feat: display number of tokens for SD models (#1831 )

fix: make parameter loading backend-aware (#1828 )
fix: null pointer dereference when loading malformed LoHa file (#1826 )
2026-09-21 21:47:49 -05:00 · 2026-07-29 22:21:54 +08:00 · 2026-07-29 22:18:44 +08:00 · 2026-07-29 21:53:26 +08:00 · 2026-07-28 00:06:46 +08:00 · 2026-07-27 23:26:34 +08:00
14 changed files with 383 additions and 251 deletions
@@ -449,8 +449,8 @@ jobs:
    runs-on: windows-2022

    env:
-      ROCM_VERSION: "7.13.0"
-      GPU_TARGETS: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
+      ROCM_VERSION: "7.14.0"
+      GPU_TARGETS: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"

    steps:
      - uses: actions/checkout@v3
@@ -472,34 +472,68 @@ jobs:
        uses: actions/cache@v4
        with:
          path: C:\TheRock\build
-          key: rocm-${{ env.ROCM_VERSION }}-gfx1151-${{ runner.os }}
+          key: rocm-wheels-${{ env.ROCM_VERSION }}-${{ runner.os }}

      - name: ccache
        uses: ggml-org/ccache-action@v1.2.16
        with:
-          key: windows-latest-rocm-${{ env.ROCM_VERSION }}-x64
+          key: windows-rocm-${{ env.ROCM_VERSION }}-x64
          evict-old-files: 1d

-      - name: Install ROCm
+      - name: Install ROCm with Wheels
        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"
+          write-host "Setting up Python virtual environment"
+
+          # Create the venv directly at the cache location to avoid relocation issues
+          New-Item -Path "C:\TheRock\build" -ItemType Directory -Force | Out-Null
+          python -m venv C:\TheRock\build\.venv
+          & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+          write-host "Upgrading pip"
+          python -m pip install --upgrade pip
+
+          write-host "Installing ROCm wheels for multi-arch support"
+          # Install ROCm wheels for multi-arch support (this may take several minutes)
+          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{env.ROCM_VERSION}}"
+
+          # Pre-expand the devel tree so it is included in the cache
+          write-host "Initializing ROCm devel tree"
+          rocm-sdk init
+          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+          write-host "Completed ROCm wheel installation to C:\TheRock\build"

      - name: Setup ROCm Environment
        run: |
-          $rocmPath = "C:\TheRock\build"
+          $ErrorActionPreference = "Stop"
+
+          # Activate venv from cache or fresh install
+          & C:\TheRock\build\.venv\Scripts\Activate.ps1
+
+          # Expand the devel tree (idempotent; no-op if already done during install)
+          rocm-sdk init
+          if ($LASTEXITCODE -ne 0) { throw "rocm-sdk init failed with exit code $LASTEXITCODE" }
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          $rocmPath = (rocm-sdk path --root)
+          if (-not $rocmPath) { throw "rocm-sdk path --root returned empty - devel package may not be installed" }
+          $rocmPath = $rocmPath.Trim()
+          $cmakePath = (rocm-sdk path --cmake).Trim()
+          $binPath = (rocm-sdk path --bin).Trim()
+          write-host "ROCm root: $rocmPath"
+          write-host "CMake path: $cmakePath"
+          write-host "Bin path: $binPath"
+
          echo "HIP_PATH=$rocmPath" >> $env:GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$cmakePath" >> $env:GITHUB_ENV
          echo "HIP_DEVICE_LIB_PATH=$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" >> $env:GITHUB_PATH
+          echo "$binPath" >> $env:GITHUB_PATH
+
+          # Keep venv in PATH for subsequent steps
+          echo "C:\TheRock\build\.venv\Scripts" >> $env:GITHUB_PATH

      - name: Build
        run: |
@@ -527,139 +561,6 @@ jobs:
      - name: Pack artifacts
        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
        run: |
-          $ErrorActionPreference = "Stop"
-          $dst = "build\bin"
-          $rocmBin = Join-Path "${env:HIP_PATH}" "bin"
-          $requiredRocmPaths = @(
-            (Join-Path $rocmBin "rocblas.dll"),
-            (Join-Path $rocmBin "rocblas\library")
-          )
-          foreach ($path in $requiredRocmPaths) {
-            if (!(Test-Path $path)) {
-              throw "Missing ROCm runtime dependency: $path"
-            }
-          }
-
-          foreach ($pattern in @("rocblas*.dll", "hipblas*.dll", "libhipblas*.dll")) {
-            Copy-Item -Path (Join-Path $rocmBin $pattern) -Destination $dst -Force -ErrorAction SilentlyContinue
-          }
-
-          foreach ($dir in @("rocblas", "hipblaslt")) {
-            $src = Join-Path $rocmBin $dir
-            if (Test-Path $src) {
-              Copy-Item -Path $src -Destination $dst -Recurse -Force
-            }
-          }
-
-          7z a sd-${{ env.BRANCH_NAME }}-${{ steps.commit.outputs.short }}-bin-win-rocm-${{ env.ROCM_VERSION }}-x64.zip .\build\bin\*
-
-      - name: Upload artifacts
-        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-${{ env.ROCM_VERSION }}-x64.zip
-
-  windows-latest-cmake-hip:
-    runs-on: windows-2022
-
-    env:
-      HIPSDK_INSTALLER_VERSION: "26.Q1"
-      ROCM_VERSION: "7.1.1"
-      GPU_TARGETS: "gfx1150;gfx1151;gfx1200;gfx1201;gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032"
-
-    steps:
-      - uses: actions/checkout@v3
-        with:
-          submodules: recursive
-
-      - name: Setup Node
-        uses: actions/setup-node@v4
-        with:
-          node-version: 20
-
-      - name: Setup pnpm
-        uses: pnpm/action-setup@v4
-        with:
-          version: 10.15.1
-
-      - name: Cache ROCm Installation
-        id: cache-rocm
-        uses: actions/cache@v4
-        with:
-          path: C:\Program Files\AMD\ROCm
-          key: rocm-${{ env.HIPSDK_INSTALLER_VERSION }}-${{ runner.os }}
-
-      - name: ccache
-        uses: ggml-org/ccache-action@v1.2.16
-        with:
-          key: windows-latest-cmake-hip-${{ env.HIPSDK_INSTALLER_VERSION }}-x64
-          evict-old-files: 1d
-
-      - name: Install ROCm
-        if: steps.cache-rocm.outputs.cache-hit != 'true'
-        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 }}-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)
-          if (-not $completed) {
-              Write-Error "ROCm installation timed out after 10 minutes. Killing the process"
-              $proc.Kill()
-              exit 1
-          }
-          if ($proc.ExitCode -ne 0) {
-              Write-Error "ROCm installation failed with exit code $($proc.ExitCode)"
-              exit 1
-          }
-          write-host "Completed AMD HIP SDK installation"
-
-      - name: Verify ROCm
-        run: |
-          # Find and test ROCm installation
-          $clangPath = Get-ChildItem 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | Select-Object -First 1
-          if (-not $clangPath) {
-            Write-Error "ROCm installation not found"
-            exit 1
-          }
-          & $clangPath.FullName --version
-          # Set HIP_PATH environment variable for later steps
-          echo "HIP_PATH=$(Resolve-Path 'C:\Program Files\AMD\ROCm\*\bin\clang.exe' | split-path | split-path)" >> $env:GITHUB_ENV
-
-      - name: Build
-        run: |
-          mkdir build
-          cd build
-          $env:CMAKE_PREFIX_PATH="${env:HIP_PATH}"
-          cmake .. `
-            -G "Unix Makefiles" `
-            -DSD_HIPBLAS=ON `
-            -DSD_BUILD_SHARED_LIBS=ON `
-            -DGGML_NATIVE=OFF `
-            -DCMAKE_C_COMPILER=clang `
-            -DCMAKE_CXX_COMPILER=clang++ `
-            -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: prompt/actions-commit-hash@v2
-
-      - name: Pack artifacts
-        if: ${{ ( github.event_name == 'push' && github.ref == 'refs/heads/master' ) || github.event.inputs.create_release == 'true' }}
-        run: |
-          md "build\bin\rocblas\library\"
-          md "build\bin\hipblaslt\library"
-          cp "${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-${{ env.ROCM_VERSION }}-x64.zip .\build\bin\*

