fix: convert f64 to f32 and i64 to i32 when loading weights

* Instruct-p2p support

* support 2 conditionings cfg

* Do not re-encode the exact same image twice

* fixes for 2-cfg

* Fix pix2pix latent inputs + improve inpainting a bit + fix naming

* prepare for other pix2pix-like models

* Support sdxl ip2p

* fix reference image embeddings

* Support 2-cond cfg properly in cli

* fix typo in help

* Support masks for ip2p models

* unify code style

* delete unused code

* use edit mode

* add img_cond

* format code

---------

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

.github/workflows/build.yml

@@ -146,7 +146,7 @@ jobs:
             sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-${{ steps.system-info.outputs.OS_NAME }}-${{ steps.system-info.outputs.OS_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}.zip
 
   windows-latest-cmake:
-    runs-on: windows-2019
+    runs-on: windows-2025
 
     env:
       VULKAN_VERSION: 1.3.261.1
@@ -155,15 +155,15 @@ 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"
+            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'
@@ -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

@@ -24,19 +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_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)
@@ -51,18 +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_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"
@@ -86,6 +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
@@ -21,7 +22,7 @@ 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
@@ -49,7 +50,7 @@ Inference of Stable Diffusion and Flux in pure C/C++
     - Linux
     - Mac OS
     - Windows
-    - Android (via Termux)
+    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))
 
 ### TODO
 
@@ -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).
@@ -206,14 +282,14 @@ usage: ./bin/sd [arguments]
 
