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.2.1.3.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.2.1.1.fc2.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.3.1.1.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.3.1.1.fc1.weight"},
+    {"pmid.qformer_perceiver.perceiver_resampler.layers.3.1.3.weight",
+     "pmid.qformer_perceiver.perceiver_resampler.layers.3.1.1.fc2.weight"},
+    {"pmid.qformer_perceiver.token_proj.0.bias",
+     "pmid.qformer_perceiver.token_proj.fc1.bias"},
+    {"pmid.qformer_perceiver.token_proj.2.bias",
+     "pmid.qformer_perceiver.token_proj.fc2.bias"},
+    {"pmid.qformer_perceiver.token_proj.0.weight",
+     "pmid.qformer_perceiver.token_proj.fc1.weight"},
+    {"pmid.qformer_perceiver.token_proj.2.weight",
+     "pmid.qformer_perceiver.token_proj.fc2.weight"},
+};
+
 std::string convert_open_clip_to_hf_clip(const std::string& name) {
     std::string new_name = name;
     std::string prefix;
@@ -212,6 +244,13 @@ std::string convert_vae_decoder_name(const std::string& name) {
     return name;
 }
 
+std::string convert_pmid_v2_name(const std::string& name) {
+    if (pmid_v2_name_map.find(name) != pmid_v2_name_map.end()) {
+        return pmid_v2_name_map[name];
+    }
+    return name;
+}
+
 /* If not a SDXL LoRA the unet" prefix will have already been replaced by this
  * point and "te2" and "te1" don't seem to appear in non-SDXL only "te_" */
 std::string convert_sdxl_lora_name(std::string tensor_name) {
@@ -443,6 +482,8 @@ std::string convert_tensor_name(std::string name) {
         new_name = convert_open_clip_to_hf_clip(name);
     } else if (starts_with(name, "first_stage_model.decoder")) {
         new_name = convert_vae_decoder_name(name);
+    } else if (starts_with(name, "pmid.qformer_perceiver")) {
+        new_name = convert_pmid_v2_name(name);
     } else if (starts_with(name, "control_model.")) {  // for controlnet pth models
         size_t pos = name.find('.');
         if (pos != std::string::npos) {
@@ -521,6 +562,26 @@ std::string convert_tensor_name(std::string name) {
     return new_name;
 }
 
+void add_preprocess_tensor_storage_types(std::map<std::string, enum ggml_type>& tensor_storages_types, std::string name, enum ggml_type type) {
+    std::string new_name = convert_tensor_name(name);
+
+    if (new_name.find("cond_stage_model") != std::string::npos && ends_with(new_name, "attn.in_proj_weight")) {
+        size_t prefix_size                                        = new_name.find("attn.in_proj_weight");
+        std::string prefix                                        = new_name.substr(0, prefix_size);
+        tensor_storages_types[prefix + "self_attn.q_proj.weight"] = type;
+        tensor_storages_types[prefix + "self_attn.k_proj.weight"] = type;
+        tensor_storages_types[prefix + "self_attn.v_proj.weight"] = type;
+    } else if (new_name.find("cond_stage_model") != std::string::npos && ends_with(new_name, "attn.in_proj_bias")) {
+        size_t prefix_size                                      = new_name.find("attn.in_proj_bias");
+        std::string prefix                                      = new_name.substr(0, prefix_size);
+        tensor_storages_types[prefix + "self_attn.q_proj.bias"] = type;
+        tensor_storages_types[prefix + "self_attn.k_proj.bias"] = type;
+        tensor_storages_types[prefix + "self_attn.v_proj.bias"] = type;
+    } else {
+        tensor_storages_types[new_name] = type;
+    }
+}
+
 void preprocess_tensor(TensorStorage tensor_storage,
                        std::vector<TensorStorage>& processed_tensor_storages) {
     std::vector<TensorStorage> result;
@@ -614,6 +675,47 @@ uint16_t f8_e4m3_to_f16(uint8_t f8) {
     return ggml_fp32_to_fp16(*reinterpret_cast<const float*>(&result));
 }
 
+uint16_t f8_e5m2_to_f16(uint8_t fp8) {
+    uint8_t sign     = (fp8 >> 7) & 0x1;
+    uint8_t exponent = (fp8 >> 2) & 0x1F;
+    uint8_t mantissa = fp8 & 0x3;
+
+    uint16_t fp16_sign = sign << 15;
+    uint16_t fp16_exponent;
+    uint16_t fp16_mantissa;
+
+    if (exponent == 0 && mantissa == 0) {  // zero
+        return fp16_sign;
+    }
+
+    if (exponent == 0x1F) {  // NAN and INF
+        fp16_exponent = 0x1F;
+        fp16_mantissa = mantissa ? (mantissa << 8) : 0;
+        return fp16_sign | (fp16_exponent << 10) | fp16_mantissa;
+    }
+
+    if (exponent == 0) {  // subnormal numbers
+        fp16_exponent = 0;
+        fp16_mantissa = (mantissa << 8);
+        return fp16_sign | fp16_mantissa;
+    }
+
+    // normal numbers
+    int16_t true_exponent = (int16_t)exponent - 15 + 15;
+    if (true_exponent <= 0) {
+        fp16_exponent = 0;
+        fp16_mantissa = (mantissa << 8);
+    } else if (true_exponent >= 0x1F) {
+        fp16_exponent = 0x1F;
+        fp16_mantissa = 0;
+    } else {
+        fp16_exponent = (uint16_t)true_exponent;
+        fp16_mantissa = mantissa << 8;
+    }
+
+    return fp16_sign | (fp16_exponent << 10) | fp16_mantissa;
+}
+
 void bf16_to_f32_vec(uint16_t* src, float* dst, int64_t n) {
     // support inplace op
     for (int64_t i = n - 1; i >= 0; i--) {
@@ -627,6 +729,12 @@ void f8_e4m3_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
         dst[i] = f8_e4m3_to_f16(src[i]);
     }
 }
+void f8_e5m2_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
+    // support inplace op
+    for (int64_t i = n - 1; i >= 0; i--) {
+        dst[i] = f8_e5m2_to_f16(src[i]);
+    }
+}
 
 void convert_tensor(void* src,
                     ggml_type src_type,
@@ -650,25 +758,25 @@ void convert_tensor(void* src,
         if (src_type == GGML_TYPE_F16) {
             ggml_fp16_to_fp32_row((ggml_fp16_t*)src, (float*)dst, n);
         } else {
-            auto qtype = ggml_internal_get_type_traits(src_type);
-            if (qtype.to_float == NULL) {
+            auto qtype = ggml_get_type_traits(src_type);
+            if (qtype->to_float == NULL) {
                 throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available",
                                                 ggml_type_name(src_type)));
             }
-            qtype.to_float(src, (float*)dst, n);
+            qtype->to_float(src, (float*)dst, n);
         }
     } else {
         // src_type == GGML_TYPE_F16 => dst_type is quantized
         // src_type is quantized => dst_type == GGML_TYPE_F16 or dst_type is quantized
-        auto qtype = ggml_internal_get_type_traits(src_type);
-        if (qtype.to_float == NULL) {
+        auto qtype = ggml_get_type_traits(src_type);
+        if (qtype->to_float == NULL) {
             throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available",
                                             ggml_type_name(src_type)));
         }
         std::vector<char> buf;
         buf.resize(sizeof(float) * n);
         char* src_data_f32 = buf.data();
-        qtype.to_float(src, (float*)src_data_f32, n);
+        qtype->to_float(src, (float*)src_data_f32, n);
         if (dst_type == GGML_TYPE_F16) {
             ggml_fp32_to_fp16_row((float*)src_data_f32, (ggml_fp16_t*)dst, n);
         } else {
@@ -843,6 +951,7 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
         GGML_ASSERT(ggml_nbytes(dummy) == tensor_storage.nbytes());
 
         tensor_storages.push_back(tensor_storage);
+        add_preprocess_tensor_storage_types(tensor_storages_types, tensor_storage.name, tensor_storage.type);
     }
 
     gguf_free(ctx_gguf_);
@@ -863,6 +972,8 @@ ggml_type str_to_ggml_type(const std::string& dtype) {
         ttype = GGML_TYPE_F32;
     } else if (dtype == "F8_E4M3") {
         ttype = GGML_TYPE_F16;
+    } else if (dtype == "F8_E5M2") {
+        ttype = GGML_TYPE_F16;
     }
     return ttype;
 }
@@ -976,11 +1087,16 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
             tensor_storage.is_f8_e4m3 = true;
             // f8 -> f16
             GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
+        } else if (dtype == "F8_E5M2") {
+            tensor_storage.is_f8_e5m2 = true;
+            // f8 -> f16
+            GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
         } else {
             GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size);
         }
 
         tensor_storages.push_back(tensor_storage);
+        add_preprocess_tensor_storage_types(tensor_storages_types, tensor_storage.name, tensor_storage.type);
 
         // LOG_DEBUG("%s %s", tensor_storage.to_string().c_str(), dtype.c_str());
     }
@@ -1206,7 +1322,7 @@ bool ModelLoader::parse_data_pkl(uint8_t* buffer,
                                  zip_t* zip,
                                  std::string dir,
                                  size_t file_index,
-                                 const std::string& prefix) {
+                                 const std::string prefix) {
     uint8_t* buffer_end = buffer + buffer_size;
     if (buffer[0] == 0x80) {  // proto
         if (buffer[1] != 2) {
@@ -1308,9 +1424,11 @@ bool ModelLoader::parse_data_pkl(uint8_t* buffer,
                         reader.tensor_storage.reverse_ne();
                         reader.tensor_storage.file_index = file_index;
                         // if(strcmp(prefix.c_str(), "scarlett") == 0)
-                        // printf(" got tensor %s \n ", reader.tensor_storage.name.c_str());
+                        // printf(" ZIP got tensor %s \n ", reader.tensor_storage.name.c_str());
                         reader.tensor_storage.name = prefix + reader.tensor_storage.name;
                         tensor_storages.push_back(reader.tensor_storage);
+                        add_preprocess_tensor_storage_types(tensor_storages_types, reader.tensor_storage.name, reader.tensor_storage.type);
+
                         // LOG_DEBUG("%s", reader.tensor_storage.name.c_str());
                         // reset
                         reader = PickleTensorReader();
@@ -1345,7 +1463,8 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
             size_t pos       = name.find("data.pkl");
             if (pos != std::string::npos) {
                 std::string dir = name.substr(0, pos);
-                void* pkl_data  = NULL;
+                printf("ZIP %d, name = %s, dir = %s \n", i, name.c_str(), dir.c_str());
+                void* pkl_data = NULL;
                 size_t pkl_size;
                 zip_entry_read(zip, &pkl_data, &pkl_size);
 
