feat: add module backend assignment support (#1500)

Co-authored-by: Stéphane du Hamel <stephduh@live.fr>

.github/pull_request_template.md

@@ -0,0 +1,15 @@
+## Summary
+
+<!-- Describe what changed and why. Keep the PR focused on one clear change. -->
+
+## Related Issue / Discussion
+
+<!-- Link related issues, discussions, or previous PRs if applicable. -->
+
+## Additional Information
+
+<!-- Add verification notes, screenshots, sample output, or other context when applicable. -->
+
+## Checklist
+
+- [ ] I have read and confirmed this PR follows the [contribution guidelines](https://github.com/leejet/stable-diffusion.cpp/blob/master/CONTRIBUTING.md).

.github/workflows/build.yml

@@ -444,12 +444,98 @@ jobs:
           path: |
             sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-${{ matrix.build }}-x64.zip
 
+  windows-latest-rocm:
+    runs-on: windows-2022
+
+    env:
+      ROCM_VERSION: "7.12.0"
+      GPU_TARGETS: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201"
+
+    steps:
+      - uses: actions/checkout@v3
+        with:
+          submodules: recursive
+
+      - name: Cache ROCm Installation
+        id: cache-rocm
+        uses: actions/cache@v4
+        with:
+          path: C:\TheRock\build
+          key: rocm-${{ env.ROCM_VERSION }}-gfx1151-${{ runner.os }}
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: windows-latest-rocm-${{ env.ROCM_VERSION }}-x64
+          evict-old-files: 1d
+
+      - name: Install ROCm
+        if: steps.cache-rocm.outputs.cache-hit != 'true'
+        run: |
+          $ErrorActionPreference = "Stop"
+          write-host "Downloading AMD ROCm ${{ env.ROCM_VERSION }} tarball"
+          Invoke-WebRequest -Uri "https://repo.amd.com/rocm/tarball/therock-dist-windows-gfx1151-${{ env.ROCM_VERSION }}.tar.gz" -OutFile "${env:RUNNER_TEMP}\rocm.tar.gz"
+          write-host "Extracting ROCm tarball"
+          mkdir C:\TheRock\build -Force
+          tar -xzf "${env:RUNNER_TEMP}\rocm.tar.gz" -C C:\TheRock\build --strip-components=1
+          write-host "Completed ROCm extraction"
+
+      - name: Setup ROCm Environment
+        run: |
+          $rocmPath = "C:\TheRock\build"
+          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
+          echo "HIP_DEVICE_LIB_PATH=$rocmPath\lib\llvm\amdgcn\bitcode" >> $env:GITHUB_ENV
+          echo "HIP_PLATFORM=amd" >> $env:GITHUB_ENV
+          echo "LLVM_PATH=$rocmPath\lib\llvm" >> $env:GITHUB_ENV
+          echo "$rocmPath\bin" >> $env:GITHUB_PATH
+          echo "$rocmPath\lib\llvm\bin" >> $env:GITHUB_PATH
+
+      - name: Build
+        run: |
+          mkdir build
+          cd build
+          cmake .. `
+            -G "Unix Makefiles" `
+            -DCMAKE_PREFIX_PATH="${env:HIP_PATH}" `
+            -DSD_HIPBLAS=ON `
+            -DSD_BUILD_SHARED_LIBS=ON `
+            -DGGML_NATIVE=OFF `
+            -DCMAKE_C_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DCMAKE_CXX_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang++.exe" `
+            -DCMAKE_HIP_COMPILER="${env:HIP_PATH}\lib\llvm\bin\clang.exe" `
+            -DHIP_PATH="${env:HIP_PATH}" `
+            -DCMAKE_BUILD_TYPE=Release `
+            -DGPU_TARGETS="${{ env.GPU_TARGETS }}"
+          cmake --build . --config Release --parallel ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Get commit hash
+        id: commit
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: pr-mpt/actions-commit-hash@v2
+
+      - name: Pack artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        run: |
+          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip .\build\bin\*
+
+      - name: Upload artifacts
+        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
+        uses: actions/upload-artifact@v4
+        with:
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
+          path: |
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
+
   windows-latest-cmake-hip:
     runs-on: windows-2022
 
     env:
-      HIPSDK_INSTALLER_VERSION: "25.Q3"
-      GPU_TARGETS: "gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
+      HIPSDK_INSTALLER_VERSION: "26.Q1"
+      ROCM_VERSION: "7.1.1"
+      GPU_TARGETS: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 
     steps:
       - uses: actions/checkout@v3
@@ -484,7 +570,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "Downloading AMD HIP SDK Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-${{ env.HIPSDK_INSTALLER_VERSION }}-Win11-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP SDK"
           $proc = Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -PassThru
           $completed = $proc.WaitForExit(600000)
@@ -537,32 +623,38 @@ jobs:
         run: |
           md "build\bin\rocblas\library\"
           md "build\bin\hipblaslt\library"
-          cp "${env:HIP_PATH}\bin\hipblas.dll" "build\bin\"
-          cp "${env:HIP_PATH}\bin\hipblaslt.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblas.dll" "build\bin\"
+          cp "${env:HIP_PATH}\bin\libhipblaslt.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas.dll" "build\bin\"
           cp "${env:HIP_PATH}\bin\rocblas\library\*" "build\bin\rocblas\library\"
           cp "${env:HIP_PATH}\bin\hipblaslt\library\*" "build\bin\hipblaslt\library\"
-          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip .\build\bin\*
+          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip .\build\bin\*
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v4
         with:
-          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
           path: |
-            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-x64.zip
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip
 
   ubuntu-latest-rocm:
-    runs-on: ubuntu-latest
-    container: rocm/dev-ubuntu-24.04:7.2
+    runs-on: ubuntu-24.04
 
     env:
-      ROCM_VERSION: "7.2"
       UBUNTU_VERSION: "24.04"
-      GPU_TARGETS: "gfx1151;gfx1150;gfx1100;gfx1101;gfx1102;gfx1200;gfx1201"
+
+    strategy:
+      matrix:
+        include:
+          - ROCM_VERSION: "7.2.1"
+            gpu_targets: "gfx908;gfx90a;gfx942;gfx1030;gfx1031;gfx1032;gfx1100;gfx1101;gfx1102;gfx1151;gfx1150;gfx1200;gfx1201"
+            build: 'x64'
+          - ROCM_VERSION: "7.12.0"
+            gpu_targets: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1200;gfx1201"
+            build: x64
 
     steps:
-      - run: apt-get update && apt-get install -y git
       - name: Clone
         id: checkout
         uses: actions/checkout@v6
@@ -579,6 +671,38 @@ jobs:
         with:
           version: 10.15.1
 
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.16
+        with:
+          key: ubuntu-rocm-cmake-${{ matrix.ROCM_VERSION }}-${{ matrix.build }}
+          evict-old-files: 1d
+
+      - name: Dependencies
+        id: depends
+        run: |
+          sudo apt install -y build-essential cmake wget zip ninja-build
+
+      - name: Setup Legacy ROCm
+        if: matrix.ROCM_VERSION == '7.2.1'
+        id: legacy_env
+        run: |
+          sudo mkdir --parents --mode=0755 /etc/apt/keyrings
+          wget https://repo.radeon.com/rocm/rocm.gpg.key -O - | \
+            gpg --dearmor | sudo tee /etc/apt/keyrings/rocm.gpg > /dev/null
+
+          sudo tee /etc/apt/sources.list.d/rocm.list << EOF
+          deb [arch=amd64 signed-by=/etc/apt/keyrings/rocm.gpg] https://repo.radeon.com/rocm/apt/${{ matrix.ROCM_VERSION }} noble main
+          EOF
+
+          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
+          Package: *
+          Pin: release o=repo.radeon.com
+          Pin-Priority: 600
+          EOF
+
+          sudo apt update
+          sudo apt-get install -y libssl-dev rocm-hip-sdk
+
       - name: Free disk space
         run: |
           # Remove preinstalled SDKs and caches not needed for this job
@@ -592,51 +716,17 @@ jobs:
           sudo rm -rf /var/lib/apt/lists/* || true
           sudo apt clean
 
-      - name: Dependencies
-        id: depends
+      - name: Setup TheRock
+        if: matrix.ROCM_VERSION != '7.2.1'
+        id: therock_env
         run: |
-          sudo apt-get update
-          sudo apt install -y \
-            cmake \
-            hip-dev \
-            hipblas-dev \
-            ninja-build \
-            rocm-dev \
-            zip
-          # Clean apt caches to recover disk space
-          sudo apt clean
-          sudo rm -rf /var/lib/apt/lists/* || true
-
-      - name: Setup ROCm Environment
-        run: |
-          # Add ROCm to PATH for current session
-          echo "/opt/rocm/bin" >> $GITHUB_PATH
-
-          # Build regex pattern from ${{ env.GPU_TARGETS }} (match target as substring)
-          TARGET_REGEX="($(printf '%s' "${{ env.GPU_TARGETS }}" | sed 's/;/|/g'))"
-
-          # Remove library files for architectures we're not building for to save disk space
-          echo "Cleaning up unneeded architecture files..."
-          cd /opt/rocm/lib/rocblas/library
-          # Keep only our target architectures
-          for file in *; do
-            if printf '%s' "$file" | grep -q 'gfx'; then
-              if ! printf '%s' "$file" | grep -Eq "$TARGET_REGEX"; then
-                echo "Removing $file" &&
-                sudo rm -f "$file";
-              fi
-            fi
-          done
-
-          cd /opt/rocm/lib/hipblaslt/library
-          for file in *; do
-            if printf '%s' "$file" | grep -q 'gfx'; then
-              if ! printf '%s' "$file" | grep -Eq "$TARGET_REGEX"; then
-                echo "Removing $file" &&
-                sudo rm -f "$file";
-              fi
-            fi
-          done
+          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ matrix.ROCM_VERSION }}.tar.gz
+          mkdir install
+          tar -xf *.tar.gz -C install
+          export ROCM_PATH=$(pwd)/install
+          echo ROCM_PATH=$ROCM_PATH >> $GITHUB_ENV
+          echo PATH=$PATH:$ROCM_PATH/bin >> $GITHUB_ENV
+          echo LD_LIBRARY_PATH=$ROCM_PATH/lib:$ROCM_PATH/llvm/lib:$ROCM_PATH/lib/rocprofiler-systems >> $GITHUB_ENV
 
       - name: Build
         id: cmake_build
@@ -644,12 +734,12 @@ jobs:
           mkdir build
           cd build
           cmake .. -G Ninja \
-            -DCMAKE_CXX_COMPILER=amdclang++ \
-            -DCMAKE_C_COMPILER=amdclang \
+            -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" \
+            -DCMAKE_HIP_FLAGS="-mllvm --amdgpu-unroll-threshold-local=600" \
             -DCMAKE_BUILD_TYPE=Release \
             -DSD_HIPBLAS=ON \
-            -DGPU_TARGETS="${{ env.GPU_TARGETS }}" \
-            -DAMDGPU_TARGETS="${{ env.GPU_TARGETS }}" \
+            -DHIP_PLATFORM=amd \
+            -DGPU_TARGETS="${{ matrix.gpu_targets }}" \
             -DCMAKE_BUILD_WITH_INSTALL_RPATH=ON \
             -DCMAKE_POSITION_INDEPENDENT_CODE=ON \
             -DSD_BUILD_SHARED_LIBS=ON
@@ -668,16 +758,6 @@ jobs:
           cp ggml/LICENSE ./build/bin/ggml.txt
           cp LICENSE ./build/bin/stable-diffusion.cpp.txt
 
-          # Move ROCm runtime libraries (to avoid double space consumption)
-          sudo mv /opt/rocm/lib/librocsparse.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhsa-runtime64.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libamdhip64.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhipblas.so* ./build/bin/
-          sudo mv /opt/rocm/lib/libhipblaslt.so* ./build/bin/
-          sudo mv /opt/rocm/lib/librocblas.so* ./build/bin/
-          sudo mv /opt/rocm/lib/rocblas/ ./build/bin/
-          sudo mv /opt/rocm/lib/hipblaslt/ ./build/bin/
-
       - name: Fetch system info
         id: system-info
         run: |
@@ -692,15 +772,15 @@ jobs:
         run: |
           cp ggml/LICENSE ./build/bin/ggml.txt
           cp LICENSE ./build/bin/stable-diffusion.cpp.txt
-          zip -y -r sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip ./build/bin
+          zip -y -r sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip ./build/bin
 
       - name: Upload artifacts
         if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
         uses: actions/upload-artifact@v4
         with:
-          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip
+          name: sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip
           path: |
-            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm.zip
+            sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-${{ steps.system-info.outputs.OS_TYPE }}-Ubuntu-${{ env.UBUNTU_VERSION }}-${{ steps.system-info.outputs.CPU_ARCH }}-rocm-${{ matrix.ROCM_VERSION }}.zip
 
   release:
     if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
@@ -715,6 +795,7 @@ jobs:
       - macOS-latest-cmake
       - windows-latest-cmake
       - windows-latest-cmake-hip
+      - windows-latest-rocm
 
     steps:
       - name: Clone

CMakeLists.txt

@@ -69,40 +69,40 @@ option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" O
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
+set(CMAKE_C_STANDARD 11)
+set(CMAKE_C_STANDARD_REQUIRED true)
+
+set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD_REQUIRED true)
+
 if(SD_CUDA)
     message("-- Use CUDA as backend stable-diffusion")
     set(GGML_CUDA ON)
-    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_METAL)
     message("-- Use Metal as backend stable-diffusion")
     set(GGML_METAL ON)
-    add_definitions(-DSD_USE_METAL)
 endif()
 
 if (SD_VULKAN)
     message("-- Use Vulkan as backend stable-diffusion")
     set(GGML_VULKAN ON)
-    add_definitions(-DSD_USE_VULKAN)
 endif ()
 
 if (SD_OPENCL)
     message("-- Use OpenCL as backend stable-diffusion")
     set(GGML_OPENCL ON)
-    add_definitions(-DSD_USE_OPENCL)
 endif ()
 
 if (SD_HIPBLAS)
     message("-- Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIP ON)
-    add_definitions(-DSD_USE_CUDA)
 endif ()
 
 if(SD_MUSA)
     message("-- Use MUSA as backend stable-diffusion")
     set(GGML_MUSA ON)
-    add_definitions(-DSD_USE_CUDA)
 endif()
 
 if(SD_WEBP)
@@ -112,19 +112,28 @@ if(SD_WEBP)
           "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBP=ON")
     endif()
     if(SD_USE_SYSTEM_WEBP)
-        find_package(WebP REQUIRED)
-        add_library(webp ALIAS WebP::webp)
-        # libwebp CMake target naming is not consistent across versions/distros.
-        # Some export WebP::libwebpmux, others export WebP::webpmux.
-        if(TARGET WebP::libwebpmux)
-            add_library(libwebpmux ALIAS WebP::libwebpmux)
-        elseif(TARGET WebP::webpmux)
-            add_library(libwebpmux ALIAS WebP::webpmux)
+        find_package(WebP)
+        if(WebP_FOUND)
+            add_library(webp ALIAS WebP::webp)
+            # libwebp CMake target naming is not consistent across versions/distros.
+            # Some export WebP::libwebpmux, others export WebP::webpmux.
+            if(TARGET WebP::libwebpmux)
+                add_library(libwebpmux ALIAS WebP::libwebpmux)
+            elseif(TARGET WebP::webpmux)
+                add_library(libwebpmux ALIAS WebP::webpmux)
+            else()
+                message(FATAL_ERROR
+                    "Could not find a compatible webpmux target in system WebP package. "
+                    "Expected WebP::libwebpmux or WebP::webpmux."
+                )
+            endif()
         else()
-            message(FATAL_ERROR
-                "Could not find a compatible webpmux target in system WebP package. "
-                "Expected WebP::libwebpmux or WebP::webpmux."
-            )
+            find_package(PkgConfig REQUIRED)
+            pkg_check_modules(WebP REQUIRED IMPORTED_TARGET GLOBAL libwebp)
+            pkg_check_modules(WebPMux REQUIRED IMPORTED_TARGET GLOBAL libwebpmux)
+            link_libraries(PkgConfig::WebP)
+            link_libraries(PkgConfig::WebPMux)
+            add_library(libwebpmux ALIAS PkgConfig::WebPMux)
         endif()
     endif()
 endif()
@@ -139,27 +148,37 @@ if(SD_WEBM)
           "Or link against system library:\n  cmake (...) -DSD_USE_SYSTEM_WEBM=ON")
     endif()
     if(SD_USE_SYSTEM_WEBM)
-        find_path(WEBM_INCLUDE_DIR
-            NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
-            PATH_SUFFIXES webm
-            REQUIRED)
-        find_library(WEBM_LIBRARY
-            NAMES webm libwebm
-            REQUIRED)
+        find_package(PkgConfig)
+        if(PkgConfig_FOUND)
+            pkg_check_modules(WebM REQUIRED IMPORTED_TARGET GLOBAL libwebm)
+        endif()
+        if(PkgConfig_FOUND AND WebM_FOUND)
+            link_libraries(PkgConfig::WebM)
+        else()
+            find_path(WEBM_INCLUDE_DIR
+                NAMES mkvmuxer/mkvmuxer.h mkvparser/mkvparser.h common/webmids.h
+                PATH_SUFFIXES webm
+                REQUIRED)
+            find_library(WEBM_LIBRARY
+                NAMES webm libwebm
+                REQUIRED)
 
-        add_library(webm UNKNOWN IMPORTED)
-        set_target_properties(webm PROPERTIES
-            IMPORTED_LOCATION "${WEBM_LIBRARY}"
-            INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+            add_library(webm UNKNOWN IMPORTED)
+            set_target_properties(webm PROPERTIES
+                IMPORTED_LOCATION "${WEBM_LIBRARY}"
+                INTERFACE_INCLUDE_DIRECTORIES "${WEBM_INCLUDE_DIR}")
+        endif()
     endif()
 endif()
 
 set(SD_LIB stable-diffusion)
 
-file(GLOB SD_LIB_SOURCES
+file(GLOB SD_LIB_SOURCES CONFIGURE_DEPENDS
     "src/*.h"
     "src/*.cpp"
     "src/*.hpp"
+    "src/model_io/*.h"
+    "src/model_io/*.cpp"
     "src/tokenizers/*.h"
     "src/tokenizers/*.cpp"
     "src/tokenizers/vocab/*.h"
@@ -220,7 +239,6 @@ if(SD_SYCL)
     message("-- Use SYCL as backend stable-diffusion")
     set(GGML_SYCL ON)
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -Wno-narrowing -fsycl")
-    add_definitions(-DSD_USE_SYCL)
     # disable fast-math on host, see:
     # https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-10/fp-model-fp.html
     if (WIN32)

CONTRIBUTING.md

@@ -0,0 +1,65 @@
+# Contributing
+
+This document collects general contribution conventions for this repository.
+
+## Before You Start
+
+Before opening a PR, please search existing PRs to avoid duplicating ongoing work.
+
+For large-scale refactors or changes with broad impact, please open an issue first to discuss the approach before submitting a PR.
+
+If you want to update a third-party dependency, please open an issue first instead of submitting a direct PR. See [Dependency Updates](#dependency-updates) for details.
+
+## Pull Requests
+
+Keep each PR focused on one clear change. Large or overly complex PRs are harder to review and may not be merged.
+
+Follow Conventional Commit-style subjects seen in history: `feat:`, `fix:`, `refactor:`, `ci:`, `docs:`, `chore:`. Keep subjects imperative and scoped.
+
+PRs should include:
+
+- What changed and why (short problem/solution summary).
+- Verification evidence when applicable (commands and key outputs).
+- Linked issue/PR context when applicable.
+- Screenshots or sample outputs for UI/visual behavior changes.
+
+## Code Style
+
+Format code according to the repository style before submitting changes.
+
+Formatting follows `.clang-format` (Chromium base, 4-space indent, no tabs). Run `format-code.sh` before opening a PR. Keep C++ standard at C++17-compatible patterns used in this repo.
+
+Naming conventions:
+
+- Use `PascalCase` for class/struct/type names.
+- In `PascalCase` names, preserve common abbreviations in uppercase, for example `SD`, `API`, `HTTP`, `JSON`, `RGB`, `VAE`, `TAE`, `LoRA`, and `WebP`.
+- Use `snake_case` for functions, methods, variables, and file names unless an existing API requires a different style.
+- Use a trailing underscore for private data member names, for example `hidden_size_` or `tokenizer_`.
+- Use `.h` for C and C++ header files. Do not introduce new `.hpp` headers.
+- Use macro-based header include guards instead of `#pragma once`.
+- Format header include guards as `__SD_{PATH}__`, where `{PATH}` is the header path in uppercase snake case without the file extension. For example, `src/sample.h` should use `__SD_SAMPLE_H__`.
+- Do not introduce anonymous namespaces in new or modified code; prefer `static` file-local functions/variables or an explicit named namespace when scoping is needed.
+- In `class`/`struct` definitions, place data members before member functions unless an existing type already clearly follows a different pattern.
+- Keep `test_*.cpp` / `test_*.py` naming for tests.
+
+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.
+
+## AI-Assisted Contributions
+
+AI tools may be used to assist development, but contributors are responsible for the quality and correctness of the submitted code.
+
+If any part of a contribution was generated with AI assistance, the contributor must perform a thorough human review before submitting the PR and understand every changed line.
+
+Do not list AI tools as co-authors. The human contributor is the sole responsible author of the submitted code.
+
+Please do not submit AI-generated code that you do not understand, and do not include meaningless experiments, temporary test code, or unrelated generated output in a PR.
+
+## Dependency Updates
+
+Do not submit PRs that update `ggml`. `ggml` updates are performed only after local validation by the maintainer.
+
+Other third-party dependencies are not updated unless necessary. If you want to update a dependency, please open an issue first instead of submitting a direct PR.
+
+## Security & Configuration
+
+Do not commit model weights, secrets, or local absolute paths. Keep large binaries out of git unless intentionally tracked release assets.

README.md

@@ -58,6 +58,7 @@ API and command-line option may change frequently.***
     - [Ovis-Image](./docs/ovis_image.md)
     - [Anima](./docs/anima.md)
     - [ERNIE-Image](./docs/ernie_image.md)
+    - [HiDream-O1-Image](./docs/hidream_o1_image.md)
   - Image Edit Models
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
@@ -77,9 +78,10 @@ API and command-line option may change frequently.***
   - OpenCL
   - SYCL
 - Supported weight formats
-  - Pytorch checkpoint (`.ckpt` or `.pth`)
+  - Pytorch checkpoint (`.ckpt` or `.pth` or `.pt`)
   - Safetensors (`.safetensors`)
   - GGUF (`.gguf`)
+- Convert mode supports converting model weights to `.gguf` or `.safetensors`
 - Supported platforms
     - Linux
     - Mac OS
@@ -131,9 +133,11 @@ API and command-line option may change frequently.***
 ## Performance
 
 If you want to improve performance or reduce VRAM/RAM usage, please refer to [performance guide](./docs/performance.md).
+For runtime and parameter backend placement, see the [backend selection guide](./docs/backend.md).
 
 ## More Guides
 
+- [Backend selection](./docs/backend.md)
 - [SD1.x/SD2.x/SDXL](./docs/sd.md)
 - [SD3/SD3.5](./docs/sd3.md)
 - [FLUX.1-dev/FLUX.1-schnell](./docs/flux.md)
@@ -147,6 +151,7 @@ If you want to improve performance or reduce VRAM/RAM usage, please refer to [pe
 - [Ovis-Image](./docs/ovis_image.md)
 - [Anima](./docs/anima.md)
 - [ERNIE-Image](./docs/ernie_image.md)
+- [HiDream-O1-Image](./docs/hidream_o1_image.md)
 - [LoRA](./docs/lora.md)
 - [LCM/LCM-LoRA](./docs/lcm.md)
 - [Using PhotoMaker to personalize image generation](./docs/photo_maker.md)
@@ -162,6 +167,7 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra
 
 * Golang (non-cgo): [seasonjs/stable-diffusion](https://github.com/seasonjs/stable-diffusion)
 * Golang (cgo): [Binozo/GoStableDiffusion](https://github.com/Binozo/GoStableDiffusion)
+* Golang (non-cgo): [l8bloom/gosd](https://github.com/l8bloom/gosd)
 * C#: [DarthAffe/StableDiffusion.NET](https://github.com/DarthAffe/StableDiffusion.NET)
 * Python: [william-murray1204/stable-diffusion-cpp-python](https://github.com/william-murray1204/stable-diffusion-cpp-python)
 * Rust: [newfla/diffusion-rs](https://github.com/newfla/diffusion-rs)

docs/backend.md

@@ -0,0 +1,122 @@
+# Backend selection
+
+`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.
+
+If `--params-backend` is not set, parameters use the same backend as their module runtime backend.
+
+## Syntax
+
+A backend assignment can be a single backend name:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cpu
+```
+
+This applies to every module that does not have a more specific assignment.
+
+Assignments can also target individual modules:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend te=cpu,vae=cuda0,diffusion=vulkan0
+```
+
+The same syntax is used for parameter placement:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend cuda0 --params-backend te=cpu,vae=cpu
+```
+
+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:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend all=cuda0,te=cpu
+```
+
+## Modules
+
+| Module | Purpose | Accepted names |
+| --- | --- | --- |
+| `diffusion` | UNet, DiT, MMDiT, Flux, Wan, Qwen Image, and other diffusion models | `diffusion`, `model`, `unet`, `dit` |
+| `te` | Text encoders and conditioners | `te`, `clip`, `text`, `textencoder`, `textencoders`, `conditioner`, `cond`, `llm`, `t5`, `t5xxl` |
+| `clip_vision` | CLIP vision encoder | `clip_vision`, `clipvision`, `clip-vision`, `vision` |
+| `vae` | VAE and TAE | `vae`, `firststage`, `autoencoder`, `tae` |
+| `controlnet` | ControlNet | `controlnet`, `control` |
+| `photomaker` | PhotoMaker ID encoder and PhotoMaker LoRA | `photomaker`, `photomakerid`, `pmid`, `photo` |
+| `upscaler` | ESRGAN upscaler | `upscaler`, `esrgan`, `hires` |
+
+`te` is the preferred module name for text encoders. `clip` is kept as an accepted alias because many existing commands and model names use CLIP terminology.
+
+## Backend names
+
+Backend names are resolved against the GGML backend device list. Matching is case-insensitive and accepts exact names or unique prefixes, so common values include names such as:
+
+- `cpu`
+- `cuda0`
+- `vulkan0`
+- `metal`
+
+The special values `auto`, `default`, and an empty backend name select the default backend. The default preference is GPU, then integrated GPU, then CPU.
+
+The special value `gpu` selects the first GPU backend, falling back to the first integrated GPU backend.
+
+## Runtime backend vs. parameter backend
+
+The runtime backend controls where graph execution runs. The parameter backend controls where model weights are allocated.
+
+For example:
+
+```shell
+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.
+
+Per-module assignments can be mixed:
+
+```shell
+sd-cli -m model.safetensors -p "a cat" --backend diffusion=cuda0,te=cpu,vae=cpu --params-backend diffusion=cuda0,te=cpu,vae=cpu
+```
+
+This keeps text encoding and VAE execution on CPU while the diffusion model runs on GPU.
+
+## Backend sharing and lifetime
+
+Backends are managed by `SDBackendManager`.
+
+Within one manager, backend instances are cached by resolved backend device name. If multiple modules request the same backend, they share the same `ggml_backend_t`.
+
+For example:
+
+```shell
+--backend te=cpu,vae=cpu
+```
+
+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.
+
+`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
+
+The older CPU placement flags are still supported:
+
+- `--clip-on-cpu`
+- `--vae-on-cpu`
+- `--control-net-cpu`
+- `--offload-to-cpu`
+
+`--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:
+
+```shell
+--params-backend cpu
+```
+
+Explicit `--backend` and `--params-backend` assignments are preferred for new commands.

docs/caching.md

@@ -131,8 +131,6 @@ sd-cli -m model.safetensors -p "a cat" --cache-mode spectrum
 | `warmup` | Steps to always compute before caching starts | 4 |
 | `stop` | Stop caching at this fraction of total steps | 0.9 |
 
-```
-
 ### Performance Tips
 
 - Start with default thresholds and adjust based on output quality

docs/hidream_o1_image.md

@@ -0,0 +1,20 @@
+# How to Use
+
+## Download weights
+
+- Download HiDream-O1-Image-Dev
+    - safetensors: https://huggingface.co/Comfy-Org/HiDream-O1-Image/tree/main/checkpoints
+- Download HiDream-O1-Image
+    - safetensors: https://huggingface.co/Comfy-Org/HiDream-O1-Image/tree/main/checkpoints
+
+## Examples
+
+### HiDream-O1-Image-Dev
+
+```
+.\bin\Release\sd-cli.exe -m  ..\..\ComfyUI\models\diffusion_models\hidream_o1_image_dev_bf16.safetensors -p "a lovely cat holding a sign says 
+'hidream o1 cpp'" --cfg-scale 1.0  -v -H 1024 -W 1024
+```
+
+<img width="256" alt="HiDream-O1-Image-Dev example" src="../assets/hidream-o1/dev_example.png" />
+

docs/hipBLAS_on_Windows.md

@@ -26,12 +26,12 @@ Fortunately, `AMD` provides complete help documentation, you can use the help do
 
 Then we must set `ROCM` as environment variables before running cmake.
 
-Usually if you install according to the official tutorial and do not modify the ROCM path, then there is a high probability that it is here `C:\Program Files\AMD\ROCm\5.5\bin`
+Usually if you install according to the official tutorial and do not modify the ROCM path, then there is a high probability that it is here `C:\Program Files\AMD\ROCm\7.1.1\bin`
 
 This is what I use to set the clang:
 ```Commandline
-set CC=C:\Program Files\AMD\ROCm\5.5\bin\clang.exe
-set CXX=C:\Program Files\AMD\ROCm\5.5\bin\clang++.exe
+set CC=C:\Program Files\AMD\ROCm\7.1.1\bin\clang.exe
+set CXX=C:\Program Files\AMD\ROCm\7.1.1\bin\clang++.exe
 ```
 
 ## Ninja
@@ -46,7 +46,7 @@ set ninja=C:\Program Files\ninja\ninja.exe
 ## Building stable-diffusion.cpp
 
 The thing different from the regular CPU build is `-DSD_HIPBLAS=ON` ,
-`-G "Ninja"`, `-DCMAKE_C_COMPILER=clang`, `-DCMAKE_CXX_COMPILER=clang++`, `-DAMDGPU_TARGETS=gfx1100`
+`-G "Ninja"`, `-DCMAKE_C_COMPILER=clang`, `-DCMAKE_CXX_COMPILER=clang++`, `-DAMDGPU_TARGETS=gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032`
 
 >**Notice**: check the `clang` and `clang++` information:
 ```Commandline
@@ -59,26 +59,29 @@ If you see like this, we can continue:
 clang version 17.0.0 (git@github.amd.com:Compute-Mirrors/llvm-project e3201662d21c48894f2156d302276eb1cf47c7be)
 Target: x86_64-pc-windows-msvc
 Thread model: posix
-InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
+InstalledDir: C:\Program Files\AMD\ROCm\7.1.1\bin
 ```
 
 ```
 clang version 17.0.0 (git@github.amd.com:Compute-Mirrors/llvm-project e3201662d21c48894f2156d302276eb1cf47c7be)
 Target: x86_64-pc-windows-msvc
 Thread model: posix
-InstalledDir: C:\Program Files\AMD\ROCm\5.5\bin
+InstalledDir: C:\Program Files\AMD\ROCm\7.1.1\bin
 ```
 
->**Notice** that the `gfx1100` is the GPU architecture of my GPU, you can change it to your GPU architecture. Click here to see your architecture [LLVM Target](https://rocm.docs.amd.com/en/latest/release/windows_support.html#windows-supported-gpus)
+>**Notice** that the GPU targets are now compatible with multiple GPU architectures (ROCm 7.1.1 targets). You can change them to match your GPU architecture. Click here to see your architecture [LLVM Target](https://rocm.docs.amd.com/en/latest/release/windows_support.html#windows-supported-gpus)
 
-My GPU is AMD Radeon™ RX 7900 XTX Graphics, so I set it to `gfx1100`.
+Examples:
+- AMD Radeon™ RX 7900 XTX Graphics: `gfx1100`
+- AMD Radeon™ RX 7900 XT Graphics: `gfx1101`
+- AMD Radeon™ RX 7900 GRE Graphics: `gfx1102`
 
 option:
 
 ```commandline
 mkdir build
 cd build
-cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS=gfx1100
+cmake .. -G "Ninja" -DCMAKE_C_COMPILER=clang -DCMAKE_CXX_COMPILER=clang++ -DSD_HIPBLAS=ON -DCMAKE_BUILD_TYPE=Release -DAMDGPU_TARGETS="gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
 cmake --build . --config Release
 ```
 

examples/cli/README.md

@@ -4,26 +4,29 @@
 usage: ./bin/sd-cli  [options]
 
 CLI Options:
-  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default:
-                              ./output.png) (eg. output_%03d.png). For video generation, single-file outputs support .avi, .webm, and animated .webp
-  --preview-path <string>     path to write preview image to (default: ./preview.png). Multi-frame previews support .avi, .webm, and animated .webp
-  --preview-interval <int>    interval in denoising steps between consecutive updates of the image preview file (default is 1, meaning updating at
-                              every step)
-  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
-  --image <string>            path to the image to inspect (for metadata mode)
-  --metadata-format <string>  metadata output format, one of [text, json] (default: text)
-  --canny                     apply canny preprocessor (edge detection)
-  --convert-name              convert tensor name (for convert mode)
-  -v, --verbose               print extra info
-  --color                     colors the logging tags according to level
-  --taesd-preview-only        prevents usage of taesd for decoding the final image. (for use with --preview tae)
-  --preview-noisy             enables previewing noisy inputs of the models rather than the denoised outputs
-  --metadata-raw              include raw hex previews for unparsed metadata payloads
-  --metadata-brief            truncate long metadata text values in text output
-  --metadata-all              include structural/container entries such as IHDR, IDAT, and non-metadata JPEG segments
-  -M, --mode                  run mode, one of [img_gen, vid_gen, upscale, convert, metadata], default: img_gen
-  --preview                   preview method. must be one of the following [none, proj, tae, vae] (default is none)
-  -h, --help                  show this help message and exit
+  -o, --output <string>         path to write result image to. you can use printf-style %d format specifiers for image
+                                sequences (default: ./output.png) (eg. output_%03d.png). Single-file video outputs
+                                support .avi, .webm, and animated .webp
+  --image <string>              path to the image to inspect (for metadata mode)
+  --metadata-format <string>    metadata output format, one of [text, json] (default: text)
+  --preview-path <string>       path to write preview image to (default: ./preview.png). Multi-frame previews support
+                                .avi, .webm, and animated .webp
+  --preview-interval <int>      interval in denoising steps between consecutive updates of the image preview file
+                                (default is 1, meaning updating at every step)
+  --output-begin-idx <int>      starting index for output image sequence, must be non-negative (default 0 if specified
+                                %d in output path, 1 otherwise)
+  --canny                       apply canny preprocessor (edge detection)
+  --convert-name                convert tensor name (for convert mode)
+  -v, --verbose                 print extra info
+  --color                       colors the logging tags according to level
+  --taesd-preview-only          prevents usage of taesd for decoding the final image. (for use with --preview tae)
+  --preview-noisy               enables previewing noisy inputs of the models rather than the denoised outputs
+  --metadata-raw                include raw hex previews for unparsed metadata payloads
+  --metadata-brief              truncate long metadata text values in text output
+  --metadata-all                include structural/container entries such as IHDR, IDAT, and non-metadata JPEG segments
+  -M, --mode                    run mode, one of [img_gen, vid_gen, upscale, convert, metadata], default: img_gen
+  --preview                     preview method. must be one of the following [none, proj, tae, vae] (default is none)
+  -h, --help                    show this help message and exit
 
 Context Options:
   -m, --model <string>                     path to full model
@@ -31,7 +34,8 @@ Context Options:
   --clip_g <string>                        path to the clip-g text encoder
   --clip_vision <string>                   path to the clip-vision encoder
   --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
+                                           mistral-small3.2 for flux2, ...)
   --llm_vision <string>                    path to the llm vit
   --qwen2vl <string>                       alias of --llm. Deprecated.
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
@@ -43,16 +47,18 @@ Context Options:
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory
+  --hires-upscalers-dir <string>           highres fix upscaler model directory
   --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --photo-maker <string>                   path to PHOTOMAKER model
   --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
-                                           CPU physical cores
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
+                                           then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --vae-tiling                             process vae in tiles to reduce memory usage
+  --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
   --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
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
+                                           when needed
   --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
@@ -67,20 +73,19 @@ Context Options:
   --chroma-disable-dit-mask                disable dit mask for chroma
   --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
-                                           type of the weight file
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
+                                           q4_K). If not specified, the default is the type of the weight file
   --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
   --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
-                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
-                                           immediately will be used.The immediately mode may have precision and
-                                           compatibility issues with quantized parameters, but it usually offers faster inference
-                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
-                                           other hand, is exactly the opposite.
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
-                                           (overrides --vae-tile-size)
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
+                                           flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
+                                           auto. In auto mode, if the model weights contain any quantized parameters,
+                                           the at_runtime mode will be used; otherwise, immediately will be used.The
+                                           immediately mode may have precision and compatibility issues with quantized
+                                           parameters, but it usually offers faster inference speed and, in some cases,
+                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
+                                           opposite.
 