 arguments:
   -h, --help                         show this help message and exit
-  -M, --mode [MODEL]                 run mode (txt2img or img2img or convert, default: txt2img)
+  -M, --mode [MODE]                  run mode, one of: [img_gen, convert], default: img_gen
   -t, --threads N                    number of threads to use during computation (default: -1)
                                      If threads <= 0, then threads will be set to the number of CPU physical cores
   -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
-  --t5xxl                            path to the the t5xxl text encoder
+  --clip_g                           path to the clip-g text encoder
+  --t5xxl                            path to the t5xxl text encoder
   --vae [VAE]                        path to vae
   --taesd [TAESD_PATH]               path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
   --control-net [CONTROL_PATH]       path to control net model
@@ -223,22 +299,34 @@ 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
+  --tensor-type-rules [EXPRESSION]   weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --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)
+  --img-cfg-scale SCALE              image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --guidance SCALE                   distilled guidance scale for models with guidance input (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%)
+  --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)
@@ -255,7 +343,10 @@ arguments:
                                      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
 ```
 
@@ -282,7 +373,7 @@ Using formats of different precisions will yield results of varying quality.
 
 
 ```
-./bin/sd --mode img2img -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
+./bin/sd -m ../models/sd-v1-4.ckpt -p "cat with blue eyes" -i ./output.png -o ./img2img_output.png --strength 0.4
 ```
 
 <p align="center">
@@ -303,10 +394,12 @@ Using formats of different precisions will yield results of varying quality.
 
 These projects wrap `stable-diffusion.cpp` for easier use in other languages/frameworks.
 
-* Golang: [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
+* Golang (non-cgo): [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
+* Golang (cgo): [Binozo/GoStableDiffusion](https://github.com/Binozo/GoStableDiffusion)
 * 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)
+* Flutter/Dart: [rmatif/Local-Diffusion](https://github.com/rmatif/Local-Diffusion)
 
 ## UIs
 
@@ -314,6 +407,9 @@ These projects use `stable-diffusion.cpp` as a backend for their image generatio
 
 - [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)
+- [Local Diffusion](https://github.com/rmatif/Local-Diffusion)
+- [sd.cpp-webui](https://github.com/daniandtheweb/sd.cpp-webui)
 
 ## Contributors
 

clip.hpp

@@ -546,7 +546,7 @@ protected:
     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 token_wtype    = (tensor_types.find(prefix + "token_embedding.weight") != tensor_types.end()) ? tensor_types[prefix + "token_embedding.weight"] : GGML_TYPE_F32;
+        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);
@@ -678,8 +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;
@@ -701,7 +701,7 @@ public:
 
     void set_clip_skip(int skip) {
         if (skip <= 0) {
-            return;
+            skip = -1;
         }
         clip_skip = skip;
     }
@@ -871,9 +871,9 @@ struct CLIPTextModelRunner : public GGMLRunner {
                         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), model(version, clip_skip_value, with_final_ln) {
+                        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);
     }
 

common.hpp

@@ -56,8 +56,8 @@ 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 = conv->forward(ctx, x);    // [N, out_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;
     }
 };

conditioner.hpp

@@ -51,7 +51,8 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
 
     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;
 
@@ -62,23 +63,24 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                       PMVersion pv      = PM_VERSION_1,
                                       int clip_skip     = -1)
         : 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;
             }
         }
-        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, clip_skip);
-        } 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, clip_skip);
-        } 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, clip_skip, false);
-            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
-        }
-    }
-
-    void set_clip_skip(int clip_skip) {
         text_model->set_clip_skip(clip_skip);
         if (sd_version_is_sdxl(version)) {
             text_model2->set_clip_skip(clip_skip);
@@ -131,28 +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, tensor_storage.type, 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(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++;
+        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;
     }
 
@@ -430,8 +459,8 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                 if (sd_version_is_sdxl(version)) {
                     text_model2->compute(n_threads,
                                          input_ids2,
-                                         0,
-                                         NULL,
+                                         num_custom_embeddings,
+                                         token_embed_custom.data(),
                                          max_token_idx,
                                          false,
                                          &chunk_hidden_states2, work_ctx);
@@ -441,8 +470,8 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                     if (chunk_idx == 0) {
                         text_model2->compute(n_threads,
                                              input_ids2,
-                                             0,
-                                             NULL,
+                                             num_custom_embeddings,
+                                             token_embed_custom.data(),
                                              max_token_idx,
                                              true,
                                              &pooled,
@@ -637,15 +666,16 @@ struct SD3CLIPEmbedder : public Conditioner {
                     std::map<std::string, enum ggml_type>& tensor_types,
                     int clip_skip = -1)
         : clip_g_tokenizer(0) {
-        if (clip_skip <= 0) {
-            clip_skip = 2;
-        }
-        clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, false);
-        clip_g = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, clip_skip, false);
+        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);
     }
@@ -719,7 +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] << ", ";
@@ -874,6 +904,7 @@ struct SD3CLIPEmbedder : public Conditioner {
 
                 t5->compute(n_threads,
                             input_ids,
+                            NULL,
                             &chunk_hidden_states_t5,
                             work_ctx);
                 {
@@ -976,18 +1007,20 @@ struct FluxCLIPEmbedder : public Conditioner {
     T5UniGramTokenizer t5_tokenizer;
     std::shared_ptr<CLIPTextModelRunner> clip_l;
     std::shared_ptr<T5Runner> t5;
+    size_t chunk_len = 256;
 
     FluxCLIPEmbedder(ggml_backend_t backend,
                      std::map<std::string, enum ggml_type>& tensor_types,
                      int clip_skip = -1) {
-        if (clip_skip <= 0) {
-            clip_skip = 2;
-        }
-        clip_l = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, clip_skip, true);
+        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);
     }
 
@@ -1049,7 +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] << ", ";
@@ -1081,7 +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
@@ -1119,6 +1151,7 @@ struct FluxCLIPEmbedder : public Conditioner {
 
                 t5->compute(n_threads,
                             input_ids,
+                            NULL,
                             &chunk_hidden_states,
                             work_ctx);
                 {
@@ -1168,7 +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);
     }
 
@@ -1190,4 +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

denoiser.hpp

@@ -168,24 +168,21 @@ struct AYSSchedule : SigmaSchedule {
         std::vector<float> inputs;
         std::vector<float> results(n + 1);
 
-        switch (version) {
-            case VERSION_SD2: /* fallthrough */
-                LOG_WARN("AYS not designed for SD2.X models");
-            case VERSION_SD1:
-                LOG_INFO("AYS using SD1.5 noise levels");
-                inputs = noise_levels[0];
-                break;
-            case VERSION_SDXL:
-                LOG_INFO("AYS using SDXL noise levels");
-                inputs = noise_levels[1];
-                break;
-            case VERSION_SVD:
-                LOG_INFO("AYS using SVD noise levels");
-                inputs = noise_levels[2];
-                break;
-            default:
-                LOG_ERROR("Version not compatable with AYS scheduler");
-                return results;
+        if (sd_version_is_sd2((SDVersion)version)) {
+            LOG_WARN("AYS not designed for SD2.X models");
+        } /* fallthrough */
+        else if (sd_version_is_sd1((SDVersion)version)) {
+            LOG_INFO("AYS using SD1.5 noise levels");
+            inputs = noise_levels[0];
+        } else if (sd_version_is_sdxl((SDVersion)version)) {
+            LOG_INFO("AYS using SDXL noise levels");
+            inputs = noise_levels[1];
+        } else if (version == VERSION_SVD) {
+            LOG_INFO("AYS using SVD noise levels");
+            inputs = noise_levels[2];
+        } else {
+            LOG_ERROR("Version not compatible with AYS scheduler");