@@ -1363,40 +1482,49 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
 }
 
 SDVersion ModelLoader::get_sd_version() {
-    TensorStorage token_embedding_weight;
-    bool is_flux    = false;
-    bool is_schnell = true;
-    bool is_lite    = true;
-    bool is_sd3     = false;
-    for (auto& tensor_storage : tensor_storages) {
-        if (tensor_storage.name.find("model.diffusion_model.guidance_in.in_layer.weight") != std::string::npos) {
-            is_schnell = false;
-        }
-        if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
-            is_flux = true;
-        }
-        if (tensor_storage.name.find("model.diffusion_model.double_blocks.8") != std::string::npos) {
-            is_lite = false;
-        }
-        if (tensor_storage.name.find("joint_blocks.0.x_block.attn2.ln_q.weight") != std::string::npos) {
-            return VERSION_SD3_5_2B;
-        }
-        if (tensor_storage.name.find("joint_blocks.37.x_block.attn.ln_q.weight") != std::string::npos) {
-            return VERSION_SD3_5_8B;
-        }
-        if (tensor_storage.name.find("model.diffusion_model.joint_blocks.23.") != std::string::npos) {
-            is_sd3 = true;
-        }
-        if (tensor_storage.name.find("conditioner.embedders.1") != std::string::npos) {
-            return VERSION_SDXL;
-        }
-        if (tensor_storage.name.find("cond_stage_model.1") != std::string::npos) {
-            return VERSION_SDXL;
-        }
-        if (tensor_storage.name.find("model.diffusion_model.input_blocks.8.0.time_mixer.mix_factor") != std::string::npos) {
-            return VERSION_SVD;
-        }
+    TensorStorage token_embedding_weight, input_block_weight;
+    bool input_block_checked = false;
 
+    bool has_multiple_encoders = false;
+    bool is_unet               = false;
+
+    bool is_xl   = false;
+    bool is_flux = false;
+
+#define found_family (is_xl || is_flux)
+    for (auto& tensor_storage : tensor_storages) {
+        if (!found_family) {
+            if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
+                is_flux = true;
+                if (input_block_checked) {
+                    break;
+                }
+            }
+            if (tensor_storage.name.find("model.diffusion_model.joint_blocks.") != std::string::npos) {
+                return VERSION_SD3;
+            }
+            if (tensor_storage.name.find("model.diffusion_model.input_blocks.") != std::string::npos) {
+                is_unet = true;
+                if (has_multiple_encoders) {
+                    is_xl = true;
+                    if (input_block_checked) {
+                        break;
+                    }
+                }
+            }
+            if (tensor_storage.name.find("conditioner.embedders.1") != std::string::npos || tensor_storage.name.find("cond_stage_model.1") != std::string::npos) {
+                has_multiple_encoders = true;
+                if (is_unet) {
+                    is_xl = true;
+                    if (input_block_checked) {
+                        break;
+                    }
+                }
+            }
+            if (tensor_storage.name.find("model.diffusion_model.input_blocks.8.0.time_mixer.mix_factor") != std::string::npos) {
+                return VERSION_SVD;
+            }
+        }
         if (tensor_storage.name == "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight" ||
             tensor_storage.name == "cond_stage_model.model.token_embedding.weight" ||
             tensor_storage.name == "text_model.embeddings.token_embedding.weight" ||
@@ -1406,23 +1534,39 @@ SDVersion ModelLoader::get_sd_version() {
             token_embedding_weight = tensor_storage;
             // break;
         }
-    }
-    if (is_flux) {
-        if (is_schnell) {
-            GGML_ASSERT(!is_lite);
-            return VERSION_FLUX_SCHNELL;
-        } else if (is_lite) {
-            return VERSION_FLUX_LITE;
-        } else {
-            return VERSION_FLUX_DEV;
+        if (tensor_storage.name == "model.diffusion_model.input_blocks.0.0.weight" || tensor_storage.name == "model.diffusion_model.img_in.weight") {
+            input_block_weight  = tensor_storage;
+            input_block_checked = true;
+            if (found_family) {
+                break;
+            }
         }
     }
-    if (is_sd3) {
-        return VERSION_SD3_2B;
+    bool is_inpaint = input_block_weight.ne[2] == 9;
+    if (is_xl) {
+        if (is_inpaint) {
+            return VERSION_SDXL_INPAINT;
+        }
+        return VERSION_SDXL;
     }
+
+    if (is_flux) {
+        is_inpaint = input_block_weight.ne[0] == 384;
+        if (is_inpaint) {
+            return VERSION_FLUX_FILL;
+        }
+        return VERSION_FLUX;
+    }
+
     if (token_embedding_weight.ne[0] == 768) {
+        if (is_inpaint) {
+            return VERSION_SD1_INPAINT;
+        }
         return VERSION_SD1;
     } else if (token_embedding_weight.ne[0] == 1024) {
+        if (is_inpaint) {
+            return VERSION_SD2_INPAINT;
+        }
         return VERSION_SD2;
     }
     return VERSION_COUNT;
@@ -1512,6 +1656,30 @@ ggml_type ModelLoader::get_vae_wtype() {
     return GGML_TYPE_COUNT;
 }
 
+void ModelLoader::set_wtype_override(ggml_type wtype, std::string prefix) {
+    for (auto& pair : tensor_storages_types) {
+        if (prefix.size() < 1 || pair.first.substr(0, prefix.size()) == prefix) {
+            bool found = false;
+            for (auto& tensor_storage : tensor_storages) {
+                std::map<std::string, ggml_type> temp;
+                add_preprocess_tensor_storage_types(temp, tensor_storage.name, tensor_storage.type);
+                for (auto& preprocessed_name : temp) {
+                    if (preprocessed_name.first == pair.first) {
+                        if (tensor_should_be_converted(tensor_storage, wtype)) {
+                            pair.second = wtype;
+                        }
+                        found = true;
+                        break;
+                    }
+                }
+                if (found) {
+                    break;
+                }
+            }
+        }
+    }
+}
+
 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;
@@ -1613,9 +1781,11 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
             }
             return true;
         };
-
+        int tensor_count = 0;
+        int64_t t1       = ggml_time_ms();
         for (auto& tensor_storage : processed_tensor_storages) {
             if (tensor_storage.file_index != file_index) {
+                ++tensor_count;
                 continue;
             }
             ggml_tensor* dst_tensor = NULL;
@@ -1627,6 +1797,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
             }
 
             if (dst_tensor == NULL) {
+                ++tensor_count;
                 continue;
             }
 
@@ -1644,6 +1815,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     } else if (tensor_storage.is_f8_e4m3) {
                         // inplace op
                         f8_e4m3_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
+                    } else if (tensor_storage.is_f8_e5m2) {
+                        // inplace op
+                        f8_e5m2_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
                     }
                 } else {
                     read_buffer.resize(tensor_storage.nbytes());
@@ -1655,6 +1829,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     } else if (tensor_storage.is_f8_e4m3) {
                         // inplace op
                         f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                    } else if (tensor_storage.is_f8_e5m2) {
+                        // inplace op
+                        f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
                     }
 
                     convert_tensor((void*)read_buffer.data(), tensor_storage.type, dst_tensor->data,
@@ -1670,6 +1847,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                 } else if (tensor_storage.is_f8_e4m3) {
                     // inplace op
                     f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                } else if (tensor_storage.is_f8_e5m2) {
+                    // inplace op
+                    f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
                 }
 
                 if (tensor_storage.type == dst_tensor->type) {
@@ -1684,6 +1864,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     ggml_backend_tensor_set(dst_tensor, convert_buffer.data(), 0, ggml_nbytes(dst_tensor));
                 }
             }
+            int64_t t2 = ggml_time_ms();
+            pretty_progress(++tensor_count, processed_tensor_storages.size(), (t2 - t1) / 1000.0f);
+            t1 = t2;
         }
 
         if (zip != NULL) {
@@ -1750,9 +1933,6 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
         if (pair.first.find("cond_stage_model.transformer.text_model.encoder.layers.23") != std::string::npos) {
             continue;
         }
-        if (pair.first.find("alphas_cumprod") != std::string::npos) {
-            continue;
-        }
 
         if (pair.first.find("alphas_cumprod") != std::string::npos) {
             continue;

model.h

@@ -12,6 +12,7 @@
 
 #include "ggml-backend.h"
 #include "ggml.h"
+#include "gguf.h"
 #include "json.hpp"
 #include "zip.h"
 
@@ -19,23 +20,78 @@
 
 enum SDVersion {
     VERSION_SD1,
+    VERSION_SD1_INPAINT,
     VERSION_SD2,
+    VERSION_SD2_INPAINT,
     VERSION_SDXL,
+    VERSION_SDXL_INPAINT,
     VERSION_SVD,
-    VERSION_SD3_2B,
-    VERSION_FLUX_DEV,
-    VERSION_FLUX_SCHNELL,
-    VERSION_SD3_5_8B,
-    VERSION_SD3_5_2B,
-    VERSION_FLUX_LITE,
+    VERSION_SD3,
+    VERSION_FLUX,
+    VERSION_FLUX_FILL,
     VERSION_COUNT,
 };
 
+static inline bool sd_version_is_flux(SDVersion version) {
+    if (version == VERSION_FLUX || version == VERSION_FLUX_FILL) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_sd3(SDVersion version) {
+    if (version == VERSION_SD3) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_sd1(SDVersion version) {
+    if (version == VERSION_SD1 || version == VERSION_SD1_INPAINT) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_sd2(SDVersion version) {
+    if (version == VERSION_SD2 || version == VERSION_SD2_INPAINT) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_sdxl(SDVersion version) {
+    if (version == VERSION_SDXL || version == VERSION_SDXL_INPAINT) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_inpaint(SDVersion version) {
+    if (version == VERSION_SD1_INPAINT || version == VERSION_SD2_INPAINT || version == VERSION_SDXL_INPAINT || version == VERSION_FLUX_FILL) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_dit(SDVersion version) {
+    if (sd_version_is_flux(version) || sd_version_is_sd3(version)) {
+        return true;
+    }
+    return false;
+}
+
+enum PMVersion {
+    PM_VERSION_1,
+    PM_VERSION_2,
+};
+
 struct TensorStorage {
     std::string name;
     ggml_type type          = GGML_TYPE_F32;
     bool is_bf16            = false;
     bool is_f8_e4m3         = false;
+    bool is_f8_e5m2         = false;
     int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
     int n_dims              = 0;
 
