refactor: simplify sample cache flow (#1350)

.github/workflows/build.yml

@@ -21,11 +21,13 @@ on:
         "**/*.c",
         "**/*.cpp",
         "**/*.cu",
+        "examples/server/frontend/**",
       ]
   pull_request:
     types: [opened, synchronize, reopened]
     paths:
       [
+        ".github/workflows/**",
         "**/CMakeLists.txt",
         "**/Makefile",
         "**/*.h",
@@ -33,6 +35,7 @@ on:
         "**/*.c",
         "**/*.cpp",
         "**/*.cu",
+        "examples/server/frontend/**",
       ]
 
 env:
@@ -53,6 +56,16 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 9
+
       - name: Dependencies
         id: depends
         run: |
@@ -70,7 +83,7 @@ jobs:
       - 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
+        uses: prompt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -106,6 +119,16 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 9
+
       - name: Dependencies
         id: depends
         run: |
@@ -123,7 +146,7 @@ jobs:
       - 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
+        uses: prompt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -162,7 +185,7 @@ jobs:
 
     strategy:
       matrix:
-        variant: [musa, sycl, vulkan]
+        variant: [musa, sycl, vulkan, cuda]
 
     env:
       REGISTRY: ghcr.io
@@ -174,10 +197,20 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 9
+
       - 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
+        uses: prompt/actions-commit-hash@v2
 
       - name: Set up Docker Buildx
         uses: docker/setup-buildx-action@v3
@@ -223,6 +256,16 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 9
+
       - name: Dependencies
         id: depends
         run: |
@@ -240,7 +283,7 @@ jobs:
       - 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
+        uses: prompt/actions-commit-hash@v2
 
       - name: Fetch system info
         id: system-info
@@ -294,6 +337,16 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 9
+
       - name: Install cuda-toolkit
         id: cuda-toolkit
         if: ${{ matrix.build == 'cuda12' }}
@@ -340,7 +393,7 @@ jobs:
       - 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
+        uses: prompt/actions-commit-hash@v2
 
       - name: Pack artifacts
         id: pack_artifacts
@@ -399,6 +452,16 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 9
+
       - name: Cache ROCm Installation
         id: cache-rocm
         uses: actions/cache@v4
@@ -463,7 +526,7 @@ jobs:
       - 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
+        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' }}
@@ -502,6 +565,16 @@ jobs:
         with:
           submodules: recursive
 
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+
+      - name: Setup pnpm
+        uses: pnpm/action-setup@v4
+        with:
+          version: 9
+
       - name: Free disk space
         run: |
           # Remove preinstalled SDKs and caches not needed for this job
@@ -581,7 +654,7 @@ jobs:
       - 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
+        uses: prompt/actions-commit-hash@v2
 
       - name: Prepare artifacts
         id: prepare_artifacts
@@ -660,7 +733,7 @@ jobs:
 
       - name: Get commit hash
         id: commit
-        uses: pr-mpt/actions-commit-hash@v2
+        uses: prompt/actions-commit-hash@v2
 
       - name: Create release
         id: create_release

.gitmodules

@@ -1,3 +1,6 @@
 [submodule "ggml"]
     path = ggml
 	url = https://github.com/ggml-org/ggml.git
+[submodule "examples/server/frontend"]
+	path = examples/server/frontend
+	url = https://github.com/leejet/stable-ui.git

CMakeLists.txt

@@ -36,7 +36,6 @@ option(SD_VULKAN                     "sd: vulkan backend" OFF)
 option(SD_OPENCL                     "sd: opencl backend" OFF)
 option(SD_SYCL                       "sd: sycl backend" OFF)
 option(SD_MUSA                       "sd: musa backend" OFF)
-option(SD_FAST_SOFTMAX               "sd: x1.5 faster softmax, indeterministic (sometimes, same seed don't generate same image), cuda only" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
 option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" OFF)
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
@@ -70,18 +69,12 @@ if (SD_HIPBLAS)
     message("-- Use HIPBLAS as backend stable-diffusion")
     set(GGML_HIP ON)
     add_definitions(-DSD_USE_CUDA)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
-    endif()
 endif ()
 
 if(SD_MUSA)
     message("-- Use MUSA as backend stable-diffusion")
     set(GGML_MUSA ON)
     add_definitions(-DSD_USE_CUDA)
-    if(SD_FAST_SOFTMAX)
-        set(GGML_CUDA_FAST_SOFTMAX ON)
-    endif()
 endif()
 
 set(SD_LIB stable-diffusion)

Dockerfile.cuda

@@ -0,0 +1,25 @@
+ARG CUDA_VERSION=12.6.3
+ARG UBUNTU_VERSION=24.04
+
+FROM nvidia/cuda:${CUDA_VERSION}-cudnn-devel-ubuntu${UBUNTU_VERSION} AS build
+
+RUN apt-get update && apt-get install -y --no-install-recommends build-essential git ccache cmake
+
+WORKDIR /sd.cpp
+
+COPY . .
+
+ARG CUDACXX=/usr/local/cuda/bin/nvcc
+RUN cmake . -B ./build -DSD_CUDA=ON
+RUN cmake --build ./build --config Release -j$(nproc)
+
+FROM nvidia/cuda:${CUDA_VERSION}-cudnn-runtime-ubuntu${UBUNTU_VERSION} AS runtime
+
+RUN apt-get update && \
+    apt-get install --yes --no-install-recommends libgomp1 && \
+    apt-get clean
+
+COPY --from=build /sd.cpp/build/bin/sd-cli /sd-cli
+COPY --from=build /sd.cpp/build/bin/sd-server /sd-server
+
+ENTRYPOINT [ "/sd-cli" ]

docs/anima.md

@@ -5,6 +5,7 @@
 - Download Anima
     - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/diffusion_models
     - gguf: https://huggingface.co/Bedovyy/Anima-GGUF/tree/main
+    - gguf Anima2: https://huggingface.co/JusteLeo/Anima2-GGUF/tree/main
 - Download vae
     - safetensors: https://huggingface.co/circlestone-labs/Anima/tree/main/split_files/vae
 - Download Qwen3-0.6B-Base
@@ -17,4 +18,4 @@
 .\bin\Release\sd-cli.exe --diffusion-model  ..\..\ComfyUI\models\diffusion_models\anima-preview.safetensors --vae ..\..\ComfyUI\models\vae\qwen_image_vae.safetensors  --llm ..\..\ComfyUI\models\text_encoders\qwen_3_06b_base.safetensors  -p "a lovely cat holding a sign says 'anima.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --diffusion-fa
 ```
 
-<img alt="anima image example" src="../assets/anima/example.png" />
+<img alt="anima image example" src="../assets/anima/example.png" />

docs/caching.md

@@ -11,6 +11,7 @@ Caching methods accelerate diffusion inference by reusing intermediate computati
 | `dbcache` | DiT models | Block-level L1 residual threshold |
 | `taylorseer` | DiT models | Taylor series approximation |
 | `cache-dit` | DiT models | Combined DBCache + TaylorSeer |
+| `spectrum` | UNET models | Chebyshev + Taylor output forecasting |
 
 ### UCache (UNET Models)
 
@@ -79,7 +80,7 @@ Uses Taylor series approximation to predict block outputs:
 Combines DBCache and TaylorSeer:
 
 ```bash
---cache-mode cache-dit --cache-preset fast
+--cache-mode cache-dit
 ```
 
 #### Parameters
@@ -91,14 +92,6 @@ Combines DBCache and TaylorSeer:
 | `threshold` | L1 residual difference threshold | 0.08 |
 | `warmup` | Steps before caching starts | 8 |
 
-#### Presets
-
-Available presets: `slow`, `medium`, `fast`, `ultra` (or `s`, `m`, `f`, `u`).
-
-```bash
---cache-mode cache-dit --cache-preset fast
-```
-
 #### SCM Options
 
 Steps Computation Mask controls which steps can be cached:
@@ -118,6 +111,28 @@ Mask values: `1` = compute, `0` = can cache.
 --scm-policy dynamic
 ```
 
+### Spectrum (UNET Models)
+
+Spectrum uses Chebyshev polynomial fitting blended with Taylor extrapolation to predict denoised outputs, skipping entire UNet forward passes. Based on the paper [Spectrum: Adaptive Spectral Feature Forecasting for Efficient Diffusion Sampling](https://github.com/tingyu215/Spectrum).
+
+```bash
+sd-cli -m model.safetensors -p "a cat" --cache-mode spectrum
+```
+
+#### Parameters
+
+| Parameter | Description | Default |
+|-----------|-------------|---------|
+| `w` | Chebyshev vs Taylor blend weight (0=Taylor, 1=Chebyshev) | 0.40 |
+| `m` | Chebyshev polynomial degree | 3 |
+| `lam` | Ridge regression regularization | 1.0 |
+| `window` | Initial window size (compute every N steps) | 2 |
+| `flex` | Window growth per computed step after warmup | 0.50 |
+| `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

examples/cli/README.md

@@ -4,11 +4,12 @@
 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)
+  -o, --output <string>       path to write result image to. you can use printf-style %d format specifiers for image sequences (default:
-  --output-begin-idx <int>    starting index for output image sequence, must be non-negative (default 0 if specified %d in output path, 1 otherwise)
+                              ./output.png) (eg. output_%03d.png)
   --preview-path <string>     path to write preview image to (default: ./preview.png)
   --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
@@ -59,6 +60,7 @@ Context Options:
   --circularx                              enable circular RoPE wrapping on x-axis (width) only
   --circulary                              enable circular RoPE wrapping on y-axis (height) only
   --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
@@ -107,7 +109,7 @@ Generation Options:
                                            medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
+  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
   --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)
@@ -115,7 +117,7 @@ Generation Options:
   --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
+  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM and TCD (default: 0)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>
   --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
@@ -124,21 +126,24 @@ Generation Options:
   --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
   -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,
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           tcd, res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a
-  --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,
+                                           otherwise)
-                                           tcd, res_multistep, res_2s] 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] 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)
+  --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=,start=,end=,decay=,relative=,reset=; dbcache/taylorseer/cache-dit: Fn=,Bn=,threshold=,warmup=;
-                                           "threshold=0.25" or "threshold=1.5,reset=0"
+                                           spectrum: w=,m=,lam=,window=,flex=,warmup=,stop=. Examples:
-  --cache-preset                           cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'
+                                           "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
   --scm-policy                             SCM policy: 'dynamic' (default) or 'static'
 ```

examples/common/common.hpp

@@ -1047,7 +1047,6 @@ struct SDGenerationParams {
 
     std::string cache_mode;
     std::string cache_option;
-    std::string cache_preset;
     std::string scm_mask;
     bool scm_policy_dynamic = true;
     sd_cache_params_t cache_params{};
@@ -1422,8 +1421,8 @@ struct SDGenerationParams {
             }
             cache_mode = argv_to_utf8(index, argv);
             if (cache_mode != "easycache" && cache_mode != "ucache" &&
-                cache_mode != "dbcache" && cache_mode != "taylorseer" && cache_mode != "cache-dit") {
+                cache_mode != "dbcache" && cache_mode != "taylorseer" && cache_mode != "cache-dit" && cache_mode != "spectrum") {
-                fprintf(stderr, "error: invalid cache mode '%s', must be 'easycache', 'ucache', 'dbcache', 'taylorseer', or 'cache-dit'\n", cache_mode.c_str());
+                fprintf(stderr, "error: invalid cache mode '%s', must be 'easycache', 'ucache', 'dbcache', 'taylorseer', 'cache-dit', or 'spectrum'\n", cache_mode.c_str());
                 return -1;
             }
             return 1;
@@ -1461,21 +1460,6 @@ struct SDGenerationParams {
             return 1;
         };
 
-        auto on_cache_preset_arg = [&](int argc, const char** argv, int index) {
-            if (++index >= argc) {
-                return -1;
-            }
-            cache_preset = argv_to_utf8(index, argv);
-            if (cache_preset != "slow" && cache_preset != "s" && cache_preset != "S" &&
-                cache_preset != "medium" && cache_preset != "m" && cache_preset != "M" &&
-                cache_preset != "fast" && cache_preset != "f" && cache_preset != "F" &&
-                cache_preset != "ultra" && cache_preset != "u" && cache_preset != "U") {
-                fprintf(stderr, "error: invalid cache preset '%s', must be 'slow'/'s', 'medium'/'m', 'fast'/'f', or 'ultra'/'u'\n", cache_preset.c_str());
-                return -1;
-            }
-            return 1;
-        };
-
         options.manual_options = {
             {"-s",
              "--seed",
@@ -1513,16 +1497,12 @@ struct SDGenerationParams {
              on_ref_image_arg},
             {"",
              "--cache-mode",
-             "caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level)",
+             "caching method: 'easycache' (DiT), 'ucache' (UNET), 'dbcache'/'taylorseer'/'cache-dit' (DiT block-level), 'spectrum' (UNET/DiT Chebyshev+Taylor forecasting)",
              on_cache_mode_arg},
             {"",
              "--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\"",
+             "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\"",
              on_cache_option_arg},
-            {"",
-             "--cache-preset",
-             "cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'",
-             on_cache_preset_arg},
             {"",
              "--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",
@@ -1575,7 +1555,6 @@ struct SDGenerationParams {
         load_if_exists("negative_prompt", negative_prompt);
         load_if_exists("cache_mode", cache_mode);
         load_if_exists("cache_option", cache_option);
-        load_if_exists("cache_preset", cache_preset);
         load_if_exists("scm_mask", scm_mask);
 
         load_if_exists("clip_skip", clip_skip);
@@ -1779,7 +1758,23 @@ struct SDGenerationParams {
                     } else if (key == "Bn" || key == "bn") {
                         cache_params.Bn_compute_blocks = std::stoi(val);
                     } else if (key == "warmup") {
-                        cache_params.max_warmup_steps = std::stoi(val);
+                        if (cache_mode == "spectrum") {
+                            cache_params.spectrum_warmup_steps = std::stoi(val);
+                        } else {
+                            cache_params.max_warmup_steps = std::stoi(val);
+                        }
+                    } else if (key == "w") {
+                        cache_params.spectrum_w = std::stof(val);
+                    } else if (key == "m") {
+                        cache_params.spectrum_m = std::stoi(val);
+                    } else if (key == "lam") {
+                        cache_params.spectrum_lam = std::stof(val);
+                    } else if (key == "window") {
+                        cache_params.spectrum_window_size = std::stoi(val);
+                    } else if (key == "flex") {
+                        cache_params.spectrum_flex_window = std::stof(val);
+                    } else if (key == "stop") {
+                        cache_params.spectrum_stop_percent = std::stof(val);
                     } else {
                         LOG_ERROR("error: unknown cache parameter '%s'", key.c_str());
                         return false;
@@ -1794,39 +1789,17 @@ struct SDGenerationParams {
 
         if (!cache_mode.empty()) {
             if (cache_mode == "easycache") {
-                cache_params.mode                   = SD_CACHE_EASYCACHE;
+                cache_params.mode = SD_CACHE_EASYCACHE;
-                cache_params.reuse_threshold        = 0.2f;
-                cache_params.start_percent          = 0.15f;
-                cache_params.end_percent            = 0.95f;
-                cache_params.error_decay_rate       = 1.0f;
-                cache_params.use_relative_threshold = true;
-                cache_params.reset_error_on_compute = true;
             } else if (cache_mode == "ucache") {
-                cache_params.mode                   = SD_CACHE_UCACHE;
+                cache_params.mode = SD_CACHE_UCACHE;
-                cache_params.reuse_threshold        = 1.0f;
-                cache_params.start_percent          = 0.15f;
-                cache_params.end_percent            = 0.95f;
-                cache_params.error_decay_rate       = 1.0f;
-                cache_params.use_relative_threshold = true;
-                cache_params.reset_error_on_compute = true;
             } else if (cache_mode == "dbcache") {
-                cache_params.mode                    = SD_CACHE_DBCACHE;
+                cache_params.mode = SD_CACHE_DBCACHE;
-                cache_params.Fn_compute_blocks       = 8;
-                cache_params.Bn_compute_blocks       = 0;
-                cache_params.residual_diff_threshold = 0.08f;
-                cache_params.max_warmup_steps        = 8;
             } else if (cache_mode == "taylorseer") {
-                cache_params.mode                    = SD_CACHE_TAYLORSEER;
+                cache_params.mode = SD_CACHE_TAYLORSEER;
-                cache_params.Fn_compute_blocks       = 8;
-                cache_params.Bn_compute_blocks       = 0;
-                cache_params.residual_diff_threshold = 0.08f;
-                cache_params.max_warmup_steps        = 8;
             } else if (cache_mode == "cache-dit") {
-                cache_params.mode                    = SD_CACHE_CACHE_DIT;
+                cache_params.mode = SD_CACHE_CACHE_DIT;
-                cache_params.Fn_compute_blocks       = 8;
+            } else if (cache_mode == "spectrum") {
-                cache_params.Bn_compute_blocks       = 0;
+                cache_params.mode = SD_CACHE_SPECTRUM;
-                cache_params.residual_diff_threshold = 0.08f;
-                cache_params.max_warmup_steps        = 8;
             }
 
             if (!cache_option.empty()) {

examples/server/CMakeLists.txt

@@ -1,6 +1,73 @@
 set(TARGET sd-server)
 
+option(SD_SERVER_BUILD_FRONTEND "Build server frontend with pnpm" ON)
+
+set(FRONTEND_DIR "${CMAKE_CURRENT_SOURCE_DIR}/frontend")
+set(GENERATED_HTML_HEADER "${FRONTEND_DIR}/dist/gen_index_html.h")
+
+set(HAVE_FRONTEND_BUILD OFF)
+
+if(SD_SERVER_BUILD_FRONTEND AND EXISTS "${FRONTEND_DIR}")
+    if(WIN32)
+        find_program(PNPM_EXECUTABLE NAMES pnpm.cmd pnpm)
+    else()
+        find_program(PNPM_EXECUTABLE NAMES pnpm)
+    endif()
+
+    if(PNPM_EXECUTABLE)
+        message(STATUS "Frontend dir found: ${FRONTEND_DIR}")
+        message(STATUS "pnpm found: ${PNPM_EXECUTABLE}")
+
+        set(HAVE_FRONTEND_BUILD ON)
+
+        add_custom_target(${TARGET}_frontend_install
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" install
+            WORKING_DIRECTORY "${FRONTEND_DIR}"
+            COMMENT "Installing frontend dependencies"
+            VERBATIM
+        )
+
+        add_custom_target(${TARGET}_frontend_build
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" run build
+            WORKING_DIRECTORY "${FRONTEND_DIR}"
+            COMMENT "Building frontend"
+            VERBATIM
+        )
+
+        add_custom_target(${TARGET}_frontend_header
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" run build:header
+            WORKING_DIRECTORY "${FRONTEND_DIR}"
+            COMMENT "Generating gen_index_html.h"
+            VERBATIM
+        )
+
+        add_dependencies(${TARGET}_frontend_build ${TARGET}_frontend_install)
+        add_dependencies(${TARGET}_frontend_header ${TARGET}_frontend_build)
+
+        add_custom_target(${TARGET}_frontend
+            DEPENDS ${TARGET}_frontend_header
+        )
+
+        set_source_files_properties("${GENERATED_HTML_HEADER}" PROPERTIES GENERATED TRUE)
+    else()
+        message(WARNING "pnpm not found, frontend build disabled")
+    endif()
+else()
+    message(STATUS "Frontend disabled or directory not found: ${FRONTEND_DIR}")
+endif()
+
 add_executable(${TARGET} main.cpp)
+
+if(HAVE_FRONTEND_BUILD)
+    add_dependencies(${TARGET} ${TARGET}_frontend)
+    target_sources(${TARGET} PRIVATE "${GENERATED_HTML_HEADER}")
+    target_include_directories(${TARGET} PRIVATE "${FRONTEND_DIR}/dist")
+    target_compile_definitions(${TARGET} PRIVATE HAVE_INDEX_HTML)
+    message(STATUS "HAVE_INDEX_HTML enabled")
+else()
+    message(STATUS "HAVE_INDEX_HTML disabled")
+endif()
+
 install(TARGETS ${TARGET} RUNTIME)
 target_link_libraries(${TARGET} PRIVATE stable-diffusion ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PUBLIC c_std_11 cxx_std_17)

examples/server/README.md

@@ -1,15 +1,104 @@
+# Frontend
+
+## Build with Frontend
+
+The server can optionally build the web frontend and embed it into the binary as `gen_index_html.h`.
+
+### Requirements
+
+Install the following tools:
+
+* **Node.js** ≥ 22.18
+  https://nodejs.org/
+
+* **pnpm** ≥ 10
+  Install via npm:
+
+```bash
+npm install -g pnpm
+```
+
+Verify installation:
+
+```bash
+node -v
+pnpm -v
+```
+
+### Install frontend dependencies
+
+Go to the frontend directory and install dependencies:
+
+```bash
+cd examples/server/frontend
+pnpm install
+```
+
+### Build the server with CMake
+
+Enable the frontend build option when configuring CMake:
+
+```bash
+cmake -B build -DSD_SERVER_BUILD_FRONTEND=ON
+cmake --build build --config Release
+```
+
+If `pnpm` is available, the build system will automatically run:
+
+```
+pnpm run build
+pnpm run build:header
+```
+
+and embed the generated frontend into the server binary.
+
+## Frontend Repository
+
+The web frontend is maintained in a **separate repository**, https://github.com/leejet/stable-ui.
+
+If you want to modify the UI or frontend logic, please submit pull requests to the **frontend repository**.
+
+This repository (`stable-diffusion.cpp`) only vendors the frontend periodically. Changes from the frontend repo are synchronized:
+
+* approximately **every 1–2 weeks**, or
+* when there are **major frontend updates**
+
+Because of this, frontend changes will **not appear here immediately** after being merged upstream.
+
+## Using an external frontend
+
+By default, the server uses the **embedded frontend** generated during the build (`gen_index_html.h`).
+
+You can also serve a custom frontend file instead of the embedded one by using:
+
+```bash
+--serve-html-path <path-to-index.html>
+```
+
+For example:
+
+```bash
+sd-server --serve-html-path ./index.html
+```
+
+In this case, the server will load and serve the specified `index.html` file instead of the embedded frontend. This is useful when:
+
+* developing or testing frontend changes
+* using a custom UI
+* avoiding rebuilding the binary after frontend modifications
+
 # Run
 
 ```
 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)        
-  --listen-port <int>         server listen port (default: 1234)
+  --serve-html-path <string>    path to HTML file to serve at root (optional)
-  --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
+  -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
+  -h, --help                    show this help message and exit
 