+            return results;
         }
 
         /* Stretches those pre-calculated reference levels out to the desired
@@ -346,6 +343,32 @@ struct CompVisVDenoiser : public CompVisDenoiser {
     }
 };
 
+struct EDMVDenoiser : public CompVisVDenoiser {
+    float min_sigma = 0.002;
+    float max_sigma = 120.0;
+
+    EDMVDenoiser(float min_sigma = 0.002, float max_sigma = 120.0)
+        : min_sigma(min_sigma), max_sigma(max_sigma) {
+        schedule = std::make_shared<ExponentialSchedule>();
+    }
+
+    float t_to_sigma(float t) {
+        return std::exp(t * 4 / (float)TIMESTEPS);
+    }
+
+    float sigma_to_t(float s) {
+        return 0.25 * std::log(s);
+    }
+
+    float sigma_min() {
+        return min_sigma;
+    }
+
+    float sigma_max() {
+        return max_sigma;
+    }
+};
+
 float time_snr_shift(float alpha, float t) {
     if (alpha == 1.0f) {
         return t;
@@ -474,7 +497,8 @@ 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 +1029,370 @@ 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,
@@ -68,6 +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,
@@ -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,
@@ -134,8 +137,9 @@ struct FluxModel : public DiffusionModel {
     FluxModel(ggml_backend_t backend,
               std::map<std::string, enum ggml_type>& tensor_types,
               SDVersion version = VERSION_FLUX,
-              bool flash_attn   = false)
-        : flux(backend, tensor_types, "model.diffusion_model", version, flash_attn) {
+              bool flash_attn   = false,
+              bool use_mask     = false)
+        : flux(backend, tensor_types, "model.diffusion_model", version, flash_attn, use_mask) {
     }
 
     void alloc_params_buffer() {
@@ -169,13 +173,14 @@ 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, c_concat, 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);
     }
 };
 

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 -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) |
+
+
+

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;
     }

flux.hpp

@@ -117,6 +117,7 @@ namespace Flux {
                                                     struct ggml_tensor* k,
                                                     struct ggml_tensor* v,
                                                     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]
@@ -124,7 +125,7 @@ namespace Flux {
         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, flash_attn);  // [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;
     }
 
@@ -167,13 +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, flash_attn);  // [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;
         }
     };
@@ -185,6 +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 {
@@ -210,19 +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()};
         }
     };
 
@@ -242,25 +243,33 @@ 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,
+                          int idx         = 0,
                           bool qkv_bias   = false,
-                          bool flash_attn = false)
-            : flash_attn(flash_attn) {
+                          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, flash_attn));
+            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, flash_attn));
 
@@ -270,17 +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"]);
 
@@ -288,7 +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"]);
 
@@ -296,10 +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];
 
@@ -324,7 +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, flash_attn);              // [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,
@@ -373,14 +410,18 @@ namespace Flux {
         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,
+                          int idx         = 0,
                           float qk_scale  = 0.f,
-                          bool flash_attn = false)
-            : hidden_size(hidden_size), num_heads(num_heads), flash_attn(flash_attn) {
+                          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) {
@@ -393,26 +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]
@@ -443,7 +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, flash_attn);  // [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]
@@ -454,13 +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,
@@ -469,17 +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);
@@ -488,6 +562,34 @@ 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;
@@ -504,6 +606,7 @@ namespace Flux {
         bool qkv_bias               = true;
         bool guidance_embed         = true;
         bool flash_attn             = true;
+        bool is_chroma              = false;
     };
 
     struct Flux : public GGMLBlock {
@@ -570,17 +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];
                 }
@@ -592,24 +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();
@@ -645,11 +783,15 @@ namespace Flux {
             : params(params) {
             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));
 
@@ -657,19 +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,
+                                                                                                                i,
                                                                                                                 params.qkv_bias,
-                                                                                                                params.flash_attn));
+                                                                                                                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,
+                                                                                                                i,
                                                                                                                 0.f,
-                                                                                                                params.flash_attn));
+                                                                                                                params.flash_attn,
+                                                                                                                params.is_chroma));
             }
 
-            blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new LastLayer(params.hidden_size, 1, params.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,
@@ -726,25 +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++) {
@@ -754,7 +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]
             }
@@ -766,7 +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]
@@ -781,7 +957,20 @@ 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* 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;
+            int pad_h          = (patch_size - H % patch_size) % patch_size;
+            int pad_w          = (patch_size - W % patch_size) % patch_size;
+            x                  = ggml_pad(ctx, x, pad_w, pad_h, 0, 0);  // [N, C, H + pad_h, W + pad_w]
+
+            // 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;
         }
 
@@ -793,7 +982,9 @@ 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<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
@@ -812,25 +1003,33 @@ namespace Flux {
             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;
-            x                  = ggml_pad(ctx, x, pad_w, pad_h, 0, 0);  // [N, C, H + pad_h, W + pad_w]
 
-            // 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]
+            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 = ggml_pad(ctx, masked, pad_w, pad_h, 0, 0);
-                mask   = ggml_pad(ctx, mask, pad_w, pad_h, 0, 0);
-
-                masked = patchify(ctx, masked, patch_size);
-                mask   = patchify(ctx, mask, patch_size);
+                masked = process_img(ctx, masked);
+                mask   = process_img(ctx, mask);
 
                 img = ggml_concat(ctx, img, ggml_concat(ctx, masked, mask, 0), 0);
             }
 
-            auto out = forward_orig(ctx, img, context, timestep, y, guidance, pe, skip_layers);  // [N, h*w, C * patch_size * patch_size]
+            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]
@@ -845,14 +1044,18 @@ namespace Flux {
     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,
                    std::map<std::string, enum ggml_type>& tensor_types = empty_tensor_types,
                    const std::string prefix                            = "",
                    SDVersion version                                   = VERSION_FLUX,
-                   bool flash_attn                                     = false)
-            : GGMLRunner(backend) {
+                   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;
@@ -868,6 +1071,10 @@ namespace Flux {
                     // 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
@@ -887,7 +1094,9 @@ namespace Flux {
             }
 
             LOG_INFO("Flux blocks: %d double, %d single", flux_params.depth, flux_params.depth_single_blocks);
-            if (!flux_params.guidance_embed) {
+            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)");
             }
 
@@ -909,22 +1118,41 @@ namespace Flux {
                                         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);
 
+            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);
             }
-            y         = to_backend(y);
+            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);
@@ -941,6 +1169,8 @@ namespace Flux {
                                                    y,
                                                    guidance,
                                                    pe,
+                                                   mod_index_arange,
+                                                   ref_latents,