@@ -65,7 +121,7 @@ struct TensorStorage {
     }
 
     int64_t nbytes_to_read() const {
-        if (is_bf16 || is_f8_e4m3) {
+        if (is_bf16 || is_f8_e4m3 || is_f8_e5m2) {
             return nbytes() / 2;
         } else {
             return nbytes();
@@ -115,6 +171,8 @@ struct TensorStorage {
             type_name = "bf16";
         } else if (is_f8_e4m3) {
             type_name = "f8_e4m3";
+        } else if (is_f8_e5m2) {
+            type_name = "f8_e5m2";
         }
         ss << name << " | " << type_name << " | ";
         ss << n_dims << " [";
@@ -141,7 +199,7 @@ protected:
                         zip_t* zip,
                         std::string dir,
                         size_t file_index,
-                        const std::string& prefix);
+                        const std::string prefix);
 
     bool init_from_gguf_file(const std::string& file_path, const std::string& prefix = "");
     bool init_from_safetensors_file(const std::string& file_path, const std::string& prefix = "");
@@ -149,16 +207,20 @@ protected:
     bool init_from_diffusers_file(const std::string& file_path, const std::string& prefix = "");
 
 public:
+    std::map<std::string, enum ggml_type> tensor_storages_types;
+
     bool init_from_file(const std::string& file_path, const std::string& prefix = "");
     SDVersion get_sd_version();
     ggml_type get_sd_wtype();
     ggml_type get_conditioner_wtype();
     ggml_type get_diffusion_model_wtype();
     ggml_type get_vae_wtype();
+    void set_wtype_override(ggml_type wtype, std::string prefix = "");
     bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend_t backend);
     bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
                       ggml_backend_t backend,
                       std::set<std::string> ignore_tensors = {});
+
     bool save_to_gguf_file(const std::string& file_path, ggml_type type);
     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);

pmid.hpp

@@ -42,6 +42,370 @@ public:
     }
 };
 
+/*
+class QFormerPerceiver(nn.Module):
+    def __init__(self, id_embeddings_dim, cross_attention_dim, num_tokens, embedding_dim=1024, use_residual=True, ratio=4):
+        super().__init__()
+
+        self.num_tokens = num_tokens
+        self.cross_attention_dim = cross_attention_dim
+        self.use_residual = use_residual
+        print(cross_attention_dim*num_tokens)
+        self.token_proj = nn.Sequential(
+            nn.Linear(id_embeddings_dim, id_embeddings_dim*ratio),
+            nn.GELU(),
+            nn.Linear(id_embeddings_dim*ratio, cross_attention_dim*num_tokens),
+        )
+        self.token_norm = nn.LayerNorm(cross_attention_dim)
+        self.perceiver_resampler = FacePerceiverResampler(
+            dim=cross_attention_dim,
+            depth=4,
+            dim_head=128,
+            heads=cross_attention_dim // 128,
+            embedding_dim=embedding_dim,
+            output_dim=cross_attention_dim,
+            ff_mult=4,
+        )
+
+    def forward(self, x, last_hidden_state):
+        x = self.token_proj(x)
+        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
+        x = self.token_norm(x) # cls token
+        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
+        if self.use_residual: # TODO: if use_residual is not true
+            out = x + 1.0 * out
+        return out
+*/
+
+struct PMFeedForward : public GGMLBlock {
+    // network hparams
+    int dim;
+
+public:
+    PMFeedForward(int d, int multi = 4)
+        : dim(d) {
+        int inner_dim = dim * multi;
+        blocks["0"]   = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["1"]   = std::shared_ptr<GGMLBlock>(new Mlp(dim, inner_dim, dim, false));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x) {
+        auto norm = std::dynamic_pointer_cast<LayerNorm>(blocks["0"]);
+        auto ff   = std::dynamic_pointer_cast<Mlp>(blocks["1"]);
+
+        x = norm->forward(ctx, x);
+        x = ff->forward(ctx, x);
+        return x;
+    }
+};
+
+struct PerceiverAttention : public GGMLBlock {
+    // network hparams
+    float scale;   // = dim_head**-0.5
+    int dim_head;  // = dim_head
+    int heads;     // = heads
+public:
+    PerceiverAttention(int dim, int dim_h = 64, int h = 8)
+        : scale(powf(dim_h, -0.5)), dim_head(dim_h), heads(h) {
+        int inner_dim    = dim_head * heads;
+        blocks["norm1"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["norm2"]  = std::shared_ptr<GGMLBlock>(new LayerNorm(dim));
+        blocks["to_q"]   = std::shared_ptr<GGMLBlock>(new Linear(dim, inner_dim, false));
+        blocks["to_kv"]  = std::shared_ptr<GGMLBlock>(new Linear(dim, inner_dim * 2, false));
+        blocks["to_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim, false));
+    }
+
+    struct ggml_tensor* reshape_tensor(struct ggml_context* ctx,
+                                       struct ggml_tensor* x,
+                                       int heads) {
+        int64_t ne[4];
+        for (int i = 0; i < 4; ++i)
+            ne[i] = x->ne[i];
+        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 0: ");
+        // printf("heads = %d \n", heads);
+        // x = ggml_view_4d(ctx, x, x->ne[0], x->ne[1], heads, x->ne[2]/heads,
+        //                          x->nb[1], x->nb[2], x->nb[3], 0);
+        x = ggml_reshape_4d(ctx, x, x->ne[0] / heads, heads, x->ne[1], x->ne[2]);
+        // x = ggml_view_4d(ctx, x, x->ne[0]/heads, heads, x->ne[1], x->ne[2],
+        //                          x->nb[1], x->nb[2], x->nb[3], 0);
+        // x = ggml_cont(ctx, x);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));
+        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 1: ");
+        // x  = ggml_reshape_4d(ctx, x, ne[0], heads, ne[1], ne[2]/heads);
+        return x;
+    }
+
+    std::vector<struct ggml_tensor*> chunk_half(struct ggml_context* ctx,
+                                                struct ggml_tensor* x) {
+        auto tlo = ggml_view_4d(ctx, x, x->ne[0] / 2, x->ne[1], x->ne[2], x->ne[3], x->nb[1], x->nb[2], x->nb[3], 0);
+        auto tli = ggml_view_4d(ctx, x, x->ne[0] / 2, x->ne[1], x->ne[2], x->ne[3], x->nb[1], x->nb[2], x->nb[3], x->nb[0] * x->ne[0] / 2);
+        return {ggml_cont(ctx, tlo),
+                ggml_cont(ctx, tli)};
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* latents) {
+        // x (torch.Tensor): image features
+        //     shape (b, n1, D)
+        // latent (torch.Tensor): latent features
+        //     shape (b, n2, D)
+        int64_t ne[4];
+        for (int i = 0; i < 4; ++i)
+            ne[i] = latents->ne[i];
+
+        auto norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
+        auto norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
+        x          = norm1->forward(ctx, x);
+        latents    = norm2->forward(ctx, latents);
+        auto to_q  = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+        auto q     = to_q->forward(ctx, latents);
+
+        auto kv_input = ggml_concat(ctx, x, latents, 1);
+        auto to_kv    = std::dynamic_pointer_cast<Linear>(blocks["to_kv"]);
+        auto kv       = to_kv->forward(ctx, kv_input);
+        auto k        = ggml_view_4d(ctx, kv, kv->ne[0] / 2, kv->ne[1], kv->ne[2], kv->ne[3], kv->nb[1] / 2, kv->nb[2] / 2, kv->nb[3] / 2, 0);
+        auto v        = ggml_view_4d(ctx, kv, kv->ne[0] / 2, kv->ne[1], kv->ne[2], kv->ne[3], kv->nb[1] / 2, kv->nb[2] / 2, kv->nb[3] / 2, kv->nb[0] * (kv->ne[0] / 2));
+        k             = ggml_cont(ctx, k);
+        v             = ggml_cont(ctx, v);
+        q             = reshape_tensor(ctx, q, heads);
+        k             = reshape_tensor(ctx, k, heads);
+        v             = reshape_tensor(ctx, v, heads);
+        scale         = 1.f / sqrt(sqrt((float)dim_head));
+        k             = ggml_scale_inplace(ctx, k, scale);
+        q             = ggml_scale_inplace(ctx, q, scale);
+        // auto weight = ggml_mul_mat(ctx, q, k);
+        auto weight = ggml_mul_mat(ctx, k, q);  // NOTE order of mul is opposite to pytorch
+
+        // GGML's softmax() is equivalent to pytorch's softmax(x, dim=-1)
+        // in this case, dimension along which Softmax will be computed is the last dim
+        // in torch and the first dim in GGML, consistent with the convention that pytorch's
+        // last dimension (varying most rapidly) corresponds to GGML's first (varying most rapidly).
+        // weight = ggml_soft_max(ctx, weight);
+        weight = ggml_soft_max_inplace(ctx, weight);
+        v      = ggml_cont(ctx, ggml_transpose(ctx, v));
+        // auto out = ggml_mul_mat(ctx, weight, v);
+        auto out    = ggml_mul_mat(ctx, v, weight);  // NOTE order of mul is opposite to pytorch
+        out         = ggml_cont(ctx, ggml_permute(ctx, out, 0, 2, 1, 3));
+        out         = ggml_reshape_3d(ctx, out, ne[0], ne[1], ggml_nelements(out) / (ne[0] * ne[1]));
+        auto to_out = std::dynamic_pointer_cast<Linear>(blocks["to_out"]);
+        out         = to_out->forward(ctx, out);
+        return out;
+    }
+};
+
+struct FacePerceiverResampler : public GGMLBlock {
+    // network hparams
+    int depth;
+
+public:
+    FacePerceiverResampler(int dim           = 768,
+                           int d             = 4,
+                           int dim_head      = 64,
+                           int heads         = 16,
+                           int embedding_dim = 1280,
+                           int output_dim    = 768,
+                           int ff_mult       = 4)
+        : depth(d) {
+        blocks["proj_in"]  = std::shared_ptr<GGMLBlock>(new Linear(embedding_dim, dim, true));
+        blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(dim, output_dim, true));
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new LayerNorm(output_dim));
+
+        for (int i = 0; i < depth; i++) {
+            std::string name = "layers." + std::to_string(i) + ".0";
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new PerceiverAttention(dim, dim_head, heads));
+            name             = "layers." + std::to_string(i) + ".1";
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new PMFeedForward(dim, ff_mult));
+        }
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* latents,
+                                struct ggml_tensor* x) {
+        // x: [N, channels, h, w]
+        auto proj_in  = std::dynamic_pointer_cast<Linear>(blocks["proj_in"]);
+        auto proj_out = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
+        auto norm_out = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_out"]);
+
+        x = proj_in->forward(ctx, x);
+        for (int i = 0; i < depth; i++) {
+            std::string name = "layers." + std::to_string(i) + ".0";
+            auto attn        = std::dynamic_pointer_cast<PerceiverAttention>(blocks[name]);
+            name             = "layers." + std::to_string(i) + ".1";
+            auto ff          = std::dynamic_pointer_cast<PMFeedForward>(blocks[name]);
+            auto t           = attn->forward(ctx, x, latents);
+            latents          = ggml_add(ctx, t, latents);
+            t                = ff->forward(ctx, latents);
+            latents          = ggml_add(ctx, t, latents);
+        }
+        latents = proj_out->forward(ctx, latents);
+        latents = norm_out->forward(ctx, latents);
+        return latents;
+    }
+};
+
+struct QFormerPerceiver : public GGMLBlock {
+    // network hparams
+    int num_tokens;
+    int cross_attention_dim;
+    bool use_residul;
+
+public:
+    QFormerPerceiver(int id_embeddings_dim, int cross_attention_d, int num_t, int embedding_dim = 1024, bool use_r = true, int ratio = 4)
+        : cross_attention_dim(cross_attention_d), num_tokens(num_t), use_residul(use_r) {
+        blocks["token_proj"]          = std::shared_ptr<GGMLBlock>(new Mlp(id_embeddings_dim,
+                                                                           id_embeddings_dim * ratio,
+                                                                           cross_attention_dim * num_tokens,
+                                                                           true));
+        blocks["token_norm"]          = std::shared_ptr<GGMLBlock>(new LayerNorm(cross_attention_d));
+        blocks["perceiver_resampler"] = std::shared_ptr<GGMLBlock>(new FacePerceiverResampler(
+            cross_attention_dim,
+            4,
+            128,
+            cross_attention_dim / 128,
+            embedding_dim,
+            cross_attention_dim,
+            4));
+    }
+
+    /*
+    def forward(self, x, last_hidden_state):
+        x = self.token_proj(x)
+        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
+        x = self.token_norm(x) # cls token
+        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
+        if self.use_residual: # TODO: if use_residual is not true
+            out = x + 1.0 * out
+        return out
+    */
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* x,
+                                struct ggml_tensor* last_hidden_state) {
+        // x: [N, channels, h, w]
+        auto token_proj          = std::dynamic_pointer_cast<Mlp>(blocks["token_proj"]);
+        auto token_norm          = std::dynamic_pointer_cast<LayerNorm>(blocks["token_norm"]);
+        auto perceiver_resampler = std::dynamic_pointer_cast<FacePerceiverResampler>(blocks["perceiver_resampler"]);
+
+        x                       = token_proj->forward(ctx, x);
+        int64_t nel             = ggml_nelements(x);
+        x                       = ggml_reshape_3d(ctx, x, cross_attention_dim, num_tokens, nel / (cross_attention_dim * num_tokens));
+        x                       = token_norm->forward(ctx, x);
+        struct ggml_tensor* out = perceiver_resampler->forward(ctx, x, last_hidden_state);
+        if (use_residul)
+            out = ggml_add(ctx, x, out);
+        return out;
+    }
+};
+
+/*
+class FacePerceiverResampler(torch.nn.Module):
+    def __init__(
+        self,
+        *,
+        dim=768,
+        depth=4,
+        dim_head=64,
+        heads=16,
+        embedding_dim=1280,
+        output_dim=768,
+        ff_mult=4,
+    ):
+        super().__init__()
+
+        self.proj_in = torch.nn.Linear(embedding_dim, dim)
+        self.proj_out = torch.nn.Linear(dim, output_dim)
+        self.norm_out = torch.nn.LayerNorm(output_dim)
+        self.layers = torch.nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                torch.nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, latents, x):
+        x = self.proj_in(x)
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+*/
+
+/*
+
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+*/
+
 struct FuseModule : public GGMLBlock {
     // network hparams
     int embed_dim;
@@ -61,12 +425,19 @@ public:
         auto mlp2       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp2"]);
         auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm"]);
 
