feat: support disk params backend (#1651)

Signed-off-by: kei-g <km.8k6ce+github@gmail.com>

.github/workflows/build.yml

@@ -14,6 +14,8 @@ on:
     paths:
       [
         ".github/workflows/**",
+        ".dockerignore",
+        "Dockerfile*",
         "**/CMakeLists.txt",
         "**/Makefile",
         "**/*.h",
@@ -29,6 +31,8 @@ on:
     paths:
       [
         ".github/workflows/**",
+        ".dockerignore",
+        "Dockerfile*",
         "**/CMakeLists.txt",
         "**/Makefile",
         "**/*.h",
@@ -177,7 +181,7 @@ jobs:
   build-and-push-docker-images:
     name: Build and push container images
     if: ${{ github.event_name != 'pull_request' }}
-    runs-on: ubuntu-latest
+    runs-on: ${{ matrix.runner }}
 
     permissions:
       contents: read
@@ -189,6 +193,20 @@ jobs:
     strategy:
       matrix:
         variant: [musa, sycl, vulkan, cuda]
+        platform: [linux/amd64]
+        runner: [ubuntu-latest]
+        build-args: [""]
+        tag-suffix: [""]
+        include:
+          - variant: cuda
+            platform: linux/arm64
+            runner: ubuntu-24.04-arm
+            tag-suffix: "-spark"
+            build-args: |
+              CUDA_VERSION=13.0.0
+              UBUNTU_VERSION=24.04
+              CUDA_ARCHITECTURES=121
+              GGML_CUDA_FA_ALL_QUANTS=ON
 
     env:
       REGISTRY: ghcr.io
@@ -243,12 +261,13 @@ jobs:
         uses: docker/build-push-action@v6
         with:
           context: .
-          platforms: linux/amd64
+          platforms: ${{ matrix.platform }}
           push: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
           file: Dockerfile.${{ matrix.variant }}
-          tags: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ env.BRANCH_NAME }}-${{ matrix.variant }}
+          tags: ${{ env.REGISTRY }}/${{ env.IMAGE_NAME }}:${{ env.BRANCH_NAME }}-${{ matrix.variant }}${{ matrix.tag-suffix }}
           labels: ${{ steps.meta.outputs.labels }}
           annotations: ${{ steps.meta.outputs.annotations }}
+          build-args: ${{ matrix.build-args }}
 
   macOS-latest-cmake:
     runs-on: macos-latest
@@ -449,13 +468,23 @@ jobs:
 
     env:
       ROCM_VERSION: "7.13.0"
-      GPU_TARGETS: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201"
+      GPU_TARGETS: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
 
     steps:
       - uses: actions/checkout@v3
         with:
           submodules: recursive
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
       - name: Cache ROCm Installation
         id: cache-rocm
         uses: actions/cache@v4
@@ -516,6 +545,30 @@ jobs:
       - name: Pack artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         run: |
+          $ErrorActionPreference = "Stop"
+          $dst = "build\bin"
+          $rocmBin = Join-Path "${env:HIP_PATH}" "bin"
+          $requiredRocmPaths = @(
+            (Join-Path $rocmBin "rocblas.dll"),
+            (Join-Path $rocmBin "rocblas\library")
+          )
+          foreach ($path in $requiredRocmPaths) {
+            if (!(Test-Path $path)) {
+              throw "Missing ROCm runtime dependency: $path"
+            }
+          }
+
+          foreach ($pattern in @("rocblas*.dll", "hipblas*.dll", "libhipblas*.dll")) {
+            Copy-Item -Path (Join-Path $rocmBin $pattern) -Destination $dst -Force -ErrorAction SilentlyContinue
+          }
+
+          foreach ($dir in @("rocblas", "hipblaslt")) {
+            $src = Join-Path $rocmBin $dir
+            if (Test-Path $src) {
+              Copy-Item -Path $src -Destination $dst -Recurse -Force
+            }
+          }
+
           7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip .\build\bin\*
 
       - name: Upload artifacts
@@ -648,7 +701,7 @@ jobs:
             gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1031;gfx1032;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201"
             build: 'x64'
           - ROCM_VERSION: "7.13.0"
-            gpu_targets: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201"
+            gpu_targets: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
             build: x64
 
     steps:
@@ -658,16 +711,6 @@ jobs:
         with:
           submodules: recursive
 
-      - name: Setup Node
-        uses: actions/setup-node@v4
-        with:
-          node-version: 20
-
-      - name: Setup pnpm
-        uses: pnpm/action-setup@v4
-        with:
-          version: 10.15.1
-
       - name: ccache
         uses: ggml-org/ccache-action@v1.2.16
         with:
@@ -725,6 +768,18 @@ jobs:
           echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
           echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
 
+      # setup-node installs into /opt/hostedtoolcache, which is removed above.
+      # Keep Node/pnpm setup after disk cleanup so the server frontend can be embedded.
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 10.15.1
+
       - name: Build
         id: cmake_build
         run: |

.github/workflows/stale-prs.yml

@@ -0,0 +1,55 @@
+name: Close inactive PRs
+
+on:
+  schedule:
+    # Run daily. GitHub cron schedules use UTC.
+    - cron: "30 1 * * *"
+  workflow_dispatch:
+    inputs:
+      debug_only:
+        description: "Dry run: log intended actions without changing PRs"
+        required: false
+        default: false
+        type: boolean
+
+permissions:
+  issues: write
+  pull-requests: write
+
+concurrency:
+  group: ${{ github.workflow }}
+  cancel-in-progress: false
+
+jobs:
+  stale-prs:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Mark and close inactive PRs
+        uses: actions/stale@v10
+        with:
+          days-before-issue-stale: -1
+          days-before-issue-close: -1
+
+          days-before-pr-stale: 365
+          days-before-pr-close: 7
+
+          stale-pr-label: pr:inactive
+          close-pr-label: pr:auto-closed
+          exempt-pr-labels: pr:keep-open
+
+          stale-pr-message: >
+            This PR has been inactive for 365 days. If there is no new activity
+            within 7 days, it will be closed automatically. Comment, push new
+            commits, or remove the pr:inactive label to keep it open. Add
+            pr:keep-open to exempt it from future inactive PR cleanup.
+
+          close-pr-message: >
+            Closing this PR because it has had no activity for 7 days after
+            being marked inactive. If this is still useful or ready to move
+            forward, feel free to reopen it with fresh context or updated
+            details. Sorry for any inconvenience.
+
+          remove-pr-stale-when-updated: true
+          delete-branch: false
+          operations-per-run: 100
+          debug-only: ${{ github.event_name == 'workflow_dispatch' && inputs.debug_only || false }}

CMakeLists.txt

@@ -22,6 +22,31 @@ endif()
 set(CMAKE_LIBRARY_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 set(CMAKE_RUNTIME_OUTPUT_DIRECTORY ${CMAKE_BINARY_DIR}/bin)
 
+if(APPLE)
+    function(sd_set_macos_rpaths target)
+        get_target_property(target_type ${target} TYPE)
+        if(target_type STREQUAL "EXECUTABLE")
+            set(runtime_paths "@executable_path" "@executable_path/../lib")
+        elseif(target_type STREQUAL "SHARED_LIBRARY" OR target_type STREQUAL "MODULE_LIBRARY")
+            set(runtime_paths "@loader_path" "@loader_path/../lib")
+            set_target_properties(${target} PROPERTIES
+                MACOSX_RPATH ON
+                INSTALL_NAME_DIR "@rpath"
+                BUILD_WITH_INSTALL_NAME_DIR ON
+            )
+        else()
+            return()
+        endif()
+
+        # Release artifacts zip the build output directly, so keep macOS rpaths relocatable.
+        set_target_properties(${target} PROPERTIES
+            BUILD_RPATH "${runtime_paths}"
+            INSTALL_RPATH "${runtime_paths}"
+            BUILD_WITH_INSTALL_RPATH ON
+        )
+    endfunction()
+endif()
+
 if(CMAKE_SOURCE_DIR STREQUAL CMAKE_CURRENT_SOURCE_DIR)
     set(SD_STANDALONE ON)
 else()
@@ -100,6 +125,12 @@ endif ()
 if (SD_HIPBLAS)
     message("-- Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIP ON)
+    # ggml-hip's device-stub objects must be position-independent, or the
+    # default-PIE sd-cli link fails with `relocation R_X86_64_32 ... cannot be
+    # used when making a PIE object` on distros that default to PIE
+    # (Ubuntu 24.04, Fedora 40+, Debian 12+). The shared-library branch below
+    # already sets this; the static build (the HIP default) did not.
+    set(CMAKE_POSITION_INDEPENDENT_CODE ON)
 endif ()
 
 if(SD_MUSA)
@@ -179,6 +210,21 @@ file(GLOB SD_LIB_SOURCES CONFIGURE_DEPENDS
     "src/*.h"
     "src/*.cpp"
     "src/*.hpp"
+    "src/conditioning/*.h"
+    "src/conditioning/*.cpp"
+    "src/conditioning/*.hpp"
+    "src/core/*.h"
+    "src/core/*.cpp"
+    "src/core/*.hpp"
+    "src/extensions/*.h"
+    "src/extensions/*.cpp"
+    "src/extensions/*.hpp"
+    "src/model/*/*.h"
+    "src/model/*/*.cpp"
+    "src/model/*/*.hpp"
+    "src/runtime/*.h"
+    "src/runtime/*.cpp"
+    "src/runtime/*.hpp"
     "src/model_io/*.h"
     "src/model_io/*.cpp"
     "src/tokenizers/*.h"
@@ -237,6 +283,10 @@ else()
     add_library(${SD_LIB} STATIC ${SD_LIB_SOURCES})
 endif()
 
+if(APPLE)
+    sd_set_macos_rpaths(${SD_LIB})
+endif()
+
 if(SD_SYCL)
     message("-- Use SYCL as backend stable-diffusion")
     set(GGML_SYCL ON)
@@ -277,6 +327,7 @@ add_subdirectory(thirdparty)
 
 target_link_libraries(${SD_LIB} PUBLIC ggml zip)
 target_include_directories(${SD_LIB} PUBLIC . src include)
+target_include_directories(${SD_LIB} PRIVATE src/core)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)
 target_compile_features(${SD_LIB} PUBLIC c_std_11 cxx_std_17)
 

CONTRIBUTING.md

@@ -44,6 +44,8 @@ Naming conventions:
 
 Some older code in the project may not fully follow the current conventions. Please do not submit PRs that only rewrite existing code to match style rules.
 
+When adding or modifying model implementations, follow the model config and weight detection conventions in [docs/model_config.md](docs/model_config.md).
+
 ## AI-Assisted Contributions
 
 AI tools may be used to assist development, but contributors are responsible for the quality and correctness of the submitted code.

Dockerfile

@@ -2,7 +2,18 @@ ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
-RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
 
 WORKDIR /sd.cpp
 
@@ -20,4 +31,4 @@ RUN apt-get update && \
 COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
 COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
 
-ENTRYPOINT [ "/sd-cli" ]
+ENTRYPOINT [ "/sd-cli" ]

Dockerfile.cuda

@@ -3,14 +3,31 @@ ARG UBUNTU_VERSION=24.04
 
 FROM nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu${UBUNTU_VERSION} AS build
 
-RUN apt-get update && apt-get install -y --no-install-recommends build-essential git ccache cmake
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git ccache cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
 
 WORKDIR /sd.cpp
 
 COPY . .
 
 ARG CUDACXX=/usr/local/cuda/bin/nvcc
-RUN cmake . -B ./build -DSD_CUDA=ON
+ARG CUDA_ARCHITECTURES=""
+ARG GGML_CUDA_FA_ALL_QUANTS=""
+
+RUN cmake . -B ./build \
+    -DSD_CUDA=ON \
+    ${CUDA_ARCHITECTURES:+-DCMAKE_CUDA_ARCHITECTURES="${CUDA_ARCHITECTURES}"} \
+    ${GGML_CUDA_FA_ALL_QUANTS:+-DGGML_CUDA_FA_ALL_QUANTS=${GGML_CUDA_FA_ALL_QUANTS}}
 RUN cmake --build ./build --config Release -j$(nproc)
 
 FROM nvidia/cuda:${CUDA_VERSION}-cudnn-runtime-ubuntu${UBUNTU_VERSION} AS runtime

Dockerfile.musa

@@ -3,7 +3,18 @@ ARG UBUNTU_VERSION=22.04
 
 FROM mthreads/musa:${MUSA_VERSION}-devel-ubuntu${UBUNTU_VERSION}-amd64 as build
 
-RUN apt-get update && apt-get install -y ccache cmake git
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends ccache cmake git ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
 
 WORKDIR /sd.cpp
 
@@ -21,4 +32,4 @@ FROM mthreads/musa:${MUSA_VERSION}-runtime-ubuntu${UBUNTU_VERSION}-amd64 as runt
 COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
 COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
 
-ENTRYPOINT [ "/sd-cli" ]
+ENTRYPOINT [ "/sd-cli" ]

Dockerfile.sycl

@@ -1,8 +1,20 @@
-ARG SYCL_VERSION=2025.1.0-0
+# ggml SYCL hardware detection uses BMG G31/WCL architecture enums added in oneAPI 2025.3.
+ARG SYCL_VERSION=2025.3.2-0
 
 FROM intel/oneapi-basekit:${SYCL_VERSION}-devel-ubuntu24.04 AS build
 
-RUN apt-get update && apt-get install -y cmake
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends cmake ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
 
 WORKDIR /sd.cpp
 

Dockerfile.vulkan

@@ -2,7 +2,18 @@ ARG UBUNTU_VERSION=24.04
 
 FROM ubuntu:$UBUNTU_VERSION AS build
 
-RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake libvulkan-dev glslc spirv-headers
+# sd-server embeds the web UI at build time, so the build image needs Node/pnpm.
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git cmake libvulkan-dev glslc spirv-headers ca-certificates curl gnupg && \
+    mkdir -p /etc/apt/keyrings && \
+    curl -fsSL https://deb.nodesource.com/gpgkey/nodesource-repo.gpg.key -o /tmp/nodesource-repo.gpg.key && \
+    gpg --dearmor -o /etc/apt/keyrings/nodesource.gpg /tmp/nodesource-repo.gpg.key && \
+    rm /tmp/nodesource-repo.gpg.key && \
+    echo "deb [signed-by=/etc/apt/keyrings/nodesource.gpg] https://deb.nodesource.com/node_20.x nodistro main" > /etc/apt/sources.list.d/nodesource.list && \
+    apt-get update && \
+    apt-get install -y --no-install-recommends nodejs && \
+    npm install -g pnpm@10.15.1 && \
+    apt-get clean && \
+    rm -rf /var/lib/apt/lists/*
 
 WORKDIR /sd.cpp
 

README.md

@@ -15,29 +15,18 @@ API and command-line option may change frequently.***
 
 ## 🔥Important News
 
+* **2026/06/04** 🚀 stable-diffusion.cpp now supports **Ideogram4**
+* **2026/05/31** 🚀 stable-diffusion.cpp now supports **PiD**
+* **2026/05/27** 🚀 stable-diffusion.cpp now supports **Lens**
+* **2026/05/17** 🚀 stable-diffusion.cpp now supports **LTX-2.3**
 * **2026/04/11** 🚀 stable-diffusion.cpp now uses a brand-new embedded web UI.  
-  👉 Details: [PR #1408](https://github.com/leejet/stable-diffusion.cpp/pull/1408)
-
 * **2026/01/18** 🚀 stable-diffusion.cpp now supports **FLUX.2-klein**  
-  👉 Details: [PR #1193](https://github.com/leejet/stable-diffusion.cpp/pull/1193)
-
 * **2025/12/01** 🚀 stable-diffusion.cpp now supports **Z-Image**  
-  👉 Details: [PR #1020](https://github.com/leejet/stable-diffusion.cpp/pull/1020)
-
 * **2025/11/30** 🚀 stable-diffusion.cpp now supports **FLUX.2-dev**  
-  👉 Details: [PR #1016](https://github.com/leejet/stable-diffusion.cpp/pull/1016)
-
 * **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
 
@@ -51,17 +40,22 @@ API and command-line option may change frequently.***
     - [SD3/SD3.5](./docs/sd3.md)
     - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
     - [FLUX.2-dev/FLUX.2-klein](./docs/flux2.md)
+    - [Lens](./docs/lens.md)
     - [Chroma](./docs/chroma.md)
     - [Chroma1-Radiance](./docs/chroma_radiance.md)
     - [Qwen Image](./docs/qwen_image.md)
+    - [PiD](./docs/pid.md)
+    - [LongCat Image](./docs/longcat_image.md)
     - [Z-Image](./docs/z_image.md)
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
     - [ERNIE-Image](./docs/ernie_image.md)
     - [HiDream-O1-Image](./docs/hidream_o1_image.md)
+    - [Ideogram4](./docs/ideogram4.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
+    - [LongCat Image Edit](./docs/longcat_image.md)
   - Video Models
     - [Wan2.1/Wan2.2](./docs/wan.md)
     - [LTX-2.3](./docs/ltx2.md)
@@ -154,6 +148,8 @@ For runtime and parameter backend placement, see the [backend selection guide](.
 - [Anima](./docs/anima.md)
 - [ERNIE-Image](./docs/ernie_image.md)
 - [HiDream-O1-Image](./docs/hidream_o1_image.md)
+- [Lens](./docs/lens.md)
+- [LongCat Image / LongCat Image Edit](./docs/longcat_image.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)

docs/backend.md

@@ -3,7 +3,7 @@
 `stable-diffusion.cpp` has two backend assignments:
 
 - `--backend` selects the runtime backend used to execute model graphs.
-- `--params-backend` selects the backend used to allocate model parameters.
+- `--params-backend` selects where model parameters are kept.
 
 If `--params-backend` is not set, parameters use the same backend as their module runtime backend.
 
@@ -29,6 +29,12 @@ The same syntax is used for parameter placement:
 sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend te=cpu,vae=cpu
 ```
 
+`--params-backend` also accepts the special value `disk`:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend disk
+```
+
 Module names are case-insensitive. Hyphens and underscores in module names are ignored, so `clip_vision`, `clip-vision`, and `clipvision` are equivalent.
 
 `all=`, `default=`, and `*=` can be used to set the default backend inside a mixed assignment:
@@ -64,9 +70,11 @@ The special values `auto`, `default`, and an empty backend name select the defau
 
 The special value `gpu` selects the first GPU backend, falling back to the first integrated GPU backend.
 
+The special value `disk` is accepted only by `--params-backend`. `--backend disk` is invalid because `disk` is a parameter residency mode, not a runtime compute backend.
+
 ## Runtime backend vs. parameter backend
 
-The runtime backend controls where graph execution runs. The parameter backend controls where model weights are allocated.
+The runtime backend controls where graph execution runs. The parameter backend controls where model weights are allocated or whether they are reloaded from disk on demand.
 
 For example:
 
@@ -76,6 +84,16 @@ sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend cpu
 
 This runs all modules on `cuda0`, but stores parameters in CPU RAM. During execution, parameters are moved to the runtime backend as needed.
 
+For example:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend disk
+```
+
+This runs all modules on `cuda0`, reloads parameters from the model file as needed, and releases those parameter buffers after use.
+
+`disk` is never selected implicitly. If `--params-backend` is not set, parameters use the runtime backend.
+
 Per-module assignments can be mixed:
 
 ```shell
@@ -100,6 +118,8 @@ uses one shared CPU backend for both `te` and `vae` runtime execution.
 
 Runtime and parameter assignments also share the same backend cache. If `--backend diffusion=cuda0` and `--params-backend diffusion=cuda0` resolve to the same device, both use the same backend instance.
 
+`--params-backend disk` does not create a separate backend instance. Parameters are loaded lazily using the module runtime backend.
+
 `SDBackendManager` owns the backend instances and frees them when the context or upscaler is destroyed. Model runners receive non-owning runtime and parameter backend pointers and do not free them.
 
 ## Compatibility flags
@@ -113,10 +133,12 @@ The older CPU placement flags are still supported:
 
 `--clip-on-cpu`, `--vae-on-cpu`, and `--control-net-cpu` affect runtime backend assignment only when `--backend` is not set. They map to `te=cpu`, `vae=cpu`, and `controlnet=cpu`.
 
-`--offload-to-cpu` affects parameter backend assignment only when `--params-backend` is not set. It is equivalent to:
+`--offload-to-cpu` prepends a CPU default to the parameter assignment before parsing:
 
 ```shell
---params-backend cpu
+--params-backend '*=cpu'
 ```
 
+Because this default is inserted first, later explicit `--params-backend` entries can still override it, for example `--offload-to-cpu --params-backend te=disk` keeps non-TE parameters on CPU and reloads TE parameters from disk.
+
 Explicit `--backend` and `--params-backend` assignments are preferred for new commands.

docs/ideogram4.md

@@ -0,0 +1,40 @@
+# How to Use
+
+## Download weights
+
+- Download Ideogram4
+    - safetensors: https://huggingface.co/ideogram-ai/ideogram-4-fp8/tree/main/transformer
+- Download Ideogram4 uncond
+    - safetensors: https://huggingface.co/ideogram-ai/ideogram-4-fp8/tree/main/unconditional_transformer
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download Qwen3-VL-8B-Instruct
+    - gguf: https://huggingface.co/unsloth/Qwen3-VL-8B-Instruct-GGUF/tree/main
+
+## Convert weights
+
+fp8 scale -> bf16
+
+```
+python .\convert_fp8_scale_to_bf16.py --input .\ideogram4_fp8.safetensors --output ideogram4_bf16.safetensors
+python .\convert_fp8_scale_to_bf16.py --input .\ideogram4_uncond_fp8.safetensors --output ideogram4_uncond_bf16.safetensors
+```
+
+bf16 -> q8
+
+```
+.\bin\Release\sd-cli.exe -M convert -m ideogram4_bf16.safetensors -o ideogram4-Q8_0.gguf --tensor-type-rules "^layers.*adaln_modulation.*weight=q8_0,layers.*attention.o.*weight=q8_0,layers.*attention.qkv.*weight=q8_0,layers.*feed_forward.*weight=q8_0" -v
+
+.\bin\Release\sd-cli.exe -M convert -m ideogram4_uncond_bf16.safetensors -o ideogram4_uncond-Q8_0.gguf --tensor-type-rules "^layers.*adaln_modulation.*weight=q8_0,layers.*attention.o.*weight=q8_0,layers.*attention.qkv.*weight=q8_0,layers.*feed_forward.*weight=q8_0" -v
+```
+
+If you want lower VRAM usage, you can change the quantization from q8_0 to a lower-level quantization, such as q4_0.
+
+
+## Examples
+
+```sh
+.\bin\Release\sd-cli.exe --diffusion-model ideogram4-Q8_0.gguf --uncond-diffusion-model ideogram4_uncond-Q8_0.gguf --llm ..\..\llm\Qwen3VL-8B-Instruct-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors -p '{"high_level_description":"A square 1024 x 1024 luxury fashion magazine cover featuring exactly one short chubby fluffy cat as the main model. The cat sits on a soft ivory studio floor, facing the viewer with a stylish calm expression, wearing tiny black sunglasses, a red silk scarf, and a small gold collar charm. In front of the cat on the floor is a wide horizontal luxury nameplate that clearly reads ideogram4.cpp. The whole design feels premium, fashionable, clean, and editorial.","style_description":{"aesthetics":"luxury fashion magazine cover, high-end pet couture campaign, minimalist editorial design, elegant studio photography, soft paper texture, refined typography, fashionable and polished","lighting":"Soft diffused studio lighting, gentle spotlight on the cat, subtle floor shadow, warm ivory highlights, clean separation between subject and background","photo":"high-resolution fashion editorial photography look, front-facing cat portrait, crisp fur details, glossy sunglasses, clear readable nameplate text, shallow depth of field","medium":"mixed media fashion photography and premium editorial graphic design","color_palette":["#F4EFE7","#111111","#D8B56D","#B73A3A","#FFFFFF","#8A7A6A"]},"compositional_deconstruction":{"canvas":"Square 1024 x 1024 canvas with a normal upright orientation. Do not rotate the poster or any text. Use a clean fashion magazine cover layout.","background":"Warm ivory studio backdrop with subtle paper grain, a soft spotlight gradient, faint floor shadow, and a few minimal gold editorial lines. The background is spacious, premium, and uncluttered.","layout":"Top center has a small elegant headline. Center area features one cat as the main fashion model. Lower foreground has a wide horizontal luxury nameplate placed on the floor in front of the cat. Bottom center has a small footer. All text is horizontal, upright, and readable left to right.","elements":[{"type":"text","desc":"Top center headline reading LOOK WHAT I FOUND in a refined high-fashion serif font. The headline is horizontal, centered, elegant, and secondary to the nameplate text."},{"type":"obj","desc":"Exactly one short chubby fluffy cat sitting in the center like a luxury fashion model. The cat has a large round head, compact body, short legs, soft detailed fur, expressive eyes, and a calm confident pose. The cat is cute and rounded, not tall, not stretched, not duplicated."},{"type":"obj","desc":"Tiny glossy black sunglasses worn naturally by the cat, slightly oversized but still showing the cat face clearly. The sunglasses add a chic fashion-editorial attitude."},{"type":"obj","desc":"A red silk scarf tied neatly around the cat neck, with soft folds and a couture feeling. The scarf must not cover the cat face or the nameplate."},{"type":"obj","desc":"A small gold collar charm or fashion accessory under the scarf, subtle and premium, adding a luxury campaign detail."},{"type":"obj","desc":"In the lower foreground, place a wide horizontal luxury nameplate on the floor in front of the cat. The nameplate is low, flat, landscape-oriented, much wider than tall, like a fashion show seat card or premium display plaque. It is centered, front-facing, level, and fully visible. It must not become vertical, tall, standing, rotated, or side-facing."},{"type":"text","desc":"Print the exact text ideogram4.cpp only on the wide horizontal nameplate. Use clean bold black lettering, perfectly spelled, lowercase, with the number 4 and .cpp extension. The text must fit completely inside the nameplate, stay horizontal, and be readable from left to right."},{"type":"obj","desc":"Add sparse premium editorial accents around the edges: thin gold lines, small code brackets, tiny cursor marks, subtle dots, and minimal geometric details. No extra cats, no stickers, no animal faces, no busy decorations."},{"type":"text","desc":"Bottom center footer reading tiny paws, big compile energy in a small refined monospace or editorial font. The footer is horizontal, centered, understated, and much smaller than the nameplate text."}]}}'  --diffusion-fa -v --offload-to-cpu -H 1024 -W 1024
+```
+
+<img alt="ideogram4 image example" src="../assets/ideogram4/example.png" />

docs/lens.md

@@ -0,0 +1,32 @@
+# How to Use
+
+Lens uses a Lens diffusion transformer, the FLUX.2 VAE, and GPT-OSS-20B as the LLM text encoder.
+
+## Download weights
+
+- Download Lens
+    - safetensors: https://huggingface.co/Comfy-Org/Lens/tree/main/diffusion_models
+- Download Lens Turbo
+    - safetensors: https://huggingface.co/Comfy-Org/Lens/tree/main/diffusion_models
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.2-dev/tree/main
+- Download GPT-OSS-20B
+    - gguf: https://huggingface.co/unsloth/gpt-oss-20b-GGUF/tree/main
+
+## Examples
+
+### Lens
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\lens_bf16.safetensors --llm "..\..\llm\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors --cfg-scale 5.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v
+```
+
+<img width="256" alt="Lens example" src="../assets/lens/example.png" />
+
+### Lens Turbo
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\lens_turbo_bf16.safetensors --llm "..\..\llm\gpt-oss-20b-UD-Q8_K_XL.gguf" --vae ..\..\ComfyUI\models\vae\flux2_ae.safetensors --cfg-scale 1.0  -p "A crystal dragon soaring through an aurora borealis sky, its entire body made of transparent faceted crystal refracting the green and purple aurora light into rainbow spectra, ice particles trailing from its wings, high fantasy digital art" --diffusion-fa -v --steps 4
+```
+
+<img width="256" alt="Lens Turbo example" src="../assets/lens/turbo_example.png" />

docs/longcat_image.md

@@ -0,0 +1,30 @@
+# How to Use
+
+LongCat-Image uses a LongCat diffusion transformer, the FLUX VAE, and Qwen2.5-VL as the LLM text encoder.
+
+## Download weights
+
+- Download LongCat Image
+    - safetensors: https://huggingface.co/Comfy-Org/LongCat-Image/tree/main/split_files/diffusion_models
+    - gguf: https://huggingface.co/vantagewithai/LongCat-Image-GGUF/tree/main/comfy
+- Download LongCat Image Edit
+    - LongCat Image Edit Turbo: https://huggingface.co/meituan-longcat/LongCat-Image-Edit-Turbo
+    - gguf: https://huggingface.co/vantagewithai/LongCat-Image-Edit-GGUF/tree/main
+- Download vae
+    - safetensors: https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/ae.safetensors
+- 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
+    - For image editing with GGUF text encoders, also download the matching mmproj file and pass it with `--llm_vision`.
+
+## Run
+
+LongCat uses quoted text for character-level text rendering. Put target text inside single quotes, double quotes, or Chinese quotes.
+
+### LongCat Image
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\LongCat-Image-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ae.sft --llm ..\..\ComfyUI\models\text_encoders\Qwen2.5-VL-7B-Instruct-Q8_0.gguf -p "a lovely cat holding a sign says 'longcat.cpp'" --cfg-scale 5.0 --sampling-method euler --flow-shift 3 -v --offload-to-cpu --diffusion-fa
+```
+
+<img alt="longcat example" src="../assets/longcat/example.png" />

docs/ltx2.md

@@ -13,6 +13,8 @@
     - safetensors: https://huggingface.co/unsloth/LTX-2.3-GGUF/tree/main/vae
 - Download audio vae
     - safetensors: https://huggingface.co/unsloth/LTX-2.3-GGUF/tree/main/vae
+- Download LTX spatial latent upscaler
+    - safetensors: https://huggingface.co/Lightricks/LTX-2.3/resolve/main/ltx-2.3-spatial-upscaler-x2-1.1.safetensors
 
 ## Examples
 
@@ -50,4 +52,26 @@
   src="../assets/ltx2/flf2v.webm"
   controls
   muted
+  style="max-width: 100%; height: auto;"></video>
+
+### LTX-2.3 spatial latent upscale
+
+LTX spatial latent upscale runs a model-backed x2 latent upsampler between the low-resolution video pass and the high-resolution refine pass. `-W` and `-H` are the pre-upscale generation size; the spatial upsampler produces x2 latent dimensions.
+
+Put `ltx-2.3-spatial-upscaler-x2-1.1.safetensors` under the directory passed to `--hires-upscalers-dir`, then use the model name without path or extension in `--hires-upscaler`.
+
+```
+.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\..\ComfyUI\models\diffusion_models\ltx-2.3-22b-dev-UD-Q4_K_M.gguf --vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_video_vae.safetensors --audio-vae ..\..\ComfyUI\models\vae\ltx-2.3-22b-dev_audio_vae.safetensors --llm ..\..\ComfyUI\models\text_encoders\gemma-3-12b-it-qat-UD-Q4_K_XL.gguf --embeddings-connectors ..\..\ComfyUI\models\text_encoders\ltx-2.3-22b-dev_embeddings_connectors.safetensors --hires-upscalers-dir ..\..\ComfyUI\models\latent_upscale_models --hires-upscaler ltx-2.3-spatial-upscaler-x2-1.1 --hires --hires-steps 4 -p "a lovely cat" --cfg-scale 6.0 --sampling-method euler -v  -W 640 -H 360 --diffusion-fa --offload-to-cpu --video-frames 33 -i ..\assets\ernie_image\turbo_example.png -o hires_i2v.webm
+```
+
+By default, the hires refine pass uses the main sampler and scheduler, then trims the second-pass sigma schedule by `--hires-denoising-strength` (`0.7` by default). To reproduce a ComfyUI-style explicit refine schedule, pass custom hires sigmas:
+
+```
+--hires-sigmas "0.85,0.725,0.421875,0.0"
+```
+
+<video
+  src="../assets/ltx2/hires_i2v.webm"
+  controls
+  muted
   style="max-width: 100%; height: auto;"></video>

docs/model_config.md

@@ -0,0 +1,118 @@
+# Model Configuration Conventions
+
+This document describes the conventions for model configuration structs and
+weight-based configuration detection.
+
+## Config Types
+
+Model configuration should live in a model-specific `*Config` struct.
+
+Examples:
+
+- `ZImageConfig`
+- `UNetConfig`
+- `MMDiTConfig`
+- `LLMConfig`
+
+Preserve established acronym casing in type names, such as `UNet`, `MMDiT`,
+`LLM`, `VAE`, and `T5`.
+
+Place the config struct near the top of the model header, before the main model
+blocks and runner types that consume it.
+
+## Config Variables
+
+Variables and members that hold a config should be named `config`.
+
+Examples:
+
+```cpp
+UNetConfig config;
+UnetModelBlock unet;
+
+MMDiTRunner(...)
+    : DiffusionModelRunner(backend, params_backend, prefix),
+      config(MMDiTConfig::detect_from_weights(tensor_storage_map, prefix)),
+      mmdit(config) {
+}
+```
+
+Avoid alternate names such as `params`, `params_cfg`, `model_params`, or
+model-specific aliases unless an existing public API requires them.
+
+## Weight Detection
+
+If a model can derive configuration from loaded weight metadata, expose that
+logic as a static method on the config type:
+
+```cpp
+static XxxConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                     const std::string& prefix);
+```
+
+Additional selector arguments are allowed when required by an existing model
+family, for example `SDVersion version` or an architecture enum:
+
+```cpp
+static UNetConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                      const std::string& prefix,
+                                      SDVersion version = VERSION_SD1);
+```
+
+Use `TensorStorage` metadata, especially `n_dims` and `ne`, to infer shapes.
+Do not load or parse tensor data for config detection.
+
+Detection should respect `prefix`. For nested weights, construct full names from
+`prefix + "." + suffix` or filter entries with `starts_with(name, prefix)`.
+
+Do not add persistent config fields such as `inferred_from_weights` only to
+record whether detection happened. If the function needs to decide whether to
+print a debug line, keep that as local control flow inside `detect_from_weights`.
+
+## Logging
+
+When config values are inferred from weights, print one `LOG_DEBUG` line at the
+end of `detect_from_weights`.
+
+Example:
+
+```cpp
+LOG_DEBUG("llm: num_layers = %" PRId64 ", vocab_size = %" PRId64 ", hidden_size = %" PRId64 ", intermediate_size = %" PRId64,
+          config.num_layers,
+          config.vocab_size,
+          config.hidden_size,
+          config.intermediate_size);
+```
+
+Only print the config detection log when the function actually inferred values
+from weights. Do not duplicate the same config summary in runner constructors or
+model loading code.
+
+Use the correct format specifiers for field types, such as `%" PRId64 "` for
+`int64_t` and `%d` for `int`.
+
+## Runner And Model Responsibilities
+
+Runners should detect the config once and pass it into the model block:
+
+```cpp
+struct XxxRunner : public DiffusionModelRunner {
+    XxxConfig config;
+    XxxModel model;
+
+    XxxRunner(..., const String2TensorStorage& tensor_storage_map, const std::string prefix)
+        : DiffusionModelRunner(backend, params_backend, prefix),
+          config(XxxConfig::detect_from_weights(tensor_storage_map, prefix)),
+          model(config) {
+        model.init(params_ctx, tensor_storage_map, prefix);
+    }
+};
+```
+
+Model blocks should consume `config` directly instead of re-scanning weights in
+their constructors. Keep config-derived behavior centralized in the config
+struct.
+
+If a model has no weight-derived config today, it may still provide
+`detect_from_weights` for API consistency, but it should not print a config
+detection log unless it actually derives values from weights.

docs/performance.md

@@ -21,6 +21,38 @@ and the compute buffer shrink in the debug log:
 
 Using `--offload-to-cpu` allows you to offload weights to the CPU, saving VRAM without reducing generation speed.
 
+## Use params backend to reduce VRAM or RAM usage.
+
+`--params-backend` controls where model parameters are kept. If it is not set, parameters use the same backend as `--backend`, so a GPU runtime backend also keeps parameters in VRAM.
+
+Use CPU params to reduce VRAM usage:
+
+```shell
+--backend cuda0 --params-backend cpu
+```
+
+This keeps model weights in system RAM and moves them to the runtime backend when needed. `--offload-to-cpu` is a compatibility shortcut that prepends `*=cpu` to `--params-backend`, so explicit module assignments can still override it:
+
+```shell
+--offload-to-cpu --params-backend te=disk
+```
+
+Use disk params to reduce both VRAM and RAM usage:
+
+```shell
+--backend cuda0 --params-backend disk
+```
+
+This reloads parameters from the model file on demand and releases them after use. It has the lowest memory residency, but can be slower because weights must be read again. `disk` is never selected implicitly; set it explicitly when RAM usage matters more than reload cost.
+
+Per-module assignments can target only the largest modules:
+
+```shell
+--backend cuda0 --params-backend diffusion=disk,te=cpu,vae=cpu
+```
+
+See [backend selection](./backend.md) for full syntax.
+
 ## Use quantization to reduce memory usage.
 
-[quantization](./quantization_and_gguf.md)
+[quantization](./quantization_and_gguf.md)

docs/pid.md

@@ -0,0 +1,39 @@
+# How to Use
+
+PiD is NVIDIA's Pixel Diffusion Decoder. It replaces the usual VAE decode or decode-then-upscale path with a pixel-space diffusion decoder conditioned on a
+source latent and text prompt.
+
+In stable-diffusion.cpp, PiD currently runs as an image edit pipeline: provide a reference image with `-r`/`--ref-image`, encode that image with a matching VAE, then let the PiD diffusion model decode/upscale directly to RGB.
+
+## Download weights
+
+- Download PiD
+    - safetensors: https://huggingface.co/Comfy-Org/PixelDiT/tree/main/diffusion_models
+- Download Gemma 2 2B
+    - safetensors: https://huggingface.co/Comfy-Org/PixelDiT/tree/main/text_encoders
+- Download the VAE that matches the PiD checkpoint backbone
+    - safetensors: https://huggingface.co/nvidia/PiD/tree/main/checkpoints
+    - Flux / Z-Image PiD: use the Flux VAE and pass `--vae-format flux`
+    - SD3 PiD: use the SD3 VAE and pass `--vae-format sd3`
+    - Flux.2 PiD: use the Flux.2 VAE and pass `--vae-format flux2`
+
+The official PiD model card should be checked before use. At the time of the initial PiD release, the official weights are under the NSCLv1 non-commercial license.
+
+## Examples
+
+```
+.\bin\Release\sd-cli.exe --diffusion-model ..\..\ComfyUI\models\diffusion_models\pid_flux1_512_to_2048_4step_bf16.safetensors --llm "..\..\ComfyUI\models\text_encoders\gemma_2_2b_it_elm_bf16.safetensors" --vae ..\..\ComfyUI\models\vae\ae.sft --vae-format flux --cfg-scale 1.0  -p "a lovely cat" -r ..\assets\ernie_image\turbo_example.png --diffusion-fa -v --steps 4 -H 2048 -W 2048 --rng cpu
+```
+
+Before:
+
+<img width="256" alt="ERNIE-Image Turbo example" src="../assets/ernie_image/turbo_example.png" />
+
+After:
+<img width="1024" alt="PiD example" src="../assets/pid/example.png" />
+
+## Notes
+
+- `-r`/`--ref-image` is required. PiD uses the first reference image as the source latent condition.
+- `--vae-format` should match the VAE latent layout used by the PiD checkpoint. This is important when using standalone VAE files because the PiD diffusion
+  checkpoint alone does not identify the VAE format.

examples/cli/CMakeLists.txt

@@ -7,6 +7,9 @@ add_executable(${TARGET}
     image_metadata.cpp
     main.cpp
 )
+if(APPLE)
+    sd_set_macos_rpaths(${TARGET})
+endif()
 target_include_directories(${TARGET} PRIVATE
     "${CMAKE_CURRENT_SOURCE_DIR}/.."
     "${PROJECT_SOURCE_DIR}/src"

examples/cli/README.md

@@ -41,6 +41,8 @@ Context Options:
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
   --diffusion-model <string>               path to the standalone diffusion model
   --high-noise-diffusion-model <string>    path to the standalone high noise diffusion model
+  --uncond-diffusion-model <string>        path to the standalone unconditional diffusion model, currently used by
+                                           Ideogram4 CFG
   --vae <string>                           path to standalone vae model
   --taesd <string>                         path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
   --tae <string>                           alias of --taesd
@@ -55,7 +57,8 @@ Context Options:
                                            then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
   --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
-                                           graph splitting; -1 auto-detects free VRAM minus 1 GiB
+                                           graph splitting; a negative value auto-detects free VRAM, sparing the
+                                           specified value (e.g. -0.5 will keep at least 0.5 GiB free)
   --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
@@ -103,9 +106,12 @@ Generation Options:
   --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
                                            (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
                                            antialiased), or a model name under --hires-upscalers-dir (default: Latent)
-  --extra-sample-args <string>             extra sampler/scheduler args, key=value list. lcm supports noise_clip_std,
-                                           noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift,
-                                           stretch, terminal; euler_ge supports gamma
+  --extra-sample-args <string>             extra sampler/scheduler/guidance args, key=value list. APG supports apg_eta,
+                                           apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports
+                                           slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end;
+                                           ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma
+  --extra-tiling-args <string>             extra VAE tiling args, key=value list. LTX video VAE supports
+                                           temporal_tile_frames (default: 4), temporal_tile_overlap (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)
@@ -125,8 +131,8 @@ Generation Options:
   --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
                                            128)
   --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)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or image edit 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
@@ -136,8 +142,8 @@ Generation Options:
                                            res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
-                                           (default: same as --cfg-scale)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or image edit 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:
@@ -147,7 +153,7 @@ Generation Options:
   --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd,
                                            res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
-  --pm-style-strength <float>              
+  --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
@@ -175,6 +181,8 @@ Generation Options:
                                            model-specific
   --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
                                            "14.61,7.8,3.5,0.0").
+  --hires-sigmas                           custom sigma values for the highres fix second pass, comma-separated (e.g.,
+                                           "0.85,0.725,0.421875,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)

examples/cli/main.cpp

@@ -169,8 +169,9 @@ struct SDCliParams {
             return 1;
         };
 
-        auto on_help_arg = [&](int argc, const char** argv, int index) {
+        auto on_help_arg = [&](int argc, const char** argv, int index, bool& valid) {
             normal_exit = true;
+            valid       = true;
             return -1;
         };
 
@@ -745,7 +746,7 @@ int main(int argc, const char* argv[]) {
         vae_decode_only = false;
     }
 
-    sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(vae_decode_only, true, cli_params.taesd_preview);
+    sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(vae_decode_only, cli_params.taesd_preview);
 
     SDImageVec results;
     int num_results             = 0;

examples/common/common.cpp

@@ -35,6 +35,22 @@ const char* const modes_str[] = {
     "metadata",
 };
 
+static sd_vae_format_t str_to_vae_format(const std::string& value) {
+    if (value == "auto") {
+        return SD_VAE_FORMAT_AUTO;
+    }
+    if (value == "flux") {
+        return SD_VAE_FORMAT_FLUX;
+    }
+    if (value == "sd3") {
+        return SD_VAE_FORMAT_SD3;
+    }
+    if (value == "flux2") {
+        return SD_VAE_FORMAT_FLUX2;
+    }
+    return SD_VAE_FORMAT_COUNT;
+}
+
 #if defined(_WIN32)
 static std::string utf16_to_utf8(const std::wstring& wstr) {
     if (wstr.empty())
@@ -229,6 +245,7 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
         return false;
     };
 
+    bool valid = false;
     for (int i = 1; i < argc; i++) {
         arg            = argv[i];
         bool found_arg = false;
@@ -271,7 +288,7 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
                 break;
 
             if (match_and_apply(options.manual_options, [&](auto& option) {
-                    int ret = option.cb(argc, argv, i);
+                    int ret = option.cb(argc, argv, i, valid);
                     if (ret < 0) {
                         invalid_arg = true;
                         return;
@@ -283,7 +300,9 @@ bool parse_options(int argc, const char** argv, const std::vector<ArgOptions>& o
         }
 
         if (invalid_arg) {
-            LOG_ERROR("error: invalid parameter for argument: %s", arg.c_str());
+            if (!valid) {
+                LOG_ERROR("error: invalid parameter for argument: %s", arg.c_str());
+            }
             return false;
         }
         if (!found_arg) {
@@ -340,6 +359,10 @@ ArgOptions SDContextParams::get_options() {
          "--high-noise-diffusion-model",
          "path to the standalone high noise diffusion model",
          &high_noise_diffusion_model_path},
+        {"",
+         "--uncond-diffusion-model",
+         "path to the standalone unconditional diffusion model, currently used by Ideogram4 CFG",
+         &uncond_diffusion_model_path},
         {"",
          "--embeddings-connectors",
          "path to LTXAV embeddings connectors",
@@ -348,6 +371,10 @@ ArgOptions SDContextParams::get_options() {
          "--vae",
          "path to standalone vae model",
          &vae_path},
+        {"",
+         "--vae-format",
+         "VAE latent format override: auto, flux, sd3, or flux2 (default: auto)",
+         &vae_format},
         {"",
          "--audio-vae",
          "path to standalone LTX audio vae model",
@@ -394,7 +421,7 @@ ArgOptions SDContextParams::get_options() {
          &backend},
         {"",
          "--params-backend",
-         "parameter backend assignment, e.g. cpu or diffusion=cpu,clip=cpu",
+         "parameter backend assignment, e.g. disk, cpu, or diffusion=disk,clip=cpu",
          &params_backend},
     };
 
@@ -413,11 +440,15 @@ ArgOptions SDContextParams::get_options() {
     options.float_options = {
         {"",
          "--max-vram",
-         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting; -1 auto-detects free VRAM minus 1 GiB",
+         "maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables graph splitting; a negative value auto-detects free VRAM, sparing the specified value (e.g. -0.5 will keep at least 0.5 GiB free)",
          &max_vram},
     };
 
     options.bool_options = {
+        {"",
+         "--stream-layers",
+         "enable residency+prefetch streaming on top of --max-vram (no effect without --max-vram; defaults to false)",
+         true, &stream_layers},
         {"",
          "--force-sdxl-vae-conv-scale",
          "force use of conv scale on sdxl vae",
@@ -639,6 +670,11 @@ bool SDContextParams::validate(SDMode mode) {
         }
     }
 
+    if (str_to_vae_format(vae_format) == SD_VAE_FORMAT_COUNT) {
+        LOG_ERROR("error: vae_format must be 'auto', 'flux', 'sd3', or 'flux2'");
+        return false;
+    }
+
     return true;
 }
 
@@ -677,8 +713,10 @@ std::string SDContextParams::to_string() const {
         << "  llm_vision_path: \"" << llm_vision_path << "\",\n"
         << "  diffusion_model_path: \"" << diffusion_model_path << "\",\n"
         << "  high_noise_diffusion_model_path: \"" << high_noise_diffusion_model_path << "\",\n"
+        << "  uncond_diffusion_model_path: \"" << uncond_diffusion_model_path << "\",\n"
         << "  embeddings_connectors_path: \"" << embeddings_connectors_path << "\",\n"
         << "  vae_path: \"" << vae_path << "\",\n"
+        << "  vae_format: \"" << vae_format << "\",\n"
         << "  audio_vae_path: \"" << audio_vae_path << "\",\n"
         << "  taesd_path: \"" << taesd_path << "\",\n"
         << "  esrgan_path: \"" << esrgan_path << "\",\n"
@@ -694,6 +732,7 @@ std::string SDContextParams::to_string() const {
         << "  sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
         << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
         << "  max_vram: " << max_vram << ",\n"
+        << "  stream_layers: " << (stream_layers ? "true" : "false") << ",\n"
         << "  backend: \"" << backend << "\",\n"
         << "  params_backend: \"" << params_backend << "\",\n"
         << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
@@ -718,7 +757,7 @@ std::string SDContextParams::to_string() const {
     return oss.str();
 }
 
-sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool free_params_immediately, bool taesd_preview) {
+sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool taesd_preview) {
     embedding_vec.clear();
     embedding_vec.reserve(embedding_map.size());
     for (const auto& kv : embedding_map) {
@@ -738,6 +777,7 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
         llm_vision_path.c_str(),
         diffusion_model_path.c_str(),
         high_noise_diffusion_model_path.c_str(),
+        uncond_diffusion_model_path.c_str(),
         embeddings_connectors_path.c_str(),
         vae_path.c_str(),
         audio_vae_path.c_str(),
@@ -748,7 +788,6 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
         photo_maker_path.c_str(),
         tensor_type_rules.c_str(),
         vae_decode_only,
-        free_params_immediately,
         n_threads,
         wtype,
         rng_type,
@@ -772,7 +811,9 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool vae_decode_only, bool f
         chroma_use_t5_mask,
         chroma_t5_mask_pad,
         qwen_image_zero_cond_t,
+        str_to_vae_format(vae_format),
         max_vram,
+        stream_layers,
         backend.c_str(),
         params_backend.c_str(),
     };
@@ -833,8 +874,12 @@ ArgOptions SDGenerationParams::get_options() {
          &hires_upscaler},
         {"",
          "--extra-sample-args",
-         "extra sampler/scheduler args, key=value list. lcm supports noise_clip_std, noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma",
+         "extra sampler/scheduler/guidance args, key=value list. APG supports apg_eta, apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma",
          &extra_sample_args},
+        {"",
+         "--extra-tiling-args",
+         "extra VAE tiling args, key=value list. LTX video VAE supports temporal_tile_frames (default: 4), temporal_tile_overlap (default: 1)",
+         &extra_tiling_args},
     };
 
     options.int_options = {
@@ -909,7 +954,7 @@ ArgOptions SDGenerationParams::get_options() {
          &sample_params.guidance.txt_cfg},
         {"",
          "--img-cfg-scale",
-         "image guidance scale for inpaint or instruct-pix2pix models: (default: same as --cfg-scale)",
+         "image guidance scale for inpaint or image edit models: (default: same as --cfg-scale)",
          &sample_params.guidance.img_cfg},
         {"",
          "--guidance",
@@ -941,7 +986,7 @@ ArgOptions SDGenerationParams::get_options() {
          &high_noise_sample_params.guidance.txt_cfg},
         {"",
          "--high-noise-img-cfg-scale",
-         "(high noise) image guidance scale for inpaint or instruct-pix2pix models (default: same as --cfg-scale)",
+         "(high noise) image guidance scale for inpaint or image edit models (default: same as --cfg-scale)",
          &high_noise_sample_params.guidance.img_cfg},
         {"",
          "--high-noise-guidance",
@@ -1134,11 +1179,11 @@ ArgOptions SDGenerationParams::get_options() {
         return 1;
     };
 
-    auto on_sigmas_arg = [&](int argc, const char** argv, int index) {
-        if (++index >= argc) {
+    auto parse_sigmas_arg = [&](const char* value, std::vector<float>* target, const char* option_name) {
+        if (target == nullptr || value == nullptr) {
             return -1;
         }
-        std::string sigmas_str = argv[index];
+        std::string sigmas_str = value;
         if (!sigmas_str.empty() && sigmas_str.front() == '[') {
             sigmas_str.erase(0, 1);
         }
@@ -1146,6 +1191,7 @@ ArgOptions SDGenerationParams::get_options() {
             sigmas_str.pop_back();
         }
 
+        size_t before = target->size();
         std::stringstream ss(sigmas_str);
         std::string item;
         while (std::getline(ss, item, ',')) {
@@ -1153,24 +1199,38 @@ ArgOptions SDGenerationParams::get_options() {
             item.erase(item.find_last_not_of(" \t\n\r\f\v") + 1);
             if (!item.empty()) {
                 try {
-                    custom_sigmas.push_back(std::stof(item));
+                    target->push_back(std::stof(item));
                 } catch (const std::invalid_argument&) {
-                    LOG_ERROR("error: invalid float value '%s' in --sigmas", item.c_str());
+                    LOG_ERROR("error: invalid float value '%s' in %s", item.c_str(), option_name);
                     return -1;
                 } catch (const std::out_of_range&) {
-                    LOG_ERROR("error: float value '%s' out of range in --sigmas", item.c_str());
+                    LOG_ERROR("error: float value '%s' out of range in %s", item.c_str(), option_name);
                     return -1;
                 }
             }
         }
 
-        if (custom_sigmas.empty() && !sigmas_str.empty()) {
-            LOG_ERROR("error: could not parse any sigma values from '%s'", argv[index]);
+        if (target->size() == before && !sigmas_str.empty()) {
+            LOG_ERROR("error: could not parse any sigma values from '%s'", value);
             return -1;
         }
         return 1;
     };
 
+    auto on_sigmas_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        return parse_sigmas_arg(argv[index], &custom_sigmas, "--sigmas");
+    };
+
+    auto on_hires_sigmas_arg = [&](int argc, const char** argv, int index) {
+        if (++index >= argc) {
+            return -1;
+        }
+        return parse_sigmas_arg(argv[index], &hires_custom_sigmas, "--hires-sigmas");
+    };
+
     auto on_ref_image_arg = [&](int argc, const char** argv, int index) {
         if (++index >= argc) {
             return -1;
@@ -1293,6 +1353,10 @@ ArgOptions SDGenerationParams::get_options() {
          "--sigmas",
          "custom sigma values for the sampler, comma-separated (e.g., \"14.61,7.8,3.5,0.0\").",
          on_sigmas_arg},
+        {"",
+         "--hires-sigmas",
+         "custom sigma values for the highres fix second pass, comma-separated (e.g., \"0.85,0.725,0.421875,0.0\").",
+         on_hires_sigmas_arg},
         {"",
          "--skip-layers",
          "layers to skip for SLG steps (default: [7,8,9])",
@@ -1525,11 +1589,31 @@ static bool resolve_model_file_from_dir(const std::string& model_name,
         LOG_ERROR("%s directory is empty", label);
         return false;
     }
+    auto ends_with_valid_ext = [&]() {
+        for (const auto& ext : valid_ext) {
+            if (model_name.size() < ext.size()) {
+                continue;
+            }
+            auto suffix = model_name.substr(model_name.size() - ext.size());
+            std::transform(suffix.begin(), suffix.end(), suffix.begin(), [](unsigned char c) {
+                return static_cast<char>(std::tolower(c));
+            });
+            std::string lower_ext = ext;
+            std::transform(lower_ext.begin(), lower_ext.end(), lower_ext.begin(), [](unsigned char c) {
+                return static_cast<char>(std::tolower(c));
+            });
+            if (suffix == lower_ext) {
+                return true;
+            }
+        }
+        return false;
+    };
+
     if (model_name.empty() ||
         model_name.find('/') != std::string::npos ||
         model_name.find('\\') != std::string::npos ||
         fs::path(model_name).has_root_path() ||
-        fs::path(model_name).has_extension()) {
+        ends_with_valid_ext()) {
         LOG_ERROR("%s must be a model name without path or extension: %s", label, model_name.c_str());
         return false;
     }
@@ -1633,6 +1717,9 @@ bool SDGenerationParams::from_json_str(
         if (hires_json.contains("denoising_strength") && hires_json["denoising_strength"].is_number()) {
             hires_denoising_strength = hires_json["denoising_strength"];
         }
+        if (hires_json.contains("custom_sigmas") && hires_json["custom_sigmas"].is_array()) {
+            hires_custom_sigmas = hires_json["custom_sigmas"].get<std::vector<float>>();
+        }
         if (hires_json.contains("upscale_tile_size") && hires_json["upscale_tile_size"].is_number_integer()) {
             hires_upscale_tile_size = hires_json["upscale_tile_size"];
         }
@@ -1738,6 +1825,9 @@ bool SDGenerationParams::from_json_str(
         if (tiling_json.contains("rel_size_y") && tiling_json["rel_size_y"].is_number()) {
             vae_tiling_params.rel_size_y = tiling_json["rel_size_y"];
         }
+        if (tiling_json.contains("extra_tiling_args") && tiling_json["extra_tiling_args"].is_string()) {
+            extra_tiling_args = tiling_json["extra_tiling_args"].get<std::string>();
+        }
     }
 
     if (!parse_lora_json_field(j, lora_path_resolver, lora_map, high_noise_lora_map)) {
@@ -1960,6 +2050,8 @@ bool SDGenerationParams::initialize_cache_params() {
 }
 
 bool SDGenerationParams::resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict) {
+    vae_tiling_params.extra_tiling_args = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
+
     if (high_noise_sample_params.sample_steps <= 0) {
         high_noise_sample_params.sample_steps = -1;
     }
@@ -2080,6 +2172,10 @@ bool SDGenerationParams::validate(SDMode mode) {
             LOG_ERROR("error: hires denoising strength must be in (0.0, 1.0]");
             return false;
         }
+        if (!hires_custom_sigmas.empty() && hires_custom_sigmas.size() < 2) {
+            LOG_ERROR("error: hires custom sigmas must contain at least two values");
+            return false;
+        }
         if (hires_upscale_tile_size < 1) {
             LOG_ERROR("error: hires upscale tile size must be positive");
             return false;
@@ -2142,6 +2238,7 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
     sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
     high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
+    vae_tiling_params.extra_tiling_args               = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     sd_pm_params_t pm_params = {
@@ -2174,15 +2271,17 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.vae_tiling_params     = vae_tiling_params;
     params.cache                 = cache_params;
 
-    params.hires.enabled            = hires_enabled;
-    params.hires.upscaler           = resolved_hires_upscaler;
-    params.hires.model_path         = hires_upscaler_model_path.empty() ? nullptr : hires_upscaler_model_path.c_str();
-    params.hires.scale              = hires_scale;
-    params.hires.target_width       = hires_width;
-    params.hires.target_height      = hires_height;
-    params.hires.steps              = hires_steps;
-    params.hires.denoising_strength = hires_denoising_strength;
-    params.hires.upscale_tile_size  = hires_upscale_tile_size;
+    params.hires.enabled             = hires_enabled;
+    params.hires.upscaler            = resolved_hires_upscaler;
+    params.hires.model_path          = hires_upscaler_model_path.empty() ? nullptr : hires_upscaler_model_path.c_str();
+    params.hires.scale               = hires_scale;
+    params.hires.target_width        = hires_width;
+    params.hires.target_height       = hires_height;
+    params.hires.steps               = hires_steps;
+    params.hires.denoising_strength  = hires_denoising_strength;
+    params.hires.upscale_tile_size   = hires_upscale_tile_size;
+    params.hires.custom_sigmas       = hires_custom_sigmas.empty() ? nullptr : hires_custom_sigmas.data();
+    params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
     return params;
 }
 
@@ -2213,29 +2312,41 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     sample_params.custom_sigmas_count                 = static_cast<int>(custom_sigmas.size());
     sample_params.extra_sample_args                   = extra_sample_args.empty() ? nullptr : extra_sample_args.c_str();
     high_noise_sample_params.extra_sample_args        = high_noise_extra_sample_args.empty() ? nullptr : high_noise_extra_sample_args.c_str();
+    vae_tiling_params.extra_tiling_args               = extra_tiling_args.empty() ? nullptr : extra_tiling_args.c_str();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
-    params.loras                    = lora_vec.empty() ? nullptr : lora_vec.data();
-    params.lora_count               = static_cast<uint32_t>(lora_vec.size());
-    params.prompt                   = prompt.c_str();
-    params.negative_prompt          = negative_prompt.c_str();
-    params.clip_skip                = clip_skip;
-    params.init_image               = init_image.get();
-    params.end_image                = end_image.get();
-    params.control_frames           = control_frame_views.empty() ? nullptr : control_frame_views.data();
-    params.control_frames_size      = static_cast<int>(control_frame_views.size());
-    params.width                    = get_resolved_width();
-    params.height                   = get_resolved_height();
-    params.sample_params            = sample_params;
-    params.high_noise_sample_params = high_noise_sample_params;
-    params.moe_boundary             = moe_boundary;
-    params.strength                 = strength;
-    params.seed                     = seed;
-    params.video_frames             = video_frames;
-    params.fps                      = fps;
-    params.vace_strength            = vace_strength;
-    params.vae_tiling_params        = vae_tiling_params;
-    params.cache                    = cache_params;
+    params.loras                     = lora_vec.empty() ? nullptr : lora_vec.data();
+    params.lora_count                = static_cast<uint32_t>(lora_vec.size());
+    params.prompt                    = prompt.c_str();
+    params.negative_prompt           = negative_prompt.c_str();
+    params.clip_skip                 = clip_skip;
+    params.init_image                = init_image.get();
+    params.end_image                 = end_image.get();
+    params.control_frames            = control_frame_views.empty() ? nullptr : control_frame_views.data();
+    params.control_frames_size       = static_cast<int>(control_frame_views.size());
+    params.width                     = get_resolved_width();
+    params.height                    = get_resolved_height();
+    params.sample_params             = sample_params;
+    params.high_noise_sample_params  = high_noise_sample_params;
+    params.moe_boundary              = moe_boundary;
+    params.strength                  = strength;
+    params.seed                      = seed;
+    params.video_frames              = video_frames;
+    params.fps                       = fps;
+    params.vace_strength             = vace_strength;
+    params.vae_tiling_params         = vae_tiling_params;
+    params.cache                     = cache_params;
+    params.hires.enabled             = hires_enabled;
+    params.hires.upscaler            = resolved_hires_upscaler;
+    params.hires.model_path          = hires_upscaler_model_path.empty() ? nullptr : hires_upscaler_model_path.c_str();
+    params.hires.scale               = hires_scale;
+    params.hires.target_width        = hires_width;
+    params.hires.target_height       = hires_height;
+    params.hires.steps               = hires_steps;
+    params.hires.denoising_strength  = hires_denoising_strength;
+    params.hires.upscale_tile_size   = hires_upscale_tile_size;
+    params.hires.custom_sigmas       = hires_custom_sigmas.empty() ? nullptr : hires_custom_sigmas.data();
+    params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
     return params;
 }
 
@@ -2318,6 +2429,7 @@ std::string SDGenerationParams::to_string() const {
         << ", target_height: " << hires_height
         << ", steps: " << hires_steps
         << ", denoising_strength: " << hires_denoising_strength
+        << ", custom_sigmas: " << vec_to_string(hires_custom_sigmas)
         << ", upscale_tile_size: " << hires_upscale_tile_size << " },\n"
         << "  vae_tiling_params: { "
         << vae_tiling_params.enabled << ", "
@@ -2326,7 +2438,8 @@ std::string SDGenerationParams::to_string() const {
         << vae_tiling_params.tile_size_y << ", "
         << vae_tiling_params.target_overlap << ", "
         << vae_tiling_params.rel_size_x << ", "
-        << vae_tiling_params.rel_size_y << " },\n"
+        << vae_tiling_params.rel_size_y << ", "
+        << "\"" << extra_tiling_args << "\" },\n"
         << "}";
     return oss.str();
 }
@@ -2414,6 +2527,7 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
     set_json_basename_if_not_empty(models, "llm_vision", ctx_params.llm_vision_path);
     set_json_basename_if_not_empty(models, "diffusion_model", ctx_params.diffusion_model_path);
     set_json_basename_if_not_empty(models, "high_noise_diffusion_model", ctx_params.high_noise_diffusion_model_path);
+    set_json_basename_if_not_empty(models, "uncond_diffusion_model", ctx_params.uncond_diffusion_model_path);
     set_json_basename_if_not_empty(models, "vae", ctx_params.vae_path);
     set_json_basename_if_not_empty(models, "taesd", ctx_params.taesd_path);
     set_json_basename_if_not_empty(models, "control_net", ctx_params.control_net_path);
@@ -2469,6 +2583,7 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
             {"target_height", gen_params.hires_height},
             {"steps", gen_params.hires_steps},
             {"denoising_strength", gen_params.hires_denoising_strength},
+            {"custom_sigmas", gen_params.hires_custom_sigmas},
             {"upscale_tile_size", gen_params.hires_upscale_tile_size},
         };
     }
@@ -2504,14 +2619,18 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
         };
     }
 
-    if (gen_params.vae_tiling_params.enabled) {
+    if (gen_params.vae_tiling_params.enabled ||
+        gen_params.vae_tiling_params.temporal_tiling ||
+        !gen_params.extra_tiling_args.empty()) {
         root["vae_tiling"] = {
             {"enabled", gen_params.vae_tiling_params.enabled},
+            {"temporal_tiling", gen_params.vae_tiling_params.temporal_tiling},
             {"tile_size_x", gen_params.vae_tiling_params.tile_size_x},
             {"tile_size_y", gen_params.vae_tiling_params.tile_size_y},
             {"target_overlap", gen_params.vae_tiling_params.target_overlap},
             {"rel_size_x", gen_params.vae_tiling_params.rel_size_x},
             {"rel_size_y", gen_params.vae_tiling_params.rel_size_y},
+            {"extra_tiling_args", gen_params.extra_tiling_args},
         };
     }
 
@@ -2576,6 +2695,9 @@ std::string get_image_params(const SDContextParams& ctx_params,
     if (!ctx_params.diffusion_model_path.empty()) {
         parameter_string += "Unet: " + sd_basename(ctx_params.diffusion_model_path) + ", ";
     }
+    if (!ctx_params.uncond_diffusion_model_path.empty()) {
+        parameter_string += "Uncond Unet: " + sd_basename(ctx_params.uncond_diffusion_model_path) + ", ";
+    }
     if (!ctx_params.vae_path.empty()) {
         parameter_string += "VAE: " + sd_basename(ctx_params.vae_path) + ", ";
     }
@@ -2588,6 +2710,9 @@ std::string get_image_params(const SDContextParams& ctx_params,
         parameter_string += "Hires resize: " + std::to_string(gen_params.hires_width) + "x" + std::to_string(gen_params.hires_height) + ", ";
         parameter_string += "Hires steps: " + std::to_string(gen_params.hires_steps) + ", ";
         parameter_string += "Denoising strength: " + std::to_string(gen_params.hires_denoising_strength) + ", ";
+        if (!gen_params.hires_custom_sigmas.empty()) {
+            parameter_string += "Hires custom sigmas: " + vec_to_string(gen_params.hires_custom_sigmas) + ", ";
+        }
     }
     parameter_string += "Version: stable-diffusion.cpp";
     parameter_string += ", SDCPP: " + build_sdcpp_image_metadata_json(ctx_params, gen_params, seed, mode);

examples/common/common.h

@@ -56,11 +56,42 @@ struct BoolOption {
     bool* target;
 };
 
+struct ManualFunction {
+    std::function<int(int, const char**, int, bool&)> _func;
+
+    ManualFunction() = default;
+
+    ManualFunction(std::function<int(int argc, const char** argv, int index, bool& valid)> func)
+        : _func(std::move(func)) {
+    }
+
+    template <typename F>
+    ManualFunction(F func)
+        : _func(make_function(func)) {
+    }
+
+    int operator()(int argc, const char** argv, int index, bool& valid) const {
+        return _func(argc, argv, index, valid);
+    }
+
+private:
+    template <typename F>
+    static std::function<int(int, const char**, int, bool&)> make_function(F func) {
+        if constexpr (std::is_invocable_v<F, int, const char**, int, bool&>) {
+            return func;
+        } else {
+            return [func](int argc, const char** argv, int index, bool&) {
+                return func(argc, argv, index);
+            };
+        }
+    }
+};
+
 struct ManualOption {
     std::string short_name;
     std::string long_name;
     std::string desc;
-    std::function<int(int argc, const char** argv, int index)> cb;
+    ManualFunction cb;
 };
 
 struct ArgOptions {
@@ -92,8 +123,10 @@ struct SDContextParams {
     std::string llm_vision_path;
     std::string diffusion_model_path;
     std::string high_noise_diffusion_model_path;
+    std::string uncond_diffusion_model_path;
     std::string embeddings_connectors_path;
     std::string vae_path;
+    std::string vae_format = "auto";
     std::string audio_vae_path;
     std::string taesd_path;
     std::string esrgan_path;
@@ -112,6 +145,7 @@ struct SDContextParams {
     rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
     bool offload_params_to_cpu  = false;
     float max_vram              = 0.f;
+    bool stream_layers          = false;
     std::string backend;
     std::string params_backend;
     bool enable_mmap           = false;
@@ -145,7 +179,7 @@ struct SDContextParams {
     bool validate(SDMode mode);
     bool resolve_and_validate(SDMode mode);
     std::string to_string() const;
-    sd_ctx_params_t to_sd_ctx_params_t(bool vae_decode_only, bool free_params_immediately, bool taesd_preview);
+    sd_ctx_params_t to_sd_ctx_params_t(bool vae_decode_only, bool taesd_preview);
 };
 
 struct SDGenerationParams {
@@ -189,7 +223,8 @@ struct SDGenerationParams {
     int video_frames                     = 1;
     int fps                              = 16;
     float vace_strength                  = 1.f;
-    sd_tiling_params_t vae_tiling_params = {false, false, 0, 0, 0.5f, 0.0f, 0.0f};
+    sd_tiling_params_t vae_tiling_params = {false, false, 0, 0, 0.5f, 0.0f, 0.0f, nullptr};
+    std::string extra_tiling_args;
 
     std::string pm_id_images_dir;
     std::string pm_id_embed_path;
@@ -207,6 +242,7 @@ struct SDGenerationParams {
     int hires_steps                = 0;
     float hires_denoising_strength = 0.7f;
     int hires_upscale_tile_size    = 128;
+    std::vector<float> hires_custom_sigmas;
 
     std::map<std::string, float> lora_map;
     std::map<std::string, float> high_noise_lora_map;

examples/server/CMakeLists.txt

@@ -74,6 +74,9 @@ add_executable(${TARGET}
     routes_sdapi.cpp
     routes_sdcpp.cpp
 )
+if(APPLE)
+    sd_set_macos_rpaths(${TARGET})
+endif()
 
 if(HAVE_FRONTEND_BUILD)
     add_dependencies(${TARGET} ${TARGET}_frontend)

examples/server/README.md

@@ -143,6 +143,8 @@ Context Options:
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
   --diffusion-model <string>               path to the standalone diffusion model
   --high-noise-diffusion-model <string>    path to the standalone high noise diffusion model
+  --uncond-diffusion-model <string>        path to the standalone unconditional diffusion model, currently used by
+                                           Ideogram4 CFG
   --vae <string>                           path to standalone vae model
   --taesd <string>                         path to taesd. Using Tiny AutoEncoder for fast decoding (low quality)
   --tae <string>                           alias of --taesd
@@ -157,7 +159,8 @@ Context Options:
                                            then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
   --max-vram <float>                       maximum VRAM budget in GiB for graph-cut segmented execution. 0 disables
-                                           graph splitting; -1 auto-detects free VRAM minus 1 GiB
+                                           graph splitting; a negative value auto-detects free VRAM, sparing the
+                                           specified value (e.g. -0.5 will keep at least 0.5 GiB free)
   --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
@@ -205,9 +208,12 @@ Default Generation Options:
   --hires-upscaler <string>                highres fix upscaler, Lanczos, Nearest, Latent, Latent (nearest), Latent
                                            (nearest-exact), Latent (antialiased), Latent (bicubic), Latent (bicubic
                                            antialiased), or a model name under --hires-upscalers-dir (default: Latent)
-  --extra-sample-args <string>             extra sampler/scheduler args, key=value list. lcm supports noise_clip_std,
-                                           noise_scale_start, noise_scale_end; ltx2 supports max_shift, base_shift,
-                                           stretch, terminal; euler_ge supports gamma
+  --extra-sample-args <string>             extra sampler/scheduler/guidance args, key=value list. APG supports apg_eta,
+                                           apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports
+                                           slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end;
+                                           ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma
+  --extra-tiling-args <string>             extra VAE tiling args, key=value list. LTX video VAE supports
+                                           temporal_tile_frames (default: 4), temporal_tile_overlap (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)
@@ -227,8 +233,8 @@ Default Generation Options:
   --hires-upscale-tile-size <int>          highres fix upscaler tile size, reserved for model-backed upscalers (default:
                                            128)
   --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)
+  --img-cfg-scale <float>                  image guidance scale for inpaint or image edit 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
@@ -238,8 +244,8 @@ Default Generation Options:
                                            res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --flow-shift <float>                     shift value for Flow models like SD3.x or WAN (default: auto)
   --high-noise-cfg-scale <float>           (high noise) unconditional guidance scale: (default: 7.0)
-  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
-                                           (default: same as --cfg-scale)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or image edit 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:
@@ -249,7 +255,7 @@ Default Generation Options:
   --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd,
                                            res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
-  --pm-style-strength <float>              
+  --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
@@ -263,6 +269,7 @@ Default Generation Options:
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
   --vae-tiling                             process vae in tiles to reduce memory usage
+  --temporal-tiling                        enable temporal tiling for LTX video VAE decode
   --hires                                  enable highres fix
   -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,
@@ -276,6 +283,8 @@ Default Generation Options:
                                            model-specific
   --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
                                            "14.61,7.8,3.5,0.0").
+  --hires-sigmas                           custom sigma values for the highres fix second pass, comma-separated (e.g.,
+                                           "0.85,0.725,0.421875,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)

examples/server/api.md

@@ -504,11 +504,13 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `sample_params.guidance.slg.scale` | `number` |
 | `vae_tiling_params` | `object` |
 | `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
 | `vae_tiling_params.tile_size_x` | `integer` |
 | `vae_tiling_params.tile_size_y` | `integer` |
 | `vae_tiling_params.target_overlap` | `number` |
 | `vae_tiling_params.rel_size_x` | `number` |
 | `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |
@@ -516,6 +518,8 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `output_format` | `string` |
 | `output_compression` | `integer` |
 
+`vae_tiling_params.extra_tiling_args` accepts a key=value list. For LTX video VAE temporal tiling, `temporal_tile_frames` defaults to `4` and `temporal_tile_overlap` defaults to `1`.
+
 `img_gen`-specific default fields:
 
 | Field | Type |
@@ -532,6 +536,7 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `hires.target_height` | `integer` |
 | `hires.steps` | `integer` |
 | `hires.denoising_strength` | `number` |
+| `hires.custom_sigmas` | `array<number>` |
 | `hires.upscale_tile_size` | `integer` |
 
 `vid_gen`-specific default fields:
@@ -685,16 +690,19 @@ Example:
     "target_height": 0,
     "steps": 0,
     "denoising_strength": 0.7,
+    "custom_sigmas": [],
     "upscale_tile_size": 128
   },
 
   "vae_tiling_params": {
     "enabled": false,
+    "temporal_tiling": false,
     "tile_size_x": 0,
     "tile_size_y": 0,
     "target_overlap": 0.5,
     "rel_size_x": 0.0,
-    "rel_size_y": 0.0
+    "rel_size_y": 0.0,
+    "extra_tiling_args": ""
   },
 
   "cache_mode": "disabled",
@@ -799,14 +807,23 @@ Other native fields:
 | `hires.target_height` | `integer` |
 | `hires.steps` | `integer` |
 | `hires.denoising_strength` | `number` |
+| `hires.custom_sigmas` | `array<number>` |
 | `hires.upscale_tile_size` | `integer` |
 | `vae_tiling_params` | `object` |
+| `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
+| `vae_tiling_params.target_overlap` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |
 | `scm_policy_dynamic` | `boolean` |
 
-For `hires.upscaler`, use `Lanczos`, `Nearest`, `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, `Latent (bicubic antialiased)`, or an `upscalers[].name` value from `GET /sdcpp/v1/capabilities`. Model-backed upscalers are resolved as `--hires-upscalers-dir / (name + ext)` and must live directly in that directory.
+For `hires.upscaler`, use `Lanczos`, `Nearest`, `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, `Latent (bicubic antialiased)`, or an `upscalers[].name` value from `GET /sdcpp/v1/capabilities`. Model-backed upscalers are resolved as `--hires-upscalers-dir / (name + ext)` and must live directly in that directory. `hires.custom_sigmas`, when present, overrides the generated second-pass hires sigma schedule; otherwise the hires schedule is trimmed by `hires.denoising_strength`.
 