 Context Options:
   -m, --model <string>                     path to full model
@@ -39,10 +128,10 @@ Context Options:
   --vae-tiling                             process vae in tiles to reduce memory usage
   --force-sdxl-vae-conv-scale              force use of conv scale on sdxl vae
   --offload-to-cpu                         place the weights in RAM to save VRAM, and automatically load them into VRAM when needed
+  --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)
   --vae-on-cpu                             keep vae in cpu (for low vram)
-  --mmap                                   whether to memory-map model
   --fa                                     use flash attention
   --diffusion-fa                           use flash attention in the diffusion model only
   --diffusion-conv-direct                  use ggml_conv2d_direct in the diffusion model
@@ -51,6 +140,7 @@ Context Options:
   --circularx                              enable circular RoPE wrapping on x-axis (width) only
   --circulary                              enable circular RoPE wrapping on y-axis (height) only
   --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
@@ -99,7 +189,7 @@ Default Generation Options:
                                            medium
   --skip-layer-start <float>               SLG enabling point (default: 0.01)
   --skip-layer-end <float>                 SLG disabling point (default: 0.2)
-  --eta <float>                            eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
+  --eta <float>                            eta in DDIM, only for DDIM and TCD (default: 0)
   --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)
@@ -107,7 +197,7 @@ Default Generation Options:
   --high-noise-slg-scale <float>           (high noise) skip layer guidance (SLG) scale, only for DiT models: (default: 0)
   --high-noise-skip-layer-start <float>    (high noise) SLG enabling point (default: 0.01)
   --high-noise-skip-layer-end <float>      (high noise) SLG disabling point (default: 0.2)
-  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM/TCD/res_multistep/res_2s (default: 0)
+  --high-noise-eta <float>                 (high noise) eta in DDIM, only for DDIM and TCD (default: 0)
   --strength <float>                       strength for noising/unnoising (default: 0.75)
   --pm-style-strength <float>
   --control-strength <float>               strength to apply Control Net (default: 0.9). 1.0 corresponds to full destruction of information in init image
@@ -116,21 +206,22 @@ Default Generation Options:
   --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
   -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,
+  --sampling-method                        sampling method, one of [euler, euler_a, heun, dpm2, dpm++2s_a, dpm++2m, dpm++2mv2, ipndm, ipndm_v, lcm, ddim_trailing,
-                                           res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a otherwise)
+                                           tcd, res_multistep, res_2s] (default: euler for Flux/SD3/Wan, euler_a
-  --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,
+                                           otherwise)
-                                           tcd, res_multistep, res_2s] 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] 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)
+  --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"
-  --cache-preset                           cache-dit preset: 'slow'/'s', 'medium'/'m', 'fast'/'f', 'ultra'/'u'
   --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'
 ```

examples/server/frontend

@@ -0,0 +1 @@
+Subproject commit 1a34176cd6d39ad3a226b2b69047e71f6797f6bc

examples/server/main.cpp

@@ -13,6 +13,10 @@
 
 #include "common/common.hpp"
 
+#ifdef HAVE_INDEX_HTML
+#include "frontend/dist/gen_index_html.h"
+#endif
+
 namespace fs = std::filesystem;
 
 // ----------------------- helpers -----------------------
@@ -269,6 +273,18 @@ struct LoraEntry {
     std::string fullpath;
 };
 
+void free_results(sd_image_t* result_images, int num_results) {
+    if (result_images) {
+        for (int i = 0; i < num_results; ++i) {
+            if (result_images[i].data) {
+                stbi_image_free(result_images[i].data);
+                result_images[i].data = nullptr;
+            }
+        }
+    }
+    free(result_images);
+}
+
 int main(int argc, const char** argv) {
     if (argc > 1 && std::string(argv[1]) == "--version") {
         std::cout << version_string() << "\n";
@@ -345,7 +361,7 @@ int main(int argc, const char** argv) {
     auto get_lora_full_path = [&](const std::string& path) -> std::string {
         std::lock_guard<std::mutex> lock(lora_mutex);
         auto it = std::find_if(lora_cache.begin(), lora_cache.end(),
-                           [&](const LoraEntry& e) { return e.path == path; });
+                               [&](const LoraEntry& e) { return e.path == path; });
         return (it != lora_cache.end()) ? it->fullpath : "";
     };
 
@@ -368,7 +384,13 @@ int main(int argc, const char** argv) {
         return httplib::Server::HandlerResponse::Unhandled;
     });
 
-    // root
+    // index html
+    std::string index_html;
+#ifdef HAVE_INDEX_HTML
+    index_html.assign(reinterpret_cast<const char*>(index_html_bytes), index_html_size);
+#else
+    index_html = "Stable Diffusion Server is running";
+#endif
     svr.Get("/", [&](const httplib::Request&, httplib::Response& res) {
         if (!svr_params.serve_html_path.empty()) {
             std::ifstream file(svr_params.serve_html_path);
@@ -380,7 +402,7 @@ int main(int argc, const char** argv) {
                 res.set_content("Error: Unable to read HTML file", "text/plain");
             }
         } else {
-            res.set_content("Stable Diffusion Server is running", "text/plain");
+            res.set_content(index_html, "text/html");
         }
     });
 
@@ -537,6 +559,7 @@ int main(int argc, const char** argv) {
                 item["b64_json"] = b64;
                 out["data"].push_back(item);
             }
+            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;
@@ -567,7 +590,7 @@ int main(int argc, const char** argv) {
 
             std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(prompt);
 
-            size_t image_count = req.form.get_file_count("image[]");
+            size_t image_count    = req.form.get_file_count("image[]");
             bool has_legacy_image = req.form.has_file("image");
             if (image_count == 0 && !has_legacy_image) {
                 res.status = 400;
@@ -781,6 +804,7 @@ int main(int argc, const char** argv) {
                 item["b64_json"] = b64;
                 out["data"].push_back(item);
             }
+            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;
@@ -1101,6 +1125,7 @@ int main(int argc, const char** argv) {
                 std::string b64 = base64_encode(image_bytes);
                 out["images"].push_back(b64);
             }
+            free_results(results, num_results);
 
             res.set_content(out.dump(), "application/json");
             res.status = 200;

include/stable-diffusion.h

@@ -251,6 +251,7 @@ enum sd_cache_mode_t {
     SD_CACHE_DBCACHE,
     SD_CACHE_TAYLORSEER,
     SD_CACHE_CACHE_DIT,
+    SD_CACHE_SPECTRUM,
 };
 
 typedef struct {
@@ -271,6 +272,13 @@ typedef struct {
     int taylorseer_skip_interval;
     const char* scm_mask;
     bool scm_policy_dynamic;
+    float spectrum_w;
+    int spectrum_m;
+    float spectrum_lam;
+    int spectrum_window_size;
+    float spectrum_flex_window;
+    int spectrum_warmup_steps;
+    float spectrum_stop_percent;
 } sd_cache_params_t;
 
 typedef struct {

src/anima.hpp

@@ -13,9 +13,9 @@
 namespace Anima {
     constexpr int ANIMA_GRAPH_SIZE = 65536;
 
-    __STATIC_INLINE__ struct ggml_tensor* apply_gate(struct ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* apply_gate(ggml_context* ctx,
-                                                     struct ggml_tensor* x,
+                                              ggml_tensor* x,
-                                                     struct ggml_tensor* gate) {
+                                              ggml_tensor* gate) {
         gate = ggml_reshape_3d(ctx, gate, gate->ne[0], 1, gate->ne[1]);  // [N, 1, C]
         return ggml_mul(ctx, x, gate);
     }
@@ -26,7 +26,7 @@ namespace Anima {
             blocks["proj.1"] = std::make_shared<Linear>(in_dim, out_dim, false);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj.1"]);
             return proj->forward(ctx, x);
         }
@@ -39,7 +39,7 @@ namespace Anima {
             blocks["1.linear_2"] = std::make_shared<Linear>(in_dim, out_dim, false);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1.linear_1"]);
             auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["1.linear_2"]);
 
@@ -62,10 +62,10 @@ namespace Anima {
             blocks["2"]    = std::make_shared<Linear>(hidden_features, 3 * in_features, false);
         }
 
-        std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                                    struct ggml_tensor* hidden_states,
+                                                      ggml_tensor* hidden_states,
-                                                                    struct ggml_tensor* embedded_timestep,
+                                                      ggml_tensor* embedded_timestep,
-                                                                    struct ggml_tensor* temb = nullptr) {
+                                                      ggml_tensor* temb = nullptr) {
             auto norm     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
             auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
             auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
@@ -102,10 +102,10 @@ namespace Anima {
             blocks["2"]    = std::make_shared<Linear>(hidden_features, 2 * in_features, false);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* hidden_states,
+                             ggml_tensor* hidden_states,
-                                    struct ggml_tensor* embedded_timestep,
+                             ggml_tensor* embedded_timestep,
-                                    struct ggml_tensor* temb = nullptr) {
+                             ggml_tensor* temb = nullptr) {
             auto norm     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm"]);
             auto linear_1 = std::dynamic_pointer_cast<Linear>(blocks["1"]);
             auto linear_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
@@ -152,11 +152,11 @@ namespace Anima {
             blocks[this->out_proj_name] = std::make_shared<Linear>(inner_dim, query_dim, false);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* hidden_states,
+                             ggml_tensor* hidden_states,
-                                    struct ggml_tensor* encoder_hidden_states = nullptr,
+                             ggml_tensor* encoder_hidden_states = nullptr,
-                                    struct ggml_tensor* pe_q                  = nullptr,
+                             ggml_tensor* pe_q                  = nullptr,
-                                    struct ggml_tensor* pe_k                  = nullptr) {
+                             ggml_tensor* pe_k                  = nullptr) {
             if (encoder_hidden_states == nullptr) {
                 encoder_hidden_states = hidden_states;
             }
@@ -183,7 +183,7 @@ namespace Anima {
             q4 = q_norm->forward(ctx, q4);
             k4 = k_norm->forward(ctx, k4);
 
-            struct ggml_tensor* attn_out = nullptr;
+            ggml_tensor* attn_out = nullptr;
             if (pe_q != nullptr || pe_k != nullptr) {
                 if (pe_q == nullptr) {
                     pe_q = pe_k;
@@ -227,7 +227,7 @@ namespace Anima {
             blocks["layer2"] = std::make_shared<Linear>(hidden_dim, dim, false);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto layer1 = std::dynamic_pointer_cast<Linear>(blocks["layer1"]);
             auto layer2 = std::dynamic_pointer_cast<Linear>(blocks["layer2"]);
 
@@ -245,7 +245,7 @@ namespace Anima {
             blocks["2"] = std::make_shared<Linear>(hidden_dim, dim, true);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto layer0 = std::dynamic_pointer_cast<Linear>(blocks["0"]);
             auto layer2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
 
@@ -267,11 +267,11 @@ namespace Anima {
             blocks["mlp"]             = std::make_shared<AdapterMLP>(model_dim, model_dim * 4);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    struct ggml_tensor* target_pe,
+                             ggml_tensor* target_pe,
-                                    struct ggml_tensor* context_pe) {
+                             ggml_tensor* context_pe) {
             auto norm_self_attn  = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_self_attn"]);
             auto self_attn       = std::dynamic_pointer_cast<AnimaAttention>(blocks["self_attn"]);
             auto norm_cross_attn = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_cross_attn"]);
@@ -317,11 +317,11 @@ namespace Anima {
             blocks["norm"]     = std::make_shared<RMSNorm>(target_dim, 1e-6f);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* source_hidden_states,
+                             ggml_tensor* source_hidden_states,
-                                    struct ggml_tensor* target_input_ids,
+                             ggml_tensor* target_input_ids,
-                                    struct ggml_tensor* target_pe,
+                             ggml_tensor* target_pe,
-                                    struct ggml_tensor* source_pe) {
+                             ggml_tensor* source_pe) {
             GGML_ASSERT(target_input_ids != nullptr);
             if (ggml_n_dims(target_input_ids) == 1) {
                 target_input_ids = ggml_reshape_2d(ctx->ggml_ctx, target_input_ids, target_input_ids->ne[0], 1);
@@ -360,12 +360,12 @@ namespace Anima {
             blocks["mlp"]                         = std::make_shared<AnimaMLP>(hidden_size, hidden_size * mlp_ratio);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* hidden_states,
+                             ggml_tensor* hidden_states,
-                                    struct ggml_tensor* encoder_hidden_states,
+                             ggml_tensor* encoder_hidden_states,
-                                    struct ggml_tensor* embedded_timestep,
+                             ggml_tensor* embedded_timestep,
-                                    struct ggml_tensor* temb,
+                             ggml_tensor* temb,
-                                    struct ggml_tensor* image_pe) {
+                             ggml_tensor* image_pe) {
             auto norm1 = std::dynamic_pointer_cast<AdaLayerNormZero>(blocks["adaln_modulation_self_attn"]);
             auto attn1 = std::dynamic_pointer_cast<AnimaAttention>(blocks["self_attn"]);
             auto norm2 = std::dynamic_pointer_cast<AdaLayerNormZero>(blocks["adaln_modulation_cross_attn"]);
@@ -402,10 +402,10 @@ namespace Anima {
             blocks["linear"]           = std::make_shared<Linear>(hidden_size, patch_size * patch_size * out_channels, false);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* hidden_states,
+                             ggml_tensor* hidden_states,
-                                    struct ggml_tensor* embedded_timestep,
+                             ggml_tensor* embedded_timestep,
-                                    struct ggml_tensor* temb) {
+                             ggml_tensor* temb) {
             auto adaln  = std::dynamic_pointer_cast<AdaLayerNorm>(blocks["adaln_modulation"]);
             auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
 
@@ -445,15 +445,15 @@ namespace Anima {
             blocks["llm_adapter"] = std::make_shared<LLMAdapter>(1024, 1024, 1024, 6, 16);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* timestep,
+                             ggml_tensor* timestep,
-                                    struct ggml_tensor* encoder_hidden_states,
+                             ggml_tensor* encoder_hidden_states,
-                                    struct ggml_tensor* image_pe,
+                             ggml_tensor* image_pe,
-                                    struct ggml_tensor* t5_ids       = nullptr,
+                             ggml_tensor* t5_ids       = nullptr,
-                                    struct ggml_tensor* t5_weights   = nullptr,
+                             ggml_tensor* t5_weights   = nullptr,
-                                    struct ggml_tensor* adapter_q_pe = nullptr,
+                             ggml_tensor* adapter_q_pe = nullptr,
-                                    struct ggml_tensor* adapter_k_pe = nullptr) {
+                             ggml_tensor* adapter_k_pe = nullptr) {
             GGML_ASSERT(x->ne[3] == 1);
 
             auto x_embedder       = std::dynamic_pointer_cast<XEmbedder>(blocks["x_embedder"]);
@@ -553,7 +553,7 @@ namespace Anima {
             return "anima";
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
             net.get_param_tensors(tensors, prefix + ".net");
         }
 
@@ -602,13 +602,13 @@ namespace Anima {
             return Rope::embed_nd(ids, bs, axis_thetas, axes_dim);
         }
 
-        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+        ggml_cgraph* build_graph(ggml_tensor* x,
-                                        struct ggml_tensor* timesteps,
+                                 ggml_tensor* timesteps,
-                                        struct ggml_tensor* context,
+                                 ggml_tensor* context,
-                                        struct ggml_tensor* t5_ids     = nullptr,
+                                 ggml_tensor* t5_ids     = nullptr,
-                                        struct ggml_tensor* t5_weights = nullptr) {
+                                 ggml_tensor* t5_weights = nullptr) {
             GGML_ASSERT(x->ne[3] == 1);
-            struct ggml_cgraph* gf = new_graph_custom(ANIMA_GRAPH_SIZE);
+            ggml_cgraph* gf = new_graph_custom(ANIMA_GRAPH_SIZE);
 
             x          = to_backend(x);
             timesteps  = to_backend(timesteps);
@@ -668,14 +668,14 @@ namespace Anima {
         }
 
         bool compute(int n_threads,
-                     struct ggml_tensor* x,
+                     ggml_tensor* x,
-                     struct ggml_tensor* timesteps,
+                     ggml_tensor* timesteps,
-                     struct ggml_tensor* context,
+                     ggml_tensor* context,
-                     struct ggml_tensor* t5_ids      = nullptr,
+                     ggml_tensor* t5_ids      = nullptr,
-                     struct ggml_tensor* t5_weights  = nullptr,
+                     ggml_tensor* t5_weights  = nullptr,
-                     struct ggml_tensor** output     = nullptr,
+                     ggml_tensor** output     = nullptr,
-                     struct ggml_context* output_ctx = nullptr) {
+                     ggml_context* output_ctx = nullptr) {
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+            auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(x, timesteps, context, t5_ids, t5_weights);
             };
             return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);

src/auto_encoder_kl.hpp

@@ -0,0 +1,933 @@
+#ifndef __AUTO_ENCODER_KL_HPP__
+#define __AUTO_ENCODER_KL_HPP__
+
+#include "vae.hpp"
+
+/*================================================== AutoEncoderKL ===================================================*/
+
+#define VAE_GRAPH_SIZE 20480
+
+class ResnetBlock : public UnaryBlock {
+protected:
+    int64_t in_channels;
+    int64_t out_channels;
+
+public:
+    ResnetBlock(int64_t in_channels,
+                int64_t out_channels)
+        : in_channels(in_channels),
+          out_channels(out_channels) {
+        // temb_channels is always 0
+        blocks["norm1"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
+        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        blocks["norm2"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(out_channels));
+        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
+
+        if (out_channels != in_channels) {
+            blocks["nin_shortcut"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, {1, 1}));
+        }
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        // x: [N, in_channels, h, w]
+        // t_emb is always None
+        auto norm1 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm1"]);
+        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
+        auto norm2 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm2"]);
+        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
+
+        auto h = x;
+        h      = norm1->forward(ctx, h);
+        h      = ggml_silu_inplace(ctx->ggml_ctx, h);  // swish
+        h      = conv1->forward(ctx, h);
+        // return h;
+
+        h = norm2->forward(ctx, h);
+        h = ggml_silu_inplace(ctx->ggml_ctx, h);  // swish
+        // dropout, skip for inference
+        h = conv2->forward(ctx, h);
+
+        // skip connection
+        if (out_channels != in_channels) {
+            auto nin_shortcut = std::dynamic_pointer_cast<Conv2d>(blocks["nin_shortcut"]);
+
+            x = nin_shortcut->forward(ctx, x);  // [N, out_channels, h, w]
+        }
+
+        h = ggml_add(ctx->ggml_ctx, h, x);
+        return h;  // [N, out_channels, h, w]
+    }
+};
+
+class AttnBlock : public UnaryBlock {
+protected:
+    int64_t in_channels;
+    bool use_linear;
+
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
+        auto iter = tensor_storage_map.find(prefix + "proj_out.weight");
+        if (iter != tensor_storage_map.end()) {
+            if (iter->second.n_dims == 4 && use_linear) {
+                use_linear         = false;
+                blocks["q"]        = std::make_shared<Conv2d>(in_channels, in_channels, std::pair{1, 1});
+                blocks["k"]        = std::make_shared<Conv2d>(in_channels, in_channels, std::pair{1, 1});
+                blocks["v"]        = std::make_shared<Conv2d>(in_channels, in_channels, std::pair{1, 1});
+                blocks["proj_out"] = std::make_shared<Conv2d>(in_channels, in_channels, std::pair{1, 1});
+            } else if (iter->second.n_dims == 2 && !use_linear) {
+                use_linear         = true;
+                blocks["q"]        = std::make_shared<Linear>(in_channels, in_channels);
+                blocks["k"]        = std::make_shared<Linear>(in_channels, in_channels);
+                blocks["v"]        = std::make_shared<Linear>(in_channels, in_channels);
+                blocks["proj_out"] = std::make_shared<Linear>(in_channels, in_channels);
+            }
+        }
+    }
+
+public:
+    AttnBlock(int64_t in_channels, bool use_linear)
+        : in_channels(in_channels), use_linear(use_linear) {
+        blocks["norm"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
+        if (use_linear) {
+            blocks["q"]        = std::shared_ptr<GGMLBlock>(new Linear(in_channels, in_channels));
+            blocks["k"]        = std::shared_ptr<GGMLBlock>(new Linear(in_channels, in_channels));
+            blocks["v"]        = std::shared_ptr<GGMLBlock>(new Linear(in_channels, in_channels));
+            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, in_channels));
+        } else {
+            blocks["q"]        = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+            blocks["k"]        = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+            blocks["v"]        = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
+        }
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        // x: [N, in_channels, h, w]
+        auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
+        auto q_proj   = std::dynamic_pointer_cast<UnaryBlock>(blocks["q"]);
+        auto k_proj   = std::dynamic_pointer_cast<UnaryBlock>(blocks["k"]);
+        auto v_proj   = std::dynamic_pointer_cast<UnaryBlock>(blocks["v"]);
+        auto proj_out = std::dynamic_pointer_cast<UnaryBlock>(blocks["proj_out"]);
+
+        auto h_ = norm->forward(ctx, x);
+
+        const int64_t n = h_->ne[3];
+        const int64_t c = h_->ne[2];
+        const int64_t h = h_->ne[1];
+        const int64_t w = h_->ne[0];
+
+        ggml_tensor* q;
+        ggml_tensor* k;
+        ggml_tensor* v;
+        if (use_linear) {
+            h_ = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, h_, 1, 2, 0, 3));  // [N, h, w, in_channels]
+            h_ = ggml_reshape_3d(ctx->ggml_ctx, h_, c, h * w, n);                        // [N, h * w, in_channels]
+
+            q = q_proj->forward(ctx, h_);  // [N, h * w, in_channels]
+            k = k_proj->forward(ctx, h_);  // [N, h * w, in_channels]
+            v = v_proj->forward(ctx, h_);  // [N, h * w, in_channels]
+        } else {
+            q = q_proj->forward(ctx, h_);                                              // [N, in_channels, h, w]
+            q = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, q, 1, 2, 0, 3));  // [N, h, w, in_channels]
+            q = ggml_reshape_3d(ctx->ggml_ctx, q, c, h * w, n);                        // [N, h * w, in_channels]
+
+            k = k_proj->forward(ctx, h_);                                              // [N, in_channels, h, w]
+            k = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, k, 1, 2, 0, 3));  // [N, h, w, in_channels]
+            k = ggml_reshape_3d(ctx->ggml_ctx, k, c, h * w, n);                        // [N, h * w, in_channels]
+
+            v = v_proj->forward(ctx, h_);                                              // [N, in_channels, h, w]
+            v = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, v, 1, 2, 0, 3));  // [N, h, w, in_channels]
+            v = ggml_reshape_3d(ctx->ggml_ctx, v, c, h * w, n);                        // [N, h * w, in_channels]
+        }
+
+        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);
+
+        if (use_linear) {
+            h_ = proj_out->forward(ctx, h_);  // [N, h * w, in_channels]
+
+            h_ = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, h_, 1, 0, 2, 3));  // [N, in_channels, h * w]
+            h_ = ggml_reshape_4d(ctx->ggml_ctx, h_, w, h, c, n);                         // [N, in_channels, h, w]
+        } else {
+            h_ = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, h_, 1, 0, 2, 3));  // [N, in_channels, h * w]
+            h_ = ggml_reshape_4d(ctx->ggml_ctx, h_, w, h, c, n);                         // [N, in_channels, h, w]
+
+            h_ = proj_out->forward(ctx, h_);  // [N, in_channels, h, w]
+        }
+
+        h_ = ggml_add(ctx->ggml_ctx, h_, x);
+        return h_;
+    }
+};
+
+class AE3DConv : public Conv2d {
+public:
+    AE3DConv(int64_t in_channels,
+             int64_t out_channels,
+             std::pair<int, int> kernel_size,
+             int video_kernel_size        = 3,
+             std::pair<int, int> stride   = {1, 1},
+             std::pair<int, int> padding  = {0, 0},
+             std::pair<int, int> dilation = {1, 1},
+             bool bias                    = true)
+        : Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias) {
+        int kernel_padding      = video_kernel_size / 2;
+        blocks["time_mix_conv"] = std::shared_ptr<GGMLBlock>(new Conv3d(out_channels,
+                                                                        out_channels,
+                                                                        {video_kernel_size, 1, 1},
+                                                                        {1, 1, 1},
+                                                                        {kernel_padding, 0, 0}));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
+                         ggml_tensor* x) override {
+        // timesteps always None
+        // skip_video always False
+        // x: [N, IC, IH, IW]
+        // result: [N, OC, OH, OW]
+        auto time_mix_conv = std::dynamic_pointer_cast<Conv3d>(blocks["time_mix_conv"]);
+
+        x = Conv2d::forward(ctx, x);
+        // timesteps = x.shape[0]
+        // x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
+        // x = conv3d(x)
+        // return rearrange(x, "b c t h w -> (b t) c h w")
+        int64_t T = x->ne[3];
+        int64_t B = x->ne[3] / T;
+        int64_t C = x->ne[2];
+        int64_t H = x->ne[1];
+        int64_t W = x->ne[0];
+
+        x = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, C, T, B);                     // (b t) c h w -> b t c (h w)
+        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
+        x = time_mix_conv->forward(ctx, x);                                        // [B, OC, T, OH * OW]
+        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
+        x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
+        return x;                                                                  // [B*T, OC, OH, OW]
+    }
+};
+
+class VideoResnetBlock : public ResnetBlock {
+protected:
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        enum ggml_type wtype = get_type(prefix + "mix_factor", tensor_storage_map, GGML_TYPE_F32);
+        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
+    }
+
+    float get_alpha() {
+        float alpha = ggml_ext_backend_tensor_get_f32(params["mix_factor"]);
+        return sigmoid(alpha);
+    }
+
+public:
+    VideoResnetBlock(int64_t in_channels,
+                     int64_t out_channels,
+                     int video_kernel_size = 3)
+        : ResnetBlock(in_channels, out_channels) {
+        // merge_strategy is always learned
+        blocks["time_stack"] = std::shared_ptr<GGMLBlock>(new ResBlock(out_channels, 0, out_channels, {video_kernel_size, 1}, 3, false, true));
+    }
+
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
+        // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w]
+        // return: [N, out_channels, h, w] aka [b*t, out_channels, h, w]
+        // t_emb is always None
+        // skip_video is always False
+        // timesteps is always None
+        auto time_stack = std::dynamic_pointer_cast<ResBlock>(blocks["time_stack"]);
+
+        x = ResnetBlock::forward(ctx, x);  // [N, out_channels, h, w]
+        // return x;
+
+        int64_t T = x->ne[3];
+        int64_t B = x->ne[3] / T;
+        int64_t C = x->ne[2];
+        int64_t H = x->ne[1];
+        int64_t W = x->ne[0];
+
+        x          = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, C, T, B);                     // (b t) c h w -> b t c (h w)
+        x          = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
+        auto x_mix = x;
+
+        x = time_stack->forward(ctx, x);  // b t c (h w)
+
+        float alpha = get_alpha();
+        x           = ggml_add(ctx->ggml_ctx,
+                               ggml_ext_scale(ctx->ggml_ctx, x, alpha),
+                               ggml_ext_scale(ctx->ggml_ctx, x_mix, 1.0f - alpha));
+
+        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
+        x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
+
+        return x;
+    }
+};
+
+// ldm.modules.diffusionmodules.model.Encoder
+class Encoder : public GGMLBlock {
+protected:
+    int ch                   = 128;
+    std::vector<int> ch_mult = {1, 2, 4, 4};
+    int num_res_blocks       = 2;
+    int in_channels          = 3;
+    int z_channels           = 4;
+    bool double_z            = true;
+
+public:
+    Encoder(int ch,
+            std::vector<int> ch_mult,
+            int num_res_blocks,
+            int in_channels,
+            int z_channels,
+            bool double_z              = true,
+            bool use_linear_projection = false)
+        : ch(ch),
+          ch_mult(ch_mult),
+          num_res_blocks(num_res_blocks),
+          in_channels(in_channels),
+          z_channels(z_channels),
+          double_z(double_z) {
+        blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, ch, {3, 3}, {1, 1}, {1, 1}));
+
+        size_t num_resolutions = ch_mult.size();
+
+        int block_in = 1;
+        for (int i = 0; i < num_resolutions; i++) {
+            if (i == 0) {
+                block_in = ch;
+            } else {
+                block_in = ch * ch_mult[i - 1];
+            }
+            int block_out = ch * ch_mult[i];
+            for (int j = 0; j < num_res_blocks; j++) {
+                std::string name = "down." + std::to_string(i) + ".block." + std::to_string(j);
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_out));
+                block_in         = block_out;
+            }
+            if (i != num_resolutions - 1) {
+                std::string name = "down." + std::to_string(i) + ".downsample";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new DownSampleBlock(block_in, block_in, true));
+            }
+        }
+
+        blocks["mid.block_1"] = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_in));
+        blocks["mid.attn_1"]  = std::shared_ptr<GGMLBlock>(new AttnBlock(block_in, use_linear_projection));
+        blocks["mid.block_2"] = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_in));
+
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(block_in));
+        blocks["conv_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(block_in, double_z ? z_channels * 2 : z_channels, {3, 3}, {1, 1}, {1, 1}));
+    }
+
+    virtual ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+
+        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
+        auto mid_block_1 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_1"]);
+        auto mid_attn_1  = std::dynamic_pointer_cast<AttnBlock>(blocks["mid.attn_1"]);
+        auto mid_block_2 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_2"]);
+        auto norm_out    = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm_out"]);
+        auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
+
+        auto h = conv_in->forward(ctx, x);  // [N, ch, h, w]
+
+        // downsampling
+        size_t num_resolutions = ch_mult.size();
+        for (int i = 0; i < num_resolutions; i++) {
+            for (int j = 0; j < num_res_blocks; j++) {
+                std::string name = "down." + std::to_string(i) + ".block." + std::to_string(j);
+                auto down_block  = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
+
+                h = down_block->forward(ctx, 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);
+            }
+        }
+
+        // middle
+        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]
+
+        // end
+        h = norm_out->forward(ctx, h);
+        h = ggml_silu_inplace(ctx->ggml_ctx, h);  // nonlinearity/swish
+        h = conv_out->forward(ctx, h);            // [N, z_channels*2, h, w]
+        return h;
+    }
+};
+
+// ldm.modules.diffusionmodules.model.Decoder
+class Decoder : public GGMLBlock {
+protected:
+    int ch                   = 128;
+    int out_ch               = 3;
+    std::vector<int> ch_mult = {1, 2, 4, 4};
+    int num_res_blocks       = 2;
+    int z_channels           = 4;
+    bool video_decoder       = false;
+    int video_kernel_size    = 3;
+
+    virtual std::shared_ptr<GGMLBlock> get_conv_out(int64_t in_channels,
+                                                    int64_t out_channels,
+                                                    std::pair<int, int> kernel_size,
+                                                    std::pair<int, int> stride  = {1, 1},
+                                                    std::pair<int, int> padding = {0, 0}) {
+        if (video_decoder) {
+            return std::shared_ptr<GGMLBlock>(new AE3DConv(in_channels, out_channels, kernel_size, video_kernel_size, stride, padding));
+        } else {
+            return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, stride, padding));
+        }
+    }
+
+    virtual std::shared_ptr<GGMLBlock> get_resnet_block(int64_t in_channels,
+                                                        int64_t out_channels) {
+        if (video_decoder) {
+            return std::shared_ptr<GGMLBlock>(new VideoResnetBlock(in_channels, out_channels, video_kernel_size));
+        } else {
+            return std::shared_ptr<GGMLBlock>(new ResnetBlock(in_channels, out_channels));
+        }
+    }
+
+public:
+    Decoder(int ch,
+            int out_ch,
+            std::vector<int> ch_mult,
+            int num_res_blocks,
+            int z_channels,
+            bool use_linear_projection = false,
+            bool video_decoder         = false,
+            int video_kernel_size      = 3)
+        : ch(ch),
+          out_ch(out_ch),
+          ch_mult(ch_mult),
+          num_res_blocks(num_res_blocks),
+          z_channels(z_channels),
+          video_decoder(video_decoder),
+          video_kernel_size(video_kernel_size) {
+        int num_resolutions = static_cast<int>(ch_mult.size());
+        int block_in        = ch * ch_mult[num_resolutions - 1];
+
+        blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, block_in, {3, 3}, {1, 1}, {1, 1}));
+
+        blocks["mid.block_1"] = get_resnet_block(block_in, block_in);
+        blocks["mid.attn_1"]  = std::shared_ptr<GGMLBlock>(new AttnBlock(block_in, use_linear_projection));
+        blocks["mid.block_2"] = get_resnet_block(block_in, block_in);
+
+        for (int i = num_resolutions - 1; i >= 0; i--) {
+            int mult      = ch_mult[i];
+            int block_out = ch * mult;
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
+                blocks[name]     = get_resnet_block(block_in, block_out);
+
+                block_in = block_out;
+            }
+            if (i != 0) {
+                std::string name = "up." + std::to_string(i) + ".upsample";
+                blocks[name]     = std::shared_ptr<GGMLBlock>(new UpSampleBlock(block_in, block_in));