 Generation Options:
   -p, --prompt <string>                    the prompt to render
@@ -89,69 +94,101 @@ Generation Options:
   --end-img <string>                       path to the end image, required by flf2v
   --mask <string>                          path to the mask image
   --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
-                                           lexicographical (character) order. For example, if the control video path is
-                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
+                                           inside should be stored as images in lexicographical (character) order. For
+                                           example, if the control video path is `frames`, the directory contain images
+                                           such as 00.png, 01.png, ... etc.
   --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
   --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
+  --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 args, key=value list. Currently lcm supports noise_clip_std,
+                                           noise_scale_start, noise_scale_end
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
   --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
-                                           will be 1 for SD1.x, 2 for SD2.x
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
+                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   -b, --batch-count <int>                  batch count
   --video-frames <int>                     video frames (default: 1)
   --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
-                                           NitroSD-Vibrant
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
+                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
   --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
+  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
+  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
+  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
+  --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 instruct-pix2pix models: (default: same
+                                           as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
-                                           medium
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
+                                           disabled, a value of 2.5 is nice for sd3.5 medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and
+                                           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-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
+                                           (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
+                                           (default: 3.5)
+  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default:
+                                           0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
+  --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>
-  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
-  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
+  --pm-style-strength <float>              
+  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full
+                                           destruction of information in init image
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
+                                           `--high-noise-steps` is set to -1
   --vace-strength <float>                  wan vace strength
-  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
+  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
+  --increase-ref-index                     automatically increase the indices of references images based on the order
+                                           they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
+  --vae-tiling                             process vae in tiles to reduce memory usage
+  --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, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd, res_multistep, res_2s, er_sde] (default: euler for Flux/SD3/Wan, euler_a
-                                           otherwise)
-  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           ddim_trailing, tcd, res_multistep, res_2s, er_sde] default: euler for Flux/SD3/Wan,
-                                           euler_a otherwise
-  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           kl_optimal, lcm, bong_tangent], default: discrete
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
+                                           dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s,
+                                           er_sde, euler_cfg_pp, euler_a_cfg_pp] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a,
+                                           dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
+                                           res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] default: euler for Flux/SD3/Wan, euler_a otherwise
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
+                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
+                                           discrete
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
+                                           "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level),
-                                           'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
+                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
+                                           Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=;
-                                           spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=. Examples:
-                                           "threshold=0.25" or "threshold=1.5,reset=0" or "w=0.4,window=2"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
+                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
+                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
+                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
+                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
 ```
 
 Metadata mode inspects PNG/JPEG container metadata without loading any model:

examples/cli/main.cpp

@@ -278,7 +278,9 @@ void parse_args(int argc, const char** argv, SDCliParams& cli_params, SDContextP
     bool valid = cli_params.resolve_and_validate();
     if (valid && cli_params.mode != METADATA) {
         valid = ctx_params.resolve_and_validate(cli_params.mode) &&
-                gen_params.resolve_and_validate(cli_params.mode, ctx_params.lora_model_dir);
+                gen_params.resolve_and_validate(cli_params.mode,
+                                                ctx_params.lora_model_dir,
+                                                ctx_params.hires_upscalers_dir);
     }
 
     if (!valid) {
@@ -431,10 +433,11 @@ bool save_results(const SDCliParams& cli_params,
         if (!img.data)
             return false;
 
-        std::string params = gen_params.embed_image_metadata
-                                 ? get_image_params(ctx_params, gen_params, gen_params.seed + idx)
-                                 : "";
-        const bool ok      = write_image_to_file(path.string(), img.data, img.width, img.height, img.channel, params, 90);
+        const int64_t metadata_seed = cli_params.mode == VID_GEN ? gen_params.seed : gen_params.seed + idx;
+        std::string params          = gen_params.embed_image_metadata
+                                          ? get_image_params(ctx_params, gen_params, metadata_seed, cli_params.mode)
+                                          : "";
+        const bool ok               = write_image_to_file(path.string(), img.data, img.width, img.height, img.channel, params, 90);
         LOG_INFO("save result image %d to '%s' (%s)", idx, path.string().c_str(), ok ? "success" : "failure");
         return ok;
     };
@@ -688,6 +691,13 @@ int main(int argc, const char* argv[]) {
         vae_decode_only = false;
     }
 
+    if (gen_params.hires_enabled &&
+        (gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_MODEL ||
+         gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_LANCZOS ||
+         gen_params.resolved_hires_upscaler == SD_HIRES_UPSCALER_NEAREST)) {
+        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);
 
     SDImageVec results;
@@ -739,7 +749,9 @@ int main(int argc, const char* argv[]) {
                                                      ctx_params.offload_params_to_cpu,
                                                      ctx_params.diffusion_conv_direct,
                                                      ctx_params.n_threads,
-                                                     gen_params.upscale_tile_size));
+                                                     gen_params.upscale_tile_size,
+                                                     ctx_params.backend.c_str(),
+                                                     ctx_params.params_backend.c_str()));
 
         if (upscaler_ctx == nullptr) {
             LOG_ERROR("new_upscaler_ctx failed");

examples/common/common.cpp

@@ -107,47 +107,60 @@ static bool is_absolute_path(const std::string& p) {
 
 std::string ArgOptions::wrap_text(const std::string& text, size_t width, size_t indent) {
     std::ostringstream oss;
-    size_t line_len = 0;
     size_t pos      = 0;
+    size_t line_len = 0;
 
     while (pos < text.size()) {
-        // Preserve manual newlines
         if (text[pos] == '\n') {
             oss << '\n'
                 << std::string(indent, ' ');
-            line_len = indent;
+            line_len = 0;
             ++pos;
             continue;
         }
 
-        // Add the character
-        oss << text[pos];
-        ++line_len;
-        ++pos;
+        if (std::isspace(static_cast<unsigned char>(text[pos]))) {
+            ++pos;
+            continue;
+        }
 
-        // If the current line exceeds width, try to break at the last space
-        if (line_len >= width) {
-            std::string current = oss.str();
-            size_t back         = current.size();
+        size_t word_start = pos;
+        while (pos < text.size() &&
+               text[pos] != '\n' &&
+               !std::isspace(static_cast<unsigned char>(text[pos]))) {
+            ++pos;
+        }
 
-            // Find the last space (for a clean break)
-            while (back > 0 && current[back - 1] != ' ' && current[back - 1] != '\n')
-                --back;
+        std::string word = text.substr(word_start, pos - word_start);
+        while (!word.empty()) {
+            size_t separator_len = line_len == 0 ? 0 : 1;
+            if (line_len + separator_len + word.size() <= width) {
+                if (separator_len > 0) {
+                    oss << ' ';
+                    ++line_len;
+                }
+                oss << word;
+                line_len += word.size();
+                word.clear();
+                continue;
+            }
 
-            // If found a space to break on
-            if (back > 0 && current[back - 1] != '\n') {
-                std::string before = current.substr(0, back - 1);
-                std::string after  = current.substr(back);
-                oss.str("");
-                oss.clear();
-                oss << before << "\n"
-                    << std::string(indent, ' ') << after;
-            } else {
-                // If no space found, just break at width
-                oss << "\n"
+            if (line_len > 0) {
+                oss << '\n'
                     << std::string(indent, ' ');
+                line_len = 0;
+                continue;
+            }
+
+            size_t chunk_len = std::min(width, word.size());
+            oss << word.substr(0, chunk_len);
+            line_len = chunk_len;
+            word.erase(0, chunk_len);
+            if (!word.empty()) {
+                oss << '\n'
+                    << std::string(indent, ' ');
+                line_len = 0;
             }
-            line_len = indent;
         }
     }
 
@@ -351,7 +364,10 @@ ArgOptions SDContextParams::get_options() {
          "--lora-model-dir",
          "lora model directory",
          &lora_model_dir},
-
+        {"",
+         "--hires-upscalers-dir",
+         "highres fix upscaler model directory",
+         &hires_upscalers_dir},
         {"",
          "--tensor-type-rules",
          "weight type per tensor pattern (example: \"^vae\\.=f16,model\\.=q8_0\")",
@@ -364,6 +380,14 @@ ArgOptions SDContextParams::get_options() {
          "--upscale-model",
          "path to esrgan model.",
          &esrgan_path},
+        {"",
+         "--backend",
+         "runtime backend assignment, e.g. cpu or clip=cpu,vae=cuda0,diffusion=vulkan0",
+         &backend},
+        {"",
+         "--params-backend",
+         "parameter backend assignment, e.g. cpu or diffusion=cpu,clip=cpu",
+         &params_backend},
     };
 
     options.int_options = {
@@ -378,7 +402,12 @@ ArgOptions SDContextParams::get_options() {
          &chroma_t5_mask_pad},
     };
 
-    options.float_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",
+         &max_vram},
+    };
 
     options.bool_options = {
         {"",
@@ -649,10 +678,14 @@ std::string SDContextParams::to_string() const {
         << "  wtype: " << sd_type_name(wtype) << ",\n"
         << "  tensor_type_rules: \"" << tensor_type_rules << "\",\n"
         << "  lora_model_dir: \"" << lora_model_dir << "\",\n"
+        << "  hires_upscalers_dir: \"" << hires_upscalers_dir << "\",\n"
         << "  photo_maker_path: \"" << photo_maker_path << "\",\n"
         << "  rng_type: " << sd_rng_type_name(rng_type) << ",\n"
         << "  sampler_rng_type: " << sd_rng_type_name(sampler_rng_type) << ",\n"
         << "  offload_params_to_cpu: " << (offload_params_to_cpu ? "true" : "false") << ",\n"
+        << "  max_vram: " << max_vram << ",\n"
+        << "  backend: \"" << backend << "\",\n"
+        << "  params_backend: \"" << params_backend << "\",\n"
         << "  enable_mmap: " << (enable_mmap ? "true" : "false") << ",\n"
         << "  control_net_cpu: " << (control_net_cpu ? "true" : "false") << ",\n"
         << "  clip_on_cpu: " << (clip_on_cpu ? "true" : "false") << ",\n"
@@ -727,6 +760,9 @@ 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,
+        max_vram,
+        backend.c_str(),
+        params_backend.c_str(),
     };
     return sd_ctx_params;
 }
@@ -777,6 +813,16 @@ ArgOptions SDGenerationParams::get_options() {
          "--pm-id-embed-path",
          "path to PHOTOMAKER v2 id embed",
          &pm_id_embed_path},
+        {"",
+         "--hires-upscaler",
+         "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)",
+         &hires_upscaler},
+        {"",
+         "--extra-sample-args",
+         "extra sampler args, key=value list. Currently lcm supports noise_clip_std, noise_scale_start, noise_scale_end",
+         &extra_sample_args},
     };
 
     options.int_options = {
@@ -826,6 +872,22 @@ ArgOptions SDGenerationParams::get_options() {
          "--upscale-tile-size",
          "tile size for ESRGAN upscaling (default: 128)",
          &upscale_tile_size},
+        {"",
+         "--hires-width",
+         "highres fix target width, 0 to use --hires-scale (default: 0)",
+         &hires_width},
+        {"",
+         "--hires-height",
+         "highres fix target height, 0 to use --hires-scale (default: 0)",
+         &hires_height},
+        {"",
+         "--hires-steps",
+         "highres fix second pass sample steps, 0 to reuse --steps (default: 0)",
+         &hires_steps},
+        {"",
+         "--hires-upscale-tile-size",
+         "highres fix upscaler tile size, reserved for model-backed upscalers (default: 128)",
+         &hires_upscale_tile_size},
     };
 
     options.float_options = {
@@ -913,6 +975,14 @@ ArgOptions SDGenerationParams::get_options() {
          "--vae-tile-overlap",
          "tile overlap for vae tiling, in fraction of tile size (default: 0.5)",
          &vae_tiling_params.target_overlap},
+        {"",
+         "--hires-scale",
+         "highres fix scale when target size is not set (default: 2.0)",
+         &hires_scale},
+        {"",
+         "--hires-denoising-strength",
+         "highres fix second pass denoising strength (default: 0.7)",
+         &hires_denoising_strength},
     };
 
     options.bool_options = {
@@ -936,6 +1006,11 @@ ArgOptions SDGenerationParams::get_options() {
          "process vae in tiles to reduce memory usage",
          true,
          &vae_tiling_params.enabled},
+        {"",
+         "--hires",
+         "enable highres fix",
+         true,
+         &hires_enabled},
     };
 
     auto on_seed_arg = [&](int argc, const char** argv, int index) {
@@ -1185,12 +1260,12 @@ ArgOptions SDGenerationParams::get_options() {
          on_seed_arg},
         {"",
          "--sampling-method",
-         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde] "
+         "sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
          "(default: euler for Flux/SD3/Wan, euler_a otherwise)",
          on_sample_method_arg},
         {"",
          "--high-noise-sampling-method",
-         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde]"
+         "(high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp]"
          " default: euler for Flux/SD3/Wan, euler_a otherwise",
          on_high_noise_sample_method_arg},
         {"",
@@ -1424,6 +1499,37 @@ static bool parse_lora_json_field(const json& parent,
     return true;
 }
 
+static bool resolve_model_file_from_dir(const std::string& model_name,
+                                        const std::string& model_dir,
+                                        const std::vector<std::string>& valid_ext,
+                                        const char* label,
+                                        std::string& resolved_path) {
+    if (model_dir.empty()) {
+        LOG_ERROR("%s directory is empty", label);
+        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()) {
+        LOG_ERROR("%s must be a model name without path or extension: %s", label, model_name.c_str());
+        return false;
+    }
+
+    fs::path model_dir_path = model_dir;
+    for (const auto& ext : valid_ext) {
+        fs::path try_path = model_dir_path / (model_name + ext);
+        if (fs::exists(try_path) && fs::is_regular_file(try_path)) {
+            resolved_path = try_path.lexically_normal().string();
+            return true;
+        }
+    }
+
+    LOG_ERROR("can not find %s %s in %s", label, model_name.c_str(), model_dir_path.lexically_normal().string().c_str());
+    return false;
+}
+
 bool SDGenerationParams::from_json_str(
     const std::string& json_str,
     const std::function<std::string(const std::string&)>& lora_path_resolver) {
@@ -1487,8 +1593,37 @@ bool SDGenerationParams::from_json_str(
     load_if_exists("increase_ref_index", increase_ref_index);
     load_if_exists("embed_image_metadata", embed_image_metadata);
 
+    if (j.contains("hires") && j["hires"].is_object()) {
+        const json& hires_json = j["hires"];
+        if (hires_json.contains("enabled") && hires_json["enabled"].is_boolean()) {
+            hires_enabled = hires_json["enabled"];
+        }
+        if (hires_json.contains("upscaler") && hires_json["upscaler"].is_string()) {
+            hires_upscaler = hires_json["upscaler"];
+        }
+        if (hires_json.contains("scale") && hires_json["scale"].is_number()) {
+            hires_scale = hires_json["scale"];
+        }
+        if (hires_json.contains("target_width") && hires_json["target_width"].is_number_integer()) {
+            hires_width = hires_json["target_width"];
+        }
+        if (hires_json.contains("target_height") && hires_json["target_height"].is_number_integer()) {
+            hires_height = hires_json["target_height"];
+        }
+        if (hires_json.contains("steps") && hires_json["steps"].is_number_integer()) {
+            hires_steps = hires_json["steps"];
+        }
+        if (hires_json.contains("denoising_strength") && hires_json["denoising_strength"].is_number()) {
+            hires_denoising_strength = hires_json["denoising_strength"];
+        }
+        if (hires_json.contains("upscale_tile_size") && hires_json["upscale_tile_size"].is_number_integer()) {
+            hires_upscale_tile_size = hires_json["upscale_tile_size"];
+        }
+    }
+
     auto parse_sample_params_json = [&](const json& sample_json,
                                         sd_sample_params_t& target_params,
+                                        std::string& target_extra_sample_args,
                                         std::vector<int>& target_skip_layers,
                                         std::vector<float>* target_custom_sigmas) {
         if (sample_json.contains("sample_steps") && sample_json["sample_steps"].is_number_integer()) {
@@ -1503,6 +1638,9 @@ bool SDGenerationParams::from_json_str(
         if (sample_json.contains("flow_shift") && sample_json["flow_shift"].is_number()) {
             target_params.flow_shift = sample_json["flow_shift"];
         }
+        if (sample_json.contains("extra_sample_args") && sample_json["extra_sample_args"].is_string()) {
+            target_extra_sample_args = sample_json["extra_sample_args"].get<std::string>();
+        }
         if (target_custom_sigmas != nullptr &&
             sample_json.contains("custom_sigmas") &&
             sample_json["custom_sigmas"].is_array()) {
@@ -1550,11 +1688,12 @@ bool SDGenerationParams::from_json_str(
     };
 
     if (j.contains("sample_params") && j["sample_params"].is_object()) {
-        parse_sample_params_json(j["sample_params"], sample_params, skip_layers, &custom_sigmas);
+        parse_sample_params_json(j["sample_params"], sample_params, extra_sample_args, skip_layers, &custom_sigmas);
     }
     if (j.contains("high_noise_sample_params") && j["high_noise_sample_params"].is_object()) {
         parse_sample_params_json(j["high_noise_sample_params"],
                                  high_noise_sample_params,
+                                 high_noise_extra_sample_args,
                                  high_noise_skip_layers,
                                  nullptr);
     }
@@ -1800,7 +1939,7 @@ bool SDGenerationParams::initialize_cache_params() {
     return true;
 }
 
-bool SDGenerationParams::resolve(const std::string& lora_model_dir, bool strict) {
+bool SDGenerationParams::resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict) {
     if (high_noise_sample_params.sample_steps <= 0) {
         high_noise_sample_params.sample_steps = -1;
     }
@@ -1819,6 +1958,27 @@ bool SDGenerationParams::resolve(const std::string& lora_model_dir, bool strict)
         sample_params.sample_steps = std::clamp(sample_params.sample_steps, 1, 100);
     }
 
+    hires_upscaler_model_path.clear();
+    if (hires_enabled) {
+        if (hires_upscaler.empty()) {
+            hires_upscaler = "Latent";
+        }
+        resolved_hires_upscaler = str_to_sd_hires_upscaler(hires_upscaler.c_str());
+        if (resolved_hires_upscaler == SD_HIRES_UPSCALER_NONE) {
+            hires_enabled = false;
+        } else if (resolved_hires_upscaler == SD_HIRES_UPSCALER_COUNT) {
+            static const std::vector<std::string> valid_ext = {".gguf", ".safetensors", ".pt", ".pth"};
+            if (!resolve_model_file_from_dir(hires_upscaler,
+                                             hires_upscalers_dir,
+                                             valid_ext,
+                                             "hires upscaler",
+                                             hires_upscaler_model_path)) {
+                return false;
+            }
+            resolved_hires_upscaler = SD_HIRES_UPSCALER_MODEL;
+        }
+    }
+
     prompt_with_lora = prompt;
     if (!lora_model_dir.empty()) {
         extract_and_remove_lora(lora_model_dir);
@@ -1883,6 +2043,29 @@ bool SDGenerationParams::validate(SDMode mode) {
         return false;
     }
 
+    if (hires_enabled) {
+        if (hires_width < 0 || hires_height < 0) {
+            LOG_ERROR("error: hires target width and height must be >= 0");
+            return false;
+        }
+        if (hires_scale <= 0.f && hires_width <= 0 && hires_height <= 0) {
+            LOG_ERROR("error: hires scale must be positive when target size is not set");
+            return false;
+        }
+        if (hires_steps < 0) {
+            LOG_ERROR("error: hires steps must be >= 0");
+            return false;
+        }
+        if (hires_denoising_strength <= 0.f || hires_denoising_strength > 1.f) {
+            LOG_ERROR("error: hires denoising strength must be in (0.0, 1.0]");
+            return false;
+        }
+        if (hires_upscale_tile_size < 1) {
+            LOG_ERROR("error: hires upscale tile size must be positive");
+            return false;
+        }
+    }
+
     if (mode == UPSCALE) {
         if (init_image_path.length() == 0) {
             LOG_ERROR("error: upscale mode needs an init image (--init-img)\n");
@@ -1893,8 +2076,11 @@ bool SDGenerationParams::validate(SDMode mode) {
     return true;
 }
 
-bool SDGenerationParams::resolve_and_validate(SDMode mode, const std::string& lora_model_dir, bool strict) {
-    if (!resolve(lora_model_dir, strict)) {
+bool SDGenerationParams::resolve_and_validate(SDMode mode,
+                                              const std::string& lora_model_dir,
+                                              const std::string& hires_upscalers_dir,
+                                              bool strict) {
+    if (!resolve(lora_model_dir, hires_upscalers_dir, strict)) {
         return false;
     }
     if (!validate(mode)) {
@@ -1934,6 +2120,8 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     high_noise_sample_params.guidance.slg.layer_count = high_noise_skip_layers.size();
     sample_params.custom_sigmas                       = custom_sigmas.empty() ? nullptr : custom_sigmas.data();
     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();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     sd_pm_params_t pm_params = {
@@ -1965,6 +2153,16 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.pm_params             = pm_params;
     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;
     return params;
 }
 
@@ -1993,6 +2191,8 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     high_noise_sample_params.guidance.slg.layer_count = high_noise_skip_layers.size();
     sample_params.custom_sigmas                       = custom_sigmas.empty() ? nullptr : custom_sigmas.data();
     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();
     cache_params.scm_mask                             = scm_mask.empty() ? nullptr : scm_mask.c_str();
 
     params.loras                    = lora_vec.empty() ? nullptr : lora_vec.data();
@@ -2089,6 +2289,15 @@ std::string SDGenerationParams::to_string() const {
         << "  seed: " << seed << ",\n"
         << "  upscale_repeats: " << upscale_repeats << ",\n"
         << "  upscale_tile_size: " << upscale_tile_size << ",\n"
+        << "  hires: { enabled: " << (hires_enabled ? "true" : "false")
+        << ", upscaler: \"" << hires_upscaler << "\""
+        << ", model_path: \"" << hires_upscaler_model_path << "\""
+        << ", scale: " << hires_scale
+        << ", target_width: " << hires_width
+        << ", target_height: " << hires_height
+        << ", steps: " << hires_steps
+        << ", denoising_strength: " << hires_denoising_strength
+        << ", upscale_tile_size: " << hires_upscale_tile_size << " },\n"
         << "  vae_tiling_params: { "
         << vae_tiling_params.enabled << ", "
         << vae_tiling_params.tile_size_x << ", "
@@ -2104,7 +2313,193 @@ std::string version_string() {
     return std::string("stable-diffusion.cpp version ") + sd_version() + ", commit " + sd_commit();
 }
 
-std::string get_image_params(const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed) {
+static std::string safe_json_string(const char* value) {
+    return value ? value : "";
+}
+
+static void set_json_basename_if_not_empty(json& target, const char* key, const std::string& path) {
+    if (!path.empty()) {
+        target[key] = sd_basename(path);
+    }
+}
+
+static json build_sampling_metadata_json(const sd_sample_params_t& sample_params,
+                                         const std::vector<int>& skip_layers,
+                                         const std::vector<float>* custom_sigmas = nullptr) {
+    json sampling = {
+        {"steps", sample_params.sample_steps},
+        {"eta", sample_params.eta},
+        {"shifted_timestep", sample_params.shifted_timestep},
+        {"flow_shift", sample_params.flow_shift},
+        {"extra_sample_args", safe_json_string(sample_params.extra_sample_args)},
+        {"guidance",
+         {
+             {"txt_cfg", sample_params.guidance.txt_cfg},
+             {"img_cfg", sample_params.guidance.img_cfg},
+             {"distilled_guidance", sample_params.guidance.distilled_guidance},
+             {"slg",
+              {
+                  {"scale", sample_params.guidance.slg.scale},
+                  {"layers", skip_layers},
+                  {"start", sample_params.guidance.slg.layer_start},
+                  {"end", sample_params.guidance.slg.layer_end},
+              }},
+         }},
+    };
+    if (sample_params.sample_method != SAMPLE_METHOD_COUNT) {
+        sampling["method"] = safe_json_string(sd_sample_method_name(sample_params.sample_method));
+    }
+    if (sample_params.scheduler != SCHEDULER_COUNT) {
+        sampling["scheduler"] = safe_json_string(sd_scheduler_name(sample_params.scheduler));
+    }
+    if (custom_sigmas != nullptr) {
+        sampling["custom_sigmas"] = *custom_sigmas;
+    }
+    return sampling;
+}
+
+std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
+                                            const SDGenerationParams& gen_params,
+                                            int64_t seed,
+                                            SDMode mode) {
+    json root;
+    root["schema"]    = "sdcpp.image.params/v1";
+    root["mode"]      = mode == VID_GEN ? "vid_gen" : "img_gen";
+    root["generator"] = {
+        {"name", "stable-diffusion.cpp"},
+        {"version", safe_json_string(sd_version())},
+        {"commit", safe_json_string(sd_commit())},
+    };
+    root["seed"]   = seed;
+    root["width"]  = gen_params.get_resolved_width();
+    root["height"] = gen_params.get_resolved_height();
+
+    root["prompt"] = {
+        {"positive", gen_params.prompt},
+        {"negative", gen_params.negative_prompt},
+    };
+    root["sampling"] = build_sampling_metadata_json(gen_params.sample_params,
+                                                    gen_params.skip_layers,
+                                                    &gen_params.custom_sigmas);
+
+    json models;
+    set_json_basename_if_not_empty(models, "model", ctx_params.model_path);
+    set_json_basename_if_not_empty(models, "clip_l", ctx_params.clip_l_path);
+    set_json_basename_if_not_empty(models, "clip_g", ctx_params.clip_g_path);
+    set_json_basename_if_not_empty(models, "clip_vision", ctx_params.clip_vision_path);
+    set_json_basename_if_not_empty(models, "t5xxl", ctx_params.t5xxl_path);
+    set_json_basename_if_not_empty(models, "llm", ctx_params.llm_path);
+    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, "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);
+    root["models"] = std::move(models);
+
+    root["clip_skip"]             = gen_params.clip_skip;
+    root["strength"]              = gen_params.strength;
+    root["control_strength"]      = gen_params.control_strength;
+    root["auto_resize_ref_image"] = gen_params.auto_resize_ref_image;
+    root["increase_ref_index"]    = gen_params.increase_ref_index;
+    if (mode == VID_GEN) {
+        root["video"] = {
+            {"frame_count", gen_params.video_frames},
+            {"fps", gen_params.fps},
+        };
+        root["moe_boundary"]        = gen_params.moe_boundary;
+        root["vace_strength"]       = gen_params.vace_strength;
+        root["high_noise_sampling"] = build_sampling_metadata_json(gen_params.high_noise_sample_params,
+                                                                   gen_params.high_noise_skip_layers);
+    }
+
+    root["rng"] = safe_json_string(sd_rng_type_name(ctx_params.rng_type));
+    if (ctx_params.sampler_rng_type != RNG_TYPE_COUNT) {
+        root["sampler_rng"] = safe_json_string(sd_rng_type_name(ctx_params.sampler_rng_type));
+    }
+
+    json loras = json::array();
+    for (const auto& entry : gen_params.lora_map) {
+        loras.push_back({
+            {"name", sd_basename(entry.first)},
+            {"multiplier", entry.second},
+            {"is_high_noise", false},
+        });
+    }
+    for (const auto& entry : gen_params.high_noise_lora_map) {
+        loras.push_back({
+            {"name", sd_basename(entry.first)},
+            {"multiplier", entry.second},
+            {"is_high_noise", true},
+        });
+    }
+    if (!loras.empty()) {
+        root["loras"] = std::move(loras);
+    }
+
+    if (gen_params.hires_enabled) {
+        root["hires"] = {
+            {"enabled", gen_params.hires_enabled},
+            {"upscaler", gen_params.hires_upscaler},
+            {"model", gen_params.hires_upscaler_model_path.empty() ? "" : sd_basename(gen_params.hires_upscaler_model_path)},
+            {"scale", gen_params.hires_scale},
+            {"target_width", gen_params.hires_width},
+            {"target_height", gen_params.hires_height},
+            {"steps", gen_params.hires_steps},
+            {"denoising_strength", gen_params.hires_denoising_strength},
+            {"upscale_tile_size", gen_params.hires_upscale_tile_size},
+        };
+    }
+
+    if (gen_params.cache_params.mode != SD_CACHE_DISABLED) {
+        root["cache"] = {
+            {"requested_mode", gen_params.cache_mode},
+            {"requested_option", gen_params.cache_option},
+            {"mode", gen_params.cache_params.mode},
+            {"scm_mask", gen_params.scm_mask},
+            {"scm_policy_dynamic", gen_params.scm_policy_dynamic},
+            {"reuse_threshold", gen_params.cache_params.reuse_threshold},
+            {"start_percent", gen_params.cache_params.start_percent},
+            {"end_percent", gen_params.cache_params.end_percent},
+            {"error_decay_rate", gen_params.cache_params.error_decay_rate},
+            {"use_relative_threshold", gen_params.cache_params.use_relative_threshold},
+            {"reset_error_on_compute", gen_params.cache_params.reset_error_on_compute},
+            {"Fn_compute_blocks", gen_params.cache_params.Fn_compute_blocks},
+            {"Bn_compute_blocks", gen_params.cache_params.Bn_compute_blocks},
+            {"residual_diff_threshold", gen_params.cache_params.residual_diff_threshold},
+            {"max_warmup_steps", gen_params.cache_params.max_warmup_steps},
+            {"max_cached_steps", gen_params.cache_params.max_cached_steps},
+            {"max_continuous_cached_steps", gen_params.cache_params.max_continuous_cached_steps},
+            {"taylorseer_n_derivatives", gen_params.cache_params.taylorseer_n_derivatives},
+            {"taylorseer_skip_interval", gen_params.cache_params.taylorseer_skip_interval},
+            {"spectrum_w", gen_params.cache_params.spectrum_w},
+            {"spectrum_m", gen_params.cache_params.spectrum_m},
+            {"spectrum_lam", gen_params.cache_params.spectrum_lam},
+            {"spectrum_window_size", gen_params.cache_params.spectrum_window_size},
+            {"spectrum_flex_window", gen_params.cache_params.spectrum_flex_window},
+            {"spectrum_warmup_steps", gen_params.cache_params.spectrum_warmup_steps},
+            {"spectrum_stop_percent", gen_params.cache_params.spectrum_stop_percent},
+        };
+    }
+
+    if (gen_params.vae_tiling_params.enabled) {
+        root["vae_tiling"] = {
+            {"enabled", gen_params.vae_tiling_params.enabled},
+            {"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},
+        };
+    }
+
+    return root.dump();
+}
+
+std::string get_image_params(const SDContextParams& ctx_params,
+                             const SDGenerationParams& gen_params,
+                             int64_t seed,
+                             SDMode mode) {
     std::string parameter_string;
     if (gen_params.prompt_with_lora.size() != 0) {
         parameter_string += gen_params.prompt_with_lora + "\n";
@@ -2117,7 +2512,7 @@ std::string get_image_params(const SDContextParams& ctx_params, const SDGenerati
     parameter_string += "Steps: " + std::to_string(gen_params.sample_params.sample_steps) + ", ";
     parameter_string += "CFG scale: " + std::to_string(gen_params.sample_params.guidance.txt_cfg) + ", ";
     if (gen_params.sample_params.guidance.slg.scale != 0 && gen_params.skip_layers.size() != 0) {
-        parameter_string += "SLG scale: " + std::to_string(gen_params.sample_params.guidance.txt_cfg) + ", ";
+        parameter_string += "SLG scale: " + std::to_string(gen_params.sample_params.guidance.slg.scale) + ", ";
         parameter_string += "Skip layers: [";
         for (const auto& layer : gen_params.skip_layers) {
             parameter_string += std::to_string(layer) + ", ";
@@ -2128,6 +2523,9 @@ std::string get_image_params(const SDContextParams& ctx_params, const SDGenerati
     }
     parameter_string += "Guidance: " + std::to_string(gen_params.sample_params.guidance.distilled_guidance) + ", ";
     parameter_string += "Eta: " + std::to_string(gen_params.sample_params.eta) + ", ";
+    if (!gen_params.extra_sample_args.empty()) {
+        parameter_string += "Extra sample args: " + gen_params.extra_sample_args + ", ";
+    }
     parameter_string += "Seed: " + std::to_string(seed) + ", ";
     parameter_string += "Size: " + std::to_string(gen_params.get_resolved_width()) + "x" + std::to_string(gen_params.get_resolved_height()) + ", ";
     parameter_string += "Model: " + sd_basename(ctx_params.model_path) + ", ";
@@ -2162,6 +2560,14 @@ std::string get_image_params(const SDContextParams& ctx_params, const SDGenerati
     if (gen_params.clip_skip != -1) {
         parameter_string += "Clip skip: " + std::to_string(gen_params.clip_skip) + ", ";
     }
+    if (gen_params.hires_enabled) {
+        parameter_string += "Hires upscale: " + gen_params.hires_upscaler + ", ";
+        parameter_string += "Hires scale: " + std::to_string(gen_params.hires_scale) + ", ";
+        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) + ", ";
+    }
     parameter_string += "Version: stable-diffusion.cpp";
+    parameter_string += ", SDCPP: " + build_sdcpp_image_metadata_json(ctx_params, gen_params, seed, mode);
     return parameter_string;
 }

examples/common/common.h

@@ -101,6 +101,7 @@ struct SDContextParams {
     sd_type_t wtype = SD_TYPE_COUNT;
     std::string tensor_type_rules;
     std::string lora_model_dir = ".";
+    std::string hires_upscalers_dir;
 
     std::map<std::string, std::string> embedding_map;
     std::vector<sd_embedding_t> embedding_vec;
@@ -108,14 +109,17 @@ struct SDContextParams {
     rng_type_t rng_type         = CUDA_RNG;
     rng_type_t sampler_rng_type = RNG_TYPE_COUNT;
     bool offload_params_to_cpu  = false;
-    bool enable_mmap            = false;
-    bool control_net_cpu        = false;
-    bool clip_on_cpu            = false;
-    bool vae_on_cpu             = false;
-    bool flash_attn             = false;
-    bool diffusion_flash_attn   = false;
-    bool diffusion_conv_direct  = false;
-    bool vae_conv_direct        = false;
+    float max_vram              = 0.f;
+    std::string backend;
+    std::string params_backend;
+    bool enable_mmap           = false;
+    bool control_net_cpu       = false;
+    bool clip_on_cpu           = false;
+    bool vae_on_cpu            = false;
+    bool flash_attn            = false;
+    bool diffusion_flash_attn  = false;
+    bool diffusion_conv_direct = false;
+    bool vae_conv_direct       = false;
 
     bool circular   = false;
     bool circular_x = false;
@@ -166,6 +170,8 @@ struct SDGenerationParams {
 
     sd_sample_params_t sample_params;
     sd_sample_params_t high_noise_sample_params;
+    std::string extra_sample_args;
+    std::string high_noise_extra_sample_args;
     std::vector<int> skip_layers            = {7, 8, 9};
     std::vector<int> high_noise_skip_layers = {7, 8, 9};
 
@@ -190,12 +196,23 @@ struct SDGenerationParams {
     int upscale_repeats   = 1;
     int upscale_tile_size = 128;
 
+    bool hires_enabled         = false;
+    std::string hires_upscaler = "Latent";
+    std::string hires_upscaler_model_path;
+    float hires_scale              = 2.f;
+    int hires_width                = 0;
+    int hires_height               = 0;
+    int hires_steps                = 0;
+    float hires_denoising_strength = 0.7f;
+    int hires_upscale_tile_size    = 128;
+
     std::map<std::string, float> lora_map;
     std::map<std::string, float> high_noise_lora_map;
 
     // Derived and normalized fields.
     std::string prompt_with_lora;  // for metadata record only
     std::vector<sd_lora_t> lora_vec;
+    sd_hires_upscaler_t resolved_hires_upscaler;
 
     // Owned execution payload.
     SDImageOwner init_image;
@@ -225,15 +242,25 @@ struct SDGenerationParams {
     void set_width_and_height_if_unset(int w, int h);
     int get_resolved_width() const;
     int get_resolved_height() const;
-    bool resolve(const std::string& lora_model_dir, bool strict = false);
+    bool resolve(const std::string& lora_model_dir, const std::string& hires_upscalers_dir, bool strict = false);
     bool validate(SDMode mode);
-    bool resolve_and_validate(SDMode mode, const std::string& lora_model_dir, bool strict = false);
+    bool resolve_and_validate(SDMode mode,
+                              const std::string& lora_model_dir,
+                              const std::string& hires_upscalers_dir,
+                              bool strict = false);
     sd_img_gen_params_t to_sd_img_gen_params_t();
     sd_vid_gen_params_t to_sd_vid_gen_params_t();
     std::string to_string() const;
 };
 
 std::string version_string();
-std::string get_image_params(const SDContextParams& ctx_params, const SDGenerationParams& gen_params, int64_t seed);
+std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
+                                            const SDGenerationParams& gen_params,
+                                            int64_t seed,
+                                            SDMode mode = IMG_GEN);
+std::string get_image_params(const SDContextParams& ctx_params,
+                             const SDGenerationParams& gen_params,
+                             int64_t seed,
+                             SDMode mode = IMG_GEN);
 
 #endif  // __EXAMPLES_COMMON_COMMON_H__

examples/server/README.md

@@ -123,11 +123,11 @@ In this case, the server will load and serve the specified `index.html` file ins
 usage: ./bin/sd-server  [options]
 
 Svr Options:
-  -l, --listen-ip <string>      server listen ip (default: 127.0.0.1)        
+  -l, --listen-ip <string>      server listen ip (default: 127.0.0.1)
   --serve-html-path <string>    path to HTML file to serve at root (optional)
   --listen-port <int>           server listen port (default: 1234)
   -v, --verbose                 print extra info
-  --color                       colors the logging tags according to level   
+  --color                       colors the logging tags according to level
   -h, --help                    show this help message and exit
 
 Context Options:
@@ -136,7 +136,8 @@ Context Options:
   --clip_g <string>                        path to the clip-g text encoder
   --clip_vision <string>                   path to the clip-vision encoder
   --t5xxl <string>                         path to the t5xxl text encoder
-  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image, mistral-small3.2 for flux2, ...)
+  --llm <string>                           path to the llm text encoder. For example: (qwenvl2.5 for qwen-image,
+                                           mistral-small3.2 for flux2, ...)
   --llm_vision <string>                    path to the llm vit
   --qwen2vl <string>                       alias of --llm. Deprecated.
   --qwen2vl_vision <string>                alias of --llm_vision. Deprecated.
@@ -148,16 +149,18 @@ Context Options:
   --control-net <string>                   path to control net model
   --embd-dir <string>                      embeddings directory
   --lora-model-dir <string>                lora model directory
+  --hires-upscalers-dir <string>           highres fix upscaler model directory
   --tensor-type-rules <string>             weight type per tensor pattern (example: "^vae\.=f16,model\.=q8_0")
   --photo-maker <string>                   path to PHOTOMAKER model
   --upscale-model <string>                 path to esrgan model.
-  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0, then threads will be set to the number of
-                                           CPU physical cores
+  -t, --threads <int>                      number of threads to use during computation (default: -1). If threads <= 0,
+                                           then threads will be set to the number of CPU physical cores
   --chroma-t5-mask-pad <int>               t5 mask pad size of chroma
-  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
-  --vae-tiling                             process vae in tiles to reduce memory usage
+  --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
   --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
+  --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM
+                                           when needed
   --mmap                                   whether to memory-map model
   --control-net-cpu                        keep controlnet in cpu (for low vram)
   --clip-on-cpu                            keep clip in cpu (for low vram)
@@ -172,20 +175,19 @@ Context Options:
   --chroma-disable-dit-mask                disable dit mask for chroma
   --qwen-image-zero-cond-t                 enable zero_cond_t for qwen image
   --chroma-enable-t5-mask                  enable t5 mask for chroma
-  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K, q4_K). If not specified, the default is the
-                                           type of the weight file
+  --type                                   weight type (examples: f32, f16, q4_0, q4_1, q5_0, q5_1, q8_0, q2_K, q3_K,
+                                           q4_K). If not specified, the default is the type of the weight file
   --rng                                    RNG, one of [std_default, cuda, cpu], default: cuda(sd-webui), cpu(comfyui)
   --sampler-rng                            sampler RNG, one of [std_default, cuda, cpu]. If not specified, use --rng
-  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow, flux2_flow]
-  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is auto. In auto mode, if the model weights
-                                           contain any quantized parameters, the at_runtime mode will be used; otherwise,
-                                           immediately will be used.The immediately mode may have precision and
-                                           compatibility issues with quantized parameters, but it usually offers faster inference
-                                           speed and, in some cases, lower memory usage. The at_runtime mode, on the
-                                           other hand, is exactly the opposite.
-  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
-  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1
-                                           (overrides --vae-tile-size)
+  --prediction                             prediction type override, one of [eps, v, edm_v, sd3_flow, flux_flow,
+                                           flux2_flow]
+  --lora-apply-mode                        the way to apply LoRA, one of [auto, immediately, at_runtime], default is
+                                           auto. In auto mode, if the model weights contain any quantized parameters,
+                                           the at_runtime mode will be used; otherwise, immediately will be used.The
+                                           immediately mode may have precision and compatibility issues with quantized
+                                           parameters, but it usually offers faster inference speed and, in some cases,
+                                           lower memory usage. The at_runtime mode, on the other hand, is exactly the
+                                           opposite.
 
 Default Generation Options:
   -p, --prompt <string>                    the prompt to render
@@ -194,65 +196,99 @@ Default Generation Options:
   --end-img <string>                       path to the end image, required by flf2v
   --mask <string>                          path to the mask image
   --control-image <string>                 path to control image, control net
-  --control-video <string>                 path to control video frames, It must be a directory path. The video frames inside should be stored as images in
-                                           lexicographical (character) order. For example, if the control video path is
-                                           `frames`, the directory contain images such as 00.png, 01.png, ... etc.
+  --control-video <string>                 path to control video frames, It must be a directory path. The video frames
+                                           inside should be stored as images in lexicographical (character) order. For
+                                           example, if the control video path is `frames`, the directory contain images
+                                           such as 00.png, 01.png, ... etc.
   --pm-id-images-dir <string>              path to PHOTOMAKER input id images dir
   --pm-id-embed-path <string>              path to PHOTOMAKER v2 id embed
+  --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 args, key=value list. Currently lcm supports noise_clip_std,
+                                           noise_scale_start, noise_scale_end
   -H, --height <int>                       image height, in pixel space (default: 512)
   -W, --width <int>                        image width, in pixel space (default: 512)
   --steps <int>                            number of sample steps (default: 20)
   --high-noise-steps <int>                 (high noise) number of sample steps (default: -1 = auto)
-  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer (default: -1). <= 0 represents unspecified,
-                                           will be 1 for SD1.x, 2 for SD2.x
+  --clip-skip <int>                        ignore last layers of CLIP network; 1 ignores none, 2 ignores one layer
+                                           (default: -1). <= 0 represents unspecified, will be 1 for SD1.x, 2 for SD2.x
   -b, --batch-count <int>                  batch count
   --video-frames <int>                     video frames (default: 1)
   --fps <int>                              fps (default: 24)
-  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for NitroSD-Realism around 250 and 500 for
-                                           NitroSD-Vibrant
+  --timestep-shift <int>                   shift timestep for NitroFusion models (default: 0). recommended N for
+                                           NitroSD-Realism around 250 and 500 for NitroSD-Vibrant
   --upscale-repeats <int>                  Run the ESRGAN upscaler this many times (default: 1)
   --upscale-tile-size <int>                tile size for ESRGAN upscaling (default: 128)
+  --hires-width <int>                      highres fix target width, 0 to use --hires-scale (default: 0)
+  --hires-height <int>                     highres fix target height, 0 to use --hires-scale (default: 0)
+  --hires-steps <int>                      highres fix second pass sample steps, 0 to reuse --steps (default: 0)
+  --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 instruct-pix2pix models: (default: same
+                                           as --cfg-scale)
   --guidance <float>                       distilled guidance scale for models with guidance input (default: 3.5)
-  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means disabled, a value of 2.5 is nice for sd3.5
-                                           medium
+  --slg-scale <float>                      skip layer guidance (SLG) scale, only for DiT models: (default: 0). 0 means
+                                           disabled, a value of 2.5 is nice for sd3.5 medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
+  --eta <float>                            noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and
+                                           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-guidance <float>            (high noise) distilled guidance scale for models with guidance input (default: 3.5)
-  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
+  --high-noise-img-cfg-scale <float>       (high noise) image guidance scale for inpaint or instruct-pix2pix models
+                                           (default: same as --cfg-scale)
+  --high-noise-guidance <float>            (high noise) distilled guidance scale for models with guidance input
+                                           (default: 3.5)
+  --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default:
+                                           0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) noise multiplier (default: 0 for ddim_trailing, tcd, res_multistep and res_2s; 1 for euler_a, er_sde and dpm++2s_a)
+  --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>
-  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
-  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if `--high-noise-steps` is set to -1
+  --pm-style-strength <float>              
+  --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full
+                                           destruction of information in init image
+  --moe-boundary <float>                   timestep boundary for Wan2.2 MoE model. (default: 0.875). Only enabled if
+                                           `--high-noise-steps` is set to -1
   --vace-strength <float>                  wan vace strength
-  --increase-ref-index                     automatically increase the indices of references images based on the order they are listed (starting with 1).
+  --vae-tile-overlap <float>               tile overlap for vae tiling, in fraction of tile size (default: 0.5)
+  --hires-scale <float>                    highres fix scale when target size is not set (default: 2.0)
+  --hires-denoising-strength <float>       highres fix second pass denoising strength (default: 0.7)
+  --increase-ref-index                     automatically increase the indices of references images based on the order
+                                           they are listed (starting with 1).
   --disable-auto-resize-ref-image          disable auto resize of ref images
   --disable-image-metadata                 do not embed generation metadata on image files
+  --vae-tiling                             process vae in tiles to reduce memory usage
+  --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, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           tcd, res_multistep, res_2s, er_sde] (default: euler for Flux/SD3/Wan, euler_a
-                                           otherwise)
-  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm,
-                                           ddim_trailing, tcd, res_multistep, res_2s, er_sde] default: euler for Flux/SD3/Wan,
-                                           euler_a otherwise
-  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits, smoothstep, sgm_uniform, simple,
-                                           kl_optimal, lcm, bong_tangent], default: discrete
-  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g., "14.61,7.8,3.5,0.0").
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m,
+                                           dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep, res_2s,
+                                           er_sde, euler_cfg_pp, euler_a_cfg_pp] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+  --high-noise-sampling-method             (high noise) sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a,
+                                           dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing, tcd, res_multistep,
+                                           res_2s, er_sde, euler_cfg_pp, euler_a_cfg_pp] default: euler for Flux/SD3/Wan, euler_a otherwise
+  --scheduler                              denoiser sigma scheduler, one of [discrete, karras, exponential, ays, gits,
+                                           smoothstep, sgm_uniform, simple, kl_optimal, lcm, bong_tangent], default:
+                                           discrete
+  --sigmas                                 custom sigma values for the sampler, comma-separated (e.g.,
+                                           "14.61,7.8,3.5,0.0").
   --skip-layers                            layers to skip for SLG steps (default: [7,8,9])
   --high-noise-skip-layers                 (high noise) layers to skip for SLG steps (default: [7,8,9])
   -r, --ref-image                          reference image for Flux Kontext models (can be used multiple times)
-  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)
+  --cache-mode                             caching method: 'easycache' (DiT), 'ucache' (UNET),
+                                           'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT
+                                           Chebyshev+Taylor forecasting)
   --cache-option                           named cache params (key=value format, comma-separated). easycache/ucache:
-                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=. Examples:
-                                           "threshold=0.25" or "threshold=1.5,reset=0"
-  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g., "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
+                                           threshold=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit:
+                                           Fn=,Bn=,threshold=,warmup=; spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=.
+                                           Examples: "threshold=0.25" or "threshold=1.5,reset=0"
+  --scm-mask                               SCM steps mask for cache-dit: comma-separated 0/1 (e.g.,
+                                           "1,1,1,0,0,1,0,0,1,0") - 1=compute, 0=can cache
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
+  --vae-tile-size                          tile size for vae tiling, format [X]x[Y] (default: 32x32)
+  --vae-relative-tile-size                 relative tile size for vae tiling, format [X]x[Y], in fraction of image size
+                                           if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)
 ```

examples/server/api.md

@@ -38,6 +38,8 @@ Current generation-related endpoints include:
 - `POST /sdapi/v1/txt2img`
 - `POST /sdapi/v1/img2img`
 - `GET /sdapi/v1/loras`
+- `GET /sdapi/v1/upscalers`
+- `GET /sdapi/v1/latent-upscale-modes`
 - `GET /sdapi/v1/samplers`
 - `GET /sdapi/v1/schedulers`
 - `GET /sdapi/v1/sd-models`
@@ -216,6 +218,13 @@ Currently supported request fields:
 | `scheduler` | `string` | Scheduler name |
 | `lora` | `array<object>` | Structured LoRA list |
 | `extra_images` | `array<string>` | Base64 or data URL images |
+| `enable_hr` | `boolean` | Enable highres fix for `txt2img` |
+| `hr_upscaler` | `string` | `Lanczos`, `Nearest`, a latent mode such as `Latent (nearest-exact)`, or an upscaler model name from `/sdapi/v1/upscalers` |
+| `hr_scale` | `number` | Highres scale when resize target is not set |
+| `hr_resize_x` | `integer` | Highres target width, `0` to use scale |
+| `hr_resize_y` | `integer` | Highres target height, `0` to use scale |
+| `hr_steps` | `integer` | Highres second-pass sample steps, `0` to reuse `steps` |
+| `denoising_strength` | `number` | Highres denoising strength for `txt2img` |
 
 Native extension fields:
 
@@ -241,6 +250,8 @@ Currently supported request fields:
 | `inpainting_mask_invert` | `integer` or `boolean` | Treated as invert flag |
 | `denoising_strength` | `number` | Clamped to `0.0..1.0` |
 
+Highres fix fields are currently handled for `txt2img`; `img2img` uses `denoising_strength` as image-to-image strength.
+
 Native extension fields:
 
 - any `sdcpp API` fields embedded through `sd_cpp_extra_args` inside `prompt`
@@ -258,6 +269,8 @@ Response fields:
 Currently exposed:
 
 - `GET /sdapi/v1/loras`
+- `GET /sdapi/v1/upscalers`
+- `GET /sdapi/v1/latent-upscale-modes`
 - `GET /sdapi/v1/samplers`
 - `GET /sdapi/v1/schedulers`
 - `GET /sdapi/v1/sd-models`
@@ -272,6 +285,26 @@ Response fields:
 | `[].name` | `string` | Display name derived from file stem |
 | `[].path` | `string` | Relative path under the configured LoRA directory |
 
+`GET /sdapi/v1/upscalers`
+
+| Field | Type | Notes |
+| --- | --- | --- |
+| `[].name` | `string` | Built-in name or model stem |
+| `[].model_name` | `string \| null` | Model family label for model-backed upscalers |
+| `[].model_path` | `string \| null` | Absolute model path for model-backed upscalers |
+| `[].model_url` | `string \| null` | Currently always null |
+| `[].scale` | `integer` | Currently `4` |
+
+Built-in entries include `None`, `Lanczos`, and `Nearest`. Model-backed entries are scanned from the top level of `--hires-upscalers-dir`; subdirectories are not scanned.
+
+`GET /sdapi/v1/latent-upscale-modes`
+
+| Field | Type | Notes |
+| --- | --- | --- |
+| `[].name` | `string` | WebUI-compatible latent upscale mode name |
+
+Built-in latent modes include `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, and `Latent (bicubic antialiased)`.
+
 `GET /sdapi/v1/samplers`
 
 | Field | Type | Notes |
@@ -388,6 +421,7 @@ Top-level fields:
 | `samplers` | `array<string>` | Available sampling methods |
 | `schedulers` | `array<string>` | Available schedulers |
 | `loras` | `array<object>` | Available LoRA entries |
+| `upscalers` | `array<object>` | Available model-backed highres upscalers |
 | `limits` | `object` | Shared queue and size limits |
 
 `model`
@@ -424,6 +458,14 @@ Shared nested fields:
 | `loras[].name` | `string` |
 | `loras[].path` | `string` |
 
+`upscalers`
+
+| Field | Type | Notes |
+| --- | --- | --- |
+| `upscalers[].name` | `string` | Built-in name or model stem; use this value in `hires.upscaler` |
+
+Built-in entries include `None`, `Lanczos`, `Nearest`, `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, and `Latent (bicubic antialiased)`. Model-backed entries are scanned from the top level of `--hires-upscalers-dir`; subdirectories are not scanned.
+
 `limits`
 
 | Field | Type |
@@ -482,6 +524,15 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `auto_resize_ref_image` | `boolean` |
 | `increase_ref_index` | `boolean` |
 | `control_strength` | `number` |
+| `hires` | `object` |
+| `hires.enabled` | `boolean` |
+| `hires.upscaler` | `string` |
+| `hires.scale` | `number` |
+| `hires.target_width` | `integer` |
+| `hires.target_height` | `integer` |
+| `hires.steps` | `integer` |
+| `hires.denoising_strength` | `number` |
+| `hires.upscale_tile_size` | `integer` |
 
 `vid_gen`-specific default fields:
 
@@ -514,6 +565,7 @@ Fields returned in `features_by_mode.img_gen`:
 - `ref_images`
 - `lora`
 - `vae_tiling`
+- `hires`
 - `cache`
 - `cancel_queued`
 - `cancel_generating`
@@ -625,6 +677,16 @@ Example:
   },
 
   "lora": [],
+  "hires": {
+    "enabled": false,
+    "upscaler": "Latent",
+    "scale": 2.0,
+    "target_width": 0,
+    "target_height": 0,
+    "steps": 0,
+    "denoising_strength": 0.7,
+    "upscale_tile_size": 128
+  },
 
   "vae_tiling_params": {
     "enabled": false,
@@ -729,12 +791,23 @@ Other native fields:
 
 | Field | Type |
 | --- | --- |
+| `hires` | `object` |
+| `hires.enabled` | `boolean` |
+| `hires.upscaler` | `string` |
+| `hires.scale` | `number` |
+| `hires.target_width` | `integer` |
+| `hires.target_height` | `integer` |
+| `hires.steps` | `integer` |
+| `hires.denoising_strength` | `number` |
+| `hires.upscale_tile_size` | `integer` |
 | `vae_tiling_params` | `object` |
 | `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.
+
 HTTP-only output fields:
 
 | Field | Type |

examples/server/main.cpp

@@ -48,7 +48,9 @@ static void parse_args(int argc,
 
     if (!svr_params.resolve_and_validate() ||
         !ctx_params.resolve_and_validate(IMG_GEN) ||
-        !default_gen_params.resolve_and_validate(IMG_GEN, ctx_params.lora_model_dir)) {
+        !default_gen_params.resolve_and_validate(IMG_GEN,
+                                                 ctx_params.lora_model_dir,
+                                                 ctx_params.hires_upscalers_dir)) {
         print_usage(argv[0], options_vec);
         exit(1);
     }
@@ -95,6 +97,8 @@ int main(int argc, const char** argv) {
 
     std::vector<LoraEntry> lora_cache;
     std::mutex lora_mutex;
+    std::vector<UpscalerEntry> upscaler_cache;
+    std::mutex upscaler_mutex;
     AsyncJobManager async_job_manager;
     ServerRuntime runtime = {
         sd_ctx.get(),
@@ -104,6 +108,8 @@ int main(int argc, const char** argv) {
         &default_gen_params,
         &lora_cache,
         &lora_mutex,
+        &upscaler_cache,
+        &upscaler_mutex,
         &async_job_manager,
     };
 