      - name: Upload artifacts
@@ -679,11 +580,8 @@ jobs:
    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: "7.13.0"
-            gpu_targets: "gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
+          - ROCM_VERSION: "7.14.0"
+            gpu_targets: "gfx900;gfx906;gfx908;gfx90a;gfx90c;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
            build: x64

    steps:
@@ -702,7 +600,7 @@ jobs:
      - name: Dependencies
        id: depends
        run: |
-          sudo apt install -y build-essential cmake wget zip ninja-build
+          sudo apt install -y build-essential git cmake wget

      - name: Free disk space
        run: |
@@ -723,38 +621,36 @@ jobs:
          sudo apt clean
          df -h

-      - 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/${{ matrix.ROCM_VERSION }} noble main
-          EOF
-
-          sudo tee /etc/apt/preferences.d/rocm-pin-600 << EOF
-          Package: *
-          Pin: release o=repo.radeon.com
-          Pin-Priority: 600
-          EOF
-
-          sudo apt update
-          sudo apt-get install -y libssl-dev rocm-hip-sdk
-
-      - name: Setup TheRock
-        if: matrix.ROCM_VERSION != '7.2.1'
+      - name: Setup TheRock with Wheels
        id: therock_env
        run: |
-          wget https://repo.amd.com/rocm/tarball/therock-dist-linux-gfx1151-${{ 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 >> $GITHUB_ENV
+          # Create Python virtual environment
+          python3 -m venv .venv
+          source .venv/bin/activate
+
+          # Install ROCm wheels for build
+          # libraries = HIP runtime and CMake configs needed for linking
+          # devel = compilers, headers, static libs
+          python -m pip install --upgrade pip
+          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{matrix.ROCM_VERSION}}"
+
+          # Get ROCm installation paths using the rocm-sdk CLI tool
+          ROCM_PATH=$(rocm-sdk path --root)
+          CMAKE_PATH=$(rocm-sdk path --cmake)
+          BIN_PATH=$(rocm-sdk path --bin)
+          echo "ROCM_PATH=$ROCM_PATH"
+          echo "CMAKE_PATH=$CMAKE_PATH"
+          echo "BIN_PATH=$BIN_PATH"
+
+          # Set environment variables
+          echo "ROCM_PATH=$ROCM_PATH" >> $GITHUB_ENV
+          echo "CMAKE_PREFIX_PATH=$CMAKE_PATH" >> $GITHUB_ENV
+          echo "HIP_PATH=$ROCM_PATH" >> $GITHUB_ENV
+          echo "PATH=$BIN_PATH:${PATH}" >> $GITHUB_ENV
+          echo "LD_LIBRARY_PATH=$ROCM_PATH/lib:${LD_LIBRARY_PATH:-}" >> $GITHUB_ENV
+
+          # Keep venv activated for subsequent steps
+          echo "$(pwd)/.venv/bin" >> $GITHUB_PATH

      # setup-node installs into /opt/hostedtoolcache, which is removed above.
      # Keep Node/pnpm setup after disk cleanup so the server frontend can be embedded.
@@ -839,7 +735,6 @@ jobs:
      - build-and-push-docker-images
      - macOS-latest-cmake
      - windows-latest-cmake
-      - windows-latest-cmake-hip
      - windows-latest-rocm

    steps:
@@ -168,6 +168,7 @@ These projects wrap `stable-diffusion.cpp` for easier use in other languages/fra

 These projects use `stable-diffusion.cpp` as a backend for their image generation.

+- [GIMP Plugins](https://github.com/themanyone/gimp-plugins)
 - [Jellybox](https://jellybox.com)
 - [Stable Diffusion GUI](https://github.com/fszontagh/sd.cpp.gui.wx)
 - [Stable Diffusion CLI-GUI](https://github.com/piallai/stable-diffusion.cpp)
@@ -1901,6 +1901,7 @@ bool SDGenerationParams::from_json_str(

    load_if_exists("strength", strength);
    load_if_exists("control_strength", control_strength);
+    load_if_exists("ip_adapter_strength", ip_adapter_strength);
    load_if_exists("moe_boundary", moe_boundary);
    load_if_exists("vace_strength", vace_strength);