+            }
+        }
+
+        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(block_in));
+        blocks["conv_out"] = get_conv_out(block_in, out_ch, {3, 3}, {1, 1}, {1, 1});
+    }
+
+    virtual ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) {
+        // z: [N, z_channels, h, w]
+        // alpha is always 0
+        // merge_strategy is always learned
+        // time_mode is always conv-only, so we need to replace conv_out_op/resnet_op to AE3DConv/VideoResBlock
+        // AttnVideoBlock will not be used
+        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
+        auto mid_block_1 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_1"]);
+        auto mid_attn_1  = std::dynamic_pointer_cast<AttnBlock>(blocks["mid.attn_1"]);
+        auto mid_block_2 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_2"]);
+        auto norm_out    = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm_out"]);
+        auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
+
+        // conv_in
+        auto h = conv_in->forward(ctx, z);  // [N, block_in, h, w]
+
+        // middle
+        h = mid_block_1->forward(ctx, h);
+        // return h;
+
+        h = mid_attn_1->forward(ctx, h);
+        h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
+
+        // upsampling
+        int num_resolutions = static_cast<int>(ch_mult.size());
+        for (int i = num_resolutions - 1; i >= 0; i--) {
+            for (int j = 0; j < num_res_blocks + 1; j++) {
+                std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
+                auto up_block    = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
+
+                h = up_block->forward(ctx, 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);
+            }
+        }
+
+        h = norm_out->forward(ctx, h);
+        h = ggml_silu_inplace(ctx->ggml_ctx, h);  // nonlinearity/swish
+        h = conv_out->forward(ctx, h);            // [N, out_ch, h*8, w*8]
+        return h;
+    }
+};
+
+// ldm.models.autoencoder.AutoencoderKL
+class AutoEncoderKLModel : public GGMLBlock {
+protected:
+    SDVersion version;
+    bool decode_only       = true;
+    bool use_video_decoder = false;
+    bool use_quant         = true;
+    int embed_dim          = 4;
+    struct {
+        int z_channels           = 4;
+        int resolution           = 256;
+        int in_channels          = 3;
+        int out_ch               = 3;
+        int ch                   = 128;
+        std::vector<int> ch_mult = {1, 2, 4, 4};
+        int num_res_blocks       = 2;
+        bool double_z            = true;
+    } dd_config;
+
+public:
+    AutoEncoderKLModel(SDVersion version          = VERSION_SD1,
+                       bool decode_only           = true,
+                       bool use_linear_projection = false,
+                       bool use_video_decoder     = false)
+        : version(version), decode_only(decode_only), use_video_decoder(use_video_decoder) {
+        if (sd_version_is_dit(version)) {
+            if (sd_version_is_flux2(version)) {
+                dd_config.z_channels = 32;
+                embed_dim            = 32;
+            } else {
+                use_quant            = false;
+                dd_config.z_channels = 16;
+            }
+        }
+        if (use_video_decoder) {
+            use_quant = false;
+        }
+        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder(dd_config.ch,
+                                                                   dd_config.out_ch,
+                                                                   dd_config.ch_mult,
+                                                                   dd_config.num_res_blocks,
+                                                                   dd_config.z_channels,
+                                                                   use_linear_projection,
+                                                                   use_video_decoder));
+        if (use_quant) {
+            blocks["post_quant_conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(dd_config.z_channels,
+                                                                              embed_dim,
+                                                                              {1, 1}));
+        }
+        if (!decode_only) {
+            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new Encoder(dd_config.ch,
+                                                                       dd_config.ch_mult,
+                                                                       dd_config.num_res_blocks,
+                                                                       dd_config.in_channels,
+                                                                       dd_config.z_channels,
+                                                                       dd_config.double_z,
+                                                                       use_linear_projection));
+            if (use_quant) {
+                int factor = dd_config.double_z ? 2 : 1;
+
+                blocks["quant_conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(embed_dim * factor,
+                                                                             dd_config.z_channels * factor,
+                                                                             {1, 1}));
+            }
+        }
+    }
+
+    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
+        // z: [N, z_channels, h, w]
+        if (sd_version_is_flux2(version)) {
+            // [N, C*p*p, h, w] -> [N, C, h*p, w*p]
+            int64_t p = 2;
+
+            int64_t N = z->ne[3];
+            int64_t C = z->ne[2] / p / p;
+            int64_t h = z->ne[1];
+            int64_t w = z->ne[0];
+            int64_t H = h * p;
+            int64_t W = w * p;
+
+            z = ggml_reshape_4d(ctx->ggml_ctx, z, w * h, p * p, C, N);                           // [N, C, p*p, h*w]
+            z = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, z, 1, 0, 2, 3));  // [N, C, h*w, p*p]
+            z = ggml_reshape_4d(ctx->ggml_ctx, z, p, p, w, h * C * N);                           // [N*C*h, w, p, p]
+            z = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, z, 0, 2, 1, 3));  // [N*C*h, p, w, p]
+            z = ggml_reshape_4d(ctx->ggml_ctx, z, W, H, C, N);                                   // [N, C, h*p, w*p]
+        }
+
+        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]
+        }
+        auto decoder = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
+
+        ggml_set_name(z, "bench-start");
+        auto h = decoder->forward(ctx, z);
+        ggml_set_name(h, "bench-end");
+        return h;
+    }
+
+    ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
+        // x: [N, in_channels, h, w]
+        auto encoder = std::dynamic_pointer_cast<Encoder>(blocks["encoder"]);
+
+        auto z = encoder->forward(ctx, x);  // [N, 2*z_channels, h/8, w/8]
+        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]
+        }
+        if (sd_version_is_flux2(version)) {
+            z = ggml_ext_chunk(ctx->ggml_ctx, z, 2, 2)[0];
+
+            // [N, C, H, W] -> [N, C*p*p, H/p, W/p]
+            int64_t p = 2;
+            int64_t N = z->ne[3];
+            int64_t C = z->ne[2];
+            int64_t H = z->ne[1];
+            int64_t W = z->ne[0];
+            int64_t h = H / p;
+            int64_t w = W / p;
+
+            z = ggml_reshape_4d(ctx->ggml_ctx, z, p, w, p, h * C * N);                 // [N*C*h, p, w, p]
+            z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 0, 2, 1, 3));  // [N*C*h, w, p, p]
+            z = ggml_reshape_4d(ctx->ggml_ctx, z, p * p, w * h, C, N);                 // [N, C, h*w, p*p]
+            z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 1, 0, 2, 3));  // [N, C, p*p, h*w]
+            z = ggml_reshape_4d(ctx->ggml_ctx, z, w, h, p * p * C, N);                 // [N, C*p*p, h*w]
+        }
+        return z;
+    }
+
+    int get_encoder_output_channels() {
+        int factor = dd_config.double_z ? 2 : 1;
+        if (sd_version_is_flux2(version)) {
+            return dd_config.z_channels * 4;
+        }
+        return dd_config.z_channels * factor;
+    }
+};
+
+struct AutoEncoderKL : public VAE {
+    float scale_factor = 1.f;
+    float shift_factor = 0.f;
+    bool decode_only   = true;
+    AutoEncoderKLModel ae;
+
+    AutoEncoderKL(ggml_backend_t backend,
+                  bool offload_params_to_cpu,
+                  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) {
+        if (sd_version_is_sd1(version) || sd_version_is_sd2(version)) {
+            scale_factor = 0.18215f;
+            shift_factor = 0.f;
+        } else if (sd_version_is_sdxl(version)) {
+            scale_factor = 0.13025f;
+            shift_factor = 0.f;
+        } else if (sd_version_is_sd3(version)) {
+            scale_factor = 1.5305f;
+            shift_factor = 0.0609f;
+        } else if (sd_version_is_flux(version) || sd_version_is_z_image(version)) {
+            scale_factor = 0.3611f;
+            shift_factor = 0.1159f;
+        } else if (sd_version_is_flux2(version)) {
+            scale_factor = 1.0f;
+            shift_factor = 0.f;
+        }
+        bool use_linear_projection = false;
+        for (const auto& [name, tensor_storage] : tensor_storage_map) {
+            if (!starts_with(name, prefix)) {
+                continue;
+            }
+            if (ends_with(name, "attn_1.proj_out.weight")) {
+                if (tensor_storage.n_dims == 2) {
+                    use_linear_projection = true;
+                }
+                break;
+            }
+        }
+        ae = AutoEncoderKLModel(version, decode_only, use_linear_projection, use_video_decoder);
+        ae.init(params_ctx, tensor_storage_map, prefix);
+    }
+
+    void set_conv2d_scale(float scale) override {
+        std::vector<GGMLBlock*> blocks;
+        ae.get_all_blocks(blocks);
+        for (auto block : blocks) {
+            if (block->get_desc() == "Conv2d") {
+                auto conv_block = (Conv2d*)block;
+                conv_block->set_scale(scale);
+            }
+        }
+    }
+
+    std::string get_desc() override {
+        return "vae";
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) override {
+        ae.get_param_tensors(tensors, prefix);
+    }
+
+    ggml_cgraph* build_graph(ggml_tensor* z, bool decode_graph) {
+        ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+
+        z = to_backend(z);
+
+        auto runner_ctx = get_context();
+
+        ggml_tensor* out = decode_graph ? ae.decode(&runner_ctx, z) : ae.encode(&runner_ctx, z);
+
+        ggml_build_forward_expand(gf, out);
+
+        return gf;
+    }
+
+    bool _compute(const int n_threads,
+                  ggml_tensor* z,
+                  bool decode_graph,
+                  ggml_tensor** output,
+                  ggml_context* output_ctx = nullptr) override {
+        GGML_ASSERT(!decode_only || decode_graph);
+        auto get_graph = [&]() -> ggml_cgraph* {
+            return build_graph(z, decode_graph);
+        };
+        // ggml_set_f32(z, 0.5f);
+        // print_ggml_tensor(z);
+        return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
+    }
+
+    ggml_tensor* gaussian_latent_sample(ggml_context* work_ctx, ggml_tensor* moments, std::shared_ptr<RNG> rng) {
+        // ldm.modules.distributions.distributions.DiagonalGaussianDistribution.sample
+        ggml_tensor* latents = ggml_new_tensor_4d(work_ctx, moments->type, moments->ne[0], moments->ne[1], moments->ne[2] / 2, moments->ne[3]);
+        ggml_tensor* noise   = ggml_dup_tensor(work_ctx, latents);
+        ggml_ext_im_set_randn_f32(noise, rng);
+        {
+            float mean   = 0;
+            float logvar = 0;
+            float value  = 0;
+            float std_   = 0;
+            for (int i = 0; i < latents->ne[3]; i++) {
+                for (int j = 0; j < latents->ne[2]; j++) {
+                    for (int k = 0; k < latents->ne[1]; k++) {
+                        for (int l = 0; l < latents->ne[0]; l++) {
+                            mean   = ggml_ext_tensor_get_f32(moments, l, k, j, i);
+                            logvar = ggml_ext_tensor_get_f32(moments, l, k, j + (int)latents->ne[2], i);
+                            logvar = std::max(-30.0f, std::min(logvar, 20.0f));
+                            std_   = std::exp(0.5f * logvar);
+                            value  = mean + std_ * ggml_ext_tensor_get_f32(noise, l, k, j, i);
+                            // printf("%d %d %d %d -> %f\n", i, j, k, l, value);
+                            ggml_ext_tensor_set_f32(latents, value, l, k, j, i);
+                        }
+                    }
+                }
+            }
+        }
+        return latents;
+    }
+
+    ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) {
+        if (sd_version_is_flux2(version)) {
+            return vae_output;
+        } else if (version == VERSION_SD1_PIX2PIX) {
+            return ggml_view_3d(work_ctx,
+                                vae_output,
+                                vae_output->ne[0],
+                                vae_output->ne[1],
+                                vae_output->ne[2] / 2,
+                                vae_output->nb[1],
+                                vae_output->nb[2],
+                                0);
+        } else {
+            return gaussian_latent_sample(work_ctx, vae_output, rng);
+        }
+    }
+
+    void get_latents_mean_std_vec(ggml_tensor* latents, int channel_dim, std::vector<float>& latents_mean_vec, std::vector<float>& latents_std_vec) {
+        // flux2
+        if (sd_version_is_flux2(version)) {
+            GGML_ASSERT(latents->ne[channel_dim] == 128);
+            latents_mean_vec = {-0.0676f, -0.0715f, -0.0753f, -0.0745f, 0.0223f, 0.0180f, 0.0142f, 0.0184f,
+                                -0.0001f, -0.0063f, -0.0002f, -0.0031f, -0.0272f, -0.0281f, -0.0276f, -0.0290f,
+                                -0.0769f, -0.0672f, -0.0902f, -0.0892f, 0.0168f, 0.0152f, 0.0079f, 0.0086f,
+                                0.0083f, 0.0015f, 0.0003f, -0.0043f, -0.0439f, -0.0419f, -0.0438f, -0.0431f,
+                                -0.0102f, -0.0132f, -0.0066f, -0.0048f, -0.0311f, -0.0306f, -0.0279f, -0.0180f,
+                                0.0030f, 0.0015f, 0.0126f, 0.0145f, 0.0347f, 0.0338f, 0.0337f, 0.0283f,
+                                0.0020f, 0.0047f, 0.0047f, 0.0050f, 0.0123f, 0.0081f, 0.0081f, 0.0146f,
+                                0.0681f, 0.0679f, 0.0767f, 0.0732f, -0.0462f, -0.0474f, -0.0392f, -0.0511f,
+                                -0.0528f, -0.0477f, -0.0470f, -0.0517f, -0.0317f, -0.0316f, -0.0345f, -0.0283f,
+                                0.0510f, 0.0445f, 0.0578f, 0.0458f, -0.0412f, -0.0458f, -0.0487f, -0.0467f,
+                                -0.0088f, -0.0106f, -0.0088f, -0.0046f, -0.0376f, -0.0432f, -0.0436f, -0.0499f,
+                                0.0118f, 0.0166f, 0.0203f, 0.0279f, 0.0113f, 0.0129f, 0.0016f, 0.0072f,
+                                -0.0118f, -0.0018f, -0.0141f, -0.0054f, -0.0091f, -0.0138f, -0.0145f, -0.0187f,
+                                0.0323f, 0.0305f, 0.0259f, 0.0300f, 0.0540f, 0.0614f, 0.0495f, 0.0590f,
+                                -0.0511f, -0.0603f, -0.0478f, -0.0524f, -0.0227f, -0.0274f, -0.0154f, -0.0255f,
+                                -0.0572f, -0.0565f, -0.0518f, -0.0496f, 0.0116f, 0.0054f, 0.0163f, 0.0104f};
+            latents_std_vec  = {
+                 1.8029f, 1.7786f, 1.7868f, 1.7837f, 1.7717f, 1.7590f, 1.7610f, 1.7479f,
+                 1.7336f, 1.7373f, 1.7340f, 1.7343f, 1.8626f, 1.8527f, 1.8629f, 1.8589f,
+                 1.7593f, 1.7526f, 1.7556f, 1.7583f, 1.7363f, 1.7400f, 1.7355f, 1.7394f,
+                 1.7342f, 1.7246f, 1.7392f, 1.7304f, 1.7551f, 1.7513f, 1.7559f, 1.7488f,
+                 1.8449f, 1.8454f, 1.8550f, 1.8535f, 1.8240f, 1.7813f, 1.7854f, 1.7945f,
+                 1.8047f, 1.7876f, 1.7695f, 1.7676f, 1.7782f, 1.7667f, 1.7925f, 1.7848f,
+                 1.7579f, 1.7407f, 1.7483f, 1.7368f, 1.7961f, 1.7998f, 1.7920f, 1.7925f,
+                 1.7780f, 1.7747f, 1.7727f, 1.7749f, 1.7526f, 1.7447f, 1.7657f, 1.7495f,
+                 1.7775f, 1.7720f, 1.7813f, 1.7813f, 1.8162f, 1.8013f, 1.8023f, 1.8033f,
+                 1.7527f, 1.7331f, 1.7563f, 1.7482f, 1.7610f, 1.7507f, 1.7681f, 1.7613f,
+                 1.7665f, 1.7545f, 1.7828f, 1.7726f, 1.7896f, 1.7999f, 1.7864f, 1.7760f,
+                 1.7613f, 1.7625f, 1.7560f, 1.7577f, 1.7783f, 1.7671f, 1.7810f, 1.7799f,
+                 1.7201f, 1.7068f, 1.7265f, 1.7091f, 1.7793f, 1.7578f, 1.7502f, 1.7455f,
+                 1.7587f, 1.7500f, 1.7525f, 1.7362f, 1.7616f, 1.7572f, 1.7444f, 1.7430f,
+                 1.7509f, 1.7610f, 1.7634f, 1.7612f, 1.7254f, 1.7135f, 1.7321f, 1.7226f,
+                 1.7664f, 1.7624f, 1.7718f, 1.7664f, 1.7457f, 1.7441f, 1.7569f, 1.7530f};
+        } else {
+            GGML_ABORT("unknown version %d", version);
+        }
+    }
+
+    ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+        ggml_tensor* vae_latents = ggml_dup(work_ctx, latents);
+        if (sd_version_is_flux2(version)) {
+            int channel_dim = 2;
+            std::vector<float> latents_mean_vec;
+            std::vector<float> latents_std_vec;
+            get_latents_mean_std_vec(latents, channel_dim, latents_mean_vec, latents_std_vec);
+
+            float mean;
+            float std_;
+            for (int i = 0; i < latents->ne[3]; i++) {
+                if (channel_dim == 3) {
+                    mean = latents_mean_vec[i];
+                    std_ = latents_std_vec[i];
+                }
+                for (int j = 0; j < latents->ne[2]; j++) {
+                    if (channel_dim == 2) {
+                        mean = latents_mean_vec[j];
+                        std_ = latents_std_vec[j];
+                    }
+                    for (int k = 0; k < latents->ne[1]; k++) {
+                        for (int l = 0; l < latents->ne[0]; l++) {
+                            float value = ggml_ext_tensor_get_f32(latents, l, k, j, i);
+                            value       = value * std_ / scale_factor + mean;
+                            ggml_ext_tensor_set_f32(vae_latents, value, l, k, j, i);
+                        }
+                    }
+                }
+            }
+        } else {
+            ggml_ext_tensor_iter(latents, [&](ggml_tensor* latents, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+                float value = ggml_ext_tensor_get_f32(latents, i0, i1, i2, i3);
+                value       = (value / scale_factor) + shift_factor;
+                ggml_ext_tensor_set_f32(vae_latents, value, i0, i1, i2, i3);
+            });
+        }
+        return vae_latents;
+    }
+
+    ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+        ggml_tensor* diffusion_latents = ggml_dup(work_ctx, latents);
+        if (sd_version_is_flux2(version)) {
+            int channel_dim = 2;
+            std::vector<float> latents_mean_vec;
+            std::vector<float> latents_std_vec;
+            get_latents_mean_std_vec(latents, channel_dim, latents_mean_vec, latents_std_vec);
+
+            float mean;
+            float std_;
+            for (int i = 0; i < latents->ne[3]; i++) {
+                if (channel_dim == 3) {
+                    mean = latents_mean_vec[i];
+                    std_ = latents_std_vec[i];
+                }
+                for (int j = 0; j < latents->ne[2]; j++) {
+                    if (channel_dim == 2) {
+                        mean = latents_mean_vec[j];
+                        std_ = latents_std_vec[j];
+                    }
+                    for (int k = 0; k < latents->ne[1]; k++) {
+                        for (int l = 0; l < latents->ne[0]; l++) {
+                            float value = ggml_ext_tensor_get_f32(latents, l, k, j, i);
+                            value       = (value - mean) * scale_factor / std_;
+                            ggml_ext_tensor_set_f32(diffusion_latents, value, l, k, j, i);
+                        }
+                    }
+                }
+            }
+        } else {
+            ggml_ext_tensor_iter(latents, [&](ggml_tensor* latents, int64_t i0, int64_t i1, int64_t i2, int64_t i3) {
+                float value = ggml_ext_tensor_get_f32(latents, i0, i1, i2, i3);
+                value       = (value - shift_factor) * scale_factor;
+                ggml_ext_tensor_set_f32(diffusion_latents, value, i0, i1, i2, i3);
+            });
+        }
+        return diffusion_latents;
+    }
+
+    int get_encoder_output_channels(int input_channels) {
+        return ae.get_encoder_output_channels();
+    }
+
+    void test() {
+        ggml_init_params params;
+        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
+        params.mem_buffer = nullptr;
+        params.no_alloc   = false;
+
+        ggml_context* work_ctx = ggml_init(params);
+        GGML_ASSERT(work_ctx != nullptr);
+
+        {
+            // CPU, x{1, 3, 64, 64}: Pass
+            // CUDA, x{1, 3, 64, 64}: Pass, but sill get wrong result for some image, may be due to interlnal nan
+            // CPU, x{2, 3, 64, 64}: Wrong result
+            // CUDA, x{2, 3, 64, 64}: Wrong result, and different from CPU result
+            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 64, 64, 3, 2);
+            ggml_set_f32(x, 0.5f);
+            print_ggml_tensor(x);
+            ggml_tensor* out = nullptr;
+
+            int64_t t0 = ggml_time_ms();
+            _compute(8, x, false, &out, work_ctx);
+            int64_t t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("encode test done in %lldms", t1 - t0);
+        }
+
+        if (false) {
+            // CPU, z{1, 4, 8, 8}: Pass
+            // CUDA, z{1, 4, 8, 8}: Pass
+            // CPU, z{3, 4, 8, 8}: Wrong result
+            // CUDA, z{3, 4, 8, 8}: Wrong result, and different from CPU result
+            auto z = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 4, 1);
+            ggml_set_f32(z, 0.5f);
+            print_ggml_tensor(z);
+            ggml_tensor* out = nullptr;
+
+            int64_t t0 = ggml_time_ms();
+            _compute(8, z, true, &out, work_ctx);
+            int64_t t1 = ggml_time_ms();
+
+            print_ggml_tensor(out);
+            LOG_DEBUG("decode test done in %lldms", t1 - t0);
+        }
+    };
+};
+
+#endif  // __AUTO_ENCODER_KL_HPP__

src/cache_dit.hpp

@@ -603,87 +603,6 @@ inline std::vector<int> generate_scm_mask(
     return mask;
 }
 
-inline std::vector<int> get_scm_preset(const std::string& preset, int total_steps) {
-    struct Preset {
-        std::vector<int> compute_bins;
-        std::vector<int> cache_bins;
-    };
-
-    Preset slow   = {{8, 3, 3, 2, 1, 1}, {1, 2, 2, 2, 3}};
-    Preset medium = {{6, 2, 2, 2, 2, 1}, {1, 3, 3, 3, 3}};
-    Preset fast   = {{6, 1, 1, 1, 1, 1}, {1, 3, 4, 5, 4}};
-    Preset ultra  = {{4, 1, 1, 1, 1}, {2, 5, 6, 7}};
-
-    Preset* p = nullptr;
-    if (preset == "slow" || preset == "s" || preset == "S")
-        p = &slow;
-    else if (preset == "medium" || preset == "m" || preset == "M")
-        p = &medium;
-    else if (preset == "fast" || preset == "f" || preset == "F")
-        p = &fast;
-    else if (preset == "ultra" || preset == "u" || preset == "U")
-        p = &ultra;
-    else
-        return {};
-
-    if (total_steps != 28 && total_steps > 0) {
-        float scale = static_cast<float>(total_steps) / 28.0f;
-        std::vector<int> scaled_compute, scaled_cache;
-
-        for (int v : p->compute_bins) {
-            scaled_compute.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
-        }
-        for (int v : p->cache_bins) {
-            scaled_cache.push_back(std::max(1, static_cast<int>(v * scale + 0.5f)));
-        }
-
-        return generate_scm_mask(scaled_compute, scaled_cache, total_steps);
-    }
-
-    return generate_scm_mask(p->compute_bins, p->cache_bins, total_steps);
-}
-
-inline float get_preset_threshold(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 0.20f;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 0.25f;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 0.30f;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 0.34f;
-    return 0.08f;
-}
-
-inline int get_preset_warmup(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 8;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 6;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 6;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 4;
-    return 8;
-}
-
-inline int get_preset_Fn(const std::string& preset) {
-    if (preset == "slow" || preset == "s" || preset == "S")
-        return 8;
-    if (preset == "medium" || preset == "m" || preset == "M")
-        return 8;
-    if (preset == "fast" || preset == "f" || preset == "F")
-        return 6;
-    if (preset == "ultra" || preset == "u" || preset == "U")
-        return 4;
-    return 8;
-}
-
-inline int get_preset_Bn(const std::string& preset) {
-    (void)preset;
-    return 0;
-}
-
 inline void parse_dbcache_options(const std::string& opts, DBCacheConfig& cfg) {
     if (opts.empty())
         return;
@@ -880,7 +799,7 @@ struct CacheDitConditionState {
         }
     }
 
-    bool before_condition(const void* cond, struct ggml_tensor* input, struct ggml_tensor* output, float sigma, int step_index) {
+    bool before_condition(const void* cond, ggml_tensor* input, ggml_tensor* output, float sigma, int step_index) {
         if (!enabled() || step_index < 0)
             return false;
 
@@ -948,7 +867,7 @@ struct CacheDitConditionState {
         return false;
     }
 
-    void after_condition(const void* cond, struct ggml_tensor* input, struct ggml_tensor* output) {
+    void after_condition(const void* cond, ggml_tensor* input, ggml_tensor* output) {
         if (!step_is_active())
             return;
 

src/clip.hpp

@@ -473,7 +473,7 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [N, n_token, d_model]
         auto fc1 = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
         auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
@@ -511,7 +511,7 @@ public:
         blocks["mlp"] = std::shared_ptr<GGMLBlock>(new CLIPMLP(d_model, intermediate_size));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, struct ggml_tensor* mask = nullptr) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* mask = nullptr) {
         // x: [N, n_token, d_model]
         auto self_attn   = std::dynamic_pointer_cast<MultiheadAttention>(blocks["self_attn"]);
         auto layer_norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm1"]);
@@ -541,10 +541,10 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* mask = nullptr,
+                         ggml_tensor* mask = nullptr,
-                                int clip_skip            = -1) {
+                         int clip_skip     = -1) {
         // x: [N, n_token, d_model]
         int layer_idx = n_layer - 1;
         // LOG_DEBUG("clip_skip %d", clip_skip);
@@ -573,7 +573,7 @@ protected:
     int64_t num_positions;
     bool force_clip_f32;
 
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
         enum ggml_type token_wtype = GGML_TYPE_F32;
         if (!force_clip_f32) {
             token_wtype = get_type(prefix + "token_embedding.weight", tensor_storage_map, GGML_TYPE_F32);
@@ -597,13 +597,13 @@ public:
           force_clip_f32(force_clip_f32) {
     }
 
-    struct ggml_tensor* get_token_embed_weight() {
+    ggml_tensor* get_token_embed_weight() {
         return params["token_embedding.weight"];
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* input_ids,
+                         ggml_tensor* input_ids,
-                                struct ggml_tensor* custom_embed_weight) {
+                         ggml_tensor* custom_embed_weight) {
         // input_ids: [N, n_token]
         auto token_embed_weight    = params["token_embedding.weight"];
         auto position_embed_weight = params["position_embedding.weight"];
@@ -630,7 +630,7 @@ protected:
     int num_patches;
     int64_t num_positions;
 
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
         enum ggml_type patch_wtype    = GGML_TYPE_F16;
         enum ggml_type class_wtype    = GGML_TYPE_F32;
         enum ggml_type position_wtype = GGML_TYPE_F32;
@@ -653,7 +653,7 @@ public:
         num_positions = num_patches + 1;
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* pixel_values) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* pixel_values) {
         // pixel_values: [N, num_channels, image_size, image_size]
         // return: [N, num_positions, embed_dim]
         GGML_ASSERT(pixel_values->ne[0] == image_size && pixel_values->ne[1] == image_size && pixel_values->ne[2] == num_channels);
@@ -663,20 +663,20 @@ public:
         auto position_embed_weight = params["position_embedding.weight"];
 
         // concat(patch_embedding, class_embedding) + position_embedding
-        struct ggml_tensor* patch_embedding;
+        ggml_tensor* patch_embedding;
         int64_t N       = pixel_values->ne[3];
         patch_embedding = ggml_ext_conv_2d(ctx->ggml_ctx, pixel_values, patch_embed_weight, nullptr, patch_size, patch_size);  // [N, embed_dim, image_size // pacht_size, image_size // pacht_size]
         patch_embedding = ggml_reshape_3d(ctx->ggml_ctx, patch_embedding, num_patches, embed_dim, N);                          // [N, embed_dim, num_patches]
         patch_embedding = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, patch_embedding, 1, 0, 2, 3));                  // [N, num_patches, embed_dim]
         patch_embedding = ggml_reshape_4d(ctx->ggml_ctx, patch_embedding, 1, embed_dim, num_patches, N);                       // [N, num_patches, embed_dim, 1]
 
-        struct ggml_tensor* class_embedding = ggml_new_tensor_2d(ctx->ggml_ctx, GGML_TYPE_F32, embed_dim, N);
+        ggml_tensor* class_embedding = ggml_new_tensor_2d(ctx->ggml_ctx, GGML_TYPE_F32, embed_dim, N);
-        class_embedding                     = ggml_repeat(ctx->ggml_ctx, class_embed_weight, class_embedding);      // [N, embed_dim]
+        class_embedding              = ggml_repeat(ctx->ggml_ctx, class_embed_weight, class_embedding);      // [N, embed_dim]
-        class_embedding                     = ggml_reshape_4d(ctx->ggml_ctx, class_embedding, 1, embed_dim, 1, N);  // [N, 1, embed_dim, 1]
+        class_embedding              = ggml_reshape_4d(ctx->ggml_ctx, class_embedding, 1, embed_dim, 1, N);  // [N, 1, embed_dim, 1]
 
-        struct ggml_tensor* x = ggml_concat(ctx->ggml_ctx, class_embedding, patch_embedding, 2);  // [N, num_positions, embed_dim, 1]
+        ggml_tensor* x = ggml_concat(ctx->ggml_ctx, class_embedding, patch_embedding, 2);  // [N, num_positions, embed_dim, 1]
-        x                     = ggml_reshape_3d(ctx->ggml_ctx, x, embed_dim, num_positions, N);   // [N, num_positions, embed_dim]
+        x              = ggml_reshape_3d(ctx->ggml_ctx, x, embed_dim, num_positions, N);   // [N, num_positions, embed_dim]
-        x                     = ggml_add(ctx->ggml_ctx, x, position_embed_weight);
+        x              = ggml_add(ctx->ggml_ctx, x, position_embed_weight);
         return x;  // [N, num_positions, embed_dim]
     }
 };
@@ -693,7 +693,7 @@ enum CLIPVersion {
 
 class CLIPTextModel : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
         if (version == OPEN_CLIP_VIT_BIGG_14) {
             enum ggml_type wtype      = GGML_TYPE_F32;
             params["text_projection"] = ggml_new_tensor_2d(ctx, wtype, projection_dim, hidden_size);
@@ -734,18 +734,18 @@ public:
         blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
     }
 
-    struct ggml_tensor* get_token_embed_weight() {
+    ggml_tensor* get_token_embed_weight() {
         auto embeddings = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
         return embeddings->get_token_embed_weight();
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* input_ids,
+                         ggml_tensor* input_ids,
-                                struct ggml_tensor* tkn_embeddings,
+                         ggml_tensor* tkn_embeddings,
-                                struct ggml_tensor* mask = nullptr,
+                         ggml_tensor* mask    = nullptr,
-                                size_t max_token_idx     = 0,
+                         size_t max_token_idx = 0,
-                                bool return_pooled       = false,
+                         bool return_pooled   = false,
-                                int clip_skip            = -1) {
+                         int clip_skip        = -1) {
         // input_ids: [N, n_token]
         auto embeddings       = std::dynamic_pointer_cast<CLIPEmbeddings>(blocks["embeddings"]);
         auto encoder          = std::dynamic_pointer_cast<CLIPEncoder>(blocks["encoder"]);
@@ -804,10 +804,10 @@ public:
         blocks["post_layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* pixel_values,
+                         ggml_tensor* pixel_values,
-                                bool return_pooled = true,
+                         bool return_pooled = true,
-                                int clip_skip      = -1) {
+                         int clip_skip      = -1) {
         // pixel_values: [N, num_channels, image_size, image_size]
         auto embeddings     = std::dynamic_pointer_cast<CLIPVisionEmbeddings>(blocks["embeddings"]);
         auto pre_layernorm  = std::dynamic_pointer_cast<LayerNorm>(blocks["pre_layernorm"]);
@@ -839,7 +839,7 @@ protected:
     int64_t out_features;
     bool transpose_weight;
 
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
         enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32);
         if (transpose_weight) {
             params["weight"] = ggml_new_tensor_2d(ctx, wtype, out_features, in_features);
@@ -856,8 +856,8 @@ public:
           out_features(out_features),
           transpose_weight(transpose_weight) {}
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
-        struct ggml_tensor* w = params["weight"];
+        ggml_tensor* w = params["weight"];
         if (transpose_weight) {
             w = ggml_cont(ctx->ggml_ctx, ggml_transpose(ctx->ggml_ctx, w));
         }
@@ -886,10 +886,10 @@ public:
         blocks["visual_projection"] = std::shared_ptr<GGMLBlock>(new CLIPProjection(hidden_size, projection_dim, transpose_proj_w));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* pixel_values,
+                         ggml_tensor* pixel_values,
-                                bool return_pooled = true,
+                         bool return_pooled = true,
-                                int clip_skip      = -1) {
+                         int clip_skip      = -1) {
         // pixel_values: [N, num_channels, image_size, image_size]
         // return: [N, projection_dim] if return_pooled else [N, n_token, hidden_size]
         auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
@@ -936,17 +936,17 @@ struct CLIPTextModelRunner : public GGMLRunner {
         return "clip";
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
         model.get_param_tensors(tensors, prefix);
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* input_ids,
+                         ggml_tensor* input_ids,
-                                struct ggml_tensor* embeddings,
+                         ggml_tensor* embeddings,
-                                struct ggml_tensor* mask,
+                         ggml_tensor* mask,
-                                size_t max_token_idx = 0,
+                         size_t max_token_idx = 0,
-                                bool return_pooled   = false,
+                         bool return_pooled   = false,
-                                int clip_skip        = -1) {
+                         int clip_skip        = -1) {
         size_t N       = input_ids->ne[1];
         size_t n_token = input_ids->ne[0];
         if (input_ids->ne[0] > model.n_token) {
@@ -957,17 +957,17 @@ struct CLIPTextModelRunner : public GGMLRunner {
         return model.forward(ctx, input_ids, embeddings, mask, max_token_idx, return_pooled, clip_skip);
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
+    ggml_cgraph* build_graph(ggml_tensor* input_ids,
-                                    int num_custom_embeddings    = 0,
+                             int num_custom_embeddings    = 0,
-                                    void* custom_embeddings_data = nullptr,
+                             void* custom_embeddings_data = nullptr,
-                                    size_t max_token_idx         = 0,
+                             size_t max_token_idx         = 0,
-                                    bool return_pooled           = false,
+                             bool return_pooled           = false,
-                                    int clip_skip                = -1) {
+                             int clip_skip                = -1) {
-        struct ggml_cgraph* gf = new_graph_custom(2048);
+        ggml_cgraph* gf = new_graph_custom(2048);
 
         input_ids = to_backend(input_ids);
 
-        struct ggml_tensor* embeddings = nullptr;
+        ggml_tensor* embeddings = nullptr;
 
         if (num_custom_embeddings > 0 && custom_embeddings_data != nullptr) {
             auto token_embed_weight = model.get_token_embed_weight();
@@ -997,7 +997,7 @@ struct CLIPTextModelRunner : public GGMLRunner {
 
         auto runner_ctx = get_context();
 
-        struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, embeddings, attention_mask, max_token_idx, return_pooled, clip_skip);
+        ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, embeddings, attention_mask, max_token_idx, return_pooled, clip_skip);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -1005,7 +1005,7 @@ struct CLIPTextModelRunner : public GGMLRunner {