-        auto prompt_embeds0 = ggml_cont(ctx, ggml_permute(ctx, prompt_embeds, 2, 0, 1, 3));
-        auto id_embeds0     = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
-        // concat is along dim 2
-        auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds0, id_embeds0, 2);
-        stacked_id_embeds      = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 1, 2, 0, 3));
+        // print_ggml_tensor(id_embeds, true, "Fuseblock id_embeds: ");
+        // print_ggml_tensor(prompt_embeds, true, "Fuseblock prompt_embeds: ");
 
+        // auto prompt_embeds0 = ggml_cont(ctx, ggml_permute(ctx, prompt_embeds, 2, 0, 1, 3));
+        // auto id_embeds0     = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
+        // print_ggml_tensor(id_embeds0, true, "Fuseblock id_embeds0: ");
+        // print_ggml_tensor(prompt_embeds0, true, "Fuseblock prompt_embeds0: ");
+        // concat is along dim 2
+        // auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds0, id_embeds0, 2);
+        auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds, id_embeds, 0);
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 0: ");
+        // stacked_id_embeds      = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 1, 2, 0, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
         // stacked_id_embeds = mlp1.forward(ctx, stacked_id_embeds);
         // stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
         // stacked_id_embeds = mlp2.forward(ctx, stacked_id_embeds);
@@ -77,6 +448,8 @@ public:
         stacked_id_embeds = mlp2->forward(ctx, stacked_id_embeds);
         stacked_id_embeds = layer_norm->forward(ctx, stacked_id_embeds);
 
+        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
+
         return stacked_id_embeds;
     }
 
@@ -98,23 +471,31 @@ public:
         // print_ggml_tensor(class_tokens_mask_pos, true, "class_tokens_mask_pos");
         struct ggml_tensor* image_token_embeds = ggml_get_rows(ctx, prompt_embeds, class_tokens_mask_pos);
         ggml_set_name(image_token_embeds, "image_token_embeds");
+        valid_id_embeds                       = ggml_reshape_2d(ctx, valid_id_embeds, valid_id_embeds->ne[0],
+                                                                ggml_nelements(valid_id_embeds) / valid_id_embeds->ne[0]);
         struct ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);
 
-        stacked_id_embeds = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // stacked_id_embeds = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "AA stacked_id_embeds");
+        // print_ggml_tensor(left, true, "AA left");
+        // print_ggml_tensor(right, true, "AA right");
         if (left && right) {
-            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 2);
-            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 2);
+            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 1);
+            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
         } else if (left) {
-            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 2);
+            stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 1);
         } else if (right) {
-            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 2);
+            stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
         }
-        stacked_id_embeds                         = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "BB stacked_id_embeds");
+        // stacked_id_embeds                         = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
+        // print_ggml_tensor(stacked_id_embeds, true, "CC stacked_id_embeds");
         class_tokens_mask                         = ggml_cont(ctx, ggml_transpose(ctx, class_tokens_mask));
         class_tokens_mask                         = ggml_repeat(ctx, class_tokens_mask, prompt_embeds);
         prompt_embeds                             = ggml_mul(ctx, prompt_embeds, class_tokens_mask);
         struct ggml_tensor* updated_prompt_embeds = ggml_add(ctx, prompt_embeds, stacked_id_embeds);
         ggml_set_name(updated_prompt_embeds, "updated_prompt_embeds");
+        // print_ggml_tensor(updated_prompt_embeds, true, "updated_prompt_embeds: ");
         return updated_prompt_embeds;
     }
 };
@@ -159,10 +540,77 @@ struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
     }
 };
 
+struct PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock : public CLIPVisionModelProjection {
+    int cross_attention_dim;
+    int num_tokens;
+
+    PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock(int id_embeddings_dim = 512)
+        : CLIPVisionModelProjection(OPENAI_CLIP_VIT_L_14),
+          cross_attention_dim(2048),
+          num_tokens(2) {
+        blocks["visual_projection_2"] = std::shared_ptr<GGMLBlock>(new Linear(1024, 1280, false));
+        blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
+        /*
+        cross_attention_dim = 2048
+        # projection
+        self.num_tokens = 2
+        self.cross_attention_dim = cross_attention_dim
+        self.qformer_perceiver = QFormerPerceiver(
+                                    id_embeddings_dim,
+                                    cross_attention_dim,
+                                    self.num_tokens,
+                                )*/
+        blocks["qformer_perceiver"] = std::shared_ptr<GGMLBlock>(new QFormerPerceiver(id_embeddings_dim,
+                                                                                      cross_attention_dim,
+                                                                                      num_tokens));
+    }
+
+    /*
+    def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds):
+        b, num_inputs, c, h, w = id_pixel_values.shape
+        id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
+
+        last_hidden_state = self.vision_model(id_pixel_values)[0]
+        id_embeds = id_embeds.view(b * num_inputs, -1)
+
+        id_embeds = self.qformer_perceiver(id_embeds, last_hidden_state)
+        id_embeds = id_embeds.view(b, num_inputs, self.num_tokens, -1)
+        updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
+    */
+
+    struct ggml_tensor* forward(struct ggml_context* ctx,
+                                struct ggml_tensor* id_pixel_values,
+                                struct ggml_tensor* prompt_embeds,
+                                struct ggml_tensor* class_tokens_mask,
+                                struct ggml_tensor* class_tokens_mask_pos,
+                                struct ggml_tensor* id_embeds,
+                                struct ggml_tensor* left,
+                                struct ggml_tensor* right) {
+        // x: [N, channels, h, w]
+        auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
+        auto fuse_module       = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);
+        auto qformer_perceiver = std::dynamic_pointer_cast<QFormerPerceiver>(blocks["qformer_perceiver"]);
+
+        // struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
+        struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values, false);  // [N, hidden_size]
+        id_embeds                             = qformer_perceiver->forward(ctx, id_embeds, last_hidden_state);
+
+        struct ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
+                                                                         prompt_embeds,
+                                                                         id_embeds,
+                                                                         class_tokens_mask,
+                                                                         class_tokens_mask_pos,
+                                                                         left, right);
+        return updated_prompt_embeds;
+    }
+};
+
 struct PhotoMakerIDEncoder : public GGMLRunner {
 public:
-    SDVersion version = VERSION_SDXL;
+    SDVersion version    = VERSION_SDXL;
+    PMVersion pm_version = PM_VERSION_1;
     PhotoMakerIDEncoderBlock id_encoder;
+    PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock id_encoder2;
     float style_strength;
 
