feat: allow models to run without all text encoder(s) (#645)

* fix: resolve precision issues in SDXL VAE under fp16

* add --force-sdxl-vae-conv-scale option

* update docs

.clang-tidy

@@ -0,0 +1,10 @@
+Checks: >
+  modernize-make-shared,
+  modernize-use-nullptr,
+  modernize-use-override,
+  modernize-pass-by-value,
+  modernize-return-braced-init-list,
+  modernize-deprecated-headers,
+HeaderFilterRegex: '^$'
+WarningsAsErrors: ''
+FormatStyle: none

.github/ISSUE_TEMPLATE/bug_report.yml

@@ -0,0 +1,73 @@
+name: 🐞 Bug Report
+description: Report a bug or unexpected behavior
+title: "[Bug] "
+labels: ["bug"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Please use this template and include as many details as possible to help us reproduce and fix the issue.
+  - type: textarea
+    id: commit
+    attributes:
+      label: Git commit
+      description: Which commit are you trying to compile?
+      placeholder: |
+        $git rev-parse HEAD
+        40a6a8710ec15b1b5db6b5a098409f6bc8f654a4
+    validations:
+      required: true
+  - type: input
+    id: os
+    attributes:
+      label: Operating System & Version
+      placeholder: e.g. “Ubuntu 22.04”, “Windows 11 23H2”, “macOS 14.3”
+    validations:
+      required: true
+  - type: dropdown
+    id: backends
+    attributes:
+        label: GGML backends
+        description: Which GGML backends do you know to be affected?
+        options: [CPU, CUDA, HIP, Metal, Musa, SYCL, Vulkan, OpenCL]
+        multiple: true
+    validations:
+      required: true
+  - type: input
+    id: cmd_arguments
+    attributes:
+      label: Command-line arguments used
+      placeholder: The full command line you ran (with all flags)
+    validations:
+      required: true
+  - type: textarea
+    id: steps_to_reproduce
+    attributes:
+      label: Steps to reproduce
+      placeholder: A step-by-step list of what you did
+    validations:
+      required: true
+  - type: textarea
+    id: expected_behavior
+    attributes:
+      label: What you expected to happen
+      placeholder: Describe the expected behavior or result
+    validations:
+      required: true
+  - type: textarea
+    id: actual_behavior
+    attributes:
+      label: What actually happened
+      placeholder: Describe what you saw instead (errors, logs, crash, etc.)
+    validations:
+      required: true
+  - type: textarea
+    id: logs_and_errors
+    attributes:
+      label: Logs / error messages / stack trace
+      placeholder: Paste complete logs or error output
+  - type: textarea
+    id: additional_info
+    attributes:
+      label: Additional context / environment details
+      placeholder: e.g. CPU model, GPU, RAM, model file versions, quantization type, etc.

.github/ISSUE_TEMPLATE/feature_request.yml

@@ -0,0 +1,33 @@
+name: 💡 Feature Request
+description: Suggest a new feature or improvement
+title: "[Feature] "
+labels: ["enhancement"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thank you for suggesting an improvement! Please fill in the fields below.
+  - type: input
+    id: summary
+    attributes:
+      label: Feature Summary
+      placeholder: A one-line summary of the feature you’d like
+    validations:
+      required: true
+  - type: textarea
+    id: description
+    attributes:
+      label: Detailed Description
+      placeholder: What problem does this solve? How do you expect it to work?
+    validations:
+      required: true
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives you considered
+      placeholder: Any alternative designs or workarounds you tried
+  - type: textarea
+    id: additional_context
+    attributes:
+      label: Additional context
+      placeholder: Any extra information (use cases, related functionalities, constraints)

.github/workflows/build.yml

@@ -149,7 +149,7 @@ jobs:
     runs-on: windows-2025
 
     env:
-      VULKAN_VERSION: 1.3.261.1
+      VULKAN_VERSION: 1.4.328.1
 
     strategy:
       matrix:
@@ -199,9 +199,9 @@ jobs:
           version: 1.11.1
       - name: Install Vulkan SDK
         id: get_vulkan
-        if: ${{ matrix.build == 'vulkan' }}
+        if: ${{ matrix.build == 'vulkan' }} https://sdk.lunarg.com/sdk/download/1.4.328.1/windows/vulkansdk-windows-X64-1.4.328.1.exe
         run: |
-          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/VulkanSDK-${env:VULKAN_VERSION}-Installer.exe"
+          curl.exe -o $env:RUNNER_TEMP/VulkanSDK-Installer.exe -L "https://sdk.lunarg.com/sdk/download/${env:VULKAN_VERSION}/windows/vulkansdk-windows-X64-${env:VULKAN_VERSION}.exe"
           & "$env:RUNNER_TEMP\VulkanSDK-Installer.exe" --accept-licenses --default-answer --confirm-command install
           Add-Content $env:GITHUB_ENV "VULKAN_SDK=C:\VulkanSDK\${env:VULKAN_VERSION}"
           Add-Content $env:GITHUB_PATH "C:\VulkanSDK\${env:VULKAN_VERSION}\bin"
@@ -254,7 +254,7 @@ jobs:
 
       - name: Copy and pack Cuda runtime
         id: pack_cuda_runtime
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' && matrix.build == 'cuda12' ) || github.event.inputs.create_release == 'true' }}
+        if: ${{ matrix.build == 'cuda12' && (github.event_name == 'push' && github.ref == 'refs/heads/master' || github.event.inputs.create_release == 'true') }}
         run: |
           echo "Cuda install location: ${{steps.cuda-toolkit.outputs.CUDA_PATH}}"
           $dst='.\build\bin\cudart\'
@@ -262,7 +262,7 @@ jobs:
           7z a cudart-sd-bin-win-cu12-x64.zip $dst\*
 
       - name: Upload Cuda runtime
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' && matrix.build == 'cuda12' ) || github.event.inputs.create_release == 'true' }}
+        if: ${{ matrix.build == 'cuda12' && (github.event_name == 'push' && github.ref == 'refs/heads/master' || github.event.inputs.create_release == 'true') }}
         uses: actions/upload-artifact@v4
         with:
           name: sd-cudart-sd-bin-win-cu12-x64.zip
@@ -288,6 +288,11 @@ jobs:
       - windows-latest-cmake
 
     steps:
+      - name: Clone
+        uses: actions/checkout@v3
+        with:
+          fetch-depth: 0
+
       - name: Download artifacts
         id: download-artifact
         uses: actions/download-artifact@v4
@@ -296,20 +301,27 @@ jobs:
           pattern: sd-*
           merge-multiple: true
 
+      - name: Get commit count
+        id: commit_count
+        run: |
+          echo "count=$(git rev-list --count HEAD)" >> $GITHUB_OUTPUT
+
       - name: Get commit hash
         id: commit
         uses: pr-mpt/actions-commit-hash@v2
 
       - name: Create release
         id: create_release
+        if: ${{ github.event_name == 'workflow_dispatch' || github.ref_name == 'master' }}
         uses: anzz1/action-create-release@v1
         env:
           GITHUB_TOKEN: ${{ secrets.GITHUB_TOKEN }}
         with:
-          tag_name: ${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}
+          tag_name: ${{ format('{0}-{1}-{2}', env.BRANCH_NAME, steps.commit_count.outputs.count, steps.commit.outputs.short) }}
 
       - name: Upload release
         id: upload_release
+        if: ${{ github.event_name == 'workflow_dispatch' || github.ref_name == 'master' }}
         uses: actions/github-script@v3
         with:
           github-token: ${{secrets.GITHUB_TOKEN}}

.gitignore

@@ -1,10 +1,10 @@
 build*/
+cmake-build-*/
 test/
 .vscode/
+.idea/
 .cache/
 *.swp
-.vscode/
-.idea/
 *.bat
 *.bin
 *.exe

CMakeLists.txt

@@ -33,6 +33,7 @@ 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_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" OFF)
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
@@ -86,18 +87,21 @@ file(GLOB SD_LIB_SOURCES
     "*.hpp"
 )
 
-# we can get only one share lib
 if(SD_BUILD_SHARED_LIBS)
     message("-- Build shared library")
     message(${SD_LIB_SOURCES})
-    set(BUILD_SHARED_LIBS OFF)
+    if(NOT SD_BUILD_SHARED_GGML_LIB)
+        set(BUILD_SHARED_LIBS OFF)
+    endif()
     add_library(${SD_LIB} SHARED ${SD_LIB_SOURCES})
     add_definitions(-DSD_BUILD_SHARED_LIB)
     target_compile_definitions(${SD_LIB} PRIVATE -DSD_BUILD_DLL)
     set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 else()
     message("-- Build static library")
-    set(BUILD_SHARED_LIBS OFF)
+    if(NOT SD_BUILD_SHARED_GGML_LIB)
+        set(BUILD_SHARED_LIBS OFF)
+    endif()
     add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
 endif()
 
@@ -149,3 +153,7 @@ if (SD_BUILD_EXAMPLES)
     add_subdirectory(examples)
 endif()
 
+set(SD_PUBLIC_HEADERS stable-diffusion.h)
+set_target_properties(${SD_LIB} PROPERTIES PUBLIC_HEADER "${SD_PUBLIC_HEADERS}")
+
+install(TARGETS ${SD_LIB} LIBRARY PUBLIC_HEADER)

Dockerfile

@@ -1,16 +1,21 @@
 ARG UBUNTU_VERSION=22.04
 
-FROM ubuntu:$UBUNTU_VERSION as build
+FROM ubuntu:$UBUNTU_VERSION AS build
 
-RUN apt-get update && apt-get install -y build-essential git cmake
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake
 
 WORKDIR /sd.cpp
 
 COPY . .
 
-RUN mkdir build && cd build && cmake .. && cmake --build . --config Release
+RUN cmake . -B ./build
+RUN cmake --build ./build --config Release --parallel
 
-FROM ubuntu:$UBUNTU_VERSION as runtime
+FROM ubuntu:$UBUNTU_VERSION AS runtime
+
+RUN apt-get update && \
+    apt-get install --yes --no-install-recommends libgomp1 && \
+    apt-get clean
 
 COPY --from=build /sd.cpp/build/bin/sd /sd
 

Dockerfile.sycl

@@ -0,0 +1,19 @@
+ARG SYCL_VERSION=2025.1.0-0
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS build
+
+RUN apt-get update && apt-get install -y cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+RUN mkdir build && cd build && \
+    cmake .. -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DSD_SYCL=ON -DCMAKE_BUILD_TYPE=Release && \
+    cmake --build . --config Release -j$(nproc)
+
+FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS runtime
+
+COPY --from=build /sd.cpp/build/bin/sd /sd
+
+ENTRYPOINT [ "/sd" ]

README.md

@@ -4,11 +4,29 @@
 
 # stable-diffusion.cpp
 
+<div align="center">
+<a href="https://trendshift.io/repositories/9714" target="_blank"><img src="https://trendshift.io/api/badge/repositories/9714" alt="leejet%2Fstable-diffusion.cpp | Trendshift" style="width: 250px; height: 55px;" width="250" height="55"/></a>
+</div>
+
 Diffusion model(SD,Flux,Wan,...) inference in pure C/C++
 
 ***Note that this project is under active development. \
 API and command-line option may change frequently.***
 
+## 🔥Important News
+
+* **2025/10/13** 🚀 stable-diffusion.cpp now supports **Qwen-Image-Edit / Qwen-Image-Edit 2509**  
+  👉 Details: [PR #877](https://github.com/leejet/stable-diffusion.cpp/pull/877)
+
+* **2025/10/12** 🚀 stable-diffusion.cpp now supports **Qwen-Image**  
+  👉 Details: [PR #851](https://github.com/leejet/stable-diffusion.cpp/pull/851)
+
+* **2025/09/14** 🚀 stable-diffusion.cpp now supports **Wan2.1 Vace**  
+  👉 Details: [PR #819](https://github.com/leejet/stable-diffusion.cpp/pull/819)
+
+* **2025/09/06** 🚀 stable-diffusion.cpp now supports **Wan2.1 / Wan2.2**  
+  👉 Details: [PR #778](https://github.com/leejet/stable-diffusion.cpp/pull/778)
+
 ## Features
 
 - Plain C/C++ implementation based on [ggml](https://github.com/ggerganov/ggml), working in the same way as [llama.cpp](https://github.com/ggerganov/llama.cpp)
@@ -17,12 +35,13 @@ API and command-line option may change frequently.***
   - Image Models
     - SD1.x, SD2.x, [SD-Turbo](https://huggingface.co/stabilityai/sd-turbo)
     - SDXL, [SDXL-Turbo](https://huggingface.co/stabilityai/sdxl-turbo)
-      - !!!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).
     - [SD3/SD3.5](./docs/sd3.md)
     - [Flux-dev/Flux-schnell](./docs/flux.md)
     - [Chroma](./docs/chroma.md)
+    - [Qwen Image](./docs/qwen_image.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
+    - [Qwen Image Edit/Qwen Image Edit 2509](./docs/qwen_image_edit.md)
   - Video Models
     - [Wan2.1/Wan2.2](./docs/wan.md)
   - [PhotoMaker](https://github.com/TencentARC/PhotoMaker) support.
@@ -31,14 +50,22 @@ API and command-line option may change frequently.***
   - Latent Consistency Models support (LCM/LCM-LoRA)
   - Faster and memory efficient latent decoding with [TAESD](https://github.com/madebyollin/taesd)
   - Upscale images generated with [ESRGAN](https://github.com/xinntao/Real-ESRGAN)
-- 16-bit, 32-bit float support
-- 2-bit, 3-bit, 4-bit, 5-bit and 8-bit integer quantization support
-- 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, 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!
+- Supported backends
+  - CPU (AVX, AVX2 and AVX512 support for x86 architectures)
+  - CUDA
+  - Vulkan
+  - Metal
+  - OpenCL
+  - SYCL
+- Supported weight formats
+  - Pytorch checkpoint (`.ckpt` or `.pth`)
+  - Safetensors (`./safetensors`)
+  - GGUF (`.gguf`)
+- Supported platforms
+    - Linux
+    - Mac OS
+    - Windows
+    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))
 - Flash Attention for memory usage optimization
 - Negative prompt
 - [stable-diffusion-webui](https://github.com/AUTOMATIC1111/stable-diffusion-webui) style tokenizer (not all the features, only token weighting for now)
@@ -54,371 +81,45 @@ API and command-line option may change frequently.***
     - [`LCM`](https://github.com/AUTOMATIC1111/stable-diffusion-webui/issues/13952)
 - Cross-platform reproducibility (`--rng cuda`, consistent with the `stable-diffusion-webui GPU RNG`)
 - Embedds generation parameters into png output as webui-compatible text string
-- Supported platforms
-    - Linux
-    - Mac OS
-    - Windows
-    - Android (via Termux, [Local Diffusion](https://github.com/rmatif/Local-Diffusion))
 
-### TODO
+## Quick Start
 
-- [ ] More sampling methods
-- [ ] Make inference faster
-    - The current implementation of ggml_conv_2d is slow and has high memory usage
-- [ ] Continuing to reduce memory usage (quantizing the weights of ggml_conv_2d)
-- [ ] Implement Inpainting support
+### Get the sd executable
 
-## Usage
+- Download pre-built binaries from the [releases page](https://github.com/leejet/stable-diffusion.cpp/releases)
+- Or build from source by following the [build guide](./docs/build.md)
 
-For most users, you can download the built executable program from the latest [release](https://github.com/leejet/stable-diffusion.cpp/releases/latest).
-If the built product does not meet your requirements, you can choose to build it manually.
+### Download model weights
 
-### Get the Code
-
-```
-git clone --recursive https://github.com/leejet/stable-diffusion.cpp
-cd stable-diffusion.cpp
-```
-
-- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
-
-```
-cd stable-diffusion.cpp
-git pull origin master
-git submodule init
-git submodule update
-```
-
-### Download weights
-
-- download original weights(.ckpt or .safetensors). For example
-    - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
+- download weights(.ckpt or .safetensors or .gguf). For example
     - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
-    - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
-    - Stable Diffusion 3 2B from https://huggingface.co/stabilityai/stable-diffusion-3-medium
 
-    ```shell
-    curl -L -O https://huggingface.co/CompVis/stable-diffusion-v-1-4-original/resolve/main/sd-v1-4.ckpt
-    # curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
-    # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-2-1/resolve/main/v2-1_768-nonema-pruned.safetensors
-    # curl -L -O https://huggingface.co/stabilityai/stable-diffusion-3-medium/resolve/main/sd3_medium_incl_clips_t5xxlfp16.safetensors
+    ```sh
+    curl -L -O https://huggingface.co/runwayml/stable-diffusion-v1-5/resolve/main/v1-5-pruned-emaonly.safetensors
     ```
 
-### Build
-
-#### Build from scratch
-
-```shell
-mkdir build
-cd build
-cmake ..
-cmake --build . --config Release
-```
-
-##### Using OpenBLAS
-
-```
-cmake .. -DGGML_OPENBLAS=ON
-cmake --build . --config Release
-```
-
-##### 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_CUDA=ON
-cmake --build . --config Release
-```
-
-##### Using HipBLAS
-This provides BLAS acceleration using the ROCm cores of your AMD GPU. Make sure to have the ROCm toolkit installed.
-
-Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
-
-```
-export GFX_NAME=$(rocminfo | grep -m 1 -E "gfx[^0]{1}" | sed -e 's/ *Name: *//' | awk '{$1=$1; print}' || echo "rocminfo missing")
-echo $GFX_NAME
-cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON
-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
-
-Using Metal makes the computation run on the GPU. Currently, there are some issues with Metal when performing operations on very large matrices, making it highly inefficient at the moment. Performance improvements are expected in the near future.
-
-```
-cmake .. -DSD_METAL=ON
-cmake --build . --config Release
-```
-
-##### Using Vulkan
-
-Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
-
-```
-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).
-
-```
-# Export relevant ENV variables
-source /opt/intel/oneapi/setvars.sh
-
-# Option 1: Use FP32 (recommended for better performance in most cases)
-cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
-
-# Option 2: Use FP16
-cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
-
-cmake --build . --config Release
-```
-
-Example of text2img by using SYCL backend:
-
-- download `stable-diffusion` model weight, refer to [download-weight](#download-weights).
-
-- run `./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors --cfg-scale 5 --steps 30 --sampling-method euler  -H 1024 -W 1024 --seed 42 -p "fantasy medieval village world inside a glass sphere , high detail, fantasy, realistic, light effect, hyper detail, volumetric lighting, cinematic, macro, depth of field, blur, red light and clouds from the back, highly detailed epic cinematic concept art cg render made in maya, blender and photoshop, octane render, excellent composition, dynamic dramatic cinematic lighting, aesthetic, very inspirational, world inside a glass sphere by james gurney by artgerm with james jean, joe fenton and tristan eaton by ross tran, fine details, 4k resolution"`
-
-<p align="center">
-  <img src="./assets/sycl_sd3_output.png" width="360x">
-</p>
-
-
-
-##### Using Flash Attention
-
-Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
-eg.:
- - flux 768x768 ~600mb
- - SD2 768x768 ~1400mb
-
-For most backends, it slows things down, but for cuda it generally speeds it up too.
-At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
-
-Run by adding `--diffusion-fa` to the arguments and watch for:
-```
-[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
-```
-and the compute buffer shrink in the debug log:
-```
-[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
-```
-
-### Run
-
-```
-usage: ./bin/sd [arguments]
-
-arguments:
-  -h, --help                         show this help message and exit
-  -M, --mode [MODE]                  run mode, one of: [img_gen, vid_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
-  --offload-to-cpu                   place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
-  -m, --model [MODEL]                path to full model
-  --diffusion-model                  path to the standalone diffusion model
-  --high-noise-diffusion-model       path to the standalone high noise diffusion model
-  --clip_l                           path to the clip-l text encoder
-  --clip_g                           path to the clip-g text encoder
-  --clip_vision                      path to the clip-vision 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
-  --embd-dir [EMBEDDING_PATH]        path to embeddings
-  --stacked-id-embd-dir [DIR]        path to PHOTOMAKER stacked id embeddings
-  --input-id-images-dir [DIR]        path to PHOTOMAKER input id images dir
-  --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 (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 init image, required by img2img
-  --mask [MASK]                      path to the mask image, required by img2img with mask
-  -i, --end-img [IMAGE]              path to the end image, required by flf2v
-  --control-image [IMAGE]            path to image condition, control net
-  -r, --ref-image [PATH]             reference image for Flux Kontext models (can be used multiple times)
-  --increase-ref-index               automatically increase the indices of references images based on the order they are listed (starting with 1).
-  -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)
-  --scheduler {discrete, karras, exponential, ays, gits} Denoiser sigma scheduler (default: discrete)
-  --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)
-  --high-noise-cfg-scale SCALE       (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale SCALE   (high noise) image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
-  --high-noise-guidance SCALE        (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-slg-scale SCALE       (high noise) 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
-  --high-noise-eta SCALE             (high noise) eta in DDIM, only for DDIM and TCD: (default: 0)
-  --high-noise-skip-layers LAYERS    (high noise) Layers to skip for SLG steps: (default: [7,8,9])
-  --high-noise-skip-layer-start      (high noise) SLG enabling point: (default: 0.01)
-  --high-noise-skip-layer-end END    (high noise) SLG disabling point: (default: 0.2)
-  --high-noise-scheduler {discrete, karras, exponential, ays, gits} Denoiser sigma scheduler (default: discrete)
-  --high-noise-sampling-method {euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd}
-                                     (high noise) sampling method (default: "euler_a")
-  --high-noise-steps  STEPS          (high noise) number of sample steps (default: -1 = auto)
-                                     SLG will be enabled at step int([STEPS]*[START]) and disabled at int([STEPS]*[END])
-  --strength STRENGTH                strength for noising/unnoising (default: 0.75)
-  --style-ratio STYLE-RATIO          strength for keeping input identity (default: 20)
-  --control-strength STRENGTH        strength to apply Control Net (default: 0.9)
-                                     1.0 corresponds to full destruction of information in init image
-  -H, --height H                     image height, in pixel space (default: 512)
-  -W, --width W                      image width, in pixel space (default: 512)
-  --rng {std_default, cuda}          RNG (default: cuda)
-  -s SEED, --seed SEED               RNG seed (default: 42, use random seed for < 0)
-  -b, --batch-count COUNT            number of images to generate
-  --clip-skip N                      ignore last_dot_pos layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1)
-                                     <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
-  --vae-tiling                       process vae in tiles to reduce memory usage
-  --vae-on-cpu                       keep vae in cpu (for low vram)
-  --clip-on-cpu                      keep clip in cpu (for low vram)
-  --diffusion-fa                     use flash attention in the diffusion model (for low vram)
-                                     Might lower quality, since it implies converting k and v to f16.
-                                     This might crash if it is not supported by the backend.
-  --diffusion-conv-direct            use Conv2d direct in the diffusion model
-                                     This might crash if it is not supported by the backend.
-  --vae-conv-direct                  use Conv2d direct in the vae model (should improve the performance)
-                                     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
-  --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
-  --video-frames                     video frames (default: 1)
-  --fps                              fps (default: 24)
-  --moe-boundary BOUNDARY            timestep boundary for Wan2.2 MoE model. (default: 0.875)
-                                     only enabled if `--high-noise-steps` is set to -1
-  --flow-shift SHIFT                 shift value for Flow models like SD3.x or WAN (default: auto)
-  -v, --verbose                      print extra info
-```
-
-#### txt2img example
+### Generate an image with just one command
 
 ```sh
-./bin/sd -m ../models/sd-v1-4.ckpt -p "a lovely cat"
-# ./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
-# ./bin/sd -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
-# ./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v
-# ./bin/sd --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v
-# ./bin/sd -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v
+./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
 ```
 
-Using formats of different precisions will yield results of varying quality.
+***For detailed command-line arguments, check out [cli doc](./examples/cli/README.md).***
 
-| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
-| ----  |----  |----  |----  |----  |----  |----  |
-| ![](./assets/f32.png) |![](./assets/f16.png) |![](./assets/q8_0.png) |![](./assets/q5_0.png) |![](./assets/q5_1.png) |![](./assets/q4_0.png) |![](./assets/q4_1.png) |
+## Performance
 
-#### img2img example
-
-- `./output.png` is the image generated from the above txt2img pipeline
-
-
-```
-./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">
-  <img src="./assets/img2img_output.png" width="256x">
-</p>
+If you want to improve performance or reduce VRAM/RAM usage, please refer to [performance guide](./docs/performance.md).
 
 ## More Guides
 
+- [SD1.x/SD2.x/SDXL](./docs/sd.md)
+- [SD3/SD3.5](./docs/sd3.md)
+- [Flux-dev/Flux-schnell](./docs/flux.md)
+- [FLUX.1-Kontext-dev](./docs/kontext.md)
+- [Chroma](./docs/chroma.md)
+- [🔥Qwen Image](./docs/qwen_image.md)
+- [🔥Qwen Image Edit/Qwen Image Edit 2509](./docs/qwen_image_edit.md)
+- [🔥Wan2.1/Wan2.2](./docs/wan.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
@@ -448,6 +149,7 @@ These projects use `stable-diffusion.cpp` as a backend for their image generatio
 - [Local Diffusion](https://github.com/rmatif/Local-Diffusion)
 - [sd.cpp-webui](https://github.com/daniandtheweb/sd.cpp-webui)
 - [LocalAI](https://github.com/mudler/LocalAI)
+- [Neural-Pixel](https://github.com/Luiz-Alcantara/Neural-Pixel)
 
 ## Contributors
 
@@ -462,6 +164,7 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 ## References
 
 - [ggml](https://github.com/ggerganov/ggml)
+- [diffusers](https://github.com/huggingface/diffusers)
 - [stable-diffusion](https://github.com/CompVis/stable-diffusion)
 - [sd3-ref](https://github.com/Stability-AI/sd3-ref)
 - [stable-diffusion-stability-ai](https://github.com/Stability-AI/stablediffusion)
@@ -472,4 +175,4 @@ Thank you to all the people who have already contributed to stable-diffusion.cpp
 - [generative-models](https://github.com/Stability-AI/generative-models/)
 - [PhotoMaker](https://github.com/TencentARC/PhotoMaker)
 - [Wan2.1](https://github.com/Wan-Video/Wan2.1)
-- [Wan2.2](https://github.com/Wan-Video/Wan2.2)
+- [Wan2.2](https://github.com/Wan-Video/Wan2.2)

clip.hpp

@@ -6,7 +6,7 @@
 
 /*================================================== CLIPTokenizer ===================================================*/
 
-std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remove_lora(std::string text) {
+__STATIC_INLINE__ std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remove_lora(std::string text) {
     std::regex re("<lora:([^:]+):([^>]+)>");
     std::smatch matches;
     std::unordered_map<std::string, float> filename2multiplier;
@@ -31,7 +31,7 @@ std::pair<std::unordered_map<std::string, float>, std::string> extract_and_remov
     return std::make_pair(filename2multiplier, text);
 }
 
-std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
+__STATIC_INLINE__ std::vector<std::pair<int, std::u32string>> bytes_to_unicode() {
     std::vector<std::pair<int, std::u32string>> byte_unicode_pairs;
     std::set<int> byte_set;
     for (int b = static_cast<int>('!'); b <= static_cast<int>('~'); ++b) {
@@ -548,11 +548,17 @@ protected:
     int64_t embed_dim;
     int64_t vocab_size;
     int64_t num_positions;
+    bool force_clip_f32;
 
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
-        enum ggml_type token_wtype    = GGML_TYPE_F32;
-        enum ggml_type position_wtype = GGML_TYPE_F32;
-
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
+        enum ggml_type token_wtype = GGML_TYPE_F32;
+        if (!force_clip_f32) {
+            token_wtype = get_type(prefix + "token_embedding.weight", tensor_types, GGML_TYPE_F32);
+            if (!support_get_rows(token_wtype)) {
+                token_wtype = GGML_TYPE_F32;
+            }
+        }
+        enum ggml_type position_wtype       = GGML_TYPE_F32;
         params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
         params["position_embedding.weight"] = ggml_new_tensor_2d(ctx, position_wtype, embed_dim, num_positions);
     }
@@ -560,10 +566,12 @@ protected:
 public:
     CLIPEmbeddings(int64_t embed_dim,
                    int64_t vocab_size    = 49408,
-                   int64_t num_positions = 77)
+                   int64_t num_positions = 77,
+                   bool force_clip_f32   = false)
         : embed_dim(embed_dim),
           vocab_size(vocab_size),
-          num_positions(num_positions) {
+          num_positions(num_positions),
+          force_clip_f32(force_clip_f32) {
     }
 
     struct ggml_tensor* get_token_embed_weight() {
@@ -579,7 +587,7 @@ public:
 
         GGML_ASSERT(input_ids->ne[0] == position_embed_weight->ne[1]);
         input_ids            = ggml_reshape_3d(ctx, input_ids, input_ids->ne[0], 1, input_ids->ne[1]);
-        auto token_embedding = ggml_get_rows(ctx, custom_embed_weight != NULL ? custom_embed_weight : token_embed_weight, input_ids);
+        auto token_embedding = ggml_get_rows(ctx, custom_embed_weight != nullptr ? custom_embed_weight : token_embed_weight, input_ids);
         token_embedding      = ggml_reshape_3d(ctx, token_embedding, token_embedding->ne[0], token_embedding->ne[1], token_embedding->ne[3]);
 
         // token_embedding + position_embedding
@@ -598,7 +606,7 @@ protected:
     int64_t image_size;
     int64_t num_patches;
     int64_t num_positions;
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
         enum ggml_type patch_wtype    = GGML_TYPE_F16;
         enum ggml_type class_wtype    = GGML_TYPE_F32;
         enum ggml_type position_wtype = GGML_TYPE_F32;
@@ -633,10 +641,10 @@ public:
         // concat(patch_embedding, class_embedding) + position_embedding
         struct ggml_tensor* patch_embedding;
         int64_t N       = pixel_values->ne[3];
-        patch_embedding = ggml_nn_conv_2d(ctx, pixel_values, patch_embed_weight, NULL, patch_size, patch_size);  // [N, embed_dim, image_size // pacht_size, image_size // pacht_size]
-        patch_embedding = ggml_reshape_3d(ctx, patch_embedding, num_patches, embed_dim, N);                      // [N, embed_dim, num_patches]
-        patch_embedding = ggml_cont(ctx, ggml_permute(ctx, patch_embedding, 1, 0, 2, 3));                        // [N, num_patches, embed_dim]
-        patch_embedding = ggml_reshape_4d(ctx, patch_embedding, 1, embed_dim, num_patches, N);                   // [N, num_patches, embed_dim, 1]
+        patch_embedding = ggml_nn_conv_2d(ctx, pixel_values, patch_embed_weight, nullptr, patch_size, patch_size);  // [N, embed_dim, image_size // pacht_size, image_size // pacht_size]
+        patch_embedding = ggml_reshape_3d(ctx, patch_embedding, num_patches, embed_dim, N);                         // [N, embed_dim, num_patches]
+        patch_embedding = ggml_cont(ctx, ggml_permute(ctx, patch_embedding, 1, 0, 2, 3));                           // [N, num_patches, embed_dim]
+        patch_embedding = ggml_reshape_4d(ctx, patch_embedding, 1, embed_dim, num_patches, N);                      // [N, num_patches, embed_dim, 1]
 
         struct ggml_tensor* class_embedding = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, embed_dim, N);
         class_embedding                     = ggml_repeat(ctx, class_embed_weight, class_embedding);      // [N, embed_dim]
@@ -661,7 +669,7 @@ enum CLIPVersion {
 
 class CLIPTextModel : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
         if (version == OPEN_CLIP_VIT_BIGG_14) {
             enum ggml_type wtype      = GGML_TYPE_F32;
             params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
@@ -678,12 +686,11 @@ public:
     int32_t n_head            = 12;
     int32_t n_layer           = 12;    // num_hidden_layers
     int32_t projection_dim    = 1280;  // only for OPEN_CLIP_VIT_BIGG_14
-    int32_t clip_skip         = -1;
     bool with_final_ln        = true;
 
     CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
                   bool with_final_ln  = true,
-                  int clip_skip_value = -1)
+                  bool force_clip_f32 = false)
         : version(version), with_final_ln(with_final_ln) {
         if (version == OPEN_CLIP_VIT_H_14) {
             hidden_size       = 1024;
@@ -696,20 +703,12 @@ public:
             n_head            = 20;
             n_layer           = 32;
         }
-        set_clip_skip(clip_skip_value);
 
-        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token));
+        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token, force_clip_f32));
         blocks["encoder"]          = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
         blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
     }
 
-    void set_clip_skip(int skip) {
-        if (skip <= 0) {
-            skip = -1;
-        }
-        clip_skip = skip;
-    }
-
     struct ggml_tensor* get_token_embed_weight() {
         auto embeddings = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
         return embeddings->get_token_embed_weight();
@@ -720,7 +719,8 @@ public:
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* tkn_embeddings,
                                 size_t max_token_idx = 0,
-                                bool return_pooled   = false) {
+                                bool return_pooled   = false,
+                                int clip_skip        = -1) {
         // input_ids: [N, n_token]
         auto embeddings       = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
         auto encoder          = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
@@ -735,8 +735,8 @@ public:
         if (return_pooled) {
             auto text_projection = params["text_projection"];
             ggml_tensor* pooled  = ggml_view_1d(ctx, x, hidden_size, x->nb[1] * max_token_idx);
-            if (text_projection != NULL) {
-                pooled = ggml_nn_linear(ctx, pooled, text_projection, NULL);
+            if (text_projection != nullptr) {
+                pooled = ggml_nn_linear(ctx, pooled, text_projection, nullptr);
             } else {
                 LOG_DEBUG("identity projection");
             }
@@ -814,7 +814,7 @@ protected:
     int64_t out_features;
     bool transpose_weight;
 
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
         enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
         if (transpose_weight) {
             params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
@@ -831,12 +831,12 @@ public:
           out_features(out_features),
           transpose_weight(transpose_weight) {}
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         struct ggml_tensor* w = params["weight"];
         if (transpose_weight) {
             w = ggml_cont(ctx, ggml_transpose(ctx, w));
         }
-        return ggml_nn_linear(ctx, x, w, NULL);
+        return ggml_nn_linear(ctx, x, w, nullptr);
     }
 };
 
@@ -889,19 +889,15 @@ struct CLIPTextModelRunner : public GGMLRunner {
                         const std::string prefix,
                         CLIPVersion version = OPENAI_CLIP_VIT_L_14,
                         bool with_final_ln  = true,
-                        int clip_skip_value = -1)
-        : GGMLRunner(backend, offload_params_to_cpu), model(version, with_final_ln, clip_skip_value) {
+                        bool force_clip_f32 = false)
+        : GGMLRunner(backend, offload_params_to_cpu), model(version, with_final_ln, force_clip_f32) {
         model.init(params_ctx, tensor_types, prefix);
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "clip";
     }
 
-    void set_clip_skip(int clip_skip) {
-        model.set_clip_skip(clip_skip);
-    }
-
     void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         model.get_param_tensors(tensors, prefix);
     }
@@ -911,7 +907,8 @@ struct CLIPTextModelRunner : public GGMLRunner {
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* embeddings,
                                 size_t max_token_idx = 0,
-                                bool return_pooled   = false) {
+                                bool return_pooled   = false,
+                                int clip_skip        = -1) {
         size_t N       = input_ids->ne[1];
         size_t n_token = input_ids->ne[0];
         if (input_ids->ne[0] > model.n_token) {
@@ -919,21 +916,22 @@ struct CLIPTextModelRunner : public GGMLRunner {
             input_ids = ggml_reshape_2d(ctx, input_ids, model.n_token, input_ids->ne[0] / model.n_token);
         }
 
-        return model.forward(ctx, backend, input_ids, embeddings, max_token_idx, return_pooled);
+        return model.forward(ctx, backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
     }
 
     struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
                                     int num_custom_embeddings    = 0,
-                                    void* custom_embeddings_data = NULL,
+                                    void* custom_embeddings_data = nullptr,
                                     size_t max_token_idx         = 0,
-                                    bool return_pooled           = false) {
+                                    bool return_pooled           = false,
+                                    int clip_skip                = -1) {
         struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         input_ids = to_backend(input_ids);
 
-        struct ggml_tensor* embeddings = NULL;
+        struct ggml_tensor* embeddings = nullptr;
 
-        if (num_custom_embeddings > 0 && custom_embeddings_data != NULL) {
+        if (num_custom_embeddings > 0 && custom_embeddings_data != nullptr) {
             auto token_embed_weight = model.get_token_embed_weight();
             auto custom_embeddings  = ggml_new_tensor_2d(compute_ctx,
                                                          token_embed_weight->type,
@@ -945,7 +943,7 @@ struct CLIPTextModelRunner : public GGMLRunner {
             embeddings = ggml_concat(compute_ctx, token_embed_weight, custom_embeddings, 1);
         }
 
-        struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, embeddings, max_token_idx, return_pooled);
+        struct ggml_tensor* hidden_states = forward(compute_ctx, runtime_backend, input_ids, embeddings, max_token_idx, return_pooled, clip_skip);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -958,10 +956,11 @@ struct CLIPTextModelRunner : public GGMLRunner {
                  void* custom_embeddings_data,
                  size_t max_token_idx,
                  bool return_pooled,
+                 int clip_skip,
                  ggml_tensor** output,
-                 ggml_context* output_ctx = NULL) {
+                 ggml_context* output_ctx = nullptr) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled);
+            return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled, clip_skip);
         };
         GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
     }

common.hpp

@@ -121,7 +121,7 @@ public:
         }
     }
 
-    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = NULL) {
+    virtual struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = nullptr) {
         // For dims==3, we reduce dimension from 5d to 4d by merging h and w, in order not to change ggml
         // [N, c, t, h, w] => [N, c, t, h * w]
         // x: [N, channels, h, w] if dims == 2 else [N, channels, t, h, w]
@@ -131,7 +131,7 @@ public:
         auto out_layers_0 = std::dynamic_pointer_cast<GroupNorm32>(blocks["out_layers.0"]);
         auto out_layers_3 = std::dynamic_pointer_cast<UnaryBlock>(blocks["out_layers.3"]);
 
-        if (emb == NULL) {
+        if (emb == nullptr) {
             GGML_ASSERT(skip_t_emb);
         }
 
@@ -177,12 +177,12 @@ public:
     }
 };
 
-class GEGLU : public GGMLBlock {
+class GEGLU : public UnaryBlock {
 protected:
     int64_t dim_in;
     int64_t dim_out;
 
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") override {
         enum ggml_type wtype      = get_type(prefix + "proj.weight", tensor_types, GGML_TYPE_F32);
         enum ggml_type bias_wtype = GGML_TYPE_F32;
         params["proj.weight"]     = ggml_new_tensor_2d(ctx, wtype, dim_in, dim_out * 2);
@@ -193,7 +193,7 @@ public:
     GEGLU(int64_t dim_in, int64_t dim_out)
         : dim_in(dim_in), dim_out(dim_out) {}
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [ne3, ne2, ne1, dim_in]
         // return: [ne3, ne2, ne1, dim_out]
         struct ggml_tensor* w = params["proj.weight"];
@@ -216,23 +216,57 @@ public:
     }
 };
 
+class GELU : public UnaryBlock {
+public:
+    GELU(int64_t dim_in, int64_t dim_out, bool bias = true) {
+        blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out, bias));
+    }
+
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
+        // x: [ne3, ne2, ne1, dim_in]
+        // return: [ne3, ne2, ne1, dim_out]
+        auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+
+        x = proj->forward(ctx, x);
+        x = ggml_gelu_inplace(ctx, x);
+        return x;
+    }
+};
+
 class FeedForward : public GGMLBlock {
 public:
+    enum class Activation {
+        GEGLU,
+        GELU
+    };
     FeedForward(int64_t dim,
                 int64_t dim_out,
-                int64_t mult = 4) {
+                int64_t mult          = 4,
+                Activation activation = Activation::GEGLU,
+                bool precision_fix    = false) {
         int64_t inner_dim = dim * mult;
+        if (activation == Activation::GELU) {
+            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GELU(dim, inner_dim));
+        } else {
+            blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
+        }
 
-        blocks["net.0"] = std::shared_ptr<GGMLBlock>(new GEGLU(dim, inner_dim));
         // net_1 is nn.Dropout(), skip for inference
-        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out));
+        float scale = 1.f;
+        if (precision_fix) {
+            scale = 1.f / 128.f;
+        }
+        // The purpose of the scale here is to prevent NaN issues in certain situations.
+        // For example, when using Vulkan without enabling force_prec_f32,
+        // or when using CUDA but the weights are k-quants.
+        blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, false, scale));
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
         // x: [ne3, ne2, ne1, dim]
         // return: [ne3, ne2, ne1, dim_out]
 
-        auto net_0 = std::dynamic_pointer_cast<GEGLU>(blocks["net.0"]);
+        auto net_0 = std::dynamic_pointer_cast<UnaryBlock>(blocks["net.0"]);
         auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
 
         x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]
@@ -291,7 +325,7 @@ public:
         auto k = to_k->forward(ctx, context);  // [N, n_context, inner_dim]
         auto v = to_v->forward(ctx, context);  // [N, n_context, inner_dim]
 
-        x = ggml_nn_attention_ext(ctx, backend, q, k, v, n_head, NULL, false, false, flash_attn);  // [N, n_token, inner_dim]
+        x = ggml_nn_attention_ext(ctx, backend, q, k, v, n_head, nullptr, false, false, flash_attn);  // [N, n_token, inner_dim]
 
         x = to_out_0->forward(ctx, x);  // [N, n_token, query_dim]
         return x;
@@ -449,7 +483,7 @@ public:
 
 class AlphaBlender : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") override {
         // Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
         enum ggml_type wtype = GGML_TYPE_F32;
         params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);

control.hpp

@@ -206,18 +206,18 @@ public:
                                              struct ggml_tensor* guided_hint,
                                              struct ggml_tensor* timesteps,
                                              struct ggml_tensor* context,
-                                             struct ggml_tensor* y = NULL) {
+                                             struct ggml_tensor* y = nullptr) {
         // x: [N, in_channels, h, w] or [N, in_channels/2, h, w]
         // timesteps: [N,]
         // context: [N, max_position, hidden_size] or [1, max_position, hidden_size]. for example, [N, 77, 768]
         // y: [N, adm_in_channels] or [1, adm_in_channels]
-        if (context != NULL) {
+        if (context != nullptr) {
             if (context->ne[2] != x->ne[3]) {
                 context = ggml_repeat(ctx, context, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, context->ne[0], context->ne[1], x->ne[3]));
             }
         }
 
-        if (y != NULL) {
+        if (y != nullptr) {
             if (y->ne[1] != x->ne[3]) {
                 y = ggml_repeat(ctx, y, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, y->ne[0], x->ne[3]));
             }
@@ -237,7 +237,7 @@ public:
         emb      = time_embed_2->forward(ctx, emb);  // [N, time_embed_dim]
 
         // SDXL/SVD
-        if (y != NULL) {
+        if (y != nullptr) {
             auto label_embed_0 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.0"]);
             auto label_embed_2 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.2"]);
 
@@ -250,7 +250,7 @@ public:
 
         std::vector<struct ggml_tensor*> outs;
 
-        if (guided_hint == NULL) {
+        if (guided_hint == nullptr) {
             guided_hint = input_hint_block_forward(ctx, hint, emb, context);
         }
         outs.push_back(guided_hint);
@@ -312,10 +312,10 @@ struct ControlNet : public GGMLRunner {
     SDVersion version = VERSION_SD1;
     ControlNetBlock control_net;
 
-    ggml_backend_buffer_t control_buffer = NULL;  // keep control output tensors in backend memory
-    ggml_context* control_ctx            = NULL;
+    ggml_backend_buffer_t control_buffer = nullptr;  // keep control output tensors in backend memory
+    ggml_context* control_ctx            = nullptr;
     std::vector<struct ggml_tensor*> controls;  // (12 input block outputs, 1 middle block output) SD 1.5
-    struct ggml_tensor* guided_hint = NULL;     // guided_hint cache, for faster inference
+    struct ggml_tensor* guided_hint = nullptr;  // guided_hint cache, for faster inference
     bool guided_hint_cached         = false;
 
     ControlNet(ggml_backend_t backend,
@@ -337,14 +337,14 @@ struct ControlNet : public GGMLRunner {
         }
     }
 
-    ~ControlNet() {
+    ~ControlNet() override {
         free_control_ctx();
     }
 
     void alloc_control_ctx(std::vector<struct ggml_tensor*> outs) {
         struct ggml_init_params params;
         params.mem_size   = static_cast<size_t>(outs.size() * ggml_tensor_overhead()) + 1024 * 1024;
-        params.mem_buffer = NULL;
+        params.mem_buffer = nullptr;
         params.no_alloc   = true;
         control_ctx       = ggml_init(params);
 