@@ -139,7 +145,7 @@ int main(int argc, const char** argv) {
     register_sdapi_endpoints(svr, runtime);
     register_sdcpp_api_endpoints(svr, runtime);
 
-    LOG_INFO("listening on: %s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
+    LOG_INFO("listening on: http://%s:%d\n", svr_params.listen_ip.c_str(), svr_params.listen_port);
     svr.listen(svr_params.listen_ip, svr_params.listen_port);
 
     {

examples/server/routes_openai.cpp

@@ -70,7 +70,7 @@ static bool build_openai_generation_request(const httplib::Request& req,
     }
 
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid params";
         return false;
     }
@@ -212,7 +212,7 @@ static bool build_openai_edit_request(const httplib::Request& req,
     }
 
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid params";
         return false;
     }

examples/server/routes_sdapi.cpp

@@ -1,6 +1,7 @@
 #include "routes.h"
 
 #include <algorithm>
+#include <cctype>
 #include <cstring>
 #include <regex>
 #include <string_view>
@@ -35,14 +36,20 @@ static fs::path resolve_display_model_path(const ServerRuntime& runtime) {
     return {};
 }
 
+static std::string lower_ascii(std::string value) {
+    std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    return value;
+}
+
 static enum sample_method_t get_sdapi_sample_method(std::string name) {
     enum sample_method_t result = str_to_sample_method(name.c_str());
     if (result != SAMPLE_METHOD_COUNT) {
         return result;
     }
 
-    std::transform(name.begin(), name.end(), name.begin(),
-                   [](unsigned char c) { return static_cast<char>(std::tolower(c)); });
+    name = lower_ascii(name);
     static const std::unordered_map<std::string_view, sample_method_t> hardcoded{
         {"euler a", EULER_A_SAMPLE_METHOD},
         {"k_euler_a", EULER_A_SAMPLE_METHOD},
@@ -60,6 +67,10 @@ static enum sample_method_t get_sdapi_sample_method(std::string name) {
         {"k_res_multistep", RES_MULTISTEP_SAMPLE_METHOD},
         {"res 2s", RES_2S_SAMPLE_METHOD},
         {"k_res_2s", RES_2S_SAMPLE_METHOD},
+        {"euler_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"k_euler_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"euler_a_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
+        {"k_euler_a_cfg_pp", EULER_CFG_PP_SAMPLE_METHOD},
     };
     auto it = hardcoded.find(name);
     return it != hardcoded.end() ? it->second : SAMPLE_METHOD_COUNT;
@@ -114,6 +125,18 @@ static bool build_sdapi_img_gen_request(const json& j,
     request.gen_params.width                          = j.value("width", -1);
     request.gen_params.height                         = j.value("height", -1);
 
+    if (!img2img && j.value("enable_hr", false)) {
+        request.gen_params.hires_enabled = true;
+        request.gen_params.hires_scale   = j.value("hr_scale", request.gen_params.hires_scale);
+        request.gen_params.hires_width   = j.value("hr_resize_x", request.gen_params.hires_width);
+        request.gen_params.hires_height  = j.value("hr_resize_y", request.gen_params.hires_height);
+        request.gen_params.hires_steps   = j.value("hr_steps", request.gen_params.hires_steps);
+        request.gen_params.hires_denoising_strength =
+            j.value("denoising_strength", request.gen_params.hires_denoising_strength);
+
+        request.gen_params.hires_upscaler = j.value("hr_upscaler", request.gen_params.hires_upscaler);
+    }
+
     std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(request.gen_params.prompt);
     if (!sd_cpp_extra_args_str.empty() && !request.gen_params.from_json_str(sd_cpp_extra_args_str)) {
         error_message = "invalid sd_cpp_extra_args";
@@ -228,7 +251,7 @@ static bool build_sdapi_img_gen_request(const json& j,
     }
 
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid params";
         return false;
     }
@@ -347,6 +370,52 @@ void register_sdapi_endpoints(httplib::Server& svr, ServerRuntime& rt) {
         res.set_content(result.dump(), "application/json");
     });
 
+    svr.Get("/sdapi/v1/upscalers", [runtime](const httplib::Request&, httplib::Response& res) {
+        refresh_upscaler_cache(*runtime);
+
+        auto make_builtin = [](const char* name) {
+            json item;
+            item["name"]       = name;
+            item["model_name"] = nullptr;
+            item["model_path"] = nullptr;
+            item["model_url"]  = nullptr;
+            item["scale"]      = 4;
+            return item;
+        };
+
+        json result = json::array();
+        result.push_back(make_builtin("None"));
+        result.push_back(make_builtin("Lanczos"));
+        result.push_back(make_builtin("Nearest"));
+
+        {
+            std::lock_guard<std::mutex> lock(*runtime->upscaler_mutex);
+            for (const auto& e : *runtime->upscaler_cache) {
+                json item;
+                item["name"]       = e.name;
+                item["model_name"] = e.model_name;
+                item["model_path"] = e.fullpath;
+                item["model_url"]  = nullptr;
+                item["scale"]      = e.scale;
+                result.push_back(item);
+            }
+        }
+
+        res.set_content(result.dump(), "application/json");
+    });
+
+    svr.Get("/sdapi/v1/latent-upscale-modes", [](const httplib::Request&, httplib::Response& res) {
+        json result = json::array({
+            {{"name", "Latent"}},
+            {{"name", "Latent (nearest)"}},
+            {{"name", "Latent (nearest-exact)"}},
+            {{"name", "Latent (antialiased)"}},
+            {{"name", "Latent (bicubic)"}},
+            {{"name", "Latent (bicubic antialiased)"}},
+        });
+        res.set_content(result.dump(), "application/json");
+    });
+
     svr.Get("/sdapi/v1/samplers", [runtime](const httplib::Request&, httplib::Response& res) {
         std::vector<std::string> sampler_names;
         sampler_names.push_back("default");

examples/server/routes_sdcpp.cpp

@@ -114,6 +114,17 @@ 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},
+         }},
         {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
         {"cache_mode", defaults.cache_mode},
         {"cache_option", defaults.cache_option},
@@ -157,6 +168,7 @@ static json make_img_gen_features_json() {
         {"ref_images", true},
         {"lora", true},
         {"vae_tiling", true},
+        {"hires", true},
         {"cache", true},
         {"cancel_queued", true},
         {"cancel_generating", false},
@@ -179,6 +191,7 @@ static json make_vid_gen_features_json() {
 
 static json make_capabilities_json(ServerRuntime& runtime) {
     refresh_lora_cache(runtime);
+    refresh_upscaler_cache(runtime);
 
     AsyncJobManager& manager  = *runtime.async_job_manager;
     const auto& defaults      = *runtime.default_gen_params;
@@ -190,6 +203,7 @@ static json make_capabilities_json(ServerRuntime& runtime) {
     json image_output_formats = supported_img_output_formats();
     json video_output_formats = supported_vid_output_formats();
     json available_loras      = json::array();
+    json available_upscalers  = json::array();
     json supported_modes      = json::array();
 
     for (int i = 0; i < SAMPLE_METHOD_COUNT; ++i) {
@@ -210,6 +224,42 @@ static json make_capabilities_json(ServerRuntime& runtime) {
         }
     }
 
+    available_upscalers.push_back({
+        {"name", "None"},
+    });
+    available_upscalers.push_back({
+        {"name", "Lanczos"},
+    });
+    available_upscalers.push_back({
+        {"name", "Nearest"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (nearest)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (nearest-exact)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (antialiased)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (bicubic)"},
+    });
+    available_upscalers.push_back({
+        {"name", "Latent (bicubic antialiased)"},
+    });
+    {
+        std::lock_guard<std::mutex> lock(*runtime.upscaler_mutex);
+        for (const auto& entry : *runtime.upscaler_cache) {
+            available_upscalers.push_back({
+                {"name", entry.name},
+            });
+        }
+    }
+
     if (supports_img) {
         supported_modes.push_back("img_gen");
     }
@@ -284,6 +334,7 @@ static json make_capabilities_json(ServerRuntime& runtime) {
     result["features"]               = top_level_features;
     result["features_by_mode"]       = features_by_mode;
     result["loras"]                  = available_loras;
+    result["upscalers"]              = available_upscalers;
     return result;
 }
 
@@ -307,7 +358,7 @@ static bool parse_img_gen_request(const json& body,
         return false;
     }
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(IMG_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid generation parameters";
         return false;
     }
@@ -334,7 +385,7 @@ static bool parse_vid_gen_request(const json& body,
         return false;
     }
     // Intentionally disable prompt-embedded LoRA tag parsing for server APIs.
-    if (!request.gen_params.resolve_and_validate(VID_GEN, "", true)) {
+    if (!request.gen_params.resolve_and_validate(VID_GEN, "", runtime.ctx_params->hires_upscalers_dir, true)) {
         error_message = "invalid generation parameters";
         return false;
     }

examples/server/runtime.cpp

@@ -1,6 +1,7 @@
 #include "runtime.h"
 
 #include <algorithm>
+#include <cctype>
 #include <chrono>
 #include <cstdlib>
 #include <filesystem>
@@ -13,6 +14,18 @@
 
 namespace fs = std::filesystem;
 
+static std::string lower_ascii(std::string value) {
+    std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    return value;
+}
+
+static bool is_supported_model_ext(const fs::path& p) {
+    auto ext = lower_ascii(p.extension().string());
+    return ext == ".gguf" || ext == ".pt" || ext == ".pth" || ext == ".safetensors";
+}
+
 static const std::string k_base64_chars =
     "ABCDEFGHIJKLMNOPQRSTUVWXYZ"
     "abcdefghijklmnopqrstuvwxyz"
@@ -241,20 +254,12 @@ 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)) {
-        auto is_lora_ext = [](const fs::path& p) {
-            auto ext = p.extension().string();
-            std::transform(ext.begin(), ext.end(), ext.begin(), [](unsigned char c) {
-                return static_cast<char>(std::tolower(c));
-            });
-            return ext == ".gguf" || ext == ".pt" || ext == ".pth" || ext == ".safetensors";
-        };
-
         for (auto& entry : fs::recursive_directory_iterator(lora_dir)) {
             if (!entry.is_regular_file()) {
                 continue;
             }
             const fs::path& p = entry.path();
-            if (!is_lora_ext(p)) {
+            if (!is_supported_model_ext(p)) {
                 continue;
             }
 
@@ -286,6 +291,40 @@ std::string get_lora_full_path(ServerRuntime& rt, const std::string& path) {
     return it != rt.lora_cache->end() ? it->fullpath : "";
 }
 
+void refresh_upscaler_cache(ServerRuntime& rt) {
+    std::vector<UpscalerEntry> new_cache;
+
+    fs::path upscaler_dir = rt.ctx_params->hires_upscalers_dir;
+    if (fs::exists(upscaler_dir) && fs::is_directory(upscaler_dir)) {
+        for (auto& entry : fs::directory_iterator(upscaler_dir)) {
+            if (!entry.is_regular_file()) {
+                continue;
+            }
+            const fs::path& p = entry.path();
+            if (!is_supported_model_ext(p)) {
+                continue;
+            }
+
+            UpscalerEntry upscaler_entry;
+            upscaler_entry.name       = p.stem().u8string();
+            upscaler_entry.fullpath   = fs::absolute(p).lexically_normal().u8string();
+            upscaler_entry.model_name = "ESRGAN_4x";
+            upscaler_entry.path       = p.filename().u8string();
+
+            new_cache.push_back(std::move(upscaler_entry));
+        }
+    }
+
+    std::sort(new_cache.begin(), new_cache.end(), [](const UpscalerEntry& a, const UpscalerEntry& b) {
+        return a.name < b.name;
+    });
+
+    {
+        std::lock_guard<std::mutex> lock(*rt.upscaler_mutex);
+        *rt.upscaler_cache = std::move(new_cache);
+    }
+}
+
 int64_t unix_timestamp_now() {
     return std::chrono::duration_cast<std::chrono::seconds>(
                std::chrono::system_clock::now().time_since_epoch())

examples/server/runtime.h

@@ -37,6 +37,14 @@ struct LoraEntry {
     std::string fullpath;
 };
 
+struct UpscalerEntry {
+    std::string name;
+    std::string path;
+    std::string fullpath;
+    std::string model_name;
+    int scale = 4;
+};
+
 struct ServerRuntime {
     sd_ctx_t* sd_ctx;
     std::mutex* sd_ctx_mutex;
@@ -45,6 +53,8 @@ struct ServerRuntime {
     const SDGenerationParams* default_gen_params;
     std::vector<LoraEntry>* lora_cache;
     std::mutex* lora_mutex;
+    std::vector<UpscalerEntry>* upscaler_cache;
+    std::mutex* upscaler_mutex;
     AsyncJobManager* async_job_manager;
 };
 
@@ -86,4 +96,5 @@ bool runtime_supports_generation_mode(const ServerRuntime& runtime, SDMode mode)
 std::string unsupported_generation_mode_error(SDMode mode);
 void refresh_lora_cache(ServerRuntime& rt);
 std::string get_lora_full_path(ServerRuntime& rt, const std::string& path);
+void refresh_upscaler_cache(ServerRuntime& rt);
 int64_t unix_timestamp_now();

format-code.sh

@@ -1,5 +1,5 @@
 for f in src/*.cpp src/*.h src/*.hpp src/tokenizers/*.h src/tokenizers/*.cpp src/tokenizers/vocab/*.h src/tokenizers/vocab/*.cpp \
-         examples/cli/*.cpp examples/cli/*.h examples/server/*.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
   [[ "$f" == vocab* ]] && continue
   echo "formatting '$f'"

include/stable-diffusion.h

@@ -51,6 +51,8 @@ enum sample_method_t {
     RES_MULTISTEP_SAMPLE_METHOD,
     RES_2S_SAMPLE_METHOD,
     ER_SDE_SAMPLE_METHOD,
+    EULER_CFG_PP_SAMPLE_METHOD,
+    EULER_A_CFG_PP_SAMPLE_METHOD,
     SAMPLE_METHOD_COUNT
 };
 
@@ -203,6 +205,9 @@ typedef struct {
     bool chroma_use_t5_mask;
     int chroma_t5_mask_pad;
     bool qwen_image_zero_cond_t;
+    float max_vram;  // GiB budget for graph-cut segmented param offload (0 = disabled, -1 = auto free VRAM minus 1 GiB)
+    const char* backend;
+    const char* params_backend;
 } sd_ctx_params_t;
 
 typedef struct {
@@ -237,6 +242,7 @@ typedef struct {
     float* custom_sigmas;
     int custom_sigmas_count;
     float flow_shift;
+    const char* extra_sample_args;
 } sd_sample_params_t;
 
 typedef struct {
@@ -289,6 +295,32 @@ typedef struct {
     const char* path;
 } sd_lora_t;
 
+enum sd_hires_upscaler_t {
+    SD_HIRES_UPSCALER_NONE,
+    SD_HIRES_UPSCALER_LATENT,
+    SD_HIRES_UPSCALER_LATENT_NEAREST,
+    SD_HIRES_UPSCALER_LATENT_NEAREST_EXACT,
+    SD_HIRES_UPSCALER_LATENT_ANTIALIASED,
+    SD_HIRES_UPSCALER_LATENT_BICUBIC,
+    SD_HIRES_UPSCALER_LATENT_BICUBIC_ANTIALIASED,
+    SD_HIRES_UPSCALER_LANCZOS,
+    SD_HIRES_UPSCALER_NEAREST,
+    SD_HIRES_UPSCALER_MODEL,
+    SD_HIRES_UPSCALER_COUNT,
+};
+
+typedef struct {
+    bool enabled;
+    enum sd_hires_upscaler_t upscaler;
+    const char* model_path;
+    float scale;
+    int target_width;
+    int target_height;
+    int steps;
+    float denoising_strength;
+    int upscale_tile_size;
+} sd_hires_params_t;
+
 typedef struct {
     const sd_lora_t* loras;
     uint32_t lora_count;
@@ -312,6 +344,7 @@ typedef struct {
     sd_pm_params_t pm_params;
     sd_tiling_params_t vae_tiling_params;
     sd_cache_params_t cache;
+    sd_hires_params_t hires;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -365,8 +398,11 @@ SD_API const char* sd_preview_name(enum preview_t preview);
 SD_API enum preview_t str_to_preview(const char* str);
 SD_API const char* sd_lora_apply_mode_name(enum lora_apply_mode_t mode);
 SD_API enum lora_apply_mode_t str_to_lora_apply_mode(const char* str);
+SD_API const char* sd_hires_upscaler_name(enum sd_hires_upscaler_t upscaler);
+SD_API enum sd_hires_upscaler_t str_to_sd_hires_upscaler(const char* str);
 
 SD_API void sd_cache_params_init(sd_cache_params_t* cache_params);
+SD_API void sd_hires_params_init(sd_hires_params_t* hires_params);
 
 SD_API void sd_ctx_params_init(sd_ctx_params_t* sd_ctx_params);
 SD_API char* sd_ctx_params_to_str(const sd_ctx_params_t* sd_ctx_params);
@@ -393,7 +429,9 @@ SD_API upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path,
                                         bool offload_params_to_cpu,
                                         bool direct,
                                         int n_threads,
-                                        int tile_size);
+                                        int tile_size,
+                                        const char* backend,
+                                        const char* params_backend);
 SD_API void free_upscaler_ctx(upscaler_ctx_t* upscaler_ctx);
 
 SD_API sd_image_t upscale(upscaler_ctx_t* upscaler_ctx,

src/anima.hpp

@@ -499,9 +499,15 @@ namespace Anima {
                 encoder_hidden_states = adapted_context;
             }
 
+            sd::ggml_graph_cut::mark_graph_cut(x, "anima.prelude", "x");
+            sd::ggml_graph_cut::mark_graph_cut(embedded_timestep, "anima.prelude", "embedded_timestep");
+            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++) {
                 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");
             }
 
             x = final_layer->forward(ctx, x, embedded_timestep, temb);  // [N, h*w, ph*pw*C]
@@ -520,10 +526,10 @@ namespace Anima {
         AnimaNet net;
 
         AnimaRunner(ggml_backend_t backend,
-                    bool offload_params_to_cpu,
+                    ggml_backend_t params_backend,
                     const String2TensorStorage& tensor_storage_map = {},
                     const std::string prefix                       = "model.diffusion_model")
-            : GGMLRunner(backend, offload_params_to_cpu) {
+            : GGMLRunner(backend, params_backend) {
             int64_t num_layers    = 0;
             std::string layer_tag = prefix + ".net.blocks.";
             for (const auto& kv : tensor_storage_map) {

src/auto_encoder_kl.hpp

@@ -328,6 +328,7 @@ public:
         auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
 
         auto h = conv_in->forward(ctx, x);  // [N, ch, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.prelude", "h");
 
         // downsampling
         size_t num_resolutions = ch_mult.size();
@@ -337,12 +338,14 @@ public:
                 auto down_block  = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
 
                 h = down_block->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.down." + std::to_string(i) + ".block." + std::to_string(j), "h");
             }
             if (i != num_resolutions - 1) {
                 std::string name = "down." + std::to_string(i) + ".downsample";
                 auto down_sample = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
 
                 h = down_sample->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.down." + std::to_string(i) + ".downsample", "h");
             }
         }
 
@@ -350,6 +353,7 @@ public:
         h = mid_block_1->forward(ctx, h);
         h = mid_attn_1->forward(ctx, h);
         h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.encoder.mid", "h");
 
         // end
         h = norm_out->forward(ctx, h);
@@ -450,6 +454,7 @@ public:
 
         // conv_in
         auto h = conv_in->forward(ctx, z);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.prelude", "h");
 
         // middle
         h = mid_block_1->forward(ctx, h);
@@ -457,6 +462,7 @@ public:
 
         h = mid_attn_1->forward(ctx, h);
         h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+        // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.mid", "h");
 
         // upsampling
         int num_resolutions = static_cast<int>(ch_mult.size());
@@ -466,12 +472,14 @@ public:
                 auto up_block    = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
 
                 h = up_block->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.up." + std::to_string(i) + ".block." + std::to_string(j), "h");
             }
             if (i != 0) {
                 std::string name = "up." + std::to_string(i) + ".upsample";
                 auto up_sample   = std::dynamic_pointer_cast<UpSampleBlock>(blocks[name]);
 
                 h = up_sample->forward(ctx, h);
+                // sd::ggml_graph_cut::mark_graph_cut(h, "vae.decoder.up." + std::to_string(i) + ".upsample", "h");
             }
         }
 
@@ -599,6 +607,7 @@ public:
         if (use_quant) {
             auto post_quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["post_quant_conv"]);
             z                    = post_quant_conv->forward(ctx, z);  // [N, z_channels, h, w]
+            // sd::ggml_graph_cut::mark_graph_cut(z, "vae.decode.prelude", "z");
         }
         auto decoder = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
 
@@ -616,6 +625,7 @@ public:
         if (use_quant) {
             auto quant_conv = std::dynamic_pointer_cast<Conv2d>(blocks["quant_conv"]);
             z               = quant_conv->forward(ctx, z);  // [N, 2*embed_dim, h/8, w/8]
+            // sd::ggml_graph_cut::mark_graph_cut(z, "vae.encode.final", "z");
         }
         if (sd_version_uses_flux2_vae(version)) {
             z = ggml_ext_chunk(ctx->ggml_ctx, z, 2, 2)[0];
@@ -654,13 +664,13 @@ struct AutoEncoderKL : public VAE {
     AutoEncoderKLModel ae;
 
     AutoEncoderKL(ggml_backend_t backend,
-                  bool offload_params_to_cpu,
+                  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, offload_params_to_cpu) {
+        : decode_only(decode_only), VAE(version, backend, params_backend) {
         if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
             scale_factor = 0.18215f;
             shift_factor = 0.f;

src/clip.hpp

@@ -95,8 +95,9 @@ public:
 
     ggml_tensor* forward(GGMLRunnerContext* ctx,
                          ggml_tensor* x,
-                         ggml_tensor* mask = nullptr,
-                         int clip_skip     = -1) {
+                         ggml_tensor* mask                   = nullptr,
+                         int clip_skip                       = -1,
+                         const std::string& graph_cut_prefix = "") {
         // x: [N, n_token, d_model]
         int layer_idx = n_layer - 1;
         // LOG_DEBUG("clip_skip %d", clip_skip);
@@ -112,6 +113,9 @@ public:
             std::string name = "layers." + std::to_string(i);
             auto layer       = std::dynamic_pointer_cast<CLIPLayer>(blocks[name]);
             x                = layer->forward(ctx, x, mask);  // [N, n_token, d_model]
+            if (!graph_cut_prefix.empty()) {
+                sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".layers." + std::to_string(i), "x");
+            }
             // LOG_DEBUG("layer %d", i);
         }
         return x;
@@ -304,7 +308,8 @@ public:
         auto final_layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["final_layer_norm"]);
 
         auto x = embeddings->forward(ctx, input_ids, tkn_embeddings);  // [N, n_token, hidden_size]
-        x      = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip);
+        sd::ggml_graph_cut::mark_graph_cut(x, "clip_text.prelude", "x");
+        x = encoder->forward(ctx, x, mask, return_pooled ? -1 : clip_skip, "clip_text");
         if (return_pooled || with_final_ln) {
             x = final_layer_norm->forward(ctx, x);
         }
@@ -368,7 +373,8 @@ public:
 
         auto x = embeddings->forward(ctx, pixel_values);  // [N, num_positions, embed_dim]
         x      = pre_layernorm->forward(ctx, x);
-        x      = encoder->forward(ctx, x, nullptr, clip_skip);
+        sd::ggml_graph_cut::mark_graph_cut(x, "clip_vision.prelude", "x");
+        x = encoder->forward(ctx, x, nullptr, clip_skip, "clip_vision");
 
         auto last_hidden_state = x;
 
@@ -463,13 +469,13 @@ struct CLIPTextModelRunner : public GGMLRunner {
     std::vector<float> attention_mask_vec;
 
     CLIPTextModelRunner(ggml_backend_t backend,
-                        bool offload_params_to_cpu,
+                        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, offload_params_to_cpu) {
+        : GGMLRunner(backend, params_backend) {
         bool proj_in = false;
         for (const auto& [name, tensor_storage] : tensor_storage_map) {
             if (!starts_with(name, prefix)) {

src/common_block.hpp

@@ -1,7 +1,9 @@
 #ifndef __COMMON_BLOCK_HPP__
 #define __COMMON_BLOCK_HPP__
 
+#include "ggml-backend.h"
 #include "ggml_extend.hpp"
+#include "util.h"
 
 class DownSampleBlock : public GGMLBlock {
 protected:
@@ -248,9 +250,6 @@ public:
         float scale         = 1.f;
         if (precision_fix) {
             scale = 1.f / 128.f;
-#ifdef SD_USE_VULKAN
-            force_prec_f32 = true;
-#endif
         }
         // The purpose of the scale here is to prevent NaN issues in certain situations.
         // For example, when using Vulkan without enabling force_prec_f32,
@@ -264,6 +263,9 @@ public:
 
         auto net_0 = std::dynamic_pointer_cast<UnaryBlock>(blocks["net.0"]);
         auto net_2 = std::dynamic_pointer_cast<Linear>(blocks["net.2"]);
+        if (sd_backend_is(ctx->backend, "Vulkan")) {
+            net_2->set_force_prec_f32(true);
+        }
 
         x = net_0->forward(ctx, x);  // [ne3, ne2, ne1, inner_dim]
         x = net_2->forward(ctx, x);  // [ne3, ne2, ne1, dim_out]

src/conditioner.hpp

@@ -14,6 +14,12 @@ struct SDCondition {
     sd::Tensor<float> c_concat;
     sd::Tensor<int32_t> c_t5_ids;
     sd::Tensor<float> c_t5_weights;
+    sd::Tensor<int32_t> c_input_ids;
+    sd::Tensor<int32_t> c_position_ids;
+    sd::Tensor<int32_t> c_token_types;
+    sd::Tensor<int32_t> c_vinput_mask;
+    std::vector<std::pair<int, sd::Tensor<float>>> c_image_embeds;
+    std::vector<sd::Tensor<float>> c_ref_images;
 
     std::vector<sd::Tensor<float>> extra_c_crossattns;
 
@@ -26,10 +32,24 @@ struct SDCondition {
 
     bool empty() const {
         if (!c_crossattn.empty() || !c_vector.empty() || !c_concat.empty() ||
-            !c_t5_ids.empty() || !c_t5_weights.empty()) {
+            !c_t5_ids.empty() || !c_t5_weights.empty() ||
+            !c_input_ids.empty() || !c_position_ids.empty() ||
+            !c_token_types.empty() || !c_vinput_mask.empty()) {
             return false;
         }
 
+        for (const auto& image_embed : c_image_embeds) {
+            if (!image_embed.second.empty()) {
+                return false;
+            }
+        }
+
+        for (const auto& tensor : c_ref_images) {
+            if (!tensor.empty()) {
+                return false;
+            }
+        }
+
         for (const auto& tensor : extra_c_crossattns) {
             if (!tensor.empty()) {
                 return false;
@@ -85,7 +105,8 @@ public:
     virtual void free_params_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_flash_attention_enabled(bool enabled)                                 = 0;
+    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes) {}
+    virtual void set_flash_attention_enabled(bool enabled) = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
     virtual std::tuple<SDCondition, std::vector<bool>> get_learned_condition_with_trigger(int n_threads,
                                                                                           const ConditionerParams& conditioner_params) {
@@ -113,7 +134,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
     std::map<std::string, std::pair<int, int>> embedding_pos_map;
 
     FrozenCLIPEmbedderWithCustomWords(ggml_backend_t backend,
-                                      bool offload_params_to_cpu,
+                                      ggml_backend_t params_backend,
                                       const String2TensorStorage& tensor_storage_map,
                                       const std::map<std::string, std::string>& orig_embedding_map,
                                       SDVersion version = VERSION_SD1,
@@ -127,12 +148,12 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         }
         bool force_clip_f32 = !embedding_map.empty();
         if (sd_version_is_sd1(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, true, force_clip_f32);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, true, force_clip_f32);
         } else if (sd_version_is_sd2(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, true, force_clip_f32);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, true, force_clip_f32);
         } else if (sd_version_is_sdxl(version)) {
-            text_model  = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false, force_clip_f32);
-            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false, force_clip_f32);
+            text_model  = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false, force_clip_f32);
+            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false, force_clip_f32);
         }
     }
 
@@ -165,6 +186,13 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        text_model->set_max_graph_vram_bytes(max_vram_bytes);
+        if (sd_version_is_sdxl(version)) {
+            text_model2->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         text_model->set_flash_attention_enabled(enabled);
         if (sd_version_is_sdxl(version)) {
@@ -642,9 +670,9 @@ struct FrozenCLIPVisionEmbedder : public GGMLRunner {
     CLIPVisionModelProjection vision_model;
 
     FrozenCLIPVisionEmbedder(ggml_backend_t backend,
-                             bool offload_params_to_cpu,
+                             ggml_backend_t params_backend,
                              const String2TensorStorage& tensor_storage_map = {})
-        : GGMLRunner(backend, offload_params_to_cpu) {
+        : GGMLRunner(backend, params_backend) {
         std::string prefix = "cond_stage_model.transformer";
         bool proj_in       = false;
         for (const auto& [name, tensor_storage] : tensor_storage_map) {
@@ -701,7 +729,7 @@ struct SD3CLIPEmbedder : public Conditioner {
     std::shared_ptr<T5Runner> t5;
 
     SD3CLIPEmbedder(ggml_backend_t backend,
-                    bool offload_params_to_cpu,
+                    ggml_backend_t params_backend,
                     const String2TensorStorage& tensor_storage_map = {})
         : clip_g_tokenizer(0) {
         bool use_clip_l = false;
@@ -721,13 +749,13 @@ struct SD3CLIPEmbedder : public Conditioner {
             return;
         }
         if (use_clip_l) {
-            clip_l = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
+            clip_l = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
         }
         if (use_clip_g) {
-            clip_g = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
+            clip_g = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "text_encoders.clip_g.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
         }
         if (use_t5) {
-            t5 = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.t5xxl.transformer");
+            t5 = std::make_shared<T5Runner>(backend, params_backend, tensor_storage_map, "text_encoders.t5xxl.transformer");
         }
     }
 
@@ -781,6 +809,18 @@ struct SD3CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        if (clip_l) {
+            clip_l->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+        if (clip_g) {
+            clip_g->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+        if (t5) {
+            t5->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -1057,7 +1097,7 @@ struct FluxCLIPEmbedder : public Conditioner {
     size_t chunk_len = 256;
 
     FluxCLIPEmbedder(ggml_backend_t backend,
-                     bool offload_params_to_cpu,
+                     ggml_backend_t params_backend,
                      const String2TensorStorage& tensor_storage_map = {}) {
         bool use_clip_l = false;
         bool use_t5     = false;
@@ -1075,12 +1115,12 @@ struct FluxCLIPEmbedder : public Conditioner {
         }
 
         if (use_clip_l) {
-            clip_l = std::make_shared<CLIPTextModelRunner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
+            clip_l = std::make_shared<CLIPTextModelRunner>(backend, params_backend, tensor_storage_map, "text_encoders.clip_l.transformer.text_model", OPENAI_CLIP_VIT_L_14, true);
         } else {
             LOG_WARN("clip_l text encoder not found! Prompt adherence might be degraded.");
         }
         if (use_t5) {
-            t5 = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.t5xxl.transformer");
+            t5 = std::make_shared<T5Runner>(backend, params_backend, tensor_storage_map, "text_encoders.t5xxl.transformer");
         } else {
             LOG_WARN("t5xxl text encoder not found! Prompt adherence might be degraded.");
         }
@@ -1124,6 +1164,15 @@ struct FluxCLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        if (clip_l) {
+            clip_l->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+        if (t5) {
+            t5->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (clip_l) {
             clip_l->set_flash_attention_enabled(enabled);
@@ -1302,7 +1351,7 @@ struct T5CLIPEmbedder : public Conditioner {
     bool is_umt5     = false;
 
     T5CLIPEmbedder(ggml_backend_t backend,
-                   bool offload_params_to_cpu,
+                   ggml_backend_t params_backend,
                    const String2TensorStorage& tensor_storage_map = {},
                    bool use_mask                                  = false,
                    int mask_pad                                   = 0,
@@ -1319,7 +1368,7 @@ struct T5CLIPEmbedder : public Conditioner {
             LOG_WARN("IMPORTANT NOTICE: No text encoders provided, cannot process prompts!");
             return;
         } else {
-            t5 = std::make_shared<T5Runner>(backend, offload_params_to_cpu, tensor_storage_map, "text_encoders.t5xxl.transformer", is_umt5);
+            t5 = std::make_shared<T5Runner>(backend, params_backend, tensor_storage_map, "text_encoders.t5xxl.transformer", is_umt5);
         }
     }
 
@@ -1349,6 +1398,12 @@ struct T5CLIPEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        if (t5) {
+            t5->set_max_graph_vram_bytes(max_vram_bytes);
+        }
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         if (t5) {
             t5->set_flash_attention_enabled(enabled);
@@ -1498,12 +1553,12 @@ struct AnimaConditioner : public Conditioner {
     std::shared_ptr<LLM::LLMRunner> llm;
 
     AnimaConditioner(ggml_backend_t backend,
-                     bool offload_params_to_cpu,
+                     ggml_backend_t params_backend,
                      const String2TensorStorage& tensor_storage_map = {}) {
         qwen_tokenizer = std::make_shared<Qwen2Tokenizer>();
         llm            = std::make_shared<LLM::LLMRunner>(LLM::LLMArch::QWEN3,
                                                backend,
-                                               offload_params_to_cpu,
+                                               params_backend,
                                                tensor_storage_map,
                                                "text_encoders.llm",
                                                false);
@@ -1525,6 +1580,10 @@ struct AnimaConditioner : public Conditioner {
         return llm->get_params_buffer_size();
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        llm->set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }
@@ -1612,7 +1671,7 @@ struct LLMEmbedder : public Conditioner {
     std::shared_ptr<LLM::LLMRunner> llm;
 
     LLMEmbedder(ggml_backend_t backend,
-                bool offload_params_to_cpu,
+                ggml_backend_t params_backend,
                 const String2TensorStorage& tensor_storage_map = {},
                 SDVersion version                              = VERSION_QWEN_IMAGE,
                 const std::string prefix                       = "",
@@ -1633,7 +1692,7 @@ struct LLMEmbedder : public Conditioner {
         }
         llm = std::make_shared<LLM::LLMRunner>(arch,
                                                backend,
-                                               offload_params_to_cpu,
+                                               params_backend,
                                                tensor_storage_map,
                                                "text_encoders.llm",
                                                enable_vision);
@@ -1657,6 +1716,10 @@ struct LLMEmbedder : public Conditioner {
         return buffer_size;
     }
 
+    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
+        llm->set_max_graph_vram_bytes(max_vram_bytes);
+    }
+
     void set_flash_attention_enabled(bool enabled) override {
         llm->set_flash_attention_enabled(enabled);
     }

src/control.hpp

@@ -319,10 +319,10 @@ struct ControlNet : public GGMLRunner {
     bool guided_hint_cached = false;
 
     ControlNet(ggml_backend_t backend,
-               bool offload_params_to_cpu,
+               ggml_backend_t params_backend,
                const String2TensorStorage& tensor_storage_map = {},
                SDVersion version                              = VERSION_SD1)
-        : GGMLRunner(backend, offload_params_to_cpu), control_net(version) {
+        : GGMLRunner(backend, params_backend), control_net(version) {
         control_net.init(params_ctx, tensor_storage_map, "");
     }
 

src/convert.cpp

@@ -0,0 +1,138 @@
+#include <cstring>
+#include <mutex>
+#include <regex>
+#include <vector>
+
+#include "model.h"
+#include "model_io/gguf_io.h"
+#include "model_io/safetensors_io.h"
+#include "util.h"
+
+#include "ggml-cpu.h"
+
+static ggml_type get_export_tensor_type(ModelLoader& model_loader,
+                                        const TensorStorage& tensor_storage,
+                                        ggml_type type,
+                                        const TensorTypeRules& tensor_type_rules) {
+    const std::string& name = tensor_storage.name;
+    ggml_type tensor_type   = tensor_storage.type;
+    ggml_type dst_type      = type;
+
+    for (const auto& tensor_type_rule : tensor_type_rules) {
+        std::regex pattern(tensor_type_rule.first);
+        if (std::regex_search(name, pattern)) {
+            dst_type = tensor_type_rule.second;
+            break;
+        }
+    }
+
+    if (model_loader.tensor_should_be_converted(tensor_storage, dst_type)) {
+        tensor_type = dst_type;
+    }
+
+    return tensor_type;
+}
+
+static bool load_tensors_for_export(ModelLoader& model_loader,
+                                    ggml_context* ggml_ctx,
+                                    ggml_type type,
+                                    const TensorTypeRules& tensor_type_rules,
+                                    std::vector<TensorWriteInfo>& tensors) {
+    std::mutex tensor_mutex;
+    auto on_new_tensor_cb = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) -> bool {
+        const std::string& name = tensor_storage.name;
+        ggml_type tensor_type   = get_export_tensor_type(model_loader, tensor_storage, type, tensor_type_rules);
+
+        std::lock_guard<std::mutex> lock(tensor_mutex);
+        ggml_tensor* tensor = ggml_new_tensor(ggml_ctx, tensor_type, tensor_storage.n_dims, tensor_storage.ne);
+        if (tensor == nullptr) {
+            LOG_ERROR("ggml_new_tensor failed");
+            return false;
+        }
+        ggml_set_name(tensor, name.c_str());
+
+        if (!tensor->data) {
+            GGML_ASSERT(ggml_nelements(tensor) == 0);
+            // Avoid crashing writers by setting a dummy pointer for zero-sized tensors.
+            LOG_DEBUG("setting dummy pointer for zero-sized tensor %s", name.c_str());
+            tensor->data = ggml_get_mem_buffer(ggml_ctx);
+        }
+
+        TensorWriteInfo write_info;
+        write_info.tensor = tensor;
+        write_info.n_dims = tensor_storage.n_dims;
+        for (int i = 0; i < tensor_storage.n_dims; ++i) {
+            write_info.ne[i] = tensor_storage.ne[i];
+        }
+
+        *dst_tensor = tensor;
+        tensors.push_back(std::move(write_info));
+
+        return true;
+    };
+
+    bool success = model_loader.load_tensors(on_new_tensor_cb);
+    LOG_INFO("load tensors done");
+    return success;
+}
+
+bool convert(const char* input_path,
+             const char* vae_path,
+             const char* output_path,
+             sd_type_t output_type,
+             const char* tensor_type_rules,
+             bool convert_name) {
+    ModelLoader model_loader;
+
+    if (!model_loader.init_from_file(input_path)) {
+        LOG_ERROR("init model loader from file failed: '%s'", input_path);
+        return false;
+    }
+
+    if (vae_path != nullptr && strlen(vae_path) > 0) {
+        if (!model_loader.init_from_file(vae_path, "vae.")) {
+            LOG_ERROR("init model loader from file failed: '%s'", vae_path);
+            return false;
+        }
+    }
+    if (convert_name) {
+        model_loader.convert_tensors_name();
+    }
+
+    ggml_type type             = (ggml_type)output_type;
+    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();
+    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);
+    LOG_INFO("model tensors mem size: %.2fMB", mem_size / 1024.f / 1024.f);
+    ggml_context* ggml_ctx = ggml_init({mem_size, nullptr, false});
+
+    if (ggml_ctx == nullptr) {
+        LOG_ERROR("ggml_init failed for converter");
+        ggml_backend_free(backend);
+        return false;
+    }
+
+    std::vector<TensorWriteInfo> tensors;
+    bool success = load_tensors_for_export(model_loader, ggml_ctx, type, type_rules, tensors);
+    ggml_backend_free(backend);
+
+    std::string error;
+    if (success) {
+        if (output_is_safetensors) {
+            success = write_safetensors_file(output_path, tensors, &error);
+        } else {
+            success = write_gguf_file(output_path, tensors, &error);
+        }
+    }
+
+    if (!success && !error.empty()) {
+        LOG_ERROR("%s", error.c_str());
+    }
+
+    ggml_free(ggml_ctx);
+    return success;
+}

src/denoiser.hpp

@@ -2,6 +2,7 @@
 #define __DENOISER_HPP__
 
 #include <cmath>
+#include <string>
 #include <utility>
 
 #include "ggml_extend.hpp"
@@ -752,7 +753,7 @@ struct Flux2FlowDenoiser : public FluxFlowDenoiser {
     }
 };
 
-typedef std::function<sd::Tensor<float>(const sd::Tensor<float>&, float, int)> denoise_cb_t;
+typedef std::function<sd::Tensor<float>(const sd::Tensor<float>&, float, int, sd::Tensor<float>*)> denoise_cb_t;
 
 static std::pair<float, float> get_ancestral_step(float sigma_from,
                                                   float sigma_to,
@@ -808,48 +809,48 @@ static std::tuple<float, float, float> get_ancestral_step_flow(float sigma_from,
     return {sigma_down, sigma_up, alpha_scale};
 }
 
+static std::tuple<float, float, float> get_ancestral_step(float sigma_from,
+                                                          float sigma_to,
+                                                          float eta,
+                                                          bool is_flow_denoiser) {
+    if (is_flow_denoiser) {
+        return get_ancestral_step_flow(sigma_from, sigma_to, eta);
+    } else {
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
+        return {sigma_down, sigma_up, 1.0f};
+    }
+}
+
 static sd::Tensor<float> sample_euler_ancestral(denoise_cb_t model,
                                                 sd::Tensor<float> x,
                                                 const std::vector<float>& sigmas,
-                                                std::shared_ptr<RNG> rng,
-                                                float eta) {
+                                                std::shared_ptr<RNG> rng = nullptr,
+                                                bool is_flow_denoiser    = false,
+                                                float eta                = 0.f) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1);
+        float sigma_to    = sigmas[i + 1];
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
-        sd::Tensor<float> denoised  = std::move(denoised_opt);
-        sd::Tensor<float> d         = (x - denoised) / sigma;
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
-        x += d * (sigma_down - sigmas[i]);
-        if (sigmas[i + 1] > 0) {
-            x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
-        }
-    }
-    return x;
-}
-
-static sd::Tensor<float> sample_euler_flow(denoise_cb_t model,
-                                           sd::Tensor<float> x,
-                                           const std::vector<float>& sigmas,
-                                           std::shared_ptr<RNG> rng,
-                                           float eta) {
-    int steps = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1);
-        if (denoised_opt.empty()) {
-            return {};
-        }
-        sd::Tensor<float> denoised               = std::move(denoised_opt);
-        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step_flow(sigma, sigmas[i + 1], eta);
-        float sigma_ratio                        = sigma_down / sigma;
-        x                                        = sigma_ratio * x + (1.0f - sigma_ratio) * denoised;
-
-        if (sigma_up > 0.0f) {
-            x = alpha_scale * x + sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+        if (sigma_to == 0.f) {
+            x = denoised;
+        } else if (eta == 0.f) {
+            float sigma_ratio = sigma_to / sigma;
+            x                 = sigma_ratio * x + (1.0 - sigma_ratio) * denoised;
+        } else {
+            auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma, sigma_to, eta, is_flow_denoiser);
+            float sigma_ratio                        = sigma_down / sigma;
+            x                                        = sigma_ratio * x + (1.0f - sigma_ratio) * denoised;
+            if (sigma_up > 0.f) {
+                if (is_flow_denoiser) {
+                    x *= alpha_scale;
+                }
+                x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
+            }
         }
     }
     return x;
@@ -861,7 +862,7 @@ static sd::Tensor<float> sample_euler(denoise_cb_t model,
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
         float sigma       = sigmas[i];
-        auto denoised_opt = model(x, sigma, i + 1);
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -877,7 +878,7 @@ static sd::Tensor<float> sample_heun(denoise_cb_t model,
                                      const std::vector<float>& sigmas) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1));
+        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -888,7 +889,7 @@ static sd::Tensor<float> sample_heun(denoise_cb_t model,
             x += d * dt;
         } else {
             sd::Tensor<float> x2 = x + d * dt;
-            auto denoised2_opt   = model(x2, sigmas[i + 1], i + 1);
+            auto denoised2_opt   = model(x2, sigmas[i + 1], i + 1, nullptr);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -905,7 +906,7 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
                                      const std::vector<float>& sigmas) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1));
+        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -918,7 +919,7 @@ static sd::Tensor<float> sample_dpm2(denoise_cb_t model,
             float dt_1           = sigma_mid - sigmas[i];
             float dt_2           = sigmas[i + 1] - sigmas[i];
             sd::Tensor<float> x2 = x + d * dt_1;
-            auto denoised2_opt   = model(x2, sigma_mid, i + 1);
+            auto denoised2_opt   = model(x2, sigma_mid, i + 1, nullptr);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -939,7 +940,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], -(i + 1));
+        auto denoised_opt = model(x, sigmas[i], -(i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -955,7 +956,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral(denoise_cb_t model,
             float s              = t + 0.5f * h;
             float sigma_s        = sigma_fn(s);
             sd::Tensor<float> x2 = (sigma_s / sigma_fn(t)) * x - (exp(-h * 0.5f) - 1) * denoised;
-            auto denoised2_opt   = model(x2, sigma_s, i + 1);
+            auto denoised2_opt   = model(x2, sigma_s, i + 1, nullptr);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -977,12 +978,12 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
                                                         float eta = 1.0f) {
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        float sigma = sigmas[i];
+        float sigma    = sigmas[i];
         float sigma_to = sigmas[i + 1];
 
         bool opt_first_step = (1.0 - sigma < 1e-6);
 