 HTTP-only output fields:
 
@@ -1009,11 +1026,13 @@ Example:
 
   "vae_tiling_params": {
     "enabled": false,
+    "temporal_tiling": false,
     "tile_size_x": 0,
     "tile_size_y": 0,
     "target_overlap": 0.5,
     "rel_size_x": 0.0,
-    "rel_size_y": 0.0
+    "rel_size_y": 0.0,
+    "extra_tiling_args": ""
   },
 
   "cache_mode": "disabled",
@@ -1131,6 +1150,14 @@ Other native fields:
 | Field | Type |
 | --- | --- |
 | `vae_tiling_params` | `object` |
+| `vae_tiling_params.enabled` | `boolean` |
+| `vae_tiling_params.temporal_tiling` | `boolean` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
+| `vae_tiling_params.target_overlap` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
+| `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
 | `scm_mask` | `string` |

examples/server/main.cpp

@@ -85,7 +85,7 @@ int main(int argc, const char** argv) {
     LOG_DEBUG("%s", ctx_params.to_string().c_str());
     LOG_DEBUG("%s", default_gen_params.to_string().c_str());
 
-    sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(false, false, false);
+    sd_ctx_params_t sd_ctx_params = ctx_params.to_sd_ctx_params_t(false, false);
     SDCtxPtr sd_ctx(new_sd_ctx(&sd_ctx_params));
 
     if (sd_ctx == nullptr) {

examples/server/routes_sdcpp.cpp

@@ -56,11 +56,13 @@ static const char* capability_sample_method_name(enum sample_method_t sample_met
 static json make_vae_tiling_json(const sd_tiling_params_t& params) {
     return {
         {"enabled", params.enabled},
+        {"temporal_tiling", params.temporal_tiling},
         {"tile_size_x", params.tile_size_x},
         {"tile_size_y", params.tile_size_y},
         {"target_overlap", params.target_overlap},
         {"rel_size_x", params.rel_size_x},
         {"rel_size_y", params.rel_size_y},
+        {"extra_tiling_args", params.extra_tiling_args ? params.extra_tiling_args : ""},
     };
 }
 
@@ -100,6 +102,20 @@ static json make_sample_params_json(const sd_sample_params_t& sample_params, con
     };
 }
 
+static json make_hires_json(const SDGenerationParams& defaults) {
+    return {
+        {"enabled", defaults.hires_enabled},
+        {"upscaler", defaults.hires_upscaler},
+        {"scale", defaults.hires_scale},
+        {"target_width", defaults.hires_width},
+        {"target_height", defaults.hires_height},
+        {"steps", defaults.hires_steps},
+        {"denoising_strength", defaults.hires_denoising_strength},
+        {"custom_sigmas", defaults.hires_custom_sigmas},
+        {"upscale_tile_size", defaults.hires_upscale_tile_size},
+    };
+}
+
 static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const std::string& output_format) {
     return {
         {"prompt", defaults.prompt},
@@ -114,17 +130,7 @@ static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const
         {"increase_ref_index", defaults.increase_ref_index},
         {"control_strength", defaults.control_strength},
         {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
-        {"hires",
-         {
-             {"enabled", defaults.hires_enabled},
-             {"upscaler", defaults.hires_upscaler},
-             {"scale", defaults.hires_scale},
-             {"target_width", defaults.hires_width},
-             {"target_height", defaults.hires_height},
-             {"steps", defaults.hires_steps},
-             {"denoising_strength", defaults.hires_denoising_strength},
-             {"upscale_tile_size", defaults.hires_upscale_tile_size},
-         }},
+        {"hires", make_hires_json(defaults)},
         {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
         {"cache_mode", defaults.cache_mode},
         {"cache_option", defaults.cache_option},
@@ -150,6 +156,7 @@ static json make_vid_gen_defaults_json(const SDGenerationParams& defaults, const
         {"vace_strength", defaults.vace_strength},
         {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
         {"high_noise_sample_params", make_sample_params_json(defaults.high_noise_sample_params, defaults.high_noise_skip_layers)},
+        {"hires", make_hires_json(defaults)},
         {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
         {"cache_mode", defaults.cache_mode},
         {"cache_option", defaults.cache_option},

examples/server/runtime.cpp

@@ -203,8 +203,9 @@ ArgOptions SDSvrParams::get_options() {
         {"", "--color", "colors the logging tags according to level", true, &color},
     };
 
-    auto on_help_arg = [&](int, const char**, int) {
+    auto on_help_arg = [&](int, const char**, int, bool& valid) {
         normal_exit = true;
+        valid       = true;
         return -1;
     };
 
@@ -254,7 +255,7 @@ void refresh_lora_cache(ServerRuntime& rt) {
 
     fs::path lora_dir = rt.ctx_params->lora_model_dir;
     if (fs::exists(lora_dir) && fs::is_directory(lora_dir)) {
-        for (auto& entry : fs::recursive_directory_iterator(lora_dir)) {
+        for (auto& entry : fs::recursive_directory_iterator(lora_dir, fs::directory_options::skip_permission_denied)) {
             if (!entry.is_regular_file()) {
                 continue;
             }

format-code.ps1

@@ -0,0 +1,54 @@
+$patterns = @(
+    "src/*.cpp"
+    "src/*.h"
+    "src/*.hpp"
+    "src/conditioning/*.cpp"
+    "src/conditioning/*.h"
+    "src/conditioning/*.hpp"
+    "src/core/*.cpp"
+    "src/core/*.h"
+    "src/core/*.hpp"
+    "src/extensions/*.cpp"
+    "src/extensions/*.h"
+    "src/extensions/*.hpp"
+    "src/runtime/*.cpp"
+    "src/runtime/*.h"
+    "src/runtime/*.hpp"
+    "src/model/*/*.cpp"
+    "src/model/*/*.h"
+    "src/model/*/*.hpp"
+    "src/tokenizers/*.h"
+    "src/tokenizers/*.cpp"
+    "src/tokenizers/vocab/*.h"
+    "src/tokenizers/vocab/*.cpp"
+    "src/model_io/*.h"
+    "src/model_io/*.cpp"
+    "examples/cli/*.cpp"
+    "examples/cli/*.h"
+    "examples/server/*.cpp"
+    "examples/common/*.hpp"
+    "examples/common/*.h"
+    "examples/common/*.cpp"
+)
+
+$root = (Get-Location).Path
+
+foreach ($pattern in $patterns) {
+    $files = Get-ChildItem -Path $pattern -File -ErrorAction SilentlyContinue | Sort-Object FullName
+
+    foreach ($file in $files) {
+        $relativePath = $file.FullName.Substring($root.Length).TrimStart('\', '/') -replace '\\', '/'
+
+        if ($relativePath -like "vocab*") {
+            continue
+        }
+
+        Write-Host "formatting '$relativePath'"
+
+        # if ($relativePath -ne "stable-diffusion.h") {
+        #     clang-tidy -fix -p build_linux/ "$relativePath"
+        # }
+
+        & clang-format -style=file -i $relativePath
+    }
+}

format-code.sh

@@ -1,10 +1,17 @@
-for f in src/*.cpp src/*.h src/*.hpp src/tokenizers/*.h src/tokenizers/*.cpp src/tokenizers/vocab/*.h src/tokenizers/vocab/*.cpp \
+for f in src/*.cpp src/*.h src/*.hpp \
+         src/conditioning/*.cpp src/conditioning/*.h src/conditioning/*.hpp \
+         src/core/*.cpp src/core/*.h src/core/*.hpp \
+         src/extensions/*.cpp src/extensions/*.h src/extensions/*.hpp \
+         src/runtime/*.cpp src/runtime/*.h src/runtime/*.hpp \
+         src/model/*/*.cpp src/model/*/*.h src/model/*/*.hpp \
+         src/tokenizers/*.h src/tokenizers/*.cpp src/tokenizers/vocab/*.h src/tokenizers/vocab/*.cpp \
          src/model_io/*.h src/model_io/*.cpp examples/cli/*.cpp examples/cli/*.h examples/server/*.cpp \
          examples/common/*.hpp examples/common/*.h examples/common/*.cpp; do
+  [[ -e "$f" ]] || continue
   [[ "$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
+done

include/stable-diffusion.h

@@ -126,7 +126,8 @@ enum sd_type_t {
     // SD_TYPE_IQ4_NL_8_8 = 38,
     SD_TYPE_MXFP4 = 39,  // MXFP4 (1 block)
     SD_TYPE_NVFP4 = 40,  // NVFP4 (4 blocks, E4M3 scale)
-    SD_TYPE_COUNT = 41,
+    SD_TYPE_Q1_0  = 41,
+    SD_TYPE_COUNT = 42,
 };
 
 enum sd_log_level_t {
@@ -159,6 +160,7 @@ typedef struct {
     float target_overlap;
     float rel_size_x;
     float rel_size_y;
+    const char* extra_tiling_args;
 } sd_tiling_params_t;
 
 typedef struct {
@@ -166,6 +168,14 @@ typedef struct {
     const char* path;
 } sd_embedding_t;
 
+enum sd_vae_format_t {
+    SD_VAE_FORMAT_AUTO = -1,
+    SD_VAE_FORMAT_FLUX,
+    SD_VAE_FORMAT_SD3,
+    SD_VAE_FORMAT_FLUX2,
+    SD_VAE_FORMAT_COUNT,
+};
+
 typedef struct {
     const char* model_path;
     const char* clip_l_path;
@@ -176,6 +186,7 @@ typedef struct {
     const char* llm_vision_path;
     const char* diffusion_model_path;
     const char* high_noise_diffusion_model_path;
+    const char* uncond_diffusion_model_path;
     const char* embeddings_connectors_path;
     const char* vae_path;
     const char* audio_vae_path;
@@ -186,7 +197,6 @@ typedef struct {
     const char* photo_maker_path;
     const char* tensor_type_rules;
     bool vae_decode_only;
-    bool free_params_immediately;
     int n_threads;
     enum sd_type_t wtype;
     enum rng_type_t rng_type;
@@ -210,7 +220,9 @@ typedef struct {
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
     bool qwen_image_zero_cond_t;
+    enum sd_vae_format_t vae_format;
     float max_vram;  // GiB budget for graph-cut segmented param offload (0 = disabled, -1 = auto free VRAM minus 1 GiB)
+    bool stream_layers;  // Enable residency+prefetch streaming on top of --max-vram (no effect without --max-vram)
     const char* backend;
     const char* params_backend;
 } sd_ctx_params_t;
@@ -331,6 +343,8 @@ typedef struct {
     int steps;
     float denoising_strength;
     int upscale_tile_size;
+    float* custom_sigmas;
+    int custom_sigmas_count;
 } sd_hires_params_t;
 
 typedef struct {
@@ -381,6 +395,7 @@ typedef struct {
     float vace_strength;
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
+    sd_hires_params_t hires;
 } sd_vid_gen_params_t;
 
 typedef struct sd_ctx_t sd_ctx_t;
@@ -475,6 +490,10 @@ SD_API bool preprocess_canny(sd_image_t image,
 SD_API const char* sd_commit(void);
 SD_API const char* sd_version(void);
 
+// for C API, caller needs to call free_sd_images to free the memory after use
+// This helps avoid CRT problems on Windows when memory is allocated in the library but freed in the caller, which may use a different CRT.
+SD_API void free_sd_images(sd_image_t* result_images, int num_images);
+
 #ifdef __cplusplus
 }
 #endif

script/convert_fp8_scale_to_bf16.py

@@ -0,0 +1,283 @@
+#!/usr/bin/env python
+import argparse
+import json
+import math
+import os
+import struct
+from collections import Counter
+from pathlib import Path
+
+import torch
+from safetensors import safe_open
+
+
+FLOAT_DTYPES = {
+    "BF16",
+    "F16",
+    "F32",
+    "F64",
+    "F8_E4M3",
+    "F8_E4M3FN",
+    "F8_E5M2",
+}
+
+FP8_DTYPES = {
+    "F8_E4M3",
+    "F8_E4M3FN",
+    "F8_E5M2",
+}
+
+DTYPE_SIZES = {
+    "BOOL": 1,
+    "U8": 1,
+    "I8": 1,
+    "F8_E4M3": 1,
+    "F8_E4M3FN": 1,
+    "F8_E5M2": 1,
+    "U16": 2,
+    "I16": 2,
+    "F16": 2,
+    "BF16": 2,
+    "U32": 4,
+    "I32": 4,
+    "F32": 4,
+    "U64": 8,
+    "I64": 8,
+    "F64": 8,
+}
+
+
+def read_safetensors_header(path: Path):
+    with path.open("rb") as f:
+        header_len = struct.unpack("<Q", f.read(8))[0]
+        header = f.read(header_len).decode("utf-8").rstrip()
+    return json.loads(header)
+
+
+def numel(shape):
+    return math.prod(shape) if shape else 1
+
+
+def scale_key_for_weight(name: str):
+    if name.endswith(".weight"):
+        return name[:-len(".weight")] + ".weight_scale"
+    if name.endswith("weight"):
+        return name + "_scale"
+    return None
+
+
+def tensor_nbytes(dtype: str, shape):
+    return numel(shape) * DTYPE_SIZES[dtype]
+
+
+def build_output_plan(header):
+    entries = {k: v for k, v in header.items() if k != "__metadata__"}
+    paired_scale_keys = set()
+    plan = []
+
+    for name, info in entries.items():
+        scale_key = scale_key_for_weight(name)
+        if info["dtype"] in FP8_DTYPES and scale_key in entries:
+            paired_scale_keys.add(scale_key)
+
+    for name, info in entries.items():
+        if name in paired_scale_keys:
+            continue
+
+        dtype = info["dtype"]
+        shape = info["shape"]
+        scale_key = scale_key_for_weight(name)
+
+        if dtype in FP8_DTYPES and scale_key in entries:
+            scale_info = entries[scale_key]
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": "BF16",
+                    "shape": shape,
+                    "mode": "fp8_scaled_weight",
+                    "scale_key": scale_key,
+                }
+            )
+            continue
+
+        if dtype in FLOAT_DTYPES:
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": "BF16",
+                    "shape": shape,
+                    "mode": "float_to_bf16",
+                }
+            )
+        else:
+            plan.append(
+                {
+                    "name": name,
+                    "source_dtype": dtype,
+                    "output_dtype": dtype,
+                    "shape": shape,
+                    "mode": "copy",
+                }
+            )
+
+    metadata = dict(header.get("__metadata__", {}) or {})
+    metadata["format"] = "pt"
+    metadata["conversion"] = "fp8_weight_scale_to_bf16"
+
+    output_header = {"__metadata__": metadata}
+    offset = 0
+    for item in plan:
+        size = tensor_nbytes(item["output_dtype"], item["shape"])
+        output_header[item["name"]] = {
+            "dtype": item["output_dtype"],
+            "shape": item["shape"],
+            "data_offsets": [offset, offset + size],
+        }
+        offset += size
+
+    return plan, output_header, offset
+
+
+def write_tensor_bytes(out, tensor):
+    tensor = tensor.detach().cpu().contiguous()
+    if tensor.numel() == 0:
+        return
+    if tensor.dtype == torch.bfloat16:
+        tensor.view(torch.uint16).numpy().tofile(out)
+    elif tensor.dtype in (getattr(torch, "float8_e4m3fn", None), getattr(torch, "float8_e5m2", None)):
+        tensor.view(torch.uint8).numpy().tofile(out)
+    else:
+        tensor.numpy().tofile(out)
+
+
+def scale_view_for_chunk(scale, chunk, first_dim_start=0, first_dim_end=None):
+    scale = scale.to(torch.float32)
+
+    if scale.numel() == 1:
+        return scale.reshape((1,) * chunk.ndim)
+
+    if chunk.ndim > 0 and scale.ndim == 1:
+        if first_dim_end is not None and scale.shape[0] >= first_dim_end:
+            scale = scale[first_dim_start:first_dim_end]
+        if scale.shape[0] == chunk.shape[0]:
+            return scale.reshape((scale.shape[0],) + (1,) * (chunk.ndim - 1))
+
+    return scale
+
+
+def write_scaled_fp8_weight(out, weight, scale, chunk_rows):
+    if weight.ndim == 0:
+        result = weight.to(torch.float32) * scale_view_for_chunk(scale, weight)
+        write_tensor_bytes(out, result.to(torch.bfloat16))
+        return
+
+    rows = weight.shape[0]
+    for start in range(0, rows, chunk_rows):
+        end = min(start + chunk_rows, rows)
+        chunk = weight[start:end].to(torch.float32)
+        scale_view = scale_view_for_chunk(scale, chunk, start, end)
+        result = chunk * scale_view
+        write_tensor_bytes(out, result.to(torch.bfloat16))
+
+
+def write_float_as_bf16(out, tensor, chunk_rows):
+    if tensor.dtype == torch.bfloat16:
+        write_tensor_bytes(out, tensor)
+        return
+
+    if tensor.ndim == 0:
+        write_tensor_bytes(out, tensor.to(torch.bfloat16))
+        return
+
+    rows = tensor.shape[0]
+    for start in range(0, rows, chunk_rows):
+        end = min(start + chunk_rows, rows)
+        write_tensor_bytes(out, tensor[start:end].to(torch.bfloat16))
+
+
+def convert(input_path: Path, output_path: Path, chunk_rows: int, dry_run: bool):
+    header = read_safetensors_header(input_path)
+    plan, output_header, data_size = build_output_plan(header)
+
+    source_counts = Counter(item["source_dtype"] for item in plan)
+    output_counts = Counter(item["output_dtype"] for item in plan)
+    scaled_count = sum(item["mode"] == "fp8_scaled_weight" for item in plan)
+    dropped_scales = sum(item["mode"] == "fp8_scaled_weight" for item in plan)
+    header_bytes = json.dumps(output_header, separators=(",", ":")).encode("utf-8")
+    expected_size = 8 + len(header_bytes) + data_size
+
+    print(f"input:  {input_path}")
+    print(f"output: {output_path}")
+    print(f"tensors written: {len(plan)}")
+    print(f"scaled fp8 weights dequantized: {scaled_count}")
+    print(f"weight_scale tensors dropped: {dropped_scales}")
+    print(f"source dtypes: {dict(sorted(source_counts.items()))}")
+    print(f"output dtypes: {dict(sorted(output_counts.items()))}")
+    print(f"expected output size: {expected_size / (1024 ** 3):.2f} GiB")
+
+    if dry_run:
+        return
+
+    if output_path.exists():
+        raise FileExistsError(f"{output_path} already exists; pass --overwrite to replace it")
+
+    tmp_path = output_path.with_suffix(output_path.suffix + ".tmp")
+    if tmp_path.exists():
+        raise FileExistsError(f"{tmp_path} already exists; remove it or choose another output")
+
+    with safe_open(str(input_path), framework="pt", device="cpu") as sf, tmp_path.open("wb") as out:
+        out.write(struct.pack("<Q", len(header_bytes)))
+        out.write(header_bytes)
+
+        for index, item in enumerate(plan, 1):
+            name = item["name"]
+            print(f"[{index:04d}/{len(plan):04d}] {name} -> {item['output_dtype']}")
+
+            tensor = sf.get_tensor(name)
+            if item["mode"] == "fp8_scaled_weight":
+                scale = sf.get_tensor(item["scale_key"])
+                write_scaled_fp8_weight(out, tensor, scale, chunk_rows)
+            elif item["mode"] == "float_to_bf16":
+                write_float_as_bf16(out, tensor, chunk_rows)
+            else:
+                write_tensor_bytes(out, tensor)
+
+        actual_size = out.tell()
+
+    if actual_size != expected_size:
+        tmp_path.unlink(missing_ok=True)
+        raise RuntimeError(f"wrote {actual_size} bytes, expected {expected_size} bytes")
+
+    tmp_path.replace(output_path)
+    print("done")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Convert an fp8 safetensors checkpoint with weight_scale tensors to bf16."
+    )
+    parser.add_argument("--input", default="ideogram4_fp8.safetensors", type=Path)
+    parser.add_argument("--output", default="ideogram4_bf16.safetensors", type=Path)
+    parser.add_argument("--chunk-rows", default=1024, type=int)
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--overwrite", action="store_true")
+    args = parser.parse_args()
+
+    input_path = args.input.resolve()
+    output_path = args.output.resolve()
+
+    if args.chunk_rows < 1:
+        raise ValueError("--chunk-rows must be >= 1")
+    if not input_path.exists():
+        raise FileNotFoundError(input_path)
+    if args.overwrite and output_path.exists():
+        output_path.unlink()
+
+    convert(input_path, output_path, args.chunk_rows, args.dry_run)
+
+
+if __name__ == "__main__":
+    main()

src/convert.cpp

@@ -3,12 +3,12 @@
 #include <regex>
 #include <vector>
 
-#include "model.h"
 #include "model_io/gguf_io.h"
 #include "model_io/safetensors_io.h"
+#include "model_loader.h"
 #include "util.h"
 
-#include "ggml-cpu.h"
+#include "ggml_extend_backend.h"
 
 static ggml_type get_export_tensor_type(ModelLoader& model_loader,
                                         const TensorStorage& tensor_storage,
@@ -103,7 +103,7 @@ bool convert(const char* input_path,
     bool output_is_safetensors = ends_with(output_path, ".safetensors");
     TensorTypeRules type_rules = parse_tensor_type_rules(tensor_type_rules);
 
-    auto backend    = ggml_backend_cpu_init();
+    auto backend    = sd_backend_cpu_init();
     size_t mem_size = 1 * 1024 * 1024;  // for padding
     mem_size += model_loader.get_tensor_storage_map().size() * ggml_tensor_overhead();
     mem_size += model_loader.get_params_mem_size(backend, type);

src/core/ggml_extend_backend.cpp

@@ -1,4 +1,4 @@
-#include "ggml_extend_backend.h"
+#include "core/ggml_extend_backend.h"
 
 #include <algorithm>
 #include <cctype>
@@ -8,7 +8,8 @@
 #include <stdexcept>
 #include <vector>
 
-#include "util.h"
+#include "core/util.h"
+#include "stable-diffusion.h"
 
 static std::string trim_copy(const std::string& value) {
     size_t begin = 0;
@@ -44,6 +45,10 @@ static bool is_default_backend_token(const std::string& name) {
     return lower.empty() || lower == "default" || lower == "auto";
 }
 
+static bool is_disk_backend_token(const std::string& name) {
+    return lower_copy(trim_copy(name)) == "disk";
+}
+
 static bool parse_backend_module(const std::string& raw_name, SDBackendModule* module) {
     std::string name = lower_copy(trim_copy(raw_name));
     name.erase(std::remove(name.begin(), name.end(), '-'), name.end());
@@ -300,6 +305,61 @@ static ggml_backend_t init_named_backend(const std::string& name) {
     return ggml_backend_init_by_name(resolved.c_str(), nullptr);
 }
 
+bool sd_backend_is_cpu(ggml_backend_t backend) {
+    if (backend == nullptr) {
+        return false;
+    }
+    auto dev = ggml_backend_get_device(backend);
+    return dev != nullptr && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU;
+}
+
+ggml_backend_t sd_backend_cpu_init() {
+    ggml_backend_load_all_once();
+    return ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);
+}
+
+bool sd_backend_cpu_set_n_threads(ggml_backend_t backend, int n_threads) {
+    if (backend == nullptr) {
+        return false;
+    }
+    auto dev = ggml_backend_get_device(backend);
+    if (dev != nullptr && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
+        auto reg                           = ggml_backend_dev_backend_reg(dev);
+        auto ggml_backend_set_n_threads_fn = (ggml_backend_set_n_threads_t)ggml_backend_reg_get_proc_address(reg, "ggml_backend_set_n_threads");
+        if (ggml_backend_set_n_threads_fn != nullptr) {
+            ggml_backend_set_n_threads_fn(backend, n_threads);
+            return true;
+        }
+    }
+    return false;
+}
+
+const char* sd_get_system_info() {
+    static std::string cache_info = []() -> std::string {
+        ggml_backend_load_all_once();
+        std::stringstream ss;
+        ss << "System Info: \n";
+        auto dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        if (dev != nullptr) {
+            auto reg                          = ggml_backend_dev_backend_reg(dev);
+            auto ggml_backend_get_features_fn = (ggml_backend_get_features_t)ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features");
+            if (ggml_backend_get_features_fn != nullptr) {
+                ggml_backend_feature* feat = ggml_backend_get_features_fn(reg);
+                while (feat->name && feat->value) {
+                    ss << "   " << feat->name << " = " << feat->value << " | ";
+                    feat++;
+                }
+            } else {
+                LOG_WARN("unable to get CPU features");
+            }
+        } else {
+            LOG_WARN("unable to get CPU features");
+        }
+        return ss.str();
+    }();
+    return cache_info.c_str();
+}
+
 static ggml_backend_t sd_get_default_backend() {
     ggml_backend_load_all_once();
     static std::once_flag once;
@@ -349,10 +409,10 @@ static ggml_backend_t sd_get_default_backend() {
 
     if (!backend) {
         LOG_WARN("loading CPU backend");
-        backend = ggml_backend_cpu_init();
+        backend = sd_backend_cpu_init();
     }
 
-    if (ggml_backend_is_cpu(backend)) {
+    if (sd_backend_is_cpu(backend)) {
         LOG_DEBUG("Using CPU backend");
     }
 
@@ -448,15 +508,22 @@ ggml_backend_t SDBackendManager::params_backend(SDBackendModule module) {
     if (name.empty()) {
         return runtime_backend(module);
     }
+    if (is_disk_backend_token(name)) {
+        return runtime_backend(module);
+    }
     return init_cached_backend(name);
 }
 
 bool SDBackendManager::runtime_backend_is_cpu(SDBackendModule module) {
-    return ggml_backend_is_cpu(runtime_backend(module));
+    return sd_backend_is_cpu(runtime_backend(module));
 }
 
 bool SDBackendManager::params_backend_is_cpu(SDBackendModule module) {
-    return ggml_backend_is_cpu(params_backend(module));
+    return sd_backend_is_cpu(params_backend(module));
+}
+
+bool SDBackendManager::params_backend_is_disk(SDBackendModule module) const {
+    return is_disk_backend_token(params_assignment_.get(module));
 }
 
 bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule module) {
@@ -464,7 +531,7 @@ bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule modu
     if (backend == nullptr) {
         return false;
     }
-    if (ggml_backend_is_cpu(backend)) {
+    if (sd_backend_is_cpu(backend)) {
         return true;
     }
     ggml_backend_dev_t dev = ggml_backend_get_device(backend);
@@ -478,7 +545,6 @@ bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule modu
 
 bool SDBackendManager::init(const char* backend_spec,
                             const char* params_backend_spec,
-                            bool offload_params_to_cpu,
                             bool keep_clip_on_cpu,
                             bool keep_vae_on_cpu,
                             bool keep_control_net_on_cpu,
@@ -504,18 +570,20 @@ bool SDBackendManager::init(const char* backend_spec,
         }
     }
 
-    if (params_assignment_.empty() && offload_params_to_cpu) {
-        params_assignment_.set_default("cpu");
-    }
-
     return validate(error);
 }
 
 bool SDBackendManager::validate(std::string* error) const {
-    auto validate_name = [&](const std::string& name) -> bool {
+    auto validate_runtime_name = [&](const std::string& name) -> bool {
         if (is_default_backend_token(name)) {
             return true;
         }
+        if (is_disk_backend_token(name)) {
+            if (error != nullptr) {
+                *error = "backend 'disk' is only supported by params_backend";
+            }
+            return false;
+        }
         if (!sd_resolve_backend_name(name).empty()) {
             return true;
         }
@@ -524,18 +592,24 @@ bool SDBackendManager::validate(std::string* error) const {
         }
         return false;
     };
+    auto validate_params_name = [&](const std::string& name) -> bool {
+        if (is_disk_backend_token(name)) {
+            return true;
+        }
+        return validate_runtime_name(name);
+    };
 
-    if (!validate_name(runtime_assignment_.default_name) ||
-        !validate_name(params_assignment_.default_name)) {
+    if (!validate_runtime_name(runtime_assignment_.default_name) ||
+        !validate_params_name(params_assignment_.default_name)) {
         return false;
     }
     for (const auto& kv : runtime_assignment_.module_names) {
-        if (!validate_name(kv.second)) {
+        if (!validate_runtime_name(kv.second)) {
             return false;
         }
     }
     for (const auto& kv : params_assignment_.module_names) {
-        if (!validate_name(kv.second)) {
+        if (!validate_params_name(kv.second)) {
             return false;
         }
     }

src/core/ggml_extend_backend.h

@@ -1,5 +1,5 @@
-#ifndef __SD_GGML_EXTEND_BACKEND_H__
-#define __SD_GGML_EXTEND_BACKEND_H__
+#ifndef __SD_CORE_GGML_EXTEND_BACKEND_H__
+#define __SD_CORE_GGML_EXTEND_BACKEND_H__
 
 #include <cstdint>
 #include <cstring>
@@ -8,7 +8,6 @@
 #include <unordered_map>
 
 #include "ggml-backend.h"
-#include "ggml-cpu.h"
 #include "ggml.h"
 
 enum class SDBackendModule {
@@ -52,7 +51,6 @@ public:
 
     bool init(const char* backend_spec,
               const char* params_backend_spec,
-              bool offload_params_to_cpu,
               bool keep_clip_on_cpu,
               bool keep_vae_on_cpu,
               bool keep_control_net_on_cpu,
@@ -64,6 +62,7 @@ public:
 
     bool runtime_backend_is_cpu(SDBackendModule module);
     bool params_backend_is_cpu(SDBackendModule module);
+    bool params_backend_is_disk(SDBackendModule module) const;
     bool runtime_backend_supports_host_buffer(SDBackendModule module);
 
 private:
@@ -72,6 +71,9 @@ private:
 };
 
 bool sd_backend_is(ggml_backend_t backend, const std::string& name);
+bool sd_backend_is_cpu(ggml_backend_t backend);
+ggml_backend_t sd_backend_cpu_init();
+bool sd_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
 const char* sd_backend_module_name(SDBackendModule module);
 void ggml_ext_im_set_f32_1d(const struct ggml_tensor* tensor, int i, float value);
-#endif
+#endif  // __SD_CORE_GGML_EXTEND_BACKEND_H__

src/core/ggml_graph_cut.cpp

@@ -1,4 +1,4 @@
-#include "ggml_graph_cut.h"
+#include "core/ggml_graph_cut.h"
 
 #include <algorithm>
 #include <cstring>
@@ -8,16 +8,15 @@
 #include <stack>
 #include <unordered_map>
 
+#include "core/util.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
-#include "util.h"
 
-#include "../ggml/src/ggml-impl.h"
+#include "ggml/src/ggml-impl.h"
 
 namespace sd::ggml_graph_cut {
 
-    static constexpr double MAX_VRAM_BYTES_PER_GIB      = 1024.0 * 1024.0 * 1024.0;
-    static constexpr size_t MAX_VRAM_AUTO_RESERVE_BYTES = 1024ULL * 1024ULL * 1024ULL;
+    static constexpr double MAX_VRAM_BYTES_PER_GIB = 1024.0 * 1024.0 * 1024.0;
 
     static std::string graph_cut_tensor_display_name(const ggml_tensor* tensor) {
         if (tensor == nullptr) {
@@ -45,7 +44,9 @@ namespace sd::ggml_graph_cut {
         if (tensor == nullptr) {
             return false;
         }
-        return params_tensor_set.find(tensor) != params_tensor_set.end();
+        return params_tensor_set.find(tensor) != params_tensor_set.end() ||
+               (tensor->view_src != nullptr &&
+                params_tensor_set.find(tensor->view_src) != params_tensor_set.end());
     }
 
     static int graph_node_index_by_name(ggml_cgraph* gf, const char* name) {
@@ -93,45 +94,65 @@ namespace sd::ggml_graph_cut {
         return static_cast<float>(static_cast<double>(max_vram_bytes) / MAX_VRAM_BYTES_PER_GIB);
     }
 
-    static size_t resolve_auto_max_vram_bytes(ggml_backend_t backend) {
+    static size_t resolve_auto_max_vram_bytes(float spare_vram, ggml_backend_t backend) {
         if (backend == nullptr) {
-            LOG_WARN("--max-vram -1 requested, but no backend is available; disabling graph splitting");
+            LOG_WARN("--max-vram < 0 requested, but no backend is available; disabling graph splitting");
             return 0;
         }
 
         ggml_backend_dev_t dev = ggml_backend_get_device(backend);
         if (dev == nullptr) {
-            LOG_WARN("--max-vram -1 requested, but no backend device is available; disabling graph splitting");
+            LOG_WARN("--max-vram < 0 requested, but no backend device is available; disabling graph splitting");
             return 0;
         }
         if (ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU) {
-            LOG_WARN("--max-vram -1 requested, but the main backend is CPU; disabling graph splitting");
+            LOG_WARN("--max-vram < 0 requested, but the main backend is CPU; disabling graph splitting");
             return 0;
         }
 
         size_t free_vram  = 0;
         size_t total_vram = 0;
         ggml_backend_dev_memory(dev, &free_vram, &total_vram);
+        size_t spare_bytes = static_cast<size_t>(MAX_VRAM_BYTES_PER_GIB * spare_vram);
 
-        if (free_vram <= MAX_VRAM_AUTO_RESERVE_BYTES) {
-            LOG_WARN("--max-vram -1 requested, but free VRAM is %.2f GiB; reserving 1.00 GiB leaves no graph budget",
-                     free_vram / MAX_VRAM_BYTES_PER_GIB);
+        if (free_vram <= spare_bytes) {
+            LOG_WARN("--max-vram < 0 requested, but free VRAM is %.2f GiB; reserving %.2f GiB leaves no graph budget",
+                     free_vram / MAX_VRAM_BYTES_PER_GIB, spare_vram);
             return 0;
         }
 
-        const size_t max_vram_bytes = free_vram - MAX_VRAM_AUTO_RESERVE_BYTES;
-        LOG_INFO("--max-vram -1 auto-detected %.2f GiB free VRAM (%.2f GiB total), reserving 1.00 GiB; using %.2f GiB",
+        const size_t max_vram_bytes = free_vram - spare_bytes;
+        LOG_INFO("--max-vram < 0 auto-detected %.2f GiB free VRAM (%.2f GiB total), reserving %.2f GiB; using %.2f GiB",
                  free_vram / MAX_VRAM_BYTES_PER_GIB,
                  total_vram / MAX_VRAM_BYTES_PER_GIB,
+                 spare_vram,
                  max_vram_bytes / MAX_VRAM_BYTES_PER_GIB);
         return max_vram_bytes;
     }
 
     float resolve_max_vram_gib(float max_vram, ggml_backend_t backend) {
-        if (max_vram != -1.f) {
+        if (max_vram >= 0.f) {
             return max_vram;
         }
-        return max_vram_bytes_to_gib(resolve_auto_max_vram_bytes(backend));
+        return max_vram_bytes_to_gib(resolve_auto_max_vram_bytes(-max_vram, backend));
+    }
+
+    static bool is_segment_output_needed_after(const Plan& plan,
+                                               size_t end_segment_index,
+                                               int output_node_index) {
+        if (end_segment_index + 1 >= plan.segments.size()) {
+            return false;
+        }
+        for (size_t seg_idx = end_segment_index + 1; seg_idx < plan.segments.size(); ++seg_idx) {
+            const auto& segment = plan.segments[seg_idx];
+            for (const auto& input_ref : segment.input_refs) {
+                if (input_ref.type == Segment::INPUT_PREVIOUS_CUT &&
+                    input_ref.node_index == output_node_index) {
+                    return true;
+                }
+            }
+        }
+        return false;
     }
 
     static Segment make_segment_seed(const Plan& plan,
@@ -146,8 +167,11 @@ namespace sd::ggml_graph_cut {
         const auto& target_segment = plan.segments[end_segment_index];
         std::unordered_set<int> seen_output_node_indices;
         for (size_t seg_idx = start_segment_index; seg_idx <= end_segment_index; ++seg_idx) {
+            const bool is_boundary_segment = seg_idx == end_segment_index;
             for (int output_node_index : plan.segments[seg_idx].output_node_indices) {
-                if (seen_output_node_indices.insert(output_node_index).second) {
+                if ((is_boundary_segment ||
+                     is_segment_output_needed_after(plan, end_segment_index, output_node_index)) &&
+                    seen_output_node_indices.insert(output_node_index).second) {
                     seed.output_node_indices.push_back(output_node_index);
                 }
             }
@@ -399,23 +423,6 @@ namespace sd::ggml_graph_cut {
         return tensors;
     }
 
-    std::vector<ggml_tensor*> runtime_param_tensors(ggml_cgraph* gf, const Segment& segment, const char* log_desc) {
-        std::vector<ggml_tensor*> tensors = param_tensors(gf, segment);
-        std::vector<ggml_tensor*> filtered_tensors;
-        filtered_tensors.reserve(tensors.size());
-        for (ggml_tensor* tensor : tensors) {
-            if (tensor_buffer(tensor) == nullptr) {
-                LOG_WARN("%s graph cut skipping param input without buffer: segment=%s tensor=%s",
-                         log_desc == nullptr ? "unknown" : log_desc,
-                         segment.group_name.c_str(),
-                         tensor->name);
-                continue;
-            }
-            filtered_tensors.push_back(tensor);
-        }
-        return filtered_tensors;
-    }
-
     std::unordered_set<std::string> collect_future_input_names(ggml_cgraph* gf,
                                                                const Plan& plan,
                                                                size_t current_segment_index) {
@@ -486,6 +493,44 @@ namespace sd::ggml_graph_cut {
             return 0;
         }
 
+        struct TensorRuntimeBinding {
+            ggml_backend_buffer_t buffer = nullptr;
+            void* data                   = nullptr;
+            void* extra                  = nullptr;
+        };
+        std::unordered_map<ggml_tensor*, TensorRuntimeBinding> saved_bindings;
+        auto mark_measurement_external = [&](ggml_tensor* tensor) {
+            if (tensor == nullptr) {
+                return;
+            }
+            auto save_tensor = [&](ggml_tensor* t) {
+                if (t == nullptr || saved_bindings.find(t) != saved_bindings.end()) {
+                    return;
+                }
+                saved_bindings[t] = {t->buffer, t->data, t->extra};
+                // During real execution params and previous-cut inputs already
+                // have backend/cache buffers, so gallocr must not reserve them.
+                t->data = reinterpret_cast<void*>(static_cast<uintptr_t>(1));
+            };
+            save_tensor(tensor);
+            save_tensor(tensor->view_src);
+        };
+        for (const auto& input : segment.input_refs) {
+            if (input.type != Segment::INPUT_PARAM &&
+                input.type != Segment::INPUT_PREVIOUS_CUT) {
+                continue;
+            }
+            mark_measurement_external(input_tensor(gf, input));
+        }
+
+        std::unordered_map<ggml_tensor*, int32_t> saved_output_flags;
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            if (output != nullptr && saved_output_flags.find(output) == saved_output_flags.end()) {
+                saved_output_flags[output] = output->flags;
+            }
+        }
+
         ggml_context* graph_ctx    = nullptr;
         ggml_cgraph* segment_graph = build_segment_graph(gf, segment, &graph_ctx);
         ggml_gallocr_t allocr      = ggml_gallocr_new(ggml_backend_get_default_buffer_type(backend));
@@ -501,6 +546,14 @@ namespace sd::ggml_graph_cut {
 
         ggml_gallocr_free(allocr);
         ggml_free(graph_ctx);
+        for (const auto& kv : saved_output_flags) {
+            kv.first->flags = kv.second;
+        }
+        for (const auto& kv : saved_bindings) {
+            kv.first->buffer = kv.second.buffer;
+            kv.first->data   = kv.second.data;
+            kv.first->extra  = kv.second.extra;
+        }
         return buffer_size;
     }
 
@@ -668,7 +721,8 @@ namespace sd::ggml_graph_cut {
                 GGML_ASSERT(!candidate_plan.segments.empty());
 
                 const auto& candidate_segment = candidate_plan.segments.back();
-                if (graph_cut_segment_vram_bytes(candidate_segment) > max_graph_vram_bytes) {
+                const size_t candidate_bytes  = graph_cut_segment_vram_bytes(candidate_segment);
+                if (candidate_bytes > max_graph_vram_bytes) {
                     break;
                 }
 
@@ -698,9 +752,9 @@ namespace sd::ggml_graph_cut {
         }
 
         if (log_desc != nullptr) {
-            LOG_INFO("%s graph cut max_vram budget merge took %lld ms",
-                     log_desc,
-                     ggml_time_ms() - t_budget_begin);
+            LOG_DEBUG("%s graph cut max_vram budget merge took %lld ms",
+                      log_desc,
+                      ggml_time_ms() - t_budget_begin);
         }
 
         return merged_plan;
@@ -752,4 +806,54 @@ namespace sd::ggml_graph_cut {
         return resolved_plan;
     }
 
+    void annotate_residency(Plan& plan, size_t max_graph_vram_bytes) {
+        // Cached plans may be reused with a smaller live budget.
+        for (auto& seg : plan.segments) {
+            seg.residency = SegmentResidency::STREAMED;
+        }
+        if (max_graph_vram_bytes == 0 || plan.segments.size() < 2) {
+            return;
+        }
+
+        bool any_param_bearing = false;
+        for (const auto& seg : plan.segments) {
+            if (seg.input_param_bytes > 0) {
+                any_param_bearing = true;
+                break;
+            }
+        }
+        if (!any_param_bearing) {
+            return;
+        }
+
+        // Leave room for the largest active streamed segment.
+        size_t worst_streamed_footprint = 0;
+        for (const auto& seg : plan.segments) {
+            const size_t seg_footprint = seg.input_param_bytes +
+                                         seg.compute_buffer_size +
+                                         seg.output_bytes +
+                                         seg.input_previous_cut_bytes +
+                                         seg.input_external_bytes;
+            if (seg_footprint > worst_streamed_footprint) {
+                worst_streamed_footprint = seg_footprint;
+            }
+        }
+        constexpr size_t safety = 512ull * 1024 * 1024;
+        const size_t reserved   = safety + worst_streamed_footprint;
+
+        if (max_graph_vram_bytes <= reserved) {
+            return;
+        }
+        const size_t available = max_graph_vram_bytes - reserved;
+
+        size_t cumulative = 0;
+        for (auto& seg : plan.segments) {
+            if (cumulative + seg.input_param_bytes > available) {
+                break;
+            }
+            seg.residency = SegmentResidency::RESIDENT;
+            cumulative += seg.input_param_bytes;
+        }
+    }
+
 }  // namespace sd::ggml_graph_cut

src/core/ggml_graph_cut.h

@@ -1,7 +1,8 @@
-#ifndef __SD_GGML_GRAPH_CUT_H__
-#define __SD_GGML_GRAPH_CUT_H__
+#ifndef __SD_CORE_GGML_GRAPH_CUT_H__
+#define __SD_CORE_GGML_GRAPH_CUT_H__
 
 #include <array>
+#include <cstdint>
 #include <string>
 #include <unordered_set>
 #include <vector>
@@ -11,6 +12,12 @@
 
 namespace sd::ggml_graph_cut {
 
+    // Streaming residency for a segment's params.
+    enum class SegmentResidency : uint8_t {
+        STREAMED = 0,
+        RESIDENT = 1,
+    };
+
     struct Segment {
         enum InputType {
             INPUT_EXTERNAL = 0,
@@ -34,6 +41,7 @@ namespace sd::ggml_graph_cut {
         std::vector<int> internal_node_indices;
         std::vector<int> output_node_indices;
         std::vector<InputRef> input_refs;
+        SegmentResidency residency = SegmentResidency::STREAMED;
     };
 
     struct Plan {
@@ -72,7 +80,6 @@ namespace sd::ggml_graph_cut {
     ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index);
     ggml_tensor* input_tensor(ggml_cgraph* gf, const Segment::InputRef& input_ref);
     std::vector<ggml_tensor*> param_tensors(ggml_cgraph* gf, const Segment& segment);
-    std::vector<ggml_tensor*> runtime_param_tensors(ggml_cgraph* gf, const Segment& segment, const char* log_desc);
     std::unordered_set<std::string> collect_future_input_names(ggml_cgraph* gf,
                                                                const Plan& plan,
                                                                size_t current_segment_index);
@@ -101,6 +108,9 @@ namespace sd::ggml_graph_cut {
                       size_t max_graph_vram_bytes,
                       const std::unordered_set<const ggml_tensor*>& params_tensor_set,
                       const char* log_desc);
+
+    // Mark leading segments resident when they fit after streamed-segment headroom.
+    void annotate_residency(Plan& plan, size_t max_graph_vram_bytes);
 }  // namespace sd::ggml_graph_cut
 
-#endif
+#endif  // __SD_CORE_GGML_GRAPH_CUT_H__

src/core/ordered_map.hpp

@@ -1,5 +1,5 @@
-#ifndef __ORDERED_MAP_HPP__
-#define __ORDERED_MAP_HPP__
+#ifndef __SD_CORE_ORDERED_MAP_HPP__
+#define __SD_CORE_ORDERED_MAP_HPP__
 
 #include <iostream>
 #include <list>
@@ -174,4 +174,4 @@ public:
     }
 };
 
-#endif  // __ORDERED_MAP_HPP__
+#endif  // __SD_CORE_ORDERED_MAP_HPP__

src/core/rng.hpp

@@ -1,5 +1,5 @@
-#ifndef __RNG_H__
-#define __RNG_H__
+#ifndef __SD_CORE_RNG_HPP__
+#define __SD_CORE_RNG_HPP__
 
 #include <random>
 #include <vector>
@@ -32,4 +32,4 @@ public:
     }
 };
 
-#endif  // __RNG_H__
+#endif  // __SD_CORE_RNG_HPP__

src/core/rng_mt19937.hpp

@@ -1,10 +1,10 @@
-#ifndef __RNG_MT19937_HPP__
-#define __RNG_MT19937_HPP__
+#ifndef __SD_CORE_RNG_MT19937_HPP__
+#define __SD_CORE_RNG_MT19937_HPP__
 
 #include <cmath>
 #include <vector>
 
-#include "rng.hpp"
+#include "core/rng.hpp"
 
 // RNG imitiating torch cpu randn on CPU.
 // Port from pytorch, original license: https://github.com/pytorch/pytorch/blob/d01a7b0241ed1c4cded7e7ca097249feb343f072/LICENSE
@@ -144,4 +144,4 @@ public:
     }
 };
 
-#endif  // __RNG_MT19937_HPP__
+#endif  // __SD_CORE_RNG_MT19937_HPP__

src/core/rng_philox.hpp

@@ -1,10 +1,10 @@
-#ifndef __RNG_PHILOX_H__
-#define __RNG_PHILOX_H__
+#ifndef __SD_CORE_RNG_PHILOX_HPP__
+#define __SD_CORE_RNG_PHILOX_HPP__
 
 #include <cmath>
 #include <vector>
 
-#include "rng.hpp"
+#include "core/rng.hpp"
 
 // RNG imitiating torch cuda randn on CPU.
 // Port from: https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob/5ef669de080814067961f28357256e8fe27544f4/modules/rng_philox.py
@@ -122,4 +122,4 @@ public:
     }
 };
 
-#endif  // __RNG_PHILOX_H__
+#endif  // __SD_CORE_RNG_PHILOX_HPP__

src/core/tensor.hpp

@@ -1,5 +1,5 @@
-#ifndef __SD_TENSOR_HPP__
-#define __SD_TENSOR_HPP__
+#ifndef __SD_CORE_TENSOR_HPP__
+#define __SD_CORE_TENSOR_HPP__
 
 #include <algorithm>
 #include <cmath>
@@ -16,7 +16,7 @@
 #include <utility>
 #include <vector>
 
-#include "rng.hpp"
+#include "core/rng.hpp"
 
 namespace sd {
 
@@ -235,6 +235,7 @@ namespace sd {
 
         Tensor& masked_fill_(const Tensor<uint8_t>& mask, const T& value);
 
+        T sum() const;
         T mean() const;
 
         static Tensor zeros(std::vector<int64_t> shape) {
@@ -327,6 +328,24 @@ namespace sd {
         std::vector<int64_t> shape_;
     };
 
+    template <typename T>
+    inline T Tensor<T>::sum() const {
+        T total = T{};
+        for (const T& value : data_) {
+            total += value;
+        }
+        return total;
+    }
+
+    template <>
+    inline float Tensor<float>::sum() const {
+        double total = 0.0;
+        for (float value : data_) {
+            total += static_cast<double>(value);
+        }
+        return static_cast<float>(total);
+    }
+
     template <typename T>
     inline T Tensor<T>::mean() const {
         if (empty()) {
@@ -1642,4 +1661,4 @@ namespace sd {
 
 }  // namespace sd
 
-#endif
+#endif  // __SD_CORE_TENSOR_HPP__

src/core/tensor_ggml.hpp

@@ -1,5 +1,5 @@
-#ifndef __SD_TENSOR_GGML_HPP__
-#define __SD_TENSOR_GGML_HPP__
+#ifndef __SD_CORE_TENSOR_GGML_HPP__
+#define __SD_CORE_TENSOR_GGML_HPP__
 
 #include <array>
 #include <cstring>
@@ -8,8 +8,8 @@
 #include <string>
 #include <type_traits>
 
+#include "core/tensor.hpp"
 #include "ggml.h"
-#include "tensor.hpp"
 
 namespace sd {
 
@@ -124,4 +124,4 @@ namespace sd {
 
 }  // namespace sd
 
-#endif
+#endif  // __SD_CORE_TENSOR_GGML_HPP__

src/core/util.cpp

@@ -1,8 +1,10 @@
-#include "util.h"
+#include "core/util.h"
 #include <algorithm>
+#include <cctype>
 #include <cmath>
 #include <codecvt>
 #include <cstdarg>
+#include <exception>
 #include <fstream>
 #include <locale>
 #include <regex>
@@ -11,7 +13,7 @@
 #include <thread>
 #include <unordered_set>
 #include <vector>
-#include "preprocessing.hpp"
+#include "runtime/preprocessing.hpp"
 
 #if defined(__APPLE__) && defined(__MACH__)
 #include <sys/sysctl.h>
@@ -23,9 +25,7 @@
 #include <unistd.h>
 #endif
 
-#include "ggml-backend.h"
 #include "ggml.h"
-#include "ggml_extend_backend.h"
 #include "stable-diffusion.h"
 
 bool ends_with(const std::string& str, const std::string& ending) {
@@ -406,7 +406,89 @@ std::vector<std::string> split_string(const std::string& str, char delimiter) {
     return result;
 }
 
-static std::string build_progress_bar(int step, int steps) {
+KeyValueArgs parse_key_value_args(const char* args, const char* context) {
+    KeyValueArgs pairs;
+
+    if (args == nullptr || args[0] == '\0') {
+        return pairs;
+    }
+
+    std::string raw(args);
+    size_t start = 0;
+    for (size_t pos = 0; pos <= raw.size(); ++pos) {
+        if (pos != raw.size() && raw[pos] != ',' && raw[pos] != ';') {
+            continue;
+        }
+
+        std::string token = trim(raw.substr(start, pos - start));
+        if (!token.empty()) {
+            size_t eq = token.find('=');
+            if (eq == std::string::npos) {
+                const char* log_context = context ? context : "key=value arg";
+                LOG_WARN("ignoring malformed %s '%s'", log_context, token.c_str());
+            } else {
+                std::string key   = trim(token.substr(0, eq));
+                std::string value = trim(token.substr(eq + 1));
+                pairs.emplace_back(std::move(key), std::move(value));
+            }
+        }
+
+        start = pos + 1;
+    }
+
+    return pairs;
+}
+
+KeyValueArgs parse_key_value_args(const std::string& args, const char* context) {
+    return parse_key_value_args(args.c_str(), context);
+}
+
+bool parse_strict_float(const std::string& text, float& value) {
+    try {
+        size_t consumed = 0;
+        float parsed    = std::stof(text, &consumed);
+        if (!trim(text.substr(consumed)).empty()) {
+            return false;
+        }
+        value = parsed;
+        return true;
+    } catch (const std::exception&) {
+        return false;
+    }
+}
+
+bool parse_strict_int(const std::string& text, int& value) {
+    try {
+        size_t consumed = 0;
+        int parsed      = std::stoi(text, &consumed);
+        if (!trim(text.substr(consumed)).empty()) {
+            return false;
+        }
+        value = parsed;
+        return true;
+    } catch (const std::exception&) {
+        return false;
+    }
+}
+
+bool parse_strict_bool(const std::string& text, bool& value) {
+    std::string lowered = trim(text);
+    std::transform(lowered.begin(), lowered.end(), lowered.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+
+    if (lowered == "1" || lowered == "true" || lowered == "yes" || lowered == "on") {
+        value = true;
+        return true;
+    }
+    if (lowered == "0" || lowered == "false" || lowered == "no" || lowered == "off") {
+        value = false;
+        return true;
+    }
+    return false;
+}
+
+static std::string build_progress_bar(int step, int steps, char progress_char = '=', bool show_head = true) {
     std::string progress = "  |";
     int max_progress     = 50;
     int32_t current      = 0;
@@ -416,21 +498,21 @@ static std::string build_progress_bar(int step, int steps) {
     for (int i = 0; i < 50; i++) {
         if (i > current) {
             progress += " ";
-        } else if (i == current && i != max_progress - 1) {
+        } else if (show_head && i == current && i != max_progress - 1) {
             progress += ">";
         } else {
-            progress += "=";
+            progress += progress_char;
         }
     }
     progress += "|";
     return progress;
 }
 
-static void print_progress_line(int step, int steps, const std::string& speed_text) {
+static void print_progress_line(int step, int steps, const std::string& speed_text, char progress_char = '=', bool show_head = true) {
     if (step == 0) {
         return;
     }
-    std::string progress = build_progress_bar(step, steps);
+    std::string progress = build_progress_bar(step, steps, progress_char, show_head);
     const char* lf       = (step == steps ? "\n" : "");
     printf("\r%s %i/%i - %s\033[K%s", progress.c_str(), step, steps, speed_text.c_str(), lf);
     fflush(stdout);  // for linux
@@ -470,9 +552,9 @@ void pretty_bytes_progress(int step, int steps, uint64_t bytes_processed, float
 
     double speed_mb = bytes_per_second / (1024.0 * 1024.0);
     if (speed_mb >= 1024.0) {
-        print_progress_line(step, steps, sd_format("%.2fGB/s", speed_mb / 1024.0));
+        print_progress_line(step, steps, sd_format("%.2fGB/s", speed_mb / 1024.0), '#', false);
     } else {
-        print_progress_line(step, steps, sd_format("%.2fMB/s", speed_mb));
+        print_progress_line(step, steps, sd_format("%.2fMB/s", speed_mb), '#', false);
     }
 }
 
@@ -758,29 +840,135 @@ std::vector<std::pair<std::string, float>> parse_prompt_attention(const std::str
     return res;
 }
 
-// namespace is needed to avoid conflicts with ggml_backend_extend.hpp
-namespace ggml_cpu {
-#include "ggml-cpu.h"
+static size_t get_utf8_char_len(char c) {
+    unsigned char uc = static_cast<unsigned char>(c);
+    if ((uc & 0x80) == 0) {
+        return 1;
+    }
+    if ((uc & 0xE0) == 0xC0) {
+        return 2;
+    }
+    if ((uc & 0xF0) == 0xE0) {
+        return 3;
+    }
+    if ((uc & 0xF8) == 0xF0) {
+        return 4;
+    }
+    return 1;
 }
 
-const char* sd_get_system_info() {
-    using namespace ggml_cpu;
-    static char buffer[1024];
-    std::stringstream ss;
-    ss << "System Info: \n";
-    ss << "    SSE3 = " << ggml_cpu_has_sse3() << " | ";
-    ss << "    AVX = " << ggml_cpu_has_avx() << " | ";
-    ss << "    AVX2 = " << ggml_cpu_has_avx2() << " | ";
-    ss << "    AVX512 = " << ggml_cpu_has_avx512() << " | ";
-    ss << "    AVX512_VBMI = " << ggml_cpu_has_avx512_vbmi() << " | ";
-    ss << "    AVX512_VNNI = " << ggml_cpu_has_avx512_vnni() << " | ";
-    ss << "    FMA = " << ggml_cpu_has_fma() << " | ";
-    ss << "    NEON = " << ggml_cpu_has_neon() << " | ";
-    ss << "    ARM_FMA = " << ggml_cpu_has_arm_fma() << " | ";
-    ss << "    F16C = " << ggml_cpu_has_f16c() << " | ";
-    ss << "    FP16_VA = " << ggml_cpu_has_fp16_va() << " | ";
-    ss << "    WASM_SIMD = " << ggml_cpu_has_wasm_simd() << " | ";
-    ss << "    VSX = " << ggml_cpu_has_vsx() << " | ";
-    snprintf(buffer, sizeof(buffer), "%s", ss.str().c_str());
-    return buffer;
+static bool is_ascii_alpha(char c) {
+    return (c >= 'a' && c <= 'z') || (c >= 'A' && c <= 'Z');
+}
+
+static bool starts_with_at(const std::string& text, size_t pos, const std::string& needle) {
+    return pos + needle.size() <= text.size() && text.compare(pos, needle.size(), needle) == 0;
+}
+
+static bool is_word_internal_apostrophe(const std::string& text, size_t pos) {
+    return pos > 0 && pos + 1 < text.size() &&
+           is_ascii_alpha(text[pos - 1]) && is_ascii_alpha(text[pos + 1]);
+}
+
+static std::vector<std::pair<std::string, bool>> split_quotation(const std::string& text) {
+    static const std::vector<std::pair<std::string, std::string>> quote_pairs = {
+        {"'", "'"},
+        {"\"", "\""},
+        {"\xE2\x80\x98", "\xE2\x80\x99"},
+        {"\xE2\x80\x9C", "\xE2\x80\x9D"},
+    };
+
+    std::vector<std::pair<std::string, bool>> result;
+    size_t segment_start = 0;
+    size_t i             = 0;
+
+    auto push_segment = [&](size_t begin, size_t end, bool matched) {
+        if (end > begin) {
+            result.emplace_back(text.substr(begin, end - begin), matched);
+        }
+    };
+
+    while (i < text.size()) {
+        bool matched_quote = false;
+        for (const auto& quote_pair : quote_pairs) {
+            const std::string& open_quote  = quote_pair.first;
+            const std::string& close_quote = quote_pair.second;
+            if (!starts_with_at(text, i, open_quote)) {
+                continue;
+            }
+            if (open_quote == "'" && is_word_internal_apostrophe(text, i)) {
+                continue;
+            }
+
+            size_t search_pos = i + open_quote.size();
+            size_t close_pos  = std::string::npos;
+            bool invalid      = false;
+            while (search_pos < text.size()) {
+                if (open_quote != close_quote && starts_with_at(text, search_pos, open_quote)) {
+                    invalid = true;
+                    break;
+                }
+                if (starts_with_at(text, search_pos, close_quote)) {
+                    if (close_quote == "'" && is_word_internal_apostrophe(text, search_pos)) {
+                        search_pos += close_quote.size();
+                        continue;
+                    }
+                    close_pos = search_pos;
+                    break;
+                }
+
+                size_t char_len = get_utf8_char_len(text[search_pos]);
+                if (search_pos + char_len > text.size()) {
+                    char_len = 1;
+                }
+                search_pos += char_len;
+            }
+            if (invalid || close_pos == std::string::npos) {
+                continue;
+            }
+
+            size_t quote_start = i;
+            push_segment(segment_start, quote_start, false);
+            i = close_pos + close_quote.size();
+            push_segment(quote_start, i, true);
+            segment_start = i;
+            matched_quote = true;
+            break;
+        }
+        if (!matched_quote) {
+            size_t char_len = get_utf8_char_len(text[i]);
+            if (i + char_len > text.size()) {
+                char_len = 1;
+            }
+            i += char_len;
+        }
+    }
+
+    push_segment(segment_start, text.size(), false);
+    return result;
+}
+
+std::vector<std::pair<std::string, float>> split_quotation_attention(
+    const std::vector<std::pair<std::string, float>>& parsed_attention) {
+    std::vector<std::pair<std::string, float>> result;
+    for (const auto& item : parsed_attention) {
+        const std::string& text = item.first;
+        float weight            = item.second;
+        for (const auto& part : split_quotation(text)) {
+            if (part.second) {
+                size_t i = 0;
+                while (i < part.first.size()) {
+                    size_t char_len = get_utf8_char_len(part.first[i]);
+                    if (i + char_len > part.first.size()) {
+                        char_len = 1;
+                    }
+                    result.emplace_back(part.first.substr(i, char_len), weight);
+                    i += char_len;
+                }
+            } else {
+                result.emplace_back(part.first, weight);
+            }
+        }
+    }
+    return result;
 }

src/core/util.h

@@ -1,14 +1,15 @@
-#ifndef __UTIL_H__
-#define __UTIL_H__
+#ifndef __SD_CORE_UTIL_H__
+#define __SD_CORE_UTIL_H__
 
 #include <cstdint>
 #include <memory>
 #include <string>
+#include <utility>
 #include <vector>
 
+#include "core/tensor.hpp"
 #include "ggml-backend.h"
 #include "stable-diffusion.h"
-#include "tensor.hpp"
 
 #define SAFE_STR(s) ((s) ? (s) : "")
 #define BOOL_STR(b) ((b) ? "true" : "false")
@@ -65,6 +66,15 @@ protected:
 
 std::string path_join(const std::string& p1, const std::string& p2);
 std::vector<std::string> split_string(const std::string& str, char delimiter);
+
+using KeyValueArgs = std::vector<std::pair<std::string, std::string>>;
+
+KeyValueArgs parse_key_value_args(const char* args, const char* context = "key=value arg");
+KeyValueArgs parse_key_value_args(const std::string& args, const char* context = "key=value arg");
+bool parse_strict_float(const std::string& text, float& value);
+bool parse_strict_int(const std::string& text, int& value);
+bool parse_strict_bool(const std::string& text, bool& value);
+
 void pretty_progress(int step, int steps, float time);
 void pretty_bytes_progress(int step, int steps, uint64_t bytes_processed, float elapsed_seconds);
 
@@ -73,6 +83,8 @@ void log_printf(sd_log_level_t level, const char* file, int line, const char* fo
 std::string trim(const std::string& s);
 
 std::vector<std::pair<std::string, float>> parse_prompt_attention(const std::string& text);
+std::vector<std::pair<std::string, float>> split_quotation_attention(
+    const std::vector<std::pair<std::string, float>>& parsed_attention);
 
 sd_progress_cb_t sd_get_progress_callback();
 void* sd_get_progress_callback_data();
@@ -91,4 +103,4 @@ bool sd_backend_is(ggml_backend_t backend, const std::string& name);
 #define LOG_INFO(format, ...) log_printf(SD_LOG_INFO, __FILE__, __LINE__, format, ##__VA_ARGS__)
 #define LOG_WARN(format, ...) log_printf(SD_LOG_WARN, __FILE__, __LINE__, format, ##__VA_ARGS__)
 #define LOG_ERROR(format, ...) log_printf(SD_LOG_ERROR, __FILE__, __LINE__, format, ##__VA_ARGS__)
-#endif  // __UTIL_H__
+#endif  // __SD_CORE_UTIL_H__

src/diffusion_model.hpp

@@ -1,775 +0,0 @@
-#ifndef __DIFFUSION_MODEL_H__
-#define __DIFFUSION_MODEL_H__
-
-#include <optional>
-#include "anima.hpp"
-#include "ernie_image.hpp"
-#include "flux.hpp"
-#include "hidream_o1.hpp"
-#include "ltxv.hpp"
-#include "mmdit.hpp"
-#include "qwen_image.hpp"
-#include "tensor_ggml.hpp"
-#include "unet.hpp"
-#include "wan.hpp"
-#include "z_image.hpp"
-
-struct DiffusionParams {
-    const sd::Tensor<float>* x                                         = nullptr;
-    const sd::Tensor<float>* timesteps                                 = nullptr;
-    const sd::Tensor<float>* audio_x                                   = nullptr;
-    const sd::Tensor<float>* audio_timesteps                           = nullptr;
-    const sd::Tensor<float>* context                                   = nullptr;
-    const sd::Tensor<float>* c_concat                                  = nullptr;
-    const sd::Tensor<float>* y                                         = nullptr;
-    const sd::Tensor<int32_t>* t5_ids                                  = nullptr;
-    const sd::Tensor<float>* t5_weights                                = nullptr;
-    const sd::Tensor<float>* guidance                                  = nullptr;
-    const std::vector<sd::Tensor<float>>* ref_latents                  = nullptr;
-    const sd::Tensor<int32_t>* input_ids                               = nullptr;
-    const sd::Tensor<int32_t>* input_pos                               = nullptr;
-    const sd::Tensor<int32_t>* token_types                             = nullptr;
-    const sd::Tensor<int32_t>* vinput_mask                             = nullptr;
-    const std::vector<sd::Tensor<float>>* vlm_images                   = nullptr;
-    const std::vector<std::pair<int, sd::Tensor<float>>>* image_embeds = nullptr;
-    bool increase_ref_index                                            = false;
-    int num_video_frames                                               = -1;
-    const std::vector<sd::Tensor<float>>* controls                     = nullptr;
-    float control_strength                                             = 0.f;
-    const sd::Tensor<float>* vace_context                              = nullptr;
-    float vace_strength                                                = 1.f;
-    int audio_length                                                   = 0;
-    float frame_rate                                                   = 24.f;
-    const sd::Tensor<float>* video_positions                           = nullptr;
-    const std::vector<int>* skip_layers                                = nullptr;
-};
-
-template <typename T>
-static inline const sd::Tensor<T>& tensor_or_empty(const sd::Tensor<T>* tensor) {
-    static const sd::Tensor<T> kEmpty;
-    return tensor != nullptr ? *tensor : kEmpty;
-}
-
-struct DiffusionModel {
-    virtual std::string get_desc()                                               = 0;
-    virtual sd::Tensor<float> compute(int n_threads,
-                                      const DiffusionParams& diffusion_params)   = 0;
-    virtual void alloc_params_buffer()                                           = 0;
-    virtual void free_params_buffer()                                            = 0;
-    virtual void free_compute_buffer()                                           = 0;
-    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) = 0;
-    virtual size_t get_params_buffer_size()                                      = 0;
-    virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
-    virtual int64_t get_adm_in_channels()                            = 0;
-    virtual void set_flash_attention_enabled(bool enabled)           = 0;
-    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes)     = 0;
-    virtual void set_circular_axes(bool circular_x, bool circular_y) = 0;
-};
-
-struct UNetModel : public DiffusionModel {
-    UNetModelRunner unet;
-
-    UNetModel(ggml_backend_t backend,
-              ggml_backend_t params_backend,
-              const String2TensorStorage& tensor_storage_map = {},
-              SDVersion version                              = VERSION_SD1)
-        : unet(backend, params_backend, tensor_storage_map, "model.diffusion_model", version) {
-    }
-
-    std::string get_desc() override {
-        return unet.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        unet.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        unet.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        unet.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        unet.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() override {
-        return unet.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        unet.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return unet.unet.adm_in_channels;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        unet.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        unet.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        unet.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_controls;
-        return unet.compute(n_threads,
-                            *diffusion_params.x,
-                            *diffusion_params.timesteps,
-                            tensor_or_empty(diffusion_params.context),
-                            tensor_or_empty(diffusion_params.c_concat),
-                            tensor_or_empty(diffusion_params.y),
-                            diffusion_params.num_video_frames,
-                            diffusion_params.controls ? *diffusion_params.controls : empty_controls,
-                            diffusion_params.control_strength);
-    }
-};
-
-struct MMDiTModel : public DiffusionModel {
-    MMDiTRunner mmdit;
-
-    MMDiTModel(ggml_backend_t backend,
-               ggml_backend_t params_backend,
-               const String2TensorStorage& tensor_storage_map = {})
-        : mmdit(backend, params_backend, tensor_storage_map, "model.diffusion_model") {
-    }
-
-    std::string get_desc() override {
-        return mmdit.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        mmdit.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        mmdit.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        mmdit.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        mmdit.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() override {
-        return mmdit.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        mmdit.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768 + 1280;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        mmdit.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        mmdit.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        mmdit.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<int> empty_skip_layers;
-        return mmdit.compute(n_threads,
-                             *diffusion_params.x,
-                             *diffusion_params.timesteps,
-                             tensor_or_empty(diffusion_params.context),
-                             tensor_or_empty(diffusion_params.y),
-                             diffusion_params.skip_layers ? *diffusion_params.skip_layers : empty_skip_layers);
-    }
-};
-
-struct FluxModel : public DiffusionModel {
-    Flux::FluxRunner flux;
-
-    FluxModel(ggml_backend_t backend,
-              ggml_backend_t params_backend,
-              const String2TensorStorage& tensor_storage_map = {},
-              SDVersion version                              = VERSION_FLUX,
-              bool use_mask                                  = false)
-        : flux(backend, params_backend, tensor_storage_map, "model.diffusion_model", version, use_mask) {
-    }
-
-    std::string get_desc() override {
-        return flux.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        flux.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        flux.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        flux.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        flux.get_param_tensors(tensors, "model.diffusion_model");
-    }
-
-    size_t get_params_buffer_size() override {
-        return flux.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        flux.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        flux.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        flux.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        flux.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_ref_latents;
-        static const std::vector<int> empty_skip_layers;
-        return flux.compute(n_threads,
-                            *diffusion_params.x,
-                            *diffusion_params.timesteps,
-                            tensor_or_empty(diffusion_params.context),
-                            tensor_or_empty(diffusion_params.c_concat),
-                            tensor_or_empty(diffusion_params.y),
-                            tensor_or_empty(diffusion_params.guidance),
-                            diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                            diffusion_params.increase_ref_index,
-                            diffusion_params.skip_layers ? *diffusion_params.skip_layers : empty_skip_layers);
-    }
-};
-
-struct AnimaModel : public DiffusionModel {
-    std::string prefix;
-    Anima::AnimaRunner anima;
-
-    AnimaModel(ggml_backend_t backend,
-               ggml_backend_t params_backend,
-               const String2TensorStorage& tensor_storage_map = {},
-               const std::string prefix                       = "model.diffusion_model")
-        : prefix(prefix), anima(backend, params_backend, tensor_storage_map, prefix) {
-    }
-
-    std::string get_desc() override {
-        return anima.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        anima.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        anima.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        anima.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        anima.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return anima.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        anima.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        anima.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        anima.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        anima.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        return anima.compute(n_threads,
-                             *diffusion_params.x,
-                             *diffusion_params.timesteps,
-                             tensor_or_empty(diffusion_params.context),
-                             tensor_or_empty(diffusion_params.t5_ids),
-                             tensor_or_empty(diffusion_params.t5_weights));
-    }
-};
-
-struct WanModel : public DiffusionModel {
-    std::string prefix;
-    WAN::WanRunner wan;
-
-    WanModel(ggml_backend_t backend,
-             ggml_backend_t params_backend,
-             const String2TensorStorage& tensor_storage_map = {},
-             const std::string prefix                       = "model.diffusion_model",
-             SDVersion version                              = VERSION_WAN2)
-        : prefix(prefix), wan(backend, params_backend, tensor_storage_map, prefix, version) {
-    }
-
-    std::string get_desc() override {
-        return wan.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        wan.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        wan.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        wan.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        wan.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return wan.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        wan.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        wan.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        wan.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        wan.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        return wan.compute(n_threads,
-                           *diffusion_params.x,
-                           *diffusion_params.timesteps,
-                           tensor_or_empty(diffusion_params.context),
-                           tensor_or_empty(diffusion_params.y),
-                           tensor_or_empty(diffusion_params.c_concat),
-                           sd::Tensor<float>(),
-                           tensor_or_empty(diffusion_params.vace_context),
-                           diffusion_params.vace_strength);
-    }
-};
-
-struct QwenImageModel : public DiffusionModel {
-    std::string prefix;
-    Qwen::QwenImageRunner qwen_image;
-
-    QwenImageModel(ggml_backend_t backend,
-                   ggml_backend_t params_backend,
-                   const String2TensorStorage& tensor_storage_map = {},
-                   const std::string prefix                       = "model.diffusion_model",
-                   SDVersion version                              = VERSION_QWEN_IMAGE,
-                   bool zero_cond_t                               = false)
-        : prefix(prefix), qwen_image(backend, params_backend, tensor_storage_map, prefix, version, zero_cond_t) {
-    }
-
-    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, 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();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        qwen_image.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        qwen_image.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        qwen_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        qwen_image.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_ref_latents;
-        return qwen_image.compute(n_threads,
-                                  *diffusion_params.x,
-                                  *diffusion_params.timesteps,
-                                  tensor_or_empty(diffusion_params.context),
-                                  diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                                  true);
-    }
-};
-
-struct HiDreamO1Model : public DiffusionModel {
-    std::string prefix;
-    HiDreamO1::HiDreamO1Runner hidream_o1;
-
-    HiDreamO1Model(ggml_backend_t backend,
-                   ggml_backend_t params_backend,
-                   const String2TensorStorage& tensor_storage_map = {},
-                   const std::string& prefix                      = "model")
-        : prefix(prefix), hidream_o1(backend, params_backend, tensor_storage_map, prefix) {
-    }
-
-    std::string get_desc() override {
-        return hidream_o1.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        hidream_o1.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        hidream_o1.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        hidream_o1.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        hidream_o1.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return hidream_o1.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        hidream_o1.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 0;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        hidream_o1.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        hidream_o1.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        hidream_o1.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        GGML_ASSERT(diffusion_params.input_ids != nullptr);
-        GGML_ASSERT(diffusion_params.input_pos != nullptr);
-        GGML_ASSERT(diffusion_params.token_types != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_images;
-        static const std::vector<std::pair<int, sd::Tensor<float>>> empty_image_embeds;
-        return hidream_o1.compute(n_threads,
-                                  *diffusion_params.x,
-                                  *diffusion_params.timesteps,
-                                  *diffusion_params.input_ids,
-                                  *diffusion_params.input_pos,
-                                  *diffusion_params.token_types,
-                                  tensor_or_empty(diffusion_params.vinput_mask),
-                                  diffusion_params.image_embeds ? *diffusion_params.image_embeds : empty_image_embeds,
-                                  diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_images);
-    }
-};
-
-struct ZImageModel : public DiffusionModel {
-    std::string prefix;
-    ZImage::ZImageRunner z_image;
-
-    ZImageModel(ggml_backend_t backend,
-                ggml_backend_t params_backend,
-                const String2TensorStorage& tensor_storage_map = {},
-                const std::string prefix                       = "model.diffusion_model",
-                SDVersion version                              = VERSION_Z_IMAGE)
-        : prefix(prefix), z_image(backend, params_backend, tensor_storage_map, prefix, version) {
-    }
-
-    std::string get_desc() override {
-        return z_image.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        z_image.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        z_image.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        z_image.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        z_image.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return z_image.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        z_image.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        z_image.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        z_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        z_image.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        static const std::vector<sd::Tensor<float>> empty_ref_latents;
-        return z_image.compute(n_threads,
-                               *diffusion_params.x,
-                               *diffusion_params.timesteps,
-                               tensor_or_empty(diffusion_params.context),
-                               diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
-                               true);
-    }
-};
-
-struct ErnieImageModel : public DiffusionModel {
-    std::string prefix;
-    ErnieImage::ErnieImageRunner ernie_image;
-
-    ErnieImageModel(ggml_backend_t backend,
-                    ggml_backend_t params_backend,
-                    const String2TensorStorage& tensor_storage_map = {},
-                    const std::string prefix                       = "model.diffusion_model")
-        : prefix(prefix), ernie_image(backend, params_backend, tensor_storage_map, prefix) {
-    }
-
-    std::string get_desc() override {
-        return ernie_image.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        ernie_image.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        ernie_image.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        ernie_image.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        ernie_image.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return ernie_image.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        ernie_image.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 768;
-    }
-
-    void set_flash_attention_enabled(bool enabled) {
-        ernie_image.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        ernie_image.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        ernie_image.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        return ernie_image.compute(n_threads,
-                                   *diffusion_params.x,
-                                   *diffusion_params.timesteps,
-                                   tensor_or_empty(diffusion_params.context));
-    }
-};
-
-struct LTXAVModel : public DiffusionModel {
-    std::string prefix;
-    LTXV::LTXAVRunner ltxav;
-
-    LTXAVModel(ggml_backend_t backend,
-               ggml_backend_t params_backend,
-               const String2TensorStorage& tensor_storage_map = {},
-               const std::string prefix                       = "model.diffusion_model")
-        : prefix(prefix), ltxav(backend, params_backend, tensor_storage_map, prefix) {
-    }
-
-    std::string get_desc() override {
-        return ltxav.get_desc();
-    }
-
-    void alloc_params_buffer() override {
-        ltxav.alloc_params_buffer();
-    }
-
-    void free_params_buffer() override {
-        ltxav.free_params_buffer();
-    }
-
-    void free_compute_buffer() override {
-        ltxav.free_compute_buffer();
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        ltxav.get_param_tensors(tensors, prefix);
-    }
-
-    size_t get_params_buffer_size() override {
-        return ltxav.get_params_buffer_size();
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        ltxav.set_weight_adapter(adapter);
-    }
-
-    int64_t get_adm_in_channels() override {
-        return 0;
-    }
-
-    void set_flash_attention_enabled(bool enabled) override {
-        ltxav.set_flash_attention_enabled(enabled);
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        ltxav.set_max_graph_vram_bytes(max_vram_bytes);
-    }
-
-    void set_circular_axes(bool circular_x, bool circular_y) override {
-        ltxav.set_circular_axes(circular_x, circular_y);
-    }
-
-    sd::Tensor<float> compute(int n_threads,
-                              const DiffusionParams& diffusion_params) override {
-        GGML_ASSERT(diffusion_params.x != nullptr);
-        GGML_ASSERT(diffusion_params.timesteps != nullptr);
-        return ltxav.compute(n_threads,
-                             *diffusion_params.x,
-                             *diffusion_params.timesteps,
-                             tensor_or_empty(diffusion_params.context),
-                             tensor_or_empty(diffusion_params.audio_x),
-                             tensor_or_empty(diffusion_params.audio_timesteps),
-                             diffusion_params.audio_length,
-                             diffusion_params.frame_rate,
-                             tensor_or_empty(diffusion_params.video_positions));
-    }
-};
-
-#endif