@@ -366,20 +366,20 @@ struct ControlNet : public GGMLRunner {
     }
 
     void free_control_ctx() {
-        if (control_buffer != NULL) {
+        if (control_buffer != nullptr) {
             ggml_backend_buffer_free(control_buffer);
-            control_buffer = NULL;
+            control_buffer = nullptr;
         }
-        if (control_ctx != NULL) {
+        if (control_ctx != nullptr) {
             ggml_free(control_ctx);
-            control_ctx = NULL;
+            control_ctx = nullptr;
         }
-        guided_hint        = NULL;
+        guided_hint        = nullptr;
         guided_hint_cached = false;
         controls.clear();
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "control_net";
     }
 
@@ -391,12 +391,12 @@ struct ControlNet : public GGMLRunner {
                                     struct ggml_tensor* hint,
                                     struct ggml_tensor* timesteps,
                                     struct ggml_tensor* context,
-                                    struct ggml_tensor* y = NULL) {
+                                    struct ggml_tensor* y = nullptr) {
         struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, CONTROL_NET_GRAPH_SIZE, false);
 
         x = to_backend(x);
         if (guided_hint_cached) {
-            hint = NULL;
+            hint = nullptr;
         } else {
             hint = to_backend(hint);
         }
@@ -408,12 +408,12 @@ struct ControlNet : public GGMLRunner {
                                         runtime_backend,
                                         x,
                                         hint,
-                                        guided_hint_cached ? guided_hint : NULL,
+                                        guided_hint_cached ? guided_hint : nullptr,
                                         timesteps,
                                         context,
                                         y);
 
-        if (control_ctx == NULL) {
+        if (control_ctx == nullptr) {
             alloc_control_ctx(outs);
         }
 
@@ -431,8 +431,8 @@ struct ControlNet : public GGMLRunner {
                  struct ggml_tensor* timesteps,
                  struct ggml_tensor* context,
                  struct ggml_tensor* y,
-                 struct ggml_tensor** output     = NULL,
-                 struct ggml_context* output_ctx = NULL) {
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
@@ -445,7 +445,7 @@ struct ControlNet : public GGMLRunner {
         guided_hint_cached = true;
     }
 
-    bool load_from_file(const std::string& file_path) {
+    bool load_from_file(const std::string& file_path, int n_threads) {
         LOG_INFO("loading control net from '%s'", file_path.c_str());
         alloc_params_buffer();
         std::map<std::string, ggml_tensor*> tensors;
@@ -458,7 +458,7 @@ struct ControlNet : public GGMLRunner {
             return false;
         }
 
-        bool success = model_loader.load_tensors(tensors, ignore_tensors);
+        bool success = model_loader.load_tensors(tensors, ignore_tensors, n_threads);
 
         if (!success) {
             LOG_ERROR("load control net tensors from model loader failed");

denoiser.hpp

@@ -19,7 +19,7 @@ struct SigmaSchedule {
 };
 
 struct DiscreteSchedule : SigmaSchedule {
-    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
         std::vector<float> result;
 
         int t_max = TIMESTEPS - 1;
@@ -43,7 +43,7 @@ struct DiscreteSchedule : SigmaSchedule {
 };
 
 struct ExponentialSchedule : SigmaSchedule {
-    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
         std::vector<float> sigmas;
 
         // Calculate step size
@@ -150,7 +150,7 @@ std::vector<float> log_linear_interpolation(std::vector<float> sigma_in,
 https://research.nvidia.com/labs/toronto-ai/AlignYourSteps/howto.html
 */
 struct AYSSchedule : SigmaSchedule {
-    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
         const std::vector<float> noise_levels[] = {
             /* SD1.5 */
             {14.6146412293f, 6.4745760956f, 3.8636745985f, 2.6946151520f,
@@ -204,7 +204,7 @@ struct AYSSchedule : SigmaSchedule {
  * GITS Scheduler: https://github.com/zju-pi/diff-sampler/tree/main/gits-main
  */
 struct GITSSchedule : SigmaSchedule {
-    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
         if (sigma_max <= 0.0f) {
             return std::vector<float>{};
         }
@@ -232,8 +232,27 @@ struct GITSSchedule : SigmaSchedule {
     }
 };
 
+struct SGMUniformSchedule : SigmaSchedule {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min_in, float sigma_max_in, t_to_sigma_t t_to_sigma_func) override {
+        std::vector<float> result;
+        if (n == 0) {
+            result.push_back(0.0f);
+            return result;
+        }
+        result.reserve(n + 1);
+        int t_max                    = TIMESTEPS - 1;
+        int t_min                    = 0;
+        std::vector<float> timesteps = linear_space(static_cast<float>(t_max), static_cast<float>(t_min), n + 1);
+        for (int i = 0; i < n; i++) {
+            result.push_back(t_to_sigma_func(timesteps[i]));
+        }
+        result.push_back(0.0f);
+        return result;
+    }
+};
+
 struct KarrasSchedule : SigmaSchedule {
-    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
         // These *COULD* be function arguments here,
         // but does anybody ever bother to touch them?
         float rho = 7.f;
@@ -251,6 +270,64 @@ struct KarrasSchedule : SigmaSchedule {
     }
 };
 
+struct SimpleSchedule : SigmaSchedule {
+    std::vector<float> get_sigmas(uint32_t n, float sigma_min, float sigma_max, t_to_sigma_t t_to_sigma) override {
+        std::vector<float> result_sigmas;
+
+        if (n == 0) {
+            return result_sigmas;
+        }
+
+        result_sigmas.reserve(n + 1);
+
+        int model_sigmas_len = TIMESTEPS;
+
+        float step_factor = static_cast<float>(model_sigmas_len) / static_cast<float>(n);
+
+        for (uint32_t i = 0; i < n; ++i) {
+            int offset_from_start_of_py_array = static_cast<int>(static_cast<float>(i) * step_factor);
+            int timestep_index                = model_sigmas_len - 1 - offset_from_start_of_py_array;
+
+            if (timestep_index < 0) {
+                timestep_index = 0;
+            }
+
+            result_sigmas.push_back(t_to_sigma(static_cast<float>(timestep_index)));
+        }
+        result_sigmas.push_back(0.0f);
+        return result_sigmas;
+    }
+};
+
+// Close to Beta Schedule, but increadably simple in code.
+struct SmoothStepSchedule : SigmaSchedule {
+    static constexpr float smoothstep(float x) {
+        return x * x * (3.0f - 2.0f * x);
+    }
+
+    std::vector<float> get_sigmas(uint32_t n, float /*sigma_min*/, float /*sigma_max*/, t_to_sigma_t t_to_sigma) override {
+        std::vector<float> result;
+        result.reserve(n + 1);
+
+        const int t_max = TIMESTEPS - 1;
+        if (n == 0) {
+            return result;
+        } else if (n == 1) {
+            result.push_back(t_to_sigma((float)t_max));
+            result.push_back(0.f);
+            return result;
+        }
+
+        for (uint32_t i = 0; i < n; i++) {
+            float u = 1.f - float(i) / float(n);
+            result.push_back(t_to_sigma(std::round(smoothstep(u) * t_max)));
+        }
+
+        result.push_back(0.f);
+        return result;
+    }
+};
+
 struct Denoiser {
     std::shared_ptr<SigmaSchedule> scheduler                                                 = std::make_shared<DiscreteSchedule>();
     virtual float sigma_min()                                                                = 0;
@@ -273,15 +350,15 @@ struct CompVisDenoiser : public Denoiser {
 
     float sigma_data = 1.0f;
 
-    float sigma_min() {
+    float sigma_min() override {
         return sigmas[0];
     }
 
-    float sigma_max() {
+    float sigma_max() override {
         return sigmas[TIMESTEPS - 1];
     }
 
-    float sigma_to_t(float sigma) {
+    float sigma_to_t(float sigma) override {
         float log_sigma = std::log(sigma);
         std::vector<float> dists;
         dists.reserve(TIMESTEPS);
@@ -307,7 +384,7 @@ struct CompVisDenoiser : public Denoiser {
         return t;
     }
 
-    float t_to_sigma(float t) {
+    float t_to_sigma(float t) override {
         int low_idx     = static_cast<int>(std::floor(t));
         int high_idx    = static_cast<int>(std::ceil(t));
         float w         = t - static_cast<float>(low_idx);
@@ -315,7 +392,7 @@ struct CompVisDenoiser : public Denoiser {
         return std::exp(log_sigma);
     }
 
-    std::vector<float> get_scalings(float sigma) {
+    std::vector<float> get_scalings(float sigma) override {
         float c_skip = 1.0f;
         float c_out  = -sigma;
         float c_in   = 1.0f / std::sqrt(sigma * sigma + sigma_data * sigma_data);
@@ -323,19 +400,19 @@ struct CompVisDenoiser : public Denoiser {
     }
 
     // this function will modify noise/latent
-    ggml_tensor* noise_scaling(float sigma, ggml_tensor* noise, ggml_tensor* latent) {
+    ggml_tensor* noise_scaling(float sigma, ggml_tensor* noise, ggml_tensor* latent) override {
         ggml_tensor_scale(noise, sigma);
         ggml_tensor_add(latent, noise);
         return latent;
     }
 
-    ggml_tensor* inverse_noise_scaling(float sigma, ggml_tensor* latent) {
+    ggml_tensor* inverse_noise_scaling(float sigma, ggml_tensor* latent) override {
         return latent;
     }
 };
 
 struct CompVisVDenoiser : public CompVisDenoiser {
-    std::vector<float> get_scalings(float sigma) {
+    std::vector<float> get_scalings(float sigma) override {
         float c_skip = sigma_data * sigma_data / (sigma * sigma + sigma_data * sigma_data);
         float c_out  = -sigma * sigma_data / std::sqrt(sigma * sigma + sigma_data * sigma_data);
         float c_in   = 1.0f / std::sqrt(sigma * sigma + sigma_data * sigma_data);
@@ -352,19 +429,19 @@ struct EDMVDenoiser : public CompVisVDenoiser {
         scheduler = std::make_shared<ExponentialSchedule>();
     }
 
-    float t_to_sigma(float t) {
+    float t_to_sigma(float t) override {
         return std::exp(t * 4 / (float)TIMESTEPS);
     }
 
-    float sigma_to_t(float s) {
+    float sigma_to_t(float s) override {
         return 0.25 * std::log(s);
     }
 
-    float sigma_min() {
+    float sigma_min() override {
         return min_sigma;
     }
 
-    float sigma_max() {
+    float sigma_max() override {
         return max_sigma;
     }
 };
@@ -393,24 +470,24 @@ struct DiscreteFlowDenoiser : public Denoiser {
         }
     }
 
-    float sigma_min() {
+    float sigma_min() override {
         return sigmas[0];
     }
 
-    float sigma_max() {
+    float sigma_max() override {
         return sigmas[TIMESTEPS - 1];
     }
 
-    float sigma_to_t(float sigma) {
+    float sigma_to_t(float sigma) override {
         return sigma * 1000.f;
     }
 
-    float t_to_sigma(float t) {
+    float t_to_sigma(float t) override {
         t = t + 1;
         return time_snr_shift(shift, t / 1000.f);
     }
 
-    std::vector<float> get_scalings(float sigma) {
+    std::vector<float> get_scalings(float sigma) override {
         float c_skip = 1.0f;
         float c_out  = -sigma;
         float c_in   = 1.0f;
@@ -418,14 +495,14 @@ struct DiscreteFlowDenoiser : public Denoiser {
     }
 
     // this function will modify noise/latent
-    ggml_tensor* noise_scaling(float sigma, ggml_tensor* noise, ggml_tensor* latent) {
+    ggml_tensor* noise_scaling(float sigma, ggml_tensor* noise, ggml_tensor* latent) override {
         ggml_tensor_scale(noise, sigma);
         ggml_tensor_scale(latent, 1.0f - sigma);
         ggml_tensor_add(latent, noise);
         return latent;
     }
 
-    ggml_tensor* inverse_noise_scaling(float sigma, ggml_tensor* latent) {
+    ggml_tensor* inverse_noise_scaling(float sigma, ggml_tensor* latent) override {
         ggml_tensor_scale(latent, 1.0f / (1.0f - sigma));
         return latent;
     }
@@ -452,24 +529,24 @@ struct FluxFlowDenoiser : public Denoiser {
         }
     }
 
-    float sigma_min() {
+    float sigma_min() override {
         return sigmas[0];
     }
 
-    float sigma_max() {
+    float sigma_max() override {
         return sigmas[TIMESTEPS - 1];
     }
 
-    float sigma_to_t(float sigma) {
+    float sigma_to_t(float sigma) override {
         return sigma;
     }
 
-    float t_to_sigma(float t) {
+    float t_to_sigma(float t) override {
         t = t + 1;
         return flux_time_shift(shift, 1.0f, t / TIMESTEPS);
     }
 
-    std::vector<float> get_scalings(float sigma) {
+    std::vector<float> get_scalings(float sigma) override {
         float c_skip = 1.0f;
         float c_out  = -sigma;
         float c_in   = 1.0f;
@@ -477,14 +554,14 @@ struct FluxFlowDenoiser : public Denoiser {
     }
 
     // this function will modify noise/latent
-    ggml_tensor* noise_scaling(float sigma, ggml_tensor* noise, ggml_tensor* latent) {
+    ggml_tensor* noise_scaling(float sigma, ggml_tensor* noise, ggml_tensor* latent) override {
         ggml_tensor_scale(noise, sigma);
         ggml_tensor_scale(latent, 1.0f - sigma);
         ggml_tensor_add(latent, noise);
         return latent;
     }
 
-    ggml_tensor* inverse_noise_scaling(float sigma, ggml_tensor* latent) {
+    ggml_tensor* inverse_noise_scaling(float sigma, ggml_tensor* latent) override {
         ggml_tensor_scale(latent, 1.0f / (1.0f - sigma));
         return latent;
     }
@@ -693,7 +770,6 @@ static void sample_k_diffusion(sample_method_t method,
         } break;
         case DPMPP2S_A: {
             struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* d     = ggml_dup_tensor(work_ctx, x);
             struct ggml_tensor* x2    = ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
@@ -708,22 +784,15 @@ static void sample_k_diffusion(sample_method_t method,
                 auto sigma_fn    = [](float t) -> float { return exp(-t); };
 
                 if (sigma_down == 0) {
-                    // Euler step
-                    float* vec_d        = (float*)d->data;
+                    // d = (x - denoised) / sigmas[i];
+                    // dt = sigma_down - sigmas[i];
+                    // x += d * dt;
+                    // => x = denoised
                     float* vec_x        = (float*)x->data;
                     float* vec_denoised = (float*)denoised->data;
 
-                    for (int j = 0; j < ggml_nelements(d); j++) {
-                        vec_d[j] = (vec_x[j] - vec_denoised[j]) / sigmas[i];
-                    }
-
-                    // TODO: If sigma_down == 0, isn't this wrong?
-                    // But
-                    // https://github.com/crowsonkb/k-diffusion/blob/master/k_diffusion/sampling.py#L525
-                    // has this exactly the same way.
-                    float dt = sigma_down - sigmas[i];
-                    for (int j = 0; j < ggml_nelements(d); j++) {
-                        vec_x[j] = vec_x[j] + vec_d[j] * dt;
+                    for (int j = 0; j < ggml_nelements(x); j++) {
+                        vec_x[j] = vec_denoised[j];
                     }
                 } else {
                     // DPM-Solver++(2S)
@@ -732,7 +801,6 @@ static void sample_k_diffusion(sample_method_t method,
                     float h      = t_next - t;
                     float s      = t + 0.5f * h;
 
-                    float* vec_d        = (float*)d->data;
                     float* vec_x        = (float*)x->data;
                     float* vec_x2       = (float*)x2->data;
                     float* vec_denoised = (float*)denoised->data;

diffusion_model.hpp

@@ -3,26 +3,33 @@
 
 #include "flux.hpp"
 #include "mmdit.hpp"
+#include "qwen_image.hpp"
 #include "unet.hpp"
 #include "wan.hpp"
 
+struct DiffusionParams {
+    struct ggml_tensor* x                     = nullptr;
+    struct ggml_tensor* timesteps             = nullptr;
+    struct ggml_tensor* context               = nullptr;
+    struct ggml_tensor* c_concat              = nullptr;
+    struct ggml_tensor* y                     = nullptr;
+    struct ggml_tensor* guidance              = nullptr;
+    std::vector<ggml_tensor*> ref_latents     = {};
+    bool increase_ref_index                   = false;
+    int num_video_frames                      = -1;
+    std::vector<struct ggml_tensor*> controls = {};
+    float control_strength                    = 0.f;
+    struct ggml_tensor* vace_context          = nullptr;
+    float vace_strength                       = 1.f;
+    std::vector<int> skip_layers              = {};
+};
+
 struct DiffusionModel {
     virtual std::string get_desc()                                                      = 0;
     virtual void compute(int n_threads,
-                         struct ggml_tensor* x,
-                         struct ggml_tensor* timesteps,
-                         struct ggml_tensor* context,
-                         struct ggml_tensor* c_concat,
-                         struct ggml_tensor* y,
-                         struct ggml_tensor* guidance,
-                         std::vector<ggml_tensor*> ref_latents     = {},
-                         bool increase_ref_index                   = false,
-                         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>())             = 0;
+                         DiffusionParams diffusion_params,
+                         struct ggml_tensor** output     = nullptr,
+                         struct ggml_context* output_ctx = nullptr)                     = 0;
     virtual void alloc_params_buffer()                                                  = 0;
     virtual void free_params_buffer()                                                   = 0;
     virtual void free_compute_buffer()                                                  = 0;
@@ -42,51 +49,47 @@ struct UNetModel : public DiffusionModel {
         : unet(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn) {
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return unet.get_desc();
     }
 
-    void alloc_params_buffer() {
+    void alloc_params_buffer() override {
         unet.alloc_params_buffer();
     }
 
-    void free_params_buffer() {
+    void free_params_buffer() override {
         unet.free_params_buffer();
     }
 
-    void free_compute_buffer() {
+    void free_compute_buffer() override {
         unet.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
         unet.get_param_tensors(tensors, "model.diffusion_model");
     }
 
-    size_t get_params_buffer_size() {
+    size_t get_params_buffer_size() override {
         return unet.get_params_buffer_size();
     }
 
-    int64_t get_adm_in_channels() {
+    int64_t get_adm_in_channels() override {
         return unet.unet.adm_in_channels;
     }
 
     void compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor* guidance,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 bool increase_ref_index                   = false,
-                 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>()) {
-        (void)skip_layers;  // SLG doesn't work with UNet models
-        return unet.compute(n_threads, x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength, output, output_ctx);
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return unet.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.num_video_frames,
+                            diffusion_params.controls,
+                            diffusion_params.control_strength, output, output_ctx);
     }
 };
 
@@ -95,54 +98,51 @@ struct MMDiTModel : public DiffusionModel {
 
     MMDiTModel(ggml_backend_t backend,
                bool offload_params_to_cpu,
+               bool flash_attn                     = false,
                const String2GGMLType& tensor_types = {})
-        : mmdit(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model") {
+        : mmdit(backend, offload_params_to_cpu, flash_attn, tensor_types, "model.diffusion_model") {
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return mmdit.get_desc();
     }
 
-    void alloc_params_buffer() {
+    void alloc_params_buffer() override {
         mmdit.alloc_params_buffer();
     }
 
-    void free_params_buffer() {
+    void free_params_buffer() override {
         mmdit.free_params_buffer();
     }
 
-    void free_compute_buffer() {
+    void free_compute_buffer() override {
         mmdit.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
         mmdit.get_param_tensors(tensors, "model.diffusion_model");
     }
 
-    size_t get_params_buffer_size() {
+    size_t get_params_buffer_size() override {
         return mmdit.get_params_buffer_size();
     }
 
-    int64_t get_adm_in_channels() {
+    int64_t get_adm_in_channels() override {
         return 768 + 1280;
     }
 
     void compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor* guidance,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 bool increase_ref_index                   = false,
-                 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 mmdit.compute(n_threads, x, timesteps, context, y, output, output_ctx, skip_layers);
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return mmdit.compute(n_threads,
+                             diffusion_params.x,
+                             diffusion_params.timesteps,
+                             diffusion_params.context,
+                             diffusion_params.y,
+                             output,
+                             output_ctx,
+                             diffusion_params.skip_layers);
     }
 };
 
@@ -158,50 +158,50 @@ struct FluxModel : public DiffusionModel {
         : flux(backend, offload_params_to_cpu, tensor_types, "model.diffusion_model", version, flash_attn, use_mask) {
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return flux.get_desc();
     }
 
-    void alloc_params_buffer() {
+    void alloc_params_buffer() override {
         flux.alloc_params_buffer();
     }
 
-    void free_params_buffer() {
+    void free_params_buffer() override {
         flux.free_params_buffer();
     }
 
-    void free_compute_buffer() {
+    void free_compute_buffer() override {
         flux.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
         flux.get_param_tensors(tensors, "model.diffusion_model");
     }
 
-    size_t get_params_buffer_size() {
+    size_t get_params_buffer_size() override {
         return flux.get_params_buffer_size();
     }
 
-    int64_t get_adm_in_channels() {
+    int64_t get_adm_in_channels() override {
         return 768;
     }
 
     void compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor* guidance,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 bool increase_ref_index                   = false,
-                 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, ref_latents, increase_ref_index, output, output_ctx, skip_layers);
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return flux.compute(n_threads,
+                            diffusion_params.x,
+                            diffusion_params.timesteps,
+                            diffusion_params.context,
+                            diffusion_params.c_concat,
+                            diffusion_params.y,
+                            diffusion_params.guidance,
+                            diffusion_params.ref_latents,
+                            diffusion_params.increase_ref_index,
+                            output,
+                            output_ctx,
+                            diffusion_params.skip_layers);
     }
 };
 
@@ -218,50 +218,105 @@ struct WanModel : public DiffusionModel {
         : prefix(prefix), wan(backend, offload_params_to_cpu, tensor_types, prefix, version, flash_attn) {
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return wan.get_desc();
     }
 
-    void alloc_params_buffer() {
+    void alloc_params_buffer() override {
         wan.alloc_params_buffer();
     }
 
-    void free_params_buffer() {
+    void free_params_buffer() override {
         wan.free_params_buffer();
     }
 
-    void free_compute_buffer() {
+    void free_compute_buffer() override {
         wan.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
         wan.get_param_tensors(tensors, prefix);
     }
 
-    size_t get_params_buffer_size() {
+    size_t get_params_buffer_size() override {
         return wan.get_params_buffer_size();
     }
 
-    int64_t get_adm_in_channels() {
+    int64_t get_adm_in_channels() override {
         return 768;
     }
 
     void compute(int n_threads,
-                 struct ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
-                 struct ggml_tensor* guidance,
-                 std::vector<ggml_tensor*> ref_latents     = {},
-                 bool increase_ref_index                   = false,
-                 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 wan.compute(n_threads, x, timesteps, context, y, c_concat, NULL, output, output_ctx);
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return wan.compute(n_threads,
+                           diffusion_params.x,
+                           diffusion_params.timesteps,
+                           diffusion_params.context,
+                           diffusion_params.y,
+                           diffusion_params.c_concat,
+                           nullptr,
+                           diffusion_params.vace_context,
+                           diffusion_params.vace_strength,
+                           output,
+                           output_ctx);
+    }
+};
+
+struct QwenImageModel : public DiffusionModel {
+    std::string prefix;
+    Qwen::QwenImageRunner qwen_image;
+
+    QwenImageModel(ggml_backend_t backend,
+                   bool offload_params_to_cpu,
+                   const String2GGMLType& tensor_types = {},
+                   const std::string prefix            = "model.diffusion_model",
+                   SDVersion version                   = VERSION_QWEN_IMAGE,
+                   bool flash_attn                     = false)
+        : prefix(prefix), qwen_image(backend, offload_params_to_cpu, tensor_types, prefix, version, flash_attn) {
+    }
+
+    std::string get_desc() override {
+        return qwen_image.get_desc();
+    }
+
+    void alloc_params_buffer() override {
+        qwen_image.alloc_params_buffer();
+    }
+
+    void free_params_buffer() override {
+        qwen_image.free_params_buffer();
+    }
+
+    void free_compute_buffer() override {
+        qwen_image.free_compute_buffer();
+    }
+
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+        qwen_image.get_param_tensors(tensors, prefix);
+    }
+
+    size_t get_params_buffer_size() override {
+        return qwen_image.get_params_buffer_size();
+    }
+
+    int64_t get_adm_in_channels() override {
+        return 768;
+    }
+
+    void compute(int n_threads,
+                 DiffusionParams diffusion_params,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr) override {
+        return qwen_image.compute(n_threads,
+                                  diffusion_params.x,
+                                  diffusion_params.timesteps,
+                                  diffusion_params.context,
+                                  diffusion_params.ref_latents,
+                                  true,  // increase_ref_index
+                                  output,
+                                  output_ctx);
     }
 };
 

docs/build.md

@@ -0,0 +1,173 @@
+# Build from scratch
+
+## Get the Code
+
+```
+git clone --recursive https://github.com/leejet/stable-diffusion.cpp
+cd stable-diffusion.cpp
+```
+
+- If you have already cloned the repository, you can use the following command to update the repository to the latest code.
+
+```
+cd stable-diffusion.cpp
+git pull origin master
+git submodule init
+git submodule update
+```
+
+## Build (CPU only)
+
+If you don't have a GPU or CUDA installed, you can build a CPU-only version.
+
+```shell
+mkdir build && cd build
+cmake ..
+cmake --build . --config Release
+```
+
+## Build with OpenBLAS
+
+```shell
+mkdir build && cd build
+cmake .. -DGGML_OPENBLAS=ON
+cmake --build . --config Release
+```
+
+## Build with CUDA
+
+This provides GPU acceleration using 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.
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_CUDA=ON
+cmake --build . --config Release
+```
+
+## Build with HipBLAS
+
+This provides GPU acceleration using AMD GPU. Make sure to have the ROCm toolkit installed.
+To build for another GPU architecture than installed in your system, set `$GFX_NAME` manually to the desired architecture (replace first command). This is also necessary if your GPU is not officially supported by ROCm, for example you have to set `$GFX_NAME` manually to `gfx1030` for consumer RDNA2 cards.
+
+Windows User Refer to [docs/hipBLAS_on_Windows.md](docs%2FhipBLAS_on_Windows.md) for a comprehensive guide.
+
+```shell
+mkdir build && cd build
+if command -v rocminfo; then export GFX_NAME=$(rocminfo | awk '/ *Name: +gfx[1-9]/ {print $2; exit}'); else echo "rocminfo missing!"; fi
+if [ -z "${GFX_NAME}" ]; then echo "Error: Couldn't detect GPU!"; else echo "Building for GPU: ${GFX_NAME}"; fi
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DGPU_TARGETS=$GFX_NAME -DAMDGPU_TARGETS=$GFX_NAME -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON -DCMAKE_POSITION_INDEPENDENT_CODE=ON
+cmake --build . --config Release
+```
+
+## Build with MUSA
+
+This provides GPU acceleration using Moore Threads GPU. Make sure to have the MUSA toolkit installed.
+
+```shell
+mkdir build && cd build
+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
+```
+
+## Build with Metal
+
+Using Metal makes the computation run on the GPU. Currently, there are some issues with Metal when performing operations on very large matrices, making it highly inefficient at the moment. Performance improvements are expected in the near future.
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_METAL=ON
+cmake --build . --config Release
+```
+
+## Build with Vulkan
+
+Install Vulkan SDK from https://www.lunarg.com/vulkan-sdk/.
+
+```shell
+mkdir build && cd build
+cmake .. -DSD_VULKAN=ON
+cmake --build . --config Release
+```
+
+## Build with 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:
+    ```shell
+    # 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:
+
+```shell
+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)*
+
+## Build with 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).
+
+```shell
+# Export relevant ENV variables
+source /opt/intel/oneapi/setvars.sh
+
+# Option 1: Use FP32 (recommended for better performance in most cases)
+cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx
+
+# Option 2: Use FP16
+cmake .. -DSD_SYCL=ON -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL_F16=ON
+
+cmake --build . --config Release
+```

docs/chroma.md

@@ -24,7 +24,7 @@ You can download the preconverted gguf weights from [silveroxides/Chroma-GGUF](h
 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
+ .\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 --clip-on-cpu
 ```
 
 ![](../assets/flux/chroma_v40.png)

docs/flux.md

@@ -28,7 +28,7 @@ Using fp16 will lead to overflow, but ggml's support for bf16 is not yet fully d
 For example:
 
 ```
- .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v
+ .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
 ```
 
 Using formats of different precisions will yield results of varying quality.
@@ -44,7 +44,7 @@ Using formats of different precisions will yield results of varying quality.
 
 
 ```
- .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4
+ .\bin\Release\sd.exe --diffusion-model  ..\models\flux1-schnell-q8_0.gguf --vae ..\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --steps 4 --clip-on-cpu
 ```
 
 | q8_0  |
@@ -60,7 +60,7 @@ Since many flux LoRA training libraries have used various LoRA naming formats, i
 - LoRA model from https://huggingface.co/XLabs-AI/flux-lora-collection/tree/main (using comfy converted version!!!)
 
 ```
-.\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models
+.\bin\Release\sd.exe --diffusion-model  ..\models\flux1-dev-q8_0.gguf --vae ...\models\ae.sft --clip_l ..\models\clip_l.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'<lora:realism_lora_comfy_converted:1>" --cfg-scale 1.0 --sampling-method euler -v --lora-model-dir ../models --clip-on-cpu
 ```
 
 ![output](../assets/flux/flux1-dev-q8_0%20with%20lora.png)

docs/kontext.md

@@ -27,7 +27,7 @@ You can download the preconverted gguf weights from [FLUX.1-Kontext-dev-GGUF](ht
 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
+ .\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 --clip-on-cpu
 ```
 
 

docs/lora.md

@@ -20,20 +20,30 @@ Here's a simple example:
 
 NOTE: The other backends may have different support.
 
-| Quant / Type | CUDA |
-|--------------|------|
-| F32          | ✔️   |
-| F16          | ✔️   |
-| BF16         | ✔️   |
-| I32          | ✔️   |
-| Q4_0         | ✔️   |
-| Q4_1         | ✔️   |
-| Q5_0         | ✔️   |
-| Q5_1         | ✔️   |
-| Q8_0         | ✔️   |
-| Q2_K         | ❌   |
-| Q3_K         | ❌   |
-| Q4_K         | ❌   |
-| Q5_K         | ❌   |
-| Q6_K         | ❌   |
-| Q8_K         | ❌   |
+| Quant / Type | CUDA | Vulkan |
+|--------------|------|--------|
+| F32          | ✔️   | ✔️   |
+| F16          | ✔️   | ✔️   |
+| BF16         | ✔️   | ✔️   |
+| I32          | ✔️   | ❌   |
+| Q4_0         | ✔️   | ✔️   |
+| Q4_1         | ✔️   | ✔️   |
+| Q5_0         | ✔️   | ✔️   |
+| Q5_1         | ✔️   | ✔️   |
+| Q8_0         | ✔️   | ✔️   |
+| Q2_K         | ❌   | ❌   |
+| Q3_K         | ❌   | ❌   |
+| Q4_K         | ❌   | ❌   |
+| Q5_K         | ❌   | ❌   |
+| Q6_K         | ❌   | ❌   |
+| Q8_K         | ❌   | ❌   |
+| IQ1_S        | ❌   | ✔️   |
+| IQ1_M        | ❌   | ✔️   |
+| IQ2_XXS      | ❌   | ✔️   |
+| IQ2_XS       | ❌   | ✔️   |
+| IQ2_S        | ❌   | ✔️   |
+| IQ3_XXS      | ❌   | ✔️   |
+| IQ3_S        | ❌   | ✔️   |
+| IQ4_XS       | ❌   | ✔️   |
+| IQ4_NL       | ❌   | ✔️   |
+| MXFP4        | ❌   | ✔️   |

docs/performance.md

@@ -0,0 +1,26 @@
+## Use Flash Attention to save memory and improve speed.
+
+Enabling flash attention for the diffusion model reduces memory usage by varying amounts of MB.
+eg.:
+ - flux 768x768 ~600mb
+ - SD2 768x768 ~1400mb
+
+For most backends, it slows things down, but for cuda it generally speeds it up too.
+At the moment, it is only supported for some models and some backends (like cpu, cuda/rocm, metal).
+
+Run by adding `--diffusion-fa` to the arguments and watch for:
+```
+[INFO ] stable-diffusion.cpp:312  - Using flash attention in the diffusion model
+```
+and the compute buffer shrink in the debug log:
+```
+[DEBUG] ggml_extend.hpp:1004 - flux compute buffer size: 650.00 MB(VRAM)
+```
+
+## Offload weights to the CPU to save VRAM without reducing generation speed.
+
+Using `--offload-to-cpu` allows you to offload weights to the CPU, saving VRAM without reducing generation speed.
+
+## Use quantization to reduce memory usage.
+
+[quantization](./quantization_and_gguf.md)

docs/photo_maker.md

@@ -6,16 +6,15 @@ You can use [PhotoMaker](https://github.com/TencentARC/PhotoMaker) to personaliz
 
 Download PhotoMaker model file (in safetensor format) [here](https://huggingface.co/bssrdf/PhotoMaker). The official release of the model file (in .bin format) does not work with ```stablediffusion.cpp```.
 
-- Specify the PhotoMaker model path using the `--stacked-id-embd-dir PATH` parameter.
-- Specify the input images path using the `--input-id-images-dir PATH` parameter.
-  - input images **must** have the same width and height for preprocessing (to be improved)
+- Specify the PhotoMaker model path using the `--photo-maker PATH` parameter.
+- Specify the input images path using the `--pm-id-images-dir PATH` parameter.
 
 In prompt, make sure you have a class word followed by the trigger word ```"img"``` (hard-coded for now). The class word could be one of ```"man, woman, girl, boy"```. If input ID images contain asian faces, add ```Asian``` before the class
 word.
 
 Another PhotoMaker specific parameter:
 
-- ```--style-ratio  (0-100)%```: default is 20 and 10-20 typically gets good results. Lower ratio means more faithfully following input ID (not necessarily better quality).
+- ```--pm-style-strength  (0-100)%```: default is 20 and 10-20 typically gets good results. Lower ratio means more faithfully following input ID (not necessarily better quality).
 
 Other parameters recommended for running Photomaker:
 
@@ -28,7 +27,7 @@ If on low memory GPUs (<= 8GB), recommend running with ```--vae-on-cpu``` option
 Example:
 
 ```bash
-bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --stacked-id-embd-dir ../models/photomaker-v1.safetensors --input-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --style-ratio 10 --vae-on-cpu -o output.png
+bin/sd -m ../models/sdxlUnstableDiffusers_v11.safetensors  --vae ../models/sdxl_vae.safetensors --photo-maker ../models/photomaker-v1.safetensors --pm-id-images-dir ../assets/photomaker_examples/scarletthead_woman -p "a girl img, retro futurism, retro game art style but extremely beautiful, intricate details, masterpiece, best quality, space-themed, cosmic, celestial, stars, galaxies, nebulas, planets, science fiction, highly detailed" -n "realistic, photo-realistic, worst quality, greyscale, bad anatomy, bad hands, error, text" --cfg-scale 5.0  --sampling-method euler -H 1024 -W 1024 --pm-style-strength 10 --vae-on-cpu --steps 50
 ```
 
 ## PhotoMaker Version 2

docs/qwen_image.md

@@ -0,0 +1,23 @@
+# How to Use
+
+## Download weights
+
+- Download Qwen Image
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/QuantStack/Qwen-Image-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
+- Download qwen_2.5_vl 7b
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-GGUF/tree/main
+
+## Examples
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\qwen-image-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --qwen2vl ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf  -p '一个穿着"QWEN"标志的T恤的中国美女正拿着黑色的马克笔面相镜头微笑。她身后的玻璃板上手写体写着 “一、Qwen-Image的技术路线： 探索视觉生成基础模型的极限，开创理解与生成一体化的未来。二、Qwen-Image的模型特色：1、复杂文字渲染。支持中英渲染、自动布局； 2、精准图像编辑。支持文字编辑、物体增减、风格变换。三、Qwen-Image的未来愿景：赋能专业内容创作、助力生成式AI发展。”' --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu -H 1024 -W 1024 --diffusion-fa --flow-shift 3
+```
+
+<img alt="qwen example" src="../assets/qwen/example.png" />
+
+
+

docs/qwen_image_edit.md

@@ -0,0 +1,35 @@
+# How to Use
+
+## Download weights
+
+- Download Qwen Image
+    - Qwen Image Edit
+        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
+        - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-GGUF/tree/main
+    - Qwen Image Edit 2509
+        - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-Edit_ComfyUI/tree/main/split_files/diffusion_models
+        - gguf: https://huggingface.co/QuantStack/Qwen-Image-Edit-2509-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/vae
+- Download qwen_2.5_vl 7b
+    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image_ComfyUI/tree/main/split_files/text_encoders
+    - gguf: https://huggingface.co/mradermacher/Qwen2.5-VL-7B-Instruct-GGUF/tree/main
+
+## Examples
+
+### Qwen Image Edit
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen_Image_Edit-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --qwen2vl ..\..\ComfyUI\models\text_encoders\qwen_2.5_vl_7b.safetensors --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'edit.cpp'" --seed 1118877715456453
+```
+
+<img alt="qwen_image_edit" src="../assets/qwen/qwen_image_edit.png" />
+
+
+### Qwen Image Edit 2509
+
+```
+.\bin\Release\sd.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Qwen-Image-Edit-2509-Q4_K_S.gguf --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --qwen2vl ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf --qwen2vl_vision ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct.mmproj-Q8_0.gguf --cfg-scale 2.5 --sampling-method euler -v --offload-to-cpu --diffusion-fa --flow-shift 3 -r ..\assets\flux\flux1-dev-q8_0.png -p "change 'flux.cpp' to 'Qwen Image Edit 2509'"
+```
+
+<img alt="qwen_image_edit_2509" src="../assets/qwen/qwen_image_edit_2509.png" />

docs/sd.md

@@ -0,0 +1,37 @@
+## Download weights
+
+- download original weights(.ckpt or .safetensors). For example
+    - Stable Diffusion v1.4 from https://huggingface.co/CompVis/stable-diffusion-v-1-4-original
+    - Stable Diffusion v1.5 from https://huggingface.co/runwayml/stable-diffusion-v1-5
+    - Stable Diffuison v2.1 from https://huggingface.co/stabilityai/stable-diffusion-2-1
+    - Stable Diffusion 3 2B from https://huggingface.co/stabilityai/stable-diffusion-3-medium
+
+### txt2img example
+
+```sh
+./bin/sd -m ../models/sd-v1-4.ckpt -p "a lovely cat"
+# ./bin/sd -m ../models/v1-5-pruned-emaonly.safetensors -p "a lovely cat"
+# ./bin/sd -m ../models/sd_xl_base_1.0.safetensors --vae ../models/sdxl_vae-fp16-fix.safetensors -H 1024 -W 1024 -p "a lovely cat" -v
+# ./bin/sd -m ../models/sd3_medium_incl_clips_t5xxlfp16.safetensors -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable Diffusion CPP\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd --diffusion-model  ../models/flux1-dev-q3_k.gguf --vae ../models/ae.sft --clip_l ../models/clip_l.safetensors --t5xxl ../models/t5xxl_fp16.safetensors  -p "a lovely cat holding a sign says 'flux.cpp'" --cfg-scale 1.0 --sampling-method euler -v --clip-on-cpu
+# ./bin/sd -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
+```
+
+Using formats of different precisions will yield results of varying quality.
+
+| f32  | f16  |q8_0  |q5_0  |q5_1  |q4_0  |q4_1  |
+| ----  |----  |----  |----  |----  |----  |----  |
+| ![](../assets/f32.png) |![](../assets/f16.png) |![](../assets/q8_0.png) |![](../assets/q5_0.png) |![](../assets/q5_1.png) |![](../assets/q4_0.png) |![](../assets/q4_1.png) |
+
+### img2img example
+
+- `./output.png` is the image generated from the above txt2img pipeline
+
+
+```
+./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">
+  <img src="../assets/img2img_output.png" width="256x">
+</p>

docs/sd3.md

@@ -14,7 +14,7 @@
 For example:
 
 ```
-.\bin\Release\sd.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v
+.\bin\Release\sd.exe -m  ..\models\sd3.5_large.safetensors --clip_l ..\models\clip_l.safetensors --clip_g ..\models\clip_g.safetensors --t5xxl ..\models\t5xxl_fp16.safetensors  -H 1024 -W 1024 -p 'a lovely cat holding a sign says \"Stable diffusion 3.5 Large\"' --cfg-scale 4.5 --sampling-method euler -v --clip-on-cpu
 ```
 
 ![](../assets/sd3.5_large.png)

docs/wan.md

@@ -18,6 +18,12 @@
         - Wan2.1 FLF2V 14B 720P
             - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
             - gguf: https://huggingface.co/city96/Wan2.1-FLF2V-14B-720P-gguf/tree/main
+        - Wan2.1 VACE 1.3B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/calcuis/wan-1.3b-gguf/tree/main
+        - Wan2.1 VACE 14B
+            - safetensors: https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged/tree/main/split_files/diffusion_models
+            - gguf: https://huggingface.co/QuantStack/Wan2.1_14B_VACE-GGUF/tree/main
     - Wan2.2
         - Wan2.2 TI2V 5B
             - safetensors: https://huggingface.co/Comfy-Org/Wan_2.2_ComfyUI_Repackaged/tree/main/split_files/diffusion_models
@@ -137,3 +143,62 @@
 ```
 
 <video src=../assets/wan/Wan2.2_14B_flf2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 VACE 1.3B
+
+#### T2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 1 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### R2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### V2V
+
+```
+mkdir post+depth
+ffmpeg -i ..\..\ComfyUI\input\post+depth.mp4 -qscale:v 1 -vf fps=8 post+depth\frame_%04d.jpg
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\wan2.1-vace-1.3b-q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_1.3B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+### Wan2.1 VACE 14B
+
+#### T2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_t2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+#### R2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa -i ..\assets\cat_with_sd_cpp_42.png --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_r2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>
+
+
+
+#### V2V
+
+```
+.\bin\Release\sd.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\Wan2.1_14B_VACE-Q8_0.gguf --vae ..\..\ComfyUI\models\vae\wan_2.1_vae.safetensors --t5xxl ..\..\ComfyUI\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "The girl is dancing in a sea of flowers, slowly moving her hands. There is a close - up shot of her upper body. The character is surrounded by other transparent glass flowers in the style of Nicoletta Ceccoli, creating a beautiful, surreal, and emotionally expressive movie scene with a white. transparent feel and a dreamyl atmosphere." --cfg-scale 6.0 --sampling-method euler -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部， 畸形的，毁容的，形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 480 -H 832 --diffusion-fa -i ..\..\ComfyUI\input\dance_girl.jpg --control-video ./post+depth --video-frames 33 --offload-to-cpu
+```
+
+<video src=../assets/wan/Wan2.1_14B_vace_v2v.mp4 controls="controls" muted="muted" type="video/mp4"></video>

esrgan.hpp

@@ -83,39 +83,44 @@ public:
 
 class RRDBNet : public GGMLBlock {
 protected:
-    int scale       = 4;  // default RealESRGAN_x4plus_anime_6B
-    int num_block   = 6;  // default RealESRGAN_x4plus_anime_6B
+    int scale       = 4;
+    int num_block   = 23;
     int num_in_ch   = 3;
     int num_out_ch  = 3;
-    int num_feat    = 64;  // default RealESRGAN_x4plus_anime_6B
-    int num_grow_ch = 32;  // default RealESRGAN_x4plus_anime_6B
+    int num_feat    = 64;
+    int num_grow_ch = 32;
 
 public:
-    RRDBNet() {
+    RRDBNet(int scale, int num_block, int num_in_ch, int num_out_ch, int num_feat, int num_grow_ch)
+        : scale(scale), num_block(num_block), num_in_ch(num_in_ch), num_out_ch(num_out_ch), num_feat(num_feat), num_grow_ch(num_grow_ch) {
         blocks["conv_first"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_in_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
         for (int i = 0; i < num_block; i++) {
             std::string name = "body." + std::to_string(i);
             blocks[name]     = std::shared_ptr<GGMLBlock>(new RRDB(num_feat, num_grow_ch));
         }
         blocks["conv_body"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
-        // upsample
-        blocks["conv_up1"]  = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
-        blocks["conv_up2"]  = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        if (scale >= 2) {
+            blocks["conv_up1"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        }
+        if (scale == 4) {
+            blocks["conv_up2"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
+        }
         blocks["conv_hr"]   = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_feat, {3, 3}, {1, 1}, {1, 1}));
         blocks["conv_last"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_out_ch, {3, 3}, {1, 1}, {1, 1}));
     }
 
+    int get_scale() { return scale; }
+    int get_num_block() { return num_block; }
+
     struct ggml_tensor* lrelu(struct ggml_context* ctx, struct ggml_tensor* x) {
         return ggml_leaky_relu(ctx, x, 0.2f, true);
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
         // x: [n, num_in_ch, h, w]
-        // return: [n, num_out_ch, h*4, w*4]
+        // return: [n, num_out_ch, h*scale, w*scale]
         auto conv_first = std::dynamic_pointer_cast<Conv2d>(blocks["conv_first"]);
         auto conv_body  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_body"]);
-        auto conv_up1   = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
-        auto conv_up2   = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
         auto conv_hr    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_hr"]);
         auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
 