     std::vector<float> ctm;
@@ -175,25 +623,38 @@ public:
     std::vector<float> zeros_right;
 
 public:
-    PhotoMakerIDEncoder(ggml_backend_t backend, ggml_type wtype, SDVersion version = VERSION_SDXL, float sty = 20.f)
-        : GGMLRunner(backend, wtype),
+    PhotoMakerIDEncoder(ggml_backend_t backend, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix, SDVersion version = VERSION_SDXL, PMVersion pm_v = PM_VERSION_1, float sty = 20.f)
+        : GGMLRunner(backend),
           version(version),
+          pm_version(pm_v),
           style_strength(sty) {
-        id_encoder.init(params_ctx, wtype);
+        if (pm_version == PM_VERSION_1) {
+            id_encoder.init(params_ctx, tensor_types, prefix);
+        } else if (pm_version == PM_VERSION_2) {
+            id_encoder2.init(params_ctx, tensor_types, prefix);
+        }
     }
 
     std::string get_desc() {
         return "pmid";
     }
 
+    PMVersion get_version() const {
+        return pm_version;
+    }
+
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
-        id_encoder.get_param_tensors(tensors, prefix);
+        if (pm_version == PM_VERSION_1)
+            id_encoder.get_param_tensors(tensors, prefix);
+        else if (pm_version == PM_VERSION_2)
+            id_encoder2.get_param_tensors(tensors, prefix);
     }
 
     struct ggml_cgraph* build_graph(  // struct ggml_allocr* allocr,
         struct ggml_tensor* id_pixel_values,
         struct ggml_tensor* prompt_embeds,
-        std::vector<bool>& class_tokens_mask) {
+        std::vector<bool>& class_tokens_mask,
+        struct ggml_tensor* id_embeds) {
         ctm.clear();
         ctmf16.clear();
         ctmpos.clear();
@@ -214,25 +675,32 @@ public:
 
         struct ggml_tensor* id_pixel_values_d = to_backend(id_pixel_values);
         struct ggml_tensor* prompt_embeds_d   = to_backend(prompt_embeds);
+        struct ggml_tensor* id_embeds_d       = to_backend(id_embeds);
 
         struct ggml_tensor* left  = NULL;
         struct ggml_tensor* right = NULL;
         for (int i = 0; i < class_tokens_mask.size(); i++) {
             if (class_tokens_mask[i]) {
+                // printf(" 1,");
                 ctm.push_back(0.f);                        // here use 0.f instead of 1.f to make a scale mask
                 ctmf16.push_back(ggml_fp32_to_fp16(0.f));  // here use 0.f instead of 1.f to make a scale mask
                 ctmpos.push_back(i);
             } else {
+                // printf(" 0,");
                 ctm.push_back(1.f);                        // here use 1.f instead of 0.f to make a scale mask
                 ctmf16.push_back(ggml_fp32_to_fp16(1.f));  // here use 0.f instead of 1.f to make a scale mask
             }
         }
+        // printf("\n");
         if (ctmpos[0] > 0) {
-            left = ggml_new_tensor_3d(ctx0, type, hidden_size, 1, ctmpos[0]);
+            // left = ggml_new_tensor_3d(ctx0, type, hidden_size, 1, ctmpos[0]);
+            left = ggml_new_tensor_3d(ctx0, type, hidden_size, ctmpos[0], 1);
         }
         if (ctmpos[ctmpos.size() - 1] < seq_length - 1) {
+            // right = ggml_new_tensor_3d(ctx0, type,
+            //                            hidden_size, 1, seq_length - ctmpos[ctmpos.size() - 1] - 1);
             right = ggml_new_tensor_3d(ctx0, type,
-                                       hidden_size, 1, seq_length - ctmpos[ctmpos.size() - 1] - 1);
+                                       hidden_size, seq_length - ctmpos[ctmpos.size() - 1] - 1, 1);
         }
         struct ggml_tensor* class_tokens_mask_pos = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ctmpos.size());
 
@@ -265,12 +733,23 @@ public:
                 }
             }
         }
-        struct ggml_tensor* updated_prompt_embeds = id_encoder.forward(ctx0,
-                                                                       id_pixel_values_d,
-                                                                       prompt_embeds_d,
-                                                                       class_tokens_mask_d,
-                                                                       class_tokens_mask_pos,
-                                                                       left, right);
+        struct ggml_tensor* updated_prompt_embeds = NULL;
+        if (pm_version == PM_VERSION_1)
+            updated_prompt_embeds = id_encoder.forward(ctx0,
+                                                       id_pixel_values_d,
+                                                       prompt_embeds_d,
+                                                       class_tokens_mask_d,
+                                                       class_tokens_mask_pos,
+                                                       left, right);
+        else if (pm_version == PM_VERSION_2)
+            updated_prompt_embeds = id_encoder2.forward(ctx0,
+                                                        id_pixel_values_d,
+                                                        prompt_embeds_d,
+                                                        class_tokens_mask_d,
+                                                        class_tokens_mask_pos,
+                                                        id_embeds_d,
+                                                        left, right);
+
         ggml_build_forward_expand(gf, updated_prompt_embeds);
 
         return gf;
@@ -279,12 +758,13 @@ public:
     void compute(const int n_threads,
                  struct ggml_tensor* id_pixel_values,
                  struct ggml_tensor* prompt_embeds,
+                 struct ggml_tensor* id_embeds,
                  std::vector<bool>& class_tokens_mask,
                  struct ggml_tensor** updated_prompt_embeds,
                  ggml_context* output_ctx) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
             // return build_graph(compute_allocr, id_pixel_values, prompt_embeds, class_tokens_mask);
-            return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask);
+            return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds);
         };
 
         // GGMLRunner::compute(get_graph, n_threads, updated_prompt_embeds);
@@ -292,4 +772,74 @@ public:
     }
 };
 
+struct PhotoMakerIDEmbed : public GGMLRunner {
+    std::map<std::string, struct ggml_tensor*> tensors;
+    std::string file_path;
+    ModelLoader* model_loader;
+    bool load_failed = false;
+    bool applied     = false;
+
+    PhotoMakerIDEmbed(ggml_backend_t backend,
+                      ModelLoader* ml,
+                      const std::string& file_path = "",
+                      const std::string& prefix    = "")
+        : file_path(file_path), GGMLRunner(backend), model_loader(ml) {
+        if (!model_loader->init_from_file(file_path, prefix)) {
+            load_failed = true;
+        }
+    }
+
+    std::string get_desc() {
+        return "id_embeds";
+    }
+
+    bool load_from_file(bool filter_tensor = false) {
+        LOG_INFO("loading PhotoMaker ID Embeds from '%s'", file_path.c_str());
+
+        if (load_failed) {
+            LOG_ERROR("init photomaker id embed from file failed: '%s'", file_path.c_str());
+            return false;
+        }
+
+        bool dry_run          = true;
+        auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+            const std::string& name = tensor_storage.name;
+
+            if (filter_tensor && !contains(name, "pmid.id_embeds")) {
+                // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
+                return true;
+            }
+            if (dry_run) {
+                struct ggml_tensor* real = ggml_new_tensor(params_ctx,
+                                                           tensor_storage.type,
+                                                           tensor_storage.n_dims,
+                                                           tensor_storage.ne);
+                tensors[name]            = real;
+            } else {
+                auto real   = tensors[name];
+                *dst_tensor = real;
+            }
+
+            return true;
+        };
+
+        model_loader->load_tensors(on_new_tensor_cb, backend);
+        alloc_params_buffer();
+
+        dry_run = false;
+        model_loader->load_tensors(on_new_tensor_cb, backend);
+
+        LOG_DEBUG("finished loading PhotoMaker ID Embeds ");
+        return true;
+    }
+
+    struct ggml_tensor* get() {
+        std::map<std::string, struct ggml_tensor*>::iterator pos;
+        pos = tensors.find("pmid.id_embeds");
+        if (pos != tensors.end())
+            return pos->second;
+        return NULL;
+    }
+};
+
 #endif  // __PMI_HPP__

stable-diffusion.h

@@ -44,6 +44,8 @@ enum sample_method_t {
     IPNDM,
     IPNDM_V,
     LCM,
+    DDIM_TRAILING,
+    TCD,
     N_SAMPLE_METHODS
 };
 
@@ -65,35 +67,40 @@ enum sd_type_t {
     SD_TYPE_Q4_1 = 3,
     // SD_TYPE_Q4_2 = 4, support has been removed
     // SD_TYPE_Q4_3 = 5, support has been removed
-    SD_TYPE_Q5_0     = 6,
-    SD_TYPE_Q5_1     = 7,
-    SD_TYPE_Q8_0     = 8,
-    SD_TYPE_Q8_1     = 9,
-    SD_TYPE_Q2_K     = 10,
-    SD_TYPE_Q3_K     = 11,
-    SD_TYPE_Q4_K     = 12,
-    SD_TYPE_Q5_K     = 13,
-    SD_TYPE_Q6_K     = 14,
-    SD_TYPE_Q8_K     = 15,
-    SD_TYPE_IQ2_XXS  = 16,
-    SD_TYPE_IQ2_XS   = 17,
-    SD_TYPE_IQ3_XXS  = 18,
-    SD_TYPE_IQ1_S    = 19,
-    SD_TYPE_IQ4_NL   = 20,
-    SD_TYPE_IQ3_S    = 21,
-    SD_TYPE_IQ2_S    = 22,
-    SD_TYPE_IQ4_XS   = 23,
-    SD_TYPE_I8       = 24,
-    SD_TYPE_I16      = 25,
-    SD_TYPE_I32      = 26,
-    SD_TYPE_I64      = 27,
-    SD_TYPE_F64      = 28,
-    SD_TYPE_IQ1_M    = 29,
-    SD_TYPE_BF16     = 30,
-    SD_TYPE_Q4_0_4_4 = 31,
-    SD_TYPE_Q4_0_4_8 = 32,
-    SD_TYPE_Q4_0_8_8 = 33,
-    SD_TYPE_COUNT,
+    SD_TYPE_Q5_0    = 6,
+    SD_TYPE_Q5_1    = 7,
+    SD_TYPE_Q8_0    = 8,
+    SD_TYPE_Q8_1    = 9,
+    SD_TYPE_Q2_K    = 10,
+    SD_TYPE_Q3_K    = 11,
+    SD_TYPE_Q4_K    = 12,
+    SD_TYPE_Q5_K    = 13,
+    SD_TYPE_Q6_K    = 14,
+    SD_TYPE_Q8_K    = 15,
+    SD_TYPE_IQ2_XXS = 16,
+    SD_TYPE_IQ2_XS  = 17,
+    SD_TYPE_IQ3_XXS = 18,
+    SD_TYPE_IQ1_S   = 19,
+    SD_TYPE_IQ4_NL  = 20,
+    SD_TYPE_IQ3_S   = 21,
+    SD_TYPE_IQ2_S   = 22,
+    SD_TYPE_IQ4_XS  = 23,
+    SD_TYPE_I8      = 24,
+    SD_TYPE_I16     = 25,
+    SD_TYPE_I32     = 26,
+    SD_TYPE_I64     = 27,
+    SD_TYPE_F64     = 28,
+    SD_TYPE_IQ1_M   = 29,
+    SD_TYPE_BF16    = 30,
+    // SD_TYPE_Q4_0_4_4 = 31, support has been removed from gguf files
+    // SD_TYPE_Q4_0_4_8 = 32,
+    // SD_TYPE_Q4_0_8_8 = 33,
+    SD_TYPE_TQ1_0 = 34,
+    SD_TYPE_TQ2_0 = 35,
+    // SD_TYPE_IQ4_NL_4_4 = 36,
+    // SD_TYPE_IQ4_NL_4_8 = 37,
+    // SD_TYPE_IQ4_NL_8_8 = 38,
+    SD_TYPE_COUNT = 39,
 };
 