                                                    skip_layers);
 
             ggml_build_forward_expand(gf, out);
@@ -955,16 +1185,17 @@ namespace Flux {
                      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, c_concat, 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);
@@ -1004,7 +1235,7 @@ namespace Flux {
                 struct ggml_tensor* out = NULL;
 
                 int t0 = ggml_time_ms();
-                compute(8, x, timesteps, context, NULL, 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);

ggml_extend.hpp

@@ -27,7 +27,7 @@
 
 #include "model.h"
 
-#ifdef SD_USE_CUBLAS
+#ifdef SD_USE_CUDA
 #include "ggml-cuda.h"
 #endif
 
@@ -39,6 +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
@@ -52,6 +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;
@@ -318,8 +388,10 @@ __STATIC_INLINE__ void sd_apply_mask(struct ggml_tensor* image_data,
     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 = ((float)(m < 254.5/255)) * (ggml_tensor_get_f32(image_data, ix, iy, k) - .5) + .5;
+                float value = (1 - m) * (ggml_tensor_get_f32(image_data, ix, iy, k) - .5) + .5;
                 ggml_tensor_set_f32(output, value, ix, iy, k);
             }
         }
@@ -530,6 +602,8 @@ typedef std::function<void(ggml_tensor*, ggml_tensor*, bool)> on_tile_process;
 
 // Tiling
 __STATIC_INLINE__ void sd_tiling(ggml_tensor* input, ggml_tensor* output, const int scale, const int tile_size, const float tile_overlap_factor, on_tile_process on_processing) {
+    output = ggml_set_f32(output, 0);
+
     int input_width   = (int)input->ne[0];
     int input_height  = (int)input->ne[1];
     int output_width  = (int)output->ne[0];
@@ -708,7 +782,7 @@ __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];
@@ -766,18 +840,34 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
 
     float scale = (1.0f / sqrt((float)d_head));
 
-    // 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);
-    // }
+    int kv_pad = 0;
+    //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 && (
+        d_head == 64 ||
+        d_head == 80 ||
+        d_head == 96 ||
+        d_head == 112 ||
+        d_head == 128 ||
+        d_head == 256
+    );
+#if 0
     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
+#else
+    if (can_use_flash_attn && L_k % 256 != 0) {
+        // TODO(Green-Sky): might be worth just padding by default
+        if (L_k == 77 || L_k == 4208 || L_k == 3952) {
+            kv_pad = GGML_PAD(L_k, 256) - L_k;
+        } else {
+            can_use_flash_attn = false;
+        }
+    }
+#endif
 
     if (mask != nullptr) {
         // TODO(Green-Sky): figure out if we can bend t5 to work too
@@ -790,13 +880,32 @@ __STATIC_INLINE__ struct ggml_tensor* ggml_nn_attention_ext(struct ggml_context*
     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");
+        //LOG_DEBUG(" uses flash attention");
+        if (kv_pad != 0) {
+            //LOG_DEBUG(" padding k and v dim1 by %d", kv_pad);
+            k = ggml_pad(ctx, k, 0, kv_pad, 0, 0);
+        }
         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]
+        if (kv_pad != 0) {
+            v = ggml_pad(ctx, v, 0, kv_pad, 0, 0);
+        }
         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);
 
@@ -809,7 +918,7 @@ __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);
@@ -864,7 +973,7 @@ __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);
@@ -987,8 +1096,8 @@ __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:

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

@@ -3,7 +3,7 @@
 
 #include "ggml_extend.hpp"
 
-#define LORA_GRAPH_SIZE 10240
+#define LORA_GRAPH_BASE_SIZE 10240
 
 struct LoraModel : public GGMLRunner {
     enum lora_t {
@@ -197,6 +197,10 @@ struct LoraModel : public GGMLRunner {
                 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) {
@@ -217,20 +221,25 @@ struct LoraModel : public GGMLRunner {
                     keys.push_back(split_blk);
                 }
             }
+            keys.push_back(blk_name);
         }
-        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);
+        size_t lora_graph_size = LORA_GRAPH_BASE_SIZE + lora_tensors.size() * 10;
+        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);
         set_backend_tensor_data(zero_index, zero_index_vec.data());
@@ -244,387 +253,544 @@ struct LoraModel : public GGMLRunner {
             std::vector<std::string> keys = to_lora_keys(k_tensor, version);
             if (keys.size() == 0)
                 continue;
-            ggml_tensor* lora_up   = NULL;
-            ggml_tensor* lora_down = NULL;
+
             for (auto& key : keys) {
-                std::string alpha_name         = "";
-                std::string scale_name         = "";
-                std::string split_q_scale_name = "";
-                std::string lora_down_name     = "";
-                std::string lora_up_name       = "";
-
-                if (starts_with(key, "SPLIT|")) {
+                bool is_qkv_split = starts_with(key, "SPLIT|");
+                if (is_qkv_split) {
                     key = key.substr(sizeof("SPLIT|") - 1);
-                    // TODO: Handle alphas
-                    std::string suffix  = "";
-                    auto split_q_d_name = lora_pre[type] + key + "q" + suffix + lora_downs[type] + ".weight";
-
-                    if (lora_tensors.find(split_q_d_name) == lora_tensors.end()) {
-                        suffix         = "_proj";
-                        split_q_d_name = lora_pre[type] + key + "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 = lora_pre[type] + key + "k" + suffix + lora_downs[type] + ".weight";
-                        auto split_v_d_name = lora_pre[type] + key + "v" + suffix + lora_downs[type] + ".weight";
-
-                        auto split_q_u_name = lora_pre[type] + key + "q" + suffix + lora_ups[type] + ".weight";
-                        auto split_k_u_name = lora_pre[type] + key + "k" + suffix + lora_ups[type] + ".weight";
-                        auto split_v_u_name = lora_pre[type] + key + "v" + suffix + lora_ups[type] + ".weight";
-
-                        auto split_q_scale_name = lora_pre[type] + key + "q" + suffix + ".scale";
-                        auto split_k_scale_name = lora_pre[type] + key + "k" + suffix + ".scale";
-                        auto split_v_scale_name = lora_pre[type] + key + "v" + suffix + ".scale";
-
-                        auto split_q_alpha_name = lora_pre[type] + key + "q" + suffix + ".alpha";
-                        auto split_k_alpha_name = lora_pre[type] + key + "k" + suffix + ".alpha";
-                        auto split_v_alpha_name = lora_pre[type] + key + "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);
-                    }
                 }
-                if (starts_with(key, "SPLIT_L|")) {
+                bool is_qkvm_split = starts_with(key, "SPLIT_L|");
+                if (is_qkvm_split) {
                     key = key.substr(sizeof("SPLIT_L|") - 1);
-
-                    auto split_q_d_name = lora_pre[type] + key + "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 = lora_pre[type] + key + "attn.to_k" + lora_downs[type] + ".weight";
-                        auto split_v_d_name = lora_pre[type] + key + "attn.to_v" + lora_downs[type] + ".weight";
-
-                        auto split_q_u_name = lora_pre[type] + key + "attn.to_q" + lora_ups[type] + ".weight";
-                        auto split_k_u_name = lora_pre[type] + key + "attn.to_k" + lora_ups[type] + ".weight";
-                        auto split_v_u_name = lora_pre[type] + key + "attn.to_v" + lora_ups[type] + ".weight";
-
-                        auto split_m_d_name = lora_pre[type] + key + "proj_mlp" + lora_downs[type] + ".weight";
-                        auto split_m_u_name = lora_pre[type] + key + "proj_mlp" + lora_ups[type] + ".weight";
-
-                        auto split_q_scale_name = lora_pre[type] + key + "attn.to_q" + ".scale";
-                        auto split_k_scale_name = lora_pre[type] + key + "attn.to_k" + ".scale";
-                        auto split_v_scale_name = lora_pre[type] + key + "attn.to_v" + ".scale";
-                        auto split_m_scale_name = lora_pre[type] + key + "proj_mlp" + ".scale";
-
-                        auto split_q_alpha_name = lora_pre[type] + key + "attn.to_q" + ".alpha";
-                        auto split_k_alpha_name = lora_pre[type] + key + "attn.to_k" + ".alpha";
-                        auto split_v_alpha_name = lora_pre[type] + key + "attn.to_v" + ".alpha";
-                        auto split_m_alpha_name = lora_pre[type] + key + "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);
-                    }
                 }
-                if (lora_up == NULL || lora_down == NULL) {
-                    lora_up_name = lora_pre[type] + key + lora_ups[type] + ".weight";
-                    if (lora_tensors.find(lora_up_name) == lora_tensors.end()) {
-                        if (key == "model_diffusion_model_output_blocks_2_2_conv") {
-                            // fix for some sdxl lora, like lcm-lora-xl
-                            key          = "model_diffusion_model_output_blocks_2_1_conv";
-                            lora_up_name = lora_pre[type] + key + lora_ups[type] + ".weight";
-                        }
+                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));
                     }
 
-                    lora_down_name = lora_pre[type] + key + lora_downs[type] + ".weight";
-                    alpha_name     = lora_pre[type] + key + ".alpha";
-                    scale_name     = lora_pre[type] + key + ".scale";
-
-                    if (lora_tensors.find(lora_up_name) != lora_tensors.end()) {
-                        lora_up = lora_tensors[lora_up_name];
+                    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));
                     }
 
-                    if (lora_tensors.find(lora_down_name) != lora_tensors.end()) {