@@ -130,15 +135,22 @@ 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, GGML_SCALE_MODE_NEAREST)));
-        feat     = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+        if (scale >= 2) {
+            auto conv_up1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
+            feat          = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+            if (scale == 4) {
+                auto conv_up2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
+                feat          = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+            }
+        }
+        // for all scales
         auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
         return out;
     }
 };
 
 struct ESRGAN : public GGMLRunner {
-    RRDBNet rrdb_net;
+    std::unique_ptr<RRDBNet> rrdb_net;
     int scale     = 4;
     int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
 
@@ -146,12 +158,14 @@ struct ESRGAN : public GGMLRunner {
            bool offload_params_to_cpu,
            const String2GGMLType& tensor_types = {})
         : GGMLRunner(backend, offload_params_to_cpu) {
-        rrdb_net.init(params_ctx, tensor_types, "");
+        // rrdb_net will be created in load_from_file
     }
 
     void enable_conv2d_direct() {
+        if (!rrdb_net)
+            return;
         std::vector<GGMLBlock*> blocks;
-        rrdb_net.get_all_blocks(blocks);
+        rrdb_net->get_all_blocks(blocks);
         for (auto block : blocks) {
             if (block->get_desc() == "Conv2d") {
                 auto conv_block = (Conv2d*)block;
@@ -160,38 +174,192 @@ struct ESRGAN : public GGMLRunner {
         }
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "esrgan";
     }
 
-    bool load_from_file(const std::string& file_path) {
+    bool load_from_file(const std::string& file_path, int n_threads) {
         LOG_INFO("loading esrgan from '%s'", file_path.c_str());
 
-        alloc_params_buffer();
-        std::map<std::string, ggml_tensor*> esrgan_tensors;
-        rrdb_net.get_param_tensors(esrgan_tensors);
-
         ModelLoader model_loader;
         if (!model_loader.init_from_file(file_path)) {
             LOG_ERROR("init esrgan model loader from file failed: '%s'", file_path.c_str());
             return false;
         }
 
-        bool success = model_loader.load_tensors(esrgan_tensors);
+        // Get tensor names
+        auto tensor_names = model_loader.get_tensor_names();
+
+        // Detect if it's ESRGAN format
+        bool is_ESRGAN = std::find(tensor_names.begin(), tensor_names.end(), "model.0.weight") != tensor_names.end();
+
+        // Detect parameters from tensor names
+        int detected_num_block = 0;
+        if (is_ESRGAN) {
+            for (const auto& name : tensor_names) {
+                if (name.find("model.1.sub.") == 0) {
+                    size_t first_dot = name.find('.', 12);
+                    if (first_dot != std::string::npos) {
+                        size_t second_dot = name.find('.', first_dot + 1);
+                        if (second_dot != std::string::npos && name.substr(first_dot + 1, 3) == "RDB") {
+                            try {
+                                int idx            = std::stoi(name.substr(12, first_dot - 12));
+                                detected_num_block = std::max(detected_num_block, idx + 1);
+                            } catch (...) {
+                            }
+                        }
+                    }
+                }
+            }
+        } else {
+            // Original format
+            for (const auto& name : tensor_names) {
+                if (name.find("body.") == 0) {
+                    size_t pos = name.find('.', 5);
+                    if (pos != std::string::npos) {
+                        try {
+                            int idx            = std::stoi(name.substr(5, pos - 5));
+                            detected_num_block = std::max(detected_num_block, idx + 1);
+                        } catch (...) {
+                        }
+                    }
+                }
+            }
+        }
+
+        int detected_scale = 4;  // default
+        if (is_ESRGAN) {
+            // For ESRGAN format, detect scale by highest model number
+            int max_model_num = 0;
+            for (const auto& name : tensor_names) {
+                if (name.find("model.") == 0) {
+                    size_t dot_pos = name.find('.', 6);
+                    if (dot_pos != std::string::npos) {
+                        try {
+                            int num       = std::stoi(name.substr(6, dot_pos - 6));
+                            max_model_num = std::max(max_model_num, num);
+                        } catch (...) {
+                        }
+                    }
+                }
+            }
+            if (max_model_num <= 4) {
+                detected_scale = 1;
+            } else if (max_model_num <= 7) {
+                detected_scale = 2;
+            } else {
+                detected_scale = 4;
+            }
+        } else {
+            // Original format
+            bool has_conv_up2 = std::any_of(tensor_names.begin(), tensor_names.end(), [](const std::string& name) {
+                return name == "conv_up2.weight";
+            });
+            bool has_conv_up1 = std::any_of(tensor_names.begin(), tensor_names.end(), [](const std::string& name) {
+                return name == "conv_up1.weight";
+            });
+            if (has_conv_up2) {
+                detected_scale = 4;
+            } else if (has_conv_up1) {
+                detected_scale = 2;
+            } else {
+                detected_scale = 1;
+            }
+        }
+
+        int detected_num_in_ch   = 3;
+        int detected_num_out_ch  = 3;
+        int detected_num_feat    = 64;
+        int detected_num_grow_ch = 32;
+
+        // Create RRDBNet with detected parameters
+        rrdb_net = std::make_unique<RRDBNet>(detected_scale, detected_num_block, detected_num_in_ch, detected_num_out_ch, detected_num_feat, detected_num_grow_ch);
+        rrdb_net->init(params_ctx, {}, "");
+
+        alloc_params_buffer();
+        std::map<std::string, ggml_tensor*> esrgan_tensors;
+        rrdb_net->get_param_tensors(esrgan_tensors);
+
+        bool success;
+        if (is_ESRGAN) {
+            // Build name mapping for ESRGAN format
+            std::map<std::string, std::string> expected_to_model;
+            expected_to_model["conv_first.weight"] = "model.0.weight";
+            expected_to_model["conv_first.bias"]   = "model.0.bias";
+
+            for (int i = 0; i < detected_num_block; i++) {
+                for (int j = 1; j <= 3; j++) {
+                    for (int k = 1; k <= 5; k++) {
+                        std::string expected_weight        = "body." + std::to_string(i) + ".rdb" + std::to_string(j) + ".conv" + std::to_string(k) + ".weight";
+                        std::string model_weight           = "model.1.sub." + std::to_string(i) + ".RDB" + std::to_string(j) + ".conv" + std::to_string(k) + ".0.weight";
+                        expected_to_model[expected_weight] = model_weight;
+
+                        std::string expected_bias        = "body." + std::to_string(i) + ".rdb" + std::to_string(j) + ".conv" + std::to_string(k) + ".bias";
+                        std::string model_bias           = "model.1.sub." + std::to_string(i) + ".RDB" + std::to_string(j) + ".conv" + std::to_string(k) + ".0.bias";
+                        expected_to_model[expected_bias] = model_bias;
+                    }
+                }
+            }
+
+            if (detected_scale == 1) {
+                expected_to_model["conv_body.weight"] = "model.1.sub." + std::to_string(detected_num_block) + ".weight";
+                expected_to_model["conv_body.bias"]   = "model.1.sub." + std::to_string(detected_num_block) + ".bias";
+                expected_to_model["conv_hr.weight"]   = "model.2.weight";
+                expected_to_model["conv_hr.bias"]     = "model.2.bias";
+                expected_to_model["conv_last.weight"] = "model.4.weight";
+                expected_to_model["conv_last.bias"]   = "model.4.bias";
+            } else {
+                expected_to_model["conv_body.weight"] = "model.1.sub." + std::to_string(detected_num_block) + ".weight";
+                expected_to_model["conv_body.bias"]   = "model.1.sub." + std::to_string(detected_num_block) + ".bias";
+                if (detected_scale >= 2) {
+                    expected_to_model["conv_up1.weight"] = "model.3.weight";
+                    expected_to_model["conv_up1.bias"]   = "model.3.bias";
+                }
+                if (detected_scale == 4) {
+                    expected_to_model["conv_up2.weight"]  = "model.6.weight";
+                    expected_to_model["conv_up2.bias"]    = "model.6.bias";
+                    expected_to_model["conv_hr.weight"]   = "model.8.weight";
+                    expected_to_model["conv_hr.bias"]     = "model.8.bias";
+                    expected_to_model["conv_last.weight"] = "model.10.weight";
+                    expected_to_model["conv_last.bias"]   = "model.10.bias";
+                } else if (detected_scale == 2) {
+                    expected_to_model["conv_hr.weight"]   = "model.5.weight";
+                    expected_to_model["conv_hr.bias"]     = "model.5.bias";
+                    expected_to_model["conv_last.weight"] = "model.7.weight";
+                    expected_to_model["conv_last.bias"]   = "model.7.bias";
+                }
+            }
+
+            std::map<std::string, ggml_tensor*> model_tensors;
+            for (auto& p : esrgan_tensors) {
+                auto it = expected_to_model.find(p.first);
+                if (it != expected_to_model.end()) {
+                    model_tensors[it->second] = p.second;
+                }
+            }
+
+            success = model_loader.load_tensors(model_tensors, {}, n_threads);
+        } else {
+            success = model_loader.load_tensors(esrgan_tensors, {}, n_threads);
+        }
 
         if (!success) {
             LOG_ERROR("load esrgan tensors from model loader failed");
             return false;
         }
 
-        LOG_INFO("esrgan model loaded");
+        scale = rrdb_net->get_scale();
+        LOG_INFO("esrgan model loaded with scale=%d, num_block=%d", scale, detected_num_block);
         return success;
     }
 
     struct ggml_cgraph* build_graph(struct ggml_tensor* x) {
-        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
-        x                       = to_backend(x);
-        struct ggml_tensor* out = rrdb_net.forward(compute_ctx, x);
+        if (!rrdb_net)
+            return nullptr;
+        constexpr int kGraphNodes = 1 << 16;  // 65k
+        struct ggml_cgraph* gf    = ggml_new_graph_custom(compute_ctx, kGraphNodes, /*grads*/ false);
+        x                         = to_backend(x);
+        struct ggml_tensor* out   = rrdb_net->forward(compute_ctx, x);
         ggml_build_forward_expand(gf, out);
         return gf;
     }
@@ -199,7 +367,7 @@ struct ESRGAN : public GGMLRunner {
     void compute(const int n_threads,
                  struct ggml_tensor* x,
                  ggml_tensor** output,
-                 ggml_context* output_ctx = NULL) {
+                 ggml_context* output_ctx = nullptr) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
             return build_graph(x);
         };

examples/cli/README.md

@@ -0,0 +1,110 @@
+# Run
+
+```
+usage: ./bin/sd  [options]
+
+Options:
+  -m, --model <string>                     path to full model
+  --clip_l <string>                        path to the clip-l text encoder
+  --clip_g <string>                        path to the clip-g text encoder
+  --clip_vision <string>                   path to the clip-vision encoder
+  --t5xxl <string>                         path to the t5xxl text encoder
+  --qwen2vl <string>                       path to the qwen2vl text encoder
+  --qwen2vl_vision <string>                path to the qwen2vl vit
+  --diffusion-model <string>               path to the standalone diffusion model
+  --high-noise-diffusion-model <string>    path to the standalone high noise diffusion model
+  --vae <string>                           path to standalone vae model
+  --taesd <string>                         path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
+  --control-net <string>                   path to control net model
+  --embd-dir <string>                      embeddings directory
+  --lora-model-dir <string>                lora model directory
+  -i, --init-img <string>                  path to the init image
+  --end-img <string>                       path to the end image, required by flf2v
+  --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
+  --photo-maker <string>                   path to PHOTOMAKER model
+  --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
+  --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
+  --mask <string>                          path to the mask image
+  --control-image <string>                 path to control image, control net
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
+                                           lexicographical (character) order. For example, if the control video path is
+                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
+  -o, --output <string>                    path to write result image to (default: ./output.png)
+  -p, --prompt <string>                    the prompt to render
+  -n, --negative-prompt <string>           the negative prompt (default: "")
+  --upscale-model <string>                 path to esrgan model.
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
+                                           CPU physical cores
+  --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
+  -H, --height <int>                       image height, in pixel space (default: 512)
+  -W, --width <int>                        image width, in pixel space (default: 512)
+  --steps <int>                            number of sample steps (default: 20)
+  --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
+                                           will be 1 for SD1.x, 2 for SD2.x
+  -b, --batch-count <int>                  batch count
+  --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
+  --video-frames <int>                     video frames (default: 1)
+  --fps <int>                              fps (default: 24)
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
+                                           NitroSD-Vibrant
+  --cfg-scale <float>                      unconditional guidance scale: (default: 7.0)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)
+  --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
+  --slg-scale <float>                      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
+  --skip-layer-start <float>               SLG enabling point (default: 0.01)
+  --skip-layer-end <float>                 SLG disabling point (default: 0.2)
+  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
+  --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
+  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+  --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
+  --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
+  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM and TCD (default: 0)
+  --strength <float>                       strength for noising/unnoising (default: 0.75)
+  --pm-style-strength <float>
+  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
+  --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
+  --vace-strength <float>                  wan vace strength
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --vae-tiling                             process vae in tiles to reduce memory usage
+  --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
+  --control-net-cpu                        keep controlnet in cpu (for low vram)
+  --clip-on-cpu                            keep clip in cpu (for low vram)
+  --vae-on-cpu                             keep vae in cpu (for low vram)
+  --diffusion-fa                           use flash attention in the diffusion model
+  --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
+  --vae-conv-direct                        use ggml_conv2d_direct in the vae model
+  --canny                                  apply canny preprocessor (edge detection)
+  -v, --verbose                            print extra info
+  --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
+  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
+  --disable-auto-resize-ref-image          disable auto resize of ref images
+  -M, --mode                               run mode, one of [img_gen, vid_gen, upscale, convert], default: img_gen
+  --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
+  --rng                                    RNG, one of [std_default, cuda], default: cuda
+  -s, --seed                               RNG seed (default: 42, use random seed for < 0)
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
+                                           tcd] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow]
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple], default:
+                                           discrete
+  --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
+  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
+                                           ddim_trailing, tcd] default: euler for Flux/SD3/Wan, euler_a otherwise
+  --high-noise-scheduler                   (high noise) denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform,
+                                           simple], default: discrete
+  --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
+  -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
+  -h, --help                               show this help message and exit
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
+                                           (overrides --vae-tile-size)
+```

examples/cli/avi_writer.h

@@ -1,10 +1,10 @@
 #ifndef __AVI_WRITER_H__
 #define __AVI_WRITER_H__
 
-#include <stdint.h>
-#include <stdio.h>
-#include <stdlib.h>
-#include <string.h>
+#include <cstdint>
+#include <cstdio>
+#include <cstdlib>
+#include <cstring>
 
 #include "stable-diffusion.h"
 
@@ -130,7 +130,7 @@ int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int
     write_u32_le(f, 0);                   // Colors important
 
     // 'movi' LIST (video frames)
-    long movi_list_pos = ftell(f);
+    // long movi_list_pos = ftell(f);
     fwrite("LIST", 4, 1, f);
     long movi_size_pos = ftell(f);
     write_u32_le(f, 0);  // Placeholder for movi size
@@ -149,7 +149,7 @@ int create_mjpg_avi_from_sd_images(const char* filename, sd_image_t* images, int
     } jpeg_data;
 
     for (int i = 0; i < num_images; i++) {
-        jpeg_data.buf  = NULL;
+        jpeg_data.buf  = nullptr;
         jpeg_data.size = 0;
 
         // Callback function to collect JPEG data into memory

flux.hpp

@@ -1,6 +1,7 @@
 #ifndef __FLUX_HPP__
 #define __FLUX_HPP__
 
+#include <memory>
 #include <vector>
 
 #include "ggml_extend.hpp"
@@ -18,7 +19,7 @@ namespace Flux {
             blocks["out_layer"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_dim, hidden_dim, true));
         }
 
-        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
             // x: [..., in_dim]
             // return: [..., hidden_dim]
             auto in_layer  = std::dynamic_pointer_cast<Linear>(blocks["in_layer"]);
@@ -36,7 +37,7 @@ namespace Flux {
         int64_t hidden_size;
         float eps;
 
-        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
             ggml_type wtype = GGML_TYPE_F32;
             params["scale"] = ggml_new_tensor_1d(ctx, wtype, hidden_size);
         }
@@ -47,7 +48,7 @@ namespace Flux {
             : hidden_size(hidden_size),
               eps(eps) {}
 
-        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
             struct ggml_tensor* w = params["scale"];
             x                     = ggml_rms_norm(ctx, x, eps);
             x                     = ggml_mul(ctx, x, w);
@@ -81,56 +82,6 @@ namespace Flux {
         }
     };
 
-    __STATIC_INLINE__ struct ggml_tensor* apply_rope(struct ggml_context* ctx,
-                                                     struct ggml_tensor* x,
-                                                     struct ggml_tensor* pe) {
-        // x: [N, L, n_head, d_head]
-        // pe: [L, d_head/2, 2, 2]
-        int64_t d_head = x->ne[0];
-        int64_t n_head = x->ne[1];
-        int64_t L      = x->ne[2];
-        int64_t N      = x->ne[3];
-        x              = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));       // [N, n_head, L, d_head]
-        x              = ggml_reshape_4d(ctx, x, 2, d_head / 2, L, n_head * N);  // [N * n_head, L, d_head/2, 2]
-        x              = ggml_cont(ctx, ggml_permute(ctx, x, 3, 0, 1, 2));       // [2, N * n_head, L, d_head/2]
-
-        int64_t offset = x->nb[2] * x->ne[2];
-        auto x_0       = ggml_view_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2], x->nb[1], x->nb[2], offset * 0);  // [N * n_head, L, d_head/2]
-        auto x_1       = ggml_view_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2], x->nb[1], x->nb[2], offset * 1);  // [N * n_head, L, d_head/2]
-        x_0            = ggml_reshape_4d(ctx, x_0, 1, x_0->ne[0], x_0->ne[1], x_0->ne[2]);                    // [N * n_head, L, d_head/2, 1]
-        x_1            = ggml_reshape_4d(ctx, x_1, 1, x_1->ne[0], x_1->ne[1], x_1->ne[2]);                    // [N * n_head, L, d_head/2, 1]
-        auto temp_x    = ggml_new_tensor_4d(ctx, x_0->type, 2, x_0->ne[1], x_0->ne[2], x_0->ne[3]);
-        x_0            = ggml_repeat(ctx, x_0, temp_x);  // [N * n_head, L, d_head/2, 2]
-        x_1            = ggml_repeat(ctx, x_1, temp_x);  // [N * n_head, L, d_head/2, 2]
-
-        pe        = ggml_cont(ctx, ggml_permute(ctx, pe, 3, 0, 1, 2));  // [2, L, d_head/2, 2]
-        offset    = pe->nb[2] * pe->ne[2];
-        auto pe_0 = ggml_view_3d(ctx, pe, pe->ne[0], pe->ne[1], pe->ne[2], pe->nb[1], pe->nb[2], offset * 0);  // [L, d_head/2, 2]
-        auto pe_1 = ggml_view_3d(ctx, pe, pe->ne[0], pe->ne[1], pe->ne[2], pe->nb[1], pe->nb[2], offset * 1);  // [L, d_head/2, 2]
-
-        auto x_out = ggml_add_inplace(ctx, ggml_mul(ctx, x_0, pe_0), ggml_mul(ctx, x_1, pe_1));  // [N * n_head, L, d_head/2, 2]
-        x_out      = ggml_reshape_3d(ctx, x_out, d_head, L, n_head * N);                         // [N*n_head, L, d_head]
-        return x_out;
-    }
-
-    __STATIC_INLINE__ struct ggml_tensor* attention(struct ggml_context* ctx,
-                                                    ggml_backend_t backend,
-                                                    struct ggml_tensor* q,
-                                                    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]
-        // return: [N, L, n_head*d_head]
-        q = apply_rope(ctx, q, pe);  // [N*n_head, L, d_head]
-        k = apply_rope(ctx, k, pe);  // [N*n_head, L, d_head]
-
-        auto x = ggml_nn_attention_ext(ctx, backend, q, k, v, v->ne[1], mask, false, true, flash_attn);  // [N, L, n_head*d_head]
-        return x;
-    }
-
     struct SelfAttention : public GGMLBlock {
     public:
         int64_t num_heads;
@@ -178,19 +129,19 @@ namespace Flux {
             // 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, backend, qkv[0], qkv[1], qkv[2], pe, mask, 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        = Rope::attention(ctx, backend, 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;
         }
     };
 
     struct ModulationOut {
-        ggml_tensor* shift = NULL;
-        ggml_tensor* scale = NULL;
-        ggml_tensor* gate  = NULL;
+        ggml_tensor* shift = nullptr;
+        ggml_tensor* scale = nullptr;
+        ggml_tensor* gate  = nullptr;
 
-        ModulationOut(ggml_tensor* shift = NULL, ggml_tensor* scale = NULL, ggml_tensor* gate = NULL)
+        ModulationOut(ggml_tensor* shift = nullptr, ggml_tensor* scale = nullptr, ggml_tensor* gate = nullptr)
             : shift(shift), scale(scale), gate(gate) {}
 
         ModulationOut(struct ggml_context* ctx, ggml_tensor* vec, int64_t offset) {
@@ -309,7 +260,7 @@ namespace Flux {
                                                                     struct ggml_tensor* txt,
                                                                     struct ggml_tensor* vec,
                                                                     struct ggml_tensor* pe,
-                                                                    struct ggml_tensor* mask = NULL) {
+                                                                    struct ggml_tensor* mask = nullptr) {
             // 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]
@@ -368,8 +319,8 @@ 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, backend, 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 attn         = Rope::attention(ctx, backend, 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,
                                              attn->ne[0],
@@ -448,7 +399,7 @@ namespace Flux {
 
         ModulationOut get_distil_mod(struct ggml_context* ctx, struct ggml_tensor* vec) {
             int64_t offset = 3 * idx;
-            return ModulationOut(ctx, vec, offset);
+            return {ctx, vec, offset};
         }
 
         struct ggml_tensor* forward(struct ggml_context* ctx,
@@ -456,7 +407,7 @@ namespace Flux {
                                     struct ggml_tensor* x,
                                     struct ggml_tensor* vec,
                                     struct ggml_tensor* pe,
-                                    struct ggml_tensor* mask = NULL) {
+                                    struct ggml_tensor* mask = nullptr) {
             // x: [N, n_token, hidden_size]
             // pe: [n_token, d_head/2, 2, 2]
             // return: [N, n_token, hidden_size]
@@ -503,7 +454,7 @@ 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, backend, q, k, v, pe, mask, flash_attn);  // [N, n_token, hidden_size]
+            auto attn        = Rope::attention(ctx, backend, 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]
@@ -535,7 +486,7 @@ namespace Flux {
             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);
+            return {shift, scale, nullptr};
         }
 
         struct ggml_tensor* forward(struct ggml_context* ctx,
@@ -615,6 +566,7 @@ namespace Flux {
         bool guidance_embed         = true;
         bool flash_attn             = true;
         bool is_chroma              = false;
+        SDVersion version           = VERSION_FLUX;
     };
 
     struct Flux : public GGMLBlock {
@@ -713,7 +665,7 @@ namespace Flux {
                                          struct ggml_tensor* y,
                                          struct ggml_tensor* guidance,
                                          struct ggml_tensor* pe,
-                                         struct ggml_tensor* mod_index_arange = NULL,
+                                         struct ggml_tensor* mod_index_arange = nullptr,
                                          std::vector<int> skip_layers         = {}) {
             auto img_in      = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
             auto txt_in      = std::dynamic_pointer_cast<Linear>(blocks["txt_in"]);
@@ -721,7 +673,7 @@ namespace Flux {
 
             img = img_in->forward(ctx, img);
             struct ggml_tensor* vec;
-            struct ggml_tensor* txt_img_mask = NULL;
+            struct ggml_tensor* txt_img_mask = nullptr;
             if (params.is_chroma) {
                 int64_t mod_index_length = 344;
                 auto approx              = std::dynamic_pointer_cast<ChromaApproximator>(blocks["distilled_guidance_layer"]);
@@ -730,7 +682,7 @@ namespace Flux {
 
                 // 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);
+                GGML_ASSERT(arange != nullptr);
                 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)
@@ -744,7 +696,7 @@ namespace Flux {
                 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) {
+                if (y != nullptr) {
                     txt_img_mask = ggml_pad(ctx, y, img->ne[1], 0, 0, 0);
                 }
             } else {
@@ -752,7 +704,7 @@ namespace Flux {
                 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);
+                    GGML_ASSERT(guidance != nullptr);
                     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);
@@ -824,14 +776,14 @@ namespace Flux {
                                     struct ggml_tensor* y,
                                     struct ggml_tensor* guidance,
                                     struct ggml_tensor* pe,
-                                    struct ggml_tensor* mod_index_arange  = NULL,
+                                    struct ggml_tensor* mod_index_arange  = nullptr,
                                     std::vector<ggml_tensor*> ref_latents = {},
                                     std::vector<int> skip_layers          = {}) {
             // Forward pass of DiT.
             // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
             // timestep: (N,) tensor of diffusion timesteps
             // context: (N, L, D)
-            // c_concat: NULL, or for (N,C+M, H, W) for Fill
+            // c_concat: nullptr, or for (N,C+M, H, W) for Fill
             // y: (N, adm_in_channels) tensor of class labels
             // guidance: (N,)
             // pe: (L, d_head/2, 2, 2)
@@ -849,7 +801,8 @@ namespace Flux {
             auto img            = process_img(ctx, x);
             uint64_t img_tokens = img->ne[1];
 
-            if (c_concat != NULL) {
+            if (params.version == VERSION_FLUX_FILL) {
+                GGML_ASSERT(c_concat != nullptr);
                 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);
 
@@ -857,6 +810,27 @@ namespace Flux {
                 mask   = process_img(ctx, mask);
 
                 img = ggml_concat(ctx, img, ggml_concat(ctx, masked, mask, 0), 0);
+            } else if (params.version == VERSION_FLEX_2) {
+                GGML_ASSERT(c_concat != nullptr);
+                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], 1, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C);
+                ggml_tensor* control = ggml_view_4d(ctx, c_concat, c_concat->ne[0], c_concat->ne[1], C, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * (C + 1));
+
+                masked  = ggml_pad(ctx, masked, pad_w, pad_h, 0, 0);
+                mask    = ggml_pad(ctx, mask, pad_w, pad_h, 0, 0);
+                control = ggml_pad(ctx, control, pad_w, pad_h, 0, 0);
+
+                masked  = patchify(ctx, masked, patch_size);
+                mask    = patchify(ctx, mask, patch_size);
+                control = patchify(ctx, control, patch_size);
+
+                img = ggml_concat(ctx, img, ggml_concat(ctx, ggml_concat(ctx, masked, mask, 0), control, 0), 0);
+            } else if (params.version == VERSION_FLUX_CONTROLS) {
+                GGML_ASSERT(c_concat != nullptr);
+
+                ggml_tensor* control = ggml_pad(ctx, c_concat, pad_w, pad_h, 0, 0);
+                control              = patchify(ctx, control, patch_size);
+                img                  = ggml_concat(ctx, img, control, 0);
             }
 
             if (ref_latents.size() > 0) {
@@ -867,6 +841,7 @@ namespace Flux {
             }
 
             auto out = forward_orig(ctx, backend, 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);
@@ -896,13 +871,18 @@ namespace Flux {
                    SDVersion version                   = VERSION_FLUX,
                    bool flash_attn                     = false,
                    bool use_mask                       = false)
-            : GGMLRunner(backend, offload_params_to_cpu), use_mask(use_mask) {
+            : GGMLRunner(backend, offload_params_to_cpu), version(version), use_mask(use_mask) {
+            flux_params.version             = version;
             flux_params.flash_attn          = flash_attn;
             flux_params.guidance_embed      = false;
             flux_params.depth               = 0;
             flux_params.depth_single_blocks = 0;
             if (version == VERSION_FLUX_FILL) {
                 flux_params.in_channels = 384;
+            } else if (version == VERSION_FLUX_CONTROLS) {
+                flux_params.in_channels = 128;
+            } else if (version == VERSION_FLEX_2) {
+                flux_params.in_channels = 196;
             }
             for (auto pair : tensor_types) {
                 std::string tensor_name = pair.first;
@@ -945,7 +925,7 @@ namespace Flux {
             flux.init(params_ctx, tensor_types, prefix);
         }
 
-        std::string get_desc() {
+        std::string get_desc() override {
             return "flux";
         }
 
@@ -965,18 +945,18 @@ namespace Flux {
             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;
+            struct ggml_tensor* mod_index_arange = nullptr;
 
             x       = to_backend(x);
             context = to_backend(context);
-            if (c_concat != NULL) {
+            if (c_concat != nullptr) {
                 c_concat = to_backend(c_concat);
             }
             if (flux_params.is_chroma) {
                 guidance = ggml_set_f32(guidance, 0);
 
                 if (!use_mask) {
-                    y = NULL;
+                    y = nullptr;
                 }
 
                 // ggml_arange is not working on some backends, precompute it
@@ -1008,7 +988,7 @@ namespace Flux {
             auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, flux_params.axes_dim_sum / 2, pos_len);
             // pe->data = pe_vec.data();
             // print_ggml_tensor(pe);
-            // pe->data = NULL;
+            // pe->data = nullptr;
             set_backend_tensor_data(pe, pe_vec.data());
 
             struct ggml_tensor* out = flux.forward(compute_ctx,
@@ -1038,8 +1018,8 @@ namespace Flux {
                      struct ggml_tensor* guidance,
                      std::vector<ggml_tensor*> ref_latents = {},
                      bool increase_ref_index               = false,
-                     struct ggml_tensor** output           = NULL,
-                     struct ggml_context* output_ctx       = NULL,
+                     struct ggml_tensor** output           = nullptr,
+                     struct ggml_context* output_ctx       = nullptr,
                      std::vector<int> skip_layers          = std::vector<int>()) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
@@ -1056,11 +1036,11 @@ namespace Flux {
         void test() {
             struct ggml_init_params params;
             params.mem_size   = static_cast<size_t>(20 * 1024 * 1024);  // 20 MB
-            params.mem_buffer = NULL;
+            params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
             struct ggml_context* work_ctx = ggml_init(params);
-            GGML_ASSERT(work_ctx != NULL);
+            GGML_ASSERT(work_ctx != nullptr);
 
             {
                 // cpu f16:
@@ -1084,10 +1064,10 @@ namespace Flux {
                 ggml_set_f32(y, 0.01f);
                 // print_ggml_tensor(y);
 
-                struct ggml_tensor* out = NULL;
+                struct ggml_tensor* out = nullptr;
 
                 int t0 = ggml_time_ms();
-                compute(8, x, timesteps, context, NULL, y, guidance, {}, false, &out, work_ctx);
+                compute(8, x, timesteps, context, nullptr, y, guidance, {}, false, &out, work_ctx);
                 int t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);
@@ -1099,7 +1079,7 @@ namespace Flux {
             // ggml_backend_t backend    = ggml_backend_cuda_init(0);
             ggml_backend_t backend           = ggml_backend_cpu_init();
             ggml_type model_data_type        = GGML_TYPE_Q8_0;
-            std::shared_ptr<FluxRunner> flux = std::shared_ptr<FluxRunner>(new FluxRunner(backend, false));
+            std::shared_ptr<FluxRunner> flux = std::make_shared<FluxRunner>(backend, false);
             {
                 LOG_INFO("loading from '%s'", file_path.c_str());
 

format-code.sh

@@ -1,5 +1,8 @@
 for f in *.cpp *.h *.hpp examples/cli/*.cpp examples/cli/*.h; do
   [[ "$f" == vocab* ]] && continue
   echo "formatting '$f'"
+  # if [ "$f" != "stable-diffusion.h" ]; then
+  #   clang-tidy -fix -p build_linux/ "$f"
+  # fi
   clang-format -style=file -i "$f"
 done

lora.hpp

@@ -1,6 +1,7 @@
 #ifndef __LORA_HPP__
 #define __LORA_HPP__
 
+#include <mutex>
 #include "ggml_extend.hpp"
 
 #define LORA_GRAPH_BASE_SIZE 10240
@@ -99,7 +100,7 @@ struct LoraModel : public GGMLRunner {
     bool load_failed                = false;
     bool applied                    = false;
     std::vector<int> zero_index_vec = {0};
-    ggml_tensor* zero_index         = NULL;
+    ggml_tensor* zero_index         = nullptr;
     enum lora_t type                = REGULAR;
 
     LoraModel(ggml_backend_t backend,
@@ -111,11 +112,11 @@ struct LoraModel : public GGMLRunner {
         }
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "lora";
     }
 
-    bool load_from_file(bool filter_tensor = false) {
+    bool load_from_file(bool filter_tensor, int n_threads) {
         LOG_INFO("loading LoRA from '%s'", file_path.c_str());
 
         if (load_failed) {
@@ -123,41 +124,53 @@ struct LoraModel : public GGMLRunner {
             return false;
         }
 
+        std::unordered_map<std::string, TensorStorage> tensors_to_create;
+        std::mutex lora_mutex;
         bool dry_run          = true;
         auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
-            const std::string& name = tensor_storage.name;
+            if (dry_run) {
+                const std::string& name = tensor_storage.name;
 
-            if (filter_tensor && !contains(name, "lora")) {
-                // LOG_INFO("skipping LoRA tesnor '%s'", name.c_str());
-                return true;
-            }
-            // LOG_INFO("lora_tensor %s", name.c_str());
-            for (int i = 0; i < LORA_TYPE_COUNT; i++) {
-                if (name.find(type_fingerprints[i]) != std::string::npos) {
-                    type = (lora_t)i;
-                    break;
+                if (filter_tensor && !contains(name, "lora")) {
+                    return true;
+                }
+
+                {
+                    std::lock_guard<std::mutex> lock(lora_mutex);
+                    for (int i = 0; i < LORA_TYPE_COUNT; i++) {
+                        if (name.find(type_fingerprints[i]) != std::string::npos) {
+                            type = (lora_t)i;
+                            break;
+                        }
+                    }
+                    tensors_to_create[name] = tensor_storage;
+                }
+            } else {
+                const std::string& name = tensor_storage.name;
+                auto iter               = lora_tensors.find(name);
+                if (iter != lora_tensors.end()) {
+                    *dst_tensor = iter->second;
                 }
             }
-
-            if (dry_run) {
-                struct ggml_tensor* real = ggml_new_tensor(params_ctx,
-                                                           tensor_storage.type,
-                                                           tensor_storage.n_dims,
-                                                           tensor_storage.ne);
-                lora_tensors[name]       = real;
-            } else {
-                auto real   = lora_tensors[name];
-                *dst_tensor = real;
-            }
-
             return true;
         };
 
-        model_loader.load_tensors(on_new_tensor_cb);
+        model_loader.load_tensors(on_new_tensor_cb, n_threads);
+
+        for (const auto& pair : tensors_to_create) {
+            const auto& name         = pair.first;
+            const auto& ts           = pair.second;
+            struct ggml_tensor* real = ggml_new_tensor(params_ctx,
+                                                       ts.type,
+                                                       ts.n_dims,
+                                                       ts.ne);
+            lora_tensors[name]       = real;
+        }
+
         alloc_params_buffer();
-        // exit(0);
+
         dry_run = false;
-        model_loader.load_tensors(on_new_tensor_cb);
+        model_loader.load_tensors(on_new_tensor_cb, n_threads);
 
         LOG_DEBUG("lora type: \"%s\"/\"%s\"", lora_downs[type].c_str(), lora_ups[type].c_str());
 
@@ -274,7 +287,7 @@ struct LoraModel : public GGMLRunner {
                 if (is_qkvm_split) {
                     key = key.substr(sizeof("SPLIT_L|") - 1);
                 }
-                struct ggml_tensor* updown = NULL;
+                struct ggml_tensor* updown = nullptr;
                 float scale_value          = 1.0f;
                 std::string full_key       = lora_pre[type] + key;
                 if (is_bias) {
@@ -301,13 +314,13 @@ struct LoraModel : public GGMLRunner {
                     }
                     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_1_mid  = nullptr;  // tau for tucker decomposition
+                    ggml_tensor* hada_1_up   = nullptr;
+                    ggml_tensor* hada_1_down = nullptr;
 
-                    ggml_tensor* hada_2_mid  = NULL;  // tau for tucker decomposition
-                    ggml_tensor* hada_2_up   = NULL;
-                    ggml_tensor* hada_2_down = NULL;
+                    ggml_tensor* hada_2_mid  = nullptr;  // tau for tucker decomposition
+                    ggml_tensor* hada_2_up   = nullptr;
+                    ggml_tensor* hada_2_down = nullptr;
 
                     std::string hada_1_mid_name  = "";
                     std::string hada_1_down_name = "";
@@ -355,7 +368,7 @@ struct LoraModel : public GGMLRunner {
                     applied_lora_tensors.insert(hada_2_up_name);
 
                     applied_lora_tensors.insert(alpha_name);
-                    if (hada_1_up == NULL || hada_1_down == NULL || hada_2_up == NULL || hada_2_down == NULL) {
+                    if (hada_1_up == nullptr || hada_1_down == nullptr || hada_2_up == nullptr || hada_2_down == nullptr) {
                         continue;
                     }
 
@@ -381,8 +394,8 @@ struct LoraModel : public GGMLRunner {
 
                     std::string alpha_name = full_key + ".alpha";
 
-                    ggml_tensor* lokr_w1 = NULL;
-                    ggml_tensor* lokr_w2 = NULL;
+                    ggml_tensor* lokr_w1 = nullptr;
+                    ggml_tensor* lokr_w2 = nullptr;
 
                     std::string lokr_w1_name = "";
                     std::string lokr_w2_name = "";
@@ -394,8 +407,8 @@ struct LoraModel : public GGMLRunner {
                         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;
+                        ggml_tensor* down     = nullptr;
+                        ggml_tensor* up       = nullptr;
                         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()) {
@@ -419,8 +432,8 @@ struct LoraModel : public GGMLRunner {
                         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;
+                        ggml_tensor* down     = nullptr;
+                        ggml_tensor* up       = nullptr;
                         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()) {
@@ -447,9 +460,9 @@ struct LoraModel : public GGMLRunner {
 
                 } else {
                     // LoRA mode
-                    ggml_tensor* lora_mid  = NULL;  // tau for tucker decomposition
-                    ggml_tensor* lora_up   = NULL;
-                    ggml_tensor* lora_down = NULL;
+                    ggml_tensor* lora_mid  = nullptr;  // tau for tucker decomposition
+                    ggml_tensor* lora_up   = nullptr;
+                    ggml_tensor* lora_down = nullptr;
 
                     std::string alpha_name         = "";
                     std::string scale_name         = "";
@@ -484,12 +497,12 @@ struct LoraModel : public GGMLRunner {
                             auto split_k_alpha_name = full_key + "k" + suffix + ".alpha";
                             auto split_v_alpha_name = full_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;
+                            ggml_tensor* lora_q_down = nullptr;
+                            ggml_tensor* lora_q_up   = nullptr;
+                            ggml_tensor* lora_k_down = nullptr;
+                            ggml_tensor* lora_k_up   = nullptr;
+                            ggml_tensor* lora_v_down = nullptr;
+                            ggml_tensor* lora_v_up   = nullptr;
 
                             lora_q_down = to_f32(compute_ctx, lora_tensors[split_q_d_name]);
 
@@ -620,15 +633,15 @@ struct LoraModel : public GGMLRunner {
                             auto split_v_alpha_name = full_key + "attn.to_v" + ".alpha";
                             auto split_m_alpha_name = full_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_q_down = nullptr;
+                            ggml_tensor* lora_q_up   = nullptr;
+                            ggml_tensor* lora_k_down = nullptr;
+                            ggml_tensor* lora_k_up   = nullptr;
+                            ggml_tensor* lora_v_down = nullptr;
+                            ggml_tensor* lora_v_up   = nullptr;
 
-                            ggml_tensor* lora_m_down = NULL;
-                            ggml_tensor* lora_m_up   = NULL;
+                            ggml_tensor* lora_m_down = nullptr;
+                            ggml_tensor* lora_m_up   = nullptr;
 
                             lora_q_up = to_f32(compute_ctx, lora_tensors[split_q_u_name]);
 
@@ -796,7 +809,7 @@ struct LoraModel : public GGMLRunner {
                         }
                     }
 
-                    if (lora_up == NULL || lora_down == NULL) {
+                    if (lora_up == nullptr || lora_down == nullptr) {
                         continue;
                     }
                     // calc_scale

ltxv.hpp

@@ -13,10 +13,10 @@ namespace LTXV {
     public:
         CausalConv3d(int64_t in_channels,
                      int64_t out_channels,
-                     int kernel_size        = 3,
-                     std::tuple<int> stride = {1, 1, 1},
-                     int dilation           = 1,
-                     bool bias              = true) {
+                     int kernel_size                  = 3,
+                     std::tuple<int, int, int> stride = {1, 1, 1},
+                     int dilation                     = 1,
+                     bool bias                        = true) {
             time_kernel_size = kernel_size / 2;
             blocks["conv"]   = std::shared_ptr<GGMLBlock>(new Conv3d(in_channels,
                                                                      out_channels,

mmdit.hpp

@@ -1,6 +1,8 @@
 #ifndef __MMDIT_HPP__
 #define __MMDIT_HPP__
 
+#include <memory>
+
 #include "ggml_extend.hpp"
 #include "model.h"
 
@@ -147,14 +149,16 @@ public:
     int64_t num_heads;
     bool pre_only;
     std::string qk_norm;
+    bool flash_attn;
 
 public:
     SelfAttention(int64_t dim,
                   int64_t num_heads   = 8,
                   std::string qk_norm = "",
                   bool qkv_bias       = false,
-                  bool pre_only       = false)
-        : num_heads(num_heads), pre_only(pre_only), qk_norm(qk_norm) {
+                  bool pre_only       = false,
+                  bool flash_attn     = false)
+        : num_heads(num_heads), pre_only(pre_only), qk_norm(qk_norm), flash_attn(flash_attn) {
         int64_t d_head = dim / num_heads;
         blocks["qkv"]  = std::shared_ptr<GGMLBlock>(new Linear(dim, dim * 3, qkv_bias));
         if (!pre_only) {
@@ -206,8 +210,8 @@ public:
                                 ggml_backend_t backend,
                                 struct ggml_tensor* x) {
         auto qkv = pre_attention(ctx, x);
-        x        = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads);  // [N, n_token, dim]
-        x        = post_attention(ctx, x);                                                  // [N, n_token, dim]
+        x        = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, false, true);  // [N, n_token, dim]
+        x        = post_attention(ctx, x);                                                                               // [N, n_token, dim]
         return x;
     }
 };
@@ -232,6 +236,7 @@ public:
     int64_t num_heads;
     bool pre_only;
     bool self_attn;
+    bool flash_attn;
 
 public:
     DismantledBlock(int64_t hidden_size,
@@ -240,16 +245,17 @@ public:
                     std::string qk_norm = "",
                     bool qkv_bias       = false,
                     bool pre_only       = false,
-                    bool self_attn      = false)
+                    bool self_attn      = false,
+                    bool flash_attn     = false)
         : num_heads(num_heads), pre_only(pre_only), self_attn(self_attn) {
         // rmsnorm is always Flase
         // scale_mod_only is always Flase
         // swiglu is always Flase
         blocks["norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size, 1e-06f, false));
-        blocks["attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, pre_only));
+        blocks["attn"]  = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, pre_only, flash_attn));
 
         if (self_attn) {
-            blocks["attn2"] = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, false));
+            blocks["attn2"] = std::shared_ptr<GGMLBlock>(new SelfAttention(hidden_size, num_heads, qk_norm, qkv_bias, false, flash_attn));
         }
 
         if (!pre_only) {
@@ -343,7 +349,7 @@ public:
             auto attn_in = modulate(ctx, norm1->forward(ctx, x), shift_msa, scale_msa);
             auto qkv     = attn->pre_attention(ctx, attn_in);
 
-            return {qkv, {NULL, NULL, NULL, NULL, NULL}};
+            return {qkv, {nullptr, nullptr, nullptr, nullptr, nullptr}};
         }
     }
 
@@ -435,8 +441,8 @@ public:
             auto qkv2          = std::get<1>(qkv_intermediates);
             auto intermediates = std::get<2>(qkv_intermediates);
 
-            auto attn_out  = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads);     // [N, n_token, dim]
-            auto attn2_out = ggml_nn_attention_ext(ctx, backend, qkv2[0], qkv2[1], qkv2[2], num_heads);  // [N, n_token, dim]
+            auto attn_out  = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, false, flash_attn);     // [N, n_token, dim]
+            auto attn2_out = ggml_nn_attention_ext(ctx, backend, qkv2[0], qkv2[1], qkv2[2], num_heads, nullptr, false, false, flash_attn);  // [N, n_token, dim]
             x              = post_attention_x(ctx,
                                               attn_out,
                                               attn2_out,
@@ -452,7 +458,7 @@ public:
             auto qkv               = qkv_intermediates.first;
             auto intermediates     = qkv_intermediates.second;
 