-        auto denoised_opt = model(x, sigma, (opt_first_step ? 1 : -1) * (i + 1));
+        auto denoised_opt = model(x, sigma, (opt_first_step ? 1 : -1) * (i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1011,8 +1012,8 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
                 // so sigma_s = 1 = sigma, and sigma_s_i_ratio = sigma_s / sigma = 1
                 // u = (x*sigma_s_i_ratio)+(denoised*(1.0f-sigma_s_i_ratio))
                 //   = (x*1)+(denoised*0) = x
-                // so D_i = model(u, sigma_s, i + 1)
-                //        = model(x, sigma,   i + 1)
+                // so D_i = model(u, sigma_s, i + 1, nullptr)
+                //        = model(x, sigma,   i + 1, nullptr)
                 //        = denoised
                 D_i = denoised;
 
@@ -1040,12 +1041,12 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
                 // and sigma_s = sigma_fn(s) = 1.0f / (exp(s) + 1.0f)
 
                 float exp_s = std::sqrt(((1 - sigma) / sigma) * ((1 - sigma_down) / sigma_down));
-                sigma_s = 1.0f / (exp_s + 1.0f);
+                sigma_s     = 1.0f / (exp_s + 1.0f);
 
                 float sigma_s_i_ratio = sigma_s / sigma;
-                sd::Tensor<float> u = (x * sigma_s_i_ratio) + (denoised * (1.0f - sigma_s_i_ratio));
+                sd::Tensor<float> u   = (x * sigma_s_i_ratio) + (denoised * (1.0f - sigma_s_i_ratio));
 
-                auto denoised2_opt = model(u, sigma_s, i + 1);
+                auto denoised2_opt = model(u, sigma_s, i + 1, nullptr);
                 if (denoised2_opt.empty()) {
                     return {};
                 }
@@ -1053,7 +1054,7 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
             }
 
             float sigma_down_i_ratio = sigma_down / sigma;
-            x = (x * sigma_down_i_ratio) + (D_i * (1.0f - sigma_down_i_ratio));
+            x                        = (x * sigma_down_i_ratio) + (D_i * (1.0f - sigma_down_i_ratio));
 
             if (sigma_to > 0.0f && eta > 0.0f) {
                 x = alpha_scale * x + sd::Tensor<float>::randn_like(x, rng) * sigma_up;
@@ -1064,8 +1065,6 @@ static sd::Tensor<float> sample_dpmpp_2s_ancestral_flow(denoise_cb_t model,
     return x;
 }
 
-
-
 static sd::Tensor<float> sample_dpmpp_2m(denoise_cb_t model,
                                          sd::Tensor<float> x,
                                          const std::vector<float>& sigmas) {
@@ -1074,7 +1073,7 @@ static sd::Tensor<float> sample_dpmpp_2m(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1);
+        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1106,7 +1105,7 @@ static sd::Tensor<float> sample_dpmpp_2m_v2(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1);
+        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1138,10 +1137,83 @@ static sd::Tensor<float> sample_lcm(denoise_cb_t model,
                                     sd::Tensor<float> x,
                                     const std::vector<float>& sigmas,
                                     std::shared_ptr<RNG> rng,
-                                    bool is_flow_denoiser) {
+                                    bool is_flow_denoiser,
+                                    const char* extra_sample_args = nullptr) {
+    struct LCMSampleArgs {
+        float noise_clip_std    = 0.0f;
+        float noise_scale_start = 1.0f;
+        float noise_scale_end   = 1.0f;
+    };
+
+    auto trim = [](std::string value) -> std::string {
+        const char* whitespace = " \t\r\n";
+        size_t begin           = value.find_first_not_of(whitespace);
+        if (begin == std::string::npos) {
+            return "";
+        }
+        size_t end = value.find_last_not_of(whitespace);
+        return value.substr(begin, end - begin + 1);
+    };
+
+    LCMSampleArgs args;
+    if (extra_sample_args != nullptr && extra_sample_args[0] != '\0') {
+        std::string raw(extra_sample_args);
+        size_t start                   = 0;
+        bool noise_scale_end_was_set   = false;
+        bool noise_scale_start_was_set = false;
+        auto parse_arg                 = [&](const std::string& item) {
+            std::string token = trim(item);
+            if (token.empty()) {
+                return;
+            }
+            size_t eq = token.find('=');
+            if (eq == std::string::npos) {
+                LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
+                return;
+            }
+
+            std::string key   = trim(token.substr(0, eq));
+            std::string value = trim(token.substr(eq + 1));
+            float parsed      = 0.0f;
+            try {
+                size_t consumed = 0;
+                parsed          = std::stof(value, &consumed);
+                if (trim(value.substr(consumed)).size() != 0) {
+                    LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
+                    return;
+                }
+            } catch (const std::exception&) {
+                LOG_WARN("ignoring invalid lcm extra sample arg '%s'", token.c_str());
+                return;
+            }
+
+            if (key == "noise_clip_std") {
+                args.noise_clip_std = parsed;
+            } else if (key == "noise_scale_start") {
+                args.noise_scale_start    = parsed;
+                noise_scale_start_was_set = true;
+            } else if (key == "noise_scale_end") {
+                args.noise_scale_end    = parsed;
+                noise_scale_end_was_set = true;
+            } else {
+                LOG_WARN("ignoring unknown lcm extra sample arg '%s'", key.c_str());
+            }
+        };
+
+        for (size_t pos = 0; pos <= raw.size(); ++pos) {
+            if (pos == raw.size() || raw[pos] == ',' || raw[pos] == ';') {
+                parse_arg(raw.substr(start, pos - start));
+                start = pos + 1;
+            }
+        }
+        if (noise_scale_start_was_set && !noise_scale_end_was_set) {
+            args.noise_scale_end = args.noise_scale_start;
+        }
+    }
+
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1);
+        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1150,7 +1222,27 @@ static sd::Tensor<float> sample_lcm(denoise_cb_t model,
             if (is_flow_denoiser) {
                 x *= (1 - sigmas[i + 1]);
             }
-            x += sd::Tensor<float>::randn_like(x, rng) * sigmas[i + 1];
+            auto noise = sd::Tensor<float>::randn_like(x, rng);
+            if (args.noise_clip_std > 0.0f && noise.numel() > 0) {
+                double mean = 0.0;
+                for (int64_t j = 0; j < noise.numel(); ++j) {
+                    mean += static_cast<double>(noise[j]);
+                }
+                mean /= static_cast<double>(noise.numel());
+
+                double variance = 0.0;
+                for (int64_t j = 0; j < noise.numel(); ++j) {
+                    double centered = static_cast<double>(noise[j]) - mean;
+                    variance += centered * centered;
+                }
+                variance /= static_cast<double>(noise.numel());
+
+                float clip_val = args.noise_clip_std * static_cast<float>(std::sqrt(variance));
+                noise          = sd::ops::clamp(noise, -clip_val, clip_val);
+            }
+            float t           = steps > 1 ? static_cast<float>(i) / static_cast<float>(steps - 1) : 0.0f;
+            float noise_scale = args.noise_scale_start + (args.noise_scale_end - args.noise_scale_start) * t;
+            x += noise * (sigmas[i + 1] * noise_scale);
         }
     }
     return x;
@@ -1167,7 +1259,7 @@ static sd::Tensor<float> sample_ipndm(denoise_cb_t model,
         float sigma      = sigmas[i];
         float sigma_next = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma, i + 1);
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1211,7 +1303,7 @@ static sd::Tensor<float> sample_ipndm_v(denoise_cb_t model,
         float sigma  = sigmas[i];
         float t_next = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma, i + 1);
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
@@ -1249,6 +1341,7 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
                                               sd::Tensor<float> x,
                                               const std::vector<float>& sigmas,
                                               std::shared_ptr<RNG> rng,
+                                              bool is_flow_denoiser,
                                               float eta) {
     sd::Tensor<float> old_denoised = x;
     bool have_old_sigma            = false;
@@ -1272,15 +1365,16 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        auto denoised_opt = model(x, sigmas[i], i + 1);
+        auto denoised_opt = model(x, sigmas[i], i + 1, nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
         sd::Tensor<float> denoised = std::move(denoised_opt);
 
-        float sigma_from            = sigmas[i];
-        float sigma_to              = sigmas[i + 1];
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
+        float sigma_from = sigmas[i];
+        float sigma_to   = sigmas[i + 1];
+
+        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma_from, sigma_to, eta, is_flow_denoiser);
 
         if (sigma_down == 0.0f || !have_old_sigma) {
             x += ((x - denoised) / sigma_from) * (sigma_down - sigma_from);
@@ -1307,7 +1401,10 @@ static sd::Tensor<float> sample_res_multistep(denoise_cb_t model,
             x = sigma_fn(h) * x + h * (b1 * denoised + b2 * old_denoised);
         }
 
-        if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
+        if (sigma_to > 0.0f && sigma_up > 0.0f) {
+            if (is_flow_denoiser) {
+                x *= alpha_scale;
+            }
             x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
         }
 
@@ -1322,6 +1419,7 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
                                        sd::Tensor<float> x,
                                        const std::vector<float>& sigmas,
                                        std::shared_ptr<RNG> rng,
+                                       bool is_flow_denoiser,
                                        float eta) {
     const float c2 = 0.5f;
     auto t_fn      = [](float sigma) -> float { return -logf(sigma); };
@@ -1344,13 +1442,13 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
         float sigma_from = sigmas[i];
         float sigma_to   = sigmas[i + 1];
 
-        auto denoised_opt = model(x, sigma_from, -(i + 1));
+        auto denoised_opt = model(x, sigma_from, -(i + 1), nullptr);
         if (denoised_opt.empty()) {
             return {};
         }
         sd::Tensor<float> denoised = std::move(denoised_opt);
 
-        auto [sigma_down, sigma_up] = get_ancestral_step(sigma_from, sigma_to, eta);
+        auto [sigma_down, sigma_up, alpha_scale] = get_ancestral_step(sigma_from, sigma_to, eta, is_flow_denoiser);
 
         sd::Tensor<float> x0 = x;
         if (sigma_down == 0.0f || sigma_from == 0.0f) {
@@ -1370,7 +1468,7 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
             sd::Tensor<float> eps1 = denoised - x0;
             sd::Tensor<float> x2   = x0 + eps1 * (h * a21);
 
-            auto denoised2_opt = model(x2, sigma_c2, i + 1);
+            auto denoised2_opt = model(x2, sigma_c2, i + 1, nullptr);
             if (denoised2_opt.empty()) {
                 return {};
             }
@@ -1379,7 +1477,10 @@ static sd::Tensor<float> sample_res_2s(denoise_cb_t model,
             x                           = x0 + h * (b1 * eps1 + b2 * eps2);
         }
 
-        if (sigmas[i + 1] > 0 && sigma_up > 0.0f) {
+        if (sigma_to > 0.0f && sigma_up > 0.0f) {
+            if (is_flow_denoiser) {
+                x *= alpha_scale;
+            }
             x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
         }
     }
@@ -1444,7 +1545,7 @@ static sd::Tensor<float> sample_er_sde(denoise_cb_t model,
 
     int steps = static_cast<int>(sigmas.size()) - 1;
     for (int i = 0; i < steps; i++) {
-        sd::Tensor<float> denoised = model(x, sigmas[i], i + 1);
+        sd::Tensor<float> denoised = model(x, sigmas[i], i + 1, nullptr);
         if (denoised.empty()) {
             return {};
         }
@@ -1520,68 +1621,6 @@ static sd::Tensor<float> sample_er_sde(denoise_cb_t model,
     return x;
 }
 
-static sd::Tensor<float> sample_ddim_trailing(denoise_cb_t model,
-                                              sd::Tensor<float> x,
-                                              const std::vector<float>& sigmas,
-                                              std::shared_ptr<RNG> rng,
-                                              float eta) {
-    float beta_start = 0.00085f;
-    float beta_end   = 0.0120f;
-    std::vector<double> alphas_cumprod(TIMESTEPS);
-    std::vector<double> compvis_sigmas(TIMESTEPS);
-    for (int i = 0; i < TIMESTEPS; i++) {
-        alphas_cumprod[i] =
-            (i == 0 ? 1.0f : alphas_cumprod[i - 1]) *
-            (1.0f -
-             std::pow(sqrtf(beta_start) +
-                          (sqrtf(beta_end) - sqrtf(beta_start)) *
-                              ((float)i / (TIMESTEPS - 1)),
-                      2));
-        compvis_sigmas[i] =
-            std::sqrt((1 - alphas_cumprod[i]) / alphas_cumprod[i]);
-    }
-
-    int steps = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        int timestep      = static_cast<int>(roundf(TIMESTEPS - i * ((float)TIMESTEPS / steps))) - 1;
-        int prev_timestep = timestep - TIMESTEPS / steps;
-        float sigma       = static_cast<float>(compvis_sigmas[timestep]);
-        if (i == 0) {
-            x *= std::sqrt(sigma * sigma + 1) / sigma;
-        } else {
-            x *= std::sqrt(sigma * sigma + 1);
-        }
-
-        auto model_output_opt = model(x, sigma, i + 1);
-        if (model_output_opt.empty()) {
-            return {};
-        }
-        sd::Tensor<float> model_output = std::move(model_output_opt);
-        model_output                   = (x - model_output) * (1.0f / sigma);
-
-        float alpha_prod_t      = static_cast<float>(alphas_cumprod[timestep]);
-        float alpha_prod_t_prev = static_cast<float>(prev_timestep >= 0 ? alphas_cumprod[prev_timestep] : alphas_cumprod[0]);
-        float beta_prod_t       = 1.0f - alpha_prod_t;
-
-        sd::Tensor<float> pred_original_sample = ((x / std::sqrt(sigma * sigma + 1)) -
-                                                  std::sqrt(beta_prod_t) * model_output) *
-                                                 (1.0f / std::sqrt(alpha_prod_t));
-
-        float beta_prod_t_prev = 1.0f - alpha_prod_t_prev;
-        float variance         = (beta_prod_t_prev / beta_prod_t) *
-                         (1.0f - alpha_prod_t / alpha_prod_t_prev);
-        float std_dev_t = eta * std::sqrt(variance);
-
-        x = std::sqrt(alpha_prod_t_prev) * pred_original_sample +
-            std::sqrt(1.0f - alpha_prod_t_prev - std::pow(std_dev_t, 2)) * model_output;
-
-        if (eta > 0) {
-            x += std_dev_t * sd::Tensor<float>::randn_like(x, rng);
-        }
-    }
-    return x;
-}
-
 static sd::Tensor<float> sample_tcd(denoise_cb_t model,
                                     sd::Tensor<float> x,
                                     const std::vector<float>& sigmas,
@@ -1603,43 +1642,96 @@ static sd::Tensor<float> sample_tcd(denoise_cb_t model,
             std::sqrt((1 - alphas_cumprod[i]) / alphas_cumprod[i]);
     }
 
-    int original_steps = 50;
-    int steps          = static_cast<int>(sigmas.size()) - 1;
-    for (int i = 0; i < steps; i++) {
-        int timestep      = TIMESTEPS - 1 - (TIMESTEPS / original_steps) * (int)floor(i * ((float)original_steps / steps));
-        int prev_timestep = i >= steps - 1 ? 0 : TIMESTEPS - 1 - (TIMESTEPS / original_steps) * (int)floor((i + 1) * ((float)original_steps / steps));
-        int timestep_s    = (int)floor((1 - eta) * prev_timestep);
-        float sigma       = static_cast<float>(compvis_sigmas[timestep]);
-
-        if (i == 0) {
-            x *= std::sqrt(sigma * sigma + 1) / sigma;
-        } else {
-            x *= std::sqrt(sigma * sigma + 1);
+    auto get_timestep_from_sigma = [&](float s) -> int {
+        auto it = std::lower_bound(compvis_sigmas.begin(), compvis_sigmas.end(), s);
+        if (it == compvis_sigmas.begin())
+            return 0;
+        if (it == compvis_sigmas.end())
+            return TIMESTEPS - 1;
+        int idx_high = static_cast<int>(std::distance(compvis_sigmas.begin(), it));
+        int idx_low  = idx_high - 1;
+        if (std::abs(compvis_sigmas[idx_high] - s) < std::abs(compvis_sigmas[idx_low] - s)) {
+            return idx_high;
         }
+        return idx_low;
+    };
 
-        auto model_output_opt = model(x, sigma, i + 1);
-        if (model_output_opt.empty()) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma_to    = sigmas[i + 1];
+        int prev_timestep = get_timestep_from_sigma(sigma_to);
+        int timestep_s    = (int)floor((1 - eta) * prev_timestep);
+        float sigma       = sigmas[i];
+
+        auto denoised_opt = model(x, sigma, i + 1, nullptr);
+        if (denoised_opt.empty()) {
             return {};
         }
-        sd::Tensor<float> model_output = std::move(model_output_opt);
-        model_output                   = (x - model_output) * (1.0f / sigma);
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> d        = (x - denoised) / sigma;
 
-        float alpha_prod_t      = static_cast<float>(alphas_cumprod[timestep]);
+        float alpha_prod_t      = 1.0f / (sigma * sigma + 1.0f);
         float beta_prod_t       = 1.0f - alpha_prod_t;
-        float alpha_prod_t_prev = static_cast<float>(prev_timestep >= 0 ? alphas_cumprod[prev_timestep] : alphas_cumprod[0]);
+        float alpha_prod_t_prev = 1.0f / (sigma_to * sigma_to + 1.0f);
         float alpha_prod_s      = static_cast<float>(alphas_cumprod[timestep_s]);
         float beta_prod_s       = 1.0f - alpha_prod_s;
 
-        sd::Tensor<float> pred_original_sample = ((x / std::sqrt(sigma * sigma + 1)) -
-                                                  std::sqrt(beta_prod_t) * model_output) *
-                                                 (1.0f / std::sqrt(alpha_prod_t));
+        x = std::sqrt(alpha_prod_s / alpha_prod_t_prev) * denoised +
+            std::sqrt(beta_prod_s / alpha_prod_t_prev) * d;
 
-        x = std::sqrt(alpha_prod_s) * pred_original_sample +
-            std::sqrt(beta_prod_s) * model_output;
-
-        if (eta > 0 && i != steps - 1) {
+        if (eta > 0 && sigma_to > 0.0f) {
             x = std::sqrt(alpha_prod_t_prev / alpha_prod_s) * x +
-                std::sqrt(1.0f - alpha_prod_t_prev / alpha_prod_s) * sd::Tensor<float>::randn_like(x, rng);
+                std::sqrt(1.0f / alpha_prod_t_prev - 1.0f / alpha_prod_s) * sd::Tensor<float>::randn_like(x, rng);
+        }
+    }
+    return x;
+}
+
+static sd::Tensor<float> sample_euler_cfg_pp(denoise_cb_t model,
+                                             sd::Tensor<float> x,
+                                             const std::vector<float>& sigmas) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma = sigmas[i];
+        sd::Tensor<float> uncond_denoised;
+
+        auto denoised_opt = model(x, sigma, i + 1, &uncond_denoised);
+        if (denoised_opt.empty() || uncond_denoised.empty()) {
+            return {};
+        }
+
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> d        = (x - uncond_denoised) / sigma;
+
+        x = denoised + d * sigmas[i + 1];
+    }
+    return x;
+}
+
+static sd::Tensor<float> sample_euler_ancestral_cfg_pp(denoise_cb_t model,
+                                                       sd::Tensor<float> x,
+                                                       const std::vector<float>& sigmas,
+                                                       std::shared_ptr<RNG> rng,
+                                                       float eta) {
+    int steps = static_cast<int>(sigmas.size()) - 1;
+    for (int i = 0; i < steps; i++) {
+        float sigma = sigmas[i];
+        sd::Tensor<float> uncond_denoised;
+
+        auto denoised_opt = model(x, sigma, i + 1, &uncond_denoised);
+        if (denoised_opt.empty() || uncond_denoised.empty()) {
+            return {};
+        }
+
+        sd::Tensor<float> denoised = std::move(denoised_opt);
+        sd::Tensor<float> d        = (x - uncond_denoised) / sigma;
+
+        auto [sigma_down, sigma_up] = get_ancestral_step(sigmas[i], sigmas[i + 1], eta);
+
+        x = denoised + d * sigma_down;
+
+        if (sigmas[i + 1] > 0) {
+            x += sd::Tensor<float>::randn_like(x, rng) * sigma_up;
         }
     }
     return x;
@@ -1652,13 +1744,11 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
                                             std::vector<float> sigmas,
                                             std::shared_ptr<RNG> rng,
                                             float eta,
-                                            bool is_flow_denoiser) {
+                                            bool is_flow_denoiser,
+                                            const char* extra_sample_args) {
     switch (method) {
         case EULER_A_SAMPLE_METHOD:
-            if (is_flow_denoiser)
-                return sample_euler_flow(model, std::move(x), sigmas, rng, eta);
-            else
-                return sample_euler_ancestral(model, std::move(x), sigmas, rng, eta);
+            return sample_euler_ancestral(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case EULER_SAMPLE_METHOD:
             return sample_euler(model, std::move(x), sigmas);
         case HEUN_SAMPLE_METHOD:
@@ -1675,21 +1765,26 @@ static sd::Tensor<float> sample_k_diffusion(sample_method_t method,
         case DPMPP2Mv2_SAMPLE_METHOD:
             return sample_dpmpp_2m_v2(model, std::move(x), sigmas);
         case LCM_SAMPLE_METHOD:
-            return sample_lcm(model, std::move(x), sigmas, rng, is_flow_denoiser);
+            return sample_lcm(model, std::move(x), sigmas, rng, is_flow_denoiser, extra_sample_args);
         case IPNDM_SAMPLE_METHOD:
             return sample_ipndm(model, std::move(x), sigmas);
         case IPNDM_V_SAMPLE_METHOD:
             return sample_ipndm_v(model, std::move(x), sigmas);
         case RES_MULTISTEP_SAMPLE_METHOD:
-            return sample_res_multistep(model, std::move(x), sigmas, rng, eta);
+            return sample_res_multistep(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case RES_2S_SAMPLE_METHOD:
-            return sample_res_2s(model, std::move(x), sigmas, rng, eta);
+            return sample_res_2s(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case ER_SDE_SAMPLE_METHOD:
             return sample_er_sde(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case DDIM_TRAILING_SAMPLE_METHOD:
-            return sample_ddim_trailing(model, std::move(x), sigmas, rng, eta);
+            // DDIM is equivalent to Euler Ancestral with the Simple scheduler
+            return sample_euler_ancestral(model, std::move(x), sigmas, rng, is_flow_denoiser, eta);
         case TCD_SAMPLE_METHOD:
             return sample_tcd(model, std::move(x), sigmas, rng, eta);
+        case EULER_CFG_PP_SAMPLE_METHOD:
+            return sample_euler_cfg_pp(model, std::move(x), sigmas);
+        case EULER_A_CFG_PP_SAMPLE_METHOD:
+            return sample_euler_ancestral_cfg_pp(model, std::move(x), sigmas, rng, eta);
         default:
             return {};
     }

src/diffusion_model.hpp

@@ -5,6 +5,7 @@
 #include "anima.hpp"
 #include "ernie_image.hpp"
 #include "flux.hpp"
+#include "hidream_o1.hpp"
 #include "mmdit.hpp"
 #include "qwen_image.hpp"
 #include "tensor_ggml.hpp"
@@ -13,22 +14,28 @@
 #include "z_image.hpp"
 
 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 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;
-    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;
-    const std::vector<int>* skip_layers               = nullptr;
+    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 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;
+    const std::vector<int>* skip_layers                                = nullptr;
 };
 
 template <typename T>
@@ -49,6 +56,7 @@ struct DiffusionModel {
     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;
 };
 
@@ -56,10 +64,10 @@ struct UNetModel : public DiffusionModel {
     UNetModelRunner unet;
 
     UNetModel(ggml_backend_t backend,
-              bool offload_params_to_cpu,
+              ggml_backend_t params_backend,
               const String2TensorStorage& tensor_storage_map = {},
               SDVersion version                              = VERSION_SD1)
-        : unet(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model", version) {
+        : unet(backend, params_backend, tensor_storage_map, "model.diffusion_model", version) {
     }
 
     std::string get_desc() override {
@@ -98,6 +106,10 @@ struct UNetModel : public DiffusionModel {
         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);
     }
@@ -123,9 +135,9 @@ struct MMDiTModel : public DiffusionModel {
     MMDiTRunner mmdit;
 
     MMDiTModel(ggml_backend_t backend,
-               bool offload_params_to_cpu,
+               ggml_backend_t params_backend,
                const String2TensorStorage& tensor_storage_map = {})
-        : mmdit(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model") {
+        : mmdit(backend, params_backend, tensor_storage_map, "model.diffusion_model") {
     }
 
     std::string get_desc() override {
@@ -164,6 +176,10 @@ struct MMDiTModel : public DiffusionModel {
         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);
     }
@@ -186,11 +202,11 @@ struct FluxModel : public DiffusionModel {
     Flux::FluxRunner flux;
 
     FluxModel(ggml_backend_t backend,
-              bool offload_params_to_cpu,
+              ggml_backend_t params_backend,
               const String2TensorStorage& tensor_storage_map = {},
               SDVersion version                              = VERSION_FLUX,
               bool use_mask                                  = false)
-        : flux(backend, offload_params_to_cpu, tensor_storage_map, "model.diffusion_model", version, use_mask) {
+        : flux(backend, params_backend, tensor_storage_map, "model.diffusion_model", version, use_mask) {
     }
 
     std::string get_desc() override {
@@ -229,6 +245,10 @@ struct FluxModel : public DiffusionModel {
         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);
     }
@@ -257,10 +277,10 @@ struct AnimaModel : public DiffusionModel {
     Anima::AnimaRunner anima;
 
     AnimaModel(ggml_backend_t backend,
-               bool offload_params_to_cpu,
+               ggml_backend_t params_backend,
                const String2TensorStorage& tensor_storage_map = {},
                const std::string prefix                       = "model.diffusion_model")
-        : prefix(prefix), anima(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
+        : prefix(prefix), anima(backend, params_backend, tensor_storage_map, prefix) {
     }
 
     std::string get_desc() override {
@@ -299,6 +319,10 @@ struct AnimaModel : public DiffusionModel {
         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);
     }
@@ -321,11 +345,11 @@ struct WanModel : public DiffusionModel {
     WAN::WanRunner wan;
 
     WanModel(ggml_backend_t backend,
-             bool offload_params_to_cpu,
+             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, offload_params_to_cpu, tensor_storage_map, prefix, version) {
+        : prefix(prefix), wan(backend, params_backend, tensor_storage_map, prefix, version) {
     }
 
     std::string get_desc() override {
@@ -364,6 +388,10 @@ struct WanModel : public DiffusionModel {
         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);
     }
@@ -389,12 +417,12 @@ struct QwenImageModel : public DiffusionModel {
     Qwen::QwenImageRunner qwen_image;
 
     QwenImageModel(ggml_backend_t backend,
-                   bool offload_params_to_cpu,
+                   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, offload_params_to_cpu, tensor_storage_map, prefix, version, zero_cond_t) {
+        : prefix(prefix), qwen_image(backend, params_backend, tensor_storage_map, prefix, version, zero_cond_t) {
     }
 
     std::string get_desc() override {
@@ -433,6 +461,10 @@ struct QwenImageModel : public DiffusionModel {
         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);
     }
@@ -451,16 +483,92 @@ struct QwenImageModel : public DiffusionModel {
     }
 };
 
+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,
-                bool offload_params_to_cpu,
+                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, offload_params_to_cpu, tensor_storage_map, prefix, version) {
+        : prefix(prefix), z_image(backend, params_backend, tensor_storage_map, prefix, version) {
     }
 
     std::string get_desc() override {
@@ -499,6 +607,10 @@ struct ZImageModel : public DiffusionModel {
         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);
     }
@@ -522,10 +634,10 @@ struct ErnieImageModel : public DiffusionModel {
     ErnieImage::ErnieImageRunner ernie_image;
 
     ErnieImageModel(ggml_backend_t backend,
-                    bool offload_params_to_cpu,
+                    ggml_backend_t params_backend,
                     const String2TensorStorage& tensor_storage_map = {},
                     const std::string prefix                       = "model.diffusion_model")
-        : prefix(prefix), ernie_image(backend, offload_params_to_cpu, tensor_storage_map, prefix) {
+        : prefix(prefix), ernie_image(backend, params_backend, tensor_storage_map, prefix) {
     }
 
     std::string get_desc() override {
@@ -564,6 +676,10 @@ struct ErnieImageModel : public DiffusionModel {
         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);
     }

src/ernie_image.hpp

@@ -295,7 +295,9 @@ namespace ErnieImage {
             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]
-            auto chunks     = ggml_ext_chunk(ctx->ggml_ctx, mod_params, 6, 0);
+            sd::ggml_graph_cut::mark_graph_cut(hidden_states, "ernie_image.prelude", "hidden_states");
+            // sd::ggml_graph_cut::mark_graph_cut(mod_params, "ernie_image.prelude", "mod_params");
+            auto chunks = ggml_ext_chunk(ctx->ggml_ctx, mod_params, 6, 0);
             std::vector<ggml_tensor*> temb;
             temb.reserve(6);
             for (auto chunk : chunks) {
@@ -305,6 +307,7 @@ namespace ErnieImage {
             for (int i = 0; i < params.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");
             }
 
             hidden_states = final_norm->forward(ctx, hidden_states, c);
@@ -328,10 +331,10 @@ namespace ErnieImage {
         std::vector<float> pe_vec;
 
         ErnieImageRunner(ggml_backend_t backend,
-                         bool offload_params_to_cpu,
+                         ggml_backend_t params_backend,
                          const String2TensorStorage& tensor_storage_map = {},
                          const std::string prefix                       = "")
-            : GGMLRunner(backend, offload_params_to_cpu) {
+            : GGMLRunner(backend, params_backend) {
             ernie_params.num_layers = 0;
             for (const auto& [name, tensor_storage] : tensor_storage_map) {
                 if (!starts_with(name, prefix)) {

src/esrgan.hpp

@@ -124,27 +124,33 @@ public:
         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);
+        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;
     }
 };
@@ -155,10 +161,10 @@ struct ESRGAN : public GGMLRunner {
     int tile_size = 128;  // avoid cuda OOM for 4gb VRAM
 
     ESRGAN(ggml_backend_t backend,
-           bool offload_params_to_cpu,
+           ggml_backend_t params_backend,
            int tile_size                                  = 128,
            const String2TensorStorage& tensor_storage_map = {})
-        : GGMLRunner(backend, offload_params_to_cpu) {
+        : GGMLRunner(backend, params_backend) {
         this->tile_size = tile_size;
     }
 

src/flux.hpp

@@ -928,6 +928,9 @@ namespace Flux {
             }
 
             txt = txt_in->forward(ctx, txt);
+            sd::ggml_graph_cut::mark_graph_cut(img, "flux.prelude", "img");
+            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++) {
                 if (skip_layers.size() > 0 && std::find(skip_layers.begin(), skip_layers.end(), i) != skip_layers.end()) {
@@ -939,6 +942,8 @@ namespace Flux {
                 auto img_txt = block->forward(ctx, img, txt, vec, pe, txt_img_mask, ds_img_mods, ds_txt_mods);
                 img          = img_txt.first;   // [N, n_img_token, hidden_size]
                 txt          = img_txt.second;  // [N, n_txt_token, hidden_size]
+                sd::ggml_graph_cut::mark_graph_cut(img, "flux.double_blocks." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "flux.double_blocks." + std::to_string(i), "txt");
             }
 
             auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);  // [N, n_txt_token + n_img_token, hidden_size]
@@ -949,6 +954,7 @@ namespace Flux {
                 auto block = std::dynamic_pointer_cast<SingleStreamBlock>(blocks["single_blocks." + std::to_string(i)]);
 
                 txt_img = block->forward(ctx, txt_img, vec, pe, txt_img_mask, ss_mods);
+                sd::ggml_graph_cut::mark_graph_cut(txt_img, "flux.single_blocks." + std::to_string(i), "txt_img");
             }
 
             img = ggml_view_3d(ctx->ggml_ctx,
@@ -1183,12 +1189,12 @@ namespace Flux {
         bool use_mask = false;
 
         FluxRunner(ggml_backend_t backend,
-                   bool offload_params_to_cpu,
+                   ggml_backend_t params_backend,
                    const String2TensorStorage& tensor_storage_map = {},
                    const std::string prefix                       = "",
                    SDVersion version                              = VERSION_FLUX,
                    bool use_mask                                  = false)
-            : GGMLRunner(backend, offload_params_to_cpu), version(version), use_mask(use_mask) {
+            : GGMLRunner(backend, params_backend), version(version), use_mask(use_mask) {
             flux_params.version             = version;
             flux_params.guidance_embed      = false;
             flux_params.depth               = 0;
@@ -1558,7 +1564,7 @@ namespace Flux {
             }
 
             std::shared_ptr<FluxRunner> flux = std::make_shared<FluxRunner>(backend,
-                                                                            false,
+                                                                            backend,
                                                                             tensor_storage_map,
                                                                             "model.diffusion_model",
                                                                             VERSION_FLUX2,

src/ggml_extend_backend.cpp

@@ -0,0 +1,600 @@
+#include "ggml_extend_backend.h"
+
+#include <algorithm>
+#include <cctype>
+#include <cstdlib>
+#include <mutex>
+#include <sstream>
+#include <stdexcept>
+#include <vector>
+
+#include "util.h"
+
+static std::string trim_copy(const std::string& value) {
+    size_t begin = 0;
+    while (begin < value.size() && std::isspace(static_cast<unsigned char>(value[begin]))) {
+        ++begin;
+    }
+    size_t end = value.size();
+    while (end > begin && std::isspace(static_cast<unsigned char>(value[end - 1]))) {
+        --end;
+    }
+    return value.substr(begin, end - begin);
+}
+
+static std::string lower_copy(std::string value) {
+    std::transform(value.begin(), value.end(), value.begin(), [](unsigned char c) {
+        return static_cast<char>(std::tolower(c));
+    });
+    return value;
+}
+
+static std::vector<std::string> split_copy(const std::string& value, char delimiter) {
+    std::vector<std::string> parts;
+    std::string part;
+    std::istringstream stream(value);
+    while (std::getline(stream, part, delimiter)) {
+        parts.push_back(part);
+    }
+    return parts;
+}
+
+static bool is_default_backend_token(const std::string& name) {
+    const std::string lower = lower_copy(trim_copy(name));
+    return lower.empty() || lower == "default" || lower == "auto";
+}
+
+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());
+    name.erase(std::remove(name.begin(), name.end(), '_'), name.end());
+
+    if (name == "diffusion" || name == "model" || name == "unet" || name == "dit") {
+        *module = SDBackendModule::DIFFUSION;
+        return true;
+    }
+    if (name == "te" || name == "clip" || name == "text" || name == "textencoder" || name == "textencoders" || name == "conditioner" || name == "cond" || name == "llm" || name == "t5" || name == "t5xxl") {
+        *module = SDBackendModule::TE;
+        return true;
+    }
+    if (name == "clipvision" || name == "vision") {
+        *module = SDBackendModule::CLIP_VISION;
+        return true;
+    }
+    if (name == "vae" || name == "firststage" || name == "autoencoder" || name == "tae") {
+        *module = SDBackendModule::VAE;
+        return true;
+    }
+    if (name == "controlnet" || name == "control") {
+        *module = SDBackendModule::CONTROL_NET;
+        return true;
+    }
+    if (name == "photomaker" || name == "photomakerid" || name == "pmid" || name == "photo") {
+        *module = SDBackendModule::PHOTOMAKER;
+        return true;
+    }
+    if (name == "upscaler" || name == "esrgan" || name == "hires") {
+        *module = SDBackendModule::UPSCALER;
+        return true;
+    }
+    return false;
+}
+
+static std::string module_assignment_name(const SDBackendAssignment& assignment, SDBackendModule module) {
+    auto it = assignment.module_names.find(module);
+    if (it != assignment.module_names.end()) {
+        return it->second;
+    }
+    return assignment.default_name;
+}
+
+static std::string backend_cache_key(ggml_backend_t backend) {
+    if (backend == nullptr) {
+        return "";
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (dev != nullptr) {
+        return lower_copy(ggml_backend_dev_name(dev));
+    }
+    const char* backend_name = ggml_backend_name(backend);
+    return backend_name != nullptr ? lower_copy(backend_name) : "";
+}
+
+static std::string resolve_first_device_by_type(enum ggml_backend_dev_type type) {
+    ggml_backend_dev_t dev = ggml_backend_dev_by_type(type);
+    if (dev == nullptr) {
+        return "";
+    }
+    return ggml_backend_dev_name(dev);
+}
+
+static ggml_backend_buffer_t ggml_backend_tensor_buffer(const struct ggml_tensor* tensor) {
+    if (tensor == nullptr) {
+        return nullptr;
+    }
+
+    return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+}
+
+static bool ggml_backend_tensor_is_host_accessible(const struct ggml_tensor* tensor) {
+    if (tensor == nullptr || tensor->data == nullptr) {
+        return false;
+    }
+
+    ggml_backend_buffer_t buffer = ggml_backend_tensor_buffer(tensor);
+    return buffer == nullptr || ggml_backend_buffer_is_host(buffer);
+}
+
+static size_t ggml_backend_tensor_offset(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+    return static_cast<size_t>(i0 * tensor->nb[0] + i1 * tensor->nb[1] + i2 * tensor->nb[2] + i3 * tensor->nb[3]);
+}
+
+template <typename T>
+static void ggml_backend_tensor_write_scalar(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, T value) {
+    const size_t offset = ggml_backend_tensor_offset(tensor, i0, i1, i2, i3);
+
+    if (ggml_backend_tensor_is_host_accessible(tensor)) {
+        auto* dst = reinterpret_cast<T*>(reinterpret_cast<char*>(tensor->data) + offset);
+        *dst      = value;
+        return;
+    }
+
+    ggml_backend_tensor_set(const_cast<struct ggml_tensor*>(tensor), &value, offset, sizeof(T));
+}
+
+static void ggml_set_f32_nd(const struct ggml_tensor* tensor, int64_t i0, int64_t i1, int64_t i2, int64_t i3, float value) {
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int8_t>(value));
+            break;
+        case GGML_TYPE_I16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int16_t>(value));
+            break;
+        case GGML_TYPE_I32:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, static_cast<int32_t>(value));
+            break;
+        case GGML_TYPE_F16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_fp16(value));
+            break;
+        case GGML_TYPE_BF16:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, ggml_fp32_to_bf16(value));
+            break;
+        case GGML_TYPE_F32:
+            ggml_backend_tensor_write_scalar(tensor, i0, i1, i2, i3, value);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+void ggml_ext_im_set_f32_1d(const struct ggml_tensor* tensor, int i, float value) {
+    if (!ggml_is_contiguous(tensor)) {
+        int64_t id[4] = {0, 0, 0, 0};
+        ggml_unravel_index(tensor, i, &id[0], &id[1], &id[2], &id[3]);
+        ggml_set_f32_nd(tensor, id[0], id[1], id[2], id[3], value);
+        return;
+    }
+
+    switch (tensor->type) {
+        case GGML_TYPE_I8:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int8_t>(value));
+            break;
+        case GGML_TYPE_I16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int16_t>(value));
+            break;
+        case GGML_TYPE_I32:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, static_cast<int32_t>(value));
+            break;
+        case GGML_TYPE_F16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_fp16(value));
+            break;
+        case GGML_TYPE_BF16:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, ggml_fp32_to_bf16(value));
+            break;
+        case GGML_TYPE_F32:
+            ggml_backend_tensor_write_scalar(tensor, i, 0, 0, 0, value);
+            break;
+        default:
+            GGML_ABORT("fatal error");
+    }
+}
+
+static void ggml_backend_load_all_once() {
+    // If the registry already has devices and the CPU backend is present,
+    // assume either static registration or explicit host-side preloading has
+    // completed and avoid rescanning the default paths.
+    if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
+        return;
+    }
+    // In dynamic-backend mode the backend modules are discovered at runtime,
+    // so we must load them before asking for the CPU backend or its proc table.
+    // If the host preloaded only a subset of backends, allow one default-path
+    // scan so missing modules can still be discovered.
+    static std::once_flag once;
+    std::call_once(once, []() {
+        if (ggml_backend_dev_count() > 0 && ggml_backend_reg_by_name("CPU") != nullptr) {
+            return;
+        }
+        ggml_backend_load_all();
+    });
+}
+
+bool sd_backend_is(ggml_backend_t backend, const std::string& name) {
+    if (!backend) {
+        return false;
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (!dev) {
+        return false;
+    }
+    std::string dev_name = ggml_backend_dev_name(dev);
+    return lower_copy(dev_name).find(lower_copy(name)) != std::string::npos;
+}
+
+static std::string get_default_backend_name() {
+    ggml_backend_load_all_once();
+    // should pick the same backend preference as ggml_backend_init_best
+    std::string name = resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+    if (!name.empty()) {
+        return name;
+    }
+    name = resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
+    if (!name.empty()) {
+        return name;
+    }
+    return resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+}
+
+static std::string sd_resolve_backend_name(const std::string& name) {
+    ggml_backend_load_all_once();
+    std::string requested = trim_copy(name);
+    std::string lower     = lower_copy(requested);
+
+    if (is_default_backend_token(lower)) {
+        return get_default_backend_name();
+    }
+    if (lower == "gpu") {
+        std::string result = resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_GPU);
+        if (!result.empty()) {
+            return result;
+        }
+        return resolve_first_device_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU);
+    }
+
+    const size_t device_count = ggml_backend_dev_count();
+    for (size_t i = 0; i < device_count; ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        std::string dev_name   = ggml_backend_dev_name(dev);
+        if (lower_copy(dev_name) == lower) {
+            return dev_name;
+        }
+    }
+
+    for (size_t i = 0; i < device_count; ++i) {
+        ggml_backend_dev_t dev = ggml_backend_dev_get(i);
+        std::string dev_name   = ggml_backend_dev_name(dev);
+        std::string dev_lower  = lower_copy(dev_name);
+        if (dev_lower.rfind(lower, 0) == 0) {
+            return dev_name;
+        }
+    }
+
+    return "";
+}
+
+static bool backend_name_exists(const std::string& name) {
+    return !sd_resolve_backend_name(name).empty();
+}
+
+static ggml_backend_t init_named_backend(const std::string& name) {
+    ggml_backend_load_all_once();
+    LOG_DEBUG("Initializing backend: %s", name.c_str());
+    if (trim_copy(name).empty()) {
+        return ggml_backend_init_best();
+    }
+
+    std::string resolved = sd_resolve_backend_name(name);
+    if (resolved.empty()) {
+        return nullptr;
+    }
+    return ggml_backend_init_by_name(resolved.c_str(), nullptr);
+}
+
+static ggml_backend_t sd_get_default_backend() {
+    ggml_backend_load_all_once();
+    static std::once_flag once;
+    std::call_once(once, []() {
+        size_t dev_count = ggml_backend_dev_count();
+        if (dev_count == 0) {
+            LOG_ERROR("No devices found!");
+        } else {
+            LOG_DEBUG("Found %zu backend devices:", dev_count);
+            for (size_t i = 0; i < dev_count; ++i) {
+                auto dev = ggml_backend_dev_get(i);
+                LOG_DEBUG("#%zu: %s", i, ggml_backend_dev_name(dev));
+            }
+        }
+    });
+
+    ggml_backend_t backend   = nullptr;
+    const char* SD_VK_DEVICE = getenv("SD_VK_DEVICE");
+    if (SD_VK_DEVICE != nullptr) {
+        std::string sd_vk_device_str = SD_VK_DEVICE;
+        try {
+            unsigned long long device  = std::stoull(sd_vk_device_str);
+            std::string vk_device_name = "Vulkan" + std::to_string(device);
+            if (backend_name_exists(vk_device_name)) {
+                LOG_INFO("Selecting %s as main device by env var SD_VK_DEVICE", vk_device_name.c_str());
+                backend = init_named_backend(vk_device_name);
+                if (!backend) {
+                    LOG_WARN("Device %s requested by SD_VK_DEVICE failed to init. Falling back to the default device.", vk_device_name.c_str());
+                }
+            } else {
+                LOG_WARN("Device %s requested by SD_VK_DEVICE was not found. Falling back to the default device.", vk_device_name.c_str());
+            }
+        } catch (const std::invalid_argument&) {
+            LOG_WARN("SD_VK_DEVICE environment variable is not a valid integer (%s). Falling back to the default device.", SD_VK_DEVICE);
+        } catch (const std::out_of_range&) {
+            LOG_WARN("SD_VK_DEVICE environment variable value is out of range for `unsigned long long` type (%s). Falling back to the default device.", SD_VK_DEVICE);
+        }
+    }
+
+    if (!backend) {
+        std::string dev_name = get_default_backend_name();
+        backend              = init_named_backend(dev_name);
+        if (!backend && !dev_name.empty()) {
+            LOG_WARN("device %s failed to init", dev_name.c_str());
+        }
+    }
+
+    if (!backend) {
+        LOG_WARN("loading CPU backend");
+        backend = ggml_backend_cpu_init();
+    }
+
+    if (ggml_backend_is_cpu(backend)) {
+        LOG_DEBUG("Using CPU backend");
+    }
+
+    return backend;
+}
+
+static bool sd_parse_backend_assignment(const std::string& spec, SDBackendAssignment* assignment, std::string* error) {
+    if (assignment == nullptr) {
+        return false;
+    }
+
+    *assignment          = {};
+    const std::string in = trim_copy(spec);
+    if (in.empty()) {
+        return true;
+    }
+
+    for (const std::string& raw_part : split_copy(in, ',')) {
+        const std::string part = trim_copy(raw_part);
+        if (part.empty()) {
+            continue;
+        }
+
+        const size_t eq = part.find('=');
+        if (eq == std::string::npos) {
+            assignment->set_default(part);
+            continue;
+        }
+
+        const std::string key   = trim_copy(part.substr(0, eq));
+        const std::string value = trim_copy(part.substr(eq + 1));
+        if (key.empty() || value.empty()) {
+            if (error != nullptr) {
+                *error = "invalid backend assignment '" + part + "'";
+            }
+            return false;
+        }
+
+        const std::string key_lower = lower_copy(key);
+        if (key_lower == "all" || key_lower == "default" || key_lower == "*") {
+            assignment->set_default(value);
+            continue;
+        }
+
+        SDBackendModule module = SDBackendModule::DIFFUSION;
+        if (!parse_backend_module(key, &module)) {
+            if (error != nullptr) {
+                *error = "unknown backend module '" + key + "'";
+            }
+            return false;
+        }
+        assignment->set_module(module, value);
+    }
+    return true;
+}
+
+bool SDBackendAssignment::empty() const {
+    return default_name.empty() && module_names.empty();
+}
+
+std::string SDBackendAssignment::get(SDBackendModule module) const {
+    return module_assignment_name(*this, module);
+}
+
+void SDBackendAssignment::set_default(const std::string& name) {
+    default_name = trim_copy(name);
+}
+
+void SDBackendAssignment::set_module(SDBackendModule module, const std::string& name) {
+    module_names[module] = trim_copy(name);
+}
+
+void SDBackendHandleDeleter::operator()(ggml_backend_t backend) const {
+    ggml_backend_free(backend);
+}
+
+SDBackendManager::~SDBackendManager() {
+    reset();
+}
+
+void SDBackendManager::reset() {
+    backends_.clear();
+    runtime_assignment_ = {};
+    params_assignment_  = {};
+}
+
+ggml_backend_t SDBackendManager::runtime_backend(SDBackendModule module) {
+    return init_cached_backend(runtime_assignment_.get(module));
+}
+
+ggml_backend_t SDBackendManager::params_backend(SDBackendModule module) {
+    std::string name = params_assignment_.get(module);
+    if (name.empty()) {
+        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));
+}
+
+bool SDBackendManager::params_backend_is_cpu(SDBackendModule module) {
+    return ggml_backend_is_cpu(params_backend(module));
+}
+
+bool SDBackendManager::runtime_backend_supports_host_buffer(SDBackendModule module) {
+    ggml_backend_t backend = runtime_backend(module);
+    if (backend == nullptr) {
+        return false;
+    }
+    if (ggml_backend_is_cpu(backend)) {
+        return true;
+    }
+    ggml_backend_dev_t dev = ggml_backend_get_device(backend);
+    if (dev == nullptr) {
+        return false;
+    }
+    ggml_backend_dev_props props;
+    ggml_backend_dev_get_props(dev, &props);
+    return props.caps.buffer_from_host_ptr;
+}
+
+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,
+                            std::string* error) {
+    reset();
+
+    if (!sd_parse_backend_assignment(SAFE_STR(backend_spec), &runtime_assignment_, error)) {
+        return false;
+    }
+    if (!sd_parse_backend_assignment(SAFE_STR(params_backend_spec), &params_assignment_, error)) {
+        return false;
+    }
+
+    if (runtime_assignment_.empty()) {
+        if (keep_clip_on_cpu) {
+            runtime_assignment_.set_module(SDBackendModule::TE, "cpu");
+        }
+        if (keep_vae_on_cpu) {
+            runtime_assignment_.set_module(SDBackendModule::VAE, "cpu");
+        }
+        if (keep_control_net_on_cpu) {
+            runtime_assignment_.set_module(SDBackendModule::CONTROL_NET, "cpu");
+        }
+    }
+
+    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 {
+        if (is_default_backend_token(name)) {
+            return true;
+        }
+        if (!sd_resolve_backend_name(name).empty()) {
+            return true;
+        }
+        if (error != nullptr) {
+            *error = "backend '" + name + "' was not found";
+        }
+        return false;
+    };
+
+    if (!validate_name(runtime_assignment_.default_name) ||
+        !validate_name(params_assignment_.default_name)) {
+        return false;
+    }
+    for (const auto& kv : runtime_assignment_.module_names) {
+        if (!validate_name(kv.second)) {
+            return false;
+        }
+    }
+    for (const auto& kv : params_assignment_.module_names) {
+        if (!validate_name(kv.second)) {
+            return false;
+        }
+    }
+    return true;
+}
+
+ggml_backend_t SDBackendManager::init_cached_backend(const std::string& name) {
+    std::string resolved   = sd_resolve_backend_name(name);
+    std::string key        = lower_copy(resolved);
+    ggml_backend_t backend = nullptr;
+
+    if (!key.empty()) {
+        auto it = backends_.find(key);
+        if (it != backends_.end()) {
+            return it->second.get();
+        }
+    } else if (!is_default_backend_token(name)) {
+        LOG_ERROR("backend '%s' was not found", name.c_str());
+        return nullptr;
+    }
+
+    backend = is_default_backend_token(name) ? sd_get_default_backend() : init_named_backend(resolved);
+    if (backend == nullptr) {
+        LOG_ERROR("failed to initialize backend '%s'", name.c_str());
+        return nullptr;
+    }
+
+    std::string actual_key = backend_cache_key(backend);
+    if (actual_key.empty()) {
+        actual_key = !key.empty() ? key : lower_copy(trim_copy(name));
+    }
+
+    auto it = backends_.find(actual_key);
+    if (it != backends_.end()) {
+        ggml_backend_free(backend);
+        return it->second.get();
+    }
+
+    SDBackendHandle handle(backend);
+    backends_.emplace(actual_key, std::move(handle));
+    return backend;
+}
+
+const char* sd_backend_module_name(SDBackendModule module) {
+    switch (module) {
+        case SDBackendModule::DIFFUSION:
+            return "diffusion";
+        case SDBackendModule::TE:
+            return "te";
+        case SDBackendModule::CLIP_VISION:
+            return "clip_vision";
+        case SDBackendModule::VAE:
+            return "vae";
+        case SDBackendModule::CONTROL_NET:
+            return "controlnet";
+        case SDBackendModule::PHOTOMAKER:
+            return "photomaker";
+        case SDBackendModule::UPSCALER:
+            return "upscaler";
+    }
+    return "unknown";
+}