-                        lora_down = lora_tensors[lora_down_name];
-                    }
-                    applied_lora_tensors.insert(lora_up_name);
-                    applied_lora_tensors.insert(lora_down_name);
+                    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);
-                    applied_lora_tensors.insert(scale_name);
-                }
+                    if (hada_1_up == NULL || hada_1_down == NULL || hada_2_up == NULL || hada_2_down == NULL) {
+                        continue;
+                    }
 
-                if (lora_up == NULL || lora_down == NULL) {
-                    continue;
-                }
-                // calc_scale
-                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;
+                    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;
-
-                // 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);
+                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;

model.cpp

@@ -26,6 +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) {
@@ -96,6 +100,7 @@ const char* unused_tensors[] = {
     "model_ema.diffusion_model",
     "embedding_manager",
     "denoiser.sigmas",
+    "text_encoders.t5xxl.transformer.encoder.embed_tokens.weight",  // only used during training
 };
 
 bool is_unused_tensor(std::string name) {
@@ -177,6 +182,64 @@ std::unordered_map<std::string, std::string> pmid_v2_name_map = {
 std::string convert_open_clip_to_hf_clip(const std::string& name) {
     std::string new_name = name;
     std::string prefix;
+    if (contains(new_name, ".enc.")) {
+        // llama.cpp naming convention for T5
+        size_t pos = new_name.find(".enc.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 5, ".encoder.");
+        }
+        pos = new_name.find("blk.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 4, "block.");
+        }
+        pos = new_name.find("output_norm.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 12, "final_layer_norm.");
+        }
+        pos = new_name.find("attn_k.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 7, "layer.0.SelfAttention.k.");
+        }
+        pos = new_name.find("attn_v.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 7, "layer.0.SelfAttention.v.");
+        }
+        pos = new_name.find("attn_o.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 7, "layer.0.SelfAttention.o.");
+        }
+        pos = new_name.find("attn_q.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 7, "layer.0.SelfAttention.q.");
+        }
+        pos = new_name.find("attn_norm.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 10, "layer.0.layer_norm.");
+        }
+        pos = new_name.find("ffn_norm.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 9, "layer.1.layer_norm.");
+        }
+        pos = new_name.find("ffn_up.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 7, "layer.1.DenseReluDense.wi_1.");
+        }
+        pos = new_name.find("ffn_down.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 9, "layer.1.DenseReluDense.wo.");
+        }
+        pos = new_name.find("ffn_gate.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 9, "layer.1.DenseReluDense.wi_0.");
+        }
+        pos = new_name.find("attn_rel_b.");
+        if (pos != std::string::npos) {
+            new_name.replace(pos, 11, "layer.0.SelfAttention.relative_attention_bias.");
+        }
+    } else if (name == "text_encoders.t5xxl.transformer.token_embd.weight") {
+        new_name = "text_encoders.t5xxl.transformer.shared.weight";
+    }
+
     if (starts_with(new_name, "conditioner.embedders.0.open_clip.")) {
         prefix   = "cond_stage_model.";
         new_name = new_name.substr(strlen("conditioner.embedders.0.open_clip."));
@@ -275,6 +338,10 @@ std::unordered_map<std::string, std::unordered_map<std::string, std::string>> su
             {"to_v", "v"},
             {"to_out_0", "proj_out"},
             {"group_norm", "norm"},
+            {"key", "k"},
+            {"query", "q"},
+            {"value", "v"},
+            {"proj_attn", "proj_out"},
         },
     },
     {
@@ -299,6 +366,10 @@ std::unordered_map<std::string, std::unordered_map<std::string, std::string>> su
             {"to_v", "v"},
             {"to_out.0", "proj_out"},
             {"group_norm", "norm"},
+            {"key", "k"},
+            {"query", "q"},
+            {"value", "v"},
+            {"proj_attn", "proj_out"},
         },
     },
     {
@@ -370,6 +441,10 @@ std::string convert_diffusers_name_to_compvis(std::string key, char seq) {
         return format("model%cdiffusion_model%ctime_embed%c", seq, seq, seq) + std::to_string(std::stoi(m[0]) * 2 - 2) + m[1];
     }
 
+    if (match(m, std::regex(format("unet%cadd_embedding%clinear_(\\d+)(.*)", seq, seq)), key)) {
+        return format("model%cdiffusion_model%clabel_emb%c0%c", seq, seq, seq, seq) + std::to_string(std::stoi(m[0]) * 2 - 2) + m[1];
+    }
+
     if (match(m, std::regex(format("unet%cdown_blocks%c(\\d+)%c(attentions|resnets)%c(\\d+)%c(.+)", seq, seq, seq, seq, seq)), key)) {
         std::string suffix = get_converted_suffix(m[1], m[3]);
         // LOG_DEBUG("%s %s %s %s", m[0].c_str(), m[1].c_str(), m[2].c_str(), m[3].c_str());
@@ -407,6 +482,19 @@ std::string convert_diffusers_name_to_compvis(std::string key, char seq) {
         return format("cond_stage_model%ctransformer%ctext_model", seq, seq) + m[0];
     }
 
+    // clip-g
+    if (match(m, std::regex(format("te%c1%ctext_model%cencoder%clayers%c(\\d+)%c(.+)", seq, seq, seq, seq, seq, seq)), key)) {
+        return format("cond_stage_model%c1%ctransformer%ctext_model%cencoder%clayers%c", seq, seq, seq, seq, seq, seq) + m[0] + seq + m[1];
+    }
+
+    if (match(m, std::regex(format("te%c1%ctext_model(.*)", seq, seq)), key)) {
+        return format("cond_stage_model%c1%ctransformer%ctext_model", seq, seq, seq) + m[0];
+    }
+
+    if (match(m, std::regex(format("te%c1%ctext_projection", seq, seq)), key)) {
+        return format("cond_stage_model%c1%ctransformer%ctext_model%ctext_projection", seq, seq, seq, seq);
+    }
+
     // vae
     if (match(m, std::regex(format("vae%c(.*)%cconv_norm_out(.*)", seq, seq)), key)) {
         return format("first_stage_model%c%s%cnorm_out%s", seq, m[0].c_str(), seq, m[1].c_str());
@@ -543,6 +631,8 @@ std::string convert_tensor_name(std::string name) {
             std::string new_key = convert_diffusers_name_to_compvis(name_without_network_parts, '.');
             if (new_key.empty()) {
                 new_name = name;
+            } else if (new_key == "cond_stage_model.1.transformer.text_model.text_projection") {
+                new_name = new_key;
             } else {
                 new_name = new_key + "." + network_part;
             }
@@ -558,6 +648,26 @@ 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;
@@ -705,6 +815,7 @@ 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--) {
@@ -712,6 +823,20 @@ void f8_e5m2_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
     }
 }
 
+void f64_to_f32_vec(double* src, float* dst, int64_t n) {
+    // support inplace op
+    for (int64_t i = 0; i < n; i++) {
+        dst[i] = (float)src[i];
+    }
+}
+
+void i64_to_i32_vec(int64_t* src, int32_t* dst, int64_t n) {
+    // support inplace op
+    for (int64_t i = 0; i < n; i++) {
+        dst[i] = (int32_t)src[i];
+    }
+}
+
 void convert_tensor(void* src,
                     ggml_type src_type,
                     void* dst,
@@ -927,7 +1052,7 @@ 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);
-        tensor_storages_types[tensor_storage.name] = tensor_storage.type;
+        add_preprocess_tensor_storage_types(tensor_storages_types, tensor_storage.name, tensor_storage.type);
     }
 
     gguf_free(ctx_gguf_);
@@ -946,10 +1071,14 @@ ggml_type str_to_ggml_type(const std::string& dtype) {
         ttype = GGML_TYPE_F32;
     } else if (dtype == "F32") {
         ttype = GGML_TYPE_F32;
+    } else if (dtype == "F64") {
+        ttype = GGML_TYPE_F32;
     } else if (dtype == "F8_E4M3") {
         ttype = GGML_TYPE_F16;
     } else if (dtype == "F8_E5M2") {
         ttype = GGML_TYPE_F16;
+    } else if (dtype == "I64") {
+        ttype = GGML_TYPE_I32;
     }
     return ttype;
 }
@@ -962,6 +1091,7 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
     std::ifstream file(file_path, std::ios::binary);
     if (!file.is_open()) {
         LOG_ERROR("failed to open '%s'", file_path.c_str());
+        file_paths_.pop_back();
         return false;
     }
 
@@ -973,6 +1103,7 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
     // read header size
     if (file_size_ <= ST_HEADER_SIZE_LEN) {
         LOG_ERROR("invalid safetensor file '%s'", file_path.c_str());
+        file_paths_.pop_back();
         return false;
     }
 
@@ -986,6 +1117,7 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
     size_t header_size_ = read_u64(header_size_buf);
     if (header_size_ >= file_size_) {
         LOG_ERROR("invalid safetensor file '%s'", file_path.c_str());
+        file_paths_.pop_back();
         return false;
     }
 
@@ -996,6 +1128,7 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
     file.read(header_buf.data(), header_size_);
     if (!file) {
         LOG_ERROR("read safetensors header failed: '%s'", file_path.c_str());
+        file_paths_.pop_back();
         return false;
     }
 
@@ -1067,12 +1200,20 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
             tensor_storage.is_f8_e5m2 = true;
             // f8 -> f16
             GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
+        } else if (dtype == "F64") {
+            tensor_storage.is_f64 = true;
+            // f64 -> f32
+            GGML_ASSERT(tensor_storage.nbytes() * 2 == tensor_data_size);
+        } else if (dtype == "I64") {
+            tensor_storage.is_i64 = true;
+            // i64 -> i32
+            GGML_ASSERT(tensor_storage.nbytes() * 2 == tensor_data_size);