src/esrgan.hpp

@@ -1,371 +0,0 @@
-#ifndef __ESRGAN_HPP__
-#define __ESRGAN_HPP__
-
-#include "ggml_extend.hpp"
-#include "model.h"
-
-/*
-    ===================================    ESRGAN  ===================================
-    References:
-    https://github.com/xinntao/Real-ESRGAN/blob/master/inference_realesrgan.py
-    https://github.com/XPixelGroup/BasicSR/blob/v1.4.2/basicsr/archs/rrdbnet_arch.py
-
-*/
-
-class ResidualDenseBlock : public GGMLBlock {
-protected:
-    int num_feat;
-    int num_grow_ch;
-
-public:
-    ResidualDenseBlock(int num_feat = 64, int num_grow_ch = 32)
-        : num_feat(num_feat), num_grow_ch(num_grow_ch) {
-        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
-        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
-        blocks["conv3"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
-        blocks["conv4"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
-        blocks["conv5"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 4 * num_grow_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
-    }
-
-    ggml_tensor* lrelu(GGMLRunnerContext* ctx, ggml_tensor* x) {
-        return ggml_leaky_relu(ctx->ggml_ctx, x, 0.2f, true);
-    }
-
-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-        // x: [n, num_feat, h, w]
-        // return: [n, num_feat, h, w]
-
-        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
-        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
-        auto conv3 = std::dynamic_pointer_cast<Conv2d>(blocks["conv3"]);
-        auto conv4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv4"]);
-        auto conv5 = std::dynamic_pointer_cast<Conv2d>(blocks["conv5"]);
-
-        auto x1    = lrelu(ctx, conv1->forward(ctx, x));
-        auto x_cat = ggml_concat(ctx->ggml_ctx, x, x1, 2);
-        auto x2    = lrelu(ctx, conv2->forward(ctx, x_cat));
-        x_cat      = ggml_concat(ctx->ggml_ctx, x_cat, x2, 2);
-        auto x3    = lrelu(ctx, conv3->forward(ctx, x_cat));
-        x_cat      = ggml_concat(ctx->ggml_ctx, x_cat, x3, 2);
-        auto x4    = lrelu(ctx, conv4->forward(ctx, x_cat));
-        x_cat      = ggml_concat(ctx->ggml_ctx, x_cat, x4, 2);
-        auto x5    = conv5->forward(ctx, x_cat);
-
-        x5 = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, x5, 0.2f), x);
-        return x5;
-    }
-};
-
-class RRDB : public GGMLBlock {
-public:
-    RRDB(int num_feat, int num_grow_ch = 32) {
-        blocks["rdb1"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
-        blocks["rdb2"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
-        blocks["rdb3"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
-    }
-
-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-        // x: [n, num_feat, h, w]
-        // return: [n, num_feat, h, w]
-
-        auto rdb1 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb1"]);
-        auto rdb2 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb2"]);
-        auto rdb3 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb3"]);
-
-        auto out = rdb1->forward(ctx, x);
-        out      = rdb2->forward(ctx, out);
-        out      = rdb3->forward(ctx, out);
-
-        out = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, out, 0.2f), x);
-        return out;
-    }
-};
-
-class RRDBNet : public GGMLBlock {
-protected:
-    int scale       = 4;
-    int num_block   = 23;
-    int num_in_ch   = 3;
-    int num_out_ch  = 3;
-    int num_feat    = 64;
-    int num_grow_ch = 32;
-
-public:
-    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}));
-        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; }
-
-    ggml_tensor* lrelu(GGMLRunnerContext* ctx, ggml_tensor* x) {
-        return ggml_leaky_relu(ctx->ggml_ctx, x, 0.2f, true);
-    }
-
-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-        // x: [n, num_in_ch, h, w]
-        // 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_hr    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_hr"]);
-        auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
-
-        auto feat = conv_first->forward(ctx, x);
-        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.prelude", "feat");
-        auto body_feat = feat;
-        for (int i = 0; i < num_block; i++) {
-            std::string name = "body." + std::to_string(i);
-            auto block       = std::dynamic_pointer_cast<RRDB>(blocks[name]);
-
-            body_feat = block->forward(ctx, body_feat);
-            sd::ggml_graph_cut::mark_graph_cut(body_feat, "esrgan.body." + std::to_string(i), "feat");
-        }
-        body_feat = conv_body->forward(ctx, body_feat);
-        feat      = ggml_add(ctx->ggml_ctx, feat, body_feat);
-        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.body.out", "feat");
-        // upsample
-        if (scale >= 2) {
-            auto conv_up1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
-            feat          = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
-            sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up1", "feat");
-            if (scale == 4) {
-                auto conv_up2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
-                feat          = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
-                sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up2", "feat");
-            }
-        }
-        // for all scales
-        auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
-        sd::ggml_graph_cut::mark_graph_cut(out, "esrgan.final", "out");
-        return out;
-    }
-};
-
-struct ESRGAN : public GGMLRunner {
-    std::unique_ptr<RRDBNet> rrdb_net;
-    int scale     = 4;
-    int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
-
-    ESRGAN(ggml_backend_t backend,
-           ggml_backend_t params_backend,
-           int tile_size                                  = 128,
-           const String2TensorStorage& tensor_storage_map = {})
-        : GGMLRunner(backend, params_backend) {
-        this->tile_size = tile_size;
-    }
-
-    std::string get_desc() override {
-        return "esrgan";
-    }
-
-    bool load_from_file(const std::string& file_path, int n_threads) {
-        LOG_INFO("loading esrgan from '%s'", file_path.c_str());
-
-        ModelLoader model_loader;
-        if (!model_loader.init_from_file_and_convert_name(file_path)) {
-            LOG_ERROR("init esrgan model loader from file failed: '%s'", file_path.c_str());
-            return false;
-        }
-
-        // 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;
-        }
-
-        scale = rrdb_net->get_scale();
-        LOG_INFO("esrgan model loaded with scale=%d, num_block=%d", scale, detected_num_block);
-        return success;
-    }
-
-    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor) {
-        if (!rrdb_net)
-            return nullptr;
-        constexpr int kGraphNodes = 1 << 16;  // 65k
-        ggml_cgraph* gf           = new_graph_custom(kGraphNodes);
-        ggml_tensor* x            = make_input(x_tensor);
-
-        auto runner_ctx  = get_context();
-        ggml_tensor* out = rrdb_net->forward(&runner_ctx, x);
-        ggml_build_forward_expand(gf, out);
-        return gf;
-    }
-
-    sd::Tensor<float> compute(const int n_threads,
-                              const sd::Tensor<float>& x) {
-        auto get_graph = [&]() -> ggml_cgraph* { return build_graph(x); };
-        auto result    = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
-        return result;
-    }
-};
-
-#endif  // __ESRGAN_HPP__

src/extensions/generation_extension.h

@@ -0,0 +1,63 @@
+#ifndef __SD_EXTENSIONS_GENERATION_EXTENSION_H__
+#define __SD_EXTENSIONS_GENERATION_EXTENSION_H__
+
+#include <functional>
+#include <map>
+#include <memory>
+#include <set>
+#include <string>
+#include <vector>
+
+#include "conditioning/conditioner.hpp"
+#include "core/ggml_extend_backend.h"
+#include "model_loader.h"
+#include "model_manager.h"
+#include "stable-diffusion.h"
+
+struct GenerationExtensionInitContext {
+    const sd_ctx_params_t* params;
+    SDVersion version;
+    const String2TensorStorage& tensor_storage_map;
+    ModelLoader& model_loader;
+    std::shared_ptr<ModelManager> model_manager;
+    int n_threads;
+    std::function<bool(SDBackendModule)> ensure_backend_pair;
+    std::function<ggml_backend_t(SDBackendModule)> backend_for;
+    std::function<ggml_backend_t(SDBackendModule)> params_backend_for;
+};
+
+struct GenerationExtensionConditionContext {
+    Conditioner* conditioner;
+    ConditionerParams& condition_params;
+    const sd_pm_params_t& pm_params;
+    int n_threads;
+    int total_steps;
+};
+
+struct GenerationExtension {
+    virtual ~GenerationExtension() = default;
+
+    virtual const char* name() const = 0;
+    virtual bool is_enabled() const {
+        return false;
+    }
+    virtual bool init(const GenerationExtensionInitContext&) {
+        return true;
+    }
+    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>&) {}
+    virtual void collect_loras(std::vector<ModelManager::LoraSpec>&) {}
+    virtual void add_ignore_tensors(std::set<std::string>&) const {}
+    virtual void runner_done() {}
+    virtual void reset_runtime_condition() {}
+    virtual bool prepare_condition(GenerationExtensionConditionContext&) {
+        return false;
+    }
+    virtual const SDCondition& before_condition(int step,
+                                                const SDCondition& condition) const {
+        return condition;
+    }
+};
+
+std::shared_ptr<GenerationExtension> create_photomaker_extension();
+
+#endif

src/extensions/photomaker_extension.cpp

@@ -0,0 +1,292 @@
+#include "extensions/generation_extension.h"
+
+#include <algorithm>
+#include <cstring>
+#include <tuple>
+#include <utility>
+
+#include "core/tensor_ggml.hpp"
+#include "core/util.h"
+#include "model/adapter/pmid.hpp"
+
+static std::tuple<std::vector<int>, std::vector<float>, std::vector<bool>>
+tokenize_photomaker_trigger(FrozenCLIPEmbedderWithCustomWords& clip_conditioner,
+                            const std::string& text,
+                            int trigger_token_count,
+                            int32_t image_token) {
+    auto tokens_and_weights           = clip_conditioner.tokenize(text);
+    std::vector<int> source_tokens    = std::move(tokens_and_weights.first);
+    std::vector<float> source_weights = std::move(tokens_and_weights.second);
+
+    if (!source_tokens.empty() && source_tokens.front() == clip_conditioner.tokenizer.BOS_TOKEN_ID) {
+        source_tokens.erase(source_tokens.begin());
+        source_weights.erase(source_weights.begin());
+    }
+    if (!source_tokens.empty() && source_tokens.back() == clip_conditioner.tokenizer.EOS_TOKEN_ID) {
+        source_tokens.pop_back();
+        source_weights.pop_back();
+    }
+
+    std::vector<int> tokens;
+    std::vector<float> weights;
+    int32_t class_idx = -1;
+    for (size_t i = 0; i < source_tokens.size(); i++) {
+        int token = source_tokens[i];
+        if (token == image_token) {
+            if (!tokens.empty()) {
+                class_idx          = static_cast<int32_t>(tokens.size()) - 1;
+                int class_token    = tokens.back();
+                float class_weight = weights.back();
+                for (int j = 1; j < trigger_token_count; j++) {
+                    tokens.push_back(class_token);
+                    weights.push_back(class_weight);
+                }
+            }
+            continue;
+        }
+        tokens.push_back(token);
+        weights.push_back(source_weights[i]);
+    }
+
+    clip_conditioner.tokenizer.pad_tokens(tokens,
+                                          &weights,
+                                          nullptr,
+                                          clip_conditioner.text_model->model.n_token,
+                                          clip_conditioner.text_model->model.n_token,
+                                          true);
+    std::vector<bool> class_token_mask;
+    for (int i = 0; i < tokens.size(); i++) {
+        class_token_mask.push_back(class_idx + 1 <= i && i < class_idx + 1 + trigger_token_count);
+    }
+
+    return std::make_tuple(tokens, weights, class_token_mask);
+}
+
+static std::tuple<SDCondition, std::vector<bool>>
+get_photomaker_condition_with_trigger(FrozenCLIPEmbedderWithCustomWords& clip_conditioner,
+                                      int n_threads,
+                                      const ConditionerParams& conditioner_params,
+                                      const std::string& trigger_word,
+                                      int trigger_token_count) {
+    auto image_tokens = clip_conditioner.convert_token_to_id(trigger_word);
+    GGML_ASSERT(image_tokens.size() == 1);
+    auto tokens_and_weights         = tokenize_photomaker_trigger(clip_conditioner,
+                                                                  conditioner_params.text,
+                                                                  trigger_token_count,
+                                                                  image_tokens[0]);
+    std::vector<int>& tokens        = std::get<0>(tokens_and_weights);
+    std::vector<float>& weights     = std::get<1>(tokens_and_weights);
+    std::vector<bool>& trigger_mask = std::get<2>(tokens_and_weights);
+    auto cond                       = clip_conditioner.get_learned_condition_common(n_threads,
+                                                                                    tokens,
+                                                                                    weights,
+                                                                                    conditioner_params.clip_skip,
+                                                                                    conditioner_params.width,
+                                                                                    conditioner_params.height,
+                                                                                    conditioner_params.zero_out_masked);
+    return std::make_tuple(std::move(cond), trigger_mask);
+}
+
+static std::string remove_photomaker_trigger_from_prompt(FrozenCLIPEmbedderWithCustomWords& clip_conditioner,
+                                                         const std::string& prompt,
+                                                         const std::string& trigger_word) {
+    auto image_tokens = clip_conditioner.convert_token_to_id(trigger_word);
+    GGML_ASSERT(image_tokens.size() == 1);
+    auto tokens_and_weights  = clip_conditioner.tokenize(prompt);
+    std::vector<int>& tokens = tokens_and_weights.first;
+    auto it                  = std::find(tokens.begin(), tokens.end(), image_tokens[0]);
+    GGML_ASSERT(it != tokens.end());
+    tokens.erase(it);
+    return clip_conditioner.decode(tokens);
+}
+
+struct PhotoMakerExtension : public GenerationExtension {
+    std::shared_ptr<PhotoMakerIDEncoder> pmid_model;
+    bool enabled = false;
+    std::string model_path;
+    std::string trigger_word = "img";
+    SDCondition id_condition;
+    int start_merge_step = -1;
+
+    const char* name() const override {
+        return "photomaker";
+    }
+
+    bool is_enabled() const override {
+        return enabled;
+    }
+
+    bool init(const GenerationExtensionInitContext& ctx) override {
+        model_path = SAFE_STR(ctx.params->photo_maker_path);
+        if (model_path.empty()) {
+            return true;
+        }
+
+        if (!ctx.ensure_backend_pair(SDBackendModule::PHOTOMAKER)) {
+            return false;
+        }
+
+        PMVersion pm_version = std::strstr(model_path.c_str(), "v2") != nullptr ? PM_VERSION_2 : PM_VERSION_1;
+        LOG_INFO("loading stacked ID embedding (PHOTOMAKER) model file from '%s'", model_path.c_str());
+        if (!ctx.model_loader.init_from_file_and_convert_name(model_path, "pmid.")) {
+            LOG_WARN("loading stacked ID embedding from '%s' failed", model_path.c_str());
+            return true;
+        }
+
+        pmid_model = std::make_shared<PhotoMakerIDEncoder>(ctx.backend_for(SDBackendModule::PHOTOMAKER),
+                                                           ctx.tensor_storage_map,
+                                                           "pmid",
+                                                           ctx.version,
+                                                           pm_version,
+                                                           20.f,
+                                                           ctx.model_manager);
+        if (pm_version == PM_VERSION_2) {
+            LOG_INFO("using PhotoMaker Version 2");
+        }
+
+        enabled = true;
+        return true;
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        if (!enabled || pmid_model == nullptr) {
+            return;
+        }
+
+        pmid_model->get_param_tensors(tensors, "pmid");
+    }
+
+    void collect_loras(std::vector<ModelManager::LoraSpec>& loras) override {
+        if (!enabled || model_path.empty()) {
+            return;
+        }
+        ModelManager::LoraSpec lora;
+        lora.path                      = model_path;
+        lora.multiplier                = 1.0f;
+        lora.tensor_name_prefix_filter = "lora.model";
+        lora.required                  = true;
+        loras.push_back(std::move(lora));
+    }
+
+    void add_ignore_tensors(std::set<std::string>& ignore_tensors) const override {
+        if (!enabled) {
+            return;
+        }
+        ignore_tensors.insert("pmid.unet.");
+    }
+
+    void runner_done() override {
+        if (pmid_model != nullptr) {
+            pmid_model->runner_done();
+        }
+    }
+
+    void reset_runtime_condition() override {
+        id_condition     = {};
+        start_merge_step = -1;
+    }
+
+    bool prepare_condition(GenerationExtensionConditionContext& ctx) override {
+        reset_runtime_condition();
+        if (!enabled || pmid_model == nullptr) {
+            return false;
+        }
+
+        bool pmv2 = pmid_model->get_version() == PM_VERSION_2;
+        if (ctx.pm_params.id_images_count <= 0 || ctx.pm_params.id_images == nullptr) {
+            LOG_WARN("Provided PhotoMaker model file, but NO input ID images");
+            LOG_WARN("Turn off PhotoMaker for this request");
+            return false;
+        }
+        auto* clip_conditioner = dynamic_cast<FrozenCLIPEmbedderWithCustomWords*>(ctx.conditioner);
+        if (clip_conditioner == nullptr) {
+            LOG_WARN("PhotoMaker requires FrozenCLIPEmbedderWithCustomWords conditioner");
+            LOG_WARN("Turn off PhotoMaker for this request");
+            return false;
+        }
+
+        int clip_image_size        = 224;
+        pmid_model->style_strength = ctx.pm_params.style_strength;
+        sd::Tensor<float> id_image_tensor;
+        for (int i = 0; i < ctx.pm_params.id_images_count; i++) {
+            auto id_image           = sd_image_to_tensor(ctx.pm_params.id_images[i]);
+            auto processed_id_image = clip_preprocess(id_image, clip_image_size, clip_image_size);
+            if (id_image_tensor.empty()) {
+                id_image_tensor = processed_id_image;
+            } else {
+                id_image_tensor = sd::ops::concat(id_image_tensor, processed_id_image, 3);
+            }
+        }
+
+        int64_t t0                        = ggml_time_ms();
+        int trigger_token_count           = pmv2 ? 2 * ctx.pm_params.id_images_count : ctx.pm_params.id_images_count;
+        auto cond_tup                     = get_photomaker_condition_with_trigger(*clip_conditioner,
+                                                                                  ctx.n_threads,
+                                                                                  ctx.condition_params,
+                                                                                  trigger_word,
+                                                                                  trigger_token_count);
+        SDCondition prepared_id_condition = std::get<0>(cond_tup);
+        auto class_tokens_mask            = std::get<1>(cond_tup);
+        if (std::find(class_tokens_mask.begin(), class_tokens_mask.end(), true) == class_tokens_mask.end()) {
+            LOG_WARN("PhotoMaker trigger word '%s' was not found in prompt", trigger_word.c_str());
+            LOG_WARN("Turn off PhotoMaker for this request");
+            return false;
+        }
+
+        sd::Tensor<float> id_embeds;
+        if (pmv2 && ctx.pm_params.id_embed_path != nullptr) {
+            try {
+                id_embeds = sd::load_tensor_from_file_as_tensor<float>(ctx.pm_params.id_embed_path);
+            } catch (const std::exception&) {
+                id_embeds = {};
+            }
+        }
+        if (pmv2 && id_embeds.empty()) {
+            LOG_WARN("Provided PhotoMaker images, but NO valid ID embeds file for PM v2");
+            LOG_WARN("Turn off PhotoMaker for this request");
+            return false;
+        }
+        if (pmv2 && ctx.pm_params.id_images_count != id_embeds.shape()[1]) {
+            LOG_WARN("PhotoMaker image count (%d) does NOT match ID embeds (%d). You should run face_detect.py again.",
+                     ctx.pm_params.id_images_count,
+                     static_cast<int>(id_embeds.shape()[1]));
+            LOG_WARN("Turn off PhotoMaker for this request");
+            return false;
+        }
+
+        auto res = pmid_model->compute(ctx.n_threads,
+                                       id_image_tensor,
+                                       prepared_id_condition.c_crossattn,
+                                       id_embeds,
+                                       class_tokens_mask);
+        if (res.empty()) {
+            LOG_ERROR("Photomaker ID Stacking failed");
+            LOG_WARN("Turn off PhotoMaker for this request");
+            return false;
+        }
+
+        prepared_id_condition.c_crossattn = std::move(res);
+        int64_t t1                        = ggml_time_ms();
+        id_condition                      = std::move(prepared_id_condition);
+        start_merge_step                  = int(ctx.pm_params.style_strength / 100.f * ctx.total_steps);
+        ctx.condition_params.text         = remove_photomaker_trigger_from_prompt(*clip_conditioner,
+                                                                                  ctx.condition_params.text,
+                                                                                  trigger_word);
+        LOG_INFO("Photomaker ID Stacking, taking %" PRId64 " ms", t1 - t0);
+        LOG_INFO("PHOTOMAKER: start_merge_step: %d", start_merge_step);
+
+        return true;
+    }
+
+    const SDCondition& before_condition(int step,
+                                        const SDCondition& condition) const override {
+        if (!id_condition.empty() && start_merge_step != -1 && step > start_merge_step) {
+            return id_condition;
+        }
+        return condition;
+    }
+};
+
+std::shared_ptr<GenerationExtension> create_photomaker_extension() {
+    return std::make_shared<PhotoMakerExtension>();
+}

src/gits_noise.inl

@@ -1,349 +0,0 @@
-#ifndef GITS_NOISE_INL
-#define GITS_NOISE_INL
-
-const std::vector<std::vector<float>> GITS_NOISE_0_80 = {
-    { 14.61464119f, 7.49001646f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 6.77309084f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 3.07277966f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 2.05039096f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 2.05039096f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 5.85520077f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.07277966f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.19567990f, 1.98035145f, 0.86115354f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.19567990f, 1.98035145f, 0.86115354f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.88507891f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.07277966f, 1.84880662f, 0.83188516f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_0_85 = {
-    { 14.61464119f, 7.49001646f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 1.84880662f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 6.77309084f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.11996698f, 3.07277966f, 1.24153244f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 2.84484982f, 0.95350921f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.09240818f, 2.84484982f, 0.95350921f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.58536053f, 3.19567990f, 1.84880662f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 5.58536053f, 3.19567990f, 1.84880662f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 8.75849152f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.65472794f, 3.07277966f, 1.84880662f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.60512662f, 2.63833880f, 1.56271636f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.90732002f, 10.31284904f, 9.75859547f, 9.24142551f, 8.75849152f, 8.30717278f, 7.88507891f, 7.49001646f, 6.77309084f, 5.85520077f, 4.65472794f, 3.46139455f, 2.45070267f, 1.56271636f, 0.72133851f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_0_90 = {
-    { 14.61464119f, 6.77309084f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 3.07277966f, 0.95350921f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.54230714f, 0.89115214f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.54230714f, 0.89115214f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 3.07277966f, 1.61558151f, 0.69515091f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.11996698f, 4.86714602f, 3.07277966f, 1.61558151f, 0.69515091f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 2.95596409f, 1.61558151f, 0.69515091f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.24153244f, 0.57119018f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.84484982f, 1.84880662f, 1.08895338f, 0.52423614f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.84484982f, 1.84880662f, 1.08895338f, 0.52423614f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.45427561f, 3.32507086f, 2.45070267f, 1.61558151f, 0.95350921f, 0.45573691f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.45427561f, 3.32507086f, 2.45070267f, 1.61558151f, 0.95350921f, 0.45573691f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 4.86714602f, 3.91689563f, 3.07277966f, 2.27973175f, 1.56271636f, 0.95350921f, 0.45573691f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.96784878f, 12.23089790f, 11.54541874f, 10.31284904f, 9.24142551f, 8.75849152f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.19988537f, 1.51179266f, 0.89115214f, 0.43325692f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_0_95 = {
-    { 14.61464119f, 6.77309084f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 2.84484982f, 0.89115214f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.36326075f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.56271636f, 0.64427125f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.95596409f, 1.56271636f, 0.64427125f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.91321158f, 1.08895338f, 0.50118381f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.19988537f, 1.41535246f, 0.803307f, 0.38853383f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.46139455f, 2.63833880f, 1.84880662f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.46139455f, 2.63833880f, 1.84880662f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 6.14220476f, 4.86714602f, 3.75677586f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 10.90732002f, 8.75849152f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.60512662f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.60512662f, 2.95596409f, 2.19988537f, 1.56271636f, 1.05362725f, 0.64427125f, 0.32104823f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.44769001f, 5.58536053f, 4.65472794f, 3.75677586f, 3.07277966f, 2.45070267f, 1.78698075f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
-    { 14.61464119f, 12.96784878f, 11.54541874f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.36326075f, 1.72759056f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.60512662f, 2.95596409f, 2.36326075f, 1.72759056f, 1.24153244f, 0.83188516f, 0.50118381f, 0.22545385f, 0.02916753f },
-    { 14.61464119f, 13.76078796f, 12.23089790f, 10.90732002f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.45427561f, 3.75677586f, 3.07277966f, 2.45070267f, 1.91321158f, 1.46270394f, 1.05362725f, 0.72133851f, 0.43325692f, 0.19894916f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_00 = {
-    { 14.61464119f, 1.56271636f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 0.95350921f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 2.36326075f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 7.11996698f, 3.07277966f, 1.56271636f, 0.59516323f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.41535246f, 0.57119018f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.86115354f, 0.38853383f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.86115354f, 0.38853383f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 4.86714602f, 3.07277966f, 1.98035145f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.07277966f, 1.98035145f, 1.24153244f, 0.72133851f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.27973175f, 1.51179266f, 0.95350921f, 0.54755926f, 0.25053367f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.12350607f, 1.56271636f, 1.08895338f, 0.72133851f, 0.41087446f, 0.17026083f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.61558151f, 1.162866f, 0.803307f, 0.50118381f, 0.27464288f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 5.85520077f, 4.65472794f, 3.75677586f, 3.07277966f, 2.45070267f, 1.84880662f, 1.36964464f, 1.01931262f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.46139455f, 2.84484982f, 2.19988537f, 1.67050016f, 1.24153244f, 0.92192322f, 0.64427125f, 0.43325692f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 8.75849152f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 9.24142551f, 8.30717278f, 7.49001646f, 6.14220476f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 12.23089790f, 9.24142551f, 8.30717278f, 7.49001646f, 6.77309084f, 5.85520077f, 5.09240818f, 4.26497746f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.12534678f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_05 = {
-    { 14.61464119f, 0.95350921f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 0.89115214f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 2.05039096f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 2.84484982f, 1.28281462f, 0.52423614f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 3.07277966f, 1.61558151f, 0.803307f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.56271636f, 0.803307f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.95350921f, 0.52423614f, 0.22545385f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 1.98035145f, 1.24153244f, 0.74807048f, 0.41087446f, 0.17026083f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.27973175f, 1.51179266f, 0.95350921f, 0.59516323f, 0.34370604f, 0.13792117f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 5.09240818f, 3.46139455f, 2.45070267f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.09240818f, 3.46139455f, 2.45070267f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.36326075f, 1.61558151f, 1.08895338f, 0.72133851f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.72759056f, 1.24153244f, 0.86115354f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.61558151f, 1.162866f, 0.83188516f, 0.59516323f, 0.38853383f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.84484982f, 2.19988537f, 1.67050016f, 1.28281462f, 0.95350921f, 0.72133851f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.36326075f, 1.84880662f, 1.41535246f, 1.08895338f, 0.83188516f, 0.61951244f, 0.45573691f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.57119018f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 8.30717278f, 7.11996698f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.57119018f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 8.30717278f, 7.11996698f, 5.85520077f, 4.65472794f, 3.60512662f, 2.95596409f, 2.45070267f, 1.98035145f, 1.61558151f, 1.32549286f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.41087446f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_10 = {
-    { 14.61464119f, 0.89115214f, 0.02916753f },
-    { 14.61464119f, 2.36326075f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 1.61558151f, 0.57119018f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 2.45070267f, 1.08895338f, 0.45573691f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 2.95596409f, 1.56271636f, 0.803307f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 3.07277966f, 1.61558151f, 0.89115214f, 0.4783645f, 0.19894916f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 3.07277966f, 1.84880662f, 1.08895338f, 0.64427125f, 0.34370604f, 0.13792117f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.84484982f, 1.61558151f, 0.95350921f, 0.54755926f, 0.27464288f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.91321158f, 1.24153244f, 0.803307f, 0.4783645f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.05039096f, 1.41535246f, 0.95350921f, 0.64427125f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.27973175f, 1.61558151f, 1.12534678f, 0.803307f, 0.54755926f, 0.36617002f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.32507086f, 2.45070267f, 1.72759056f, 1.24153244f, 0.89115214f, 0.64427125f, 0.45573691f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 5.09240818f, 3.60512662f, 2.84484982f, 2.05039096f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 5.09240818f, 3.60512662f, 2.84484982f, 2.12350607f, 1.61558151f, 1.24153244f, 0.95350921f, 0.72133851f, 0.54755926f, 0.41087446f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.19567990f, 2.45070267f, 1.91321158f, 1.51179266f, 1.20157266f, 0.95350921f, 0.74807048f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.43325692f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.86115354f, 0.69515091f, 0.54755926f, 0.43325692f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 11.54541874f, 7.49001646f, 5.85520077f, 4.45427561f, 3.46139455f, 2.84484982f, 2.19988537f, 1.72759056f, 1.36964464f, 1.08895338f, 0.89115214f, 0.72133851f, 0.59516323f, 0.4783645f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_15 = {
-    { 14.61464119f, 0.83188516f, 0.02916753f },
-    { 14.61464119f, 1.84880662f, 0.59516323f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 1.56271636f, 0.52423614f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 1.91321158f, 0.83188516f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.45070267f, 1.24153244f, 0.59516323f, 0.25053367f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.51179266f, 0.803307f, 0.41087446f, 0.17026083f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.56271636f, 0.89115214f, 0.50118381f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 3.07277966f, 1.84880662f, 1.12534678f, 0.72133851f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 3.07277966f, 1.91321158f, 1.24153244f, 0.803307f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 2.95596409f, 1.91321158f, 1.24153244f, 0.803307f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.05039096f, 1.36964464f, 0.95350921f, 0.69515091f, 0.4783645f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.43325692f, 0.29807833f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.78698075f, 1.32549286f, 1.01931262f, 0.803307f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.78698075f, 1.32549286f, 1.01931262f, 0.803307f, 0.64427125f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.12534678f, 0.89115214f, 0.72133851f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.86714602f, 3.19567990f, 2.45070267f, 1.84880662f, 1.41535246f, 1.12534678f, 0.89115214f, 0.72133851f, 0.59516323f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_20 = {
-    { 14.61464119f, 0.803307f, 0.02916753f },
-    { 14.61464119f, 1.56271636f, 0.52423614f, 0.02916753f },
-    { 14.61464119f, 2.36326075f, 0.92192322f, 0.36617002f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.24153244f, 0.59516323f, 0.25053367f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.05039096f, 0.95350921f, 0.45573691f, 0.17026083f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.45070267f, 1.24153244f, 0.64427125f, 0.29807833f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.803307f, 0.45573691f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 0.95350921f, 0.59516323f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.67050016f, 1.08895338f, 0.74807048f, 0.50118381f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.24153244f, 0.83188516f, 0.59516323f, 0.41087446f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 3.07277966f, 1.98035145f, 1.36964464f, 0.95350921f, 0.69515091f, 0.50118381f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 3.46139455f, 2.36326075f, 1.56271636f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 6.77309084f, 3.46139455f, 2.45070267f, 1.61558151f, 1.162866f, 0.86115354f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.12350607f, 1.51179266f, 1.08895338f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.20157266f, 0.92192322f, 0.72133851f, 0.57119018f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 7.49001646f, 4.65472794f, 3.07277966f, 2.19988537f, 1.61558151f, 1.24153244f, 0.95350921f, 0.74807048f, 0.59516323f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_25 = {
-    { 14.61464119f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 1.56271636f, 0.50118381f, 0.02916753f },
-    { 14.61464119f, 2.05039096f, 0.803307f, 0.32104823f, 0.02916753f },
-    { 14.61464119f, 2.36326075f, 0.95350921f, 0.43325692f, 0.17026083f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.24153244f, 0.59516323f, 0.27464288f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 3.07277966f, 1.51179266f, 0.803307f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.36326075f, 1.24153244f, 0.72133851f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.83188516f, 0.52423614f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 0.98595673f, 0.64427125f, 0.43325692f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.67050016f, 1.08895338f, 0.74807048f, 0.52423614f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.803307f, 0.59516323f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.24153244f, 0.86115354f, 0.64427125f, 0.4783645f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.95596409f, 1.84880662f, 1.28281462f, 0.92192322f, 0.69515091f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.72133851f, 0.54755926f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.72133851f, 0.57119018f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.95596409f, 1.91321158f, 1.32549286f, 0.95350921f, 0.74807048f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.41535246f, 1.05362725f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.41535246f, 1.05362725f, 0.803307f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 3.07277966f, 2.05039096f, 1.46270394f, 1.08895338f, 0.83188516f, 0.66947293f, 0.54755926f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_30 = {
-    { 14.61464119f, 0.72133851f, 0.02916753f },
-    { 14.61464119f, 1.24153244f, 0.43325692f, 0.02916753f },
-    { 14.61464119f, 1.56271636f, 0.59516323f, 0.22545385f, 0.02916753f },
-    { 14.61464119f, 1.84880662f, 0.803307f, 0.36617002f, 0.13792117f, 0.02916753f },
-    { 14.61464119f, 2.36326075f, 1.01931262f, 0.52423614f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.36964464f, 0.74807048f, 0.41087446f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 3.07277966f, 1.56271636f, 0.89115214f, 0.54755926f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 3.07277966f, 1.61558151f, 0.95350921f, 0.61951244f, 0.41087446f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.45070267f, 1.36964464f, 0.83188516f, 0.54755926f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.45070267f, 1.41535246f, 0.92192322f, 0.64427125f, 0.45573691f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.6383388f, 1.56271636f, 1.01931262f, 0.72133851f, 0.50118381f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.05362725f, 0.74807048f, 0.54755926f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.77538133f, 0.57119018f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.83188516f, 0.64427125f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.72759056f, 1.162866f, 0.83188516f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.78698075f, 1.24153244f, 0.92192322f, 0.72133851f, 0.57119018f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.84484982f, 1.78698075f, 1.24153244f, 0.92192322f, 0.72133851f, 0.57119018f, 0.4783645f, 0.41087446f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_35 = {
-    { 14.61464119f, 0.69515091f, 0.02916753f },
-    { 14.61464119f, 0.95350921f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 1.56271636f, 0.57119018f, 0.19894916f, 0.02916753f },
-    { 14.61464119f, 1.61558151f, 0.69515091f, 0.29807833f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.84880662f, 0.83188516f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.162866f, 0.64427125f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.36964464f, 0.803307f, 0.50118381f, 0.32104823f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.41535246f, 0.83188516f, 0.54755926f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.56271636f, 0.95350921f, 0.64427125f, 0.45573691f, 0.32104823f, 0.22545385f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.56271636f, 0.95350921f, 0.64427125f, 0.45573691f, 0.34370604f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 3.07277966f, 1.61558151f, 1.01931262f, 0.72133851f, 0.52423614f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 3.07277966f, 1.61558151f, 1.01931262f, 0.72133851f, 0.52423614f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 3.07277966f, 1.61558151f, 1.05362725f, 0.74807048f, 0.54755926f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 3.07277966f, 1.72759056f, 1.12534678f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 3.07277966f, 1.72759056f, 1.12534678f, 0.803307f, 0.59516323f, 0.4783645f, 0.38853383f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.45070267f, 1.51179266f, 1.01931262f, 0.74807048f, 0.57119018f, 0.45573691f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 5.85520077f, 2.6383388f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_40 = {
-    { 14.61464119f, 0.59516323f, 0.02916753f },
-    { 14.61464119f, 0.95350921f, 0.34370604f, 0.02916753f },
-    { 14.61464119f, 1.08895338f, 0.43325692f, 0.13792117f, 0.02916753f },
-    { 14.61464119f, 1.56271636f, 0.64427125f, 0.27464288f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.61558151f, 0.803307f, 0.43325692f, 0.22545385f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.05039096f, 0.95350921f, 0.54755926f, 0.34370604f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.24153244f, 0.72133851f, 0.43325692f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.52423614f, 0.36617002f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.54755926f, 0.38853383f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.41535246f, 0.86115354f, 0.59516323f, 0.43325692f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.64427125f, 0.45573691f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.64427125f, 0.4783645f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.56271636f, 0.98595673f, 0.69515091f, 0.52423614f, 0.41087446f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.72133851f, 0.54755926f, 0.43325692f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.61558151f, 1.05362725f, 0.74807048f, 0.57119018f, 0.45573691f, 0.38853383f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.41087446f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.61951244f, 0.50118381f, 0.43325692f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.61558151f, 1.08895338f, 0.803307f, 0.64427125f, 0.52423614f, 0.45573691f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_45 = {
-    { 14.61464119f, 0.59516323f, 0.02916753f },
-    { 14.61464119f, 0.803307f, 0.25053367f, 0.02916753f },
-    { 14.61464119f, 0.95350921f, 0.34370604f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.24153244f, 0.54755926f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.56271636f, 0.72133851f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.61558151f, 0.803307f, 0.45573691f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.91321158f, 0.95350921f, 0.57119018f, 0.36617002f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.19988537f, 1.08895338f, 0.64427125f, 0.41087446f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.34370604f, 0.25053367f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.24153244f, 0.74807048f, 0.50118381f, 0.36617002f, 0.27464288f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.54755926f, 0.41087446f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.28281462f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.28281462f, 0.83188516f, 0.59516323f, 0.45573691f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.28281462f, 0.83188516f, 0.59516323f, 0.45573691f, 0.36617002f, 0.32104823f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.69515091f, 0.52423614f, 0.41087446f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.51179266f, 0.95350921f, 0.69515091f, 0.52423614f, 0.43325692f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.56271636f, 0.98595673f, 0.72133851f, 0.54755926f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.74807048f, 0.57119018f, 0.4783645f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.84484982f, 1.56271636f, 1.01931262f, 0.74807048f, 0.59516323f, 0.50118381f, 0.43325692f, 0.38853383f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<std::vector<float>> GITS_NOISE_1_50 = {
-    { 14.61464119f, 0.54755926f, 0.02916753f },
-    { 14.61464119f, 0.803307f, 0.25053367f, 0.02916753f },
-    { 14.61464119f, 0.86115354f, 0.32104823f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.24153244f, 0.54755926f, 0.25053367f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.56271636f, 0.72133851f, 0.36617002f, 0.19894916f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.61558151f, 0.803307f, 0.45573691f, 0.27464288f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.61558151f, 0.83188516f, 0.52423614f, 0.34370604f, 0.25053367f, 0.17026083f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.84880662f, 0.95350921f, 0.59516323f, 0.38853383f, 0.27464288f, 0.19894916f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.84880662f, 0.95350921f, 0.59516323f, 0.41087446f, 0.29807833f, 0.22545385f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 1.84880662f, 0.95350921f, 0.61951244f, 0.43325692f, 0.32104823f, 0.25053367f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.19988537f, 1.12534678f, 0.72133851f, 0.50118381f, 0.36617002f, 0.27464288f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.19988537f, 1.12534678f, 0.72133851f, 0.50118381f, 0.36617002f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.29807833f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.57119018f, 0.43325692f, 0.34370604f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.59516323f, 0.45573691f, 0.36617002f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.36326075f, 1.24153244f, 0.803307f, 0.59516323f, 0.45573691f, 0.38853383f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.32549286f, 0.86115354f, 0.64427125f, 0.50118381f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.36964464f, 0.92192322f, 0.69515091f, 0.54755926f, 0.45573691f, 0.41087446f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f },
-    { 14.61464119f, 2.45070267f, 1.41535246f, 0.95350921f, 0.72133851f, 0.57119018f, 0.4783645f, 0.43325692f, 0.38853383f, 0.36617002f, 0.34370604f, 0.32104823f, 0.29807833f, 0.27464288f, 0.25053367f, 0.22545385f, 0.19894916f, 0.17026083f, 0.13792117f, 0.09824532f, 0.02916753f }
-};
-
-const std::vector<const std::vector<std::vector<float>>*> GITS_NOISE = {
-    &GITS_NOISE_0_80,
-    &GITS_NOISE_0_85,
-    &GITS_NOISE_0_90,
-    &GITS_NOISE_0_95,
-    &GITS_NOISE_1_00,
-    &GITS_NOISE_1_05,
-    &GITS_NOISE_1_10,
-    &GITS_NOISE_1_15,
-    &GITS_NOISE_1_20,
-    &GITS_NOISE_1_25,
-    &GITS_NOISE_1_30,
-    &GITS_NOISE_1_35,
-    &GITS_NOISE_1_40,
-    &GITS_NOISE_1_45,
-    &GITS_NOISE_1_50
-};
-
-#endif // GITS_NOISE_INL

src/guidance.cpp

@@ -1,89 +0,0 @@
-#include "guidance.h"
-
-#include <utility>
-
-namespace sd::guidance {
-
-    static bool has_tensor(const sd::Tensor<float>* tensor) {
-        return tensor != nullptr && !tensor->empty();
-    }
-
-    ClassifierFreeGuidance::ClassifierFreeGuidance(float guidance_scale,
-                                                   float image_guidance_scale)
-        : guidance_scale_(guidance_scale),
-          image_guidance_scale_(image_guidance_scale) {
-    }
-
-    GuiderOutput ClassifierFreeGuidance::forward(const GuidanceInput& input,
-                                                 GuiderOutput previous) const {
-        (void)previous;
-
-        GuiderOutput output;
-        if (!has_tensor(input.pred_cond)) {
-            return output;
-        }
-
-        const sd::Tensor<float>& pred_cond = *input.pred_cond;
-        output.pred                        = pred_cond;
-        if (has_tensor(input.pred_uncond)) {
-            const sd::Tensor<float>& pred_uncond = *input.pred_uncond;
-            if (has_tensor(input.pred_img_cond)) {
-                const sd::Tensor<float>& pred_img_cond = *input.pred_img_cond;
-                output.pred                            = pred_uncond +
-                              image_guidance_scale_ * (pred_img_cond - pred_uncond) +
-                              guidance_scale_ * (pred_cond - pred_img_cond);
-            } else {
-                output.pred = pred_uncond + guidance_scale_ * (pred_cond - pred_uncond);
-            }
-        } else if (has_tensor(input.pred_img_cond)) {
-            const sd::Tensor<float>& pred_img_cond = *input.pred_img_cond;
-            output.pred                            = pred_img_cond + guidance_scale_ * (pred_cond - pred_img_cond);
-        }
-
-        return output;
-    }
-
-    SkipLayerGuidance::SkipLayerGuidance(std::vector<int> layers,
-                                         float scale,
-                                         float start,
-                                         float stop)
-        : layers_(std::move(layers)),
-          scale_(scale),
-          start_(start),
-          stop_(stop) {
-    }
-
-    bool SkipLayerGuidance::is_enabled_for_step(const GuidanceInput& input) const {
-        if (scale_ == 0.0f || layers_.empty() || input.schedule_size == 0) {
-            return false;
-        }
-
-        int start_step = static_cast<int>(start_ * static_cast<float>(input.schedule_size));
-        int stop_step  = static_cast<int>(stop_ * static_cast<float>(input.schedule_size));
-        return input.step > start_step && input.step < stop_step;
-    }
-
-    const std::vector<int>& SkipLayerGuidance::layers() const {
-        return layers_;
-    }
-
-    GuiderOutput SkipLayerGuidance::forward(const GuidanceInput& input,
-                                            GuiderOutput output) const {
-        if (!is_enabled_for_step(input) || !input.predict_skip_layer) {
-            return output;
-        }
-
-        if (output.pred.empty() || !has_tensor(input.pred_cond)) {
-            return GuiderOutput();
-        }
-
-        output.pred_skip_layer = input.predict_skip_layer();
-        if (output.pred_skip_layer.empty()) {
-            return GuiderOutput();
-        }
-
-        output.pred += (*input.pred_cond - output.pred_skip_layer) * scale_;
-        return output;
-    }
-
-}  // namespace sd::guidance

src/model.h

@@ -1,17 +1,14 @@
 #ifndef __MODEL_H__
 #define __MODEL_H__
 
-#include <functional>
-#include <map>
-#include <memory>
-#include <set>
 #include <string>
+#include <utility>
 #include <vector>
 
+#include "core/ordered_map.hpp"
 #include "ggml-backend.h"
 #include "ggml.h"
 #include "model_io/tensor_storage.h"
-#include "ordered_map.hpp"
 
 enum SDVersion {
     VERSION_SD1,
@@ -47,6 +44,11 @@ enum SDVersion {
     VERSION_Z_IMAGE,
     VERSION_OVIS_IMAGE,
     VERSION_ERNIE_IMAGE,
+    VERSION_LENS,
+    VERSION_LONGCAT,
+    VERSION_PID,
+    VERSION_IDEOGRAM4,
+    VERSION_ESRGAN,
     VERSION_COUNT,
 };
 
@@ -141,6 +143,13 @@ static inline bool sd_version_is_z_image(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_longcat(SDVersion version) {
+    if (version == VERSION_LONGCAT) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_is_ernie_image(SDVersion version) {
     if (version == VERSION_ERNIE_IMAGE) {
         return true;
@@ -148,8 +157,29 @@ static inline bool sd_version_is_ernie_image(SDVersion version) {
     return false;
 }
 
+static inline bool sd_version_is_lens(SDVersion version) {
+    if (version == VERSION_LENS) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_pid(SDVersion version) {
+    if (version == VERSION_PID) {
+        return true;
+    }
+    return false;
+}
+
+static inline bool sd_version_is_ideogram4(SDVersion version) {
+    if (version == VERSION_IDEOGRAM4) {
+        return true;
+    }
+    return false;
+}
+
 static inline bool sd_version_uses_flux2_vae(SDVersion version) {
-    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version)) {
+    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version) || sd_version_is_lens(version) || sd_version_is_ideogram4(version)) {
         return true;
     }
     return false;
@@ -176,7 +206,11 @@ static inline bool sd_version_is_dit(SDVersion version) {
         version == VERSION_HIDREAM_O1 ||
         sd_version_is_anima(version) ||
         sd_version_is_z_image(version) ||
-        sd_version_is_ernie_image(version)) {
+        sd_version_is_ernie_image(version) ||
+        sd_version_is_lens(version) ||
+        sd_version_is_longcat(version) ||
+        sd_version_is_pid(version) ||
+        sd_version_is_ideogram4(version)) {
         return true;
     }
     return false;
@@ -202,73 +236,4 @@ enum PMVersion {
 typedef OrderedMap<std::string, TensorStorage> String2TensorStorage;
 using TensorTypeRules = std::vector<std::pair<std::string, ggml_type>>;
 
-TensorTypeRules parse_tensor_type_rules(const std::string& tensor_type_rules);
-
-class MmapWrapper;
-
-struct ModelFileData {
-    std::string path;
-    std::vector<TensorStorage> tensors;
-    std::shared_ptr<MmapWrapper> mmapped;
-    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
-    bool is_zip;
-};
-
-struct MmapTensorStore {
-    std::shared_ptr<MmapWrapper> mmapped;
-    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
-};
-
-class ModelLoader {
-protected:
-    SDVersion version_ = VERSION_COUNT;
-    std::vector<std::string> file_paths_;
-    std::vector<ModelFileData> file_data;
-    bool model_files_processed = false;
-    String2TensorStorage tensor_storage_map;
-
-    void add_tensor_storage(const TensorStorage& tensor_storage);
-
-    bool init_from_gguf_file(const std::string& file_path, const std::string& prefix = "");
-    bool init_from_safetensors_file(const std::string& file_path, const std::string& prefix = "");
-    bool init_from_torch_zip_file(const std::string& file_path, const std::string& prefix = "");
-    bool init_from_torch_legacy_file(const std::string& file_path, const std::string& prefix = "");
-    bool init_from_diffusers_file(const std::string& file_path, const std::string& prefix = "");
-
-public:
-    bool init_from_file(const std::string& file_path, const std::string& prefix = "");
-    void convert_tensors_name();
-    bool init_from_file_and_convert_name(const std::string& file_path,
-                                         const std::string& prefix = "",
-                                         SDVersion version         = VERSION_COUNT);
-    SDVersion get_sd_version();
-    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();
-    String2TensorStorage& get_tensor_storage_map() { return tensor_storage_map; }
-    void set_wtype_override(ggml_type wtype, std::string tensor_type_rules = "");
-    void process_model_files(bool enable_mmap = false, bool writable_mmap = true);
-    std::vector<MmapTensorStore> mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
-                                              std::set<std::string> ignore_tensors = {},
-                                              bool writable                        = true);
-    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads = 0, bool use_mmap = false);
-    bool load_tensors(std::map<std::string, ggml_tensor*>& tensors,
-                      std::set<std::string> ignore_tensors = {},
-                      int n_threads                        = 0,
-                      bool use_mmap                        = false);
-
-    std::vector<std::string> get_tensor_names() const {
-        std::vector<std::string> names;
-        for (const auto& [name, tensor_storage] : tensor_storage_map) {
-            names.push_back(name);
-        }
-        return names;
-    }
-
-    bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
-    int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
-    ~ModelLoader() = default;
-};
-
 #endif  // __MODEL_H__

src/model/adapter/lora.hpp

@@ -1,8 +1,10 @@
-#ifndef __LORA_HPP__
-#define __LORA_HPP__
+#ifndef __SD_MODEL_ADAPTER_LORA_HPP__
+#define __SD_MODEL_ADAPTER_LORA_HPP__
 
 #include <mutex>
-#include "ggml_extend.hpp"
+#include "core/ggml_extend.hpp"
+#include "model_loader.h"
+#include "model_manager.h"
 
 #define LORA_GRAPH_BASE_SIZE 10240
 
@@ -13,22 +15,24 @@ struct LoraModel : public GGMLRunner {
     std::map<ggml_tensor*, ggml_tensor*> original_tensor_to_final_tensor;
     std::set<std::string> applied_lora_tensors;
     std::string file_path;
-    ModelLoader model_loader;
-    bool load_failed         = false;
-    bool applied             = false;
-    bool tensor_preprocessed = false;
+    std::shared_ptr<ModelManager> model_manager;
+    ggml_backend_t params_backend = nullptr;
+    bool load_failed              = false;
+    bool applied                  = false;
+    bool tensor_preprocessed      = false;
 
     typedef std::function<bool(const std::string&)> filter_t;
 
     LoraModel(const std::string& lora_id,
               ggml_backend_t backend,
-              ggml_backend_t params_backend,
-              const std::string& file_path = "",
-              std::string prefix           = "",
-              SDVersion version            = VERSION_COUNT)
-        : lora_id(lora_id), file_path(file_path), GGMLRunner(backend, params_backend) {
+              ggml_backend_t params_backend_,
+              const std::string& file_path          = "",
+              std::string prefix                    = "",
+              SDVersion version                     = VERSION_COUNT,
+              std::shared_ptr<ModelManager> manager = std::make_shared<ModelManager>())
+        : GGMLRunner(backend, manager), lora_id(lora_id), file_path(file_path), model_manager(std::move(manager)), params_backend(params_backend_) {
         prefix = "lora." + prefix;
-        if (!model_loader.init_from_file_and_convert_name(file_path, prefix, version)) {
+        if (model_manager == nullptr || !model_manager->loader().init_from_file_and_convert_name(file_path, prefix, version)) {
             load_failed = true;
         }
     }
@@ -70,7 +74,11 @@ struct LoraModel : public GGMLRunner {
             return true;
         };
 
-        model_loader.load_tensors(on_new_tensor_cb, n_threads);
+        if (model_manager != nullptr) {
+            model_manager->set_n_threads(n_threads);
+        }
+        ModelLoader& model_loader = model_manager->loader();
+        model_loader.load_tensors(on_new_tensor_cb);
 
         if (tensors_to_create.empty()) {
             return true;
@@ -86,22 +94,64 @@ struct LoraModel : public GGMLRunner {
             lora_tensors[name] = real;
         }
 
-        alloc_params_buffer();
-
-        dry_run = false;
-        model_loader.load_tensors(on_new_tensor_cb, n_threads);
+        std::map<std::string, ggml_tensor*> tensors;
+        for (const auto& pair : lora_tensors) {
+            tensors[pair.first] = pair.second;
+        }
+        if (model_manager == nullptr ||
+            !model_manager->register_param_tensors("LoRA",
+                                                   std::move(tensors),
+                                                   ModelManager::ResidencyMode::ParamBackend,
+                                                   runtime_backend,
+                                                   params_backend) ||
+            !model_manager->validate_registered_tensors()) {
+            LOG_ERROR("lora model manager registration failed");
+            return false;
+        }
+        std::vector<ggml_tensor*> lora_params;
+        lora_params.reserve(lora_tensors.size());
+        for (const auto& pair : lora_tensors) {
+            lora_params.push_back(pair.second);
+        }
+        if (!model_manager->prepare_params(lora_params)) {
+            LOG_ERROR("lora model manager prepare params failed");
+            return false;
+        }
 
         LOG_DEBUG("finished loaded lora");
         return true;
     }
 
-    void preprocess_lora_tensors(const std::map<std::string, ggml_tensor*>& model_tensors) {
+    void release_loaded_tensors() {
+        runner_done();
+        free_compute_buffer();
+        model_manager.reset();
+        free_params_ctx();
+        alloc_params_ctx();
+        model_manager  = std::make_shared<ModelManager>();
+        weight_manager = model_manager;
+        lora_tensors.clear();
+        original_tensor_to_final_tensor.clear();
+        applied_lora_tensors.clear();
+        applied             = false;
+        tensor_preprocessed = false;
+    }
+
+    static std::set<std::string> tensor_names(const std::map<std::string, ggml_tensor*>& model_tensors) {
+        std::set<std::string> names;
+        for (const auto& item : model_tensors) {
+            names.insert(item.first);
+        }
+        return names;
+    }
+
+    void preprocess_lora_tensors(const std::set<std::string>& model_tensor_names) {
         if (tensor_preprocessed) {
             return;
         }
         tensor_preprocessed = true;
         // I really hate these hardcoded processes.
-        if (model_tensors.find("cond_stage_model.1.transformer.text_model.encoder.layers.0.self_attn.in_proj.weight") != model_tensors.end()) {
+        if (model_tensor_names.find("cond_stage_model.1.transformer.text_model.encoder.layers.0.self_attn.in_proj.weight") != model_tensor_names.end()) {
             std::unordered_map<std::string, ggml_tensor*> new_lora_tensors;
             for (auto& [old_name, tensor] : lora_tensors) {
                 std::string new_name = old_name;
@@ -608,7 +658,7 @@ struct LoraModel : public GGMLRunner {
                 if (lokr_w2)
                     applied_lora_tensors.insert(lokr_w2_name);
                 if (lokr_w2_a)
-                    applied_lora_tensors.insert(lokr_w2_name);
+                    applied_lora_tensors.insert(lokr_w2_a_name);
                 if (lokr_w2_b)
                     applied_lora_tensors.insert(lokr_w2_b_name);
                 applied_lora_tensors.insert(alpha_name);
@@ -749,11 +799,13 @@ struct LoraModel : public GGMLRunner {
         return out_diff;
     }
 
-    ggml_cgraph* build_lora_graph(const std::map<std::string, ggml_tensor*>& model_tensors, SDVersion version) {
+    ggml_cgraph* build_lora_graph(const std::map<std::string, ggml_tensor*>& model_tensors,
+                                  const std::set<std::string>& model_tensor_names,
+                                  SDVersion version) {
         size_t lora_graph_size = LORA_GRAPH_BASE_SIZE + lora_tensors.size() * 10;
         ggml_cgraph* gf        = ggml_new_graph_custom(compute_ctx, lora_graph_size, false);
 
-        preprocess_lora_tensors(model_tensors);
+        preprocess_lora_tensors(model_tensor_names);
 
         original_tensor_to_final_tensor.clear();
         applied_lora_tensors.clear();
@@ -769,7 +821,7 @@ struct LoraModel : public GGMLRunner {
             }
 
             ggml_tensor* original_tensor = model_tensor;
-            if (!ggml_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
+            if (!sd_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
                 model_tensor = ggml_dup_tensor(compute_ctx, model_tensor);
                 set_backend_tensor_data(model_tensor, original_tensor->data);
             }
@@ -783,19 +835,23 @@ struct LoraModel : public GGMLRunner {
                 final_tensor = ggml_add_inplace(compute_ctx, model_tensor, diff);
             }
             ggml_build_forward_expand(gf, final_tensor);
-            if (!ggml_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
+            if (!sd_backend_is_cpu(runtime_backend) && ggml_backend_buffer_is_host(original_tensor->buffer)) {
                 original_tensor_to_final_tensor[original_tensor] = final_tensor;
             }
         }
         return gf;
     }
 
-    void apply(std::map<std::string, ggml_tensor*> model_tensors, SDVersion version, int n_threads) {
+    void apply(std::map<std::string, ggml_tensor*> model_tensors,
+               const std::set<std::string>& model_tensor_names,
+               SDVersion version,
+               int n_threads,
+               bool warn_unused = true) {
         auto get_graph = [&]() -> ggml_cgraph* {
-            return build_lora_graph(model_tensors, version);
+            return build_lora_graph(model_tensors, model_tensor_names, version);
         };
-        GGMLRunner::compute<float>(get_graph, n_threads, false, true);
-        stat();
+        GGMLRunner::compute<float>(get_graph, n_threads, false, false, false, true);
+        stat(!warn_unused);
         for (auto item : original_tensor_to_final_tensor) {
             ggml_tensor* original_tensor = item.first;
             ggml_tensor* final_tensor    = item.second;
@@ -806,6 +862,10 @@ struct LoraModel : public GGMLRunner {
         GGMLRunner::free_compute_buffer();
     }
 
+    void apply(std::map<std::string, ggml_tensor*> model_tensors, SDVersion version, int n_threads, bool warn_unused = true) {
+        apply(model_tensors, tensor_names(model_tensors), version, n_threads, warn_unused);
+    }
+
     void stat(bool at_runntime = false) {
         size_t total_lora_tensors_count   = 0;
         size_t applied_lora_tensors_count = 0;
@@ -911,4 +971,4 @@ public:
     }
 };
 
-#endif  // __LORA_HPP__
+#endif  // __SD_MODEL_ADAPTER_LORA_HPP__

src/model/adapter/pmid.hpp

@@ -1,10 +1,12 @@
-#ifndef __PMI_HPP__
-#define __PMI_HPP__
+#ifndef __SD_MODEL_ADAPTER_PMID_HPP__
+#define __SD_MODEL_ADAPTER_PMID_HPP__
 
-#include "ggml_extend.hpp"
+#include "core/ggml_extend.hpp"
 
-#include "clip.hpp"
-#include "lora.hpp"
+#include "model/adapter/lora.hpp"
+#include "model/common/block.hpp"
+#include "model/te/clip.hpp"
+#include "model_loader.h"
 
 struct FuseBlock : public GGMLBlock {
     // network hparams
@@ -411,13 +413,13 @@ public:
 
 public:
     PhotoMakerIDEncoder(ggml_backend_t backend,
-                        ggml_backend_t params_backend,
                         const String2TensorStorage& tensor_storage_map,
                         const std::string prefix,
-                        SDVersion version = VERSION_SDXL,
-                        PMVersion pm_v    = PM_VERSION_1,
-                        float sty         = 20.f)
-        : GGMLRunner(backend, params_backend),
+                        SDVersion version                                   = VERSION_SDXL,
+                        PMVersion pm_v                                      = PM_VERSION_1,
+                        float sty                                           = 20.f,
+                        std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : GGMLRunner(backend, weight_manager),
           version(version),
           pm_version(pm_v),
           style_strength(sty) {
@@ -556,24 +558,25 @@ public:
             return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds);
         };
 
-        return take_or_empty(GGMLRunner::compute<float>(get_graph, n_threads, true));
+        return take_or_empty(GGMLRunner::compute<float>(get_graph, n_threads, true, true, true));
     }
 };
 
 struct PhotoMakerIDEmbed : public GGMLRunner {
     std::map<std::string, ggml_tensor*> tensors;
     std::string file_path;
-    ModelLoader* model_loader;
-    bool load_failed = false;
-    bool applied     = false;
+    std::shared_ptr<ModelManager> model_manager;
+    ggml_backend_t params_backend = nullptr;
+    bool load_failed              = false;
+    bool applied                  = false;
 
     PhotoMakerIDEmbed(ggml_backend_t backend,
-                      ggml_backend_t params_backend,
-                      ModelLoader* ml,
-                      const std::string& file_path = "",
-                      const std::string& prefix    = "")
-        : file_path(file_path), GGMLRunner(backend, params_backend), model_loader(ml) {
-        if (!model_loader->init_from_file_and_convert_name(file_path, prefix)) {
+                      ggml_backend_t params_backend_,
+                      std::shared_ptr<ModelManager> manager = std::make_shared<ModelManager>(),
+                      const std::string& file_path          = "",
+                      const std::string& prefix             = "")
+        : GGMLRunner(backend, manager), file_path(file_path), model_manager(std::move(manager)), params_backend(params_backend_) {
+        if (model_manager == nullptr || !model_manager->loader().init_from_file_and_convert_name(file_path, prefix)) {
             load_failed = true;
         }
     }
@@ -614,11 +617,27 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
             return true;
         };
 
-        model_loader->load_tensors(on_new_tensor_cb, n_threads);
-        alloc_params_buffer();
-
-        dry_run = false;
-        model_loader->load_tensors(on_new_tensor_cb, n_threads);
+        model_manager->set_n_threads(n_threads);
+        ModelLoader& model_loader = model_manager->loader();
+        model_loader.load_tensors(on_new_tensor_cb);
+        if (!model_manager->register_param_tensors("PhotoMaker ID embeds",
+                                                   tensors,
+                                                   ModelManager::ResidencyMode::ParamBackend,
+                                                   runtime_backend,
+                                                   params_backend) ||
+            !model_manager->validate_registered_tensors()) {
+            LOG_ERROR("PhotoMaker ID embeds model manager registration failed");
+            return false;
+        }
+        std::vector<ggml_tensor*> id_embed_params;
+        id_embed_params.reserve(tensors.size());
+        for (const auto& pair : tensors) {
+            id_embed_params.push_back(pair.second);
+        }
+        if (!model_manager->prepare_params(id_embed_params)) {
+            LOG_ERROR("PhotoMaker ID embeds model manager prepare params failed");
+            return false;
+        }
 
         LOG_DEBUG("finished loading PhotoMaker ID Embeds ");
         return true;
@@ -633,4 +652,4 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
     }
 };
 
-#endif  // __PMI_HPP__
+#endif  // __SD_MODEL_ADAPTER_PMID_HPP__

src/model/common/block.hpp

@@ -1,9 +1,9 @@
-#ifndef __COMMON_BLOCK_HPP__
-#define __COMMON_BLOCK_HPP__
+#ifndef __SD_MODEL_COMMON_BLOCK_HPP__
+#define __SD_MODEL_COMMON_BLOCK_HPP__
 
+#include "core/ggml_extend.hpp"
+#include "core/util.h"
 #include "ggml-backend.h"
-#include "ggml_extend.hpp"
-#include "util.h"
 
 class DownSampleBlock : public GGMLBlock {
 protected:
@@ -227,6 +227,37 @@ public:
     }
 };
 
+struct Mlp : public GGMLBlock {
+public:
+    Mlp(int64_t in_features,
+        int64_t hidden_features = -1,
+        int64_t out_features    = -1,
+        bool bias               = true) {
+        // act_layer is always lambda: nn.GELU(approximate="tanh")
+        // norm_layer is always None
+        // use_conv is always False
+        if (hidden_features == -1) {
+            hidden_features = in_features;
+        }
+        if (out_features == -1) {
+            out_features = in_features;
+        }
+        blocks["fc1"] = std::shared_ptr<GGMLBlock>(new Linear(in_features, hidden_features, bias));
+        blocks["fc2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_features, out_features, bias));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        // x: [N, n_token, in_features]
+        auto fc1 = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
+        auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
+
+        x = fc1->forward(ctx, x);
+        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
+        x = fc2->forward(ctx, x);
+        return x;
+    }
+};
+
 class FeedForward : public GGMLBlock {
 public:
     enum class Activation {
@@ -529,11 +560,11 @@ protected:
         params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
     }
 
-    float get_alpha() {
+    ggml_tensor* get_alpha(GGMLRunnerContext* ctx) {
         // image_only_indicator is always tensor([0.]) and since mix_factor.shape is [1,]
         // so learned_with_images is same as learned
-        float alpha = ggml_ext_backend_tensor_get_f32(params["mix_factor"]);
-        return sigmoid(alpha);
+        auto mix_factor = ggml_ext_cast_f32(ctx->ggml_ctx, ctx->backend, params["mix_factor"]);
+        return ggml_sigmoid(ctx->ggml_ctx, mix_factor);
     }
 
 public:
@@ -547,11 +578,12 @@ public:
                          ggml_tensor* x_spatial,
                          ggml_tensor* x_temporal) {
         // image_only_indicator is always tensor([0.])
-        float alpha = get_alpha();
-        auto x      = ggml_add(ctx->ggml_ctx,
-                               ggml_ext_scale(ctx->ggml_ctx, x_spatial, alpha),
-                               ggml_ext_scale(ctx->ggml_ctx, x_temporal, 1.0f - alpha));
-        return x;
+        auto alpha = get_alpha(ctx);
+        return ggml_add(ctx->ggml_ctx,
+                        x_temporal,
+                        ggml_mul(ctx->ggml_ctx,
+                                 ggml_sub(ctx->ggml_ctx, x_spatial, x_temporal),
+                                 alpha));
     }
 };
 
@@ -603,4 +635,4 @@ public:
     }
 };
 
-#endif  // __COMMON_BLOCK_HPP__
+#endif  // __SD_MODEL_COMMON_BLOCK_HPP__

src/model/common/rope.hpp

@@ -1,10 +1,10 @@
-#ifndef __ROPE_HPP__
-#define __ROPE_HPP__
+#ifndef __SD_MODEL_COMMON_ROPE_HPP__
+#define __SD_MODEL_COMMON_ROPE_HPP__
 
 #include <algorithm>
 #include <cmath>
 #include <vector>
-#include "ggml_extend.hpp"
+#include "core/ggml_extend.hpp"
 
 namespace Rope {
     enum class EmbedNDLayout {
@@ -111,6 +111,16 @@ namespace Rope {
         return txt_ids;
     }
 
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_longcat_txt_ids(int bs, int context_len, int axes_dim_num) {
+        auto txt_ids = std::vector<std::vector<float>>(bs * context_len, std::vector<float>(axes_dim_num, 0.0f));
+        for (int i = 0; i < bs * context_len; i++) {
+            float token_index = static_cast<float>(i % context_len);
+            txt_ids[i][1]     = token_index;
+            txt_ids[i][2]     = token_index;
+        }
+        return txt_ids;
+    }
+
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_flux_img_ids(int h,
                                                                        int w,
                                                                        int patch_size,
@@ -122,7 +132,6 @@ namespace Rope {
                                                                        bool scale_rope = false) {
         int h_len = (h + (patch_size / 2)) / patch_size;
         int w_len = (w + (patch_size / 2)) / patch_size;
-
         std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(axes_dim_num, 0.0));
 
         int h_start = h_offset;
@@ -135,7 +144,6 @@ namespace Rope {
 
         std::vector<float> row_ids = linspace<float>(1.f * h_start, 1.f * h_start + h_len - 1, h_len);
         std::vector<float> col_ids = linspace<float>(1.f * w_start, 1.f * w_start + w_len - 1, w_len);
-
         for (int i = 0; i < h_len; ++i) {
             for (int j = 0; j < w_len; ++j) {
                 img_ids[i * w_len + j][0] = 1.f * index;
@@ -241,17 +249,111 @@ namespace Rope {
         return embed_nd(ids, bs, axis_thetas, axes_dim, wrap_dims, layout);
     }
 
+    __STATIC_INLINE__ std::vector<float> embed_interleaved_mrope(const std::vector<std::vector<float>>& ids,
+                                                                 int bs,
+                                                                 float theta,
+                                                                 int head_dim,
+                                                                 const std::vector<int>& mrope_section,
+                                                                 const std::vector<std::vector<int>>& axis_wrap_dims = {}) {
+        GGML_ASSERT(bs > 0);
+        GGML_ASSERT(head_dim % 2 == 0);
+        GGML_ASSERT(mrope_section.size() >= 3);
+
+        std::vector<std::vector<float>> trans_ids = transpose(ids);
+        size_t pos_len                            = ids.size() / bs;
+        int half_dim                              = head_dim / 2;
+
+        std::vector<std::vector<std::vector<float>>> axis_embs;
+        axis_embs.reserve(3);
+        for (int axis = 0; axis < 3; ++axis) {
+            std::vector<int> axis_wrap;
+            if (axis < static_cast<int>(axis_wrap_dims.size())) {
+                axis_wrap = axis_wrap_dims[axis];
+            }
+            axis_embs.push_back(rope(trans_ids[axis], head_dim, theta, axis_wrap));
+        }
+
+        std::vector<std::vector<float>> emb = axis_embs[0];
+        for (int axis = 1; axis < 3; ++axis) {
+            int length = std::min<int>(mrope_section[axis] * 3, half_dim);
+            for (int freq_idx = axis; freq_idx < length; freq_idx += 3) {
+                for (size_t pos_idx = 0; pos_idx < bs * pos_len; ++pos_idx) {
+                    for (int k = 0; k < 4; ++k) {
+                        emb[pos_idx][4 * freq_idx + k] = axis_embs[axis][pos_idx][4 * freq_idx + k];
+                    }
+                }
+            }
+        }
+
+        return flatten(emb);
+    }
+
+    __STATIC_INLINE__ std::vector<float> embed_2d_interleaved(int height,
+                                                              int width,
+                                                              int dim,
+                                                              float theta    = 10000.f,
+                                                              float scale    = 16.f,
+                                                              int ref_grid_h = 0,
+                                                              int ref_grid_w = 0) {
+        assert(dim % 4 == 0);
+        int half_dim      = dim / 2;
+        int dim_axis      = dim / 2;
+        int axis_half_dim = dim_axis / 2;
+
+        float h_ntk = 1.f;
+        float w_ntk = 1.f;
+        if (ref_grid_h > 0 && ref_grid_w > 0 && dim_axis > 2) {
+            float power = static_cast<float>(dim_axis) / static_cast<float>(dim_axis - 2);
+            h_ntk       = std::pow(static_cast<float>(height) / static_cast<float>(ref_grid_h), power);
+            w_ntk       = std::pow(static_cast<float>(width) / static_cast<float>(ref_grid_w), power);
+        }
+
+        std::vector<float> x_pos;
+        std::vector<float> y_pos;
+        x_pos.reserve(static_cast<size_t>(height) * width);
+        y_pos.reserve(static_cast<size_t>(height) * width);
+        for (int iy = 0; iy < height; ++iy) {
+            float y = height == 1 ? 0.f : scale * static_cast<float>(iy) / static_cast<float>(height - 1);
+            for (int ix = 0; ix < width; ++ix) {
+                float x = width == 1 ? 0.f : scale * static_cast<float>(ix) / static_cast<float>(width - 1);
+                x_pos.push_back(x);
+                y_pos.push_back(y);
+            }
+        }
+
+        auto x_emb = rope(x_pos, dim_axis, theta * w_ntk);
+        auto y_emb = rope(y_pos, dim_axis, theta * h_ntk);
+
+        std::vector<float> out(static_cast<size_t>(height) * width * half_dim * 4);
+        for (int pos = 0; pos < height * width; ++pos) {
+            for (int i = 0; i < axis_half_dim; ++i) {
+                int jx        = 2 * i;
+                int jy        = 2 * i + 1;
+                size_t base_x = static_cast<size_t>(pos) * half_dim * 4 + static_cast<size_t>(jx) * 4;
+                size_t base_y = static_cast<size_t>(pos) * half_dim * 4 + static_cast<size_t>(jy) * 4;
+                size_t axis   = static_cast<size_t>(i) * 4;
+                for (int k = 0; k < 4; ++k) {
+                    out[base_x + k] = x_emb[pos][axis + k];
+                    out[base_y + k] = y_emb[pos][axis + k];
+                }
+            }
+        }
+        return out;
+    }
+
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_refs_ids(int patch_size,
                                                                    int bs,
                                                                    int axes_dim_num,
+                                                                   int start_index,
                                                                    const std::vector<ggml_tensor*>& ref_latents,
                                                                    bool increase_ref_index,
                                                                    float ref_index_scale,
-                                                                   bool scale_rope) {
+                                                                   bool scale_rope,
+                                                                   int base_offset = 0) {
         std::vector<std::vector<float>> ids;
         int curr_h_offset = 0;
         int curr_w_offset = 0;
-        int index         = 1;
+        int index         = start_index;
         for (ggml_tensor* ref : ref_latents) {
             int h_offset = 0;
             int w_offset = 0;
@@ -270,8 +372,8 @@ namespace Rope {
                                             bs,
                                             axes_dim_num,
                                             static_cast<int>(index * ref_index_scale),
-                                            h_offset,
-                                            w_offset,
+                                            h_offset + base_offset,
+                                            w_offset + base_offset,
                                             scale_rope);
             ids          = concat_ids(ids, ref_ids, bs);
 