 SD_API const char* sd_type_name(enum sd_type_t type);
@@ -142,7 +149,11 @@ SD_API sd_ctx_t* new_sd_ctx(const char* model_path,
                             enum schedule_t s,
                             bool keep_clip_on_cpu,
                             bool keep_control_net_cpu,
-                            bool keep_vae_on_cpu);
+                            bool keep_vae_on_cpu,
+                            bool diffusion_flash_attn,
+                            bool chroma_use_dit_mask,
+                            bool chroma_use_t5_mask,
+                            int chroma_t5_mask_pad);
 
 SD_API void free_sd_ctx(sd_ctx_t* sd_ctx);
 
@@ -152,6 +163,7 @@ SD_API sd_image_t* txt2img(sd_ctx_t* sd_ctx,
                            int clip_skip,
                            float cfg_scale,
                            float guidance,
+                           float eta,
                            int width,
                            int height,
                            enum sample_method_t sample_method,
@@ -163,18 +175,21 @@ SD_API sd_image_t* txt2img(sd_ctx_t* sd_ctx,
                            float style_strength,
                            bool normalize_input,
                            const char* input_id_images_path,
-                           std::vector<int> skip_layers = {},
-                           float slg_scale              = 2.5,
-                           float skip_layer_start       = 0.01,
-                           float skip_layer_end         = 0.2);
+                           int* skip_layers,
+                           size_t skip_layers_count,
+                           float slg_scale,
+                           float skip_layer_start,
+                           float skip_layer_end);
 
 SD_API sd_image_t* img2img(sd_ctx_t* sd_ctx,
                            sd_image_t init_image,
+                           sd_image_t mask_image,
                            const char* prompt,
                            const char* negative_prompt,
                            int clip_skip,
                            float cfg_scale,
                            float guidance,
+                           float eta,
                            int width,
                            int height,
                            enum sample_method_t sample_method,
@@ -187,10 +202,11 @@ SD_API sd_image_t* img2img(sd_ctx_t* sd_ctx,
                            float style_strength,
                            bool normalize_input,
                            const char* input_id_images_path,
-                           std::vector<int> skip_layers = {},
-                           float slg_scale              = 2.5,
-                           float skip_layer_start       = 0.01,
-                           float skip_layer_end         = 0.2);
+                           int* skip_layers,
+                           size_t skip_layers_count,
+                           float slg_scale,
+                           float skip_layer_start,
+                           float skip_layer_end);
 
 SD_API sd_image_t* img2vid(sd_ctx_t* sd_ctx,
                            sd_image_t init_image,
@@ -207,11 +223,36 @@ SD_API sd_image_t* img2vid(sd_ctx_t* sd_ctx,
                            float strength,
                            int64_t seed);
 
+SD_API sd_image_t* edit(sd_ctx_t* sd_ctx,
+                        sd_image_t* ref_images,
+                        int ref_images_count,
+                        const char* prompt,
+                        const char* negative_prompt,
+                        int clip_skip,
+                        float cfg_scale,
+                        float guidance,
+                        float eta,
+                        int width,
+                        int height,
+                        enum sample_method_t sample_method,
+                        int sample_steps,
+                        float strength,
+                        int64_t seed,
+                        int batch_count,
+                        const sd_image_t* control_cond,
+                        float control_strength,
+                        float style_strength,
+                        bool normalize_input,
+                        int* skip_layers,
+                        size_t skip_layers_count,
+                        float slg_scale,
+                        float skip_layer_start,
+                        float skip_layer_end);
+
 typedef struct upscaler_ctx_t upscaler_ctx_t;
 
 SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
-                                        int n_threads,
-                                        enum sd_type_t wtype);
+                                        int n_threads);
 SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);
 
 SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_t upscale_factor);

t5.hpp

@@ -385,6 +385,7 @@ public:
 
     void pad_tokens(std::vector<int>& tokens,
                     std::vector<float>& weights,
+                    std::vector<float>* attention_mask,
                     size_t max_length = 0,
                     bool padding      = false) {
         if (max_length > 0 && padding) {
@@ -397,11 +398,15 @@ public:
             LOG_DEBUG("token length: %llu", length);
             std::vector<int> new_tokens;
             std::vector<float> new_weights;
+            std::vector<float> new_attention_mask;
             int token_idx = 0;
             for (int i = 0; i < length; i++) {
                 if (token_idx >= orig_token_num) {
                     break;
                 }
+                if (attention_mask != nullptr) {
+                    new_attention_mask.push_back(0.0);
+                }
                 if (i % max_length == max_length - 1) {
                     new_tokens.push_back(eos_id_);
                     new_weights.push_back(1.0);
@@ -414,13 +419,24 @@ public:
 
             new_tokens.push_back(eos_id_);
             new_weights.push_back(1.0);
+            if (attention_mask != nullptr) {
+                new_attention_mask.push_back(0.0);
+            }
+
             tokens  = new_tokens;
             weights = new_weights;
+            if (attention_mask != nullptr) {
+                *attention_mask = new_attention_mask;
+            }
 
             if (padding) {
                 int pad_token_id = pad_id_;
                 tokens.insert(tokens.end(), length - tokens.size(), pad_token_id);
                 weights.insert(weights.end(), length - weights.size(), 1.0);
+                if (attention_mask != nullptr) {
+                    // maybe keep some padding tokens unmasked?
+                    attention_mask->insert(attention_mask->end(), length - attention_mask->size(), -HUGE_VALF);
+                }
             }
         }
     }
@@ -441,8 +457,9 @@ protected:
     int64_t hidden_size;
     float eps;
 
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = GGML_TYPE_F32;  //(tensor_types.find(prefix + "weight") != tensor_types.end()) ? tensor_types[prefix + "weight"] : GGML_TYPE_F32;
+        params["weight"]     = ggml_new_tensor_1d(ctx, wtype, hidden_size);
     }
 
 public:
@@ -578,6 +595,7 @@ public:
         }
         if (past_bias != NULL) {
             if (mask != NULL) {
+                mask = ggml_repeat(ctx, mask, past_bias);
                 mask = ggml_add(ctx, mask, past_bias);
             } else {
                 mask = past_bias;
@@ -717,14 +735,15 @@ struct T5Runner : public GGMLRunner {
     std::vector<int> relative_position_bucket_vec;
 
     T5Runner(ggml_backend_t backend,
-             ggml_type wtype,
+             std::map<std::string, enum ggml_type>& tensor_types,
+             const std::string prefix,
              int64_t num_layers = 24,
              int64_t model_dim  = 4096,
              int64_t ff_dim     = 10240,
              int64_t num_heads  = 64,
              int64_t vocab_size = 32128)
-        : GGMLRunner(backend, wtype), model(num_layers, model_dim, ff_dim, num_heads, vocab_size) {
-        model.init(params_ctx, wtype);
+        : GGMLRunner(backend), model(num_layers, model_dim, ff_dim, num_heads, vocab_size) {
+        model.init(params_ctx, tensor_types, prefix);
     }
 
     std::string get_desc() {
@@ -737,15 +756,17 @@ struct T5Runner : public GGMLRunner {
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
                                 struct ggml_tensor* input_ids,
-                                struct ggml_tensor* relative_position_bucket) {
+                                struct ggml_tensor* relative_position_bucket,
+                                struct ggml_tensor* attention_mask = NULL) {
         size_t N       = input_ids->ne[1];
         size_t n_token = input_ids->ne[0];
 
-        auto hidden_states = model.forward(ctx, input_ids, NULL, NULL, relative_position_bucket);  // [N, n_token, model_dim]
+        auto hidden_states = model.forward(ctx, input_ids, NULL, attention_mask, relative_position_bucket);  // [N, n_token, model_dim]
         return hidden_states;
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids) {
+    struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
+                                    struct ggml_tensor* attention_mask = NULL) {
         struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         input_ids = to_backend(input_ids);
@@ -765,7 +786,7 @@ struct T5Runner : public GGMLRunner {
                                                            input_ids->ne[0]);
         set_backend_tensor_data(relative_position_bucket, relative_position_bucket_vec.data());
 
-        struct ggml_tensor* hidden_states = forward(compute_ctx, input_ids, relative_position_bucket);
+        struct ggml_tensor* hidden_states = forward(compute_ctx, input_ids, relative_position_bucket, attention_mask);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -774,10 +795,11 @@ struct T5Runner : public GGMLRunner {
 
     void compute(const int n_threads,
                  struct ggml_tensor* input_ids,
+                 struct ggml_tensor* attention_mask,
                  ggml_tensor** output,
                  ggml_context* output_ctx = NULL) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(input_ids);
+            return build_graph(input_ids, attention_mask);
         };
         GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
     }
@@ -854,14 +876,17 @@ struct T5Embedder {
     T5UniGramTokenizer tokenizer;
     T5Runner model;
 
+    static std::map<std::string, enum ggml_type> empty_tensor_types;
+
     T5Embedder(ggml_backend_t backend,
-               ggml_type wtype,
-               int64_t num_layers = 24,
-               int64_t model_dim  = 4096,
-               int64_t ff_dim     = 10240,
-               int64_t num_heads  = 64,
-               int64_t vocab_size = 32128)
-        : model(backend, wtype, num_layers, model_dim, ff_dim, num_heads, vocab_size) {
+               std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
+               const std::string prefix                            = "",
+               int64_t num_layers                                  = 24,
+               int64_t model_dim                                   = 4096,
+               int64_t ff_dim                                      = 10240,
+               int64_t num_heads                                   = 64,
+               int64_t vocab_size                                  = 32128)
+        : model(backend, tensor_types, prefix, num_layers, model_dim, ff_dim, num_heads, vocab_size) {
     }
 