         } else {
             GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size);
         }
 
         tensor_storages.push_back(tensor_storage);
-        tensor_storages_types[tensor_storage.name] = tensor_storage.type;
+        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());
     }
@@ -1083,18 +1224,45 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
 /*================================================= DiffusersModelLoader ==================================================*/
 
 bool ModelLoader::init_from_diffusers_file(const std::string& file_path, const std::string& prefix) {
-    std::string unet_path = path_join(file_path, "unet/diffusion_pytorch_model.safetensors");
-    std::string vae_path  = path_join(file_path, "vae/diffusion_pytorch_model.safetensors");
-    std::string clip_path = path_join(file_path, "text_encoder/model.safetensors");
+    std::string unet_path   = path_join(file_path, "unet/diffusion_pytorch_model.safetensors");
+    std::string vae_path    = path_join(file_path, "vae/diffusion_pytorch_model.safetensors");
+    std::string clip_path   = path_join(file_path, "text_encoder/model.safetensors");
+    std::string clip_g_path = path_join(file_path, "text_encoder_2/model.safetensors");
 
     if (!init_from_safetensors_file(unet_path, "unet.")) {
         return false;
     }
+    for (auto ts : tensor_storages) {
+        if (ts.name.find("add_embedding") != std::string::npos || ts.name.find("label_emb") != std::string::npos) {
+            // probably SDXL
+            LOG_DEBUG("Fixing name for SDXL output blocks.2.2");
+            for (auto& tensor_storage : tensor_storages) {
+                int len  = 34;
+                auto pos = tensor_storage.name.find("unet.up_blocks.0.upsamplers.0.conv");
+                if (pos == std::string::npos) {
+                    len = 44;
+                    pos = tensor_storage.name.find("model.diffusion_model.output_blocks.2.1.conv");
+                }
+                if (pos != std::string::npos) {
+                    tensor_storage.name = "model.diffusion_model.output_blocks.2.2.conv" + tensor_storage.name.substr(len);
+                    LOG_DEBUG("NEW NAME: %s", tensor_storage.name.c_str());
+                    add_preprocess_tensor_storage_types(tensor_storages_types, tensor_storage.name, tensor_storage.type);
+                }
+            }
+            break;
+        }
+    }
+
     if (!init_from_safetensors_file(vae_path, "vae.")) {
-        return false;
+        LOG_WARN("Couldn't find working VAE in %s", file_path.c_str());
+        // return false;
     }
     if (!init_from_safetensors_file(clip_path, "te.")) {
-        return false;
+        LOG_WARN("Couldn't find working text encoder in %s", file_path.c_str());
+        // return false;
+    }
+    if (!init_from_safetensors_file(clip_g_path, "te.1.")) {
+        LOG_DEBUG("Couldn't find working second text encoder in %s", file_path.c_str());
     }
     return true;
 }
@@ -1403,7 +1571,7 @@ bool ModelLoader::parse_data_pkl(uint8_t* buffer,
                         // 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);
-                        tensor_storages_types[reader.tensor_storage.name] = reader.tensor_storage.type;
+                        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
@@ -1457,14 +1625,23 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
     return true;
 }
 
+bool ModelLoader::model_is_unet() {
+    for (auto& tensor_storage : tensor_storages) {
+        if (tensor_storage.name.find("model.diffusion_model.input_blocks.") != std::string::npos) {
+            return true;
+        }
+    }
+    return false;
+}
+
 SDVersion ModelLoader::get_sd_version() {
     TensorStorage token_embedding_weight, input_block_weight;
     bool input_block_checked = false;
 
-    bool has_multiple_encoders   = false;
-    bool is_unet = false;
+    bool has_multiple_encoders = false;
+    bool is_unet               = false;
 
-    bool is_xl = false;
+    bool is_xl   = false;
     bool is_flux = false;
 
 #define found_family (is_xl || is_flux)
@@ -1479,18 +1656,18 @@ SDVersion ModelLoader::get_sd_version() {
             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) {
+            if (tensor_storage.name.find("model.diffusion_model.input_blocks.") != std::string::npos || tensor_storage.name.find("unet.down_blocks.") != std::string::npos) {
                 is_unet = true;
-                if(has_multiple_encoders){
+                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) {
+            if (tensor_storage.name.find("conditioner.embedders.1") != std::string::npos || tensor_storage.name.find("cond_stage_model.1") != std::string::npos || tensor_storage.name.find("te.1") != std::string::npos) {
                 has_multiple_encoders = true;
-                if(is_unet){
+                if (is_unet) {
                     is_xl = true;
                     if (input_block_checked) {
                         break;
@@ -1510,7 +1687,7 @@ SDVersion ModelLoader::get_sd_version() {
             token_embedding_weight = tensor_storage;
             // break;
         }
-        if (tensor_storage.name == "model.diffusion_model.input_blocks.0.0.weight" || tensor_storage.name == "model.diffusion_model.img_in.weight") {
+        if (tensor_storage.name == "model.diffusion_model.input_blocks.0.0.weight" || tensor_storage.name == "model.diffusion_model.img_in.weight" || tensor_storage.name == "unet.conv_in.weight") {
             input_block_weight  = tensor_storage;
             input_block_checked = true;
             if (found_family) {
@@ -1519,10 +1696,14 @@ SDVersion ModelLoader::get_sd_version() {
         }
     }
     bool is_inpaint = input_block_weight.ne[2] == 9;
+    bool is_ip2p    = input_block_weight.ne[2] == 8;
     if (is_xl) {
         if (is_inpaint) {
             return VERSION_SDXL_INPAINT;
         }
+        if (is_ip2p) {
+            return VERSION_SDXL_PIX2PIX;
+        }
         return VERSION_SDXL;
     }
 
@@ -1538,6 +1719,9 @@ SDVersion ModelLoader::get_sd_version() {
         if (is_inpaint) {
             return VERSION_SD1_INPAINT;
         }
+        if (is_ip2p) {
+            return VERSION_SD1_PIX2PIX;
+        }
         return VERSION_SD1;
     } else if (token_embedding_weight.ne[0] == 1024) {
         if (is_inpaint) {
@@ -1595,7 +1779,7 @@ ggml_type ModelLoader::get_diffusion_model_wtype() {
             continue;
         }
 
-        if (tensor_storage.name.find("model.diffusion_model.") == std::string::npos) {
+        if (tensor_storage.name.find("model.diffusion_model.") == std::string::npos && tensor_storage.name.find("unet.") == std::string::npos) {
             continue;
         }
 
@@ -1635,11 +1819,20 @@ ggml_type ModelLoader::get_vae_wtype() {
 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) {
-                if (tensor_storage.name == pair.first) {
-                    if (tensor_should_be_converted(tensor_storage, wtype)) {
-                        pair.second = wtype;
+                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;
                 }
             }
@@ -1748,9 +1941,12 @@ 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();
+        bool partial     = false;
         for (auto& tensor_storage : processed_tensor_storages) {
             if (tensor_storage.file_index != file_index) {
+                ++tensor_count;
                 continue;
             }
             ggml_tensor* dst_tensor = NULL;
@@ -1762,6 +1958,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
             }
 
             if (dst_tensor == NULL) {
+                ++tensor_count;
                 continue;
             }
 
@@ -1771,7 +1968,12 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                 // for the CPU and Metal backend, we can copy directly into the tensor
                 if (tensor_storage.type == dst_tensor->type) {
                     GGML_ASSERT(ggml_nbytes(dst_tensor) == tensor_storage.nbytes());
-                    read_data(tensor_storage, (char*)dst_tensor->data, nbytes_to_read);
+                    if (tensor_storage.is_f64 || tensor_storage.is_i64) {
+                        read_buffer.resize(tensor_storage.nbytes_to_read());
+                        read_data(tensor_storage, (char*)read_buffer.data(), nbytes_to_read);
+                    } else {
+                        read_data(tensor_storage, (char*)dst_tensor->data, nbytes_to_read);
+                    }
 
                     if (tensor_storage.is_bf16) {
                         // inplace op
@@ -1782,9 +1984,13 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     } 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 if (tensor_storage.is_f64) {
+                        f64_to_f32_vec((double*)read_buffer.data(), (float*)dst_tensor->data, tensor_storage.nelements());
+                    } else if (tensor_storage.is_i64) {
+                        i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)dst_tensor->data, tensor_storage.nelements());
                     }
                 } else {
-                    read_buffer.resize(tensor_storage.nbytes());
+                    read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
                     read_data(tensor_storage, (char*)read_buffer.data(), nbytes_to_read);
 
                     if (tensor_storage.is_bf16) {
@@ -1796,13 +2002,19 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     } 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());
+                    } else if (tensor_storage.is_f64) {
+                        // inplace op
+                        f64_to_f32_vec((double*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                    } else if (tensor_storage.is_i64) {
+                        // inplace op
+                        i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)read_buffer.data(), tensor_storage.nelements());
                     }
 