-            auto attn_out = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads);  // [N, n_token, dim]
+            auto attn_out = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, false, flash_attn);  // [N, n_token, dim]
             x             = post_attention(ctx,
                                            attn_out,
                                            intermediates[0],
@@ -468,6 +474,7 @@ public:
 __STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*>
 block_mixing(struct ggml_context* ctx,
              ggml_backend_t backend,
+             bool flash_attn,
              struct ggml_tensor* context,
              struct ggml_tensor* x,
              struct ggml_tensor* c,
@@ -497,8 +504,8 @@ block_mixing(struct ggml_context* ctx,
         qkv.push_back(ggml_concat(ctx, context_qkv[i], x_qkv[i], 1));
     }
 
-    auto attn         = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], x_block->num_heads);  // [N, n_context + n_token, hidden_size]
-    attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));                              // [n_context + n_token, N, hidden_size]
+    auto attn         = ggml_nn_attention_ext(ctx, backend, qkv[0], qkv[1], qkv[2], x_block->num_heads, nullptr, false, false, flash_attn);  // [N, n_context + n_token, hidden_size]
+    attn              = ggml_cont(ctx, ggml_permute(ctx, attn, 0, 2, 1, 3));                                                                 // [n_context + n_token, N, hidden_size]
     auto context_attn = ggml_view_3d(ctx,
                                      attn,
                                      attn->ne[0],
@@ -527,7 +534,7 @@ block_mixing(struct ggml_context* ctx,
                                                 context_intermediates[3],
                                                 context_intermediates[4]);
     } else {
-        context = NULL;
+        context = nullptr;
     }
 
     if (x_block->self_attn) {
@@ -556,6 +563,8 @@ block_mixing(struct ggml_context* ctx,
 }
 
 struct JointBlock : public GGMLBlock {
+    bool flash_attn;
+
 public:
     JointBlock(int64_t hidden_size,
                int64_t num_heads,
@@ -563,9 +572,11 @@ public:
                std::string qk_norm = "",
                bool qkv_bias       = false,
                bool pre_only       = false,
-               bool self_attn_x    = false) {
-        blocks["context_block"] = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, pre_only));
-        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x));
+               bool self_attn_x    = false,
+               bool flash_attn     = false)
+        : flash_attn(flash_attn) {
+        blocks["context_block"] = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, pre_only, false, flash_attn));
+        blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x, flash_attn));
     }
 
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
@@ -576,7 +587,7 @@ public:
         auto context_block = std::dynamic_pointer_cast<DismantledBlock>(blocks["context_block"]);
         auto x_block       = std::dynamic_pointer_cast<DismantledBlock>(blocks["x_block"]);
 
-        return block_mixing(ctx, backend, context, x, c, context_block, x_block);
+        return block_mixing(ctx, backend, flash_attn, context, x, c, context_block, x_block);
     }
 };
 
@@ -634,14 +645,16 @@ protected:
     int64_t context_embedder_out_dim = 1536;
     int64_t hidden_size;
     std::string qk_norm;
+    bool flash_attn = false;
 
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, std::string prefix = "") override {
         enum ggml_type wtype = GGML_TYPE_F32;
         params["pos_embed"]  = ggml_new_tensor_3d(ctx, wtype, hidden_size, num_patchs, 1);
     }
 
 public:
-    MMDiT(const String2GGMLType& tensor_types = {}) {
+    MMDiT(bool flash_attn = false, const String2GGMLType& tensor_types = {})
+        : flash_attn(flash_attn) {
         // input_size is always None
         // learn_sigma is always False
         // register_length is alwalys 0
@@ -709,7 +722,8 @@ public:
                                                                                                     qk_norm,
                                                                                                     true,
                                                                                                     i == depth - 1,
-                                                                                                    i <= d_self));
+                                                                                                    i <= d_self,
+                                                                                                    flash_attn));
         }
 
         blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new FinalLayer(hidden_size, patch_size, out_channels));
@@ -811,8 +825,8 @@ public:
                                 ggml_backend_t backend,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* t,
-                                struct ggml_tensor* y        = NULL,
-                                struct ggml_tensor* context  = NULL,
+                                struct ggml_tensor* y        = nullptr,
+                                struct ggml_tensor* context  = nullptr,
                                 std::vector<int> skip_layers = std::vector<int>()) {
         // Forward pass of DiT.
         // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
@@ -831,14 +845,14 @@ public:
         x                = ggml_add(ctx, patch_embed, pos_embed);  // [N, H*W, hidden_size]
 
         auto c = t_embedder->forward(ctx, t);  // [N, hidden_size]
-        if (y != NULL && adm_in_channels != -1) {
+        if (y != nullptr && adm_in_channels != -1) {
             auto y_embedder = std::dynamic_pointer_cast<VectorEmbedder>(blocks["y_embedder"]);
 
             y = y_embedder->forward(ctx, y);  // [N, hidden_size]
             c = ggml_add(ctx, c, y);
         }
 
-        if (context != NULL) {
+        if (context != nullptr) {
             auto context_embedder = std::dynamic_pointer_cast<Linear>(blocks["context_embedder"]);
 
             context = context_embedder->forward(ctx, context);  // [N, L, D] aka [N, L, 1536]
@@ -856,13 +870,14 @@ struct MMDiTRunner : public GGMLRunner {
 
     MMDiTRunner(ggml_backend_t backend,
                 bool offload_params_to_cpu,
+                bool flash_attn,
                 const String2GGMLType& tensor_types = {},
                 const std::string prefix            = "")
-        : GGMLRunner(backend, offload_params_to_cpu), mmdit(tensor_types) {
+        : GGMLRunner(backend, offload_params_to_cpu), mmdit(flash_attn, tensor_types) {
         mmdit.init(params_ctx, tensor_types, prefix);
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "mmdit";
     }
 
@@ -900,8 +915,8 @@ struct MMDiTRunner : public GGMLRunner {
                  struct ggml_tensor* timesteps,
                  struct ggml_tensor* context,
                  struct ggml_tensor* y,
-                 struct ggml_tensor** output     = NULL,
-                 struct ggml_context* output_ctx = NULL,
+                 struct ggml_tensor** output     = nullptr,
+                 struct ggml_context* output_ctx = nullptr,
                  std::vector<int> skip_layers    = std::vector<int>()) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
@@ -917,11 +932,11 @@ struct MMDiTRunner : public GGMLRunner {
     void test() {
         struct ggml_init_params params;
         params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-        params.mem_buffer = NULL;
+        params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
         struct ggml_context* work_ctx = ggml_init(params);
-        GGML_ASSERT(work_ctx != NULL);
+        GGML_ASSERT(work_ctx != nullptr);
 
         {
             // cpu f16: pass
@@ -942,7 +957,7 @@ struct MMDiTRunner : public GGMLRunner {
             ggml_set_f32(y, 0.01f);
             // print_ggml_tensor(y);
 
-            struct ggml_tensor* out = NULL;
+            struct ggml_tensor* out = nullptr;
 
             int t0 = ggml_time_ms();
             compute(8, x, timesteps, context, y, &out, work_ctx);
@@ -957,7 +972,7 @@ struct MMDiTRunner : public GGMLRunner {
         // ggml_backend_t backend    = ggml_backend_cuda_init(0);
         ggml_backend_t backend             = ggml_backend_cpu_init();
         ggml_type model_data_type          = GGML_TYPE_F16;
-        std::shared_ptr<MMDiTRunner> mmdit = std::shared_ptr<MMDiTRunner>(new MMDiTRunner(backend, false));
+        std::shared_ptr<MMDiTRunner> mmdit = std::make_shared<MMDiTRunner>(backend, false, false);
         {
             LOG_INFO("loading from '%s'", file_path.c_str());
 

model.cpp

@@ -1,8 +1,14 @@
-#include <stdarg.h>
+#include <algorithm>
+#include <atomic>
+#include <chrono>
+#include <cstdarg>
 #include <fstream>
+#include <functional>
+#include <mutex>
 #include <regex>
 #include <set>
 #include <string>
+#include <thread>
 #include <unordered_map>
 #include <vector>
 
@@ -11,6 +17,7 @@
 #include "stable-diffusion.h"
 #include "util.h"
 #include "vocab.hpp"
+#include "vocab_qwen.hpp"
 #include "vocab_umt5.hpp"
 
 #include "ggml-alloc.h"
@@ -104,10 +111,12 @@ const char* unused_tensors[] = {
     "embedding_manager",
     "denoiser.sigmas",
     "text_encoders.t5xxl.transformer.encoder.embed_tokens.weight",  // only used during training
+    "text_encoders.qwen2vl.output.weight",
+    "text_encoders.qwen2vl.lm_head.",
 };
 
 bool is_unused_tensor(std::string name) {
-    for (int i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
+    for (size_t i = 0; i < sizeof(unused_tensors) / sizeof(const char*); i++) {
         if (starts_with(name, unused_tensors[i])) {
             return true;
         }
@@ -187,6 +196,39 @@ std::unordered_map<std::string, std::string> pmid_v2_name_map = {
      "pmid.qformer_perceiver.token_proj.fc2.weight"},
 };
 
+std::unordered_map<std::string, std::string> qwenvl_name_map{
+    {"token_embd.", "model.embed_tokens."},
+    {"blk.", "model.layers."},
+    {"attn_q.", "self_attn.q_proj."},
+    {"attn_k.", "self_attn.k_proj."},
+    {"attn_v.", "self_attn.v_proj."},
+    {"attn_output.", "self_attn.o_proj."},
+    {"attn_norm.", "input_layernorm."},
+    {"ffn_down.", "mlp.down_proj."},
+    {"ffn_gate.", "mlp.gate_proj."},
+    {"ffn_up.", "mlp.up_proj."},
+    {"ffn_norm.", "post_attention_layernorm."},
+    {"output_norm.", "model.norm."},
+};
+
+std::unordered_map<std::string, std::string> qwenvl_vision_name_map{
+    {"mm.", "merger.mlp."},
+    {"v.post_ln.", "merger.ln_q."},
+    {"v.patch_embd.weight", "patch_embed.proj.0.weight"},
+    {"patch_embed.proj.0.weight.1", "patch_embed.proj.1.weight"},
+    {"v.patch_embd.weight.1", "patch_embed.proj.1.weight"},
+    {"v.blk.", "blocks."},
+    {"attn_q.", "attn.q_proj."},
+    {"attn_k.", "attn.k_proj."},
+    {"attn_v.", "attn.v_proj."},
+    {"attn_out.", "attn.proj."},
+    {"ffn_down.", "mlp.down_proj."},
+    {"ffn_gate.", "mlp.gate_proj."},
+    {"ffn_up.", "mlp.up_proj."},
+    {"ln1.", "norm1."},
+    {"ln2.", "norm2."},
+};
+
 std::string convert_cond_model_name(const std::string& name) {
     std::string new_name = name;
     std::string prefix;
@@ -244,6 +286,22 @@ std::string convert_cond_model_name(const std::string& name) {
         if (pos != std::string::npos) {
             new_name.replace(pos, 11, "layer.0.SelfAttention.relative_attention_bias.");
         }
+    } else if (contains(name, "qwen2vl")) {
+        if (contains(name, "qwen2vl.visual")) {
+            for (auto kv : qwenvl_vision_name_map) {
+                size_t pos = new_name.find(kv.first);
+                if (pos != std::string::npos) {
+                    new_name.replace(pos, kv.first.size(), kv.second);
+                }
+            }
+        } else {
+            for (auto kv : qwenvl_name_map) {
+                size_t pos = new_name.find(kv.first);
+                if (pos != std::string::npos) {
+                    new_name.replace(pos, kv.first.size(), kv.second);
+                }
+            }
+        }
     } else if (name == "text_encoders.t5xxl.transformer.token_embd.weight") {
         new_name = "text_encoders.t5xxl.transformer.shared.weight";
     }
@@ -574,7 +632,11 @@ std::string convert_tensor_name(std::string name) {
     //     name.replace(pos, strlen("lora_B"), "lora_down");
     // }
     std::string new_name = name;
-    if (starts_with(name, "cond_stage_model.") || starts_with(name, "conditioner.embedders.") || starts_with(name, "text_encoders.") || ends_with(name, ".vision_model.visual_projection.weight")) {
+    if (starts_with(name, "cond_stage_model.") ||
+        starts_with(name, "conditioner.embedders.") ||
+        starts_with(name, "text_encoders.") ||
+        ends_with(name, ".vision_model.visual_projection.weight") ||
+        starts_with(name, "qwen2vl")) {
         new_name = convert_cond_model_name(name);
     } else if (starts_with(name, "first_stage_model.decoder")) {
         new_name = convert_vae_decoder_name(name);
@@ -693,6 +755,7 @@ void preprocess_tensor(TensorStorage tensor_storage,
 
     // convert unet transformer linear to conv2d 1x1
     if (starts_with(new_name, "model.diffusion_model.") &&
+        !starts_with(new_name, "model.diffusion_model.proj_out.") &&
         (ends_with(new_name, "proj_in.weight") || ends_with(new_name, "proj_out.weight"))) {
         tensor_storage.unsqueeze();
     }
@@ -806,7 +869,6 @@ uint16_t f8_e5m2_to_f16(uint8_t fp8) {
     }
 
     if (exponent == 0) {  // subnormal numbers
-        fp16_exponent = 0;
         fp16_mantissa = (mantissa << 8);
         return fp16_sign | fp16_mantissa;
     }
@@ -885,7 +947,7 @@ void convert_tensor(void* src,
             ggml_fp16_to_fp32_row((ggml_fp16_t*)src, (float*)dst, n);
         } else {
             auto qtype = ggml_get_type_traits(src_type);
-            if (qtype->to_float == NULL) {
+            if (qtype->to_float == nullptr) {
                 throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available",
                                                 ggml_type_name(src_type)));
             }
@@ -895,7 +957,7 @@ void convert_tensor(void* src,
         // src_type == GGML_TYPE_F16 => dst_type is quantized
         // src_type is quantized => dst_type == GGML_TYPE_F16 or dst_type is quantized
         auto qtype = ggml_get_type_traits(src_type);
-        if (qtype->to_float == NULL) {
+        if (qtype->to_float == nullptr) {
             throw std::runtime_error(format("type %s unsupported for integer quantization: no dequantization available",
                                             ggml_type_name(src_type)));
         }
@@ -957,7 +1019,7 @@ std::map<char, int> unicode_to_byte() {
 
 bool is_zip_file(const std::string& file_path) {
     struct zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
-    if (zip == NULL) {
+    if (zip == nullptr) {
         return false;
     }
     zip_close(zip);
@@ -1053,8 +1115,8 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
     file_paths_.push_back(file_path);
     size_t file_index = file_paths_.size() - 1;
 
-    gguf_context* ctx_gguf_ = NULL;
-    ggml_context* ctx_meta_ = NULL;
+    gguf_context* ctx_gguf_ = nullptr;
+    ggml_context* ctx_meta_ = nullptr;
 
     ctx_gguf_ = gguf_init_from_file(file_path.c_str(), {true, &ctx_meta_});
     if (!ctx_gguf_) {
@@ -1663,7 +1725,7 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
     size_t file_index = file_paths_.size() - 1;
 
     struct zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
-    if (zip == NULL) {
+    if (zip == nullptr) {
         LOG_ERROR("failed to open '%s'", file_path.c_str());
         return false;
     }
@@ -1676,7 +1738,7 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
             if (pos != std::string::npos) {
                 std::string dir = name.substr(0, pos);
                 printf("ZIP %d, name = %s, dir = %s \n", i, name.c_str(), dir.c_str());
-                void* pkl_data = NULL;
+                void* pkl_data = nullptr;
                 size_t pkl_size;
                 zip_entry_read(zip, &pkl_data, &pkl_size);
 
@@ -1726,6 +1788,9 @@ 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.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
+                return VERSION_QWEN_IMAGE;
+            }
             if (tensor_storage.name.find("model.diffusion_model.blocks.0.cross_attn.norm_k.weight") != std::string::npos) {
                 is_wan = true;
             }
@@ -1797,10 +1862,15 @@ SDVersion ModelLoader::get_sd_version() {
     }
 
     if (is_flux) {
-        is_inpaint = input_block_weight.ne[0] == 384;
-        if (is_inpaint) {
+        if (input_block_weight.ne[0] == 384) {
             return VERSION_FLUX_FILL;
         }
+        if (input_block_weight.ne[0] == 128) {
+            return VERSION_FLUX_CONTROLS;
+        }
+        if (input_block_weight.ne[0] == 196) {
+            return VERSION_FLEX_2;
+        }
         return VERSION_FLUX;
     }
 
@@ -1821,24 +1891,25 @@ SDVersion ModelLoader::get_sd_version() {
     return VERSION_COUNT;
 }
 
-ggml_type ModelLoader::get_sd_wtype() {
+std::map<ggml_type, uint32_t> ModelLoader::get_wtype_stat() {
+    std::map<ggml_type, uint32_t> wtype_stat;
     for (auto& tensor_storage : tensor_storages) {
         if (is_unused_tensor(tensor_storage.name)) {
             continue;
         }
 
-        if (ggml_is_quantized(tensor_storage.type)) {
-            return tensor_storage.type;
-        }
-
-        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
-            return tensor_storage.type;
+        auto iter = wtype_stat.find(tensor_storage.type);
+        if (iter != wtype_stat.end()) {
+            iter->second++;
+        } else {
+            wtype_stat[tensor_storage.type] = 1;
         }
     }
-    return GGML_TYPE_COUNT;
+    return wtype_stat;
 }
 
-ggml_type ModelLoader::get_conditioner_wtype() {
+std::map<ggml_type, uint32_t> ModelLoader::get_conditioner_wtype_stat() {
+    std::map<ggml_type, uint32_t> wtype_stat;
     for (auto& tensor_storage : tensor_storages) {
         if (is_unused_tensor(tensor_storage.name)) {
             continue;
@@ -1851,18 +1922,18 @@ ggml_type ModelLoader::get_conditioner_wtype() {
             continue;
         }
 
-        if (ggml_is_quantized(tensor_storage.type)) {
-            return tensor_storage.type;
-        }
-
-        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
-            return tensor_storage.type;
+        auto iter = wtype_stat.find(tensor_storage.type);
+        if (iter != wtype_stat.end()) {
+            iter->second++;
+        } else {
+            wtype_stat[tensor_storage.type] = 1;
         }
     }
-    return GGML_TYPE_COUNT;
+    return wtype_stat;
 }
 
-ggml_type ModelLoader::get_diffusion_model_wtype() {
+std::map<ggml_type, uint32_t> ModelLoader::get_diffusion_model_wtype_stat() {
+    std::map<ggml_type, uint32_t> wtype_stat;
     for (auto& tensor_storage : tensor_storages) {
         if (is_unused_tensor(tensor_storage.name)) {
             continue;
@@ -1872,18 +1943,18 @@ ggml_type ModelLoader::get_diffusion_model_wtype() {
             continue;
         }
 
-        if (ggml_is_quantized(tensor_storage.type)) {
-            return tensor_storage.type;
-        }
-
-        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
-            return tensor_storage.type;
+        auto iter = wtype_stat.find(tensor_storage.type);
+        if (iter != wtype_stat.end()) {
+            iter->second++;
+        } else {
+            wtype_stat[tensor_storage.type] = 1;
         }
     }
-    return GGML_TYPE_COUNT;
+    return wtype_stat;
 }
 
-ggml_type ModelLoader::get_vae_wtype() {
+std::map<ggml_type, uint32_t> ModelLoader::get_vae_wtype_stat() {
+    std::map<ggml_type, uint32_t> wtype_stat;
     for (auto& tensor_storage : tensor_storages) {
         if (is_unused_tensor(tensor_storage.name)) {
             continue;
@@ -1894,15 +1965,14 @@ ggml_type ModelLoader::get_vae_wtype() {
             continue;
         }
 
-        if (ggml_is_quantized(tensor_storage.type)) {
-            return tensor_storage.type;
-        }
-
-        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
-            return tensor_storage.type;
+        auto iter = wtype_stat.find(tensor_storage.type);
+        if (iter != wtype_stat.end()) {
+            iter->second++;
+        } else {
+            wtype_stat[tensor_storage.type] = 1;
         }
     }
-    return GGML_TYPE_COUNT;
+    return wtype_stat;
 }
 
 void ModelLoader::set_wtype_override(ggml_type wtype, std::string prefix) {
@@ -1934,6 +2004,11 @@ std::string ModelLoader::load_merges() {
     return merges_utf8_str;
 }
 
+std::string ModelLoader::load_qwen2_merges() {
+    std::string merges_utf8_str(reinterpret_cast<const char*>(qwen2_merges_utf8_c_str), sizeof(qwen2_merges_utf8_c_str));
+    return merges_utf8_str;
+}
+
 std::string ModelLoader::load_t5_tokenizer_json() {
     std::string json_str(reinterpret_cast<const char*>(t5_tokenizer_json_str), sizeof(t5_tokenizer_json_str));
     return json_str;
@@ -1944,292 +2019,357 @@ std::string ModelLoader::load_umt5_tokenizer_json() {
     return json_str;
 }
 
-std::vector<TensorStorage> remove_duplicates(const std::vector<TensorStorage>& vec) {
-    std::vector<TensorStorage> res;
-    std::unordered_map<std::string, size_t> name_to_index_map;
+bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p) {
+    int64_t process_time_ms = 0;
+    std::atomic<int64_t> read_time_ms(0);
+    std::atomic<int64_t> memcpy_time_ms(0);
+    std::atomic<int64_t> copy_to_backend_time_ms(0);
+    std::atomic<int64_t> convert_time_ms(0);
 
-    for (size_t i = 0; i < vec.size(); ++i) {
-        const std::string& current_name = vec[i].name;
-        auto it                         = name_to_index_map.find(current_name);
+    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : get_num_physical_cores();
+    LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
 
-        if (it != name_to_index_map.end()) {
-            res[it->second] = vec[i];
-        } else {
-            name_to_index_map[current_name] = i;
-            res.push_back(vec[i]);
-        }
-    }
-
-    // vec.resize(name_to_index_map.size());
-
-    return res;
-}
-
-bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb) {
-    int64_t process_time_ms         = 0;
-    int64_t read_time_ms            = 0;
-    int64_t memcpy_time_ms          = 0;
-    int64_t copy_to_backend_time_ms = 0;
-    int64_t convert_time_ms         = 0;
-
-    int64_t prev_time_ms = 0;
-    int64_t curr_time_ms = 0;
-    int64_t start_time   = ggml_time_ms();
-    prev_time_ms         = start_time;
+    int64_t start_time = ggml_time_ms();
     std::vector<TensorStorage> processed_tensor_storages;
-    for (auto& tensor_storage : tensor_storages) {
-        // LOG_DEBUG("%s", name.c_str());
 
-        if (is_unused_tensor(tensor_storage.name)) {
-            continue;
+    {
+        struct IndexedStorage {
+            size_t index;
+            TensorStorage ts;
+        };
+
+        std::mutex vec_mutex;
+        std::vector<IndexedStorage> all_results;
+
+        int n_threads = std::min(num_threads_to_use, (int)tensor_storages.size());
+        if (n_threads < 1) {
+            n_threads = 1;
+        }
+        std::vector<std::thread> workers;
+
+        for (int i = 0; i < n_threads; ++i) {
+            workers.emplace_back([&, thread_id = i]() {
+                std::vector<IndexedStorage> local_results;
+                std::vector<TensorStorage> temp_storages;
+
+                for (size_t j = thread_id; j < tensor_storages.size(); j += n_threads) {
+                    const auto& tensor_storage = tensor_storages[j];
+                    if (is_unused_tensor(tensor_storage.name)) {
+                        continue;
+                    }
+
+                    temp_storages.clear();
+                    preprocess_tensor(tensor_storage, temp_storages);
+
+                    for (const auto& ts : temp_storages) {
+                        local_results.push_back({j, ts});
+                    }
+                }
+
+                if (!local_results.empty()) {
+                    std::lock_guard<std::mutex> lock(vec_mutex);
+                    all_results.insert(all_results.end(),
+                                       local_results.begin(), local_results.end());
+                }
+            });
+        }
+        for (auto& w : workers) {
+            w.join();
         }
 
-        preprocess_tensor(tensor_storage, processed_tensor_storages);
-    }
-    std::vector<TensorStorage> dedup = remove_duplicates(processed_tensor_storages);
-    processed_tensor_storages        = dedup;
-    curr_time_ms                     = ggml_time_ms();
-    process_time_ms                  = curr_time_ms - prev_time_ms;
-    prev_time_ms                     = curr_time_ms;
+        std::vector<IndexedStorage> deduplicated;
+        deduplicated.reserve(all_results.size());
+        std::unordered_map<std::string, size_t> name_to_pos;
+        for (auto& entry : all_results) {
+            auto it = name_to_pos.find(entry.ts.name);
+            if (it == name_to_pos.end()) {
+                name_to_pos.emplace(entry.ts.name, deduplicated.size());
+                deduplicated.push_back(entry);
+            } else if (deduplicated[it->second].index < entry.index) {
+                deduplicated[it->second] = entry;
+            }
+        }
+
+        std::sort(deduplicated.begin(), deduplicated.end(), [](const IndexedStorage& a, const IndexedStorage& b) {
+            return a.index < b.index;
+        });
+
+        processed_tensor_storages.reserve(deduplicated.size());
+        for (auto& entry : deduplicated) {
+            processed_tensor_storages.push_back(entry.ts);
+        }
+    }
+
+    process_time_ms = ggml_time_ms() - start_time;
+
+    bool success                          = true;
+    size_t total_tensors_processed        = 0;
+    const size_t total_tensors_to_process = processed_tensor_storages.size();
+    const int64_t t_start                 = ggml_time_ms();
+    int last_n_threads                    = 1;
 
-    bool success = true;
     for (size_t file_index = 0; file_index < file_paths_.size(); file_index++) {
         std::string file_path = file_paths_[file_index];
         LOG_DEBUG("loading tensors from %s", file_path.c_str());
 
-        std::ifstream file(file_path, std::ios::binary);
-        if (!file.is_open()) {
-            LOG_ERROR("failed to open '%s'", file_path.c_str());
-            return false;
+        std::vector<const TensorStorage*> file_tensors;
+        for (const auto& ts : processed_tensor_storages) {
+            if (ts.file_index == file_index) {
+                file_tensors.push_back(&ts);
+            }
+        }
+        if (file_tensors.empty()) {
+            continue;
         }
 
         bool is_zip = false;
-        for (auto& tensor_storage : tensor_storages) {
-            if (tensor_storage.file_index != file_index) {
-                continue;
-            }
-            if (tensor_storage.index_in_zip >= 0) {
+        for (auto const& ts : file_tensors) {
+            if (ts->index_in_zip >= 0) {
                 is_zip = true;
                 break;
             }
         }
 
-        struct zip_t* zip = NULL;
-        if (is_zip) {
-            zip = zip_open(file_path.c_str(), 0, 'r');
-            if (zip == NULL) {
-                LOG_ERROR("failed to open zip '%s'", file_path.c_str());
-                return false;
-            }
+        int n_threads = is_zip ? 1 : std::min(num_threads_to_use, (int)file_tensors.size());
+        if (n_threads < 1) {
+            n_threads = 1;
+        }
+        last_n_threads = n_threads;
+
+        std::atomic<size_t> tensor_idx(0);
+        std::atomic<bool> failed(false);
+        std::vector<std::thread> workers;
+
+        for (int i = 0; i < n_threads; ++i) {
+            workers.emplace_back([&, file_path, is_zip]() {
+                std::ifstream file;
+                struct zip_t* zip = nullptr;
+                if (is_zip) {
+                    zip = zip_open(file_path.c_str(), 0, 'r');
+                    if (zip == nullptr) {
+                        LOG_ERROR("failed to open zip '%s'", file_path.c_str());
+                        failed = true;
+                        return;
+                    }
+                } else {
+                    file.open(file_path, std::ios::binary);
+                    if (!file.is_open()) {
+                        LOG_ERROR("failed to open '%s'", file_path.c_str());
+                        failed = true;
+                        return;
+                    }
+                }
+
+                std::vector<uint8_t> read_buffer;
+                std::vector<uint8_t> convert_buffer;
+
+                while (true) {
+                    int64_t t0, t1;
+                    size_t idx = tensor_idx.fetch_add(1);
+                    if (idx >= file_tensors.size() || failed) {
+                        break;
+                    }
+
+                    const TensorStorage& tensor_storage = *file_tensors[idx];
+                    ggml_tensor* dst_tensor             = nullptr;
+
+                    t0 = ggml_time_ms();
+
+                    if (!on_new_tensor_cb(tensor_storage, &dst_tensor)) {
+                        LOG_WARN("process tensor failed: '%s'", tensor_storage.name.c_str());
+                        failed = true;
+                        break;
+                    }
+
+                    if (dst_tensor == nullptr) {
+                        t1 = ggml_time_ms();
+                        read_time_ms.fetch_add(t1 - t0);
+                        continue;
+                    }
+
+                    size_t nbytes_to_read = tensor_storage.nbytes_to_read();
+
+                    auto read_data = [&](char* buf, size_t n) {
+                        if (zip != nullptr) {
+                            zip_entry_openbyindex(zip, tensor_storage.index_in_zip);
+                            size_t entry_size = zip_entry_size(zip);
+                            if (entry_size != n) {
+                                int64_t t_memcpy_start;
+                                read_buffer.resize(entry_size);
+                                zip_entry_noallocread(zip, (void*)read_buffer.data(), entry_size);
+                                t_memcpy_start = ggml_time_ms();
+                                memcpy((void*)buf, (void*)(read_buffer.data() + tensor_storage.offset), n);
+                                memcpy_time_ms.fetch_add(ggml_time_ms() - t_memcpy_start);
+                            } else {
+                                zip_entry_noallocread(zip, (void*)buf, n);
+                            }
+                            zip_entry_close(zip);
+                        } else {
+                            file.seekg(tensor_storage.offset);
+                            file.read(buf, n);
+                            if (!file) {
+                                LOG_ERROR("read tensor data failed: '%s'", file_path.c_str());
+                                failed = true;
+                            }
+                        }
+                    };
+
+                    if (dst_tensor->buffer == nullptr || ggml_backend_buffer_is_host(dst_tensor->buffer)) {
+                        if (tensor_storage.type == dst_tensor->type) {
+                            GGML_ASSERT(ggml_nbytes(dst_tensor) == tensor_storage.nbytes());
+                            if (tensor_storage.is_f64 || tensor_storage.is_i64) {
+                                read_buffer.resize(tensor_storage.nbytes_to_read());
+                                read_data((char*)read_buffer.data(), nbytes_to_read);
+                            } else {
+                                read_data((char*)dst_tensor->data, nbytes_to_read);
+                            }
+                            t1 = ggml_time_ms();
+                            read_time_ms.fetch_add(t1 - t0);
+
+                            t0 = ggml_time_ms();
+                            if (tensor_storage.is_bf16) {
+                                // inplace op
+                                bf16_to_f32_vec((uint16_t*)dst_tensor->data, (float*)dst_tensor->data, tensor_storage.nelements());
+                            } else if (tensor_storage.is_f8_e4m3) {
+                                // inplace op
+                                f8_e4m3_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
+                            } else if (tensor_storage.is_f8_e5m2) {
+                                // inplace op
+                                f8_e5m2_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
+                            } else 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());
+                            }
+                            t1 = ggml_time_ms();
+                            convert_time_ms.fetch_add(t1 - t0);
+                        } else {
+                            read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
+                            read_data((char*)read_buffer.data(), nbytes_to_read);
+                            t1 = ggml_time_ms();
+                            read_time_ms.fetch_add(t1 - t0);
+
+                            t0 = ggml_time_ms();
+                            if (tensor_storage.is_bf16) {
+                                // inplace op
+                                bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                            } else if (tensor_storage.is_f8_e4m3) {
+                                // inplace op
+                                f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                            } else if (tensor_storage.is_f8_e5m2) {
+                                // inplace op
+                                f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                            } 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]);
+                            t1 = ggml_time_ms();
+                            convert_time_ms.fetch_add(t1 - t0);
+                        }
+                    } else {
+                        read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
+                        read_data((char*)read_buffer.data(), nbytes_to_read);
+                        t1 = ggml_time_ms();
+                        read_time_ms.fetch_add(t1 - t0);
+
+                        t0 = ggml_time_ms();
+                        if (tensor_storage.is_bf16) {
+                            // inplace op
+                            bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                        } else if (tensor_storage.is_f8_e4m3) {
+                            // inplace op
+                            f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                        } else if (tensor_storage.is_f8_e5m2) {
+                            // inplace op
+                            f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
+                        } 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) {
+                            // copy to device memory
+                            t1 = ggml_time_ms();
+                            convert_time_ms.fetch_add(t1 - t0);
+                            t0 = ggml_time_ms();
+                            ggml_backend_tensor_set(dst_tensor, read_buffer.data(), 0, ggml_nbytes(dst_tensor));
+                            t1 = ggml_time_ms();
+                            copy_to_backend_time_ms.fetch_add(t1 - t0);
+                        } else {
+                            // convert first, then copy to device memory
+
+                            convert_buffer.resize(ggml_nbytes(dst_tensor));
+                            convert_tensor((void*)read_buffer.data(), tensor_storage.type, (void*)convert_buffer.data(), dst_tensor->type, (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
+                            t1 = ggml_time_ms();
+                            convert_time_ms.fetch_add(t1 - t0);
+                            t0 = ggml_time_ms();
+                            ggml_backend_tensor_set(dst_tensor, convert_buffer.data(), 0, ggml_nbytes(dst_tensor));
+                            t1 = ggml_time_ms();
+                            copy_to_backend_time_ms.fetch_add(t1 - t0);
+                        }
+                    }
+                }
+                if (zip != nullptr) {
+                    zip_close(zip);
+                }
+            });
         }
 
-        std::vector<uint8_t> read_buffer;
-        std::vector<uint8_t> convert_buffer;
-
-        auto read_data = [&](const TensorStorage& tensor_storage, char* buf, size_t n) {
-            if (zip != NULL) {
-                zip_entry_openbyindex(zip, tensor_storage.index_in_zip);
-                size_t entry_size = zip_entry_size(zip);
-                if (entry_size != n) {
-                    read_buffer.resize(entry_size);
-                    prev_time_ms = ggml_time_ms();
-                    zip_entry_noallocread(zip, (void*)read_buffer.data(), entry_size);
-                    curr_time_ms = ggml_time_ms();
-                    read_time_ms += curr_time_ms - prev_time_ms;
-                    prev_time_ms = curr_time_ms;
-                    memcpy((void*)buf, (void*)(read_buffer.data() + tensor_storage.offset), n);
-                    curr_time_ms = ggml_time_ms();
-                    memcpy_time_ms += curr_time_ms - prev_time_ms;
-                } else {
-                    prev_time_ms = ggml_time_ms();
-                    zip_entry_noallocread(zip, (void*)buf, n);
-                    curr_time_ms = ggml_time_ms();
-                    read_time_ms += curr_time_ms - prev_time_ms;
-                }
-                zip_entry_close(zip);
-            } else {
-                prev_time_ms = ggml_time_ms();
-                file.seekg(tensor_storage.offset);
-                file.read(buf, n);
-                curr_time_ms = ggml_time_ms();
-                read_time_ms += curr_time_ms - prev_time_ms;
-                if (!file) {
-                    LOG_ERROR("read tensor data failed: '%s'", file_path.c_str());
-                    return false;
-                }
-            }
-            return true;
-        };
-        int tensor_count = 0;
-        int64_t t0       = ggml_time_ms();
-        int64_t t1       = t0;
-        bool partial     = true;
-        int tensor_max   = (int)processed_tensor_storages.size();
-        pretty_progress(0, tensor_max, 0.0f);
-        for (auto& tensor_storage : processed_tensor_storages) {
-            if (tensor_storage.file_index != file_index) {
-                ++tensor_count;
-                continue;
-            }
-            ggml_tensor* dst_tensor = NULL;
-
-            success = on_new_tensor_cb(tensor_storage, &dst_tensor);
-            if (!success) {
-                LOG_WARN("process tensor failed: '%s'", tensor_storage.name.c_str());
+        while (true) {
+            size_t current_idx = tensor_idx.load();
+            if (current_idx >= file_tensors.size() || failed) {
                 break;
             }
-
-            if (dst_tensor == NULL) {
-                ++tensor_count;
-                continue;
-            }
-
-            size_t nbytes_to_read = tensor_storage.nbytes_to_read();
-
-            if (dst_tensor->buffer == NULL || ggml_backend_buffer_is_host(dst_tensor->buffer)) {
-                // 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());
-                    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);
-                    }
-
-                    prev_time_ms = ggml_time_ms();
-                    if (tensor_storage.is_bf16) {
-                        // inplace op
-                        bf16_to_f32_vec((uint16_t*)dst_tensor->data, (float*)dst_tensor->data, tensor_storage.nelements());
-                    } else if (tensor_storage.is_f8_e4m3) {
-                        // inplace op
-                        f8_e4m3_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
-                    } else if (tensor_storage.is_f8_e5m2) {
-                        // inplace op
-                        f8_e5m2_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
-                    } else 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());
-                    }
-                    curr_time_ms = ggml_time_ms();
-                    convert_time_ms += curr_time_ms - prev_time_ms;
-                } else {
-                    read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
-                    read_data(tensor_storage, (char*)read_buffer.data(), nbytes_to_read);
-
-                    prev_time_ms = ggml_time_ms();
-                    if (tensor_storage.is_bf16) {
-                        // inplace op
-                        bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
-                    } else if (tensor_storage.is_f8_e4m3) {
-                        // inplace op
-                        f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
-                    } else if (tensor_storage.is_f8_e5m2) {
-                        // inplace op
-                        f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
-                    } 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]);
-                    curr_time_ms = ggml_time_ms();
-                    convert_time_ms += curr_time_ms - prev_time_ms;
-                }
-            } else {
-                read_buffer.resize(std::max(tensor_storage.nbytes(), tensor_storage.nbytes_to_read()));
-                read_data(tensor_storage, (char*)read_buffer.data(), nbytes_to_read);
-
-                prev_time_ms = ggml_time_ms();
-                if (tensor_storage.is_bf16) {
-                    // inplace op
-                    bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
-                } else if (tensor_storage.is_f8_e4m3) {
-                    // inplace op
-                    f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
-                } else if (tensor_storage.is_f8_e5m2) {
-                    // inplace op
-                    f8_e5m2_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
-                } 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) {
-                    // copy to device memory
-                    curr_time_ms = ggml_time_ms();
-                    convert_time_ms += curr_time_ms - prev_time_ms;
-                    prev_time_ms = curr_time_ms;
-                    ggml_backend_tensor_set(dst_tensor, read_buffer.data(), 0, ggml_nbytes(dst_tensor));
-                    curr_time_ms = ggml_time_ms();
-                    copy_to_backend_time_ms += curr_time_ms - prev_time_ms;
-                } else {
-                    // convert first, then copy to device memory
-                    convert_buffer.resize(ggml_nbytes(dst_tensor));
-                    convert_tensor((void*)read_buffer.data(), tensor_storage.type,
-                                   (void*)convert_buffer.data(), dst_tensor->type,
-                                   (int)tensor_storage.nelements() / (int)tensor_storage.ne[0], (int)tensor_storage.ne[0]);
-                    curr_time_ms = ggml_time_ms();
-                    convert_time_ms += curr_time_ms - prev_time_ms;
-                    prev_time_ms = curr_time_ms;
-                    ggml_backend_tensor_set(dst_tensor, convert_buffer.data(), 0, ggml_nbytes(dst_tensor));
-                    curr_time_ms = ggml_time_ms();
-                    copy_to_backend_time_ms += curr_time_ms - prev_time_ms;
-                }
-            }
-            ++tensor_count;
-            int64_t t2 = ggml_time_ms();
-            if ((t2 - t1) >= 200) {
-                t1 = t2;
-                pretty_progress(tensor_count, tensor_max, (t1 - t0) / (1000.0f * tensor_count));
-                partial = tensor_count != tensor_max;
-            }
+            size_t curr_num = total_tensors_processed + current_idx;
+            pretty_progress(curr_num, total_tensors_to_process, (ggml_time_ms() - t_start) / 1000.0f / (curr_num + 1e-6f));
+            std::this_thread::sleep_for(std::chrono::milliseconds(200));
         }
 
-        if (partial) {
-            if (tensor_count >= 1) {
-                t1 = ggml_time_ms();
-                pretty_progress(tensor_count, tensor_max, (t1 - t0) / (1000.0f * tensor_count));
-            }
-            if (tensor_count < tensor_max) {
-                printf("\n");
-            }
+        for (auto& w : workers) {
+            w.join();
         }
 
-        if (zip != NULL) {
-            zip_close(zip);
-        }
-
-        if (!success) {
+        if (failed) {
+            success = false;
             break;
         }
+        total_tensors_processed += file_tensors.size();
+        pretty_progress(total_tensors_processed, total_tensors_to_process, (ggml_time_ms() - t_start) / 1000.0f / (total_tensors_processed + 1e-6f));
+        if (total_tensors_processed < total_tensors_to_process) {
+            printf("\n");
+        }
     }
+
     int64_t end_time = ggml_time_ms();
     LOG_INFO("loading tensors completed, taking %.2fs (process: %.2fs, read: %.2fs, memcpy: %.2fs, convert: %.2fs, copy_to_backend: %.2fs)",
              (end_time - start_time) / 1000.f,
              process_time_ms / 1000.f,
-             read_time_ms / 1000.f,
-             memcpy_time_ms / 1000.f,
-             convert_time_ms / 1000.f,
-             copy_to_backend_time_ms / 1000.f);
+             (read_time_ms.load() / (float)last_n_threads) / 1000.f,
+             (memcpy_time_ms.load() / (float)last_n_threads) / 1000.f,
+             (convert_time_ms.load() / (float)last_n_threads) / 1000.f,
+             (copy_to_backend_time_ms.load() / (float)last_n_threads) / 1000.f);
     return success;
 }
 
 bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
-                               std::set<std::string> ignore_tensors) {
+                               std::set<std::string> ignore_tensors,
+                               int n_threads) {
     std::set<std::string> tensor_names_in_file;
+    std::mutex tensor_names_mutex;
     auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
         const std::string& name = tensor_storage.name;
         // LOG_DEBUG("%s", tensor_storage.to_string().c_str());
-        tensor_names_in_file.insert(name);
+        {
+            std::lock_guard<std::mutex> lock(tensor_names_mutex);
+            tensor_names_in_file.insert(name);
+        }
 
         struct ggml_tensor* real;
         if (tensors.find(name) != tensors.end()) {
@@ -2263,7 +2403,7 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
         return true;
     };
 
-    bool success = load_tensors(on_new_tensor_cb);
+    bool success = load_tensors(on_new_tensor_cb, n_threads);
     if (!success) {
         LOG_ERROR("load tensors from file failed");
         return false;
@@ -2310,7 +2450,7 @@ std::vector<std::pair<std::string, ggml_type>> parse_tensor_type_rules(const std
         if (type_name == "f32") {
             tensor_type = GGML_TYPE_F32;
         } else {
-            for (size_t i = 0; i < SD_TYPE_COUNT; i++) {
+            for (size_t i = 0; i < GGML_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;
@@ -2351,6 +2491,8 @@ bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage
             // Pass, do not convert. For MMDiT
         } else if (contains(name, "time_embed.") || contains(name, "label_emb.")) {
             // Pass, do not convert. For Unet
+        } else if (contains(name, "embedding")) {
+            // Pass, do not convert embedding
         } else {
             return true;
         }
@@ -2364,12 +2506,13 @@ bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type
     mem_size += tensor_storages.size() * ggml_tensor_overhead();
     mem_size += get_params_mem_size(backend, type);
     LOG_INFO("model tensors mem size: %.2fMB", mem_size / 1024.f / 1024.f);
-    ggml_context* ggml_ctx = ggml_init({mem_size, NULL, false});
+    ggml_context* ggml_ctx = ggml_init({mem_size, nullptr, false});
 
     gguf_context* gguf_ctx = gguf_init_empty();
 
     auto tensor_type_rules = parse_tensor_type_rules(tensor_type_rules_str);
 
+    std::mutex tensor_mutex;
     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;
@@ -2387,8 +2530,9 @@ bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type
             tensor_type = dst_type;
         }
 
+        std::lock_guard<std::mutex> lock(tensor_mutex);
         ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
-        if (tensor == NULL) {
+        if (tensor == nullptr) {
             LOG_ERROR("ggml_new_tensor failed");
             return false;
         }
@@ -2421,7 +2565,7 @@ bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type type
 
 int64_t ModelLoader::get_params_mem_size(ggml_backend_t backend, ggml_type type) {
     size_t alignment = 128;
-    if (backend != NULL) {
+    if (backend != nullptr) {
         alignment = ggml_backend_get_alignment(backend);
     }
     int64_t mem_size = 0;
@@ -2451,7 +2595,7 @@ bool convert(const char* input_path, const char* vae_path, const char* output_pa
         return false;
     }
 