@@ -2072,6 +2073,10 @@ bool SDGenerationParams::from_json_str(
        LOG_ERROR("invalid control_image");
        return false;
    }
+    if (!parse_image_json_field(j, "ip_adapter_image", 3, width, height, ip_adapter_image)) {
+        LOG_ERROR("invalid ip_adapter_image");
+        return false;
+    }

    return true;
 }
@@ -2807,6 +2812,7 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
    root["clip_skip"]             = gen_params.clip_skip;
    root["strength"]              = gen_params.strength;
    root["control_strength"]      = gen_params.control_strength;
+    root["ip_adapter_strength"]   = gen_params.ip_adapter_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) {
@@ -528,6 +528,7 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `auto_resize_ref_image` | `boolean` |
 | `increase_ref_index` | `boolean` |
 | `control_strength` | `number` |
+| `ip_adapter_strength` | `number` |
 | `hires` | `object` |
 | `hires.enabled` | `boolean` |
 | `hires.upscaler` | `string` |
@@ -567,6 +568,7 @@ Fields returned in `features_by_mode.img_gen`:
 - `init_image`
 - `mask_image`
 - `control_image`
+- `ip_adapter_image`
 - `ref_images`
 - `lora`
 - `vae_tiling`
@@ -653,12 +655,14 @@ Example:
  "auto_resize_ref_image": true,
  "increase_ref_index": false,
  "control_strength": 0.9,
+  "ip_adapter_strength": 1.0,
  "embed_image_metadata": true,

  "init_image": null,
  "ref_images": [],
  "mask_image": null,
  "control_image": null,
+  "ip_adapter_image": null,

  "sample_params": {
    "scheduler": "discrete",
@@ -733,6 +737,7 @@ Channel expectations:
 - `init_image`: 3 channels
 - `ref_images[]`: 3 channels
 - `control_image`: 3 channels
+- `ip_adapter_image`: 3 channels
 - `mask_image`: 1 channel

 If omitted or null:
@@ -757,6 +762,7 @@ Top-level scalar fields:
 | `auto_resize_ref_image` | `boolean` |
 | `increase_ref_index` | `boolean` |
 | `control_strength` | `number` |
+| `ip_adapter_strength` | `number` |
 | `embed_image_metadata` | `boolean` |

 Image fields:
@@ -767,6 +773,7 @@ Image fields:
 | `ref_images` | `array<string>` |
 | `mask_image` | `string \| null` |
 | `control_image` | `string \| null` |
+| `ip_adapter_image` | `string \| null` |

 LoRA fields:

@@ -958,7 +965,7 @@ Response fields:
 Compared with `img_gen`, the `vid_gen` request body:

 - `vid_gen` is a single video sequence job, so `batch_count` is not part of the request schema
- `ref_images`, `mask_image`, `control_image`, `control_strength`, and `embed_image_metadata` are not part of the request schema
+- `ref_images`, `mask_image`, `control_image`, `control_strength`, `ip_adapter_image`, `ip_adapter_strength`, and `embed_image_metadata` are not part of the request schema
 - `vid_gen` adds `end_image`, `control_frames`, `high_noise_sample_params`, `video_frames`, `fps`, `moe_boundary`, and `vace_strength`

 Example:
@@ -130,6 +130,7 @@ static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const
        {"auto_resize_ref_image", defaults.auto_resize_ref_image},
        {"increase_ref_index", defaults.increase_ref_index},
        {"control_strength", defaults.control_strength},
+        {"ip_adapter_strength", defaults.ip_adapter_strength},
        {"sample_params", make_sample_params_json(defaults.sample_params, defaults.skip_layers)},
        {"hires", make_hires_json(defaults)},
        {"vae_tiling_params", make_vae_tiling_json(defaults.vae_tiling_params)},
@@ -173,6 +174,7 @@ static json make_img_gen_features_json() {
        {"init_image", true},
        {"mask_image", true},
        {"control_image", true},
+        {"ip_adapter_image", true},
        {"ref_images", true},
        {"lora", true},
        {"vae_tiling", true},
@@ -117,6 +117,7 @@ public:
    virtual SDCondition get_learned_condition(int n_threads,
                                              const ConditionerParams& conditioner_params) = 0;
    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors)           = 0;
+    virtual void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& tensor_ops) {}
    virtual void set_max_graph_vram_bytes(size_t max_vram_bytes) {}
    virtual void set_stream_layers_enabled(bool enabled) {}
    virtual void set_runtime_backends(const std::vector<ggml_backend_t>& backends) {}
@@ -1664,6 +1665,10 @@ struct AnimaConditioner : public Conditioner {
        llm->get_param_tensors(tensors, "text_encoders.llm");
    }

+    void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& tensor_ops) override {
+        llm->get_param_tensor_ops(tensor_ops);
+    }
+
    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
        llm->set_max_graph_vram_bytes(max_vram_bytes);
    }
@@ -1847,6 +1852,10 @@ struct LLMEmbedder : public Conditioner {
        }
    }

+    void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& tensor_ops) override {
+        llm->get_param_tensor_ops(tensor_ops);
+    }
+
    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
        llm->set_max_graph_vram_bytes(max_vram_bytes);
        if (byt5) {
@@ -2828,6 +2837,10 @@ struct LTXAVEmbedder : public Conditioner {
        projector->get_param_tensors(tensors, "text_embedding_projection");
    }

+    void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& tensor_ops) override {
+        llm->get_param_tensor_ops(tensor_ops);
+    }
+
    void set_flash_attention_enabled(bool enabled) override {
        llm->set_flash_attention_enabled(enabled);
        projector->set_flash_attention_enabled(enabled);
@@ -1753,7 +1753,7 @@ protected:
    std::vector<size_t> graph_cut_layer_split_backend_vram_limits_;

    std::vector<ggml_backend_t> extra_runtime_backends;  // borrowed (SDBackendManager-owned)
-    ggml_backend_sched_t sched             = nullptr;    // owned, multi-device only
+    ggml_backend_sched_t sched             = nullptr;    // owned
    ggml_backend_t cpu_fallback_backend    = nullptr;    // owned, sched requires a trailing CPU backend
    bool multi_device_eval_callback_warned = false;

@@ -2147,8 +2147,22 @@ protected:
        return !extra_runtime_backends.empty();
    }

+    bool graph_requires_backend_fallback(ggml_cgraph* gf) const {
+        if (gf == nullptr || sd_backend_is_cpu(runtime_backend)) {
+            return false;
+        }
+        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 && !ggml_backend_supports_op(runtime_backend, node)) {
+                return true;
+            }
+        }
+        return false;
+    }
+
    bool alloc_compute_buffer(ggml_cgraph* gf) {
-        if (is_multi_device()) {
+        if (sched != nullptr || is_multi_device() || graph_requires_backend_fallback(gf)) {
            // The sched replaces the gallocr. Do NOT ggml_backend_sched_reserve
            // the graph here: reserve runs split_graph, which rewires the
            // graph's src pointers to sched-internal copy tensors, and the
@@ -2156,6 +2170,10 @@ protected:
            // rewired graph, silently corrupting every cross-backend input. A
            // graph must be split at most once; the alloc in execute_graph
            // performs the real allocation.
+            if (compute_allocr != nullptr) {
+                ggml_gallocr_free(compute_allocr);
+                compute_allocr = nullptr;
+            }
            return ensure_sched(gf);
        }
        if (compute_allocr != nullptr) {
@@ -2753,7 +2771,7 @@ protected:
        };
        ComputeBufferGuard compute_buffer_guard(this, free_compute_buffer);