src/ggml_extend_backend.h

@@ -0,0 +1,77 @@
+#ifndef __SD_GGML_EXTEND_BACKEND_H__
+#define __SD_GGML_EXTEND_BACKEND_H__
+
+#include <cstdint>
+#include <cstring>
+#include <memory>
+#include <string>
+#include <unordered_map>
+
+#include "ggml-backend.h"
+#include "ggml-cpu.h"
+#include "ggml.h"
+
+enum class SDBackendModule {
+    DIFFUSION,
+    TE,
+    CLIP_VISION,
+    VAE,
+    CONTROL_NET,
+    PHOTOMAKER,
+    UPSCALER,
+};
+
+struct SDBackendAssignment {
+    std::string default_name;
+    std::unordered_map<SDBackendModule, std::string> module_names;
+
+    bool empty() const;
+    std::string get(SDBackendModule module) const;
+    void set_default(const std::string& name);
+    void set_module(SDBackendModule module, const std::string& name);
+};
+
+struct SDBackendHandleDeleter {
+    void operator()(ggml_backend_t backend) const;
+};
+
+using SDBackendHandle = std::unique_ptr<struct ggml_backend, SDBackendHandleDeleter>;
+
+class SDBackendManager {
+private:
+    SDBackendAssignment runtime_assignment_;
+    SDBackendAssignment params_assignment_;
+    std::unordered_map<std::string, SDBackendHandle> backends_;
+
+public:
+    SDBackendManager() = default;
+    ~SDBackendManager();
+
+    SDBackendManager(const SDBackendManager&)            = delete;
+    SDBackendManager& operator=(const SDBackendManager&) = delete;
+
+    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,
+              std::string* error);
+    void reset();
+
+    ggml_backend_t runtime_backend(SDBackendModule module);
+    ggml_backend_t params_backend(SDBackendModule module);
+
+    bool runtime_backend_is_cpu(SDBackendModule module);
+    bool params_backend_is_cpu(SDBackendModule module);
+    bool runtime_backend_supports_host_buffer(SDBackendModule module);
+
+private:
+    bool validate(std::string* error) const;
+    ggml_backend_t init_cached_backend(const std::string& name);
+};
+
+bool sd_backend_is(ggml_backend_t backend, const std::string& name);
+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

src/ggml_graph_cut.cpp

@@ -0,0 +1,755 @@
+#include "ggml_graph_cut.h"
+
+#include <algorithm>
+#include <cstring>
+#include <map>
+#include <set>
+#include <sstream>
+#include <stack>
+#include <unordered_map>
+
+#include "ggml-alloc.h"
+#include "ggml-backend.h"
+#include "util.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 std::string graph_cut_tensor_display_name(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return "<null>";
+        }
+        if (tensor->name[0] != '\0') {
+            return tensor->name;
+        }
+        return sd_format("<tensor@%p>", (const void*)tensor);
+    }
+
+    static int graph_leaf_index(ggml_cgraph* gf, const ggml_tensor* tensor) {
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(tensor != nullptr);
+        for (int i = 0; i < gf->n_leafs; ++i) {
+            if (gf->leafs[i] == tensor) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    static bool is_params_tensor(const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                                 const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return false;
+        }
+        return params_tensor_set.find(tensor) != params_tensor_set.end();
+    }
+
+    static int graph_node_index_by_name(ggml_cgraph* gf, const char* name) {
+        GGML_ASSERT(gf != nullptr);
+        if (name == nullptr || name[0] == '\0') {
+            return -1;
+        }
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor* node = ggml_graph_node(gf, i);
+            if (node != nullptr && std::strcmp(node->name, name) == 0) {
+                return i;
+            }
+        }
+        return -1;
+    }
+
+    static Plan::InputShape input_shape(const ggml_tensor* tensor) {
+        Plan::InputShape shape;
+        if (tensor == nullptr) {
+            return shape;
+        }
+        shape.type = tensor->type;
+        for (int i = 0; i < GGML_MAX_DIMS; ++i) {
+            shape.ne[static_cast<size_t>(i)] = tensor->ne[i];
+        }
+        return shape;
+    }
+
+    static size_t graph_cut_segment_vram_bytes(const Segment& segment) {
+        return segment.compute_buffer_size +
+               segment.input_param_bytes +
+               segment.input_previous_cut_bytes +
+               segment.output_bytes;
+    }
+
+    size_t max_vram_gib_to_bytes(float max_vram) {
+        if (max_vram <= 0.f) {
+            return 0;
+        }
+        return static_cast<size_t>(static_cast<double>(max_vram) * MAX_VRAM_BYTES_PER_GIB);
+    }
+
+    static float max_vram_bytes_to_gib(size_t max_vram_bytes) {
+        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) {
+        if (backend == nullptr) {
+            LOG_WARN("--max-vram -1 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");
+            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");
+            return 0;
+        }
+
+        size_t free_vram  = 0;
+        size_t total_vram = 0;
+        ggml_backend_dev_memory(dev, &free_vram, &total_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);
+            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",
+                 free_vram / MAX_VRAM_BYTES_PER_GIB,
+                 total_vram / MAX_VRAM_BYTES_PER_GIB,
+                 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) {
+            return max_vram;
+        }
+        return max_vram_bytes_to_gib(resolve_auto_max_vram_bytes(backend));
+    }
+
+    static Segment make_segment_seed(const Plan& plan,
+                                     size_t start_segment_index,
+                                     size_t end_segment_index) {
+        GGML_ASSERT(start_segment_index < plan.segments.size());
+        GGML_ASSERT(end_segment_index < plan.segments.size());
+        GGML_ASSERT(start_segment_index <= end_segment_index);
+
+        Segment seed;
+        const auto& start_segment  = plan.segments[start_segment_index];
+        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) {
+            for (int output_node_index : plan.segments[seg_idx].output_node_indices) {
+                if (seen_output_node_indices.insert(output_node_index).second) {
+                    seed.output_node_indices.push_back(output_node_index);
+                }
+            }
+        }
+        if (start_segment_index == end_segment_index) {
+            seed.group_name = target_segment.group_name;
+        } else {
+            seed.group_name = sd_format("%s..%s",
+                                        start_segment.group_name.c_str(),
+                                        target_segment.group_name.c_str());
+        }
+        return seed;
+    }
+
+    static void build_segment(ggml_cgraph* gf,
+                              Plan& plan,
+                              Segment& segment,
+                              const std::unordered_map<const ggml_tensor*, int>& producer_index,
+                              std::unordered_set<int>& available_cut_output_node_indices,
+                              ggml_backend_t backend,
+                              const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                              const char* log_desc) {
+        std::set<int> internal_nodes;
+        std::unordered_set<const ggml_tensor*> input_seen;
+        std::vector<Segment::InputRef> input_refs;
+
+        std::stack<ggml_tensor*> work_stack;
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            if (output != nullptr) {
+                work_stack.push(output);
+            }
+        }
+
+        while (!work_stack.empty()) {
+            ggml_tensor* tensor = work_stack.top();
+            work_stack.pop();
+
+            if (tensor == nullptr) {
+                continue;
+            }
+
+            auto producer_it = producer_index.find(tensor);
+            if (producer_it == producer_index.end()) {
+                if (input_seen.insert(tensor).second) {
+                    Segment::InputRef input_ref;
+                    input_ref.type         = is_params_tensor(params_tensor_set, tensor) ? Segment::INPUT_PARAM : Segment::INPUT_EXTERNAL;
+                    input_ref.display_name = graph_cut_tensor_display_name(tensor);
+                    input_ref.leaf_index   = graph_leaf_index(gf, tensor);
+                    input_refs.push_back(std::move(input_ref));
+                }
+                continue;
+            }
+
+            int node_idx = producer_it->second;
+            if (available_cut_output_node_indices.find(node_idx) != available_cut_output_node_indices.end()) {
+                if (input_seen.insert(tensor).second) {
+                    Segment::InputRef input_ref;
+                    input_ref.type         = Segment::INPUT_PREVIOUS_CUT;
+                    input_ref.display_name = graph_cut_tensor_display_name(tensor);
+                    input_ref.node_index   = node_idx;
+                    input_refs.push_back(std::move(input_ref));
+                }
+                continue;
+            }
+
+            if (!internal_nodes.insert(node_idx).second) {
+                continue;
+            }
+
+            ggml_tensor* node = ggml_graph_node(gf, node_idx);
+            for (int src_idx = 0; src_idx < GGML_MAX_SRC; ++src_idx) {
+                if (node->src[src_idx] != nullptr) {
+                    work_stack.push(node->src[src_idx]);
+                }
+            }
+        }
+
+        if (!internal_nodes.empty()) {
+            segment.internal_node_indices.assign(internal_nodes.begin(), internal_nodes.end());
+        }
+
+        std::sort(input_refs.begin(),
+                  input_refs.end(),
+                  [](const Segment::InputRef& a, const Segment::InputRef& b) {
+                      if (a.type != b.type) {
+                          return a.type < b.type;
+                      }
+                      return a.display_name < b.display_name;
+                  });
+        segment.input_refs = input_refs;
+        for (const auto& input : input_refs) {
+            ggml_tensor* current_input = input_tensor(gf, input);
+            size_t tensor_bytes        = current_input == nullptr
+                                             ? 0
+                                             : (input.type == Segment::INPUT_PREVIOUS_CUT
+                                                    ? cache_tensor_bytes(current_input)
+                                                    : ggml_nbytes(current_input));
+            switch (input.type) {
+                case Segment::INPUT_PREVIOUS_CUT:
+                    segment.input_previous_cut_bytes += tensor_bytes;
+                    break;
+                case Segment::INPUT_PARAM:
+                    segment.input_param_bytes += tensor_bytes;
+                    break;
+                case Segment::INPUT_EXTERNAL:
+                default:
+                    segment.input_external_bytes += tensor_bytes;
+                    break;
+            }
+        }
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            segment.output_bytes += cache_tensor_bytes(output);
+        }
+        segment.compute_buffer_size = measure_segment_compute_buffer(backend, gf, segment, log_desc);
+
+        for (int output_node_index : segment.output_node_indices) {
+            available_cut_output_node_indices.insert(output_node_index);
+        }
+        plan.segments.push_back(std::move(segment));
+    }
+
+    bool is_graph_cut_tensor(const ggml_tensor* tensor) {
+        if (tensor == nullptr || tensor->name[0] == '\0') {
+            return false;
+        }
+        return std::strncmp(tensor->name, GGML_RUNNER_CUT_PREFIX, std::strlen(GGML_RUNNER_CUT_PREFIX)) == 0;
+    }
+
+    std::string make_graph_cut_name(const std::string& group, const std::string& output) {
+        return std::string(GGML_RUNNER_CUT_PREFIX) + group + "|" + output;
+    }
+
+    void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output) {
+        if (tensor == nullptr) {
+            return;
+        }
+        auto name = make_graph_cut_name(group, output);
+        ggml_set_name(tensor, name.c_str());
+    }
+
+    int leaf_count(ggml_cgraph* gf) {
+        GGML_ASSERT(gf != nullptr);
+        return gf->n_leafs;
+    }
+
+    ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index) {
+        GGML_ASSERT(gf != nullptr);
+        if (leaf_index < 0 || leaf_index >= gf->n_leafs) {
+            return nullptr;
+        }
+        return gf->leafs[leaf_index];
+    }
+
+    ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return nullptr;
+        }
+        return tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
+    }
+
+    ggml_tensor* cache_source_tensor(ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return nullptr;
+        }
+        if (tensor_buffer(tensor) == nullptr && tensor->src[0] != nullptr &&
+            ggml_nelements(tensor->src[0]) == ggml_nelements(tensor) &&
+            ggml_nbytes(tensor->src[0]) == ggml_nbytes(tensor)) {
+            return cache_source_tensor(tensor->src[0]);
+        }
+        return tensor->view_src ? tensor->view_src : tensor;
+    }
+
+    size_t cache_tensor_bytes(const ggml_tensor* tensor) {
+        if (tensor == nullptr) {
+            return 0;
+        }
+        const ggml_tensor* cache_src = tensor->view_src ? tensor->view_src : tensor;
+        return ggml_nbytes(cache_src);
+    }
+
+    bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan) {
+        GGML_ASSERT(gf != nullptr);
+        if (ggml_graph_n_nodes(gf) != plan.n_nodes || gf->n_leafs != plan.n_leafs) {
+            return false;
+        }
+        for (const auto& input_shape_ref : plan.input_shapes) {
+            if (input_shape_ref.leaf_index < 0 || input_shape_ref.leaf_index >= gf->n_leafs) {
+                return false;
+            }
+            ggml_tensor* leaf = gf->leafs[input_shape_ref.leaf_index];
+            if (leaf == nullptr || input_shape_ref.type != leaf->type) {
+                return false;
+            }
+            for (int d = 0; d < GGML_MAX_DIMS; ++d) {
+                if (input_shape_ref.ne[static_cast<size_t>(d)] != leaf->ne[d]) {
+                    return false;
+                }
+            }
+        }
+        return true;
+    }
+
+    ggml_tensor* output_tensor(ggml_cgraph* gf, const Segment& segment, size_t output_index) {
+        GGML_ASSERT(gf != nullptr);
+        if (output_index >= segment.output_node_indices.size()) {
+            return nullptr;
+        }
+        int node_index = segment.output_node_indices[output_index];
+        if (node_index < 0 || node_index >= ggml_graph_n_nodes(gf)) {
+            return nullptr;
+        }
+        return ggml_graph_node(gf, node_index);
+    }
+
+    ggml_tensor* input_tensor(ggml_cgraph* gf, const Segment::InputRef& input_ref) {
+        GGML_ASSERT(gf != nullptr);
+        if (input_ref.type == Segment::INPUT_PREVIOUS_CUT) {
+            if (input_ref.node_index < 0 || input_ref.node_index >= ggml_graph_n_nodes(gf)) {
+                return nullptr;
+            }
+            return ggml_graph_node(gf, input_ref.node_index);
+        }
+        if (input_ref.leaf_index < 0 || input_ref.leaf_index >= gf->n_leafs) {
+            return nullptr;
+        }
+        return leaf_tensor(gf, input_ref.leaf_index);
+    }
+
+    std::vector<ggml_tensor*> param_tensors(ggml_cgraph* gf, const Segment& segment) {
+        GGML_ASSERT(gf != nullptr);
+        std::vector<ggml_tensor*> tensors;
+        std::unordered_set<ggml_tensor*> seen_tensors;
+        tensors.reserve(segment.input_refs.size());
+        seen_tensors.reserve(segment.input_refs.size());
+        for (const auto& input_ref : segment.input_refs) {
+            if (input_ref.type != Segment::INPUT_PARAM) {
+                continue;
+            }
+            ggml_tensor* tensor = input_tensor(gf, input_ref);
+            if (tensor == nullptr) {
+                continue;
+            }
+            if (seen_tensors.insert(tensor).second) {
+                tensors.push_back(tensor);
+            }
+        }
+        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) {
+        GGML_ASSERT(gf != nullptr);
+        std::unordered_set<std::string> future_input_names;
+        for (size_t seg_idx = current_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) {
+                    continue;
+                }
+                ggml_tensor* current_input = input_tensor(gf, input_ref);
+                if (current_input != nullptr && current_input->name[0] != '\0') {
+                    future_input_names.insert(current_input->name);
+                }
+            }
+        }
+        return future_input_names;
+    }
+
+    ggml_cgraph* build_segment_graph(ggml_cgraph* gf,
+                                     const Segment& segment,
+                                     ggml_context** graph_ctx_out) {
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(graph_ctx_out != nullptr);
+
+        const size_t graph_size = segment.internal_node_indices.size() + segment.input_refs.size() + 8;
+        ggml_init_params params = {
+            /*.mem_size   =*/ggml_graph_overhead_custom(graph_size, false) + 1024,
+            /*.mem_buffer =*/nullptr,
+            /*.no_alloc   =*/true,
+        };
+        ggml_context* graph_ctx = ggml_init(params);
+        GGML_ASSERT(graph_ctx != nullptr);
+        ggml_cgraph* segment_graph = ggml_new_graph_custom(graph_ctx, graph_size, false);
+        GGML_ASSERT(segment_graph != nullptr);
+
+        for (const auto& input : segment.input_refs) {
+            ggml_tensor* current_input = input_tensor(gf, input);
+            if (current_input == nullptr) {
+                continue;
+            }
+            GGML_ASSERT(segment_graph->n_leafs < segment_graph->size);
+            segment_graph->leafs[segment_graph->n_leafs++] = current_input;
+        }
+
+        for (int output_node_index : segment.output_node_indices) {
+            ggml_tensor* output = ggml_graph_node(gf, output_node_index);
+            if (output == nullptr) {
+                continue;
+            }
+            ggml_set_output(output);
+        }
+        for (int node_idx : segment.internal_node_indices) {
+            ggml_graph_add_node(segment_graph, ggml_graph_node(gf, node_idx));
+        }
+        *graph_ctx_out = graph_ctx;
+        return segment_graph;
+    }
+
+    size_t measure_segment_compute_buffer(ggml_backend_t backend,
+                                          ggml_cgraph* gf,
+                                          const Segment& segment,
+                                          const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        if (segment.internal_node_indices.empty()) {
+            return 0;
+        }
+
+        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));
+
+        size_t sizes[1] = {0};
+        ggml_gallocr_reserve_n_size(
+            allocr,
+            segment_graph,
+            nullptr,
+            nullptr,
+            sizes);
+        size_t buffer_size = sizes[0];
+
+        ggml_gallocr_free(allocr);
+        ggml_free(graph_ctx);
+        return buffer_size;
+    }
+
+    Plan build_plan(ggml_backend_t backend,
+                    ggml_cgraph* gf,
+                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                    const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        Plan plan;
+        plan.available    = true;
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        if (n_nodes <= 0) {
+            return plan;
+        }
+        plan.n_nodes = n_nodes;
+        plan.n_leafs = gf->n_leafs;
+        for (int i = 0; i < gf->n_leafs; ++i) {
+            ggml_tensor* leaf = gf->leafs[i];
+            if (is_params_tensor(params_tensor_set, leaf)) {
+                continue;
+            }
+            auto shape       = input_shape(leaf);
+            shape.leaf_index = i;
+            plan.input_shapes.push_back(shape);
+        }
+
+        std::unordered_map<const ggml_tensor*, int> producer_index;
+        producer_index.reserve(static_cast<size_t>(n_nodes));
+        for (int i = 0; i < n_nodes; ++i) {
+            producer_index[ggml_graph_node(gf, i)] = i;
+        }
+
+        std::vector<Segment> grouped_segments;
+        std::unordered_map<std::string, size_t> group_to_segment;
+        for (int i = 0; i < n_nodes; ++i) {
+            ggml_tensor* node = ggml_graph_node(gf, i);
+            if (!is_graph_cut_tensor(node)) {
+                continue;
+            }
+
+            plan.has_cuts = true;
+            std::string full_name(node->name);
+            std::string payload = full_name.substr(std::strlen(GGML_RUNNER_CUT_PREFIX));
+            size_t sep          = payload.find('|');
+            std::string group   = sep == std::string::npos ? payload : payload.substr(0, sep);
+
+            auto it = group_to_segment.find(group);
+            if (it == group_to_segment.end()) {
+                Segment segment;
+                segment.group_name = group;
+                segment.output_node_indices.push_back(i);
+                group_to_segment[group] = grouped_segments.size();
+                grouped_segments.push_back(std::move(segment));
+            } else {
+                auto& segment = grouped_segments[it->second];
+                segment.output_node_indices.push_back(i);
+            }
+        }
+
+        if (!plan.has_cuts) {
+            return plan;
+        }
+
+        std::unordered_set<int> available_cut_output_node_indices;
+        available_cut_output_node_indices.reserve(static_cast<size_t>(n_nodes));
+        for (auto& segment : grouped_segments) {
+            build_segment(gf,
+                          plan,
+                          segment,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+        }
+
+        int final_output_index = graph_node_index_by_name(gf, "ggml_runner_final_result_tensor");
+        if (final_output_index < 0) {
+            final_output_index = n_nodes - 1;
+        }
+        ggml_tensor* final_output = final_output_index >= 0 ? ggml_graph_node(gf, final_output_index) : nullptr;
+        if (final_output != nullptr && available_cut_output_node_indices.find(final_output_index) == available_cut_output_node_indices.end()) {
+            Segment final_segment;
+            final_segment.group_name = "ggml_runner.final";
+            final_segment.output_node_indices.push_back(final_output_index);
+            build_segment(gf,
+                          plan,
+                          final_segment,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+        }
+
+        return plan;
+    }
+
+    Plan apply_max_vram_budget(ggml_cgraph* gf,
+                               const Plan& base_plan,
+                               size_t max_graph_vram_bytes,
+                               ggml_backend_t backend,
+                               const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                               const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        int64_t t_budget_begin = ggml_time_ms();
+        if (max_graph_vram_bytes == 0 || !base_plan.has_cuts || base_plan.segments.size() <= 1) {
+            return base_plan;
+        }
+
+        const int n_nodes = ggml_graph_n_nodes(gf);
+        std::unordered_map<const ggml_tensor*, int> producer_index;
+        producer_index.reserve(static_cast<size_t>(n_nodes));
+        for (int i = 0; i < n_nodes; ++i) {
+            producer_index[ggml_graph_node(gf, i)] = i;
+        }
+
+        Plan merged_plan;
+        merged_plan.available = true;
+        merged_plan.has_cuts  = base_plan.has_cuts;
+        merged_plan.valid     = base_plan.valid;
+        merged_plan.n_nodes   = base_plan.n_nodes;
+        merged_plan.n_leafs   = base_plan.n_leafs;
+
+        std::unordered_set<int> available_cut_output_node_indices;
+        available_cut_output_node_indices.reserve(static_cast<size_t>(n_nodes));
+
+        size_t start_segment_index = 0;
+        while (start_segment_index < base_plan.segments.size()) {
+            Plan single_plan;
+            auto single_available_cut_output_node_indices = available_cut_output_node_indices;
+            auto single_seed                              = make_segment_seed(base_plan,
+                                                                              start_segment_index,
+                                                                              start_segment_index);
+            build_segment(gf,
+                          single_plan,
+                          single_seed,
+                          producer_index,
+                          single_available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+            GGML_ASSERT(!single_plan.segments.empty());
+
+            size_t best_end_segment_index = start_segment_index;
+            bool can_merge_next_segment   = graph_cut_segment_vram_bytes(single_plan.segments.back()) <= max_graph_vram_bytes;
+
+            while (can_merge_next_segment && best_end_segment_index + 1 < base_plan.segments.size()) {
+                const size_t next_end_segment_index = best_end_segment_index + 1;
+                Plan candidate_plan;
+                auto candidate_available_cut_output_node_indices = available_cut_output_node_indices;
+                auto candidate_seed                              = make_segment_seed(base_plan,
+                                                                                     start_segment_index,
+                                                                                     next_end_segment_index);
+                build_segment(gf,
+                              candidate_plan,
+                              candidate_seed,
+                              producer_index,
+                              candidate_available_cut_output_node_indices,
+                              backend,
+                              params_tensor_set,
+                              log_desc);
+                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) {
+                    break;
+                }
+
+                best_end_segment_index = next_end_segment_index;
+            }
+
+            auto best_seed = make_segment_seed(base_plan,
+                                               start_segment_index,
+                                               best_end_segment_index);
+            build_segment(gf,
+                          merged_plan,
+                          best_seed,
+                          producer_index,
+                          available_cut_output_node_indices,
+                          backend,
+                          params_tensor_set,
+                          log_desc);
+            start_segment_index = best_end_segment_index + 1;
+        }
+
+        if (log_desc != nullptr && merged_plan.segments.size() != base_plan.segments.size()) {
+            LOG_INFO("%s graph cut max_vram=%.2f MB merged %zu segments -> %zu segments",
+                     log_desc,
+                     max_graph_vram_bytes / 1024.0 / 1024.0,
+                     base_plan.segments.size(),
+                     merged_plan.segments.size());
+        }
+
+        if (log_desc != nullptr) {
+            LOG_INFO("%s graph cut max_vram budget merge took %lld ms",
+                     log_desc,
+                     ggml_time_ms() - t_budget_begin);
+        }
+
+        return merged_plan;
+    }
+
+    Plan resolve_plan(ggml_backend_t backend,
+                      ggml_cgraph* gf,
+                      PlanCache* cache,
+                      size_t max_graph_vram_bytes,
+                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                      const char* log_desc) {
+        GGML_ASSERT(backend != nullptr);
+        GGML_ASSERT(gf != nullptr);
+        GGML_ASSERT(cache != nullptr);
+
+        int64_t t_prepare_begin = ggml_time_ms();
+        Plan base_plan;
+        int64_t t_plan_begin = ggml_time_ms();
+        if (cache->graph_cut_plan.available && plan_matches_graph(gf, cache->graph_cut_plan)) {
+            base_plan = cache->graph_cut_plan;
+        } else {
+            base_plan                                = build_plan(backend, gf, params_tensor_set, log_desc);
+            cache->graph_cut_plan                    = base_plan;
+            cache->graph_cut_plan.available          = true;
+            cache->budgeted_graph_cut_plan.available = false;
+            if (log_desc != nullptr) {
+                LOG_INFO("%s build cached graph cut plan done (taking %lld ms)", log_desc, ggml_time_ms() - t_plan_begin);
+            }
+        }
+
+        Plan resolved_plan = base_plan;
+        if (max_graph_vram_bytes > 0 && base_plan.has_cuts) {
+            if (cache->budgeted_graph_cut_plan.available &&
+                cache->budgeted_graph_cut_plan_max_vram_bytes == max_graph_vram_bytes &&
+                plan_matches_graph(gf, cache->budgeted_graph_cut_plan)) {
+                resolved_plan = cache->budgeted_graph_cut_plan;
+            } else {
+                resolved_plan                                 = apply_max_vram_budget(gf,
+                                                                                      base_plan,
+                                                                                      max_graph_vram_bytes,
+                                                                                      backend,
+                                                                                      params_tensor_set,
+                                                                                      log_desc);
+                cache->budgeted_graph_cut_plan                = resolved_plan;
+                cache->budgeted_graph_cut_plan.available      = true;
+                cache->budgeted_graph_cut_plan_max_vram_bytes = max_graph_vram_bytes;
+            }
+        }
+        return resolved_plan;
+    }
+
+}  // namespace sd::ggml_graph_cut

src/ggml_graph_cut.h

@@ -0,0 +1,106 @@
+#ifndef __SD_GGML_GRAPH_CUT_H__
+#define __SD_GGML_GRAPH_CUT_H__
+
+#include <array>
+#include <string>
+#include <unordered_set>
+#include <vector>
+
+#include "ggml-backend.h"
+#include "ggml.h"
+
+namespace sd::ggml_graph_cut {
+
+    struct Segment {
+        enum InputType {
+            INPUT_EXTERNAL = 0,
+            INPUT_PREVIOUS_CUT,
+            INPUT_PARAM,
+        };
+
+        struct InputRef {
+            InputType type = INPUT_EXTERNAL;
+            std::string display_name;
+            int leaf_index = -1;
+            int node_index = -1;
+        };
+
+        size_t compute_buffer_size      = 0;
+        size_t output_bytes             = 0;
+        size_t input_external_bytes     = 0;
+        size_t input_previous_cut_bytes = 0;
+        size_t input_param_bytes        = 0;
+        std::string group_name;
+        std::vector<int> internal_node_indices;
+        std::vector<int> output_node_indices;
+        std::vector<InputRef> input_refs;
+    };
+
+    struct Plan {
+        struct InputShape {
+            int leaf_index                        = -1;
+            ggml_type type                        = GGML_TYPE_COUNT;
+            std::array<int64_t, GGML_MAX_DIMS> ne = {0, 0, 0, 0};
+        };
+
+        bool available = false;
+        bool has_cuts  = false;
+        bool valid     = true;
+        int n_nodes    = 0;
+        int n_leafs    = 0;
+        std::vector<InputShape> input_shapes;
+        std::vector<Segment> segments;
+    };
+
+    struct PlanCache {
+        Plan graph_cut_plan;
+        Plan budgeted_graph_cut_plan;
+        size_t budgeted_graph_cut_plan_max_vram_bytes = 0;
+    };
+
+    static constexpr const char* GGML_RUNNER_CUT_PREFIX = "ggml_runner_cut:";
+
+    bool is_graph_cut_tensor(const ggml_tensor* tensor);
+    std::string make_graph_cut_name(const std::string& group, const std::string& output);
+    void mark_graph_cut(ggml_tensor* tensor, const std::string& group, const std::string& output);
+    int leaf_count(ggml_cgraph* gf);
+    ggml_tensor* leaf_tensor(ggml_cgraph* gf, int leaf_index);
+    ggml_backend_buffer_t tensor_buffer(const ggml_tensor* tensor);
+    ggml_tensor* cache_source_tensor(ggml_tensor* tensor);
+    size_t cache_tensor_bytes(const ggml_tensor* tensor);
+    bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan);
+    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);
+    ggml_cgraph* build_segment_graph(ggml_cgraph* gf,
+                                     const Segment& segment,
+                                     ggml_context** graph_ctx_out);
+    size_t measure_segment_compute_buffer(ggml_backend_t backend,
+                                          ggml_cgraph* gf,
+                                          const Segment& segment,
+                                          const char* log_desc);
+    size_t max_vram_gib_to_bytes(float max_vram);
+    float resolve_max_vram_gib(float max_vram, ggml_backend_t backend);
+    Plan build_plan(ggml_backend_t backend,
+                    ggml_cgraph* gf,
+                    const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                    const char* log_desc);
+    Plan apply_max_vram_budget(ggml_cgraph* gf,
+                               const Plan& base_plan,
+                               size_t max_graph_vram_bytes,
+                               ggml_backend_t backend,
+                               const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                               const char* log_desc);
+    Plan resolve_plan(ggml_backend_t backend,
+                      ggml_cgraph* gf,
+                      PlanCache* cache,
+                      size_t max_graph_vram_bytes,
+                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                      const char* log_desc);
+}  // namespace sd::ggml_graph_cut
+
+#endif