     }
 
     bool compute(const int n_threads,
-                 struct ggml_tensor* input_ids,
+                 ggml_tensor* input_ids,
                  int num_custom_embeddings,
                  void* custom_embeddings_data,
                  size_t max_token_idx,
@@ -1013,7 +1013,7 @@ struct CLIPTextModelRunner : public GGMLRunner {
                  int clip_skip,
                  ggml_tensor** output,
                  ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(input_ids, num_custom_embeddings, custom_embeddings_data, max_token_idx, return_pooled, clip_skip);
         };
         return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);

src/common_block.hpp

@@ -23,7 +23,7 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [N, channels, h, w]
         if (vae_downsample) {
             auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
@@ -52,7 +52,7 @@ public:
         blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [N, channels, h, w]
         auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
 
@@ -121,7 +121,7 @@ public:
         }
     }
 
-    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, struct ggml_tensor* emb = nullptr) {
+    virtual ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* emb = nullptr) {
         // For dims==3, we reduce dimension from 5d to 4d by merging h and w, in order not to change ggml
         // [N, c, t, h, w] => [N, c, t, h * w]
         // x: [N, channels, h, w] if dims == 2 else [N, channels, t, h, w]
@@ -188,7 +188,7 @@ public:
         blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out * 2));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         // x: [ne3, ne2, ne1, dim_in]
         // return: [ne3, ne2, ne1, dim_out]
         auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
@@ -214,7 +214,7 @@ public:
         blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(dim_in, dim_out, bias));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         // x: [ne3, ne2, ne1, dim_in]
         // return: [ne3, ne2, ne1, dim_out]
         auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
@@ -258,7 +258,7 @@ public:
         blocks["net.2"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim_out, true, false, force_prec_f32, scale));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [ne3, ne2, ne1, dim]
         // return: [ne3, ne2, ne1, dim_out]
 
@@ -297,9 +297,9 @@ public:
         // to_out_1 is nn.Dropout(), skip for inference
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* context) {
+                         ggml_tensor* context) {
         // x: [N, n_token, query_dim]
         // context: [N, n_context, context_dim]
         // return: [N, n_token, query_dim]
@@ -355,9 +355,9 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* context) {
+                         ggml_tensor* context) {
         // x: [N, n_token, query_dim]
         // context: [N, n_context, context_dim]
         // return: [N, n_token, query_dim]
@@ -406,7 +406,7 @@ protected:
     int64_t context_dim = 768;  // hidden_size, 1024 for VERSION_SD2
     bool use_linear     = false;
 
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
         auto iter = tensor_storage_map.find(prefix + "proj_out.weight");
         if (iter != tensor_storage_map.end()) {
             int64_t inner_dim = n_head * d_head;
@@ -456,9 +456,9 @@ public:
         }
     }
 
-    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    virtual ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                        struct ggml_tensor* x,
+                                 ggml_tensor* x,
-                                        struct ggml_tensor* context) {
+                                 ggml_tensor* context) {
         // x: [N, in_channels, h, w]
         // context: [N, max_position(aka n_token), hidden_size(aka context_dim)]
         auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
@@ -510,7 +510,7 @@ public:
 
 class AlphaBlender : public GGMLBlock {
 protected:
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
         // Get the type of the "mix_factor" tensor from the input tensors map with the specified prefix
         enum ggml_type wtype = GGML_TYPE_F32;
         params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
@@ -530,9 +530,9 @@ public:
         // since mix_factor.shape is [1,], we don't need rearrange using  rearrange_pattern
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x_spatial,
+                         ggml_tensor* x_spatial,
-                                struct ggml_tensor* x_temporal) {
+                         ggml_tensor* x_temporal) {
         // image_only_indicator is always tensor([0.])
         float alpha = get_alpha();
         auto x      = ggml_add(ctx->ggml_ctx,
@@ -555,10 +555,10 @@ public:
         blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* emb,
+                         ggml_tensor* emb,
-                                int num_video_frames) {
+                         int num_video_frames) {
         // x: [N, channels, h, w] aka [b*t, channels, h, w]
         // emb: [N, emb_channels] aka [b*t, emb_channels]
         // image_only_indicator is always tensor([0.])

src/conditioner.hpp

@@ -6,17 +6,17 @@
 #include "t5.hpp"
 
 struct SDCondition {
-    struct ggml_tensor* c_crossattn = nullptr;  // aka context
+    ggml_tensor* c_crossattn = nullptr;  // aka context
-    struct ggml_tensor* c_vector    = nullptr;  // aka y
+    ggml_tensor* c_vector    = nullptr;  // aka y
-    struct ggml_tensor* c_concat    = nullptr;
+    ggml_tensor* c_concat    = nullptr;
 
-    std::vector<struct ggml_tensor*> extra_c_crossattns;
+    std::vector<ggml_tensor*> extra_c_crossattns;
 
     SDCondition() = default;
-    SDCondition(struct ggml_tensor* c_crossattn,
+    SDCondition(ggml_tensor* c_crossattn,
-                struct ggml_tensor* c_vector,
+                ggml_tensor* c_vector,
-                struct ggml_tensor* c_concat,
+                ggml_tensor* c_concat,
-                const std::vector<struct ggml_tensor*>& extra_c_crossattns = {})
+                const std::vector<ggml_tensor*>& extra_c_crossattns = {})
         : c_crossattn(c_crossattn), c_vector(c_vector), c_concat(c_concat), extra_c_crossattns(extra_c_crossattns) {}
 };
 
@@ -37,7 +37,7 @@ struct Conditioner {
                                               const ConditionerParams& conditioner_params) = 0;
     virtual void alloc_params_buffer()                                                     = 0;
     virtual void free_params_buffer()                                                      = 0;
-    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors)    = 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_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) {}
@@ -92,7 +92,7 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
         }
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         text_model->get_param_tensors(tensors, "cond_stage_model.transformer.text_model");
         if (sd_version_is_sdxl(version)) {
             text_model2->get_param_tensors(tensors, "cond_stage_model.1.transformer.text_model");
@@ -149,14 +149,14 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             }
             return true;
         }
-        struct ggml_init_params params;
+        ggml_init_params params;
-        params.mem_size               = 100 * 1024 * 1024;  // max for custom embeddings 100 MB
+        params.mem_size        = 100 * 1024 * 1024;  // max for custom embeddings 100 MB
-        params.mem_buffer             = nullptr;
+        params.mem_buffer      = nullptr;
-        params.no_alloc               = false;
+        params.no_alloc        = false;
-        struct ggml_context* embd_ctx = ggml_init(params);
+        ggml_context* embd_ctx = ggml_init(params);
-        struct ggml_tensor* embd      = nullptr;
+        ggml_tensor* embd      = nullptr;
-        struct ggml_tensor* embd2     = nullptr;
+        ggml_tensor* embd2     = nullptr;
-        auto on_load                  = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) {
+        auto on_load           = [&](const TensorStorage& tensor_storage, ggml_tensor** dst_tensor) {
             if (tensor_storage.ne[0] != text_model->model.hidden_size) {
                 if (text_model2) {
                     if (tensor_storage.ne[0] == text_model2->model.hidden_size) {
@@ -435,12 +435,12 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                              int height,
                                              int adm_in_channels  = -1,
                                              bool zero_out_masked = false) {
-        int64_t t0                               = ggml_time_ms();
+        int64_t t0                        = ggml_time_ms();
-        struct ggml_tensor* hidden_states        = nullptr;  // [N, n_token, hidden_size]
+        ggml_tensor* hidden_states        = nullptr;  // [N, n_token, hidden_size]
-        struct ggml_tensor* chunk_hidden_states  = nullptr;  // [n_token, hidden_size] or [n_token, hidden_size + hidden_size2]
+        ggml_tensor* chunk_hidden_states  = nullptr;  // [n_token, hidden_size] or [n_token, hidden_size + hidden_size2]
-        struct ggml_tensor* chunk_hidden_states1 = nullptr;  // [n_token, hidden_size]
+        ggml_tensor* chunk_hidden_states1 = nullptr;  // [n_token, hidden_size]
-        struct ggml_tensor* chunk_hidden_states2 = nullptr;  // [n_token, hidden_size2]
+        ggml_tensor* chunk_hidden_states2 = nullptr;  // [n_token, hidden_size2]
-        struct ggml_tensor* pooled               = nullptr;
+        ggml_tensor* pooled               = nullptr;
         std::vector<float> hidden_states_vec;
 
         if (clip_skip <= 0) {
@@ -455,9 +455,9 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
             std::vector<float> chunk_weights(weights.begin() + chunk_idx * chunk_len,
                                              weights.begin() + (chunk_idx + 1) * chunk_len);
 
-            auto input_ids                 = vector_to_ggml_tensor_i32(work_ctx, chunk_tokens);
+            auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, chunk_tokens);
-            struct ggml_tensor* input_ids2 = nullptr;
+            ggml_tensor* input_ids2 = nullptr;
-            size_t max_token_idx           = 0;
+            size_t max_token_idx    = 0;
             if (sd_version_is_sdxl(version)) {
                 auto it = std::find(chunk_tokens.begin(), chunk_tokens.end(), tokenizer.EOS_TOKEN_ID);
                 if (it != chunk_tokens.end()) {
@@ -676,18 +676,18 @@ struct FrozenCLIPVisionEmbedder : public GGMLRunner {
         return "clip_vision";
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) {
         vision_model.get_param_tensors(tensors, "cond_stage_model.transformer");
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* pixel_values, bool return_pooled, int clip_skip) {
+    ggml_cgraph* build_graph(ggml_tensor* pixel_values, bool return_pooled, int clip_skip) {
-        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+        ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         pixel_values = to_backend(pixel_values);
 
         auto runner_ctx = get_context();
 
-        struct ggml_tensor* hidden_states = vision_model.forward(&runner_ctx, pixel_values, return_pooled, clip_skip);
+        ggml_tensor* hidden_states = vision_model.forward(&runner_ctx, pixel_values, return_pooled, clip_skip);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -700,7 +700,7 @@ struct FrozenCLIPVisionEmbedder : public GGMLRunner {
                  int clip_skip,
                  ggml_tensor** output,
                  ggml_context* output_ctx) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(pixel_values, return_pooled, clip_skip);
         };
         return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
@@ -746,7 +746,7 @@ struct SD3CLIPEmbedder : public Conditioner {
         }
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         if (clip_l) {
             clip_l->get_param_tensors(tensors, "text_encoders.clip_l.transformer.text_model");
         }
@@ -909,15 +909,15 @@ struct SD3CLIPEmbedder : public Conditioner {
             clip_skip = 2;
         }
 
-        int64_t t0                                 = ggml_time_ms();
+        int64_t t0                          = ggml_time_ms();
-        struct ggml_tensor* hidden_states          = nullptr;  // [N, n_token*2, 4096]
+        ggml_tensor* hidden_states          = nullptr;  // [N, n_token*2, 4096]
-        struct ggml_tensor* chunk_hidden_states    = nullptr;  // [n_token*2, 4096]
+        ggml_tensor* chunk_hidden_states    = nullptr;  // [n_token*2, 4096]
-        struct ggml_tensor* chunk_hidden_states_l  = nullptr;  // [n_token, hidden_size_l]
+        ggml_tensor* chunk_hidden_states_l  = nullptr;  // [n_token, hidden_size_l]
-        struct ggml_tensor* chunk_hidden_states_g  = nullptr;  // [n_token, hidden_size_g]
+        ggml_tensor* chunk_hidden_states_g  = nullptr;  // [n_token, hidden_size_g]
-        struct ggml_tensor* chunk_hidden_states_t5 = nullptr;  // [n_token, hidden_size_t5]
+        ggml_tensor* chunk_hidden_states_t5 = nullptr;  // [n_token, hidden_size_t5]
-        struct ggml_tensor* pooled                 = nullptr;
+        ggml_tensor* pooled                 = nullptr;
-        struct ggml_tensor* pooled_l               = nullptr;  // [768,]
+        ggml_tensor* pooled_l               = nullptr;  // [768,]
-        struct ggml_tensor* pooled_g               = nullptr;  // [1280,]
+        ggml_tensor* pooled_g               = nullptr;  // [1280,]
         std::vector<float> hidden_states_vec;
 
         size_t chunk_len   = 77;
@@ -1178,7 +1178,7 @@ struct FluxCLIPEmbedder : public Conditioner {
         }
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         if (clip_l) {
             clip_l->get_param_tensors(tensors, "text_encoders.clip_l.transformer.text_model");
         }
@@ -1306,10 +1306,10 @@ struct FluxCLIPEmbedder : public Conditioner {
             clip_skip = 2;
         }
 
-        int64_t t0                              = ggml_time_ms();
+        int64_t t0                       = ggml_time_ms();
-        struct ggml_tensor* hidden_states       = nullptr;  // [N, n_token, 4096]
+        ggml_tensor* hidden_states       = nullptr;  // [N, n_token, 4096]
-        struct ggml_tensor* chunk_hidden_states = nullptr;  // [n_token, 4096]
+        ggml_tensor* chunk_hidden_states = nullptr;  // [n_token, 4096]
-        struct ggml_tensor* pooled              = nullptr;  // [768,]
+        ggml_tensor* pooled              = nullptr;  // [768,]
         std::vector<float> hidden_states_vec;
 
         size_t chunk_count = std::max(clip_l_tokens.size() > 0 ? chunk_len : 0, t5_tokens.size()) / chunk_len;
@@ -1448,7 +1448,7 @@ struct T5CLIPEmbedder : public Conditioner {
         }
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         if (t5) {
             t5->get_param_tensors(tensors, "text_encoders.t5xxl.transformer");
         }
@@ -1523,7 +1523,7 @@ struct T5CLIPEmbedder : public Conditioner {
         return {t5_tokens, t5_weights, t5_mask};
     }
 
-    void modify_mask_to_attend_padding(struct ggml_tensor* mask, int max_seq_length, int num_extra_padding = 8) {
+    void modify_mask_to_attend_padding(ggml_tensor* mask, int max_seq_length, int num_extra_padding = 8) {
         float* mask_data = (float*)mask->data;
         int num_pad      = 0;
         for (int64_t i = 0; i < max_seq_length; i++) {
@@ -1554,11 +1554,11 @@ struct T5CLIPEmbedder : public Conditioner {
         auto& t5_weights       = std::get<1>(token_and_weights);
         auto& t5_attn_mask_vec = std::get<2>(token_and_weights);
 
-        int64_t t0                              = ggml_time_ms();
+        int64_t t0                       = ggml_time_ms();
-        struct ggml_tensor* hidden_states       = nullptr;  // [N, n_token, 4096]
+        ggml_tensor* hidden_states       = nullptr;  // [N, n_token, 4096]
-        struct ggml_tensor* chunk_hidden_states = nullptr;  // [n_token, 4096]
+        ggml_tensor* chunk_hidden_states = nullptr;  // [n_token, 4096]
-        struct ggml_tensor* pooled              = nullptr;
+        ggml_tensor* pooled              = nullptr;
-        struct ggml_tensor* t5_attn_mask        = vector_to_ggml_tensor(work_ctx, t5_attn_mask_vec);  // [n_token]
+        ggml_tensor* t5_attn_mask        = vector_to_ggml_tensor(work_ctx, t5_attn_mask_vec);  // [n_token]
 
         std::vector<float> hidden_states_vec;
 
@@ -1658,7 +1658,7 @@ struct AnimaConditioner : public Conditioner {
                                                false);
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         llm->get_param_tensors(tensors, "text_encoders.llm");
     }
 
@@ -1736,7 +1736,7 @@ struct AnimaConditioner : public Conditioner {
 
         auto input_ids = vector_to_ggml_tensor_i32(work_ctx, qwen_tokens);
 
-        struct ggml_tensor* hidden_states = nullptr;  // [N, n_token, 1024]
+        ggml_tensor* hidden_states = nullptr;  // [N, n_token, 1024]
         llm->compute(n_threads,
                      input_ids,
                      nullptr,
@@ -1763,8 +1763,8 @@ struct AnimaConditioner : public Conditioner {
             }
         }
 
-        struct ggml_tensor* t5_ids_tensor    = nullptr;
+        ggml_tensor* t5_ids_tensor    = nullptr;
-        struct ggml_tensor* t5_weight_tensor = nullptr;
+        ggml_tensor* t5_weight_tensor = nullptr;
         if (!t5_tokens.empty()) {
             t5_ids_tensor    = vector_to_ggml_tensor_i32(work_ctx, t5_tokens);
             t5_weight_tensor = vector_to_ggml_tensor(work_ctx, t5_weights);
@@ -1808,7 +1808,7 @@ struct LLMEmbedder : public Conditioner {
                                                enable_vision);
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         llm->get_param_tensors(tensors, "text_encoders.llm");
     }
 
@@ -1904,7 +1904,7 @@ struct LLMEmbedder : public Conditioner {
             tokenizer->pad_tokens(tokens, weights, max_length, true);
         }
 
-        struct ggml_tensor* hidden_states = nullptr;  // [N, n_token, hidden_size]
+        ggml_tensor* hidden_states = nullptr;  // [N, n_token, hidden_size]
 
         auto input_ids = vector_to_ggml_tensor_i32(work_ctx, tokens);
 

src/control.hpp

@@ -164,26 +164,26 @@ public:
         blocks["middle_block_out.0"] = std::shared_ptr<GGMLBlock>(make_zero_conv(ch));
     }
 
-    struct ggml_tensor* resblock_forward(std::string name,
+    ggml_tensor* resblock_forward(std::string name,
-                                         GGMLRunnerContext* ctx,
+                                  GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
+                                  ggml_tensor* x,
-                                         struct ggml_tensor* emb) {
+                                  ggml_tensor* emb) {
         auto block = std::dynamic_pointer_cast<ResBlock>(blocks[name]);
         return block->forward(ctx, x, emb);
     }
 
-    struct ggml_tensor* attention_layer_forward(std::string name,
+    ggml_tensor* attention_layer_forward(std::string name,
-                                                GGMLRunnerContext* ctx,
+                                         GGMLRunnerContext* ctx,
-                                                struct ggml_tensor* x,
+                                         ggml_tensor* x,
-                                                struct ggml_tensor* context) {
+                                         ggml_tensor* context) {
         auto block = std::dynamic_pointer_cast<SpatialTransformer>(blocks[name]);
         return block->forward(ctx, x, context);
     }
 
-    struct ggml_tensor* input_hint_block_forward(GGMLRunnerContext* ctx,
+    ggml_tensor* input_hint_block_forward(GGMLRunnerContext* ctx,
-                                                 struct ggml_tensor* hint,
+                                          ggml_tensor* hint,
-                                                 struct ggml_tensor* emb,
+                                          ggml_tensor* emb,
-                                                 struct ggml_tensor* context) {
+                                          ggml_tensor* context) {
         int num_input_blocks = 15;
         auto h               = hint;
         for (int i = 0; i < num_input_blocks; i++) {
@@ -198,13 +198,13 @@ public:
         return h;
     }
 
-    std::vector<struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
+    std::vector<ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                             struct ggml_tensor* x,
+                                      ggml_tensor* x,
-                                             struct ggml_tensor* hint,
+                                      ggml_tensor* hint,
-                                             struct ggml_tensor* guided_hint,
+                                      ggml_tensor* guided_hint,
-                                             struct ggml_tensor* timesteps,
+                                      ggml_tensor* timesteps,
-                                             struct ggml_tensor* context,
+                                      ggml_tensor* context,
-                                             struct ggml_tensor* y = nullptr) {
+                                      ggml_tensor* y = nullptr) {
         // x: [N, in_channels, h, w] or [N, in_channels/2, h, w]
         // timesteps: [N,]
         // context: [N, max_position, hidden_size] or [1, max_position, hidden_size]. for example, [N, 77, 768]
@@ -246,7 +246,7 @@ public:
             emb = ggml_add(ctx->ggml_ctx, emb, label_emb);  // [N, time_embed_dim]
         }
 
-        std::vector<struct ggml_tensor*> outs;
+        std::vector<ggml_tensor*> outs;
 
         if (guided_hint == nullptr) {
             guided_hint = input_hint_block_forward(ctx, hint, emb, context);
@@ -312,9 +312,9 @@ struct ControlNet : public GGMLRunner {
 
     ggml_backend_buffer_t control_buffer = nullptr;  // keep control output tensors in backend memory
     ggml_context* control_ctx            = nullptr;
-    std::vector<struct ggml_tensor*> controls;  // (12 input block outputs, 1 middle block output) SD 1.5
+    std::vector<ggml_tensor*> controls;  // (12 input block outputs, 1 middle block output) SD 1.5
-    struct ggml_tensor* guided_hint = nullptr;  // guided_hint cache, for faster inference
+    ggml_tensor* guided_hint = nullptr;  // guided_hint cache, for faster inference
-    bool guided_hint_cached         = false;
+    bool guided_hint_cached  = false;
 
     ControlNet(ggml_backend_t backend,
                bool offload_params_to_cpu,
@@ -328,8 +328,8 @@ struct ControlNet : public GGMLRunner {
         free_control_ctx();
     }
 
-    void alloc_control_ctx(std::vector<struct ggml_tensor*> outs) {
+    void alloc_control_ctx(std::vector<ggml_tensor*> outs) {
-        struct ggml_init_params params;
+        ggml_init_params params;
         params.mem_size   = static_cast<size_t>(outs.size() * ggml_tensor_overhead()) + 1024 * 1024;
         params.mem_buffer = nullptr;
         params.no_alloc   = true;
@@ -370,16 +370,16 @@ struct ControlNet : public GGMLRunner {
         return "control_net";
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
         control_net.get_param_tensors(tensors, prefix);
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+    ggml_cgraph* build_graph(ggml_tensor* x,
-                                    struct ggml_tensor* hint,
+                             ggml_tensor* hint,
-                                    struct ggml_tensor* timesteps,
+                             ggml_tensor* timesteps,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    struct ggml_tensor* y = nullptr) {
+                             ggml_tensor* y = nullptr) {
-        struct ggml_cgraph* gf = new_graph_custom(CONTROL_NET_GRAPH_SIZE);
+        ggml_cgraph* gf = new_graph_custom(CONTROL_NET_GRAPH_SIZE);
 
         x = to_backend(x);
         if (guided_hint_cached) {
@@ -414,18 +414,18 @@ struct ControlNet : public GGMLRunner {
     }
 
     bool compute(int n_threads,
-                 struct ggml_tensor* x,
+                 ggml_tensor* x,
-                 struct ggml_tensor* hint,
+                 ggml_tensor* hint,
-                 struct ggml_tensor* timesteps,
+                 ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
+                 ggml_tensor* context,
-                 struct ggml_tensor* y,
+                 ggml_tensor* y,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output     = nullptr,
-                 struct ggml_context* output_ctx = nullptr) {
+                 ggml_context* output_ctx = nullptr) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
         // y: [N, adm_in_channels] or [1, adm_in_channels]
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(x, hint, timesteps, context, y);
         };
 

src/denoiser.hpp

@@ -773,8 +773,8 @@ static bool sample_k_diffusion(sample_method_t method,
     // sample_euler_ancestral
     switch (method) {
         case EULER_A_SAMPLE_METHOD: {
-            struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* d     = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* d     = ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
                 float sigma = sigmas[i];
@@ -830,7 +830,7 @@ static bool sample_k_diffusion(sample_method_t method,
         } break;
         case EULER_SAMPLE_METHOD:  // Implemented without any sigma churn
         {
-            struct ggml_tensor* d = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* d = ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
                 float sigma = sigmas[i];
@@ -865,8 +865,8 @@ static bool sample_k_diffusion(sample_method_t method,
             }
         } break;
         case HEUN_SAMPLE_METHOD: {
-            struct ggml_tensor* d  = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* d  = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* x2 = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* x2 = ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
                 // denoise
@@ -921,8 +921,8 @@ static bool sample_k_diffusion(sample_method_t method,
             }
         } break;
         case DPM2_SAMPLE_METHOD: {
-            struct ggml_tensor* d  = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* d  = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* x2 = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* x2 = ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
                 // denoise
@@ -979,8 +979,8 @@ static bool sample_k_diffusion(sample_method_t method,
 
         } break;
         case DPMPP2S_A_SAMPLE_METHOD: {
-            struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* x2    = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* x2    = ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
                 // denoise
@@ -1050,7 +1050,7 @@ static bool sample_k_diffusion(sample_method_t method,
         } break;
         case DPMPP2M_SAMPLE_METHOD:  // DPM++ (2M) from Karras et al (2022)
         {
-            struct ggml_tensor* old_denoised = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* old_denoised = ggml_dup_tensor(work_ctx, x);
 
             auto t_fn = [](float sigma) -> float { return -log(sigma); };
 
@@ -1092,7 +1092,7 @@ static bool sample_k_diffusion(sample_method_t method,
         } break;
         case DPMPP2Mv2_SAMPLE_METHOD:  // Modified DPM++ (2M) from https://github.com/AUTOMATIC1111/stable-diffusion-webui/discussions/8457
         {
-            struct ggml_tensor* old_denoised = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* old_denoised = ggml_dup_tensor(work_ctx, x);
 
             auto t_fn = [](float sigma) -> float { return -log(sigma); };
 
@@ -1157,8 +1157,8 @@ static bool sample_k_diffusion(sample_method_t method,
                 }
                 float* vec_denoised = (float*)denoised->data;
                 // d_cur = (x_cur - denoised) / sigma
-                struct ggml_tensor* d_cur = ggml_dup_tensor(work_ctx, x_cur);
+                ggml_tensor* d_cur = ggml_dup_tensor(work_ctx, x_cur);
-                float* vec_d_cur          = (float*)d_cur->data;
+                float* vec_d_cur   = (float*)d_cur->data;
 
                 for (int j = 0; j < ggml_nelements(d_cur); j++) {
                     vec_d_cur[j] = (vec_x_cur[j] - vec_denoised[j]) / sigma;
@@ -1225,11 +1225,11 @@ static bool sample_k_diffusion(sample_method_t method,
                 float t_next = sigmas[i + 1];
 
                 // Denoising step
-                ggml_tensor* denoised     = model(x, sigma, i + 1);
+                ggml_tensor* denoised = model(x, sigma, i + 1);
-                float* vec_denoised       = (float*)denoised->data;
+                float* vec_denoised   = (float*)denoised->data;
-                struct ggml_tensor* d_cur = ggml_dup_tensor(work_ctx, x);
+                ggml_tensor* d_cur    = ggml_dup_tensor(work_ctx, x);
-                float* vec_d_cur          = (float*)d_cur->data;
+                float* vec_d_cur      = (float*)d_cur->data;
-                float* vec_x              = (float*)x->data;
+                float* vec_x          = (float*)x->data;
 
                 // d_cur = (x - denoised) / sigma
                 for (int j = 0; j < ggml_nelements(d_cur); j++) {
@@ -1290,8 +1290,8 @@ static bool sample_k_diffusion(sample_method_t method,
         } break;
         case LCM_SAMPLE_METHOD:  // Latent Consistency Models
         {
-            struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* d     = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* d     = ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
                 float sigma = sigmas[i];
@@ -1358,9 +1358,9 @@ static bool sample_k_diffusion(sample_method_t method,
                               alphas_cumprod[i]);
             }
 
-            struct ggml_tensor* pred_original_sample =
+            ggml_tensor* pred_original_sample =
                 ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* variance_noise =
+            ggml_tensor* variance_noise =
                 ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
@@ -1422,7 +1422,7 @@ static bool sample_k_diffusion(sample_method_t method,
                 // model_output = model() is the D(x, sigma) as
                 // defined in Karras et al. (2022), p. 3, Table 1 and
                 // p. 8 (7), compare also p. 38 (226) therein.
-                struct ggml_tensor* model_output =
+                ggml_tensor* model_output =
                     model(x, sigma, i + 1);
                 // Here model_output is still the k-diffusion denoiser
                 // output, not the U-net output F_theta(c_in(sigma) x;
@@ -1545,9 +1545,9 @@ static bool sample_k_diffusion(sample_method_t method,
             }
             int original_steps = 50;
 
-            struct ggml_tensor* pred_original_sample =
+            ggml_tensor* pred_original_sample =
                 ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* noise =
+            ggml_tensor* noise =
                 ggml_dup_tensor(work_ctx, x);
 
             for (int i = 0; i < steps; i++) {
@@ -1581,7 +1581,7 @@ static bool sample_k_diffusion(sample_method_t method,
                         vec_x[j] *= std::sqrt(sigma * sigma + 1);
                     }
                 }
-                struct ggml_tensor* model_output =
+                ggml_tensor* model_output =
                     model(x, sigma, i + 1);
                 {
                     float* vec_x = (float*)x->data;
@@ -1689,8 +1689,8 @@ static bool sample_k_diffusion(sample_method_t method,
         } break;
         case RES_MULTISTEP_SAMPLE_METHOD:  // Res Multistep sampler
         {
-            struct ggml_tensor* noise        = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* noise        = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* old_denoised = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* old_denoised = ggml_dup_tensor(work_ctx, x);
 
             bool have_old_sigma  = false;
             float old_sigma_down = 0.0f;
@@ -1797,9 +1797,9 @@ static bool sample_k_diffusion(sample_method_t method,
         } break;
         case RES_2S_SAMPLE_METHOD:  // Res 2s sampler
         {
-            struct ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* noise = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* x0    = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* x0    = ggml_dup_tensor(work_ctx, x);
-            struct ggml_tensor* x2    = ggml_dup_tensor(work_ctx, x);
+            ggml_tensor* x2    = ggml_dup_tensor(work_ctx, x);
 
             const float c2 = 0.5f;
             auto t_fn      = [](float sigma) -> float { return -logf(sigma); };

src/diffusion_model.hpp

@@ -10,33 +10,33 @@
 #include "z_image.hpp"
 
 struct DiffusionParams {
-    struct ggml_tensor* x                     = nullptr;
+    ggml_tensor* x                        = nullptr;
-    struct ggml_tensor* timesteps             = nullptr;
+    ggml_tensor* timesteps                = nullptr;
-    struct ggml_tensor* context               = nullptr;
+    ggml_tensor* context                  = nullptr;
-    struct ggml_tensor* c_concat              = nullptr;
+    ggml_tensor* c_concat                 = nullptr;
-    struct ggml_tensor* y                     = nullptr;
+    ggml_tensor* y                        = nullptr;
-    struct ggml_tensor* guidance              = nullptr;
+    ggml_tensor* guidance                 = nullptr;
-    std::vector<ggml_tensor*> ref_latents     = {};
+    std::vector<ggml_tensor*> ref_latents = {};
-    bool increase_ref_index                   = false;
+    bool increase_ref_index               = false;
-    int num_video_frames                      = -1;
+    int num_video_frames                  = -1;
-    std::vector<struct ggml_tensor*> controls = {};
+    std::vector<ggml_tensor*> controls    = {};
-    float control_strength                    = 0.f;
+    float control_strength                = 0.f;
-    struct ggml_tensor* vace_context          = nullptr;
+    ggml_tensor* vace_context             = nullptr;
-    float vace_strength                       = 1.f;
+    float vace_strength                   = 1.f;
-    std::vector<int> skip_layers              = {};
+    std::vector<int> skip_layers          = {};
 };
 
 struct DiffusionModel {
-    virtual std::string get_desc()                                                      = 0;
+    virtual std::string get_desc()                                               = 0;
     virtual bool compute(int n_threads,
                          DiffusionParams diffusion_params,
-                         struct ggml_tensor** output     = nullptr,
+                         ggml_tensor** output     = nullptr,
-                         struct ggml_context* output_ctx = nullptr)                     = 0;
+                         ggml_context* output_ctx = nullptr)                     = 0;
-    virtual void alloc_params_buffer()                                                  = 0;
+    virtual void alloc_params_buffer()                                           = 0;
-    virtual void free_params_buffer()                                                   = 0;
+    virtual void free_params_buffer()                                            = 0;
-    virtual void free_compute_buffer()                                                  = 0;
+    virtual void free_compute_buffer()                                           = 0;
-    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) = 0;
+    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) = 0;
-    virtual size_t get_params_buffer_size()                                             = 0;
+    virtual size_t get_params_buffer_size()                                      = 0;
     virtual void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter){};
     virtual int64_t get_adm_in_channels()                            = 0;
     virtual void set_flash_attention_enabled(bool enabled)           = 0;
@@ -69,7 +69,7 @@ struct UNetModel : public DiffusionModel {
         unet.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         unet.get_param_tensors(tensors, "model.diffusion_model");
     }
 
@@ -95,8 +95,8 @@ struct UNetModel : public DiffusionModel {
 
     bool compute(int n_threads,
                  DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output     = nullptr,
-                 struct ggml_context* output_ctx = nullptr) override {
+                 ggml_context* output_ctx = nullptr) override {
         return unet.compute(n_threads,
                             diffusion_params.x,
                             diffusion_params.timesteps,
@@ -134,7 +134,7 @@ struct MMDiTModel : public DiffusionModel {
         mmdit.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         mmdit.get_param_tensors(tensors, "model.diffusion_model");
     }
 
@@ -160,8 +160,8 @@ struct MMDiTModel : public DiffusionModel {
 
     bool compute(int n_threads,
                  DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output     = nullptr,
-                 struct ggml_context* output_ctx = nullptr) override {
+                 ggml_context* output_ctx = nullptr) override {
         return mmdit.compute(n_threads,
                              diffusion_params.x,
                              diffusion_params.timesteps,
@@ -200,7 +200,7 @@ struct FluxModel : public DiffusionModel {
         flux.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         flux.get_param_tensors(tensors, "model.diffusion_model");
     }
 
@@ -226,8 +226,8 @@ struct FluxModel : public DiffusionModel {
 
     bool compute(int n_threads,
                  DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output     = nullptr,