-        if (is_multi_device()) {
+        if (sched != nullptr) {
            ggml_backend_sched_reset(sched);
            pin_multi_device_nodes(gf);  // reset clears the pins; re-apply before alloc
            if (!ggml_backend_sched_alloc_graph(sched, gf)) {
@@ -2774,9 +2792,9 @@ protected:
        }

        ggml_status status;
-        if (is_multi_device()) {
+        if (sched != nullptr) {
            if (sd_get_backend_eval_callback() != nullptr && !multi_device_eval_callback_warned) {
-                LOG_WARN("%s: eval callback is not supported with multiple runtime backends; ignoring",
+                LOG_WARN("%s: eval callback is not supported with the backend scheduler; ignoring",
                         get_desc().c_str());
                multi_device_eval_callback_warned = true;
            }
@@ -3018,12 +3036,9 @@ public:

    // do copy after alloc graph
    void set_backend_tensor_data(ggml_tensor* tensor, const void* data) {
-        if (is_multi_device()) {
-            // The sched only assigns a backend (and thus a buffer) to tensors
-            // that participate in the graph; flag standalone data tensors as
-            // inputs so they get one.
-            ggml_set_input(tensor);
-        }
+        // The scheduler only allocates standalone data tensors when they are
+        // marked as graph inputs. The flag is harmless for single-backend graphs.
+        ggml_set_input(tensor);
        backend_tensor_data_map[tensor] = data;
    }

@@ -3240,6 +3255,11 @@ protected:

    virtual void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {}

+    virtual enum ggml_op param_usage_op(const std::string& name) const {
+        (void)name;
+        return GGML_OP_NONE;
+    }
+
 public:
    void init(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") {
        if (prefix.size() > 0) {
@@ -3290,6 +3310,18 @@ public:
        }
    }

+    void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& tensor_ops) {
+        for (auto& pair : blocks) {
+            pair.second->get_param_tensor_ops(tensor_ops);
+        }
+        for (auto& pair : params) {
+            enum ggml_op op = param_usage_op(pair.first);
+            if (op != GGML_OP_NONE) {
+                tensor_ops[pair.second] = op;
+            }
+        }
+    }
+
    virtual std::string get_desc() {
        return "GGMLBlock";
    }
@@ -3417,6 +3449,10 @@ protected:
        params["weight"] = ggml_new_tensor_2d(ctx, wtype, embedding_dim, num_embeddings);
    }

+    enum ggml_op param_usage_op(const std::string& name) const override {
+        return name == "weight" ? GGML_OP_GET_ROWS : GGML_OP_NONE;
+    }
+
 public:
    Embedding(int64_t num_embeddings, int64_t embedding_dim)
        : embedding_dim(embedding_dim),
@@ -14,6 +14,8 @@ struct LoraModel : public GGMLRunner {
    std::unordered_map<std::string, ggml_tensor*> lora_tensors;
    std::map<ggml_tensor*, ggml_tensor*> original_tensor_to_final_tensor;
    std::set<std::string> applied_lora_tensors;
+    std::set<std::string> skipped_incompatible_lora_tensors;
+    std::set<std::string> warned_incompatible_model_tensors;
    std::string file_path;
    std::shared_ptr<ModelManager> model_manager;
    ggml_backend_t params_backend = nullptr;
@@ -133,6 +135,8 @@ struct LoraModel : public GGMLRunner {
        lora_tensors.clear();
        original_tensor_to_final_tensor.clear();
        applied_lora_tensors.clear();
+        skipped_incompatible_lora_tensors.clear();
+        warned_incompatible_model_tensors.clear();
        applied             = false;
        tensor_preprocessed = false;
    }
@@ -338,7 +342,9 @@ struct LoraModel : public GGMLRunner {
            iter = lora_tensors.find(hada_1_mid_name);
            if (iter != lora_tensors.end()) {
                hada_1_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
-                hada_1_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_1_up));
+                if (hada_1_up != nullptr) {
+                    hada_1_up = ggml_cont(ctx, ggml_transpose(ctx, hada_1_up));
+                }
            }

            iter = lora_tensors.find(hada_2_down_name);
@@ -354,7 +360,9 @@ struct LoraModel : public GGMLRunner {
            iter = lora_tensors.find(hada_2_mid_name);
            if (iter != lora_tensors.end()) {
                hada_2_mid = ggml_ext_cast_f32(ctx, backend, iter->second);
-                hada_2_up  = ggml_cont(ctx, ggml_transpose(ctx, hada_2_up));
+                if (hada_2_up != nullptr) {
+                    hada_2_up = ggml_cont(ctx, ggml_transpose(ctx, hada_2_up));
+                }
            }

            if (hada_1_up == nullptr || hada_1_down == nullptr || hada_2_up == nullptr || hada_2_down == nullptr) {
@@ -546,7 +554,27 @@ struct LoraModel : public GGMLRunner {
                }
            }

-            GGML_ASSERT(ggml_nelements(diff) == ggml_nelements(model_tensor));
+            if (ggml_nelements(diff) != ggml_nelements(model_tensor)) {
+                const std::string lora_tensor_prefix = "lora." + model_tensor_name + ".";
+                for (const auto& tensor_name : applied_lora_tensors) {
+                    if (starts_with(tensor_name, lora_tensor_prefix)) {
+                        skipped_incompatible_lora_tensors.insert(tensor_name);
+                    }
+                }
+                if (warned_incompatible_model_tensors.insert(model_tensor_name).second) {
+                    LOG_WARN("skip incompatible LoRA tensor |%s|: model shape = [%lld, %lld, %lld, %lld], LoRA shape = [%lld, %lld, %lld, %lld]",
+                             model_tensor_name.c_str(),
+                             static_cast<long long>(model_tensor->ne[0]),
+                             static_cast<long long>(model_tensor->ne[1]),
+                             static_cast<long long>(model_tensor->ne[2]),
+                             static_cast<long long>(model_tensor->ne[3]),
+                             static_cast<long long>(diff->ne[0]),
+                             static_cast<long long>(diff->ne[1]),
+                             static_cast<long long>(diff->ne[2]),
+                             static_cast<long long>(diff->ne[3]));
+                }
+                return nullptr;
+            }
            diff = ggml_reshape(ctx, diff, model_tensor);
        }
        return diff;
@@ -555,6 +583,7 @@ struct LoraModel : public GGMLRunner {
    ggml_tensor* get_out_diff(ggml_context* ctx,
                              ggml_backend_t backend,
                              ggml_tensor* x,
+                              ggml_tensor* model_weight,
                              WeightAdapter::ForwardParams forward_params,
                              const std::string& model_tensor_name) {
        ggml_tensor* out_diff = nullptr;
@@ -707,6 +736,43 @@ struct LoraModel : public GGMLRunner {
                break;
            }