src/hidream_o1.hpp

@@ -0,0 +1,653 @@
+#ifndef __SD_HIDREAM_O1_H__
+#define __SD_HIDREAM_O1_H__
+
+#include <algorithm>
+#include <array>
+#include <cmath>
+#include <cstring>
+#include <memory>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "common_dit.hpp"
+#include "conditioner.hpp"
+#include "llm.hpp"
+#include "util.h"
+
+namespace HiDreamO1 {
+    constexpr int HIDREAM_O1_GRAPH_SIZE = 32768;
+    constexpr int PATCH_SIZE            = 32;
+    constexpr int TIMESTEP_TOKEN_NUM    = 1;
+    constexpr int IMAGE_TOKEN_ID        = 151655;
+    constexpr int VISION_START_TOKEN_ID = 151652;
+
+    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());
+        for (int64_t i = 0; i < count; ++i) {
+            out += token;
+        }
+        return out;
+    }
+
+    static inline std::pair<int, int> calculate_dimensions(int max_size, double ratio) {
+        int width  = static_cast<int>(std::sqrt(max_size * max_size * ratio));
+        int height = static_cast<int>(width / ratio);
+        width      = (width / PATCH_SIZE) * PATCH_SIZE;
+        height     = (height / PATCH_SIZE) * PATCH_SIZE;
+        width      = std::max(width, PATCH_SIZE);
+        height     = std::max(height, PATCH_SIZE);
+        return {width, height};
+    }
+
+    static inline sd::Tensor<float> resize_to_area(const sd::Tensor<float>& image, int image_size) {
+        int64_t width  = image.shape()[0];
+        int64_t height = image.shape()[1];
+        int64_t s_max  = static_cast<int64_t>(image_size) * image_size;
+        double scale   = std::sqrt(static_cast<double>(s_max) / static_cast<double>(width * height));
+
+        std::vector<std::pair<int64_t, int64_t>> sizes = {
+            {(static_cast<int64_t>(std::llround(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::llround(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::llround(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::floor(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::floor(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::llround(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+            {(static_cast<int64_t>(std::floor(width * scale)) / PATCH_SIZE) * PATCH_SIZE, (static_cast<int64_t>(std::floor(height * scale)) / PATCH_SIZE) * PATCH_SIZE},
+        };
+        std::sort(sizes.begin(), sizes.end(), [](const auto& a, const auto& b) {
+            return a.first * a.second > b.first * b.second;
+        });
+
+        std::pair<int64_t, int64_t> new_size = sizes.back();
+        for (const auto& size : sizes) {
+            if (size.first > 0 && size.second > 0 && size.first * size.second <= s_max) {
+                new_size = size;
+                break;
+            }
+        }
+
+        double s1 = static_cast<double>(width) / static_cast<double>(new_size.first);
+        double s2 = static_cast<double>(height) / static_cast<double>(new_size.second);
+        sd::Tensor<float> resized;
+        if (s1 < s2) {
+            int64_t resized_h = static_cast<int64_t>(std::llround(height / s1));
+            resized           = sd::ops::interpolate(image,
+                                                     {new_size.first, resized_h, image.shape()[2], image.shape()[3]},
+                                                     sd::ops::InterpolateMode::Bicubic);
+            int64_t top       = (resized_h - new_size.second) / 2;
+            resized           = sd::ops::slice(resized, 1, top, top + new_size.second);
+        } else {
+            int64_t resized_w = static_cast<int64_t>(std::llround(width / s2));
+            resized           = sd::ops::interpolate(image,
+                                                     {resized_w, new_size.second, image.shape()[2], image.shape()[3]},
+                                                     sd::ops::InterpolateMode::Bicubic);
+            int64_t left      = (resized_w - new_size.first) / 2;
+            resized           = sd::ops::slice(resized, 0, left, left + new_size.first);
+        }
+        return resized;
+    }
+
+    static inline std::vector<int32_t> build_position_ids(const std::vector<int32_t>& input_ids,
+                                                          const std::vector<std::array<int32_t, 3>>& image_grids,
+                                                          const std::vector<int32_t>& skip_vision_start_token) {
+        std::vector<int32_t> position_ids(4 * input_ids.size(), 0);
+        int image_index = 0;
+        int st          = 0;
+        int fix_point   = 4096;
+        std::vector<int32_t> out_t;
+        std::vector<int32_t> out_h;
+        std::vector<int32_t> out_w;
+
+        while (st < static_cast<int>(input_ids.size())) {
+            int ed = st;
+            while (ed < static_cast<int>(input_ids.size()) && input_ids[ed] != IMAGE_TOKEN_ID) {
+                ed++;
+            }
+
+            if (ed >= static_cast<int>(input_ids.size())) {
+                int st_idx = out_t.empty() ? 0 : (*std::max_element(out_t.begin(), out_t.end()) + 1);
+                for (int i = 0; i < static_cast<int>(input_ids.size()) - st; ++i) {
+                    out_t.push_back(st_idx + i);
+                    out_h.push_back(st_idx + i);
+                    out_w.push_back(st_idx + i);
+                }
+                break;
+            }
+
+            int text_len = std::max(0, ed - st - skip_vision_start_token[image_index]);
+            int st_idx   = out_t.empty() ? 0 : (*std::max_element(out_t.begin(), out_t.end()) + 1);
+            for (int i = 0; i < text_len; ++i) {
+                out_t.push_back(st_idx + i);
+                out_h.push_back(st_idx + i);
+                out_w.push_back(st_idx + i);
+            }
+
+            auto grid = image_grids[image_index];
+            int base;
+            if (skip_vision_start_token[image_index]) {
+                if (fix_point > 0) {
+                    base      = fix_point;
+                    fix_point = 0;
+                } else {
+                    base = st_idx;
+                }
+            } else {
+                base = text_len + st_idx;
+            }
+            for (int32_t ti = 0; ti < grid[0]; ++ti) {
+                for (int32_t hi = 0; hi < grid[1]; ++hi) {
+                    for (int32_t wi = 0; wi < grid[2]; ++wi) {
+                        out_t.push_back(base + ti);
+                        out_h.push_back(base + hi);
+                        out_w.push_back(base + wi);
+                    }
+                }
+            }
+
+            st = ed + grid[0] * grid[1] * grid[2];
+            image_index++;
+        }
+
+        GGML_ASSERT(out_t.size() == input_ids.size());
+        for (size_t i = 0; i < input_ids.size(); ++i) {
+            // ggml IMROPE consumes 4 flattened position streams:
+            //   [t, h, w, e]
+            // llama.cpp's generic Qwen-VL fallback expands text positions as
+            // [pos, pos, pos, 0]. Keep the extra stream zeroed here too.
+            position_ids[i]                        = out_t[i];
+            position_ids[input_ids.size() + i]     = out_h[i];
+            position_ids[input_ids.size() * 2 + i] = out_w[i];
+            position_ids[input_ids.size() * 3 + i] = 0;
+        }
+        return position_ids;
+    }
+
+    struct TimestepEmbedder : public GGMLBlock {
+        int frequency_embedding_size = 256;
+
+        TimestepEmbedder(int64_t hidden_size) {
+            blocks["mlp.0"] = std::make_shared<Linear>(frequency_embedding_size, hidden_size, true);
+            blocks["mlp.2"] = std::make_shared<Linear>(hidden_size, hidden_size, 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 emb   = ggml_ext_timestep_embedding(ctx->ggml_ctx, t, frequency_embedding_size, 10000, 1000.0f);
+            emb        = mlp_0->forward(ctx, emb);
+            emb        = ggml_silu_inplace(ctx->ggml_ctx, emb);
+            emb        = mlp_2->forward(ctx, emb);
+            return emb;
+        }
+    };
+
+    struct BottleneckPatchEmbed : public GGMLBlock {
+        BottleneckPatchEmbed(int64_t in_dim, int64_t pca_dim, int64_t embed_dim) {
+            blocks["proj1"] = std::make_shared<Linear>(in_dim, pca_dim, false);
+            blocks["proj2"] = std::make_shared<Linear>(pca_dim, embed_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto proj1 = std::dynamic_pointer_cast<Linear>(blocks["proj1"]);
+            auto proj2 = std::dynamic_pointer_cast<Linear>(blocks["proj2"]);
+            return proj2->forward(ctx, proj1->forward(ctx, x));
+        }
+    };
+
+    struct FinalLayer : public GGMLBlock {
+        FinalLayer(int64_t hidden_size, int64_t out_dim) {
+            blocks["linear"] = std::make_shared<Linear>(hidden_size, out_dim, true);
+        }
+
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+            auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
+            return linear->forward(ctx, x);
+        }
+    };
+
+    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;
+
+        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);
+        }
+
+        std::shared_ptr<LLM::TextModel> text_model() {
+            return std::dynamic_pointer_cast<LLM::TextModel>(blocks["language_model"]);
+        }
+
+        std::shared_ptr<TimestepEmbedder> timestep_embedder() {
+            return std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder1"]);
+        }
+
+        std::shared_ptr<BottleneckPatchEmbed> patch_embedder() {
+            return std::dynamic_pointer_cast<BottleneckPatchEmbed>(blocks["x_embedder"]);
+        }
+
+        std::shared_ptr<FinalLayer> final_layer() {
+            return std::dynamic_pointer_cast<FinalLayer>(blocks["final_layer2"]);
+        }
+    };
+
+    struct HiDreamO1VisionRunner : public GGMLRunner {
+        HiDreamO1Params params;
+        std::shared_ptr<LLM::VisionModel> model;
+
+        std::vector<int> window_index_vec;
+        std::vector<int> window_inverse_index_vec;
+        std::vector<float> window_mask_vec;
+        std::vector<float> pe_vec;
+        std::array<std::vector<int32_t>, 4> pos_embed_idx_data_;
+        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)) {
+            model->init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "hidream_o1_vision";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix = "model.visual") {
+            model->get_param_tensors(tensors, prefix);
+        }
+
+        ggml_tensor* encode_image(GGMLRunnerContext* runner_ctx, ggml_tensor* image) {
+            return LLM::LLMRunner::encode_image_common(this,
+                                                       compute_ctx,
+                                                       runner_ctx,
+                                                       image,
+                                                       params.llm.vision,
+                                                       model,
+                                                       window_index_vec,
+                                                       window_inverse_index_vec,
+                                                       window_mask_vec,
+                                                       pe_vec,
+                                                       pos_embed_idx_data_,
+                                                       pos_embed_weight_data_);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& image_tensor) {
+            ggml_cgraph* gf    = new_graph_custom(HIDREAM_O1_GRAPH_SIZE);
+            ggml_tensor* image = make_input(image_tensor);
+            auto runner_ctx    = get_context();
+            auto image_embeds  = encode_image(&runner_ctx, image);
+            ggml_build_forward_expand(gf, image_embeds);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& image) {
+            auto get_graph = [&]() {
+                return build_graph(image);
+            };
+            auto output = GGMLRunner::compute<float>(get_graph, n_threads, false);
+            return output.has_value() ? std::move(output.value()) : sd::Tensor<float>();
+        }
+    };
+
+    struct HiDreamO1Runner : public GGMLRunner {
+        HiDreamO1Params params;
+        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);
+            model.init(params_ctx, tensor_storage_map, prefix);
+        }
+
+        std::string get_desc() override {
+            return "hidream_o1";
+        }
+
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
+            model.get_param_tensors(tensors, prefix);
+        }
+
+        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
+                                 const sd::Tensor<float>& timestep_tensor,
+                                 const sd::Tensor<int32_t>& input_ids_tensor,
+                                 const sd::Tensor<int32_t>& input_pos_tensor,
+                                 const sd::Tensor<int32_t>& token_types_tensor,
+                                 const sd::Tensor<int32_t>& vinput_mask_tensor,
+                                 const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds_tensor,
+                                 const std::vector<sd::Tensor<float>>& ref_images) {
+            ggml_cgraph* gf        = new_graph_custom(HIDREAM_O1_GRAPH_SIZE);
+            ggml_tensor* x         = make_input(x_tensor);
+            ggml_tensor* timestep  = make_input(timestep_tensor);
+            ggml_tensor* input_ids = make_input(input_ids_tensor);
+            ggml_tensor* input_pos = make_input(input_pos_tensor);
+
+            auto text_model   = model.text_model();
+            auto t_embedder1  = model.timestep_embedder();
+            auto x_embedder   = model.patch_embedder();
+            auto final_layer2 = model.final_layer();
+
+            std::vector<ggml_tensor*> ref_image_tensors;
+            for (const auto& image : ref_images) {
+                ref_image_tensors.push_back(make_input(image));
+            }
+
+            attention_mask_vec    = std::vector<float>(static_cast<size_t>(token_types_tensor.shape()[0] * token_types_tensor.shape()[0]), 0.0f);
+            int64_t total_seq_len = token_types_tensor.shape()[0];
+            for (int64_t query = 0; query < total_seq_len; ++query) {
+                bool is_gen = token_types_tensor.values()[static_cast<size_t>(query)] > 0;
+                for (int64_t key = 0; key < total_seq_len; ++key) {
+                    if (!is_gen && key > query) {
+                        attention_mask_vec[static_cast<size_t>(query * total_seq_len + key)] = -INFINITY;
+                    }
+                }
+            }
+            auto attention_mask = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, total_seq_len, total_seq_len);
+            set_backend_tensor_data(attention_mask, attention_mask_vec.data());
+
+            auto runner_ctx = get_context();
+            auto txt        = text_model->embed(&runner_ctx, input_ids);
+            std::vector<std::pair<int, ggml_tensor*>> image_embeds;
+            image_embeds.reserve(image_embeds_tensor.size());
+            for (const auto& image_embed : image_embeds_tensor) {
+                image_embeds.emplace_back(image_embed.first, make_input(image_embed.second));
+            }
+            txt = LLM::splice_image_embeds(&runner_ctx, txt, image_embeds);
+
+            auto t_emb          = t_embedder1->forward(&runner_ctx, timestep);
+            int64_t txt_seq_len = input_ids->ne[0];
+            if (txt_seq_len > 1) {
+                auto prefix = ggml_ext_slice(compute_ctx, txt, 1, 0, txt_seq_len - 1);
+                txt         = ggml_concat(compute_ctx, prefix, ggml_reshape_3d(compute_ctx, t_emb, t_emb->ne[0], 1, 1), 1);
+            } else {
+                txt = ggml_reshape_3d(compute_ctx, t_emb, t_emb->ne[0], 1, 1);
+            }
+
+            auto vinputs          = DiT::pad_and_patchify(&runner_ctx, x, PATCH_SIZE, PATCH_SIZE);
+            int64_t target_tokens = vinputs->ne[1];
+            for (ggml_tensor* ref_image : ref_image_tensors) {
+                auto ref = DiT::pad_and_patchify(&runner_ctx, ref_image, PATCH_SIZE, PATCH_SIZE);
+                vinputs  = ggml_concat(compute_ctx, vinputs, ref, 1);
+            }
+            auto vis = x_embedder->forward(&runner_ctx, vinputs);
+
+            auto inputs_embeds = ggml_concat(compute_ctx, txt, vis, 1);
+            auto hidden_states = text_model->forward_embeds(&runner_ctx, inputs_embeds, input_pos, attention_mask, {});
+            auto x_pred_all    = final_layer2->forward(&runner_ctx, hidden_states);
+
+            int64_t x_pred_start = txt_seq_len;
+            if (!vinput_mask_tensor.empty()) {
+                int64_t seq_len      = static_cast<int64_t>(vinput_mask_tensor.shape()[0]);
+                int64_t first_vinput = 0;
+                while (first_vinput < seq_len && vinput_mask_tensor.values()[static_cast<size_t>(first_vinput)] == 0) {
+                    first_vinput++;
+                }
+                x_pred_start = first_vinput;
+            }
+            auto x_pred = ggml_ext_slice(compute_ctx, x_pred_all, 1, x_pred_start, x_pred_start + target_tokens);
+            x_pred      = DiT::unpatchify_and_crop(compute_ctx, x_pred, x->ne[1], x->ne[0], PATCH_SIZE, PATCH_SIZE);
+
+            float sigma = 1.0f - timestep_tensor.values()[0];
+            sigma       = std::max(1e-6f, sigma);
+            auto out    = ggml_scale(compute_ctx, ggml_sub(compute_ctx, x, x_pred), 1.0f / sigma);
+
+            ggml_build_forward_expand(gf, out);
+            return gf;
+        }
+
+        sd::Tensor<float> compute(int n_threads,
+                                  const sd::Tensor<float>& x,
+                                  const sd::Tensor<float>& timestep,
+                                  const sd::Tensor<int32_t>& input_ids,
+                                  const sd::Tensor<int32_t>& input_pos,
+                                  const sd::Tensor<int32_t>& token_types,
+                                  const sd::Tensor<int32_t>& vinput_mask,
+                                  const std::vector<std::pair<int, sd::Tensor<float>>>& image_embeds,
+                                  const std::vector<sd::Tensor<float>>& ref_images) {
+            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());
+        }
+    };
+
+    struct HiDreamO1Conditioner : public Conditioner {
+        Qwen2Tokenizer tokenizer;
+        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)) {}
+
+        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);
+        }
+
+        void set_flash_attention_enabled(bool enabled) override {
+            vision_runner->set_flash_attention_enabled(enabled);
+        }
+
+        void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
+            vision_runner->set_weight_adapter(adapter);
+        }
+
+        SDCondition get_learned_condition(int n_threads,
+                                          const ConditionerParams& conditioner_params) override {
+            SDCondition result;
+
+            int width                = conditioner_params.width;
+            int height               = conditioner_params.height;
+            int64_t target_image_len = static_cast<int64_t>(width / PATCH_SIZE) * static_cast<int64_t>(height / PATCH_SIZE);
+
+            std::vector<sd::Tensor<float>> ref_images;
+            if (conditioner_params.ref_images != nullptr) {
+                ref_images = *conditioner_params.ref_images;
+            }
+
+            std::vector<std::pair<int, sd::Tensor<float>>> vlm_images;
+            std::vector<std::array<int32_t, 3>> image_grids;
+            std::vector<int32_t> skip_vision_start;
+
+            std::string prompt = "<|im_start|>user\n";
+
+            if (ref_images.empty()) {
+                prompt += conditioner_params.text;
+                prompt += "<|im_end|>\n<|im_start|>assistant\n<|boi_token|><|tms_token|>";
+                auto input_ids = tokenizer.encode(prompt, nullptr);
+
+                std::vector<int32_t> input_ids_pad = input_ids;
+                input_ids_pad.push_back(VISION_START_TOKEN_ID);
+                input_ids_pad.insert(input_ids_pad.end(), target_image_len - 1, IMAGE_TOKEN_ID);
+
+                image_grids.push_back({1, static_cast<int32_t>(height / PATCH_SIZE), static_cast<int32_t>(width / PATCH_SIZE)});
+                skip_vision_start.push_back(1);
+
+                std::vector<int32_t> token_types(input_ids_pad.size(), 0);
+                int txt_seq_len = static_cast<int>(input_ids.size());
+                int bgn         = txt_seq_len - TIMESTEP_TOKEN_NUM;
+                for (int i = bgn; i < static_cast<int>(token_types.size()); ++i) {
+                    token_types[i] = 1;
+                }
+
+                auto position_ids = build_position_ids(input_ids_pad, image_grids, skip_vision_start);
+
+                std::vector<int64_t> input_shape{static_cast<int64_t>(input_ids.size())};
+                std::vector<int64_t> position_shape{static_cast<int64_t>(input_ids_pad.size() * 4)};
+                std::vector<int64_t> token_type_shape{static_cast<int64_t>(token_types.size())};
+                std::vector<int32_t> vinput_mask(token_types.size(), 0);
+                for (int64_t i = txt_seq_len; i < static_cast<int64_t>(vinput_mask.size()); ++i) {
+                    vinput_mask[static_cast<size_t>(i)] = 1;
+                }
+                std::vector<int64_t> vinput_mask_shape{static_cast<int64_t>(vinput_mask.size())};
+
+                result.c_input_ids    = sd::Tensor<int32_t>(input_shape, std::move(input_ids));
+                result.c_position_ids = sd::Tensor<int32_t>(position_shape, position_ids);
+                result.c_token_types  = sd::Tensor<int32_t>(token_type_shape, std::move(token_types));
+                result.c_vinput_mask  = sd::Tensor<int32_t>(vinput_mask_shape, std::move(vinput_mask));
+                return result;
+            }
+
+            int K = static_cast<int>(ref_images.size());
+            int max_size;
+            if (K == 1) {
+                max_size = std::max(height, width);
+            } else if (K == 2) {
+                max_size = std::max(height, width) * 48 / 64;
+            } else if (K <= 4) {
+                max_size = std::max(height, width) / 2;
+            } else if (K <= 8) {
+                max_size = std::max(height, width) * 24 / 64;
+            } else {
+                max_size = std::max(height, width) / 4;
+            }
+
+            int cond_img_size;
+            if (K <= 4) {
+                cond_img_size = 384;
+            } else if (K <= 8) {
+                cond_img_size = 384 * 48 / 64;
+            } else {
+                cond_img_size = 384 / 2;
+            }
+
+            for (const auto& ref_image : ref_images) {
+                auto resized_ref = resize_to_area(ref_image, max_size);
+                resized_ref      = sd::ops::clamp(resized_ref, 0.0f, 1.0f);
+
+                // VLM image: Qwen3-VL expects mean=[0.5]/std=[0.5] (i.e. range [-1,1]),
+                // not CLIP normalization. Resize the already-resized ref directly to
+                // (cond_w, cond_h) to match the Python pipeline's pil_r.resize().
+                auto dims                   = calculate_dimensions(cond_img_size,
+                                                                   static_cast<double>(resized_ref.shape()[0]) / static_cast<double>(resized_ref.shape()[1]));
+                sd::Tensor<float> vlm_image = sd::ops::interpolate(
+                    resized_ref,
+                    {dims.first, dims.second, resized_ref.shape()[2], resized_ref.shape()[3]});
+                vlm_image            = vlm_image * 2.0f - 1.0f;
+                int64_t image_tokens = static_cast<int64_t>(dims.first / PATCH_SIZE) * static_cast<int64_t>(dims.second / PATCH_SIZE);
+
+                auto patch_img = resized_ref * 2.0f - 1.0f;
+                result.c_ref_images.push_back(std::move(patch_img));
+                int64_t prompt_start = static_cast<int64_t>(tokenizer.encode(prompt + "<|vision_start|>", nullptr).size());
+                prompt += "<|vision_start|>";
+                prompt += repeat_special_token("<|image_pad|>", image_tokens);
+                prompt += "<|vision_end|>";
+                vlm_images.emplace_back(static_cast<int>(prompt_start), std::move(vlm_image));
+                image_grids.push_back({1, dims.second / PATCH_SIZE, dims.first / PATCH_SIZE});
+                skip_vision_start.push_back(0);
+            }
+
+            prompt += conditioner_params.text;
+            prompt += "<|im_end|>\n<|im_start|>assistant\n<|boi_token|><|tms_token|>";
+            auto input_ids = tokenizer.encode(prompt, nullptr);
+
+            std::vector<int32_t> input_ids_pad = input_ids;
+            input_ids_pad.push_back(VISION_START_TOKEN_ID);
+            input_ids_pad.insert(input_ids_pad.end(), target_image_len - 1, IMAGE_TOKEN_ID);
+            image_grids.push_back({1, static_cast<int32_t>(height / PATCH_SIZE), static_cast<int32_t>(width / PATCH_SIZE)});
+            skip_vision_start.push_back(1);
+
+            for (const auto& ref_image : result.c_ref_images) {
+                int64_t ref_len = static_cast<int64_t>(ref_image.shape()[0] / PATCH_SIZE) * static_cast<int64_t>(ref_image.shape()[1] / PATCH_SIZE);
+                input_ids_pad.push_back(VISION_START_TOKEN_ID);
+                input_ids_pad.insert(input_ids_pad.end(), ref_len - 1, IMAGE_TOKEN_ID);
+                image_grids.push_back({1, static_cast<int32_t>(ref_image.shape()[1] / PATCH_SIZE), static_cast<int32_t>(ref_image.shape()[0] / PATCH_SIZE)});
+                skip_vision_start.push_back(1);
+            }
+
+            std::vector<int32_t> token_types(input_ids_pad.size(), 0);
+            int txt_seq_len = static_cast<int>(input_ids.size());
+            int bgn         = txt_seq_len - TIMESTEP_TOKEN_NUM;
+            for (int i = bgn; i < static_cast<int>(token_types.size()); ++i) {
+                token_types[i] = 1;
+            }
+
+            std::vector<int64_t> input_shape{static_cast<int64_t>(input_ids.size())};
+            std::vector<int64_t> position_shape{static_cast<int64_t>(input_ids_pad.size() * 4)};
+            std::vector<int64_t> token_type_shape{static_cast<int64_t>(token_types.size())};
+            std::vector<int32_t> vinput_mask(token_types.size(), 0);
+            for (int i = txt_seq_len; i < static_cast<int>(vinput_mask.size()); ++i) {
+                vinput_mask[static_cast<size_t>(i)] = 1;
+            }
+            std::vector<int64_t> vinput_mask_shape{static_cast<int64_t>(vinput_mask.size())};
+
+            result.c_input_ids    = sd::Tensor<int32_t>(input_shape, std::move(input_ids));
+            result.c_position_ids = sd::Tensor<int32_t>(position_shape, build_position_ids(input_ids_pad, image_grids, skip_vision_start));
+            result.c_token_types  = sd::Tensor<int32_t>(token_type_shape, std::move(token_types));
+            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);
+                if (image_embed.empty()) {
+                    LOG_ERROR("hidream_o1 conditioner: encode VLM image failed");
+                    return SDCondition();
+                }
+                result.c_image_embeds.emplace_back(vlm_image.first, std::move(image_embed));
+            }
+            return result;
+        }
+    };
+}  // namespace HiDreamO1
+
+#endif  // __SD_HIDREAM_O1_H__

src/lora.hpp

@@ -22,10 +22,11 @@ struct LoraModel : public GGMLRunner {
 
     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, false) {
+        : lora_id(lora_id), file_path(file_path), GGMLRunner(backend, params_backend) {
         prefix = "lora." + prefix;
         if (!model_loader.init_from_file_and_convert_name(file_path, prefix, version)) {
             load_failed = true;
@@ -129,7 +130,7 @@ struct LoraModel : public GGMLRunner {
         }
     }
 
-    ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -152,17 +153,17 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(lora_up_name);
             if (iter != lora_tensors.end()) {
-                lora_up = ggml_ext_cast_f32(ctx, iter->second);
+                lora_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lora_mid_name);
             if (iter != lora_tensors.end()) {
-                lora_mid = ggml_ext_cast_f32(ctx, iter->second);
+                lora_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lora_down_name);
             if (iter != lora_tensors.end()) {
-                lora_down = ggml_ext_cast_f32(ctx, iter->second);
+                lora_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             if (lora_up == nullptr || lora_down == nullptr) {
@@ -208,7 +209,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_raw_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_raw_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -225,7 +226,7 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(diff_name);
             if (iter != lora_tensors.end()) {
-                curr_updown = ggml_ext_cast_f32(ctx, iter->second);
+                curr_updown = ggml_ext_cast_f32(ctx, backend, iter->second);
             } else {
                 break;
             }
@@ -248,7 +249,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_loha_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_loha_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -276,33 +277,33 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(hada_1_down_name);
             if (iter != lora_tensors.end()) {
-                hada_1_down = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_1_up_name);
             if (iter != lora_tensors.end()) {
-                hada_1_up = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_1_mid_name);
             if (iter != lora_tensors.end()) {
-                hada_1_mid = ggml_ext_cast_f32(ctx, iter->second);
+                hada_1_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
                 hada_1_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_1_up));
             }
 
             iter = lora_tensors.find(hada_2_down_name);
             if (iter != lora_tensors.end()) {
-                hada_2_down = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_down = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_2_up_name);
             if (iter != lora_tensors.end()) {
-                hada_2_up = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_up = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(hada_2_mid_name);
             if (iter != lora_tensors.end()) {
-                hada_2_mid = ggml_ext_cast_f32(ctx, iter->second);
+                hada_2_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
                 hada_2_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_2_up));
             }
 
@@ -351,7 +352,7 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_lokr_weight_diff(const std::string& model_tensor_name, ggml_context* ctx) {
+    ggml_tensor* get_lokr_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {
         ggml_tensor* updown = nullptr;
         int index           = 0;
         while (true) {
@@ -378,24 +379,24 @@ struct LoraModel : public GGMLRunner {
 
             auto iter = lora_tensors.find(lokr_w1_name);
             if (iter != lora_tensors.end()) {
-                lokr_w1 = ggml_ext_cast_f32(ctx, iter->second);
+                lokr_w1 = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             iter = lora_tensors.find(lokr_w2_name);
             if (iter != lora_tensors.end()) {
-                lokr_w2 = ggml_ext_cast_f32(ctx, iter->second);
+                lokr_w2 = ggml_ext_cast_f32(ctx, backend, iter->second);
             }
 
             int64_t rank = 1;
             if (lokr_w1 == nullptr) {
                 iter = lora_tensors.find(lokr_w1_a_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w1_a = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w1_a = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 iter = lora_tensors.find(lokr_w1_b_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w1_b = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w1_b = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 if (lokr_w1_a == nullptr || lokr_w1_b == nullptr) {
@@ -410,12 +411,12 @@ struct LoraModel : public GGMLRunner {
             if (lokr_w2 == nullptr) {
                 iter = lora_tensors.find(lokr_w2_a_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w2_a = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w2_a = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 iter = lora_tensors.find(lokr_w2_b_name);
                 if (iter != lora_tensors.end()) {
-                    lokr_w2_b = ggml_ext_cast_f32(ctx, iter->second);
+                    lokr_w2_b = ggml_ext_cast_f32(ctx, backend, iter->second);
                 }
 
                 if (lokr_w2_a == nullptr || lokr_w2_b == nullptr) {
@@ -468,23 +469,23 @@ struct LoraModel : public GGMLRunner {
         return updown;
     }
 
-    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
+    ggml_tensor* get_weight_diff(const std::string& model_tensor_name, ggml_backend_t backend, ggml_context* ctx, ggml_tensor* model_tensor, bool with_lora_and_lokr = true) {
         // lora
         ggml_tensor* diff = nullptr;
         if (with_lora_and_lokr) {
-            diff = get_lora_weight_diff(model_tensor_name, ctx);
+            diff = get_lora_weight_diff(model_tensor_name, ctx, backend);
         }
         // diff
         if (diff == nullptr) {
-            diff = get_raw_weight_diff(model_tensor_name, ctx);
+            diff = get_raw_weight_diff(model_tensor_name, ctx, backend);
         }
         // loha
         if (diff == nullptr) {
-            diff = get_loha_weight_diff(model_tensor_name, ctx);
+            diff = get_loha_weight_diff(model_tensor_name, ctx, backend);
         }
         // lokr
         if (diff == nullptr && with_lora_and_lokr) {
-            diff = get_lokr_weight_diff(model_tensor_name, ctx);
+            diff = get_lokr_weight_diff(model_tensor_name, ctx, backend);
         }
         if (diff != nullptr) {
             if (ggml_nelements(diff) < ggml_nelements(model_tensor)) {
@@ -502,6 +503,7 @@ struct LoraModel : public GGMLRunner {
     }
 
     ggml_tensor* get_out_diff(ggml_context* ctx,
+                              ggml_backend_t backend,
                               ggml_tensor* x,
                               WeightAdapter::ForwardParams forward_params,
                               const std::string& model_tensor_name) {
@@ -590,7 +592,7 @@ struct LoraModel : public GGMLRunner {
                 }
                 scale_value *= multiplier;
 
-                auto curr_out_diff = ggml_ext_lokr_forward(ctx, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
+                auto curr_out_diff = ggml_ext_lokr_forward(ctx, backend, x, lokr_w1, lokr_w1_a, lokr_w1_b, lokr_w2, lokr_w2_a, lokr_w2_b, is_conv2d, forward_params.conv2d, scale_value);
                 if (out_diff == nullptr) {
                     out_diff = curr_out_diff;
                 } else {
@@ -761,7 +763,7 @@ struct LoraModel : public GGMLRunner {
             ggml_tensor* model_tensor     = it.second;
 
             // lora
-            ggml_tensor* diff = get_weight_diff(model_tensor_name, compute_ctx, model_tensor);
+            ggml_tensor* diff = get_weight_diff(model_tensor_name, runtime_backend, compute_ctx, model_tensor);
             if (diff == nullptr) {
                 continue;
             }
@@ -774,7 +776,7 @@ struct LoraModel : public GGMLRunner {
 
             ggml_tensor* final_tensor;
             if (model_tensor->type != GGML_TYPE_F32 && model_tensor->type != GGML_TYPE_F16) {
-                final_tensor = ggml_ext_cast_f32(compute_ctx, model_tensor);
+                final_tensor = ggml_ext_cast_f32(compute_ctx, runtime_backend, model_tensor);
                 final_tensor = ggml_add_inplace(compute_ctx, final_tensor, diff);
                 final_tensor = ggml_cpy(compute_ctx, final_tensor, model_tensor);
             } else {
@@ -841,34 +843,35 @@ public:
         : lora_models(lora_models) {
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* weight, const std::string& weight_name, bool with_lora_and_lokr) {
         for (auto& lora_model : lora_models) {
-            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, ctx, weight, with_lora_and_lokr);
+            ggml_tensor* diff = lora_model->get_weight_diff(weight_name, backend, ctx, weight, with_lora_and_lokr);
             if (diff == nullptr) {
                 continue;
             }
 
             if (weight->type != GGML_TYPE_F32 && weight->type != GGML_TYPE_F16) {
-                weight = ggml_ext_cast_f32(ctx, weight);
+                weight = ggml_ext_cast_f32(ctx, backend, weight);
             }
             weight = ggml_add(ctx, weight, diff);
         }
         return weight;
     }
 
-    ggml_tensor* patch_weight(ggml_context* ctx, ggml_tensor* weight, const std::string& weight_name) override {
-        return patch_weight(ctx, weight, weight_name, true);
+    ggml_tensor* patch_weight(ggml_context* ctx, ggml_backend_t backend, ggml_tensor* weight, const std::string& weight_name) override {
+        return patch_weight(ctx, backend, weight, weight_name, true);
     }
 
     ggml_tensor* forward_with_lora(ggml_context* ctx,
+                                   ggml_backend_t backend,
                                    ggml_tensor* x,
                                    ggml_tensor* w,
                                    ggml_tensor* b,
                                    const std::string& prefix,
                                    WeightAdapter::ForwardParams forward_params) override {
-        w = patch_weight(ctx, w, prefix + "weight", false);
+        w = patch_weight(ctx, backend, w, prefix + "weight", false);
         if (b) {
-            b = patch_weight(ctx, b, prefix + "bias", false);
+            b = patch_weight(ctx, backend, b, prefix + "bias", false);
         }
         ggml_tensor* out;
         if (forward_params.op_type == ForwardParams::op_type_t::OP_LINEAR) {
@@ -890,7 +893,7 @@ public:
                                    forward_params.conv2d.scale);
         }
         for (auto& lora_model : lora_models) {
-            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, x, forward_params, prefix + "weight");
+            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, backend, x, forward_params, prefix + "weight");
             if (out_diff == nullptr) {
                 continue;
             }

src/mmdit.hpp

@@ -767,6 +767,8 @@ public:
             auto context_x = block->forward(ctx, context, x, c_mod);
             context        = context_x.first;
             x              = context_x.second;
+            sd::ggml_graph_cut::mark_graph_cut(context, "mmdit.joint_blocks." + std::to_string(i), "context");
+            sd::ggml_graph_cut::mark_graph_cut(x, "mmdit.joint_blocks." + std::to_string(i), "x");
         }
 
         x = final_layer->forward(ctx, x, c_mod);  // (N, T, patch_size ** 2 * out_channels)
@@ -809,6 +811,11 @@ public:
 
             context = context_embedder->forward(ctx, context);  // [N, L, D] aka [N, L, 1536]
         }
+        sd::ggml_graph_cut::mark_graph_cut(x, "mmdit.prelude", "x");
+        sd::ggml_graph_cut::mark_graph_cut(c, "mmdit.prelude", "c");
+        if (context != nullptr) {
+            sd::ggml_graph_cut::mark_graph_cut(context, "mmdit.prelude", "context");
+        }
 
         x = forward_core_with_concat(ctx, x, c, context, skip_layers);  // (N, H*W, patch_size ** 2 * out_channels)
 
@@ -821,10 +828,10 @@ struct MMDiTRunner : public GGMLRunner {
     MMDiT mmdit;
 
     MMDiTRunner(ggml_backend_t backend,
-                bool offload_params_to_cpu,
+                ggml_backend_t params_backend,
                 const String2TensorStorage& tensor_storage_map = {},
                 const std::string prefix                       = "")
-        : GGMLRunner(backend, offload_params_to_cpu), mmdit(tensor_storage_map) {
+        : GGMLRunner(backend, params_backend), mmdit(tensor_storage_map) {
         mmdit.init(params_ctx, tensor_storage_map, prefix);
     }
 
@@ -927,7 +934,7 @@ struct MMDiTRunner : public GGMLRunner {
         // ggml_backend_t backend    = ggml_backend_cuda_init(0);
         ggml_backend_t backend             = ggml_backend_cpu_init();
         ggml_type model_data_type          = GGML_TYPE_F16;
-        std::shared_ptr<MMDiTRunner> mmdit = std::make_shared<MMDiTRunner>(backend, false);
+        std::shared_ptr<MMDiTRunner> mmdit = std::make_shared<MMDiTRunner>(backend, backend);
         {
             LOG_INFO("loading from '%s'", file_path.c_str());
 

src/model.h

@@ -5,20 +5,13 @@
 #include <map>
 #include <memory>
 #include <set>
-#include <sstream>
 #include <string>
-#include <tuple>
-#include <utility>
 #include <vector>
 
 #include "ggml-backend.h"
 #include "ggml.h"
-#include "gguf.h"
-#include "json.hpp"
+#include "model_io/tensor_storage.h"
 #include "ordered_map.hpp"
-#include "zip.h"
-
-#define SD_MAX_DIMS 5
 
 enum SDVersion {
     VERSION_SD1,
@@ -49,6 +42,7 @@ enum SDVersion {
     VERSION_ANIMA,
     VERSION_FLUX2,
     VERSION_FLUX2_KLEIN,
+    VERSION_HIDREAM_O1,
     VERSION_Z_IMAGE,
     VERSION_OVIS_IMAGE,
     VERSION_ERNIE_IMAGE,
@@ -170,6 +164,7 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_sd3(version) ||
         sd_version_is_wan(version) ||
         sd_version_is_qwen_image(version) ||
+        version == VERSION_HIDREAM_O1 ||
         sd_version_is_anima(version) ||
         sd_version_is_z_image(version) ||
         sd_version_is_ernie_image(version)) {
@@ -195,135 +190,40 @@ enum PMVersion {
     PM_VERSION_2,
 };
 
-struct TensorStorage {
-    std::string name;
-    ggml_type type          = GGML_TYPE_F32;
-    ggml_type expected_type = GGML_TYPE_COUNT;
-    bool is_f8_e4m3         = false;
-    bool is_f8_e5m2         = false;
-    bool is_f64             = false;
-    bool is_i64             = false;
-    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
-    int n_dims              = 0;
+typedef OrderedMap<std::string, TensorStorage> String2TensorStorage;
+using TensorTypeRules = std::vector<std::pair<std::string, ggml_type>>;
 
-    size_t file_index = 0;
-    int index_in_zip  = -1;  // >= means stored in a zip file
-    uint64_t offset   = 0;   // offset in file
+TensorTypeRules parse_tensor_type_rules(const std::string& tensor_type_rules);
 
-    TensorStorage() = default;
+class MmapWrapper;
 
-    TensorStorage(std::string name, ggml_type type, const int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
-        : name(std::move(name)), type(type), n_dims(n_dims), file_index(file_index), offset(offset) {
-        for (int i = 0; i < n_dims; i++) {
-            this->ne[i] = ne[i];
-        }
-    }
-
-    int64_t nelements() const {
-        int64_t n = 1;
-        for (int i = 0; i < SD_MAX_DIMS; i++) {
-            n *= ne[i];
-        }
-        return n;
-    }
-
-    int64_t nbytes() const {
-        return nelements() * ggml_type_size(type) / ggml_blck_size(type);
-    }
-
-    int64_t nbytes_to_read() const {
-        if (is_f8_e4m3 || is_f8_e5m2) {
-            return nbytes() / 2;
-        } else if (is_f64 || is_i64) {
-            return nbytes() * 2;
-        } else {
-            return nbytes();
-        }
-    }
-
-    void unsqueeze() {
-        if (n_dims == 2) {
-            n_dims = 4;
-            ne[3]  = ne[1];
-            ne[2]  = ne[0];
-            ne[1]  = 1;
-            ne[0]  = 1;
-        }
-    }
-
-    std::vector<TensorStorage> chunk(size_t n) {
-        std::vector<TensorStorage> chunks;
-        uint64_t chunk_size = nbytes_to_read() / n;
-        // printf("%d/%d\n", chunk_size, nbytes_to_read());
-        reverse_ne();
-        for (size_t i = 0; i < n; i++) {
-            TensorStorage chunk_i = *this;
-            chunk_i.ne[0]         = ne[0] / n;
-            chunk_i.offset        = offset + i * chunk_size;
-            chunk_i.reverse_ne();
-            chunks.push_back(chunk_i);
-        }
-        reverse_ne();
-        return chunks;
-    }
-
-    void reverse_ne() {
-        int64_t new_ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
-        for (int i = 0; i < n_dims; i++) {
-            new_ne[i] = ne[n_dims - 1 - i];
-        }
-        for (int i = 0; i < n_dims; i++) {
-            ne[i] = new_ne[i];
-        }
-    }
-
-    std::string to_string() const {
-        std::stringstream ss;
-        const char* type_name = ggml_type_name(type);
-        if (is_f8_e4m3) {
-            type_name = "f8_e4m3";
-        } else if (is_f8_e5m2) {
-            type_name = "f8_e5m2";
-        } else if (is_f64) {
-            type_name = "f64";
-        } else if (is_i64) {
-            type_name = "i64";
-        }
-        ss << name << " | " << type_name << " | ";
-        ss << n_dims << " [";
-        for (int i = 0; i < SD_MAX_DIMS; i++) {
-            ss << ne[i];
-            if (i != SD_MAX_DIMS - 1) {
-                ss << ", ";
-            }
-        }
-        ss << "]";
-        return ss.str();
-    }
+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;
 };
 
-typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
-
-typedef OrderedMap<std::string, TensorStorage> String2TensorStorage;
+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 parse_data_pkl(uint8_t* buffer,
-                        size_t buffer_size,
-                        zip_t* zip,
-                        std::string dir,
-                        size_t file_index,
-                        const std::string prefix);
-
     bool init_from_gguf_file(const std::string& file_path, const std::string& prefix = "");
     bool init_from_safetensors_file(const std::string& file_path, const std::string& prefix = "");
-    bool init_from_ckpt_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:
@@ -339,6 +239,10 @@ public:
     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 = {},
@@ -353,7 +257,6 @@ public:
         return names;
     }
 
-    bool save_to_gguf_file(const std::string& file_path, ggml_type type, const std::string& tensor_type_rules);
     bool tensor_should_be_converted(const TensorStorage& tensor_storage, ggml_type type);
     int64_t get_params_mem_size(ggml_backend_t backend, ggml_type type = GGML_TYPE_COUNT);
     ~ModelLoader() = default;

src/model_io/binary_io.h

@@ -0,0 +1,57 @@
+#ifndef __SD_MODEL_IO_BINARY_IO_H__
+#define __SD_MODEL_IO_BINARY_IO_H__
+
+#include <cstdint>
+#include <ostream>
+
+namespace model_io {
+
+    inline int32_t read_int(const uint8_t* buffer) {
+        uint32_t value = 0;
+        value |= static_cast<uint32_t>(buffer[3]) << 24;
+        value |= static_cast<uint32_t>(buffer[2]) << 16;
+        value |= static_cast<uint32_t>(buffer[1]) << 8;
+        value |= static_cast<uint32_t>(buffer[0]);
+        return static_cast<int32_t>(value);
+    }
+
+    inline uint16_t read_short(const uint8_t* buffer) {
+        uint16_t value = 0;
+        value |= static_cast<uint16_t>(buffer[1]) << 8;
+        value |= static_cast<uint16_t>(buffer[0]);
+        return value;
+    }
+
+    inline uint64_t read_u64(const uint8_t* buffer) {
+        uint64_t value = 0;
+        value |= static_cast<uint64_t>(buffer[7]) << 56;
+        value |= static_cast<uint64_t>(buffer[6]) << 48;
+        value |= static_cast<uint64_t>(buffer[5]) << 40;
+        value |= static_cast<uint64_t>(buffer[4]) << 32;
+        value |= static_cast<uint64_t>(buffer[3]) << 24;
+        value |= static_cast<uint64_t>(buffer[2]) << 16;
+        value |= static_cast<uint64_t>(buffer[1]) << 8;
+        value |= static_cast<uint64_t>(buffer[0]);
+        return value;
+    }
+
+    inline void write_u64(std::ostream& stream, uint64_t value) {
+        uint8_t buffer[8];
+        for (int i = 0; i < 8; ++i) {
+            buffer[i] = static_cast<uint8_t>((value >> (8 * i)) & 0xFF);
+        }
+        stream.write((const char*)buffer, sizeof(buffer));
+    }
+
+    inline int find_char(const uint8_t* buffer, int len, char c) {
+        for (int pos = 0; pos < len; pos++) {
+            if (buffer[pos] == (uint8_t)c) {
+                return pos;
+            }
+        }
+        return -1;
+    }
+
+}  // namespace model_io
+
+#endif  // __SD_MODEL_IO_BINARY_IO_H__

src/model_io/gguf_io.cpp

@@ -0,0 +1,123 @@
+#include "gguf_io.h"
+
+#include <cstdint>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#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) {
+        *error = message;
+    }
+}
+
+bool is_gguf_file(const std::string& file_path) {
+    std::ifstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        return false;
+    }
+
+    char magic[4];
+
+    file.read(magic, sizeof(magic));
+    if (!file) {
+        return false;
+    }
+    for (uint32_t i = 0; i < sizeof(magic); i++) {
+        if (magic[i] != GGUF_MAGIC[i]) {
+            return false;
+        }
+    }
+
+    return true;
+}
+
+bool read_gguf_file(const std::string& file_path,
+                    std::vector<TensorStorage>& tensor_storages,
+                    std::string* error) {
+    tensor_storages.clear();
+
+    gguf_context* ctx_gguf_ = nullptr;
+    ggml_context* ctx_meta_ = nullptr;
+
+    ctx_gguf_ = gguf_init_from_file(file_path.c_str(), {true, &ctx_meta_});
+    if (!ctx_gguf_) {
+        GGUFReader gguf_reader;
+        if (!gguf_reader.load(file_path)) {
+            set_error(error, "failed to open '" + file_path + "' with GGUFReader");
+            return false;
+        }
+
+        size_t data_offset = gguf_reader.data_offset();
+        for (const auto& gguf_tensor_info : gguf_reader.tensors()) {
+            TensorStorage tensor_storage(
+                gguf_tensor_info.name,
+                gguf_tensor_info.type,
+                gguf_tensor_info.shape.data(),
+                static_cast<int>(gguf_tensor_info.shape.size()),
+                0,
+                data_offset + gguf_tensor_info.offset);
+
+            tensor_storages.push_back(tensor_storage);
+        }
+
+        return true;
+    }
+
+    int n_tensors = static_cast<int>(gguf_get_n_tensors(ctx_gguf_));
+
+    size_t data_offset = gguf_get_data_offset(ctx_gguf_);
+    for (int i = 0; i < n_tensors; i++) {
+        std::string name   = gguf_get_tensor_name(ctx_gguf_, i);
+        ggml_tensor* dummy = ggml_get_tensor(ctx_meta_, name.c_str());
+        size_t offset      = data_offset + gguf_get_tensor_offset(ctx_gguf_, i);
+
+        TensorStorage tensor_storage(name, dummy->type, dummy->ne, ggml_n_dims(dummy), 0, offset);
+
+        if (ggml_nbytes(dummy) != tensor_storage.nbytes()) {
+            gguf_free(ctx_gguf_);
+            ggml_free(ctx_meta_);
+            set_error(error, "size mismatch for tensor '" + name + "'");
+            return false;
+        }
+
+        tensor_storages.push_back(tensor_storage);
+    }
+
+    gguf_free(ctx_gguf_);
+    ggml_free(ctx_meta_);
+
+    return true;
+}
+
+bool write_gguf_file(const std::string& file_path,
+                     const std::vector<TensorWriteInfo>& tensors,
+                     std::string* error) {
+    gguf_context* gguf_ctx = gguf_init_empty();
+    if (gguf_ctx == nullptr) {
+        set_error(error, "gguf_init_empty failed");
+        return false;
+    }
+
+    for (const TensorWriteInfo& write_tensor : tensors) {
+        ggml_tensor* tensor = write_tensor.tensor;
+        if (tensor == nullptr) {
+            set_error(error, "null tensor cannot be written to GGUF");
+            gguf_free(gguf_ctx);
+            return false;
+        }
+        gguf_add_tensor(gguf_ctx, tensor);
+    }
+
+    LOG_INFO("trying to save tensors to %s", file_path.c_str());
+    bool success = gguf_write_to_file(gguf_ctx, file_path.c_str(), false);
+    if (!success) {
+        set_error(error, "failed to write GGUF file '" + file_path + "'");
+    }
+    gguf_free(gguf_ctx);
+    return success;
+}

src/model_io/gguf_io.h

@@ -0,0 +1,17 @@
+#ifndef __SD_MODEL_IO_GGUF_IO_H__
+#define __SD_MODEL_IO_GGUF_IO_H__
+
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool is_gguf_file(const std::string& file_path);
+bool read_gguf_file(const std::string& file_path,
+                    std::vector<TensorStorage>& tensor_storages,
+                    std::string* error = nullptr);
+bool write_gguf_file(const std::string& file_path,
+                     const std::vector<TensorWriteInfo>& tensors,
+                     std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_GGUF_IO_H__

src/model_io/gguf_reader_ext.h

@@ -1,5 +1,5 @@
-#ifndef __GGUF_READER_HPP__
-#define __GGUF_READER_HPP__
+#ifndef __SD_MODEL_IO_GGUF_READER_EXT_H__
+#define __SD_MODEL_IO_GGUF_READER_EXT_H__
 
 #include <cstdint>
 #include <fstream>
@@ -231,4 +231,4 @@ public:
     size_t data_offset() const { return data_offset_; }
 };
 