                     convert_tensor((void*)read_buffer.data(), tensor_storage.type, dst_tensor->data,
                                    dst_tensor->type, (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
                 }
             } else {
-                read_buffer.resize(tensor_storage.nbytes());
+                read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
                 read_data(tensor_storage, (char*)read_buffer.data(), nbytes_to_read);
 
                 if (tensor_storage.is_bf16) {
@@ -1814,6 +2026,12 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                 } 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());
+                } else if (tensor_storage.is_f64) {
+                    // inplace op
+                    f64_to_f32_vec((double*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                } else if (tensor_storage.is_i64) {
+                    // inplace op
+                    i64_to_i32_vec((int64_t*)read_buffer.data(), (int32_t*)read_buffer.data(), tensor_storage.nelements());
                 }
 
                 if (tensor_storage.type == dst_tensor->type) {
@@ -1828,12 +2046,21 @@ 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));
                 }
             }
+            size_t tensor_max = processed_tensor_storages.size();
+            int64_t t2        = ggml_time_ms();
+            pretty_progress(++tensor_count, tensor_max, (t2 - t1) / 1000.0f);
+            t1      = t2;
+            partial = tensor_count != tensor_max;
         }
 
         if (zip != NULL) {
             zip_close(zip);
         }
 
+        if (partial) {
+            printf("\n");
+        }
+
         if (!success) {
             break;
         }
@@ -1894,9 +2121,6 @@ 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;
@@ -1914,6 +2138,41 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
     return true;
 }
 
+std::vector<std::pair<std::string, ggml_type>> parse_tensor_type_rules(const std::string& tensor_type_rules) {
+    std::vector<std::pair<std::string, ggml_type>> result;
+    for (const auto& item : splitString(tensor_type_rules, ',')) {
+        if (item.size() == 0)
+            continue;
+        std::string::size_type pos = item.find('=');
+        if (pos == std::string::npos) {
+            LOG_WARN("ignoring invalid quant override \"%s\"", item.c_str());
+            continue;
+        }
+        std::string tensor_pattern = item.substr(0, pos);
+        std::string type_name      = item.substr(pos + 1);
+
+        ggml_type tensor_type = GGML_TYPE_COUNT;
+
+        if (type_name == "f32") {
+            tensor_type = GGML_TYPE_F32;
+        } else {
+            for (size_t i = 0; i < SD_TYPE_COUNT; i++) {
+                auto trait = ggml_get_type_traits((ggml_type)i);
+                if (trait->to_float && trait->type_size && type_name == trait->type_name) {
+                    tensor_type = (ggml_type)i;
+                }
+            }
+        }
+
+        if (tensor_type != GGML_TYPE_COUNT) {
+            result.emplace_back(tensor_pattern, tensor_type);
+        } else {
+            LOG_WARN("ignoring invalid quant override \"%s\"", item.c_str());
+        }
+    }
+    return result;
+}
+
 bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type) {
     const std::string& name = tensor_storage.name;
     if (type != GGML_TYPE_COUNT) {
@@ -1945,7 +2204,7 @@ bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage
     return false;
 }
 
-bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type) {
+bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type, const std::string& tensor_type_rules_str) {
     auto backend    = ggml_backend_cpu_init();
     size_t mem_size = 1 * 1024 * 1024;  // for padding
     mem_size += tensor_storages.size() * ggml_tensor_overhead();
@@ -1955,12 +2214,23 @@ bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type
 
     gguf_context* gguf_ctx = gguf_init_empty();
 
+    auto tensor_type_rules = parse_tensor_type_rules(tensor_type_rules_str);
+
     auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
         const std::string& name = tensor_storage.name;
+        ggml_type tensor_type   = tensor_storage.type;
+        ggml_type dst_type      = type;
 
-        ggml_type tensor_type = tensor_storage.type;
-        if (tensor_should_be_converted(tensor_storage, type)) {
-            tensor_type = type;
+        for (const auto& tensor_type_rule : tensor_type_rules) {
+            std::regex pattern(tensor_type_rule.first);
+            if (std::regex_search(name, pattern)) {
+                dst_type = tensor_type_rule.second;
+                break;
+            }
+        }
+
+        if (tensor_should_be_converted(tensor_storage, dst_type)) {
+            tensor_type = dst_type;
         }
 
         ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
@@ -2019,7 +2289,7 @@ int64_t ModelLoader::get_params_mem_size(ggml_backend_t backend, ggml_type type)
     return mem_size;
 }
 
-bool convert(const char* input_path, const char* vae_path, const char* output_path, sd_type_t output_type) {
+bool convert(const char* input_path, const char* vae_path, const char* output_path, sd_type_t output_type, const char* tensor_type_rules) {
     ModelLoader model_loader;
 
     if (!model_loader.init_from_file(input_path)) {
@@ -2033,6 +2303,6 @@ bool convert(const char* input_path, const char* vae_path, const char* output_pa
             return false;
         }
     }
-    bool success = model_loader.save_to_gguf_file(output_path, (ggml_type)output_type);
+    bool success = model_loader.save_to_gguf_file(output_path, (ggml_type)output_type, tensor_type_rules);
     return success;
 }

model.h

@@ -12,6 +12,7 @@
 
 #include "ggml-backend.h"
 #include "ggml.h"
+#include "gguf.h"
 #include "json.hpp"
 #include "zip.h"
 
@@ -20,10 +21,12 @@
 enum SDVersion {
     VERSION_SD1,
     VERSION_SD1_INPAINT,
+    VERSION_SD1_PIX2PIX,
     VERSION_SD2,
     VERSION_SD2_INPAINT,
     VERSION_SDXL,
     VERSION_SDXL_INPAINT,
+    VERSION_SDXL_PIX2PIX,
     VERSION_SVD,
     VERSION_SD3,
     VERSION_FLUX,
@@ -46,7 +49,7 @@ static inline bool sd_version_is_sd3(SDVersion version) {
 }
 
 static inline bool sd_version_is_sd1(SDVersion version) {
-    if (version == VERSION_SD1 || version == VERSION_SD1_INPAINT) {
+    if (version == VERSION_SD1 || version == VERSION_SD1_INPAINT || version == VERSION_SD1_PIX2PIX) {
         return true;
     }
     return false;
@@ -60,7 +63,7 @@ static inline bool sd_version_is_sd2(SDVersion version) {
 }
 
 static inline bool sd_version_is_sdxl(SDVersion version) {
-    if (version == VERSION_SDXL || version == VERSION_SDXL_INPAINT) {
+    if (version == VERSION_SDXL || version == VERSION_SDXL_INPAINT || version == VERSION_SDXL_PIX2PIX) {
         return true;
     }
     return false;
@@ -80,6 +83,14 @@ static inline bool sd_version_is_dit(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_unet_edit(SDVersion version) {
+    return version == VERSION_SD1_PIX2PIX || version == VERSION_SDXL_PIX2PIX;
+}
+
+static bool sd_version_is_inpaint_or_unet_edit(SDVersion version) {
+    return sd_version_is_unet_edit(version) || sd_version_is_inpaint(version);
+}
+
 enum PMVersion {
     PM_VERSION_1,
     PM_VERSION_2,
@@ -91,6 +102,8 @@ struct TensorStorage {
     bool is_bf16            = false;
     bool is_f8_e4m3         = false;
     bool is_f8_e5m2         = false;
+    bool is_f64             = false;
+    bool is_i64             = false;
     int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
     int n_dims              = 0;
 
@@ -122,6 +135,8 @@ struct TensorStorage {
     int64_t nbytes_to_read() const {
         if (is_bf16 || is_f8_e4m3 || is_f8_e5m2) {
             return nbytes() / 2;
+        } else if (is_f64 || is_i64) {
+            return nbytes() * 2;
         } else {
             return nbytes();
         }
@@ -172,6 +187,10 @@ struct TensorStorage {
             type_name = "f8_e4m3";
         } else if (is_f8_e5m2) {
             type_name = "f8_e5m2";
+        } else if (is_f64) {
+            type_name = "f64";
+        } else if (is_i64) {
+            type_name = "i64";
         }
         ss << name << " | " << type_name << " | ";
         ss << n_dims << " [";
@@ -209,6 +228,7 @@ public:
     std::map<std::string, enum ggml_type> tensor_storages_types;
 
     bool init_from_file(const std::string& file_path, const std::string& prefix = "");
+    bool model_is_unet();
     SDVersion get_sd_version();
     ggml_type get_sd_wtype();
     ggml_type get_conditioner_wtype();
@@ -220,7 +240,7 @@ public:
                       ggml_backend_t backend,
                       std::set<std::string> ignore_tensors = {});
 
-    bool save_to_gguf_file(const std::string& file_path, ggml_type type);
+    bool save_to_gguf_file(const std::string& file_path, ggml_type type, const std::string& tensor_type_rules);
     bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
     int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
     ~ModelLoader() = default;

stable-diffusion.h

@@ -30,7 +30,8 @@ extern "C" {
 
 enum rng_type_t {
     STD_DEFAULT_RNG,
-    CUDA_RNG
+    CUDA_RNG,
+    RNG_TYPE_COUNT
 };
 
 enum sample_method_t {
@@ -44,7 +45,9 @@ enum sample_method_t {
     IPNDM,
     IPNDM_V,
     LCM,
-    N_SAMPLE_METHODS
+    DDIM_TRAILING,
+    TCD,
+    SAMPLE_METHOD_COUNT
 };
 
 enum schedule_t {
@@ -54,7 +57,7 @@ enum schedule_t {
     EXPONENTIAL,
     AYS,
     GITS,
-    N_SCHEDULES
+    SCHEDULE_COUNT
 };
 
 // same as enum ggml_type
@@ -65,41 +68,42 @@ 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_TQ1_0    = 34,
-    SD_TYPE_TQ2_0    = 35,
-    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);
-
 enum sd_log_level_t {
     SD_LOG_DEBUG,
     SD_LOG_INFO,
@@ -107,6 +111,98 @@ enum sd_log_level_t {
     SD_LOG_ERROR
 };
 
+typedef struct {
+    const char* model_path;
+    const char* clip_l_path;
+    const char* clip_g_path;
+    const char* t5xxl_path;