-    if (vae_path != NULL && strlen(vae_path) > 0) {
+    if (vae_path != nullptr && strlen(vae_path) > 0) {
         if (!model_loader.init_from_file(vae_path, "vae.")) {
             LOG_ERROR("init model loader from file failed: '%s'", vae_path);
             return false;

model.h

@@ -8,6 +8,7 @@
 #include <sstream>
 #include <string>
 #include <tuple>
+#include <utility>
 #include <vector>
 
 #include "ggml-backend.h"
@@ -31,9 +32,12 @@ enum SDVersion {
     VERSION_SD3,
     VERSION_FLUX,
     VERSION_FLUX_FILL,
+    VERSION_FLUX_CONTROLS,
+    VERSION_FLEX_2,
     VERSION_WAN2,
     VERSION_WAN2_2_I2V,
     VERSION_WAN2_2_TI2V,
+    VERSION_QWEN_IMAGE,
     VERSION_COUNT,
 };
 
@@ -66,7 +70,7 @@ static inline bool sd_version_is_sd3(SDVersion version) {
 }
 
 static inline bool sd_version_is_flux(SDVersion version) {
-    if (version == VERSION_FLUX || version == VERSION_FLUX_FILL) {
+    if (version == VERSION_FLUX || version == VERSION_FLUX_FILL || version == VERSION_FLUX_CONTROLS || version == VERSION_FLEX_2) {
         return true;
     }
     return false;
@@ -79,15 +83,25 @@ static inline bool sd_version_is_wan(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_qwen_image(SDVersion version) {
+    if (version == VERSION_QWEN_IMAGE) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_inpaint(SDVersion version) {
-    if (version == VERSION_SD1_INPAINT || version == VERSION_SD2_INPAINT || version == VERSION_SDXL_INPAINT || version == VERSION_FLUX_FILL) {
+    if (version == VERSION_SD1_INPAINT || version == VERSION_SD2_INPAINT || version == VERSION_SDXL_INPAINT || version == VERSION_FLUX_FILL || version == VERSION_FLEX_2) {
         return true;
     }
     return false;
 }
 
 static inline bool sd_version_is_dit(SDVersion version) {
-    if (sd_version_is_flux(version) || sd_version_is_sd3(version) || sd_version_is_wan(version)) {
+    if (sd_version_is_flux(version) ||
+        sd_version_is_sd3(version) ||
+        sd_version_is_wan(version) ||
+        sd_version_is_qwen_image(version)) {
         return true;
     }
     return false;
@@ -97,8 +111,12 @@ static inline bool sd_version_is_unet_edit(SDVersion version) {
     return version == VERSION_SD1_PIX2PIX || version == VERSION_SDXL_PIX2PIX;
 }
 
+static inline bool sd_version_is_control(SDVersion version) {
+    return version == VERSION_FLUX_CONTROLS || version == VERSION_FLEX_2;
+}
+
 static bool sd_version_is_inpaint_or_unet_edit(SDVersion version) {
-    return sd_version_is_unet_edit(version) || sd_version_is_inpaint(version);
+    return sd_version_is_unet_edit(version) || sd_version_is_inpaint(version) || sd_version_is_control(version);
 }
 
 enum PMVersion {
@@ -119,12 +137,12 @@ struct TensorStorage {
 
     size_t file_index = 0;
     int index_in_zip  = -1;  // >= means stored in a zip file
-    size_t offset     = 0;   // offset in file
+    uint64_t offset   = 0;   // offset in file
 
     TensorStorage() = default;
 
-    TensorStorage(const std::string& name, ggml_type type, const int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
-        : name(name), type(type), n_dims(n_dims), file_index(file_index), offset(offset) {
+    TensorStorage(std::string name, ggml_type type, const int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
+        : name(std::move(name)), type(type), n_dims(n_dims), file_index(file_index), offset(offset) {
         for (int i = 0; i < n_dims; i++) {
             this->ne[i] = ne[i];
         }
@@ -164,10 +182,10 @@ struct TensorStorage {
 
     std::vector<TensorStorage> chunk(size_t n) {
         std::vector<TensorStorage> chunks;
-        size_t chunk_size = nbytes_to_read() / n;
+        uint64_t chunk_size = nbytes_to_read() / n;
         // printf("%d/%d\n", chunk_size, nbytes_to_read());
         reverse_ne();
-        for (int i = 0; i < n; i++) {
+        for (size_t i = 0; i < n; i++) {
             TensorStorage chunk_i = *this;
             chunk_i.ne[0]         = ne[0] / n;
             chunk_i.offset        = offset + i * chunk_size;
@@ -242,14 +260,23 @@ public:
     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();
-    ggml_type get_diffusion_model_wtype();
-    ggml_type get_vae_wtype();
+    std::map<ggml_type, uint32_t> get_wtype_stat();
+    std::map<ggml_type, uint32_t> get_conditioner_wtype_stat();
+    std::map<ggml_type, uint32_t> get_diffusion_model_wtype_stat();
+    std::map<ggml_type, uint32_t> get_vae_wtype_stat();
     void set_wtype_override(ggml_type wtype, std::string prefix = "");
-    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb);
+    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0);
     bool load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
-                      std::set<std::string> ignore_tensors = {});
+                      std::set<std::string> ignore_tensors = {},
+                      int n_threads                        = 0);
+
+    std::vector<std::string> get_tensor_names() const {
+        std::vector<std::string> names;
+        for (const auto& ts : tensor_storages) {
+            names.push_back(ts.name);
+        }
+        return names;
+    }
 
     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);
@@ -257,6 +284,7 @@ public:
     ~ModelLoader() = default;
 
     static std::string load_merges();
+    static std::string load_qwen2_merges();
     static std::string load_t5_tokenizer_json();
     static std::string load_umt5_tokenizer_json();
 };

pmid.hpp

@@ -42,41 +42,6 @@ public:
     }
 };
 
-/*
-class QFormerPerceiver(nn.Module):
-    def __init__(self, id_embeddings_dim, cross_attention_dim, num_tokens, embedding_dim=1024, use_residual=True, ratio=4):
-        super().__init__()
-
-        self.num_tokens = num_tokens
-        self.cross_attention_dim = cross_attention_dim
-        self.use_residual = use_residual
-        print(cross_attention_dim*num_tokens)
-        self.token_proj = nn.Sequential(
-            nn.Linear(id_embeddings_dim, id_embeddings_dim*ratio),
-            nn.GELU(),
-            nn.Linear(id_embeddings_dim*ratio, cross_attention_dim*num_tokens),
-        )
-        self.token_norm = nn.LayerNorm(cross_attention_dim)
-        self.perceiver_resampler = FacePerceiverResampler(
-            dim=cross_attention_dim,
-            depth=4,
-            dim_head=128,
-            heads=cross_attention_dim // 128,
-            embedding_dim=embedding_dim,
-            output_dim=cross_attention_dim,
-            ff_mult=4,
-        )
-
-    def forward(self, x, last_hidden_state):
-        x = self.token_proj(x)
-        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
-        x = self.token_norm(x) # cls token
-        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
-        if self.use_residual: # TODO: if use_residual is not true
-            out = x + 1.0 * out
-        return out
-*/
-
 struct PMFeedForward : public GGMLBlock {
     // network hparams
     int dim;
@@ -122,17 +87,8 @@ public:
         int64_t ne[4];
         for (int i = 0; i < 4; ++i)
             ne[i] = x->ne[i];
-        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 0: ");
-        // printf("heads = %d \n", heads);
-        // x = ggml_view_4d(ctx, x, x->ne[0], x->ne[1], heads, x->ne[2]/heads,
-        //                          x->nb[1], x->nb[2], x->nb[3], 0);
         x = ggml_reshape_4d(ctx, x, x->ne[0] / heads, heads, x->ne[1], x->ne[2]);
-        // x = ggml_view_4d(ctx, x, x->ne[0]/heads, heads, x->ne[1], x->ne[2],
-        //                          x->nb[1], x->nb[2], x->nb[3], 0);
-        // x = ggml_cont(ctx, x);
         x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));
-        // print_ggml_tensor(x, true, "PerceiverAttention reshape x 1: ");
-        // x  = ggml_reshape_4d(ctx, x, ne[0], heads, ne[1], ne[2]/heads);
         return x;
     }
 
@@ -269,17 +225,6 @@ public:
             4));
     }
 
-    /*
-    def forward(self, x, last_hidden_state):
-        x = self.token_proj(x)
-        x = x.reshape(-1, self.num_tokens, self.cross_attention_dim)
-        x = self.token_norm(x) # cls token
-        out = self.perceiver_resampler(x, last_hidden_state) # retrieve from patch tokens
-        if self.use_residual: # TODO: if use_residual is not true
-            out = x + 1.0 * out
-        return out
-    */
-
     struct ggml_tensor* forward(struct ggml_context* ctx,
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* last_hidden_state) {
@@ -299,113 +244,6 @@ public:
     }
 };
 
-/*
-class FacePerceiverResampler(torch.nn.Module):
-    def __init__(
-        self,
-        *,
-        dim=768,
-        depth=4,
-        dim_head=64,
-        heads=16,
-        embedding_dim=1280,
-        output_dim=768,
-        ff_mult=4,
-    ):
-        super().__init__()
-
-        self.proj_in = torch.nn.Linear(embedding_dim, dim)
-        self.proj_out = torch.nn.Linear(dim, output_dim)
-        self.norm_out = torch.nn.LayerNorm(output_dim)
-        self.layers = torch.nn.ModuleList([])
-        for _ in range(depth):
-            self.layers.append(
-                torch.nn.ModuleList(
-                    [
-                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
-                        FeedForward(dim=dim, mult=ff_mult),
-                    ]
-                )
-            )
-
-    def forward(self, latents, x):
-        x = self.proj_in(x)
-        for attn, ff in self.layers:
-            latents = attn(x, latents) + latents
-            latents = ff(latents) + latents
-        latents = self.proj_out(latents)
-        return self.norm_out(latents)
-*/
-
-/*
-
-def FeedForward(dim, mult=4):
-    inner_dim = int(dim * mult)
-    return nn.Sequential(
-        nn.LayerNorm(dim),
-        nn.Linear(dim, inner_dim, bias=False),
-        nn.GELU(),
-        nn.Linear(inner_dim, dim, bias=False),
-    )
-
-def reshape_tensor(x, heads):
-    bs, length, width = x.shape
-    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
-    x = x.view(bs, length, heads, -1)
-    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
-    x = x.transpose(1, 2)
-    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
-    x = x.reshape(bs, heads, length, -1)
-    return x
-
-class PerceiverAttention(nn.Module):
-    def __init__(self, *, dim, dim_head=64, heads=8):
-        super().__init__()
-        self.scale = dim_head**-0.5
-        self.dim_head = dim_head
-        self.heads = heads
-        inner_dim = dim_head * heads
-
-        self.norm1 = nn.LayerNorm(dim)
-        self.norm2 = nn.LayerNorm(dim)
-
-        self.to_q = nn.Linear(dim, inner_dim, bias=False)
-        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
-        self.to_out = nn.Linear(inner_dim, dim, bias=False)
-
-    def forward(self, x, latents):
-        """
-        Args:
-            x (torch.Tensor): image features
-                shape (b, n1, D)
-            latent (torch.Tensor): latent features
-                shape (b, n2, D)
-        """
-        x = self.norm1(x)
-        latents = self.norm2(latents)
-
-        b, l, _ = latents.shape
-
-        q = self.to_q(latents)
-        kv_input = torch.cat((x, latents), dim=-2)
-        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
-
-        q = reshape_tensor(q, self.heads)
-        k = reshape_tensor(k, self.heads)
-        v = reshape_tensor(v, self.heads)
-
-        # attention
-        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
-        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
-        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
-        out = weight @ v
-
-        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
-
-        return self.to_out(out)
-
-*/
-
 struct FuseModule : public GGMLBlock {
     // network hparams
     int embed_dim;
@@ -425,31 +263,13 @@ public:
         auto mlp2       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp2"]);
         auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm"]);
 
-        // print_ggml_tensor(id_embeds, true, "Fuseblock id_embeds: ");
-        // print_ggml_tensor(prompt_embeds, true, "Fuseblock prompt_embeds: ");
-
-        // auto prompt_embeds0 = ggml_cont(ctx, ggml_permute(ctx, prompt_embeds, 2, 0, 1, 3));
-        // auto id_embeds0     = ggml_cont(ctx, ggml_permute(ctx, id_embeds, 2, 0, 1, 3));
-        // print_ggml_tensor(id_embeds0, true, "Fuseblock id_embeds0: ");
-        // print_ggml_tensor(prompt_embeds0, true, "Fuseblock prompt_embeds0: ");
-        // concat is along dim 2
-        // auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds0, id_embeds0, 2);
         auto stacked_id_embeds = ggml_concat(ctx, prompt_embeds, id_embeds, 0);
-        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 0: ");
-        // stacked_id_embeds      = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 1, 2, 0, 3));
-        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
-        // stacked_id_embeds = mlp1.forward(ctx, stacked_id_embeds);
-        // stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
-        // stacked_id_embeds = mlp2.forward(ctx, stacked_id_embeds);
-        // stacked_id_embeds = ggml_nn_layer_norm(ctx, stacked_id_embeds, ln_w, ln_b);
 
         stacked_id_embeds = mlp1->forward(ctx, stacked_id_embeds);
         stacked_id_embeds = ggml_add(ctx, stacked_id_embeds, prompt_embeds);
         stacked_id_embeds = mlp2->forward(ctx, stacked_id_embeds);
         stacked_id_embeds = layer_norm->forward(ctx, stacked_id_embeds);
 
-        // print_ggml_tensor(stacked_id_embeds, true, "Fuseblock stacked_id_embeds 1: ");
-
         return stacked_id_embeds;
     }
 
@@ -464,21 +284,14 @@ public:
 
         struct ggml_tensor* valid_id_embeds = id_embeds;
         // # slice out the image token embeddings
-        // print_ggml_tensor(class_tokens_mask_pos, false);
         ggml_set_name(class_tokens_mask_pos, "class_tokens_mask_pos");
         ggml_set_name(prompt_embeds, "prompt_embeds");
-        // print_ggml_tensor(valid_id_embeds, true, "valid_id_embeds");
-        // print_ggml_tensor(class_tokens_mask_pos, true, "class_tokens_mask_pos");
         struct ggml_tensor* image_token_embeds = ggml_get_rows(ctx, prompt_embeds, class_tokens_mask_pos);
         ggml_set_name(image_token_embeds, "image_token_embeds");
         valid_id_embeds                       = ggml_reshape_2d(ctx, valid_id_embeds, valid_id_embeds->ne[0],
                                                                 ggml_nelements(valid_id_embeds) / valid_id_embeds->ne[0]);
         struct ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);
 
-        // stacked_id_embeds = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
-        // print_ggml_tensor(stacked_id_embeds, true, "AA stacked_id_embeds");
-        // print_ggml_tensor(left, true, "AA left");
-        // print_ggml_tensor(right, true, "AA right");
         if (left && right) {
             stacked_id_embeds = ggml_concat(ctx, left, stacked_id_embeds, 1);
             stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
@@ -487,15 +300,12 @@ public:
         } else if (right) {
             stacked_id_embeds = ggml_concat(ctx, stacked_id_embeds, right, 1);
         }
-        // print_ggml_tensor(stacked_id_embeds, true, "BB stacked_id_embeds");
-        // stacked_id_embeds                         = ggml_cont(ctx, ggml_permute(ctx, stacked_id_embeds, 0, 2, 1, 3));
-        // print_ggml_tensor(stacked_id_embeds, true, "CC stacked_id_embeds");
+
         class_tokens_mask                         = ggml_cont(ctx, ggml_transpose(ctx, class_tokens_mask));
         class_tokens_mask                         = ggml_repeat(ctx, class_tokens_mask, prompt_embeds);
         prompt_embeds                             = ggml_mul(ctx, prompt_embeds, class_tokens_mask);
         struct ggml_tensor* updated_prompt_embeds = ggml_add(ctx, prompt_embeds, stacked_id_embeds);
         ggml_set_name(updated_prompt_embeds, "updated_prompt_embeds");
-        // print_ggml_tensor(updated_prompt_embeds, true, "updated_prompt_embeds: ");
         return updated_prompt_embeds;
     }
 };
@@ -551,34 +361,11 @@ struct PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock : public CLIPVisionMo
           num_tokens(2) {
         blocks["visual_projection_2"] = std::shared_ptr<GGMLBlock>(new Linear(1024, 1280, false));
         blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
-        /*
-        cross_attention_dim = 2048
-        # projection
-        self.num_tokens = 2
-        self.cross_attention_dim = cross_attention_dim
-        self.qformer_perceiver = QFormerPerceiver(
-                                    id_embeddings_dim,
-                                    cross_attention_dim,
-                                    self.num_tokens,
-                                )*/
-        blocks["qformer_perceiver"] = std::shared_ptr<GGMLBlock>(new QFormerPerceiver(id_embeddings_dim,
-                                                                                      cross_attention_dim,
-                                                                                      num_tokens));
+        blocks["qformer_perceiver"]   = std::shared_ptr<GGMLBlock>(new QFormerPerceiver(id_embeddings_dim,
+                                                                                        cross_attention_dim,
+                                                                                        num_tokens));
     }
 
-    /*
-    def forward(self, id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds):
-        b, num_inputs, c, h, w = id_pixel_values.shape
-        id_pixel_values = id_pixel_values.view(b * num_inputs, c, h, w)
-
-        last_hidden_state = self.vision_model(id_pixel_values)[0]
-        id_embeds = id_embeds.view(b * num_inputs, -1)
-
-        id_embeds = self.qformer_perceiver(id_embeds, last_hidden_state)
-        id_embeds = id_embeds.view(b, num_inputs, self.num_tokens, -1)
-        updated_prompt_embeds = self.fuse_module(prompt_embeds, id_embeds, class_tokens_mask)
-    */
-
     struct ggml_tensor* forward(struct ggml_context* ctx,
                                 ggml_backend_t backend,
                                 struct ggml_tensor* id_pixel_values,
@@ -685,8 +472,8 @@ public:
         struct ggml_tensor* prompt_embeds_d   = to_backend(prompt_embeds);
         struct ggml_tensor* id_embeds_d       = to_backend(id_embeds);
 
-        struct ggml_tensor* left  = NULL;
-        struct ggml_tensor* right = NULL;
+        struct ggml_tensor* left  = nullptr;
+        struct ggml_tensor* right = nullptr;
         for (int i = 0; i < class_tokens_mask.size(); i++) {
             if (class_tokens_mask[i]) {
                 // printf(" 1,");
@@ -741,7 +528,7 @@ public:
                 }
             }
         }
-        struct ggml_tensor* updated_prompt_embeds = NULL;
+        struct ggml_tensor* updated_prompt_embeds = nullptr;
         if (pm_version == PM_VERSION_1)
             updated_prompt_embeds = id_encoder.forward(ctx0,
                                                        runtime_backend,
@@ -804,7 +591,7 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
         return "id_embeds";
     }
 
-    bool load_from_file(bool filter_tensor = false) {
+    bool load_from_file(bool filter_tensor, int n_threads) {
         LOG_INFO("loading PhotoMaker ID Embeds from '%s'", file_path.c_str());
 
         if (load_failed) {
@@ -812,7 +599,8 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
             return false;
         }
 
-        bool dry_run          = true;
+        bool dry_run = true;
+        std::mutex tensor_mutex;
         auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
             const std::string& name = tensor_storage.name;
 
@@ -821,6 +609,7 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
                 return true;
             }
             if (dry_run) {
+                std::lock_guard<std::mutex> lock(tensor_mutex);
                 struct ggml_tensor* real = ggml_new_tensor(params_ctx,
                                                            tensor_storage.type,
                                                            tensor_storage.n_dims,
@@ -834,11 +623,11 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
             return true;
         };
 
-        model_loader->load_tensors(on_new_tensor_cb);
+        model_loader->load_tensors(on_new_tensor_cb, n_threads);
         alloc_params_buffer();
 
         dry_run = false;
-        model_loader->load_tensors(on_new_tensor_cb);
+        model_loader->load_tensors(on_new_tensor_cb, n_threads);
 
         LOG_DEBUG("finished loading PhotoMaker ID Embeds ");
         return true;
@@ -849,7 +638,7 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
         pos = tensors.find("pmid.id_embeds");
         if (pos != tensors.end())
             return pos->second;
-        return NULL;
+        return nullptr;
     }
 };
 

preprocessing.hpp

@@ -6,8 +6,8 @@
 
 void convolve(struct ggml_tensor* input, struct ggml_tensor* output, struct ggml_tensor* kernel, int padding) {
     struct ggml_init_params params;
-    params.mem_size                 = 20 * 1024 * 1024;  // 10
-    params.mem_buffer               = NULL;
+    params.mem_size                 = 80 * input->ne[0] * input->ne[1];  // 20M for 512x512
+    params.mem_buffer               = nullptr;
     params.no_alloc                 = false;
     struct ggml_context* ctx0       = ggml_init(params);
     struct ggml_tensor* kernel_fp16 = ggml_new_tensor_4d(ctx0, GGML_TYPE_F16, kernel->ne[0], kernel->ne[1], 1, 1);
@@ -162,16 +162,16 @@ void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float lo
     }
 }
 
-uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
+bool preprocess_canny(sd_image_t img, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
     struct ggml_init_params params;
-    params.mem_size               = static_cast<size_t>(10 * 1024 * 1024);  // 10
-    params.mem_buffer             = NULL;
+    params.mem_size               = static_cast<size_t>(40 * img.width * img.height);  // 10MB for 512x512
+    params.mem_buffer             = nullptr;
     params.no_alloc               = false;
     struct ggml_context* work_ctx = ggml_init(params);
 
     if (!work_ctx) {
         LOG_ERROR("ggml_init() failed");
-        return NULL;
+        return false;
     }
 
     float kX[9] = {
@@ -192,8 +192,8 @@ uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_thresh
     struct ggml_tensor* sf_ky = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
     memcpy(sf_ky->data, kY, ggml_nbytes(sf_ky));
     gaussian_kernel(gkernel);
-    struct ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, height, 3, 1);
-    struct ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, width, height, 1, 1);
+    struct ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 3, 1);
+    struct ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 1, 1);
     struct ggml_tensor* iX         = ggml_dup_tensor(work_ctx, image_gray);
     struct ggml_tensor* iY         = ggml_dup_tensor(work_ctx, image_gray);
     struct ggml_tensor* G          = ggml_dup_tensor(work_ctx, image_gray);
@@ -209,8 +209,8 @@ uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_thresh
     non_max_supression(image_gray, G, tetha);
     threshold_hystersis(image_gray, high_threshold, low_threshold, weak, strong);
     // to RGB channels
-    for (int iy = 0; iy < height; iy++) {
-        for (int ix = 0; ix < width; ix++) {
+    for (int iy = 0; iy < img.height; iy++) {
+        for (int ix = 0; ix < img.width; ix++) {
             float gray = ggml_tensor_get_f32(image_gray, ix, iy);
             gray       = inverse ? 1.0f - gray : gray;
             ggml_tensor_set_f32(image, gray, ix, iy);
@@ -218,10 +218,9 @@ uint8_t* preprocess_canny(uint8_t* img, int width, int height, float high_thresh
             ggml_tensor_set_f32(image, gray, ix, iy, 2);
         }
     }
-    free(img);
-    uint8_t* output = sd_tensor_to_image(image);
+    sd_tensor_to_image(image, img.data);
     ggml_free(work_ctx);
-    return output;
+    return true;
 }
 
 #endif  // __PREPROCESSING_HPP__

qwen_image.hpp

@@ -0,0 +1,696 @@
+#ifndef __QWEN_IMAGE_HPP__
+#define __QWEN_IMAGE_HPP__
+
+#include <memory>
+
+#include "common.hpp"
+#include "flux.hpp"
+#include "ggml_extend.hpp"
+
+namespace Qwen {
+    constexpr int QWEN_IMAGE_GRAPH_SIZE = 20480;
+
+    struct TimestepEmbedding : public GGMLBlock {
+    public:
+        TimestepEmbedding(int64_t in_channels,
+                          int64_t time_embed_dim,
+                          int64_t out_dim       = 0,
+                          int64_t cond_proj_dim = 0,
+                          bool sample_proj_bias = true) {
+            blocks["linear_1"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, time_embed_dim, sample_proj_bias));
+            if (cond_proj_dim > 0) {
+                blocks["cond_proj"] = std::shared_ptr<GGMLBlock>(new Linear(cond_proj_dim, in_channels, false));
+            }
+            if (out_dim <= 0) {
+                out_dim = time_embed_dim;
+            }
+            blocks["linear_2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, out_dim, sample_proj_bias));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* sample,
+                                    struct ggml_tensor* condition = nullptr) {
+            if (condition != nullptr) {
+                auto cond_proj = std::dynamic_pointer_cast<Linear>(blocks["cond_proj"]);
+                sample         = ggml_add(ctx, sample, cond_proj->forward(ctx, condition));
+            }
+            auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["linear_1"]);
+            auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["linear_2"]);
+
+            sample = linear_1->forward(ctx, sample);
+            sample = ggml_silu_inplace(ctx, sample);
+            sample = linear_2->forward(ctx, sample);
+            return sample;
+        }
+    };
+
+    struct QwenTimestepProjEmbeddings : public GGMLBlock {
+    public:
+        QwenTimestepProjEmbeddings(int64_t embedding_dim) {
+            blocks["timestep_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedding(256, embedding_dim));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* timesteps) {
+            // timesteps: [N,]
+            // return: [N, embedding_dim]
+            auto timestep_embedder = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["timestep_embedder"]);
+
+            auto timesteps_proj = ggml_nn_timestep_embedding(ctx, timesteps, 256, 10000, 1.f);
+            auto timesteps_emb  = timestep_embedder->forward(ctx, timesteps_proj);
+            return timesteps_emb;
+        }
+    };
+
+    struct QwenImageAttention : public GGMLBlock {
+    protected:
+        int64_t dim_head;
+        bool flash_attn;
+
+    public:
+        QwenImageAttention(int64_t query_dim,
+                           int64_t dim_head,
+                           int64_t num_heads,
+                           int64_t out_dim         = 0,
+                           int64_t out_context_dim = 0,
+                           bool bias               = true,
+                           bool out_bias           = true,
+                           float eps               = 1e-6,
+                           bool flash_attn         = false)
+            : dim_head(dim_head), flash_attn(flash_attn) {
+            int64_t inner_dim = out_dim > 0 ? out_dim : dim_head * num_heads;
+            out_dim           = out_dim > 0 ? out_dim : query_dim;
+            out_context_dim   = out_context_dim > 0 ? out_context_dim : query_dim;
+
+            blocks["to_q"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["to_k"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["to_v"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+
+            blocks["norm_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+            blocks["norm_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+
+            blocks["add_q_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["add_k_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+            blocks["add_v_proj"] = std::shared_ptr<GGMLBlock>(new Linear(query_dim, inner_dim, bias));
+
+            blocks["norm_added_q"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+            blocks["norm_added_k"] = std::shared_ptr<GGMLBlock>(new RMSNorm(dim_head, eps));
+
+            float scale = 1.f / 32.f;
+            // The purpose of the scale here is to prevent NaN issues in certain situations.
+            // For example when using CUDA but the weights are k-quants (not all prompts).
+            blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_dim, out_bias, false, false, scale));
+            // to_out.1 is nn.Dropout
+
+            blocks["to_add_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_context_dim, out_bias, false, false, scale));
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,
+                                                      ggml_backend_t backend,
+                                                      struct ggml_tensor* img,
+                                                      struct ggml_tensor* txt,
+                                                      struct ggml_tensor* pe,
+                                                      struct ggml_tensor* mask = nullptr) {
+            // 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 norm_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_q"]);
+            auto norm_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_k"]);
+
+            auto to_q     = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
+            auto to_k     = std::dynamic_pointer_cast<Linear>(blocks["to_k"]);
+            auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
+            auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+
+            auto norm_added_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_q"]);
+            auto norm_added_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_k"]);
+
+            auto add_q_proj = std::dynamic_pointer_cast<Linear>(blocks["add_q_proj"]);
+            auto add_k_proj = std::dynamic_pointer_cast<Linear>(blocks["add_k_proj"]);
+            auto add_v_proj = std::dynamic_pointer_cast<Linear>(blocks["add_v_proj"]);
+            auto to_add_out = std::dynamic_pointer_cast<Linear>(blocks["to_add_out"]);
+
+            int64_t N           = img->ne[2];
+            int64_t n_img_token = img->ne[1];
+            int64_t n_txt_token = txt->ne[1];
+
+            auto img_q        = to_q->forward(ctx, img);
+            int64_t num_heads = img_q->ne[0] / dim_head;
+            img_q             = ggml_reshape_4d(ctx, img_q, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+            auto img_k        = to_k->forward(ctx, img);
+            img_k             = ggml_reshape_4d(ctx, img_k, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+            auto img_v        = to_v->forward(ctx, img);
+            img_v             = ggml_reshape_4d(ctx, img_v, dim_head, num_heads, n_img_token, N);  // [N, n_img_token, n_head, d_head]
+
+            img_q = norm_q->forward(ctx, img_q);
+            img_k = norm_k->forward(ctx, img_k);
+
+            auto txt_q = add_q_proj->forward(ctx, txt);
+            txt_q      = ggml_reshape_4d(ctx, txt_q, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+            auto txt_k = add_k_proj->forward(ctx, txt);
+            txt_k      = ggml_reshape_4d(ctx, txt_k, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+            auto txt_v = add_v_proj->forward(ctx, txt);
+            txt_v      = ggml_reshape_4d(ctx, txt_v, dim_head, num_heads, n_txt_token, N);  // [N, n_txt_token, n_head, d_head]
+
+            txt_q = norm_added_q->forward(ctx, txt_q);
+            txt_k = norm_added_k->forward(ctx, txt_k);
+
+            auto q = ggml_concat(ctx, txt_q, img_q, 2);  // [N, n_txt_token + n_img_token, n_head, d_head]
+            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         = Rope::attention(ctx, backend, q, k, v, pe, mask, flash_attn, (1.0f / 128.f));  // [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,
+                                             attn->ne[0],
+                                             attn->ne[1],
+                                             txt->ne[1],
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             0);                                              // [n_txt_token, N, hidden_size]
+            txt_attn_out      = ggml_cont(ctx, ggml_permute(ctx, txt_attn_out, 0, 2, 1, 3));  // [N, n_txt_token, hidden_size]
+            auto img_attn_out = ggml_view_3d(ctx,
+                                             attn,
+                                             attn->ne[0],
+                                             attn->ne[1],
+                                             img->ne[1],
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             attn->nb[2] * txt->ne[1]);                       // [n_img_token, N, hidden_size]
+            img_attn_out      = ggml_cont(ctx, ggml_permute(ctx, img_attn_out, 0, 2, 1, 3));  // [N, n_img_token, hidden_size]
+
+            img_attn_out = to_out_0->forward(ctx, img_attn_out);
+            txt_attn_out = to_add_out->forward(ctx, txt_attn_out);
+
+            return {img_attn_out, txt_attn_out};
+        }
+    };
+
+    class QwenImageTransformerBlock : public GGMLBlock {
+    public:
+        QwenImageTransformerBlock(int64_t dim,
+                                  int64_t num_attention_heads,
+                                  int64_t attention_head_dim,
+                                  float eps       = 1e-6,
+                                  bool flash_attn = false) {
+            // img_mod.0 is nn.SiLU()
+            blocks["img_mod.1"] = std::shared_ptr<GGMLBlock>(new Linear(dim, 6 * dim, true));
+
+            blocks["img_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["img_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["img_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU, true));
+
+            // txt_mod.0 is nn.SiLU()
+            blocks["txt_mod.1"] = std::shared_ptr<GGMLBlock>(new Linear(dim, 6 * dim, true));
+
+            blocks["txt_norm1"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["txt_norm2"] = std::shared_ptr<GGMLBlock>(new LayerNorm(dim, eps, false));
+            blocks["txt_mlp"]   = std::shared_ptr<GGMLBlock>(new FeedForward(dim, dim, 4, FeedForward::Activation::GELU));
+
+            blocks["attn"] = std::shared_ptr<GGMLBlock>(new QwenImageAttention(dim,
+                                                                               attention_head_dim,
+                                                                               num_attention_heads,
+                                                                               0,     // out_dim
+                                                                               0,     // out_context-dim
+                                                                               true,  // bias
+                                                                               true,  // out_bias
+                                                                               eps,
+                                                                               flash_attn));
+        }
+
+        virtual std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,
+                                                              ggml_backend_t backend,
+                                                              struct ggml_tensor* img,
+                                                              struct ggml_tensor* txt,
+                                                              struct ggml_tensor* t_emb,
+                                                              struct ggml_tensor* pe) {
+            // 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_1 = std::dynamic_pointer_cast<Linear>(blocks["img_mod.1"]);
+            auto img_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm1"]);
+            auto img_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm2"]);
+            auto img_mlp   = std::dynamic_pointer_cast<FeedForward>(blocks["img_mlp"]);
+
+            auto txt_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["txt_mod.1"]);
+            auto txt_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm1"]);
+            auto txt_norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["txt_norm2"]);
+            auto txt_mlp   = std::dynamic_pointer_cast<FeedForward>(blocks["txt_mlp"]);
+
+            auto attn = std::dynamic_pointer_cast<QwenImageAttention>(blocks["attn"]);
+
+            auto img_mod_params    = ggml_silu(ctx, t_emb);
+            img_mod_params         = img_mod_1->forward(ctx, img_mod_params);
+            auto img_mod_param_vec = ggml_chunk(ctx, img_mod_params, 6, 0);
+
+            auto txt_mod_params    = ggml_silu(ctx, t_emb);
+            txt_mod_params         = txt_mod_1->forward(ctx, txt_mod_params);
+            auto txt_mod_param_vec = ggml_chunk(ctx, txt_mod_params, 6, 0);
+
+            auto img_normed    = img_norm1->forward(ctx, img);
+            auto img_modulated = Flux::modulate(ctx, img_normed, img_mod_param_vec[0], img_mod_param_vec[1]);
+            auto img_gate1     = img_mod_param_vec[2];
+
+            auto txt_normed    = txt_norm1->forward(ctx, txt);
+            auto txt_modulated = Flux::modulate(ctx, txt_normed, txt_mod_param_vec[0], txt_mod_param_vec[1]);
+            auto txt_gate1     = txt_mod_param_vec[2];
+
+            auto [img_attn_output, txt_attn_output] = attn->forward(ctx, backend, img_modulated, txt_modulated, pe);
+
+            img = ggml_add(ctx, img, ggml_mul(ctx, img_attn_output, img_gate1));
+            txt = ggml_add(ctx, txt, ggml_mul(ctx, txt_attn_output, txt_gate1));
+
+            auto img_normed2    = img_norm2->forward(ctx, img);
+            auto img_modulated2 = Flux::modulate(ctx, img_normed2, img_mod_param_vec[3], img_mod_param_vec[4]);
+            auto img_gate2      = img_mod_param_vec[5];
+
+            auto txt_normed2    = txt_norm2->forward(ctx, txt);
+            auto txt_modulated2 = Flux::modulate(ctx, txt_normed2, txt_mod_param_vec[3], txt_mod_param_vec[4]);
+            auto txt_gate2      = txt_mod_param_vec[5];
+
+            auto img_mlp_out = img_mlp->forward(ctx, img_modulated2);
+            auto txt_mlp_out = txt_mlp->forward(ctx, txt_modulated2);
+
+            img = ggml_add(ctx, img, ggml_mul(ctx, img_mlp_out, img_gate2));
+            txt = ggml_add(ctx, txt, ggml_mul(ctx, txt_mlp_out, txt_gate2));
+
+            return {img, txt};
+        }
+    };
+
+    struct AdaLayerNormContinuous : public GGMLBlock {
+    public:
+        AdaLayerNormContinuous(int64_t embedding_dim,
+                               int64_t conditioning_embedding_dim,
+                               bool elementwise_affine = true,
+                               float eps               = 1e-5f,
+                               bool bias               = true) {
+            blocks["norm"]   = std::shared_ptr<GGMLBlock>(new LayerNorm(conditioning_embedding_dim, eps, elementwise_affine, bias));
+            blocks["linear"] = std::shared_ptr<GGMLBlock>(new Linear(conditioning_embedding_dim, embedding_dim * 2, bias));
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* c) {
+            // x: [N, n_token, hidden_size]
+            // c: [N, hidden_size]
+            // return: [N, n_token, patch_size * patch_size * out_channels]
+
+            auto norm   = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+
+            auto emb   = linear->forward(ctx, ggml_silu(ctx, c));
+            auto mods  = ggml_chunk(ctx, emb, 2, 0);
+            auto scale = mods[0];
+            auto shift = mods[1];
+
+            x = norm->forward(ctx, x);
+            x = Flux::modulate(ctx, x, shift, scale);
+
+            return x;
+        }
+    };
+
+    struct QwenImageParams {
+        int64_t patch_size          = 2;
+        int64_t in_channels         = 64;
+        int64_t out_channels        = 16;
+        int64_t num_layers          = 60;
+        int64_t attention_head_dim  = 128;
+        int64_t num_attention_heads = 24;
+        int64_t joint_attention_dim = 3584;
+        float theta                 = 10000;
+        std::vector<int> axes_dim   = {16, 56, 56};
+        int64_t axes_dim_sum        = 128;
+        bool flash_attn             = false;
+    };
+
+    class QwenImageModel : public GGMLBlock {
+    protected:
+        QwenImageParams params;
+
+    public:
+        QwenImageModel() {}
+        QwenImageModel(QwenImageParams params)
+            : params(params) {
+            int64_t inner_dim         = params.num_attention_heads * params.attention_head_dim;
+            blocks["time_text_embed"] = std::shared_ptr<GGMLBlock>(new QwenTimestepProjEmbeddings(inner_dim));
+            blocks["txt_norm"]        = std::shared_ptr<GGMLBlock>(new RMSNorm(params.joint_attention_dim, 1e-6f));
+            blocks["img_in"]          = std::shared_ptr<GGMLBlock>(new Linear(params.in_channels, inner_dim));
+            blocks["txt_in"]          = std::shared_ptr<GGMLBlock>(new Linear(params.joint_attention_dim, inner_dim));
+
+            // blocks
+            for (int i = 0; i < params.num_layers; i++) {
+                auto block                                        = std::shared_ptr<GGMLBlock>(new QwenImageTransformerBlock(inner_dim,
+                                                                                                                             params.num_attention_heads,
+                                                                                                                             params.attention_head_dim,
+                                                                                                                             1e-6f,
+                                                                                                                             params.flash_attn));
+                blocks["transformer_blocks." + std::to_string(i)] = block;
+            }
+
+            blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new AdaLayerNormContinuous(inner_dim, inner_dim, false, 1e-6f));
+            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, params.patch_size * params.patch_size * params.out_channels));
+        }
+
+        struct ggml_tensor* pad_to_patch_size(struct ggml_context* ctx,
+                                              struct ggml_tensor* x) {
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+
+            int pad_h = (params.patch_size - H % params.patch_size) % params.patch_size;
+            int pad_w = (params.patch_size - W % params.patch_size) % params.patch_size;
+            x         = ggml_pad(ctx, x, pad_w, pad_h, 0, 0);  // [N, C, H + pad_h, W + pad_w]
+            return x;
+        }
+
+        struct ggml_tensor* patchify(struct ggml_context* ctx,
+                                     struct ggml_tensor* x) {
+            // x: [N, C, H, W]
+            // return: [N, h*w, C * patch_size * patch_size]
+            int64_t N = x->ne[3];
+            int64_t C = x->ne[2];
+            int64_t H = x->ne[1];
+            int64_t W = x->ne[0];
+            int64_t p = params.patch_size;
+            int64_t h = H / params.patch_size;
+            int64_t w = W / params.patch_size;
+
+            GGML_ASSERT(h * p == H && w * p == W);
+
+            x = ggml_reshape_4d(ctx, x, p, w, p, h * C * N);       // [N*C*h, p, w, p]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, w, p, p]
+            x = ggml_reshape_4d(ctx, x, p * p, w * h, C, N);       // [N, C, h*w, p*p]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, h*w, C, p*p]
+            x = ggml_reshape_3d(ctx, x, p * p * C, w * h, N);      // [N, h*w, C*p*p]
+            return x;
+        }
+
+        struct ggml_tensor* process_img(struct ggml_context* ctx,
+                                        struct ggml_tensor* x) {
+            x = pad_to_patch_size(ctx, x);
+            x = patchify(ctx, x);
+            return x;
+        }
+
+        struct ggml_tensor* unpatchify(struct ggml_context* ctx,
+                                       struct ggml_tensor* x,
+                                       int64_t h,
+                                       int64_t w) {
+            // x: [N, h*w, C*patch_size*patch_size]
+            // return: [N, C, H, W]
+            int64_t N = x->ne[2];
+            int64_t C = x->ne[0] / params.patch_size / params.patch_size;
+            int64_t H = h * params.patch_size;
+            int64_t W = w * params.patch_size;
+            int64_t p = params.patch_size;
+
+            GGML_ASSERT(C * p * p == x->ne[0]);
+
+            x = ggml_reshape_4d(ctx, x, p * p, C, w * h, N);       // [N, h*w, C, p*p]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, C, h*w, p*p]
+            x = ggml_reshape_4d(ctx, x, p, p, w, h * C * N);       // [N*C*h, w, p, p]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N*C*h, p, w, p]
+            x = ggml_reshape_4d(ctx, x, W, H, C, N);               // [N, C, h*p, w*p]
+
+            return x;
+        }
+
+        struct ggml_tensor* forward_orig(struct ggml_context* ctx,
+                                         ggml_backend_t backend,
+                                         struct ggml_tensor* x,
+                                         struct ggml_tensor* timestep,
+                                         struct ggml_tensor* context,
+                                         struct ggml_tensor* pe) {
+            auto time_text_embed = std::dynamic_pointer_cast<QwenTimestepProjEmbeddings>(blocks["time_text_embed"]);
+            auto txt_norm        = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm"]);
+            auto img_in          = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
+            auto txt_in          = std::dynamic_pointer_cast<Linear>(blocks["txt_in"]);
+            auto norm_out        = std::dynamic_pointer_cast<AdaLayerNormContinuous>(blocks["norm_out"]);
+            auto proj_out        = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
+
+            auto t_emb = time_text_embed->forward(ctx, timestep);
+            auto img   = img_in->forward(ctx, x);
+            auto txt   = txt_norm->forward(ctx, context);
+            txt        = txt_in->forward(ctx, txt);
+
+            for (int i = 0; i < params.num_layers; i++) {
+                auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
+
+                auto result = block->forward(ctx, backend, img, txt, t_emb, pe);
+                img         = result.first;
+                txt         = result.second;
+            }
+
+            img = norm_out->forward(ctx, img, t_emb);
+            img = proj_out->forward(ctx, img);
+
+            return img;
+        }
+
+        struct ggml_tensor* forward(struct ggml_context* ctx,
+                                    ggml_backend_t backend,
+                                    struct ggml_tensor* x,
+                                    struct ggml_tensor* timestep,
+                                    struct ggml_tensor* context,
+                                    struct ggml_tensor* pe,
+                                    std::vector<ggml_tensor*> ref_latents = {}) {
+            // Forward pass of DiT.
+            // x: [N, C, H, W]
+            // timestep: [N,]
+            // context: [N, L, D]
+            // pe: [L, d_head/2, 2, 2]
+            // return: [N, C, H, W]
+
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t C = x->ne[2];
+            int64_t N = x->ne[3];
+
+            auto img            = process_img(ctx, x);
+            uint64_t img_tokens = img->ne[1];
+
+            if (ref_latents.size() > 0) {
+                for (ggml_tensor* ref : ref_latents) {
+                    ref = process_img(ctx, ref);
+                    img = ggml_concat(ctx, img, ref, 1);
+                }
+            }
+
+            int64_t h_len = ((H + (params.patch_size / 2)) / params.patch_size);
+            int64_t w_len = ((W + (params.patch_size / 2)) / params.patch_size);
+
+            auto out = forward_orig(ctx, backend, img, timestep, context, pe);  // [N, h_len*w_len, ph*pw*C]
+
+            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]
+            }
+
+            out = unpatchify(ctx, out, h_len, w_len);  // [N, C, H + pad_h, W + pad_w]
+
+            // slice
+            out = ggml_slice(ctx, out, 1, 0, H);  // [N, C, H, W + pad_w]
+            out = ggml_slice(ctx, out, 0, 0, W);  // [N, C, H, W]
+
+            return out;
+        }
+    };
+
+    struct QwenImageRunner : public GGMLRunner {
+    public:
+        QwenImageParams qwen_image_params;
+        QwenImageModel qwen_image;
+        std::vector<float> pe_vec;
+        SDVersion version;
+
+        QwenImageRunner(ggml_backend_t backend,
+                        bool offload_params_to_cpu,
+                        const String2GGMLType& tensor_types = {},
+                        const std::string prefix            = "",
+                        SDVersion version                   = VERSION_QWEN_IMAGE,
+                        bool flash_attn                     = false)
+            : GGMLRunner(backend, offload_params_to_cpu) {
+            qwen_image_params.flash_attn = flash_attn;
+            qwen_image_params.num_layers = 0;
+            for (auto pair : tensor_types) {
+                std::string tensor_name = pair.first;
+                if (tensor_name.find(prefix) == std::string::npos)
+                    continue;
+                size_t pos = tensor_name.find("transformer_blocks.");
+                if (pos != std::string::npos) {
+                    tensor_name = tensor_name.substr(pos);  // remove prefix
+                    auto items  = split_string(tensor_name, '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        if (block_index + 1 > qwen_image_params.num_layers) {
+                            qwen_image_params.num_layers = block_index + 1;
+                        }
+                    }
+                    continue;
+                }
+            }
+            LOG_INFO("qwen_image_params.num_layers: %ld", qwen_image_params.num_layers);
+            qwen_image = QwenImageModel(qwen_image_params);
+            qwen_image.init(params_ctx, tensor_types, prefix);
+        }
+
+        std::string get_desc() override {
+            return "qwen_image";
+        }
+
+        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+            qwen_image.get_param_tensors(tensors, prefix);
+        }
+
+        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+                                        struct ggml_tensor* timesteps,
+                                        struct ggml_tensor* context,
+                                        std::vector<ggml_tensor*> ref_latents = {},
+                                        bool increase_ref_index               = false) {
+            GGML_ASSERT(x->ne[3] == 1);
+            struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, QWEN_IMAGE_GRAPH_SIZE, false);
+
+            x         = to_backend(x);
+            context   = to_backend(context);
+            timesteps = to_backend(timesteps);
+
+            for (int i = 0; i < ref_latents.size(); i++) {
+                ref_latents[i] = to_backend(ref_latents[i]);
+            }
+
+            pe_vec      = Rope::gen_qwen_image_pe(x->ne[1],
+                                                  x->ne[0],
+                                                  qwen_image_params.patch_size,
+                                                  x->ne[3],
+                                                  context->ne[1],
+                                                  ref_latents,
+                                                  increase_ref_index,
+                                                  qwen_image_params.theta,
+                                                  qwen_image_params.axes_dim);
+            int pos_len = pe_vec.size() / qwen_image_params.axes_dim_sum / 2;
+            // LOG_DEBUG("pos_len %d", pos_len);
+            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, qwen_image_params.axes_dim_sum / 2, pos_len);
+            // pe->data = pe_vec.data();
+            // print_ggml_tensor(pe, true, "pe");
+            // pe->data = nullptr;
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            struct ggml_tensor* out = qwen_image.forward(compute_ctx,
+                                                         runtime_backend,
+                                                         x,
+                                                         timesteps,
+                                                         context,
+                                                         pe,
+                                                         ref_latents);
+
+            ggml_build_forward_expand(gf, out);
+
+            return gf;
+        }
+
+        void compute(int n_threads,
+                     struct ggml_tensor* x,
+                     struct ggml_tensor* timesteps,
+                     struct ggml_tensor* context,
+                     std::vector<ggml_tensor*> ref_latents = {},
+                     bool increase_ref_index               = false,
+                     struct ggml_tensor** output           = nullptr,
+                     struct ggml_context* output_ctx       = nullptr) {
+            // x: [N, in_channels, h, w]
+            // timesteps: [N, ]
+            // context: [N, max_position, hidden_size]
+            auto get_graph = [&]() -> struct ggml_cgraph* {
+                return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
+            };
+
+            GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+        }
+
+        void test() {
+            struct ggml_init_params params;
+            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
+            params.mem_buffer = nullptr;
+            params.no_alloc   = false;
+
+            struct ggml_context* work_ctx = ggml_init(params);
+            GGML_ASSERT(work_ctx != nullptr);
+
+            {
+                // auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 16, 16, 16, 1);
+                // ggml_set_f32(x, 0.01f);
+                auto x = load_tensor_from_file(work_ctx, "./qwen_image_x.bin");
+                print_ggml_tensor(x);
+
+                std::vector<float> timesteps_vec(1, 1000.f);
+                auto timesteps = vector_to_ggml_tensor(work_ctx, timesteps_vec);
+
+                // auto context = ggml_new_tensor_3d(work_ctx, GGML_TYPE_F32, 3584, 256, 1);
+                // ggml_set_f32(context, 0.01f);
+                auto context = load_tensor_from_file(work_ctx, "./qwen_image_context.bin");
+                print_ggml_tensor(context);
+
+                struct ggml_tensor* out = nullptr;
+
+                int t0 = ggml_time_ms();
+                compute(8, x, timesteps, context, {}, false, &out, work_ctx);
+                int t1 = ggml_time_ms();
+
+                print_ggml_tensor(out);
+                LOG_DEBUG("qwen_image test done in %dms", t1 - t0);
+            }
+        }
+
+        static void load_from_file_and_test(const std::string& file_path) {
+            // cuda q8: pass
+            // cuda q8 fa: nan
+            // ggml_backend_t backend    = ggml_backend_cuda_init(0);
+            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_type model_data_type = GGML_TYPE_Q8_0;
+
+            ModelLoader model_loader;
+            if (!model_loader.init_from_file(file_path, "model.diffusion_model.")) {
+                LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
+                return;
+            }
+
+            auto tensor_types = model_loader.tensor_storages_types;
+            for (auto& item : tensor_types) {
+                // LOG_DEBUG("%s %u", item.first.c_str(), item.second);
+                if (ends_with(item.first, "weight")) {
+                    item.second = model_data_type;
+                }
+            }
+
+            std::shared_ptr<QwenImageRunner> qwen_image = std::make_shared<QwenImageRunner>(backend,
+                                                                                            false,
+                                                                                            tensor_types,
+                                                                                            "model.diffusion_model",
+                                                                                            VERSION_QWEN_IMAGE,
+                                                                                            true);
+
+            qwen_image->alloc_params_buffer();
+            std::map<std::string, ggml_tensor*> tensors;
+            qwen_image->get_param_tensors(tensors, "model.diffusion_model");
+
+            bool success = model_loader.load_tensors(tensors);
+
+            if (!success) {
+                LOG_ERROR("load tensors from model loader failed");
+                return;
+            }
+
+            LOG_INFO("qwen_image model loaded");
+            qwen_image->test();
+        }
+    };
+
+}  // namespace name
+
+#endif  // __QWEN_IMAGE_HPP__

rng.hpp

@@ -15,11 +15,11 @@ private:
     std::default_random_engine generator;
 
 public:
-    void manual_seed(uint64_t seed) {
+    void manual_seed(uint64_t seed) override {
         generator.seed((unsigned int)seed);
     }
 