@@ -294,13 +396,17 @@ namespace Rope {
                                                                    std::set<int> txt_arange_dims,
                                                                    const std::vector<ggml_tensor*>& ref_latents,
                                                                    bool increase_ref_index,
-                                                                   float ref_index_scale) {
-        auto txt_ids = gen_flux_txt_ids(bs, context_len, axes_dim_num, txt_arange_dims);
-        auto img_ids = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num);
+                                                                   float ref_index_scale,
+                                                                   bool is_longcat) {
+        int x_index = is_longcat ? 1 : 0;
+
+        auto txt_ids = is_longcat ? gen_longcat_txt_ids(bs, context_len, axes_dim_num) : gen_flux_txt_ids(bs, context_len, axes_dim_num, txt_arange_dims);
+        int offset   = is_longcat ? context_len : 0;
+        auto img_ids = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, x_index, offset, offset);
 
         auto ids = concat_ids(txt_ids, img_ids, bs);
         if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, ref_latents, increase_ref_index, ref_index_scale, false);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, x_index + 1, ref_latents, increase_ref_index, ref_index_scale, false, offset);
             ids           = concat_ids(ids, refs_ids, bs);
         }
         return ids;
@@ -319,7 +425,8 @@ namespace Rope {
                                                      int theta,
                                                      bool circular_h,
                                                      bool circular_w,
-                                                     const std::vector<int>& axes_dim) {
+                                                     const std::vector<int>& axes_dim,
+                                                     bool is_longcat) {
         std::vector<std::vector<float>> ids = gen_flux_ids(h,
                                                            w,
                                                            patch_size,
@@ -329,7 +436,8 @@ namespace Rope {
                                                            txt_arange_dims,
                                                            ref_latents,
                                                            increase_ref_index,
-                                                           ref_index_scale);
+                                                           ref_index_scale,
+                                                           is_longcat);
         std::vector<std::vector<int>> wrap_dims;
         if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
             int h_len = (h + (patch_size / 2)) / patch_size;
@@ -394,7 +502,7 @@ namespace Rope {
         auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, 0, 0, 0, true);
         auto ids         = concat_ids(txt_ids_repeated, img_ids, bs);
         if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, ref_latents, increase_ref_index, 1.f, true);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, 1, ref_latents, increase_ref_index, 1.f, true);
             ids           = concat_ids(ids, refs_ids, bs);
         }
         return ids;
@@ -462,6 +570,52 @@ namespace Rope {
         return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
+    __STATIC_INLINE__ std::vector<std::vector<float>> gen_lens_ids(int h,
+                                                                   int w,
+                                                                   int bs,
+                                                                   int context_len,
+                                                                   bool scale_rope = true) {
+        auto img_ids_repeated = gen_flux_img_ids(h, w, 1, bs, 3, 0, 0, 0, scale_rope);
+
+        int txt_id_start = scale_rope ? std::max(h / 2, w / 2) : 0;
+        auto txt_ids     = linspace<float>(1.f * txt_id_start, 1.f * 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]};
+            }
+        }
+
+        return concat_ids(img_ids_repeated, txt_ids_repeated, bs);
+    }
+
+    __STATIC_INLINE__ std::vector<float> gen_lens_pe(int h,
+                                                     int w,
+                                                     int bs,
+                                                     int context_len,
+                                                     int theta,
+                                                     bool circular_h,
+                                                     bool circular_w,
+                                                     const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_lens_ids(h, w, bs, context_len, true);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            size_t pos_len = ids.size() / bs;
+            wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+            const size_t img_tokens = static_cast<size_t>(h) * static_cast<size_t>(w);
+            for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                if (circular_h) {
+                    wrap_dims[1][token_i] = h;
+                }
+                if (circular_w) {
+                    wrap_dims[2][token_i] = w;
+                }
+            }
+        }
+
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
+    }
+
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_ernie_image_ids(int h,
                                                                           int w,
                                                                           int patch_size,
@@ -753,4 +907,4 @@ namespace Rope {
     }
 };  // namespace Rope
 
-#endif  // __ROPE_HPP__
+#endif  // __SD_MODEL_COMMON_ROPE_HPP__

src/model/diffusion/anima.hpp

@@ -1,18 +1,61 @@
-#ifndef __ANIMA_HPP__
-#define __ANIMA_HPP__
+#ifndef __SD_MODEL_DIFFUSION_ANIMA_HPP__
+#define __SD_MODEL_DIFFUSION_ANIMA_HPP__
 
+#include <algorithm>
 #include <cmath>
 #include <memory>
 #include <utility>
 #include <vector>
 
-#include "common_block.hpp"
-#include "flux.hpp"
-#include "rope.hpp"
+#include "model/common/block.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/flux.hpp"
+#include "model/diffusion/model.hpp"
 
 namespace Anima {
     constexpr int ANIMA_GRAPH_SIZE = 65536;
 
+    struct AnimaConfig {
+        int64_t in_channels       = 16;
+        int64_t out_channels      = 16;
+        int64_t hidden_size       = 2048;
+        int64_t text_embed_dim    = 1024;
+        int64_t num_heads         = 16;
+        int64_t head_dim          = 128;
+        int patch_size            = 2;
+        int64_t num_layers        = 28;
+        std::vector<int> axes_dim = {44, 42, 42};
+        int theta                 = 10000;
+
+        static AnimaConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            AnimaConfig config;
+            int64_t detected_layers = 0;
+            std::string layer_tag   = prefix.empty() ? "blocks." : prefix + ".blocks.";
+            for (const auto& [name, _] : tensor_storage_map) {
+                size_t pos = name.find(layer_tag);
+                if (pos == std::string::npos) {
+                    continue;
+                }
+                size_t start = pos + layer_tag.size();
+                size_t end   = name.find('.', start);
+                if (end == std::string::npos) {
+                    continue;
+                }
+                int64_t layer_id = atoll(name.substr(start, end - start).c_str());
+                detected_layers  = std::max(detected_layers, layer_id + 1);
+            }
+            if (detected_layers > 0) {
+                config.num_layers = detected_layers;
+                LOG_DEBUG("anima: num_layers = %" PRId64 ", hidden_size = %" PRId64 ", num_heads = %" PRId64 ", head_dim = %" PRId64,
+                          config.num_layers,
+                          config.hidden_size,
+                          config.num_heads,
+                          config.head_dim);
+            }
+            return config;
+        }
+    };
+
     __STATIC_INLINE__ ggml_tensor* apply_gate(ggml_context* ctx,
                                               ggml_tensor* x,
                                               ggml_tensor* gate) {
@@ -417,31 +460,22 @@ namespace Anima {
 
     struct AnimaNet : public GGMLBlock {
     public:
-        int64_t in_channels       = 16;
-        int64_t out_channels      = 16;
-        int64_t hidden_size       = 2048;
-        int64_t text_embed_dim    = 1024;
-        int64_t num_heads         = 16;
-        int64_t head_dim          = 128;
-        int patch_size            = 2;
-        int64_t num_layers        = 28;
-        std::vector<int> axes_dim = {44, 42, 42};
-        int theta                 = 10000;
+        AnimaConfig config;
 
     public:
         AnimaNet() = default;
-        explicit AnimaNet(int64_t num_layers)
-            : num_layers(num_layers) {
-            blocks["x_embedder"]       = std::make_shared<XEmbedder>((in_channels + 1) * patch_size * patch_size, hidden_size);
-            blocks["t_embedder"]       = std::make_shared<TimestepEmbedder>(hidden_size, hidden_size * 3);
-            blocks["t_embedding_norm"] = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
-            for (int i = 0; i < num_layers; i++) {
-                blocks["blocks." + std::to_string(i)] = std::make_shared<TransformerBlock>(hidden_size,
-                                                                                           text_embed_dim,
-                                                                                           num_heads,
-                                                                                           head_dim);
+        explicit AnimaNet(AnimaConfig config)
+            : config(config) {
+            blocks["x_embedder"]       = std::make_shared<XEmbedder>((config.in_channels + 1) * config.patch_size * config.patch_size, config.hidden_size);
+            blocks["t_embedder"]       = std::make_shared<TimestepEmbedder>(config.hidden_size, config.hidden_size * 3);
+            blocks["t_embedding_norm"] = std::make_shared<RMSNorm>(config.hidden_size, 1e-6f);
+            for (int i = 0; i < config.num_layers; i++) {
+                blocks["blocks." + std::to_string(i)] = std::make_shared<TransformerBlock>(config.hidden_size,
+                                                                                           config.text_embed_dim,
+                                                                                           config.num_heads,
+                                                                                           config.head_dim);
             }
-            blocks["final_layer"] = std::make_shared<FinalLayer>(hidden_size, patch_size, out_channels);
+            blocks["final_layer"] = std::make_shared<FinalLayer>(config.hidden_size, config.patch_size, config.out_channels);
             blocks["llm_adapter"] = std::make_shared<LLMAdapter>(1024, 1024, 1024, 6, 16);
         }
 
@@ -468,11 +502,11 @@ namespace Anima {
             auto padding_mask = ggml_ext_zeros(ctx->ggml_ctx, x->ne[0], x->ne[1], 1, x->ne[3]);
             x                 = ggml_concat(ctx->ggml_ctx, x, padding_mask, 2);  // [N, C + 1, H, W]
 
-            x = DiT::pad_and_patchify(ctx, x, patch_size, patch_size);  // [N, h*w, (C+1)*ph*pw]
+            x = DiT::pad_and_patchify(ctx, x, config.patch_size, config.patch_size);  // [N, h*w, (C+1)*ph*pw]
 
             x = x_embedder->forward(ctx, x);
 
-            auto timestep_proj     = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep, static_cast<int>(hidden_size));
+            auto timestep_proj     = ggml_ext_timestep_embedding(ctx->ggml_ctx, timestep, static_cast<int>(config.hidden_size));
             auto temb              = t_embedder->forward(ctx, timestep_proj);
             auto embedded_timestep = t_embedding_norm->forward(ctx, timestep_proj);
 
@@ -504,7 +538,7 @@ namespace Anima {
             sd::ggml_graph_cut::mark_graph_cut(temb, "anima.prelude", "temb");
             sd::ggml_graph_cut::mark_graph_cut(encoder_hidden_states, "anima.prelude", "context");
 
-            for (int i = 0; i < num_layers; i++) {
+            for (int i = 0; i < config.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<TransformerBlock>(blocks["blocks." + std::to_string(i)]);
                 x          = block->forward(ctx, x, encoder_hidden_states, embedded_timestep, temb, image_pe);
                 sd::ggml_graph_cut::mark_graph_cut(x, "anima.blocks." + std::to_string(i), "x");
@@ -512,46 +546,27 @@ namespace Anima {
 
             x = final_layer->forward(ctx, x, embedded_timestep, temb);  // [N, h*w, ph*pw*C]
 
-            x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, patch_size, patch_size, false);  // [N, C, H, W]
+            x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, config.patch_size, config.patch_size, false);  // [N, C, H, W]
 
             return x;
         }
     };
 
-    struct AnimaRunner : public GGMLRunner {
+    struct AnimaRunner : public DiffusionModelRunner {
     public:
         std::vector<float> image_pe_vec;
         std::vector<float> adapter_q_pe_vec;
         std::vector<float> adapter_k_pe_vec;
+        AnimaConfig config;
         AnimaNet net;
 
         AnimaRunner(ggml_backend_t backend,
-                    ggml_backend_t params_backend,
-                    const String2TensorStorage& tensor_storage_map = {},
-                    const std::string prefix                       = "model.diffusion_model")
-            : GGMLRunner(backend, params_backend) {
-            int64_t num_layers    = 0;
-            std::string layer_tag = prefix + ".net.blocks.";
-            for (const auto& kv : tensor_storage_map) {
-                const std::string& tensor_name = kv.first;
-                size_t pos                     = tensor_name.find(layer_tag);
-                if (pos == std::string::npos) {
-                    continue;
-                }
-                size_t start = pos + layer_tag.size();
-                size_t end   = tensor_name.find('.', start);
-                if (end == std::string::npos) {
-                    continue;
-                }
-                int64_t layer_id = atoll(tensor_name.substr(start, end - start).c_str());
-                num_layers       = std::max(num_layers, layer_id + 1);
-            }
-            if (num_layers <= 0) {
-                num_layers = 28;
-            }
-            LOG_INFO("anima net layers: %" PRId64, num_layers);
-
-            net = AnimaNet(num_layers);
+                    const String2TensorStorage& tensor_storage_map      = {},
+                    const std::string prefix                            = "model.diffusion_model",
+                    std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(AnimaConfig::detect_from_weights(tensor_storage_map, prefix + ".net")) {
+            net = AnimaNet(config);
             net.init(params_ctx, tensor_storage_map, prefix + ".net");
         }
 
@@ -559,7 +574,7 @@ namespace Anima {
             return "anima";
         }
 
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
             net.get_param_tensors(tensors, prefix + ".net");
         }
 
@@ -598,7 +613,8 @@ namespace Anima {
                                           {},
                                           empty_ref_latents,
                                           false,
-                                          1.0f);
+                                          1.0f,
+                                          false);
 
             std::vector<float> axis_thetas = {
                 static_cast<float>(theta) * calc_ntk_factor(t_extrapolation_ratio, axes_dim[0]),
@@ -621,22 +637,22 @@ namespace Anima {
             GGML_ASSERT(x->ne[3] == 1);
             ggml_cgraph* gf = new_graph_custom(ANIMA_GRAPH_SIZE);
 
-            int64_t pad_h = (net.patch_size - x->ne[1] % net.patch_size) % net.patch_size;
-            int64_t pad_w = (net.patch_size - x->ne[0] % net.patch_size) % net.patch_size;
+            int64_t pad_h = (config.patch_size - x->ne[1] % config.patch_size) % config.patch_size;
+            int64_t pad_w = (config.patch_size - x->ne[0] % config.patch_size) % config.patch_size;
             int64_t h_pad = x->ne[1] + pad_h;
             int64_t w_pad = x->ne[0] + pad_w;
 
             image_pe_vec          = gen_anima_image_pe_vec(1,
                                                            static_cast<int>(h_pad),
                                                            static_cast<int>(w_pad),
-                                                           static_cast<int>(net.patch_size),
-                                                           net.theta,
-                                                           net.axes_dim,
+                                                           static_cast<int>(config.patch_size),
+                                                           config.theta,
+                                                           config.axes_dim,
                                                            4.0f,
                                                            4.0f,
                                                            1.0f);
-            int64_t image_pos_len = static_cast<int64_t>(image_pe_vec.size()) / (2 * 2 * (net.head_dim / 2));
-            auto image_pe         = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, net.head_dim / 2, image_pos_len);
+            int64_t image_pos_len = static_cast<int64_t>(image_pe_vec.size()) / (2 * 2 * (config.head_dim / 2));
+            auto image_pe         = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2, image_pos_len);
             set_backend_tensor_data(image_pe, image_pe_vec.data());
 
             ggml_tensor* adapter_q_pe = nullptr;
@@ -681,9 +697,22 @@ namespace Anima {
             auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(x, timesteps, context, t5_ids, t5_weights);
             };
-            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            const auto* extra = diffusion_extra_as<AnimaDiffusionExtra>(diffusion_params);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           tensor_or_empty(extra->t5_ids),
+                           tensor_or_empty(extra->t5_weights));
         }
     };
 }  // namespace Anima
 
-#endif  // __ANIMA_HPP__
+#endif  // __SD_MODEL_DIFFUSION_ANIMA_HPP__

src/model/diffusion/control.hpp

@@ -1,8 +1,9 @@
-#ifndef __CONTROL_HPP__
-#define __CONTROL_HPP__
+#ifndef __SD_MODEL_DIFFUSION_CONTROL_HPP__
+#define __SD_MODEL_DIFFUSION_CONTROL_HPP__
 
-#include "common_block.hpp"
-#include "model.h"
+#include "model/common/block.hpp"
+#include "model_loader.h"
+#include "model_manager.h"
 
 #define CONTROL_NET_GRAPH_SIZE 1536
 
@@ -309,6 +310,7 @@ public:
 struct ControlNet : public GGMLRunner {
     SDVersion version = VERSION_SD1;
     ControlNetBlock control_net;
+    std::string weight_prefix;
 
     ggml_backend_buffer_t control_buffer = nullptr;
     ggml_context* control_ctx            = nullptr;
@@ -317,13 +319,17 @@ struct ControlNet : public GGMLRunner {
     std::vector<sd::Tensor<float>> controls;
     sd::Tensor<float> guided_hint;
     bool guided_hint_cached = false;
+    std::shared_ptr<ModelManager> owned_model_manager;
+    ggml_backend_t params_backend = nullptr;
 
     ControlNet(ggml_backend_t backend,
-               ggml_backend_t params_backend,
-               const String2TensorStorage& tensor_storage_map = {},
-               SDVersion version                              = VERSION_SD1)
-        : GGMLRunner(backend, params_backend), control_net(version) {
-        control_net.init(params_ctx, tensor_storage_map, "");
+               ggml_backend_t params_backend_,
+               const String2TensorStorage& tensor_storage_map      = {},
+               SDVersion version                                   = VERSION_SD1,
+               const std::string& prefix                           = "",
+               std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : GGMLRunner(backend, weight_manager), version(version), control_net(version), weight_prefix(prefix), params_backend(params_backend_) {
+        control_net.init(params_ctx, tensor_storage_map, prefix);
     }
 
     ~ControlNet() override {
@@ -374,8 +380,8 @@ struct ControlNet : public GGMLRunner {
         return "control_net";
     }
 
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
-        control_net.get_param_tensors(tensors, prefix);
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {
+        control_net.get_param_tensors(tensors, weight_prefix);
     }
 
     ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
@@ -435,7 +441,7 @@ struct ControlNet : public GGMLRunner {
             return build_graph(x, hint, timesteps, context, y);
         };
 
-        auto compute_result = GGMLRunner::compute<float>(get_graph, n_threads, false);
+        auto compute_result = GGMLRunner::compute<float>(get_graph, n_threads, false, false, false);
         if (!compute_result.has_value()) {
             return std::nullopt;
         }
@@ -457,27 +463,36 @@ struct ControlNet : public GGMLRunner {
 
     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;
         control_net.get_param_tensors(tensors);
-        std::set<std::string> ignore_tensors;
 
-        ModelLoader model_loader;
+        auto manager = std::dynamic_pointer_cast<ModelManager>(weight_manager.lock());
+        if (manager == nullptr) {
+            owned_model_manager = std::make_shared<ModelManager>();
+            weight_manager      = owned_model_manager;
+            manager             = owned_model_manager;
+        }
+
+        ModelLoader& model_loader = manager->loader();
         if (!model_loader.init_from_file_and_convert_name(file_path)) {
             LOG_ERROR("init control net model loader from file failed: '%s'", file_path.c_str());
             return false;
         }
 
-        bool success = model_loader.load_tensors(tensors, ignore_tensors, n_threads);
-
-        if (!success) {
-            LOG_ERROR("load control net tensors from model loader failed");
+        manager->set_n_threads(n_threads);
+        if (!manager->register_param_tensors("ControlNet",
+                                             std::move(tensors),
+                                             ModelManager::ResidencyMode::ParamBackend,
+                                             runtime_backend,
+                                             params_backend) ||
+            !manager->validate_registered_tensors()) {
+            LOG_ERROR("register control net tensors with model manager failed");
             return false;
         }
 
         LOG_INFO("control net model loaded");
-        return success;
+        return true;
     }
 };
 
-#endif  // __CONTROL_HPP__
+#endif  // __SD_MODEL_DIFFUSION_CONTROL_HPP__

src/model/diffusion/dit.hpp

@@ -1,7 +1,7 @@
-#ifndef __COMMON_DIT_HPP__
-#define __COMMON_DIT_HPP__
+#ifndef __SD_MODEL_DIFFUSION_DIT_HPP__
+#define __SD_MODEL_DIFFUSION_DIT_HPP__
 
-#include "ggml_extend.hpp"
+#include "core/ggml_extend.hpp"
 
 namespace DiT {
     inline ggml_tensor* patchify(ggml_context* ctx,
@@ -163,4 +163,4 @@ namespace DiT {
     }
 }  // namespace DiT
 
-#endif  // __COMMON_DIT_HPP__
+#endif  // __SD_MODEL_DIFFUSION_DIT_HPP__

src/model/diffusion/ernie_image.hpp

@@ -1,17 +1,88 @@
-#ifndef __SD_ERNIE_IMAGE_HPP__
-#define __SD_ERNIE_IMAGE_HPP__
+#ifndef __SD_MODEL_DIFFUSION_ERNIE_IMAGE_HPP__
+#define __SD_MODEL_DIFFUSION_ERNIE_IMAGE_HPP__
 
 #include <memory>
 #include <vector>
 
-#include "common_dit.hpp"
-#include "flux.hpp"
-#include "qwen_image.hpp"
-#include "rope.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/flux.hpp"
+#include "model/diffusion/model.hpp"
+#include "model/diffusion/qwen_image.hpp"
 
 namespace ErnieImage {
     constexpr int ERNIE_IMAGE_GRAPH_SIZE = 40960;
 
+    struct ErnieImageConfig {
+        int64_t hidden_size       = 4096;
+        int64_t num_heads         = 32;
+        int64_t num_layers        = 36;
+        int64_t ffn_hidden_size   = 12288;
+        int64_t in_channels       = 128;
+        int64_t out_channels      = 128;
+        int patch_size            = 1;
+        int64_t text_in_dim       = 3072;
+        int theta                 = 256;
+        std::vector<int> axes_dim = {32, 48, 48};
+        int axes_dim_sum          = 128;
+        float eps                 = 1e-6f;
+
+        static ErnieImageConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            ErnieImageConfig config;
+            config.num_layers         = 0;
+            int64_t detected_head_dim = 0;
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "x_embedder.proj.weight") && tensor_storage.n_dims == 4) {
+                    config.patch_size  = static_cast<int>(tensor_storage.ne[0]);
+                    config.in_channels = tensor_storage.ne[2];
+                    config.hidden_size = tensor_storage.ne[3];
+                } else if (ends_with(name, "text_proj.weight") && tensor_storage.n_dims == 2) {
+                    config.text_in_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "layers.0.self_attention.norm_q.weight")) {
+                    detected_head_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "layers.0.mlp.gate_proj.weight") && tensor_storage.n_dims == 2) {
+                    config.ffn_hidden_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "final_linear.weight") && tensor_storage.n_dims == 2) {
+                    int64_t out_dim     = tensor_storage.ne[1];
+                    int64_t patch_area  = config.patch_size * config.patch_size;
+                    config.out_channels = out_dim / patch_area;
+                }
+
+                size_t pos = name.find("layers.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        if (block_index + 1 > config.num_layers) {
+                            config.num_layers = block_index + 1;
+                        }
+                    }
+                }
+            }
+            if (config.num_layers == 0) {
+                config.num_layers = 36;
+            }
+            if (detected_head_dim > 0) {
+                config.num_heads = config.hidden_size / detected_head_dim;
+            }
+            config.axes_dim_sum = 0;
+            for (int axis_dim : config.axes_dim) {
+                config.axes_dim_sum += axis_dim;
+            }
+            LOG_DEBUG("ernie_image: num_layers = %" PRId64 ", hidden_size = %" PRId64 ", num_heads = %" PRId64 ", ffn_hidden_size = %" PRId64 ", in_channels = %" PRId64 ", out_channels = %" PRId64,
+                      config.num_layers,
+                      config.hidden_size,
+                      config.num_heads,
+                      config.ffn_hidden_size,
+                      config.in_channels,
+                      config.out_channels);
+            return config;
+        }
+    };
+
     __STATIC_INLINE__ ggml_tensor* timestep_embedding_sin_cos(ggml_context* ctx,
                                                               ggml_tensor* timesteps,
                                                               int dim,
@@ -207,51 +278,36 @@ namespace ErnieImage {
         }
     };
 
-    struct ErnieImageParams {
-        int64_t hidden_size       = 4096;
-        int64_t num_heads         = 32;
-        int64_t num_layers        = 36;
-        int64_t ffn_hidden_size   = 12288;
-        int64_t in_channels       = 128;
-        int64_t out_channels      = 128;
-        int patch_size            = 1;
-        int64_t text_in_dim       = 3072;
-        int theta                 = 256;
-        std::vector<int> axes_dim = {32, 48, 48};
-        int axes_dim_sum          = 128;
-        float eps                 = 1e-6f;
-    };
-
     class ErnieImageModel : public GGMLBlock {
     public:
-        ErnieImageParams params;
+        ErnieImageConfig config;
 
         ErnieImageModel() = default;
-        ErnieImageModel(ErnieImageParams params)
-            : params(params) {
-            blocks["x_embedder.proj"] = std::make_shared<Conv2d>(params.in_channels,
-                                                                 params.hidden_size,
-                                                                 std::pair<int, int>{params.patch_size, params.patch_size},
-                                                                 std::pair<int, int>{params.patch_size, params.patch_size},
+        ErnieImageModel(ErnieImageConfig config)
+            : config(config) {
+            blocks["x_embedder.proj"] = std::make_shared<Conv2d>(config.in_channels,
+                                                                 config.hidden_size,
+                                                                 std::pair<int, int>{config.patch_size, config.patch_size},
+                                                                 std::pair<int, int>{config.patch_size, config.patch_size},
                                                                  std::pair<int, int>{0, 0},
                                                                  std::pair<int, int>{1, 1},
                                                                  true);
-            if (params.text_in_dim != params.hidden_size) {
-                blocks["text_proj"] = std::make_shared<Linear>(params.text_in_dim, params.hidden_size, false);
+            if (config.text_in_dim != config.hidden_size) {
+                blocks["text_proj"] = std::make_shared<Linear>(config.text_in_dim, config.hidden_size, false);
             }
-            blocks["time_embedding"]     = std::make_shared<Qwen::TimestepEmbedding>(params.hidden_size, params.hidden_size);
-            blocks["adaLN_modulation.1"] = std::make_shared<Linear>(params.hidden_size, 6 * params.hidden_size, true);
+            blocks["time_embedding"]     = std::make_shared<Qwen::TimestepEmbedding>(config.hidden_size, config.hidden_size);
+            blocks["adaLN_modulation.1"] = std::make_shared<Linear>(config.hidden_size, 6 * config.hidden_size, true);
 
-            for (int i = 0; i < params.num_layers; i++) {
-                blocks["layers." + std::to_string(i)] = std::make_shared<ErnieImageSharedAdaLNBlock>(params.hidden_size,
-                                                                                                     params.num_heads,
-                                                                                                     params.ffn_hidden_size,
-                                                                                                     params.eps);
+            for (int i = 0; i < config.num_layers; i++) {
+                blocks["layers." + std::to_string(i)] = std::make_shared<ErnieImageSharedAdaLNBlock>(config.hidden_size,
+                                                                                                     config.num_heads,
+                                                                                                     config.ffn_hidden_size,
+                                                                                                     config.eps);
             }
 
-            blocks["final_norm"]   = std::make_shared<ErnieImageAdaLNContinuous>(params.hidden_size, params.eps);
-            blocks["final_linear"] = std::make_shared<Linear>(params.hidden_size,
-                                                              params.patch_size * params.patch_size * params.out_channels,
+            blocks["final_norm"]   = std::make_shared<ErnieImageAdaLNContinuous>(config.hidden_size, config.eps);
+            blocks["final_linear"] = std::make_shared<Linear>(config.hidden_size,
+                                                              config.patch_size * config.patch_size * config.out_channels,
                                                               true);
         }
 
@@ -264,12 +320,12 @@ namespace ErnieImage {
             // context: [N, text_tokens, 3072]
             // pe: [image_tokens + text_tokens, head_dim/2, 2, 2]
             GGML_ASSERT(context != nullptr);
-            GGML_ASSERT(x->ne[1] % params.patch_size == 0 && x->ne[0] % params.patch_size == 0);
+            GGML_ASSERT(x->ne[1] % config.patch_size == 0 && x->ne[0] % config.patch_size == 0);
 
             int64_t W     = x->ne[0];
             int64_t H     = x->ne[1];
-            int64_t Hp    = H / params.patch_size;
-            int64_t Wp    = W / params.patch_size;
+            int64_t Hp    = H / config.patch_size;
+            int64_t Wp    = W / config.patch_size;
             int64_t n_img = Hp * Wp;
             int64_t N     = x->ne[3];
 
@@ -291,7 +347,7 @@ namespace ErnieImage {
 
             auto hidden_states = ggml_concat(ctx->ggml_ctx, img, txt, 1);  // [N, image_tokens + text_tokens, hidden_size]
 
-            auto sample = timestep_embedding_sin_cos(ctx->ggml_ctx, timestep, static_cast<int>(params.hidden_size));
+            auto sample = timestep_embedding_sin_cos(ctx->ggml_ctx, timestep, static_cast<int>(config.hidden_size));
             auto c      = time_embedding->forward(ctx, sample);  // [N, hidden_size]
 
             auto mod_params = adaLN_mod->forward(ctx, ggml_silu(ctx->ggml_ctx, c));  // [N, 6 * hidden_size]
@@ -304,7 +360,7 @@ namespace ErnieImage {
                 temb.push_back(ggml_reshape_3d(ctx->ggml_ctx, chunk, chunk->ne[0], 1, chunk->ne[1]));  // [N, 1, hidden_size]
             }
 
-            for (int i = 0; i < params.num_layers; i++) {
+            for (int i = 0; i < config.num_layers; i++) {
                 auto layer    = std::dynamic_pointer_cast<ErnieImageSharedAdaLNBlock>(blocks["layers." + std::to_string(i)]);
                 hidden_states = layer->forward(ctx, hidden_states, pe, temb);
                 sd::ggml_graph_cut::mark_graph_cut(hidden_states, "ernie_image.layers." + std::to_string(i), "hidden_states");
@@ -318,74 +374,25 @@ namespace ErnieImage {
                                        patches,
                                        Hp,
                                        Wp,
-                                       params.patch_size,
-                                       params.patch_size,
+                                       config.patch_size,
+                                       config.patch_size,
                                        false);  // [N, out_channels, H, W]
             return out;
         }
     };
 
-    struct ErnieImageRunner : public GGMLRunner {
-        ErnieImageParams ernie_params;
+    struct ErnieImageRunner : public DiffusionModelRunner {
+        ErnieImageConfig config;
         ErnieImageModel ernie_image;
         std::vector<float> pe_vec;
 
         ErnieImageRunner(ggml_backend_t backend,
-                         ggml_backend_t params_backend,
-                         const String2TensorStorage& tensor_storage_map = {},
-                         const std::string prefix                       = "")
-            : GGMLRunner(backend, params_backend) {
-            ernie_params.num_layers = 0;
-            for (const auto& [name, tensor_storage] : tensor_storage_map) {
-                if (!starts_with(name, prefix)) {
-                    continue;
-                }
-                if (ends_with(name, "x_embedder.proj.weight") && tensor_storage.n_dims == 4) {
-                    ernie_params.patch_size  = static_cast<int>(tensor_storage.ne[0]);
-                    ernie_params.in_channels = tensor_storage.ne[2];
-                    ernie_params.hidden_size = tensor_storage.ne[3];
-                } else if (ends_with(name, "text_proj.weight") && tensor_storage.n_dims == 2) {
-                    ernie_params.text_in_dim = tensor_storage.ne[0];
-                } else if (ends_with(name, "layers.0.self_attention.norm_q.weight")) {
-                    int64_t head_dim       = tensor_storage.ne[0];
-                    ernie_params.num_heads = ernie_params.hidden_size / head_dim;
-                } else if (ends_with(name, "layers.0.mlp.gate_proj.weight") && tensor_storage.n_dims == 2) {
-                    ernie_params.ffn_hidden_size = tensor_storage.ne[1];
-                } else if (ends_with(name, "final_linear.weight") && tensor_storage.n_dims == 2) {
-                    int64_t out_dim           = tensor_storage.ne[1];
-                    ernie_params.out_channels = out_dim / ernie_params.patch_size / ernie_params.patch_size;
-                }
-
-                size_t pos = name.find("layers.");
-                if (pos != std::string::npos) {
-                    std::string layer_name = name.substr(pos);
-                    auto items             = split_string(layer_name, '.');
-                    if (items.size() > 1) {
-                        int block_index = atoi(items[1].c_str());
-                        if (block_index + 1 > ernie_params.num_layers) {
-                            ernie_params.num_layers = block_index + 1;
-                        }
-                    }
-                }
-            }
-            if (ernie_params.num_layers == 0) {
-                ernie_params.num_layers = 36;
-            }
-            ernie_params.axes_dim_sum = 0;
-            for (int axis_dim : ernie_params.axes_dim) {
-                ernie_params.axes_dim_sum += axis_dim;
-            }
-
-            LOG_INFO("ernie_image: layers = %" PRId64 ", hidden_size = %" PRId64 ", heads = %" PRId64
-                     ", ffn_hidden_size = %" PRId64 ", in_channels = %" PRId64 ", out_channels = %" PRId64,
-                     ernie_params.num_layers,
-                     ernie_params.hidden_size,
-                     ernie_params.num_heads,
-                     ernie_params.ffn_hidden_size,
-                     ernie_params.in_channels,
-                     ernie_params.out_channels);
-
-            ernie_image = ErnieImageModel(ernie_params);
+                         const String2TensorStorage& tensor_storage_map      = {},
+                         const std::string prefix                            = "",
+                         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(ErnieImageConfig::detect_from_weights(tensor_storage_map, prefix)) {
+            ernie_image = ErnieImageModel(config);
             ernie_image.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -393,7 +400,7 @@ namespace ErnieImage {
             return "ernie_image";
         }
 
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
             ernie_image.get_param_tensors(tensors, prefix);
         }
 
@@ -409,15 +416,15 @@ namespace ErnieImage {
 
             pe_vec      = Rope::gen_ernie_image_pe(static_cast<int>(x->ne[1]),
                                                    static_cast<int>(x->ne[0]),
-                                                   ernie_params.patch_size,
+                                                   config.patch_size,
                                                    static_cast<int>(x->ne[3]),
                                                    static_cast<int>(context->ne[1]),
-                                                   ernie_params.theta,
+                                                   config.theta,
                                                    circular_y_enabled,
                                                    circular_x_enabled,
-                                                   ernie_params.axes_dim);
-            int pos_len = static_cast<int>(pe_vec.size() / ernie_params.axes_dim_sum / 2);
-            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, ernie_params.axes_dim_sum, 1, pos_len, 2);
+                                                   config.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
+            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, config.axes_dim_sum, 1, pos_len, 2);
             set_backend_tensor_data(pe, pe_vec.data());
 
             auto runner_ctx  = get_context();
@@ -433,9 +440,19 @@ namespace ErnieImage {
             auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(x, timesteps, context);
             };
-            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context));
         }
     };
 }  // namespace ErnieImage
 
-#endif  // __SD_ERNIE_IMAGE_HPP__
+#endif  // __SD_MODEL_DIFFUSION_ERNIE_IMAGE_HPP__

src/model/diffusion/flux.hpp

@@ -1,17 +1,167 @@
-#ifndef __FLUX_HPP__
-#define __FLUX_HPP__
+#ifndef __SD_MODEL_DIFFUSION_FLUX_HPP__
+#define __SD_MODEL_DIFFUSION_FLUX_HPP__
 
 #include <memory>
 #include <vector>
 
-#include "common_dit.hpp"
-#include "model.h"
-#include "rope.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model_loader.h"
 
 #define FLUX_GRAPH_SIZE 10240
 
 namespace Flux {
 
+    struct ChromaRadianceConfig {
+        int64_t nerf_hidden_size = 64;
+        int nerf_mlp_ratio       = 4;
+        int nerf_depth           = 4;
+        int nerf_max_freqs       = 8;
+        bool use_x0              = false;
+        bool fake_patch_size_x2  = false;
+    };
+
+    struct FluxConfig {
+        SDVersion version         = VERSION_FLUX;
+        bool is_chroma            = false;
+        int patch_size            = 2;
+        int64_t in_channels       = 64;
+        int64_t out_channels      = 64;
+        int64_t vec_in_dim        = 768;
+        int64_t context_in_dim    = 4096;
+        int64_t hidden_size       = 3072;
+        float mlp_ratio           = 4.0f;
+        int num_heads             = 24;
+        int depth                 = 19;
+        int depth_single_blocks   = 38;
+        std::vector<int> axes_dim = {16, 56, 56};
+        int axes_dim_sum          = 128;
+        int theta                 = 10000;
+        bool qkv_bias             = true;
+        bool guidance_embed       = true;
+        int64_t in_dim            = 64;
+        bool disable_bias         = false;
+        bool share_modulation     = false;
+        bool semantic_txt_norm    = false;
+        bool use_yak_mlp          = false;
+        bool use_mlp_silu_act     = false;
+        float ref_index_scale     = 1.f;
+        ChromaRadianceConfig chroma_radiance_params;
+
+        static FluxConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                              const std::string& prefix,
+                                              SDVersion version = VERSION_FLUX) {
+            FluxConfig config;
+            config.version             = version;
+            config.guidance_embed      = false;
+            config.depth               = 0;
+            config.depth_single_blocks = 0;
+            if (version == VERSION_FLUX_FILL) {
+                config.in_channels = 384;
+            } else if (version == VERSION_FLUX_CONTROLS) {
+                config.in_channels = 128;
+            } else if (version == VERSION_FLEX_2) {
+                config.in_channels = 196;
+            } else if (version == VERSION_CHROMA_RADIANCE) {
+                config.in_channels = 3;
+                config.patch_size  = 16;
+            } else if (version == VERSION_OVIS_IMAGE) {
+                config.semantic_txt_norm = true;
+                config.use_yak_mlp       = true;
+                config.vec_in_dim        = 0;
+            } else if (sd_version_is_flux2(version)) {
+                config.in_channels      = 128;
+                config.patch_size       = 1;
+                config.out_channels     = 128;
+                config.mlp_ratio        = 3.f;
+                config.theta            = 2000;
+                config.axes_dim         = {32, 32, 32, 32};
+                config.vec_in_dim       = 0;
+                config.qkv_bias         = false;
+                config.disable_bias     = true;
+                config.share_modulation = true;
+                config.ref_index_scale  = 10.f;
+                config.use_mlp_silu_act = true;
+            } else if (sd_version_is_longcat(version)) {
+                config.context_in_dim = 3584;
+                config.vec_in_dim     = 0;
+            }
+
+            int64_t head_dim                   = 0;
+            int64_t actual_radiance_patch_size = -1;
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (name.find("guidance_in.in_layer.weight") != std::string::npos) {
+                    config.guidance_embed = true;
+                }
+                if (name.find("__x0__") != std::string::npos) {
+                    LOG_DEBUG("using x0 prediction");
+                    config.chroma_radiance_params.use_x0 = true;
+                }
+                if (name.find("__32x32__") != std::string::npos) {
+                    LOG_DEBUG("using patch size 32");
+                    config.patch_size = 32;
+                }
+                if (name.find("img_in_patch.weight") != std::string::npos) {
+                    actual_radiance_patch_size = tensor_storage.ne[0];
+                    LOG_DEBUG("actual radiance patch size: %" PRId64, actual_radiance_patch_size);
+                }
+                if (name.find("distilled_guidance_layer.in_proj.weight") != std::string::npos) {
+                    config.is_chroma = true;
+                }
+                size_t db = name.find("double_blocks.");
+                if (db != std::string::npos) {
+                    std::string block_name = name.substr(db);
+                    int block_depth        = atoi(block_name.substr(14, block_name.find(".", 14)).c_str());
+                    if (block_depth + 1 > config.depth) {
+                        config.depth = block_depth + 1;
+                    }
+                }
+                size_t sb = name.find("single_blocks.");
+                if (sb != std::string::npos) {
+                    std::string block_name = name.substr(sb);
+                    int block_depth        = atoi(block_name.substr(14, block_name.find(".", 14)).c_str());
+                    if (block_depth + 1 > config.depth_single_blocks) {
+                        config.depth_single_blocks = block_depth + 1;
+                    }
+                }
+                if (ends_with(name, "txt_in.weight")) {
+                    config.context_in_dim = tensor_storage.ne[0];
+                    config.hidden_size    = tensor_storage.ne[1];
+                }
+                if (ends_with(name, "single_blocks.0.norm.key_norm.scale")) {
+                    head_dim = tensor_storage.ne[0];
+                }
+                if (ends_with(name, "double_blocks.0.txt_attn.norm.key_norm.scale")) {
+                    head_dim = tensor_storage.ne[0];
+                }
+            }
+            if (actual_radiance_patch_size > 0 && actual_radiance_patch_size != config.patch_size) {
+                GGML_ASSERT(config.patch_size == 2 * actual_radiance_patch_size);
+                LOG_DEBUG("using fake x2 patch size");
+                config.chroma_radiance_params.fake_patch_size_x2 = true;
+            }
+            if (head_dim > 0) {
+                config.num_heads = static_cast<int>(config.hidden_size / head_dim);
+            }
+            config.axes_dim_sum = 0;
+            for (int axis_dim : config.axes_dim) {
+                config.axes_dim_sum += axis_dim;
+            }
+            LOG_DEBUG("flux: depth = %d, depth_single_blocks = %d, guidance_embed = %s, context_in_dim = %" PRId64 ", hidden_size = %" PRId64 ", num_heads = %d",
+                      config.depth,
+                      config.depth_single_blocks,
+                      config.guidance_embed ? "true" : "false",
+                      config.context_in_dim,
+                      config.hidden_size,
+                      config.num_heads);
+            return config;
+        }
+    };
+
     struct MLPEmbedder : public UnaryBlock {
     public:
         MLPEmbedder(int64_t in_dim, int64_t hidden_dim, bool bias = true) {
@@ -446,7 +596,6 @@ namespace Flux {
             if (use_yak_mlp || use_mlp_silu_act) {
                 mlp_mult_factor = 2;
             }
-
             blocks["linear1"]  = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size * 3 + mlp_hidden_dim * mlp_mult_factor, mlp_proj_bias));
             blocks["linear2"]  = std::shared_ptr<GGMLBlock>(new Linear(hidden_size + mlp_hidden_dim, hidden_size, mlp_proj_bias));
             blocks["norm"]     = std::shared_ptr<GGMLBlock>(new QKNorm(head_dim));
@@ -723,127 +872,90 @@ namespace Flux {
         }
     };
 
-    struct ChromaRadianceParams {
-        int64_t nerf_hidden_size = 64;
-        int nerf_mlp_ratio       = 4;
-        int nerf_depth           = 4;
-        int nerf_max_freqs       = 8;
-        bool use_x0              = false;
-        bool fake_patch_size_x2  = false;
-    };
-
-    struct FluxParams {
-        SDVersion version         = VERSION_FLUX;
-        bool is_chroma            = false;
-        int patch_size            = 2;
-        int64_t in_channels       = 64;
-        int64_t out_channels      = 64;
-        int64_t vec_in_dim        = 768;
-        int64_t context_in_dim    = 4096;
-        int64_t hidden_size       = 3072;
-        float mlp_ratio           = 4.0f;
-        int num_heads             = 24;
-        int depth                 = 19;
-        int depth_single_blocks   = 38;
-        std::vector<int> axes_dim = {16, 56, 56};
-        int axes_dim_sum          = 128;
-        int theta                 = 10000;
-        bool qkv_bias             = true;
-        bool guidance_embed       = true;
-        int64_t in_dim            = 64;
-        bool disable_bias         = false;
-        bool share_modulation     = false;
-        bool semantic_txt_norm    = false;
-        bool use_yak_mlp          = false;
-        bool use_mlp_silu_act     = false;
-        float ref_index_scale     = 1.f;
-        ChromaRadianceParams chroma_radiance_params;
-    };
-
     struct Flux : public GGMLBlock {
     public:
-        FluxParams params;
+        FluxConfig config;
         Flux() {}
-        Flux(FluxParams params)
-            : params(params) {
-            if (params.version == VERSION_CHROMA_RADIANCE) {
-                std::pair<int, int> kernel_size = {params.patch_size, params.patch_size};
-                if (params.chroma_radiance_params.fake_patch_size_x2) {
-                    kernel_size = {params.patch_size / 2, params.patch_size / 2};
+        Flux(FluxConfig config)
+            : config(config) {
+            if (config.version == VERSION_CHROMA_RADIANCE) {
+                std::pair<int, int> kernel_size = {config.patch_size, config.patch_size};
+                if (config.chroma_radiance_params.fake_patch_size_x2) {
+                    kernel_size = {config.patch_size / 2, config.patch_size / 2};
                 }
                 std::pair<int, int> stride = kernel_size;
 
-                blocks["img_in_patch"] = std::make_shared<Conv2d>(params.in_channels,
-                                                                  params.hidden_size,
+                blocks["img_in_patch"] = std::make_shared<Conv2d>(config.in_channels,
+                                                                  config.hidden_size,
                                                                   kernel_size,
                                                                   stride);
             } else {
-                blocks["img_in"] = std::make_shared<Linear>(params.in_channels, params.hidden_size, !params.disable_bias);
+                blocks["img_in"] = std::make_shared<Linear>(config.in_channels, config.hidden_size, !config.disable_bias);
             }
-            if (params.is_chroma) {
-                blocks["distilled_guidance_layer"] = std::make_shared<ChromaApproximator>(params.in_dim, params.hidden_size);
+            if (config.is_chroma) {
+                blocks["distilled_guidance_layer"] = std::make_shared<ChromaApproximator>(config.in_dim, config.hidden_size);
             } else {
-                blocks["time_in"] = std::make_shared<MLPEmbedder>(256, params.hidden_size, !params.disable_bias);
-                if (params.vec_in_dim > 0) {
-                    blocks["vector_in"] = std::make_shared<MLPEmbedder>(params.vec_in_dim, params.hidden_size, !params.disable_bias);
+                blocks["time_in"] = std::make_shared<MLPEmbedder>(256, config.hidden_size, !config.disable_bias);
+                if (config.vec_in_dim > 0) {
+                    blocks["vector_in"] = std::make_shared<MLPEmbedder>(config.vec_in_dim, config.hidden_size, !config.disable_bias);
                 }
-                if (params.guidance_embed) {
-                    blocks["guidance_in"] = std::make_shared<MLPEmbedder>(256, params.hidden_size, !params.disable_bias);
+                if (config.guidance_embed) {
+                    blocks["guidance_in"] = std::make_shared<MLPEmbedder>(256, config.hidden_size, !config.disable_bias);
                 }
             }
-            if (params.semantic_txt_norm) {
-                blocks["txt_norm"] = std::make_shared<RMSNorm>(params.context_in_dim);
+            if (config.semantic_txt_norm) {
+                blocks["txt_norm"] = std::make_shared<RMSNorm>(config.context_in_dim);
             }
-            blocks["txt_in"] = std::make_shared<Linear>(params.context_in_dim, params.hidden_size, !params.disable_bias);
+            blocks["txt_in"] = std::make_shared<Linear>(config.context_in_dim, config.hidden_size, !config.disable_bias);
 
-            for (int i = 0; i < params.depth; i++) {
-                blocks["double_blocks." + std::to_string(i)] = std::make_shared<DoubleStreamBlock>(params.hidden_size,
-                                                                                                   params.num_heads,
-                                                                                                   params.mlp_ratio,
+            for (int i = 0; i < config.depth; i++) {
+                blocks["double_blocks." + std::to_string(i)] = std::make_shared<DoubleStreamBlock>(config.hidden_size,
+                                                                                                   config.num_heads,
+                                                                                                   config.mlp_ratio,
                                                                                                    i,
-                                                                                                   params.qkv_bias,
-                                                                                                   params.is_chroma,
-                                                                                                   params.share_modulation,
-                                                                                                   !params.disable_bias,
-                                                                                                   params.use_yak_mlp,
-                                                                                                   params.use_mlp_silu_act);
+                                                                                                   config.qkv_bias,
+                                                                                                   config.is_chroma,
+                                                                                                   config.share_modulation,
+                                                                                                   !config.disable_bias,
+                                                                                                   config.use_yak_mlp,
+                                                                                                   config.use_mlp_silu_act);
             }
 
-            for (int i = 0; i < params.depth_single_blocks; i++) {
-                blocks["single_blocks." + std::to_string(i)] = std::make_shared<SingleStreamBlock>(params.hidden_size,
-                                                                                                   params.num_heads,
-                                                                                                   params.mlp_ratio,
+            for (int i = 0; i < config.depth_single_blocks; i++) {
+                blocks["single_blocks." + std::to_string(i)] = std::make_shared<SingleStreamBlock>(config.hidden_size,
+                                                                                                   config.num_heads,
+                                                                                                   config.mlp_ratio,
                                                                                                    i,
                                                                                                    0.f,
-                                                                                                   params.is_chroma,
-                                                                                                   params.share_modulation,
-                                                                                                   !params.disable_bias,
-                                                                                                   params.use_yak_mlp,
-                                                                                                   params.use_mlp_silu_act);
+                                                                                                   config.is_chroma,
+                                                                                                   config.share_modulation,
+                                                                                                   !config.disable_bias,
+                                                                                                   config.use_yak_mlp,
+                                                                                                   config.use_mlp_silu_act);
             }
 
-            if (params.version == VERSION_CHROMA_RADIANCE) {
-                blocks["nerf_image_embedder"] = std::make_shared<NerfEmbedder>(params.in_channels,
-                                                                               params.chroma_radiance_params.nerf_hidden_size,
-                                                                               params.chroma_radiance_params.nerf_max_freqs);
+            if (config.version == VERSION_CHROMA_RADIANCE) {
+                blocks["nerf_image_embedder"] = std::make_shared<NerfEmbedder>(config.in_channels,
+                                                                               config.chroma_radiance_params.nerf_hidden_size,
+                                                                               config.chroma_radiance_params.nerf_max_freqs);
 
-                for (int i = 0; i < params.chroma_radiance_params.nerf_depth; i++) {
-                    blocks["nerf_blocks." + std::to_string(i)] = std::make_shared<NerfGLUBlock>(params.hidden_size,
-                                                                                                params.chroma_radiance_params.nerf_hidden_size,
-                                                                                                params.chroma_radiance_params.nerf_mlp_ratio);
+                for (int i = 0; i < config.chroma_radiance_params.nerf_depth; i++) {
+                    blocks["nerf_blocks." + std::to_string(i)] = std::make_shared<NerfGLUBlock>(config.hidden_size,
+                                                                                                config.chroma_radiance_params.nerf_hidden_size,
+                                                                                                config.chroma_radiance_params.nerf_mlp_ratio);
                 }
 
-                blocks["nerf_final_layer_conv"] = std::make_shared<NerfFinalLayerConv>(params.chroma_radiance_params.nerf_hidden_size,
-                                                                                       params.in_channels);
+                blocks["nerf_final_layer_conv"] = std::make_shared<NerfFinalLayerConv>(config.chroma_radiance_params.nerf_hidden_size,
+                                                                                       config.in_channels);
 
             } else {
-                blocks["final_layer"] = std::make_shared<LastLayer>(params.hidden_size, 1, params.out_channels, params.is_chroma, !params.disable_bias);
+                blocks["final_layer"] = std::make_shared<LastLayer>(config.hidden_size, 1, config.out_channels, config.is_chroma, !config.disable_bias);
             }
 
-            if (params.share_modulation) {
-                blocks["double_stream_modulation_img"] = std::make_shared<Modulation>(params.hidden_size, true, !params.disable_bias);
-                blocks["double_stream_modulation_txt"] = std::make_shared<Modulation>(params.hidden_size, true, !params.disable_bias);
-                blocks["single_stream_modulation"]     = std::make_shared<Modulation>(params.hidden_size, false, !params.disable_bias);
+            if (config.share_modulation) {
+                blocks["double_stream_modulation_img"] = std::make_shared<Modulation>(config.hidden_size, true, !config.disable_bias);
+                blocks["double_stream_modulation_txt"] = std::make_shared<Modulation>(config.hidden_size, true, !config.disable_bias);
+                blocks["single_stream_modulation"]     = std::make_shared<Modulation>(config.hidden_size, false, !config.disable_bias);
             }
         }
 
@@ -866,7 +978,7 @@ namespace Flux {
 
             ggml_tensor* vec;
             ggml_tensor* txt_img_mask = nullptr;
-            if (params.is_chroma) {
+            if (config.is_chroma) {
                 int64_t mod_index_length = 344;
                 auto approx              = std::dynamic_pointer_cast<ChromaApproximator>(blocks["distilled_guidance_layer"]);
                 auto distill_timestep    = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, 16, 10000, 1000.f);
@@ -894,7 +1006,7 @@ namespace Flux {
             } else {
                 auto time_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["time_in"]);
                 vec          = time_in->forward(ctx, ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, 256, 10000, 1000.f));
-                if (params.guidance_embed) {
+                if (config.guidance_embed) {
                     GGML_ASSERT(guidance != nullptr);
                     auto guidance_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["guidance_in"]);
                     // bf16 and fp16 result is different
@@ -902,7 +1014,7 @@ namespace Flux {
                     vec       = ggml_add(ctx->ggml_ctx, vec, guidance_in->forward(ctx, g_in));
                 }
 
-                if (params.vec_in_dim > 0) {
+                if (config.vec_in_dim > 0) {
                     auto vector_in = std::dynamic_pointer_cast<MLPEmbedder>(blocks["vector_in"]);
                     vec            = ggml_add(ctx->ggml_ctx, vec, vector_in->forward(ctx, y));
                 }
@@ -911,7 +1023,7 @@ namespace Flux {
             std::vector<ModulationOut> ds_img_mods;
             std::vector<ModulationOut> ds_txt_mods;
             std::vector<ModulationOut> ss_mods;
-            if (params.share_modulation) {
+            if (config.share_modulation) {
                 auto double_stream_modulation_img = std::dynamic_pointer_cast<Modulation>(blocks["double_stream_modulation_img"]);
                 auto double_stream_modulation_txt = std::dynamic_pointer_cast<Modulation>(blocks["double_stream_modulation_txt"]);
                 auto single_stream_modulation     = std::dynamic_pointer_cast<Modulation>(blocks["single_stream_modulation"]);
@@ -921,7 +1033,7 @@ namespace Flux {
                 ss_mods     = single_stream_modulation->forward(ctx, vec);
             }
 
-            if (params.semantic_txt_norm) {
+            if (config.semantic_txt_norm) {
                 auto semantic_txt_norm = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm"]);
 
                 txt = semantic_txt_norm->forward(ctx, txt);
@@ -932,7 +1044,7 @@ namespace Flux {
             sd::ggml_graph_cut::mark_graph_cut(txt, "flux.prelude", "txt");
             sd::ggml_graph_cut::mark_graph_cut(vec, "flux.prelude", "vec");
 
-            for (int i = 0; i < params.depth; i++) {
+            for (int i = 0; i < config.depth; i++) {
                 if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
                     continue;
                 }
@@ -947,8 +1059,8 @@ namespace Flux {
             }
 
             auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
-            for (int i = 0; i < params.depth_single_blocks; i++) {
-                if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i + params.depth) != skip_layers.end()) {
+            for (int i = 0; i < config.depth_single_blocks; i++) {
+                if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i + config.depth) != skip_layers.end()) {
                     continue;
                 }
                 auto block = std::dynamic_pointer_cast<SingleStreamBlock>(blocks["single_blocks." + std::to_string(i)]);
@@ -999,14 +1111,14 @@ namespace Flux {
             int64_t W      = x->ne[0];
             int64_t H      = x->ne[1];
             int64_t C      = x->ne[2];
-            int patch_size = params.patch_size;
+            int patch_size = config.patch_size;
             int pad_h      = (patch_size - H % patch_size) % patch_size;
             int pad_w      = (patch_size - W % patch_size) % patch_size;
 
-            auto img      = DiT::pad_to_patch_size(ctx, x, params.patch_size, params.patch_size);
+            auto img      = DiT::pad_to_patch_size(ctx, x, config.patch_size, config.patch_size);
             auto orig_img = img;
 
-            if (params.chroma_radiance_params.fake_patch_size_x2) {
+            if (config.chroma_radiance_params.fake_patch_size_x2) {
                 // It's supposed to be using GGML_SCALE_MODE_NEAREST, but this seems more stable
                 // Maybe the implementation of nearest-neighbor interpolation in ggml behaves differently than the one in PyTorch?
                 // img = F.interpolate(img, size=(H//2, W//2), mode="nearest")
@@ -1037,7 +1149,7 @@ namespace Flux {
             auto nerf_hidden = ggml_reshape_2d(ctx->ggml_ctx, out, out->ne[0], out->ne[1] * out->ne[2]);  // [N*num_patches, hidden_size]
             auto img_dct     = nerf_image_embedder->forward(ctx, nerf_pixels, dct);                       // [N*num_patches, patch_size*patch_size, nerf_hidden_size]
 
-            for (int i = 0; i < params.chroma_radiance_params.nerf_depth; i++) {
+            for (int i = 0; i < config.chroma_radiance_params.nerf_depth; i++) {
                 auto block = std::dynamic_pointer_cast<NerfGLUBlock>(blocks["nerf_blocks." + std::to_string(i)]);
 
                 img_dct = block->forward(ctx, img_dct, nerf_hidden);
@@ -1049,7 +1161,7 @@ namespace Flux {
 
             out = nerf_final_layer_conv->forward(ctx, img_dct);  // [N, C, H, W]
 
-            if (params.chroma_radiance_params.use_x0) {
+            if (config.chroma_radiance_params.use_x0) {
                 out = _apply_x0_residual(ctx, out, orig_img, timestep);
             }
 
@@ -1073,14 +1185,14 @@ namespace Flux {
             int64_t W      = x->ne[0];
             int64_t H      = x->ne[1];
             int64_t C      = x->ne[2];
-            int patch_size = params.patch_size;
+            int patch_size = config.patch_size;
             int pad_h      = (patch_size - H % patch_size) % patch_size;
             int pad_w      = (patch_size - W % patch_size) % patch_size;
 
             auto img           = DiT::pad_and_patchify(ctx, x, patch_size, patch_size);
             int64_t img_tokens = img->ne[1];
 
-            if (params.version == VERSION_FLUX_FILL) {
+            if (config.version == VERSION_FLUX_FILL) {
                 GGML_ASSERT(c_concat != nullptr);
                 ggml_tensor* masked = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], C, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], 0);
                 ggml_tensor* mask   = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], 8 * 8, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C);
@@ -1089,7 +1201,7 @@ namespace Flux {
                 mask   = DiT::pad_and_patchify(ctx, mask, patch_size, patch_size);
 
                 img = ggml_concat(ctx->ggml_ctx, img, ggml_concat(ctx->ggml_ctx, masked, mask, 0), 0);
-            } else if (params.version == VERSION_FLEX_2) {
+            } else if (config.version == VERSION_FLEX_2) {
                 GGML_ASSERT(c_concat != nullptr);
                 ggml_tensor* masked  = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], C, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], 0);
                 ggml_tensor* mask    = ggml_view_4d(ctx->ggml_ctx, c_concat, c_concat->ne[0], c_concat->ne[1], 1, 1, c_concat->nb[1], c_concat->nb[2], c_concat->nb[3], c_concat->nb[2] * C);
@@ -1100,7 +1212,7 @@ namespace Flux {
                 control = DiT::pad_and_patchify(ctx, control, patch_size, patch_size);
 
                 img = ggml_concat(ctx->ggml_ctx, img, ggml_concat(ctx->ggml_ctx, ggml_concat(ctx->ggml_ctx, masked, mask, 0), control, 0), 0);
-            } else if (params.version == VERSION_FLUX_CONTROLS) {
+            } else if (config.version == VERSION_FLUX_CONTROLS) {
                 GGML_ASSERT(c_concat != nullptr);
 
                 auto control = DiT::pad_and_patchify(ctx, c_concat, patch_size, patch_size);
@@ -1147,7 +1259,7 @@ namespace Flux {
             // pe: (L, d_head/2, 2, 2)
             // return: (N, C, H, W)
 
-            if (params.version == VERSION_CHROMA_RADIANCE) {
+            if (config.version == VERSION_CHROMA_RADIANCE) {
                 return forward_chroma_radiance(ctx,
                                                x,
                                                timestep,
@@ -1177,9 +1289,9 @@ namespace Flux {
         }
     };
 
-    struct FluxRunner : public GGMLRunner {
+    struct FluxRunner : public DiffusionModelRunner {
     public:
-        FluxParams flux_params;
+        FluxConfig config;
         Flux flux;
         std::vector<float> pe_vec;
         std::vector<float> mod_index_arange_vec;
@@ -1189,116 +1301,20 @@ namespace Flux {
         bool use_mask = false;
 
         FluxRunner(ggml_backend_t backend,
-                   ggml_backend_t params_backend,
-                   const String2TensorStorage& tensor_storage_map = {},
-                   const std::string prefix                       = "",
-                   SDVersion version                              = VERSION_FLUX,
-                   bool use_mask                                  = false)
-            : GGMLRunner(backend, params_backend), version(version), use_mask(use_mask) {
-            flux_params.version             = version;
-            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;
-            } else if (version == VERSION_CHROMA_RADIANCE) {
-                flux_params.in_channels = 3;
-                flux_params.patch_size  = 16;
-            } else if (version == VERSION_OVIS_IMAGE) {
-                flux_params.semantic_txt_norm = true;
-                flux_params.use_yak_mlp       = true;
-                flux_params.vec_in_dim        = 0;
-            } else if (sd_version_is_flux2(version)) {
-                flux_params.in_channels      = 128;
-                flux_params.patch_size       = 1;
-                flux_params.out_channels     = 128;
-                flux_params.mlp_ratio        = 3.f;
-                flux_params.theta            = 2000;
-                flux_params.axes_dim         = {32, 32, 32, 32};
-                flux_params.vec_in_dim       = 0;
-                flux_params.qkv_bias         = false;
-                flux_params.disable_bias     = true;
-                flux_params.share_modulation = true;
-                flux_params.ref_index_scale  = 10.f;
-                flux_params.use_mlp_silu_act = true;
-            }
-            int64_t head_dim                   = 0;
-            int64_t actual_radiance_patch_size = -1;
-            for (auto pair : tensor_storage_map) {
-                std::string tensor_name = pair.first;
-                if (!starts_with(tensor_name, prefix))
-                    continue;
-                if (tensor_name.find("guidance_in.in_layer.weight") != std::string::npos) {
-                    flux_params.guidance_embed = true;
-                }
-                if (tensor_name.find("__x0__") != std::string::npos) {
-                    LOG_DEBUG("using x0 prediction");
-                    flux_params.chroma_radiance_params.use_x0 = true;
-                }
-                if (tensor_name.find("__32x32__") != std::string::npos) {
-                    LOG_DEBUG("using patch size 32");
-                    flux_params.patch_size = 32;
-                }
-                if (tensor_name.find("img_in_patch.weight") != std::string::npos) {
-                    actual_radiance_patch_size = pair.second.ne[0];
-                    LOG_DEBUG("actual radiance patch size: %d", actual_radiance_patch_size);
-                }
-                if (tensor_name.find("distilled_guidance_layer.in_proj.weight") != std::string::npos) {
-                    // Chroma
-                    flux_params.is_chroma = true;
-                }
-                size_t db = tensor_name.find("double_blocks.");
-                if (db != std::string::npos) {
-                    tensor_name     = tensor_name.substr(db);  // remove prefix
-                    int block_depth = atoi(tensor_name.substr(14, tensor_name.find(".", 14)).c_str());
-                    if (block_depth + 1 > flux_params.depth) {
-                        flux_params.depth = block_depth + 1;
-                    }
-                }
-                size_t sb = tensor_name.find("single_blocks.");
-                if (sb != std::string::npos) {
-                    tensor_name     = tensor_name.substr(sb);  // remove prefix
-                    int block_depth = atoi(tensor_name.substr(14, tensor_name.find(".", 14)).c_str());
-                    if (block_depth + 1 > flux_params.depth_single_blocks) {
-                        flux_params.depth_single_blocks = block_depth + 1;
-                    }
-                }
-                if (ends_with(tensor_name, "txt_in.weight")) {
-                    flux_params.context_in_dim = pair.second.ne[0];
-                    flux_params.hidden_size    = pair.second.ne[1];
-                }
-                if (ends_with(tensor_name, "single_blocks.0.norm.key_norm.scale")) {
-                    head_dim = pair.second.ne[0];
-                }
-                if (ends_with(tensor_name, "double_blocks.0.txt_attn.norm.key_norm.scale")) {
-                    head_dim = pair.second.ne[0];
-                }
-            }
-            if (actual_radiance_patch_size > 0 && actual_radiance_patch_size != flux_params.patch_size) {
-                GGML_ASSERT(flux_params.patch_size == 2 * actual_radiance_patch_size);
-                LOG_DEBUG("using fake x2 patch size");
-                flux_params.chroma_radiance_params.fake_patch_size_x2 = true;
-            }
-
-            flux_params.num_heads = static_cast<int>(flux_params.hidden_size / head_dim);
-
-            LOG_INFO("flux: depth = %d, depth_single_blocks = %d, guidance_embed = %s, context_in_dim = %" PRId64
-                     ", hidden_size = %" PRId64 ", num_heads = %d",
-                     flux_params.depth,
-                     flux_params.depth_single_blocks,
-                     flux_params.guidance_embed ? "true" : "false",
-                     flux_params.context_in_dim,
-                     flux_params.hidden_size,
-                     flux_params.num_heads);
-            if (flux_params.is_chroma) {
+                   const String2TensorStorage& tensor_storage_map      = {},
+                   const std::string prefix                            = "",
+                   SDVersion version                                   = VERSION_FLUX,
+                   bool use_mask                                       = false,
+                   std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(FluxConfig::detect_from_weights(tensor_storage_map, prefix, version)),
+              version(version),
+              use_mask(use_mask) {
+            if (config.is_chroma) {
                 LOG_INFO("Using pruned modulation (Chroma)");
             }
 
-            flux = Flux(flux_params);
+            flux = Flux(config);
             flux.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -1306,7 +1322,7 @@ namespace Flux {
             return "flux";
         }
 
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
             flux.get_param_tensors(tensors, prefix);
         }
 
@@ -1374,10 +1390,10 @@ namespace Flux {
             ggml_tensor* context   = make_optional_input(context_tensor);
             ggml_tensor* c_concat  = make_optional_input(c_concat_tensor);
             ggml_tensor* y         = make_optional_input(y_tensor);
-            if (flux_params.guidance_embed || flux_params.is_chroma) {
+            if (config.guidance_embed || config.is_chroma) {
                 if (!guidance_tensor.empty()) {
                     this->guidance_tensor = guidance_tensor;
-                    if (flux_params.is_chroma) {
+                    if (config.is_chroma) {
                         this->guidance_tensor.fill_(0.f);
                     }
                 }
@@ -1395,7 +1411,7 @@ namespace Flux {
             ggml_tensor* mod_index_arange = nullptr;
             ggml_tensor* dct              = nullptr;  // for chroma radiance
 
-            if (flux_params.is_chroma) {
+            if (config.is_chroma) {
                 if (!use_mask) {
                     y = nullptr;
                 }
@@ -1412,31 +1428,31 @@ namespace Flux {
             } else if (version == VERSION_OVIS_IMAGE) {
                 txt_arange_dims = {1, 2};
             }
-
             pe_vec      = Rope::gen_flux_pe(static_cast<int>(x->ne[1]),
                                             static_cast<int>(x->ne[0]),
-                                            flux_params.patch_size,
+                                            config.patch_size,
                                             static_cast<int>(x->ne[3]),
                                             static_cast<int>(context->ne[1]),
                                             txt_arange_dims,
                                             ref_latents,
                                             increase_ref_index,
-                                            flux_params.ref_index_scale,
-                                            flux_params.theta,
+                                            config.ref_index_scale,
+                                            config.theta,
                                             circular_y_enabled,
                                             circular_x_enabled,
-                                            flux_params.axes_dim);
-            int pos_len = static_cast<int>(pe_vec.size() / flux_params.axes_dim_sum / 2);
+                                            config.axes_dim,
+                                            sd_version_is_longcat(version));
+            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             // LOG_DEBUG("pos_len %d", pos_len);
-            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, flux_params.axes_dim_sum / 2, pos_len);
+            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
             // pe->data = pe_vec.data();
             // print_ggml_tensor(pe);
             // pe->data = nullptr;
             set_backend_tensor_data(pe, pe_vec.data());
 
             if (version == VERSION_CHROMA_RADIANCE) {
-                int patch_size     = flux_params.patch_size;
-                int nerf_max_freqs = flux_params.chroma_radiance_params.nerf_max_freqs;
+                int patch_size     = config.patch_size;
+                int nerf_max_freqs = config.chroma_radiance_params.nerf_max_freqs;
                 dct_vec            = fetch_dct_pos(patch_size, nerf_max_freqs);
                 dct                = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, nerf_max_freqs * nerf_max_freqs, patch_size * patch_size);
                 // dct->data = dct_vec.data();
@@ -1484,10 +1500,29 @@ namespace Flux {
                 return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, increase_ref_index, skip_layers);
             };
 
-            auto result = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+            auto result = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
             return result;
         }
 
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            const auto* extra = diffusion_extra_as<FluxDiffusionExtra>(diffusion_params);
+            static const std::vector<sd::Tensor<float>> empty_ref_latents;
+            static const std::vector<int> empty_skip_layers;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           tensor_or_empty(diffusion_params.c_concat),
+                           tensor_or_empty(diffusion_params.y),
+                           tensor_or_empty(extra->guidance),
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
+                           diffusion_params.increase_ref_index,
+                           extra->skip_layers ? *extra->skip_layers : empty_skip_layers);
+        }
+
         void test() {
             ggml_init_params params;
             params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
@@ -1545,10 +1580,11 @@ namespace Flux {
 
         static void load_from_file_and_test(const std::string& file_path) {
             // ggml_backend_t backend = ggml_backend_cuda_init(0);
-            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_backend_t backend    = sd_backend_cpu_init();
             ggml_type model_data_type = GGML_TYPE_COUNT;
 
-            ModelLoader model_loader;
+            auto model_manager        = std::make_shared<ModelManager>();
+            ModelLoader& model_loader = model_manager->loader();
             if (!model_loader.init_from_file_and_convert_name(file_path, "model.diffusion_model.")) {
                 LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
                 return;
@@ -1564,20 +1600,20 @@ namespace Flux {
             }
 
             std::shared_ptr<FluxRunner> flux = std::make_shared<FluxRunner>(backend,
-                                                                            backend,
                                                                             tensor_storage_map,
                                                                             "model.diffusion_model",
                                                                             VERSION_FLUX2,
-                                                                            false);
+                                                                            false,
+                                                                            model_manager);
 
-            flux->alloc_params_buffer();
-            std::map<std::string, ggml_tensor*> tensors;
-            flux->get_param_tensors(tensors, "model.diffusion_model");
-
-            bool success = model_loader.load_tensors(tensors);
-
-            if (!success) {
-                LOG_ERROR("load tensors from model loader failed");
+            if (!model_manager->register_runner_params("Flux test",
+                                                       *flux,
+                                                       "model.diffusion_model",
+                                                       ModelManager::ResidencyMode::ParamBackend,
+                                                       backend,
+                                                       backend) ||
+                !model_manager->validate_registered_tensors()) {
+                LOG_ERROR("register flux tensors with model manager failed");
                 return;
             }
 
@@ -1588,4 +1624,4 @@ namespace Flux {
 
 }  // namespace Flux
 
-#endif  // __FLUX_HPP__
+#endif  // __SD_MODEL_DIFFUSION_FLUX_HPP__

src/model/diffusion/hidream_o1.hpp

@@ -1,5 +1,5 @@
-#ifndef __SD_HIDREAM_O1_H__
-#define __SD_HIDREAM_O1_H__
+#ifndef __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__
+#define __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__
 
 #include <algorithm>
 #include <array>
@@ -10,10 +10,11 @@
 #include <utility>
 #include <vector>
 
-#include "common_dit.hpp"
-#include "conditioner.hpp"
-#include "llm.hpp"
-#include "util.h"
+#include "conditioning/conditioner.hpp"
+#include "core/util.h"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model/te/llm.hpp"
 
 namespace HiDreamO1 {
     constexpr int HIDREAM_O1_GRAPH_SIZE = 32768;
@@ -22,6 +23,39 @@ namespace HiDreamO1 {
     constexpr int IMAGE_TOKEN_ID        = 151655;
     constexpr int VISION_START_TOKEN_ID = 151652;
 
+    struct HiDreamO1Config {
+        LLM::LLMConfig llm;
+        int patch_size = PATCH_SIZE;
+
+        static HiDreamO1Config detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            (void)tensor_storage_map;
+            (void)prefix;
+            HiDreamO1Config config;
+            config.llm.arch                           = LLM::LLMArch::QWEN3_VL;
+            config.llm.hidden_size                    = 4096;
+            config.llm.intermediate_size              = 12288;
+            config.llm.num_layers                     = 36;
+            config.llm.num_heads                      = 32;
+            config.llm.num_kv_heads                   = 8;
+            config.llm.head_dim                       = 128;
+            config.llm.qkv_bias                       = false;
+            config.llm.qk_norm                        = true;
+            config.llm.vocab_size                     = 151936;
+            config.llm.rms_norm_eps                   = 1e-6f;
+            config.llm.vision.arch                    = LLM::LLMVisionArch::QWEN3_VL;
+            config.llm.vision.num_layers              = 27;
+            config.llm.vision.hidden_size             = 1152;
+            config.llm.vision.intermediate_size       = 4304;
+            config.llm.vision.num_heads               = 16;
+            config.llm.vision.out_hidden_size         = 4096;
+            config.llm.vision.patch_size              = 16;
+            config.llm.vision.spatial_merge_size      = 2;
+            config.llm.vision.temporal_patch_size     = 2;
+            config.llm.vision.num_position_embeddings = 2304;
+            return config;
+        }
+    };
+
     static inline std::string repeat_special_token(const std::string& token, int64_t count) {
         std::string out;
         out.reserve(static_cast<size_t>(count) * token.size());
@@ -204,50 +238,19 @@ namespace HiDreamO1 {
         }
     };
 
-    struct HiDreamO1Params {
-        LLM::LLMParams llm;
-        int patch_size = PATCH_SIZE;
-    };
-
-    static inline HiDreamO1Params make_hidream_o1_params() {
-        HiDreamO1Params params;
-        params.llm.arch                           = LLM::LLMArch::QWEN3_VL;
-        params.llm.hidden_size                    = 4096;
-        params.llm.intermediate_size              = 12288;
-        params.llm.num_layers                     = 36;
-        params.llm.num_heads                      = 32;
-        params.llm.num_kv_heads                   = 8;
-        params.llm.head_dim                       = 128;
-        params.llm.qkv_bias                       = false;
-        params.llm.qk_norm                        = true;
-        params.llm.vocab_size                     = 151936;
-        params.llm.rms_norm_eps                   = 1e-6f;
-        params.llm.vision.arch                    = LLM::LLMVisionArch::QWEN3_VL;
-        params.llm.vision.num_layers              = 27;
-        params.llm.vision.hidden_size             = 1152;
-        params.llm.vision.intermediate_size       = 4304;
-        params.llm.vision.num_heads               = 16;
-        params.llm.vision.out_hidden_size         = 4096;
-        params.llm.vision.patch_size              = 16;
-        params.llm.vision.spatial_merge_size      = 2;
-        params.llm.vision.temporal_patch_size     = 2;
-        params.llm.vision.num_position_embeddings = 2304;
-        return params;
-    }
-
     struct HiDreamO1Model : public GGMLBlock {
-        HiDreamO1Params params;
+        HiDreamO1Config config;
 
         HiDreamO1Model() = default;
-        explicit HiDreamO1Model(HiDreamO1Params params)
-            : params(std::move(params)) {
-            blocks["language_model"] = std::make_shared<LLM::TextModel>(this->params.llm);
-            blocks["t_embedder1"]    = std::make_shared<TimestepEmbedder>(this->params.llm.hidden_size);
-            blocks["x_embedder"]     = std::make_shared<BottleneckPatchEmbed>(this->params.patch_size * this->params.patch_size * 3,
-                                                                          this->params.llm.hidden_size / 4,
-                                                                          this->params.llm.hidden_size);
-            blocks["final_layer2"]   = std::make_shared<FinalLayer>(this->params.llm.hidden_size,
-                                                                  this->params.patch_size * this->params.patch_size * 3);
+        explicit HiDreamO1Model(HiDreamO1Config config)
+            : config(std::move(config)) {
+            blocks["language_model"] = std::make_shared<LLM::TextModel>(this->config.llm);
+            blocks["t_embedder1"]    = std::make_shared<TimestepEmbedder>(this->config.llm.hidden_size);
+            blocks["x_embedder"]     = std::make_shared<BottleneckPatchEmbed>(this->config.patch_size * this->config.patch_size * 3,
+                                                                          this->config.llm.hidden_size / 4,
+                                                                          this->config.llm.hidden_size);
+            blocks["final_layer2"]   = std::make_shared<FinalLayer>(this->config.llm.hidden_size,
+                                                                  this->config.patch_size * this->config.patch_size * 3);
         }
 
         std::shared_ptr<LLM::TextModel> text_model() {
@@ -268,7 +271,7 @@ namespace HiDreamO1 {
     };
 
     struct HiDreamO1VisionRunner : public GGMLRunner {
-        HiDreamO1Params params;
+        HiDreamO1Config config;
         std::shared_ptr<LLM::VisionModel> model;
 
         std::vector<int> window_index_vec;
@@ -279,12 +282,12 @@ namespace HiDreamO1 {
         std::array<std::vector<float>, 4> pos_embed_weight_data_;
 