-#endif  // __GGUF_READER_HPP__
+#endif  // __SD_MODEL_IO_GGUF_READER_EXT_H__

src/model_io/pickle_io.h

@@ -0,0 +1,21 @@
+#ifndef __SD_MODEL_IO_PICKLE_IO_H__
+#define __SD_MODEL_IO_PICKLE_IO_H__
+
+#include <cstddef>
+#include <cstdint>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool skip_pickle_object(const uint8_t* buffer, size_t buffer_size, size_t* object_size);
+bool pickle_object_is_torch_magic_number(const uint8_t* buffer, size_t buffer_size);
+bool parse_pickle_uint32_object(const uint8_t* buffer, size_t buffer_size, uint32_t* value);
+bool parse_torch_state_dict_pickle(const uint8_t* buffer,
+                                   size_t buffer_size,
+                                   std::vector<TensorStorage>& tensor_storages,
+                                   std::unordered_map<std::string, uint64_t>& storage_nbytes,
+                                   std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_PICKLE_IO_H__

src/model_io/safetensors_io.cpp

@@ -0,0 +1,316 @@
+#include "safetensors_io.h"
+
+#include <cstdint>
+#include <exception>
+#include <fstream>
+#include <string>
+#include <vector>
+
+#include "binary_io.h"
+#include "json.hpp"
+#include "util.h"
+
+static constexpr size_t ST_HEADER_SIZE_LEN = 8;
+
+static void set_error(std::string* error, const std::string& message) {
+    if (error != nullptr) {
+        *error = message;
+    }
+}
+
+bool is_safetensors_file(const std::string& file_path) {
+    std::ifstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        return false;
+    }
+
+    // get file size
+    file.seekg(0, file.end);
+    size_t file_size_ = file.tellg();
+    file.seekg(0, file.beg);
+
+    // read header size
+    if (file_size_ <= ST_HEADER_SIZE_LEN) {
+        return false;
+    }
+
+    uint8_t header_size_buf[ST_HEADER_SIZE_LEN];
+    file.read((char*)header_size_buf, ST_HEADER_SIZE_LEN);
+    if (!file) {
+        return false;
+    }
+
+    size_t header_size_ = model_io::read_u64(header_size_buf);
+    if (header_size_ >= file_size_ || header_size_ <= 2) {
+        return false;
+    }
+
+    // read header
+    std::vector<char> header_buf;
+    header_buf.resize(header_size_ + 1);
+    header_buf[header_size_] = '\0';
+    file.read(header_buf.data(), header_size_);
+    if (!file) {
+        return false;
+    }
+    try {
+        nlohmann::json header_ = nlohmann::json::parse(header_buf.data());
+    } catch (const std::exception&) {
+        return false;
+    }
+    return true;
+}
+
+static ggml_type safetensors_dtype_to_ggml_type(const std::string& dtype) {
+    ggml_type ttype = GGML_TYPE_COUNT;
+    if (dtype == "F16") {
+        ttype = GGML_TYPE_F16;
+    } else if (dtype == "BF16") {
+        ttype = GGML_TYPE_BF16;
+    } else if (dtype == "F32") {
+        ttype = GGML_TYPE_F32;
+    } else if (dtype == "F64") {
+        ttype = GGML_TYPE_F32;
+    } else if (dtype == "F8_E4M3") {
+        ttype = GGML_TYPE_F16;
+    } else if (dtype == "F8_E5M2") {
+        ttype = GGML_TYPE_F16;
+    } else if (dtype == "I32") {
+        ttype = GGML_TYPE_I32;
+    } else if (dtype == "I64") {
+        ttype = GGML_TYPE_I32;
+    }
+    return ttype;
+}
+
+// https://huggingface.co/docs/safetensors/index
+bool read_safetensors_file(const std::string& file_path,
+                           std::vector<TensorStorage>& tensor_storages,
+                           std::string* error) {
+    std::ifstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        set_error(error, "failed to open '" + file_path + "'");
+        return false;
+    }
+
+    // get file size
+    file.seekg(0, file.end);
+    size_t file_size_ = file.tellg();
+    file.seekg(0, file.beg);
+
+    // read header size
+    if (file_size_ <= ST_HEADER_SIZE_LEN) {
+        set_error(error, "invalid safetensor file '" + file_path + "'");
+        return false;
+    }
+
+    uint8_t header_size_buf[ST_HEADER_SIZE_LEN];
+    file.read((char*)header_size_buf, ST_HEADER_SIZE_LEN);
+    if (!file) {
+        set_error(error, "read safetensors header size failed: '" + file_path + "'");
+        return false;
+    }
+
+    size_t header_size_ = model_io::read_u64(header_size_buf);
+    if (header_size_ >= file_size_) {
+        set_error(error, "invalid safetensor file '" + file_path + "'");
+        return false;
+    }
+
+    // read header
+    std::vector<char> header_buf;
+    header_buf.resize(header_size_ + 1);
+    header_buf[header_size_] = '\0';
+    file.read(header_buf.data(), header_size_);
+    if (!file) {
+        set_error(error, "read safetensors header failed: '" + file_path + "'");
+        return false;
+    }
+
+    nlohmann::json header_;
+    try {
+        header_ = nlohmann::json::parse(header_buf.data());
+    } catch (const std::exception&) {
+        set_error(error, "parsing safetensors header failed: '" + file_path + "'");
+        return false;
+    }
+
+    tensor_storages.clear();
+    for (auto& item : header_.items()) {
+        std::string name           = item.key();
+        nlohmann::json tensor_info = item.value();
+        // LOG_DEBUG("%s %s\n", name.c_str(), tensor_info.dump().c_str());
+
+        if (name == "__metadata__") {
+            continue;
+        }
+
+        std::string dtype    = tensor_info["dtype"];
+        nlohmann::json shape = tensor_info["shape"];
+
+        if (dtype == "U8") {
+            continue;
+        }
+
+        size_t begin = tensor_info["data_offsets"][0].get<size_t>();
+        size_t end   = tensor_info["data_offsets"][1].get<size_t>();
+
+        ggml_type type = safetensors_dtype_to_ggml_type(dtype);
+        if (type == GGML_TYPE_COUNT) {
+            set_error(error, "unsupported dtype '" + dtype + "' (tensor '" + name + "')");
+            return false;
+        }
+
+        if (shape.size() > SD_MAX_DIMS) {
+            set_error(error, "invalid tensor '" + name + "'");
+            return false;
+        }
+
+        int n_dims              = (int)shape.size();
+        int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+        for (int i = 0; i < n_dims; i++) {
+            ne[i] = shape[i].get<int64_t>();
+        }
+
+        if (n_dims == 5) {
+            n_dims = 4;
+            ne[0]  = ne[0] * ne[1];
+            ne[1]  = ne[2];
+            ne[2]  = ne[3];
+            ne[3]  = ne[4];
+        }
+
+        // ggml_n_dims returns 1 for scalars
+        if (n_dims == 0) {
+            n_dims = 1;
+        }
+
+        TensorStorage tensor_storage(name, type, ne, n_dims, 0, ST_HEADER_SIZE_LEN + header_size_ + begin);
+        tensor_storage.reverse_ne();
+
+        size_t tensor_data_size = end - begin;
+
+        bool tensor_size_ok;
+        if (dtype == "F8_E4M3") {
+            tensor_storage.is_f8_e4m3 = true;
+            // f8 -> f16
+            tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size * 2);
+        } else if (dtype == "F8_E5M2") {
+            tensor_storage.is_f8_e5m2 = true;
+            // f8 -> f16
+            tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size * 2);
+        } else if (dtype == "F64") {
+            tensor_storage.is_f64 = true;
+            // f64 -> f32
+            tensor_size_ok = (tensor_storage.nbytes() * 2 == tensor_data_size);
+        } else if (dtype == "I64") {
+            tensor_storage.is_i64 = true;
+            // i64 -> i32
+            tensor_size_ok = (tensor_storage.nbytes() * 2 == tensor_data_size);
+        } else {
+            tensor_size_ok = (tensor_storage.nbytes() == tensor_data_size);
+        }
+        if (!tensor_size_ok) {
+            set_error(error, "size mismatch for tensor '" + name + "' (" + dtype + ")");
+            return false;
+        }
+
+        tensor_storages.push_back(tensor_storage);
+
+        // LOG_DEBUG("%s %s", tensor_storage.to_string().c_str(), dtype.c_str());
+    }
+
+    return true;
+}
+
+static bool ggml_type_to_safetensors_dtype(ggml_type type, std::string* dtype) {
+    switch (type) {
+        case GGML_TYPE_F16:
+            *dtype = "F16";
+            return true;
+        case GGML_TYPE_BF16:
+            *dtype = "BF16";
+            return true;
+        case GGML_TYPE_F32:
+            *dtype = "F32";
+            return true;
+        case GGML_TYPE_I32:
+            *dtype = "I32";
+            return true;
+        default:
+            return false;
+    }
+}
+
+bool write_safetensors_file(const std::string& file_path,
+                            const std::vector<TensorWriteInfo>& tensors,
+                            std::string* error) {
+    nlohmann::ordered_json header = nlohmann::ordered_json::object();
+
+    uint64_t data_offset = 0;
+    for (const TensorWriteInfo& write_tensor : tensors) {
+        ggml_tensor* tensor = write_tensor.tensor;
+        if (tensor == nullptr) {
+            set_error(error, "null tensor cannot be written to safetensors");
+            return false;
+        }
+
+        const std::string name = ggml_get_name(tensor);
+        std::string dtype;
+        if (!ggml_type_to_safetensors_dtype(tensor->type, &dtype)) {
+            set_error(error,
+                      "unsupported safetensors dtype '" + std::string(ggml_type_name(tensor->type)) +
+                          "' for tensor '" + name + "'");
+            return false;
+        }
+
+        const uint64_t tensor_nbytes = ggml_nbytes(tensor);
+
+        nlohmann::ordered_json json_tensor_info = nlohmann::ordered_json::object();
+        json_tensor_info["dtype"]               = dtype;
+
+        nlohmann::ordered_json shape = nlohmann::ordered_json::array();
+        for (int i = 0; i < write_tensor.n_dims; ++i) {
+            shape.push_back(write_tensor.ne[write_tensor.n_dims - 1 - i]);
+        }
+        json_tensor_info["shape"] = shape;
+
+        nlohmann::ordered_json data_offsets = nlohmann::ordered_json::array();
+        data_offsets.push_back(data_offset);
+        data_offsets.push_back(data_offset + tensor_nbytes);
+        json_tensor_info["data_offsets"] = data_offsets;
+
+        header[name] = json_tensor_info;
+        data_offset += tensor_nbytes;
+    }
+
+    const std::string header_str = header.dump();
+
+    std::ofstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        set_error(error, "failed to open '" + file_path + "' for writing");
+        return false;
+    }
+
+    LOG_INFO("trying to save tensors to %s", file_path.c_str());
+    model_io::write_u64(file, header_str.size());
+    file.write(header_str.data(), header_str.size());
+    if (!file) {
+        set_error(error, "failed to write safetensors header to '" + file_path + "'");
+        return false;
+    }
+
+    for (const TensorWriteInfo& write_tensor : tensors) {
+        ggml_tensor* tensor        = write_tensor.tensor;
+        const std::string name     = ggml_get_name(tensor);
+        const size_t tensor_nbytes = ggml_nbytes(tensor);
+        file.write((const char*)tensor->data, tensor_nbytes);
+        if (!file) {
+            set_error(error,
+                      "failed to write tensor '" + name + "' to '" + file_path + "'");
+            return false;
+        }
+    }
+
+    return true;
+}

src/model_io/safetensors_io.h

@@ -0,0 +1,17 @@
+#ifndef __SD_MODEL_IO_SAFETENSORS_IO_H__
+#define __SD_MODEL_IO_SAFETENSORS_IO_H__
+
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool is_safetensors_file(const std::string& file_path);
+bool read_safetensors_file(const std::string& file_path,
+                           std::vector<TensorStorage>& tensor_storages,
+                           std::string* error = nullptr);
+bool write_safetensors_file(const std::string& file_path,
+                            const std::vector<TensorWriteInfo>& tensors,
+                            std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_SAFETENSORS_IO_H__

src/model_io/tensor_storage.h

@@ -0,0 +1,132 @@
+#ifndef __SD_TENSOR_STORAGE_H__
+#define __SD_TENSOR_STORAGE_H__
+
+#include <cstddef>
+#include <cstdint>
+#include <functional>
+#include <sstream>
+#include <string>
+#include <utility>
+#include <vector>
+
+#include "ggml.h"
+
+#define SD_MAX_DIMS 5
+
+struct TensorStorage {
+    std::string name;
+    ggml_type type          = GGML_TYPE_F32;
+    ggml_type expected_type = GGML_TYPE_COUNT;
+    bool is_f8_e4m3         = false;
+    bool is_f8_e5m2         = false;
+    bool is_f64             = false;
+    bool is_i64             = false;
+    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+    int n_dims              = 0;
+
+    std::string storage_key;
+    size_t file_index = 0;
+    int index_in_zip  = -1;  // >= means stored in a zip file
+    uint64_t offset   = 0;   // offset in file
+
+    TensorStorage() = default;
+
+    TensorStorage(std::string name, ggml_type type, const int64_t* ne, int n_dims, size_t file_index, size_t offset = 0)
+        : name(std::move(name)), type(type), n_dims(n_dims), file_index(file_index), offset(offset) {
+        for (int i = 0; i < n_dims; i++) {
+            this->ne[i] = ne[i];
+        }
+    }
+
+    int64_t nelements() const {
+        int64_t n = 1;
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            n *= ne[i];
+        }
+        return n;
+    }
+
+    int64_t nbytes() const {
+        return nelements() * ggml_type_size(type) / ggml_blck_size(type);
+    }
+
+    int64_t nbytes_to_read() const {
+        if (is_f8_e4m3 || is_f8_e5m2) {
+            return nbytes() / 2;
+        } else if (is_f64 || is_i64) {
+            return nbytes() * 2;
+        } else {
+            return nbytes();
+        }
+    }
+
+    void unsqueeze() {
+        if (n_dims == 2) {
+            n_dims = 4;
+            ne[3]  = ne[1];
+            ne[2]  = ne[0];
+            ne[1]  = 1;
+            ne[0]  = 1;
+        }
+    }
+
+    std::vector<TensorStorage> chunk(size_t n) {
+        std::vector<TensorStorage> chunks;
+        uint64_t chunk_size = nbytes_to_read() / n;
+        // printf("%d/%d\n", chunk_size, nbytes_to_read());
+        reverse_ne();
+        for (size_t i = 0; i < n; i++) {
+            TensorStorage chunk_i = *this;
+            chunk_i.ne[0]         = ne[0] / n;
+            chunk_i.offset        = offset + i * chunk_size;
+            chunk_i.reverse_ne();
+            chunks.push_back(chunk_i);
+        }
+        reverse_ne();
+        return chunks;
+    }
+
+    void reverse_ne() {
+        int64_t new_ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+        for (int i = 0; i < n_dims; i++) {
+            new_ne[i] = ne[n_dims - 1 - i];
+        }
+        for (int i = 0; i < n_dims; i++) {
+            ne[i] = new_ne[i];
+        }
+    }
+
+    std::string to_string() const {
+        std::stringstream ss;
+        const char* type_name = ggml_type_name(type);
+        if (is_f8_e4m3) {
+            type_name = "f8_e4m3";
+        } else if (is_f8_e5m2) {
+            type_name = "f8_e5m2";
+        } else if (is_f64) {
+            type_name = "f64";
+        } else if (is_i64) {
+            type_name = "i64";
+        }
+        ss << name << " | " << type_name << " | ";
+        ss << n_dims << " [";
+        for (int i = 0; i < SD_MAX_DIMS; i++) {
+            ss << ne[i];
+            if (i != SD_MAX_DIMS - 1) {
+                ss << ", ";
+            }
+        }
+        ss << "]";
+        return ss.str();
+    }
+};
+
+struct TensorWriteInfo {
+    int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
+    int n_dims              = 0;
+    ggml_tensor* tensor     = nullptr;
+};
+
+typedef std::function<bool(const TensorStorage&, ggml_tensor**)> on_new_tensor_cb_t;
+
+#endif  // __SD_TENSOR_STORAGE_H__

src/model_io/torch_legacy_io.cpp

@@ -0,0 +1,252 @@
+#include "torch_legacy_io.h"
+
+#include <algorithm>
+#include <cstdint>
+#include <fstream>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "pickle_io.h"
+#include "util.h"
+
+// torch.save format background:
+//
+//   - Before PyTorch 1.6.0, torch.save used this legacy non-zip format by
+//     default.
+//   - Since PyTorch 1.6.0, torch.save defaults to an uncompressed ZIP64 archive
+//     containing data.pkl, data/, version, and, since PyTorch 2.1.0, byteorder.
+//   - The old format can still be produced explicitly with:
+//       torch.save(obj, path, _use_new_zipfile_serialization=False)
+//
+// Whether obj is a state_dict or a whole nn.Module does not change the outer
+// container format selected by torch.save. It changes the pickled object inside:
+//
+//   - state_dict: usually an OrderedDict[str, Tensor]. pickle_io.cpp supports a
+//     restricted subset of this layout because tensor metadata and raw storages
+//     can be recovered without executing pickle callables.
+//   - whole module/checkpoint object: arbitrary Python object graph. This may
+//     require importing user classes and executing pickle GLOBAL/REDUCE rebuild
+//     logic, so it is intentionally not supported here.
+//
+// Legacy non-zip PyTorch files are not a single pickle object:
+//
+//   1. pickle object: PyTorch legacy magic number
+//   2. pickle object: legacy protocol version, expected to be 1001
+//   3. pickle object: sys_info metadata, ignored by this reader
+//   4. pickle object: state_dict metadata, parsed by pickle_io.cpp
+//   5. pickle object: serialized storage key list, skipped here
+//   6. raw storage data payloads
+//      - PyTorch writes storages after the pickles, ordered by storage key
+//      - each storage has an 8-byte legacy storage header followed by raw bytes
+static constexpr size_t LEGACY_STORAGE_HEADER_SIZE = 8;
+
+static void set_error(std::string* error, const std::string& message) {
+    if (error != nullptr) {
+        *error = message;
+    }
+}
+
+static std::string bytes_to_hex(const std::vector<uint8_t>& bytes) {
+    static const char* hex = "0123456789ABCDEF";
+    std::string result;
+    result.reserve(bytes.size() * 3);
+    for (size_t i = 0; i < bytes.size(); ++i) {
+        if (i > 0) {
+            result.push_back('-');
+        }
+        result.push_back(hex[(bytes[i] >> 4) & 0x0F]);
+        result.push_back(hex[bytes[i] & 0x0F]);
+    }
+    return result;
+}
+
+static bool is_probably_tar_file(const std::vector<uint8_t>& header) {
+    return header.size() >= 262 &&
+           header[257] == 'u' &&
+           header[258] == 's' &&
+           header[259] == 't' &&
+           header[260] == 'a' &&
+           header[261] == 'r';
+}
+
+static std::string torch_legacy_diagnostics(const std::string& file_path, const std::vector<uint8_t>& buffer) {
+    if (!ends_with(file_path, ".pt") && !ends_with(file_path, ".pth")) {
+        return "";
+    }
+    if (buffer.empty()) {
+        return "unsupported PyTorch file '" + file_path + "': empty file";
+    }
+
+    size_t short_len = std::min<size_t>(buffer.size(), 32);
+    std::vector<uint8_t> short_header(buffer.begin(), buffer.begin() + short_len);
+    const bool raw_pickle = buffer[0] == 0x80;
+    const bool tar_file   = is_probably_tar_file(buffer);
+
+    std::string message = "unsupported PyTorch file '" + file_path + "': first bytes " +
+                          bytes_to_hex(short_header) +
+                          ", raw_pickle=" + (raw_pickle ? "true" : "false") +
+                          ", tar=" + (tar_file ? "true" : "false");
+    if (raw_pickle) {
+        message += "; raw pickle did not match the restricted state_dict layouts currently supported";
+    } else if (tar_file) {
+        message += "; legacy tar PyTorch checkpoints are not supported yet";
+    }
+    return message;
+}
+
+bool read_torch_legacy_file(const std::string& file_path,
+                            std::vector<TensorStorage>& tensor_storages,
+                            std::string* error) {
+    std::ifstream file(file_path, std::ios::binary);
+    if (!file.is_open()) {
+        set_error(error, "failed to open '" + file_path + "'");
+        return false;
+    }
+
+    file.seekg(0, file.end);
+    size_t file_size = (size_t)file.tellg();
+    file.seekg(0, file.beg);
+    if (file_size == 0) {
+        set_error(error, "empty file '" + file_path + "'");
+        return false;
+    }
+
+    std::vector<uint8_t> buffer(file_size);
+    file.read((char*)buffer.data(), file_size);
+    if (!file) {
+        set_error(error, "failed to read '" + file_path + "'");
+        return false;
+    }
+
+    auto finalize_tensor_offsets = [&](size_t storage_data_offset,
+                                       const std::unordered_map<std::string, uint64_t>& legacy_storage_map) -> bool {
+        if (storage_data_offset > file_size) {
+            return false;
+        }
+
+        std::vector<std::string> storage_keys;
+        storage_keys.reserve(legacy_storage_map.size());
+        for (const auto& [storage_key, _] : legacy_storage_map) {
+            storage_keys.push_back(storage_key);
+        }
+        std::sort(storage_keys.begin(), storage_keys.end());
+
+        std::unordered_map<std::string, uint64_t> storage_offsets;
+        uint64_t current_offset = storage_data_offset;
+        for (const auto& storage_key : storage_keys) {
+            auto it = legacy_storage_map.find(storage_key);
+            if (it == legacy_storage_map.end()) {
+                return false;
+            }
+            if (current_offset + LEGACY_STORAGE_HEADER_SIZE + it->second > file_size) {
+                return false;
+            }
+            storage_offsets[storage_key] = current_offset + LEGACY_STORAGE_HEADER_SIZE;
+            current_offset += LEGACY_STORAGE_HEADER_SIZE + it->second;
+        }
+
+        for (auto& tensor_storage : tensor_storages) {
+            if (tensor_storage.storage_key.empty()) {
+                continue;
+            }
+
+            auto it_offset = storage_offsets.find(tensor_storage.storage_key);
+            auto it_size   = legacy_storage_map.find(tensor_storage.storage_key);
+            if (it_offset == storage_offsets.end() || it_size == legacy_storage_map.end()) {
+                return false;
+            }
+
+            uint64_t base_offset    = it_offset->second;
+            uint64_t storage_nbytes = it_size->second;
+            uint64_t tensor_nbytes  = tensor_storage.nbytes_to_read();
+            if (tensor_storage.offset + tensor_nbytes > storage_nbytes) {
+                return false;
+            }
+
+            tensor_storage.offset = base_offset + tensor_storage.offset;
+            tensor_storage.storage_key.clear();
+        }
+
+        return true;
+    };
+
+    auto parse_state_dict_at = [&](size_t state_dict_offset, size_t state_dict_size, size_t* storage_data_offset) -> bool {
+        tensor_storages.clear();
+        std::unordered_map<std::string, uint64_t> legacy_storage_map;
+        if (!parse_torch_state_dict_pickle(buffer.data() + state_dict_offset,
+                                           state_dict_size,
+                                           tensor_storages,
+                                           legacy_storage_map,
+                                           error)) {
+            return false;
+        }
+
+        size_t offset_after_state_dict = state_dict_offset + state_dict_size;
+        size_t storage_keys_size       = 0;
+        if (!skip_pickle_object(buffer.data() + offset_after_state_dict,
+                                buffer.size() - offset_after_state_dict,
+                                &storage_keys_size)) {
+            return false;
+        }
+
+        *storage_data_offset = offset_after_state_dict + storage_keys_size;
+        return finalize_tensor_offsets(*storage_data_offset, legacy_storage_map);
+    };
+
+    size_t object_size_1 = 0;
+    size_t offset        = 0;
+
+    if (skip_pickle_object(buffer.data(), buffer.size(), &object_size_1) &&
+        pickle_object_is_torch_magic_number(buffer.data(), object_size_1)) {
+        offset += object_size_1;
+
+        size_t object_size_2 = 0;
+        if (!skip_pickle_object(buffer.data() + offset, buffer.size() - offset, &object_size_2)) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+            return false;
+        }
+        uint32_t protocol_version = 0;
+        if (!parse_pickle_uint32_object(buffer.data() + offset, object_size_2, &protocol_version) || protocol_version != 1001) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+            return false;
+        }
+        offset += object_size_2;
+
+        size_t object_size_3 = 0;
+        if (!skip_pickle_object(buffer.data() + offset, buffer.size() - offset, &object_size_3)) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+            return false;
+        }
+        offset += object_size_3;
+
+        size_t state_dict_size = 0;
+        if (!skip_pickle_object(buffer.data() + offset, buffer.size() - offset, &state_dict_size)) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+            return false;
+        }
+
+        size_t storage_data_offset = 0;
+        if (parse_state_dict_at(offset, state_dict_size, &storage_data_offset)) {
+            return true;
+        }
+
+        if (error != nullptr && error->empty()) {
+            set_error(error, torch_legacy_diagnostics(file_path, buffer));
+        }
+        return false;
+    }
+
+    size_t state_dict_size = 0;
+    if (skip_pickle_object(buffer.data(), buffer.size(), &state_dict_size)) {
+        size_t storage_data_offset = 0;
+        if (parse_state_dict_at(0, state_dict_size, &storage_data_offset)) {
+            return true;
+        }
+    }
+
+    if (error != nullptr && error->empty()) {
+        set_error(error, torch_legacy_diagnostics(file_path, buffer));
+    }
+    return false;
+}

src/model_io/torch_legacy_io.h

@@ -0,0 +1,13 @@
+#ifndef __SD_MODEL_IO_TORCH_LEGACY_IO_H__
+#define __SD_MODEL_IO_TORCH_LEGACY_IO_H__
+
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool read_torch_legacy_file(const std::string& file_path,
+                            std::vector<TensorStorage>& tensor_storages,
+                            std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_TORCH_LEGACY_IO_H__

src/model_io/torch_zip_io.cpp

@@ -0,0 +1,140 @@
+#include "torch_zip_io.h"
+
+#include <cstdint>
+#include <cstdlib>
+#include <string>
+#include <unordered_map>
+#include <vector>
+
+#include "pickle_io.h"
+
+#include "zip.h"
+
+static void set_error(std::string* error, const std::string& message) {
+    if (error != nullptr) {
+        *error = message;
+    }
+}
+
+bool is_torch_zip_file(const std::string& file_path) {
+    zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
+    if (zip == nullptr) {
+        return false;
+    }
+    zip_close(zip);
+    return true;
+}
+
+static bool find_zip_entry(zip_t* zip, const std::string& entry_name, int* index, uint64_t* size) {
+    size_t n = zip_entries_total(zip);
+    for (size_t i = 0; i < n; ++i) {
+        zip_entry_openbyindex(zip, i);
+        std::string name = zip_entry_name(zip);
+        if (name == entry_name) {
+            *index = (int)i;
+            *size  = zip_entry_size(zip);
+            zip_entry_close(zip);
+            return true;
+        }
+        zip_entry_close(zip);
+    }
+    return false;
+}
+
+static bool parse_zip_data_pkl(const uint8_t* buffer,
+                               size_t buffer_size,
+                               zip_t* zip,
+                               const std::string& dir,
+                               std::vector<TensorStorage>& tensor_storages,
+                               std::string* error) {
+    std::vector<TensorStorage> parsed_tensors;
+    std::unordered_map<std::string, uint64_t> storage_nbytes;
+    if (!parse_torch_state_dict_pickle(buffer, buffer_size, parsed_tensors, storage_nbytes, error)) {
+        if (error != nullptr && error->empty()) {
+            *error = "failed to parse torch zip pickle metadata";
+        }
+        return false;
+    }
+
+    for (auto& tensor_storage : parsed_tensors) {
+        if (tensor_storage.storage_key.empty()) {
+            set_error(error, "tensor '" + tensor_storage.name + "' has no storage key");
+            return false;
+        }
+
+        const std::string entry_name = dir + "data/" + tensor_storage.storage_key;
+        int zip_index                = -1;
+        uint64_t entry_size          = 0;
+        if (!find_zip_entry(zip, entry_name, &zip_index, &entry_size)) {
+            set_error(error, "storage entry '" + entry_name + "' was not found");
+            return false;
+        }
+
+        auto it_storage_size = storage_nbytes.find(tensor_storage.storage_key);
+        if (it_storage_size != storage_nbytes.end() && entry_size < it_storage_size->second) {
+            set_error(error, "storage entry '" + entry_name + "' is smaller than pickle metadata");
+            return false;
+        }
+
+        uint64_t tensor_nbytes = tensor_storage.nbytes_to_read();
+        if (tensor_storage.offset + tensor_nbytes > entry_size) {
+            set_error(error, "tensor '" + tensor_storage.name + "' exceeds storage entry '" + entry_name + "'");
+            return false;
+        }
+
+        tensor_storage.index_in_zip = zip_index;
+        tensor_storage.storage_key.clear();
+        tensor_storages.push_back(tensor_storage);
+    }
+
+    return true;
+}
+
+bool read_torch_zip_file(const std::string& file_path,
+                         std::vector<TensorStorage>& tensor_storages,
+                         std::string* error) {
+    zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
+    if (zip == nullptr) {
+        set_error(error, "failed to open '" + file_path + "'");
+        return false;
+    }
+
+    tensor_storages.clear();
+    bool success        = true;
+    bool found_data_pkl = false;
+    int n               = (int)zip_entries_total(zip);
+    for (int i = 0; i < n; ++i) {
+        zip_entry_openbyindex(zip, i);
+        std::string name = zip_entry_name(zip);
+        size_t pos       = name.find("data.pkl");
+        if (pos != std::string::npos) {
+            found_data_pkl  = true;
+            std::string dir = name.substr(0, pos);
+            void* pkl_data  = nullptr;
+            size_t pkl_size = 0;
+            zip_entry_read(zip, &pkl_data, &pkl_size);
+
+            if (pkl_data == nullptr || pkl_size == 0) {
+                set_error(error, "failed to read '" + name + "' from '" + file_path + "'");
+                success = false;
+            } else if (!parse_zip_data_pkl((const uint8_t*)pkl_data, pkl_size, zip, dir, tensor_storages, error)) {
+                success = false;
+            }
+
+            free(pkl_data);
+        }
+        zip_entry_close(zip);
+
+        if (!success) {
+            break;
+        }
+    }
+
+    if (success && !found_data_pkl) {
+        set_error(error, "data.pkl was not found in '" + file_path + "'");
+        success = false;
+    }
+
+    zip_close(zip);
+    return success;
+}

src/model_io/torch_zip_io.h

@@ -0,0 +1,14 @@
+#ifndef __SD_MODEL_IO_TORCH_ZIP_IO_H__
+#define __SD_MODEL_IO_TORCH_ZIP_IO_H__
+
+#include <string>
+#include <vector>
+
+#include "tensor_storage.h"
+
+bool is_torch_zip_file(const std::string& file_path);
+bool read_torch_zip_file(const std::string& file_path,
+                         std::vector<TensorStorage>& tensor_storages,
+                         std::string* error = nullptr);
+
+#endif  // __SD_MODEL_IO_TORCH_ZIP_IO_H__

src/pmid.hpp

@@ -411,13 +411,13 @@ public:
 
 public:
     PhotoMakerIDEncoder(ggml_backend_t backend,
-                        bool offload_params_to_cpu,
+                        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, offload_params_to_cpu),
+        : GGMLRunner(backend, params_backend),
           version(version),
           pm_version(pm_v),
           style_strength(sty) {
@@ -568,11 +568,11 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
     bool applied     = false;
 
     PhotoMakerIDEmbed(ggml_backend_t backend,
-                      bool offload_params_to_cpu,
+                      ggml_backend_t params_backend,
                       ModelLoader* ml,
                       const std::string& file_path = "",
                       const std::string& prefix    = "")
-        : file_path(file_path), GGMLRunner(backend, offload_params_to_cpu), model_loader(ml) {
+        : file_path(file_path), GGMLRunner(backend, params_backend), model_loader(ml) {
         if (!model_loader->init_from_file_and_convert_name(file_path, prefix)) {
             load_failed = true;
         }

src/preprocessing.hpp

@@ -24,6 +24,75 @@ static inline void preprocessing_set_4d(sd::Tensor<float>& tensor, float value,
     tensor.values()[static_cast<size_t>(preprocessing_offset_4d(tensor, i0, i1, i2, i3))] = value;
 }
 
+static inline uint8_t preprocessing_float_to_u8(float value) {
+    if (value <= 0.0f) {
+        return 0;
+    }
+    if (value >= 1.0f) {
+        return 255;
+    }
+    return static_cast<uint8_t>(value * 255.0f + 0.5f);
+}
+
+static inline void preprocessing_tensor_frame_to_sd_image(const sd::Tensor<float>& tensor, int frame_index, uint8_t* image_data) {
+    const auto& shape = tensor.shape();
+    GGML_ASSERT(shape.size() == 4 || shape.size() == 5);
+    GGML_ASSERT(image_data != nullptr);
+
+    const int width     = static_cast<int>(shape[0]);
+    const int height    = static_cast<int>(shape[1]);
+    const int channel   = static_cast<int>(shape[shape.size() == 5 ? 3 : 2]);
+    const size_t pixels = static_cast<size_t>(width) * static_cast<size_t>(height);
+    const float* src    = tensor.data();
+
+    if (shape.size() == 4) {
+        GGML_ASSERT(frame_index >= 0 && frame_index < shape[3]);
+        const size_t frame_stride = pixels * static_cast<size_t>(channel);
+        const float* frame_ptr    = src + static_cast<size_t>(frame_index) * frame_stride;
+        if (channel == 3) {
+            const float* c0 = frame_ptr;
+            const float* c1 = frame_ptr + pixels;
+            const float* c2 = frame_ptr + pixels * 2;
+            for (size_t i = 0; i < pixels; ++i) {
+                image_data[i * 3 + 0] = preprocessing_float_to_u8(c0[i]);
+                image_data[i * 3 + 1] = preprocessing_float_to_u8(c1[i]);
+                image_data[i * 3 + 2] = preprocessing_float_to_u8(c2[i]);
+            }
+            return;
+        }
+
+        for (size_t i = 0; i < pixels; ++i) {
+            for (int c = 0; c < channel; ++c) {
+                image_data[i * static_cast<size_t>(channel) + static_cast<size_t>(c)] =
+                    preprocessing_float_to_u8(frame_ptr[i + pixels * static_cast<size_t>(c)]);
+            }
+        }
+        return;
+    }
+
+    GGML_ASSERT(frame_index >= 0 && frame_index < shape[2]);
+    const size_t channel_stride = pixels * static_cast<size_t>(shape[2]);
+    const float* frame_ptr      = src + static_cast<size_t>(frame_index) * pixels;
+    if (channel == 3) {
+        const float* c0 = frame_ptr;
+        const float* c1 = frame_ptr + channel_stride;
+        const float* c2 = frame_ptr + channel_stride * 2;
+        for (size_t i = 0; i < pixels; ++i) {
+            image_data[i * 3 + 0] = preprocessing_float_to_u8(c0[i]);
+            image_data[i * 3 + 1] = preprocessing_float_to_u8(c1[i]);
+            image_data[i * 3 + 2] = preprocessing_float_to_u8(c2[i]);
+        }
+        return;
+    }
+
+    for (size_t i = 0; i < pixels; ++i) {
+        for (int c = 0; c < channel; ++c) {
+            image_data[i * static_cast<size_t>(channel) + static_cast<size_t>(c)] =
+                preprocessing_float_to_u8(frame_ptr[i + channel_stride * static_cast<size_t>(c)]);
+        }
+    }
+}
+
 static inline sd::Tensor<float> sd_image_to_preprocessing_tensor(sd_image_t image) {
     sd::Tensor<float> tensor({static_cast<int64_t>(image.width), static_cast<int64_t>(image.height), static_cast<int64_t>(image.channel), 1});
     for (uint32_t y = 0; y < image.height; ++y) {
@@ -39,20 +108,7 @@ static inline sd::Tensor<float> sd_image_to_preprocessing_tensor(sd_image_t imag
 static inline void preprocessing_tensor_to_sd_image(const sd::Tensor<float>& tensor, uint8_t* image_data) {
     GGML_ASSERT(tensor.dim() == 4);
     GGML_ASSERT(tensor.shape()[3] == 1);
-    GGML_ASSERT(image_data != nullptr);
-
-    int width   = static_cast<int>(tensor.shape()[0]);
-    int height  = static_cast<int>(tensor.shape()[1]);
-    int channel = static_cast<int>(tensor.shape()[2]);
-    for (int y = 0; y < height; ++y) {
-        for (int x = 0; x < width; ++x) {
-            for (int c = 0; c < channel; ++c) {
-                float value                               = preprocessing_get_4d(tensor, x, y, c, 0);
-                value                                     = std::min(1.0f, std::max(0.0f, value));
-                image_data[(y * width + x) * channel + c] = static_cast<uint8_t>(std::round(value * 255.0f));
-            }
-        }
-    }
+    preprocessing_tensor_frame_to_sd_image(tensor, 0, image_data);
 }
 
 static inline sd::Tensor<float> gaussian_kernel_tensor(int kernel_size) {

src/qwen_image.hpp

@@ -95,9 +95,7 @@ namespace Qwen {
 
             float scale         = 1.f / 32.f;
             bool force_prec_f32 = false;
-#ifdef SD_USE_VULKAN
-            force_prec_f32 = true;
-#endif
+
             // The purpose of the scale here is to prevent NaN issues in certain situations.
             // For example when using CUDA but the weights are k-quants (not all prompts).
             blocks["to_out.0"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, out_dim, out_bias, false, force_prec_f32, scale));
@@ -124,6 +122,10 @@ namespace Qwen {
             auto to_v     = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
             auto to_out_0 = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
 
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                to_out_0->set_force_prec_f32(true);
+            }
+
             auto norm_added_q = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_q"]);
             auto norm_added_k = std::dynamic_pointer_cast<UnaryBlock>(blocks["norm_added_k"]);
 
@@ -410,6 +412,9 @@ namespace Qwen {
             auto img = img_in->forward(ctx, x);
             auto txt = txt_norm->forward(ctx, context);
             txt      = txt_in->forward(ctx, txt);
+            sd::ggml_graph_cut::mark_graph_cut(img, "qwen_image.prelude", "img");
+            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++) {
                 auto block = std::dynamic_pointer_cast<QwenImageTransformerBlock>(blocks["transformer_blocks." + std::to_string(i)]);
@@ -417,6 +422,8 @@ namespace Qwen {
                 auto result = block->forward(ctx, img, txt, t_emb, pe, modulate_index);
                 img         = result.first;
                 txt         = result.second;
+                sd::ggml_graph_cut::mark_graph_cut(img, "qwen_image.transformer_blocks." + std::to_string(i), "img");
+                sd::ggml_graph_cut::mark_graph_cut(txt, "qwen_image.transformer_blocks." + std::to_string(i), "txt");
             }
 
             if (params.zero_cond_t) {
@@ -481,12 +488,12 @@ namespace Qwen {
         SDVersion version;
 
         QwenImageRunner(ggml_backend_t backend,
-                        bool offload_params_to_cpu,
+                        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, offload_params_to_cpu) {
+            : 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) {
@@ -679,7 +686,7 @@ namespace Qwen {
             }
 
             std::shared_ptr<QwenImageRunner> qwen_image = std::make_shared<QwenImageRunner>(backend,
-                                                                                            false,
+                                                                                            backend,
                                                                                             tensor_storage_map,
                                                                                             "model.diffusion_model",
                                                                                             VERSION_QWEN_IMAGE);

src/t5.hpp

@@ -251,7 +251,8 @@ public:
                          ggml_tensor* x,
                          ggml_tensor* past_bias                = nullptr,
                          ggml_tensor* attention_mask           = nullptr,
-                         ggml_tensor* relative_position_bucket = nullptr) {
+                         ggml_tensor* relative_position_bucket = nullptr,
+                         const std::string& graph_cut_prefix   = "") {
         // x: [N, n_token, model_dim]
         for (int i = 0; i < num_layers; i++) {
             auto block = std::dynamic_pointer_cast<T5Block>(blocks["block." + std::to_string(i)]);
@@ -259,6 +260,9 @@ public:
             auto ret  = block->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
             x         = ret.first;
             past_bias = ret.second;
+            if (!graph_cut_prefix.empty()) {
+                sd::ggml_graph_cut::mark_graph_cut(x, graph_cut_prefix + ".block." + std::to_string(i), "x");
+            }
         }
 
         auto final_layer_norm = std::dynamic_pointer_cast<T5LayerNorm>(blocks["final_layer_norm"]);
@@ -305,7 +309,8 @@ public:
         auto encoder = std::dynamic_pointer_cast<T5Stack>(blocks["encoder"]);
 
         auto x = shared->forward(ctx, input_ids);
-        x      = encoder->forward(ctx, x, past_bias, attention_mask, relative_position_bucket);
+        sd::ggml_graph_cut::mark_graph_cut(x, "t5.prelude", "x");
+        x = encoder->forward(ctx, x, past_bias, attention_mask, relative_position_bucket, "t5");
         return x;
     }
 };
@@ -316,11 +321,11 @@ struct T5Runner : public GGMLRunner {
     std::vector<int> relative_position_bucket_vec;
 
     T5Runner(ggml_backend_t backend,
-             bool offload_params_to_cpu,
+             ggml_backend_t params_backend,
              const String2TensorStorage& tensor_storage_map,
              const std::string prefix,
              bool is_umt5 = false)
-        : GGMLRunner(backend, offload_params_to_cpu) {
+        : GGMLRunner(backend, params_backend) {
         if (is_umt5) {
             params.vocab_size         = 256384;
             params.relative_attention = false;
@@ -459,11 +464,11 @@ struct T5Embedder {
     T5Runner model;
 
     T5Embedder(ggml_backend_t backend,
-               bool offload_params_to_cpu,
+               ggml_backend_t params_backend,
                const String2TensorStorage& tensor_storage_map = {},
                const std::string prefix                       = "",
                bool is_umt5                                   = false)
-        : model(backend, offload_params_to_cpu, tensor_storage_map, prefix, is_umt5), tokenizer(is_umt5) {
+        : model(backend, params_backend, tensor_storage_map, prefix, is_umt5), tokenizer(is_umt5) {
     }
 
     void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
@@ -571,7 +576,7 @@ struct T5Embedder {
             }
         }
 
-        std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, false, tensor_storage_map, "", true);
+        std::shared_ptr<T5Embedder> t5 = std::make_shared<T5Embedder>(backend, backend, tensor_storage_map, "", true);
 
         t5->alloc_params_buffer();
         std::map<std::string, ggml_tensor*> tensors;

src/tae.hpp

@@ -542,14 +542,14 @@ struct TinyImageAutoEncoder : public VAE {
     bool decode_only = false;
 
     TinyImageAutoEncoder(ggml_backend_t backend,
-                         bool offload_params_to_cpu,
+                         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, offload_params_to_cpu) {
+          VAE(version, backend, params_backend) {
         scale_input = false;
         taesd.init(params_ctx, tensor_storage_map, prefix);
     }
@@ -604,14 +604,14 @@ struct TinyVideoAutoEncoder : public VAE {
     bool decode_only = false;
 
     TinyVideoAutoEncoder(ggml_backend_t backend,
-                         bool offload_params_to_cpu,
+                         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, offload_params_to_cpu) {
+          VAE(version, backend, params_backend) {
         scale_input = false;
         taehv.init(params_ctx, tensor_storage_map, prefix);
     }

src/tensor.hpp

@@ -815,11 +815,202 @@ namespace sd {
     namespace ops {
         enum class InterpolateMode {
             Nearest,
+            NearestExact,
             NearestMax,
             NearestMin,
             NearestAvg,
+            Bilinear,
+            Bicubic,
+            Lanczos,
         };
 
+        inline bool is_nearest_like_interpolate_mode(InterpolateMode mode) {
+            return mode == InterpolateMode::Nearest ||
+                   mode == InterpolateMode::NearestExact ||
+                   mode == InterpolateMode::NearestMax ||
+                   mode == InterpolateMode::NearestMin ||
+                   mode == InterpolateMode::NearestAvg;
+        }
+
+        inline bool is_2d_filter_interpolate_mode(InterpolateMode mode) {
+            return mode == InterpolateMode::Bilinear ||
+                   mode == InterpolateMode::Bicubic ||
+                   mode == InterpolateMode::Lanczos;
+        }
+
+        inline int64_t nearest_exact_interpolate_index(int64_t output_index,
+                                                       int64_t input_size,
+                                                       int64_t output_size) {
+            const double scale  = static_cast<double>(input_size) / static_cast<double>(output_size);
+            const double center = (static_cast<double>(output_index) + 0.5) * scale - 0.5;
+            return std::min(std::max<int64_t>(static_cast<int64_t>(std::floor(center + 0.5)), 0), input_size - 1);
+        }
+
+        inline double linear_interpolate_weight(double x) {
+            x = std::abs(x);
+            return x < 1.0 ? 1.0 - x : 0.0;
+        }
+
+        inline double cubic_interpolate_weight(double x) {
+            constexpr double a = -0.75;  // Match PyTorch bicubic interpolation.
+            x                  = std::abs(x);
+            if (x <= 1.0) {
+                return ((a + 2.0) * x - (a + 3.0)) * x * x + 1.0;
+            }
+            if (x < 2.0) {
+                return ((a * x - 5.0 * a) * x + 8.0 * a) * x - 4.0 * a;
+            }
+            return 0.0;
+        }
+
+        inline double sinc(double x) {
+            constexpr double pi = 3.14159265358979323846;
+            if (std::abs(x) < 1e-12) {
+                return 1.0;
+            }
+            const double pix = pi * x;
+            return std::sin(pix) / pix;
+        }
+
+        inline double lanczos_interpolate_weight(double x) {
+            constexpr double radius = 3.0;
+            x                       = std::abs(x);
+            if (x >= radius) {
+                return 0.0;
+            }
+            return sinc(x) * sinc(x / radius);
+        }
+
+        struct InterpolateContributor {
+            int64_t index;
+            double weight;
+        };
+
+        inline std::vector<std::vector<InterpolateContributor>> make_interpolate_contributors(
+            int64_t input_size,
+            int64_t output_size,
+            InterpolateMode mode,
+            bool antialias) {
+            std::vector<std::vector<InterpolateContributor>> contributors(static_cast<size_t>(output_size));
+            const double scale        = static_cast<double>(input_size) / static_cast<double>(output_size);
+            const double filter_scale = antialias ? std::max(1.0, scale) : 1.0;
+
+            for (int64_t out = 0; out < output_size; ++out) {
+                const double center = (static_cast<double>(out) + 0.5) * scale - 0.5;
+                int64_t start       = 0;
+                int64_t end         = 0;
+
+                if (mode == InterpolateMode::Bilinear) {
+                    const double support = filter_scale;
+                    start                = static_cast<int64_t>(std::ceil(center - support));
+                    end                  = static_cast<int64_t>(std::floor(center + support));
+                } else if (mode == InterpolateMode::Bicubic) {
+                    const double support = 2.0 * filter_scale;
+                    start                = static_cast<int64_t>(std::ceil(center - support));
+                    end                  = static_cast<int64_t>(std::floor(center + support));
+                } else if (mode == InterpolateMode::Lanczos) {
+                    const double support = 3.0 * filter_scale;
+                    start                = static_cast<int64_t>(std::ceil(center - support));
+                    end                  = static_cast<int64_t>(std::floor(center + support));
+                } else {
+                    tensor_throw_invalid_argument("Unsupported 2D filter interpolate mode: mode=" +
+                                                  std::to_string(static_cast<int>(mode)));
+                }
+
+                double weight_sum                                      = 0.0;
+                std::vector<InterpolateContributor>& axis_contributors = contributors[static_cast<size_t>(out)];
+                axis_contributors.reserve(static_cast<size_t>(end - start + 1));
+
+                for (int64_t in = start; in <= end; ++in) {
+                    double weight = 0.0;
+                    if (mode == InterpolateMode::Bilinear) {
+                        weight = linear_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
+                    } else if (mode == InterpolateMode::Bicubic) {
+                        weight = cubic_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
+                    } else {
+                        weight = lanczos_interpolate_weight((center - static_cast<double>(in)) / filter_scale);
+                    }
+
+                    if (weight == 0.0) {
+                        continue;
+                    }
+
+                    const int64_t clamped_index = std::min(std::max<int64_t>(in, 0), input_size - 1);
+                    axis_contributors.push_back({clamped_index, weight});
+                    weight_sum += weight;
+                }
+
+                if ((antialias || mode == InterpolateMode::Lanczos) &&
+                    std::abs(weight_sum) > 1e-12) {
+                    for (auto& contributor : axis_contributors) {
+                        contributor.weight /= weight_sum;
+                    }
+                }
+
+                if (axis_contributors.empty()) {
+                    const int64_t nearest = std::min(
+                        std::max<int64_t>(static_cast<int64_t>(std::floor(center + 0.5)), 0),
+                        input_size - 1);
+                    axis_contributors.push_back({nearest, 1.0});
+                }
+            }
+
+            return contributors;
+        }
+
+        template <typename T>
+        inline Tensor<T> interpolate_2d_filter(const Tensor<T>& input,
+                                               const std::vector<int64_t>& output_shape,
+                                               InterpolateMode mode,
+                                               bool antialias) {
+            if (input.dim() < 2) {
+                tensor_throw_invalid_argument("2D filter interpolate requires rank >= 2: input_shape=" +
+                                              tensor_shape_to_string(input.shape()) + ", output_shape=" +
+                                              tensor_shape_to_string(output_shape));
+            }
+            for (size_t i = 2; i < output_shape.size(); ++i) {
+                if (input.shape()[i] != output_shape[i]) {
+                    tensor_throw_invalid_argument("2D filter interpolate only supports resizing dimensions 0 and 1: input_shape=" +
+                                                  tensor_shape_to_string(input.shape()) + ", output_shape=" +
+                                                  tensor_shape_to_string(output_shape));
+                }
+            }
+
+            Tensor<T> output(output_shape);
+            const int64_t input_width   = input.shape()[0];
+            const int64_t input_height  = input.shape()[1];
+            const int64_t output_width  = output_shape[0];
+            const int64_t output_height = output_shape[1];
+            const int64_t input_plane   = input_width * input_height;
+            const int64_t output_plane  = output_width * output_height;
+            const int64_t plane_count   = input.numel() / input_plane;
+
+            auto x_contributors = make_interpolate_contributors(input_width, output_width, mode, antialias);
+            auto y_contributors = make_interpolate_contributors(input_height, output_height, mode, antialias);
+
+            for (int64_t plane = 0; plane < plane_count; ++plane) {
+                const int64_t input_plane_offset  = plane * input_plane;
+                const int64_t output_plane_offset = plane * output_plane;
+                for (int64_t y = 0; y < output_height; ++y) {
+                    const auto& y_axis = y_contributors[static_cast<size_t>(y)];
+                    for (int64_t x = 0; x < output_width; ++x) {
+                        const auto& x_axis = x_contributors[static_cast<size_t>(x)];
+                        double value       = 0.0;
+                        for (const auto& yc : y_axis) {
+                            const int64_t input_row_offset = input_plane_offset + yc.index * input_width;
+                            for (const auto& xc : x_axis) {
+                                value += static_cast<double>(input.data()[input_row_offset + xc.index]) *
+                                         xc.weight * yc.weight;
+                            }
+                        }
+                        output.data()[output_plane_offset + y * output_width + x] = static_cast<T>(value);
+                    }
+                }
+            }
+
+            return output;
+        }
+
         inline int64_t normalize_slice_bound(int64_t index, int64_t dim_size) {
             if (index < 0) {
                 index += dim_size;
@@ -1014,17 +1205,20 @@ namespace sd {
         inline Tensor<T> interpolate(const Tensor<T>& input,
                                      std::vector<int64_t> output_shape,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false) {
-            const bool is_nearest_like_mode = (mode == InterpolateMode::Nearest ||
-                                               mode == InterpolateMode::NearestMax ||
-                                               mode == InterpolateMode::NearestMin ||
-                                               mode == InterpolateMode::NearestAvg);
-            if (!is_nearest_like_mode) {
-                tensor_throw_invalid_argument("Only nearest-like interpolate modes are implemented, got mode=" +
+                                     bool align_corners   = false,
+                                     bool antialias       = false) {
+            const bool is_nearest_like_mode = is_nearest_like_interpolate_mode(mode);
+            const bool is_2d_filter_mode    = is_2d_filter_interpolate_mode(mode);
+            if (!is_nearest_like_mode && !is_2d_filter_mode) {
+                tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
+                                              std::to_string(static_cast<int>(mode)));
+            }
+            if (antialias && !is_2d_filter_mode) {
+                tensor_throw_invalid_argument("Tensor interpolate antialias requires a 2D filter mode: mode=" +
                                               std::to_string(static_cast<int>(mode)));
             }
             if (align_corners) {
-                tensor_throw_invalid_argument("align_corners is not supported for nearest-like interpolate: input_shape=" +
+                tensor_throw_invalid_argument("align_corners is not supported for tensor interpolate: input_shape=" +
                                               tensor_shape_to_string(input.shape()) + ", output_shape=" +
                                               tensor_shape_to_string(output_shape));
             }
@@ -1051,6 +1245,10 @@ namespace sd {
                 }
             }
 