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
@@ -872,9 +897,9 @@ struct T5Embedder {
         model.alloc_params_buffer();
     }
 
-    std::pair<std::vector<int>, std::vector<float>> tokenize(std::string text,
-                                                             size_t max_length = 0,
-                                                             bool padding      = false) {
+    std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
+                                                                                  size_t max_length = 0,
+                                                                                  bool padding      = false) {
         auto parsed_attention = parse_prompt_attention(text);
 
         {
@@ -901,14 +926,16 @@ struct T5Embedder {
         tokens.push_back(EOS_TOKEN_ID);
         weights.push_back(1.0);
 
-        tokenizer.pad_tokens(tokens, weights, max_length, padding);
+        std::vector<float> attention_mask;
+
+        tokenizer.pad_tokens(tokens, weights, &attention_mask, max_length, padding);
 
         // for (int i = 0; i < tokens.size(); i++) {
         //     std::cout << tokens[i] << ":" << weights[i] << ", ";
         // }
         // std::cout << std::endl;
 
-        return {tokens, weights};
+        return {tokens, weights, attention_mask};
     }
 
     void test() {
@@ -929,8 +956,8 @@ struct T5Embedder {
             // TODO: fix cuda nan
             std::string text("a lovely cat");
             auto tokens_and_weights     = tokenize(text, 77, true);
-            std::vector<int>& tokens    = tokens_and_weights.first;
-            std::vector<float>& weights = tokens_and_weights.second;
+            std::vector<int>& tokens    = std::get<0>(tokens_and_weights);
+            std::vector<float>& weights = std::get<1>(tokens_and_weights);
             for (auto token : tokens) {
                 printf("%d ", token);
             }
@@ -939,7 +966,7 @@ struct T5Embedder {
             struct ggml_tensor* out = NULL;
 
             int t0 = ggml_time_ms();
-            model.compute(8, input_ids, &out, work_ctx);
+            model.compute(8, input_ids, NULL, &out, work_ctx);
             int t1 = ggml_time_ms();
 
             print_ggml_tensor(out);
@@ -951,7 +978,7 @@ struct T5Embedder {
         // ggml_backend_t backend    = ggml_backend_cuda_init(0);
         ggml_backend_t backend         = ggml_backend_cpu_init();
         ggml_type model_data_type      = GGML_TYPE_F32;
-        std::shared_ptr<T5Embedder> t5 = std::shared_ptr<T5Embedder>(new T5Embedder(backend, model_data_type));
+        std::shared_ptr<T5Embedder> t5 = std::shared_ptr<T5Embedder>(new T5Embedder(backend));
         {
             LOG_INFO("loading from '%s'", file_path.c_str());
 

tae.hpp

@@ -62,7 +62,8 @@ class TinyEncoder : public UnaryBlock {
     int num_blocks  = 3;
 
 public:
-    TinyEncoder() {
+    TinyEncoder(int z_channels = 4)
+        : z_channels(z_channels) {
         int index                       = 0;
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, channels, {3, 3}, {1, 1}, {1, 1}));
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TAEBlock(channels, channels));
@@ -106,7 +107,10 @@ class TinyDecoder : public UnaryBlock {
     int num_blocks   = 3;
 
 public:
-    TinyDecoder(int index = 0) {
+    TinyDecoder(int z_channels = 4)
+        : z_channels(z_channels) {
+        int index = 0;
+
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels, {3, 3}, {1, 1}, {1, 1}));
         index++;  // nn.ReLU()
 
@@ -145,7 +149,7 @@ public:
                 if (i == 1) {
                     h = ggml_relu_inplace(ctx, h);
                 } else {
-                    h = ggml_upscale(ctx, h, 2);
+                    h = ggml_upscale(ctx, h, 2, GGML_SCALE_MODE_NEAREST);
                 }
                 continue;
             }
@@ -163,12 +167,16 @@ protected:
     bool decode_only;
 
 public:
-    TAESD(bool decode_only = true)
+    TAESD(bool decode_only = true, SDVersion version = VERSION_SD1)
         : decode_only(decode_only) {
-        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder());
+        int z_channels = 4;
+        if (sd_version_is_dit(version)) {
+            z_channels = 16;
+        }
+        blocks["decoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyDecoder(z_channels));
 
         if (!decode_only) {
-            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder());
+            blocks["encoder.layers"] = std::shared_ptr<GGMLBlock>(new TinyEncoder(z_channels));
         }
     }
 
@@ -188,12 +196,14 @@ struct TinyAutoEncoder : public GGMLRunner {
     bool decode_only = false;
 
     TinyAutoEncoder(ggml_backend_t backend,
-                    ggml_type wtype,
-                    bool decoder_only = true)
+                    std::map<std::string, enum ggml_type>& tensor_types,
+                    const std::string prefix,
+                    bool decoder_only = true,
+                    SDVersion version = VERSION_SD1)
         : decode_only(decoder_only),
-          taesd(decode_only),
-          GGMLRunner(backend, wtype) {
-        taesd.init(params_ctx, wtype);
+          taesd(decoder_only, version),
+          GGMLRunner(backend) {
+        taesd.init(params_ctx, tensor_types, prefix);
     }
 
     std::string get_desc() {

thirdparty/stb_image_write.h

@@ -177,7 +177,7 @@ STBIWDEF int stbi_write_png(char const *filename, int w, int h, int comp, const
 STBIWDEF int stbi_write_bmp(char const *filename, int w, int h, int comp, const void  *data);
 STBIWDEF int stbi_write_tga(char const *filename, int w, int h, int comp, const void  *data);
 STBIWDEF int stbi_write_hdr(char const *filename, int w, int h, int comp, const float *data);
-STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void  *data, int quality);
+STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void  *data, int quality, const char* parameters = NULL);
 
 #ifdef STBIW_WINDOWS_UTF8
 STBIWDEF int stbiw_convert_wchar_to_utf8(char *buffer, size_t bufferlen, const wchar_t* input);
@@ -1412,7 +1412,7 @@ static int stbiw__jpg_processDU(stbi__write_context *s, int *bitBuf, int *bitCnt
    return DU[0];
 }
 
-static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, int comp, const void* data, int quality) {
+static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, int comp, const void* data, int quality, const char* parameters) {
    // Constants that don't pollute global namespace
    static const unsigned char std_dc_luminance_nrcodes[] = {0,0,1,5,1,1,1,1,1,1,0,0,0,0,0,0,0};
    static const unsigned char std_dc_luminance_values[] = {0,1,2,3,4,5,6,7,8,9,10,11};
@@ -1521,6 +1521,20 @@ static int stbi_write_jpg_core(stbi__write_context *s, int width, int height, in
       s->func(s->context, (void*)YTable, sizeof(YTable));
       stbiw__putc(s, 1);
       s->func(s->context, UVTable, sizeof(UVTable));
+
+      // comment block with parameters of generation
+      if(parameters != NULL) {
+         stbiw__putc(s, 0xFF /* comnent */ );
+         stbiw__putc(s, 0xFE /* marker  */ );
+         size_t param_length = std::min(2 + strlen("parameters") + 1 + strlen(parameters) + 1, (size_t) 0xFFFF);
+         stbiw__putc(s, param_length >> 8); // no need to mask, length < 65536
+         stbiw__putc(s, param_length & 0xFF);
+         s->func(s->context, (void*)"parameters", strlen("parameters") + 1); // std::string is zero-terminated
+         s->func(s->context, (void*)parameters, std::min(param_length, (size_t) 65534) - 2 - strlen("parameters") - 1);
+         if(param_length > 65534) stbiw__putc(s, 0); // always zero-terminate for safety
+         if(param_length & 1) stbiw__putc(s, 0xFF); // pad to even length
+      }
+
       s->func(s->context, (void*)head1, sizeof(head1));
       s->func(s->context, (void*)(std_dc_luminance_nrcodes+1), sizeof(std_dc_luminance_nrcodes)-1);
       s->func(s->context, (void*)std_dc_luminance_values, sizeof(std_dc_luminance_values));
@@ -1625,16 +1639,16 @@ STBIWDEF int stbi_write_jpg_to_func(stbi_write_func *func, void *context, int x,
 {
    stbi__write_context s = { 0 };
    stbi__start_write_callbacks(&s, func, context);
-   return stbi_write_jpg_core(&s, x, y, comp, (void *) data, quality);
+   return stbi_write_jpg_core(&s, x, y, comp, (void *) data, quality, NULL);
 }
 
 
 #ifndef STBI_WRITE_NO_STDIO
-STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void *data, int quality)
+STBIWDEF int stbi_write_jpg(char const *filename, int x, int y, int comp, const void *data, int quality, const char* parameters)
 {
    stbi__write_context s = { 0 };
    if (stbi__start_write_file(&s,filename)) {
-      int r = stbi_write_jpg_core(&s, x, y, comp, data, quality);
+      int r = stbi_write_jpg_core(&s, x, y, comp, data, quality, parameters);
       stbi__end_write_file(&s);
       return r;
    } else

unet.hpp

@@ -166,6 +166,7 @@ public:
 // ldm.modules.diffusionmodules.openaimodel.UNetModel
 class UnetModelBlock : public GGMLBlock {
 protected:
+    static std::map<std::string, enum ggml_type> empty_tensor_types;
     SDVersion version = VERSION_SD1;
     // network hparams
     int in_channels                        = 4;
@@ -183,13 +184,13 @@ public:
     int model_channels  = 320;
     int adm_in_channels = 2816;  // only for VERSION_SDXL/SVD
 
-    UnetModelBlock(SDVersion version = VERSION_SD1)
+    UnetModelBlock(SDVersion version = VERSION_SD1, std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types, bool flash_attn = false)
         : version(version) {
-        if (version == VERSION_SD2) {
+        if (sd_version_is_sd2(version)) {
             context_dim       = 1024;
             num_head_channels = 64;
             num_heads         = -1;
-        } else if (version == VERSION_SDXL) {
+        } else if (sd_version_is_sdxl(version)) {
             context_dim           = 2048;
             attention_resolutions = {4, 2};
             channel_mult          = {1, 2, 4};
@@ -204,6 +205,10 @@ public:
             num_head_channels = 64;
             num_heads         = -1;
         }
+        if (sd_version_is_inpaint(version)) {
+            in_channels = 9;
+        }
+
         // dims is always 2
         // use_temporal_attention is always True for SVD
 