         HiDreamO1VisionRunner(ggml_backend_t backend,
-                              ggml_backend_t params_backend,
-                              const String2TensorStorage& tensor_storage_map = {},
-                              const std::string& prefix                      = "model.visual")
-            : GGMLRunner(backend, params_backend),
-              params(make_hidream_o1_params()),
-              model(std::make_shared<LLM::VisionModel>(false, params.llm.vision)) {
+                              const String2TensorStorage& tensor_storage_map      = {},
+                              const std::string& prefix                           = "model.visual",
+                              std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : GGMLRunner(backend, weight_manager),
+              config(HiDreamO1Config::detect_from_weights(tensor_storage_map, prefix)),
+              model(std::make_shared<LLM::VisionModel>(false, config.llm.vision)) {
             model->init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -301,7 +304,7 @@ namespace HiDreamO1 {
                                                        compute_ctx,
                                                        runner_ctx,
                                                        image,
-                                                       params.llm.vision,
+                                                       config.llm.vision,
                                                        model,
                                                        window_index_vec,
                                                        window_inverse_index_vec,
@@ -320,28 +323,32 @@ namespace HiDreamO1 {
             return gf;
         }
 
-        sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& image) {
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& image,
+                                  bool auto_free           = true,
+                                  bool free_compute_buffer = true,
+                                  bool free_compute_params = true) {
             auto get_graph = [&]() {
                 return build_graph(image);
             };
-            auto output = GGMLRunner::compute<float>(get_graph, n_threads, false);
+            auto output = GGMLRunner::compute<float>(get_graph, n_threads, auto_free, free_compute_buffer, free_compute_params);
             return output.has_value() ? std::move(output.value()) : sd::Tensor<float>();
         }
     };
 
-    struct HiDreamO1Runner : public GGMLRunner {
-        HiDreamO1Params params;
+    struct HiDreamO1Runner : public DiffusionModelRunner {
+        HiDreamO1Config config;
         HiDreamO1Model model;
 
         std::vector<float> attention_mask_vec;
 
         HiDreamO1Runner(ggml_backend_t backend,
-                        ggml_backend_t params_backend,
-                        const String2TensorStorage& tensor_storage_map = {},
-                        const std::string& prefix                      = "model")
-            : GGMLRunner(backend, params_backend),
-              params(make_hidream_o1_params()) {
-            model = HiDreamO1Model(params);
+                        const String2TensorStorage& tensor_storage_map      = {},
+                        const std::string& prefix                           = "model",
+                        std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(HiDreamO1Config::detect_from_weights(tensor_storage_map, prefix)) {
+            model = HiDreamO1Model(config);
             model.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -349,7 +356,7 @@ namespace HiDreamO1 {
             return "hidream_o1";
         }
 
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
             model.get_param_tensors(tensors, prefix);
         }
 
@@ -452,7 +459,29 @@ namespace HiDreamO1 {
             auto get_graph = [&]() {
                 return build_graph(x, timestep, input_ids, input_pos, token_types, vinput_mask, image_embeds, ref_images);
             };
-            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            const auto* extra = diffusion_extra_as<HiDreamO1DiffusionExtra>(diffusion_params);
+            GGML_ASSERT(extra != nullptr);
+            GGML_ASSERT(extra->input_ids != nullptr);
+            GGML_ASSERT(extra->input_pos != nullptr);
+            GGML_ASSERT(extra->token_types != nullptr);
+            static const std::vector<sd::Tensor<float>> empty_images;
+            static const std::vector<std::pair<int, sd::Tensor<float>>> empty_image_embeds;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           *extra->input_ids,
+                           *extra->input_pos,
+                           *extra->token_types,
+                           tensor_or_empty(extra->vinput_mask),
+                           extra->image_embeds ? *extra->image_embeds : empty_image_embeds,
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_images);
         }
     };
 
@@ -461,26 +490,14 @@ namespace HiDreamO1 {
         std::shared_ptr<HiDreamO1VisionRunner> vision_runner;
 
         HiDreamO1Conditioner(ggml_backend_t backend,
-                             ggml_backend_t params_backend,
-                             const String2TensorStorage& tensor_storage_map = {})
-            : vision_runner(std::make_shared<HiDreamO1VisionRunner>(backend, params_backend, tensor_storage_map)) {}
+                             const String2TensorStorage& tensor_storage_map      = {},
+                             std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : vision_runner(std::make_shared<HiDreamO1VisionRunner>(backend, tensor_storage_map, "model.visual", weight_manager)) {}
 
         void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
             vision_runner->get_param_tensors(tensors);
         }
 
-        void alloc_params_buffer() override {
-            vision_runner->alloc_params_buffer();
-        }
-
-        void free_params_buffer() override {
-            vision_runner->free_params_buffer();
-        }
-
-        size_t get_params_buffer_size() override {
-            return vision_runner->get_params_buffer_size();
-        }
-
         void set_max_graph_vram_bytes(size_t max_graph_vram_bytes) override {
             vision_runner->set_max_graph_vram_bytes(max_graph_vram_bytes);
         }
@@ -493,6 +510,10 @@ namespace HiDreamO1 {
             vision_runner->set_weight_adapter(adapter);
         }
 
+        void runner_done() override {
+            vision_runner->runner_done();
+        }
+
         SDCondition get_learned_condition(int n_threads,
                                           const ConditionerParams& conditioner_params) override {
             SDCondition result;
@@ -638,7 +659,7 @@ namespace HiDreamO1 {
             result.c_vinput_mask  = sd::Tensor<int32_t>(vinput_mask_shape, std::move(vinput_mask));
             result.c_image_embeds.reserve(vlm_images.size());
             for (const auto& vlm_image : vlm_images) {
-                auto image_embed = vision_runner->compute(n_threads, vlm_image.second);
+                auto image_embed = vision_runner->compute(n_threads, vlm_image.second, false, true, true);
                 if (image_embed.empty()) {
                     LOG_ERROR("hidream_o1 conditioner: encode VLM image failed");
                     return SDCondition();
@@ -650,4 +671,4 @@ namespace HiDreamO1 {
     };
 }  // namespace HiDreamO1
 
-#endif  // __SD_HIDREAM_O1_H__
+#endif  // __SD_MODEL_DIFFUSION_HIDREAM_O1_HPP__

src/model/diffusion/ideogram4.hpp

@@ -0,0 +1,557 @@
+#ifndef __SD_MODEL_DIFFUSION_IDEOGRAM4_HPP__
+#define __SD_MODEL_DIFFUSION_IDEOGRAM4_HPP__
+
+#include <algorithm>
+#include <cmath>
+#include <cstdlib>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "core/ggml_graph_cut.h"
+#include "model/common/rope.hpp"
+#include "model/diffusion/model.hpp"
+
+namespace Ideogram4 {
+    constexpr int IDEOGRAM4_GRAPH_SIZE    = 65536;
+    constexpr int OUTPUT_IMAGE_INDICATOR  = 2;
+    constexpr int IMAGE_POSITION_OFFSET   = 65536;
+    constexpr int DEFAULT_MROPE_SECTION_T = 24;
+    constexpr int DEFAULT_MROPE_SECTION_H = 20;
+    constexpr int DEFAULT_MROPE_SECTION_W = 20;
+    constexpr int TIMESTEP_MAX_PERIOD     = 10000;
+    constexpr int LLM_HIDDEN_STATE_LAYERS = 13;
+
+    struct Ideogram4Config {
+        int64_t emb_dim                = 4608;
+        int64_t num_layers             = 34;
+        int64_t num_heads              = 18;
+        int64_t intermediate_size      = 12288;
+        int64_t adanln_dim             = 512;
+        int64_t in_channels            = 128;
+        int64_t llm_features_dim       = 53248;
+        int64_t rope_theta             = 5000000;
+        float norm_eps                 = 1e-5f;
+        int patch_size                 = 2;
+        int ae_channels                = 32;
+        std::vector<int> mrope_section = {DEFAULT_MROPE_SECTION_T,
+                                          DEFAULT_MROPE_SECTION_H,
+                                          DEFAULT_MROPE_SECTION_W};
+
+        static Ideogram4Config detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                                   const std::string& prefix) {
+            Ideogram4Config config;
+            int64_t detected_layers  = 0;
+            std::string layer_prefix = prefix.empty() ? "layers." : prefix + ".layers.";
+            for (const auto& [name, _] : tensor_storage_map) {
+                if (name.find(layer_prefix) != 0) {
+                    continue;
+                }
+                std::string tail = name.substr(layer_prefix.size());
+                size_t dot       = tail.find('.');
+                if (dot == std::string::npos) {
+                    continue;
+                }
+                int layer_idx   = std::atoi(tail.substr(0, dot).c_str());
+                detected_layers = std::max<int64_t>(detected_layers, layer_idx + 1);
+            }
+            if (detected_layers > 0) {
+                config.num_layers = detected_layers;
+                LOG_DEBUG("ideogram4: num_layers = %" PRId64 ", emb_dim = %" PRId64 ", num_heads = %" PRId64 ", intermediate_size = %" PRId64,
+                          config.num_layers,
+                          config.emb_dim,
+                          config.num_heads,
+                          config.intermediate_size);
+            }
+            return config;
+        }
+    };
+
+    __STATIC_INLINE__ ggml_tensor* timestep_embedding_sin_cos(ggml_context* ctx,
+                                                              ggml_tensor* timesteps,
+                                                              int dim) {
+        GGML_ASSERT(dim % 2 == 0);
+        auto embedding = ggml_ext_timestep_embedding(ctx, timesteps, dim, TIMESTEP_MAX_PERIOD, 10.f);
+        auto chunks    = ggml_ext_chunk(ctx, embedding, 2, 0);
+        return ggml_concat(ctx, chunks[1], chunks[0], 0);
+    }
+
+    __STATIC_INLINE__ ggml_tensor* to_token_modulation(ggml_context* ctx, ggml_tensor* x) {
+        // [N, C] -> [N, 1, C] in PyTorch layout.
+        if (ggml_n_dims(x) < 3 || x->ne[1] != 1) {
+            x = ggml_reshape_3d(ctx, x, x->ne[0], 1, x->ne[1]);
+        }
+        return x;
+    }
+
+    __STATIC_INLINE__ ggml_tensor* interleave_hidden_state_layers(ggml_context* ctx, ggml_tensor* x) {
+        // Match upstream stack(...).permute(1, 2, 3, 0).reshape(...):
+        // [layers * hidden, tokens, batch] -> [hidden * layers, tokens, batch].
+        GGML_ASSERT(x->ne[0] % LLM_HIDDEN_STATE_LAYERS == 0);
+        const int64_t hidden_size = x->ne[0] / LLM_HIDDEN_STATE_LAYERS;
+        const int64_t token_count = x->ne[1];
+        const int64_t batch_count = x->ne[2];
+
+        x = ggml_reshape_4d(ctx, x, hidden_size, LLM_HIDDEN_STATE_LAYERS, token_count, batch_count);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 1, 0, 2, 3));
+        return ggml_reshape_3d(ctx, x, hidden_size * LLM_HIDDEN_STATE_LAYERS, token_count, batch_count);
+    }
+
+    __STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* scale) {
+        scale = to_token_modulation(ctx, scale);
+        return ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+    }
+
+    __STATIC_INLINE__ ggml_tensor* patchify(ggml_context* ctx, ggml_tensor* x, const Ideogram4Config& config) {
+        // x: [N, 128, H, W] with channel order [ae, ph, pw].
+        // return: [N, H*W, 128] with token channel order [ph, pw, ae].
+        const int64_t W = x->ne[0];
+        const int64_t H = x->ne[1];
+        const int64_t C = x->ne[2];
+        const int64_t N = x->ne[3];
+
+        GGML_ASSERT(N == 1);
+        GGML_ASSERT(C == config.ae_channels * config.patch_size * config.patch_size);
+
+        x = ggml_cont(ctx, x);
+        x = ggml_reshape_4d(ctx, x, W * H, config.patch_size, config.patch_size, config.ae_channels);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 3, 1, 2, 0));
+        x = ggml_reshape_3d(ctx, x, C, W * H, N);
+        return x;
+    }
+
+    __STATIC_INLINE__ ggml_tensor* unpatchify(ggml_context* ctx,
+                                              ggml_tensor* x,
+                                              int64_t H,
+                                              int64_t W,
+                                              const Ideogram4Config& config) {
+        const int64_t C = x->ne[0];
+        const int64_t N = x->ne[2];
+
+        GGML_ASSERT(N == 1);
+        GGML_ASSERT(C == config.ae_channels * config.patch_size * config.patch_size);
+        GGML_ASSERT(x->ne[1] == H * W);
+
+        x = ggml_reshape_4d(ctx, x, config.ae_channels, config.patch_size, config.patch_size, H * W);
+        x = ggml_cont(ctx, ggml_permute(ctx, x, 3, 1, 2, 0));
+        x = ggml_reshape_4d(ctx, x, W, H, C, N);
+        return x;
+    }
+
+    __STATIC_INLINE__ std::shared_ptr<Linear> make_linear(int64_t in_features,
+                                                          int64_t out_features,
+                                                          bool bias = true) {
+        return std::make_shared<Linear>(in_features, out_features, bias, false, false, 1.f, true);
+    }
+
+    __STATIC_INLINE__ std::vector<float> gen_ideogram4_pe(int grid_h,
+                                                          int grid_w,
+                                                          int bs,
+                                                          int context_len,
+                                                          int head_dim,
+                                                          int rope_theta,
+                                                          const std::vector<int>& mrope_section,
+                                                          bool circular_x = false,
+                                                          bool circular_y = false) {
+        GGML_ASSERT(bs == 1);
+        std::vector<std::vector<float>> ids(static_cast<size_t>(bs) * (context_len + grid_h * grid_w),
+                                            std::vector<float>(3, 0.f));
+
+        for (int i = 0; i < context_len; ++i) {
+            ids[i] = {static_cast<float>(i), static_cast<float>(i), static_cast<float>(i)};
+        }
+
+        int cursor = context_len;
+        for (int y = 0; y < grid_h; ++y) {
+            for (int x = 0; x < grid_w; ++x) {
+                ids[cursor++] = {static_cast<float>(IMAGE_POSITION_OFFSET),
+                                 static_cast<float>(IMAGE_POSITION_OFFSET + y),
+                                 static_cast<float>(IMAGE_POSITION_OFFSET + x)};
+            }
+        }
+
+        std::vector<std::vector<int>> axis_wrap_dims(3);
+        if (circular_y || circular_x) {
+            size_t total_len = static_cast<size_t>(bs) * (context_len + grid_h * grid_w);
+            axis_wrap_dims[1].assign(total_len, 0);
+            axis_wrap_dims[2].assign(total_len, 0);
+            if (circular_y) {
+                for (size_t idx = static_cast<size_t>(context_len); idx < total_len; ++idx) {
+                    axis_wrap_dims[1][idx] = grid_h;
+                }
+            }
+            if (circular_x) {
+                for (size_t idx = static_cast<size_t>(context_len); idx < total_len; ++idx) {
+                    axis_wrap_dims[2][idx] = grid_w;
+                }
+            }
+        }
+
+        return Rope::embed_interleaved_mrope(ids,
+                                             bs,
+                                             static_cast<float>(rope_theta),
+                                             head_dim,
+                                             mrope_section,
+                                             axis_wrap_dims);
+    }
+
+    class Ideogram4Attention : public GGMLBlock {
+    protected:
+        int64_t hidden_size;
+        int64_t num_heads;
+        int64_t head_dim;
+
+    public:
+        Ideogram4Attention(int64_t hidden_size, int64_t num_heads, float eps)
+            : hidden_size(hidden_size), num_heads(num_heads), head_dim(hidden_size / num_heads) {
+            GGML_ASSERT(hidden_size % num_heads == 0);
+            blocks["qkv"]    = make_linear(hidden_size, hidden_size * 3, false);
+            blocks["norm_q"] = std::make_shared<RMSNorm>(head_dim, eps);
+            blocks["norm_k"] = std::make_shared<RMSNorm>(head_dim, eps);
+            blocks["o"]      = make_linear(hidden_size, hidden_size, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe,
+                             ggml_tensor* mask = nullptr) {
+            int64_t n_token = x->ne[1];
+            int64_t N       = x->ne[2];
+
+            auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
+            auto norm_q   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto out_proj = std::dynamic_pointer_cast<Linear>(blocks["o"]);
+
+            auto qkv     = qkv_proj->forward(ctx, x);
+            auto qkv_vec = split_qkv(ctx->ggml_ctx, qkv);
+            auto q       = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[0], head_dim, num_heads, n_token, N);
+            auto k       = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[1], head_dim, num_heads, n_token, N);
+            auto v       = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[2], head_dim, num_heads, n_token, N);
+
+            q = norm_q->forward(ctx, q);
+            k = norm_k->forward(ctx, k);
+
+            x = Rope::attention(ctx, q, k, v, pe, mask, 1.f / 128.f, false);
+            x = out_proj->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Ideogram4MLP : public GGMLBlock {
+    public:
+        Ideogram4MLP(int64_t dim, int64_t hidden_dim) {
+            blocks["w1"] = make_linear(dim, hidden_dim, false);
+            blocks["w2"] = make_linear(hidden_dim, dim, false);
+            blocks["w3"] = make_linear(dim, hidden_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto w1 = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
+            auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
+            auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
+
+            auto x1 = ggml_silu(ctx->ggml_ctx, w1->forward(ctx, x));
+            auto x3 = w3->forward(ctx, x);
+            x       = ggml_mul(ctx->ggml_ctx, x1, x3);
+            x       = w2->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Ideogram4TransformerBlock : public GGMLBlock {
+    public:
+        Ideogram4TransformerBlock(const Ideogram4Config& config) {
+            blocks["attention"]        = std::make_shared<Ideogram4Attention>(config.emb_dim, config.num_heads, config.norm_eps);
+            blocks["feed_forward"]     = std::make_shared<Ideogram4MLP>(config.emb_dim, config.intermediate_size);
+            blocks["attention_norm1"]  = std::make_shared<RMSNorm>(config.emb_dim, config.norm_eps);
+            blocks["ffn_norm1"]        = std::make_shared<RMSNorm>(config.emb_dim, config.norm_eps);
+            blocks["attention_norm2"]  = std::make_shared<RMSNorm>(config.emb_dim, config.norm_eps);
+            blocks["ffn_norm2"]        = std::make_shared<RMSNorm>(config.emb_dim, config.norm_eps);
+            blocks["adaln_modulation"] = make_linear(config.adanln_dim, 4 * config.emb_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* pe,
+                             ggml_tensor* adaln_input,
+                             ggml_tensor* mask = nullptr) {
+            auto attention        = std::dynamic_pointer_cast<Ideogram4Attention>(blocks["attention"]);
+            auto feed_forward     = std::dynamic_pointer_cast<Ideogram4MLP>(blocks["feed_forward"]);
+            auto attention_norm1  = std::dynamic_pointer_cast<RMSNorm>(blocks["attention_norm1"]);
+            auto ffn_norm1        = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm1"]);
+            auto attention_norm2  = std::dynamic_pointer_cast<RMSNorm>(blocks["attention_norm2"]);
+            auto ffn_norm2        = std::dynamic_pointer_cast<RMSNorm>(blocks["ffn_norm2"]);
+            auto adaln_modulation = std::dynamic_pointer_cast<Linear>(blocks["adaln_modulation"]);
+
+            auto mod       = adaln_modulation->forward(ctx, adaln_input);
+            auto mods      = ggml_ext_chunk(ctx->ggml_ctx, mod, 4, 0);
+            auto scale_msa = mods[0];
+            auto gate_msa  = to_token_modulation(ctx->ggml_ctx, ggml_tanh(ctx->ggml_ctx, mods[1]));
+            auto scale_mlp = mods[2];
+            auto gate_mlp  = to_token_modulation(ctx->ggml_ctx, ggml_tanh(ctx->ggml_ctx, mods[3]));
+
+            auto attn_out = attention_norm1->forward(ctx, x);
+            attn_out      = modulate(ctx->ggml_ctx, attn_out, scale_msa);
+            attn_out      = attention->forward(ctx, attn_out, pe, mask);
+            attn_out      = attention_norm2->forward(ctx, attn_out);
+            x             = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out, gate_msa));
+
+            auto ffn_out = ffn_norm1->forward(ctx, x);
+            ffn_out      = modulate(ctx->ggml_ctx, ffn_out, scale_mlp);
+            ffn_out      = feed_forward->forward(ctx, ffn_out);
+            ffn_out      = ffn_norm2->forward(ctx, ffn_out);
+            x            = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, ffn_out, gate_mlp));
+
+            return x;
+        }
+    };
+
+    class Ideogram4EmbedScalar : public GGMLBlock {
+    protected:
+        int64_t dim;
+
+    public:
+        Ideogram4EmbedScalar(int64_t dim)
+            : dim(dim) {
+            blocks["mlp_in"]  = make_linear(dim, dim, true);
+            blocks["mlp_out"] = make_linear(dim, dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto mlp_in  = std::dynamic_pointer_cast<Linear>(blocks["mlp_in"]);
+            auto mlp_out = std::dynamic_pointer_cast<Linear>(blocks["mlp_out"]);
+
+            x = timestep_embedding_sin_cos(ctx->ggml_ctx, x, static_cast<int>(dim));
+            x = ggml_silu(ctx->ggml_ctx, mlp_in->forward(ctx, x));
+            x = mlp_out->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Ideogram4FinalLayer : public GGMLBlock {
+    public:
+        Ideogram4FinalLayer(const Ideogram4Config& config) {
+            blocks["norm_final"]       = std::make_shared<LayerNorm>(config.emb_dim, 1e-6f, false);
+            blocks["linear"]           = make_linear(config.emb_dim, config.in_channels, true);
+            blocks["adaln_modulation"] = make_linear(config.adanln_dim, config.emb_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* c) {
+            auto norm_final       = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_final"]);
+            auto linear           = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            auto adaln_modulation = std::dynamic_pointer_cast<Linear>(blocks["adaln_modulation"]);
+
+            auto scale = adaln_modulation->forward(ctx, ggml_silu(ctx->ggml_ctx, c));
+            x          = norm_final->forward(ctx, x);
+            x          = modulate(ctx->ggml_ctx, x, scale);
+            x          = linear->forward(ctx, x);
+            return x;
+        }
+    };
+
+    class Ideogram4Transformer : public GGMLBlock {
+    protected:
+        Ideogram4Config config;
+
+    public:
+        Ideogram4Transformer() = default;
+        explicit Ideogram4Transformer(Ideogram4Config config)
+            : config(std::move(config)) {
+            blocks["input_proj"]            = make_linear(this->config.in_channels, this->config.emb_dim, true);
+            blocks["llm_cond_norm"]         = std::make_shared<RMSNorm>(this->config.llm_features_dim, 1e-6f);
+            blocks["llm_cond_proj"]         = make_linear(this->config.llm_features_dim, this->config.emb_dim, true);
+            blocks["t_embedding"]           = std::make_shared<Ideogram4EmbedScalar>(this->config.emb_dim);
+            blocks["adaln_proj"]            = make_linear(this->config.emb_dim, this->config.adanln_dim, true);
+            blocks["embed_image_indicator"] = std::make_shared<Embedding>(2, this->config.emb_dim);
+
+            for (int i = 0; i < this->config.num_layers; ++i) {
+                blocks["layers." + std::to_string(i)] = std::make_shared<Ideogram4TransformerBlock>(this->config);
+            }
+            blocks["final_layer"] = std::make_shared<Ideogram4FinalLayer>(this->config);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timestep,
+                             ggml_tensor* context,
+                             ggml_tensor* pe,
+                             ggml_tensor* image_indicator_ids) {
+            int64_t W = x->ne[0];
+            int64_t H = x->ne[1];
+            int64_t N = x->ne[3];
+            GGML_ASSERT(N == 1);
+
+            auto input_proj            = std::dynamic_pointer_cast<Linear>(blocks["input_proj"]);
+            auto llm_cond_norm         = std::dynamic_pointer_cast<RMSNorm>(blocks["llm_cond_norm"]);
+            auto llm_cond_proj         = std::dynamic_pointer_cast<Linear>(blocks["llm_cond_proj"]);
+            auto t_embedding           = std::dynamic_pointer_cast<Ideogram4EmbedScalar>(blocks["t_embedding"]);
+            auto adaln_proj            = std::dynamic_pointer_cast<Linear>(blocks["adaln_proj"]);
+            auto embed_image_indicator = std::dynamic_pointer_cast<Embedding>(blocks["embed_image_indicator"]);
+            auto final_layer           = std::dynamic_pointer_cast<Ideogram4FinalLayer>(blocks["final_layer"]);
+
+            auto img = patchify(ctx->ggml_ctx, x, config);
+            img      = input_proj->forward(ctx, img);
+
+            ggml_tensor* h      = img;
+            int64_t context_len = 0;
+            if (context != nullptr) {
+                if (ggml_n_dims(context) < 3) {
+                    context = ggml_reshape_3d(ctx->ggml_ctx, context, context->ne[0], context->ne[1], 1);
+                }
+                context     = interleave_hidden_state_layers(ctx->ggml_ctx, context);
+                context_len = context->ne[1];
+                auto txt    = llm_cond_norm->forward(ctx, context);
+                txt         = llm_cond_proj->forward(ctx, txt);
+                h           = ggml_concat(ctx->ggml_ctx, txt, img, 1);
+            }
+
+            auto indicator_embedding = embed_image_indicator->forward(ctx, image_indicator_ids);
+            h                        = ggml_add(ctx->ggml_ctx, h, indicator_embedding);
+
+            auto t_cond      = t_embedding->forward(ctx, timestep);
+            auto adaln_input = ggml_silu(ctx->ggml_ctx, adaln_proj->forward(ctx, t_cond));
+
+            for (int i = 0; i < config.num_layers; ++i) {
+                auto block = std::dynamic_pointer_cast<Ideogram4TransformerBlock>(blocks["layers." + std::to_string(i)]);
+                h          = block->forward(ctx, h, pe, adaln_input, nullptr);
+                sd::ggml_graph_cut::mark_graph_cut(h, "ideogram4.layers." + std::to_string(i), "hidden");
+            }
+
+            h = final_layer->forward(ctx, h, adaln_input);
+            if (context_len > 0) {
+                h = ggml_ext_slice(ctx->ggml_ctx, h, 1, context_len, h->ne[1]);
+            }
+
+            h = unpatchify(ctx->ggml_ctx, h, H, W, config);
+            h = ggml_ext_scale(ctx->ggml_ctx, h, -1.f);
+            return h;
+        }
+    };
+
+    class Ideogram4Runner : public DiffusionModelRunner {
+    protected:
+        bool should_use_uncond_model(const DiffusionParams& diffusion_params) const {
+            return has_uncond_model &&
+                   diffusion_params.context == nullptr &&
+                   diffusion_params.y != nullptr &&
+                   !diffusion_params.y->empty();
+        }
+
+    public:
+        Ideogram4Config config;
+        Ideogram4Transformer model;
+        Ideogram4Transformer uncond_model;
+        bool has_uncond_model = false;
+        std::string uncond_prefix;
+        std::vector<float> pe_vec;
+        std::vector<int32_t> image_indicator_vec;
+
+        Ideogram4Runner(ggml_backend_t backend,
+                        const String2TensorStorage& tensor_storage_map      = {},
+                        const std::string prefix                            = "",
+                        std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(Ideogram4Config::detect_from_weights(tensor_storage_map, prefix)),
+              uncond_prefix(prefix + ".uncond") {
+            model = Ideogram4Transformer(config);
+            model.init(params_ctx, tensor_storage_map, prefix);
+            for (const auto& pair : tensor_storage_map) {
+                const std::string& name = pair.first;
+                if (starts_with(name, uncond_prefix)) {
+                    has_uncond_model = true;
+                    break;
+                }
+            }
+            if (has_uncond_model) {
+                LOG_DEBUG("using uncond model");
+                uncond_model = Ideogram4Transformer(config);
+                uncond_model.init(params_ctx, tensor_storage_map, uncond_prefix);
+            }
+        }
+
+        std::string get_desc() override {
+            return "ideogram4";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+            if (has_uncond_model) {
+                uncond_model.get_param_tensors(tensors, this->uncond_prefix);
+            }
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 bool use_uncond_model = false) {
+            ggml_cgraph* gf        = new_graph_custom(IDEOGRAM4_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            Ideogram4Transformer& active_model = use_uncond_model ? uncond_model : model;
+
+            ggml_tensor* context = nullptr;
+            int64_t context_len  = 0;
+            if (!context_tensor.empty()) {
+                context     = make_input(context_tensor);
+                context_len = context->ne[1];
+            }
+
+            int64_t grid_w   = x->ne[0];
+            int64_t grid_h   = x->ne[1];
+            int64_t pos_len  = context_len + grid_h * grid_w;
+            int64_t head_dim = config.emb_dim / config.num_heads;
+
+            auto runner_ctx = get_context();
+            pe_vec          = gen_ideogram4_pe(static_cast<int>(grid_h),
+                                               static_cast<int>(grid_w),
+                                               static_cast<int>(x->ne[3]),
+                                               static_cast<int>(context_len),
+                                               static_cast<int>(head_dim),
+                                               static_cast<int>(config.rope_theta),
+                                               config.mrope_section,
+                                               runner_ctx.circular_x_enabled,
+                                               runner_ctx.circular_y_enabled);
+            auto pe         = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, head_dim / 2, pos_len);
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            image_indicator_vec.assign(static_cast<size_t>(pos_len), 1);
+            for (int64_t i = 0; i < context_len; ++i) {
+                image_indicator_vec[static_cast<size_t>(i)] = 0;
+            }
+            auto indicator = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_I32, pos_len, x->ne[3]);
+            set_backend_tensor_data(indicator, image_indicator_vec.data());
+
+            ggml_tensor* out = active_model.forward(&runner_ctx, x, timesteps, context, pe, indicator);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  bool use_uncond_model = false) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, use_uncond_model);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            bool use_uncond_model = should_use_uncond_model(diffusion_params);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           use_uncond_model);
+        }
+    };
+}  // namespace Ideogram4
+
+#endif  // __SD_MODEL_DIFFUSION_IDEOGRAM4_HPP__

src/model/diffusion/lens.hpp

@@ -0,0 +1,426 @@
+#ifndef __SD_MODEL_DIFFUSION_LENS_HPP__
+#define __SD_MODEL_DIFFUSION_LENS_HPP__
+
+#include <memory>
+#include <vector>
+
+#include "model/common/block.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/flux.hpp"
+#include "model/diffusion/model.hpp"
+#include "model/diffusion/qwen_image.hpp"
+
+namespace Lens {
+    constexpr int LENS_GRAPH_SIZE = 40960;
+
+    struct LensConfig {
+        int patch_size              = 2;
+        int64_t in_channels         = 128;
+        int64_t out_channels        = 32;
+        int num_layers              = 48;
+        int64_t attention_head_dim  = 64;
+        int64_t num_attention_heads = 24;
+        int64_t joint_attention_dim = 2880;
+        int selected_layer_count    = 4;
+        int theta                   = 10000;
+        std::vector<int> axes_dim   = {8, 28, 28};
+        int axes_dim_sum            = 64;
+
+        static LensConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            LensConfig config;
+            config.num_layers = 0;
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "img_in.weight") && tensor_storage.n_dims == 2) {
+                    config.in_channels = tensor_storage.ne[0];
+                    int64_t inner_dim  = tensor_storage.ne[1];
+                    if (config.attention_head_dim > 0) {
+                        config.num_attention_heads = inner_dim / config.attention_head_dim;
+                    }
+                } else if (ends_with(name, "txt_in.weight") && tensor_storage.n_dims == 2) {
+                    config.selected_layer_count = static_cast<int>(tensor_storage.ne[0] / config.joint_attention_dim);
+                } else if (ends_with(name, "proj_out.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area  = config.patch_size * config.patch_size;
+                    config.out_channels = tensor_storage.ne[1] / patch_area;
+                } else if (ends_with(name, "transformer_blocks.0.attn.norm_q.weight") && tensor_storage.n_dims == 1) {
+                    config.attention_head_dim = tensor_storage.ne[0];
+                }
+
+                size_t pos = name.find("transformer_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        if (block_index + 1 > config.num_layers) {
+                            config.num_layers = block_index + 1;
+                        }
+                    }
+                }
+            }
+            if (config.num_layers == 0) {
+                config.num_layers = 48;
+            }
+            config.axes_dim_sum = 0;
+            for (int axis_dim : config.axes_dim) {
+                config.axes_dim_sum += axis_dim;
+            }
+            LOG_DEBUG("lens: num_layers = %d, selected_layer_count = %d, hidden_size = %" PRId64 ", num_attention_heads = %" PRId64 ", attention_head_dim = %" PRId64 ", in_channels = %" PRId64 ", out_channels = %" PRId64,
+                      config.num_layers,
+                      config.selected_layer_count,
+                      config.num_attention_heads * config.attention_head_dim,
+                      config.num_attention_heads,
+                      config.attention_head_dim,
+                      config.in_channels,
+                      config.out_channels);
+            return config;
+        }
+    };
+
+    struct LensTimestepProjEmbeddings : public GGMLBlock {
+        LensTimestepProjEmbeddings(int64_t embedding_dim) {
+            blocks["timestep_embedder"] = std::make_shared<Qwen::TimestepEmbedding>(256, embedding_dim);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* timesteps) {
+            auto timestep_embedder = std::dynamic_pointer_cast<Qwen::TimestepEmbedding>(blocks["timestep_embedder"]);
+            auto timesteps_proj    = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, 256, 10000, 1000.f);
+            return timestep_embedder->forward(ctx, timesteps_proj);
+        }
+    };
+
+    struct LensGateMLP : public GGMLBlock {
+        LensGateMLP(int64_t dim, int64_t hidden_dim) {
+            blocks["w1"] = std::make_shared<Linear>(dim, hidden_dim, false);
+            blocks["w2"] = std::make_shared<Linear>(hidden_dim, dim, false);
+            blocks["w3"] = std::make_shared<Linear>(dim, hidden_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto w1 = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
+            auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
+            auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
+
+            auto gate = ggml_silu(ctx->ggml_ctx, w1->forward(ctx, x));
+            auto up   = w3->forward(ctx, x);
+            x         = ggml_mul(ctx->ggml_ctx, gate, up);
+            return w2->forward(ctx, x);
+        }
+    };
+
+    struct LensJointAttention : public GGMLBlock {
+        int64_t dim_head;
+        int64_t num_heads;
+
+        LensJointAttention(int64_t query_dim,
+                           int64_t dim_head,
+                           int64_t num_heads,
+                           float eps = 1e-5f)
+            : dim_head(dim_head), num_heads(num_heads) {
+            int64_t inner_dim = dim_head * num_heads;
+            blocks["img_qkv"] = std::make_shared<Linear>(query_dim, inner_dim * 3, true);
+            blocks["txt_qkv"] = std::make_shared<Linear>(query_dim, inner_dim * 3, true);
+
+            blocks["norm_q"]       = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_k"]       = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_added_q"] = std::make_shared<RMSNorm>(dim_head, eps);
+            blocks["norm_added_k"] = std::make_shared<RMSNorm>(dim_head, eps);
+
+            blocks["to_out.0"]   = std::make_shared<Linear>(inner_dim, query_dim, true);
+            blocks["to_add_out"] = std::make_shared<Linear>(inner_dim, query_dim, true);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* img,
+                                                      ggml_tensor* txt,
+                                                      ggml_tensor* pe,
+                                                      ggml_tensor* mask = nullptr) {
+            auto img_qkv    = std::dynamic_pointer_cast<Linear>(blocks["img_qkv"]);
+            auto txt_qkv    = std::dynamic_pointer_cast<Linear>(blocks["txt_qkv"]);
+            auto norm_q     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"]);
+            auto norm_k     = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"]);
+            auto norm_add_q = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_added_q"]);
+            auto norm_add_k = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_added_k"]);
+            auto to_out_0   = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
+            auto to_add_out = std::dynamic_pointer_cast<Linear>(blocks["to_add_out"]);
+            int64_t n_img   = img->ne[1];
+            int64_t n_txt   = txt->ne[1];
+            int64_t N       = img->ne[2];
+            int64_t inner   = dim_head * num_heads;
+
+            auto img_qkv_vec = split_qkv(ctx->ggml_ctx, img_qkv->forward(ctx, img));
+            auto txt_qkv_vec = split_qkv(ctx->ggml_ctx, txt_qkv->forward(ctx, txt));
+
+            auto img_q = ggml_reshape_4d(ctx->ggml_ctx, img_qkv_vec[0], dim_head, num_heads, n_img, N);
+            auto img_k = ggml_reshape_4d(ctx->ggml_ctx, img_qkv_vec[1], dim_head, num_heads, n_img, N);
+            auto img_v = ggml_reshape_4d(ctx->ggml_ctx, img_qkv_vec[2], dim_head, num_heads, n_img, N);
+
+            img_q = norm_q->forward(ctx, img_q);
+            img_k = norm_k->forward(ctx, img_k);
+
+            auto txt_q = ggml_reshape_4d(ctx->ggml_ctx, txt_qkv_vec[0], dim_head, num_heads, n_txt, N);
+            auto txt_k = ggml_reshape_4d(ctx->ggml_ctx, txt_qkv_vec[1], dim_head, num_heads, n_txt, N);
+            auto txt_v = ggml_reshape_4d(ctx->ggml_ctx, txt_qkv_vec[2], dim_head, num_heads, n_txt, N);
+
+            txt_q = norm_add_q->forward(ctx, txt_q);
+            txt_k = norm_add_k->forward(ctx, txt_k);
+
+            auto q = ggml_concat(ctx->ggml_ctx, img_q, txt_q, 2);
+            auto k = ggml_concat(ctx->ggml_ctx, img_k, txt_k, 2);
+            auto v = ggml_concat(ctx->ggml_ctx, img_v, txt_v, 2);
+
+            auto attn = Rope::attention(ctx, q, k, v, pe, mask, (1.0f / 128.f));
+
+            auto img_attn_out = ggml_view_3d(ctx->ggml_ctx,
+                                             attn,
+                                             inner,
+                                             n_img,
+                                             N,
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             0);
+            auto txt_attn_out = ggml_view_3d(ctx->ggml_ctx,
+                                             attn,
+                                             inner,
+                                             n_txt,
+                                             N,
+                                             attn->nb[1],
+                                             attn->nb[2],
+                                             n_img * attn->nb[1]);
+
+            img_attn_out = to_out_0->forward(ctx, ggml_cont(ctx->ggml_ctx, img_attn_out));
+            txt_attn_out = to_add_out->forward(ctx, ggml_cont(ctx->ggml_ctx, txt_attn_out));
+            return {img_attn_out, txt_attn_out};
+        }
+    };
+
+    struct LensTransformerBlock : public GGMLBlock {
+        LensTransformerBlock(int64_t dim,
+                             int64_t num_attention_heads,
+                             int64_t attention_head_dim,
+                             float eps = 1e-6f) {
+            int64_t mlp_hidden_dim = dim / 3 * 8;
+            blocks["img_mod.1"]    = std::make_shared<Linear>(dim, 6 * dim, true);
+            blocks["txt_mod.1"]    = std::make_shared<Linear>(dim, 6 * dim, true);
+            blocks["img_norm1"]    = std::make_shared<RMSNorm>(dim, eps);
+            blocks["img_norm2"]    = std::make_shared<RMSNorm>(dim, eps);
+            blocks["txt_norm1"]    = std::make_shared<RMSNorm>(dim, eps);
+            blocks["txt_norm2"]    = std::make_shared<RMSNorm>(dim, eps);
+            blocks["img_mlp"]      = std::make_shared<LensGateMLP>(dim, mlp_hidden_dim);
+            blocks["txt_mlp"]      = std::make_shared<LensGateMLP>(dim, mlp_hidden_dim);
+            blocks["attn"]         = std::make_shared<LensJointAttention>(dim, attention_head_dim, num_attention_heads);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* img,
+                                                      ggml_tensor* txt,
+                                                      ggml_tensor* t_emb,
+                                                      ggml_tensor* pe) {
+            auto img_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["img_mod.1"]);
+            auto txt_mod_1 = std::dynamic_pointer_cast<Linear>(blocks["txt_mod.1"]);
+            auto img_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["img_norm1"]);
+            auto img_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["img_norm2"]);
+            auto txt_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm1"]);
+            auto txt_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm2"]);
+            auto img_mlp   = std::dynamic_pointer_cast<LensGateMLP>(blocks["img_mlp"]);
+            auto txt_mlp   = std::dynamic_pointer_cast<LensGateMLP>(blocks["txt_mlp"]);
+            auto attn      = std::dynamic_pointer_cast<LensJointAttention>(blocks["attn"]);
+
+            auto temb = ggml_silu(ctx->ggml_ctx, t_emb);
+
+            auto img_mod_params = img_mod_1->forward(ctx, temb);
+            auto img_mod_vec    = ggml_ext_chunk(ctx->ggml_ctx, img_mod_params, 6, 0);
+            auto txt_mod_params = txt_mod_1->forward(ctx, temb);
+            auto txt_mod_vec    = ggml_ext_chunk(ctx->ggml_ctx, txt_mod_params, 6, 0);
+
+            auto img_normed    = img_norm1->forward(ctx, img);
+            auto img_modulated = Flux::modulate(ctx->ggml_ctx, img_normed, img_mod_vec[0], img_mod_vec[1]);
+            auto txt_normed    = txt_norm1->forward(ctx, txt);
+            auto txt_modulated = Flux::modulate(ctx->ggml_ctx, txt_normed, txt_mod_vec[0], txt_mod_vec[1]);
+
+            auto [img_attn_output, txt_attn_output] = attn->forward(ctx, img_modulated, txt_modulated, pe);
+
+            img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_attn_output, img_mod_vec[2]));
+            txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_attn_output, txt_mod_vec[2]));
+
+            auto img_normed2    = img_norm2->forward(ctx, img);
+            auto img_modulated2 = Flux::modulate(ctx->ggml_ctx, img_normed2, img_mod_vec[3], img_mod_vec[4]);
+            auto txt_normed2    = txt_norm2->forward(ctx, txt);
+            auto txt_modulated2 = Flux::modulate(ctx->ggml_ctx, txt_normed2, txt_mod_vec[3], txt_mod_vec[4]);
+
+            img = ggml_add(ctx->ggml_ctx, img, ggml_mul(ctx->ggml_ctx, img_mlp->forward(ctx, img_modulated2), img_mod_vec[5]));
+            txt = ggml_add(ctx->ggml_ctx, txt, ggml_mul(ctx->ggml_ctx, txt_mlp->forward(ctx, txt_modulated2), txt_mod_vec[5]));
+            return {img, txt};
+        }
+    };
+
+    struct LensAdaLayerNormContinuous : public GGMLBlock {
+        int64_t hidden_size;
+        float eps;
+
+        LensAdaLayerNormContinuous(int64_t hidden_size, float eps = 1e-6f)
+            : hidden_size(hidden_size), eps(eps) {
+            blocks["linear"] = std::make_shared<Linear>(hidden_size, hidden_size * 2, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* conditioning) {
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            auto mods   = ggml_ext_chunk(ctx->ggml_ctx, linear->forward(ctx, ggml_silu(ctx->ggml_ctx, conditioning)), 2, 0);
+            auto scale  = mods[0];
+            auto shift  = mods[1];
+            x           = ggml_norm(ctx->ggml_ctx, x, eps);
+            return Flux::modulate(ctx->ggml_ctx, x, shift, scale);
+        }
+    };
+
+    class LensModel : public GGMLBlock {
+    public:
+        LensConfig config;
+
+        LensModel() = default;
+        LensModel(LensConfig config)
+            : config(config) {
+            int64_t inner_dim         = config.num_attention_heads * config.attention_head_dim;
+            blocks["time_text_embed"] = std::make_shared<LensTimestepProjEmbeddings>(inner_dim);
+            blocks["img_in"]          = std::make_shared<Linear>(config.in_channels, inner_dim, true);
+            blocks["txt_in"]          = std::make_shared<Linear>(config.joint_attention_dim * config.selected_layer_count, inner_dim, true);
+            for (int i = 0; i < config.selected_layer_count; ++i) {
+                blocks["txt_norm." + std::to_string(i)] = std::make_shared<RMSNorm>(config.joint_attention_dim, 1e-5f);
+            }
+            for (int i = 0; i < config.num_layers; ++i) {
+                blocks["transformer_blocks." + std::to_string(i)] = std::make_shared<LensTransformerBlock>(inner_dim,
+                                                                                                           config.num_attention_heads,
+                                                                                                           config.attention_head_dim);
+            }
+            blocks["norm_out"] = std::make_shared<LensAdaLayerNormContinuous>(inner_dim, 1e-6f);
+            blocks["proj_out"] = std::make_shared<Linear>(inner_dim, config.patch_size * config.patch_size * config.out_channels, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timestep,
+                             ggml_tensor* context,
+                             ggml_tensor* pe) {
+            GGML_ASSERT(context != nullptr);
+            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 time_text_embed = std::dynamic_pointer_cast<LensTimestepProjEmbeddings>(blocks["time_text_embed"]);
+            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<LensAdaLayerNormContinuous>(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 = ggml_reshape_3d(ctx->ggml_ctx, x, W * H, C, N);
+            img      = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, img, 1, 0, 2, 3));
+            img      = img_in->forward(ctx, img);
+
+            std::vector<ggml_tensor*> txt_chunks = ggml_ext_chunk(ctx->ggml_ctx, context, config.selected_layer_count, 0);
+            ggml_tensor* txt                     = nullptr;
+            for (int i = 0; i < config.selected_layer_count; ++i) {
+                auto txt_norm = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm." + std::to_string(i)]);
+                auto chunk    = txt_norm->forward(ctx, txt_chunks[i]);
+                txt           = txt == nullptr ? chunk : ggml_concat(ctx->ggml_ctx, txt, chunk, 0);
+            }
+            txt = txt_in->forward(ctx, txt);
+
+            sd::ggml_graph_cut::mark_graph_cut(img, "lens.prelude", "img");
+            sd::ggml_graph_cut::mark_graph_cut(txt, "lens.prelude", "txt");
+
+            for (int i = 0; i < config.num_layers; ++i) {
+                auto block = std::dynamic_pointer_cast<LensTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
+                auto out   = block->forward(ctx, img, txt, t_emb, pe);
+                img        = out.first;
+                txt        = out.second;
+                sd::ggml_graph_cut::mark_graph_cut(img, "lens.transformer_blocks." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "lens.transformer_blocks." + std::to_string(i), "txt");
+            }
+
+            img = norm_out->forward(ctx, img, t_emb);
+            img = proj_out->forward(ctx, img);
+
+            auto out = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, img, 1, 0, 2, 3));
+            out      = ggml_reshape_4d(ctx->ggml_ctx, out, W, H, config.patch_size * config.patch_size * config.out_channels, N);
+            return out;
+        }
+    };
+
+    struct LensRunner : public DiffusionModelRunner {
+        LensConfig config;
+        LensModel lens;
+        std::vector<float> pe_vec;
+
+        LensRunner(ggml_backend_t backend,
+                   const String2TensorStorage& tensor_storage_map      = {},
+                   const std::string prefix                            = "",
+                   std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(LensConfig::detect_from_weights(tensor_storage_map, prefix)) {
+            lens = LensModel(config);
+            lens.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "lens";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            lens.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor) {
+            ggml_cgraph* gf        = new_graph_custom(LENS_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timesteps = make_input(timesteps_tensor);
+            GGML_ASSERT(x->ne[3] == 1);
+            GGML_ASSERT(!context_tensor.empty());
+            ggml_tensor* context = make_input(context_tensor);
+
+            pe_vec      = Rope::gen_lens_pe(static_cast<int>(x->ne[1]),
+                                            static_cast<int>(x->ne[0]),
+                                            static_cast<int>(x->ne[3]),
+                                            static_cast<int>(context->ne[1]),
+                                            config.theta,
+                                            circular_y_enabled,
+                                            circular_x_enabled,
+                                            config.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
+            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
+            set_backend_tensor_data(pe, pe_vec.data());
+
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = lens.forward(&runner_ctx, x, timesteps, context, pe);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context));
+        }
+    };
+}  // namespace Lens
+
+#endif  // __SD_MODEL_DIFFUSION_LENS_HPP__

src/model/diffusion/mmdit.hpp

@@ -1,41 +1,137 @@
-#ifndef __MMDIT_HPP__
-#define __MMDIT_HPP__
+#ifndef __SD_MODEL_DIFFUSION_MMDIT_HPP__
+#define __SD_MODEL_DIFFUSION_MMDIT_HPP__
 
+#include <algorithm>
 #include <memory>
+#include <string>
+#include <vector>
 
-#include "ggml_extend.hpp"
-#include "model.h"
+#include "core/ggml_extend.hpp"
+#include "model/common/block.hpp"
+#include "model/diffusion/model.hpp"
+#include "model_loader.h"
 
 #define MMDIT_GRAPH_SIZE 10240
 
-struct Mlp : public GGMLBlock {
-public:
-    Mlp(int64_t in_features,
-        int64_t hidden_features = -1,
-        int64_t out_features    = -1,
-        bool bias               = true) {
-        // act_layer is always lambda: nn.GELU(approximate="tanh")
-        // norm_layer is always None
-        // use_conv is always False
-        if (hidden_features == -1) {
-            hidden_features = in_features;
-        }
-        if (out_features == -1) {
-            out_features = in_features;
-        }
-        blocks["fc1"] = std::shared_ptr<GGMLBlock>(new Linear(in_features, hidden_features, bias));
-        blocks["fc2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_features, out_features, bias));
-    }
+struct MMDiTConfig {
+    int64_t input_size               = -1;
+    int patch_size                   = 2;
+    int64_t in_channels              = 16;
+    int64_t d_self                   = -1;  // >=0 for MMdiT-X
+    int64_t depth                    = 24;
+    float mlp_ratio                  = 4.0f;
+    int64_t adm_in_channels          = 2048;
+    int64_t out_channels             = 16;
+    int64_t pos_embed_max_size       = 192;
+    int64_t num_patches              = 36864;  // 192 * 192
+    int64_t context_size             = 4096;
+    int64_t context_embedder_out_dim = 1536;
+    int64_t hidden_size              = 1536;
+    std::string qk_norm;
 
-    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-        // x: [N, n_token, in_features]
-        auto fc1 = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
-        auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
+    static MMDiTConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+        MMDiTConfig config;
+        bool has_weight_config = false;
+        bool has_pos_embed     = false;
+        bool has_hidden_size   = false;
+        bool has_context_embed = false;
 
-        x = fc1->forward(ctx, x);
-        x = ggml_ext_gelu(ctx->ggml_ctx, x, true);
-        x = fc2->forward(ctx, x);
-        return x;
+        for (const auto& [name, tensor_storage] : tensor_storage_map) {
+            if (!starts_with(name, prefix)) {
+                continue;
+            }
+
+            if (name.find("x_embedder.proj.weight") != std::string::npos && tensor_storage.n_dims == 4) {
+                has_weight_config  = true;
+                has_hidden_size    = true;
+                config.patch_size  = static_cast<int>(tensor_storage.ne[0]);
+                config.in_channels = tensor_storage.ne[2];
+                config.hidden_size = tensor_storage.ne[3];
+            } else if (name.find("t_embedder.mlp.0.weight") != std::string::npos && tensor_storage.n_dims == 2) {
+                has_weight_config  = true;
+                has_hidden_size    = true;
+                config.hidden_size = tensor_storage.ne[1];
+            } else if (name.find("y_embedder.mlp.0.weight") != std::string::npos && tensor_storage.n_dims == 2) {
+                has_weight_config      = true;
+                has_hidden_size        = true;
+                config.adm_in_channels = tensor_storage.ne[0];
+                config.hidden_size     = tensor_storage.ne[1];
+            } else if (name.find("context_embedder.weight") != std::string::npos && tensor_storage.n_dims == 2) {
+                has_weight_config               = true;
+                has_context_embed               = true;
+                config.context_size             = tensor_storage.ne[0];
+                config.context_embedder_out_dim = tensor_storage.ne[1];
+            } else if (name.find("final_layer.linear.weight") != std::string::npos && tensor_storage.n_dims == 2) {
+                has_weight_config  = true;
+                has_hidden_size    = true;
+                config.hidden_size = tensor_storage.ne[0];
+                int64_t patch_area = static_cast<int64_t>(config.patch_size) * config.patch_size;
+                if (patch_area > 0) {
+                    config.out_channels = tensor_storage.ne[1] / patch_area;
+                }
+            } else if (name.find("pos_embed") != std::string::npos && tensor_storage.n_dims == 3) {
+                has_weight_config  = true;
+                has_pos_embed      = true;
+                has_hidden_size    = true;
+                config.hidden_size = tensor_storage.ne[0];
+                config.num_patches = tensor_storage.ne[1];
+                for (int64_t size = 1; size * size <= config.num_patches; size++) {
+                    if (size * size == config.num_patches) {
+                        config.pos_embed_max_size = size;
+                        break;
+                    }
+                }
+            }
+
+            size_t jb = name.find("joint_blocks.");
+            if (jb == std::string::npos) {
+                continue;
+            }
+
+            has_weight_config      = true;
+            std::string block_name = name.substr(jb);
+            int64_t block_depth    = atoi(block_name.substr(13, block_name.find(".", 13)).c_str());
+            if (block_depth + 1 > config.depth) {
+                config.depth = block_depth + 1;
+            }
+            if (block_name.find("attn.ln") != std::string::npos) {
+                if (block_name.find(".bias") != std::string::npos) {
+                    config.qk_norm = "ln";
+                } else {
+                    config.qk_norm = "rms";
+                }
+            }
+            if (block_name.find("attn2") != std::string::npos) {
+                if (block_depth > config.d_self) {
+                    config.d_self = block_depth;
+                }
+            }
+        }
+
+        if (!has_pos_embed && config.d_self >= 0) {
+            config.pos_embed_max_size *= 2;
+            config.num_patches *= 4;
+        }
+        if (!has_hidden_size || config.hidden_size <= 0) {
+            config.hidden_size = 64 * config.depth;
+        }
+        if (!has_context_embed || config.context_embedder_out_dim <= 0) {
+            config.context_embedder_out_dim = config.hidden_size;
+        }
+
+        if (has_weight_config) {
+            LOG_DEBUG("mmdit: num_layers = %" PRId64 ", num_mmdit_x_layers = %" PRId64 ", hidden_size = %" PRId64 ", patch_size = %d, in_channels = %" PRId64 ", out_channels = %" PRId64 ", context_size = %" PRId64 ", adm_in_channels = %" PRId64 ", qk_norm = %s",
+                      config.depth,
+                      config.d_self + 1,
+                      config.hidden_size,
+                      config.patch_size,
+                      config.in_channels,
+                      config.out_channels,
+                      config.context_size,
+                      config.adm_in_channels,
+                      config.qk_norm.empty() ? "none" : config.qk_norm.c_str());
+        }
+        return config;
     }
 };
 
@@ -611,28 +707,16 @@ public:
 struct MMDiT : public GGMLBlock {
     // Diffusion model with a Transformer backbone.
 protected:
-    int64_t input_size               = -1;
-    int patch_size                   = 2;
-    int64_t in_channels              = 16;
-    int64_t d_self                   = -1;  // >=0 for MMdiT-X
-    int64_t depth                    = 24;
-    float mlp_ratio                  = 4.0f;
-    int64_t adm_in_channels          = 2048;
-    int64_t out_channels             = 16;
-    int64_t pos_embed_max_size       = 192;
-    int64_t num_patchs               = 36864;  // 192 * 192
-    int64_t context_size             = 4096;
-    int64_t context_embedder_out_dim = 1536;
-    int64_t hidden_size;
-    std::string qk_norm;
-
     void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, 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);
+        params["pos_embed"]  = ggml_new_tensor_3d(ctx, wtype, config.hidden_size, config.num_patches, 1);
     }
 
 public:
-    MMDiT(const String2TensorStorage& tensor_storage_map = {}) {
+    MMDiTConfig config;
+
+    explicit MMDiT(MMDiTConfig config = {})
+        : config(config) {
         // input_size is always None
         // learn_sigma is always False
         // register_length is alwalys 0
@@ -645,64 +729,30 @@ public:
         // pos_embed_offset is not used
         // context_embedder_config is always {'target': 'torch.nn.Linear', 'params': {'in_features': 4096, 'out_features': 1536}}
 
-        for (auto pair : tensor_storage_map) {
-            std::string tensor_name = pair.first;
-            if (tensor_name.find("model.diffusion_model.") == std::string::npos)
-                continue;
-            size_t jb = tensor_name.find("joint_blocks.");
-            if (jb != std::string::npos) {
-                tensor_name     = tensor_name.substr(jb);  // remove prefix
-                int block_depth = atoi(tensor_name.substr(13, tensor_name.find(".", 13)).c_str());
-                if (block_depth + 1 > depth) {
-                    depth = block_depth + 1;
-                }
-                if (tensor_name.find("attn.ln") != std::string::npos) {
-                    if (tensor_name.find(".bias") != std::string::npos) {
-                        qk_norm = "ln";
-                    } else {
-                        qk_norm = "rms";
-                    }
-                }
-                if (tensor_name.find("attn2") != std::string::npos) {
-                    if (block_depth > d_self) {
-                        d_self = block_depth;
-                    }
-                }
-            }
+        blocks["x_embedder"] = std::shared_ptr<GGMLBlock>(new PatchEmbed(config.input_size,
+                                                                         config.patch_size,
+                                                                         config.in_channels,
+                                                                         config.hidden_size,
+                                                                         true));
+        blocks["t_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedder(config.hidden_size));
+
+        if (config.adm_in_channels != -1) {
+            blocks["y_embedder"] = std::shared_ptr<GGMLBlock>(new VectorEmbedder(config.adm_in_channels, config.hidden_size));
         }
 
-        if (d_self >= 0) {
-            pos_embed_max_size *= 2;
-            num_patchs *= 4;
-        }
+        blocks["context_embedder"] = std::shared_ptr<GGMLBlock>(new Linear(config.context_size, config.context_embedder_out_dim, true, true));
 
-        LOG_INFO("MMDiT layers: %d (including %d MMDiT-x layers)", depth, d_self + 1);
-
-        int64_t default_out_channels = in_channels;
-        hidden_size                  = 64 * depth;
-        context_embedder_out_dim     = 64 * depth;
-        int64_t num_heads            = depth;
-
-        blocks["x_embedder"] = std::shared_ptr<GGMLBlock>(new PatchEmbed(input_size, patch_size, in_channels, hidden_size, true));
-        blocks["t_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedder(hidden_size));
-
-        if (adm_in_channels != -1) {
-            blocks["y_embedder"] = std::shared_ptr<GGMLBlock>(new VectorEmbedder(adm_in_channels, hidden_size));
-        }
-
-        blocks["context_embedder"] = std::shared_ptr<GGMLBlock>(new Linear(4096, context_embedder_out_dim, true, true));
-
-        for (int i = 0; i < depth; i++) {
-            blocks["joint_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new JointBlock(hidden_size,
-                                                                                                    num_heads,
-                                                                                                    mlp_ratio,
-                                                                                                    qk_norm,
+        for (int i = 0; i < config.depth; i++) {
+            blocks["joint_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new JointBlock(config.hidden_size,
+                                                                                                    config.depth,
+                                                                                                    config.mlp_ratio,
+                                                                                                    config.qk_norm,
                                                                                                     true,
-                                                                                                    i == depth - 1,
-                                                                                                    i <= d_self));
+                                                                                                    i == config.depth - 1,
+                                                                                                    i <= config.d_self));
         }
 
-        blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new FinalLayer(hidden_size, patch_size, out_channels));
+        blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new FinalLayer(config.hidden_size, config.patch_size, config.out_channels));
     }
 
     ggml_tensor*
@@ -711,22 +761,22 @@ public:
                       int64_t w) {
         auto pos_embed = params["pos_embed"];
 
-        h = (h + 1) / patch_size;
-        w = (w + 1) / patch_size;
+        h = (h + 1) / config.patch_size;
+        w = (w + 1) / config.patch_size;
 
-        GGML_ASSERT(h <= pos_embed_max_size && h > 0);
-        GGML_ASSERT(w <= pos_embed_max_size && w > 0);
+        GGML_ASSERT(h <= config.pos_embed_max_size && h > 0);
+        GGML_ASSERT(w <= config.pos_embed_max_size && w > 0);
 
-        int64_t top  = (pos_embed_max_size - h) / 2;
-        int64_t left = (pos_embed_max_size - w) / 2;
+        int64_t top  = (config.pos_embed_max_size - h) / 2;
+        int64_t left = (config.pos_embed_max_size - w) / 2;
 
-        auto spatial_pos_embed = ggml_reshape_3d(ctx, pos_embed, hidden_size, pos_embed_max_size, pos_embed_max_size);
+        auto spatial_pos_embed = ggml_reshape_3d(ctx, pos_embed, config.hidden_size, config.pos_embed_max_size, config.pos_embed_max_size);
 
         // spatial_pos_embed = spatial_pos_embed[:, top : top + h, left : left + w, :]
         spatial_pos_embed = ggml_view_3d(ctx,
                                          spatial_pos_embed,
-                                         hidden_size,
-                                         pos_embed_max_size,
+                                         config.hidden_size,
+                                         config.pos_embed_max_size,
                                          h,
                                          spatial_pos_embed->nb[1],
                                          spatial_pos_embed->nb[2],
@@ -734,14 +784,14 @@ public:
         spatial_pos_embed = ggml_cont(ctx, ggml_permute(ctx, spatial_pos_embed, 0, 2, 1, 3));  // [pos_embed_max_size, h, hidden_size]
         spatial_pos_embed = ggml_view_3d(ctx,
                                          spatial_pos_embed,
-                                         hidden_size,
+                                         config.hidden_size,
                                          h,
                                          w,
                                          spatial_pos_embed->nb[1],
                                          spatial_pos_embed->nb[2],
-                                         spatial_pos_embed->nb[2] * left);                     // [w, h, hidden_size]
-        spatial_pos_embed = ggml_cont(ctx, ggml_permute(ctx, spatial_pos_embed, 0, 2, 1, 3));  // [h, w, hidden_size]
-        spatial_pos_embed = ggml_reshape_3d(ctx, spatial_pos_embed, hidden_size, h * w, 1);    // [1, h*w, hidden_size]
+                                         spatial_pos_embed->nb[2] * left);                          // [w, h, hidden_size]
+        spatial_pos_embed = ggml_cont(ctx, ggml_permute(ctx, spatial_pos_embed, 0, 2, 1, 3));       // [h, w, hidden_size]
+        spatial_pos_embed = ggml_reshape_3d(ctx, spatial_pos_embed, config.hidden_size, h * w, 1);  // [1, h*w, hidden_size]
         return spatial_pos_embed;
     }
 
@@ -756,7 +806,7 @@ public:
         // return: [N, N*W, patch_size * patch_size * out_channels]
         auto final_layer = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
 
-        for (int i = 0; i < depth; i++) {
+        for (int i = 0; i < config.depth; i++) {
             // skip iteration if i is in skip_layers
             if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
                 continue;
@@ -799,7 +849,7 @@ public:
         x                = ggml_add(ctx->ggml_ctx, patch_embed, pos_embed);  // [N, H*W, hidden_size]
 
         auto c = t_embedder->forward(ctx, t);  // [N, hidden_size]
-        if (y != nullptr && adm_in_channels != -1) {
+        if (y != nullptr && config.adm_in_channels != -1) {
             auto y_embedder = std::dynamic_pointer_cast<VectorEmbedder>(blocks["y_embedder"]);
 
             y = y_embedder->forward(ctx, y);  // [N, hidden_size]
@@ -819,19 +869,22 @@ public:
 
         x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 
-        x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, patch_size, patch_size, /*patch_last*/ false);  // [N, C, H, W]
+        x = DiT::unpatchify_and_crop(ctx->ggml_ctx, x, H, W, config.patch_size, config.patch_size, /*patch_last*/ false);  // [N, C, H, W]
 
         return x;
     }
 };
-struct MMDiTRunner : public GGMLRunner {
+struct MMDiTRunner : public DiffusionModelRunner {
+    MMDiTConfig config;
     MMDiT mmdit;
 
     MMDiTRunner(ggml_backend_t backend,
-                ggml_backend_t params_backend,
-                const String2TensorStorage& tensor_storage_map = {},
-                const std::string prefix                       = "")
-        : GGMLRunner(backend, params_backend), mmdit(tensor_storage_map) {
+                const String2TensorStorage& tensor_storage_map      = {},
+                const std::string prefix                            = "",
+                std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : DiffusionModelRunner(backend, prefix, weight_manager),
+          config(MMDiTConfig::detect_from_weights(tensor_storage_map, prefix)),
+          mmdit(config) {
         mmdit.init(params_ctx, tensor_storage_map, prefix);
     }
 
@@ -839,7 +892,7 @@ struct MMDiTRunner : public GGMLRunner {
         return "mmdit";
     }
 
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
         mmdit.get_param_tensors(tensors, prefix);
     }
 
@@ -882,7 +935,21 @@ struct MMDiTRunner : public GGMLRunner {
             return build_graph(x, timesteps, context, y, skip_layers);
         };
 
-        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+    }
+
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        const auto* extra = diffusion_extra_as<SkipLayerDiffusionExtra>(diffusion_params);
+        static const std::vector<int> empty_skip_layers;
+        return compute(n_threads,
+                       *diffusion_params.x,
+                       *diffusion_params.timesteps,
+                       tensor_or_empty(diffusion_params.context),
+                       tensor_or_empty(diffusion_params.y),
+                       extra->skip_layers ? *extra->skip_layers : empty_skip_layers);
     }
 
     void test() {
@@ -932,26 +999,27 @@ struct MMDiTRunner : public GGMLRunner {
 
     static void load_from_file_and_test(const std::string& file_path) {
         // ggml_backend_t backend    = ggml_backend_cuda_init(0);
-        ggml_backend_t backend             = ggml_backend_cpu_init();
+        ggml_backend_t backend             = sd_backend_cpu_init();
         ggml_type model_data_type          = GGML_TYPE_F16;
-        std::shared_ptr<MMDiTRunner> mmdit = std::make_shared<MMDiTRunner>(backend, backend);
+        auto model_manager                 = std::make_shared<ModelManager>();
+        std::shared_ptr<MMDiTRunner> mmdit = std::make_shared<MMDiTRunner>(backend, String2TensorStorage{}, "", model_manager);
         {
             LOG_INFO("loading from '%s'", file_path.c_str());
 
-            mmdit->alloc_params_buffer();
-            std::map<std::string, ggml_tensor*> tensors;
-            mmdit->get_param_tensors(tensors, "model.diffusion_model");
-
-            ModelLoader model_loader;
+            ModelLoader& model_loader = model_manager->loader();
             if (!model_loader.init_from_file_and_convert_name(file_path)) {
                 LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
                 return;
             }
 
-            bool success = model_loader.load_tensors(tensors);
-
-            if (!success) {
-                LOG_ERROR("load tensors from model loader failed");
+            if (!model_manager->register_runner_params("MMDiT test",
+                                                       *mmdit,
+                                                       "model.diffusion_model",
+                                                       ModelManager::ResidencyMode::ParamBackend,
+                                                       backend,
+                                                       backend) ||
+                !model_manager->validate_registered_tensors()) {
+                LOG_ERROR("register mmdit tensors with model manager failed");
                 return;
             }
 
@@ -961,4 +1029,4 @@ struct MMDiTRunner : public GGMLRunner {
     }
 };
 