-                 struct ggml_context* output_ctx = nullptr) override {
+                 ggml_context* output_ctx = nullptr) override {
         return flux.compute(n_threads,
                             diffusion_params.x,
                             diffusion_params.timesteps,
@@ -270,7 +270,7 @@ struct AnimaModel : public DiffusionModel {
         anima.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         anima.get_param_tensors(tensors, prefix);
     }
 
@@ -296,8 +296,8 @@ struct AnimaModel : public DiffusionModel {
 
     bool compute(int n_threads,
                  DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output     = nullptr,
-                 struct ggml_context* output_ctx = nullptr) override {
+                 ggml_context* output_ctx = nullptr) override {
         return anima.compute(n_threads,
                              diffusion_params.x,
                              diffusion_params.timesteps,
@@ -337,7 +337,7 @@ struct WanModel : public DiffusionModel {
         wan.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         wan.get_param_tensors(tensors, prefix);
     }
 
@@ -363,8 +363,8 @@ struct WanModel : public DiffusionModel {
 
     bool compute(int n_threads,
                  DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output     = nullptr,
-                 struct ggml_context* output_ctx = nullptr) override {
+                 ggml_context* output_ctx = nullptr) override {
         return wan.compute(n_threads,
                            diffusion_params.x,
                            diffusion_params.timesteps,
@@ -408,7 +408,7 @@ struct QwenImageModel : public DiffusionModel {
         qwen_image.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         qwen_image.get_param_tensors(tensors, prefix);
     }
 
@@ -434,8 +434,8 @@ struct QwenImageModel : public DiffusionModel {
 
     bool compute(int n_threads,
                  DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output     = nullptr,
-                 struct ggml_context* output_ctx = nullptr) override {
+                 ggml_context* output_ctx = nullptr) override {
         return qwen_image.compute(n_threads,
                                   diffusion_params.x,
                                   diffusion_params.timesteps,
@@ -475,7 +475,7 @@ struct ZImageModel : public DiffusionModel {
         z_image.free_compute_buffer();
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors) override {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
         z_image.get_param_tensors(tensors, prefix);
     }
 
@@ -501,8 +501,8 @@ struct ZImageModel : public DiffusionModel {
 
     bool compute(int n_threads,
                  DiffusionParams diffusion_params,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output     = nullptr,
-                 struct ggml_context* output_ctx = nullptr) override {
+                 ggml_context* output_ctx = nullptr) override {
         return z_image.compute(n_threads,
                                diffusion_params.x,
                                diffusion_params.timesteps,

src/esrgan.hpp

@@ -27,11 +27,11 @@ public:
         blocks["conv5"] = std::shared_ptr<GGMLBlock>(new Conv2d(num_feat + 4 * num_grow_ch, num_feat, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* lrelu(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* lrelu(GGMLRunnerContext* ctx, ggml_tensor* x) {
         return ggml_leaky_relu(ctx->ggml_ctx, x, 0.2f, true);
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [n, num_feat, h, w]
         // return: [n, num_feat, h, w]
 
@@ -64,7 +64,7 @@ public:
         blocks["rdb3"] = std::shared_ptr<GGMLBlock>(new ResidualDenseBlock(num_feat, num_grow_ch));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [n, num_feat, h, w]
         // return: [n, num_feat, h, w]
 
@@ -112,11 +112,11 @@ public:
     int get_scale() { return scale; }
     int get_num_block() { return num_block; }
 
-    struct ggml_tensor* lrelu(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* lrelu(GGMLRunnerContext* ctx, ggml_tensor* x) {
         return ggml_leaky_relu(ctx->ggml_ctx, x, 0.2f, true);
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [n, num_in_ch, h, w]
         // return: [n, num_out_ch, h*scale, w*scale]
         auto conv_first = std::dynamic_pointer_cast<Conv2d>(blocks["conv_first"]);
@@ -341,24 +341,24 @@ struct ESRGAN : public GGMLRunner {
         return success;
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* x) {
+    ggml_cgraph* build_graph(ggml_tensor* x) {
         if (!rrdb_net)
             return nullptr;
         constexpr int kGraphNodes = 1 << 16;  // 65k
-        struct ggml_cgraph* gf    = new_graph_custom(kGraphNodes);
+        ggml_cgraph* gf           = new_graph_custom(kGraphNodes);
         x                         = to_backend(x);
 
-        auto runner_ctx         = get_context();
+        auto runner_ctx  = get_context();
-        struct ggml_tensor* out = rrdb_net->forward(&runner_ctx, x);
+        ggml_tensor* out = rrdb_net->forward(&runner_ctx, x);
         ggml_build_forward_expand(gf, out);
         return gf;
     }
 
     bool compute(const int n_threads,
-                 struct ggml_tensor* x,
+                 ggml_tensor* x,
                  ggml_tensor** output,
                  ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(x);
         };
         return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);

src/flux.hpp

@@ -19,7 +19,7 @@ namespace Flux {
             blocks["out_layer"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_dim, hidden_dim, bias));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
             // x: [..., in_dim]
             // return: [..., hidden_dim]
             auto in_layer  = std::dynamic_pointer_cast<Linear>(blocks["in_layer"]);
@@ -37,7 +37,7 @@ namespace Flux {
         int64_t hidden_size;
         float eps;
 
-        void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             ggml_type wtype = GGML_TYPE_F32;
             params["scale"] = ggml_new_tensor_1d(ctx, wtype, hidden_size);
         }
@@ -48,10 +48,10 @@ namespace Flux {
             : hidden_size(hidden_size),
               eps(eps) {}
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
-            struct ggml_tensor* w = params["scale"];
+            ggml_tensor* w = params["scale"];
-            x                     = ggml_rms_norm(ctx->ggml_ctx, x, eps);
+            x              = ggml_rms_norm(ctx->ggml_ctx, x, eps);
-            x                     = ggml_mul(ctx->ggml_ctx, x, w);
+            x              = ggml_mul(ctx->ggml_ctx, x, w);
             return x;
         }
     };
@@ -63,7 +63,7 @@ namespace Flux {
             blocks["key_norm"]   = std::shared_ptr<GGMLBlock>(new RMSNorm(dim));
         }
 
-        struct ggml_tensor* query_norm(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* query_norm(GGMLRunnerContext* ctx, ggml_tensor* x) {
             // x: [..., dim]
             // return: [..., dim]
             auto norm = std::dynamic_pointer_cast<RMSNorm>(blocks["query_norm"]);
@@ -72,7 +72,7 @@ namespace Flux {
             return x;
         }
 
-        struct ggml_tensor* key_norm(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* key_norm(GGMLRunnerContext* ctx, ggml_tensor* x) {
             // x: [..., dim]
             // return: [..., dim]
             auto norm = std::dynamic_pointer_cast<RMSNorm>(blocks["key_norm"]);
@@ -98,32 +98,34 @@ namespace Flux {
             blocks["proj"]   = std::shared_ptr<GGMLBlock>(new Linear(dim, dim, proj_bias));
         }
 
-        std::vector<struct ggml_tensor*> pre_attention(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        std::vector<ggml_tensor*> pre_attention(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
             auto norm     = std::dynamic_pointer_cast<QKNorm>(blocks["norm"]);
 
             auto qkv         = qkv_proj->forward(ctx, x);
-            auto qkv_vec     = ggml_ext_chunk(ctx->ggml_ctx, qkv, 3, 0, true);
+            int64_t head_dim = qkv->ne[0] / 3 / num_heads;
-            int64_t head_dim = qkv_vec[0]->ne[0] / num_heads;
+            auto q           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
-            auto q           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[0], head_dim, num_heads, qkv_vec[0]->ne[1], qkv_vec[0]->ne[2]);
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], 0);
-            auto k           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[1], head_dim, num_heads, qkv_vec[1]->ne[1], qkv_vec[1]->ne[2]);
+            auto k           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
-            auto v           = ggml_reshape_4d(ctx->ggml_ctx, qkv_vec[2], head_dim, num_heads, qkv_vec[2]->ne[1], qkv_vec[2]->ne[2]);
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], (qkv->nb[0]) * qkv->ne[0] / 3);
+            auto v           = ggml_view_4d(ctx->ggml_ctx, qkv, head_dim, num_heads, qkv->ne[1], qkv->ne[2],
+                                            qkv->nb[0] * head_dim, qkv->nb[1], qkv->nb[2], (qkv->nb[0]) * 2 * qkv->ne[0] / 3);
             q                = norm->query_norm(ctx, q);
             k                = norm->key_norm(ctx, k);
             return {q, k, v};
         }
 
-        struct ggml_tensor* post_attention(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* post_attention(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
 
             x = proj->forward(ctx, x);  // [N, n_token, dim]
             return x;
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* mask) {
+                             ggml_tensor* mask) {
             // x: [N, n_token, dim]
             // pe: [n_token, d_head/2, 2, 2]
             // return [N, n_token, dim]
@@ -145,7 +147,7 @@ namespace Flux {
             blocks["2"]             = std::make_shared<Linear>(intermediate_size, hidden_size, bias);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["0"]);
             auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["2"]);
 
@@ -168,7 +170,7 @@ namespace Flux {
             blocks["down_proj"] = std::make_shared<Linear>(intermediate_size, hidden_size, bias);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto gate_proj = std::dynamic_pointer_cast<Linear>(blocks["gate_proj"]);
             auto up_proj   = std::dynamic_pointer_cast<Linear>(blocks["up_proj"]);
             auto down_proj = std::dynamic_pointer_cast<Linear>(blocks["down_proj"]);
@@ -210,7 +212,7 @@ namespace Flux {
             blocks["lin"] = std::shared_ptr<GGMLBlock>(new Linear(dim, dim * multiplier, bias));
         }
 
-        std::vector<ModulationOut> forward(GGMLRunnerContext* ctx, struct ggml_tensor* vec) {
+        std::vector<ModulationOut> forward(GGMLRunnerContext* ctx, ggml_tensor* vec) {
             // x: [N, dim]
             // return: [ModulationOut, ModulationOut]
             auto lin = std::dynamic_pointer_cast<Linear>(blocks["lin"]);
@@ -230,11 +232,11 @@ namespace Flux {
         }
     };
 
-    __STATIC_INLINE__ struct ggml_tensor* modulate(struct ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx,
-                                                   struct ggml_tensor* x,
+                                            ggml_tensor* x,
-                                                   struct ggml_tensor* shift,
+                                            ggml_tensor* shift,
-                                                   struct ggml_tensor* scale,
+                                            ggml_tensor* scale,
-                                                   bool skip_reshape = false) {
+                                            bool skip_reshape = false) {
         // x: [N, L, C]
         // scale: [N, C]
         // shift: [N, C]
@@ -292,7 +294,7 @@ namespace Flux {
             }
         }
 
-        std::vector<ModulationOut> get_distil_img_mod(GGMLRunnerContext* ctx, struct ggml_tensor* vec) {
+        std::vector<ModulationOut> get_distil_img_mod(GGMLRunnerContext* ctx, ggml_tensor* vec) {
             // TODO: not hardcoded?
             const int single_blocks_count = 38;
             const int double_blocks_count = 19;
@@ -301,7 +303,7 @@ namespace Flux {
             return {ModulationOut(ctx, vec, offset), ModulationOut(ctx, vec, offset + 3)};
         }
 
-        std::vector<ModulationOut> get_distil_txt_mod(GGMLRunnerContext* ctx, struct ggml_tensor* vec) {
+        std::vector<ModulationOut> get_distil_txt_mod(GGMLRunnerContext* ctx, ggml_tensor* vec) {
             // TODO: not hardcoded?
             const int single_blocks_count = 38;
             const int double_blocks_count = 19;
@@ -310,14 +312,14 @@ namespace Flux {
             return {ModulationOut(ctx, vec, offset), ModulationOut(ctx, vec, offset + 3)};
         }
 
-        std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
+        std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                                    struct ggml_tensor* img,
+                                                      ggml_tensor* img,
-                                                                    struct ggml_tensor* txt,
+                                                      ggml_tensor* txt,
-                                                                    struct ggml_tensor* vec,
+                                                      ggml_tensor* vec,
-                                                                    struct ggml_tensor* pe,
+                                                      ggml_tensor* pe,
-                                                                    struct ggml_tensor* mask            = nullptr,
+                                                      ggml_tensor* mask                   = nullptr,
-                                                                    std::vector<ModulationOut> img_mods = {},
+                                                      std::vector<ModulationOut> img_mods = {},
-                                                                    std::vector<ModulationOut> txt_mods = {}) {
+                                                      std::vector<ModulationOut> txt_mods = {}) {
             // img: [N, n_img_token, hidden_size]
             // txt: [N, n_txt_token, hidden_size]
             // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
@@ -455,17 +457,17 @@ namespace Flux {
             }
         }
 
-        ModulationOut get_distil_mod(GGMLRunnerContext* ctx, struct ggml_tensor* vec) {
+        ModulationOut get_distil_mod(GGMLRunnerContext* ctx, ggml_tensor* vec) {
             int64_t offset = 3 * idx;
             return ModulationOut(ctx, vec, offset);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* vec,
+                             ggml_tensor* vec,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* mask        = nullptr,
+                             ggml_tensor* mask               = nullptr,
-                                    std::vector<ModulationOut> mods = {}) {
+                             std::vector<ModulationOut> mods = {}) {
             // x: [N, n_token, hidden_size]
             // pe: [n_token, d_head/2, 2, 2]
             // return: [N, n_token, hidden_size]
@@ -491,15 +493,14 @@ namespace Flux {
             auto x_mod   = Flux::modulate(ctx->ggml_ctx, pre_norm->forward(ctx, x), mod.shift, mod.scale);
             auto qkv_mlp = linear1->forward(ctx, x_mod);  // [N, n_token, hidden_size * 3 + mlp_hidden_dim*mlp_mult_factor]
 
-            auto q = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], 0);
-            auto k = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], hidden_size * qkv_mlp->nb[0]);
-            auto v = ggml_view_3d(ctx->ggml_ctx, qkv_mlp, hidden_size, qkv_mlp->ne[1], qkv_mlp->ne[2], qkv_mlp->nb[1], qkv_mlp->nb[2], hidden_size * 2 * qkv_mlp->nb[0]);
-
             int64_t head_dim = hidden_size / num_heads;
 
-            q = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, q), head_dim, num_heads, q->ne[1], q->ne[2]);  // [N, n_token, n_head, d_head]
+            auto q = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
-            k = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, k), head_dim, num_heads, k->ne[1], k->ne[2]);  // [N, n_token, n_head, d_head]
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], 0);
-            v = ggml_reshape_4d(ctx->ggml_ctx, ggml_cont(ctx->ggml_ctx, v), head_dim, num_heads, v->ne[1], v->ne[2]);  // [N, n_token, n_head, d_head]
+            auto k = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], (qkv_mlp->nb[0]) * hidden_size);
+            auto v = ggml_view_4d(ctx->ggml_ctx, qkv_mlp, head_dim, num_heads, qkv_mlp->ne[1], qkv_mlp->ne[2],
+                                  qkv_mlp->nb[0] * head_dim, qkv_mlp->nb[1], qkv_mlp->nb[2], (qkv_mlp->nb[0]) * 2 * hidden_size);
 
             q         = norm->query_norm(ctx, q);
             k         = norm->key_norm(ctx, k);
@@ -538,7 +539,7 @@ namespace Flux {
             }
         }
 
-        ModulationOut get_distil_mod(GGMLRunnerContext* ctx, struct ggml_tensor* vec) {
+        ModulationOut get_distil_mod(GGMLRunnerContext* ctx, ggml_tensor* vec) {
             int64_t offset = vec->ne[2] - 2;
             int64_t stride = vec->nb[1] * vec->ne[1];
             auto shift     = ggml_view_2d(ctx->ggml_ctx, vec, vec->ne[0], vec->ne[1], vec->nb[1], stride * (offset + 0));  // [N, dim]
@@ -547,15 +548,15 @@ namespace Flux {
             return {shift, scale, nullptr};
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* c) {
+                             ggml_tensor* c) {
             // x: [N, n_token, hidden_size]
             // c: [N, hidden_size]
             // return: [N, n_token, patch_size * patch_size * out_channels]
             auto norm_final = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_final"]);
             auto linear     = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
-            struct ggml_tensor *shift, *scale;
+            ggml_tensor *shift, *scale;
             if (prune_mod) {
                 auto mod = get_distil_mod(ctx, c);
                 shift    = mod.shift;
@@ -588,7 +589,7 @@ namespace Flux {
             blocks["out_proj"] = std::shared_ptr<GGMLBlock>(new Linear(inner_size, hidden_size, true));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto in_proj  = std::dynamic_pointer_cast<Linear>(blocks["in_proj"]);
             auto out_proj = std::dynamic_pointer_cast<Linear>(blocks["out_proj"]);
 
@@ -611,9 +612,9 @@ namespace Flux {
             blocks["embedder.0"] = std::make_shared<Linear>(in_channels + max_freqs * max_freqs, hidden_size_input);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* dct) {
+                             ggml_tensor* dct) {
             // x: (B, P^2, C)
             // dct: (1, P^2, max_freqs^2)
             // return: (B, P^2, hidden_size_input)
@@ -638,9 +639,9 @@ namespace Flux {
             blocks["norm"]            = std::make_shared<RMSNorm>(hidden_size_x);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* s) {
+                             ggml_tensor* s) {
             // x: (batch_size, n_token, hidden_size_x)
             // s: (batch_size, hidden_size_s)
             // return: (batch_size, n_token, hidden_size_x)
@@ -688,8 +689,8 @@ namespace Flux {
             blocks["linear"] = std::make_shared<Linear>(hidden_size, out_channels);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x) {
+                             ggml_tensor* x) {
             auto norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
             auto linear = std::dynamic_pointer_cast<Linear>(blocks["linear"]);
 
@@ -707,8 +708,8 @@ namespace Flux {
             blocks["conv"] = std::make_shared<Conv2d>(hidden_size, out_channels, std::pair{3, 3}, std::pair{1, 1}, std::pair{1, 1});
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x) {
+                             ggml_tensor* x) {
             // x: [N, C, H, W]
             auto norm = std::dynamic_pointer_cast<RMSNorm>(blocks["norm"]);
             auto conv = std::dynamic_pointer_cast<Conv2d>(blocks["conv"]);
@@ -846,15 +847,15 @@ namespace Flux {
             }
         }
 
-        struct ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
+        ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* img,
+                                  ggml_tensor* img,
-                                         struct ggml_tensor* txt,
+                                  ggml_tensor* txt,
-                                         struct ggml_tensor* timesteps,
+                                  ggml_tensor* timesteps,
-                                         struct ggml_tensor* y,
+                                  ggml_tensor* y,
-                                         struct ggml_tensor* guidance,
+                                  ggml_tensor* guidance,
-                                         struct ggml_tensor* pe,
+                                  ggml_tensor* pe,
-                                         struct ggml_tensor* mod_index_arange = nullptr,
+                                  ggml_tensor* mod_index_arange = nullptr,
-                                         std::vector<int> skip_layers         = {}) {
+                                  std::vector<int> skip_layers  = {}) {
             auto img_in      = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
             auto txt_in      = std::dynamic_pointer_cast<Linear>(blocks["txt_in"]);
             auto final_layer = std::dynamic_pointer_cast<LastLayer>(blocks["final_layer"]);
@@ -863,8 +864,8 @@ namespace Flux {
                 img = img_in->forward(ctx, img);
             }
 
-            struct ggml_tensor* vec;
+            ggml_tensor* vec;
-            struct ggml_tensor* txt_img_mask = nullptr;
+            ggml_tensor* txt_img_mask = nullptr;
             if (params.is_chroma) {
                 int64_t mod_index_length = 344;
                 auto approx              = std::dynamic_pointer_cast<ChromaApproximator>(blocks["distilled_guidance_layer"]);
@@ -966,27 +967,27 @@ namespace Flux {
             return img;
         }
 
-        struct ggml_tensor* _apply_x0_residual(GGMLRunnerContext* ctx,
+        ggml_tensor* _apply_x0_residual(GGMLRunnerContext* ctx,
-                                               struct ggml_tensor* predicted,
+                                        ggml_tensor* predicted,
-                                               struct ggml_tensor* noisy,
+                                        ggml_tensor* noisy,
-                                               struct ggml_tensor* timesteps) {
+                                        ggml_tensor* timesteps) {
             auto x = ggml_sub(ctx->ggml_ctx, noisy, predicted);
             x      = ggml_div(ctx->ggml_ctx, x, timesteps);
             return x;
         }
 
-        struct ggml_tensor* forward_chroma_radiance(GGMLRunnerContext* ctx,
+        ggml_tensor* forward_chroma_radiance(GGMLRunnerContext* ctx,
-                                                    struct ggml_tensor* x,
+                                             ggml_tensor* x,
-                                                    struct ggml_tensor* timestep,
+                                             ggml_tensor* timestep,
-                                                    struct ggml_tensor* context,
+                                             ggml_tensor* context,
-                                                    struct ggml_tensor* c_concat,
+                                             ggml_tensor* c_concat,
-                                                    struct ggml_tensor* y,
+                                             ggml_tensor* y,
-                                                    struct ggml_tensor* guidance,
+                                             ggml_tensor* guidance,
-                                                    struct ggml_tensor* pe,
+                                             ggml_tensor* pe,
-                                                    struct ggml_tensor* mod_index_arange  = nullptr,
+                                             ggml_tensor* mod_index_arange         = nullptr,
-                                                    struct ggml_tensor* dct               = nullptr,
+                                             ggml_tensor* dct                      = nullptr,
-                                                    std::vector<ggml_tensor*> ref_latents = {},
+                                             std::vector<ggml_tensor*> ref_latents = {},
-                                                    std::vector<int> skip_layers          = {}) {
+                                             std::vector<int> skip_layers          = {}) {
             GGML_ASSERT(x->ne[3] == 1);
 
             int64_t W      = x->ne[0];
@@ -1049,18 +1050,18 @@ namespace Flux {
             return out;
         }
 
-        struct ggml_tensor* forward_flux_chroma(GGMLRunnerContext* ctx,
+        ggml_tensor* forward_flux_chroma(GGMLRunnerContext* ctx,
-                                                struct ggml_tensor* x,
+                                         ggml_tensor* x,
-                                                struct ggml_tensor* timestep,
+                                         ggml_tensor* timestep,
-                                                struct ggml_tensor* context,
+                                         ggml_tensor* context,
-                                                struct ggml_tensor* c_concat,
+                                         ggml_tensor* c_concat,
-                                                struct ggml_tensor* y,
+                                         ggml_tensor* y,
-                                                struct ggml_tensor* guidance,
+                                         ggml_tensor* guidance,
-                                                struct ggml_tensor* pe,
+                                         ggml_tensor* pe,
-                                                struct ggml_tensor* mod_index_arange  = nullptr,
+                                         ggml_tensor* mod_index_arange         = nullptr,
-                                                struct ggml_tensor* dct               = nullptr,
+                                         ggml_tensor* dct                      = nullptr,
-                                                std::vector<ggml_tensor*> ref_latents = {},
+                                         std::vector<ggml_tensor*> ref_latents = {},
-                                                std::vector<int> skip_layers          = {}) {
+                                         std::vector<int> skip_layers          = {}) {
             GGML_ASSERT(x->ne[3] == 1);
 
             int64_t W      = x->ne[0];
@@ -1118,18 +1119,18 @@ namespace Flux {
             return out;
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* timestep,
+                             ggml_tensor* timestep,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    struct ggml_tensor* c_concat,
+                             ggml_tensor* c_concat,
-                                    struct ggml_tensor* y,
+                             ggml_tensor* y,
-                                    struct ggml_tensor* guidance,
+                             ggml_tensor* guidance,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* mod_index_arange  = nullptr,
+                             ggml_tensor* mod_index_arange         = nullptr,
-                                    struct ggml_tensor* dct               = nullptr,
+                             ggml_tensor* dct                      = nullptr,
-                                    std::vector<ggml_tensor*> ref_latents = {},
+                             std::vector<ggml_tensor*> ref_latents = {},
-                                    std::vector<int> skip_layers          = {}) {
+                             std::vector<int> skip_layers          = {}) {
             // Forward pass of DiT.
             // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
             // timestep: (N,) tensor of diffusion timesteps
@@ -1298,7 +1299,7 @@ namespace Flux {
             return "flux";
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
             flux.get_param_tensors(tensors, prefix);
         }
 
@@ -1352,20 +1353,20 @@ namespace Flux {
             return dct;
         }
 
-        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+        ggml_cgraph* build_graph(ggml_tensor* x,
-                                        struct ggml_tensor* timesteps,
+                                 ggml_tensor* timesteps,
-                                        struct ggml_tensor* context,
+                                 ggml_tensor* context,
-                                        struct ggml_tensor* c_concat,
+                                 ggml_tensor* c_concat,
-                                        struct ggml_tensor* y,
+                                 ggml_tensor* y,
-                                        struct ggml_tensor* guidance,
+                                 ggml_tensor* guidance,
-                                        std::vector<ggml_tensor*> ref_latents = {},
+                                 std::vector<ggml_tensor*> ref_latents = {},
-                                        bool increase_ref_index               = false,
+                                 bool increase_ref_index               = false,
-                                        std::vector<int> skip_layers          = {}) {
+                                 std::vector<int> skip_layers          = {}) {
             GGML_ASSERT(x->ne[3] == 1);
-            struct ggml_cgraph* gf = new_graph_custom(FLUX_GRAPH_SIZE);
+            ggml_cgraph* gf = new_graph_custom(FLUX_GRAPH_SIZE);
 
-            struct ggml_tensor* mod_index_arange = nullptr;
+            ggml_tensor* mod_index_arange = nullptr;
-            struct ggml_tensor* dct              = nullptr;  // for chroma radiance
+            ggml_tensor* dct              = nullptr;  // for chroma radiance
 
             x       = to_backend(x);
             context = to_backend(context);
@@ -1436,18 +1437,18 @@ namespace Flux {
 
             auto runner_ctx = get_context();
 
-            struct ggml_tensor* out = flux.forward(&runner_ctx,
+            ggml_tensor* out = flux.forward(&runner_ctx,
-                                                   x,
+                                            x,
-                                                   timesteps,
+                                            timesteps,
-                                                   context,
+                                            context,
-                                                   c_concat,
+                                            c_concat,
-                                                   y,
+                                            y,
-                                                   guidance,
+                                            guidance,
-                                                   pe,
+                                            pe,
-                                                   mod_index_arange,
+                                            mod_index_arange,
-                                                   dct,
+                                            dct,
-                                                   ref_latents,
+                                            ref_latents,
-                                                   skip_layers);
+                                            skip_layers);
 
             ggml_build_forward_expand(gf, out);
 
@@ -1455,23 +1456,23 @@ namespace Flux {
         }
 
         bool compute(int n_threads,
-                     struct ggml_tensor* x,
+                     ggml_tensor* x,
-                     struct ggml_tensor* timesteps,
+                     ggml_tensor* timesteps,
-                     struct ggml_tensor* context,
+                     ggml_tensor* context,
-                     struct ggml_tensor* c_concat,
+                     ggml_tensor* c_concat,
-                     struct ggml_tensor* y,
+                     ggml_tensor* y,
-                     struct ggml_tensor* guidance,
+                     ggml_tensor* guidance,
                      std::vector<ggml_tensor*> ref_latents = {},
                      bool increase_ref_index               = false,
-                     struct ggml_tensor** output           = nullptr,
+                     ggml_tensor** output                  = nullptr,
-                     struct ggml_context* output_ctx       = nullptr,
+                     ggml_context* output_ctx              = nullptr,
                      std::vector<int> skip_layers          = std::vector<int>()) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
             // y: [N, adm_in_channels] or [1, adm_in_channels]
             // guidance: [N, ]
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+            auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(x, timesteps, context, c_concat, y, guidance, ref_latents, increase_ref_index, skip_layers);
             };
 
@@ -1479,12 +1480,12 @@ namespace Flux {
         }
 
         void test() {
-            struct ggml_init_params params;
+            ggml_init_params params;
             params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            struct ggml_context* work_ctx = ggml_init(params);
+            ggml_context* work_ctx = ggml_init(params);
             GGML_ASSERT(work_ctx != nullptr);
 
             {
@@ -1512,7 +1513,7 @@ namespace Flux {
                 auto y = nullptr;
                 // print_ggml_tensor(y);
 
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 compute(8, x, timesteps, context, nullptr, y, guidance, {}, false, &out, work_ctx);

src/latent-preview.h

@@ -163,7 +163,7 @@ const float sd_latent_rgb_proj[4][3] = {
     {-0.178022f, -0.200862f, -0.678514f}};
 float sd_latent_rgb_bias[3] = {-0.017478f, -0.055834f, -0.105825f};
 
-void preview_latent_video(uint8_t* buffer, struct ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
+void preview_latent_video(uint8_t* buffer, ggml_tensor* latents, const float (*latent_rgb_proj)[3], const float latent_rgb_bias[3], int patch_size) {
     size_t buffer_head = 0;
 
     uint32_t latent_width  = static_cast<uint32_t>(latents->ne[0]);

src/llm.hpp

@@ -522,7 +522,7 @@ namespace LLM {
             blocks["down_proj"] = std::shared_ptr<GGMLBlock>(new Linear(intermediate_size, hidden_size, bias));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             // x: [N, n_token, hidden_size]
             auto gate_proj = std::dynamic_pointer_cast<Linear>(blocks["gate_proj"]);
             auto up_proj   = std::dynamic_pointer_cast<Linear>(blocks["up_proj"]);
@@ -582,7 +582,7 @@ namespace LLM {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             // x: [N*grid_t*grid_h*grid_w, in_channels, temporal_patch_size*patch_size*patch_size]
             // return: [N*grid_t*grid_h*grid_w, embed_dim]
             x = ggml_reshape_4d(ctx->ggml_ctx,
@@ -631,7 +631,7 @@ namespace LLM {
             blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, dim));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto ln_q  = std::dynamic_pointer_cast<RMSNorm>(blocks["ln_q"]);
             auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
             auto mlp_2 = std::dynamic_pointer_cast<Linear>(blocks["mlp.2"]);
@@ -668,10 +668,10 @@ namespace LLM {
             blocks["proj"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* mask = nullptr) {
+                             ggml_tensor* mask = nullptr) {
             // x: [N, n_token, hidden_size]
             int64_t n_token = x->ne[1];
             int64_t N       = x->ne[2];
@@ -718,10 +718,10 @@ namespace LLM {
             blocks["norm2"] = std::shared_ptr<GGMLBlock>(new RMSNorm(hidden_size, eps));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* mask = nullptr) {
+                             ggml_tensor* mask = nullptr) {
             // x: [N, n_token, hidden_size]
             auto attn  = std::dynamic_pointer_cast<VisionAttention>(blocks["attn"]);
             auto mlp   = std::dynamic_pointer_cast<MLP>(blocks["mlp"]);
@@ -778,12 +778,12 @@ namespace LLM {
             blocks["merger"] = std::shared_ptr<GGMLBlock>(new PatchMerger(out_hidden_size, hidden_size, spatial_merge_size));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* pixel_values,
+                             ggml_tensor* pixel_values,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* window_index,
+                             ggml_tensor* window_index,
-                                    struct ggml_tensor* window_inverse_index,
+                             ggml_tensor* window_inverse_index,
-                                    struct ggml_tensor* window_mask) {
+                             ggml_tensor* window_mask) {
             // pixel_values: [grid_t*(H/mh/ph)*(W/mw/pw)*mh*mw, C*pt*ph*pw]
             // window_index: [grid_t*(H/mh/ph)*(W/mw/pw)]
             // window_inverse_index: [grid_t*(H/mh/ph)*(W/mw/pw)]
@@ -836,10 +836,10 @@ namespace LLM {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* input_pos,
+                             ggml_tensor* input_pos,
-                                    struct ggml_tensor* attention_mask = nullptr) {
+                             ggml_tensor* attention_mask = nullptr) {
             // x: [N, n_token, hidden_size]
             int64_t n_token = x->ne[1];
             int64_t N       = x->ne[2];
@@ -898,10 +898,10 @@ namespace LLM {
             blocks["post_attention_layernorm"] = std::make_shared<RMSNorm>(params.hidden_size, params.rms_norm_eps);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* input_pos,
+                             ggml_tensor* input_pos,