+            if (!is_conv2d) {
+                const int64_t down_in  = lora_down->ne[0];
+                const int64_t down_out = lora_down->ne[1];
+                const int64_t up_in    = lora_up->ne[0];
+                const int64_t up_out   = lora_up->ne[1];
+
+                bool compatible = down_in == model_weight->ne[0] &&
+                                  up_out == model_weight->ne[1];
+                if (lora_mid != nullptr) {
+                    compatible = compatible &&
+                                 lora_mid->ne[0] == down_out &&
+                                 up_in == lora_mid->ne[1];
+                } else {
+                    compatible = compatible && up_in == down_out;
+                }
+
+                if (!compatible) {
+                    skipped_incompatible_lora_tensors.insert(lora_down_name);
+                    skipped_incompatible_lora_tensors.insert(lora_up_name);
+                    skipped_incompatible_lora_tensors.insert(lora_mid_name);
+                    skipped_incompatible_lora_tensors.insert(scale_name);
+                    skipped_incompatible_lora_tensors.insert(alpha_name);
+                    if (warned_incompatible_model_tensors.insert(model_tensor_name).second) {
+                        LOG_WARN("skip incompatible LoRA tensor |%s|: model shape = [%lld, %lld], down shape = [%lld, %lld], up shape = [%lld, %lld]",
+                                 model_tensor_name.c_str(),
+                                 static_cast<long long>(model_weight->ne[0]),
+                                 static_cast<long long>(model_weight->ne[1]),
+                                 static_cast<long long>(down_in),
+                                 static_cast<long long>(down_out),
+                                 static_cast<long long>(up_in),
+                                 static_cast<long long>(up_out));
+                    }
+                    index++;
+                    continue;
+                }
+            }
+
            applied_lora_tensors.insert(lora_up_name);
            applied_lora_tensors.insert(lora_down_name);

@@ -869,10 +935,13 @@ struct LoraModel : public GGMLRunner {
    void stat(bool at_runntime = false) {
        size_t total_lora_tensors_count   = 0;
        size_t applied_lora_tensors_count = 0;
+        size_t skipped_lora_tensors_count = 0;

        for (auto& kv : lora_tensors) {
            total_lora_tensors_count++;
-            if (applied_lora_tensors.find(kv.first) == applied_lora_tensors.end()) {
+            if (skipped_incompatible_lora_tensors.find(kv.first) != skipped_incompatible_lora_tensors.end()) {
+                skipped_lora_tensors_count++;
+            } else if (applied_lora_tensors.find(kv.first) == applied_lora_tensors.end()) {
                if (!at_runntime) {
                    LOG_WARN("unused lora tensor |%s|", kv.first.c_str());
                    print_ggml_tensor(kv.second, true);
@@ -884,12 +953,17 @@ struct LoraModel : public GGMLRunner {
        /* Don't worry if this message shows up twice in the logs per LoRA,
         * this function is called once to calculate the required buffer size
         * and then again to actually generate a graph to be used */
-        if (!at_runntime && applied_lora_tensors_count != total_lora_tensors_count) {
+        size_t compatible_lora_tensors_count = total_lora_tensors_count - skipped_lora_tensors_count;
+        if (!at_runntime && applied_lora_tensors_count != compatible_lora_tensors_count) {
            LOG_WARN("Only (%lu / %lu) LoRA tensors have been applied, lora_file_path = %s",
-                     applied_lora_tensors_count, total_lora_tensors_count, file_path.c_str());
+                     applied_lora_tensors_count, compatible_lora_tensors_count, file_path.c_str());
        } else {
            LOG_INFO("(%lu / %lu) LoRA tensors have been applied, lora_file_path = %s",
-                     applied_lora_tensors_count, total_lora_tensors_count, file_path.c_str());
+                     applied_lora_tensors_count, compatible_lora_tensors_count, file_path.c_str());
+        }
+        if (skipped_lora_tensors_count > 0) {
+            LOG_WARN("(%lu / %lu) incompatible LoRA tensors have been skipped, lora_file_path = %s",
+                     skipped_lora_tensors_count, total_lora_tensors_count, file_path.c_str());
        }
    }
 };
@@ -953,7 +1027,7 @@ public:
                                   forward_params.conv2d.scale);
        }
        for (auto& lora_model : lora_models) {
-            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, backend, x, forward_params, prefix + "weight");
+            ggml_tensor* out_diff = lora_model->get_out_diff(ctx, backend, x, w, forward_params, prefix + "weight");
            if (out_diff == nullptr) {
                continue;
            }
@@ -1657,6 +1657,10 @@ namespace LLM {
            model.get_param_tensors(tensors, prefix);
        }

+        void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& tensor_ops) {
+            model.get_param_tensor_ops(tensor_ops);
+        }
+
        ggml_tensor* forward(GGMLRunnerContext* ctx,
                             ggml_tensor* input_ids,
                             ggml_tensor* input_pos,
@@ -1053,7 +1053,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
        std::atomic<size_t> tensor_idx(0);
        std::atomic<bool> failed(false);
        std::vector<std::thread> workers;
-        std::mutex rpc_backend_mutex;
+        std::mutex backend_tensor_set_mutex;

        for (int i = 0; i < n_threads; ++i) {
            workers.emplace_back([&, file_path, is_zip]() {
@@ -1214,17 +1214,8 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb,
                    if (dst_tensor->buffer != nullptr && !ggml_backend_buffer_is_host(dst_tensor->buffer)) {
                        t0 = ggml_time_ms();