+    const char* diffusion_model_path;
+    const char* vae_path;
+    const char* taesd_path;
+    const char* control_net_path;
+    const char* lora_model_dir;
+    const char* embedding_dir;
+    const char* stacked_id_embed_dir;
+    bool vae_decode_only;
+    bool vae_tiling;
+    bool free_params_immediately;
+    int n_threads;
+    enum sd_type_t wtype;
+    enum rng_type_t rng_type;
+    enum schedule_t schedule;
+    bool keep_clip_on_cpu;
+    bool keep_control_net_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_ctx_params_t;
+
+typedef struct {
+    uint32_t width;
+    uint32_t height;
+    uint32_t channel;
+    uint8_t* data;
+} sd_image_t;
+
+typedef struct {
+    int* layers;
+    size_t layer_count;
+    float layer_start;
+    float layer_end;
+    float scale;
+} sd_slg_params_t;
+
+typedef struct {
+    float txt_cfg;
+    float img_cfg;
+    float min_cfg;
+    float distilled_guidance;
+    sd_slg_params_t slg;
+} sd_guidance_params_t;
+
+typedef struct {
+    const char* prompt;
+    const char* negative_prompt;
+    int clip_skip;
+    sd_guidance_params_t guidance;
+    sd_image_t init_image;
+    sd_image_t* ref_images;
+    int ref_images_count;
+    sd_image_t mask_image;
+    int width;
+    int height;
+    enum sample_method_t sample_method;
+    int sample_steps;
+    float eta;
+    float strength;
+    int64_t seed;
+    int batch_count;
+    const sd_image_t* control_cond;
+    float control_strength;
+    float style_strength;
+    bool normalize_input;
+    const char* input_id_images_path;
+} sd_img_gen_params_t;
+
+typedef struct {
+    sd_image_t init_image;
+    int width;
+    int height;
+    sd_guidance_params_t guidance;
+    enum sample_method_t sample_method;
+    int sample_steps;
+    float strength;
+    int64_t seed;
+    int video_frames;
+    int motion_bucket_id;
+    int fps;
+    float augmentation_level;
+} sd_vid_gen_params_t;
+
+typedef struct sd_ctx_t sd_ctx_t;
+
 typedef void (*sd_log_cb_t)(enum sd_log_level_t level, const char* text, void* data);
 typedef void (*sd_progress_cb_t)(int step, int steps, float time, void* data);
 
@@ -115,103 +211,27 @@ SD_API void sd_set_progress_callback(sd_progress_cb_t cb, void* data);
 SD_API int32_t get_num_physical_cores();
 SD_API const char* sd_get_system_info();
 
-typedef struct {
-    uint32_t width;
-    uint32_t height;
-    uint32_t channel;
-    uint8_t* data;
-} sd_image_t;
+SD_API const char* sd_type_name(enum sd_type_t type);
+SD_API enum sd_type_t str_to_sd_type(const char* str);
+SD_API const char* sd_rng_type_name(enum rng_type_t rng_type);
+SD_API enum rng_type_t str_to_rng_type(const char* str);
+SD_API const char* sd_sample_method_name(enum sample_method_t sample_method);
+SD_API enum sample_method_t str_to_sample_method(const char* str);
+SD_API const char* sd_schedule_name(enum schedule_t schedule);
+SD_API enum schedule_t str_to_schedule(const char* str);
 
-typedef struct sd_ctx_t sd_ctx_t;
-
-SD_API sd_ctx_t* new_sd_ctx(const char* model_path,
-                            const char* clip_l_path,
-                            const char* clip_g_path,
-                            const char* t5xxl_path,
-                            const char* diffusion_model_path,
-                            const char* vae_path,
-                            const char* taesd_path,
-                            const char* control_net_path_c_str,
-                            const char* lora_model_dir,
-                            const char* embed_dir_c_str,
-                            const char* stacked_id_embed_dir_c_str,
-                            bool vae_decode_only,
-                            bool vae_tiling,
-                            bool free_params_immediately,
-                            int n_threads,
-                            enum sd_type_t wtype,
-                            enum rng_type_t rng_type,
-                            enum schedule_t s,
-                            bool keep_clip_on_cpu,
-                            bool keep_control_net_cpu,
-                            bool keep_vae_on_cpu,
-                            bool diffusion_flash_attn);
+SD_API void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params);
+SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);
 
+SD_API sd_ctx_t* new_sd_ctx(const sd_ctx_params_t* sd_ctx_params);
 SD_API void free_sd_ctx(sd_ctx_t* sd_ctx);
 
-SD_API sd_image_t* txt2img(sd_ctx_t* sd_ctx,
-                           const char* prompt,
-                           const char* negative_prompt,
-                           int clip_skip,
-                           float cfg_scale,
-                           float guidance,
-                           int width,
-                           int height,
-                           enum sample_method_t sample_method,
-                           int sample_steps,
-                           int64_t seed,
-                           int batch_count,
-                           const sd_image_t* control_cond,
-                           float control_strength,
-                           float style_strength,
-                           bool normalize_input,
-                           const char* input_id_images_path,
-                           int* skip_layers,
-                           size_t skip_layers_count,
-                           float slg_scale,
-                           float skip_layer_start,
-                           float skip_layer_end);
+SD_API void sd_img_gen_params_init(sd_img_gen_params_t* sd_img_gen_params);
+SD_API char* sd_img_gen_params_to_str(const sd_img_gen_params_t* sd_img_gen_params);
+SD_API sd_image_t* generate_image(sd_ctx_t* sd_ctx, const sd_img_gen_params_t* sd_img_gen_params);
 
-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,
-                           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,
-                           const char* input_id_images_path,
-                           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,
-                           int width,
-                           int height,
-                           int video_frames,
-                           int motion_bucket_id,
-                           int fps,
-                           float augmentation_level,
-                           float min_cfg,
-                           float cfg_scale,
-                           enum sample_method_t sample_method,
-                           int sample_steps,
-                           float strength,
-                           int64_t seed);
+SD_API void sd_vid_gen_params_init(sd_vid_gen_params_t* sd_vid_gen_params);
+SD_API sd_image_t* generate_video(sd_ctx_t* sd_ctx, const sd_vid_gen_params_t* sd_vid_gen_params);  // broken
 
 typedef struct upscaler_ctx_t upscaler_ctx_t;
 
@@ -221,7 +241,11 @@ 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);
 
-SD_API bool convert(const char* input_path, const char* vae_path, const char* output_path, enum sd_type_t output_type);
+SD_API bool convert(const char* input_path,
+                    const char* vae_path,
+                    const char* output_path,
+                    enum sd_type_t output_type,
+                    const char* tensor_type_rules);
 
 SD_API uint8_t* preprocess_canny(uint8_t* img,
                                  int width,

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);
+                }
             }
         }
     }
@@ -579,6 +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;
@@ -739,15 +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);
@@ -767,7 +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);
 
@@ -776,10 +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);
     }
@@ -877,9 +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);
 
         {
@@ -906,14 +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() {
@@ -934,8 +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);
             }
@@ -944,7 +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);

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));
         }
     }
 
@@ -190,9 +198,10 @@ struct TinyAutoEncoder : public GGMLRunner {
     TinyAutoEncoder(ggml_backend_t backend,
                     std::map<std::string, enum ggml_type>& tensor_types,
                     const std::string prefix,
-                    bool decoder_only = true)
+                    bool decoder_only = true,
+                    SDVersion version = VERSION_SD1)
         : decode_only(decoder_only),
-          taesd(decode_only),
+          taesd(decoder_only, version),
           GGMLRunner(backend) {
         taesd.init(params_ctx, tensor_types, prefix);
     }

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

@@ -207,6 +207,8 @@ public:
         }
         if (sd_version_is_inpaint(version)) {
             in_channels = 9;
+        } else if (sd_version_is_unet_edit(version)) {
+            in_channels = 8;
         }
 
         // dims is always 2

upscaler.cpp

@@ -15,7 +15,7 @@ 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
@@ -28,6 +28,10 @@ struct UpscalerGGML {
         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);

util.cpp

@@ -112,19 +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);
 
@@ -348,7 +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
@@ -428,10 +441,6 @@ const char* sd_get_system_info() {
     return buffer;
 }
 
-const char* sd_type_name(enum sd_type_t type) {
-    return ggml_type_name((ggml_type)type);
-}
-
 sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image) {
     sd_image_f32_t converted_image;
     converted_image.width   = image.width;

util.h

@@ -7,6 +7,9 @@
 
 #include "stable-diffusion.h"
 
+#define SAFE_STR(s) ((s) ? (s) : "")
+#define BOOL_STR(b) ((b) ? "true" : "false")
+
 bool ends_with(const std::string& str, const std::string& ending);
 bool starts_with(const std::string& str, const std::string& start);
 bool contains(const std::string& str, const std::string& substr);