-    std::vector<float> randn(uint32_t n) {
+    std::vector<float> randn(uint32_t n) override {
         std::vector<float> result;
         float mean   = 0.0;
         float stddev = 1.0;

rng_philox.hpp

@@ -93,12 +93,12 @@ public:
         this->offset = 0;
     }
 
-    void manual_seed(uint64_t seed) {
+    void manual_seed(uint64_t seed) override {
         this->seed   = seed;
         this->offset = 0;
     }
 
-    std::vector<float> randn(uint32_t n) {
+    std::vector<float> randn(uint32_t n) override {
         std::vector<std::vector<uint32_t>> counter(4, std::vector<uint32_t>(n, 0));
         for (uint32_t i = 0; i < n; i++) {
             counter[0][i] = this->offset;

rope.hpp

@@ -4,9 +4,9 @@
 #include <vector>
 #include "ggml_extend.hpp"
 
-struct Rope {
+namespace Rope {
     template <class T>
-    static std::vector<T> linspace(T start, T end, int num) {
+    __STATIC_INLINE__ std::vector<T> linspace(T start, T end, int num) {
         std::vector<T> result(num);
         if (num == 1) {
             result[0] = start;
@@ -19,7 +19,7 @@ struct Rope {
         return result;
     }
 
-    static std::vector<std::vector<float>> transpose(const std::vector<std::vector<float>>& mat) {
+    __STATIC_INLINE__ std::vector<std::vector<float>> transpose(const std::vector<std::vector<float>>& mat) {
         int rows = mat.size();
         int cols = mat[0].size();
         std::vector<std::vector<float>> transposed(cols, std::vector<float>(rows));
@@ -31,7 +31,7 @@ struct Rope {
         return transposed;
     }
 
-    static std::vector<float> flatten(const std::vector<std::vector<float>>& vec) {
+    __STATIC_INLINE__ std::vector<float> flatten(const std::vector<std::vector<float>>& vec) {
         std::vector<float> flat_vec;
         for (const auto& sub_vec : vec) {
             flat_vec.insert(flat_vec.end(), sub_vec.begin(), sub_vec.end());
@@ -39,7 +39,7 @@ struct Rope {
         return flat_vec;
     }
 
-    static std::vector<std::vector<float>> rope(const std::vector<float>& pos, int dim, int theta) {
+    __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos, int dim, int theta) {
         assert(dim % 2 == 0);
         int half_dim = dim / 2;
 
@@ -72,11 +72,11 @@ struct Rope {
     }
 
     // Generate IDs for image patches and text
-    static std::vector<std::vector<float>> gen_txt_ids(int bs, int context_len) {
+    __STATIC_INLINE__ 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));
     }
 
-    static 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) {
+    __STATIC_INLINE__ 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;
 
@@ -102,9 +102,9 @@ struct Rope {
         return img_ids_repeated;
     }
 
-    static std::vector<std::vector<float>> concat_ids(const std::vector<std::vector<float>>& a,
-                                                      const std::vector<std::vector<float>>& b,
-                                                      int bs) {
+    __STATIC_INLINE__ 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));
@@ -119,10 +119,10 @@ struct Rope {
         return ids;
     }
 
-    static std::vector<float> embed_nd(const std::vector<std::vector<float>>& ids,
-                                       int bs,
-                                       int theta,
-                                       const std::vector<int>& axes_dim) {
+    __STATIC_INLINE__ std::vector<float> embed_nd(const std::vector<std::vector<float>>& ids,
+                                                  int bs,
+                                                  int theta,
+                                                  const std::vector<int>& axes_dim) {
         std::vector<std::vector<float>> trans_ids = transpose(ids);
         size_t pos_len                            = ids.size() / bs;
         int num_axes                              = axes_dim.size();
@@ -151,17 +151,11 @@ struct Rope {
         return flatten(emb);
     }
 
-    static std::vector<std::vector<float>> gen_flux_ids(int h,
-                                                        int w,
-                                                        int patch_size,
-                                                        int bs,
-                                                        int context_len,
-                                                        std::vector<ggml_tensor*> ref_latents,
-                                                        bool increase_ref_index) {
-        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);
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_refs_ids(int patch_size,
+                                                                   int bs,
+                                                                   const std::vector<ggml_tensor*>& ref_latents,
+                                                                   bool increase_ref_index) {
+        std::vector<std::vector<float>> ids;
         uint64_t curr_h_offset = 0;
         uint64_t curr_w_offset = 0;
         int index              = 1;
@@ -189,30 +183,88 @@ struct Rope {
         return ids;
     }
 
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_flux_ids(int h,
+                                                                   int w,
+                                                                   int patch_size,
+                                                                   int bs,
+                                                                   int context_len,
+                                                                   const std::vector<ggml_tensor*>& ref_latents,
+                                                                   bool increase_ref_index) {
+        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);
+        if (ref_latents.size() > 0) {
+            auto refs_ids = gen_refs_ids(patch_size, bs, ref_latents, increase_ref_index);
+            ids           = concat_ids(ids, refs_ids, bs);
+        }
+        return ids;
+    }
+
     // Generate flux positional embeddings
-    static std::vector<float> gen_flux_pe(int h,
-                                          int w,
-                                          int patch_size,
-                                          int bs,
-                                          int context_len,
-                                          std::vector<ggml_tensor*> ref_latents,
-                                          bool increase_ref_index,
-                                          int theta,
-                                          const std::vector<int>& axes_dim) {
+    __STATIC_INLINE__ std::vector<float> gen_flux_pe(int h,
+                                                     int w,
+                                                     int patch_size,
+                                                     int bs,
+                                                     int context_len,
+                                                     const std::vector<ggml_tensor*>& ref_latents,
+                                                     bool increase_ref_index,
+                                                     int theta,
+                                                     const std::vector<int>& axes_dim) {
         std::vector<std::vector<float>> ids = gen_flux_ids(h, w, patch_size, bs, context_len, ref_latents, increase_ref_index);
         return embed_nd(ids, bs, theta, axes_dim);
     }
 
-    static std::vector<std::vector<float>> gen_vid_ids(int t,
-                                                       int h,
-                                                       int w,
-                                                       int pt,
-                                                       int ph,
-                                                       int pw,
-                                                       int bs,
-                                                       int t_offset = 0,
-                                                       int h_offset = 0,
-                                                       int w_offset = 0) {
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen_image_ids(int h,
+                                                                         int w,
+                                                                         int patch_size,
+                                                                         int bs,
+                                                                         int context_len,
+                                                                         const std::vector<ggml_tensor*>& ref_latents,
+                                                                         bool increase_ref_index) {
+        int h_len        = (h + (patch_size / 2)) / patch_size;
+        int w_len        = (w + (patch_size / 2)) / patch_size;
+        int txt_id_start = std::max(h_len, w_len);
+        auto txt_ids     = linspace<float>(txt_id_start, context_len + txt_id_start, context_len);
+        std::vector<std::vector<float>> txt_ids_repeated(bs * context_len, std::vector<float>(3));
+        for (int i = 0; i < bs; ++i) {
+            for (int j = 0; j < txt_ids.size(); ++j) {
+                txt_ids_repeated[i * txt_ids.size() + j] = {txt_ids[j], txt_ids[j], txt_ids[j]};
+            }
+        }
+        auto img_ids = gen_img_ids(h, w, patch_size, bs);
+        auto ids     = concat_ids(txt_ids_repeated, img_ids, bs);
+        if (ref_latents.size() > 0) {
+            auto refs_ids = gen_refs_ids(patch_size, bs, ref_latents, increase_ref_index);
+            ids           = concat_ids(ids, refs_ids, bs);
+        }
+        return ids;
+    }
+
+    // Generate qwen_image positional embeddings
+    __STATIC_INLINE__ std::vector<float> gen_qwen_image_pe(int h,
+                                                           int w,
+                                                           int patch_size,
+                                                           int bs,
+                                                           int context_len,
+                                                           const std::vector<ggml_tensor*>& ref_latents,
+                                                           bool increase_ref_index,
+                                                           int theta,
+                                                           const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_qwen_image_ids(h, w, patch_size, bs, context_len, ref_latents, increase_ref_index);
+        return embed_nd(ids, bs, theta, axes_dim);
+    }
+
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
+                                                                  int h,
+                                                                  int w,
+                                                                  int pt,
+                                                                  int ph,
+                                                                  int pw,
+                                                                  int bs,
+                                                                  int t_offset = 0,
+                                                                  int h_offset = 0,
+                                                                  int w_offset = 0) {
         int t_len = (t + (pt / 2)) / pt;
         int h_len = (h + (ph / 2)) / ph;
         int w_len = (w + (pw / 2)) / pw;
@@ -244,18 +296,115 @@ struct Rope {
     }
 
     // Generate wan positional embeddings
-    static std::vector<float> gen_wan_pe(int t,
-                                         int h,
-                                         int w,
-                                         int pt,
-                                         int ph,
-                                         int pw,
-                                         int bs,
-                                         int theta,
-                                         const std::vector<int>& axes_dim) {
+    __STATIC_INLINE__ std::vector<float> gen_wan_pe(int t,
+                                                    int h,
+                                                    int w,
+                                                    int pt,
+                                                    int ph,
+                                                    int pw,
+                                                    int bs,
+                                                    int theta,
+                                                    const std::vector<int>& axes_dim) {
         std::vector<std::vector<float>> ids = gen_vid_ids(t, h, w, pt, ph, pw, bs);
         return embed_nd(ids, bs, theta, axes_dim);
     }
-};  // struct Rope
+
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_qwen2vl_ids(int grid_h,
+                                                                      int grid_w,
+                                                                      int merge_size,
+                                                                      const std::vector<int>& window_index) {
+        std::vector<std::vector<float>> ids(grid_h * grid_w, std::vector<float>(2, 0.0));
+        int index = 0;
+        for (int ih = 0; ih < grid_h; ih += merge_size) {
+            for (int iw = 0; iw < grid_w; iw += merge_size) {
+                for (int iy = 0; iy < merge_size; iy++) {
+                    for (int ix = 0; ix < merge_size; ix++) {
+                        int inverse_index = window_index[index / (merge_size * merge_size)];
+                        int i             = inverse_index * (merge_size * merge_size) + index % (merge_size * merge_size);
+
+                        GGML_ASSERT(i < grid_h * grid_w);
+
+                        ids[i][0] = ih + iy;
+                        ids[i][1] = iw + ix;
+                        index++;
+                    }
+                }
+            }
+        }
+        return ids;
+    }
+
+    // Generate qwen2vl positional embeddings
+    __STATIC_INLINE__ std::vector<float> gen_qwen2vl_pe(int grid_h,
+                                                        int grid_w,
+                                                        int merge_size,
+                                                        const std::vector<int>& window_index,
+                                                        int theta,
+                                                        const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_qwen2vl_ids(grid_h, grid_w, merge_size, window_index);
+        return embed_nd(ids, 1, theta, axes_dim);
+    }
+
+    __STATIC_INLINE__ struct ggml_tensor* apply_rope(struct ggml_context* ctx,
+                                                     struct ggml_tensor* x,
+                                                     struct ggml_tensor* pe,
+                                                     bool rope_interleaved = true) {
+        // x: [N, L, n_head, d_head]
+        // pe: [L, d_head/2, 2, 2], [[cos, -sin], [sin, cos]]
+        int64_t d_head = x->ne[0];
+        int64_t n_head = x->ne[1];
+        int64_t L      = x->ne[2];
+        int64_t N      = x->ne[3];
+        x              = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));  // [N, n_head, L, d_head]
+        if (rope_interleaved) {
+            x = ggml_reshape_4d(ctx, x, 2, d_head / 2, L, n_head * N);  // [N * n_head, L, d_head/2, 2]
+            x = ggml_cont(ctx, ggml_permute(ctx, x, 3, 0, 1, 2));       // [2, N * n_head, L, d_head/2]
+        } else {
+            x = ggml_reshape_4d(ctx, x, d_head / 2, 2, L, n_head * N);   // [N * n_head, L, 2, d_head/2]
+            x = ggml_cont(ctx, ggml_torch_permute(ctx, x, 0, 2, 3, 1));  // [2, N * n_head, L, d_head/2]
+        }
+
+        int64_t offset = x->nb[2] * x->ne[2];
+        auto x_0       = ggml_view_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2], x->nb[1], x->nb[2], offset * 0);  // [N * n_head, L, d_head/2]
+        auto x_1       = ggml_view_3d(ctx, x, x->ne[0], x->ne[1], x->ne[2], x->nb[1], x->nb[2], offset * 1);  // [N * n_head, L, d_head/2]
+        x_0            = ggml_reshape_4d(ctx, x_0, 1, x_0->ne[0], x_0->ne[1], x_0->ne[2]);                    // [N * n_head, L, d_head/2, 1]
+        x_1            = ggml_reshape_4d(ctx, x_1, 1, x_1->ne[0], x_1->ne[1], x_1->ne[2]);                    // [N * n_head, L, d_head/2, 1]
+        auto temp_x    = ggml_new_tensor_4d(ctx, x_0->type, 2, x_0->ne[1], x_0->ne[2], x_0->ne[3]);
+        x_0            = ggml_repeat(ctx, x_0, temp_x);  // [N * n_head, L, d_head/2, 2]
+        x_1            = ggml_repeat(ctx, x_1, temp_x);  // [N * n_head, L, d_head/2, 2]
+
+        pe        = ggml_cont(ctx, ggml_permute(ctx, pe, 3, 0, 1, 2));  // [2, L, d_head/2, 2]
+        offset    = pe->nb[2] * pe->ne[2];
+        auto pe_0 = ggml_view_3d(ctx, pe, pe->ne[0], pe->ne[1], pe->ne[2], pe->nb[1], pe->nb[2], offset * 0);  // [L, d_head/2, 2]
+        auto pe_1 = ggml_view_3d(ctx, pe, pe->ne[0], pe->ne[1], pe->ne[2], pe->nb[1], pe->nb[2], offset * 1);  // [L, d_head/2, 2]
+
+        auto x_out = ggml_add_inplace(ctx, ggml_mul(ctx, x_0, pe_0), ggml_mul(ctx, x_1, pe_1));  // [N * n_head, L, d_head/2, 2]
+        if (!rope_interleaved) {
+            x_out = ggml_cont(ctx, ggml_permute(ctx, x_out, 1, 0, 2, 3));  // [N * n_head, L, x, d_head/2]
+        }
+        x_out = ggml_reshape_3d(ctx, x_out, d_head, L, n_head * N);  // [N*n_head, L, d_head]
+        return x_out;
+    }
+
+    __STATIC_INLINE__ struct ggml_tensor* attention(struct ggml_context* ctx,
+                                                    ggml_backend_t backend,
+                                                    struct ggml_tensor* q,
+                                                    struct ggml_tensor* k,
+                                                    struct ggml_tensor* v,
+                                                    struct ggml_tensor* pe,
+                                                    struct ggml_tensor* mask,
+                                                    bool flash_attn,
+                                                    float kv_scale        = 1.0f,
+                                                    bool rope_interleaved = true) {
+        // q,k,v: [N, L, n_head, d_head]
+        // pe: [L, d_head/2, 2, 2]
+        // return: [N, L, n_head*d_head]
+        q = apply_rope(ctx, q, pe, rope_interleaved);  // [N*n_head, L, d_head]
+        k = apply_rope(ctx, k, pe, rope_interleaved);  // [N*n_head, L, d_head]
+
+        auto x = ggml_nn_attention_ext(ctx, backend, q, k, v, v->ne[1], mask, false, true, flash_attn, kv_scale);  // [N, L, n_head*d_head]
+        return x;
+    }
+};  // namespace Rope
 
 #endif  // __ROPE_HPP__

stable-diffusion.h

@@ -35,7 +35,7 @@ enum rng_type_t {
 };
 
 enum sample_method_t {
-    EULER_A,
+    SAMPLE_METHOD_DEFAULT,
     EULER,
     HEUN,
     DPM2,
@@ -47,6 +47,7 @@ enum sample_method_t {
     LCM,
     DDIM_TRAILING,
     TCD,
+    EULER_A,
     SAMPLE_METHOD_COUNT
 };
 
@@ -57,9 +58,22 @@ enum scheduler_t {
     EXPONENTIAL,
     AYS,
     GITS,
+    SGM_UNIFORM,
+    SIMPLE,
+    SMOOTHSTEP,
     SCHEDULE_COUNT
 };
 
+enum prediction_t {
+    DEFAULT_PRED,
+    EPS_PRED,
+    V_PRED,
+    EDM_V_PRED,
+    SD3_FLOW_PRED,
+    FLUX_FLOW_PRED,
+    PREDICTION_COUNT
+};
+
 // same as enum ggml_type
 enum sd_type_t {
     SD_TYPE_F32  = 0,
@@ -112,12 +126,23 @@ enum sd_log_level_t {
     SD_LOG_ERROR
 };
 
+typedef struct {
+    bool enabled;
+    int tile_size_x;
+    int tile_size_y;
+    float target_overlap;
+    float rel_size_x;
+    float rel_size_y;
+} sd_tiling_params_t;
+
 typedef struct {
     const char* model_path;
     const char* clip_l_path;
     const char* clip_g_path;
     const char* clip_vision_path;
     const char* t5xxl_path;
+    const char* qwen2vl_path;
+    const char* qwen2vl_vision_path;
     const char* diffusion_model_path;
     const char* high_noise_diffusion_model_path;
     const char* vae_path;
@@ -125,13 +150,13 @@ typedef struct {
     const char* control_net_path;
     const char* lora_model_dir;
     const char* embedding_dir;
-    const char* stacked_id_embed_dir;
+    const char* photo_maker_path;
     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 prediction_t prediction;
     bool offload_params_to_cpu;
     bool keep_clip_on_cpu;
     bool keep_control_net_on_cpu;
@@ -139,6 +164,7 @@ typedef struct {
     bool diffusion_flash_attn;
     bool diffusion_conv_direct;
     bool vae_conv_direct;
+    bool force_sdxl_vae_conv_scale;
     bool chroma_use_dit_mask;
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
@@ -173,8 +199,16 @@ typedef struct {
     enum sample_method_t sample_method;
     int sample_steps;
     float eta;
+    int shifted_timestep;
 } sd_sample_params_t;
 
+typedef struct {
+    sd_image_t* id_images;
+    int id_images_count;
+    const char* id_embed_path;
+    float style_strength;
+} sd_pm_params_t;  // photo maker
+
 typedef struct {
     const char* prompt;
     const char* negative_prompt;
@@ -182,6 +216,7 @@ typedef struct {
     sd_image_t init_image;
     sd_image_t* ref_images;
     int ref_images_count;
+    bool auto_resize_ref_image;
     bool increase_ref_index;
     sd_image_t mask_image;
     int width;
@@ -192,9 +227,8 @@ typedef struct {
     int batch_count;
     sd_image_t control_image;
     float control_strength;
-    float style_strength;
-    bool normalize_input;
-    const char* input_id_images_path;
+    sd_pm_params_t pm_params;
+    sd_tiling_params_t vae_tiling_params;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -203,6 +237,8 @@ typedef struct {
     int clip_skip;
     sd_image_t init_image;
     sd_image_t end_image;
+    sd_image_t* control_frames;
+    int control_frames_size;
     int width;
     int height;
     sd_sample_params_t sample_params;
@@ -211,6 +247,7 @@ typedef struct {
     float strength;
     int64_t seed;
     int video_frames;
+    float vace_strength;
 } sd_vid_gen_params_t;
 
 typedef struct sd_ctx_t sd_ctx_t;
@@ -231,12 +268,15 @@ 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 scheduler_t scheduler);
 SD_API enum scheduler_t str_to_schedule(const char* str);
+SD_API const char* sd_prediction_name(enum prediction_t prediction);
+SD_API enum prediction_t str_to_prediction(const char* str);
 
 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 enum sample_method_t sd_get_default_sample_method(const sd_ctx_t* sd_ctx);
 
 SD_API void sd_sample_params_init(sd_sample_params_t* sample_params);
 SD_API char* sd_sample_params_to_str(const sd_sample_params_t* sample_params);
@@ -260,20 +300,20 @@ SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx,
                           sd_image_t input_image,
                           uint32_t upscale_factor);
 
+SD_API int get_upscale_factor(upscaler_ctx_t* upscaler_ctx);
+
 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,
-                                 int height,
-                                 float high_threshold,
-                                 float low_threshold,
-                                 float weak,
-                                 float strong,
-                                 bool inverse);
+SD_API bool preprocess_canny(sd_image_t image,
+                             float high_threshold,
+                             float low_threshold,
+                             float weak,
+                             float strong,
+                             bool inverse);
 
 #ifdef __cplusplus
 }

t5.hpp

@@ -1,7 +1,7 @@
 #ifndef __T5_HPP__
 #define __T5_HPP__
 
-#include <float.h>
+#include <cfloat>
 #include <limits>
 #include <map>
 #include <memory>
@@ -461,7 +461,7 @@ protected:
     int64_t hidden_size;
     float eps;
 
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
         enum ggml_type wtype = GGML_TYPE_F32;
         params["weight"]     = ggml_new_tensor_1d(ctx, wtype, hidden_size);
     }
@@ -472,7 +472,7 @@ public:
         : hidden_size(hidden_size),
           eps(eps) {}
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         struct ggml_tensor* w = params["weight"];
         x                     = ggml_rms_norm(ctx, x, eps);
         x                     = ggml_mul(ctx, x, w);
@@ -487,7 +487,7 @@ public:
         blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [N, n_token, model_dim]
         auto wi = std::dynamic_pointer_cast<Linear>(blocks["wi"]);
         auto wo = std::dynamic_pointer_cast<Linear>(blocks["wo"]);
@@ -504,10 +504,12 @@ public:
     T5DenseGatedActDense(int64_t model_dim, int64_t ff_dim) {
         blocks["wi_0"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
         blocks["wi_1"] = std::shared_ptr<GGMLBlock>(new Linear(model_dim, ff_dim, false));
-        blocks["wo"]   = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false));
+        float scale    = 1.f / 32.f;
+        // The purpose of the scale here is to prevent NaN issues on some backends(CUDA, ...).
+        blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false, false, false, scale));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [N, n_token, model_dim]
         auto wi_0 = std::dynamic_pointer_cast<Linear>(blocks["wi_0"]);
         auto wi_1 = std::dynamic_pointer_cast<Linear>(blocks["wi_1"]);
@@ -528,7 +530,7 @@ public:
         blocks["layer_norm"]     = std::shared_ptr<GGMLBlock>(new T5LayerNorm(model_dim));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [N, n_token, model_dim]
         auto DenseReluDense = std::dynamic_pointer_cast<T5DenseGatedActDense>(blocks["DenseReluDense"]);
         auto layer_norm     = std::dynamic_pointer_cast<T5LayerNorm>(blocks["layer_norm"]);
@@ -580,9 +582,9 @@ public:
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
                                                                 ggml_backend_t backend,
                                                                 struct ggml_tensor* x,
-                                                                struct ggml_tensor* past_bias                = NULL,
-                                                                struct ggml_tensor* mask                     = NULL,
-                                                                struct ggml_tensor* relative_position_bucket = NULL) {
+                                                                struct ggml_tensor* past_bias                = nullptr,
+                                                                struct ggml_tensor* mask                     = nullptr,
+                                                                struct ggml_tensor* relative_position_bucket = nullptr) {
         auto q_proj   = std::dynamic_pointer_cast<Linear>(blocks["q"]);
         auto k_proj   = std::dynamic_pointer_cast<Linear>(blocks["k"]);
         auto v_proj   = std::dynamic_pointer_cast<Linear>(blocks["v"]);
@@ -595,11 +597,11 @@ public:
         auto k = k_proj->forward(ctx, x);
         auto v = v_proj->forward(ctx, x);
 
-        if (using_relative_attention_bias && relative_position_bucket != NULL) {
+        if (using_relative_attention_bias && relative_position_bucket != nullptr) {
             past_bias = compute_bias(ctx, relative_position_bucket);
         }
-        if (past_bias != NULL) {
-            if (mask != NULL) {
+        if (past_bias != nullptr) {
+            if (mask != nullptr) {
                 mask = ggml_repeat(ctx, mask, past_bias);
                 mask = ggml_add(ctx, mask, past_bias);
             } else {
@@ -630,9 +632,9 @@ public:
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
                                                                 ggml_backend_t backend,
                                                                 struct ggml_tensor* x,
-                                                                struct ggml_tensor* past_bias                = NULL,
-                                                                struct ggml_tensor* mask                     = NULL,
-                                                                struct ggml_tensor* relative_position_bucket = NULL) {
+                                                                struct ggml_tensor* past_bias                = nullptr,
+                                                                struct ggml_tensor* mask                     = nullptr,
+                                                                struct ggml_tensor* relative_position_bucket = nullptr) {
         // x: [N, n_token, model_dim]
         auto SelfAttention = std::dynamic_pointer_cast<T5Attention>(blocks["SelfAttention"]);
         auto layer_norm    = std::dynamic_pointer_cast<T5LayerNorm>(blocks["layer_norm"]);
@@ -657,9 +659,9 @@ public:
     std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(struct ggml_context* ctx,
                                                                 ggml_backend_t backend,
                                                                 struct ggml_tensor* x,
-                                                                struct ggml_tensor* past_bias                = NULL,
-                                                                struct ggml_tensor* mask                     = NULL,
-                                                                struct ggml_tensor* relative_position_bucket = NULL) {
+                                                                struct ggml_tensor* past_bias                = nullptr,
+                                                                struct ggml_tensor* mask                     = nullptr,
+                                                                struct ggml_tensor* relative_position_bucket = nullptr) {
         // x: [N, n_token, model_dim]
         auto layer_0 = std::dynamic_pointer_cast<T5LayerSelfAttention>(blocks["layer.0"]);
         auto layer_1 = std::dynamic_pointer_cast<T5LayerFF>(blocks["layer.1"]);
@@ -693,9 +695,9 @@ public:
     struct ggml_tensor* forward(struct ggml_context* ctx,
                                 ggml_backend_t backend,
                                 struct ggml_tensor* x,
-                                struct ggml_tensor* past_bias                = NULL,
-                                struct ggml_tensor* attention_mask           = NULL,
-                                struct ggml_tensor* relative_position_bucket = NULL) {
+                                struct ggml_tensor* past_bias                = nullptr,
+                                struct ggml_tensor* attention_mask           = nullptr,
+                                struct ggml_tensor* relative_position_bucket = nullptr) {
         // x: [N, n_token, model_dim]
         for (int i = 0; i < num_layers; i++) {
             auto block = std::dynamic_pointer_cast<T5Block>(blocks["block." + std::to_string(i)]);
@@ -741,9 +743,9 @@ public:
     struct ggml_tensor* forward(struct ggml_context* ctx,
                                 ggml_backend_t backend,
                                 struct ggml_tensor* input_ids,
-                                struct ggml_tensor* past_bias                = NULL,
-                                struct ggml_tensor* attention_mask           = NULL,
-                                struct ggml_tensor* relative_position_bucket = NULL) {
+                                struct ggml_tensor* past_bias                = nullptr,
+                                struct ggml_tensor* attention_mask           = nullptr,
+                                struct ggml_tensor* relative_position_bucket = nullptr) {
         // input_ids: [N, n_token]
 
         auto shared  = std::dynamic_pointer_cast<Embedding>(blocks["shared"]);
@@ -774,7 +776,7 @@ struct T5Runner : public GGMLRunner {
         model.init(params_ctx, tensor_types, prefix);
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "t5";
     }
 
@@ -786,16 +788,16 @@ struct T5Runner : public GGMLRunner {
                                 ggml_backend_t backend,
                                 struct ggml_tensor* input_ids,
                                 struct ggml_tensor* relative_position_bucket,
-                                struct ggml_tensor* attention_mask = NULL) {
+                                struct ggml_tensor* attention_mask = nullptr) {
         size_t N       = input_ids->ne[1];
         size_t n_token = input_ids->ne[0];
 
-        auto hidden_states = model.forward(ctx, backend, input_ids, NULL, attention_mask, relative_position_bucket);  // [N, n_token, model_dim]
+        auto hidden_states = model.forward(ctx, backend, input_ids, nullptr, 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_tensor* attention_mask = NULL) {
+                                    struct ggml_tensor* attention_mask = nullptr) {
         struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         input_ids      = to_backend(input_ids);
@@ -827,7 +829,7 @@ struct T5Runner : public GGMLRunner {
                  struct ggml_tensor* input_ids,
                  struct ggml_tensor* attention_mask,
                  ggml_tensor** output,
-                 ggml_context* output_ctx = NULL) {
+                 ggml_context* output_ctx = nullptr) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
             return build_graph(input_ids, attention_mask);
         };
@@ -966,11 +968,11 @@ struct T5Embedder {
     void test() {
         struct ggml_init_params params;
         params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-        params.mem_buffer = NULL;
+        params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
         struct ggml_context* work_ctx = ggml_init(params);
-        GGML_ASSERT(work_ctx != NULL);
+        GGML_ASSERT(work_ctx != nullptr);
 
         {
             std::string text("a lovely cat");
@@ -985,7 +987,7 @@ struct T5Embedder {
             printf("\n");
             auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
             auto attention_mask     = vector_to_ggml_tensor(work_ctx, masks);
-            struct ggml_tensor* out = NULL;
+            struct ggml_tensor* out = nullptr;
 
             int t0 = ggml_time_ms();
             model.compute(8, input_ids, attention_mask, &out, work_ctx);
@@ -1020,7 +1022,7 @@ struct T5Embedder {
             }
         }
 
-        std::shared_ptr<T5Embedder> t5 = std::shared_ptr<T5Embedder>(new T5Embedder(backend, false, tensor_types, "", true));
+        std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, false, tensor_types, "", true);
 
         t5->alloc_params_buffer();
         std::map<std::string, ggml_tensor*> tensors;

tae.hpp

@@ -29,7 +29,7 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [n, n_in, h, w]
         // return: [n, n_out, h, w]
 
@@ -86,7 +86,7 @@ public:
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, z_channels, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [n, in_channels, h, w]
         // return: [n, z_channels, h/8, w/8]
 
@@ -136,7 +136,7 @@ public:
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* z) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* z) override {
         // z: [n, z_channels, h, w]
         // return: [n, out_channels, h*8, w*8]
 
@@ -218,11 +218,11 @@ struct TinyAutoEncoder : public GGMLRunner {
         }
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "taesd";
     }
 
-    bool load_from_file(const std::string& file_path) {
+    bool load_from_file(const std::string& file_path, int n_threads) {
         LOG_INFO("loading taesd from '%s', decode_only = %s", file_path.c_str(), decode_only ? "true" : "false");
         alloc_params_buffer();
         std::map<std::string, ggml_tensor*> taesd_tensors;
@@ -238,7 +238,7 @@ struct TinyAutoEncoder : public GGMLRunner {
             return false;
         }
 
-        bool success = model_loader.load_tensors(taesd_tensors, ignore_tensors);
+        bool success = model_loader.load_tensors(taesd_tensors, ignore_tensors, n_threads);
 
         if (!success) {
             LOG_ERROR("load tae tensors from model loader failed");
@@ -261,7 +261,7 @@ struct TinyAutoEncoder : public GGMLRunner {
                  struct ggml_tensor* z,
                  bool decode_graph,
                  struct ggml_tensor** output,
-                 struct ggml_context* output_ctx = NULL) {
+                 struct ggml_context* output_ctx = nullptr) {
         auto get_graph = [&]() -> struct ggml_cgraph* {
             return build_graph(z, decode_graph);
         };