-#endif
+#endif  // __SD_MODEL_DIFFUSION_MMDIT_HPP__

src/model/diffusion/model.hpp

@@ -0,0 +1,108 @@
+#ifndef __SD_MODEL_DIFFUSION_MODEL_HPP__
+#define __SD_MODEL_DIFFUSION_MODEL_HPP__
+
+#include <string>
+#include <utility>
+#include <variant>
+
+#include "core/ggml_extend.hpp"
+#include "core/tensor_ggml.hpp"
+#include "model_manager.h"
+
+struct UNetDiffusionExtra {
+    int num_video_frames                           = -1;
+    const std::vector<sd::Tensor<float>>* controls = nullptr;
+    float control_strength                         = 0.f;
+};
+
+struct SkipLayerDiffusionExtra {
+    const std::vector<int>* skip_layers = nullptr;
+};
+
+struct FluxDiffusionExtra {
+    const sd::Tensor<float>* guidance   = nullptr;
+    const std::vector<int>* skip_layers = nullptr;
+};
+
+struct AnimaDiffusionExtra {
+    const sd::Tensor<int32_t>* t5_ids   = nullptr;
+    const sd::Tensor<float>* t5_weights = nullptr;
+};
+
+struct WanDiffusionExtra {
+    const sd::Tensor<float>* vace_context = nullptr;
+    float vace_strength                   = 1.f;
+};
+
+struct HiDreamO1DiffusionExtra {
+    const sd::Tensor<int32_t>* input_ids                               = nullptr;
+    const sd::Tensor<int32_t>* input_pos                               = nullptr;
+    const sd::Tensor<int32_t>* token_types                             = nullptr;
+    const sd::Tensor<int32_t>* vinput_mask                             = nullptr;
+    const std::vector<std::pair<int, sd::Tensor<float>>>* image_embeds = nullptr;
+};
+
+struct LTXAVDiffusionExtra {
+    const sd::Tensor<float>* audio_x         = nullptr;
+    const sd::Tensor<float>* audio_timesteps = nullptr;
+    int audio_length                         = 0;
+    float frame_rate                         = 24.f;
+    const sd::Tensor<float>* video_positions = nullptr;
+};
+
+using DiffusionExtraParams = std::variant<std::monostate,
+                                          UNetDiffusionExtra,
+                                          SkipLayerDiffusionExtra,
+                                          FluxDiffusionExtra,
+                                          AnimaDiffusionExtra,
+                                          WanDiffusionExtra,
+                                          HiDreamO1DiffusionExtra,
+                                          LTXAVDiffusionExtra>;
+
+struct DiffusionParams {
+    const sd::Tensor<float>* x                        = nullptr;
+    const sd::Tensor<float>* timesteps                = nullptr;
+    const sd::Tensor<float>* context                  = nullptr;
+    const sd::Tensor<float>* c_concat                 = nullptr;
+    const sd::Tensor<float>* y                        = nullptr;
+    const std::vector<sd::Tensor<float>>* ref_latents = nullptr;
+    bool increase_ref_index                           = false;
+    DiffusionExtraParams extra                        = std::monostate{};
+};
+
+template <typename T>
+static inline const T* diffusion_extra_as(const DiffusionParams& params) {
+    const auto* extra = std::get_if<T>(&params.extra);
+    GGML_ASSERT(extra != nullptr);
+    return extra;
+}
+
+template <typename T>
+static inline const sd::Tensor<T>& tensor_or_empty(const sd::Tensor<T>* tensor) {
+    static const sd::Tensor<T> kEmpty;
+    return tensor != nullptr ? *tensor : kEmpty;
+}
+
+struct DiffusionModelRunner : public GGMLRunner {
+protected:
+    std::string prefix;
+
+public:
+    DiffusionModelRunner(ggml_backend_t backend,
+                         const std::string& prefix,
+                         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : GGMLRunner(backend, weight_manager),
+          prefix(prefix) {}
+
+    virtual sd::Tensor<float> compute(int n_threads,
+                                      const DiffusionParams& diffusion_params) = 0;
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {
+        get_param_tensors(tensors, prefix);
+    }
+
+    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                                   const std::string& prefix) = 0;
+};
+
+#endif  // __SD_MODEL_DIFFUSION_MODEL_HPP__

src/model/diffusion/pid.hpp

@@ -0,0 +1,847 @@
+#ifndef __SD_MODEL_DIFFUSION_PID_HPP__
+#define __SD_MODEL_DIFFUSION_PID_HPP__
+
+#include <cmath>
+#include <cstdlib>
+#include <memory>
+#include <string>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "model/common/rope.hpp"
+#include "model/diffusion/dit.hpp"
+#include "model/diffusion/mmdit.hpp"
+
+namespace Pid {
+    constexpr int PID_GRAPH_SIZE = 196608;
+    constexpr float PID_PI       = 3.14159265358979323846f;
+
+    struct PixelDiTConfig {
+        int64_t in_channels            = 3;
+        int64_t hidden_size            = 1536;
+        int64_t num_groups             = 24;
+        int64_t patch_mlp_hidden_dim   = 4096;
+        int64_t pixel_hidden_size      = 16;
+        int64_t pixel_attn_hidden_size = 1152;
+        int64_t pixel_num_groups       = 16;
+        int64_t patch_depth            = 14;
+        int64_t pixel_depth            = 2;
+        int64_t patch_size             = 16;
+        int64_t txt_embed_dim          = 2304;
+        int64_t txt_max_length         = 300;
+        float text_rope_theta          = 10000.f;
+        int64_t lq_latent_channels     = 16;
+        int64_t lq_hidden_dim          = 512;
+        int64_t lq_num_res_blocks      = 4;
+        int64_t lq_interval            = 2;
+        int64_t lq_sr_scale            = 4;
+        int64_t lq_latent_down_factor  = 8;
+        int64_t rope_ref_grid_h        = 64;
+        int64_t rope_ref_grid_w        = 64;
+
+        static PixelDiTConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            PixelDiTConfig config;
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                size_t pos = name.find("patch_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index    = atoi(items[1].c_str());
+                        config.patch_depth = std::max<int64_t>(config.patch_depth, block_index + 1);
+                    }
+                }
+                pos = name.find("pixel_blocks.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index    = atoi(items[1].c_str());
+                        config.pixel_depth = std::max<int64_t>(config.pixel_depth, block_index + 1);
+                    }
+                }
+                if (name.find("lq_proj.latent_proj.0.weight") != std::string::npos) {
+                    config.lq_latent_channels    = tensor_storage.ne[2];
+                    config.lq_latent_down_factor = config.lq_latent_channels >= 64 ? 16 : 8;
+                }
+                if (name.find("patch_blocks.0.mlp_x.w1.weight") != std::string::npos) {
+                    config.patch_mlp_hidden_dim = tensor_storage.ne[1];
+                }
+            }
+            LOG_DEBUG("pid: patch_depth = %" PRId64 ", pixel_depth = %" PRId64 ", patch_mlp_hidden_dim = %" PRId64 ", lq_latent_channels = %" PRId64 ", lq_latent_down_factor = %" PRId64,
+                      config.patch_depth,
+                      config.pixel_depth,
+                      config.patch_mlp_hidden_dim,
+                      config.lq_latent_channels,
+                      config.lq_latent_down_factor);
+            return config;
+        }
+    };
+
+    inline std::vector<float> make_rope_1d(int length,
+                                           int dim,
+                                           float theta) {
+        GGML_ASSERT(dim % 2 == 0);
+        return Rope::flatten(Rope::rope(Rope::linspace(0.f, static_cast<float>(length - 1), length), dim, theta));
+    }
+
+    inline std::vector<float> make_rope_2d(int height,
+                                           int width,
+                                           int dim,
+                                           float theta    = 10000.f,
+                                           float scale    = 16.f,
+                                           int ref_grid_h = 0,
+                                           int ref_grid_w = 0) {
+        GGML_ASSERT(dim % 4 == 0);
+        return Rope::embed_2d_interleaved(height, width, dim, theta, scale, ref_grid_h, ref_grid_w);
+    }
+
+    inline std::vector<float> make_pixel_abs_pos(int height,
+                                                 int width,
+                                                 int dim) {
+        GGML_ASSERT(dim % 4 == 0);
+        int half_dim = dim / 2;
+        std::vector<float> x_pos;
+        std::vector<float> y_pos;
+        x_pos.reserve(static_cast<size_t>(height) * width);
+        y_pos.reserve(static_cast<size_t>(height) * width);
+        for (int iy = 0; iy < height; ++iy) {
+            for (int ix = 0; ix < width; ++ix) {
+                x_pos.push_back(static_cast<float>(ix));
+                y_pos.push_back(static_cast<float>(iy));
+            }
+        }
+
+        auto x_emb = timestep_embedding(x_pos, half_dim, 10000, false);
+        auto y_emb = timestep_embedding(y_pos, half_dim, 10000, false);
+
+        std::vector<float> out(static_cast<size_t>(dim) * height * width);
+        for (int pos = 0; pos < height * width; ++pos) {
+            size_t out_base = static_cast<size_t>(pos) * dim;
+            size_t emb_base = static_cast<size_t>(pos) * half_dim;
+            for (int i = 0; i < half_dim; ++i) {
+                out[out_base + i]            = x_emb[emb_base + i];
+                out[out_base + half_dim + i] = y_emb[emb_base + i];
+            }
+        }
+        return out;
+    }
+
+    inline ggml_tensor* apply_adaln(ggml_context* ctx,
+                                    ggml_tensor* x,
+                                    ggml_tensor* shift,
+                                    ggml_tensor* scale) {
+        return ggml_add(ctx, ggml_add(ctx, x, ggml_mul(ctx, x, scale)), shift);
+    }
+
+    struct PatchTokenEmbedder : public GGMLBlock {
+        bool use_rms_norm;
+
+        PatchTokenEmbedder(int64_t in_chans,
+                           int64_t embed_dim,
+                           bool use_rms_norm = false,
+                           bool bias         = true)
+            : use_rms_norm(use_rms_norm) {
+            blocks["proj"] = std::make_shared<Linear>(in_chans, embed_dim, bias);
+            if (use_rms_norm) {
+                blocks["norm"] = std::make_shared<RMSNorm>(embed_dim, 1e-6f);
+            }
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+            x         = proj->forward(ctx, x);
+            if (use_rms_norm) {
+                auto norm = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+                x         = norm->forward(ctx, x);
+            }
+            return x;
+        }
+    };
+
+    struct PixelDiTTimestepEmbedder : public GGMLBlock {
+        int frequency_embedding_size;
+
+        PixelDiTTimestepEmbedder(int64_t hidden_size,
+                                 int frequency_embedding_size = 256)
+            : frequency_embedding_size(frequency_embedding_size) {
+            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true, true);
+            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, true, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t) {
+            auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
+            auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
+            auto t_emb = ggml_ext_timestep_embedding(ctx->ggml_ctx, t, frequency_embedding_size, 10);
+            t_emb      = mlp_0->forward(ctx, t_emb);
+            t_emb      = ggml_silu_inplace(ctx->ggml_ctx, t_emb);
+            return mlp_2->forward(ctx, t_emb);
+        }
+    };
+
+    struct FeedForward : public GGMLBlock {
+        FeedForward(int64_t dim, int64_t hidden_dim) {
+            blocks["w1"] = std::make_shared<Linear>(dim, hidden_dim, false);
+            blocks["w2"] = std::make_shared<Linear>(hidden_dim, dim, false);
+            blocks["w3"] = std::make_shared<Linear>(dim, hidden_dim, false);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto w1 = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
+            auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
+            auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
+            auto h  = ggml_silu_inplace(ctx->ggml_ctx, w1->forward(ctx, x));
+            h       = ggml_mul_inplace(ctx->ggml_ctx, h, w3->forward(ctx, x));
+            return w2->forward(ctx, h);
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+        FinalLayer(int64_t hidden_size, int64_t out_channels) {
+            blocks["norm"]   = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["linear"] = std::make_shared<Linear>(hidden_size, out_channels, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            return linear->forward(ctx, norm->forward(ctx, x));
+        }
+    };
+
+    struct RotaryAttention : public GGMLBlock {
+        int64_t dim;
+        int64_t num_heads;
+
+        RotaryAttention(int64_t dim, int64_t num_heads)
+            : dim(dim), num_heads(num_heads) {
+            int64_t head_dim = dim / num_heads;
+            blocks["qkv"]    = std::make_shared<Linear>(dim, dim * 3, false);
+            blocks["q_norm"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["proj"]   = std::make_shared<Linear>(dim, dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* pos) {
+            auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
+            auto q_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
+            auto k_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
+            auto proj     = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+
+            auto qkv         = qkv_proj->forward(ctx, x);
+            auto qkv_vec     = split_qkv(ctx->ggml_ctx, qkv);
+            int64_t L        = x->ne[1];
+            int64_t N        = x->ne[2];
+            int64_t head_dim = dim / num_heads;
+            auto q           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[0], head_dim, num_heads, L, N);
+            auto k           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[1], head_dim, num_heads, L, N);
+            auto v           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[2], head_dim, num_heads, L, N);
+            q                = q_norm->forward(ctx, q);
+            k                = k_norm->forward(ctx, k);
+            x                = Rope::attention(ctx, q, k, v, pos, nullptr, 1.0f / 128.f, true);
+            return proj->forward(ctx, x);
+        }
+    };
+
+    struct MMDiTJointAttention : public GGMLBlock {
+        int64_t dim;
+        int64_t num_heads;
+
+        MMDiTJointAttention(int64_t dim, int64_t num_heads)
+            : dim(dim), num_heads(num_heads) {
+            int64_t head_dim   = dim / num_heads;
+            blocks["qkv_x"]    = std::make_shared<Linear>(dim, dim * 3, false);
+            blocks["qkv_y"]    = std::make_shared<Linear>(dim, dim * 3, false);
+            blocks["q_norm_x"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm_x"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["q_norm_y"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["k_norm_y"] = std::make_shared<RMSNorm>(head_dim, 1e-6f);
+            blocks["proj_x"]   = std::make_shared<Linear>(dim, dim, true);
+            blocks["proj_y"]   = std::make_shared<Linear>(dim, dim, true);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* x,
+                                                      ggml_tensor* y,
+                                                      ggml_tensor* pos_img,
+                                                      ggml_tensor* pos_txt) {
+            auto qkv_x_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv_x"]);
+            auto qkv_y_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv_y"]);
+            auto q_norm_x   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm_x"]);
+            auto k_norm_x   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm_x"]);
+            auto q_norm_y   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm_y"]);
+            auto k_norm_y   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm_y"]);
+            auto proj_x     = std::dynamic_pointer_cast<Linear>(blocks["proj_x"]);
+            auto proj_y     = std::dynamic_pointer_cast<Linear>(blocks["proj_y"]);
+
+            int64_t Nx       = x->ne[1];
+            int64_t Ny       = y->ne[1];
+            int64_t N        = x->ne[2];
+            int64_t head_dim = dim / num_heads;
+
+            auto qkv_x = split_qkv(ctx->ggml_ctx, qkv_x_proj->forward(ctx, x));
+            auto qx    = ggml_reshape_4d(ctx->ggml_ctx, qkv_x[0], head_dim, num_heads, Nx, N);
+            auto kx    = ggml_reshape_4d(ctx->ggml_ctx, qkv_x[1], head_dim, num_heads, Nx, N);
+            auto vx    = ggml_reshape_4d(ctx->ggml_ctx, qkv_x[2], head_dim, num_heads, Nx, N);
+            qx         = q_norm_x->forward(ctx, qx);
+            kx         = k_norm_x->forward(ctx, kx);
+
+            auto qkv_y = split_qkv(ctx->ggml_ctx, qkv_y_proj->forward(ctx, y));
+            auto qy    = ggml_reshape_4d(ctx->ggml_ctx, qkv_y[0], head_dim, num_heads, Ny, N);
+            auto ky    = ggml_reshape_4d(ctx->ggml_ctx, qkv_y[1], head_dim, num_heads, Ny, N);
+            auto vy    = ggml_reshape_4d(ctx->ggml_ctx, qkv_y[2], head_dim, num_heads, Ny, N);
+            qy         = q_norm_y->forward(ctx, qy);
+            ky         = k_norm_y->forward(ctx, ky);
+
+            auto q_joint   = ggml_concat(ctx->ggml_ctx, qy, qx, 2);
+            auto k_joint   = ggml_concat(ctx->ggml_ctx, ky, kx, 2);
+            auto v_joint   = ggml_concat(ctx->ggml_ctx, vy, vx, 2);
+            auto pos_joint = ggml_concat(ctx->ggml_ctx, pos_txt, pos_img, 3);
+            auto out       = Rope::attention(ctx, q_joint, k_joint, v_joint, pos_joint, nullptr, 1.0f, true);
+
+            auto out_y = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, Ny);
+            auto out_x = ggml_ext_slice(ctx->ggml_ctx, out, 1, Ny, Ny + Nx);
+            return {proj_x->forward(ctx, out_x), proj_y->forward(ctx, out_y)};
+        }
+    };
+
+    struct MMDiTBlockT2I : public GGMLBlock {
+        int64_t hidden_size;
+
+        MMDiTBlockT2I(int64_t hidden_size, int64_t groups, int64_t mlp_hidden_dim)
+            : hidden_size(hidden_size) {
+            blocks["norm_x1"]                = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["norm_y1"]                = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["attn"]                   = std::make_shared<MMDiTJointAttention>(hidden_size, groups);
+            blocks["norm_x2"]                = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["norm_y2"]                = std::make_shared<RMSNorm>(hidden_size, 1e-6f);
+            blocks["mlp_x"]                  = std::make_shared<FeedForward>(hidden_size, mlp_hidden_dim);
+            blocks["mlp_y"]                  = std::make_shared<FeedForward>(hidden_size, mlp_hidden_dim);
+            blocks["adaLN_modulation_img.0"] = std::make_shared<Linear>(hidden_size, 6 * hidden_size, true);
+            blocks["adaLN_modulation_txt.0"] = std::make_shared<Linear>(hidden_size, 6 * hidden_size, true);
+        }
+
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                                      ggml_tensor* x,
+                                                      ggml_tensor* y,
+                                                      ggml_tensor* c,
+                                                      ggml_tensor* pos_img,
+                                                      ggml_tensor* pos_txt) {
+            auto norm_x1 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_x1"]);
+            auto norm_y1 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_y1"]);
+            auto attn    = std::dynamic_pointer_cast<MMDiTJointAttention>(blocks["attn"]);
+            auto norm_x2 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_x2"]);
+            auto norm_y2 = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_y2"]);
+            auto mlp_x   = std::dynamic_pointer_cast<FeedForward>(blocks["mlp_x"]);
+            auto mlp_y   = std::dynamic_pointer_cast<FeedForward>(blocks["mlp_y"]);
+            auto ada_img = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation_img.0"]);
+            auto ada_txt = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation_txt.0"]);
+
+            auto mx = ggml_ext_chunk(ctx->ggml_ctx, ada_img->forward(ctx, c), 6, 0);
+            auto my = ggml_ext_chunk(ctx->ggml_ctx, ada_txt->forward(ctx, c), 6, 0);
+
+            auto x_norm   = apply_adaln(ctx->ggml_ctx, norm_x1->forward(ctx, x), mx[0], mx[1]);
+            auto y_norm   = apply_adaln(ctx->ggml_ctx, norm_y1->forward(ctx, y), my[0], my[1]);
+            auto attn_out = attn->forward(ctx, x_norm, y_norm, pos_img, pos_txt);
+
+            x = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_out.first, mx[2]));
+            y = ggml_add(ctx->ggml_ctx, y, ggml_mul(ctx->ggml_ctx, attn_out.second, my[2]));
+
+            auto x_mlp = mlp_x->forward(ctx, apply_adaln(ctx->ggml_ctx, norm_x2->forward(ctx, x), mx[3], mx[4]));
+            auto y_mlp = mlp_y->forward(ctx, apply_adaln(ctx->ggml_ctx, norm_y2->forward(ctx, y), my[3], my[4]));
+            x          = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, x_mlp, mx[5]));
+            y          = ggml_add(ctx->ggml_ctx, y, ggml_mul(ctx->ggml_ctx, y_mlp, my[5]));
+            return {x, y};
+        }
+    };
+
+    struct PixelTokenEmbedder : public GGMLBlock {
+        int64_t in_channels;
+        int64_t hidden_size_output;
+
+        PixelTokenEmbedder(int64_t in_channels, int64_t hidden_size_output)
+            : in_channels(in_channels), hidden_size_output(hidden_size_output) {
+            blocks["proj"] = std::make_shared<Linear>(in_channels, hidden_size_output, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* inputs,
+                             int64_t patch_size,
+                             ggml_tensor* pos_full) {
+            auto proj  = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
+            int64_t W  = inputs->ne[0];
+            int64_t H  = inputs->ne[1];
+            int64_t B  = inputs->ne[3];
+            int64_t L  = (W / patch_size) * (H / patch_size);
+            int64_t P2 = patch_size * patch_size;
+
+            auto x = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, inputs, 2, 0, 1, 3));
+            x      = ggml_reshape_3d(ctx->ggml_ctx, x, in_channels, W * H, B);
+            x      = proj->forward(ctx, x);
+            x      = ggml_add(ctx->ggml_ctx, x, pos_full);
+            x      = ggml_reshape_4d(ctx->ggml_ctx, x, hidden_size_output, W, H, B);
+            x      = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 1, 2, 0, 3));
+            x      = DiT::patchify(ctx->ggml_ctx, x, static_cast<int>(patch_size), static_cast<int>(patch_size), false);
+            x      = ggml_reshape_3d(ctx->ggml_ctx, x, hidden_size_output, P2, L * B);
+            return x;
+        }
+    };
+
+    struct PiTBlock : public GGMLBlock {
+        int64_t pixel_dim;
+        int64_t context_dim;
+        int64_t attn_dim;
+        int64_t num_heads;
+        int64_t patch_size;
+
+        PiTBlock(int64_t pixel_dim,
+                 int64_t context_dim,
+                 int64_t patch_size,
+                 int64_t attn_dim,
+                 int64_t num_heads)
+            : pixel_dim(pixel_dim),
+              context_dim(context_dim),
+              attn_dim(attn_dim),
+              num_heads(num_heads),
+              patch_size(patch_size) {
+            int64_t p2                   = patch_size * patch_size;
+            blocks["compress_to_attn"]   = std::make_shared<Linear>(p2 * pixel_dim, attn_dim, true);
+            blocks["expand_from_attn"]   = std::make_shared<Linear>(attn_dim, p2 * pixel_dim, true);
+            blocks["norm1"]              = std::make_shared<RMSNorm>(pixel_dim, 1e-6f);
+            blocks["attn"]               = std::make_shared<RotaryAttention>(attn_dim, num_heads);
+            blocks["norm2"]              = std::make_shared<RMSNorm>(pixel_dim, 1e-6f);
+            blocks["mlp"]                = std::make_shared<Mlp>(pixel_dim, pixel_dim * 4);
+            blocks["adaLN_modulation.0"] = std::make_shared<Linear>(context_dim, 6 * pixel_dim * p2, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* s_cond,
+                             int64_t image_height,
+                             int64_t image_width,
+                             ggml_tensor* pos_comp) {
+            auto compress = std::dynamic_pointer_cast<Linear>(blocks["compress_to_attn"]);
+            auto expand   = std::dynamic_pointer_cast<Linear>(blocks["expand_from_attn"]);
+            auto norm1    = std::dynamic_pointer_cast<RMSNorm>(blocks["norm1"]);
+            auto attn     = std::dynamic_pointer_cast<RotaryAttention>(blocks["attn"]);
+            auto norm2    = std::dynamic_pointer_cast<RMSNorm>(blocks["norm2"]);
+            auto mlp      = std::dynamic_pointer_cast<Mlp>(blocks["mlp"]);
+            auto ada      = std::dynamic_pointer_cast<Linear>(blocks["adaLN_modulation.0"]);
+
+            int64_t Hs = image_height / patch_size;
+            int64_t Ws = image_width / patch_size;
+            int64_t L  = Hs * Ws;
+            int64_t BL = x->ne[2];
+            int64_t B  = BL / L;
+            int64_t P2 = patch_size * patch_size;
+
+            auto ada_params = ada->forward(ctx, s_cond);
+            ada_params      = ggml_reshape_3d(ctx->ggml_ctx, ada_params, 6 * pixel_dim, P2, BL);
+            auto mod        = ggml_ext_chunk(ctx->ggml_ctx, ada_params, 6, 0);
+
+            auto x_norm    = apply_adaln(ctx->ggml_ctx, norm1->forward(ctx, x), mod[0], mod[1]);
+            auto x_flat    = ggml_reshape_2d(ctx->ggml_ctx, x_norm, P2 * pixel_dim, BL);
+            auto x_comp    = compress->forward(ctx, x_flat);
+            x_comp         = ggml_reshape_3d(ctx->ggml_ctx, x_comp, attn_dim, L, B);
+            auto attn_out  = attn->forward(ctx, x_comp, pos_comp);
+            auto attn_flat = expand->forward(ctx, ggml_reshape_2d(ctx->ggml_ctx, attn_out, attn_dim, BL));
+            auto attn_exp  = ggml_reshape_3d(ctx->ggml_ctx, attn_flat, pixel_dim, P2, BL);
+            x              = ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, attn_exp, mod[2]));
+
+            auto mlp_out = mlp->forward(ctx, apply_adaln(ctx->ggml_ctx, norm2->forward(ctx, x), mod[3], mod[4]));
+            return ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, mlp_out, mod[5]));
+        }
+    };
+
+    struct SigmaAwareGate : public GGMLBlock {
+        int64_t dim;
+
+        SigmaAwareGate(int64_t dim)
+            : dim(dim) {
+            blocks["content_proj"] = std::make_shared<Linear>(dim * 2, dim, true);
+        }
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         std::string prefix                             = "") override {
+            params["log_alpha"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 1);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* lq,
+                             ggml_tensor* sigma) {
+            auto content_proj = std::dynamic_pointer_cast<Linear>(blocks["content_proj"]);
+
+            auto content_logit = content_proj->forward(ctx, ggml_concat(ctx->ggml_ctx, x, lq, 0));
+            sigma              = ggml_reshape_3d(ctx->ggml_ctx, sigma, 1, 1, sigma->ne[0]);
+            auto alpha         = ggml_exp(ctx->ggml_ctx, params["log_alpha"]);
+            auto offset        = ggml_neg(ctx->ggml_ctx, ggml_mul(ctx->ggml_ctx, alpha, sigma));
+            auto gate          = ggml_sigmoid(ctx->ggml_ctx, ggml_add(ctx->ggml_ctx, content_logit, offset));
+            return ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, gate, lq));
+        }
+    };
+
+    struct PiDResBlock : public GGMLBlock {
+        PiDResBlock(int64_t channels) {
+            blocks["block.0"] = std::make_shared<GroupNorm>(4, channels, 1e-5f);
+            blocks["block.2"] = std::make_shared<Conv2d>(channels, channels, std::pair<int, int>{3, 3}, std::pair<int, int>{1, 1}, std::pair<int, int>{1, 1});
+            blocks["block.3"] = std::make_shared<GroupNorm>(4, channels, 1e-5f);
+            blocks["block.5"] = std::make_shared<Conv2d>(channels, channels, std::pair<int, int>{3, 3}, std::pair<int, int>{1, 1}, std::pair<int, int>{1, 1});
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto norm1 = std::dynamic_pointer_cast<GroupNorm>(blocks["block.0"]);
+            auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["block.2"]);
+            auto norm2 = std::dynamic_pointer_cast<GroupNorm>(blocks["block.3"]);
+            auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["block.5"]);
+            auto h     = ggml_silu_inplace(ctx->ggml_ctx, norm1->forward(ctx, x));
+            h          = conv1->forward(ctx, h);
+            h          = ggml_silu_inplace(ctx->ggml_ctx, norm2->forward(ctx, h));
+            h          = conv2->forward(ctx, h);
+            return ggml_add(ctx->ggml_ctx, x, h);
+        }
+    };
+
+    struct LQProjection2D : public GGMLBlock {
+        PixelDiTConfig config;
+
+        LQProjection2D(const PixelDiTConfig& config)
+            : config(config) {
+            blocks["latent_proj.0"] = std::make_shared<Conv2d>(config.lq_latent_channels, config.lq_hidden_dim, std::pair<int, int>{3, 3}, std::pair<int, int>{1, 1}, std::pair<int, int>{1, 1});
+            blocks["latent_proj.2"] = std::make_shared<Conv2d>(config.lq_hidden_dim, config.lq_hidden_dim, std::pair<int, int>{3, 3}, std::pair<int, int>{1, 1}, std::pair<int, int>{1, 1});
+            for (int i = 0; i < config.lq_num_res_blocks; ++i) {
+                blocks["latent_proj." + std::to_string(3 + i)] = std::make_shared<PiDResBlock>(config.lq_hidden_dim);
+            }
+
+            int num_outputs = static_cast<int>((config.patch_depth + config.lq_interval - 1) / config.lq_interval);
+            for (int i = 0; i < num_outputs; ++i) {
+                blocks["output_heads." + std::to_string(i)] = std::make_shared<Linear>(config.lq_hidden_dim, config.hidden_size, true);
+                blocks["gate_modules." + std::to_string(i)] = std::make_shared<SigmaAwareGate>(config.hidden_size);
+            }
+        }
+
+        bool is_gate_active(int block_idx) const {
+            return block_idx % config.lq_interval == 0;
+        }
+
+        int get_output_index(int block_idx) const {
+            return block_idx / static_cast<int>(config.lq_interval);
+        }
+
+        ggml_tensor* gate(GGMLRunnerContext* ctx,
+                          ggml_tensor* x,
+                          ggml_tensor* lq,
+                          ggml_tensor* sigma,
+                          int out_idx) {
+            auto gate_module = std::dynamic_pointer_cast<SigmaAwareGate>(blocks["gate_modules." + std::to_string(out_idx)]);
+            return gate_module->forward(ctx, x, lq, sigma);
+        }
+
+        std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx,
+                                          ggml_tensor* lq_latent,
+                                          int64_t target_pH,
+                                          int64_t target_pW) {
+            auto conv0             = std::dynamic_pointer_cast<Conv2d>(blocks["latent_proj.0"]);
+            auto conv2             = std::dynamic_pointer_cast<Conv2d>(blocks["latent_proj.2"]);
+            float z_to_patch_ratio = static_cast<float>(config.lq_sr_scale * config.lq_latent_down_factor) /
+                                     static_cast<float>(config.patch_size);
+            GGML_ASSERT(z_to_patch_ratio >= 1.0f);
+            if (lq_latent->ne[0] != target_pW || lq_latent->ne[1] != target_pH) {
+                lq_latent = ggml_interpolate(ctx->ggml_ctx,
+                                             lq_latent,
+                                             target_pW,
+                                             target_pH,
+                                             lq_latent->ne[2],
+                                             lq_latent->ne[3],
+                                             GGML_SCALE_MODE_NEAREST);
+            }
+
+            auto feat = conv0->forward(ctx, lq_latent);
+            feat      = ggml_silu_inplace(ctx->ggml_ctx, feat);
+            feat      = conv2->forward(ctx, feat);
+            for (int i = 0; i < config.lq_num_res_blocks; ++i) {
+                auto block = std::dynamic_pointer_cast<PiDResBlock>(blocks["latent_proj." + std::to_string(3 + i)]);
+                feat       = block->forward(ctx, feat);
+            }
+
+            int64_t B   = feat->ne[3];
+            int64_t C   = feat->ne[2];
+            int64_t L   = target_pH * target_pW;
+            auto tokens = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, feat, 2, 0, 1, 3));
+            tokens      = ggml_reshape_3d(ctx->ggml_ctx, tokens, C, L, B);
+
+            int num_outputs = static_cast<int>((config.patch_depth + config.lq_interval - 1) / config.lq_interval);
+            std::vector<ggml_tensor*> outputs;
+            outputs.reserve(num_outputs);
+            for (int i = 0; i < num_outputs; ++i) {
+                auto head = std::dynamic_pointer_cast<Linear>(blocks["output_heads." + std::to_string(i)]);
+                outputs.push_back(head->forward(ctx, tokens));
+            }
+            return outputs;
+        }
+    };
+
+    struct PixelDiT : public GGMLBlock {
+        PixelDiTConfig config;
+
+        PixelDiT() = default;
+
+        PixelDiT(const PixelDiTConfig& config)
+            : config(config) {
+            blocks["pixel_embedder"] = std::make_shared<PixelTokenEmbedder>(config.in_channels, config.pixel_hidden_size);
+            blocks["s_embedder"]     = std::make_shared<PatchTokenEmbedder>(config.in_channels * config.patch_size * config.patch_size, config.hidden_size, false, true);
+            blocks["t_embedder"]     = std::make_shared<PixelDiTTimestepEmbedder>(config.hidden_size);
+            blocks["y_embedder"]     = std::make_shared<PatchTokenEmbedder>(config.txt_embed_dim, config.hidden_size, true, true);
+            for (int i = 0; i < config.patch_depth; ++i) {
+                blocks["patch_blocks." + std::to_string(i)] = std::make_shared<MMDiTBlockT2I>(config.hidden_size, config.num_groups, config.patch_mlp_hidden_dim);
+            }
+            for (int i = 0; i < config.pixel_depth; ++i) {
+                blocks["pixel_blocks." + std::to_string(i)] = std::make_shared<PiTBlock>(config.pixel_hidden_size,
+                                                                                         config.hidden_size,
+                                                                                         config.patch_size,
+                                                                                         config.pixel_attn_hidden_size,
+                                                                                         config.pixel_num_groups);
+            }
+            blocks["final_layer"] = std::make_shared<FinalLayer>(config.pixel_hidden_size, config.in_channels);
+            blocks["lq_proj"]     = std::make_shared<LQProjection2D>(config);
+        }
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         std::string prefix                             = "") override {
+            params["y_pos_embedding"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, config.hidden_size, config.txt_max_length, 1);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
+                             ggml_tensor* x,
+                             ggml_tensor* timesteps,
+                             ggml_tensor* context,
+                             ggml_tensor* lq_latent,
+                             ggml_tensor* degrade_sigma,
+                             ggml_tensor* pos_img,
+                             ggml_tensor* pos_txt,
+                             ggml_tensor* pixel_pos_full,
+                             ggml_tensor* pixel_pos_comp) {
+            auto pixel_embedder = std::dynamic_pointer_cast<PixelTokenEmbedder>(blocks["pixel_embedder"]);
+            auto s_embedder     = std::dynamic_pointer_cast<PatchTokenEmbedder>(blocks["s_embedder"]);
+            auto t_embedder     = std::dynamic_pointer_cast<PixelDiTTimestepEmbedder>(blocks["t_embedder"]);
+            auto y_embedder     = std::dynamic_pointer_cast<PatchTokenEmbedder>(blocks["y_embedder"]);
+            auto final_layer    = std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer"]);
+            auto lq_proj        = std::dynamic_pointer_cast<LQProjection2D>(blocks["lq_proj"]);
+
+            int64_t W_orig = x->ne[0];
+            int64_t H_orig = x->ne[1];
+            x              = DiT::pad_to_patch_size(ctx, x, static_cast<int>(config.patch_size), static_cast<int>(config.patch_size));
+            int64_t W      = x->ne[0];
+            int64_t H      = x->ne[1];
+            int64_t B      = x->ne[3];
+            int64_t Hs     = H / config.patch_size;
+            int64_t Ws     = W / config.patch_size;
+            int64_t L      = Hs * Ws;
+            int64_t P2     = config.patch_size * config.patch_size;
+
+            auto x_patches = DiT::patchify(ctx->ggml_ctx, x, static_cast<int>(config.patch_size), static_cast<int>(config.patch_size), true);
+            auto t_emb     = t_embedder->forward(ctx, timesteps);
+            auto condition = ggml_silu(ctx->ggml_ctx, t_emb);
+
+            GGML_ASSERT(context != nullptr);
+            int64_t Ltxt = std::min<int64_t>(context->ne[1], config.txt_max_length);
+            auto y       = ggml_ext_slice(ctx->ggml_ctx, context, 1, 0, Ltxt);
+            auto y_emb   = y_embedder->forward(ctx, y);
+            auto y_pos   = ggml_ext_slice(ctx->ggml_ctx, params["y_pos_embedding"], 1, 0, Ltxt);
+            y_emb        = ggml_add(ctx->ggml_ctx, y_emb, y_pos);
+
+            std::vector<ggml_tensor*> lq_features = lq_proj->forward(ctx, lq_latent, Hs, Ws);
+
+            auto s = s_embedder->forward(ctx, x_patches);
+
+            for (int i = 0; i < config.patch_depth; ++i) {
+                if (lq_proj->is_gate_active(i)) {
+                    int out_idx = lq_proj->get_output_index(i);
+                    if (out_idx < static_cast<int>(lq_features.size())) {
+                        s = lq_proj->gate(ctx, s, lq_features[out_idx], degrade_sigma, out_idx);
+                    }
+                }
+                auto block = std::dynamic_pointer_cast<MMDiTBlockT2I>(blocks["patch_blocks." + std::to_string(i)]);
+                auto out   = block->forward(ctx,
+                                            s,
+                                            y_emb,
+                                            condition,
+                                            pos_img,
+                                            pos_txt);
+                s          = out.first;
+                y_emb      = out.second;
+                sd::ggml_graph_cut::mark_graph_cut(s, "pid.patch_blocks." + std::to_string(i), "s");
+                sd::ggml_graph_cut::mark_graph_cut(y_emb, "pid.patch_blocks." + std::to_string(i), "y");
+            }
+            s = ggml_silu(ctx->ggml_ctx, ggml_add(ctx->ggml_ctx, s, t_emb));
+
+            auto s_cond = ggml_reshape_2d(ctx->ggml_ctx, s, config.hidden_size, L * B);
+            auto pixels = pixel_embedder->forward(ctx, x, config.patch_size, pixel_pos_full);
+            for (int i = 0; i < config.pixel_depth; ++i) {
+                auto block = std::dynamic_pointer_cast<PiTBlock>(blocks["pixel_blocks." + std::to_string(i)]);
+                pixels     = block->forward(ctx, pixels, s_cond, H, W, pixel_pos_comp);
+                sd::ggml_graph_cut::mark_graph_cut(pixels, "pid.pixel_blocks." + std::to_string(i), "pixels");
+            }
+
+            pixels   = final_layer->forward(ctx, pixels);
+            pixels   = ggml_reshape_3d(ctx->ggml_ctx, pixels, config.in_channels * P2, L, B);
+            auto out = DiT::unpatchify(ctx->ggml_ctx,
+                                       pixels,
+                                       Hs,
+                                       Ws,
+                                       static_cast<int>(config.patch_size),
+                                       static_cast<int>(config.patch_size),
+                                       false);
+            out      = ggml_ext_slice(ctx->ggml_ctx, out, 1, 0, H_orig);
+            out      = ggml_ext_slice(ctx->ggml_ctx, out, 0, 0, W_orig);
+            return out;
+        }
+    };
+
+    struct PiDRunner : public DiffusionModelRunner {
+        PixelDiTConfig config;
+        PixelDiT model;
+        std::vector<float> pos_img_vec;
+        std::vector<float> pos_txt_vec;
+        std::vector<float> pixel_pos_vec;
+        std::vector<float> pixel_pos_comp_vec;
+
+        PiDRunner(ggml_backend_t backend,
+                  const String2TensorStorage& tensor_storage_map,
+                  const std::string prefix                            = "model.diffusion_model",
+                  std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(PixelDiTConfig::detect_from_weights(tensor_storage_map, prefix)) {
+            model = PixelDiT(config);
+            model.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "PiD";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timesteps_tensor,
+                                 const sd::Tensor<float>& context_tensor,
+                                 const sd::Tensor<float>& lq_latent_tensor,
+                                 const sd::Tensor<float>& degrade_sigma_tensor) {
+            ggml_cgraph* gf            = new_graph_custom(PID_GRAPH_SIZE);
+            ggml_tensor* x             = make_input(x_tensor);
+            ggml_tensor* timesteps     = make_input(timesteps_tensor);
+            ggml_tensor* context       = make_input(context_tensor);
+            ggml_tensor* lq_latent     = make_input(lq_latent_tensor);
+            ggml_tensor* degrade_sigma = make_input(degrade_sigma_tensor);
+
+            int64_t W  = x->ne[0];
+            int64_t H  = x->ne[1];
+            int64_t B  = x->ne[3];
+            int64_t Wp = align_up(static_cast<int>(W), static_cast<int>(config.patch_size));
+            int64_t Hp = align_up(static_cast<int>(H), static_cast<int>(config.patch_size));
+            int64_t Hs = Hp / config.patch_size;
+            int64_t Ws = Wp / config.patch_size;
+
+            pos_img_vec  = make_rope_2d(static_cast<int>(Hs),
+                                        static_cast<int>(Ws),
+                                        static_cast<int>(config.hidden_size / config.num_groups),
+                                        10000.f,
+                                        16.f,
+                                        static_cast<int>(config.rope_ref_grid_h),
+                                        static_cast<int>(config.rope_ref_grid_w));
+            auto pos_img = ggml_new_tensor_4d(compute_ctx,
+                                              GGML_TYPE_F32,
+                                              2,
+                                              2,
+                                              config.hidden_size / config.num_groups / 2,
+                                              Hs * Ws);
+            set_backend_tensor_data(pos_img, pos_img_vec.data());
+
+            int64_t Ltxt = std::min<int64_t>(context->ne[1], config.txt_max_length);
+            pos_txt_vec  = make_rope_1d(static_cast<int>(Ltxt),
+                                        static_cast<int>(config.hidden_size / config.num_groups),
+                                        config.text_rope_theta);
+            auto pos_txt = ggml_new_tensor_4d(compute_ctx,
+                                              GGML_TYPE_F32,
+                                              2,
+                                              2,
+                                              config.hidden_size / config.num_groups / 2,
+                                              Ltxt);
+            set_backend_tensor_data(pos_txt, pos_txt_vec.data());
+
+            pixel_pos_vec  = make_pixel_abs_pos(static_cast<int>(Hp),
+                                                static_cast<int>(Wp),
+                                                static_cast<int>(config.pixel_hidden_size));
+            auto pixel_pos = ggml_new_tensor_3d(compute_ctx,
+                                                GGML_TYPE_F32,
+                                                config.pixel_hidden_size,
+                                                Wp * Hp,
+                                                1);
+            set_backend_tensor_data(pixel_pos, pixel_pos_vec.data());
+
+            pixel_pos_comp_vec  = make_rope_2d(static_cast<int>(Hs),
+                                               static_cast<int>(Ws),
+                                               static_cast<int>(config.pixel_attn_hidden_size / config.pixel_num_groups),
+                                               10000.f,
+                                               16.f,
+                                               static_cast<int>(config.rope_ref_grid_h),
+                                               static_cast<int>(config.rope_ref_grid_w));
+            auto pixel_pos_comp = ggml_new_tensor_4d(compute_ctx,
+                                                     GGML_TYPE_F32,
+                                                     2,
+                                                     2,
+                                                     config.pixel_attn_hidden_size / config.pixel_num_groups / 2,
+                                                     Hs * Ws);
+            set_backend_tensor_data(pixel_pos_comp, pixel_pos_comp_vec.data());
+
+            auto runner_ctx = get_context();
+            auto out        = model.forward(&runner_ctx,
+                                            x,
+                                            timesteps,
+                                            context,
+                                            lq_latent,
+                                            degrade_sigma,
+                                            pos_img,
+                                            pos_txt,
+                                            pixel_pos,
+                                            pixel_pos_comp);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timesteps,
+                                  const sd::Tensor<float>& context,
+                                  const sd::Tensor<float>& lq_latent,
+                                  const sd::Tensor<float>& degrade_sigma) {
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_graph(x, timesteps, context, lq_latent, degrade_sigma);
+            };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            GGML_ASSERT(diffusion_params.context != nullptr);
+            GGML_ASSERT(diffusion_params.ref_latents != nullptr);
+            GGML_ASSERT(!diffusion_params.ref_latents->empty());
+            auto degrade_sigma = sd::Tensor<float>::from_vector({0.0f});
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           *diffusion_params.context,
+                           diffusion_params.ref_latents->front(),
+                           degrade_sigma);
+        }
+    };
+}  // namespace Pid
+
+#endif  // __SD_MODEL_DIFFUSION_PID_HPP__

src/model/diffusion/qwen_image.hpp

@@ -1,14 +1,58 @@
-#ifndef __QWEN_IMAGE_HPP__
-#define __QWEN_IMAGE_HPP__
+#ifndef __SD_MODEL_DIFFUSION_QWEN_IMAGE_HPP__
+#define __SD_MODEL_DIFFUSION_QWEN_IMAGE_HPP__
 
 #include <memory>
 
-#include "common_block.hpp"
-#include "flux.hpp"
+#include "model/common/block.hpp"
+#include "model/diffusion/flux.hpp"
+#include "model/diffusion/model.hpp"
+#include "model_loader.h"
 
 namespace Qwen {
     constexpr int QWEN_IMAGE_GRAPH_SIZE = 20480;
 
+    struct QwenImageConfig {
+        int patch_size              = 2;
+        int64_t in_channels         = 64;
+        int64_t out_channels        = 16;
+        int num_layers              = 60;
+        int64_t attention_head_dim  = 128;
+        int64_t num_attention_heads = 24;
+        int64_t joint_attention_dim = 3584;
+        int theta                   = 10000;
+        std::vector<int> axes_dim   = {16, 56, 56};
+        int axes_dim_sum            = 128;
+        bool zero_cond_t            = false;
+
+        static QwenImageConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            QwenImageConfig config;
+            config.num_layers = 0;
+            for (const auto& [name, _] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (name.find("__index_timestep_zero__") != std::string::npos) {
+                    config.zero_cond_t = true;
+                }
+                size_t pos = name.find("transformer_blocks.");
+                if (pos == std::string::npos) {
+                    continue;
+                }
+                auto items = split_string(name.substr(pos), '.');
+                if (items.size() > 1) {
+                    int block_index = atoi(items[1].c_str());
+                    if (block_index + 1 > config.num_layers) {
+                        config.num_layers = block_index + 1;
+                    }
+                }
+            }
+            LOG_DEBUG("qwen_image: num_layers = %d, zero_cond_t = %s",
+                      config.num_layers,
+                      config.zero_cond_t ? "true" : "false");
+            return config;
+        }
+    };
+
     struct TimestepEmbedding : public GGMLBlock {
     public:
         TimestepEmbedding(int64_t in_channels,
@@ -349,46 +393,32 @@ namespace Qwen {
         }
     };
 
-    struct QwenImageParams {
-        int patch_size              = 2;
-        int64_t in_channels         = 64;
-        int64_t out_channels        = 16;
-        int num_layers              = 60;
-        int64_t attention_head_dim  = 128;
-        int64_t num_attention_heads = 24;
-        int64_t joint_attention_dim = 3584;
-        int theta                   = 10000;
-        std::vector<int> axes_dim   = {16, 56, 56};
-        int axes_dim_sum            = 128;
-        bool zero_cond_t            = false;
-    };
-
     class QwenImageModel : public GGMLBlock {
     protected:
-        QwenImageParams params;
+        QwenImageConfig config;
 
     public:
         QwenImageModel() {}
-        QwenImageModel(QwenImageParams params)
-            : params(params) {
-            int64_t inner_dim         = params.num_attention_heads * params.attention_head_dim;
+        QwenImageModel(QwenImageConfig config)
+            : config(config) {
+            int64_t inner_dim         = config.num_attention_heads * config.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["txt_norm"]        = std::shared_ptr<GGMLBlock>(new RMSNorm(config.joint_attention_dim, 1e-6f));
+            blocks["img_in"]          = std::shared_ptr<GGMLBlock>(new Linear(config.in_channels, inner_dim));
+            blocks["txt_in"]          = std::shared_ptr<GGMLBlock>(new Linear(config.joint_attention_dim, inner_dim));
 
             // blocks
-            for (int i = 0; i < params.num_layers; i++) {
+            for (int i = 0; i < config.num_layers; i++) {
                 auto block                                        = std::shared_ptr<GGMLBlock>(new QwenImageTransformerBlock(inner_dim,
-                                                                                                                             params.num_attention_heads,
-                                                                                                                             params.attention_head_dim,
+                                                                                                                             config.num_attention_heads,
+                                                                                                                             config.attention_head_dim,
                                                                                                                              1e-6f,
-                                                                                                                             params.zero_cond_t));
+                                                                                                                             config.zero_cond_t));
                 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));
+            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, config.patch_size * config.patch_size * config.out_channels));
         }
 
         ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
@@ -405,7 +435,7 @@ namespace Qwen {
             auto proj_out        = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
 
             auto t_emb = time_text_embed->forward(ctx, timestep);
-            if (params.zero_cond_t) {
+            if (config.zero_cond_t) {
                 auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep));
                 t_emb        = ggml_concat(ctx->ggml_ctx, t_emb, t_emb_0, 1);
             }
@@ -416,7 +446,7 @@ namespace Qwen {
             sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.prelude", "txt");
             // sd::ggml_graph_cut::mark_graph_cut(t_emb, "qwen_image.prelude", "t_emb");
 
-            for (int i = 0; i < params.num_layers; i++) {
+            for (int i = 0; i < config.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
 
                 auto result = block->forward(ctx, img, txt, t_emb, pe, modulate_index);
@@ -426,7 +456,7 @@ namespace Qwen {
                 sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.transformer_blocks." + std::to_string(i), "txt");
             }
 
-            if (params.zero_cond_t) {
+            if (config.zero_cond_t) {
                 t_emb = ggml_ext_chunk(ctx->ggml_ctx, t_emb, 2, 1)[0];
             }
 
@@ -455,12 +485,12 @@ namespace Qwen {
             int64_t C = x->ne[2];
             int64_t N = x->ne[3];
 
-            auto img           = DiT::pad_and_patchify(ctx, x, params.patch_size, params.patch_size);
+            auto img           = DiT::pad_and_patchify(ctx, x, config.patch_size, config.patch_size);
             int64_t img_tokens = img->ne[1];
 
             if (ref_latents.size() > 0) {
                 for (ggml_tensor* ref : ref_latents) {
-                    ref = DiT::pad_and_patchify(ctx, ref, params.patch_size, params.patch_size);
+                    ref = DiT::pad_and_patchify(ctx, ref, config.patch_size, config.patch_size);
                     img = ggml_concat(ctx->ggml_ctx, img, ref, 1);
                 }
             }
@@ -473,54 +503,30 @@ namespace Qwen {
                 out = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, out, 0, 2, 1, 3));  // [N, h*w, C * patch_size * patch_size]
             }
 
-            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, params.patch_size, params.patch_size);  // [N, C, H, W]
+            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, config.patch_size, config.patch_size);  // [N, C, H, W]
 
             return out;
         }
     };
 
-    struct QwenImageRunner : public GGMLRunner {
+    struct QwenImageRunner : public DiffusionModelRunner {
     public:
-        QwenImageParams qwen_image_params;
+        QwenImageConfig config;
         QwenImageModel qwen_image;
         std::vector<float> pe_vec;
         std::vector<float> modulate_index_vec;
         SDVersion version;
 
         QwenImageRunner(ggml_backend_t backend,
-                        ggml_backend_t params_backend,
-                        const String2TensorStorage& tensor_storage_map = {},
-                        const std::string prefix                       = "",
-                        SDVersion version                              = VERSION_QWEN_IMAGE,
-                        bool zero_cond_t                               = false)
-            : GGMLRunner(backend, params_backend) {
-            qwen_image_params.num_layers  = 0;
-            qwen_image_params.zero_cond_t = zero_cond_t;
-            for (auto pair : tensor_storage_map) {
-                std::string tensor_name = pair.first;
-                if (tensor_name.find(prefix) == std::string::npos)
-                    continue;
-                if (tensor_name.find("__index_timestep_zero__") != std::string::npos) {
-                    qwen_image_params.zero_cond_t = true;
-                }
-                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);
-            if (qwen_image_params.zero_cond_t) {
-                LOG_INFO("use zero_cond_t");
-            }
-            qwen_image = QwenImageModel(qwen_image_params);
+                        const String2TensorStorage& tensor_storage_map      = {},
+                        const std::string prefix                            = "",
+                        SDVersion version                                   = VERSION_QWEN_IMAGE,
+                        bool zero_cond_t                                    = false,
+                        std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(QwenImageConfig::detect_from_weights(tensor_storage_map, prefix)) {
+            config.zero_cond_t = config.zero_cond_t || zero_cond_t;
+            qwen_image         = QwenImageModel(config);
             qwen_image.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -528,7 +534,7 @@ namespace Qwen {
             return "qwen_image";
         }
 
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
             qwen_image.get_param_tensors(tensors, prefix);
         }
 
@@ -551,36 +557,36 @@ namespace Qwen {
 
             pe_vec      = Rope::gen_qwen_image_pe(static_cast<int>(x->ne[1]),
                                                   static_cast<int>(x->ne[0]),
-                                                  qwen_image_params.patch_size,
+                                                  config.patch_size,
                                                   static_cast<int>(x->ne[3]),
                                                   static_cast<int>(context->ne[1]),
                                                   ref_latents,
                                                   increase_ref_index,
-                                                  qwen_image_params.theta,
+                                                  config.theta,
                                                   circular_y_enabled,
                                                   circular_x_enabled,
-                                                  qwen_image_params.axes_dim);
-            int pos_len = static_cast<int>(pe_vec.size() / qwen_image_params.axes_dim_sum / 2);
+                                                  config.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / config.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);
+            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.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());
 
             ggml_tensor* modulate_index = nullptr;
-            if (qwen_image_params.zero_cond_t) {
+            if (config.zero_cond_t) {
                 modulate_index_vec.clear();
 
-                int64_t h_len          = ((x->ne[1] + (qwen_image_params.patch_size / 2)) / qwen_image_params.patch_size);
-                int64_t w_len          = ((x->ne[0] + (qwen_image_params.patch_size / 2)) / qwen_image_params.patch_size);
+                int64_t h_len          = ((x->ne[1] + (config.patch_size / 2)) / config.patch_size);
+                int64_t w_len          = ((x->ne[0] + (config.patch_size / 2)) / config.patch_size);
                 int64_t num_img_tokens = h_len * w_len;
 
                 modulate_index_vec.insert(modulate_index_vec.end(), num_img_tokens, 0.f);
                 int64_t num_ref_img_tokens = 0;
                 for (ggml_tensor* ref : ref_latents) {
-                    int64_t h_len = ((ref->ne[1] + (qwen_image_params.patch_size / 2)) / qwen_image_params.patch_size);
-                    int64_t w_len = ((ref->ne[0] + (qwen_image_params.patch_size / 2)) / qwen_image_params.patch_size);
+                    int64_t h_len = ((ref->ne[1] + (config.patch_size / 2)) / config.patch_size);
+                    int64_t w_len = ((ref->ne[0] + (config.patch_size / 2)) / config.patch_size);
 
                     num_ref_img_tokens += h_len * w_len;
                 }
@@ -621,7 +627,20 @@ namespace Qwen {
                 return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
             };
 
-            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            static const std::vector<sd::Tensor<float>> empty_ref_latents;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
+                           diffusion_params.increase_ref_index);
         }
 
         void test() {
@@ -669,10 +688,11 @@ namespace Qwen {
             // cuda q8: pass
             // cuda q8 fa: pass
             // ggml_backend_t backend    = ggml_backend_cuda_init(0);
-            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_backend_t backend    = sd_backend_cpu_init();
             ggml_type model_data_type = GGML_TYPE_Q8_0;
 
-            ModelLoader model_loader;
+            auto model_manager        = std::make_shared<ModelManager>();
+            ModelLoader& model_loader = model_manager->loader();
             if (!model_loader.init_from_file_and_convert_name(file_path, "model.diffusion_model.")) {
                 LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
                 return;
@@ -686,19 +706,20 @@ namespace Qwen {
             }
 
             std::shared_ptr<QwenImageRunner> qwen_image = std::make_shared<QwenImageRunner>(backend,
-                                                                                            backend,
                                                                                             tensor_storage_map,
                                                                                             "model.diffusion_model",
-                                                                                            VERSION_QWEN_IMAGE);
+                                                                                            VERSION_QWEN_IMAGE,
+                                                                                            false,
+                                                                                            model_manager);
 
-            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");
+            if (!model_manager->register_runner_params("Qwen image test",
+                                                       *qwen_image,
+                                                       "model.diffusion_model",
+                                                       ModelManager::ResidencyMode::ParamBackend,
+                                                       backend,
+                                                       backend) ||
+                !model_manager->validate_registered_tensors()) {
+                LOG_ERROR("register qwen_image tensors with model manager failed");
                 return;
             }
 
@@ -709,4 +730,4 @@ namespace Qwen {
 
 }  // namespace name
 
-#endif  // __QWEN_IMAGE_HPP__
+#endif  // __SD_MODEL_DIFFUSION_QWEN_IMAGE_HPP__

src/model/diffusion/unet.hpp

@@ -1,13 +1,136 @@
-#ifndef __UNET_HPP__
-#define __UNET_HPP__
+#ifndef __SD_MODEL_DIFFUSION_UNET_HPP__
+#define __SD_MODEL_DIFFUSION_UNET_HPP__
+
+#include <algorithm>
+#include <vector>
 
-#include "common_block.hpp"
 #include "model.h"
+#include "model/common/block.hpp"
+#include "model/diffusion/model.hpp"
 
 /*==================================================== UnetModel =====================================================*/
 
 #define UNET_GRAPH_SIZE 102400
 
+struct UNetConfig {
+    SDVersion version = VERSION_SD1;
+    // network hparams
+    int in_channels                        = 4;
+    int out_channels                       = 4;
+    int num_res_blocks                     = 2;
+    std::vector<int> attention_resolutions = {4, 2, 1};
+    std::vector<int> channel_mult          = {1, 2, 4, 4};
+    std::vector<int> transformer_depth     = {1, 1, 1, 1};
+    int time_embed_dim                     = 1280;  // model_channels*4
+    int num_heads                          = 8;
+    int num_head_channels                  = -1;   // channels // num_heads
+    int context_dim                        = 768;  // 1024 for VERSION_SD2, 2048 for VERSION_SDXL
+    bool use_linear_projection             = false;
+    bool tiny_unet                         = false;
+    int model_channels                     = 320;
+    int adm_in_channels                    = 2816;  // only for VERSION_SDXL/SVD
+
+    static UNetConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                          const std::string& prefix,
+                                          SDVersion version = VERSION_SD1) {
+        UNetConfig config;
+        config.version = version;
+
+        if (sd_version_is_sd2(version)) {
+            config.context_dim           = 1024;
+            config.num_head_channels     = 64;
+            config.num_heads             = -1;
+            config.use_linear_projection = true;
+        } else if (sd_version_is_sdxl(version)) {
+            config.context_dim           = 2048;
+            config.attention_resolutions = {4, 2};
+            config.channel_mult          = {1, 2, 4};
+            config.transformer_depth     = {1, 2, 10};
+            config.num_head_channels     = 64;
+            config.num_heads             = -1;
+            config.use_linear_projection = true;
+            if (version == VERSION_SDXL_VEGA) {
+                config.transformer_depth = {1, 1, 2};
+            }
+        } else if (version == VERSION_SVD) {
+            config.in_channels           = 8;
+            config.out_channels          = 4;
+            config.context_dim           = 1024;
+            config.adm_in_channels       = 768;
+            config.num_head_channels     = 64;
+            config.num_heads             = -1;
+            config.use_linear_projection = true;
+        }
+        if (sd_version_is_inpaint(version)) {
+            config.in_channels = 9;
+        } else if (sd_version_is_unet_edit(version)) {
+            config.in_channels = 8;
+        }
+        if (version == VERSION_SD1_TINY_UNET || version == VERSION_SD2_TINY_UNET || version == VERSION_SDXS_512_DS || version == VERSION_SDXS_09) {
+            config.num_res_blocks = 1;
+            config.channel_mult   = {1, 2, 4};
+            config.tiny_unet      = true;
+            if (version == VERSION_SDXS_512_DS) {
+                config.attention_resolutions = {4, 2};  // here just like SDXL
+            }
+        }
+
+        auto find_weight = [&](const std::string& suffix) -> const TensorStorage* {
+            std::string name = prefix.empty() ? suffix : prefix + "." + suffix;
+            auto it          = tensor_storage_map.find(name);
+            if (it == tensor_storage_map.end()) {
+                return nullptr;
+            }
+            return &it->second;
+        };
+
+        if (const TensorStorage* input = find_weight("input_blocks.0.0.weight")) {
+            if (input->n_dims == 4) {
+                config.in_channels    = static_cast<int>(input->ne[2]);
+                config.model_channels = static_cast<int>(input->ne[3]);
+                config.time_embed_dim = config.model_channels * 4;
+            }
+        }
+        if (const TensorStorage* time_embed = find_weight("time_embed.0.weight")) {
+            if (time_embed->n_dims == 2) {
+                config.model_channels = static_cast<int>(time_embed->ne[0]);
+                config.time_embed_dim = static_cast<int>(time_embed->ne[1]);
+            }
+        }
+        if (const TensorStorage* label_emb = find_weight("label_emb.0.0.weight")) {
+            if (label_emb->n_dims == 2) {
+                config.adm_in_channels = static_cast<int>(label_emb->ne[0]);
+                config.time_embed_dim  = static_cast<int>(label_emb->ne[1]);
+            }
+        }
+        if (const TensorStorage* out = find_weight("out.2.weight")) {
+            if (out->n_dims == 4) {
+                config.out_channels = static_cast<int>(out->ne[3]);
+            }
+        }
+        for (const auto& [name, tensor_storage] : tensor_storage_map) {
+            if (!starts_with(name, prefix)) {
+                continue;
+            }
+            if (name.find("attn2.to_k.weight") != std::string::npos && tensor_storage.n_dims == 2) {
+                config.context_dim = static_cast<int>(tensor_storage.ne[0]);
+                break;
+            }
+        }
+
+        LOG_DEBUG("unet: in_channels = %d, out_channels = %d, model_channels = %d, time_embed_dim = %d, context_dim = %d, adm_in_channels = %d, num_res_blocks = %d, tiny_unet = %s",
+                  config.in_channels,
+                  config.out_channels,
+                  config.model_channels,
+                  config.time_embed_dim,
+                  config.context_dim,
+                  config.adm_in_channels,
+                  config.num_res_blocks,
+                  config.tiny_unet ? "true" : "false");
+        return config;
+    }
+};
+
 class SpatialVideoTransformer : public SpatialTransformer {
 protected:
     int64_t time_depth;
@@ -165,66 +288,26 @@ public:
 
 // ldm.modules.diffusionmodules.openaimodel.UNetModel
 class UnetModelBlock : public GGMLBlock {
-protected:
-    SDVersion version = VERSION_SD1;
-    // network hparams
-    int in_channels                        = 4;
-    int out_channels                       = 4;
-    int num_res_blocks                     = 2;
-    std::vector<int> attention_resolutions = {4, 2, 1};
-    std::vector<int> channel_mult          = {1, 2, 4, 4};
-    std::vector<int> transformer_depth     = {1, 1, 1, 1};
-    int time_embed_dim                     = 1280;  // model_channels*4
-    int num_heads                          = 8;
-    int num_head_channels                  = -1;   // channels // num_heads
-    int context_dim                        = 768;  // 1024 for VERSION_SD2, 2048 for VERSION_SDXL
-    bool use_linear_projection             = false;
-    bool tiny_unet                         = false;
-
 public:
-    int model_channels  = 320;
-    int adm_in_channels = 2816;  // only for VERSION_SDXL/SVD
+    UNetConfig config;
 
-    UnetModelBlock(SDVersion version = VERSION_SD1, const String2TensorStorage& tensor_storage_map = {})
-        : version(version) {
-        if (sd_version_is_sd2(version)) {
-            context_dim           = 1024;
-            num_head_channels     = 64;
-            num_heads             = -1;
-            use_linear_projection = true;
-        } else if (sd_version_is_sdxl(version)) {
-            context_dim           = 2048;
-            attention_resolutions = {4, 2};
-            channel_mult          = {1, 2, 4};
-            transformer_depth     = {1, 2, 10};
-            num_head_channels     = 64;
-            num_heads             = -1;
-            use_linear_projection = true;
-            if (version == VERSION_SDXL_VEGA) {
-                transformer_depth = {1, 1, 2};
-            }
-        } else if (version == VERSION_SVD) {
-            in_channels           = 8;
-            out_channels          = 4;
-            context_dim           = 1024;
-            adm_in_channels       = 768;
-            num_head_channels     = 64;
-            num_heads             = -1;
-            use_linear_projection = true;
-        }
-        if (sd_version_is_inpaint(version)) {
-            in_channels = 9;
-        } else if (sd_version_is_unet_edit(version)) {
-            in_channels = 8;
-        }
-        if (version == VERSION_SD1_TINY_UNET || version == VERSION_SD2_TINY_UNET || version == VERSION_SDXS_512_DS || version == VERSION_SDXS_09) {
-            num_res_blocks = 1;
-            channel_mult   = {1, 2, 4};
-            tiny_unet      = true;
-            if (version == VERSION_SDXS_512_DS) {
-                attention_resolutions = {4, 2};  // here just like SDXL
-            }
-        }
+    explicit UnetModelBlock(UNetConfig config = {})
+        : config(config) {
+        const SDVersion version           = this->config.version;
+        const int in_channels             = this->config.in_channels;
+        const int out_channels            = this->config.out_channels;
+        const int num_res_blocks          = this->config.num_res_blocks;
+        const auto& attention_resolutions = this->config.attention_resolutions;
+        const auto& channel_mult          = this->config.channel_mult;
+        const auto& transformer_depth     = this->config.transformer_depth;
+        const int time_embed_dim          = this->config.time_embed_dim;
+        const int num_heads               = this->config.num_heads;
+        const int num_head_channels       = this->config.num_head_channels;
+        const int context_dim             = this->config.context_dim;
+        const bool use_linear_projection  = this->config.use_linear_projection;
+        const bool tiny_unet              = this->config.tiny_unet;
+        const int model_channels          = this->config.model_channels;
+        const int adm_in_channels         = this->config.adm_in_channels;
 
         // dims is always 2
         // use_temporal_attention is always True for SVD
@@ -397,7 +480,7 @@ public:
                                   ggml_tensor* x,
                                   ggml_tensor* emb,
                                   int num_video_frames) {
-        if (version == VERSION_SVD) {
+        if (config.version == VERSION_SVD) {
             auto block = std::dynamic_pointer_cast<VideoResBlock>(blocks[name]);
 
             return block->forward(ctx, x, emb, num_video_frames);
@@ -413,7 +496,7 @@ public:
                                          ggml_tensor* x,
                                          ggml_tensor* context,
                                          int timesteps) {
-        if (version == VERSION_SVD) {
+        if (config.version == VERSION_SVD) {
             auto block = std::dynamic_pointer_cast<SpatialVideoTransformer>(blocks[name]);
 
             return block->forward(ctx, x, context, timesteps);
@@ -439,6 +522,13 @@ 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]
+        const SDVersion version           = config.version;
+        const int model_channels          = config.model_channels;
+        const int num_res_blocks          = config.num_res_blocks;
+        const auto& attention_resolutions = config.attention_resolutions;
+        const auto& channel_mult          = config.channel_mult;
+        const bool tiny_unet              = config.tiny_unet;
+
         if (context != nullptr) {
             if (context->ne[2] != x->ne[3]) {
                 context = ggml_repeat(ctx->ggml_ctx, context, ggml_new_tensor_3d(ctx->ggml_ctx, GGML_TYPE_F32, context->ne[0], context->ne[1], x->ne[3]));
@@ -599,15 +689,18 @@ public:
     }
 };
 
-struct UNetModelRunner : public GGMLRunner {
+struct UNetModelRunner : public DiffusionModelRunner {
+    UNetConfig config;
     UnetModelBlock unet;
 
     UNetModelRunner(ggml_backend_t backend,
-                    ggml_backend_t params_backend,
                     const String2TensorStorage& tensor_storage_map,
                     const std::string prefix,
-                    SDVersion version = VERSION_SD1)
-        : GGMLRunner(backend, params_backend), unet(version, tensor_storage_map) {
+                    SDVersion version                                   = VERSION_SD1,
+                    std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : DiffusionModelRunner(backend, prefix, weight_manager),
+          config(UNetConfig::detect_from_weights(tensor_storage_map, prefix, version)),
+          unet(config) {
         unet.init(params_ctx, tensor_storage_map, prefix);
     }
 
@@ -615,7 +708,7 @@ struct UNetModelRunner : public GGMLRunner {
         return "unet";
     }
 
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
         unet.get_param_tensors(tensors, prefix);
     }
 
@@ -679,7 +772,24 @@ struct UNetModelRunner : public GGMLRunner {
             return build_graph(x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength);
         };
 
-        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+    }
+
+    sd::Tensor<float> compute(int n_threads,
+                              const DiffusionParams& diffusion_params) override {
+        GGML_ASSERT(diffusion_params.x != nullptr);
+        GGML_ASSERT(diffusion_params.timesteps != nullptr);
+        const auto* extra = diffusion_extra_as<UNetDiffusionExtra>(diffusion_params);
+        static const std::vector<sd::Tensor<float>> empty_controls;
+        return compute(n_threads,
+                       *diffusion_params.x,
+                       *diffusion_params.timesteps,
+                       tensor_or_empty(diffusion_params.context),
+                       tensor_or_empty(diffusion_params.c_concat),
+                       tensor_or_empty(diffusion_params.y),
+                       extra->num_video_frames,
+                       extra->controls ? *extra->controls : empty_controls,
+                       extra->control_strength);
     }
 
     void test() {
@@ -734,4 +844,4 @@ struct UNetModelRunner : public GGMLRunner {
     }
 };
 
-#endif  // __UNET_HPP__
+#endif  // __SD_MODEL_DIFFUSION_UNET_HPP__

src/model/diffusion/z_image.hpp

@@ -1,13 +1,15 @@
-#ifndef __Z_IMAGE_HPP__
-#define __Z_IMAGE_HPP__
+#ifndef __SD_MODEL_DIFFUSION_Z_IMAGE_HPP__
+#define __SD_MODEL_DIFFUSION_Z_IMAGE_HPP__
 
 #include <algorithm>
 
-#include "flux.hpp"
-#include "ggml_extend.hpp"
-#include "mmdit.hpp"
+#include "core/ggml_extend.hpp"
+#include "model/diffusion/flux.hpp"
+#include "model/diffusion/mmdit.hpp"
+#include "model/diffusion/model.hpp"
+#include "model_loader.h"
 
-// Ref: https://github.com/Alpha-VLLM/Lumina-Image-2.0/blob/main/models/model.py
+// Ref: https://github.com/Alpha-VLLM/Lumina-Image-2.0/blob/main/model/model.py
 // Ref: https://github.com/huggingface/diffusers/pull/12703
 
 #ifndef MIN
@@ -19,6 +21,104 @@ namespace ZImage {
     constexpr int ADALN_EMBED_DIM    = 256;
     constexpr int SEQ_MULTI_OF       = 32;
 
+    struct ZImageConfig {
+        int patch_size             = 2;
+        int64_t hidden_size        = 3840;
+        int64_t in_channels        = 16;
+        int64_t out_channels       = 16;
+        int64_t num_layers         = 30;
+        int64_t num_refiner_layers = 2;
+        int64_t head_dim           = 128;
+        int64_t num_heads          = 30;
+        int64_t num_kv_heads       = 30;
+        int64_t multiple_of        = 256;
+        float ffn_dim_multiplier   = 8.0f / 3.0f;
+        float norm_eps             = 1e-5f;
+        bool qk_norm               = true;
+        int64_t cap_feat_dim       = 2560;
+        int theta                  = 256;
+        std::vector<int> axes_dim  = {32, 48, 48};
+        int64_t axes_dim_sum       = 128;
+
+        static ZImageConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
+            ZImageConfig config;
+            int64_t detected_layers          = 0;
+            int64_t detected_refiner_layers  = 0;
+            int64_t detected_context_refiner = 0;
+            int64_t detected_head_dim        = 0;
+            int64_t detected_qkv_dim         = 0;
+
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (!starts_with(name, prefix)) {
+                    continue;
+                }
+                if (ends_with(name, "x_embedder.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area = config.patch_size * config.patch_size;
+                    config.in_channels = tensor_storage.ne[0] / patch_area;
+                    config.hidden_size = tensor_storage.ne[1];
+                } else if (ends_with(name, "cap_embedder.1.weight") && tensor_storage.n_dims == 2) {
+                    config.cap_feat_dim = tensor_storage.ne[0];
+                    config.hidden_size  = tensor_storage.ne[1];
+                } else if (ends_with(name, "layers.0.attention.q_norm.weight") && tensor_storage.n_dims == 1) {
+                    detected_head_dim = tensor_storage.ne[0];
+                } else if (ends_with(name, "layers.0.attention.qkv.weight") && tensor_storage.n_dims == 2) {
+                    detected_qkv_dim = tensor_storage.ne[1];
+                } else if (ends_with(name, "final_layer.linear.weight") && tensor_storage.n_dims == 2) {
+                    int64_t patch_area  = config.patch_size * config.patch_size;
+                    config.out_channels = tensor_storage.ne[1] / patch_area;
+                }
+
+                size_t pos = name.find("layers.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index = atoi(items[1].c_str());
+                        detected_layers = std::max<int64_t>(detected_layers, block_index + 1);
+                    }
+                }
+                pos = name.find("noise_refiner.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index         = atoi(items[1].c_str());
+                        detected_refiner_layers = std::max<int64_t>(detected_refiner_layers, block_index + 1);
+                    }
+                }
+                pos = name.find("context_refiner.");
+                if (pos != std::string::npos) {
+                    auto items = split_string(name.substr(pos), '.');
+                    if (items.size() > 1) {
+                        int block_index          = atoi(items[1].c_str());
+                        detected_context_refiner = std::max<int64_t>(detected_context_refiner, block_index + 1);
+                    }
+                }
+            }
+            if (detected_layers > 0) {
+                config.num_layers = detected_layers;
+            }
+            if (detected_refiner_layers > 0 || detected_context_refiner > 0) {
+                config.num_refiner_layers = std::max(detected_refiner_layers, detected_context_refiner);
+            }
+            if (detected_head_dim > 0) {
+                config.head_dim  = detected_head_dim;
+                config.num_heads = config.hidden_size / config.head_dim;
+                if (detected_qkv_dim > 0) {
+                    int64_t qkv_heads   = detected_qkv_dim / config.head_dim;
+                    config.num_kv_heads = std::max<int64_t>(1, (qkv_heads - config.num_heads) / 2);
+                }
+            }
+            LOG_DEBUG("z_image: num_layers = %" PRId64 ", num_refiner_layers = %" PRId64 ", hidden_size = %" PRId64 ", num_heads = %" PRId64 ", num_kv_heads = %" PRId64 ", in_channels = %" PRId64 ", out_channels = %" PRId64,
+                      config.num_layers,
+                      config.num_refiner_layers,
+                      config.hidden_size,
+                      config.num_heads,
+                      config.num_kv_heads,
+                      config.in_channels,
+                      config.out_channels);
+            return config;
+        }
+    };
+
     struct JointAttention : public GGMLBlock {
     protected:
         int64_t head_dim;
@@ -262,90 +362,70 @@ namespace ZImage {
         }
     };
 
-    struct ZImageParams {
-        int patch_size             = 2;
-        int64_t hidden_size        = 3840;
-        int64_t in_channels        = 16;
-        int64_t out_channels       = 16;
-        int64_t num_layers         = 30;
-        int64_t num_refiner_layers = 2;
-        int64_t head_dim           = 128;
-        int64_t num_heads          = 30;
-        int64_t num_kv_heads       = 30;
-        int64_t multiple_of        = 256;
-        float ffn_dim_multiplier   = 8.0f / 3.0f;
-        float norm_eps             = 1e-5f;
-        bool qk_norm               = true;
-        int64_t cap_feat_dim       = 2560;
-        int theta                  = 256;
-        std::vector<int> axes_dim  = {32, 48, 48};
-        int64_t axes_dim_sum       = 128;
-    };
-
     class ZImageModel : public GGMLBlock {
     protected:
-        ZImageParams z_image_params;
+        ZImageConfig config;
 
         void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
-            params["cap_pad_token"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, z_image_params.hidden_size);
-            params["x_pad_token"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, z_image_params.hidden_size);
+            params["cap_pad_token"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, config.hidden_size);
+            params["x_pad_token"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, config.hidden_size);
         }
 
     public:
         ZImageModel() = default;
-        ZImageModel(ZImageParams z_image_params)
-            : z_image_params(z_image_params) {
-            blocks["x_embedder"]     = std::make_shared<Linear>(z_image_params.patch_size * z_image_params.patch_size * z_image_params.in_channels, z_image_params.hidden_size);
-            blocks["t_embedder"]     = std::make_shared<TimestepEmbedder>(MIN(z_image_params.hidden_size, 1024), 256, 256);
-            blocks["cap_embedder.0"] = std::make_shared<RMSNorm>(z_image_params.cap_feat_dim, z_image_params.norm_eps);
-            blocks["cap_embedder.1"] = std::make_shared<Linear>(z_image_params.cap_feat_dim, z_image_params.hidden_size);
+        ZImageModel(ZImageConfig config)
+            : config(config) {
+            blocks["x_embedder"]     = std::make_shared<Linear>(config.patch_size * config.patch_size * config.in_channels, config.hidden_size);
+            blocks["t_embedder"]     = std::make_shared<TimestepEmbedder>(MIN(config.hidden_size, 1024), 256, 256);
+            blocks["cap_embedder.0"] = std::make_shared<RMSNorm>(config.cap_feat_dim, config.norm_eps);
+            blocks["cap_embedder.1"] = std::make_shared<Linear>(config.cap_feat_dim, config.hidden_size);
 
-            for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
+            for (int i = 0; i < config.num_refiner_layers; i++) {
                 auto block = std::make_shared<JointTransformerBlock>(i,
-                                                                     z_image_params.hidden_size,
-                                                                     z_image_params.head_dim,
-                                                                     z_image_params.num_heads,
-                                                                     z_image_params.num_kv_heads,
-                                                                     z_image_params.multiple_of,
-                                                                     z_image_params.ffn_dim_multiplier,
-                                                                     z_image_params.norm_eps,
-                                                                     z_image_params.qk_norm,
+                                                                     config.hidden_size,
+                                                                     config.head_dim,
+                                                                     config.num_heads,
+                                                                     config.num_kv_heads,
+                                                                     config.multiple_of,
+                                                                     config.ffn_dim_multiplier,
+                                                                     config.norm_eps,
+                                                                     config.qk_norm,
                                                                      true);
 
                 blocks["noise_refiner." + std::to_string(i)] = block;
             }
 
-            for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
+            for (int i = 0; i < config.num_refiner_layers; i++) {
                 auto block = std::make_shared<JointTransformerBlock>(i,
-                                                                     z_image_params.hidden_size,
-                                                                     z_image_params.head_dim,
-                                                                     z_image_params.num_heads,
-                                                                     z_image_params.num_kv_heads,
-                                                                     z_image_params.multiple_of,
-                                                                     z_image_params.ffn_dim_multiplier,
-                                                                     z_image_params.norm_eps,
-                                                                     z_image_params.qk_norm,
+                                                                     config.hidden_size,
+                                                                     config.head_dim,
+                                                                     config.num_heads,
+                                                                     config.num_kv_heads,
+                                                                     config.multiple_of,
+                                                                     config.ffn_dim_multiplier,
+                                                                     config.norm_eps,
+                                                                     config.qk_norm,
                                                                      false);
 
                 blocks["context_refiner." + std::to_string(i)] = block;
             }
 
-            for (int i = 0; i < z_image_params.num_layers; i++) {
+            for (int i = 0; i < config.num_layers; i++) {
                 auto block = std::make_shared<JointTransformerBlock>(i,
-                                                                     z_image_params.hidden_size,
-                                                                     z_image_params.head_dim,
-                                                                     z_image_params.num_heads,
-                                                                     z_image_params.num_kv_heads,
-                                                                     z_image_params.multiple_of,
-                                                                     z_image_params.ffn_dim_multiplier,
-                                                                     z_image_params.norm_eps,
-                                                                     z_image_params.qk_norm,
+                                                                     config.hidden_size,
+                                                                     config.head_dim,
+                                                                     config.num_heads,
+                                                                     config.num_kv_heads,
+                                                                     config.multiple_of,
+                                                                     config.ffn_dim_multiplier,
+                                                                     config.norm_eps,
+                                                                     config.qk_norm,
                                                                      true);
 
                 blocks["layers." + std::to_string(i)] = block;
             }
 
-            blocks["final_layer"] = std::make_shared<FinalLayer>(z_image_params.hidden_size, z_image_params.patch_size, z_image_params.out_channels);
+            blocks["final_layer"] = std::make_shared<FinalLayer>(config.hidden_size, config.patch_size, config.out_channels);
         }
 
         ggml_tensor* forward_core(GGMLRunnerContext* ctx,
@@ -392,14 +472,14 @@ namespace ZImage {
             auto txt_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, 0, txt->ne[1]);
             auto img_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt->ne[1], pe->ne[3]);
 
-            for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
+            for (int i = 0; i < config.num_refiner_layers; i++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["context_refiner." + std::to_string(i)]);
 
                 txt = block->forward(ctx, txt, txt_pe, nullptr, nullptr);
                 sd::ggml_graph_cut::mark_graph_cut(txt, "z_image.context_refiner." + std::to_string(i), "txt");
             }
 
-            for (int i = 0; i < z_image_params.num_refiner_layers; i++) {
+            for (int i = 0; i < config.num_refiner_layers; i++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["noise_refiner." + std::to_string(i)]);
 
                 img = block->forward(ctx, img, img_pe, nullptr, t_emb);
@@ -409,7 +489,7 @@ namespace ZImage {
             auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_txt_pad_token + n_img_token + n_img_pad_token, hidden_size]
             sd::ggml_graph_cut::mark_graph_cut(txt_img, "z_image.prelude", "txt_img");
 
-            for (int i = 0; i < z_image_params.num_layers; i++) {
+            for (int i = 0; i < config.num_layers; i++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["layers." + std::to_string(i)]);
 
                 txt_img = block->forward(ctx, txt_img, pe, nullptr, t_emb);
@@ -441,7 +521,7 @@ namespace ZImage {
             int64_t C = x->ne[2];
             int64_t N = x->ne[3];
 
-            int patch_size = z_image_params.patch_size;
+            int patch_size = config.patch_size;
 
             auto img             = DiT::pad_and_patchify(ctx, x, patch_size, patch_size, false);
             uint64_t n_img_token = img->ne[1];
@@ -464,21 +544,22 @@ namespace ZImage {
         }
     };
 
-    struct ZImageRunner : public GGMLRunner {
+    struct ZImageRunner : public DiffusionModelRunner {
     public:
-        ZImageParams z_image_params;
+        ZImageConfig config;
         ZImageModel z_image;
         std::vector<float> pe_vec;
         std::vector<float> timestep_vec;
         SDVersion version;
 
         ZImageRunner(ggml_backend_t backend,
-                     ggml_backend_t params_backend,
-                     const String2TensorStorage& tensor_storage_map = {},
-                     const std::string prefix                       = "",
-                     SDVersion version                              = VERSION_Z_IMAGE)
-            : GGMLRunner(backend, params_backend) {
-            z_image = ZImageModel(z_image_params);
+                     const String2TensorStorage& tensor_storage_map      = {},
+                     const std::string prefix                            = "",
+                     SDVersion version                                   = VERSION_Z_IMAGE,
+                     std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : DiffusionModelRunner(backend, prefix, weight_manager),
+              config(ZImageConfig::detect_from_weights(tensor_storage_map, prefix)) {
+            z_image = ZImageModel(config);
             z_image.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -486,7 +567,7 @@ namespace ZImage {
             return "z_image";
         }
 
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
             z_image.get_param_tensors(tensors, prefix);
         }
 
@@ -509,19 +590,19 @@ namespace ZImage {
 
             pe_vec      = Rope::gen_z_image_pe(static_cast<int>(x->ne[1]),
                                                static_cast<int>(x->ne[0]),
-                                               z_image_params.patch_size,
+                                               config.patch_size,
                                                static_cast<int>(x->ne[3]),
                                                static_cast<int>(context->ne[1]),
                                                SEQ_MULTI_OF,
                                                ref_latents,
                                                increase_ref_index,
-                                               z_image_params.theta,
+                                               config.theta,
                                                circular_y_enabled,
                                                circular_x_enabled,
-                                               z_image_params.axes_dim);
-            int pos_len = static_cast<int>(pe_vec.size() / z_image_params.axes_dim_sum / 2);
+                                               config.axes_dim);
+            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             // LOG_DEBUG("pos_len %d", pos_len);
-            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, z_image_params.axes_dim_sum / 2, pos_len);
+            auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
             // pe->data = pe_vec.data();
             // print_ggml_tensor(pe, true, "pe");
             // pe->data = nullptr;
@@ -553,7 +634,20 @@ namespace ZImage {
                 return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
             };
 