-                        // RPC backends require serialized access to prevent concurrency issues
-                        const char* buffer_type_name = ggml_backend_buft_name(ggml_backend_buffer_get_type(dst_tensor->buffer));
-                        bool is_rpc_buffer           = buffer_type_name != nullptr &&
-                                             std::string(buffer_type_name).find("RPC") != std::string::npos;
-
-                        if (is_rpc_buffer) {
-                            std::lock_guard<std::mutex> lock(rpc_backend_mutex);
-                            ggml_backend_tensor_set(dst_tensor, convert_buf, 0, ggml_nbytes(dst_tensor));
-                        } else {
-                            ggml_backend_tensor_set(dst_tensor, convert_buf, 0, ggml_nbytes(dst_tensor));
-                        }
+                        std::lock_guard<std::mutex> lock(backend_tensor_set_mutex);
+                        ggml_backend_tensor_set(dst_tensor, convert_buf, 0, ggml_nbytes(dst_tensor));

                        t1 = ggml_time_ms();
                        copy_to_backend_time_ms.fetch_add(t1 - t0);
@@ -53,6 +53,48 @@ static bool backend_supports_host_buffer(ggml_backend_t backend) {
    return props.caps.buffer_from_host_ptr;
 }

+static bool device_supports_param_op(ggml_backend_dev_t device,
+                                      ggml_tensor* weight,
+                                      enum ggml_op op,
+                                      ggml_backend_buffer_type_t buft) {
+    if (op == GGML_OP_NONE) {
+        return true;
+    }
+    if (device == nullptr || weight == nullptr || buft == nullptr || weight->buffer != nullptr) {
+        return false;
+    }
+
+    ggml_init_params params;
+    params.mem_size   = ggml_tensor_overhead() * 2;
+    params.mem_buffer = nullptr;
+    params.no_alloc   = true;
+    ggml_context* ctx = ggml_init(params);
+    if (ctx == nullptr) {
+        return false;
+    }
+
+    ggml_tensor* op_tensor = nullptr;
+    if (op == GGML_OP_GET_ROWS) {
+        ggml_tensor* indices = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, 1);
+        op_tensor            = ggml_get_rows(ctx, weight, indices);
+    }
+    if (op_tensor == nullptr) {
+        ggml_free(ctx);
+        return false;
+    }
+
+    weight->buffer = ggml_backend_buft_alloc_buffer(buft, 0);
+    if (weight->buffer == nullptr) {
+        ggml_free(ctx);
+        return false;
+    }
+    bool supported = ggml_backend_dev_supports_op(device, op_tensor);
+    ggml_backend_buffer_free(weight->buffer);
+    weight->buffer = nullptr;
+    ggml_free(ctx);
+    return supported;
+}
+
 ModelManager::~ModelManager() {
    release_all();
 }
@@ -135,7 +177,8 @@ bool ModelManager::register_param_tensors(const std::string& desc,
                                          ggml_backend_t params_backend,
                                          size_t* registered_tensor_size,
                                          bool allow_split_buffer,
-                                          bool params_follow_compute_backend) {
+                                          bool params_follow_compute_backend,
+                                          const std::map<ggml_tensor*, enum ggml_op>* tensor_ops) {
    if (desc.empty()) {
        LOG_ERROR("model manager tensor desc is empty");
        return false;
@@ -168,6 +211,12 @@ bool ModelManager::register_param_tensors(const std::string& desc,
        state->params_backend                = params_backend;
        state->allow_split_buffer            = allow_split_buffer;
        state->params_follow_compute_backend = params_follow_compute_backend;
+        if (tensor_ops != nullptr) {
+            auto op_it = tensor_ops->find(tensor);
+            if (op_it != tensor_ops->end()) {
+                state->usage_op = op_it->second;
+            }
+        }
        new_states.push_back(std::move(state));
    }

@@ -844,6 +893,22 @@ ggml_backend_buffer_type_t ModelManager::params_buffer_type_for(const TensorStat
    if (params_buft == nullptr) {
        params_buft = ggml_backend_get_default_buffer_type(state.params_backend);
    }
+    if (state.usage_op != GGML_OP_NONE &&
+        state.compute_backend != nullptr) {
+        ggml_backend_dev_t compute_dev = ggml_backend_get_device(state.compute_backend);
+        if (device_supports_param_op(compute_dev, state.tensor, state.usage_op, params_buft)) {
+            return params_buft;
+        }
+
+        ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
+        params_buft                = cpu_dev != nullptr ? ggml_backend_dev_buffer_type(cpu_dev) : nullptr;
+        if (!device_supports_param_op(cpu_dev, state.tensor, state.usage_op, params_buft)) {
+            LOG_ERROR("model manager has no compatible buffer for tensor '%s' used by %s",
+                      state.name.c_str(),
+                      ggml_op_name(state.usage_op));
+            return nullptr;
+        }
+    }
    return params_buft;
 }

@@ -39,6 +39,7 @@ private:
        bool allow_split_buffer            = false;
        bool params_follow_compute_backend = false;
        bool metadata_validated            = false;
+        enum ggml_op usage_op              = GGML_OP_NONE;

        int active_prepare_count = 0;

@@ -132,7 +133,8 @@ public:
                                ggml_backend_t params_backend,
                                size_t* registered_tensor_size     = nullptr,
                                bool allow_split_buffer            = false,
-                                bool params_follow_compute_backend = false);
+                                bool params_follow_compute_backend = false,
+                                const std::map<ggml_tensor*, enum ggml_op>* tensor_ops = nullptr);

    bool unregister_param_tensors(const std::string& desc,
                                  size_t* registered_tensor_size = nullptr);
@@ -141,6 +141,8 @@ const char* sampling_methods_str[] = {
    "Euler CFG++",
    "Euler A CFG++",
    "Euler GE",
+    "DPM++ (2M) SDE",
+    "DPM++ (2M) SDE BT",
 };