tokenize_util.cpp

@@ -0,0 +1,985 @@
+#include <algorithm>
+#include <iostream>
+#include <string>
+#include <vector>
+
+#include "tokenize_util.h"
+
+bool is_number(char32_t ch) {
+    return (ch >= U'0' && ch <= U'9');
+}
+
+bool is_letter(char32_t ch) {
+    static const struct { char32_t start, end; } ranges[] = {
+        {0x41, 0x5A},
+        {0x61, 0x7A},
+        {0xAA, 0xAA},
+        {0xB5, 0xB5},
+        {0xBA, 0xBA},
+        {0xC0, 0xD6},
+        {0xD8, 0xF6},
+        {0xF8, 0x2C1},
+        {0x2C6, 0x2D1},
+        {0x2E0, 0x2E4},
+        {0x2EC, 0x2EC},
+        {0x2EE, 0x2EE},
+        {0x370, 0x374},
+        {0x376, 0x377},
+        {0x37A, 0x37D},
+        {0x37F, 0x37F},
+        {0x386, 0x386},
+        {0x388, 0x38A},
+        {0x38C, 0x38C},
+        {0x38E, 0x3A1},
+        {0x3A3, 0x3F5},
+        {0x3F7, 0x481},
+        {0x48A, 0x52F},
+        {0x531, 0x556},
+        {0x559, 0x559},
+        {0x560, 0x588},
+        {0x5D0, 0x5EA},
+        {0x5EF, 0x5F2},
+        {0x620, 0x64A},
+        {0x66E, 0x66F},
+        {0x671, 0x6D3},
+        {0x6D5, 0x6D5},
+        {0x6E5, 0x6E6},
+        {0x6EE, 0x6EF},
+        {0x6FA, 0x6FC},
+        {0x6FF, 0x6FF},
+        {0x710, 0x710},
+        {0x712, 0x72F},
+        {0x74D, 0x7A5},
+        {0x7B1, 0x7B1},
+        {0x7CA, 0x7EA},
+        {0x7F4, 0x7F5},
+        {0x7FA, 0x7FA},
+        {0x800, 0x815},
+        {0x81A, 0x81A},
+        {0x824, 0x824},
+        {0x828, 0x828},
+        {0x840, 0x858},
+        {0x860, 0x86A},
+        {0x870, 0x887},
+        {0x889, 0x88F},
+        {0x8A0, 0x8C9},
+        {0x904, 0x939},
+        {0x93D, 0x93D},
+        {0x950, 0x950},
+        {0x958, 0x961},
+        {0x971, 0x980},
+        {0x985, 0x98C},
+        {0x98F, 0x990},
+        {0x993, 0x9A8},
+        {0x9AA, 0x9B0},
+        {0x9B2, 0x9B2},
+        {0x9B6, 0x9B9},
+        {0x9BD, 0x9BD},
+        {0x9CE, 0x9CE},
+        {0x9DC, 0x9DD},
+        {0x9DF, 0x9E1},
+        {0x9F0, 0x9F1},
+        {0x9FC, 0x9FC},
+        {0xA05, 0xA0A},
+        {0xA0F, 0xA10},
+        {0xA13, 0xA28},
+        {0xA2A, 0xA30},
+        {0xA32, 0xA33},
+        {0xA35, 0xA36},
+        {0xA38, 0xA39},
+        {0xA59, 0xA5C},
+        {0xA5E, 0xA5E},
+        {0xA72, 0xA74},
+        {0xA85, 0xA8D},
+        {0xA8F, 0xA91},
+        {0xA93, 0xAA8},
+        {0xAAA, 0xAB0},
+        {0xAB2, 0xAB3},
+        {0xAB5, 0xAB9},
+        {0xABD, 0xABD},
+        {0xAD0, 0xAD0},
+        {0xAE0, 0xAE1},
+        {0xAF9, 0xAF9},
+        {0xB05, 0xB0C},
+        {0xB0F, 0xB10},
+        {0xB13, 0xB28},
+        {0xB2A, 0xB30},
+        {0xB32, 0xB33},
+        {0xB35, 0xB39},
+        {0xB3D, 0xB3D},
+        {0xB5C, 0xB5D},
+        {0xB5F, 0xB61},
+        {0xB71, 0xB71},
+        {0xB83, 0xB83},
+        {0xB85, 0xB8A},
+        {0xB8E, 0xB90},
+        {0xB92, 0xB95},
+        {0xB99, 0xB9A},
+        {0xB9C, 0xB9C},
+        {0xB9E, 0xB9F},
+        {0xBA3, 0xBA4},
+        {0xBA8, 0xBAA},
+        {0xBAE, 0xBB9},
+        {0xBD0, 0xBD0},
+        {0xC05, 0xC0C},
+        {0xC0E, 0xC10},
+        {0xC12, 0xC28},
+        {0xC2A, 0xC39},
+        {0xC3D, 0xC3D},
+        {0xC58, 0xC5A},
+        {0xC5C, 0xC5D},
+        {0xC60, 0xC61},
+        {0xC80, 0xC80},
+        {0xC85, 0xC8C},
+        {0xC8E, 0xC90},
+        {0xC92, 0xCA8},
+        {0xCAA, 0xCB3},
+        {0xCB5, 0xCB9},
+        {0xCBD, 0xCBD},
+        {0xCDC, 0xCDE},
+        {0xCE0, 0xCE1},
+        {0xCF1, 0xCF2},
+        {0xD04, 0xD0C},
+        {0xD0E, 0xD10},
+        {0xD12, 0xD3A},
+        {0xD3D, 0xD3D},
+        {0xD4E, 0xD4E},
+        {0xD54, 0xD56},
+        {0xD5F, 0xD61},
+        {0xD7A, 0xD7F},
+        {0xD85, 0xD96},
+        {0xD9A, 0xDB1},
+        {0xDB3, 0xDBB},
+        {0xDBD, 0xDBD},
+        {0xDC0, 0xDC6},
+        {0xE01, 0xE30},
+        {0xE32, 0xE33},
+        {0xE40, 0xE46},
+        {0xE81, 0xE82},
+        {0xE84, 0xE84},
+        {0xE86, 0xE8A},
+        {0xE8C, 0xEA3},
+        {0xEA5, 0xEA5},
+        {0xEA7, 0xEB0},
+        {0xEB2, 0xEB3},
+        {0xEBD, 0xEBD},
+        {0xEC0, 0xEC4},
+        {0xEC6, 0xEC6},
+        {0xEDC, 0xEDF},
+        {0xF00, 0xF00},
+        {0xF40, 0xF47},
+        {0xF49, 0xF6C},
+        {0xF88, 0xF8C},
+        {0x1000, 0x102A},
+        {0x103F, 0x103F},
+        {0x1050, 0x1055},
+        {0x105A, 0x105D},
+        {0x1061, 0x1061},
+        {0x1065, 0x1066},
+        {0x106E, 0x1070},
+        {0x1075, 0x1081},
+        {0x108E, 0x108E},
+        {0x10A0, 0x10C5},
+        {0x10C7, 0x10C7},
+        {0x10CD, 0x10CD},
+        {0x10D0, 0x10FA},
+        {0x10FC, 0x1248},
+        {0x124A, 0x124D},
+        {0x1250, 0x1256},
+        {0x1258, 0x1258},
+        {0x125A, 0x125D},
+        {0x1260, 0x1288},
+        {0x128A, 0x128D},
+        {0x1290, 0x12B0},
+        {0x12B2, 0x12B5},
+        {0x12B8, 0x12BE},
+        {0x12C0, 0x12C0},
+        {0x12C2, 0x12C5},
+        {0x12C8, 0x12D6},
+        {0x12D8, 0x1310},
+        {0x1312, 0x1315},
+        {0x1318, 0x135A},
+        {0x1380, 0x138F},
+        {0x13A0, 0x13F5},
+        {0x13F8, 0x13FD},
+        {0x1401, 0x166C},
+        {0x166F, 0x167F},
+        {0x1681, 0x169A},
+        {0x16A0, 0x16EA},
+        {0x16F1, 0x16F8},
+        {0x1700, 0x1711},
+        {0x171F, 0x1731},
+        {0x1740, 0x1751},
+        {0x1760, 0x176C},
+        {0x176E, 0x1770},
+        {0x1780, 0x17B3},
+        {0x17D7, 0x17D7},
+        {0x17DC, 0x17DC},
+        {0x1820, 0x1878},
+        {0x1880, 0x1884},
+        {0x1887, 0x18A8},
+        {0x18AA, 0x18AA},
+        {0x18B0, 0x18F5},
+        {0x1900, 0x191E},
+        {0x1950, 0x196D},
+        {0x1970, 0x1974},
+        {0x1980, 0x19AB},
+        {0x19B0, 0x19C9},
+        {0x1A00, 0x1A16},
+        {0x1A20, 0x1A54},
+        {0x1AA7, 0x1AA7},
+        {0x1B05, 0x1B33},
+        {0x1B45, 0x1B4C},
+        {0x1B83, 0x1BA0},
+        {0x1BAE, 0x1BAF},
+        {0x1BBA, 0x1BE5},
+        {0x1C00, 0x1C23},
+        {0x1C4D, 0x1C4F},
+        {0x1C5A, 0x1C7D},
+        {0x1C80, 0x1C8A},
+        {0x1C90, 0x1CBA},
+        {0x1CBD, 0x1CBF},
+        {0x1CE9, 0x1CEC},
+        {0x1CEE, 0x1CF3},
+        {0x1CF5, 0x1CF6},
+        {0x1CFA, 0x1CFA},
+        {0x1D00, 0x1DBF},
+        {0x1E00, 0x1F15},
+        {0x1F18, 0x1F1D},
+        {0x1F20, 0x1F45},
+        {0x1F48, 0x1F4D},
+        {0x1F50, 0x1F57},
+        {0x1F59, 0x1F59},
+        {0x1F5B, 0x1F5B},
+        {0x1F5D, 0x1F5D},
+        {0x1F5F, 0x1F7D},
+        {0x1F80, 0x1FB4},
+        {0x1FB6, 0x1FBC},
+        {0x1FBE, 0x1FBE},
+        {0x1FC2, 0x1FC4},
+        {0x1FC6, 0x1FCC},
+        {0x1FD0, 0x1FD3},
+        {0x1FD6, 0x1FDB},
+        {0x1FE0, 0x1FEC},
+        {0x1FF2, 0x1FF4},
+        {0x1FF6, 0x1FFC},
+        {0x2071, 0x2071},
+        {0x207F, 0x207F},
+        {0x2090, 0x209C},
+        {0x2102, 0x2102},
+        {0x2107, 0x2107},
+        {0x210A, 0x2113},
+        {0x2115, 0x2115},
+        {0x2119, 0x211D},
+        {0x2124, 0x2124},
+        {0x2126, 0x2126},
+        {0x2128, 0x2128},
+        {0x212A, 0x212D},
+        {0x212F, 0x2139},
+        {0x213C, 0x213F},
+        {0x2145, 0x2149},
+        {0x214E, 0x214E},
+        {0x2183, 0x2184},
+        {0x2C00, 0x2CE4},
+        {0x2CEB, 0x2CEE},
+        {0x2CF2, 0x2CF3},
+        {0x2D00, 0x2D25},
+        {0x2D27, 0x2D27},
+        {0x2D2D, 0x2D2D},
+        {0x2D30, 0x2D67},
+        {0x2D6F, 0x2D6F},
+        {0x2D80, 0x2D96},
+        {0x2DA0, 0x2DA6},
+        {0x2DA8, 0x2DAE},
+        {0x2DB0, 0x2DB6},
+        {0x2DB8, 0x2DBE},
+        {0x2DC0, 0x2DC6},
+        {0x2DC8, 0x2DCE},
+        {0x2DD0, 0x2DD6},
+        {0x2DD8, 0x2DDE},
+        {0x2E2F, 0x2E2F},
+        {0x3005, 0x3006},
+        {0x3031, 0x3035},
+        {0x303B, 0x303C},
+        {0x3041, 0x3096},
+        {0x309D, 0x309F},
+        {0x30A1, 0x30FA},
+        {0x30FC, 0x30FF},
+        {0x3105, 0x312F},
+        {0x3131, 0x318E},
+        {0x31A0, 0x31BF},
+        {0x31F0, 0x31FF},
+        {0x3400, 0x4DBF},
+        {0x4E00, 0xA48C},
+        {0xA4D0, 0xA4FD},
+        {0xA500, 0xA60C},
+        {0xA610, 0xA61F},
+        {0xA62A, 0xA62B},
+        {0xA640, 0xA66E},
+        {0xA67F, 0xA69D},
+        {0xA6A0, 0xA6E5},
+        {0xA717, 0xA71F},
+        {0xA722, 0xA788},
+        {0xA78B, 0xA7DC},
+        {0xA7F1, 0xA801},
+        {0xA803, 0xA805},
+        {0xA807, 0xA80A},
+        {0xA80C, 0xA822},
+        {0xA840, 0xA873},
+        {0xA882, 0xA8B3},
+        {0xA8F2, 0xA8F7},
+        {0xA8FB, 0xA8FB},
+        {0xA8FD, 0xA8FE},
+        {0xA90A, 0xA925},
+        {0xA930, 0xA946},
+        {0xA960, 0xA97C},
+        {0xA984, 0xA9B2},
+        {0xA9CF, 0xA9CF},
+        {0xA9E0, 0xA9E4},
+        {0xA9E6, 0xA9EF},
+        {0xA9FA, 0xA9FE},
+        {0xAA00, 0xAA28},
+        {0xAA40, 0xAA42},
+        {0xAA44, 0xAA4B},
+        {0xAA60, 0xAA76},
+        {0xAA7A, 0xAA7A},
+        {0xAA7E, 0xAAAF},
+        {0xAAB1, 0xAAB1},
+        {0xAAB5, 0xAAB6},
+        {0xAAB9, 0xAABD},
+        {0xAAC0, 0xAAC0},
+        {0xAAC2, 0xAAC2},
+        {0xAADB, 0xAADD},
+        {0xAAE0, 0xAAEA},
+        {0xAAF2, 0xAAF4},
+        {0xAB01, 0xAB06},
+        {0xAB09, 0xAB0E},
+        {0xAB11, 0xAB16},
+        {0xAB20, 0xAB26},
+        {0xAB28, 0xAB2E},
+        {0xAB30, 0xAB5A},
+        {0xAB5C, 0xAB69},
+        {0xAB70, 0xABE2},
+        {0xAC00, 0xD7A3},
+        {0xD7B0, 0xD7C6},
+        {0xD7CB, 0xD7FB},
+        {0xF900, 0xFA6D},
+        {0xFA70, 0xFAD9},
+        {0xFB00, 0xFB06},
+        {0xFB13, 0xFB17},
+        {0xFB1D, 0xFB1D},
+        {0xFB1F, 0xFB28},
+        {0xFB2A, 0xFB36},
+        {0xFB38, 0xFB3C},
+        {0xFB3E, 0xFB3E},
+        {0xFB40, 0xFB41},
+        {0xFB43, 0xFB44},
+        {0xFB46, 0xFBB1},
+        {0xFBD3, 0xFD3D},
+        {0xFD50, 0xFD8F},
+        {0xFD92, 0xFDC7},
+        {0xFDF0, 0xFDFB},
+        {0xFE70, 0xFE74},
+        {0xFE76, 0xFEFC},
+        {0xFF21, 0xFF3A},
+        {0xFF41, 0xFF5A},
+        {0xFF66, 0xFFBE},
+        {0xFFC2, 0xFFC7},
+        {0xFFCA, 0xFFCF},
+        {0xFFD2, 0xFFD7},
+        {0xFFDA, 0xFFDC},
+        {0x10000, 0x1000B},
+        {0x1000D, 0x10026},
+        {0x10028, 0x1003A},
+        {0x1003C, 0x1003D},
+        {0x1003F, 0x1004D},
+        {0x10050, 0x1005D},
+        {0x10080, 0x100FA},
+        {0x10280, 0x1029C},
+        {0x102A0, 0x102D0},
+        {0x10300, 0x1031F},
+        {0x1032D, 0x10340},
+        {0x10342, 0x10349},
+        {0x10350, 0x10375},
+        {0x10380, 0x1039D},
+        {0x103A0, 0x103C3},
+        {0x103C8, 0x103CF},
+        {0x10400, 0x1049D},
+        {0x104B0, 0x104D3},
+        {0x104D8, 0x104FB},
+        {0x10500, 0x10527},
+        {0x10530, 0x10563},
+        {0x10570, 0x1057A},
+        {0x1057C, 0x1058A},
+        {0x1058C, 0x10592},
+        {0x10594, 0x10595},
+        {0x10597, 0x105A1},
+        {0x105A3, 0x105B1},
+        {0x105B3, 0x105B9},
+        {0x105BB, 0x105BC},
+        {0x105C0, 0x105F3},
+        {0x10600, 0x10736},
+        {0x10740, 0x10755},
+        {0x10760, 0x10767},
+        {0x10780, 0x10785},
+        {0x10787, 0x107B0},
+        {0x107B2, 0x107BA},
+        {0x10800, 0x10805},
+        {0x10808, 0x10808},
+        {0x1080A, 0x10835},
+        {0x10837, 0x10838},
+        {0x1083C, 0x1083C},
+        {0x1083F, 0x10855},
+        {0x10860, 0x10876},
+        {0x10880, 0x1089E},
+        {0x108E0, 0x108F2},
+        {0x108F4, 0x108F5},
+        {0x10900, 0x10915},
+        {0x10920, 0x10939},
+        {0x10940, 0x10959},
+        {0x10980, 0x109B7},
+        {0x109BE, 0x109BF},
+        {0x10A00, 0x10A00},
+        {0x10A10, 0x10A13},
+        {0x10A15, 0x10A17},
+        {0x10A19, 0x10A35},
+        {0x10A60, 0x10A7C},
+        {0x10A80, 0x10A9C},
+        {0x10AC0, 0x10AC7},
+        {0x10AC9, 0x10AE4},
+        {0x10B00, 0x10B35},
+        {0x10B40, 0x10B55},
+        {0x10B60, 0x10B72},
+        {0x10B80, 0x10B91},
+        {0x10C00, 0x10C48},
+        {0x10C80, 0x10CB2},
+        {0x10CC0, 0x10CF2},
+        {0x10D00, 0x10D23},
+        {0x10D4A, 0x10D65},
+        {0x10D6F, 0x10D85},
+        {0x10E80, 0x10EA9},
+        {0x10EB0, 0x10EB1},
+        {0x10EC2, 0x10EC7},
+        {0x10F00, 0x10F1C},
+        {0x10F27, 0x10F27},
+        {0x10F30, 0x10F45},
+        {0x10F70, 0x10F81},
+        {0x10FB0, 0x10FC4},
+        {0x10FE0, 0x10FF6},
+        {0x11003, 0x11037},
+        {0x11071, 0x11072},
+        {0x11075, 0x11075},
+        {0x11083, 0x110AF},
+        {0x110D0, 0x110E8},
+        {0x11103, 0x11126},
+        {0x11144, 0x11144},
+        {0x11147, 0x11147},
+        {0x11150, 0x11172},
+        {0x11176, 0x11176},
+        {0x11183, 0x111B2},
+        {0x111C1, 0x111C4},
+        {0x111DA, 0x111DA},
+        {0x111DC, 0x111DC},
+        {0x11200, 0x11211},
+        {0x11213, 0x1122B},
+        {0x1123F, 0x11240},
+        {0x11280, 0x11286},
+        {0x11288, 0x11288},
+        {0x1128A, 0x1128D},
+        {0x1128F, 0x1129D},
+        {0x1129F, 0x112A8},
+        {0x112B0, 0x112DE},
+        {0x11305, 0x1130C},
+        {0x1130F, 0x11310},
+        {0x11313, 0x11328},
+        {0x1132A, 0x11330},
+        {0x11332, 0x11333},
+        {0x11335, 0x11339},
+        {0x1133D, 0x1133D},
+        {0x11350, 0x11350},
+        {0x1135D, 0x11361},
+        {0x11380, 0x11389},
+        {0x1138B, 0x1138B},
+        {0x1138E, 0x1138E},
+        {0x11390, 0x113B5},
+        {0x113B7, 0x113B7},
+        {0x113D1, 0x113D1},
+        {0x113D3, 0x113D3},
+        {0x11400, 0x11434},
+        {0x11447, 0x1144A},
+        {0x1145F, 0x11461},
+        {0x11480, 0x114AF},
+        {0x114C4, 0x114C5},
+        {0x114C7, 0x114C7},
+        {0x11580, 0x115AE},
+        {0x115D8, 0x115DB},
+        {0x11600, 0x1162F},
+        {0x11644, 0x11644},
+        {0x11680, 0x116AA},
+        {0x116B8, 0x116B8},
+        {0x11700, 0x1171A},
+        {0x11740, 0x11746},
+        {0x11800, 0x1182B},
+        {0x118A0, 0x118DF},
+        {0x118FF, 0x11906},
+        {0x11909, 0x11909},
+        {0x1190C, 0x11913},
+        {0x11915, 0x11916},
+        {0x11918, 0x1192F},
+        {0x1193F, 0x1193F},
+        {0x11941, 0x11941},
+        {0x119A0, 0x119A7},
+        {0x119AA, 0x119D0},
+        {0x119E1, 0x119E1},
+        {0x119E3, 0x119E3},
+        {0x11A00, 0x11A00},
+        {0x11A0B, 0x11A32},
+        {0x11A3A, 0x11A3A},
+        {0x11A50, 0x11A50},
+        {0x11A5C, 0x11A89},
+        {0x11A9D, 0x11A9D},
+        {0x11AB0, 0x11AF8},
+        {0x11BC0, 0x11BE0},
+        {0x11C00, 0x11C08},
+        {0x11C0A, 0x11C2E},
+        {0x11C40, 0x11C40},
+        {0x11C72, 0x11C8F},
+        {0x11D00, 0x11D06},
+        {0x11D08, 0x11D09},
+        {0x11D0B, 0x11D30},
+        {0x11D46, 0x11D46},
+        {0x11D60, 0x11D65},
+        {0x11D67, 0x11D68},
+        {0x11D6A, 0x11D89},
+        {0x11D98, 0x11D98},
+        {0x11DB0, 0x11DDB},
+        {0x11EE0, 0x11EF2},
+        {0x11F02, 0x11F02},
+        {0x11F04, 0x11F10},
+        {0x11F12, 0x11F33},
+        {0x11FB0, 0x11FB0},
+        {0x12000, 0x12399},
+        {0x12480, 0x12543},
+        {0x12F90, 0x12FF0},
+        {0x13000, 0x1342F},
+        {0x13441, 0x13446},
+        {0x13460, 0x143FA},
+        {0x14400, 0x14646},
+        {0x16100, 0x1611D},
+        {0x16800, 0x16A38},
+        {0x16A40, 0x16A5E},
+        {0x16A70, 0x16ABE},
+        {0x16AD0, 0x16AED},
+        {0x16B00, 0x16B2F},
+        {0x16B40, 0x16B43},
+        {0x16B63, 0x16B77},
+        {0x16B7D, 0x16B8F},
+        {0x16D40, 0x16D6C},
+        {0x16E40, 0x16E7F},
+        {0x16EA0, 0x16EB8},
+        {0x16EBB, 0x16ED3},
+        {0x16F00, 0x16F4A},
+        {0x16F50, 0x16F50},
+        {0x16F93, 0x16F9F},
+        {0x16FE0, 0x16FE1},
+        {0x16FE3, 0x16FE3},
+        {0x16FF2, 0x16FF3},
+        {0x17000, 0x18CD5},
+        {0x18CFF, 0x18D1E},
+        {0x18D80, 0x18DF2},
+        {0x1AFF0, 0x1AFF3},
+        {0x1AFF5, 0x1AFFB},
+        {0x1AFFD, 0x1AFFE},
+        {0x1B000, 0x1B122},
+        {0x1B132, 0x1B132},
+        {0x1B150, 0x1B152},
+        {0x1B155, 0x1B155},
+        {0x1B164, 0x1B167},
+        {0x1B170, 0x1B2FB},
+        {0x1BC00, 0x1BC6A},
+        {0x1BC70, 0x1BC7C},
+        {0x1BC80, 0x1BC88},
+        {0x1BC90, 0x1BC99},
+        {0x1D400, 0x1D454},
+        {0x1D456, 0x1D49C},
+        {0x1D49E, 0x1D49F},
+        {0x1D4A2, 0x1D4A2},
+        {0x1D4A5, 0x1D4A6},
+        {0x1D4A9, 0x1D4AC},
+        {0x1D4AE, 0x1D4B9},
+        {0x1D4BB, 0x1D4BB},
+        {0x1D4BD, 0x1D4C3},
+        {0x1D4C5, 0x1D505},
+        {0x1D507, 0x1D50A},
+        {0x1D50D, 0x1D514},
+        {0x1D516, 0x1D51C},
+        {0x1D51E, 0x1D539},
+        {0x1D53B, 0x1D53E},
+        {0x1D540, 0x1D544},
+        {0x1D546, 0x1D546},
+        {0x1D54A, 0x1D550},
+        {0x1D552, 0x1D6A5},
+        {0x1D6A8, 0x1D6C0},
+        {0x1D6C2, 0x1D6DA},
+        {0x1D6DC, 0x1D6FA},
+        {0x1D6FC, 0x1D714},
+        {0x1D716, 0x1D734},
+        {0x1D736, 0x1D74E},
+        {0x1D750, 0x1D76E},
+        {0x1D770, 0x1D788},
+        {0x1D78A, 0x1D7A8},
+        {0x1D7AA, 0x1D7C2},
+        {0x1D7C4, 0x1D7CB},
+        {0x1DF00, 0x1DF1E},
+        {0x1DF25, 0x1DF2A},
+        {0x1E030, 0x1E06D},
+        {0x1E100, 0x1E12C},
+        {0x1E137, 0x1E13D},
+        {0x1E14E, 0x1E14E},
+        {0x1E290, 0x1E2AD},
+        {0x1E2C0, 0x1E2EB},
+        {0x1E4D0, 0x1E4EB},
+        {0x1E5D0, 0x1E5ED},
+        {0x1E5F0, 0x1E5F0},
+        {0x1E6C0, 0x1E6DE},
+        {0x1E6E0, 0x1E6E2},
+        {0x1E6E4, 0x1E6E5},
+        {0x1E6E7, 0x1E6ED},
+        {0x1E6F0, 0x1E6F4},
+        {0x1E6FE, 0x1E6FF},
+        {0x1E7E0, 0x1E7E6},
+        {0x1E7E8, 0x1E7EB},
+        {0x1E7ED, 0x1E7EE},
+        {0x1E7F0, 0x1E7FE},
+        {0x1E800, 0x1E8C4},
+        {0x1E900, 0x1E943},
+        {0x1E94B, 0x1E94B},
+        {0x1EE00, 0x1EE03},
+        {0x1EE05, 0x1EE1F},
+        {0x1EE21, 0x1EE22},
+        {0x1EE24, 0x1EE24},
+        {0x1EE27, 0x1EE27},
+        {0x1EE29, 0x1EE32},
+        {0x1EE34, 0x1EE37},
+        {0x1EE39, 0x1EE39},
+        {0x1EE3B, 0x1EE3B},
+        {0x1EE42, 0x1EE42},
+        {0x1EE47, 0x1EE47},
+        {0x1EE49, 0x1EE49},
+        {0x1EE4B, 0x1EE4B},
+        {0x1EE4D, 0x1EE4F},
+        {0x1EE51, 0x1EE52},
+        {0x1EE54, 0x1EE54},
+        {0x1EE57, 0x1EE57},
+        {0x1EE59, 0x1EE59},
+        {0x1EE5B, 0x1EE5B},
+        {0x1EE5D, 0x1EE5D},
+        {0x1EE5F, 0x1EE5F},
+        {0x1EE61, 0x1EE62},
+        {0x1EE64, 0x1EE64},
+        {0x1EE67, 0x1EE6A},
+        {0x1EE6C, 0x1EE72},
+        {0x1EE74, 0x1EE77},
+        {0x1EE79, 0x1EE7C},
+        {0x1EE7E, 0x1EE7E},
+        {0x1EE80, 0x1EE89},
+        {0x1EE8B, 0x1EE9B},
+        {0x1EEA1, 0x1EEA3},
+        {0x1EEA5, 0x1EEA9},
+        {0x1EEAB, 0x1EEBB},
+        {0x20000, 0x2A6DF},
+        {0x2A700, 0x2B81D},
+        {0x2B820, 0x2CEAD},
+        {0x2CEB0, 0x2EBE0},
+        {0x2EBF0, 0x2EE5D},
+        {0x2F800, 0x2FA1D},
+        {0x30000, 0x3134A},
+        {0x31350, 0x33479},
+    };
+
+    for (const auto& r : ranges) {
+        if (ch >= r.start && ch <= r.end)
+            return true;
+    }
+    return false;
+}
+
+bool is_space(char32_t cp) {
+    switch (cp) {
+        case 0x0009:  // TAB \t
+        case 0x000A:  // LF \n
+        case 0x000B:  // VT
+        case 0x000C:  // FF
+        case 0x000D:  // CR \r
+        case 0x0020:  // Space
+        case 0x00A0:  // No-Break Space
+        case 0x1680:  // Ogham Space Mark
+        case 0x2000:  // En Quad
+        case 0x2001:  // Em Quad
+        case 0x2002:  // En Space
+        case 0x2003:  // Em Space
+        case 0x2004:  // Three-Per-Em Space
+        case 0x2005:  // Four-Per-Em Space
+        case 0x2006:  // Six-Per-Em Space
+        case 0x2007:  // Figure Space
+        case 0x2008:  // Punctuation Space
+        case 0x2009:  // Thin Space
+        case 0x200A:  // Hair Space
+        case 0x202F:  // Narrow No-Break Space
+        case 0x205F:  // Medium Mathematical Space
+        case 0x3000:  // Ideographic Space
+            return true;
+        default:
+            return false;
+    }
+}
+
+std::string str_to_lower(const std::string& input) {
+    std::string result = input;
+    std::transform(result.begin(), result.end(), result.begin(),
+                   [](unsigned char c) { return std::tolower(c); });
+    return result;
+}
+
+// UTF-8 -> Unicode code points
+std::vector<char32_t> utf8_to_codepoints(const std::string& str) {
+    std::vector<char32_t> codepoints;
+    size_t i = 0;
+    while (i < str.size()) {
+        unsigned char c    = str[i];
+        char32_t cp        = 0;
+        size_t extra_bytes = 0;
+
+        if ((c & 0x80) == 0)
+            cp = c;
+        else if ((c & 0xE0) == 0xC0) {
+            cp          = c & 0x1F;
+            extra_bytes = 1;
+        } else if ((c & 0xF0) == 0xE0) {
+            cp          = c & 0x0F;
+            extra_bytes = 2;
+        } else if ((c & 0xF8) == 0xF0) {
+            cp          = c & 0x07;
+            extra_bytes = 3;
+        } else {
+            ++i;
+            continue;
+        }  // Invalid UTF-8
+
+        if (i + extra_bytes >= str.size())
+            break;
+
+        for (size_t j = 1; j <= extra_bytes; ++j)
+            cp = (cp << 6) | (str[i + j] & 0x3F);
+
+        codepoints.push_back(cp);
+        i += 1 + extra_bytes;
+    }
+    return codepoints;
+}
+
+// Unicode code point -> UTF-8
+std::string codepoint_to_utf8(char32_t cp) {
+    std::string out;
+    if (cp <= 0x7F)
+        out.push_back(static_cast<char>(cp));
+    else if (cp <= 0x7FF) {
+        out.push_back(static_cast<char>(0xC0 | (cp >> 6)));
+        out.push_back(static_cast<char>(0x80 | (cp & 0x3F)));
+    } else if (cp <= 0xFFFF) {
+        out.push_back(static_cast<char>(0xE0 | (cp >> 12)));
+        out.push_back(static_cast<char>(0x80 | ((cp >> 6) & 0x3F)));
+        out.push_back(static_cast<char>(0x80 | (cp & 0x3F)));
+    } else {
+        out.push_back(static_cast<char>(0xF0 | (cp >> 18)));
+        out.push_back(static_cast<char>(0x80 | ((cp >> 12) & 0x3F)));
+        out.push_back(static_cast<char>(0x80 | ((cp >> 6) & 0x3F)));
+        out.push_back(static_cast<char>(0x80 | (cp & 0x3F)));
+    }
+    return out;
+}
+
+bool starts_with(const std::vector<char32_t>& text,
+                 const std::vector<char32_t>& prefix,
+                 std::size_t index) {
+    if (index > text.size()) {
+        return false;
+    }
+    if (prefix.size() > text.size() - index) {
+        return false;
+    }
+    return std::equal(prefix.begin(), prefix.end(), text.begin() + index);
+}
+
+std::vector<std::string> token_split(const std::string& text) {
+    std::vector<std::string> tokens;
+    auto cps = utf8_to_codepoints(text);
+    size_t i = 0;
+
+    while (i < cps.size()) {
+        char32_t cp = cps[i];
+
+        // `(?i:'s|'t|'re|'ve|'m|'ll|'d)`
+        if (cp == U'\'' && i + 1 < cps.size()) {
+            std::string next = str_to_lower(codepoint_to_utf8(cps[i + 1]));
+            if (next == "s" || next == "t" || next == "m") {
+                tokens.push_back("'" + next);
+                i += 2;
+                continue;
+            }
+            if (i + 2 < cps.size()) {
+                next += str_to_lower(codepoint_to_utf8(cps[i + 2]));
+                if (next == "re" || next == "ve" || next == "ll" || next == "d") {
+                    tokens.push_back("'" + next);
+                    i += 3;
+                    continue;
+                }
+            }
+        }
+
+        // `\p{N}`
+        if (is_number(cp)) {
+            tokens.push_back(codepoint_to_utf8(cp));
+            ++i;
+            continue;
+        }
+
+        // `[^\r\n\p{L}\p{N}]?\p{L}+`
+        {
+            // `[^\r\n\p{L}\p{N}]\p{L}+`
+            if (!is_letter(cp) && cp != U'\r' && cp != U'\n' && i + 1 < cps.size() && is_letter(cps[i + 1])) {
+                std::string token = codepoint_to_utf8(cp);
+                ++i;
+
+                while (i < cps.size() && is_letter(cps[i])) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+                tokens.push_back(token);
+                continue;
+            }
+
+            // `\p{L}+`
+            if (is_letter(cp)) {
+                std::string token = codepoint_to_utf8(cp);
+                ++i;
+                while (i < cps.size() && is_letter(cps[i])) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+                tokens.push_back(token);
+                continue;
+            }
+        }
+
+        // ` ?[^\s\p{L}\p{N}]+[\r\n]*`
+        {
+            // ` [^\s\p{L}\p{N}]+[\r\n]*`
+            if (cp == U' ' && i + 1 < cps.size() && !isspace(cps[i + 1]) && !is_letter(cps[i + 1]) && !is_number(cps[i + 1])) {
+                std::string token = codepoint_to_utf8(cp);
+                token += codepoint_to_utf8(cps[i + 1]);
+                i += 2;
+
+                while (i < cps.size() && !is_letter(cps[i]) && !is_number(cps[i]) && !isspace(cps[i])) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+
+                while (i < cps.size() && (cps[i] == U'\r' || cps[i] == U'\n')) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+
+                tokens.push_back(token);
+                continue;
+            }
+
+            // `[^\s\p{L}\p{N}]+[\r\n]*`
+            std::string token;
+            if (!is_letter(cps[i]) && !is_number(cps[i]) && !isspace(cps[i])) {
+                std::string token = codepoint_to_utf8(cp);
+                ++i;
+
+                while (i < cps.size() && !is_letter(cps[i]) && !is_number(cps[i]) && !isspace(cps[i])) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+
+                while (i < cps.size() && (cps[i] == U'\r' || cps[i] == U'\n')) {
+                    token += codepoint_to_utf8(cps[i]);
+                    ++i;
+                }
+
+                tokens.push_back(token);
+                continue;
+            }
+        }
+
+        // `\s*[\r\n]+|\s+(?!\S)|\s+`
+        if (is_space(cp)) {
+            std::string token = codepoint_to_utf8(cp);
+            ++i;
+
+            while (i < cps.size() && is_space(cps[i])) {
+                token += codepoint_to_utf8(cps[i]);
+                ++i;
+                if (cps[i] == U'\r' || cps[i] == U'\n') {
+                    break;
+                }
+            }
+
+            tokens.push_back(token);
+            continue;
+        }
+
+        // skip
+        ++i;
+    }
+
+    return tokens;
+}
+
+std::vector<std::string> split_with_special_tokens(
+    const std::string& text,
+    const std::vector<std::string>& special_tokens) {
+    std::vector<std::string> result;
+    size_t pos      = 0;
+    size_t text_len = text.size();
+
+    while (pos < text_len) {
+        size_t next_pos = text_len;
+        std::string matched_token;
+
+        for (const auto& token : special_tokens) {
+            size_t token_pos = text.find(token, pos);
+            if (token_pos != std::string::npos && token_pos < next_pos) {
+                next_pos      = token_pos;
+                matched_token = token;
+            }
+        }
+
+        if (next_pos > pos) {
+            result.push_back(text.substr(pos, next_pos - pos));
+        }
+
+        if (!matched_token.empty()) {
+            result.push_back(matched_token);
+            pos = next_pos + matched_token.size();
+        } else {
+            break;
+        }
+    }
+
+    return result;
+}
+
+// int main() {
+//     std::string text = "I'm testing C++ token_split function. 你好，世界! 123";
+//     auto tokens = token_split(text);
+
+//     for (const auto& t : tokens) {
+//         std::cout << "[" << t << "] ";
+//     }
+//     std::cout << "\n";
+//     return 0;
+// }

tokenize_util.h

@@ -0,0 +1,10 @@
+#ifndef __TOKENIZE_UTIL__
+#define __TOKENIZE_UTIL__
+
+#include <string>
+#include <vector>
+
+std::vector<std::string> token_split(const std::string& text);
+std::vector<std::string> split_with_special_tokens(const std::string& text, const std::vector<std::string>& special_tokens);
+
+#endif  // __TOKENIZE_UTIL__

unet.hpp

@@ -384,8 +384,8 @@ public:
                                 struct ggml_tensor* x,
                                 struct ggml_tensor* timesteps,
                                 struct ggml_tensor* context,
-                                struct ggml_tensor* c_concat              = NULL,
-                                struct ggml_tensor* y                     = NULL,
+                                struct ggml_tensor* c_concat              = nullptr,
+                                struct ggml_tensor* y                     = nullptr,
                                 int num_video_frames                      = -1,
                                 std::vector<struct ggml_tensor*> controls = {},
                                 float control_strength                    = 0.f) {
@@ -395,20 +395,20 @@ public:
         // c_concat: [N, in_channels, h, w] or [1, in_channels, h, w]
         // y: [N, adm_in_channels] or [1, adm_in_channels]
         // return: [N, out_channels, h, w]
-        if (context != NULL) {
+        if (context != nullptr) {
             if (context->ne[2] != x->ne[3]) {
                 context = ggml_repeat(ctx, context, ggml_new_tensor_3d(ctx, GGML_TYPE_F32, context->ne[0], context->ne[1], x->ne[3]));
             }
         }
 
-        if (c_concat != NULL) {
+        if (c_concat != nullptr) {
             if (c_concat->ne[3] != x->ne[3]) {
                 c_concat = ggml_repeat(ctx, c_concat, x);
             }
             x = ggml_concat(ctx, x, c_concat, 2);
         }
 
-        if (y != NULL) {
+        if (y != nullptr) {
             if (y->ne[1] != x->ne[3]) {
                 y = ggml_repeat(ctx, y, ggml_new_tensor_2d(ctx, GGML_TYPE_F32, y->ne[0], x->ne[3]));
             }
@@ -428,7 +428,7 @@ public:
         emb      = time_embed_2->forward(ctx, emb);  // [N, time_embed_dim]
 
         // SDXL/SVD
-        if (y != NULL) {
+        if (y != nullptr) {
             auto label_embed_0 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.0"]);
             auto label_embed_2 = std::dynamic_pointer_cast<Linear>(blocks["label_emb.0.2"]);
 
@@ -562,7 +562,7 @@ struct UNetModelRunner : public GGMLRunner {
         }
     }
 
-    std::string get_desc() {
+    std::string get_desc() override {
         return "unet";
     }
 
@@ -573,8 +573,8 @@ struct UNetModelRunner : public GGMLRunner {
     struct ggml_cgraph* build_graph(struct ggml_tensor* x,
                                     struct ggml_tensor* timesteps,
                                     struct ggml_tensor* context,
-                                    struct ggml_tensor* c_concat              = NULL,
-                                    struct ggml_tensor* y                     = NULL,
+                                    struct ggml_tensor* c_concat              = nullptr,
+                                    struct ggml_tensor* y                     = nullptr,
                                     int num_video_frames                      = -1,
                                     std::vector<struct ggml_tensor*> controls = {},
                                     float control_strength                    = 0.f) {
@@ -619,8 +619,8 @@ struct UNetModelRunner : public GGMLRunner {
                  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) {
+                 struct ggml_tensor** output               = nullptr,
+                 struct ggml_context* output_ctx           = nullptr) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
@@ -636,11 +636,11 @@ struct UNetModelRunner : public GGMLRunner {
     void test() {
         struct ggml_init_params params;
         params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-        params.mem_buffer = NULL;
+        params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
         struct ggml_context* work_ctx = ggml_init(params);
-        GGML_ASSERT(work_ctx != NULL);
+        GGML_ASSERT(work_ctx != nullptr);
 
         {
             // CPU, num_video_frames = 1, x{num_video_frames, 8, 8, 8}: Pass
@@ -663,10 +663,10 @@ struct UNetModelRunner : public GGMLRunner {
             ggml_set_f32(y, 0.5f);
             // print_ggml_tensor(y);
 
-            struct ggml_tensor* out = NULL;
+            struct ggml_tensor* out = nullptr;
 
             int t0 = ggml_time_ms();
-            compute(8, x, timesteps, context, NULL, y, num_video_frames, {}, 0.f, &out, work_ctx);
+            compute(8, x, timesteps, context, nullptr, y, num_video_frames, {}, 0.f, &out, work_ctx);
             int t1 = ggml_time_ms();
 
             print_ggml_tensor(out);

upscaler.cpp

@@ -4,7 +4,7 @@
 #include "stable-diffusion.h"
 
 struct UpscalerGGML {
-    ggml_backend_t backend    = NULL;  // general backend
+    ggml_backend_t backend    = nullptr;  // general backend
     ggml_type model_data_type = GGML_TYPE_F16;
     std::shared_ptr<ESRGAN> esrgan_upscaler;
     std::string esrgan_path;
@@ -18,14 +18,15 @@ struct UpscalerGGML {
     }
 
     bool load_from_file(const std::string& esrgan_path,
-                        bool offload_params_to_cpu) {
+                        bool offload_params_to_cpu,
+                        int n_threads) {
+        ggml_log_set(ggml_log_callback_default, nullptr);
 #ifdef SD_USE_CUDA
         LOG_DEBUG("Using CUDA backend");
         backend = ggml_backend_cuda_init(0);
 #endif
 #ifdef SD_USE_METAL
         LOG_DEBUG("Using Metal backend");
-        ggml_log_set(ggml_log_callback_default, nullptr);
         backend = ggml_backend_metal_init();
 #endif
 #ifdef SD_USE_VULKAN
@@ -54,7 +55,7 @@ struct UpscalerGGML {
         if (direct) {
             esrgan_upscaler->enable_conv2d_direct();
         }
-        if (!esrgan_upscaler->load_from_file(esrgan_path)) {
+        if (!esrgan_upscaler->load_from_file(esrgan_path, n_threads)) {
             return false;
         }
         return true;
@@ -62,16 +63,15 @@ struct UpscalerGGML {
 
     sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor) {
         // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
-        sd_image_t upscaled_image = {0, 0, 0, NULL};
+        sd_image_t upscaled_image = {0, 0, 0, nullptr};
         int output_width          = (int)input_image.width * esrgan_upscaler->scale;
         int output_height         = (int)input_image.height * esrgan_upscaler->scale;
         LOG_INFO("upscaling from (%i x %i) to (%i x %i)",
                  input_image.width, input_image.height, output_width, output_height);
 
         struct ggml_init_params params;
-        params.mem_size = output_width * output_height * 3 * sizeof(float) * 2;
-        params.mem_size += 2 * ggml_tensor_overhead();
-        params.mem_buffer = NULL;
+        params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
+        params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
         // draft context
@@ -80,9 +80,9 @@ struct UpscalerGGML {
             LOG_ERROR("ggml_init() failed");
             return upscaled_image;
         }
-        LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
+        // LOG_DEBUG("upscale work buffer size: %.2f MB", params.mem_size / 1024.f / 1024.f);
         ggml_tensor* input_image_tensor = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, input_image.width, input_image.height, 3, 1);
-        sd_image_to_tensor(input_image.data, input_image_tensor);
+        sd_image_to_tensor(input_image, input_image_tensor);
 
         ggml_tensor* upscaled = ggml_new_tensor_4d(upscale_ctx, GGML_TYPE_F32, output_width, output_height, 3, 1);
         auto on_tiling        = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
@@ -107,7 +107,7 @@ struct UpscalerGGML {
 };
 
 struct upscaler_ctx_t {
-    UpscalerGGML* upscaler = NULL;
+    UpscalerGGML* upscaler = nullptr;
 };
 
 upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
@@ -115,21 +115,21 @@ upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
                                  bool direct,
                                  int n_threads) {
     upscaler_ctx_t* upscaler_ctx = (upscaler_ctx_t*)malloc(sizeof(upscaler_ctx_t));
-    if (upscaler_ctx == NULL) {
-        return NULL;
+    if (upscaler_ctx == nullptr) {
+        return nullptr;
     }
     std::string esrgan_path(esrgan_path_c_str);
 
     upscaler_ctx->upscaler = new UpscalerGGML(n_threads, direct);
-    if (upscaler_ctx->upscaler == NULL) {
-        return NULL;
+    if (upscaler_ctx->upscaler == nullptr) {
+        return nullptr;
     }
 
-    if (!upscaler_ctx->upscaler->load_from_file(esrgan_path, offload_params_to_cpu)) {
+    if (!upscaler_ctx->upscaler->load_from_file(esrgan_path, offload_params_to_cpu, n_threads)) {
         delete upscaler_ctx->upscaler;
-        upscaler_ctx->upscaler = NULL;
+        upscaler_ctx->upscaler = nullptr;
         free(upscaler_ctx);
-        return NULL;
+        return nullptr;
     }
     return upscaler_ctx;
 }
@@ -138,10 +138,17 @@ sd_image_t upscale(upscaler_ctx_t* upscaler_ctx, sd_image_t input_image, uint32_
     return upscaler_ctx->upscaler->upscale(input_image, upscale_factor);
 }
 
+int get_upscale_factor(upscaler_ctx_t* upscaler_ctx) {
+    if (upscaler_ctx == nullptr || upscaler_ctx->upscaler == nullptr || upscaler_ctx->upscaler->esrgan_upscaler == nullptr) {
+        return 1;
+    }
+    return upscaler_ctx->upscaler->esrgan_upscaler->scale;
+}
+
 void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx) {
-    if (upscaler_ctx->upscaler != NULL) {
+    if (upscaler_ctx->upscaler != nullptr) {
         delete upscaler_ctx->upscaler;
-        upscaler_ctx->upscaler = NULL;
+        upscaler_ctx->upscaler = nullptr;
     }
     free(upscaler_ctx);
 }

util.cpp

@@ -1,8 +1,8 @@
 #include "util.h"
-#include <stdarg.h>
 #include <algorithm>
 #include <cmath>
 #include <codecvt>
+#include <cstdarg>
 #include <fstream>
 #include <locale>
 #include <sstream>
@@ -64,7 +64,7 @@ std::string format(const char* fmt, ...) {
     va_list ap2;
     va_start(ap, fmt);
     va_copy(ap2, ap);
-    int size = vsnprintf(NULL, 0, fmt, ap);
+    int size = vsnprintf(nullptr, 0, fmt, ap);
     std::vector<char> buf(size + 1);
     int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
     va_end(ap2);
@@ -84,6 +84,7 @@ int round_up_to(int value, int base) {
 }
 
 #ifdef _WIN32  // code for windows
+#define NOMINMAX
 #include <windows.h>
 
 bool file_exists(const std::string& filename) {
@@ -110,56 +111,6 @@ std::string get_full_path(const std::string& dir, const std::string& filename) {
     }
 }
 
-std::vector<std::string> get_files_from_dir(const std::string& dir) {
-    std::vector<std::string> files;
-
-    WIN32_FIND_DATA findFileData;
-    HANDLE hFind;
-
-    char currentDirectory[MAX_PATH];
-    GetCurrentDirectory(MAX_PATH, currentDirectory);
-
-    char directoryPath[MAX_PATH];  // this is absolute path
-    sprintf(directoryPath, "%s\\%s\\*", currentDirectory, dir.c_str());
-
-    // Find the first file in the directory
-    hFind               = FindFirstFile(directoryPath, &findFileData);
-    bool isAbsolutePath = false;
-    // Check if the directory was found
-    if (hFind == INVALID_HANDLE_VALUE) {
-        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)) {
-            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);
-
-    // Close the handle
-    FindClose(hFind);
-
-    sort(files.begin(), files.end());
-
-    return files;
-}
-
 #else  // Unix
 #include <dirent.h>
 #include <sys/stat.h>
@@ -194,27 +145,6 @@ std::string get_full_path(const std::string& dir, const std::string& filename) {
     return "";
 }
 
-std::vector<std::string> get_files_from_dir(const std::string& dir) {
-    std::vector<std::string> files;
-
-    DIR* dp = opendir(dir.c_str());
-
-    if (dp != nullptr) {
-        struct dirent* entry;
-
-        while ((entry = readdir(dp)) != nullptr) {
-            std::string fname = dir + "/" + entry->d_name;
-            if (!is_directory(fname))
-                files.push_back(fname);
-        }
-        closedir(dp);
-    }
-
-    sort(files.begin(), files.end());
-
-    return files;
-}
-
 #endif
 
 // get_num_physical_cores is copy from
@@ -240,11 +170,11 @@ int32_t get_num_physical_cores() {
 #elif defined(__APPLE__) && defined(__MACH__)
     int32_t num_physical_cores;
     size_t len = sizeof(num_physical_cores);
-    int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    int result = sysctlbyname("hw.perflevel0.physicalcpu", &num_physical_cores, &len, nullptr, 0);
     if (result == 0) {
         return num_physical_cores;
     }
-    result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, NULL, 0);
+    result = sysctlbyname("hw.physicalcpu", &num_physical_cores, &len, nullptr, 0);
     if (result == 0) {
         return num_physical_cores;
     }
@@ -255,8 +185,8 @@ int32_t get_num_physical_cores() {
     return n_threads > 0 ? (n_threads <= 4 ? n_threads : n_threads / 2) : 4;
 }
 
-static sd_progress_cb_t sd_progress_cb = NULL;
-void* sd_progress_cb_data              = NULL;
+static sd_progress_cb_t sd_progress_cb = nullptr;
+void* sd_progress_cb_data              = nullptr;
 
 std::u32string utf8_to_utf32(const std::string& utf8_str) {
     std::wstring_convert<std::codecvt_utf8<char32_t>, char32_t> converter;
@@ -318,39 +248,6 @@ std::vector<std::string> split_string(const std::string& str, char delimiter) {
     return result;
 }
 