-                                    struct ggml_tensor* attention_mask = nullptr) {
+                             ggml_tensor* attention_mask = nullptr) {
             // x: [N, n_token, hidden_size]
             auto self_attn                = std::dynamic_pointer_cast<Attention>(blocks["self_attn"]);
             auto mlp                      = std::dynamic_pointer_cast<MLP>(blocks["mlp"]);
@@ -936,12 +936,12 @@ namespace LLM {
             blocks["norm"] = std::shared_ptr<GGMLBlock>(new RMSNorm(params.hidden_size, params.rms_norm_eps));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* input_ids,
+                             ggml_tensor* input_ids,
-                                    struct ggml_tensor* input_pos,
+                             ggml_tensor* input_pos,
-                                    struct ggml_tensor* attention_mask,
+                             ggml_tensor* attention_mask,
-                                    std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+                             std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                                    std::set<int> out_layers) {
+                             std::set<int> out_layers) {
             // input_ids: [N, n_token]
             // return: [N, n_token, hidden_size]
 
@@ -1037,12 +1037,12 @@ namespace LLM {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* input_ids,
+                             ggml_tensor* input_ids,
-                                    struct ggml_tensor* input_pos,
+                             ggml_tensor* input_pos,
-                                    struct ggml_tensor* attention_mask,
+                             ggml_tensor* attention_mask,
-                                    std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+                             std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                                    std::set<int> out_layers) {
+                             std::set<int> out_layers) {
             // input_ids: [N, n_token]
             auto model = std::dynamic_pointer_cast<TextModel>(blocks["model"]);
 
@@ -1050,12 +1050,12 @@ namespace LLM {
             return x;
         }
 
-        struct ggml_tensor* vision_forward(GGMLRunnerContext* ctx,
+        ggml_tensor* vision_forward(GGMLRunnerContext* ctx,
-                                           struct ggml_tensor* pixel_values,
+                                    ggml_tensor* pixel_values,
-                                           struct ggml_tensor* pe,
+                                    ggml_tensor* pe,
-                                           struct ggml_tensor* window_index,
+                                    ggml_tensor* window_index,
-                                           struct ggml_tensor* window_inverse_index,
+                                    ggml_tensor* window_inverse_index,
-                                           struct ggml_tensor* window_mask) {
+                                    ggml_tensor* window_mask) {
             GGML_ASSERT(enable_vision);
             auto vision_model = std::dynamic_pointer_cast<VisionModel>(blocks["visual"]);
             return vision_model->forward(ctx, pixel_values, pe, window_index, window_inverse_index, window_mask);
@@ -1156,35 +1156,35 @@ namespace LLM {
             return llm_arch_to_str[static_cast<int>(params.arch)];
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
             model.get_param_tensors(tensors, prefix);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* input_ids,
+                             ggml_tensor* input_ids,
-                                    struct ggml_tensor* input_pos,
+                             ggml_tensor* input_pos,
-                                    struct ggml_tensor* attention_mask,
+                             ggml_tensor* attention_mask,
-                                    std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+                             std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                                    std::set<int> out_layers) {
+                             std::set<int> out_layers) {
             auto hidden_states = model.forward(ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);  // [N, n_token, hidden_size]
             return hidden_states;
         }
 
-        struct ggml_tensor* vision_forward(GGMLRunnerContext* ctx,
+        ggml_tensor* vision_forward(GGMLRunnerContext* ctx,
-                                           struct ggml_tensor* pixel_values,
+                                    ggml_tensor* pixel_values,
-                                           struct ggml_tensor* input_pos,
+                                    ggml_tensor* input_pos,
-                                           struct ggml_tensor* window_index,
+                                    ggml_tensor* window_index,
-                                           struct ggml_tensor* window_inverse_index,
+                                    ggml_tensor* window_inverse_index,
-                                           struct ggml_tensor* window_mask) {
+                                    ggml_tensor* window_mask) {
             auto hidden_states = model.vision_forward(ctx, pixel_values, input_pos, window_index, window_inverse_index, window_mask);
             return hidden_states;
         }
 
-        struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
+        ggml_cgraph* build_graph(ggml_tensor* input_ids,
-                                        struct ggml_tensor* attention_mask,
+                                 ggml_tensor* attention_mask,
-                                        std::vector<std::pair<int, ggml_tensor*>> image_embeds,
+                                 std::vector<std::pair<int, ggml_tensor*>> image_embeds,
-                                        std::set<int> out_layers) {
+                                 std::set<int> out_layers) {
-            struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+            ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
             input_ids = to_backend(input_ids);
 
@@ -1232,7 +1232,7 @@ namespace LLM {
 
             auto runner_ctx = get_context();
 
-            struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);
+            ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, input_pos, attention_mask, image_embeds, out_layers);
 
             ggml_build_forward_expand(gf, hidden_states);
 
@@ -1240,13 +1240,13 @@ namespace LLM {
         }
 
         bool compute(const int n_threads,
-                     struct ggml_tensor* input_ids,
+                     ggml_tensor* input_ids,
-                     struct ggml_tensor* attention_mask,
+                     ggml_tensor* attention_mask,
                      std::vector<std::pair<int, ggml_tensor*>> image_embeds,
                      std::set<int> out_layers,
                      ggml_tensor** output,
                      ggml_context* output_ctx = nullptr) {
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+            auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(input_ids, attention_mask, image_embeds, out_layers);
             };
             return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
@@ -1261,7 +1261,7 @@ namespace LLM {
             return grid_t * grid_h * grid_w;
         }
 
-        struct ggml_tensor* process_image(struct ggml_context* ctx, struct ggml_tensor* image) {
+        ggml_tensor* process_image(ggml_context* ctx, ggml_tensor* image) {
             // image: [C, H, W]
             // return: [grid_t*(H/mh/ph)*(W/mw/pw)*mh*mw, C*pt*ph*pw], grid_t == 1
             int64_t C  = image->ne[2];
@@ -1288,8 +1288,8 @@ namespace LLM {
             return image;
         }
 
-        struct ggml_cgraph* build_encode_image_graph(struct ggml_tensor* image) {
+        ggml_cgraph* build_encode_image_graph(ggml_tensor* image) {
-            struct ggml_cgraph* gf = new_graph_custom(LLM_GRAPH_SIZE);
+            ggml_cgraph* gf = new_graph_custom(LLM_GRAPH_SIZE);
 
             GGML_ASSERT(image->ne[1] % (params.vision.patch_size * params.vision.spatial_merge_size) == 0);
             GGML_ASSERT(image->ne[0] % (params.vision.patch_size * params.vision.spatial_merge_size) == 0);
@@ -1399,23 +1399,23 @@ namespace LLM {
             // pe->data = nullptr;
             set_backend_tensor_data(pe, pe_vec.data());
 
-            auto runnter_ctx                  = get_context();
+            auto runnter_ctx           = get_context();
-            struct ggml_tensor* hidden_states = vision_forward(&runnter_ctx,
+            ggml_tensor* hidden_states = vision_forward(&runnter_ctx,
-                                                               pixel_values,
+                                                        pixel_values,
-                                                               pe,
+                                                        pe,
-                                                               window_index,
+                                                        window_index,
-                                                               window_inverse_index,
+                                                        window_inverse_index,
-                                                               window_mask);
+                                                        window_mask);
             ggml_build_forward_expand(gf, hidden_states);
 
             return gf;
         }
 
         void encode_image(const int n_threads,
-                          struct ggml_tensor* image,
+                          ggml_tensor* image,
                           ggml_tensor** output,
                           ggml_context* output_ctx = nullptr) {
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+            auto get_graph = [&]() -> ggml_cgraph* {
                 return build_encode_image_graph(image);
             };
             GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
@@ -1440,7 +1440,7 @@ namespace LLM {
             }
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
             model.get_param_tensors(tensors, prefix);
         }
 
@@ -1492,12 +1492,12 @@ namespace LLM {
         }
 
         void test() {
-            struct ggml_init_params params;
+            ggml_init_params params;
             params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            struct ggml_context* work_ctx = ggml_init(params);
+            ggml_context* work_ctx = ggml_init(params);
             GGML_ASSERT(work_ctx != nullptr);
             bool test_mistral          = false;
             bool test_qwen3            = true;
@@ -1509,7 +1509,7 @@ namespace LLM {
                 {
                     auto image = load_tensor_from_file(work_ctx, "qwen2vl_normalized.bin");
                     print_ggml_tensor(image, false, "image");
-                    struct ggml_tensor* out = nullptr;
+                    ggml_tensor* out = nullptr;
 
                     int64_t t0 = ggml_time_ms();
                     model.encode_image(8, image, &out, work_ctx);
@@ -1547,8 +1547,8 @@ namespace LLM {
                     printf("%d ", token);
                 }
                 printf("\n");
-                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
+                auto input_ids   = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 model.compute(8, input_ids, nullptr, image_embeds, {}, &out, work_ctx);
@@ -1561,7 +1561,7 @@ namespace LLM {
                 // ggml_set_f32(image, 0.f);
                 auto image = load_tensor_from_file(work_ctx, "qwen2vl_normalized.bin");
                 print_ggml_tensor(image, false, "image");
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 model.encode_image(8, image, &out, work_ctx);
@@ -1587,8 +1587,8 @@ namespace LLM {
                     printf("%d ", token);
                 }
                 printf("\n");
-                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
+                auto input_ids   = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 model.compute(8, input_ids, nullptr, {}, {10, 20, 30}, &out, work_ctx);
@@ -1610,8 +1610,8 @@ namespace LLM {
                     printf("%d ", token);
                 }
                 printf("\n");
-                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
+                auto input_ids   = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 model.compute(8, input_ids, nullptr, {}, {35}, &out, work_ctx);
@@ -1633,8 +1633,8 @@ namespace LLM {
                     printf("%d ", token);
                 }
                 printf("\n");
-                auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
+                auto input_ids   = vector_to_ggml_tensor_i32(work_ctx, tokens);
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 model.compute(8, input_ids, nullptr, {}, {}, &out, work_ctx);

src/lora.hpp

@@ -9,7 +9,7 @@
 struct LoraModel : public GGMLRunner {
     std::string lora_id;
     float multiplier = 1.0f;
-    std::unordered_map<std::string, struct ggml_tensor*> lora_tensors;
+    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::string file_path;
@@ -76,13 +76,13 @@ struct LoraModel : public GGMLRunner {
         }
 
         for (const auto& pair : tensors_to_create) {
-            const auto& name         = pair.first;
+            const auto& name   = pair.first;
-            const auto& ts           = pair.second;
+            const auto& ts     = pair.second;
-            struct ggml_tensor* real = ggml_new_tensor(params_ctx,
+            ggml_tensor* real  = ggml_new_tensor(params_ctx,
-                                                       ts.type,
+                                                 ts.type,
-                                                       ts.n_dims,
+                                                 ts.n_dims,
-                                                       ts.ne);
+                                                 ts.ne);
-            lora_tensors[name]       = real;
+            lora_tensors[name] = real;
         }
 
         alloc_params_buffer();
@@ -337,10 +337,10 @@ struct LoraModel : public GGMLRunner {
             }
             scale_value *= multiplier;
 
-            struct ggml_tensor* updown_1 = ggml_ext_merge_lora(ctx, hada_1_down, hada_1_up, hada_1_mid);
+            ggml_tensor* updown_1 = ggml_ext_merge_lora(ctx, hada_1_down, hada_1_up, hada_1_mid);
-            struct ggml_tensor* updown_2 = ggml_ext_merge_lora(ctx, hada_2_down, hada_2_up, hada_2_mid);
+            ggml_tensor* updown_2 = ggml_ext_merge_lora(ctx, hada_2_down, hada_2_up, hada_2_mid);
-            auto curr_updown             = ggml_mul_inplace(ctx, updown_1, updown_2);
+            auto curr_updown      = ggml_mul_inplace(ctx, updown_1, updown_2);
-            curr_updown                  = ggml_ext_scale(ctx, curr_updown, scale_value, true);
+            curr_updown           = ggml_ext_scale(ctx, curr_updown, scale_value, true);
             if (updown == nullptr) {
                 updown = curr_updown;
             } else {
@@ -747,9 +747,9 @@ struct LoraModel : public GGMLRunner {
         return out_diff;
     }
 
-    struct ggml_cgraph* build_lora_graph(const std::map<std::string, ggml_tensor*>& model_tensors, SDVersion version) {
+    ggml_cgraph* build_lora_graph(const std::map<std::string, ggml_tensor*>& model_tensors, SDVersion version) {
         size_t lora_graph_size = LORA_GRAPH_BASE_SIZE + lora_tensors.size() * 10;
-        struct ggml_cgraph* gf = ggml_new_graph_custom(compute_ctx, lora_graph_size, false);
+        ggml_cgraph* gf        = ggml_new_graph_custom(compute_ctx, lora_graph_size, false);
 
         preprocess_lora_tensors(model_tensors);
 
@@ -788,8 +788,8 @@ struct LoraModel : public GGMLRunner {
         return gf;
     }
 
-    void apply(std::map<std::string, struct ggml_tensor*> model_tensors, SDVersion version, int n_threads) {
+    void apply(std::map<std::string, ggml_tensor*> model_tensors, SDVersion version, int n_threads) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_lora_graph(model_tensors, version);
         };
         GGMLRunner::compute(get_graph, n_threads, false);

src/ltxv.hpp

@@ -26,9 +26,9 @@ namespace LTXV {
                                                                      bias));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    bool causal = true) {
+                             bool causal = true) {
             // x: [N*IC, ID, IH, IW]
             // result: [N*OC, OD, OH, OW]
             auto conv = std::dynamic_pointer_cast<Conv3d>(blocks["conv"]);

src/mmdit.hpp

@@ -27,7 +27,7 @@ public:
         blocks["fc2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_features, out_features, bias));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [N, n_token, in_features]
         auto fc1 = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
         auto fc2 = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
@@ -72,7 +72,7 @@ public:
                                                                bias));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [N, C, H, W]
         // return: [N, H*W, embed_dim]
         auto proj = std::dynamic_pointer_cast<Conv2d>(blocks["proj"]);
@@ -111,7 +111,7 @@ public:
         blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, out_channels, true, true));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* t) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* t) {
         // t: [N, ]
         // return: [N, hidden_size]
         auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
@@ -135,7 +135,7 @@ public:
         blocks["mlp.2"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, hidden_size, true, true));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [N, input_dim]
         // return: [N, hidden_size]
         auto mlp_0 = std::dynamic_pointer_cast<Linear>(blocks["mlp.0"]);
@@ -175,7 +175,7 @@ public:
         }
     }
 
-    std::vector<struct ggml_tensor*> pre_attention(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    std::vector<ggml_tensor*> pre_attention(GGMLRunnerContext* ctx, ggml_tensor* x) {
         auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
 
         auto qkv         = qkv_proj->forward(ctx, x);
@@ -198,7 +198,7 @@ public:
         return {q, k, v};
     }
 
-    struct ggml_tensor* post_attention(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* post_attention(GGMLRunnerContext* ctx, ggml_tensor* x) {
         GGML_ASSERT(!pre_only);
 
         auto proj = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
@@ -208,8 +208,8 @@ public:
     }
 
     // x: [N, n_token, dim]
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x) {
+                         ggml_tensor* x) {
         auto qkv = pre_attention(ctx, x);
         x        = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, qkv[0], qkv[1], qkv[2], num_heads, nullptr, false, ctx->flash_attn_enabled);  // [N, n_token, dim]
         x        = post_attention(ctx, x);                                                                                                           // [N, n_token, dim]
@@ -217,10 +217,10 @@ public:
     }
 };
 
-__STATIC_INLINE__ struct ggml_tensor* modulate(struct ggml_context* ctx,
+__STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx,
-                                               struct ggml_tensor* x,
+                                        ggml_tensor* x,
-                                               struct ggml_tensor* shift,
+                                        ggml_tensor* shift,
-                                               struct ggml_tensor* scale) {
+                                        ggml_tensor* scale) {
     // x: [N, L, C]
     // scale: [N, C]
     // shift: [N, C]
@@ -274,8 +274,8 @@ public:
     }
 
     std::tuple<std::vector<ggml_tensor*>, std::vector<ggml_tensor*>, std::vector<ggml_tensor*>> pre_attention_x(GGMLRunnerContext* ctx,
-                                                                                                                struct ggml_tensor* x,
+                                                                                                                ggml_tensor* x,
-                                                                                                                struct ggml_tensor* c) {
+                                                                                                                ggml_tensor* c) {
         GGML_ASSERT(self_attn);
         // x: [N, n_token, hidden_size]
         // c: [N, hidden_size]
@@ -309,9 +309,9 @@ public:
         return {qkv, qkv2, {x, gate_msa, shift_mlp, scale_mlp, gate_mlp, gate_msa2}};
     }
 
-    std::pair<std::vector<struct ggml_tensor*>, std::vector<struct ggml_tensor*>> pre_attention(GGMLRunnerContext* ctx,
+    std::pair<std::vector<ggml_tensor*>, std::vector<ggml_tensor*>> pre_attention(GGMLRunnerContext* ctx,
-                                                                                                struct ggml_tensor* x,
+                                                                                  ggml_tensor* x,
-                                                                                                struct ggml_tensor* c) {
+                                                                                  ggml_tensor* c) {
         // x: [N, n_token, hidden_size]
         // c: [N, hidden_size]
         auto norm1              = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
@@ -346,15 +346,15 @@ public:
         }
     }
 
-    struct ggml_tensor* post_attention_x(GGMLRunnerContext* ctx,
+    ggml_tensor* post_attention_x(GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* attn_out,
+                                  ggml_tensor* attn_out,
-                                         struct ggml_tensor* attn2_out,
+                                  ggml_tensor* attn2_out,
-                                         struct ggml_tensor* x,
+                                  ggml_tensor* x,
-                                         struct ggml_tensor* gate_msa,
+                                  ggml_tensor* gate_msa,
-                                         struct ggml_tensor* shift_mlp,
+                                  ggml_tensor* shift_mlp,
-                                         struct ggml_tensor* scale_mlp,
+                                  ggml_tensor* scale_mlp,
-                                         struct ggml_tensor* gate_mlp,
+                                  ggml_tensor* gate_mlp,
-                                         struct ggml_tensor* gate_msa2) {
+                                  ggml_tensor* gate_msa2) {
         // attn_out: [N, n_token, hidden_size]
         // x: [N, n_token, hidden_size]
         // gate_msa: [N, hidden_size]
@@ -384,13 +384,13 @@ public:
         return x;
     }
 
-    struct ggml_tensor* post_attention(GGMLRunnerContext* ctx,
+    ggml_tensor* post_attention(GGMLRunnerContext* ctx,
-                                       struct ggml_tensor* attn_out,
+                                ggml_tensor* attn_out,
-                                       struct ggml_tensor* x,
+                                ggml_tensor* x,
-                                       struct ggml_tensor* gate_msa,
+                                ggml_tensor* gate_msa,
-                                       struct ggml_tensor* shift_mlp,
+                                ggml_tensor* shift_mlp,
-                                       struct ggml_tensor* scale_mlp,
+                                ggml_tensor* scale_mlp,
-                                       struct ggml_tensor* gate_mlp) {
+                                ggml_tensor* gate_mlp) {
         // attn_out: [N, n_token, hidden_size]
         // x: [N, n_token, hidden_size]
         // gate_msa: [N, hidden_size]
@@ -416,9 +416,9 @@ public:
         return x;
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* c) {
+                         ggml_tensor* c) {
         // x: [N, n_token, hidden_size]
         // c: [N, hidden_size]
         // return: [N, n_token, hidden_size]
@@ -463,11 +463,11 @@ public:
     }
 };
 
-__STATIC_INLINE__ std::pair<struct ggml_tensor*, struct ggml_tensor*>
+__STATIC_INLINE__ std::pair<ggml_tensor*, ggml_tensor*>
 block_mixing(GGMLRunnerContext* ctx,
-             struct ggml_tensor* context,
+             ggml_tensor* context,
-             struct ggml_tensor* x,
+             ggml_tensor* x,
-             struct ggml_tensor* c,
+             ggml_tensor* c,
              std::shared_ptr<DismantledBlock> context_block,
              std::shared_ptr<DismantledBlock> x_block) {
     // context: [N, n_context, hidden_size]
@@ -489,7 +489,7 @@ block_mixing(GGMLRunnerContext* ctx,
         x_qkv                    = x_qkv_intermediates.first;
         x_intermediates          = x_qkv_intermediates.second;
     }
-    std::vector<struct ggml_tensor*> qkv;
+    std::vector<ggml_tensor*> qkv;
     for (int i = 0; i < 3; i++) {
         qkv.push_back(ggml_concat(ctx->ggml_ctx, context_qkv[i], x_qkv[i], 1));
     }
@@ -563,10 +563,10 @@ public:
         blocks["x_block"]       = std::shared_ptr<GGMLBlock>(new DismantledBlock(hidden_size, num_heads, mlp_ratio, qk_norm, qkv_bias, false, self_attn_x));
     }
 
-    std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
+    std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                                struct ggml_tensor* context,
+                                                  ggml_tensor* context,
-                                                                struct ggml_tensor* x,
+                                                  ggml_tensor* x,
-                                                                struct ggml_tensor* c) {
+                                                  ggml_tensor* c) {
         auto context_block = std::dynamic_pointer_cast<DismantledBlock>(blocks["context_block"]);
         auto x_block       = std::dynamic_pointer_cast<DismantledBlock>(blocks["x_block"]);
 
@@ -586,9 +586,9 @@ public:
         blocks["adaLN_modulation.1"] = std::shared_ptr<GGMLBlock>(new Linear(hidden_size, 2 * hidden_size));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* c) {
+                         ggml_tensor* c) {
         // x: [N, n_token, hidden_size]
         // c: [N, hidden_size]
         // return: [N, n_token, patch_size * patch_size * out_channels]
@@ -626,7 +626,7 @@ protected:
     int64_t hidden_size;
     std::string qk_norm;
 
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
         enum ggml_type wtype = GGML_TYPE_F32;
         params["pos_embed"]  = ggml_new_tensor_3d(ctx, wtype, hidden_size, num_patchs, 1);
     }
@@ -705,8 +705,8 @@ public:
         blocks["final_layer"] = std::shared_ptr<GGMLBlock>(new FinalLayer(hidden_size, patch_size, out_channels));
     }
 
-    struct ggml_tensor*
+    ggml_tensor*
-    cropped_pos_embed(struct ggml_context* ctx,
+    cropped_pos_embed(ggml_context* ctx,
                       int64_t h,
                       int64_t w) {
         auto pos_embed = params["pos_embed"];
@@ -745,11 +745,11 @@ public:
         return spatial_pos_embed;
     }
 
-    struct ggml_tensor* forward_core_with_concat(GGMLRunnerContext* ctx,
+    ggml_tensor* forward_core_with_concat(GGMLRunnerContext* ctx,
-                                                 struct ggml_tensor* x,
+                                          ggml_tensor* x,
-                                                 struct ggml_tensor* c_mod,
+                                          ggml_tensor* c_mod,
-                                                 struct ggml_tensor* context,
+                                          ggml_tensor* context,
-                                                 std::vector<int> skip_layers = std::vector<int>()) {
+                                          std::vector<int> skip_layers = std::vector<int>()) {
         // x: [N, H*W, hidden_size]
         // context: [N, n_context, d_context]
         // c: [N, hidden_size]
@@ -774,12 +774,12 @@ public:
         return x;
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* t,
+                         ggml_tensor* t,
-                                struct ggml_tensor* y        = nullptr,
+                         ggml_tensor* y               = nullptr,
-                                struct ggml_tensor* context  = nullptr,
+                         ggml_tensor* context         = nullptr,
-                                std::vector<int> skip_layers = std::vector<int>()) {
+                         std::vector<int> skip_layers = std::vector<int>()) {
         // Forward pass of DiT.
         // x: (N, C, H, W) tensor of spatial inputs (images or latent representations of images)
         // t: (N,) tensor of diffusion timesteps
@@ -832,29 +832,29 @@ struct MMDiTRunner : public GGMLRunner {
         return "mmdit";
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
         mmdit.get_param_tensors(tensors, prefix);
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+    ggml_cgraph* build_graph(ggml_tensor* x,
-                                    struct ggml_tensor* timesteps,
+                             ggml_tensor* timesteps,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    struct ggml_tensor* y,
+                             ggml_tensor* y,
-                                    std::vector<int> skip_layers = std::vector<int>()) {
+                             std::vector<int> skip_layers = std::vector<int>()) {
-        struct ggml_cgraph* gf = new_graph_custom(MMDIT_GRAPH_SIZE);
+        ggml_cgraph* gf = new_graph_custom(MMDIT_GRAPH_SIZE);
 
         x         = to_backend(x);
         context   = to_backend(context);
         y         = to_backend(y);
         timesteps = to_backend(timesteps);
 
-        auto runner_ctx         = get_context();
+        auto runner_ctx  = get_context();
-        struct ggml_tensor* out = mmdit.forward(&runner_ctx,
+        ggml_tensor* out = mmdit.forward(&runner_ctx,
-                                                x,
+                                         x,
-                                                timesteps,
+                                         timesteps,
-                                                y,
+                                         y,
-                                                context,
+                                         context,
-                                                skip_layers);
+                                         skip_layers);
 
         ggml_build_forward_expand(gf, out);
 
@@ -862,18 +862,18 @@ struct MMDiTRunner : public GGMLRunner {
     }
 
     bool compute(int n_threads,
-                 struct ggml_tensor* x,
+                 ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
+                 ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
+                 ggml_tensor* context,
-                 struct ggml_tensor* y,
+                 ggml_tensor* y,
-                 struct ggml_tensor** output     = nullptr,
+                 ggml_tensor** output         = nullptr,
-                 struct ggml_context* output_ctx = nullptr,
+                 ggml_context* output_ctx     = nullptr,
-                 std::vector<int> skip_layers    = std::vector<int>()) {
+                 std::vector<int> skip_layers = std::vector<int>()) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 154, 4096]) or [1, max_position, hidden_size]
         // y: [N, adm_in_channels] or [1, adm_in_channels]
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(x, timesteps, context, y, skip_layers);
         };
 
@@ -881,12 +881,12 @@ struct MMDiTRunner : public GGMLRunner {
     }
 
     void test() {
-        struct ggml_init_params params;
+        ggml_init_params params;
         params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
         params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
-        struct ggml_context* work_ctx = ggml_init(params);
+        ggml_context* work_ctx = ggml_init(params);
         GGML_ASSERT(work_ctx != nullptr);
 
         {
@@ -908,7 +908,7 @@ struct MMDiTRunner : public GGMLRunner {
             ggml_set_f32(y, 0.01f);
             // print_ggml_tensor(y);
 
-            struct ggml_tensor* out = nullptr;
+            ggml_tensor* out = nullptr;
 
             int64_t t0 = ggml_time_ms();
             compute(8, x, timesteps, context, y, &out, work_ctx);

src/model.cpp

@@ -287,7 +287,7 @@ void ModelLoader::add_tensor_storage(const TensorStorage& tensor_storage) {
 }
 
 bool is_zip_file(const std::string& file_path) {
-    struct zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
+    zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
     if (zip == nullptr) {
         return false;
     }
@@ -453,9 +453,9 @@ bool ModelLoader::init_from_gguf_file(const std::string& file_path, const std::s
     size_t total_size  = 0;
     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);
+        std::string name   = gguf_get_tensor_name(ctx_gguf_, i);
-        struct ggml_tensor* dummy = ggml_get_tensor(ctx_meta_, name.c_str());
+        ggml_tensor* dummy = ggml_get_tensor(ctx_meta_, name.c_str());
-        size_t offset             = data_offset + gguf_get_tensor_offset(ctx_gguf_, i);
+        size_t offset      = data_offset + gguf_get_tensor_offset(ctx_gguf_, i);
 
         // LOG_DEBUG("%s", name.c_str());
 
@@ -812,7 +812,7 @@ struct PickleTensorReader {
         }
     }
 
-    void read_string(const std::string& str, struct zip_t* zip, std::string dir) {
+    void read_string(const std::string& str, zip_t* zip, std::string dir) {
         if (str == "storage") {
             read_global_type = true;
         } else if (str != "state_dict") {
@@ -995,7 +995,7 @@ bool ModelLoader::init_from_ckpt_file(const std::string& file_path, const std::s
     file_paths_.push_back(file_path);
     size_t file_index = file_paths_.size() - 1;
 
-    struct zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
+    zip_t* zip = zip_open(file_path.c_str(), 0, 'r');
     if (zip == nullptr) {
         LOG_ERROR("failed to open '%s'", file_path.c_str());
         return false;
@@ -1104,10 +1104,12 @@ SDVersion ModelLoader::get_sd_version() {
             tensor_storage.name.find("unet.mid_block.resnets.1.") != std::string::npos) {
             has_middle_block_1 = true;
         }
-        if (tensor_storage.name.find("model.diffusion_model.output_blocks.3.1.transformer_blocks.1") != std::string::npos) {
+        if (tensor_storage.name.find("model.diffusion_model.output_blocks.3.1.transformer_blocks.1") != std::string::npos ||
+            tensor_storage.name.find("unet.up_blocks.1.attentions.0.transformer_blocks.1") != std::string::npos) {
             has_output_block_311 = true;
         }
-        if (tensor_storage.name.find("model.diffusion_model.output_blocks.7.1") != std::string::npos) {
+        if (tensor_storage.name.find("model.diffusion_model.output_blocks.7.1") != std::string::npos ||
+            tensor_storage.name.find("unet.up_blocks.2.attentions.1") != std::string::npos) {
             has_output_block_71 = true;
         }
         if (tensor_storage.name == "cond_stage_model.transformer.text_model.embeddings.token_embedding.weight" ||
@@ -1411,7 +1413,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
         for (int i = 0; i < n_threads; ++i) {
             workers.emplace_back([&, file_path, is_zip]() {
                 std::ifstream file;
-                struct zip_t* zip = nullptr;
+                zip_t* zip = nullptr;
                 if (is_zip) {
                     zip = zip_open(file_path.c_str(), 0, 'r');
                     if (zip == nullptr) {
@@ -1599,7 +1601,7 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, int n_thread
     return success;
 }
 
-bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tensors,
+bool ModelLoader::load_tensors(std::map<std::string, ggml_tensor*>& tensors,
                                std::set<std::string> ignore_tensors,
                                int n_threads,
                                bool enable_mmap) {
@@ -1613,7 +1615,7 @@ bool ModelLoader::load_tensors(std::map<std::string, struct ggml_tensor*>& tenso
             tensor_names_in_file.insert(name);
         }
 
-        struct ggml_tensor* real;
+        ggml_tensor* real;
         if (tensors.find(name) != tensors.end()) {
             real = tensors[name];
         } else {

src/model.h

@@ -323,7 +323,7 @@ public:
     String2TensorStorage& get_tensor_storage_map() { return tensor_storage_map; }
     void set_wtype_override(ggml_type wtype, std::string tensor_type_rules = "");
     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, struct ggml_tensor*>& tensors,
+    bool load_tensors(std::map<std::string, ggml_tensor*>& tensors,
                       std::set<std::string> ignore_tensors = {},
                       int n_threads                        = 0,
                       bool use_mmap                        = false);

src/name_conversion.cpp

@@ -1120,7 +1120,11 @@ std::string convert_tensor_name(std::string name, SDVersion version) {
         for (const auto& prefix : first_stage_model_prefix_vec) {
             if (starts_with(name, prefix)) {
                 name = convert_first_stage_model_name(name.substr(prefix.size()), prefix);
-                name = prefix + name;
+                if (version == VERSION_SDXS) {
+                    name = "tae." + name;
+                } else {
+                    name = prefix + name;
+                }
                 break;
             }
         }

src/pmid.hpp

@@ -21,14 +21,14 @@ public:
         blocks["layernorm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(in_dim));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
         // x: [N, channels, h, w]
 
         auto fc1        = std::dynamic_pointer_cast<Linear>(blocks["fc1"]);