 /*================================================== Helper Functions ================================================*/
@@ -224,6 +226,7 @@ public:
    std::shared_ptr<ControlNet> control_net;
    std::shared_ptr<IPAdapter::IPAdapterRunner> ip_adapter;
    sd::Tensor<float> ip_adapter_tokens;
+    sd::Tensor<float> ip_adapter_uncond_tokens;
    float ip_adapter_strength = 1.0f;
    std::vector<std::shared_ptr<GenerationExtension>> generation_extensions;
    std::vector<std::shared_ptr<LoraModel>> runtime_lora_models;
@@ -317,7 +320,11 @@ public:
            return true;
        }
        std::map<std::string, ggml_tensor*> group_tensors;
+        std::map<ggml_tensor*, enum ggml_op> tensor_ops;
        model->get_param_tensors(group_tensors);
+        if constexpr (std::is_base_of_v<Conditioner, T>) {
+            model->get_param_tensor_ops(tensor_ops);
+        }
        if (model_manager == nullptr) {
            return true;
        }
@@ -334,6 +341,7 @@ public:
                                                                module,
                                                                module_backends,
                                                                std::move(group_tensors),
+                                                                tensor_ops,
                                                                residency_mode,
                                                                params_mem_size);
                    }
@@ -342,6 +350,7 @@ public:
                                                              module,
                                                              module_backends,
                                                              std::move(group_tensors),
+                                                              tensor_ops,
                                                              residency_mode,
                                                              params_mem_size);
                }
@@ -355,7 +364,10 @@ public:
                                                     residency_mode,
                                                     backend_for(module),
                                                     params_backend_for(module),
-                                                     params_mem_size);
+                                                     params_mem_size,
+                                                     false,
+                                                     false,
+                                                     &tensor_ops);
    }

    template <typename T>
@@ -364,6 +376,7 @@ public:
                                          SDBackendModule module,
                                          const std::vector<ggml_backend_t>& module_backends,
                                          std::map<std::string, ggml_tensor*> group_tensors,
+                                          const std::map<ggml_tensor*, enum ggml_op>& tensor_ops,
                                          ModelManager::ResidencyMode residency_mode,
                                          size_t* params_mem_size) {
        ggml_backend_t main_backend = module_backends[0];
@@ -375,6 +388,7 @@ public:
                                                      module,
                                                      module_backends,
                                                      std::move(group_tensors),
+                                                      tensor_ops,
                                                      residency_mode,
                                                      params_mem_size);
        };
@@ -449,7 +463,9 @@ public:
                                                   main_backend,
                                                   params_backend_for(module),
                                                   params_mem_size,
-                                                   /*allow_split_buffer=*/true)) {
+                                                   /*allow_split_buffer=*/true,
+                                                   false,
+                                                   &tensor_ops)) {
            return false;
        }
        return model_manager->register_param_tensors(desc,
@@ -457,7 +473,10 @@ public:
                                                     residency_mode,
                                                     main_backend,
                                                     params_backend_for(module),
-                                                     params_mem_size);
+                                                     params_mem_size,
+                                                     false,
+                                                     false,
+                                                     &tensor_ops);
    }

    // Register graph-cut layer-split tensors on the primary backend first.
@@ -469,6 +488,7 @@ public:
                                            SDBackendModule module,
                                            const std::vector<ggml_backend_t>& module_backends,
                                            std::map<std::string, ggml_tensor*> group_tensors,
+                                            const std::map<ggml_tensor*, enum ggml_op>& tensor_ops,
                                            ModelManager::ResidencyMode residency_mode,
                                            size_t* params_mem_size) {
        bool has_cpu_device = false;
@@ -490,7 +510,10 @@ public:
                                                         residency_mode,
                                                         module_backends[0],
                                                         params_backend_for(module),
-                                                         params_mem_size);
+                                                         params_mem_size,
+                                                         false,
+                                                         false,
+                                                         &tensor_ops);
        }

        model->set_runtime_backends(module_backends);
@@ -515,7 +538,8 @@ public:
                                                     initial_params_backend,
                                                     params_mem_size,
                                                     false,
-                                                     params_follow_runtime);
+                                                     params_follow_runtime,
+                                                     &tensor_ops);
    }

    bool unload_control_net() {
@@ -2100,8 +2124,9 @@ public:
    }

    void compute_ip_adapter_tokens(const sd_image_t& image, float strength) {
-        ip_adapter_tokens   = {};
-        ip_adapter_strength = strength;
+        ip_adapter_tokens        = {};
+        ip_adapter_uncond_tokens = {};
+        ip_adapter_strength      = strength;
        if (ip_adapter == nullptr || clip_vision == nullptr || image.data == nullptr) {
            return;
        }
@@ -2111,10 +2136,19 @@ public:
            return;
        }
        ip_adapter_tokens = ip_adapter->compute(n_threads, embed);
-        if (!ip_adapter_tokens.empty()) {
-            LOG_INFO("IP-Adapter: %lld image tokens, strength %.2f",
-                     (long long)ip_adapter_tokens.shape()[1], strength);
+        if (ip_adapter_tokens.empty()) {
+            LOG_ERROR("IP-Adapter conditional image projection failed");
+            return;
        }
+        auto uncond_embed        = sd::Tensor<float>::zeros_like(embed);
+        ip_adapter_uncond_tokens = ip_adapter->compute(n_threads, uncond_embed);
+        if (ip_adapter_uncond_tokens.empty()) {
+            LOG_ERROR("IP-Adapter unconditional image projection failed");
+            ip_adapter_tokens = {};
+            return;
+        }
+        LOG_INFO("IP-Adapter: %lld image tokens, strength %.2f",
+                 (long long)ip_adapter_tokens.shape()[1], strength);
    }