-sd_image_t* preprocess_id_image(sd_image_t* img) {
-    int shortest_edge   = 224;
-    int size            = shortest_edge;
-    sd_image_t* resized = NULL;
-    uint32_t w          = img->width;
-    uint32_t h          = img->height;
-    uint32_t c          = img->channel;
-
-    // 1. do resize using stb_resize functions
-
-    unsigned char* buf = (unsigned char*)malloc(sizeof(unsigned char) * 3 * size * size);
-    if (!stbir_resize_uint8(img->data, w, h, 0,
-                            buf, size, size, 0,
-                            c)) {
-        fprintf(stderr, "%s: resize operation failed \n ", __func__);
-        return resized;
-    }
-
-    // 2. do center crop (likely unnecessary due to step 1)
-
-    // 3. do rescale
-
-    // 4. do normalize
-
-    // 3 and 4 will need to be done in float format.
-
-    resized = new sd_image_t{(uint32_t)shortest_edge,
-                             (uint32_t)shortest_edge,
-                             3,
-                             buf};
-    return resized;
-}
-
 void pretty_progress(int step, int steps, float time) {
     if (sd_progress_cb) {
         sd_progress_cb(step, steps, time, sd_progress_cb_data);
@@ -399,10 +296,10 @@ std::string trim(const std::string& s) {
     return rtrim(ltrim(s));
 }
 
-static sd_log_cb_t sd_log_cb = NULL;
-void* sd_log_cb_data         = NULL;
+static sd_log_cb_t sd_log_cb = nullptr;
+void* sd_log_cb_data         = nullptr;
 
-#define LOG_BUFFER_SIZE 1024
+#define LOG_BUFFER_SIZE 4096
 
 void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
     va_list args;
@@ -414,7 +311,10 @@ void log_printf(sd_log_level_t level, const char* file, int line, const char* fo
     if (written >= 0 && written < LOG_BUFFER_SIZE) {
         vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
     }
-    strncat(log_buffer, "\n", LOG_BUFFER_SIZE - strlen(log_buffer));
+    size_t len = strlen(log_buffer);
+    if (log_buffer[len - 1] != '\n') {
+        strncat(log_buffer, "\n", LOG_BUFFER_SIZE - len);
+    }
 
     if (sd_log_cb) {
         sd_log_cb(level, log_buffer, sd_log_cb_data);
@@ -488,10 +388,10 @@ sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int
             float original_x = (float)x * image.width / target_width;
             float original_y = (float)y * image.height / target_height;
 
-            int x1 = (int)original_x;
-            int y1 = (int)original_y;
-            int x2 = x1 + 1;
-            int y2 = y1 + 1;
+            uint32_t x1 = (uint32_t)original_x;
+            uint32_t y1 = (uint32_t)original_y;
+            uint32_t x2 = std::min(x1 + 1, image.width - 1);
+            uint32_t y2 = std::min(y1 + 1, image.height - 1);
 
             for (int k = 0; k < image.channel; k++) {
                 float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
@@ -528,23 +428,26 @@ float means[3] = {0.48145466, 0.4578275, 0.40821073};
 float stds[3]  = {0.26862954, 0.26130258, 0.27577711};
 
 // Function to clip and preprocess sd_image_f32_t
-sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size) {
-    float scale = (float)size / fmin(image.width, image.height);
+sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int target_height) {
+    float width_scale  = (float)target_width / image.width;
+    float height_scale = (float)target_height / image.height;
+
+    float scale = std::fmax(width_scale, height_scale);
 
     // Interpolation
-    int new_width       = (int)(scale * image.width);
-    int new_height      = (int)(scale * image.height);
-    float* resized_data = (float*)malloc(new_width * new_height * image.channel * sizeof(float));
+    int resized_width   = (int)(scale * image.width);
+    int resized_height  = (int)(scale * image.height);
+    float* resized_data = (float*)malloc(resized_width * resized_height * image.channel * sizeof(float));
 
-    for (int y = 0; y < new_height; y++) {
-        for (int x = 0; x < new_width; x++) {
-            float original_x = (float)x * image.width / new_width;
-            float original_y = (float)y * image.height / new_height;
+    for (int y = 0; y < resized_height; y++) {
+        for (int x = 0; x < resized_width; x++) {
+            float original_x = (float)x * image.width / resized_width;
+            float original_y = (float)y * image.height / resized_height;
 
-            int x1 = (int)original_x;
-            int y1 = (int)original_y;
-            int x2 = x1 + 1;
-            int y2 = y1 + 1;
+            uint32_t x1 = (uint32_t)original_x;
+            uint32_t y1 = (uint32_t)original_y;
+            uint32_t x2 = std::min(x1 + 1, image.width - 1);
+            uint32_t y2 = std::min(y1 + 1, image.height - 1);
 
             for (int k = 0; k < image.channel; k++) {
                 float v1 = *(image.data + y1 * image.width * image.channel + x1 * image.channel + k);
@@ -557,26 +460,28 @@ sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size) {
 
                 float value = interpolate(v1, v2, v3, v4, x_ratio, y_ratio);
 
-                *(resized_data + y * new_width * image.channel + x * image.channel + k) = value;
+                *(resized_data + y * resized_width * image.channel + x * image.channel + k) = value;
             }
         }
     }
 
     // Clip and preprocess
-    int h = (new_height - size) / 2;
-    int w = (new_width - size) / 2;
+    int h_offset = std::max((int)(resized_height - target_height) / 2, 0);
+    int w_offset = std::max((int)(resized_width - target_width) / 2, 0);
 
     sd_image_f32_t result;
-    result.width   = size;
-    result.height  = size;
+    result.width   = target_width;
+    result.height  = target_height;
     result.channel = image.channel;
-    result.data    = (float*)malloc(size * size * image.channel * sizeof(float));
+    result.data    = (float*)malloc(target_height * target_width * image.channel * sizeof(float));
 
     for (int k = 0; k < image.channel; k++) {
-        for (int i = 0; i < size; i++) {
-            for (int j = 0; j < size; j++) {
-                *(result.data + i * size * image.channel + j * image.channel + k) =
-                    fmin(fmax(*(resized_data + (i + h) * new_width * image.channel + (j + w) * image.channel + k), 0.0f), 255.0f) / 255.0f;
+        for (int i = 0; i < result.height; i++) {
+            for (int j = 0; j < result.width; j++) {
+                int src_y = std::min(i + h_offset, resized_height - 1);
+                int src_x = std::min(j + w_offset, resized_width - 1);
+                *(result.data + i * result.width * image.channel + j * image.channel + k) =
+                    fmin(fmax(*(resized_data + src_y * resized_width * image.channel + src_x * image.channel + k), 0.0f), 255.0f) / 255.0f;
             }
         }
     }
@@ -586,10 +491,10 @@ sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size) {
 
     // Normalize
     for (int k = 0; k < image.channel; k++) {
-        for (int i = 0; i < size; i++) {
-            for (int j = 0; j < size; j++) {
+        for (int i = 0; i < result.height; i++) {
+            for (int j = 0; j < result.width; j++) {
                 // *(result.data + i * size * image.channel + j * image.channel + k) = 0.5f;
-                int offset  = i * size * image.channel + j * image.channel + k;
+                int offset  = i * result.width * image.channel + j * image.channel + k;
                 float value = *(result.data + offset);
                 value       = (value - means[k]) / stds[k];
                 // value = 0.5f;

util.h

@@ -24,14 +24,9 @@ bool file_exists(const std::string& filename);
 bool is_directory(const std::string& path);
 std::string get_full_path(const std::string& dir, const std::string& filename);
 
-std::vector<std::string> get_files_from_dir(const std::string& dir);
-
 std::u32string utf8_to_utf32(const std::string& utf8_str);
 std::string utf32_to_utf8(const std::u32string& utf32_str);
 std::u32string unicode_value_to_utf32(int unicode_value);
-
-sd_image_t* preprocess_id_image(sd_image_t* img);
-
 // std::string sd_basename(const std::string& path);
 
 typedef struct {
@@ -47,7 +42,7 @@ sd_image_f32_t sd_image_t_to_sd_image_f32_t(sd_image_t image);
 
 sd_image_f32_t resize_sd_image_f32_t(sd_image_f32_t image, int target_width, int target_height);
 
-sd_image_f32_t clip_preprocess(sd_image_f32_t image, int size);
+sd_image_f32_t clip_preprocess(sd_image_f32_t image, int target_width, int target_height);
 
 std::string path_join(const std::string& p1, const std::string& p2);
 std::vector<std::string> split_string(const std::string& str, char delimiter);

vae.hpp

@@ -30,7 +30,7 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [N, in_channels, h, w]
         // t_emb is always None
         auto norm1 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm1"]);
@@ -76,7 +76,7 @@ public:
         blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [N, in_channels, h, w]
         auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
         auto q_proj   = std::dynamic_pointer_cast<Conv2d>(blocks["q"]);
@@ -134,7 +134,7 @@ public:
     }
 
     struct ggml_tensor* forward(struct ggml_context* ctx,
-                                struct ggml_tensor* x) {
+                                struct ggml_tensor* x) override {
         // timesteps always None
         // skip_video always False
         // x: [N, IC, IH, IW]
@@ -163,7 +163,7 @@ public:
 
 class VideoResnetBlock : public ResnetBlock {
 protected:
-    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+    void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
         enum ggml_type wtype = get_type(prefix + "mix_factor", tensor_types, GGML_TYPE_F32);
         params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
     }
@@ -182,7 +182,7 @@ public:
         blocks["time_stack"] = std::shared_ptr<GGMLBlock>(new ResBlock(out_channels, 0, out_channels, {video_kernel_size, 1}, 3, false, true));
     }
 
-    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+    struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
         // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w]
         // return: [N, out_channels, h, w] aka [b*t, out_channels, h, w]
         // t_emb is always None
@@ -530,6 +530,7 @@ struct VAE : public GGMLRunner {
                          struct ggml_context* output_ctx)                                                         = 0;
     virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) = 0;
     virtual void enable_conv2d_direct(){};
+    virtual void set_conv2d_scale(float scale) { SD_UNUSED(scale); };
 };
 
 struct AutoEncoderKL : public VAE {
@@ -547,7 +548,7 @@ struct AutoEncoderKL : public VAE {
         ae.init(params_ctx, tensor_types, prefix);
     }
 
-    void enable_conv2d_direct() {
+    void enable_conv2d_direct() override {
         std::vector<GGMLBlock*> blocks;
         ae.get_all_blocks(blocks);
         for (auto block : blocks) {
@@ -558,11 +559,22 @@ struct AutoEncoderKL : public VAE {
         }
     }
 
-    std::string get_desc() {
+    void set_conv2d_scale(float scale) override {
+        std::vector<GGMLBlock*> blocks;
+        ae.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->set_scale(scale);
+            }
+        }
+    }
+
+    std::string get_desc() override {
         return "vae";
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) override {
         ae.get_param_tensors(tensors, prefix);
     }
 
@@ -582,23 +594,24 @@ struct AutoEncoderKL : public VAE {
                  struct ggml_tensor* z,
                  bool decode_graph,
                  struct ggml_tensor** output,
-                 struct ggml_context* output_ctx = NULL) {
+                 struct ggml_context* output_ctx = nullptr) override {
+        GGML_ASSERT(!decode_only || decode_graph);
         auto get_graph = [&]() -> struct ggml_cgraph* {
             return build_graph(z, decode_graph);
         };
         // ggml_set_f32(z, 0.5f);
         // print_ggml_tensor(z);
-        GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
+        GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
     }
 
     void test() {
         struct ggml_init_params params;
         params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-        params.mem_buffer = NULL;
+        params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
         struct ggml_context* work_ctx = ggml_init(params);
-        GGML_ASSERT(work_ctx != NULL);
+        GGML_ASSERT(work_ctx != nullptr);
 
         {
             // CPU, x{1, 3, 64, 64}: Pass
@@ -608,7 +621,7 @@ struct AutoEncoderKL : public VAE {
             auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 64, 64, 3, 2);
             ggml_set_f32(x, 0.5f);
             print_ggml_tensor(x);
-            struct ggml_tensor* out = NULL;
+            struct ggml_tensor* out = nullptr;
 
             int t0 = ggml_time_ms();
             compute(8, x, false, &out, work_ctx);
@@ -626,7 +639,7 @@ struct AutoEncoderKL : public VAE {
             auto z = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 4, 1);
             ggml_set_f32(z, 0.5f);
             print_ggml_tensor(z);
-            struct ggml_tensor* out = NULL;
+            struct ggml_tensor* out = nullptr;
 
             int t0 = ggml_time_ms();
             compute(8, z, true, &out, work_ctx);

wan.hpp

@@ -2,6 +2,8 @@
 #define __WAN_HPP__
 
 #include <map>
+#include <memory>
+#include <utility>
 
 #include "common.hpp"
 #include "flux.hpp"
@@ -24,7 +26,7 @@ namespace WAN {
         std::tuple<int, int, int> dilation;
         bool bias;
 
-        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
             params["weight"] = ggml_new_tensor_4d(ctx,
                                                   GGML_TYPE_F16,
                                                   std::get<2>(kernel_size),
@@ -46,17 +48,17 @@ namespace WAN {
                      bool bias                          = true)
             : in_channels(in_channels),
               out_channels(out_channels),
-              kernel_size(kernel_size),
-              stride(stride),
-              padding(padding),
-              dilation(dilation),
+              kernel_size(std::move(kernel_size)),
+              stride(std::move(stride)),
+              padding(std::move(padding)),
+              dilation(std::move(dilation)),
               bias(bias) {}
 
-        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* cache_x = NULL) {
+        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* cache_x = nullptr) {
             // x: [N*IC, ID, IH, IW]
             // result: x: [N*OC, ID, IH, IW]
             struct ggml_tensor* w = params["weight"];
-            struct ggml_tensor* b = NULL;
+            struct ggml_tensor* b = nullptr;
             if (bias) {
                 b = params["bias"];
             }
@@ -68,7 +70,7 @@ namespace WAN {
             int lp2 = 2 * std::get<0>(padding);
             int rp2 = 0;
 
-            if (cache_x != NULL && lp2 > 0) {
+            if (cache_x != nullptr && lp2 > 0) {
                 x = ggml_concat(ctx, cache_x, x, 2);
                 lp2 -= (int)cache_x->ne[2];
             }
@@ -85,7 +87,7 @@ namespace WAN {
     protected:
         int64_t dim;
 
-        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
             ggml_type wtype = GGML_TYPE_F32;
             params["gamma"] = ggml_new_tensor_1d(ctx, wtype, dim);
         }
@@ -94,7 +96,7 @@ namespace WAN {
         RMS_norm(int64_t dim)
             : dim(dim) {}
 
-        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) {
+        struct ggml_tensor* forward(struct ggml_context* ctx, struct ggml_tensor* x) override {
             // x: [N*IC, ID, IH, IW], IC == dim
             // assert N == 1
 
@@ -159,12 +161,12 @@ namespace WAN {
                     int idx = feat_idx;
                     feat_idx += 1;
                     if (chunk_idx == 0) {
-                        // feat_cache[idx] == NULL, pass
+                        // feat_cache[idx] == nullptr, pass
                     } else {
                         auto time_conv = std::dynamic_pointer_cast<CausalConv3d>(blocks["time_conv"]);
 
                         auto cache_x = ggml_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
-                        if (cache_x->ne[2] < 2 && feat_cache[idx] != NULL) {  // chunk_idx >= 2
+                        if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {  // chunk_idx >= 2
                             // cache last frame of last two chunk
                             cache_x = ggml_concat(ctx,
                                                   ggml_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
@@ -209,7 +211,7 @@ namespace WAN {
             if (mode == "downsample3d") {
                 if (feat_cache.size() > 0) {
                     int idx = feat_idx;
-                    if (feat_cache[idx] == NULL) {
+                    if (feat_cache[idx] == nullptr) {
                         feat_cache[idx] = x;
                         feat_idx += 1;
                     } else {
@@ -373,7 +375,7 @@ namespace WAN {
                     if (feat_cache.size() > 0) {
                         int idx      = feat_idx;
                         auto cache_x = ggml_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
-                        if (cache_x->ne[2] < 2 && feat_cache[idx] != NULL) {
+                        if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
                             // cache last frame of last two chunk
                             cache_x = ggml_concat(ctx,
                                                   ggml_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
@@ -566,7 +568,7 @@ namespace WAN {
 
             x = ggml_nn_attention(ctx, q, k, v, false);  // [t, h * w, c]
             // v      = ggml_cont(ctx, ggml_torch_permute(ctx, v, 1, 0, 2, 3));  // [t, h * w, c]
-            // x = ggml_nn_attention_ext(ctx, q, k, v, q->ne[2], NULL, false, false, true);
+            // x = ggml_nn_attention_ext(ctx, q, k, v, q->ne[2], nullptr, false, false, true);
 
             x = ggml_nn_cont(ctx, ggml_permute(ctx, x, 1, 0, 2, 3));  // [t, c, h * w]
             x = ggml_reshape_4d(ctx, x, w, h, c, n);                  // [t, c, h, w]
@@ -672,7 +674,7 @@ namespace WAN {
             if (feat_cache.size() > 0) {
                 int idx      = feat_idx;
                 auto cache_x = ggml_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
-                if (cache_x->ne[2] < 2 && feat_cache[idx] != NULL) {
+                if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
                     // cache last frame of last two chunk
                     cache_x = ggml_concat(ctx,
                                           ggml_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
@@ -724,7 +726,7 @@ namespace WAN {
             if (feat_cache.size() > 0) {
                 int idx      = feat_idx;
                 auto cache_x = ggml_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
-                if (cache_x->ne[2] < 2 && feat_cache[idx] != NULL) {
+                if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
                     // cache last frame of last two chunk
                     cache_x = ggml_concat(ctx,
                                           ggml_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
@@ -843,7 +845,7 @@ namespace WAN {
             if (feat_cache.size() > 0) {
                 int idx      = feat_idx;
                 auto cache_x = ggml_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
-                if (cache_x->ne[2] < 2 && feat_cache[idx] != NULL) {
+                if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
                     // cache last frame of last two chunk
                     cache_x = ggml_concat(ctx,
                                           ggml_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
@@ -895,7 +897,7 @@ namespace WAN {
             if (feat_cache.size() > 0) {
                 int idx      = feat_idx;
                 auto cache_x = ggml_slice(ctx, x, 2, -CACHE_T, x->ne[2]);
-                if (cache_x->ne[2] < 2 && feat_cache[idx] != NULL) {
+                if (cache_x->ne[2] < 2 && feat_cache[idx] != nullptr) {
                     // cache last frame of last two chunk
                     cache_x = ggml_concat(ctx,
                                           ggml_slice(ctx, feat_cache[idx], 2, -1, feat_cache[idx]->ne[2]),
@@ -935,9 +937,9 @@ namespace WAN {
 
         void clear_cache() {
             _conv_idx     = 0;
-            _feat_map     = std::vector<struct ggml_tensor*>(_conv_num, NULL);
+            _feat_map     = std::vector<struct ggml_tensor*>(_conv_num, nullptr);
             _enc_conv_idx = 0;
-            _enc_feat_map = std::vector<struct ggml_tensor*>(_enc_conv_num, NULL);
+            _enc_feat_map = std::vector<struct ggml_tensor*>(_enc_conv_num, nullptr);
         }
 
     public:
@@ -1116,11 +1118,11 @@ namespace WAN {
             ae.init(params_ctx, tensor_types, prefix);
         }
 
-        std::string get_desc() {
+        std::string get_desc() override {
             return "wan_vae";
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) override {
             ae.get_param_tensors(tensors, prefix);
         }
 
@@ -1152,7 +1154,7 @@ namespace WAN {
 
             for (int64_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) {
                 ggml_tensor* feat_cache = ae._feat_map[feat_idx];
-                if (feat_cache != NULL) {
+                if (feat_cache != nullptr) {
                     cache("feat_idx:" + std::to_string(feat_idx), feat_cache);
                     ggml_build_forward_expand(gf, feat_cache);
                 }
@@ -1167,7 +1169,7 @@ namespace WAN {
                      struct ggml_tensor* z,
                      bool decode_graph,
                      struct ggml_tensor** output,
-                     struct ggml_context* output_ctx = NULL) {
+                     struct ggml_context* output_ctx = nullptr) override {
             if (true) {
                 auto get_graph = [&]() -> struct ggml_cgraph* {
                     return build_graph(z, decode_graph);
@@ -1180,7 +1182,7 @@ namespace WAN {
                 auto get_graph = [&]() -> struct ggml_cgraph* {
                     return build_graph_partial(z, decode_graph, i);
                 };
-                struct ggml_tensor* out = NULL;
+                struct ggml_tensor* out = nullptr;
                 GGMLRunner::compute(get_graph, n_threads, true, &out, output_ctx);
                 ae.clear_cache();
                 if (t == 1) {
@@ -1219,12 +1221,12 @@ namespace WAN {
 
         void test() {
             struct ggml_init_params params;
-            params.mem_size   = static_cast<size_t>(1000 * 1024 * 1024);  // 10 MB
-            params.mem_buffer = NULL;
+            params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
+            params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
             struct ggml_context* work_ctx = ggml_init(params);
-            GGML_ASSERT(work_ctx != NULL);
+            GGML_ASSERT(work_ctx != nullptr);
 
             if (true) {
                 // cpu f32, pass
@@ -1235,7 +1237,7 @@ namespace WAN {
                 ggml_set_f32(z, 0.5f);
                 z = load_tensor_from_file(work_ctx, "wan_vae_z.bin");
                 print_ggml_tensor(z);
-                struct ggml_tensor* out = NULL;
+                struct ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 compute(8, z, true, &out, work_ctx);
@@ -1250,7 +1252,7 @@ namespace WAN {
             // ggml_backend_t backend = ggml_backend_cuda_init(0);
             ggml_backend_t backend            = ggml_backend_cpu_init();
             ggml_type model_data_type         = GGML_TYPE_F16;
-            std::shared_ptr<WanVAERunner> vae = std::shared_ptr<WanVAERunner>(new WanVAERunner(backend, false, {}, "", false, VERSION_WAN2_2_TI2V));
+            std::shared_ptr<WanVAERunner> vae = std::make_shared<WanVAERunner>(backend, false, String2GGMLType{}, "", false, VERSION_WAN2_2_TI2V);
             {
                 LOG_INFO("loading from '%s'", file_path.c_str());
 
@@ -1309,7 +1311,7 @@ namespace WAN {
                                             ggml_backend_t backend,
                                             struct ggml_tensor* x,
                                             struct ggml_tensor* pe,
-                                            struct ggml_tensor* mask = NULL) {
+                                            struct ggml_tensor* mask = nullptr) {
             // x: [N, n_token, dim]
             // pe: [n_token, d_head/2, 2, 2]
             // return [N, n_token, dim]
@@ -1333,7 +1335,7 @@ namespace WAN {
             k = ggml_reshape_4d(ctx, k, head_dim, num_heads, n_token, N);  // [N, n_token, n_head, d_head]
             v = ggml_reshape_4d(ctx, v, head_dim, num_heads, n_token, N);  // [N, n_token, n_head, d_head]
 
-            x = Flux::attention(ctx, backend, q, k, v, pe, mask, flash_attn);  // [N, n_token, dim]
+            x = Rope::attention(ctx, backend, q, k, v, pe, mask, flash_attn);  // [N, n_token, dim]
 
             x = o_proj->forward(ctx, x);  // [N, n_token, dim]
             return x;
@@ -1367,7 +1369,7 @@ namespace WAN {
                                     ggml_backend_t backend,
                                     struct ggml_tensor* x,
                                     struct ggml_tensor* context,
-                                    int64_t context_img_len) {
+                                    int64_t context_img_len) override {
             // x: [N, n_token, dim]
             // context: [N, n_context, dim]
             // context_img_len: unused
@@ -1388,7 +1390,7 @@ namespace WAN {
             k      = norm_k->forward(ctx, k);
             auto v = v_proj->forward(ctx, context);  // [N, n_context, dim]
 
-            x = ggml_nn_attention_ext(ctx, backend, q, k, v, num_heads, NULL, false, false, flash_attn);  // [N, n_token, dim]
+            x = ggml_nn_attention_ext(ctx, backend, q, k, v, num_heads, nullptr, false, false, flash_attn);  // [N, n_token, dim]
 
             x = o_proj->forward(ctx, x);  // [N, n_token, dim]
             return x;
@@ -1417,7 +1419,7 @@ namespace WAN {
                                     ggml_backend_t backend,
                                     struct ggml_tensor* x,
                                     struct ggml_tensor* context,
-                                    int64_t context_img_len) {
+                                    int64_t context_img_len) override {
             // x: [N, n_token, dim]
             // context: [N, context_img_len + context_txt_len, dim]
             // return [N, n_token, dim]
@@ -1455,8 +1457,8 @@ namespace WAN {
             k_img      = norm_k_img->forward(ctx, k_img);
             auto v_img = v_img_proj->forward(ctx, context_img);  // [N, context_img_len, dim]
 
-            auto img_x = ggml_nn_attention_ext(ctx, backend, q, k_img, v_img, num_heads, NULL, false, false, flash_attn);  // [N, n_token, dim]
-            x          = ggml_nn_attention_ext(ctx, backend, q, k, v, num_heads, NULL, false, false, flash_attn);          // [N, n_token, dim]
+            auto img_x = ggml_nn_attention_ext(ctx, backend, q, k_img, v_img, num_heads, nullptr, false, false, flash_attn);  // [N, n_token, dim]
+            x          = ggml_nn_attention_ext(ctx, backend, q, k, v, num_heads, nullptr, false, false, flash_attn);          // [N, n_token, dim]
 
             x = ggml_add(ctx, x, img_x);
 
@@ -1497,7 +1499,7 @@ namespace WAN {
     protected:
         int dim;
 
-        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
             enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
             params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 6, 1);
         }
@@ -1532,13 +1534,13 @@ namespace WAN {
             blocks["ffn.2"] = std::shared_ptr<GGMLBlock>(new Linear(ffn_dim, dim));
         }
 
-        struct ggml_tensor* forward(struct ggml_context* ctx,
-                                    ggml_backend_t backend,
-                                    struct ggml_tensor* x,
-                                    struct ggml_tensor* e,
-                                    struct ggml_tensor* pe,
-                                    struct ggml_tensor* context,
-                                    int64_t context_img_len = 257) {
+        virtual struct ggml_tensor* forward(struct ggml_context* ctx,
+                                            ggml_backend_t backend,
+                                            struct ggml_tensor* x,
+                                            struct ggml_tensor* e,
+                                            struct ggml_tensor* pe,
+                                            struct ggml_tensor* context,
+                                            int64_t context_img_len = 257) {
             // x: [N, n_token, dim]
             // e: [N, 6, dim] or [N, T, 6, dim]
             // context: [N, context_img_len + context_txt_len, dim]
@@ -1584,11 +1586,64 @@ namespace WAN {
         }
     };
 
+    class VaceWanAttentionBlock : public WanAttentionBlock {
+    protected:
+        int block_id;
+        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
+            enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
+            params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 6, 1);
+        }
+
+    public:
+        VaceWanAttentionBlock(bool t2v_cross_attn,
+                              int64_t dim,
+                              int64_t ffn_dim,
+                              int64_t num_heads,
+                              bool qk_norm         = true,
+                              bool cross_attn_norm = false,
+                              float eps            = 1e-6,
+                              int block_id         = 0,
+                              bool flash_attn      = false)
+            : WanAttentionBlock(t2v_cross_attn, dim, ffn_dim, num_heads, qk_norm, cross_attn_norm, eps, flash_attn), block_id(block_id) {
+            if (block_id == 0) {
+                blocks["before_proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
+            }
+            blocks["after_proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim));
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(struct ggml_context* ctx,
+                                                      ggml_backend_t backend,
+                                                      struct ggml_tensor* c,
+                                                      struct ggml_tensor* x,
+                                                      struct ggml_tensor* e,
+                                                      struct ggml_tensor* pe,
+                                                      struct ggml_tensor* context,
+                                                      int64_t context_img_len = 257) {
+            // x: [N, n_token, dim]
+            // e: [N, 6, dim] or [N, T, 6, dim]
+            // context: [N, context_img_len + context_txt_len, dim]
+            // return [N, n_token, dim]
+            if (block_id == 0) {
+                auto before_proj = std::dynamic_pointer_cast<Linear>(blocks["before_proj"]);
+
+                c = before_proj->forward(ctx, c);
+                c = ggml_add(ctx, c, x);
+            }
+
+            auto after_proj = std::dynamic_pointer_cast<Linear>(blocks["after_proj"]);
+
+            c           = WanAttentionBlock::forward(ctx, backend, c, e, pe, context, context_img_len);
+            auto c_skip = after_proj->forward(ctx, c);
+
+            return {c_skip, c};
+        }
+    };
+
     class Head : public GGMLBlock {
     protected:
         int dim;
 
-        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
             enum ggml_type wtype = get_type(prefix + "weight", tensor_types, GGML_TYPE_F32);
             params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 2, 1);
         }
@@ -1635,7 +1690,7 @@ namespace WAN {
         int in_dim;
         int flf_pos_embed_token_number;
 
-        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
+        void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") override {
             if (flf_pos_embed_token_number > 0) {
                 params["emb_pos"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, in_dim, flf_pos_embed_token_number, 1);
             }
@@ -1680,22 +1735,25 @@ namespace WAN {
     };
 
     struct WanParams {
-        std::string model_type               = "t2v";
-        std::tuple<int, int, int> patch_size = {1, 2, 2};
-        int64_t text_len                     = 512;
-        int64_t in_dim                       = 16;
-        int64_t dim                          = 2048;
-        int64_t ffn_dim                      = 8192;
-        int64_t freq_dim                     = 256;
-        int64_t text_dim                     = 4096;
-        int64_t out_dim                      = 16;
-        int64_t num_heads                    = 16;
-        int64_t num_layers                   = 32;
-        bool qk_norm                         = true;
-        bool cross_attn_norm                 = true;
-        float eps                            = 1e-6;
-        int64_t flf_pos_embed_token_number   = 0;
-        int theta                            = 10000;
+        std::string model_type                 = "t2v";
+        std::tuple<int, int, int> patch_size   = {1, 2, 2};
+        int64_t text_len                       = 512;
+        int64_t in_dim                         = 16;
+        int64_t dim                            = 2048;
+        int64_t ffn_dim                        = 8192;
+        int64_t freq_dim                       = 256;
+        int64_t text_dim                       = 4096;
+        int64_t out_dim                        = 16;
+        int64_t num_heads                      = 16;
+        int64_t num_layers                     = 32;
+        int64_t vace_layers                    = 0;
+        int64_t vace_in_dim                    = 96;
+        std::map<int, int> vace_layers_mapping = {};
+        bool qk_norm                           = true;
+        bool cross_attn_norm                   = true;
+        float eps                              = 1e-6;
+        int64_t flf_pos_embed_token_number     = 0;
+        int theta                              = 10000;
         // wan2.1 1.3B: 1536/12, wan2.1/2.2 14B: 5120/40, wan2.2 5B: 3074/24
         std::vector<int> axes_dim = {44, 42, 42};
         int64_t axes_dim_sum      = 128;
@@ -1746,13 +1804,38 @@ namespace WAN {
             if (params.model_type == "i2v") {
                 blocks["img_emb"] = std::shared_ptr<GGMLBlock>(new MLPProj(1280, params.dim, params.flf_pos_embed_token_number));
             }
+
+            // vace
+            if (params.vace_layers > 0) {
+                for (int i = 0; i < params.vace_layers; i++) {
+                    auto block                                 = std::shared_ptr<GGMLBlock>(new VaceWanAttentionBlock(params.model_type == "t2v",
+                                                                                                                      params.dim,
+                                                                                                                      params.ffn_dim,
+                                                                                                                      params.num_heads,
+                                                                                                                      params.qk_norm,
+                                                                                                                      params.cross_attn_norm,
+                                                                                                                      params.eps,
+                                                                                                                      i,
+                                                                                                                      params.flash_attn));
+                    blocks["vace_blocks." + std::to_string(i)] = block;
+                }
+
+                int step = params.num_layers / params.vace_layers;
+                int n    = 0;
+                for (int i = 0; i < params.num_layers; i += step) {
+                    this->params.vace_layers_mapping[i] = n;
+                    n++;
+                }
+
+                blocks["vace_patch_embedding"] = std::shared_ptr<GGMLBlock>(new Conv3d(params.vace_in_dim, params.dim, params.patch_size, params.patch_size));
+            }
         }
 
         struct ggml_tensor* pad_to_patch_size(struct ggml_context* ctx,
                                               struct ggml_tensor* x) {
             int64_t W = x->ne[0];
             int64_t H = x->ne[1];
-            int64_t T = x->ne[1];
+            int64_t T = x->ne[2];
 
             int pad_t = (std::get<0>(params.patch_size) - T % std::get<0>(params.patch_size)) % std::get<0>(params.patch_size);
             int pad_h = (std::get<1>(params.patch_size) - H % std::get<1>(params.patch_size)) % std::get<1>(params.patch_size);
@@ -1795,9 +1878,12 @@ namespace WAN {
                                          struct ggml_tensor* timestep,
                                          struct ggml_tensor* context,
                                          struct ggml_tensor* pe,
-                                         struct ggml_tensor* clip_fea = NULL,
-                                         int64_t N                    = 1) {
+                                         struct ggml_tensor* clip_fea     = nullptr,
+                                         struct ggml_tensor* vace_context = nullptr,
+                                         float vace_strength              = 1.f,
+                                         int64_t N                        = 1) {
             // x: [N*C, T, H, W], C => in_dim
+            // vace_context: [N*vace_in_dim, T, H, W]
             // timestep: [N,] or [T]
             // context: [N, L, text_dim]
             // return: [N, t_len*h_len*w_len, out_dim*pt*ph*pw]
@@ -1836,7 +1922,7 @@ namespace WAN {
             context = text_embedding_2->forward(ctx, context);  // [N, context_txt_len, dim]
 
             int64_t context_img_len = 0;
-            if (clip_fea != NULL) {
+            if (clip_fea != nullptr) {
                 if (params.model_type == "i2v") {
                     auto img_emb     = std::dynamic_pointer_cast<MLPProj>(blocks["img_emb"]);
                     auto context_img = img_emb->forward(ctx, clip_fea);            // [N, context_img_len, dim]
@@ -1845,10 +1931,35 @@ namespace WAN {
                 context_img_len = clip_fea->ne[1];  // 257
             }
 
+            // vace_patch_embedding
+            ggml_tensor* c = nullptr;
+            if (params.vace_layers > 0) {
+                auto vace_patch_embedding = std::dynamic_pointer_cast<Conv3d>(blocks["vace_patch_embedding"]);
+
+                c = vace_patch_embedding->forward(ctx, vace_context);                          // [N*dim, t_len, h_len, w_len]
+                c = ggml_reshape_3d(ctx, c, c->ne[0] * c->ne[1] * c->ne[2], c->ne[3] / N, N);  // [N, dim, t_len*h_len*w_len]
+                c = ggml_nn_cont(ctx, ggml_torch_permute(ctx, c, 1, 0, 2, 3));                 // [N, t_len*h_len*w_len, dim]
+            }
+
+            auto x_orig = x;
+
             for (int i = 0; i < params.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<WanAttentionBlock>(blocks["blocks." + std::to_string(i)]);
 
                 x = block->forward(ctx, backend, x, e0, pe, context, context_img_len);
+
+                auto iter = params.vace_layers_mapping.find(i);
+                if (iter != params.vace_layers_mapping.end()) {
+                    int n = iter->second;
+
+                    auto vace_block = std::dynamic_pointer_cast<VaceWanAttentionBlock>(blocks["vace_blocks." + std::to_string(n)]);
+
+                    auto result = vace_block->forward(ctx, backend, c, x_orig, e0, pe, context, context_img_len);
+                    auto c_skip = result.first;
+                    c           = result.second;
+                    c_skip      = ggml_scale(ctx, c_skip, vace_strength);
+                    x           = ggml_add(ctx, x, c_skip);
+                }
             }
 
             x = head->forward(ctx, x, e);  // [N, t_len*h_len*w_len, pt*ph*pw*out_dim]
@@ -1862,8 +1973,10 @@ namespace WAN {
                                     struct ggml_tensor* timestep,
                                     struct ggml_tensor* context,
                                     struct ggml_tensor* pe,
-                                    struct ggml_tensor* clip_fea        = NULL,
-                                    struct ggml_tensor* time_dim_concat = NULL,
+                                    struct ggml_tensor* clip_fea        = nullptr,
+                                    struct ggml_tensor* time_dim_concat = nullptr,
+                                    struct ggml_tensor* vace_context    = nullptr,
+                                    float vace_strength                 = 1.f,
                                     int64_t N                           = 1) {
             // Forward pass of DiT.
             // x: [N*C, T, H, W]
@@ -1886,13 +1999,13 @@ namespace WAN {
             int64_t h_len = ((H + (std::get<1>(params.patch_size) / 2)) / std::get<1>(params.patch_size));
             int64_t w_len = ((W + (std::get<2>(params.patch_size) / 2)) / std::get<2>(params.patch_size));
 
-            if (time_dim_concat != NULL) {
+            if (time_dim_concat != nullptr) {
                 time_dim_concat = pad_to_patch_size(ctx, time_dim_concat);
                 x               = ggml_concat(ctx, x, time_dim_concat, 2);  // [N*C, (T+pad_t) + (T2+pad_t2), H + pad_h, W + pad_w]
                 t_len           = ((x->ne[2] + (std::get<0>(params.patch_size) / 2)) / std::get<0>(params.patch_size));
             }
 
-            auto out = forward_orig(ctx, backend, x, timestep, context, pe, clip_fea, N);  // [N, t_len*h_len*w_len, pt*ph*pw*C]
+            auto out = forward_orig(ctx, backend, x, timestep, context, pe, clip_fea, vace_context, vace_strength, N);  // [N, t_len*h_len*w_len, pt*ph*pw*C]
 
             out = unpatchify(ctx, out, t_len, h_len, w_len);  // [N*C, (T+pad_t) + (T2+pad_t2), H + pad_h, W + pad_w]
 
@@ -1927,7 +2040,19 @@ namespace WAN {
                 std::string tensor_name = pair.first;
                 if (tensor_name.find(prefix) == std::string::npos)
                     continue;
-                size_t pos = tensor_name.find("blocks.");
+                size_t pos = tensor_name.find("vace_blocks.");
+                if (pos != std::string::npos) {
+                    tensor_name = tensor_name.substr(pos);  // remove prefix
+                    auto items  = split_string(tensor_name, '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        if (block_index + 1 > wan_params.vace_layers) {
+                            wan_params.vace_layers = block_index + 1;
+                        }
+                    }
+                    continue;
+                }
+                pos = tensor_name.find("blocks.");
                 if (pos != std::string::npos) {
                     tensor_name = tensor_name.substr(pos);  // remove prefix
                     auto items  = split_string(tensor_name, '.');
@@ -1937,6 +2062,7 @@ namespace WAN {
                             wan_params.num_layers = block_index + 1;
                         }
                     }
+                    continue;
                 }
                 if (tensor_name.find("img_emb") != std::string::npos) {
                     wan_params.model_type = "i2v";
@@ -1958,7 +2084,11 @@ namespace WAN {
                     wan_params.out_dim   = 48;
                     wan_params.text_len  = 512;
                 } else {
-                    desc                 = "Wan2.1-T2V-1.3B";
+                    if (wan_params.vace_layers > 0) {
+                        desc = "Wan2.1-VACE-1.3B";
+                    } else {
+                        desc = "Wan2.1-T2V-1.3B";
+                    }
                     wan_params.dim       = 1536;
                     wan_params.eps       = 1e-06;
                     wan_params.ffn_dim   = 8960;
@@ -1974,7 +2104,11 @@ namespace WAN {
                         desc              = "Wan2.2-I2V-14B";
                         wan_params.in_dim = 36;
                     } else {
-                        desc              = "Wan2.x-T2V-14B";
+                        if (wan_params.vace_layers > 0) {
+                            desc = "Wan2.x-VACE-14B";
+                        } else {
+                            desc = "Wan2.x-T2V-14B";
+                        }
                         wan_params.in_dim = 16;
                     }
                 } else {
@@ -2002,7 +2136,7 @@ namespace WAN {
             wan.init(params_ctx, tensor_types, prefix);
         }
 
-        std::string get_desc() {
+        std::string get_desc() override {
             return desc;
         }
 
@@ -2013,9 +2147,11 @@ namespace WAN {
         struct ggml_cgraph* build_graph(struct ggml_tensor* x,
                                         struct ggml_tensor* timesteps,
                                         struct ggml_tensor* context,
-                                        struct ggml_tensor* clip_fea        = NULL,
-                                        struct ggml_tensor* c_concat        = NULL,
-                                        struct ggml_tensor* time_dim_concat = NULL) {
+                                        struct ggml_tensor* clip_fea        = nullptr,
+                                        struct ggml_tensor* c_concat        = nullptr,
+                                        struct ggml_tensor* time_dim_concat = nullptr,
+                                        struct ggml_tensor* vace_context    = nullptr,
+                                        float vace_strength                 = 1.f) {
             struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, WAN_GRAPH_SIZE, false);
 
             x               = to_backend(x);
@@ -2024,6 +2160,7 @@ namespace WAN {
             clip_fea        = to_backend(clip_fea);
             c_concat        = to_backend(c_concat);
             time_dim_concat = to_backend(time_dim_concat);
+            vace_context    = to_backend(vace_context);
 
             pe_vec      = Rope::gen_wan_pe(x->ne[2],
                                            x->ne[1],
@@ -2039,10 +2176,10 @@ namespace WAN {
             auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, wan_params.axes_dim_sum / 2, pos_len);
             // pe->data = pe_vec.data();
             // print_ggml_tensor(pe);
-            // pe->data = NULL;
+            // pe->data = nullptr;
             set_backend_tensor_data(pe, pe_vec.data());
 
-            if (c_concat != NULL) {
+            if (c_concat != nullptr) {
                 x = ggml_concat(compute_ctx, x, c_concat, 3);
             }
 
@@ -2053,7 +2190,9 @@ namespace WAN {
                                                   context,
                                                   pe,
                                                   clip_fea,
-                                                  time_dim_concat);
+                                                  time_dim_concat,
+                                                  vace_context,
+                                                  vace_strength);
 
             ggml_build_forward_expand(gf, out);
 
@@ -2064,13 +2203,15 @@ namespace WAN {
                      struct ggml_tensor* x,
                      struct ggml_tensor* timesteps,
                      struct ggml_tensor* context,
-                     struct ggml_tensor* clip_fea        = NULL,
-                     struct ggml_tensor* c_concat        = NULL,
-                     struct ggml_tensor* time_dim_concat = NULL,
-                     struct ggml_tensor** output         = NULL,
-                     struct ggml_context* output_ctx     = NULL) {
+                     struct ggml_tensor* clip_fea        = nullptr,
+                     struct ggml_tensor* c_concat        = nullptr,
+                     struct ggml_tensor* time_dim_concat = nullptr,
+                     struct ggml_tensor* vace_context    = nullptr,
+                     float vace_strength                 = 1.f,
+                     struct ggml_tensor** output         = nullptr,
+                     struct ggml_context* output_ctx     = nullptr) {
             auto get_graph = [&]() -> struct ggml_cgraph* {
-                return build_graph(x, timesteps, context, clip_fea, c_concat, time_dim_concat);
+                return build_graph(x, timesteps, context, clip_fea, c_concat, time_dim_concat, vace_context, vace_strength);
             };
 
             GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
@@ -2079,11 +2220,11 @@ namespace WAN {
         void test() {
             struct ggml_init_params params;
             params.mem_size   = static_cast<size_t>(200 * 1024 * 1024);  // 200 MB
-            params.mem_buffer = NULL;
+            params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
             struct ggml_context* work_ctx = ggml_init(params);
-            GGML_ASSERT(work_ctx != NULL);
+            GGML_ASSERT(work_ctx != nullptr);
 
             {
                 // cpu f16: pass
@@ -2105,10 +2246,10 @@ namespace WAN {
                 // auto clip_fea = load_tensor_from_file(work_ctx, "wan_dit_clip_fea.bin");
                 // print_ggml_tensor(clip_fea);
 
-                struct ggml_tensor* out = NULL;
+                struct ggml_tensor* out = nullptr;
 
                 int t0 = ggml_time_ms();
-                compute(8, x, timesteps, context, NULL, NULL, NULL, &out, work_ctx);
+                compute(8, x, timesteps, context, nullptr, nullptr, nullptr, nullptr, 1.f, &out, work_ctx);
                 int t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);
@@ -2136,12 +2277,12 @@ namespace WAN {
                 }
             }
 
-            std::shared_ptr<WanRunner> wan = std::shared_ptr<WanRunner>(new WanRunner(backend,
-                                                                                      false,
-                                                                                      tensor_types,
-                                                                                      "model.diffusion_model",
-                                                                                      VERSION_WAN2_2_TI2V,
-                                                                                      true));
+            std::shared_ptr<WanRunner> wan = std::make_shared<WanRunner>(backend,
+                                                                         false,
+                                                                         tensor_types,
+                                                                         "model.diffusion_model",
+                                                                         VERSION_WAN2_2_TI2V,
+                                                                         true);
 
             wan->alloc_params_buffer();
             std::map<std::string, ggml_tensor*> tensors;