-            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), x.dim());
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const DiffusionParams& diffusion_params) override {
+            GGML_ASSERT(diffusion_params.x != nullptr);
+            GGML_ASSERT(diffusion_params.timesteps != nullptr);
+            static const std::vector<sd::Tensor<float>> empty_ref_latents;
+            return compute(n_threads,
+                           *diffusion_params.x,
+                           *diffusion_params.timesteps,
+                           tensor_or_empty(diffusion_params.context),
+                           diffusion_params.ref_latents ? *diffusion_params.ref_latents : empty_ref_latents,
+                           diffusion_params.increase_ref_index);
         }
 
         void test() {
@@ -601,10 +695,11 @@ namespace ZImage {
             // cuda q8: pass
             // cuda q8 fa: pass
             // ggml_backend_t backend = ggml_backend_cuda_init(0);
-            ggml_backend_t backend    = ggml_backend_cpu_init();
+            ggml_backend_t backend    = sd_backend_cpu_init();
             ggml_type model_data_type = GGML_TYPE_Q8_0;
 
-            ModelLoader model_loader;
+            auto model_manager        = std::make_shared<ModelManager>();
+            ModelLoader& model_loader = model_manager->loader();
             if (!model_loader.init_from_file_and_convert_name(file_path, "model.diffusion_model.")) {
                 LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
                 return;
@@ -620,19 +715,19 @@ namespace ZImage {
             }
 
             std::shared_ptr<ZImageRunner> z_image = std::make_shared<ZImageRunner>(backend,
-                                                                                   backend,
                                                                                    tensor_storage_map,
                                                                                    "model.diffusion_model",
-                                                                                   VERSION_QWEN_IMAGE);
+                                                                                   VERSION_QWEN_IMAGE,
+                                                                                   model_manager);
 
-            z_image->alloc_params_buffer();
-            std::map<std::string, ggml_tensor*> tensors;
-            z_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");
+            if (!model_manager->register_runner_params("ZImage test",
+                                                       *z_image,
+                                                       "model.diffusion_model",
+                                                       ModelManager::ResidencyMode::ParamBackend,
+                                                       backend,
+                                                       backend) ||
+                !model_manager->validate_registered_tensors()) {
+                LOG_ERROR("register z_image tensors with model manager failed");
                 return;
             }
 
@@ -643,4 +738,4 @@ namespace ZImage {
 
 }  // namespace ZImage
 
-#endif  // __Z_IMAGE_HPP__
+#endif  // __SD_MODEL_DIFFUSION_Z_IMAGE_HPP__

src/model/te/clip.hpp

@@ -1,13 +1,13 @@
-#ifndef __CLIP_HPP__
-#define __CLIP_HPP__
+#ifndef __SD_MODEL_TE_CLIP_HPP__
+#define __SD_MODEL_TE_CLIP_HPP__
 
-#include "ggml_extend.hpp"
+#include "core/ggml_extend.hpp"
 #include "model.h"
 #include "tokenizers/clip_tokenizer.h"
 
 /*================================================ FrozenCLIPEmbedder ================================================*/
 
-// Ref: https://github.com/huggingface/transformers/blob/main/src/transformers/models/clip/modeling_clip.py
+// Ref: https://github.com/huggingface/transformers/blob/main/src/transformers/model/clip/modeling_clip.py
 
 struct CLIPMLP : public GGMLBlock {
 protected:
@@ -469,13 +469,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
     std::vector<float> attention_mask_vec;
 
     CLIPTextModelRunner(ggml_backend_t backend,
-                        ggml_backend_t params_backend,
                         const String2TensorStorage& tensor_storage_map,
                         const std::string prefix,
-                        CLIPVersion version = OPENAI_CLIP_VIT_L_14,
-                        bool with_final_ln  = true,
-                        bool force_clip_f32 = false)
-        : GGMLRunner(backend, params_backend) {
+                        CLIPVersion version                                 = OPENAI_CLIP_VIT_L_14,
+                        bool with_final_ln                                  = true,
+                        bool force_clip_f32                                 = false,
+                        std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : GGMLRunner(backend, weight_manager) {
         bool proj_in = false;
         for (const auto& [name, tensor_storage] : tensor_storage_map) {
             if (!starts_with(name, prefix)) {
@@ -567,11 +567,14 @@ struct CLIPTextModelRunner : public GGMLRunner {
                               void* custom_embeddings_data,
                               size_t max_token_idx,
                               bool return_pooled,
-                              int clip_skip) {
+                              int clip_skip,
+                              bool auto_free           = true,
+                              bool free_compute_buffer = true,
+                              bool free_compute_params = true) {
         auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled, clip_skip);
         };
-        auto result = GGMLRunner::compute<float>(get_graph, n_threads, true);
+        auto result = GGMLRunner::compute<float>(get_graph, n_threads, auto_free, free_compute_buffer, free_compute_params);
         if (return_pooled) {
             return take_or_empty(std::move(result));
         }
@@ -579,4 +582,4 @@ struct CLIPTextModelRunner : public GGMLRunner {
     }
 };
 
-#endif  // __CLIP_HPP__
+#endif  // __SD_MODEL_TE_CLIP_HPP__

src/model/te/t5.hpp

@@ -1,5 +1,5 @@
-#ifndef __T5_HPP__
-#define __T5_HPP__
+#ifndef __SD_MODEL_TE_T5_HPP__
+#define __SD_MODEL_TE_T5_HPP__
 
 #include <cfloat>
 #include <limits>
@@ -10,10 +10,33 @@
 #include <string>
 #include <unordered_map>
 
-#include "ggml_extend.hpp"
-#include "model.h"
+#include "core/ggml_extend.hpp"
+#include "model_loader.h"
+#include "model_manager.h"
 #include "tokenizers/t5_unigram_tokenizer.h"
 
+struct T5Config {
+    int64_t num_layers      = 24;
+    int64_t model_dim       = 4096;
+    int64_t ff_dim          = 10240;
+    int64_t num_heads       = 64;
+    int64_t vocab_size      = 32128;
+    bool relative_attention = true;
+
+    static T5Config detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                        const std::string& prefix,
+                                        bool is_umt5 = false) {
+        (void)tensor_storage_map;
+        (void)prefix;
+        T5Config config;
+        if (is_umt5) {
+            config.vocab_size         = 256384;
+            config.relative_attention = false;
+        }
+        return config;
+    }
+};
+
 class T5LayerNorm : public UnaryBlock {
 protected:
     int64_t hidden_size;
@@ -272,30 +295,21 @@ public:
     }
 };
 
-struct T5Params {
-    int64_t num_layers      = 24;
-    int64_t model_dim       = 4096;
-    int64_t ff_dim          = 10240;
-    int64_t num_heads       = 64;
-    int64_t vocab_size      = 32128;
-    bool relative_attention = true;
-};
-
 struct T5 : public GGMLBlock {
-    T5Params params;
+    T5Config config;
 
 public:
     T5() {}
-    T5(T5Params params)
-        : params(params) {
-        blocks["encoder"] = std::shared_ptr<GGMLBlock>(new T5Stack(params.num_layers,
-                                                                   params.model_dim,
-                                                                   params.model_dim,
-                                                                   params.ff_dim,
-                                                                   params.num_heads,
-                                                                   params.relative_attention));
-        blocks["shared"]  = std::shared_ptr<GGMLBlock>(new Embedding(params.vocab_size,
-                                                                     params.model_dim));
+    T5(T5Config config)
+        : config(config) {
+        blocks["encoder"] = std::shared_ptr<GGMLBlock>(new T5Stack(config.num_layers,
+                                                                   config.model_dim,
+                                                                   config.model_dim,
+                                                                   config.ff_dim,
+                                                                   config.num_heads,
+                                                                   config.relative_attention));
+        blocks["shared"]  = std::shared_ptr<GGMLBlock>(new Embedding(config.vocab_size,
+                                                                     config.model_dim));
     }
 
     ggml_tensor* forward(GGMLRunnerContext* ctx,
@@ -316,21 +330,18 @@ public:
 };
 
 struct T5Runner : public GGMLRunner {
-    T5Params params;
+    T5Config config;
     T5 model;
     std::vector<int> relative_position_bucket_vec;
 
     T5Runner(ggml_backend_t backend,
-             ggml_backend_t params_backend,
              const String2TensorStorage& tensor_storage_map,
              const std::string prefix,
-             bool is_umt5 = false)
-        : GGMLRunner(backend, params_backend) {
-        if (is_umt5) {
-            params.vocab_size         = 256384;
-            params.relative_attention = false;
-        }
-        model = T5(params);
+             bool is_umt5                                        = false,
+             std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : GGMLRunner(backend, weight_manager),
+          config(T5Config::detect_from_weights(tensor_storage_map, prefix, is_umt5)) {
+        model = T5(config);
         model.init(params_ctx, tensor_storage_map, prefix);
     }
 
@@ -384,11 +395,14 @@ struct T5Runner : public GGMLRunner {
 
     sd::Tensor<float> compute(const int n_threads,
                               const sd::Tensor<int32_t>& input_ids,
-                              const sd::Tensor<float>& attention_mask) {
+                              const sd::Tensor<float>& attention_mask,
+                              bool auto_free           = true,
+                              bool free_compute_buffer = true,
+                              bool free_compute_params = true) {
         auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(input_ids, attention_mask);
         };
-        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, true), 3);
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, auto_free, free_compute_buffer, free_compute_params), 3);
     }
 
     static std::vector<int> _relative_position_bucket(const std::vector<int>& relative_position,
@@ -464,21 +478,17 @@ struct T5Embedder {
     T5Runner model;
 
     T5Embedder(ggml_backend_t backend,
-               ggml_backend_t params_backend,
-               const String2TensorStorage& tensor_storage_map = {},
-               const std::string prefix                       = "",
-               bool is_umt5                                   = false)
-        : model(backend, params_backend, tensor_storage_map, prefix, is_umt5), tokenizer(is_umt5) {
+               const String2TensorStorage& tensor_storage_map      = {},
+               const std::string prefix                            = "",
+               bool is_umt5                                        = false,
+               std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : model(backend, tensor_storage_map, prefix, is_umt5, weight_manager), tokenizer(is_umt5) {
     }
 
     void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
         model.get_param_tensors(tensors, prefix);
     }
 
-    void alloc_params_buffer() {
-        model.alloc_params_buffer();
-    }
-
     std::tuple<std::vector<int>, std::vector<float>, std::vector<float>> tokenize(std::string text,
                                                                                   size_t max_length = 0,
                                                                                   bool padding      = false) {
@@ -560,10 +570,11 @@ struct T5Embedder {
         // cuda f32: pass
         // cuda q8_0: pass
         // ggml_backend_t backend = ggml_backend_cuda_init(0);
-        ggml_backend_t backend    = ggml_backend_cpu_init();
+        ggml_backend_t backend    = sd_backend_cpu_init();
         ggml_type model_data_type = GGML_TYPE_F16;
 
-        ModelLoader model_loader;
+        auto model_manager        = std::make_shared<ModelManager>();
+        ModelLoader& model_loader = model_manager->loader();
         if (!model_loader.init_from_file_and_convert_name(file_path)) {
             LOG_ERROR("init model loader from file failed: '%s'", file_path.c_str());
             return;
@@ -576,16 +587,16 @@ struct T5Embedder {
             }
         }
 
-        std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, backend, tensor_storage_map, "", true);
+        std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, tensor_storage_map, "", true, model_manager);
 
-        t5->alloc_params_buffer();
-        std::map<std::string, ggml_tensor*> tensors;
-        t5->get_param_tensors(tensors, "");
-
-        bool success = model_loader.load_tensors(tensors);
-
-        if (!success) {
-            LOG_ERROR("load tensors from model loader failed");
+        if (!model_manager->register_runner_params("T5 test",
+                                                   *t5,
+                                                   "",
+                                                   ModelManager::ResidencyMode::ParamBackend,
+                                                   backend,
+                                                   backend) ||
+            !model_manager->validate_registered_tensors()) {
+            LOG_ERROR("register t5 tensors with model manager failed");
             return;
         }
 
@@ -594,4 +605,4 @@ struct T5Embedder {
     }
 };
 
-#endif  // __T5_HPP__
+#endif  // __SD_MODEL_TE_T5_HPP__

src/model/upscaler/esrgan.hpp

@@ -0,0 +1,273 @@
+#ifndef __SD_MODEL_UPSCALER_ESRGAN_HPP__
+#define __SD_MODEL_UPSCALER_ESRGAN_HPP__
+
+#include <algorithm>
+#include <map>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "core/util.h"
+
+/*
+    ===================================    ESRGAN  ===================================
+    References:
+    https://github.com/xinntao/Real-ESRGAN/blob/master/inference_realesrgan.py
+    https://github.com/XPixelGroup/BasicSR/blob/v1.4.2/basicsr/archs/rrdbnet_arch.py
+
+*/
+
+struct ESRGANConfig {
+    int scale       = 4;
+    int num_block   = 23;
+    int num_in_ch   = 3;
+    int num_out_ch  = 3;
+    int num_feat    = 64;
+    int num_grow_ch = 32;
+
+    static ESRGANConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                            const std::string& prefix = "") {
+        ESRGANConfig config;
+        auto find_weight = [&](const std::string& suffix) -> const TensorStorage* {
+            std::string name = prefix.empty() ? suffix : prefix + "." + suffix;
+            auto iter        = tensor_storage_map.find(name);
+            if (iter == tensor_storage_map.end()) {
+                return nullptr;
+            }
+            return &iter->second;
+        };
+
+        int detected_num_block        = 0;
+        const std::string body_prefix = prefix.empty() ? "body." : prefix + ".body.";
+        for (const auto& [name, _] : tensor_storage_map) {
+            if (!starts_with(name, body_prefix)) {
+                continue;
+            }
+            size_t pos = name.find('.', body_prefix.size());
+            if (pos == std::string::npos) {
+                continue;
+            }
+            try {
+                int idx            = std::stoi(name.substr(body_prefix.size(), pos - body_prefix.size()));
+                detected_num_block = std::max(detected_num_block, idx + 1);
+            } catch (...) {
+            }
+        }
+        if (detected_num_block > 0) {
+            config.num_block = detected_num_block;
+        }
+
+        bool has_conv_up2 = find_weight("conv_up2.weight") != nullptr;
+        bool has_conv_up1 = find_weight("conv_up1.weight") != nullptr;
+        bool has_model_tensor =
+            detected_num_block > 0 ||
+            find_weight("conv_first.weight") != nullptr ||
+            find_weight("conv_hr.weight") != nullptr ||
+            find_weight("conv_last.weight") != nullptr;
+        if (has_conv_up2) {
+            config.scale = 4;
+        } else if (has_conv_up1) {
+            config.scale = 2;
+        } else if (has_model_tensor) {
+            config.scale = 1;
+        }
+
+        if (has_model_tensor || has_conv_up1 || has_conv_up2) {
+            LOG_DEBUG("esrgan: scale = %d, num_block = %d, num_in_ch = %d, num_out_ch = %d, num_feat = %d, num_grow_ch = %d",
+                      config.scale,
+                      config.num_block,
+                      config.num_in_ch,
+                      config.num_out_ch,
+                      config.num_feat,
+                      config.num_grow_ch);
+        }
+        return config;
+    }
+};
+
+class ResidualDenseBlock : public GGMLBlock {
+protected:
+    int num_feat;
+    int num_grow_ch;
+
+public:
+    ResidualDenseBlock(int num_feat = 64, int num_grow_ch = 32)
+        : num_feat(num_feat), num_grow_ch(num_grow_ch) {
+        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv3"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 2 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv4"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 3 * num_grow_ch, num_grow_ch, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv5"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 4 * num_grow_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    ggml_tensor* lrelu(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        return ggml_leaky_relu(ctx->ggml_ctx, x, 0.2f, true);
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
+
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
+        auto conv3 = std::dynamic_pointer_cast<Conv2d>(blocks["conv3"]);
+        auto conv4 = std::dynamic_pointer_cast<Conv2d>(blocks["conv4"]);
+        auto conv5 = std::dynamic_pointer_cast<Conv2d>(blocks["conv5"]);
+
+        auto x1    = lrelu(ctx, conv1->forward(ctx, x));
+        auto x_cat = ggml_concat(ctx->ggml_ctx, x, x1, 2);
+        auto x2    = lrelu(ctx, conv2->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx->ggml_ctx, x_cat, x2, 2);
+        auto x3    = lrelu(ctx, conv3->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx->ggml_ctx, x_cat, x3, 2);
+        auto x4    = lrelu(ctx, conv4->forward(ctx, x_cat));
+        x_cat      = ggml_concat(ctx->ggml_ctx, x_cat, x4, 2);
+        auto x5    = conv5->forward(ctx, x_cat);
+
+        x5 = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, x5, 0.2f), x);
+        return x5;
+    }
+};
+
+class RRDB : public GGMLBlock {
+public:
+    RRDB(int num_feat, int num_grow_ch = 32) {
+        blocks["rdb1"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+        blocks["rdb2"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+        blocks["rdb3"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        // x: [n, num_feat, h, w]
+        // return: [n, num_feat, h, w]
+
+        auto rdb1 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb1"]);
+        auto rdb2 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb2"]);
+        auto rdb3 = std::dynamic_pointer_cast<ResidualDenseBlock>(blocks["rdb3"]);
+
+        auto out = rdb1->forward(ctx, x);
+        out      = rdb2->forward(ctx, out);
+        out      = rdb3->forward(ctx, out);
+
+        out = ggml_add(ctx->ggml_ctx, ggml_ext_scale(ctx->ggml_ctx, out, 0.2f), x);
+        return out;
+    }
+};
+
+class RRDBNet : public GGMLBlock {
+protected:
+    ESRGANConfig config;
+
+public:
+    explicit RRDBNet(ESRGANConfig config)
+        : config(std::move(config)) {
+        blocks["conv_first"] = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.num_in_ch, this->config.num_feat, {3, 3}, {1, 1}, {1, 1}));
+        for (int i = 0; i < this->config.num_block; i++) {
+            std::string name = "body." + std::to_string(i);
+            blocks[name]     = std::shared_ptr<GGMLBlock>(new RRDB(this->config.num_feat, this->config.num_grow_ch));
+        }
+        blocks["conv_body"] = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.num_feat, this->config.num_feat, {3, 3}, {1, 1}, {1, 1}));
+        if (this->config.scale >= 2) {
+            blocks["conv_up1"] = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.num_feat, this->config.num_feat, {3, 3}, {1, 1}, {1, 1}));
+        }
+        if (this->config.scale == 4) {
+            blocks["conv_up2"] = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.num_feat, this->config.num_feat, {3, 3}, {1, 1}, {1, 1}));
+        }
+        blocks["conv_hr"]   = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.num_feat, this->config.num_feat, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv_last"] = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.num_feat, this->config.num_out_ch, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    int get_scale() { return config.scale; }
+    int get_num_block() { return config.num_block; }
+
+    ggml_tensor* lrelu(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        return ggml_leaky_relu(ctx->ggml_ctx, x, 0.2f, true);
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        // x: [n, num_in_ch, h, w]
+        // 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_hr    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_hr"]);
+        auto conv_last  = std::dynamic_pointer_cast<Conv2d>(blocks["conv_last"]);
+
+        auto feat = conv_first->forward(ctx, x);
+        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.prelude", "feat");
+        auto body_feat = feat;
+        for (int i = 0; i < config.num_block; i++) {
+            std::string name = "body." + std::to_string(i);
+            auto block       = std::dynamic_pointer_cast<RRDB>(blocks[name]);
+
+            body_feat = block->forward(ctx, body_feat);
+            sd::ggml_graph_cut::mark_graph_cut(body_feat, "esrgan.body." + std::to_string(i), "feat");
+        }
+        body_feat = conv_body->forward(ctx, body_feat);
+        feat      = ggml_add(ctx->ggml_ctx, feat, body_feat);
+        sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.body.out", "feat");
+        // upsample
+        if (config.scale >= 2) {
+            auto conv_up1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up1"]);
+            feat          = lrelu(ctx, conv_up1->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+            sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up1", "feat");
+            if (config.scale == 4) {
+                auto conv_up2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv_up2"]);
+                feat          = lrelu(ctx, conv_up2->forward(ctx, ggml_upscale(ctx->ggml_ctx, feat, 2, GGML_SCALE_MODE_NEAREST)));
+                sd::ggml_graph_cut::mark_graph_cut(feat, "esrgan.up2", "feat");
+            }
+        }
+        // for all scales
+        auto out = conv_last->forward(ctx, lrelu(ctx, conv_hr->forward(ctx, feat)));
+        sd::ggml_graph_cut::mark_graph_cut(out, "esrgan.final", "out");
+        return out;
+    }
+};
+
+struct ESRGAN : public GGMLRunner {
+    ESRGANConfig config;
+    std::unique_ptr<RRDBNet> rrdb_net;
+
+    ESRGAN(ggml_backend_t backend,
+           const String2TensorStorage& tensor_storage_map      = {},
+           std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : GGMLRunner(backend, weight_manager),
+          config(ESRGANConfig::detect_from_weights(tensor_storage_map)),
+          rrdb_net(std::make_unique<RRDBNet>(config)) {
+        rrdb_net->init(params_ctx, tensor_storage_map, "");
+    }
+
+    std::string get_desc() override {
+        return "esrgan";
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {
+        if (!rrdb_net) {
+            return;
+        }
+
+        rrdb_net->get_param_tensors(tensors);
+    }
+
+    ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor) {
+        if (!rrdb_net)
+            return nullptr;
+        constexpr int kGraphNodes = 1 << 16;  // 65k
+        ggml_cgraph* gf           = new_graph_custom(kGraphNodes);
+        ggml_tensor* x            = make_input(x_tensor);
+
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = rrdb_net->forward(&runner_ctx, x);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    sd::Tensor<float> compute(const int n_threads,
+                              const sd::Tensor<float>& x) {
+        auto get_graph = [&]() -> ggml_cgraph* { return build_graph(x); };
+        auto result    = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), x.dim());
+        return result;
+    }
+};
+
+#endif  // __SD_MODEL_UPSCALER_ESRGAN_HPP__

src/model/upscaler/ltx_latent_upscaler.hpp

@@ -0,0 +1,508 @@
+#ifndef __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__
+#define __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__
+
+#include <algorithm>
+#include <cinttypes>
+#include <cmath>
+#include <map>
+#include <memory>
+#include <set>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "core/ggml_extend.hpp"
+#include "core/ggml_graph_cut.h"
+#include "core/util.h"
+#include "model/diffusion/dit.hpp"
+#include "model_loader.h"
+
+namespace LTXVUpsampler {
+    constexpr int LTX_UPSAMPLER_GRAPH_SIZE = 10240;
+
+    struct LatentUpsamplerConfig {
+        int64_t in_channels      = 128;
+        int64_t mid_channels     = 1024;
+        int num_blocks_per_stage = 4;
+        int dims                 = 3;
+        bool spatial_upsample    = true;
+        bool temporal_upsample   = false;
+        bool rational_resampler  = false;
+        float spatial_scale      = 2.f;
+        int spatial_up_num       = 2;
+        int spatial_down_den     = 1;
+        int temporal_up_factor   = 1;
+
+        static LatentUpsamplerConfig detect_from_weights(const String2TensorStorage& tensor_storage_map,
+                                                         const std::string& prefix = "") {
+            LatentUpsamplerConfig config;
+            auto find_weight = [&](const std::string& suffix) -> const TensorStorage* {
+                std::string name = prefix.empty() ? suffix : prefix + "." + suffix;
+                auto iter        = tensor_storage_map.find(name);
+                if (iter == tensor_storage_map.end()) {
+                    return nullptr;
+                }
+                return &iter->second;
+            };
+
+            bool inferred = false;
+
+            const TensorStorage* initial_norm = find_weight("initial_norm.weight");
+            if (initial_norm != nullptr) {
+                config.mid_channels = initial_norm->ne[0];
+                inferred            = true;
+            }
+
+            const TensorStorage* final_conv = find_weight("final_conv.bias");
+            if (final_conv != nullptr) {
+                config.in_channels = final_conv->ne[0];
+                inferred           = true;
+            }
+
+            int detected_blocks                 = 0;
+            const std::string res_blocks_prefix = prefix.empty() ? "res_blocks." : prefix + ".res_blocks.";
+            for (const auto& [name, _] : tensor_storage_map) {
+                if (!starts_with(name, res_blocks_prefix)) {
+                    continue;
+                }
+                size_t begin = res_blocks_prefix.size();
+                size_t end   = name.find('.', begin);
+                if (end == std::string::npos) {
+                    continue;
+                }
+                try {
+                    int idx         = std::stoi(name.substr(begin, end - begin));
+                    detected_blocks = std::max(detected_blocks, idx + 1);
+                } catch (...) {
+                }
+            }
+            if (detected_blocks > 0) {
+                config.num_blocks_per_stage = detected_blocks;
+                inferred                    = true;
+            }
+
+            const TensorStorage* rational_upsampler_weight = find_weight("upsampler.conv.weight");
+            const TensorStorage* upsampler_bias            = find_weight("upsampler.0.bias");
+            config.rational_resampler                      = rational_upsampler_weight != nullptr;
+            int64_t upsampler_out_channels                 = upsampler_bias == nullptr ? 0 : upsampler_bias->ne[0];
+            config.spatial_upsample                        = config.rational_resampler || upsampler_out_channels == 4 * config.mid_channels;
+            config.temporal_upsample                       = upsampler_out_channels == 2 * config.mid_channels;
+            if (config.rational_resampler || upsampler_out_channels > 0) {
+                inferred = true;
+            }
+            if (config.temporal_upsample) {
+                config.temporal_up_factor = 2;
+            }
+            if (rational_upsampler_weight != nullptr) {
+                int64_t out_channels = rational_upsampler_weight->ne[3];
+                if (config.mid_channels > 0 && out_channels % config.mid_channels == 0) {
+                    int64_t ratio = out_channels / config.mid_channels;
+                    int num       = static_cast<int>(std::round(std::sqrt(static_cast<double>(ratio))));
+                    if (num > 0 && static_cast<int64_t>(num) * num == ratio) {
+                        config.spatial_up_num = num;
+                    }
+                }
+                if (config.spatial_up_num == 3) {
+                    config.spatial_down_den = 2;
+                    config.spatial_scale    = 1.5f;
+                } else if (config.spatial_up_num == 4) {
+                    config.spatial_down_den = 1;
+                    config.spatial_scale    = 4.f;
+                } else {
+                    config.spatial_down_den = 1;
+                    config.spatial_scale    = static_cast<float>(config.spatial_up_num);
+                }
+            }
+
+            if (inferred) {
+                LOG_DEBUG("ltx latent upsampler: in_channels = %" PRId64 ", mid_channels = %" PRId64 ", num_blocks_per_stage = %d, spatial_scale = %.3f, temporal_up_factor = %d, rational_resampler = %d",
+                          config.in_channels,
+                          config.mid_channels,
+                          config.num_blocks_per_stage,
+                          config.spatial_scale,
+                          config.temporal_up_factor,
+                          config.rational_resampler);
+            }
+            return config;
+        }
+    };
+
+    class VideoGroupNorm : public GGMLBlock {
+    protected:
+        int num_groups;
+        int64_t num_channels;
+        float eps;
+        std::string prefix;
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            SD_UNUSED(tensor_storage_map);
+            this->prefix     = prefix;
+            params["weight"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_channels);
+            params["bias"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_channels);
+        }
+
+    public:
+        VideoGroupNorm(int num_groups, int64_t num_channels, float eps = 1e-05f)
+            : num_groups(num_groups),
+              num_channels(num_channels),
+              eps(eps) {}
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            // LTX video latent layout is [W, H, T, C]. ggml_group_norm treats ne[2]
+            // as channels, so fold only H/T internally and restore the same layout.
+            GGML_ASSERT(x->ne[3] == num_channels);
+            const int64_t W = x->ne[0];
+            const int64_t H = x->ne[1];
+            const int64_t T = x->ne[2];
+            x               = ggml_ext_cont(ctx->ggml_ctx, x);
+            x               = ggml_reshape_4d(ctx->ggml_ctx, x, W, H * T, num_channels, 1);
+            x               = ggml_group_norm(ctx->ggml_ctx, x, num_groups, eps);
+
+            ggml_tensor* weight = params["weight"];
+            ggml_tensor* bias   = params["bias"];
+            if (ctx->weight_adapter) {
+                weight = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, weight, prefix + "weight");
+                bias   = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, bias, prefix + "bias");
+            }
+            weight = ggml_reshape_4d(ctx->ggml_ctx, weight, 1, 1, num_channels, 1);
+            bias   = ggml_reshape_4d(ctx->ggml_ctx, bias, 1, 1, num_channels, 1);
+            x      = ggml_mul_inplace(ctx->ggml_ctx, x, weight);
+            x      = ggml_add_inplace(ctx->ggml_ctx, x, bias);
+            return ggml_reshape_4d(ctx->ggml_ctx, x, W, H, T, num_channels);
+        }
+    };
+
+    class ResBlock : public GGMLBlock {
+    public:
+        ResBlock(int64_t channels, int dims = 3) {
+            GGML_ASSERT(dims == 3);
+            blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv3d(channels, channels, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
+            blocks["norm1"] = std::shared_ptr<GGMLBlock>(new VideoGroupNorm(32, channels));
+            blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv3d(channels, channels, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
+            blocks["norm2"] = std::shared_ptr<GGMLBlock>(new VideoGroupNorm(32, channels));
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto conv1 = std::dynamic_pointer_cast<Conv3d>(blocks["conv1"]);
+            auto norm1 = std::dynamic_pointer_cast<VideoGroupNorm>(blocks["norm1"]);
+            auto conv2 = std::dynamic_pointer_cast<Conv3d>(blocks["conv2"]);
+            auto norm2 = std::dynamic_pointer_cast<VideoGroupNorm>(blocks["norm2"]);
+
+            ggml_tensor* residual = x;
+
+            x = conv1->forward(ctx, x);
+            x = norm1->forward(ctx, x);
+            x = ggml_silu_inplace(ctx->ggml_ctx, x);
+            x = conv2->forward(ctx, x);
+            x = norm2->forward(ctx, x);
+            x = ggml_add(ctx->ggml_ctx, x, residual);
+            return ggml_silu(ctx->ggml_ctx, x);
+        }
+    };
+
+    class PixelShuffleND : public UnaryBlock {
+    protected:
+        int upscale_factor;
+
+    public:
+        explicit PixelShuffleND(int upscale_factor)
+            : upscale_factor(upscale_factor) {}
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            GGML_ASSERT(upscale_factor > 0);
+            int64_t h = x->ne[1];
+            int64_t w = x->ne[0];
+            GGML_ASSERT(x->ne[2] % (upscale_factor * upscale_factor) == 0);
+            // x: [b*f, c*p1*p2, h, w] -> [b*f, c, h*p1, w*p2]
+            x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 2, 0, 1, 3));  // [b*f, h, w, c*p1*p2]
+            x = ggml_reshape_3d(ctx->ggml_ctx, x, x->ne[0], x->ne[1] * x->ne[2], x->ne[3]);          // [b*f, h*w, c*p1*p2]
+            return DiT::unpatchify(ctx->ggml_ctx, x, h, w, upscale_factor, upscale_factor, true);
+        }
+    };
+
+    class TemporalPixelShuffleND : public UnaryBlock {
+    protected:
+        int upscale_factor;
+
+    public:
+        explicit TemporalPixelShuffleND(int upscale_factor)
+            : upscale_factor(upscale_factor) {}
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+            GGML_ASSERT(upscale_factor > 0);
+            GGML_ASSERT(x->ne[3] % upscale_factor == 0);
+            const int64_t W = x->ne[0];
+            const int64_t H = x->ne[1];
+            const int64_t F = x->ne[2];
+            const int64_t C = x->ne[3] / upscale_factor;
+
+            // x: [b, c*p, f, h, w] -> [b, c, f*p, h, w]
+            x = ggml_ext_cont(ctx->ggml_ctx, x);
+            x = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, F, upscale_factor, C);
+            x = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));
+            return ggml_reshape_4d(ctx->ggml_ctx, x, W, H, F * upscale_factor, C);
+        }
+    };
+
+    class BlurDownsample : public GGMLBlock {
+    protected:
+        int64_t channels;
+        int stride;
+        std::vector<float> kernel_data;
+        std::string kernel_name;
+
+        void init_params(ggml_context* ctx,
+                         const String2TensorStorage& tensor_storage_map = {},
+                         const std::string prefix                       = "") override {
+            SD_UNUSED(ctx);
+            SD_UNUSED(tensor_storage_map);
+            if (stride == 1) {
+                return;
+            }
+            kernel_name = prefix + "kernel";
+        }
+
+    public:
+        BlurDownsample(int64_t channels, int stride)
+            : channels(channels),
+              stride(stride) {
+            GGML_ASSERT(stride >= 1);
+            static const float binomial[5] = {1.f, 4.f, 6.f, 4.f, 1.f};
+            kernel_data.resize(static_cast<size_t>(5 * 5 * channels));
+            for (int64_t c = 0; c < channels; ++c) {
+                for (int y = 0; y < 5; ++y) {
+                    for (int x = 0; x < 5; ++x) {
+                        kernel_data[static_cast<size_t>(x + 5 * (y + 5 * c))] =
+                            binomial[y] * binomial[x] / 256.f;
+                    }
+                }
+            }
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            if (stride == 1) {
+                return x;
+            }
+            GGML_ASSERT(ctx != nullptr);
+            GGML_ASSERT(!kernel_data.empty());
+            GGML_ASSERT(x->ne[2] == channels);
+            ggml_tensor* kernel = ggml_new_tensor_4d(ctx->ggml_ctx, GGML_TYPE_F32, 5, 5, 1, channels);
+            ggml_set_name(kernel, kernel_name.empty() ? "blur_down.kernel" : kernel_name.c_str());
+            ctx->bind_backend_tensor_data(kernel, kernel_data.data());
+            if (ctx->conv2d_direct_enabled) {
+                return ggml_conv_2d_dw_direct(ctx->ggml_ctx, kernel, x, stride, stride, 2, 2, 1, 1);
+            }
+            return ggml_conv_2d_dw(ctx->ggml_ctx, kernel, x, stride, stride, 2, 2, 1, 1);
+        }
+    };
+
+    class SpatialRationalResampler : public GGMLBlock {
+    protected:
+        int64_t mid_channels;
+        int num;
+        int den;
+
+    public:
+        SpatialRationalResampler(int64_t mid_channels, int num, int den)
+            : mid_channels(mid_channels),
+              num(num),
+              den(den) {
+            GGML_ASSERT(num >= 1);
+            GGML_ASSERT(den >= 1);
+            blocks["conv"]          = std::shared_ptr<GGMLBlock>(new Conv2d(mid_channels, num * num * mid_channels, {3, 3}, {1, 1}, {1, 1}));
+            blocks["pixel_shuffle"] = std::shared_ptr<GGMLBlock>(new PixelShuffleND(num));
+            blocks["blur_down"]     = std::shared_ptr<GGMLBlock>(new BlurDownsample(mid_channels, den));
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto conv          = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
+            auto pixel_shuffle = std::dynamic_pointer_cast<PixelShuffleND>(blocks["pixel_shuffle"]);
+            auto blur_down     = std::dynamic_pointer_cast<BlurDownsample>(blocks["blur_down"]);
+
+            // rearrange(x, "b c f h w -> (b f) c h w")
+            x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+            x = conv->forward(ctx, x);
+            x = pixel_shuffle->forward(ctx, x);
+            x = blur_down->forward(ctx, x);
+            return ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
+        }
+    };
+
+    class LatentUpsampler : public GGMLBlock {
+    public:
+        LatentUpsamplerConfig config;
+
+        explicit LatentUpsampler(LatentUpsamplerConfig config)
+            : config(std::move(config)) {
+            GGML_ASSERT(this->config.dims == 3);
+            GGML_ASSERT(this->config.spatial_upsample || this->config.temporal_upsample);
+
+            blocks["initial_conv"] = std::shared_ptr<GGMLBlock>(new Conv3d(this->config.in_channels,
+                                                                           this->config.mid_channels,
+                                                                           {3, 3, 3},
+                                                                           {1, 1, 1},
+                                                                           {1, 1, 1}));
+            blocks["initial_norm"] = std::shared_ptr<GGMLBlock>(new VideoGroupNorm(32, this->config.mid_channels));
+            for (int i = 0; i < this->config.num_blocks_per_stage; ++i) {
+                blocks["res_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new ResBlock(this->config.mid_channels, this->config.dims));
+            }
+            if (this->config.rational_resampler) {
+                blocks["upsampler"] = std::shared_ptr<GGMLBlock>(new SpatialRationalResampler(this->config.mid_channels,
+                                                                                              this->config.spatial_up_num,
+                                                                                              this->config.spatial_down_den));
+            } else if (this->config.temporal_upsample) {
+                blocks["upsampler.0"] = std::shared_ptr<GGMLBlock>(new Conv3d(this->config.mid_channels,
+                                                                              this->config.temporal_up_factor * this->config.mid_channels,
+                                                                              {3, 3, 3},
+                                                                              {1, 1, 1},
+                                                                              {1, 1, 1}));
+                blocks["upsampler.1"] = std::shared_ptr<GGMLBlock>(new TemporalPixelShuffleND(this->config.temporal_up_factor));
+            } else {
+                blocks["upsampler.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(this->config.mid_channels,
+                                                                              4 * this->config.mid_channels,
+                                                                              {3, 3},
+                                                                              {1, 1},
+                                                                              {1, 1}));
+                blocks["upsampler.1"] = std::shared_ptr<GGMLBlock>(new PixelShuffleND(2));
+            }
+            for (int i = 0; i < this->config.num_blocks_per_stage; ++i) {
+                blocks["post_upsample_res_blocks." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new ResBlock(this->config.mid_channels, this->config.dims));
+            }
+            blocks["final_conv"] = std::shared_ptr<GGMLBlock>(new Conv3d(this->config.mid_channels,
+                                                                         this->config.in_channels,
+                                                                         {3, 3, 3},
+                                                                         {1, 1, 1},
+                                                                         {1, 1, 1}));
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            // x: [b, c, f, h, w]
+            // return: [b, c, scaled_f, scaled_h, scaled_w]
+            auto initial_conv = std::dynamic_pointer_cast<Conv3d>(blocks["initial_conv"]);
+            auto initial_norm = std::dynamic_pointer_cast<VideoGroupNorm>(blocks["initial_norm"]);
+            auto final_conv   = std::dynamic_pointer_cast<Conv3d>(blocks["final_conv"]);
+
+            x = initial_conv->forward(ctx, x);
+            x = initial_norm->forward(ctx, x);
+            x = ggml_silu(ctx->ggml_ctx, x);
+            sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.initial", "x");
+
+            for (int i = 0; i < config.num_blocks_per_stage; ++i) {
+                auto block = std::dynamic_pointer_cast<ResBlock>(blocks["res_blocks." + std::to_string(i)]);
+                x          = block->forward(ctx, x);
+                sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.res_blocks." + std::to_string(i), "x");
+            }
+
+            if (config.rational_resampler) {
+                auto upsampler = std::dynamic_pointer_cast<SpatialRationalResampler>(blocks["upsampler"]);
+                x              = upsampler->forward(ctx, x);
+            } else if (config.temporal_upsample) {
+                auto upsample_conv = std::dynamic_pointer_cast<Conv3d>(blocks["upsampler.0"]);
+                auto pixel_shuffle = std::dynamic_pointer_cast<TemporalPixelShuffleND>(blocks["upsampler.1"]);
+                x                  = upsample_conv->forward(ctx, x);                    // [b, c*2, f, h, w]
+                x                  = pixel_shuffle->forward(ctx, x);                    // [b, c, f*2, h, w]
+                x                  = ggml_ext_slice(ctx->ggml_ctx, x, 2, 1, x->ne[2]);  // x[:, :, 1:, :, :]
+            } else {
+                auto upsample_conv = std::dynamic_pointer_cast<Conv2d>(blocks["upsampler.0"]);
+                auto pixel_shuffle = std::dynamic_pointer_cast<PixelShuffleND>(blocks["upsampler.1"]);
+
+                // rearrange(x, "b c f h w -> (b f) c h w"),
+                x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));  // [b*f, c, h, w]
+                x = upsample_conv->forward(ctx, x);                                                      // [b*f, c*4, h, w]
+                x = pixel_shuffle->forward(ctx, x);                                                      // [b*f, c, h*2, w*2]
+                x = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));  // [b*c, f, h, w]
+            }
+            sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.spatial_up", "x");
+
+            for (int i = 0; i < config.num_blocks_per_stage; ++i) {
+                auto block = std::dynamic_pointer_cast<ResBlock>(blocks["post_upsample_res_blocks." + std::to_string(i)]);
+                x          = block->forward(ctx, x);
+                sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.post_blocks." + std::to_string(i), "x");
+            }
+
+            x = final_conv->forward(ctx, x);
+            sd::ggml_graph_cut::mark_graph_cut(x, "ltx_latent_upsampler.final", "x");
+            return x;
+        }
+    };
+
+    struct LatentUpsamplerRunner : public GGMLRunner {
+        LatentUpsamplerConfig config;
+        std::unique_ptr<LatentUpsampler> model;
+
+        LatentUpsamplerRunner(ggml_backend_t backend,
+                              const String2TensorStorage& tensor_storage_map,
+                              std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : GGMLRunner(backend, weight_manager),
+              config(LatentUpsamplerConfig::detect_from_weights(tensor_storage_map)) {
+            if (config.dims != 3 || (!config.spatial_upsample && !config.temporal_upsample) ||
+                config.spatial_up_num < 1 || config.spatial_down_den < 1 || config.temporal_up_factor < 1) {
+                LOG_ERROR("unsupported LTX latent upsampler config: dims=%d spatial=%d temporal=%d rational=%d scale=%.3f temporal_factor=%d",
+                          config.dims,
+                          config.spatial_upsample,
+                          config.temporal_upsample,
+                          config.rational_resampler,
+                          config.spatial_scale,
+                          config.temporal_up_factor);
+                return;
+            }
+
+            model = std::make_unique<LatentUpsampler>(config);
+            model->init(params_ctx, tensor_storage_map, "");
+        }
+
+        std::string get_desc() override {
+            return "ltx_latent_upsampler";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {
+            if (model) {
+                model->get_param_tensors(tensors);
+            }
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor) {
+            if (!model) {
+                return nullptr;
+            }
+            ggml_cgraph* gf  = new_graph_custom(LTX_UPSAMPLER_GRAPH_SIZE);
+            ggml_tensor* x   = make_input(x_tensor);
+            auto runner_ctx  = get_context();
+            ggml_tensor* out = model->forward(&runner_ctx, x);
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(const int n_threads,
+                                  const sd::Tensor<float>& x) {
+            if (!model) {
+                LOG_ERROR("LTX latent upsampler is not loaded");
+                return {};
+            }
+            if (x.dim() != 4 && x.dim() != 5) {
+                LOG_ERROR("LTX latent upsampler expects 4D or 5D video latent, got dim=%lld",
+                          (long long)x.dim());
+                return {};
+            }
+            if (x.dim() == 5 && x.shape()[4] != 1) {
+                LOG_ERROR("LTX latent upsampler currently supports batch size 1, got batch=%lld",
+                          (long long)x.shape()[4]);
+                return {};
+            }
+            if (x.shape()[3] != config.in_channels) {
+                LOG_ERROR("LTX latent upsampler expected %" PRId64 " channels, got %lld",
+                          config.in_channels,
+                          (long long)x.shape()[3]);
+                return {};
+            }
+            size_t expected_dim = static_cast<size_t>(x.dim());
+            auto get_graph      = [&]() -> ggml_cgraph* { return build_graph(x); };
+            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), expected_dim);
+        }
+    };
+
+}  // namespace LTXVUpsampler
+
+#endif  // __SD_MODEL_UPSCALER_LTX_LATENT_UPSCALER_HPP__

src/model/vae/auto_encoder_kl.hpp

@@ -1,7 +1,7 @@
-#ifndef __AUTO_ENCODER_KL_HPP__
-#define __AUTO_ENCODER_KL_HPP__
+#ifndef __SD_MODEL_VAE_AUTO_ENCODER_KL_HPP__
+#define __SD_MODEL_VAE_AUTO_ENCODER_KL_HPP__
 
-#include "vae.hpp"
+#include "model/vae/vae.hpp"
 
 /*================================================== AutoEncoderKL ===================================================*/
 
@@ -213,9 +213,9 @@ protected:
         params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
     }
 
-    float get_alpha() {
-        float alpha = ggml_ext_backend_tensor_get_f32(params["mix_factor"]);
-        return sigmoid(alpha);
+    ggml_tensor* get_alpha(GGMLRunnerContext* ctx) {
+        auto mix_factor = ggml_ext_cast_f32(ctx->ggml_ctx, ctx->backend, params["mix_factor"]);
+        return ggml_sigmoid(ctx->ggml_ctx, mix_factor);
     }
 
 public:
@@ -250,10 +250,12 @@ public:
 
         x = time_stack->forward(ctx, x);  // b t c (h w)
 
-        float alpha = get_alpha();
-        x           = ggml_add(ctx->ggml_ctx,
-                               ggml_ext_scale(ctx->ggml_ctx, x, alpha),
-                               ggml_ext_scale(ctx->ggml_ctx, x_mix, 1.0f - alpha));
+        auto alpha = get_alpha(ctx);
+        x          = ggml_add(ctx->ggml_ctx,
+                              x_mix,
+                              ggml_mul(ctx->ggml_ctx,
+                                       ggml_sub(ctx->ggml_ctx, x, x_mix),
+                                       alpha));
 
         x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
         x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
@@ -664,13 +666,13 @@ struct AutoEncoderKL : public VAE {
     AutoEncoderKLModel ae;
 
     AutoEncoderKL(ggml_backend_t backend,
-                  ggml_backend_t params_backend,
                   const String2TensorStorage& tensor_storage_map,
                   const std::string prefix,
-                  bool decode_only       = false,
-                  bool use_video_decoder = false,
-                  SDVersion version      = VERSION_SD1)
-        : decode_only(decode_only), VAE(version, backend, params_backend) {
+                  bool decode_only                                    = false,
+                  bool use_video_decoder                              = false,
+                  SDVersion version                                   = VERSION_SD1,
+                  std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : VAE(version, backend, prefix, weight_manager), decode_only(decode_only) {
         if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
             scale_factor = 0.18215f;
             shift_factor = 0.f;
@@ -680,7 +682,7 @@ struct AutoEncoderKL : public VAE {
         } else if (sd_version_is_sd3(version)) {
             scale_factor = 1.5305f;
             shift_factor = 0.0609f;
-        } else if (sd_version_is_flux(version) || sd_version_is_z_image(version)) {
+        } else if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_longcat(version)) {
             scale_factor = 0.3611f;
             shift_factor = 0.1159f;
         } else if (sd_version_uses_flux2_vae(version)) {
@@ -718,8 +720,8 @@ struct AutoEncoderKL : public VAE {
         return "vae";
     }
 
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) override {
-        ae.get_param_tensors(tensors, prefix);
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        ae.get_param_tensors(tensors, weight_prefix);
     }
 
     ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
@@ -742,7 +744,7 @@ struct AutoEncoderKL : public VAE {
         auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(z, decode_graph);
         };
-        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), z.dim());
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), z.dim());
     }
 
     sd::Tensor<float> gaussian_latent_sample(const sd::Tensor<float>& moments, std::shared_ptr<RNG> rng) {
@@ -886,4 +888,4 @@ struct AutoEncoderKL : public VAE {
     };
 };
 
-#endif  // __AUTO_ENCODER_KL_HPP__
+#endif  // __SD_MODEL_VAE_AUTO_ENCODER_KL_HPP__

src/model/vae/ltx_audio_vae.hpp

@@ -1,12 +1,15 @@
-#ifndef __SD_LTX_AUDIO_VAE_H__
-#define __SD_LTX_AUDIO_VAE_H__
+#ifndef __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__
+#define __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__
 
 #include <cmath>
+#include <limits>
 #include <numeric>
 #include <string>
 #include <vector>
 
-#include "ggml_extend.hpp"
+#include "core/ggml_extend.hpp"
+#include "model_loader.h"
+#include "model_manager.h"
 
 namespace LTXV {
 
@@ -57,11 +60,12 @@ namespace LTXV {
             return base_output_sample_rate();
         }
 
-        static LTXAudioVAEConfig detect_from_weights(const String2TensorStorage& tensor_storage_map) {
+        static LTXAudioVAEConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix = "") {
             LTXAudioVAEConfig config;
 
             auto require = [&](const std::string& name) -> const TensorStorage* {
-                auto iter = tensor_storage_map.find(name);
+                std::string tensor_name = prefix.empty() ? name : prefix + "." + name;
+                auto iter               = tensor_storage_map.find(tensor_name);
                 if (iter == tensor_storage_map.end()) {
                     return nullptr;
                 }
@@ -167,94 +171,69 @@ namespace LTXV {
             if (config.audio_channels != 2 || config.latent_channels != 8 || config.mel_bins != 64) {
                 return config;
             }
+            LOG_DEBUG("ltx_audio_vae: sample_rate = %d, mel_bins = %d, latent_channels = %d, latent_frequency_bins = %d, has_bwe = %s",
+                      config.sample_rate,
+                      config.mel_bins,
+                      config.latent_channels,
+                      config.latent_frequency_bins,
+                      config.has_bwe ? "true" : "false");
             return config;
         }
     };
 
-    static sd::Tensor<float> squeeze_trailing_singleton_dims(sd::Tensor<float> tensor) {
-        while (tensor.dim() > 0 && tensor.shape().back() == 1) {
-            tensor = tensor.squeeze(static_cast<size_t>(tensor.dim() - 1));
-        }
-        return tensor;
-    }
+    static ggml_tensor* compute_log_mel_spectrogram(GGMLRunnerContext* runner_ctx,
+                                                    ggml_tensor* waveform,
+                                                    ggml_tensor* forward_basis,
+                                                    ggml_tensor* mel_basis,
+                                                    int hop_length) {
+        auto ctx = runner_ctx->ggml_ctx;
+        GGML_ASSERT(ctx != nullptr);
+        GGML_ASSERT(waveform != nullptr);
+        GGML_ASSERT(forward_basis != nullptr);
+        GGML_ASSERT(mel_basis != nullptr);
+        GGML_ASSERT(waveform->type == GGML_TYPE_F32);
+        GGML_ASSERT(forward_basis->type == GGML_TYPE_F32);
+        GGML_ASSERT(mel_basis->type == GGML_TYPE_F32);
+        GGML_ASSERT(forward_basis->ne[1] == 1);
 
-    static sd::Tensor<float> normalize_waveform_for_host(sd::Tensor<float> waveform) {
-        waveform = squeeze_trailing_singleton_dims(std::move(waveform));
-        if (waveform.empty()) {
-            return waveform;
-        }
-        if (waveform.dim() == 1) {
-            return waveform.reshape({waveform.shape()[0], 1, 1});
-        }
-        if (waveform.dim() == 2) {
-            return waveform.reshape({waveform.shape()[0], waveform.shape()[1], 1});
-        }
-        if (waveform.dim() == 3) {
-            return waveform;
-        }
-        throw std::runtime_error("Unsupported waveform rank for host processing: rank=" + std::to_string(waveform.dim()));
-    }
+        const int64_t time          = waveform->ne[0];
+        const int64_t channels      = waveform->ne[1];
+        const int64_t batch         = waveform->ne[2];
+        const int64_t filter_len    = forward_basis->ne[0];
+        const int64_t stft_channels = forward_basis->ne[2];
+        const int64_t n_freqs       = stft_channels / 2;
+        const int64_t n_mels        = mel_basis->ne[1];
+        const int64_t left_pad      = std::max<int64_t>(0, filter_len - hop_length);
+        const int64_t padded_time   = time + left_pad;
+        const int64_t frame_count   = padded_time < filter_len ? 0 : 1 + (padded_time - filter_len) / hop_length;
 
-    static sd::Tensor<float> load_param_tensor_f32(ggml_tensor* tensor) {
-        GGML_ASSERT(tensor != nullptr);
-        return squeeze_trailing_singleton_dims(sd::make_sd_tensor_from_ggml<float>(tensor));
-    }
+        GGML_ASSERT(stft_channels % 2 == 0);
+        GGML_ASSERT(mel_basis->ne[0] == n_freqs);
+        GGML_ASSERT(waveform->ne[3] == 1);
+        GGML_ASSERT(frame_count > 0);
 
-    static sd::Tensor<float> compute_log_mel_spectrogram(const sd::Tensor<float>& waveform_in,
-                                                         const sd::Tensor<float>& forward_basis,
-                                                         const sd::Tensor<float>& mel_basis,
-                                                         int hop_length) {
-        auto waveform = normalize_waveform_for_host(waveform_in);
-        GGML_ASSERT(forward_basis.dim() >= 3);
-        GGML_ASSERT(mel_basis.dim() >= 2);
-
-        const int64_t time        = waveform.shape()[0];
-        const int64_t channels    = waveform.shape()[1];
-        const int64_t batch       = waveform.shape()[2];
-        const int64_t filter_len  = forward_basis.shape()[0];
-        const int64_t basis_freq2 = forward_basis.shape().back();
-        const int64_t n_freqs     = basis_freq2 / 2;
-        const int64_t n_mels      = mel_basis.shape()[1];
-        const int64_t left_pad    = std::max<int64_t>(0, filter_len - hop_length);
-        const int64_t padded_time = time + left_pad;
-        const int64_t frame_count = padded_time < filter_len ? 0 : 1 + (padded_time - filter_len) / hop_length;
-
-        sd::Tensor<float> log_mel({n_mels, frame_count, channels, batch});
-        std::vector<float> padded(static_cast<size_t>(padded_time), 0.0f);
-        std::vector<float> magnitude(static_cast<size_t>(n_freqs), 0.0f);
-
-        for (int64_t b = 0; b < batch; ++b) {
-            for (int64_t c = 0; c < channels; ++c) {
-                std::fill(padded.begin(), padded.end(), 0.0f);
-                for (int64_t t = 0; t < time; ++t) {
-                    padded[static_cast<size_t>(t + left_pad)] = waveform.index(t, c, b);
-                }
-
-                for (int64_t frame = 0; frame < frame_count; ++frame) {
-                    const int64_t frame_offset = frame * hop_length;
-                    for (int64_t f = 0; f < n_freqs; ++f) {
-                        double real = 0.0;
-                        double imag = 0.0;
-                        for (int64_t k = 0; k < filter_len; ++k) {
-                            const float sample = padded[static_cast<size_t>(frame_offset + k)];
-                            real += static_cast<double>(sample) * static_cast<double>(forward_basis.index(k, 0, f));
-                            imag += static_cast<double>(sample) * static_cast<double>(forward_basis.index(k, 0, f + n_freqs));
-                        }
-                        magnitude[static_cast<size_t>(f)] = static_cast<float>(std::sqrt(real * real + imag * imag));
-                    }
-
-                    for (int64_t m = 0; m < n_mels; ++m) {
-                        double mel_value = 0.0;
-                        for (int64_t f = 0; f < n_freqs; ++f) {
-                            mel_value += static_cast<double>(mel_basis.index(f, m)) * static_cast<double>(magnitude[static_cast<size_t>(f)]);
-                        }
-                        log_mel.index(m, frame, c, b) = static_cast<float>(std::log(std::max(mel_value, 1e-5)));
-                    }
-                }
-            }
+        auto x = ggml_reshape_3d(ctx, waveform, time, 1, channels * batch);
+        if (left_pad > 0) {
+            x = ggml_pad_ext(ctx, x, static_cast<int>(left_pad), 0, 0, 0, 0, 0, 0, 0);
         }
 
-        return log_mel;
+        auto frames = ggml_conv_1d(ctx, forward_basis, x, hop_length, 0, 1);
+        GGML_ASSERT(frames->ne[0] == frame_count);
+        GGML_ASSERT(frames->ne[1] == stft_channels);
+        GGML_ASSERT(frames->ne[2] == channels * batch);
+
+        auto real      = ggml_ext_slice(ctx, frames, 1, 0, n_freqs);
+        auto imag      = ggml_ext_slice(ctx, frames, 1, n_freqs, stft_channels);
+        auto magnitude = ggml_sqrt(ctx,
+                                   ggml_add(ctx,
+                                            ggml_sqr(ctx, real),
+                                            ggml_sqr(ctx, imag)));
+
+        magnitude = ggml_cont(ctx, ggml_permute(ctx, magnitude, 1, 0, 2, 3));
+        auto mel  = ggml_mul_mat(ctx, mel_basis, magnitude);
+        mel       = ggml_log(ctx, ggml_clamp(ctx, mel, 1e-5f, std::numeric_limits<float>::max()));
+
+        return ggml_reshape_4d(ctx, mel, n_mels, frame_count, channels, batch);
     }
 
     static std::vector<float> build_hann_resample_filter(int ratio) {
@@ -276,8 +255,135 @@ namespace LTXV {
         return filter;
     }
 
-    static sd::Tensor<float> upsample_waveform_hann(const sd::Tensor<float>& waveform_in, int ratio) {
-        auto waveform = normalize_waveform_for_host(waveform_in);
+    static ggml_type audio_conv_weight_type(ggml_type type) {
+        return type == GGML_TYPE_BF16 ? GGML_TYPE_F16 : type;
+    }
+
+    static ggml_tensor* repeat_with_vulkan_f32_workaround(ggml_backend_t backend,
+                                                          ggml_context* ctx,
+                                                          ggml_tensor* x,
+                                                          int64_t ne0,
+                                                          int64_t ne1,
+                                                          int64_t ne2,
+                                                          int64_t ne3) {
+        if (x->type != GGML_TYPE_F32 &&
+            (x->type == GGML_TYPE_F16 || x->type == GGML_TYPE_BF16) &&
+            sd_backend_is(backend, "vulkan")) {
+            auto x_f32    = ggml_cast(ctx, x, GGML_TYPE_F32);
+            auto repeated = ggml_repeat_4d(ctx,
+                                           x_f32,
+                                           ne0,
+                                           ne1,
+                                           ne2,
+                                           ne3);
+            return ggml_cast(ctx, repeated, x->type);
+        }
+        return ggml_repeat_4d(ctx, x, ne0, ne1, ne2, ne3);
+    }
+
+    static ggml_tensor* repeat_1d_value(GGMLRunnerContext* runner_ctx, ggml_tensor* x, int64_t count) {
+        auto ctx = runner_ctx->ggml_ctx;
+        GGML_ASSERT(x->ne[0] == 1);
+        return repeat_with_vulkan_f32_workaround(runner_ctx->backend, ctx, x, count, x->ne[1], x->ne[2], x->ne[3]);
+    }
+
+    static ggml_tensor* replicate_pad_1d(GGMLRunnerContext* runner_ctx, ggml_tensor* x, int64_t left, int64_t right) {
+        auto ctx = runner_ctx->ggml_ctx;
+        if (left > 0) {
+            auto first = ggml_ext_slice(ctx, x, 0, 0, 1);
+            x          = ggml_concat(ctx, repeat_1d_value(runner_ctx, first, left), x, 0);
+        }
+        if (right > 0) {
+            auto last = ggml_ext_slice(ctx, x, 0, x->ne[0] - 1, x->ne[0]);
+            x         = ggml_concat(ctx, x, repeat_1d_value(runner_ctx, last, right), 0);
+        }
+        return x;
+    }
+
+    static ggml_tensor* tile_depthwise_filter_1d(GGMLRunnerContext* runner_ctx, ggml_tensor* filter, int64_t channels) {
+        auto ctx          = runner_ctx->ggml_ctx;
+        ggml_tensor* base = filter;
+        if (ggml_n_dims(base) == 3) {
+            base = ggml_reshape_4d(ctx, base, base->ne[0], 1, 1, 1);
+        } else if (ggml_n_dims(base) == 1) {
+            base = ggml_reshape_4d(ctx, base, base->ne[0], 1, 1, 1);
+        }
+        return repeat_with_vulkan_f32_workaround(runner_ctx->backend, ctx, base, base->ne[0], 1, channels, 1);
+    }
+
+    static ggml_tensor* depthwise_conv1d(GGMLRunnerContext* runner_ctx,
+                                         ggml_tensor* x,
+                                         ggml_tensor* filter,
+                                         int stride,
+                                         int padding) {
+        auto ctx = runner_ctx->ggml_ctx;
+        GGML_ASSERT(x->ne[3] == 1);
+        auto tiled = tile_depthwise_filter_1d(runner_ctx, filter, x->ne[1]);
+        auto out   = ggml_conv_1d_dw(ctx, tiled, x, stride, padding, 1);
+        return ggml_reshape_4d(ctx, out, out->ne[0], out->ne[1], 1, 1);
+    }
+
+    static ggml_tensor* reverse_1d_filter(ggml_context* ctx, ggml_tensor* filter) {
+        GGML_ASSERT(ctx != nullptr);
+        GGML_ASSERT(filter != nullptr);
+        GGML_ASSERT(filter->ne[1] == 1);
+        GGML_ASSERT(filter->ne[2] == 1);
+        GGML_ASSERT(filter->ne[3] == 1);
+
+        ggml_tensor* reversed = nullptr;
+        for (int64_t k = filter->ne[0] - 1; k >= 0; --k) {
+            auto slice = ggml_ext_slice(ctx, filter, 0, k, k + 1);
+            reversed   = reversed == nullptr ? slice : ggml_concat(ctx, reversed, slice, 0);
+        }
+        return reversed;
+    }
+
+    static ggml_tensor* depthwise_conv_transpose1d(ggml_context* ctx,
+                                                   ggml_tensor* x,
+                                                   ggml_tensor* filter,
+                                                   int stride) {
+        GGML_ASSERT(x->ne[2] == 1 && x->ne[3] == 1);
+        GGML_ASSERT(filter->ne[1] == 1);
+        GGML_ASSERT(filter->ne[2] == 1 && filter->ne[3] == 1);
+
+        const int64_t time        = x->ne[0];
+        const int64_t channels    = x->ne[1];
+        const int64_t kernel_size = filter->ne[0];
+        const int64_t out_time    = (time - 1) * stride + kernel_size;
+
+        auto x_flat = ggml_reshape_3d(ctx, x, 1, time, channels);
+        if (stride > 1) {
+            auto zero_unit = ggml_ext_scale(ctx, x_flat, 0.0f);
+            auto zero_tail = zero_unit;
+            for (int i = 1; i < stride - 1; ++i) {
+                zero_tail = ggml_concat(ctx, zero_tail, zero_unit, 0);
+            }
+            x_flat = ggml_concat(ctx, x_flat, zero_tail, 0);
+        }
+        x_flat = ggml_reshape_3d(ctx, x_flat, time * stride, 1, channels);
+
+        auto reversed_filter = reverse_1d_filter(ctx, filter);
+        auto out             = ggml_conv_1d(ctx, reversed_filter, x_flat, 1, static_cast<int>(kernel_size - 1), 1);
+        if (out->ne[0] > out_time) {
+            out = ggml_ext_slice(ctx, out, 0, 0, out_time);
+        }
+        GGML_ASSERT(out->ne[0] == out_time);
+        GGML_ASSERT(out->ne[1] == 1);
+        GGML_ASSERT(out->ne[2] == channels);
+
+        out = ggml_ext_scale(ctx, out, static_cast<float>(stride));
+        return ggml_reshape_4d(ctx, out, out_time, channels, 1, 1);
+    }
+
+    static ggml_tensor* upsample_waveform_hann(GGMLRunnerContext* runner_ctx,
+                                               ggml_tensor* waveform,
+                                               ggml_tensor* filter,
+                                               int ratio) {
+        auto ctx = runner_ctx->ggml_ctx;
+        GGML_ASSERT(ctx != nullptr);
+        GGML_ASSERT(waveform != nullptr);
+        GGML_ASSERT(filter != nullptr);
+        GGML_ASSERT(waveform->ne[3] == 1);
         if (ratio <= 1) {
             return waveform;
         }
@@ -289,122 +395,29 @@ namespace LTXV {
         const int pad                  = width;
         const int pad_left             = 2 * width * ratio;
         const int pad_right            = kernel_size - ratio;
-        const int64_t time             = waveform.shape()[0];
-        const int64_t channels         = waveform.shape()[1];
-        const int64_t batch            = waveform.shape()[2];
-        const int64_t padded_time      = time + 2 * pad;
-        const int64_t conv_out_time    = (padded_time - 1) * ratio + kernel_size;
-        const int64_t cropped_time     = conv_out_time - pad_left - pad_right;
-        auto filter                    = build_hann_resample_filter(ratio);
+        const int64_t time             = waveform->ne[0];
+        const int64_t channels         = waveform->ne[1];
+        const int64_t batch            = waveform->ne[2];
 
-        sd::Tensor<float> output({cropped_time, channels, batch});
-        std::vector<float> padded(static_cast<size_t>(padded_time), 0.0f);
-        std::vector<float> conv_out(static_cast<size_t>(conv_out_time), 0.0f);
+        GGML_ASSERT(filter->ne[0] == kernel_size);
 
-        for (int64_t b = 0; b < batch; ++b) {
-            for (int64_t c = 0; c < channels; ++c) {
-                std::fill(padded.begin(), padded.end(), 0.0f);
-                const float first = waveform.index(0, c, b);
-                const float last  = waveform.index(time - 1, c, b);
-                for (int i = 0; i < pad; ++i) {
-                    padded[static_cast<size_t>(i)]              = first;
-                    padded[static_cast<size_t>(pad + time + i)] = last;
-                }
-                for (int64_t t = 0; t < time; ++t) {
-                    padded[static_cast<size_t>(pad + t)] = waveform.index(t, c, b);
-                }
-
-                std::fill(conv_out.begin(), conv_out.end(), 0.0f);
-                for (int64_t t = 0; t < padded_time; ++t) {
-                    const double sample    = static_cast<double>(padded[static_cast<size_t>(t)]) * ratio;
-                    const int64_t out_base = t * ratio;
-                    for (int k = 0; k < kernel_size; ++k) {
-                        conv_out[static_cast<size_t>(out_base + k)] += static_cast<float>(sample * filter[static_cast<size_t>(k)]);
-                    }
-                }
-
-                for (int64_t t = 0; t < cropped_time; ++t) {
-                    output.index(t, c, b) = conv_out[static_cast<size_t>(t + pad_left)];
-                }
-            }
-        }
-
-        return output;
+        auto x = ggml_reshape_3d(ctx, waveform, time, channels * batch, 1);
+        x      = replicate_pad_1d(runner_ctx, x, pad, pad);
+        x      = depthwise_conv_transpose1d(ctx, x, filter, ratio);
+        x      = ggml_ext_slice(ctx, x, 0, pad_left, x->ne[0] - pad_right);
+        return ggml_reshape_3d(ctx, x, x->ne[0], channels, batch);
     }
 
-    static sd::Tensor<float> crop_waveform_samples(const sd::Tensor<float>& waveform_in, int64_t target_samples) {
-        auto waveform = normalize_waveform_for_host(waveform_in);
-        if (waveform.shape()[0] == target_samples) {
+    static ggml_tensor* crop_waveform_samples(ggml_context* ctx,
+                                              ggml_tensor* waveform,
+                                              int64_t target_samples) {
+        GGML_ASSERT(ctx != nullptr);
+        GGML_ASSERT(waveform != nullptr);
+        if (waveform->ne[0] == target_samples) {
             return waveform;
         }
-        if (waveform.shape()[0] > target_samples) {
-            return sd::ops::slice(waveform, 0, 0, target_samples);
-        }
-        sd::Tensor<float> output({target_samples, waveform.shape()[1], waveform.shape()[2]});
-        sd::ops::slice_assign(&output, 0, 0, waveform.shape()[0], waveform);
-        return output;
-    }
-
-    static ggml_type audio_conv_weight_type(ggml_type type) {
-        return type == GGML_TYPE_BF16 ? GGML_TYPE_F16 : type;
-    }
-
-    static ggml_tensor* repeat_1d_value(ggml_context* ctx, ggml_tensor* x, int64_t count) {
-        GGML_ASSERT(x->ne[0] == 1);
-        ggml_tensor* target = ggml_new_tensor_4d(ctx, x->type, count, x->ne[1], x->ne[2], x->ne[3]);
-        return ggml_repeat(ctx, x, target);
-    }
-
-    static ggml_tensor* replicate_pad_1d(ggml_context* ctx, ggml_tensor* x, int64_t left, int64_t right) {
-        if (left > 0) {
-            auto first = ggml_ext_slice(ctx, x, 0, 0, 1);
-            x          = ggml_concat(ctx, repeat_1d_value(ctx, first, left), x, 0);
-        }
-        if (right > 0) {
-            auto last = ggml_ext_slice(ctx, x, 0, x->ne[0] - 1, x->ne[0]);
-            x         = ggml_concat(ctx, x, repeat_1d_value(ctx, last, right), 0);
-        }
-        return x;
-    }
-
-    static ggml_tensor* tile_depthwise_filter_1d(ggml_context* ctx, ggml_tensor* filter, int64_t channels) {
-        ggml_tensor* base = filter;
-        if (ggml_n_dims(base) == 3) {
-            base = ggml_reshape_4d(ctx, base, base->ne[0], 1, 1, 1);
-        } else if (ggml_n_dims(base) == 1) {
-            base = ggml_reshape_4d(ctx, base, base->ne[0], 1, 1, 1);
-        }
-        ggml_tensor* target = ggml_new_tensor_4d(ctx, base->type, base->ne[0], 1, channels, 1);
-        return ggml_repeat(ctx, base, target);
-    }
-
-    static ggml_tensor* depthwise_conv1d(ggml_context* ctx,
-                                         ggml_tensor* x,
-                                         ggml_tensor* filter,
-                                         int stride,
-                                         int padding) {
-        GGML_ASSERT(x->ne[3] == 1);
-        auto tiled = tile_depthwise_filter_1d(ctx, filter, x->ne[1]);
-        auto out   = ggml_conv_1d_dw(ctx, tiled, x, stride, padding, 1);
-        return ggml_reshape_4d(ctx, out, out->ne[0], out->ne[1], 1, 1);
-    }
-
-    static ggml_tensor* depthwise_conv_transpose1d(ggml_context* ctx,
-                                                   ggml_tensor* x,
-                                                   ggml_tensor* filter,
-                                                   int stride) {
-        GGML_ASSERT(x->ne[2] == 1 && x->ne[3] == 1);
-        GGML_ASSERT(filter->ne[1] == 1);
-
-        ggml_tensor* out = nullptr;
-        for (int64_t c = 0; c < x->ne[1]; ++c) {
-            auto xi = ggml_ext_slice(ctx, x, 1, c, c + 1);
-            auto yi = ggml_conv_transpose_1d(ctx, filter, xi, stride, 0, 1);
-            yi      = ggml_ext_scale(ctx, yi, static_cast<float>(stride));
-            yi      = ggml_reshape_4d(ctx, yi, yi->ne[0], 1, 1, 1);
-            out     = out == nullptr ? yi : ggml_concat(ctx, out, yi, 1);
-        }
-        return out;
+        GGML_ASSERT(waveform->ne[0] > target_samples);
+        return ggml_ext_slice(ctx, waveform, 0, 0, target_samples);
     }
 
     struct PixelNorm2D : public UnaryBlock {
@@ -654,7 +667,7 @@ namespace LTXV {
             int up_pad_left  = up_pad * up_ratio + (up_kernel_size - up_ratio) / 2;
             int up_pad_right = up_pad * up_ratio + (up_kernel_size - up_ratio + 1) / 2;
 
-            x = replicate_pad_1d(ctx->ggml_ctx, x, up_pad, up_pad);
+            x = replicate_pad_1d(ctx, x, up_pad, up_pad);
             x = depthwise_conv_transpose1d(ctx->ggml_ctx, x, up_filter, up_ratio);
             x = ggml_ext_slice(ctx->ggml_ctx, x, 0, up_pad_left, x->ne[0] - up_pad_right);
 
@@ -662,8 +675,8 @@ namespace LTXV {
 
             int down_pad_left  = down_kernel_size / 2 - (down_kernel_size % 2 == 0 ? 1 : 0);
             int down_pad_right = down_kernel_size / 2;
-            x                  = replicate_pad_1d(ctx->ggml_ctx, x, down_pad_left, down_pad_right);
-            x                  = depthwise_conv1d(ctx->ggml_ctx, x, down_filter, down_ratio, 0);
+            x                  = replicate_pad_1d(ctx, x, down_pad_left, down_pad_right);
+            x                  = depthwise_conv1d(ctx, x, down_filter, down_ratio, 0);
             return x;
         }
     };
@@ -926,57 +939,88 @@ namespace LTXV {
             }
         }
 
-        ggml_tensor* decode_to_mel(GGMLRunnerContext* ctx,
-                                   ggml_tensor* latent,
-                                   int target_time,
-                                   int target_freq) {
-            auto mean    = params["audio_vae.per_channel_statistics.mean-of-means"];
-            auto stddev  = params["audio_vae.per_channel_statistics.std-of-means"];
-            auto decoder = std::dynamic_pointer_cast<AudioDecoder>(blocks["audio_vae.decoder"]);
-            return decoder->forward(ctx, latent, mean, stddev, target_time, target_freq);
-        }
+        ggml_tensor* decode(GGMLRunnerContext* ctx,
+                            ggml_tensor* latent,
+                            ggml_tensor* bwe_skip_filter) {
+            int target_time = static_cast<int>(latent->ne[1]) * config.latent_downsample_factor() -
+                              (config.latent_downsample_factor() - 1);
+            int target_freq = config.mel_bins;
 
-        ggml_tensor* run_vocoder(GGMLRunnerContext* ctx, ggml_tensor* mel) {
-            auto vocoder = std::dynamic_pointer_cast<Vocoder>(blocks["vocoder.vocoder"]);
-            return vocoder->forward(ctx, mel);
-        }
+            auto decoder  = std::dynamic_pointer_cast<AudioDecoder>(blocks["audio_vae.decoder"]);
+            auto mean     = params["audio_vae.per_channel_statistics.mean-of-means"];
+            auto stddev   = params["audio_vae.per_channel_statistics.std-of-means"];
+            auto mel      = decoder->forward(ctx, latent, mean, stddev, target_time, target_freq);
+            auto vocoder  = std::dynamic_pointer_cast<Vocoder>(blocks["vocoder.vocoder"]);
+            auto waveform = vocoder->forward(ctx, mel);
 
-        ggml_tensor* run_bwe_generator(GGMLRunnerContext* ctx, ggml_tensor* mel) {
-            GGML_ASSERT(config.has_bwe);
-            auto bwe_generator = std::dynamic_pointer_cast<Vocoder>(blocks["vocoder.bwe_generator"]);
-            return bwe_generator->forward(ctx, mel);
-        }
+            if (config.has_bwe) {
+                GGML_ASSERT(bwe_skip_filter != nullptr);
+                const int bwe_ratio    = config.bwe_output_sample_rate / config.bwe_input_sample_rate;
+                const int64_t low_time = waveform->ne[0];
+                const int64_t out_time = low_time * bwe_ratio;
+                int64_t remainder      = low_time % config.bwe_hop_length;
+                auto bwe_waveform      = waveform;
+                if (remainder != 0) {
+                    bwe_waveform = ggml_pad_ext(ctx->ggml_ctx,
+                                                bwe_waveform,
+                                                0,
+                                                static_cast<int>(config.bwe_hop_length - remainder),
+                                                0,
+                                                0,
+                                                0,
+                                                0,
+                                                0,
+                                                0);
+                }
 
-        ggml_tensor* mel_basis_tensor() const {
-            auto iter = params.find("vocoder.mel_stft.mel_basis");
-            return iter == params.end() ? nullptr : iter->second;
-        }
+                auto mel_basis  = params["vocoder.mel_stft.mel_basis"];
+                auto stft_basis = params["vocoder.mel_stft.stft_fn.forward_basis"];
+                GGML_ASSERT(mel_basis != nullptr && stft_basis != nullptr);
+                auto bwe_mel       = compute_log_mel_spectrogram(ctx, bwe_waveform, stft_basis, mel_basis, config.bwe_hop_length);
+                auto bwe_generator = std::dynamic_pointer_cast<Vocoder>(blocks["vocoder.bwe_generator"]);
+                auto residual      = bwe_generator->forward(ctx, bwe_mel);
 