         auto fc2        = std::dynamic_pointer_cast<Linear>(blocks["fc2"]);
         auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layernorm"]);
 
-        struct ggml_tensor* r = x;
+        ggml_tensor* r = x;
         // x = ggml_ext_layer_norm(ctx, x, ln_w, ln_b);
         x = layer_norm->forward(ctx, x);
         // x = ggml_add(ctx, ggml_mul_mat(ctx, fc1_w, x),  fc1_b);
@@ -54,8 +54,8 @@ public:
         blocks["1"]   = std::shared_ptr<GGMLBlock>(new Mlp(dim, inner_dim, dim, false));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x) {
+                         ggml_tensor* x) {
         auto norm = std::dynamic_pointer_cast<LayerNorm>(blocks["0"]);
         auto ff   = std::dynamic_pointer_cast<Mlp>(blocks["1"]);
 
@@ -81,9 +81,9 @@ public:
         blocks["to_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, dim, false));
     }
 
-    struct ggml_tensor* reshape_tensor(struct ggml_context* ctx,
+    ggml_tensor* reshape_tensor(ggml_context* ctx,
-                                       struct ggml_tensor* x,
+                                ggml_tensor* x,
-                                       int heads) {
+                                int heads) {
         int64_t ne[4];
         for (int i = 0; i < 4; ++i)
             ne[i] = x->ne[i];
@@ -92,17 +92,17 @@ public:
         return x;
     }
 
-    std::vector<struct ggml_tensor*> chunk_half(struct ggml_context* ctx,
+    std::vector<ggml_tensor*> chunk_half(ggml_context* ctx,
-                                                struct ggml_tensor* x) {
+                                         ggml_tensor* x) {
         auto tlo = ggml_view_4d(ctx, x, x->ne[0] / 2, x->ne[1], x->ne[2], x->ne[3], x->nb[1], x->nb[2], x->nb[3], 0);
         auto tli = ggml_view_4d(ctx, x, x->ne[0] / 2, x->ne[1], x->ne[2], x->ne[3], x->nb[1], x->nb[2], x->nb[3], x->nb[0] * x->ne[0] / 2);
         return {ggml_cont(ctx, tlo),
                 ggml_cont(ctx, tli)};
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* latents) {
+                         ggml_tensor* latents) {
         // x (torch.Tensor): image features
         //     shape (b, n1, D)
         // latent (torch.Tensor): latent features
@@ -176,9 +176,9 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* latents,
+                         ggml_tensor* latents,
-                                struct ggml_tensor* x) {
+                         ggml_tensor* x) {
         // x: [N, channels, h, w]
         auto proj_in  = std::dynamic_pointer_cast<Linear>(blocks["proj_in"]);
         auto proj_out = std::dynamic_pointer_cast<Linear>(blocks["proj_out"]);
@@ -225,19 +225,19 @@ public:
             4));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* last_hidden_state) {
+                         ggml_tensor* last_hidden_state) {
         // x: [N, channels, h, w]
         auto token_proj          = std::dynamic_pointer_cast<Mlp>(blocks["token_proj"]);
         auto token_norm          = std::dynamic_pointer_cast<LayerNorm>(blocks["token_norm"]);
         auto perceiver_resampler = std::dynamic_pointer_cast<FacePerceiverResampler>(blocks["perceiver_resampler"]);
 
-        x                       = token_proj->forward(ctx, x);
+        x                = token_proj->forward(ctx, x);
-        int64_t nel             = ggml_nelements(x);
+        int64_t nel      = ggml_nelements(x);
-        x                       = ggml_reshape_3d(ctx->ggml_ctx, x, cross_attention_dim, num_tokens, nel / (cross_attention_dim * num_tokens));
+        x                = ggml_reshape_3d(ctx->ggml_ctx, x, cross_attention_dim, num_tokens, nel / (cross_attention_dim * num_tokens));
-        x                       = token_norm->forward(ctx, x);
+        x                = token_norm->forward(ctx, x);
-        struct ggml_tensor* out = perceiver_resampler->forward(ctx, x, last_hidden_state);
+        ggml_tensor* out = perceiver_resampler->forward(ctx, x, last_hidden_state);
         if (use_residul)
             out = ggml_add(ctx->ggml_ctx, x, out);
         return out;
@@ -256,9 +256,9 @@ public:
         blocks["layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(embed_dim));
     }
 
-    struct ggml_tensor* fuse_fn(GGMLRunnerContext* ctx,
+    ggml_tensor* fuse_fn(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* prompt_embeds,
+                         ggml_tensor* prompt_embeds,
-                                struct ggml_tensor* id_embeds) {
+                         ggml_tensor* id_embeds) {
         auto mlp1       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp1"]);
         auto mlp2       = std::dynamic_pointer_cast<FuseBlock>(blocks["mlp2"]);
         auto layer_norm = std::dynamic_pointer_cast<LayerNorm>(blocks["layer_norm"]);
@@ -273,24 +273,24 @@ public:
         return stacked_id_embeds;
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* prompt_embeds,
+                         ggml_tensor* prompt_embeds,
-                                struct ggml_tensor* id_embeds,
+                         ggml_tensor* id_embeds,
-                                struct ggml_tensor* class_tokens_mask,
+                         ggml_tensor* class_tokens_mask,
-                                struct ggml_tensor* class_tokens_mask_pos,
+                         ggml_tensor* class_tokens_mask_pos,
-                                struct ggml_tensor* left,
+                         ggml_tensor* left,
-                                struct ggml_tensor* right) {
+                         ggml_tensor* right) {
         // x: [N, channels, h, w]
 
-        struct ggml_tensor* valid_id_embeds = id_embeds;
+        ggml_tensor* valid_id_embeds = id_embeds;
         // # slice out the image token embeddings
         ggml_set_name(class_tokens_mask_pos, "class_tokens_mask_pos");
         ggml_set_name(prompt_embeds, "prompt_embeds");
-        struct ggml_tensor* image_token_embeds = ggml_get_rows(ctx->ggml_ctx, prompt_embeds, class_tokens_mask_pos);
+        ggml_tensor* image_token_embeds = ggml_get_rows(ctx->ggml_ctx, prompt_embeds, class_tokens_mask_pos);
         ggml_set_name(image_token_embeds, "image_token_embeds");
-        valid_id_embeds                       = ggml_reshape_2d(ctx->ggml_ctx, valid_id_embeds, valid_id_embeds->ne[0],
+        valid_id_embeds                = ggml_reshape_2d(ctx->ggml_ctx, valid_id_embeds, valid_id_embeds->ne[0],
-                                                                ggml_nelements(valid_id_embeds) / valid_id_embeds->ne[0]);
+                                                         ggml_nelements(valid_id_embeds) / valid_id_embeds->ne[0]);
-        struct ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);
+        ggml_tensor* stacked_id_embeds = fuse_fn(ctx, image_token_embeds, valid_id_embeds);
 
         if (left && right) {
             stacked_id_embeds = ggml_concat(ctx->ggml_ctx, left, stacked_id_embeds, 1);
@@ -301,10 +301,10 @@ public:
             stacked_id_embeds = ggml_concat(ctx->ggml_ctx, stacked_id_embeds, right, 1);
         }
 
-        class_tokens_mask                         = ggml_cont(ctx->ggml_ctx, ggml_transpose(ctx->ggml_ctx, class_tokens_mask));
+        class_tokens_mask                  = ggml_cont(ctx->ggml_ctx, ggml_transpose(ctx->ggml_ctx, class_tokens_mask));
-        class_tokens_mask                         = ggml_repeat(ctx->ggml_ctx, class_tokens_mask, prompt_embeds);
+        class_tokens_mask                  = ggml_repeat(ctx->ggml_ctx, class_tokens_mask, prompt_embeds);
-        prompt_embeds                             = ggml_mul(ctx->ggml_ctx, prompt_embeds, class_tokens_mask);
+        prompt_embeds                      = ggml_mul(ctx->ggml_ctx, prompt_embeds, class_tokens_mask);
-        struct ggml_tensor* updated_prompt_embeds = ggml_add(ctx->ggml_ctx, prompt_embeds, stacked_id_embeds);
+        ggml_tensor* updated_prompt_embeds = ggml_add(ctx->ggml_ctx, prompt_embeds, stacked_id_embeds);
         ggml_set_name(updated_prompt_embeds, "updated_prompt_embeds");
         return updated_prompt_embeds;
     }
@@ -317,22 +317,22 @@ struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
         blocks["fuse_module"]         = std::shared_ptr<GGMLBlock>(new FuseModule(2048));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* id_pixel_values,
+                         ggml_tensor* id_pixel_values,
-                                struct ggml_tensor* prompt_embeds,
+                         ggml_tensor* prompt_embeds,
-                                struct ggml_tensor* class_tokens_mask,
+                         ggml_tensor* class_tokens_mask,
-                                struct ggml_tensor* class_tokens_mask_pos,
+                         ggml_tensor* class_tokens_mask_pos,
-                                struct ggml_tensor* left,
+                         ggml_tensor* left,
-                                struct ggml_tensor* right) {
+                         ggml_tensor* right) {
         // x: [N, channels, h, w]
         auto vision_model        = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
         auto visual_projection   = std::dynamic_pointer_cast<CLIPProjection>(blocks["visual_projection"]);
         auto visual_projection_2 = std::dynamic_pointer_cast<Linear>(blocks["visual_projection_2"]);
         auto fuse_module         = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);
 
-        struct ggml_tensor* shared_id_embeds = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
+        ggml_tensor* shared_id_embeds = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
-        struct ggml_tensor* id_embeds        = visual_projection->forward(ctx, shared_id_embeds);    // [N, proj_dim(768)]
+        ggml_tensor* id_embeds        = visual_projection->forward(ctx, shared_id_embeds);    // [N, proj_dim(768)]
-        struct ggml_tensor* id_embeds_2      = visual_projection_2->forward(ctx, shared_id_embeds);  // [N, 1280]
+        ggml_tensor* id_embeds_2      = visual_projection_2->forward(ctx, shared_id_embeds);  // [N, 1280]
 
         id_embeds   = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, id_embeds, 2, 0, 1, 3));
         id_embeds_2 = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, id_embeds_2, 2, 0, 1, 3));
@@ -340,12 +340,12 @@ struct PhotoMakerIDEncoderBlock : public CLIPVisionModelProjection {
         id_embeds = ggml_concat(ctx->ggml_ctx, id_embeds, id_embeds_2, 2);  // [batch_size, seq_length, 1, 2048] check whether concat at dim 2 is right
         id_embeds = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, id_embeds, 1, 2, 0, 3));
 
-        struct ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
+        ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
-                                                                         prompt_embeds,
+                                                                  prompt_embeds,
-                                                                         id_embeds,
+                                                                  id_embeds,
-                                                                         class_tokens_mask,
+                                                                  class_tokens_mask,
-                                                                         class_tokens_mask_pos,
+                                                                  class_tokens_mask_pos,
-                                                                         left, right);
+                                                                  left, right);
         return updated_prompt_embeds;
     }
 };
@@ -365,29 +365,29 @@ struct PhotoMakerIDEncoder_CLIPInsightfaceExtendtokenBlock : public CLIPVisionMo
                                                                                         num_tokens));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* id_pixel_values,
+                         ggml_tensor* id_pixel_values,
-                                struct ggml_tensor* prompt_embeds,
+                         ggml_tensor* prompt_embeds,
-                                struct ggml_tensor* class_tokens_mask,
+                         ggml_tensor* class_tokens_mask,
-                                struct ggml_tensor* class_tokens_mask_pos,
+                         ggml_tensor* class_tokens_mask_pos,
-                                struct ggml_tensor* id_embeds,
+                         ggml_tensor* id_embeds,
-                                struct ggml_tensor* left,
+                         ggml_tensor* left,
-                                struct ggml_tensor* right) {
+                         ggml_tensor* right) {
         // x: [N, channels, h, w]
         auto vision_model      = std::dynamic_pointer_cast<CLIPVisionModel>(blocks["vision_model"]);
         auto fuse_module       = std::dynamic_pointer_cast<FuseModule>(blocks["fuse_module"]);
         auto qformer_perceiver = std::dynamic_pointer_cast<QFormerPerceiver>(blocks["qformer_perceiver"]);
 
-        // struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
+        // ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values);          // [N, hidden_size]
-        struct ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values, false);  // [N, hidden_size]
+        ggml_tensor* last_hidden_state = vision_model->forward(ctx, id_pixel_values, false);  // [N, hidden_size]
-        id_embeds                             = qformer_perceiver->forward(ctx, id_embeds, last_hidden_state);
+        id_embeds                      = qformer_perceiver->forward(ctx, id_embeds, last_hidden_state);
 
-        struct ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
+        ggml_tensor* updated_prompt_embeds = fuse_module->forward(ctx,
-                                                                         prompt_embeds,
+                                                                  prompt_embeds,
-                                                                         id_embeds,
+                                                                  id_embeds,
-                                                                         class_tokens_mask,
+                                                                  class_tokens_mask,
-                                                                         class_tokens_mask_pos,
+                                                                  class_tokens_mask_pos,
-                                                                         left, right);
+                                                                  left, right);
         return updated_prompt_embeds;
     }
 };
@@ -436,18 +436,18 @@ public:
         return pm_version;
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
         if (pm_version == PM_VERSION_1)
             id_encoder.get_param_tensors(tensors, prefix);
         else if (pm_version == PM_VERSION_2)
             id_encoder2.get_param_tensors(tensors, prefix);
     }
 
-    struct ggml_cgraph* build_graph(  // struct ggml_allocr* allocr,
+    ggml_cgraph* build_graph(  // ggml_allocr* allocr,
-        struct ggml_tensor* id_pixel_values,
+        ggml_tensor* id_pixel_values,
-        struct ggml_tensor* prompt_embeds,
+        ggml_tensor* prompt_embeds,
         std::vector<bool>& class_tokens_mask,
-        struct ggml_tensor* id_embeds) {
+        ggml_tensor* id_embeds) {
         ctm.clear();
         ctmf16.clear();
         ctmpos.clear();
@@ -458,20 +458,20 @@ public:
 
         auto runner_ctx = get_context();
 
-        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+        ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         int64_t hidden_size = prompt_embeds->ne[0];
         int64_t seq_length  = prompt_embeds->ne[1];
         ggml_type type      = GGML_TYPE_F32;
 
-        struct ggml_tensor* class_tokens_mask_d = ggml_new_tensor_1d(runner_ctx.ggml_ctx, type, class_tokens_mask.size());
+        ggml_tensor* class_tokens_mask_d = ggml_new_tensor_1d(runner_ctx.ggml_ctx, type, class_tokens_mask.size());
 
-        struct ggml_tensor* id_pixel_values_d = to_backend(id_pixel_values);
+        ggml_tensor* id_pixel_values_d = to_backend(id_pixel_values);
-        struct ggml_tensor* prompt_embeds_d   = to_backend(prompt_embeds);
+        ggml_tensor* prompt_embeds_d   = to_backend(prompt_embeds);
-        struct ggml_tensor* id_embeds_d       = to_backend(id_embeds);
+        ggml_tensor* id_embeds_d       = to_backend(id_embeds);
 
-        struct ggml_tensor* left  = nullptr;
+        ggml_tensor* left  = nullptr;
-        struct ggml_tensor* right = nullptr;
+        ggml_tensor* right = nullptr;
         for (int i = 0; i < class_tokens_mask.size(); i++) {
             if (class_tokens_mask[i]) {
                 // printf(" 1,");
@@ -495,7 +495,7 @@ public:
             right = ggml_new_tensor_3d(runner_ctx.ggml_ctx, type,
                                        hidden_size, seq_length - ctmpos[ctmpos.size() - 1] - 1, 1);
         }
-        struct ggml_tensor* class_tokens_mask_pos = ggml_new_tensor_1d(runner_ctx.ggml_ctx, GGML_TYPE_I32, ctmpos.size());
+        ggml_tensor* class_tokens_mask_pos = ggml_new_tensor_1d(runner_ctx.ggml_ctx, GGML_TYPE_I32, ctmpos.size());
 
         {
             if (type == GGML_TYPE_F16)
@@ -526,7 +526,7 @@ public:
                 }
             }
         }
-        struct ggml_tensor* updated_prompt_embeds = nullptr;
+        ggml_tensor* updated_prompt_embeds = nullptr;
         if (pm_version == PM_VERSION_1)
             updated_prompt_embeds = id_encoder.forward(&runner_ctx,
                                                        id_pixel_values_d,
@@ -549,13 +549,13 @@ public:
     }
 
     bool compute(const int n_threads,
-                 struct ggml_tensor* id_pixel_values,
+                 ggml_tensor* id_pixel_values,
-                 struct ggml_tensor* prompt_embeds,
+                 ggml_tensor* prompt_embeds,
-                 struct ggml_tensor* id_embeds,
+                 ggml_tensor* id_embeds,
                  std::vector<bool>& class_tokens_mask,
-                 struct ggml_tensor** updated_prompt_embeds,
+                 ggml_tensor** updated_prompt_embeds,
                  ggml_context* output_ctx) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             // return build_graph(compute_allocr, id_pixel_values, prompt_embeds, class_tokens_mask);
             return build_graph(id_pixel_values, prompt_embeds, class_tokens_mask, id_embeds);
         };
@@ -566,7 +566,7 @@ public:
 };
 
 struct PhotoMakerIDEmbed : public GGMLRunner {
-    std::map<std::string, struct ggml_tensor*> tensors;
+    std::map<std::string, ggml_tensor*> tensors;
     std::string file_path;
     ModelLoader* model_loader;
     bool load_failed = false;
@@ -606,11 +606,11 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
             }
             if (dry_run) {
                 std::lock_guard<std::mutex> lock(tensor_mutex);
-                struct ggml_tensor* real = ggml_new_tensor(params_ctx,
+                ggml_tensor* real = ggml_new_tensor(params_ctx,
-                                                           tensor_storage.type,
+                                                    tensor_storage.type,
-                                                           tensor_storage.n_dims,
+                                                    tensor_storage.n_dims,
-                                                           tensor_storage.ne);
+                                                    tensor_storage.ne);
-                tensors[name]            = real;
+                tensors[name]     = real;
             } else {
                 auto real   = tensors[name];
                 *dst_tensor = real;
@@ -629,8 +629,8 @@ struct PhotoMakerIDEmbed : public GGMLRunner {
         return true;
     }
 
-    struct ggml_tensor* get() {
+    ggml_tensor* get() {
-        std::map<std::string, struct ggml_tensor*>::iterator pos;
+        std::map<std::string, ggml_tensor*>::iterator pos;
         pos = tensors.find("pmid.id_embeds");
         if (pos != tensors.end())
             return pos->second;

src/preprocessing.hpp

@@ -4,13 +4,13 @@
 #include "ggml_extend.hpp"
 #define M_PI_ 3.14159265358979323846f
 
-void convolve(struct ggml_tensor* input, struct ggml_tensor* output, struct ggml_tensor* kernel, int padding) {
+void convolve(ggml_tensor* input, ggml_tensor* output, ggml_tensor* kernel, int padding) {
-    struct ggml_init_params params;
+    ggml_init_params params;
-    params.mem_size                 = 80 * input->ne[0] * input->ne[1];  // 20M for 512x512
+    params.mem_size          = 80 * input->ne[0] * input->ne[1];  // 20M for 512x512
-    params.mem_buffer               = nullptr;
+    params.mem_buffer        = nullptr;
-    params.no_alloc                 = false;
+    params.no_alloc          = false;
-    struct ggml_context* ctx0       = ggml_init(params);
+    ggml_context* ctx0       = ggml_init(params);
-    struct ggml_tensor* kernel_fp16 = ggml_new_tensor_4d(ctx0, GGML_TYPE_F16, kernel->ne[0], kernel->ne[1], 1, 1);
+    ggml_tensor* kernel_fp16 = ggml_new_tensor_4d(ctx0, GGML_TYPE_F16, kernel->ne[0], kernel->ne[1], 1, 1);
     ggml_fp32_to_fp16_row((float*)kernel->data, (ggml_fp16_t*)kernel_fp16->data, ggml_nelements(kernel));
     ggml_tensor* h  = ggml_conv_2d(ctx0, kernel_fp16, input, 1, 1, padding, padding, 1, 1);
     ggml_cgraph* gf = ggml_new_graph(ctx0);
@@ -19,7 +19,7 @@ void convolve(struct ggml_tensor* input, struct ggml_tensor* output, struct ggml
     ggml_free(ctx0);
 }
 
-void gaussian_kernel(struct ggml_tensor* kernel) {
+void gaussian_kernel(ggml_tensor* kernel) {
     int ks_mid   = static_cast<int>(kernel->ne[0] / 2);
     float sigma  = 1.4f;
     float normal = 1.f / (2.0f * M_PI_ * powf(sigma, 2.0f));
@@ -33,7 +33,7 @@ void gaussian_kernel(struct ggml_tensor* kernel) {
     }
 }
 
-void grayscale(struct ggml_tensor* rgb_img, struct ggml_tensor* grayscale) {
+void grayscale(ggml_tensor* rgb_img, ggml_tensor* grayscale) {
     for (int iy = 0; iy < rgb_img->ne[1]; iy++) {
         for (int ix = 0; ix < rgb_img->ne[0]; ix++) {
             float r    = ggml_ext_tensor_get_f32(rgb_img, ix, iy);
@@ -45,7 +45,7 @@ void grayscale(struct ggml_tensor* rgb_img, struct ggml_tensor* grayscale) {
     }
 }
 
-void prop_hypot(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor* h) {
+void prop_hypot(ggml_tensor* x, ggml_tensor* y, ggml_tensor* h) {
     int n_elements = static_cast<int>(ggml_nelements(h));
     float* dx      = (float*)x->data;
     float* dy      = (float*)y->data;
@@ -55,7 +55,7 @@ void prop_hypot(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor
     }
 }
 
-void prop_arctan2(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tensor* h) {
+void prop_arctan2(ggml_tensor* x, ggml_tensor* y, ggml_tensor* h) {
     int n_elements = static_cast<int>(ggml_nelements(h));
     float* dx      = (float*)x->data;
     float* dy      = (float*)y->data;
@@ -65,7 +65,7 @@ void prop_arctan2(struct ggml_tensor* x, struct ggml_tensor* y, struct ggml_tens
     }
 }
 
-void normalize_tensor(struct ggml_tensor* g) {
+void normalize_tensor(ggml_tensor* g) {
     int n_elements = static_cast<int>(ggml_nelements(g));
     float* dg      = (float*)g->data;
     float max      = -INFINITY;
@@ -78,7 +78,7 @@ void normalize_tensor(struct ggml_tensor* g) {
     }
 }
 
-void non_max_supression(struct ggml_tensor* result, struct ggml_tensor* G, struct ggml_tensor* D) {
+void non_max_supression(ggml_tensor* result, ggml_tensor* G, ggml_tensor* D) {
     for (int iy = 1; iy < result->ne[1] - 1; iy++) {
         for (int ix = 1; ix < result->ne[0] - 1; ix++) {
             float angle = ggml_ext_tensor_get_f32(D, ix, iy) * 180.0f / M_PI_;
@@ -117,7 +117,7 @@ void non_max_supression(struct ggml_tensor* result, struct ggml_tensor* G, struc
     }
 }
 
-void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float low_threshold, float weak, float strong) {
+void threshold_hystersis(ggml_tensor* img, float high_threshold, float low_threshold, float weak, float strong) {
     int n_elements = static_cast<int>(ggml_nelements(img));
     float* imd     = (float*)img->data;
     float max      = -INFINITY;
@@ -163,11 +163,11 @@ void threshold_hystersis(struct ggml_tensor* img, float high_threshold, float lo
 }
 
 bool preprocess_canny(sd_image_t img, float high_threshold, float low_threshold, float weak, float strong, bool inverse) {
-    struct ggml_init_params params;
+    ggml_init_params params;
-    params.mem_size               = static_cast<size_t>(40 * img.width * img.height);  // 10MB for 512x512
+    params.mem_size        = static_cast<size_t>(40 * img.width * img.height);  // 10MB for 512x512
-    params.mem_buffer             = nullptr;
+    params.mem_buffer      = nullptr;
-    params.no_alloc               = false;
+    params.no_alloc        = false;
-    struct ggml_context* work_ctx = ggml_init(params);
+    ggml_context* work_ctx = ggml_init(params);
 
     if (!work_ctx) {
         LOG_ERROR("ggml_init() failed");
@@ -185,19 +185,19 @@ bool preprocess_canny(sd_image_t img, float high_threshold, float low_threshold,
         -1, -2, -1};
 
     // generate kernel
-    int kernel_size             = 5;
+    int kernel_size      = 5;
-    struct ggml_tensor* gkernel = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, kernel_size, kernel_size, 1, 1);
+    ggml_tensor* gkernel = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, kernel_size, kernel_size, 1, 1);
-    struct ggml_tensor* sf_kx   = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
+    ggml_tensor* sf_kx   = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
     memcpy(sf_kx->data, kX, ggml_nbytes(sf_kx));
-    struct ggml_tensor* sf_ky = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
+    ggml_tensor* sf_ky = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 3, 3, 1, 1);
     memcpy(sf_ky->data, kY, ggml_nbytes(sf_ky));
     gaussian_kernel(gkernel);
-    struct ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 3, 1);
+    ggml_tensor* image      = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 3, 1);
-    struct ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 1, 1);
+    ggml_tensor* image_gray = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, img.width, img.height, 1, 1);
-    struct ggml_tensor* iX         = ggml_dup_tensor(work_ctx, image_gray);
+    ggml_tensor* iX         = ggml_dup_tensor(work_ctx, image_gray);
-    struct ggml_tensor* iY         = ggml_dup_tensor(work_ctx, image_gray);
+    ggml_tensor* iY         = ggml_dup_tensor(work_ctx, image_gray);
-    struct ggml_tensor* G          = ggml_dup_tensor(work_ctx, image_gray);
+    ggml_tensor* G          = ggml_dup_tensor(work_ctx, image_gray);
-    struct ggml_tensor* tetha      = ggml_dup_tensor(work_ctx, image_gray);
+    ggml_tensor* tetha      = ggml_dup_tensor(work_ctx, image_gray);
     sd_image_to_ggml_tensor(img, image);
     grayscale(image, image_gray);
     convolve(image_gray, image_gray, gkernel, 2);

src/qwen_image.hpp

@@ -26,9 +26,9 @@ namespace Qwen {
             blocks["linear_2"] = std::shared_ptr<GGMLBlock>(new Linear(time_embed_dim, out_dim, sample_proj_bias));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* sample,
+                             ggml_tensor* sample,
-                                    struct ggml_tensor* condition = nullptr) {
+                             ggml_tensor* condition = nullptr) {
             if (condition != nullptr) {
                 auto cond_proj = std::dynamic_pointer_cast<Linear>(blocks["cond_proj"]);
                 sample         = ggml_add(ctx->ggml_ctx, sample, cond_proj->forward(ctx, condition));
@@ -49,8 +49,8 @@ namespace Qwen {
             blocks["timestep_embedder"] = std::shared_ptr<GGMLBlock>(new TimestepEmbedding(256, embedding_dim));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* timesteps) {
+                             ggml_tensor* timesteps) {
             // timesteps: [N,]
             // return: [N, embedding_dim]
             auto timestep_embedder = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["timestep_embedder"]);
@@ -107,10 +107,10 @@ namespace Qwen {
         }
 
         std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                      struct ggml_tensor* img,
+                                                      ggml_tensor* img,
-                                                      struct ggml_tensor* txt,
+                                                      ggml_tensor* txt,
-                                                      struct ggml_tensor* pe,
+                                                      ggml_tensor* pe,
-                                                      struct ggml_tensor* mask = nullptr) {
+                                                      ggml_tensor* mask = nullptr) {
             // img: [N, n_img_token, hidden_size]
             // txt: [N, n_txt_token, hidden_size]
             // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
@@ -249,11 +249,11 @@ namespace Qwen {
         }
 
         virtual std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                              struct ggml_tensor* img,
+                                                              ggml_tensor* img,
-                                                              struct ggml_tensor* txt,
+                                                              ggml_tensor* txt,
-                                                              struct ggml_tensor* t_emb,
+                                                              ggml_tensor* t_emb,
-                                                              struct ggml_tensor* pe,
+                                                              ggml_tensor* pe,
-                                                              struct ggml_tensor* modulate_index = nullptr) {
+                                                              ggml_tensor* modulate_index = nullptr) {
             // img: [N, n_img_token, hidden_size]
             // txt: [N, n_txt_token, hidden_size]
             // pe: [n_img_token + n_txt_token, d_head/2, 2, 2]
@@ -325,9 +325,9 @@ namespace Qwen {
             blocks["linear"] = std::shared_ptr<GGMLBlock>(new Linear(conditioning_embedding_dim, embedding_dim * 2, bias));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* c) {
+                             ggml_tensor* c) {
             // x: [N, n_token, hidden_size]
             // c: [N, hidden_size]
             // return: [N, n_token, patch_size * patch_size * out_channels]
@@ -389,12 +389,12 @@ namespace Qwen {
             blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(inner_dim, params.patch_size * params.patch_size * params.out_channels));
         }
 
-        struct ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
+        ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
+                                  ggml_tensor* x,
-                                         struct ggml_tensor* timestep,
+                                  ggml_tensor* timestep,
-                                         struct ggml_tensor* context,
+                                  ggml_tensor* context,
-                                         struct ggml_tensor* pe,
+                                  ggml_tensor* pe,
-                                         struct ggml_tensor* modulate_index = nullptr) {
+                                  ggml_tensor* modulate_index = nullptr) {
             auto time_text_embed = std::dynamic_pointer_cast<QwenTimestepProjEmbeddings>(blocks["time_text_embed"]);
             auto txt_norm        = std::dynamic_pointer_cast<RMSNorm>(blocks["txt_norm"]);
             auto img_in          = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
@@ -404,7 +404,7 @@ namespace Qwen {
 
             auto t_emb = time_text_embed->forward(ctx, timestep);
             if (params.zero_cond_t) {
-                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros(ctx->ggml_ctx, timestep->ne[0], timestep->ne[1], timestep->ne[2], timestep->ne[3]));
+                auto t_emb_0 = time_text_embed->forward(ctx, ggml_ext_zeros_like(ctx->ggml_ctx, timestep));
                 t_emb        = ggml_concat(ctx->ggml_ctx, t_emb, t_emb_0, 1);
             }
             auto img = img_in->forward(ctx, x);
@@ -429,13 +429,13 @@ namespace Qwen {
             return img;
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* timestep,
+                             ggml_tensor* timestep,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    std::vector<ggml_tensor*> ref_latents = {},
+                             std::vector<ggml_tensor*> ref_latents = {},
-                                    struct ggml_tensor* modulate_index    = nullptr) {
+                             ggml_tensor* modulate_index           = nullptr) {
             // Forward pass of DiT.
             // x: [N, C, H, W]
             // timestep: [N,]
@@ -521,17 +521,17 @@ namespace Qwen {
             return "qwen_image";
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
             qwen_image.get_param_tensors(tensors, prefix);
         }
 
-        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+        ggml_cgraph* build_graph(ggml_tensor* x,
-                                        struct ggml_tensor* timesteps,
+                                 ggml_tensor* timesteps,
-                                        struct ggml_tensor* context,
+                                 ggml_tensor* context,
-                                        std::vector<ggml_tensor*> ref_latents = {},
+                                 std::vector<ggml_tensor*> ref_latents = {},
-                                        bool increase_ref_index               = false) {
+                                 bool increase_ref_index               = false) {
             GGML_ASSERT(x->ne[3] == 1);
-            struct ggml_cgraph* gf = new_graph_custom(QWEN_IMAGE_GRAPH_SIZE);
+            ggml_cgraph* gf = new_graph_custom(QWEN_IMAGE_GRAPH_SIZE);
 
             x         = to_backend(x);
             context   = to_backend(context);
@@ -587,13 +587,13 @@ namespace Qwen {
 
             auto runner_ctx = get_context();
 
-            struct ggml_tensor* out = qwen_image.forward(&runner_ctx,
+            ggml_tensor* out = qwen_image.forward(&runner_ctx,
-                                                         x,
+                                                  x,
-                                                         timesteps,
+                                                  timesteps,
-                                                         context,
+                                                  context,
-                                                         pe,
+                                                  pe,
-                                                         ref_latents,
+                                                  ref_latents,
-                                                         modulate_index);
+                                                  modulate_index);
 
             ggml_build_forward_expand(gf, out);
 
@@ -601,17 +601,17 @@ namespace Qwen {
         }
 
         bool compute(int n_threads,
-                     struct ggml_tensor* x,
+                     ggml_tensor* x,
-                     struct ggml_tensor* timesteps,
+                     ggml_tensor* timesteps,
-                     struct ggml_tensor* context,
+                     ggml_tensor* context,
                      std::vector<ggml_tensor*> ref_latents = {},
                      bool increase_ref_index               = false,
-                     struct ggml_tensor** output           = nullptr,
+                     ggml_tensor** output                  = nullptr,
-                     struct ggml_context* output_ctx       = nullptr) {
+                     ggml_context* output_ctx              = nullptr) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+            auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
             };
 
@@ -619,12 +619,12 @@ namespace Qwen {
         }
 
         void test() {
-            struct ggml_init_params params;
+            ggml_init_params params;
             params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            struct ggml_context* work_ctx = ggml_init(params);
+            ggml_context* work_ctx = ggml_init(params);
             GGML_ASSERT(work_ctx != nullptr);
 
             {
@@ -641,7 +641,7 @@ namespace Qwen {
                 auto context = load_tensor_from_file(work_ctx, "./qwen_image_context.bin");
                 print_ggml_tensor(context);
 
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 compute(8, x, timesteps, context, {}, false, &out, work_ctx);

src/rope.hpp

@@ -600,10 +600,10 @@ namespace Rope {
         return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
-    __STATIC_INLINE__ struct ggml_tensor* apply_rope(struct ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* apply_rope(ggml_context* ctx,
-                                                     struct ggml_tensor* x,
+                                              ggml_tensor* x,
-                                                     struct ggml_tensor* pe,
+                                              ggml_tensor* pe,
-                                                     bool rope_interleaved = true) {
+                                              bool rope_interleaved = true) {
         // x: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2], [[cos, -sin], [sin, cos]]
         int64_t d_head = x->ne[0];
@@ -641,14 +641,14 @@ namespace Rope {
         return x_out;
     }
 
-    __STATIC_INLINE__ struct ggml_tensor* attention(GGMLRunnerContext* ctx,
+    __STATIC_INLINE__ ggml_tensor* attention(GGMLRunnerContext* ctx,
-                                                    struct ggml_tensor* q,