@@ -211,7 +216,7 @@ public:
         // time_embed_1 is nn.SiLU()
         blocks["time_embed.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim));
 
-        if (version == VERSION_SDXL || version == VERSION_SVD) {
+        if (sd_version_is_sdxl(version) || version == VERSION_SVD) {
             blocks["label_emb.0.0"] = std::shared_ptr<GGMLBlock>(new Linear(adm_in_channels, time_embed_dim));
             // label_emb_1 is nn.SiLU()
             blocks["label_emb.0.2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, time_embed_dim));
@@ -242,7 +247,7 @@ public:
             if (version == VERSION_SVD) {
                 return new SpatialVideoTransformer(in_channels, n_head, d_head, depth, context_dim);
             } else {
-                return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim);
+                return new SpatialTransformer(in_channels, n_head, d_head, depth, context_dim, flash_attn);
             }
         };
 
@@ -532,10 +537,12 @@ struct UNetModelRunner : public GGMLRunner {
     UnetModelBlock unet;
 
     UNetModelRunner(ggml_backend_t backend,
-                    ggml_type wtype,
-                    SDVersion version = VERSION_SD1)
-        : GGMLRunner(backend, wtype), unet(version) {
-        unet.init(params_ctx, wtype);
+                    std::map<std::string, enum ggml_type>& tensor_types,
+                    const std::string prefix,
+                    SDVersion version = VERSION_SD1,
+                    bool flash_attn   = false)
+        : GGMLRunner(backend), unet(version, tensor_types, flash_attn) {
+        unet.init(params_ctx, tensor_types, prefix);
     }
 
     std::string get_desc() {
@@ -564,6 +571,7 @@ struct UNetModelRunner : public GGMLRunner {
         context   = to_backend(context);
         y         = to_backend(y);
         timesteps = to_backend(timesteps);
+        c_concat  = to_backend(c_concat);
 
         for (int i = 0; i < controls.size(); i++) {
             controls[i] = to_backend(controls[i]);
@@ -649,4 +657,4 @@ struct UNetModelRunner : public GGMLRunner {
     }
 };
 
-#endif  // __UNET_HPP__
+#endif  // __UNET_HPP__

upscaler.cpp

@@ -15,30 +15,38 @@ struct UpscalerGGML {
     }
 
     bool load_from_file(const std::string& esrgan_path) {
-#ifdef SD_USE_CUBLAS
+#ifdef SD_USE_CUDA
         LOG_DEBUG("Using CUDA backend");
         backend = ggml_backend_cuda_init(0);
 #endif
 #ifdef SD_USE_METAL
         LOG_DEBUG("Using Metal backend");
-        ggml_backend_metal_log_set_callback(ggml_log_callback_default, nullptr);
+        ggml_log_set(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
 #ifdef SD_USE_VULKAN
         LOG_DEBUG("Using Vulkan backend");
         backend = ggml_backend_vk_init(0);
 #endif
+#ifdef SD_USE_OPENCL
+        LOG_DEBUG("Using OpenCL backend");
+        backend = ggml_backend_opencl_init();
+#endif
 #ifdef SD_USE_SYCL
         LOG_DEBUG("Using SYCL backend");
         backend = ggml_backend_sycl_init(0);
 #endif
-
+        ModelLoader model_loader;
+        if (!model_loader.init_from_file(esrgan_path)) {
+            LOG_ERROR("init model loader from file failed: '%s'", esrgan_path.c_str());
+        }
+        model_loader.set_wtype_override(model_data_type);
         if (!backend) {
             LOG_DEBUG("Using CPU backend");
             backend = ggml_backend_cpu_init();
         }
         LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
-        esrgan_upscaler = std::make_shared<ESRGAN>(backend, model_data_type);
+        esrgan_upscaler = std::make_shared<ESRGAN>(backend, model_loader.tensor_storages_types);
         if (!esrgan_upscaler->load_from_file(esrgan_path)) {
             return false;
         }
@@ -96,8 +104,7 @@ struct upscaler_ctx_t {
 };
 
 upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
-                                 int n_threads,
-                                 enum sd_type_t wtype) {
+                                 int n_threads) {
     upscaler_ctx_t* upscaler_ctx = (upscaler_ctx_t*)malloc(sizeof(upscaler_ctx_t));
     if (upscaler_ctx == NULL) {
         return NULL;

util.cpp

@@ -22,6 +22,7 @@
 #include <unistd.h>
 #endif
 
+#include "ggml-cpu.h"
 #include "ggml.h"
 #include "stable-diffusion.h"
 
@@ -111,19 +112,32 @@ std::vector<std::string> get_files_from_dir(const std::string& dir) {
     sprintf(directoryPath, "%s\\%s\\*", currentDirectory, dir.c_str());
 
     // Find the first file in the directory
-    hFind = FindFirstFile(directoryPath, &findFileData);
-
+    hFind               = FindFirstFile(directoryPath, &findFileData);
+    bool isAbsolutePath = false;
     // Check if the directory was found
     if (hFind == INVALID_HANDLE_VALUE) {
-        printf("Unable to find directory.\n");
-        return files;
+        printf("Unable to find directory. Try with original path \n");
+
+        char directoryPathAbsolute[MAX_PATH];
+        sprintf(directoryPathAbsolute, "%s*", dir.c_str());
+
+        hFind          = FindFirstFile(directoryPathAbsolute, &findFileData);
+        isAbsolutePath = true;
+        if (hFind == INVALID_HANDLE_VALUE) {
+            printf("Absolute path was also wrong.\n");
+            return files;
+        }
     }
 
     // Loop through all files in the directory
     do {
         // Check if the found file is a regular file (not a directory)
         if (!(findFileData.dwFileAttributes & FILE_ATTRIBUTE_DIRECTORY)) {
-            files.push_back(std::string(currentDirectory) + "\\" + dir + "\\" + std::string(findFileData.cFileName));
+            if (isAbsolutePath) {
+                files.push_back(dir + "\\" + std::string(findFileData.cFileName));
+            } else {
+                files.push_back(std::string(currentDirectory) + "\\" + dir + "\\" + std::string(findFileData.cFileName));
+            }
         }
     } while (FindNextFile(hFind, &findFileData) != 0);
 
@@ -276,6 +290,23 @@ std::string path_join(const std::string& p1, const std::string& p2) {
     return p1 + "/" + p2;
 }
 
+std::vector<std::string> splitString(const std::string& str, char delimiter) {
+    std::vector<std::string> result;
+    size_t start = 0;
+    size_t end   = str.find(delimiter);
+
+    while (end != std::string::npos) {
+        result.push_back(str.substr(start, end - start));
+        start = end + 1;
+        end   = str.find(delimiter, start);
+    }
+
+    // Add the last segment after the last delimiter
+    result.push_back(str.substr(start));
+
+    return result;
+}
+
 sd_image_t* preprocess_id_image(sd_image_t* img) {
     int shortest_edge   = 224;
     int size            = shortest_edge;
@@ -330,7 +361,7 @@ void pretty_progress(int step, int steps, float time) {
         }
     }
     progress += "|";
-    printf(time > 1.0f ? "\r%s %i/%i - %.2fs/it" : "\r%s %i/%i - %.2fit/s",
+    printf(time > 1.0f ? "\r%s %i/%i - %.2fs/it" : "\r%s %i/%i - %.2fit/s\033[K",
            progress.c_str(), step, steps,
            time > 1.0f || time == 0 ? time : (1.0f / time));
     fflush(stdout);  // for linux
@@ -393,7 +424,6 @@ const char* sd_get_system_info() {
     static char buffer[1024];
     std::stringstream ss;
     ss << "System Info: \n";
-    ss << "    BLAS = " << ggml_cpu_has_blas() << std::endl;
     ss << "    SSE3 = " << ggml_cpu_has_sse3() << std::endl;
     ss << "    AVX = " << ggml_cpu_has_avx() << std::endl;
     ss << "    AVX2 = " << ggml_cpu_has_avx2() << std::endl;

util.h

@@ -45,7 +45,7 @@ sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int
 sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size);
 
 std::string path_join(const std::string& p1, const std::string& p2);
-
+std::vector<std::string> splitString(const std::string& str, char delimiter);
 void pretty_progress(int step, int steps, float time);
 
 void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...);

vae.hpp

@@ -163,8 +163,9 @@ public:
 
 class VideoResnetBlock : public ResnetBlock {
 protected:
-    void init_params(struct ggml_context* ctx, ggml_type wtype) {
-        params["mix_factor"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+    void init_params(struct ggml_context* ctx, std::map<std::string, enum ggml_type>& tensor_types, const std::string prefix = "") {
+        enum ggml_type wtype = (tensor_types.find(prefix + "mix_factor") != tensor_types.end()) ? tensor_types[prefix + "mix_factor"] : GGML_TYPE_F32;
+        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
     }
 
     float get_alpha() {
@@ -457,7 +458,7 @@ public:
                        bool use_video_decoder = false,
                        SDVersion version      = VERSION_SD1)
         : decode_only(decode_only), use_video_decoder(use_video_decoder) {
-        if (version == VERSION_SD3_2B || version == VERSION_SD3_5_8B || version == VERSION_SD3_5_2B || version == VERSION_FLUX_DEV || version == VERSION_FLUX_SCHNELL || version == VERSION_FLUX_LITE) {
+        if (sd_version_is_dit(version)) {
             dd_config.z_channels = 16;
             use_quant            = false;
         }
@@ -524,12 +525,13 @@ struct AutoEncoderKL : public GGMLRunner {
     AutoencodingEngine ae;
 
     AutoEncoderKL(ggml_backend_t backend,
-                  ggml_type wtype,
+                  std::map<std::string, enum ggml_type>& tensor_types,
+                  const std::string prefix,
                   bool decode_only       = false,
                   bool use_video_decoder = false,
                   SDVersion version      = VERSION_SD1)
-        : decode_only(decode_only), ae(decode_only, use_video_decoder, version), GGMLRunner(backend, wtype) {
-        ae.init(params_ctx, wtype);
+        : decode_only(decode_only), ae(decode_only, use_video_decoder, version), GGMLRunner(backend) {
+        ae.init(params_ctx, tensor_types, prefix);
     }
 
     std::string get_desc() {
@@ -612,4 +614,4 @@ struct AutoEncoderKL : public GGMLRunner {
     };
 };
 
-#endif
+#endif