    std::vector<float> process_timesteps(const std::vector<float>& timesteps,
@@ -2606,7 +2640,7 @@ public:
                                     const sd::Tensor<float>* c_concat_override                 = nullptr,
                                     const std::vector<int>* local_skip_layers                  = nullptr,
                                     const std::vector<sd::Tensor<float>>* ref_latents_override = nullptr,
-                                     bool apply_ip                                              = true) -> sd::Tensor<float> {
+                                     bool use_uncond_ip                                         = false) -> sd::Tensor<float> {
                diffusion_params.context     = condition.c_crossattn.empty() ? nullptr : &condition.c_crossattn;
                diffusion_params.c_concat    = c_concat_override != nullptr ? c_concat_override : (condition.c_concat.empty() ? nullptr : &condition.c_concat);
                diffusion_params.y           = condition.c_vector.empty() ? nullptr : &condition.c_vector;
@@ -2618,8 +2652,9 @@ public:
                        nvf = static_cast<int>(noised_input.shape()[3]);
                    }
                    UNetDiffusionExtra unet_extra{nvf, &controls, control_strength};
-                    if (apply_ip && !ip_adapter_tokens.empty()) {
-                        unet_extra.ip_context = &ip_adapter_tokens;
+                    const auto& ip_tokens = use_uncond_ip ? ip_adapter_uncond_tokens : ip_adapter_tokens;
+                    if (!ip_tokens.empty()) {
+                        unet_extra.ip_context = &ip_tokens;
                        unet_extra.ip_scale   = ip_adapter_strength;
                    }
                    diffusion_params.extra = unet_extra;
@@ -2715,7 +2750,7 @@ public:
                                           uncond.c_concat.empty() ? nullptr : &uncond.c_concat,
                                           uncond_skip_layers,
                                           nullptr,
-                                           false);
+                                           true);
                if (uncond_out.empty()) {
                    return {};
                }
@@ -2725,7 +2760,7 @@ public:
                                               img_uncond.c_concat.empty() ? nullptr : &img_uncond.c_concat,
                                               nullptr,
                                               uncond_without_ref_latents ? &empty_ref_latents : nullptr,
-                                               false);
+                                               true);
                if (img_uncond_out.empty()) {
                    return {};
                }
@@ -3573,21 +3608,22 @@ char* sd_sample_params_to_str(const sd_sample_params_t* sample_params) {
 void sd_img_gen_params_init(sd_img_gen_params_t* sd_img_gen_params) {
    *sd_img_gen_params = {};
    sd_sample_params_init(&sd_img_gen_params->sample_params);
-    sd_img_gen_params->clip_skip         = -1;
-    sd_img_gen_params->ref_images_count  = 0;
-    sd_img_gen_params->ref_image_args    = "";
-    sd_img_gen_params->width             = 512;
-    sd_img_gen_params->height            = 512;
-    sd_img_gen_params->strength          = 0.75f;
-    sd_img_gen_params->seed              = -1;
-    sd_img_gen_params->batch_count       = 1;
-    sd_img_gen_params->control_strength  = 0.9f;
-    sd_img_gen_params->qwen_image_layers = 3;
-    sd_img_gen_params->circular_x        = false;
-    sd_img_gen_params->circular_y        = false;
-    sd_img_gen_params->pm_params         = {nullptr, 0, nullptr, 20.f};
-    sd_img_gen_params->pulid_params      = {nullptr, 1.0f};
-    sd_img_gen_params->vae_tiling_params = {false, false, 0, 0, 0.5f, 0.0f, 0.0f, nullptr};
+    sd_img_gen_params->clip_skip           = -1;
+    sd_img_gen_params->ref_images_count    = 0;
+    sd_img_gen_params->ref_image_args      = "";
+    sd_img_gen_params->width               = 512;
+    sd_img_gen_params->height              = 512;
+    sd_img_gen_params->strength            = 0.75f;
+    sd_img_gen_params->seed                = -1;
+    sd_img_gen_params->batch_count         = 1;
+    sd_img_gen_params->control_strength    = 0.9f;
+    sd_img_gen_params->ip_adapter_strength = 1.0f;
+    sd_img_gen_params->qwen_image_layers   = 3;
+    sd_img_gen_params->circular_x          = false;
+    sd_img_gen_params->circular_y          = false;
+    sd_img_gen_params->pm_params           = {nullptr, 0, nullptr, 20.f};
+    sd_img_gen_params->pulid_params        = {nullptr, 1.0f};
+    sd_img_gen_params->vae_tiling_params   = {false, false, 0, 0, 0.5f, 0.0f, 0.0f, nullptr};
    sd_cache_params_init(&sd_img_gen_params->cache);
    sd_hires_params_init(&sd_img_gen_params->hires);
 }
@@ -5071,8 +5107,8 @@ static std::optional<ImageGenerationEmbeds> prepare_image_generation_embeds(sd_c
            }
            condition_params.text            = request->negative_prompt;
            condition_params.zero_out_masked = zero_out_masked;
+            std::vector<sd::Tensor<float>> empty_ref_images;
            if (use_ref_latent_img_cfg) {
-                std::vector<sd::Tensor<float>> empty_ref_images;
                condition_params.ref_images = &empty_ref_images;
            }
            img_uncond = sd_ctx->sd->cond_stage_model->get_learned_condition(sd_ctx->sd->n_threads,
@@ -5776,7 +5812,7 @@ static std::optional<ImageGenerationLatents> prepare_video_generation_latents(sd
        auto encode_condition_frame = [&](const sd::Tensor<float>& image,
                                          int64_t latent_frame,
                                          const char* name) -> bool {
-            auto encoded = sd_ctx->sd->encode_first_stage(image);
+            auto encoded = sd_ctx->sd->encode_first_stage(image.unsqueeze(2));
            if (encoded.empty()) {
                LOG_ERROR("failed to encode Hunyuan Video %s conditioning frame", name);
                return false;
@@ -205,7 +205,7 @@ std::vector<int> BPETokenizer::encode(const std::string& text, on_new_token_cb_t
        ss << "\"" << token << "\", ";
    }
    ss << "]";
-    LOG_DEBUG("split prompt \"%s\" to tokens %s", text.c_str(), ss.str().c_str());
+    LOG_DEBUG("split prompt \"%s\" to %zu tokens %s", text.c_str(), bpe_tokens.size(), ss.str().c_str());
    return bpe_tokens;
 }
Author	SHA1	Message	Date
vmobilis	9cfe2af8f9	feat: display number of tokens for SD models (#1831 )	2026-07-29 22:21:54 +08:00
yzyyzyhhh	2993b7fb43	fix: make parameter loading backend-aware (#1828 )	2026-07-29 22:18:44 +08:00
Nicky Mouha	53856e7ec8	fix: null pointer dereference when loading malformed LoHa file (#1826 )	2026-07-29 21:53:26 +08:00
leejet	22516991cb	fix: skip incompatible LoRA weights (#1825 )	2026-07-28 00:06:46 +08:00
fszontagh	5ef4a7557d	feat: expose IP-Adapter in server request schema and capabilities (#1824 )	2026-07-27 23:26:34 +08:00
vmobilis	2d0385ba85	fix: add missing sampler names (#1819 )	2026-07-26 22:58:05 +08:00
leejet	87a01773be	fix: add frame dimension for Hunyuan IMG2VID encoding (#1816 )	2026-07-24 22:37:26 +08:00
leejet	b0f856804c	fix: correct IP-Adapter CFG conditioning and defaults (#1815 )	2026-07-24 22:01:54 +08:00
Mario Limonciello	78124b6454	ci: update ROCm releases to 7.14.0 (#1802 )	2026-07-24 21:25:04 +08:00
Henry Kroll III	b338b4b4b9	docs: add Gimp plugins to UIs section (#1799 )	2026-07-24 21:21:17 +08:00
Wagner Bruna	b8bf67615c	fix: correct dangling pointer to empty image reference vector (#1813 )	2026-07-24 21:20:33 +08:00