+            if (is_2d_filter_mode) {
+                return interpolate_2d_filter(input, output_shape, mode, antialias);
+            }
+
             bool has_downsampling = false;
             for (int64_t i = 0; i < input.dim(); ++i) {
                 if (input.shape()[i] > output_shape[i]) {
@@ -1060,12 +1258,20 @@ namespace sd {
             }
 
             Tensor<T> output(std::move(output_shape));
-            if (mode == InterpolateMode::Nearest || !has_downsampling) {
+            if (mode == InterpolateMode::Nearest ||
+                mode == InterpolateMode::NearestExact ||
+                !has_downsampling) {
                 for (int64_t flat = 0; flat < output.numel(); ++flat) {
                     std::vector<int64_t> output_coord = tensor_unravel_index(flat, output.shape());
                     std::vector<int64_t> input_coord(static_cast<size_t>(input.dim()), 0);
                     for (size_t i = 0; i < static_cast<size_t>(input.dim()); ++i) {
-                        input_coord[i] = output_coord[i] * input.shape()[i] / output.shape()[i];
+                        if (mode == InterpolateMode::NearestExact) {
+                            input_coord[i] = nearest_exact_interpolate_index(output_coord[i],
+                                                                             input.shape()[i],
+                                                                             output.shape()[i]);
+                        } else {
+                            input_coord[i] = output_coord[i] * input.shape()[i] / output.shape()[i];
+                        }
                     }
                     output[flat] = input.index(input_coord);
                 }
@@ -1083,6 +1289,12 @@ namespace sd {
                         return T(0);
                     case InterpolateMode::Nearest:
                         return T(0);
+                    case InterpolateMode::NearestExact:
+                        return T(0);
+                    case InterpolateMode::Bilinear:
+                    case InterpolateMode::Bicubic:
+                    case InterpolateMode::Lanczos:
+                        break;
                 }
 
                 tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
@@ -1102,6 +1314,12 @@ namespace sd {
                         break;
                     case InterpolateMode::Nearest:
                         break;
+                    case InterpolateMode::NearestExact:
+                        break;
+                    case InterpolateMode::Bilinear:
+                    case InterpolateMode::Bicubic:
+                    case InterpolateMode::Lanczos:
+                        break;
                 }
             };
 
@@ -1157,17 +1375,20 @@ namespace sd {
                                      const std::optional<std::vector<int64_t>>& size,
                                      const std::optional<std::vector<double>>& scale_factor,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false) {
-            const bool is_nearest_like_mode = (mode == InterpolateMode::Nearest ||
-                                               mode == InterpolateMode::NearestMax ||
-                                               mode == InterpolateMode::NearestMin ||
-                                               mode == InterpolateMode::NearestAvg);
-            if (!is_nearest_like_mode) {
-                tensor_throw_invalid_argument("Only nearest-like interpolate modes are implemented, got mode=" +
+                                     bool align_corners   = false,
+                                     bool antialias       = false) {
+            const bool is_nearest_like_mode = is_nearest_like_interpolate_mode(mode);
+            const bool is_2d_filter_mode    = is_2d_filter_interpolate_mode(mode);
+            if (!is_nearest_like_mode && !is_2d_filter_mode) {
+                tensor_throw_invalid_argument("Unsupported interpolate mode: mode=" +
+                                              std::to_string(static_cast<int>(mode)));
+            }
+            if (antialias && !is_2d_filter_mode) {
+                tensor_throw_invalid_argument("Tensor interpolate antialias requires a 2D filter mode: mode=" +
                                               std::to_string(static_cast<int>(mode)));
             }
             if (align_corners) {
-                tensor_throw_invalid_argument("align_corners is not supported for nearest-like interpolate: input_shape=" +
+                tensor_throw_invalid_argument("align_corners is not supported for tensor interpolate: input_shape=" +
                                               tensor_shape_to_string(input.shape()));
             }
             if (size.has_value() == scale_factor.has_value()) {
@@ -1211,7 +1432,7 @@ namespace sd {
                 }
             }
 
-            return interpolate(input, std::move(output_shape), mode, align_corners);
+            return interpolate(input, std::move(output_shape), mode, align_corners, antialias);
         }
 
         template <typename T>
@@ -1219,12 +1440,14 @@ namespace sd {
                                      const std::optional<std::vector<int64_t>>& size,
                                      double scale_factor,
                                      InterpolateMode mode = InterpolateMode::Nearest,
-                                     bool align_corners   = false) {
+                                     bool align_corners   = false,
+                                     bool antialias       = false) {
             return interpolate(input,
                                size,
                                std::vector<double>(size.has_value() ? size->size() : input.dim(), scale_factor),
                                mode,
-                               align_corners);
+                               align_corners,
+                               antialias);
         }
 
         template <typename T>

src/tokenizers/clip_tokenizer.cpp

@@ -62,7 +62,7 @@ void CLIPTokenizer::load_from_merges(const std::string& merges_utf8_str) {
     }
     vocab.push_back(utf8_to_utf32("<|startoftext|>"));
     vocab.push_back(utf8_to_utf32("<|endoftext|>"));
-    LOG_DEBUG("vocab size: %llu", vocab.size());
+    LOG_DEBUG("vocab size: %zu", vocab.size());
     int i = 0;
     for (const auto& token : vocab) {
         encoder[token] = i;

src/tokenizers/mistral_tokenizer.cpp

@@ -28,7 +28,7 @@ void MistralTokenizer::load_from_merges(const std::string& merges_utf8_str, cons
         byte_decoder[pair.second] = pair.first;
     }
     std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
-    LOG_DEBUG("merges size %llu", merges.size());
+    LOG_DEBUG("merges size %zu", merges.size());
     std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
     for (const auto& merge : merges) {
         size_t space_pos = merge.find(' ');

src/tokenizers/qwen2_tokenizer.cpp

@@ -11,7 +11,7 @@ void Qwen2Tokenizer::load_from_merges(const std::string& merges_utf8_str) {
     }
 
     std::vector<std::u32string> merges = split_utf32(merges_utf8_str);
-    LOG_DEBUG("merges size %llu", merges.size());
+    LOG_DEBUG("merges size %zu", merges.size());
     std::vector<std::pair<std::u32string, std::u32string>> merge_pairs;
     for (const auto& merge : merges) {
         size_t space_pos = merge.find(' ');
@@ -81,6 +81,11 @@ Qwen2Tokenizer::Qwen2Tokenizer(const std::string& merges_utf8_str) {
         "</tool_response>",
         "<think>",
         "</think>",
+        "<|boi_token|>",
+        "<|bor_token|>",
+        "<|eor_token|>",
+        "<|bot_token|>",
+        "<|tms_token|>",
     };
 
     if (merges_utf8_str.size() > 0) {

src/unet.hpp

@@ -482,12 +482,14 @@ public:
 
             emb = ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
         }
+        // sd::ggml_graph_cut::mark_graph_cut(emb, "unet.prelude", "emb");
 
         // input_blocks
         std::vector<ggml_tensor*> hs;
 
         // input block 0
         auto h = input_blocks_0_0->forward(ctx, x);
+        sd::ggml_graph_cut::mark_graph_cut(h, "unet.input_blocks.0", "h");
 
         ggml_set_name(h, "bench-start");
         hs.push_back(h);
@@ -505,6 +507,7 @@ public:
                     std::string name = "input_blocks." + std::to_string(input_block_idx) + ".1";
                     h                = attention_layer_forward(name, ctx, h, context, num_video_frames);  // [N, mult*model_channels, h, w]
                 }
+                sd::ggml_graph_cut::mark_graph_cut(h, "unet.input_blocks." + std::to_string(input_block_idx), "h");
                 hs.push_back(h);
             }
             if (tiny_unet) {
@@ -518,6 +521,7 @@ public:
                 auto block       = std::dynamic_pointer_cast<DownSampleBlock>(blocks[name]);
 
                 h = block->forward(ctx, h);  // [N, mult*model_channels, h/(2^(i+1)), w/(2^(i+1))]
+                // sd::ggml_graph_cut::mark_graph_cut(h, "unet.input_blocks." + std::to_string(input_block_idx), "h");
                 hs.push_back(h);
             }
         }
@@ -531,6 +535,7 @@ public:
                 h = resblock_forward("middle_block.2", ctx, h, emb, num_video_frames);             // [N, 4*model_channels, h/8, w/8]
             }
         }
+        sd::ggml_graph_cut::mark_graph_cut(h, "unet.middle_block", "h");
         if (controls.size() > 0) {
             auto cs = ggml_ext_scale(ctx->ggml_ctx, controls[controls.size() - 1], control_strength, true);
             h       = ggml_add(ctx->ggml_ctx, h, cs);  // middle control
@@ -581,6 +586,7 @@ public:
                 }
 
                 output_block_idx += 1;
+                sd::ggml_graph_cut::mark_graph_cut(h, "unet.output_blocks." + std::to_string(output_block_idx - 1), "h");
             }
         }
 
@@ -597,11 +603,11 @@ struct UNetModelRunner : public GGMLRunner {
     UnetModelBlock unet;
 
     UNetModelRunner(ggml_backend_t backend,
-                    bool offload_params_to_cpu,
+                    ggml_backend_t params_backend,
                     const String2TensorStorage& tensor_storage_map,
                     const std::string prefix,
                     SDVersion version = VERSION_SD1)
-        : GGMLRunner(backend, offload_params_to_cpu), unet(version, tensor_storage_map) {
+        : GGMLRunner(backend, params_backend), unet(version, tensor_storage_map) {
         unet.init(params_ctx, tensor_storage_map, prefix);
     }
 

src/upscaler.cpp

@@ -1,125 +1,144 @@
-#include "esrgan.hpp"
+#include "upscaler.h"
 #include "ggml_extend.hpp"
 #include "model.h"
 #include "stable-diffusion.h"
 #include "util.h"
 
-struct UpscalerGGML {
-    ggml_backend_t backend    = nullptr;  // general backend
-    ggml_type model_data_type = GGML_TYPE_F16;
-    std::shared_ptr<ESRGAN> esrgan_upscaler;
-    std::string esrgan_path;
-    int n_threads;
-    bool direct   = false;
-    int tile_size = 128;
+#include <utility>
 
-    UpscalerGGML(int n_threads,
-                 bool direct   = false,
-                 int tile_size = 128)
-        : n_threads(n_threads),
-          direct(direct),
-          tile_size(tile_size) {
+UpscalerGGML::UpscalerGGML(int n_threads,
+                           bool direct,
+                           int tile_size,
+                           std::string backend_spec,
+                           std::string params_backend_spec)
+    : n_threads(n_threads),
+      direct(direct),
+      tile_size(tile_size),
+      backend_spec(std::move(backend_spec)),
+      params_backend_spec(std::move(params_backend_spec)) {
+}
+
+void UpscalerGGML::set_max_graph_vram_bytes(size_t max_vram_bytes) {
+    max_graph_vram_bytes = max_vram_bytes;
+    if (esrgan_upscaler) {
+        esrgan_upscaler->set_max_graph_vram_bytes(max_vram_bytes);
     }
+}
 
-    bool load_from_file(const std::string& esrgan_path,
-                        bool offload_params_to_cpu,
-                        int n_threads) {
-        ggml_log_set(ggml_log_callback_default, nullptr);
-#ifdef SD_USE_CUDA
-        LOG_DEBUG("Using CUDA backend");
-        backend = ggml_backend_cuda_init(0);
-#endif
-#ifdef SD_USE_METAL
-        LOG_DEBUG("Using Metal backend");
-        backend = ggml_backend_metal_init();
-#endif
-#ifdef SD_USE_VULKAN
-        LOG_DEBUG("Using Vulkan backend");
-        backend = ggml_backend_vk_init(0);
-#endif
-#ifdef SD_USE_OPENCL
-        LOG_DEBUG("Using OpenCL backend");
-        backend = ggml_backend_opencl_init();
-#endif
-#ifdef SD_USE_SYCL
-        LOG_DEBUG("Using SYCL backend");
-        backend = ggml_backend_sycl_init(0);
-#endif
-        ModelLoader model_loader;
-        if (!model_loader.init_from_file_and_convert_name(esrgan_path)) {
-            LOG_ERROR("init model loader from file failed: '%s'", esrgan_path.c_str());
+bool UpscalerGGML::load_from_file(const std::string& esrgan_path,
+                                  bool offload_params_to_cpu,
+                                  int n_threads) {
+    ggml_log_set(ggml_log_callback_default, nullptr);
+
+    std::string error;
+    if (!backend_manager.init(backend_spec.c_str(),
+                              params_backend_spec.c_str(),
+                              offload_params_to_cpu,
+                              false,
+                              false,
+                              false,
+                              &error)) {
+        LOG_ERROR("upscaler backend config failed: %s", error.c_str());
+        return false;
+    }
+    auto backend_for = [&](SDBackendModule module) {
+        ggml_backend_t module_backend = backend_manager.runtime_backend(module);
+        if (module_backend == nullptr) {
+            LOG_ERROR("failed to initialize %s backend", sd_backend_module_name(module));
         }
-        model_loader.set_wtype_override(model_data_type);
-        if (!backend) {
-            LOG_DEBUG("Using CPU backend");
-            backend = ggml_backend_cpu_init();
+        return module_backend;
+    };
+    auto params_backend_for = [&](SDBackendModule module) {
+        ggml_backend_t module_backend = backend_manager.params_backend(module);
+        if (module_backend == nullptr) {
+            LOG_ERROR("failed to initialize %s params backend", sd_backend_module_name(module));
         }
-        LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
-        esrgan_upscaler = std::make_shared<ESRGAN>(backend, offload_params_to_cpu, tile_size, model_loader.get_tensor_storage_map());
-        if (direct) {
-            esrgan_upscaler->set_conv2d_direct_enabled(true);
-        }
-        if (!esrgan_upscaler->load_from_file(esrgan_path, n_threads)) {
+        return module_backend;
+    };
+    auto ensure_backend_pair = [&](SDBackendModule module) {
+        if (backend_for(module) == nullptr) {
             return false;
         }
-        return true;
+        return params_backend_for(module) != nullptr;
+    };
+    if (!ensure_backend_pair(SDBackendModule::UPSCALER)) {
+        return false;
     }
 
-    sd::Tensor<float> upscale_tensor(const sd::Tensor<float>& input_tensor) {
-        sd::Tensor<float> upscaled;
-        if (tile_size <= 0 || (input_tensor.shape()[0] <= tile_size && input_tensor.shape()[1] <= tile_size)) {
-            upscaled = esrgan_upscaler->compute(n_threads, input_tensor);
-        } else {
-            auto on_processing = [&](const sd::Tensor<float>& input_tile) -> sd::Tensor<float> {
-                auto output_tile = esrgan_upscaler->compute(n_threads, input_tile);
-                if (output_tile.empty()) {
-                    LOG_ERROR("esrgan compute failed while processing a tile");
-                    return {};
-                }
-                return output_tile;
-            };
-
-            upscaled = process_tiles_2d(input_tensor,
-                                        static_cast<int>(input_tensor.shape()[0] * esrgan_upscaler->scale),
-                                        static_cast<int>(input_tensor.shape()[1] * esrgan_upscaler->scale),
-                                        esrgan_upscaler->scale,
-                                        tile_size,
-                                        tile_size,
-                                        0.25f,
-                                        false,
-                                        false,
-                                        on_processing);
-        }
-        esrgan_upscaler->free_compute_buffer();
-        if (upscaled.empty()) {
-            LOG_ERROR("esrgan compute failed");
-            return {};
-        }
-        return upscaled;
+    ModelLoader model_loader;
+    if (!model_loader.init_from_file_and_convert_name(esrgan_path)) {
+        LOG_ERROR("init model loader from file failed: '%s'", esrgan_path.c_str());
     }
+    model_loader.set_wtype_override(model_data_type);
+    LOG_INFO("Upscaler weight type: %s", ggml_type_name(model_data_type));
+    esrgan_upscaler = std::make_shared<ESRGAN>(backend_for(SDBackendModule::UPSCALER),
+                                               params_backend_for(SDBackendModule::UPSCALER),
+                                               tile_size,
+                                               model_loader.get_tensor_storage_map());
+    esrgan_upscaler->set_max_graph_vram_bytes(max_graph_vram_bytes);
+    if (direct) {
+        esrgan_upscaler->set_conv2d_direct_enabled(true);
+    }
+    if (!esrgan_upscaler->load_from_file(esrgan_path, n_threads)) {
+        return false;
+    }
+    return true;
+}
 
-    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor) {
-        // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
-        sd_image_t upscaled_image = {0, 0, 0, nullptr};
-        int output_width          = (int)input_image.width * esrgan_upscaler->scale;
-        int output_height         = (int)input_image.height * esrgan_upscaler->scale;
-        LOG_INFO("upscaling from (%i x %i) to (%i x %i)",
-                 input_image.width, input_image.height, output_width, output_height);
+sd::Tensor<float> UpscalerGGML::upscale_tensor(const sd::Tensor<float>& input_tensor) {
+    sd::Tensor<float> upscaled;
+    if (tile_size <= 0 || (input_tensor.shape()[0] <= tile_size && input_tensor.shape()[1] <= tile_size)) {
+        upscaled = esrgan_upscaler->compute(n_threads, input_tensor);
+    } else {
+        auto on_processing = [&](const sd::Tensor<float>& input_tile) -> sd::Tensor<float> {
+            auto output_tile = esrgan_upscaler->compute(n_threads, input_tile);
+            if (output_tile.empty()) {
+                LOG_ERROR("esrgan compute failed while processing a tile");
+                return {};
+            }
+            return output_tile;
+        };
 
-        sd::Tensor<float> input_tensor = sd_image_to_tensor(input_image);
-        sd::Tensor<float> upscaled;
-        int64_t t0 = ggml_time_ms();
-        upscaled   = upscale_tensor(input_tensor);
-        if (upscaled.empty()) {
-            return upscaled_image;
-        }
-        sd_image_t upscaled_data = tensor_to_sd_image(upscaled);
-        int64_t t3               = ggml_time_ms();
-        LOG_INFO("input_image_tensor upscaled, taking %.2fs", (t3 - t0) / 1000.0f);
-        upscaled_image = upscaled_data;
+        upscaled = process_tiles_2d(input_tensor,
+                                    static_cast<int>(input_tensor.shape()[0] * esrgan_upscaler->scale),
+                                    static_cast<int>(input_tensor.shape()[1] * esrgan_upscaler->scale),
+                                    esrgan_upscaler->scale,
+                                    tile_size,
+                                    tile_size,
+                                    0.25f,
+                                    false,
+                                    false,
+                                    on_processing);
+    }
+    esrgan_upscaler->free_compute_buffer();
+    if (upscaled.empty()) {
+        LOG_ERROR("esrgan compute failed");
+        return {};
+    }
+    return upscaled;
+}
+
+sd_image_t UpscalerGGML::upscale(sd_image_t input_image, uint32_t upscale_factor) {
+    // upscale_factor, unused for RealESRGAN_x4plus_anime_6B.pth
+    sd_image_t upscaled_image = {0, 0, 0, nullptr};
+    int output_width          = (int)input_image.width * esrgan_upscaler->scale;
+    int output_height         = (int)input_image.height * esrgan_upscaler->scale;
+    LOG_INFO("upscaling from (%i x %i) to (%i x %i)",
+             input_image.width, input_image.height, output_width, output_height);
+
+    sd::Tensor<float> input_tensor = sd_image_to_tensor(input_image);
+    sd::Tensor<float> upscaled;
+    int64_t t0 = ggml_time_ms();
+    upscaled   = upscale_tensor(input_tensor);
+    if (upscaled.empty()) {
         return upscaled_image;
     }
-};
+    sd_image_t upscaled_data = tensor_to_sd_image(upscaled);
+    int64_t t3               = ggml_time_ms();
+    LOG_INFO("input_image_tensor upscaled, taking %.2fs", (t3 - t0) / 1000.0f);
+    upscaled_image = upscaled_data;
+    return upscaled_image;
+}
 
 struct upscaler_ctx_t {
     UpscalerGGML* upscaler = nullptr;
@@ -129,14 +148,16 @@ upscaler_ctx_t* new_upscaler_ctx(const char* esrgan_path_c_str,
                                  bool offload_params_to_cpu,
                                  bool direct,
                                  int n_threads,
-                                 int tile_size) {
+                                 int tile_size,
+                                 const char* backend,
+                                 const char* params_backend) {
     upscaler_ctx_t* upscaler_ctx = (upscaler_ctx_t*)malloc(sizeof(upscaler_ctx_t));
     if (upscaler_ctx == nullptr) {
         return nullptr;
     }
     std::string esrgan_path(esrgan_path_c_str);
 
-    upscaler_ctx->upscaler = new UpscalerGGML(n_threads, direct, tile_size);
+    upscaler_ctx->upscaler = new UpscalerGGML(n_threads, direct, tile_size, SAFE_STR(backend), SAFE_STR(params_backend));
     if (upscaler_ctx->upscaler == nullptr) {
         return nullptr;
     }

src/upscaler.h

@@ -0,0 +1,38 @@
+#ifndef __SD_UPSCALER_H__
+#define __SD_UPSCALER_H__
+
+#include "esrgan.hpp"
+#include "ggml_extend_backend.h"
+#include "stable-diffusion.h"
+#include "tensor.hpp"
+
+#include <memory>
+#include <string>
+
+struct UpscalerGGML {
+    SDBackendManager backend_manager;
+    ggml_type model_data_type = GGML_TYPE_F16;
+    std::shared_ptr<ESRGAN> esrgan_upscaler;
+    std::string esrgan_path;
+    int n_threads;
+    bool direct                 = false;
+    int tile_size               = 128;
+    size_t max_graph_vram_bytes = 0;
+    std::string backend_spec;
+    std::string params_backend_spec;
+
+    UpscalerGGML(int n_threads,
+                 bool direct                     = false,
+                 int tile_size                   = 128,
+                 std::string backend_spec        = "",
+                 std::string params_backend_spec = "");
+
+    bool load_from_file(const std::string& esrgan_path,
+                        bool offload_params_to_cpu,
+                        int n_threads);
+    void set_max_graph_vram_bytes(size_t max_vram_bytes);
+    sd::Tensor<float> upscale_tensor(const sd::Tensor<float>& input_tensor);
+    sd_image_t upscale(sd_image_t input_image, uint32_t upscale_factor);
+};
+
+#endif  // __SD_UPSCALER_H__

src/util.cpp

@@ -23,8 +23,9 @@
 #include <unistd.h>
 #endif
 
-#include "ggml-cpu.h"
+#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) {
@@ -111,7 +112,7 @@ private:
     HANDLE hmapping_;
 };
 
-std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
+std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bool writable) {
     void* mapped_data = nullptr;
     size_t file_size  = 0;
 
@@ -119,10 +120,10 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
         filename.c_str(),
         GENERIC_READ,
         FILE_SHARE_READ,
-        NULL,
+        nullptr,
         OPEN_EXISTING,
         FILE_ATTRIBUTE_NORMAL,
-        NULL);
+        nullptr);
 
     if (file_handle == INVALID_HANDLE_VALUE) {
         return nullptr;
@@ -136,16 +137,20 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
 
     file_size = static_cast<size_t>(size.QuadPart);
 
-    HANDLE mapping_handle = CreateFileMapping(file_handle, NULL, PAGE_READONLY, 0, 0, NULL);
+    DWORD page_prot = writable ? PAGE_WRITECOPY : PAGE_READONLY;
 
-    if (mapping_handle == NULL) {
+    HANDLE mapping_handle = CreateFileMapping(file_handle, nullptr, page_prot, 0, 0, nullptr);
+
+    if (mapping_handle == nullptr) {
         CloseHandle(file_handle);
         return nullptr;
     }
 
-    mapped_data = MapViewOfFile(mapping_handle, FILE_MAP_READ, 0, 0, file_size);
+    DWORD view_access = writable ? FILE_MAP_COPY : FILE_MAP_READ;
 
-    if (mapped_data == NULL) {
+    mapped_data = MapViewOfFile(mapping_handle, view_access, 0, 0, file_size);
+
+    if (mapped_data == nullptr) {
         CloseHandle(mapping_handle);
         CloseHandle(file_handle);
         return nullptr;
@@ -171,28 +176,85 @@ bool is_directory(const std::string& path) {
     return (stat(path.c_str(), &buffer) == 0 && S_ISDIR(buffer.st_mode));
 }
 
-class MmapWrapperImpl : public MmapWrapper {
-public:
-    MmapWrapperImpl(void* data, size_t size)
-        : MmapWrapper(data, size) {}
-
-    ~MmapWrapperImpl() override {
-        munmap(data_, size_);
-    }
+struct MmapFlags {
+    bool sequential;
+    bool populate;
+    bool willneed;
+    bool dontneed;
 };
 
-std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
+static MmapFlags get_mmap_flags() {
+    MmapFlags result          = {};
+    const char* SD_MMAP_FLAGS = std::getenv("SD_MMAP_FLAGS");
+    if (SD_MMAP_FLAGS && *SD_MMAP_FLAGS) {
+        std::stringstream ss(SD_MMAP_FLAGS);
+        std::string token;
+        while (std::getline(ss, token, ',')) {
+            std::string ntoken = trim(token);
+            std::transform(ntoken.begin(), ntoken.end(), ntoken.begin(), ::tolower);
+            if (ntoken == "sequential") {
+                result.sequential = true;
+            } else if (ntoken == "populate") {
+                result.populate = true;
+            } else if (ntoken == "willneed") {
+                result.willneed = true;
+            } else if (ntoken == "dontneed") {
+                result.dontneed = true;
+            }
+        }
+    }
+    return result;
+}
+
+class MmapWrapperImpl : public MmapWrapper {
+public:
+    MmapWrapperImpl(void* data, size_t size, int fd)
+        : MmapWrapper(data, size), fd_(fd) {}
+
+    ~MmapWrapperImpl() override {
+#ifdef __linux__
+        auto cfg_flags = get_mmap_flags();
+
+        // Drop the kernel pagecache pages for this file. madvise(DONTNEED)
+        // alone only unmaps from the process address space; pagecache
+        // entries persist (`free` reports them as buff/cache and the OOM
+        // killer doesn't touch them, but they ARE counted against
+        // overcommit and can starve other allocations on tight-RAM
+        // systems). posix_fadvise(POSIX_FADV_DONTNEED) is the documented
+        // way to evict pagecache for a specific fd's pages.
+        if (cfg_flags.dontneed) {
+            madvise(data_, size_, MADV_DONTNEED);
+            posix_fadvise(fd_, 0, 0, POSIX_FADV_DONTNEED);
+        }
+#endif
+        munmap(data_, size_);
+        close(fd_);
+    }
+
+private:
+    int fd_;
+};
+
+std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename, bool writable) {
     int file_descriptor = open(filename.c_str(), O_RDONLY);
     if (file_descriptor == -1) {
         return nullptr;
     }
 
+    auto cfg_flags = get_mmap_flags();
+
     int mmap_flags = MAP_PRIVATE;
 
 #ifdef __linux__
-    // performance flags used by llama.cpp
-    // posix_fadvise(file_descriptor, 0, 0, POSIX_FADV_SEQUENTIAL);
-    // mmap_flags |= MAP_POPULATE;
+    // Sequential access hint helps the kernel read-ahead efficiently and
+    // also encourages eviction of already-read pages (the kernel keeps
+    // a smaller working set when this is set).
+    if (cfg_flags.sequential) {
+        posix_fadvise(file_descriptor, 0, 0, POSIX_FADV_SEQUENTIAL);
+    }
+    if (cfg_flags.populate) {
+        mmap_flags |= MAP_POPULATE;
+    }
 #endif
 
     struct stat sb;
@@ -203,20 +265,27 @@ std::unique_ptr<MmapWrapper> MmapWrapper::create(const std::string& filename) {
 
     size_t file_size = sb.st_size;
 
-    void* mapped_data = mmap(NULL, file_size, PROT_READ, mmap_flags, file_descriptor, 0);
+    if (file_size == 0) {
+        close(file_descriptor);
+        return nullptr;
+    }
 
-    close(file_descriptor);
+    int mmap_prot = PROT_READ | (writable ? PROT_WRITE : 0);
+
+    void* mapped_data = mmap(nullptr, file_size, mmap_prot, mmap_flags, file_descriptor, 0);
 
     if (mapped_data == MAP_FAILED) {
+        close(file_descriptor);
         return nullptr;
     }
 
 #ifdef __linux__
-    // performance flags used by llama.cpp
-    // posix_madvise(mapped_data, file_size, POSIX_MADV_WILLNEED);
+    if (cfg_flags.willneed) {
+        posix_madvise(mapped_data, file_size, POSIX_MADV_WILLNEED);
+    }
 #endif
 
-    return std::make_unique<MmapWrapperImpl>(mapped_data, file_size);
+    return std::make_unique<MmapWrapperImpl>(mapped_data, file_size, file_descriptor);
 }
 
 #endif
@@ -495,26 +564,6 @@ sd_progress_cb_t sd_get_progress_callback() {
 void* sd_get_progress_callback_data() {
     return sd_progress_cb_data;
 }
-const char* sd_get_system_info() {
-    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;
-}
 
 sd_image_t tensor_to_sd_image(const sd::Tensor<float>& tensor, int frame_index) {
     const auto& shape = tensor.shape();
@@ -524,17 +573,7 @@ sd_image_t tensor_to_sd_image(const sd::Tensor<float>& tensor, int frame_index)
     int channel   = static_cast<int>(shape[shape.size() == 5 ? 3 : 2]);
     uint8_t* data = (uint8_t*)malloc(static_cast<size_t>(width * height * channel));
     GGML_ASSERT(data != nullptr);
-
-    for (int iw = 0; iw < width; ++iw) {
-        for (int ih = 0; ih < height; ++ih) {
-            for (int ic = 0; ic < channel; ++ic) {
-                float value                            = shape.size() == 5 ? tensor.index(iw, ih, frame_index, ic, 0)
-                                                                           : tensor.index(iw, ih, ic, frame_index);
-                value                                  = std::clamp(value, 0.0f, 1.0f);
-                data[(ih * width + iw) * channel + ic] = static_cast<uint8_t>(std::round(value * 255.0f));
-            }
-        }
-    }
+    preprocessing_tensor_frame_to_sd_image(tensor, frame_index, data);
     return {
         static_cast<uint32_t>(width),
         static_cast<uint32_t>(height),
@@ -718,3 +757,30 @@ 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"
+}
+
+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;
+}

src/util.h

@@ -6,6 +6,7 @@
 #include <string>
 #include <vector>
 
+#include "ggml-backend.h"
 #include "stable-diffusion.h"
 #include "tensor.hpp"
 
@@ -41,7 +42,7 @@ sd::Tensor<float> clip_preprocess(const sd::Tensor<float>& image, int target_wid
 
 class MmapWrapper {
 public:
-    static std::unique_ptr<MmapWrapper> create(const std::string& filename);
+    static std::unique_ptr<MmapWrapper> create(const std::string& filename, bool writable = false);
 
     virtual ~MmapWrapper() = default;
 
@@ -51,6 +52,7 @@ public:
     MmapWrapper& operator=(MmapWrapper&&)      = delete;
 
     const uint8_t* data() const { return static_cast<uint8_t*>(data_); }
+    uint8_t* writable_data() { return static_cast<uint8_t*>(data_); }
     size_t size() const { return size_; }
     bool copy_data(void* buf, size_t n, size_t offset) const;
 
@@ -82,6 +84,9 @@ int sd_get_preview_interval();
 bool sd_should_preview_denoised();
 bool sd_should_preview_noisy();
 
+// test if the backend is a specific one, e.g. "CUDA", "ROCm", "Vulkan" etc.
+bool sd_backend_is(ggml_backend_t backend, const std::string& name);
+
 #define LOG_DEBUG(format, ...) log_printf(SD_LOG_DEBUG, __FILE__, __LINE__, format, ##__VA_ARGS__)
 #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__)

src/vae.hpp

@@ -62,8 +62,8 @@ protected:
     }
 
 public:
-    VAE(SDVersion version, ggml_backend_t backend, bool offload_params_to_cpu)
-        : version(version), GGMLRunner(backend, offload_params_to_cpu) {}
+    VAE(SDVersion version, ggml_backend_t backend, ggml_backend_t params_backend)
+        : version(version), GGMLRunner(backend, params_backend) {}
 
     int get_scale_factor() {
         int scale_factor = 8;
@@ -71,7 +71,7 @@ public:
             scale_factor = 16;
         } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 16;
-        } else if (version == VERSION_CHROMA_RADIANCE) {
+        } else if (version == VERSION_CHROMA_RADIANCE || version == VERSION_HIDREAM_O1) {
             scale_factor = 1;
         }
         return scale_factor;
@@ -142,9 +142,10 @@ public:
                                    "vae encode compute failed while processing a tile");
         } else {
             output = _compute(n_threads, input, false);
-            free_compute_buffer();
         }
 
+        free_compute_buffer();
+
         if (output.empty()) {
             LOG_ERROR("vae encode compute failed");
             return {};
@@ -215,8 +216,8 @@ public:
 };
 
 struct FakeVAE : public VAE {
-    FakeVAE(SDVersion version, ggml_backend_t backend, bool offload_params_to_cpu)
-        : VAE(version, backend, offload_params_to_cpu) {}
+    FakeVAE(SDVersion version, ggml_backend_t backend, ggml_backend_t params_backend)
+        : VAE(version, backend, params_backend) {}
 
     int get_encoder_output_channels(int input_channels) {
         return input_channels;

src/wan.hpp

@@ -692,6 +692,7 @@ namespace WAN {
             } else {
                 x = conv1->forward(ctx, x);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.prelude", "x");
 
             // downsamples
             std::vector<int64_t> dims = {dim};
@@ -717,12 +718,14 @@ namespace WAN {
                         x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
                     }
                 }
+                // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.down." + std::to_string(i), "x");
             }
 
             // middle
             x = middle_0->forward(ctx, x, b, feat_cache, feat_idx);
             x = middle_1->forward(ctx, x, b);
             x = middle_2->forward(ctx, x, b, feat_cache, feat_idx);
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encoder.mid", "x");
 
             // head
             x = head_0->forward(ctx, x);
@@ -863,11 +866,13 @@ namespace WAN {
             } else {
                 x = conv1->forward(ctx, x);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.prelude", "x");
 
             // middle
             x = middle_0->forward(ctx, x, b, feat_cache, feat_idx);
             x = middle_1->forward(ctx, x, b);
             x = middle_2->forward(ctx, x, b, feat_cache, feat_idx);
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.mid", "x");
 
             // upsamples
             std::vector<int64_t> dims = {dim_mult[dim_mult.size() - 1] * dim};
@@ -893,6 +898,7 @@ namespace WAN {
                         x = layer->forward(ctx, x, b, feat_cache, feat_idx, chunk_idx);
                     }
                 }
+                // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decoder.up." + std::to_string(i), "x");
             }
 
             // head
@@ -1031,6 +1037,7 @@ namespace WAN {
             if (wan2_2) {
                 x = patchify(ctx->ggml_ctx, x, 2, b);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.encode.prelude", "x");
 
             auto encoder = std::dynamic_pointer_cast<Encoder3d>(blocks["encoder"]);
             auto conv1   = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
@@ -1051,6 +1058,7 @@ namespace WAN {
             }
             out     = conv1->forward(ctx, out);
             auto mu = ggml_ext_chunk(ctx->ggml_ctx, out, 2, 3)[0];
+            // sd::ggml_graph_cut::mark_graph_cut(mu, "wan_vae.encode.final", "mu");
             clear_cache();
             return mu;
         }
@@ -1068,6 +1076,7 @@ namespace WAN {
 
             int64_t iter_ = z->ne[2];
             auto x        = conv2->forward(ctx, z);
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decode.prelude", "x");
             ggml_tensor* out;
             for (int i = 0; i < iter_; i++) {
                 _conv_idx = 0;
@@ -1083,6 +1092,7 @@ namespace WAN {
             if (wan2_2) {
                 out = unpatchify(ctx->ggml_ctx, out, 2, b);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(out, "wan_vae.decode.final", "out");
             clear_cache();
             return out;
         }
@@ -1097,13 +1107,15 @@ namespace WAN {
             auto decoder = std::dynamic_pointer_cast<Decoder3d>(blocks["decoder"]);
             auto conv2   = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv2"]);
 
-            auto x    = conv2->forward(ctx, z);
+            auto x = conv2->forward(ctx, z);
+            // sd::ggml_graph_cut::mark_graph_cut(x, "wan_vae.decode_partial.prelude", "x");
             auto in   = ggml_ext_slice(ctx->ggml_ctx, x, 2, i, i + 1);  // [b*c, 1, h, w]
             _conv_idx = 0;
             auto out  = decoder->forward(ctx, in, b, _feat_map, _conv_idx, i);
             if (wan2_2) {
                 out = unpatchify(ctx->ggml_ctx, out, 2, b);
             }
+            // sd::ggml_graph_cut::mark_graph_cut(out, "wan_vae.decode_partial.final", "out");
             return out;
         }
     };
@@ -1114,12 +1126,12 @@ namespace WAN {
         WanVAE ae;
 
         WanVAERunner(ggml_backend_t backend,
-                     bool offload_params_to_cpu,
+                     ggml_backend_t params_backend,
                      const String2TensorStorage& tensor_storage_map = {},
                      const std::string prefix                       = "",
                      bool decode_only                               = false,
                      SDVersion version                              = VERSION_WAN2)
-            : decode_only(decode_only), ae(decode_only, version == VERSION_WAN2_2_TI2V), VAE(version, backend, offload_params_to_cpu) {
+            : decode_only(decode_only), ae(decode_only, version == VERSION_WAN2_2_TI2V), VAE(version, backend, params_backend) {
             ae.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -1317,7 +1329,7 @@ namespace WAN {
             // ggml_backend_t backend = ggml_backend_cuda_init(0);
             ggml_backend_t backend            = ggml_backend_cpu_init();
             ggml_type model_data_type         = GGML_TYPE_F16;
-            std::shared_ptr<WanVAERunner> vae = std::make_shared<WanVAERunner>(backend, false, String2TensorStorage{}, "", false, VERSION_WAN2_2_TI2V);
+            std::shared_ptr<WanVAERunner> vae = std::make_shared<WanVAERunner>(backend, backend, String2TensorStorage{}, "", false, VERSION_WAN2_2_TI2V);
             {
                 LOG_INFO("loading from '%s'", file_path.c_str());
 
@@ -1984,6 +1996,13 @@ namespace WAN {
                 c = ggml_reshape_3d(ctx->ggml_ctx, c, c->ne[0] * c->ne[1] * c->ne[2], c->ne[3] / N, N);  // [N, dim, t_len*h_len*w_len]
                 c = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, c, 1, 0, 2, 3));  // [N, t_len*h_len*w_len, dim]
             }
+            sd::ggml_graph_cut::mark_graph_cut(x, "wan.prelude", "x");
+            // sd::ggml_graph_cut::mark_graph_cut(e, "wan.prelude", "e");
+            // sd::ggml_graph_cut::mark_graph_cut(e0, "wan.prelude", "e0");
+            // sd::ggml_graph_cut::mark_graph_cut(context, "wan.prelude", "context");
+            if (c != nullptr) {
+                sd::ggml_graph_cut::mark_graph_cut(c, "wan.prelude", "c");
+            }
 
             auto x_orig = x;
 
@@ -2004,6 +2023,10 @@ namespace WAN {
                     c_skip      = ggml_ext_scale(ctx->ggml_ctx, c_skip, vace_strength);
                     x           = ggml_add(ctx->ggml_ctx, x, c_skip);
                 }
+                sd::ggml_graph_cut::mark_graph_cut(x, "wan.blocks." + std::to_string(i), "x");
+                if (c != nullptr) {
+                    sd::ggml_graph_cut::mark_graph_cut(c, "wan.blocks." + std::to_string(i), "c");
+                }
             }
 
             x = head->forward(ctx, x, e);  // [N, t_len*h_len*w_len, pt*ph*pw*out_dim]
@@ -2071,11 +2094,11 @@ namespace WAN {
         SDVersion version;
 
         WanRunner(ggml_backend_t backend,
-                  bool offload_params_to_cpu,
+                  ggml_backend_t params_backend,
                   const String2TensorStorage& tensor_storage_map = {},
                   const std::string prefix                       = "",
                   SDVersion version                              = VERSION_WAN2)
-            : GGMLRunner(backend, offload_params_to_cpu) {
+            : GGMLRunner(backend, params_backend) {
             wan_params.num_layers = 0;
             for (auto pair : tensor_storage_map) {
                 std::string tensor_name = pair.first;
@@ -2323,7 +2346,7 @@ namespace WAN {
             }
 
             std::shared_ptr<WanRunner> wan = std::make_shared<WanRunner>(backend,
-                                                                         false,
+                                                                         backend,
                                                                          tensor_storage_map,
                                                                          "model.diffusion_model",
                                                                          VERSION_WAN2_2_TI2V);

src/z_image.hpp

@@ -31,10 +31,6 @@ namespace ZImage {
             : head_dim(head_dim), num_heads(num_heads), num_kv_heads(num_kv_heads), qk_norm(qk_norm) {
             blocks["qkv"] = std::make_shared<Linear>(hidden_size, (num_heads + num_kv_heads * 2) * head_dim, false);
             float scale   = 1.f;
-#if GGML_USE_HIP
-            // Prevent NaN issues with certain ROCm setups
-            scale = 1.f / 16.f;
-#endif
             blocks["out"] = std::make_shared<Linear>(num_heads * head_dim, hidden_size, false, false, false, scale);
             if (qk_norm) {
                 blocks["q_norm"] = std::make_shared<RMSNorm>(head_dim);
@@ -52,6 +48,10 @@ namespace ZImage {
             auto qkv_proj   = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
             auto out_proj   = std::dynamic_pointer_cast<Linear>(blocks["out"]);
 
+            if (sd_backend_is(ctx->backend, "ROCm")) {
+                out_proj->set_scale(1.f / 16.f);
+            }
+
             auto qkv = qkv_proj->forward(ctx, x);                                                                            // [N, n_token, (num_heads + num_kv_heads*2)*head_dim]
             qkv      = ggml_reshape_4d(ctx->ggml_ctx, qkv, head_dim, num_heads + num_kv_heads * 2, qkv->ne[1], qkv->ne[2]);  // [N, n_token, num_heads + num_kv_heads*2, head_dim]
 
@@ -115,9 +115,7 @@ namespace ZImage {
 
             bool force_prec_f32 = false;
             float scale         = 1.f / 128.f;
-#ifdef SD_USE_VULKAN
-            force_prec_f32 = true;
-#endif
+
             // The purpose of the scale here is to prevent NaN issues in certain situations.
             // For example, when using CUDA but the weights are k-quants.
             blocks["w2"] = std::make_shared<Linear>(hidden_dim, dim, false, false, force_prec_f32, scale);
@@ -129,6 +127,10 @@ namespace ZImage {
             auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
             auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
 
+            if (sd_backend_is(ctx->backend, "Vulkan")) {
+                w2->set_force_prec_f32(true);
+            }
+
             auto x1 = w1->forward(ctx, x);
             auto x3 = w3->forward(ctx, x);
             x       = ggml_swiglu_split(ctx->ggml_ctx, x1, x3);
@@ -369,6 +371,9 @@ namespace ZImage {
 
             auto txt = cap_embedder_1->forward(ctx, cap_embedder_0->forward(ctx, context));  // [N, n_txt_token, hidden_size]
             auto img = x_embedder->forward(ctx, x);                                          // [N, n_img_token, hidden_size]
+            sd::ggml_graph_cut::mark_graph_cut(txt, "z_image.prelude", "txt");
+            sd::ggml_graph_cut::mark_graph_cut(img, "z_image.prelude", "img");
+            sd::ggml_graph_cut::mark_graph_cut(t_emb, "z_image.prelude", "t_emb");
 
             int64_t n_txt_pad_token = Rope::bound_mod(static_cast<int>(n_txt_token), SEQ_MULTI_OF);
             if (n_txt_pad_token > 0) {
@@ -391,20 +396,24 @@ namespace ZImage {
                 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++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["noise_refiner." + std::to_string(i)]);
 
                 img = block->forward(ctx, img, img_pe, nullptr, t_emb);
+                sd::ggml_graph_cut::mark_graph_cut(img, "z_image.noise_refiner." + std::to_string(i), "img");
             }
 
             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++) {
                 auto block = std::dynamic_pointer_cast<JointTransformerBlock>(blocks["layers." + std::to_string(i)]);
 
                 txt_img = block->forward(ctx, txt_img, pe, nullptr, t_emb);
+                sd::ggml_graph_cut::mark_graph_cut(txt_img, "z_image.layers." + std::to_string(i), "txt_img");
             }
 
             txt_img = final_layer->forward(ctx, txt_img, t_emb);  // [N, n_txt_token + n_txt_pad_token + n_img_token + n_img_pad_token, ph*pw*C]
@@ -464,11 +473,11 @@ namespace ZImage {
         SDVersion version;
 
         ZImageRunner(ggml_backend_t backend,
-                     bool offload_params_to_cpu,
+                     ggml_backend_t params_backend,
                      const String2TensorStorage& tensor_storage_map = {},
                      const std::string prefix                       = "",
                      SDVersion version                              = VERSION_Z_IMAGE)
-            : GGMLRunner(backend, offload_params_to_cpu) {
+            : GGMLRunner(backend, params_backend) {
             z_image = ZImageModel(z_image_params);
             z_image.init(params_ctx, tensor_storage_map, prefix);
         }
@@ -611,7 +620,7 @@ namespace ZImage {
             }
 
             std::shared_ptr<ZImageRunner> z_image = std::make_shared<ZImageRunner>(backend,
-                                                                                   false,
+                                                                                   backend,
                                                                                    tensor_storage_map,
                                                                                    "model.diffusion_model",
                                                                                    VERSION_QWEN_IMAGE);