-        ggml_tensor* stft_forward_basis_tensor() const {
-            auto iter = params.find("vocoder.mel_stft.stft_fn.forward_basis");
-            return iter == params.end() ? nullptr : iter->second;
+                auto skip = upsample_waveform_hann(ctx,
+                                                   bwe_waveform,
+                                                   bwe_skip_filter,
+                                                   bwe_ratio);
+                waveform  = ggml_clamp(ctx->ggml_ctx,
+                                       ggml_add(ctx->ggml_ctx, residual, skip),
+                                       -1.0f,
+                                       1.0f);
+                waveform  = crop_waveform_samples(ctx->ggml_ctx, waveform, out_time);
+            }
+
+            return waveform;
         }
     };
 
     struct LTXAudioVAERunner : public GGMLRunner {
         LTXAudioVAEConfig config;
         LTXAudioVAE model;
+        std::string weight_prefix;
+        sd::Tensor<float> bwe_skip_filter_tensor;
 
         LTXAudioVAERunner(ggml_backend_t backend,
-                          ggml_backend_t params_backend,
                           const String2TensorStorage& tensor_storage_map,
-                          const std::string& prefix = "")
-            : GGMLRunner(backend, params_backend),
+                          const std::string& prefix                           = "",
+                          std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+            : GGMLRunner(backend, weight_manager),
+              weight_prefix(prefix),
               config(LTXAudioVAEConfig::detect_from_weights(tensor_storage_map)),
               model(config) {
             model.init(params_ctx, tensor_storage_map, prefix);
+            if (config.has_bwe) {
+                const int bwe_ratio    = config.bwe_output_sample_rate / config.bwe_input_sample_rate;
+                bwe_skip_filter_tensor = sd::Tensor<float>::from_vector(build_hann_resample_filter(bwe_ratio));
+            }
         }
 
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
-            model.get_param_tensors(tensors, prefix);
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {
+            model.get_param_tensors(tensors, weight_prefix);
         }
 
-        size_t get_params_buffer_size() {
+        size_t get_params_mem_size() {
             return model.get_params_mem_size();
         }
 
@@ -984,77 +1028,22 @@ namespace LTXV {
             return "ltx_audio_vae";
         }
 
-        ggml_cgraph* build_base_graph(const sd::Tensor<float>& latent_tensor) {
-            auto latent     = make_input(latent_tensor);
-            int target_time = static_cast<int>(latent_tensor.shape()[1]) * config.latent_downsample_factor() -
-                              (config.latent_downsample_factor() - 1);
-            int target_freq = config.mel_bins;
-
-            ggml_cgraph* gf = new_graph_custom(655360);
-            auto runner_ctx = GGMLRunner::get_context();
-            auto mel        = model.decode_to_mel(&runner_ctx, latent, target_time, target_freq);
-            auto waveform   = model.run_vocoder(&runner_ctx, mel);
-            ggml_build_forward_expand(gf, waveform);
-            return gf;
-        }
-
-        ggml_cgraph* build_bwe_graph(const sd::Tensor<float>& mel_tensor) {
-            auto mel        = make_input(mel_tensor);
-            ggml_cgraph* gf = new_graph_custom(655360);
-            auto runner_ctx = GGMLRunner::get_context();
-            auto residual   = model.run_bwe_generator(&runner_ctx, mel);
-            ggml_build_forward_expand(gf, residual);
-            return gf;
-        }
-
-        sd::Tensor<float> compute_base_waveform(int n_threads,
-                                                const sd::Tensor<float>& latent_tensor) {
-            auto get_graph = [&]() -> ggml_cgraph* {
-                return build_base_graph(latent_tensor);
-            };
-            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), 4);
-        }
-
-        sd::Tensor<float> compute_bwe_residual(int n_threads,
-                                               const sd::Tensor<float>& mel_tensor) {
-            auto get_graph = [&]() -> ggml_cgraph* {
-                return build_bwe_graph(mel_tensor);
-            };
-            return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), 4);
-        }
-
         sd::Tensor<float> decode(int n_threads,
                                  const sd::Tensor<float>& latent_tensor) {
-            auto waveform = compute_base_waveform(n_threads, latent_tensor);
-            if (!config.has_bwe || waveform.empty()) {
-                return waveform;
-            }
-
-            auto waveform_host     = normalize_waveform_for_host(waveform);
-            const int64_t low_time = waveform_host.shape()[0];
-            const int64_t out_time = low_time * config.bwe_output_sample_rate / config.bwe_input_sample_rate;
-            int64_t remainder      = low_time % config.bwe_hop_length;
-            if (remainder != 0) {
-                sd::Tensor<float> padded({low_time + (config.bwe_hop_length - remainder), waveform_host.shape()[1], waveform_host.shape()[2]});
-                sd::ops::slice_assign(&padded, 0, 0, low_time, waveform_host);
-                waveform_host = std::move(padded);
-            }
-
-            auto mel_basis_tensor  = model.mel_basis_tensor();
-            auto stft_basis_tensor = model.stft_forward_basis_tensor();
-            GGML_ASSERT(mel_basis_tensor != nullptr && stft_basis_tensor != nullptr);
-            auto mel_basis     = load_param_tensor_f32(mel_basis_tensor);
-            auto forward_basis = load_param_tensor_f32(stft_basis_tensor);
-            auto bwe_mel       = compute_log_mel_spectrogram(waveform_host, forward_basis, mel_basis, config.bwe_hop_length);
-            auto residual_raw  = compute_bwe_residual(n_threads, bwe_mel);
-            if (residual_raw.empty()) {
-                return waveform;
-            }
-            auto residual = normalize_waveform_for_host(residual_raw);
-            auto skip     = upsample_waveform_hann(waveform_host, config.bwe_output_sample_rate / config.bwe_input_sample_rate);
-            auto combined = sd::ops::clamp(residual + skip, -1.0f, 1.0f);
-            auto cropped  = crop_waveform_samples(combined, out_time);
-            return restore_trailing_singleton_dims(cropped, 4);
+            int64_t t0     = ggml_time_ms();
+            auto get_graph = [&]() -> ggml_cgraph* {
+                auto latent                  = make_input(latent_tensor);
+                ggml_tensor* bwe_skip_filter = config.has_bwe ? make_input(bwe_skip_filter_tensor) : nullptr;
+                ggml_cgraph* gf              = new_graph_custom(655360);
+                auto runner_ctx              = GGMLRunner::get_context();
+                auto waveform                = model.decode(&runner_ctx, latent, bwe_skip_filter);
+                ggml_build_forward_expand(gf, waveform);
+                return gf;
+            };
+            auto result = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), 4);
+            int64_t t1  = ggml_time_ms();
+            LOG_INFO("ltx audio vae decode completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+            return result;
         }
 
         void test(const std::string& input_path) {
@@ -1074,11 +1063,12 @@ namespace LTXV {
         static void load_from_file_and_test(const std::string& model_path,
                                             const std::string& input_path,
                                             const std::string& prefix = "") {
-            ggml_backend_t backend = ggml_backend_cpu_init();
+            ggml_backend_t backend = sd_backend_cpu_init();
             // ggml_backend_t backend = ggml_backend_cuda_init(0);
             LOG_INFO("loading ltx audio vae from '%s'", model_path.c_str());
 
-            ModelLoader model_loader;
+            auto model_manager        = std::make_shared<ModelManager>();
+            ModelLoader& model_loader = model_manager->loader();
             if (!model_loader.init_from_file(model_path)) {
                 LOG_ERROR("init model loader from file failed: '%s'", model_path.c_str());
                 return;
@@ -1086,16 +1076,17 @@ namespace LTXV {
 
             auto& tensor_storage_map = model_loader.get_tensor_storage_map();
             auto ltx_audio_vae       = std::make_shared<LTXAudioVAERunner>(backend,
-                                                                     backend,
                                                                      tensor_storage_map,
-                                                                     prefix);
+                                                                     prefix,
+                                                                     model_manager);
 
-            ltx_audio_vae->alloc_params_buffer();
-            std::map<std::string, ggml_tensor*> tensors;
-            ltx_audio_vae->get_param_tensors(tensors, "");
-
-            if (!model_loader.load_tensors(tensors)) {
-                LOG_ERROR("load tensors from model loader failed");
+            if (!model_manager->register_runner_params("LTX audio VAE test",
+                                                       *ltx_audio_vae,
+                                                       ModelManager::ResidencyMode::ParamBackend,
+                                                       backend,
+                                                       backend) ||
+                !model_manager->validate_registered_tensors()) {
+                LOG_ERROR("register ltx audio vae tensors with model manager failed");
                 return;
             }
 
@@ -1106,4 +1097,4 @@ namespace LTXV {
 
 }  // namespace LTXV
 
-#endif  // __SD_LTX_AUDIO_VAE_H__
+#endif  // __SD_MODEL_VAE_LTX_AUDIO_VAE_HPP__

src/model/vae/ltx_vae.hpp

@@ -1,6 +1,7 @@
-#ifndef __SD_LTX_VAE_HPP__
-#define __SD_LTX_VAE_HPP__
+#ifndef __SD_MODEL_VAE_LTX_VAE_HPP__
+#define __SD_MODEL_VAE_LTX_VAE_HPP__
 
+#include <algorithm>
 #include <fstream>
 #include <memory>
 #include <string>
@@ -8,9 +9,10 @@
 #include <utility>
 #include <vector>
 
-#include "ltxv.hpp"
-#include "vae.hpp"
-#include "wan.hpp"
+#include "model/diffusion/ltxv.hpp"
+#include "model/vae/vae.hpp"
+#include "model/vae/wan_vae.hpp"
+#include "model_loader.h"
 
 namespace LTXVAE {
 
@@ -143,16 +145,25 @@ namespace LTXVAE {
                              std::vector<ggml_tensor*>& feat_map,
                              int& feat_idx,
                              int chunk_idx,
-                             bool causal = true) {
+                             bool causal      = true,
+                             int temporal_pad = 0) {
             auto conv     = std::dynamic_pointer_cast<Conv3d>(blocks["conv"]);
             const int pad = causal ? (time_kernel_size - 1) : (time_kernel_size - 1) / 2;
 
             ggml_tensor* prev = (feat_idx < (int)feat_map.size()) ? feat_map[feat_idx] : nullptr;
 
+            GGML_ASSERT(x->ne[2] >= temporal_pad);
+
+            int end_idx   = (int)x->ne[2] - temporal_pad;
+            int start_idx = std::max(end_idx - pad, 0);
+
             // Save a contiguous copy of the last `pad` frames so the large `x`
             // tensor is not kept alive across iterations by a dangling view.
-            if (feat_idx < (int)feat_map.size() && pad > 0 && x->ne[2] >= pad) {
-                auto slice         = ggml_ext_slice(ctx->ggml_ctx, x, 2, x->ne[2] - pad, x->ne[2]);
+            if (feat_idx < (int)feat_map.size() && end_idx - start_idx > 0) {
+                GGML_ASSERT(start_idx >= 0);
+                GGML_ASSERT(end_idx > 0);
+
+                auto slice         = ggml_ext_slice(ctx->ggml_ctx, x, 2, start_idx, end_idx);
                 feat_map[feat_idx] = ggml_cont(ctx->ggml_ctx, slice);
             }
             feat_idx++;
@@ -284,7 +295,8 @@ namespace LTXVAE {
                              bool causal,
                              std::vector<ggml_tensor*>& feat_map,
                              int& feat_idx,
-                             int chunk_idx) {
+                             int chunk_idx,
+                             int temporal_pad = 0) {
             auto norm1 = std::dynamic_pointer_cast<PixelNorm3D>(blocks["norm1"]);
             auto conv1 = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
             auto norm2 = std::dynamic_pointer_cast<PixelNorm3D>(blocks["norm2"]);
@@ -311,14 +323,14 @@ namespace LTXVAE {
                 h = apply_scale_shift(ctx->ggml_ctx, h, scale1, shift1);
             }
             h = ggml_silu_inplace(ctx->ggml_ctx, h);
-            h = conv1->forward(ctx, h, feat_map, feat_idx, chunk_idx, causal);
+            h = conv1->forward(ctx, h, feat_map, feat_idx, chunk_idx, causal, temporal_pad);
 
             h = norm2->forward(ctx, h);
             if (timestep_conditioning) {
                 h = apply_scale_shift(ctx->ggml_ctx, h, scale2, shift2);
             }
             h = ggml_silu_inplace(ctx->ggml_ctx, h);
-            h = conv2->forward(ctx, h, feat_map, feat_idx, chunk_idx, causal);
+            h = conv2->forward(ctx, h, feat_map, feat_idx, chunk_idx, causal, temporal_pad);
 
             return ggml_add(ctx->ggml_ctx, h, x);
         }
@@ -367,7 +379,8 @@ namespace LTXVAE {
                              bool causal,
                              std::vector<ggml_tensor*>& feat_map,
                              int& feat_idx,
-                             int chunk_idx) {
+                             int chunk_idx,
+                             int temporal_pad = 0) {
             ggml_tensor* timestep_embed = nullptr;
             if (timestep_conditioning) {
                 GGML_ASSERT(timestep != nullptr);
@@ -376,7 +389,7 @@ namespace LTXVAE {
             }
             for (int i = 0; i < num_layers; i++) {
                 auto resnet = std::dynamic_pointer_cast<ResnetBlock3D>(blocks["res_blocks." + std::to_string(i)]);
-                x           = resnet->forward(ctx, x, timestep_embed, causal, feat_map, feat_idx, chunk_idx);
+                x           = resnet->forward(ctx, x, timestep_embed, causal, feat_map, feat_idx, chunk_idx, temporal_pad);
             }
             return x;
         }
@@ -437,7 +450,8 @@ namespace LTXVAE {
                              bool causal,
                              std::vector<ggml_tensor*>& feat_map,
                              int& feat_idx,
-                             int chunk_idx) {
+                             int chunk_idx,
+                             int temporal_pad = 0) {
             auto conv = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv"]);
 
             bool drop_first = (chunk_idx == 0) && (factor_t > 1);
@@ -453,7 +467,7 @@ namespace LTXVAE {
                 x_in = res;
             }
 
-            x = conv->forward(ctx, x, feat_map, feat_idx, chunk_idx, causal);
+            x = conv->forward(ctx, x, feat_map, feat_idx, chunk_idx, causal, temporal_pad);
             x = depth_to_space_3d(ctx->ggml_ctx, x, get_output_channels(), factor_t, factor_s, drop_first);
             if (residual) {
                 x = ggml_add(ctx->ggml_ctx, x, x_in);
@@ -943,8 +957,8 @@ namespace LTXVAE {
 
             ggml_tensor* scaled_timestep = timestep;
             if (timestep_conditioning) {
-                auto multiplier = ggml_ext_backend_tensor_get_f32(params["timestep_scale_multiplier"]);
-                scaled_timestep = ggml_ext_scale(ctx->ggml_ctx, timestep, multiplier);
+                auto multiplier = ggml_ext_cast_f32(ctx->ggml_ctx, ctx->backend, params["timestep_scale_multiplier"]);
+                scaled_timestep = ggml_mul(ctx->ggml_ctx, timestep, multiplier);
             }
 
             x = conv_in->forward(ctx, x, causal_decoder);
@@ -986,19 +1000,20 @@ namespace LTXVAE {
                                          ggml_tensor* timestep,
                                          std::vector<ggml_tensor*>& feat_map,
                                          int& feat_idx,
-                                         int chunk_idx) {
+                                         int chunk_idx,
+                                         int& temporal_pad) {
             auto conv_in       = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv_in"]);
             auto conv_norm_out = std::dynamic_pointer_cast<PixelNorm3D>(blocks["conv_norm_out"]);
             auto conv_out      = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv_out"]);
 
             ggml_tensor* scaled_timestep = timestep;
             if (timestep_conditioning && timestep != nullptr) {
-                auto multiplier = ggml_ext_backend_tensor_get_f32(params["timestep_scale_multiplier"]);
-                scaled_timestep = ggml_ext_scale(ctx->ggml_ctx, timestep, multiplier);
+                auto multiplier = ggml_ext_cast_f32(ctx->ggml_ctx, ctx->backend, params["timestep_scale_multiplier"]);
+                scaled_timestep = ggml_mul(ctx->ggml_ctx, timestep, multiplier);
             }
 
             // conv_in with feat_map for left temporal context
-            x = conv_in->forward(ctx, x, feat_map, feat_idx, chunk_idx, causal_decoder);
+            x = conv_in->forward(ctx, x, feat_map, feat_idx, chunk_idx, causal_decoder, temporal_pad);
 
             // up_blocks
             int block_idx = 0;
@@ -1006,12 +1021,13 @@ namespace LTXVAE {
                 auto mid_block = std::dynamic_pointer_cast<UNetMidBlock3D>(blocks["up_blocks." + std::to_string(block_idx)]);
                 if (mid_block) {
                     x = mid_block->forward(ctx, x, scaled_timestep, causal_decoder,
-                                           feat_map, feat_idx, chunk_idx);
+                                           feat_map, feat_idx, chunk_idx, temporal_pad);
                 } else {
                     auto upsample = std::dynamic_pointer_cast<DepthToSpaceUpsample>(
                         blocks["up_blocks." + std::to_string(block_idx)]);
                     x = upsample->forward(ctx, x, causal_decoder,
-                                          feat_map, feat_idx, chunk_idx);
+                                          feat_map, feat_idx, chunk_idx, temporal_pad);
+                    temporal_pad *= upsample->factor_t;
                 }
                 block_idx++;
             }
@@ -1028,7 +1044,7 @@ namespace LTXVAE {
                 x                       = apply_scale_shift(ctx->ggml_ctx, x, scale, shift);
             }
             x = ggml_silu_inplace(ctx->ggml_ctx, x);
-            x = conv_out->forward(ctx, x, feat_map, feat_idx, chunk_idx, causal_decoder);
+            x = conv_out->forward(ctx, x, feat_map, feat_idx, chunk_idx, causal_decoder, temporal_pad);
             return x;
         }
     };
@@ -1084,7 +1100,9 @@ namespace LTXVAE {
         // tensors can be freed by GGML before the next iteration starts.
         ggml_tensor* decode_tiled(GGMLRunnerContext* ctx,
                                   ggml_tensor* z,
-                                  ggml_tensor* timestep) {
+                                  ggml_tensor* timestep,
+                                  int temporal_window_size  = 1,
+                                  int temporal_tile_overlap = 0) {
             auto decoder   = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
             auto processor = std::dynamic_pointer_cast<PerChannelStatistics>(blocks["per_channel_statistics"]);
             auto latents   = processor->un_normalize(ctx, z);
@@ -1099,18 +1117,69 @@ namespace LTXVAE {
             // 128 slots is generous enough for any supported decoder configuration.
             std::vector<ggml_tensor*> feat_map(128, nullptr);
 
+            // Ensure window size is at least 1
+            int window  = std::max(1, temporal_window_size);
+            int overlap = std::max(0, temporal_tile_overlap);
+
+            if (overlap >= window) {
+                LOG_WARN("temporal_tile_overlap (%d) is greater than or equal to temporal_tile_frames (%d), adjusting values to avoid empty decode windows",
+                         overlap, window);
+                overlap = window - 1;
+            }
+            LOG_DEBUG("Using temporal tiling: temporal_tile_frames = %d, temporal_tile_overlap = %d, total frames = %d, resulting in %d tiles",
+                      window,
+                      overlap,
+                      (int)T,
+                      (T + window - overlap - 1) / (window - overlap));
             ggml_tensor* out = nullptr;
-            for (int i = 0; i < (int)T; i++) {
+            for (int i = 0; i < (int)T - overlap; i += (window - overlap)) {
                 int feat_idx = 0;
-                auto z_i     = ggml_ext_slice(ctx->ggml_ctx, latents, 2, i, i + 1);
-                auto out_i   = decoder->forward_tiled_frame(ctx, z_i, timestep,
-                                                            feat_map, feat_idx, i);
-                out          = (out == nullptr) ? out_i : ggml_concat(ctx->ggml_ctx, out, out_i, 2);
+
+                // Calculate the end index for the current temporal chunk
+                int end_i = std::min((int)T, i + window);
+                if (end_i >= (int)T) {
+                    overlap = 0;  // avoid overlap issues in the last chunk
+                }
+
+                int chunk_overlap = overlap;  // modified by forward_tiled_frame temporal inflation
+
+                auto z_chunk = ggml_ext_slice(ctx->ggml_ctx, latents, 2, i, end_i);
+
+                auto out_chunk = decoder->forward_tiled_frame(ctx, z_chunk, timestep,
+                                                              feat_map, feat_idx, i, chunk_overlap);
+
+                // discard overlap frames if it's not the final chunk
+                if (overlap > 0 && end_i < (int)T) {
+                    out_chunk = ggml_ext_slice(ctx->ggml_ctx, out_chunk, 2, 0, out_chunk->ne[2] - chunk_overlap);
+                }
+
+                out = (out == nullptr) ? out_chunk : ggml_concat(ctx->ggml_ctx, out, out_chunk, 2);
             }
 
             return WAN::WanVAE::unpatchify(ctx->ggml_ctx, out, patch_size, 1);
         }
 
+        ggml_tensor* decode_tiled_chunk(GGMLRunnerContext* ctx,
+                                        ggml_tensor* z,
+                                        ggml_tensor* timestep,
+                                        std::vector<ggml_tensor*>& feat_map,
+                                        int chunk_idx,
+                                        int temporal_tile_overlap,
+                                        int& feat_idx) {
+            auto decoder   = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
+            auto processor = std::dynamic_pointer_cast<PerChannelStatistics>(blocks["per_channel_statistics"]);
+            auto latents   = processor->un_normalize(ctx, z);
+
+            feat_idx          = 0;
+            int chunk_overlap = temporal_tile_overlap;  // modified by forward_tiled_frame temporal inflation
+            auto out_chunk    = decoder->forward_tiled_frame(ctx, latents, timestep,
+                                                             feat_map, feat_idx, chunk_idx, chunk_overlap);
+            if (chunk_overlap > 0) {
+                out_chunk = ggml_ext_slice(ctx->ggml_ctx, out_chunk, 2, 0, out_chunk->ne[2] - chunk_overlap);
+            }
+            return WAN::WanVAE::unpatchify(ctx->ggml_ctx, out_chunk, patch_size, 1);
+        }
+
         ggml_tensor* encode(GGMLRunnerContext* ctx,
                             ggml_tensor* x) {
             GGML_ASSERT(!decode_only);
@@ -1123,13 +1192,30 @@ namespace LTXVAE {
             mean      = ggml_cont(ctx->ggml_ctx, mean);
             return processor->normalize(ctx, mean);
         }
+
+        ggml_tensor* normalize_latents(GGMLRunnerContext* ctx,
+                                       ggml_tensor* x) {
+            auto processor = std::dynamic_pointer_cast<PerChannelStatistics>(blocks["per_channel_statistics"]);
+            return processor->normalize(ctx, x);
+        }
+
+        ggml_tensor* un_normalize_latents(GGMLRunnerContext* ctx,
+                                          ggml_tensor* x) {
+            auto processor = std::dynamic_pointer_cast<PerChannelStatistics>(blocks["per_channel_statistics"]);
+            return processor->un_normalize(ctx, x);
+        }
     };
 
 }  // namespace LTXVAE
 
 struct LTXVideoVAE : public VAE {
+    static constexpr int DEFAULT_TEMPORAL_TILE_FRAMES  = 4;
+    static constexpr int DEFAULT_TEMPORAL_TILE_OVERLAP = 1;
+
     bool decode_only;
     bool temporal_tiling_enabled = false;
+    int temporal_tile_frames     = DEFAULT_TEMPORAL_TILE_FRAMES;
+    int temporal_tile_overlap    = DEFAULT_TEMPORAL_TILE_OVERLAP;
     int ltx_vae_version;
     bool timestep_conditioning;
     int patch_size;
@@ -1137,11 +1223,11 @@ struct LTXVideoVAE : public VAE {
     LTXVAE::VideoVAE vae;
 
     LTXVideoVAE(ggml_backend_t backend,
-                ggml_backend_t params_backend,
                 const String2TensorStorage& tensor_storage_map,
                 const std::string& prefix,
-                bool decode_only  = true,
-                SDVersion version = VERSION_LTXAV)
+                bool decode_only                                    = true,
+                SDVersion version                                   = VERSION_LTXAV,
+                std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
         : decode_only(decode_only),
           ltx_vae_version(LTXVAE::detect_ltx_vae_version(tensor_storage_map, prefix)),
           timestep_conditioning(LTXVAE::detect_ltx_vae_timestep_conditioning(tensor_storage_map, prefix)),
@@ -1153,7 +1239,7 @@ struct LTXVideoVAE : public VAE {
               patch_size,
               tensor_storage_map,
               prefix),
-          VAE(version, backend, params_backend) {
+          VAE(version, backend, prefix, weight_manager) {
         vae.init(params_ctx, tensor_storage_map, prefix);
         decode_timestep_tensor.values()[0] = vae.decode_timestep;
     }
@@ -1166,8 +1252,62 @@ struct LTXVideoVAE : public VAE {
         temporal_tiling_enabled = enabled;
     }
 
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) override {
-        vae.get_param_tensors(tensors, prefix);
+    void set_tiling_params(const sd_tiling_params_t& params) override {
+        temporal_tiling_enabled = params.temporal_tiling;
+        temporal_tile_frames    = DEFAULT_TEMPORAL_TILE_FRAMES;
+        temporal_tile_overlap   = DEFAULT_TEMPORAL_TILE_OVERLAP;
+
+        for (const auto& [key, value] : parse_key_value_args(params.extra_tiling_args, "LTX VAE extra tiling arg")) {
+            int parsed = 0;
+            if (!parse_strict_int(value, parsed)) {
+                LOG_WARN("ignoring invalid LTX VAE extra tiling arg '%s=%s'", key.c_str(), value.c_str());
+            } else if (key == "temporal_tile_frames") {
+                temporal_tile_frames = std::max(1, parsed);
+            } else if (key == "temporal_tile_overlap") {
+                temporal_tile_overlap = std::max(0, parsed);
+            } else {
+                LOG_WARN("ignoring unknown LTX VAE extra tiling arg '%s'", key.c_str());
+            }
+        }
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        vae.get_param_tensors(tensors, weight_prefix);
+    }
+
+    struct TemporalTilePlan {
+        int frames    = 1;
+        int overlap   = 0;
+        int stride    = 1;
+        int num_tiles = 1;
+    };
+
+    TemporalTilePlan resolve_temporal_tile_plan(int64_t total_frames) const {
+        TemporalTilePlan plan;
+        plan.frames  = std::max(1, temporal_tile_frames);
+        plan.overlap = std::max(0, temporal_tile_overlap);
+
+        if (plan.overlap >= plan.frames) {
+            LOG_WARN("temporal_tile_overlap (%d) is greater than or equal to temporal_tile_frames (%d), adjusting values to avoid empty decode windows",
+                     plan.overlap,
+                     plan.frames);
+            plan.overlap = plan.frames - 1;
+        }
+        if (total_frames > 1 && plan.overlap >= total_frames) {
+            LOG_WARN("temporal_tile_overlap (%d) is greater than or equal to total latent frames (%lld), adjusting values to decode at least one tile",
+                     plan.overlap,
+                     (long long)total_frames);
+            plan.overlap = static_cast<int>(total_frames - 1);
+        }
+
+        plan.stride          = std::max(1, plan.frames - plan.overlap);
+        int64_t tiled_frames = std::max<int64_t>(1, total_frames - plan.overlap);
+        plan.num_tiles       = total_frames > 0 ? static_cast<int>((tiled_frames + plan.stride - 1) / plan.stride) : 0;
+        return plan;
+    }
+
+    std::string temporal_feat_cache_name(size_t feat_idx) const {
+        return "ltx_vae_temporal_feat:" + std::to_string(feat_idx);
     }
 
     ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
@@ -1180,18 +1320,108 @@ struct LTXVideoVAE : public VAE {
 
         auto runner_ctx = get_context();
         ggml_tensor* out;
-        bool use_tiled = decode_graph && temporal_tiling_enabled &&
-                         z_tensor.dim() == 5 && z_tensor.shape()[2] > 1;
-        if (use_tiled) {
-            out = vae.decode_tiled(&runner_ctx, z, timestep);
-        } else {
-            out = decode_graph ? vae.decode(&runner_ctx, z, timestep) : vae.encode(&runner_ctx, z);
-        }
+        out = decode_graph ? vae.decode(&runner_ctx, z, timestep) : vae.encode(&runner_ctx, z);
         ggml_build_forward_expand(gf, out);
 
         return gf;
     }
 
+    ggml_cgraph* build_temporal_tile_graph(const sd::Tensor<float>& z_chunk_tensor,
+                                           int chunk_idx,
+                                           int chunk_overlap) {
+        ggml_cgraph* gf       = new_graph_custom(20480);
+        ggml_tensor* z        = make_input(z_chunk_tensor);
+        ggml_tensor* timestep = nullptr;
+        if (timestep_conditioning) {
+            timestep = make_input(decode_timestep_tensor);
+        }
+
+        std::vector<ggml_tensor*> feat_map(128, nullptr);
+        for (size_t feat_idx = 0; feat_idx < feat_map.size(); ++feat_idx) {
+            feat_map[feat_idx] = get_cache_tensor_by_name(temporal_feat_cache_name(feat_idx));
+        }
+
+        auto runner_ctx  = get_context();
+        int feat_count   = 0;
+        ggml_tensor* out = vae.decode_tiled_chunk(&runner_ctx,
+                                                  z,
+                                                  timestep,
+                                                  feat_map,
+                                                  chunk_idx,
+                                                  chunk_overlap,
+                                                  feat_count);
+
+        for (int feat_idx = 0; feat_idx < feat_count && feat_idx < static_cast<int>(feat_map.size()); ++feat_idx) {
+            ggml_tensor* feat_cache = feat_map[static_cast<size_t>(feat_idx)];
+            if (feat_cache != nullptr) {
+                cache(temporal_feat_cache_name(static_cast<size_t>(feat_idx)), feat_cache);
+                ggml_build_forward_expand(gf, feat_cache);
+            }
+        }
+
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
+    sd::Tensor<float> decode_temporal_tiled_streaming(const int n_threads,
+                                                      const sd::Tensor<float>& input,
+                                                      size_t expected_dim) {
+        const int64_t total_frames = input.shape()[2];
+        TemporalTilePlan plan      = resolve_temporal_tile_plan(total_frames);
+
+        LOG_DEBUG("Using streaming temporal tiling: temporal_tile_frames=%d, temporal_tile_overlap=%d, total latent frames=%lld, resulting in %d tiles",
+                  plan.frames,
+                  plan.overlap,
+                  (long long)total_frames,
+                  plan.num_tiles);
+
+        free_cache_ctx_and_buffer();
+        cache_tensor_map.clear();
+
+        sd::Tensor<float> output;
+        for (int64_t start = 0; start < total_frames - plan.overlap; start += plan.stride) {
+            const int64_t end       = std::min<int64_t>(total_frames, start + plan.frames);
+            const int chunk_overlap = end < total_frames ? plan.overlap : 0;
+            auto z_chunk            = sd::ops::slice(input, 2, start, end);
+
+            LOG_DEBUG("LTX VAE temporal tile %lld/%d: latent frames [%lld, %lld), overlap=%d",
+                      (long long)(start / plan.stride + 1),
+                      plan.num_tiles,
+                      (long long)start,
+                      (long long)end,
+                      chunk_overlap);
+
+            auto get_graph = [&]() -> ggml_cgraph* {
+                return build_temporal_tile_graph(z_chunk,
+                                                 static_cast<int>(start),
+                                                 chunk_overlap);
+            };
+            auto chunk = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, true, true, true),
+                                                         expected_dim);
+            if (chunk.empty()) {
+                free_cache_ctx_and_buffer();
+                cache_tensor_map.clear();
+                return {};
+            }
+            output = output.empty() ? std::move(chunk) : sd::ops::concat(output, chunk, 2);
+        }
+
+        free_cache_ctx_and_buffer();
+        cache_tensor_map.clear();
+        return output;
+    }
+
+    ggml_cgraph* build_latent_statistics_graph(const sd::Tensor<float>& z_tensor, bool normalize) {
+        ggml_cgraph* gf = new_graph_custom(1024);
+        ggml_tensor* z  = make_input(z_tensor);
+
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = normalize ? vae.normalize_latents(&runner_ctx, z)
+                                     : vae.un_normalize_latents(&runner_ctx, z);
+        ggml_build_forward_expand(gf, out);
+        return gf;
+    }
+
     sd::Tensor<float> _compute(const int n_threads,
                                const sd::Tensor<float>& z,
                                bool decode_graph) override {
@@ -1216,16 +1446,39 @@ struct LTXVideoVAE : public VAE {
                 input = sd::ops::slice(input, 2, 0, cropped_t);
             }
         }
+        if (decode_graph && temporal_tiling_enabled && input.dim() == 5 && input.shape()[2] > 1) {
+            return decode_temporal_tiled_streaming(n_threads, input, expected_dim);
+        }
         auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(input, decode_graph);
         };
-        auto result = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), expected_dim);
+        auto result = restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), expected_dim);
         if (result.empty()) {
             return {};
         }
         return result;
     }
 
+    sd::Tensor<float> apply_latent_statistics(const int n_threads,
+                                              const sd::Tensor<float>& z,
+                                              bool normalize) {
+        auto get_graph = [&]() -> ggml_cgraph* {
+            return build_latent_statistics_graph(z, normalize);
+        };
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false),
+                                               static_cast<size_t>(z.dim()));
+    }
+
+    sd::Tensor<float> normalize_latents(const int n_threads,
+                                        const sd::Tensor<float>& z) {
+        return apply_latent_statistics(n_threads, z, true);
+    }
+
+    sd::Tensor<float> un_normalize_latents(const int n_threads,
+                                           const sd::Tensor<float>& z) {
+        return apply_latent_statistics(n_threads, z, false);
+    }
+
     int get_encoder_output_channels(int input_channels) override {
         SD_UNUSED(input_channels);
         return 256;
@@ -1265,10 +1518,11 @@ struct LTXVideoVAE : public VAE {
     static void load_from_file_and_test(const std::string& model_path,
                                         const std::string& input_path) {
         // ggml_backend_t backend = ggml_backend_cuda_init(0);
-        ggml_backend_t backend = ggml_backend_cpu_init();
+        ggml_backend_t backend = sd_backend_cpu_init();
         LOG_INFO("loading ltx vae from '%s'", model_path.c_str());
 
-        ModelLoader model_loader;
+        auto model_manager        = std::make_shared<ModelManager>();
+        ModelLoader& model_loader = model_manager->loader();
         if (!model_loader.init_from_file_and_convert_name(model_path, "vae.")) {
             LOG_ERROR("init model loader from file failed: '%s'", model_path.c_str());
             return;
@@ -1276,18 +1530,19 @@ struct LTXVideoVAE : public VAE {
 
         auto& tensor_storage_map         = model_loader.get_tensor_storage_map();
         std::shared_ptr<LTXVideoVAE> vae = std::make_shared<LTXVideoVAE>(backend,
-                                                                         backend,
                                                                          tensor_storage_map,
                                                                          "first_stage_model",
                                                                          true,
-                                                                         VERSION_LTXAV);
+                                                                         VERSION_LTXAV,
+                                                                         model_manager);
 
-        vae->alloc_params_buffer();
-        std::map<std::string, ggml_tensor*> tensors;
-        vae->get_param_tensors(tensors, "first_stage_model");
-
-        if (!model_loader.load_tensors(tensors)) {
-            LOG_ERROR("load tensors from model loader failed");
+        if (!model_manager->register_runner_params("LTX VAE test",
+                                                   *vae,
+                                                   ModelManager::ResidencyMode::ParamBackend,
+                                                   backend,
+                                                   backend) ||
+            !model_manager->validate_registered_tensors()) {
+            LOG_ERROR("register ltx vae tensors with model manager failed");
             return;
         }
 
@@ -1296,4 +1551,4 @@ struct LTXVideoVAE : public VAE {
     }
 };
 
-#endif  // __SD_LTX_VAE_HPP__
+#endif  // __SD_MODEL_VAE_LTX_VAE_HPP__

src/model/vae/tae.hpp

@@ -1,13 +1,13 @@
-#ifndef __TAE_HPP__
-#define __TAE_HPP__
+#ifndef __SD_MODEL_VAE_TAE_HPP__
+#define __SD_MODEL_VAE_TAE_HPP__
 
-#include "ggml_extend.hpp"
+#include "core/ggml_extend.hpp"
 #include "model.h"
 
 /*
     ===================================    TinyAutoEncoder  ===================================
     References:
-    https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/autoencoders/vae.py
+    https://github.com/huggingface/diffusers/blob/main/src/diffusers/model/autoencoders/vae.py
     https://github.com/madebyollin/taesd/blob/main/taesd.py
 
 */
@@ -259,10 +259,54 @@ public:
     }
 };
 
-ggml_tensor* patchify(ggml_context* ctx,
-                      ggml_tensor* x,
-                      int64_t patch_size,
-                      int64_t b = 1) {
+class WideMemBlock : public GGMLBlock {
+    bool has_skip_conv = false;
+
+public:
+    WideMemBlock(int channels, int out_channels)
+        : has_skip_conv(channels != out_channels) {
+        int groups       = std::max(1, out_channels / 64);
+        blocks["conv.0"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * 2, out_channels, {1, 1}, {1, 1}));
+        blocks["conv.2"] = std::shared_ptr<GGMLBlock>(new Conv2d_grouped(out_channels, out_channels, groups, {3, 3}, {1, 1}, {1, 1}));
+        blocks["conv.4"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {1, 1}, {1, 1}));
+        blocks["conv.6"] = std::shared_ptr<GGMLBlock>(new Conv2d_grouped(out_channels, out_channels, groups, {3, 3}, {1, 1}, {1, 1}));
+        if (has_skip_conv) {
+            blocks["skip"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
+        }
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* past) {
+        // x: [n, channels, h, w]
+        auto conv0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
+        auto conv1 = std::dynamic_pointer_cast<Conv2d_grouped>(blocks["conv.2"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.4"]);
+        auto conv3 = std::dynamic_pointer_cast<Conv2d_grouped>(blocks["conv.6"]);
+
+        auto h = ggml_concat(ctx->ggml_ctx, x, past, 2);
+        h      = conv0->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv1->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv2->forward(ctx, h);
+        h      = ggml_relu_inplace(ctx->ggml_ctx, h);
+        h      = conv3->forward(ctx, h);
+
+        auto skip = x;
+        if (has_skip_conv) {
+            auto skip_conv = std::dynamic_pointer_cast<Conv2d>(blocks["skip"]);
+            skip           = skip_conv->forward(ctx, x);
+        }
+        h = ggml_add_inplace(ctx->ggml_ctx, h, skip);
+        h = ggml_relu_inplace(ctx->ggml_ctx, h);
+        return h;
+    }
+};
+
+ggml_tensor*
+patchify(ggml_context* ctx,
+         ggml_tensor* x,
+         int64_t patch_size,
+         int64_t b = 1) {
     // x: [f, b*c, h*q, w*r]
     // return: [f, b*c*r*q, h, w]
     if (patch_size == 1) {
@@ -322,13 +366,20 @@ class TinyVideoEncoder : public UnaryBlock {
     int patch_size  = 1;
 
 public:
-    TinyVideoEncoder(int z_channels = 4, int patch_size = 1)
+    int t_downscale = 1;
+    TinyVideoEncoder(int z_channels = 4, int patch_size = 1, std::vector<bool> time_downscale = {true, true, false})
         : z_channels(z_channels), patch_size(patch_size) {
+        t_downscale = 1;
+        for (bool downscale : time_downscale) {
+            if (downscale) {
+                t_downscale *= 2;
+            }
+        }
         int index                       = 0;
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels * patch_size * patch_size, hidden, {3, 3}, {1, 1}, {1, 1}));
         index++;  // nn.ReLU()
         for (int i = 0; i < num_layers; i++) {
-            int stride                      = i == num_layers - 1 ? 1 : 2;
+            int stride                      = time_downscale[i] ? 2 : 1;
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TPool(hidden, stride));
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, hidden, {3, 3}, {2, 2}, {1, 1}, {1, 1}, false));
             for (int j = 0; j < num_blocks; j++) {
@@ -375,17 +426,35 @@ class TinyVideoDecoder : public UnaryBlock {
     static const int num_layers  = 3;
     int channels[num_layers + 1] = {256, 128, 64, 64};
     int patch_size               = 1;
+    int t_upscale                = 1;
+    bool is_wide                 = false;
 
 public:
-    TinyVideoDecoder(int z_channels = 4, int patch_size = 1)
-        : z_channels(z_channels), patch_size(patch_size) {
+    TinyVideoDecoder(int z_channels = 4, int patch_size = 1, std::vector<bool> time_upscale = {false, true, true}, bool is_wide = false)
+        : z_channels(z_channels), patch_size(patch_size), is_wide(is_wide) {
+        t_upscale = 1;
+        if (is_wide) {
+            channels[0] = 1024;
+            channels[1] = 512;
+            channels[2] = 256;
+        }
+
+        for (bool upscale : time_upscale) {
+            if (upscale) {
+                t_upscale *= 2;
+            }
+        }
         int index                       = 1;  // Clamp()
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, channels[0], {3, 3}, {1, 1}, {1, 1}));
         index++;  // nn.ReLU()
         for (int i = 0; i < num_layers; i++) {
-            int stride = i == 0 ? 1 : 2;
+            int stride = time_upscale[i] ? 2 : 1;
             for (int j = 0; j < num_blocks; j++) {
-                blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new MemBlock(channels[i], channels[i]));
+                if (is_wide) {
+                    blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new WideMemBlock(channels[i], channels[i]));
+                } else {
+                    blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new MemBlock(channels[i], channels[i]));
+                }
             }
             index++;  // nn.Upsample()
             blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new TGrow(channels[i], stride));
@@ -410,10 +479,15 @@ public:
         int index = 3;
         for (int i = 0; i < num_layers; i++) {
             for (int j = 0; j < num_blocks; j++) {
-                auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
-                auto mem   = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
-                mem        = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
-                h          = block->forward(ctx, h, mem);
+                auto mem = ggml_pad_ext(ctx->ggml_ctx, h, 0, 0, 0, 0, 0, 0, 1, 0);
+                mem      = ggml_view_4d(ctx->ggml_ctx, mem, h->ne[0], h->ne[1], h->ne[2], h->ne[3], h->nb[1], h->nb[2], h->nb[3], 0);
+                if (is_wide) {
+                    auto block = std::dynamic_pointer_cast<WideMemBlock>(blocks[std::to_string(index++)]);
+                    h          = block->forward(ctx, h, mem);
+                } else {
+                    auto block = std::dynamic_pointer_cast<MemBlock>(blocks[std::to_string(index++)]);
+                    h          = block->forward(ctx, h, mem);
+                }
             }
             // upsample
             index++;
@@ -430,8 +504,8 @@ public:
         if (patch_size > 1) {
             h = unpatchify(ctx->ggml_ctx, h, patch_size, 1);
         }
-        // shape(W, H, 3, 3 + T) => shape(W, H, 3, T)
-        h = ggml_view_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2], h->ne[3] - 3, h->nb[1], h->nb[2], h->nb[3], 3 * h->nb[3]);
+        // shape(W, H, 3, (t_upscale - 1) + T) => shape(W, H, 3, T)
+        h = ggml_view_4d(ctx->ggml_ctx, h, h->ne[0], h->ne[1], h->ne[2], h->ne[3] - (t_upscale - 1), h->nb[1], h->nb[2], h->nb[3], (t_upscale - 1) * h->nb[3]);
         return h;
     }
 };
@@ -440,32 +514,40 @@ class TAEHV : public GGMLBlock {
 protected:
     bool decode_only;
     SDVersion version;
+    bool is_wide;
 
 public:
-    int z_channels = 16;
+    int z_channels                   = 16;
+    std::vector<bool> time_downscale = {true, true, false};
+    std::vector<bool> time_upscale   = {false, true, true};
 
 public:
-    TAEHV(bool decode_only = true, SDVersion version = VERSION_WAN2)
-        : decode_only(decode_only), version(version) {
+    TAEHV(bool decode_only = true, SDVersion version = VERSION_WAN2, bool is_wide = false)
+        : decode_only(decode_only), version(version), is_wide(is_wide) {
         int patch = 1;
         if (version == VERSION_WAN2_2_TI2V) {
             z_channels = 48;
             patch      = 2;
+        } else if (sd_version_is_ltxav(version)) {
+            z_channels     = 128;
+            patch          = 4;
+            time_downscale = {true, true, true};
+            time_upscale   = {true, true, true};
         }
-        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoDecoder(z_channels, patch));
+        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoDecoder(z_channels, patch, time_upscale, is_wide));
         if (!decode_only) {
-            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoEncoder(z_channels, patch));
+            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new TinyVideoEncoder(z_channels, patch, time_downscale));
         }
     }
 
     ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
         auto decoder = std::dynamic_pointer_cast<TinyVideoDecoder>(blocks["decoder"]);
-        if (sd_version_is_wan(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
             // (W, H, C, T) -> (W, H, T, C)
             z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 0, 1, 3, 2));
         }
         auto result = decoder->forward(ctx, z);
-        if (sd_version_is_wan(version)) {
+        if (sd_version_is_wan(version) || sd_version_is_ltxav(version)) {
             // (W, H, C, T) -> (W, H, T, C)
             result = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, result, 0, 1, 3, 2));
         }
@@ -477,10 +559,10 @@ public:
         // (W, H, T, C) -> (W, H, C, T)
         x                  = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
         int64_t num_frames = x->ne[3];
-        if (num_frames % 4) {
-            // pad to multiple of 4 at the end
+        if (num_frames % encoder->t_downscale) {
+            // pad to multiple of encoder->t_downscale at the end
             auto last_frame = ggml_view_4d(ctx->ggml_ctx, x, x->ne[0], x->ne[1], x->ne[2], 1, x->nb[1], x->nb[2], x->nb[3], (num_frames - 1) * x->nb[3]);
-            for (int i = 0; i < 4 - num_frames % 4; i++) {
+            for (int i = 0; i < encoder->t_downscale - num_frames % encoder->t_downscale; i++) {
                 x = ggml_concat(ctx->ggml_ctx, x, last_frame, 3);
             }
         }
@@ -502,7 +584,7 @@ public:
     TAESD(bool decode_only = true, SDVersion version = VERSION_SD1)
         : decode_only(decode_only) {
         bool use_midblock_gn = false;
-        taef2                = sd_version_is_flux2(version);
+        taef2                = sd_version_uses_flux2_vae(version);
 
         if (sd_version_is_dit(version)) {
             z_channels = 16;
@@ -541,14 +623,14 @@ struct TinyImageAutoEncoder : public VAE {
     bool decode_only = false;
 
     TinyImageAutoEncoder(ggml_backend_t backend,
-                         ggml_backend_t params_backend,
                          const String2TensorStorage& tensor_storage_map,
                          const std::string prefix,
-                         bool decoder_only = true,
-                         SDVersion version = VERSION_SD1)
-        : decode_only(decoder_only),
-          taesd(decoder_only, version),
-          VAE(version, backend, params_backend) {
+                         bool decoder_only                                   = true,
+                         SDVersion version                                   = VERSION_SD1,
+                         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : VAE(version, backend, "tae", weight_manager),
+          decode_only(decoder_only),
+          taesd(decoder_only, version) {
         scale_input = false;
         taesd.init(params_ctx, tensor_storage_map, prefix);
     }
@@ -557,8 +639,8 @@ struct TinyImageAutoEncoder : public VAE {
         return "taesd";
     }
 
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
-        taesd.get_param_tensors(tensors, prefix);
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        taesd.get_param_tensors(tensors, weight_prefix);
     }
 
     sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
@@ -594,23 +676,30 @@ struct TinyImageAutoEncoder : public VAE {
             return build_graph(z_tensor, decode_graph);
         };
 
-        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), z_tensor.dim());
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), z_tensor.dim());
     }
 };
 
 struct TinyVideoAutoEncoder : public VAE {
     TAEHV taehv;
     bool decode_only = false;
+    bool is_wide     = false;
 
     TinyVideoAutoEncoder(ggml_backend_t backend,
-                         ggml_backend_t params_backend,
                          const String2TensorStorage& tensor_storage_map,
                          const std::string prefix,
-                         bool decoder_only = true,
-                         SDVersion version = VERSION_WAN2)
-        : decode_only(decoder_only),
-          taehv(decoder_only, version),
-          VAE(version, backend, params_backend) {
+                         bool decoder_only                                   = true,
+                         SDVersion version                                   = VERSION_WAN2,
+                         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : VAE(version, backend, "tae", weight_manager),
+          decode_only(decoder_only) {
+        for (auto tensor_storage : tensor_storage_map) {
+            if (tensor_storage.first.find(prefix + ".3.conv.6.weight") != std::string::npos) {
+                is_wide = true;
+                break;
+            }
+        }
+        taehv       = TAEHV(decoder_only, version, is_wide);
         scale_input = false;
         taehv.init(params_ctx, tensor_storage_map, prefix);
     }
@@ -619,8 +708,8 @@ struct TinyVideoAutoEncoder : public VAE {
         return "taehv";
     }
 
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
-        taehv.get_param_tensors(tensors, prefix);
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
+        taehv.get_param_tensors(tensors, weight_prefix);
     }
 
     sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) override {
@@ -641,7 +730,8 @@ struct TinyVideoAutoEncoder : public VAE {
     }
 
     ggml_cgraph* build_graph(const sd::Tensor<float>& z_tensor, bool decode_graph) {
-        ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        ggml_cgraph* gf  = decode_graph && is_wide ? ggml_new_graph_custom(compute_ctx, 4096, false)
+                                                   : ggml_new_graph(compute_ctx);
         ggml_tensor* z   = make_input(z_tensor);
         auto runner_ctx  = get_context();
         ggml_tensor* out = decode_graph ? taehv.decode(&runner_ctx, z) : taehv.encode(&runner_ctx, z);
@@ -656,8 +746,8 @@ struct TinyVideoAutoEncoder : public VAE {
             return build_graph(z_tensor, decode_graph);
         };
 
-        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false), z_tensor.dim());
+        return restore_trailing_singleton_dims(GGMLRunner::compute<float>(get_graph, n_threads, false, false, false), z_tensor.dim());
     }
 };
 
-#endif  // __TAE_HPP__
+#endif  // __SD_MODEL_VAE_TAE_HPP__

src/model/vae/vae.hpp

@@ -1,12 +1,14 @@
-#ifndef __VAE_HPP__
-#define __VAE_HPP__
+#ifndef __SD_MODEL_VAE_VAE_HPP__
+#define __SD_MODEL_VAE_VAE_HPP__
 
-#include "common_block.hpp"
-#include "tensor_ggml.hpp"
+#include "core/tensor_ggml.hpp"
+#include "model/common/block.hpp"
+#include "model_manager.h"
 
 struct VAE : public GGMLRunner {
 protected:
     SDVersion version;
+    std::string weight_prefix;
     bool scale_input                                      = true;
     virtual sd::Tensor<float> _compute(const int n_threads,
                                        const sd::Tensor<float>& z,
@@ -62,8 +64,11 @@ protected:
     }
 
 public:
-    VAE(SDVersion version, ggml_backend_t backend, ggml_backend_t params_backend)
-        : version(version), GGMLRunner(backend, params_backend) {}
+    VAE(SDVersion version,
+        ggml_backend_t backend,
+        const std::string& weight_prefix                    = "",
+        std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : version(version), weight_prefix(weight_prefix), GGMLRunner(backend, weight_manager) {}
 
     int get_scale_factor() {
         int scale_factor = 8;
@@ -167,6 +172,7 @@ public:
         int64_t t0              = ggml_time_ms();
         sd::Tensor<float> input = x;
         sd::Tensor<float> output;
+        set_tiling_params(tiling_params);
 
         if (tiling_params.enabled) {
             const int scale_factor = get_scale_factor();
@@ -213,14 +219,19 @@ public:
     virtual sd::Tensor<float> vae_output_to_latents(const sd::Tensor<float>& vae_output, std::shared_ptr<RNG> rng) = 0;
     virtual sd::Tensor<float> diffusion_to_vae_latents(const sd::Tensor<float>& latents)                           = 0;
     virtual sd::Tensor<float> vae_to_diffusion_latents(const sd::Tensor<float>& latents)                           = 0;
-    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix)         = 0;
+    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors)                                   = 0;
     virtual void set_conv2d_scale(float scale) { SD_UNUSED(scale); };
     virtual void set_temporal_tiling_enabled(bool enabled) { SD_UNUSED(enabled); };
+    virtual void set_tiling_params(const sd_tiling_params_t& params) {
+        set_temporal_tiling_enabled(params.temporal_tiling);
+    };
 };
 
 struct FakeVAE : public VAE {
-    FakeVAE(SDVersion version, ggml_backend_t backend, ggml_backend_t params_backend)
-        : VAE(version, backend, params_backend) {}
+    FakeVAE(SDVersion version,
+            ggml_backend_t backend,
+            std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
+        : VAE(version, backend, "", weight_manager) {}
 
     int get_encoder_output_channels(int input_channels) {
         return input_channels;
@@ -247,11 +258,11 @@ struct FakeVAE : public VAE {
         return latents;
     }
 
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) override {}
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {}
 
     std::string get_desc() override {
         return "fake_vae";
     }
 };
 
-#endif  // __VAE_HPP__
+#endif  // __SD_MODEL_VAE_VAE_HPP__

src/model_io/gguf_io.cpp

@@ -5,9 +5,9 @@
 #include <string>
 #include <vector>
 
+#include "core/util.h"
 #include "gguf.h"
 #include "gguf_reader_ext.h"
-#include "util.h"
 
 static void set_error(std::string* error, const std::string& message) {
     if (error != nullptr) {

src/model_io/gguf_reader_ext.h

@@ -6,8 +6,8 @@
 #include <string>
 #include <vector>
 
+#include "core/util.h"
 #include "ggml.h"
-#include "util.h"
 
 struct GGUFTensorInfo {
     std::string name;

src/model_io/pickle_io.cpp

@@ -8,7 +8,7 @@
 #include <vector>
 
 #include "binary_io.h"
-#include "util.h"
+#include "core/util.h"
 
 // $ python -m pickletools sd-v1-4/archive/data.pkl | head -n 100
 //     0: \x80 PROTO      2

src/model_io/safetensors_io.cpp

@@ -7,8 +7,8 @@
 #include <vector>
 
 #include "binary_io.h"
+#include "core/util.h"
 #include "json.hpp"
-#include "util.h"
 
 static constexpr size_t ST_HEADER_SIZE_LEN = 8;
 

src/model_io/torch_legacy_io.cpp

@@ -7,8 +7,8 @@
 #include <unordered_map>
 #include <vector>
 
+#include "core/util.h"
 #include "pickle_io.h"
-#include "util.h"
 
 // torch.save format background:
 //

src/model_loader.cpp

@@ -1,6 +1,7 @@
 #include <algorithm>
 #include <atomic>
 #include <chrono>
+#include <cinttypes>
 #include <cstdarg>
 #include <cstdlib>
 #include <fstream>
@@ -13,18 +14,18 @@
 #include <unordered_map>
 #include <vector>
 
-#include "model.h"
+#include "core/util.h"
 #include "model_io/gguf_io.h"
 #include "model_io/safetensors_io.h"
 #include "model_io/torch_legacy_io.h"
 #include "model_io/torch_zip_io.h"
+#include "model_loader.h"
 #include "stable-diffusion.h"
-#include "util.h"
 
+#include "core/ggml_extend_backend.h"
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 #include "ggml.h"
-#include "ggml_extend_backend.h"
 #include "zip.h"
 
 #include "name_conversion.h"
@@ -204,10 +205,28 @@ void convert_tensor(void* src,
 
 /*================================================= ModelLoader ==================================================*/
 
+ModelLoader::ModelLoader()
+    : n_threads_(sd_get_num_physical_cores()) {
+}
+
+size_t ModelLoader::add_file_path(const std::string& file_path) {
+    if (model_files_processed) {
+        file_data.clear();
+        model_files_processed = false;
+    }
+    file_paths_.push_back(file_path);
+    return file_paths_.size() - 1;
+}
+
 void ModelLoader::add_tensor_storage(const TensorStorage& tensor_storage) {
     tensor_storage_map[tensor_storage.name] = tensor_storage;
 }
 
+void ModelLoader::set_n_threads(int n_threads) {
+    n_threads_ = n_threads > 0 ? n_threads : sd_get_num_physical_cores();
+    LOG_DEBUG("using %d threads for model loading", n_threads_);
+}
+
 bool ModelLoader::init_from_file(const std::string& file_path, const std::string& prefix) {
     if (is_directory(file_path)) {
         LOG_INFO("load %s using diffusers format", file_path.c_str());
@@ -271,8 +290,7 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
         return false;
     }
 
-    file_paths_.push_back(file_path);
-    size_t file_index = file_paths_.size() - 1;
+    size_t file_index = add_file_path(file_path);
 
     for (auto& tensor_storage : tensor_storages) {
         // LOG_DEBUG("%s", tensor_storage.name.c_str());
@@ -300,8 +318,7 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
         return false;
     }
 
-    file_paths_.push_back(file_path);
-    size_t file_index = file_paths_.size() - 1;
+    size_t file_index = add_file_path(file_path);
 
     for (auto& tensor_storage : tensor_storages) {
         if (is_unused_tensor(tensor_storage.name)) {
@@ -335,8 +352,7 @@ bool ModelLoader::init_from_torch_legacy_file(const std::string& file_path, cons
         return false;
     }
 
-    file_paths_.push_back(file_path);
-    size_t file_index = file_paths_.size() - 1;
+    size_t file_index = add_file_path(file_path);
 
     for (auto& tensor_storage : tensor_storages) {
         if (is_unused_tensor(tensor_storage.name)) {
@@ -366,8 +382,7 @@ bool ModelLoader::init_from_torch_zip_file(const std::string& file_path, const s
         return false;
     }
 
-    file_paths_.push_back(file_path);
-    size_t file_index = file_paths_.size() - 1;
+    size_t file_index = add_file_path(file_path);
 
     for (auto& tensor_storage : tensor_storages) {
         if (!starts_with(tensor_storage.name, prefix)) {
@@ -410,7 +425,7 @@ bool ModelLoader::init_from_diffusers_file(const std::string& file_path, const s
 }
 
 SDVersion ModelLoader::get_sd_version() {
-    TensorStorage token_embedding_weight, input_block_weight;
+    TensorStorage token_embedding_weight, input_block_weight, context_ebedding_weight;
 
     bool has_multiple_encoders = false;
     bool is_unet               = false;
@@ -428,9 +443,16 @@ SDVersion ModelLoader::get_sd_version() {
     bool has_attn_1024               = false;
 
     for (auto& [name, tensor_storage] : tensor_storage_map) {
-        if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos) {
+        if (tensor_storage.name.find("model.diffusion_model.double_blocks.") != std::string::npos ||
+            tensor_storage.name.find("model.diffusion_model.single_transformer_blocks.") != std::string::npos) {
             is_flux = true;
         }
+        if (tensor_storage.name.find("model.diffusion_model.net.lq_proj.latent_proj.0.weight") != std::string::npos) {
+            return VERSION_PID;
+        }
+        if (tensor_storage.name.find("embed_image_indicator.weight") != std::string::npos) {
+            return VERSION_IDEOGRAM4;
+        }
         if (tensor_storage.name.find("model.diffusion_model.nerf_final_layer_conv.") != std::string::npos) {
             return VERSION_CHROMA_RADIANCE;
         }
@@ -441,6 +463,10 @@ SDVersion ModelLoader::get_sd_version() {
             tensor_storage_map.find("model.language_model.layers.0.self_attn.q_proj.weight") != tensor_storage_map.end()) {
             return VERSION_HIDREAM_O1;
         }
+        if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.attn.norm_added_q.weight") != std::string::npos &&
+            tensor_storage_map.find("model.diffusion_model.transformer_blocks.0.img_mlp.w1.weight") != tensor_storage_map.end()) {
+            return VERSION_LENS;
+        }
         if (tensor_storage.name.find("model.diffusion_model.transformer_blocks.0.img_mod.1.weight") != std::string::npos) {
             return VERSION_QWEN_IMAGE;
         }
@@ -522,6 +548,9 @@ SDVersion ModelLoader::get_sd_version() {
             tensor_storage.name == "unet.conv_in.weight") {
             input_block_weight = tensor_storage;
         }
+        if (tensor_storage.name == "model.diffusion_model.txt_in.weight" || tensor_storage.name == "model.diffusion_model.context_embedder.weight") {
+            context_ebedding_weight = tensor_storage;
+        }
     }
     if (is_wan) {
         LOG_DEBUG("patch_embedding_channels %d", patch_embedding_channels);
@@ -552,16 +581,20 @@ SDVersion ModelLoader::get_sd_version() {
     }
 
     if (is_flux && !is_flux2) {
-        if (input_block_weight.ne[0] == 384) {
-            return VERSION_FLUX_FILL;
+        if (context_ebedding_weight.ne[0] == 3584) {
+            return VERSION_LONGCAT;
+        } else {
+            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;
         }
-        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;
     }
 
     if (is_flux2) {
@@ -742,8 +775,6 @@ void ModelLoader::process_model_files(bool enable_mmap, bool writable_mmap) {
         return;
     }
 
-    int64_t start_time = ggml_time_ms();
-
     std::vector<TensorStorage> processed_tensor_storages;
     for (const auto& [name, tensor_storage] : tensor_storage_map) {
         if (is_unused_tensor(tensor_storage.name)) {
@@ -794,20 +825,12 @@ void ModelLoader::process_model_files(bool enable_mmap, bool writable_mmap) {
             } else {
                 LOG_WARN("failed to memory-map '%s' (falling back to read())", file_path.c_str());
             }
-        } else if (!is_zip) {
-            LOG_INFO("NOT using mmap for '%s' (mmap disabled by caller)",
-                     file_path.c_str());
         }
 
         file_data.push_back(std::move(fdata));
     }
 
     model_files_processed = true;
-
-    int64_t end_time        = ggml_time_ms();
-    int64_t process_time_ms = end_time - start_time;
-
-    LOG_INFO("model files processing completed in %.2fs", process_time_ms / 1000.f);
 }
 
 std::vector<MmapTensorStore> ModelLoader::mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
@@ -853,8 +876,13 @@ std::vector<MmapTensorStore> ModelLoader::mmap_tensors(std::map<std::string, ggm
             if (dst_tensor == nullptr)
                 continue;
 
-            if (tensor_storage.type != dst_tensor->type)
+            if (tensor_storage.is_f8_e4m3 ||
+                tensor_storage.is_f8_e5m2 ||
+                tensor_storage.is_f64 ||
+                tensor_storage.is_i64 ||
+                tensor_storage.type != dst_tensor->type) {
                 continue;
+            }
 
             size_t tensor_size   = tensor_storage.nbytes();
             size_t tensor_offset = tensor_storage.offset;
@@ -896,7 +924,9 @@ std::vector<MmapTensorStore> ModelLoader::mmap_tensors(std::map<std::string, ggm
     return result;
 }
 
-bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_threads_p, bool enable_mmap) {
+bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
+                               bool enable_mmap,
+                               const std::set<std::string>* target_tensor_names) {
     process_model_files(enable_mmap, false);
 
     std::atomic<int64_t> read_time_ms(0);
@@ -905,14 +935,26 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
     std::atomic<int64_t> convert_time_ms(0);
     std::atomic<uint64_t> bytes_processed(0);
 
-    int num_threads_to_use = n_threads_p > 0 ? n_threads_p : sd_get_num_physical_cores();
-    LOG_DEBUG("using %d threads for model loading", num_threads_to_use);
+    int num_threads_to_use = n_threads_;
 
     int64_t start_time = ggml_time_ms();
 
     size_t total_tensors_to_process = 0;
+    std::vector<size_t> file_tensors_to_process_counts;
+    file_tensors_to_process_counts.reserve(file_data.size());
     for (const auto& fdata : file_data) {
-        total_tensors_to_process += fdata.tensors.size();
+        size_t file_tensors_to_process = 0;
+        if (target_tensor_names == nullptr) {
+            file_tensors_to_process = fdata.tensors.size();
+        } else {
+            for (const TensorStorage& tensor_storage : fdata.tensors) {
+                if (target_tensor_names->find(tensor_storage.name) != target_tensor_names->end()) {
+                    file_tensors_to_process++;
+                }
+            }
+        }
+        file_tensors_to_process_counts.push_back(file_tensors_to_process);
+        total_tensors_to_process += file_tensors_to_process;
     }
 
     bool success                   = true;
@@ -920,17 +962,38 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
     const int64_t t_start          = start_time;
     int last_n_threads             = 1;
 
-    for (auto& fdata : file_data) {
+    for (size_t file_index = 0; file_index < file_data.size(); ++file_index) {
+        auto& fdata                  = file_data[file_index];
         const std::string& file_path = fdata.path;
-        LOG_DEBUG("loading tensors from %s", file_path.c_str());
 
         const std::vector<TensorStorage>& file_tensors = fdata.tensors;
+        std::vector<const TensorStorage*> tensors_to_process;
+        size_t file_tensors_to_process = file_tensors_to_process_counts[file_index];
+        tensors_to_process.reserve(file_tensors_to_process);
+        if (target_tensor_names == nullptr) {
+            for (const TensorStorage& tensor_storage : file_tensors) {
+                tensors_to_process.push_back(&tensor_storage);
+            }
+        } else {
+            for (const TensorStorage& tensor_storage : file_tensors) {
+                if (target_tensor_names->find(tensor_storage.name) != target_tensor_names->end()) {
+                    tensors_to_process.push_back(&tensor_storage);
+                }
+            }
+        }
+        if (tensors_to_process.empty()) {
+            continue;
+        }
+        LOG_DEBUG("loading %zu/%zu tensors from %s",
+                  tensors_to_process.size(),
+                  file_tensors.size(),
+                  file_path.c_str());
 
         bool is_zip = fdata.is_zip;
 
         std::shared_ptr<MmapWrapper> mmapped = fdata.mmapped;
 
-        int n_threads = is_zip ? 1 : std::min(num_threads_to_use, (int)file_tensors.size());
+        int n_threads = is_zip ? 1 : std::min(num_threads_to_use, (int)tensors_to_process.size());
         if (n_threads < 1) {
             n_threads = 1;
         }
@@ -966,11 +1029,11 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                 while (true) {
                     int64_t t0, t1;
                     size_t idx = tensor_idx.fetch_add(1);
-                    if (idx >= file_tensors.size() || failed) {
+                    if (idx >= tensors_to_process.size() || failed) {
                         break;
                     }
 
-                    const TensorStorage& tensor_storage = file_tensors[idx];
+                    const TensorStorage& tensor_storage = *tensors_to_process[idx];
                     ggml_tensor* dst_tensor             = nullptr;
 
                     t0 = ggml_time_ms();
@@ -987,6 +1050,12 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
                         continue;
                     }
 
+                    if (dst_tensor->data == nullptr) {
+                        LOG_ERROR("process tensor data failed: '%s'", tensor_storage.name.c_str());
+                        failed = true;
+                        break;
+                    }
+
                     // skip mmapped tensors
                     if (dst_tensor->buffer != nullptr && dst_tensor->buffer == fdata.mmbuffer.get()) {
                         continue;
@@ -1104,16 +1173,18 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
 
         while (true) {
             size_t current_idx = tensor_idx.load();
-            if (current_idx >= file_tensors.size() || failed) {
+            if (current_idx >= tensors_to_process.size() || failed) {
                 break;
             }
             size_t curr_num       = total_tensors_processed + current_idx;
             float elapsed_seconds = (ggml_time_ms() - t_start) / 1000.0f;
-            pretty_bytes_progress(static_cast<int>(curr_num),
-                                  static_cast<int>(total_tensors_to_process),
-                                  bytes_processed.load(),
-                                  elapsed_seconds);
-            std::this_thread::sleep_for(std::chrono::milliseconds(200));
+            if (total_tensors_to_process > 0) {
+                pretty_bytes_progress(static_cast<int>(curr_num),
+                                      static_cast<int>(total_tensors_to_process),
+                                      bytes_processed.load(),
+                                      elapsed_seconds);
+            }
+            std::this_thread::sleep_for(std::chrono::milliseconds(total_tensors_to_process <= 4 ? 10 : 200));
         }
 
         for (auto& w : workers) {
@@ -1124,12 +1195,14 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
             success = false;
             break;
         }
-        total_tensors_processed += file_tensors.size();
-        pretty_bytes_progress(static_cast<int>(total_tensors_processed),
-                              static_cast<int>(total_tensors_to_process),
-                              bytes_processed.load(),
-                              (ggml_time_ms() - t_start) / 1000.0f);
-        if (total_tensors_processed < total_tensors_to_process) {
+        total_tensors_processed += tensors_to_process.size();
+        if (total_tensors_to_process > 0) {
+            pretty_bytes_progress(static_cast<int>(total_tensors_processed),
+                                  static_cast<int>(total_tensors_to_process),
+                                  bytes_processed.load(),
+                                  (ggml_time_ms() - t_start) / 1000.0f);
+        }
+        if (total_tensors_processed < total_tensors_to_process && total_tensors_to_process > 0) {
             printf("\n");
         }
     }
@@ -1144,9 +1217,77 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
     return success;
 }
 
+bool ModelLoader::load_float_tensor(const std::string& name,
+                                    std::vector<float>& data,
+                                    int n_threads,
+                                    bool use_mmap) {
+    data.clear();
+
+    auto tensor_storage_it = tensor_storage_map.find(name);
+    if (tensor_storage_it == tensor_storage_map.end()) {
+        return false;
+    }
+
+    const TensorStorage& tensor_storage = tensor_storage_it->second;
+    int64_t n_elements                  = tensor_storage.nelements();
+    if (n_elements <= 0) {
+        LOG_ERROR("tensor '%s' has invalid element count: %" PRId64, name.c_str(), n_elements);
+        return false;
+    }
+    if (tensor_storage.n_dims <= 0 || tensor_storage.n_dims > GGML_MAX_DIMS) {
+        LOG_ERROR("tensor '%s' has unsupported dims: %d", name.c_str(), tensor_storage.n_dims);
+        return false;
+    }
+
+    std::vector<float> loaded_data(static_cast<size_t>(n_elements));
+    ggml_init_params params;
+    params.mem_size   = ggml_tensor_overhead();
+    params.mem_buffer = nullptr;
+    params.no_alloc   = true;
+
+    ggml_context* ctx = ggml_init(params);
+    if (ctx == nullptr) {
+        LOG_ERROR("failed to create context for tensor '%s'", name.c_str());
+        return false;
+    }
+
+    ggml_tensor* tensor = ggml_new_tensor(ctx, GGML_TYPE_F32, tensor_storage.n_dims, tensor_storage.ne);
+    ggml_set_name(tensor, name.c_str());
+    tensor->data = loaded_data.data();
+
+    bool loaded           = false;
+    auto on_new_tensor_cb = [&](const TensorStorage& current_tensor_storage, ggml_tensor** dst_tensor) -> bool {
+        *dst_tensor = nullptr;
+        if (current_tensor_storage.name != name) {
+            return true;
+        }
+        if (current_tensor_storage.nelements() != n_elements) {
+            LOG_ERROR("tensor '%s' element count changed during load", name.c_str());
+            return false;
+        }
+        *dst_tensor = tensor;
+        loaded      = true;
+        return true;
+    };
+
+    std::set<std::string> target_tensor_names{name};
+    if (n_threads > 0) {
+        set_n_threads(n_threads);
+    }
+    bool success = load_tensors(on_new_tensor_cb, use_mmap, &target_tensor_names);
+    ggml_free(ctx);
+
+    if (!success || !loaded) {
+        data.clear();
+        return false;
+    }
+
+    data = std::move(loaded_data);
+    return true;
+}
+
 bool ModelLoader::load_tensors(std::map<std::string, ggml_tensor*>& tensors,
                                std::set<std::string> ignore_tensors,
-                               int n_threads,
                                bool enable_mmap) {
     std::set<std::string> tensor_names_in_file;
     std::mutex tensor_names_mutex;
@@ -1190,7 +1331,7 @@ bool ModelLoader::load_tensors(std::map<std::string, ggml_tensor*>& tensors,
         return true;
     };
 
-    bool success = load_tensors(on_new_tensor_cb, n_threads, enable_mmap);
+    bool success = load_tensors(on_new_tensor_cb, enable_mmap);
     if (!success) {
         LOG_ERROR("load tensors from file failed");
         return false;
@@ -1228,6 +1369,8 @@ bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage
             // Pass, do not convert
         } else if (ends_with(name, ".scale")) {
             // Pass, do not convert
+        } else if (ends_with(name, ".weight_scale")) {
+            // Pass, do not convert
         } else if (contains(name, "img_in.") ||
                    contains(name, "txt_in.") ||
                    contains(name, "time_in.") ||

src/model_loader.h

@@ -0,0 +1,93 @@
+#ifndef __MODEL_LOADER_H__
+#define __MODEL_LOADER_H__
+
+#include <cstdint>
+#include <map>
+#include <memory>
+#include <set>
+#include <string>
+#include <vector>
+
+#include "model.h"
+
+TensorTypeRules parse_tensor_type_rules(const std::string& tensor_type_rules);
+
+class MmapWrapper;
+
+struct ModelFileData {
+    std::string path;
+    std::vector<TensorStorage> tensors;
+    std::shared_ptr<MmapWrapper> mmapped;
+    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
+    bool is_zip;
+};
+
+struct MmapTensorStore {
+    std::shared_ptr<MmapWrapper> mmapped;
+    std::shared_ptr<struct ggml_backend_buffer> mmbuffer;
+};
+
+class ModelLoader {
+protected:
+    SDVersion version_ = VERSION_COUNT;
+    std::vector<std::string> file_paths_;
+    std::vector<ModelFileData> file_data;
+    bool model_files_processed = false;
+    String2TensorStorage tensor_storage_map;
+    int n_threads_;
+
+    size_t add_file_path(const std::string& file_path);
+    void add_tensor_storage(const TensorStorage& tensor_storage);
+
+    bool init_from_gguf_file(const std::string& file_path, const std::string& prefix = "");
+    bool init_from_safetensors_file(const std::string& file_path, const std::string& prefix = "");
+    bool init_from_torch_zip_file(const std::string& file_path, const std::string& prefix = "");
+    bool init_from_torch_legacy_file(const std::string& file_path, const std::string& prefix = "");
+    bool init_from_diffusers_file(const std::string& file_path, const std::string& prefix = "");
+
+public:
+    ModelLoader();
+
+    bool init_from_file(const std::string& file_path, const std::string& prefix = "");
+    void convert_tensors_name();
+    bool init_from_file_and_convert_name(const std::string& file_path,
+                                         const std::string& prefix = "",
+                                         SDVersion version         = VERSION_COUNT);
+    SDVersion get_sd_version();
+    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();
+    String2TensorStorage& get_tensor_storage_map() { return tensor_storage_map; }
+    const String2TensorStorage& get_tensor_storage_map() const { return tensor_storage_map; }
+    void set_n_threads(int n_threads);
+    void set_wtype_override(ggml_type wtype, std::string tensor_type_rules = "");
+    void process_model_files(bool enable_mmap = false, bool writable_mmap = true);
+    std::vector<MmapTensorStore> mmap_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                                              std::set<std::string> ignore_tensors = {},
+                                              bool writable                        = true);
+    bool load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
+                      bool use_mmap                                    = false,
+                      const std::set<std::string>* target_tensor_names = nullptr);
+    bool load_tensors(std::map<std::string, ggml_tensor*>& tensors,
+                      std::set<std::string> ignore_tensors = {},
+                      bool use_mmap                        = false);
+    bool load_float_tensor(const std::string& name,
+                           std::vector<float>& data,
+                           int n_threads = 0,
+                           bool use_mmap = false);
+
+    std::vector<std::string> get_tensor_names() const {
+        std::vector<std::string> names;
+        for (const auto& [name, tensor_storage] : tensor_storage_map) {
+            names.push_back(name);
+        }
+        return names;
+    }
+
+    bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
+    int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
+    ~ModelLoader() = default;
+};
+
+#endif  // __MODEL_LOADER_H__