+                                             ggml_tensor* q,
-                                                    struct ggml_tensor* k,
+                                             ggml_tensor* k,
-                                                    struct ggml_tensor* v,
+                                             ggml_tensor* v,
-                                                    struct ggml_tensor* pe,
+                                             ggml_tensor* pe,
-                                                    struct ggml_tensor* mask,
+                                             ggml_tensor* mask,
-                                                    float kv_scale        = 1.0f,
+                                             float kv_scale        = 1.0f,
-                                                    bool rope_interleaved = true) {
+                                             bool rope_interleaved = true) {
         // q,k,v: [N, L, n_head, d_head]
         // pe: [L, d_head/2, 2, 2]
         // return: [N, L, n_head*d_head]

src/spectrum.hpp

@@ -0,0 +1,195 @@
+#ifndef __SPECTRUM_HPP__
+#define __SPECTRUM_HPP__
+
+#include <cmath>
+#include <cstring>
+#include <vector>
+
+#include "ggml_extend.hpp"
+
+struct SpectrumConfig {
+    float w            = 0.40f;
+    int m              = 3;
+    float lam          = 1.0f;
+    int window_size    = 2;
+    float flex_window  = 0.50f;
+    int warmup_steps   = 4;
+    float stop_percent = 0.9f;
+};
+
+struct SpectrumState {
+    SpectrumConfig config;
+    int cnt                 = 0;
+    int num_cached          = 0;
+    float curr_ws           = 2.0f;
+    int K                   = 6;
+    int stop_step           = 0;
+    int total_steps_skipped = 0;
+
+    std::vector<std::vector<float>> H_buf;
+    std::vector<float> T_buf;
+
+    void init(const SpectrumConfig& cfg, size_t total_steps) {
+        config              = cfg;
+        cnt                 = 0;
+        num_cached          = 0;
+        curr_ws             = (float)cfg.window_size;
+        K                   = std::max(cfg.m + 1, 6);
+        stop_step           = (int)(cfg.stop_percent * (float)total_steps);
+        total_steps_skipped = 0;
+        H_buf.clear();
+        T_buf.clear();
+    }
+
+    float taus(int step_cnt) const {
+        return (step_cnt / 50.0f) * 2.0f - 1.0f;
+    }
+
+    bool should_predict() {
+        if (cnt < config.warmup_steps)
+            return false;
+        if (stop_step > 0 && cnt >= stop_step)
+            return false;
+        if ((int)H_buf.size() < 2)
+            return false;
+
+        int ws = std::max(1, (int)std::floor(curr_ws));
+        return (num_cached + 1) % ws != 0;
+    }
+
+    void update(const ggml_tensor* denoised) {
+        int64_t ne        = ggml_nelements(denoised);
+        const float* data = (const float*)denoised->data;
+
+        H_buf.emplace_back(data, data + ne);
+        T_buf.push_back(taus(cnt));
+
+        while ((int)H_buf.size() > K) {
+            H_buf.erase(H_buf.begin());
+            T_buf.erase(T_buf.begin());
+        }
+
+        if (cnt >= config.warmup_steps)
+            curr_ws += config.flex_window;
+
+        num_cached = 0;
+        cnt++;
+    }
+
+    void predict(ggml_tensor* denoised) {
+        int64_t F    = (int64_t)H_buf[0].size();
+        int K_curr   = (int)H_buf.size();
+        int M1       = config.m + 1;
+        float tau_at = taus(cnt);
+
+        // Design matrix X: K_curr x M1 (Chebyshev basis)
+        std::vector<float> X(K_curr * M1);
+        for (int i = 0; i < K_curr; i++) {
+            X[i * M1] = 1.0f;
+            if (M1 > 1)
+                X[i * M1 + 1] = T_buf[i];
+            for (int j = 2; j < M1; j++)
+                X[i * M1 + j] = 2.0f * T_buf[i] * X[i * M1 + j - 1] - X[i * M1 + j - 2];
+        }
+
+        // x_star: Chebyshev basis at current tau
+        std::vector<float> x_star(M1);
+        x_star[0] = 1.0f;
+        if (M1 > 1)
+            x_star[1] = tau_at;
+        for (int j = 2; j < M1; j++)
+            x_star[j] = 2.0f * tau_at * x_star[j - 1] - x_star[j - 2];
+
+        // XtX = X^T X + lambda I
+        std::vector<float> XtX(M1 * M1, 0.0f);
+        for (int i = 0; i < M1; i++) {
+            for (int j = 0; j < M1; j++) {
+                float sum = 0.0f;
+                for (int k = 0; k < K_curr; k++)
+                    sum += X[k * M1 + i] * X[k * M1 + j];
+                XtX[i * M1 + j] = sum + (i == j ? config.lam : 0.0f);
+            }
+        }
+
+        // Cholesky decomposition
+        std::vector<float> L(M1 * M1, 0.0f);
+        if (!cholesky_decompose(XtX.data(), L.data(), M1)) {
+            float trace = 0.0f;
+            for (int i = 0; i < M1; i++)
+                trace += XtX[i * M1 + i];
+            for (int i = 0; i < M1; i++)
+                XtX[i * M1 + i] += 1e-4f * trace / M1;
+            cholesky_decompose(XtX.data(), L.data(), M1);
+        }
+
+        // Solve XtX v = x_star
+        std::vector<float> v(M1);
+        cholesky_solve(L.data(), x_star.data(), v.data(), M1);
+
+        // Prediction weights per history entry
+        std::vector<float> weights(K_curr, 0.0f);
+        for (int k = 0; k < K_curr; k++)
+            for (int j = 0; j < M1; j++)
+                weights[k] += X[k * M1 + j] * v[j];
+
+        // Blend Chebyshev and Taylor predictions
+        float* out          = (float*)denoised->data;
+        float w_cheb        = config.w;
+        float w_taylor      = 1.0f - w_cheb;
+        const float* h_last = H_buf.back().data();
+        const float* h_prev = H_buf[H_buf.size() - 2].data();
+
+        for (int64_t f = 0; f < F; f++) {
+            float pred_cheb = 0.0f;
+            for (int k = 0; k < K_curr; k++)
+                pred_cheb += weights[k] * H_buf[k][f];
+
+            float pred_taylor = h_last[f] + 0.5f * (h_last[f] - h_prev[f]);
+
+            out[f] = w_taylor * pred_taylor + w_cheb * pred_cheb;
+        }
+
+        num_cached++;
+        total_steps_skipped++;
+        cnt++;
+    }
+
+private:
+    static bool cholesky_decompose(const float* A, float* L, int n) {
+        std::memset(L, 0, n * n * sizeof(float));
+        for (int i = 0; i < n; i++) {
+            for (int j = 0; j <= i; j++) {
+                float sum = 0.0f;
+                for (int k = 0; k < j; k++)
+                    sum += L[i * n + k] * L[j * n + k];
+                if (i == j) {
+                    float diag = A[i * n + i] - sum;
+                    if (diag <= 0.0f)
+                        return false;
+                    L[i * n + j] = std::sqrt(diag);
+                } else {
+                    L[i * n + j] = (A[i * n + j] - sum) / L[j * n + j];
+                }
+            }
+        }
+        return true;
+    }
+
+    static void cholesky_solve(const float* L, const float* b, float* x, int n) {
+        std::vector<float> y(n);
+        for (int i = 0; i < n; i++) {
+            float sum = 0.0f;
+            for (int j = 0; j < i; j++)
+                sum += L[i * n + j] * y[j];
+            y[i] = (b[i] - sum) / L[i * n + i];
+        }
+        for (int i = n - 1; i >= 0; i--) {
+            float sum = 0.0f;
+            for (int j = i + 1; j < n; j++)
+                sum += L[j * n + i] * x[j];
+            x[i] = (y[i] - sum) / L[i * n + i];
+        }
+    }
+};
+
+#endif  // __SPECTRUM_HPP__

src/t5.hpp

@@ -211,9 +211,9 @@ protected:
         // implementation. It's based on the following three ideas:
         //
         // 1. Because it uses the *unigram* model:
-        //     best_score(x1, x2, …, xt) = best_score(x1, x2, …, x{t-1}) + score(xt)
+        //     best_score(x1, x2, <EFBFBD>? xt) = best_score(x1, x2, <20>? x{t-1}) + score(xt)
         // Deciding the best path (and score) can be decoupled into two isolated
-        // terms: (a) the best path ended before the last token `best_score(x1, x2, …,
+        // terms: (a) the best path ended before the last token `best_score(x1, x2, <EFBFBD>?
         // x{t-1})`, and (b) the last token and its `score(xt)`. The two terms are
         // not related to each other at all.
         //
@@ -462,7 +462,7 @@ protected:
     int64_t hidden_size;
     float eps;
 
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+    void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
         enum ggml_type wtype = GGML_TYPE_F32;
         params["weight"]     = ggml_new_tensor_1d(ctx, wtype, hidden_size);
     }
@@ -473,10 +473,10 @@ public:
         : hidden_size(hidden_size),
           eps(eps) {}
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
-        struct ggml_tensor* w = params["weight"];
+        ggml_tensor* w = params["weight"];
-        x                     = ggml_rms_norm(ctx->ggml_ctx, x, eps);
+        x              = ggml_rms_norm(ctx->ggml_ctx, x, eps);
-        x                     = ggml_mul(ctx->ggml_ctx, x, w);
+        x              = ggml_mul(ctx->ggml_ctx, x, w);
         return x;
     }
 };
@@ -488,7 +488,7 @@ public:
         blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         // x: [N, n_token, model_dim]
         auto wi = std::dynamic_pointer_cast<Linear>(blocks["wi"]);
         auto wo = std::dynamic_pointer_cast<Linear>(blocks["wo"]);
@@ -510,7 +510,7 @@ public:
         blocks["wo"] = std::shared_ptr<GGMLBlock>(new Linear(ff_dim, model_dim, false, false, false, scale));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         // x: [N, n_token, model_dim]
         auto wi_0 = std::dynamic_pointer_cast<Linear>(blocks["wi_0"]);
         auto wi_1 = std::dynamic_pointer_cast<Linear>(blocks["wi_1"]);
@@ -531,7 +531,7 @@ public:
         blocks["layer_norm"]     = std::shared_ptr<GGMLBlock>(new T5LayerNorm(model_dim));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         // x: [N, n_token, model_dim]
         auto DenseReluDense = std::dynamic_pointer_cast<T5DenseGatedActDense>(blocks["DenseReluDense"]);
         auto layer_norm     = std::dynamic_pointer_cast<T5LayerNorm>(blocks["layer_norm"]);
@@ -570,8 +570,8 @@ public:
         }
     }
 
-    struct ggml_tensor* compute_bias(GGMLRunnerContext* ctx,
+    ggml_tensor* compute_bias(GGMLRunnerContext* ctx,
-                                     struct ggml_tensor* relative_position_bucket) {
+                              ggml_tensor* relative_position_bucket) {
         auto relative_attention_bias = std::dynamic_pointer_cast<Embedding>(blocks["relative_attention_bias"]);
 
         auto values = relative_attention_bias->forward(ctx, relative_position_bucket);            // shape (query_length, key_length, num_heads)
@@ -580,11 +580,11 @@ public:
     }
 
     // x: [N, n_token, model_dim]
-    std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
+    std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                                struct ggml_tensor* x,
+                                                  ggml_tensor* x,
-                                                                struct ggml_tensor* past_bias                = nullptr,
+                                                  ggml_tensor* past_bias                = nullptr,
-                                                                struct ggml_tensor* mask                     = nullptr,
+                                                  ggml_tensor* mask                     = nullptr,
-                                                                struct ggml_tensor* relative_position_bucket = nullptr) {
+                                                  ggml_tensor* relative_position_bucket = nullptr) {
         auto q_proj   = std::dynamic_pointer_cast<Linear>(blocks["q"]);
         auto k_proj   = std::dynamic_pointer_cast<Linear>(blocks["k"]);
         auto v_proj   = std::dynamic_pointer_cast<Linear>(blocks["v"]);
@@ -629,11 +629,11 @@ public:
         blocks["layer_norm"]    = std::shared_ptr<GGMLBlock>(new T5LayerNorm(model_dim));
     }
 
-    std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
+    std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                                struct ggml_tensor* x,
+                                                  ggml_tensor* x,
-                                                                struct ggml_tensor* past_bias                = nullptr,
+                                                  ggml_tensor* past_bias                = nullptr,
-                                                                struct ggml_tensor* mask                     = nullptr,
+                                                  ggml_tensor* mask                     = nullptr,
-                                                                struct ggml_tensor* relative_position_bucket = nullptr) {
+                                                  ggml_tensor* relative_position_bucket = nullptr) {
         // x: [N, n_token, model_dim]
         auto SelfAttention = std::dynamic_pointer_cast<T5Attention>(blocks["SelfAttention"]);
         auto layer_norm    = std::dynamic_pointer_cast<T5LayerNorm>(blocks["layer_norm"]);
@@ -655,11 +655,11 @@ public:
         blocks["layer.1"] = std::shared_ptr<GGMLBlock>(new T5LayerFF(model_dim, ff_dim));
     }
 
-    std::pair<struct ggml_tensor*, struct ggml_tensor*> forward(GGMLRunnerContext* ctx,
+    std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                                struct ggml_tensor* x,
+                                                  ggml_tensor* x,
-                                                                struct ggml_tensor* past_bias                = nullptr,
+                                                  ggml_tensor* past_bias                = nullptr,
-                                                                struct ggml_tensor* mask                     = nullptr,
+                                                  ggml_tensor* mask                     = nullptr,
-                                                                struct ggml_tensor* relative_position_bucket = nullptr) {
+                                                  ggml_tensor* relative_position_bucket = nullptr) {
         // x: [N, n_token, model_dim]
         auto layer_0 = std::dynamic_pointer_cast<T5LayerSelfAttention>(blocks["layer.0"]);
         auto layer_1 = std::dynamic_pointer_cast<T5LayerFF>(blocks["layer.1"]);
@@ -690,11 +690,11 @@ public:
         blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new T5LayerNorm(model_dim));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* past_bias                = nullptr,
+                         ggml_tensor* past_bias                = nullptr,
-                                struct ggml_tensor* attention_mask           = nullptr,
+                         ggml_tensor* attention_mask           = nullptr,
-                                struct ggml_tensor* relative_position_bucket = nullptr) {
+                         ggml_tensor* relative_position_bucket = nullptr) {
         // x: [N, n_token, model_dim]
         for (int i = 0; i < num_layers; i++) {
             auto block = std::dynamic_pointer_cast<T5Block>(blocks["block." + std::to_string(i)]);
@@ -737,11 +737,11 @@ public:
                                                                      params.model_dim));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* input_ids,
+                         ggml_tensor* input_ids,
-                                struct ggml_tensor* past_bias                = nullptr,
+                         ggml_tensor* past_bias                = nullptr,
-                                struct ggml_tensor* attention_mask           = nullptr,
+                         ggml_tensor* attention_mask           = nullptr,
-                                struct ggml_tensor* relative_position_bucket = nullptr) {
+                         ggml_tensor* relative_position_bucket = nullptr) {
         // input_ids: [N, n_token]
 
         auto shared  = std::dynamic_pointer_cast<Embedding>(blocks["shared"]);
@@ -776,14 +776,14 @@ struct T5Runner : public GGMLRunner {
         return "t5";
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
         model.get_param_tensors(tensors, prefix);
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* input_ids,
+                         ggml_tensor* input_ids,
-                                struct ggml_tensor* relative_position_bucket,
+                         ggml_tensor* relative_position_bucket,
-                                struct ggml_tensor* attention_mask = nullptr) {
+                         ggml_tensor* attention_mask = nullptr) {
         size_t N       = input_ids->ne[1];
         size_t n_token = input_ids->ne[0];
 
@@ -791,9 +791,9 @@ struct T5Runner : public GGMLRunner {
         return hidden_states;
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* input_ids,
+    ggml_cgraph* build_graph(ggml_tensor* input_ids,
-                                    struct ggml_tensor* attention_mask = nullptr) {
+                             ggml_tensor* attention_mask = nullptr) {
-        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
+        ggml_cgraph* gf = ggml_new_graph(compute_ctx);
 
         input_ids      = to_backend(input_ids);
         attention_mask = to_backend(attention_mask);
@@ -813,8 +813,8 @@ struct T5Runner : public GGMLRunner {
                                                            input_ids->ne[0]);
         set_backend_tensor_data(relative_position_bucket, relative_position_bucket_vec.data());
 
-        auto runner_ctx                   = get_context();
+        auto runner_ctx            = get_context();
-        struct ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, relative_position_bucket, attention_mask);
+        ggml_tensor* hidden_states = forward(&runner_ctx, input_ids, relative_position_bucket, attention_mask);
 
         ggml_build_forward_expand(gf, hidden_states);
 
@@ -822,11 +822,11 @@ struct T5Runner : public GGMLRunner {
     }
 
     bool compute(const int n_threads,
-                 struct ggml_tensor* input_ids,
+                 ggml_tensor* input_ids,
-                 struct ggml_tensor* attention_mask,
+                 ggml_tensor* attention_mask,
                  ggml_tensor** output,
                  ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(input_ids, attention_mask);
         };
         return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
@@ -912,7 +912,7 @@ struct T5Embedder {
         : model(backend, offload_params_to_cpu, tensor_storage_map, prefix, is_umt5), tokenizer(is_umt5) {
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
         model.get_param_tensors(tensors, prefix);
     }
 
@@ -962,17 +962,17 @@ struct T5Embedder {
     }
 
     void test() {
-        struct ggml_init_params params;
+        ggml_init_params params;
         params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
         params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
-        struct ggml_context* work_ctx = ggml_init(params);
+        ggml_context* work_ctx = ggml_init(params);
         GGML_ASSERT(work_ctx != nullptr);
 
         {
             std::string text("a lovely cat");
-            // std::string text("一只可爱的猫"); // umt5 chinease test
+            // std::string text("一只可爱的<EFBFBD>?); // umt5 chinease test
             auto tokens_and_weights     = tokenize(text, 512, true);
             std::vector<int>& tokens    = std::get<0>(tokens_and_weights);
             std::vector<float>& weights = std::get<1>(tokens_and_weights);
@@ -981,9 +981,9 @@ struct T5Embedder {
                 printf("%d ", token);
             }
             printf("\n");
-            auto input_ids          = vector_to_ggml_tensor_i32(work_ctx, tokens);
+            auto input_ids      = vector_to_ggml_tensor_i32(work_ctx, tokens);
-            auto attention_mask     = vector_to_ggml_tensor(work_ctx, masks);
+            auto attention_mask = vector_to_ggml_tensor(work_ctx, masks);
-            struct ggml_tensor* out = nullptr;
+            ggml_tensor* out    = nullptr;
 
             int64_t t0 = ggml_time_ms();
             model.compute(8, input_ids, attention_mask, &out, work_ctx);

src/tae.hpp

@@ -37,7 +37,7 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         // x: [n, n_in, h, w]
         // return: [n, n_out, h, w]
 
@@ -107,7 +107,7 @@ public:
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, z_channels, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         // x: [n, in_channels, h, w]
         // return: [n, z_channels, h/8, w/8]
 
@@ -157,7 +157,7 @@ public:
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
         // z: [n, z_channels, h, w]
         // return: [n, out_channels, h*8, w*8]
 
@@ -192,7 +192,7 @@ public:
         blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels * stride, channels, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks["conv"]);
         auto h    = x;
         if (stride != 1) {
@@ -212,7 +212,7 @@ public:
         blocks["conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(channels, channels * stride, {1, 1}, {1, 1}, {0, 0}, {1, 1}, false));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
         auto conv = std::dynamic_pointer_cast<UnaryBlock>(blocks["conv"]);
         auto h    = conv->forward(ctx, x);
         if (stride != 1) {
@@ -236,7 +236,7 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, struct ggml_tensor* past) {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* past) {
         // x: [n, channels, h, w]
         auto conv0 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.0"]);
         auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv.2"]);
@@ -260,10 +260,10 @@ public:
     }
 };
 
-struct ggml_tensor* patchify(struct ggml_context* ctx,
+ggml_tensor* patchify(ggml_context* ctx,
-                             struct ggml_tensor* x,
+                      ggml_tensor* x,
-                             int64_t patch_size,
+                      int64_t patch_size,
-                             int64_t b = 1) {
+                      int64_t b = 1) {
     // x: [f, b*c, h*q, w*r]
     // return: [f, b*c*r*q, h, w]
     if (patch_size == 1) {
@@ -289,10 +289,10 @@ struct ggml_tensor* patchify(struct ggml_context* ctx,
     return x;
 }
 
-struct ggml_tensor* unpatchify(struct ggml_context* ctx,
+ggml_tensor* unpatchify(ggml_context* ctx,
-                               struct ggml_tensor* x,
+                        ggml_tensor* x,
-                               int64_t patch_size,
+                        int64_t patch_size,
-                               int64_t b = 1) {
+                        int64_t b = 1) {
     // x: [f, b*c*r*q, h, w]
     // return: [f, b*c, h*q, w*r]
     if (patch_size == 1) {
@@ -339,7 +339,7 @@ public:
         blocks[std::to_string(index)] = std::shared_ptr<GGMLBlock>(new Conv2d(hidden, z_channels, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
         auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["0"]);
 
         if (patch_size > 1) {
@@ -396,7 +396,7 @@ public:
         blocks[std::to_string(index++)] = std::shared_ptr<GGMLBlock>(new Conv2d(channels[num_layers], out_channels * patch_size * patch_size, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) override {
+    ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* z) override {
         auto first_conv = std::dynamic_pointer_cast<Conv2d>(blocks["1"]);
 
         // Clamp()
@@ -442,11 +442,13 @@ protected:
     bool decode_only;
     SDVersion version;
 
+public:
+    int z_channels = 16;
+
 public:
     TAEHV(bool decode_only = true, SDVersion version = VERSION_WAN2)
         : decode_only(decode_only), version(version) {
-        int z_channels = 16;
+        int patch = 1;
-        int patch      = 1;
         if (version == VERSION_WAN2_2_TI2V) {
             z_channels = 48;
             patch      = 2;
@@ -457,7 +459,7 @@ public:
         }
     }
 
-    struct ggml_tensor* decode(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
+    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
         auto decoder = std::dynamic_pointer_cast<TinyVideoDecoder>(blocks["decoder"]);
         if (sd_version_is_wan(version)) {
             // (W, H, C, T) -> (W, H, T, C)
@@ -471,7 +473,7 @@ public:
         return result;
     }
 
-    struct ggml_tensor* encode(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
         auto encoder = std::dynamic_pointer_cast<TinyVideoEncoder>(blocks["encoder"]);
         // (W, H, T, C) -> (W, H, C, T)
         x                  = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 1, 3, 2));
@@ -494,10 +496,12 @@ protected:
     bool decode_only;
     bool taef2 = false;
 
+public:
+    int z_channels = 4;
+
 public:
     TAESD(bool decode_only = true, SDVersion version = VERSION_SD1)
         : decode_only(decode_only) {
-        int z_channels       = 4;
         bool use_midblock_gn = false;
         taef2                = sd_version_is_flux2(version);
 
@@ -515,7 +519,7 @@ public:
         }
     }
 
-    struct ggml_tensor* decode(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
+    ggml_tensor* decode(GGMLRunnerContext* ctx, ggml_tensor* z) {
         auto decoder = std::dynamic_pointer_cast<TinyDecoder>(blocks["decoder.layers"]);
         if (taef2) {
             z = unpatchify(ctx->ggml_ctx, z, 2);
@@ -523,7 +527,7 @@ public:
         return decoder->forward(ctx, z);
     }
 
-    struct ggml_tensor* encode(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+    ggml_tensor* encode(GGMLRunnerContext* ctx, ggml_tensor* x) {
         auto encoder = std::dynamic_pointer_cast<TinyEncoder>(blocks["encoder.layers"]);
         auto z       = encoder->forward(ctx, x);
         if (taef2) {
@@ -533,20 +537,7 @@ public:
     }
 };
 
-struct TinyAutoEncoder : public GGMLRunner {
+struct TinyImageAutoEncoder : public VAE {
-    TinyAutoEncoder(ggml_backend_t backend, bool offload_params_to_cpu)
-        : GGMLRunner(backend, offload_params_to_cpu) {}
-    virtual bool compute(const int n_threads,
-                         struct ggml_tensor* z,
-                         bool decode_graph,
-                         struct ggml_tensor** output,
-                         struct ggml_context* output_ctx = nullptr) = 0;
-
-    virtual bool load_from_file(const std::string& file_path, int n_threads)                                      = 0;
-    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) = 0;
-};
-
-struct TinyImageAutoEncoder : public TinyAutoEncoder {
     TAESD taesd;
     bool decode_only = false;
 
@@ -558,7 +549,8 @@ struct TinyImageAutoEncoder : public TinyAutoEncoder {
                          SDVersion version = VERSION_SD1)
         : decode_only(decoder_only),
           taesd(decoder_only, version),
-          TinyAutoEncoder(backend, offload_params_to_cpu) {
+          VAE(version, backend, offload_params_to_cpu) {
+        scale_input = false;
         taesd.init(params_ctx, tensor_storage_map, prefix);
     }
 
@@ -566,52 +558,41 @@ struct TinyImageAutoEncoder : public TinyAutoEncoder {
         return "taesd";
     }
 
-    bool load_from_file(const std::string& file_path, int n_threads) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
-        LOG_INFO("loading taesd from '%s', decode_only = %s", file_path.c_str(), decode_only ? "true" : "false");
-        alloc_params_buffer();
-        std::map<std::string, ggml_tensor*> taesd_tensors;
-        taesd.get_param_tensors(taesd_tensors);
-        std::set<std::string> ignore_tensors;
-        if (decode_only) {
-            ignore_tensors.insert("encoder.");
-        }
-
-        ModelLoader model_loader;
-        if (!model_loader.init_from_file_and_convert_name(file_path)) {
-            LOG_ERROR("init taesd model loader from file failed: '%s'", file_path.c_str());
-            return false;
-        }
-
-        bool success = model_loader.load_tensors(taesd_tensors, ignore_tensors, n_threads);
-
-        if (!success) {
-            LOG_ERROR("load tae tensors from model loader failed");
-            return false;
-        }
-
-        LOG_INFO("taesd model loaded");
-        return success;
-    }
-
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         taesd.get_param_tensors(tensors, prefix);
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
+    ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) {
-        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        return vae_output;
-        z                       = to_backend(z);
+    }
-        auto runner_ctx         = get_context();
+
-        struct ggml_tensor* out = decode_graph ? taesd.decode(&runner_ctx, z) : taesd.encode(&runner_ctx, z);
+    ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    }
+
+    ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    }
+
+    int get_encoder_output_channels(int input_channels) {
+        return taesd.z_channels;
+    }
+
+    ggml_cgraph* build_graph(ggml_tensor* z, bool decode_graph) {
+        ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        z                = to_backend(z);
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = decode_graph ? taesd.decode(&runner_ctx, z) : taesd.encode(&runner_ctx, z);
         ggml_build_forward_expand(gf, out);
         return gf;
     }
 
-    bool compute(const int n_threads,
+    bool _compute(const int n_threads,
-                 struct ggml_tensor* z,
+                  ggml_tensor* z,
-                 bool decode_graph,
+                  bool decode_graph,
-                 struct ggml_tensor** output,
+                  ggml_tensor** output,
-                 struct ggml_context* output_ctx = nullptr) {
+                  ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(z, decode_graph);
         };
 
@@ -619,7 +600,7 @@ struct TinyImageAutoEncoder : public TinyAutoEncoder {
     }
 };
 
-struct TinyVideoAutoEncoder : public TinyAutoEncoder {
+struct TinyVideoAutoEncoder : public VAE {
     TAEHV taehv;
     bool decode_only = false;
 
@@ -631,7 +612,8 @@ struct TinyVideoAutoEncoder : public TinyAutoEncoder {
                          SDVersion version = VERSION_WAN2)
         : decode_only(decoder_only),
           taehv(decoder_only, version),
-          TinyAutoEncoder(backend, offload_params_to_cpu) {
+          VAE(version, backend, offload_params_to_cpu) {
+        scale_input = false;
         taehv.init(params_ctx, tensor_storage_map, prefix);
     }
 
@@ -639,52 +621,41 @@ struct TinyVideoAutoEncoder : public TinyAutoEncoder {
         return "taehv";
     }
 
-    bool load_from_file(const std::string& file_path, int n_threads) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
-        LOG_INFO("loading taehv from '%s', decode_only = %s", file_path.c_str(), decode_only ? "true" : "false");
-        alloc_params_buffer();
-        std::map<std::string, ggml_tensor*> taehv_tensors;
-        taehv.get_param_tensors(taehv_tensors);
-        std::set<std::string> ignore_tensors;
-        if (decode_only) {
-            ignore_tensors.insert("encoder.");
-        }
-
-        ModelLoader model_loader;
-        if (!model_loader.init_from_file(file_path)) {
-            LOG_ERROR("init taehv model loader from file failed: '%s'", file_path.c_str());
-            return false;
-        }
-
-        bool success = model_loader.load_tensors(taehv_tensors, ignore_tensors, n_threads);
-
-        if (!success) {
-            LOG_ERROR("load tae tensors from model loader failed");
-            return false;
-        }
-
-        LOG_INFO("taehv model loaded");
-        return success;
-    }
-
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
         taehv.get_param_tensors(tensors, prefix);
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
+    ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) {
-        struct ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        return vae_output;
-        z                       = to_backend(z);
+    }
-        auto runner_ctx         = get_context();
+
-        struct ggml_tensor* out = decode_graph ? taehv.decode(&runner_ctx, z) : taehv.encode(&runner_ctx, z);
+    ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    }
+
+    ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    }
+
+    int get_encoder_output_channels(int input_channels) {
+        return taehv.z_channels;
+    }
+
+    ggml_cgraph* build_graph(ggml_tensor* z, bool decode_graph) {
+        ggml_cgraph* gf  = ggml_new_graph(compute_ctx);
+        z                = to_backend(z);
+        auto runner_ctx  = get_context();
+        ggml_tensor* out = decode_graph ? taehv.decode(&runner_ctx, z) : taehv.encode(&runner_ctx, z);
         ggml_build_forward_expand(gf, out);
         return gf;
     }
 
-    bool compute(const int n_threads,
+    bool _compute(const int n_threads,
-                 struct ggml_tensor* z,
+                  ggml_tensor* z,
-                 bool decode_graph,
+                  bool decode_graph,
-                 struct ggml_tensor** output,
+                  ggml_tensor** output,
-                 struct ggml_context* output_ctx = nullptr) {
+                  ggml_context* output_ctx = nullptr) {
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(z, decode_graph);
         };
 

src/ucache.hpp

@@ -19,6 +19,7 @@ struct UCacheConfig {
     bool adaptive_threshold     = true;
     float early_step_multiplier = 0.5f;
     float late_step_multiplier  = 1.5f;
+    float relative_norm_gain    = 1.6f;
     bool reset_error_on_compute = true;
 };
 
@@ -45,14 +46,16 @@ struct UCacheState {
     bool has_output_prev_norm             = false;
     bool has_relative_transformation_rate = false;
     float relative_transformation_rate    = 0.0f;
-    float cumulative_change_rate          = 0.0f;
     float last_input_change               = 0.0f;
     bool has_last_input_change            = false;
+    float output_change_ema               = 0.0f;
+    bool has_output_change_ema            = false;
     int total_steps_skipped               = 0;
     int current_step_index                = -1;
     int steps_computed_since_active       = 0;
+    int expected_total_steps              = 0;
+    int consecutive_skipped_steps         = 0;
     float accumulated_error               = 0.0f;
-    float reference_output_norm           = 0.0f;
 
     struct BlockMetrics {
         float sum_transformation_rate = 0.0f;
@@ -106,14 +109,16 @@ struct UCacheState {
         has_output_prev_norm             = false;
         has_relative_transformation_rate = false;
         relative_transformation_rate     = 0.0f;
-        cumulative_change_rate           = 0.0f;
         last_input_change                = 0.0f;
         has_last_input_change            = false;
+        output_change_ema                = 0.0f;
+        has_output_change_ema            = false;
         total_steps_skipped              = 0;
         current_step_index               = -1;
         steps_computed_since_active      = 0;
+        expected_total_steps             = 0;
+        consecutive_skipped_steps        = 0;
         accumulated_error                = 0.0f;
-        reference_output_norm            = 0.0f;
         block_metrics.reset();
         total_active_steps = 0;
     }
@@ -133,7 +138,8 @@ struct UCacheState {
         if (!initialized || sigmas.size() < 2) {
             return;
         }
-        size_t n_steps = sigmas.size() - 1;
+        size_t n_steps       = sigmas.size() - 1;
+        expected_total_steps = static_cast<int>(n_steps);
 
         size_t start_step = static_cast<size_t>(config.start_percent * n_steps);
         size_t end_step   = static_cast<size_t>(config.end_percent * n_steps);
@@ -207,11 +213,15 @@ struct UCacheState {
         }
 
         int effective_total = estimated_total_steps;
+        if (effective_total <= 0) {
+            effective_total = expected_total_steps;
+        }
         if (effective_total <= 0) {
             effective_total = std::max(20, steps_computed_since_active * 2);
         }
 
         float progress = (effective_total > 0) ? (static_cast<float>(steps_computed_since_active) / effective_total) : 0.0f;
+        progress       = std::max(0.0f, std::min(1.0f, progress));
 
         float multiplier = 1.0f;
         if (progress < 0.2f) {
@@ -309,17 +319,31 @@ struct UCacheState {
 
         if (has_output_prev_norm && has_relative_transformation_rate &&
             last_input_change > 0.0f && output_prev_norm > 0.0f) {
-            float approx_output_change_rate = (relative_transformation_rate * last_input_change) / output_prev_norm;
+            float approx_output_change = relative_transformation_rate * last_input_change;
-            accumulated_error               = accumulated_error * config.error_decay_rate + approx_output_change_rate;
+            float approx_output_change_rate;
+            if (config.use_relative_threshold) {
+                float base_scale          = std::max(output_prev_norm, 1e-6f);
+                float dyn_scale           = has_output_change_ema
+                                                ? std::max(output_change_ema * std::max(1.0f, config.relative_norm_gain), 1e-6f)
+                                                : base_scale;
+                float scale               = std::sqrt(base_scale * dyn_scale);
+                approx_output_change_rate = approx_output_change / scale;
+            } else {
+                approx_output_change_rate = approx_output_change;
+            }
+            // Increase estimated error with skip horizon to avoid long extrapolation streaks
+            approx_output_change_rate *= (1.0f + 0.50f * consecutive_skipped_steps);
+            accumulated_error = accumulated_error * config.error_decay_rate + approx_output_change_rate;
 
             float effective_threshold = get_adaptive_threshold();
-            if (config.use_relative_threshold && reference_output_norm > 0.0f) {
+            if (!config.use_relative_threshold && output_prev_norm > 0.0f) {
-                effective_threshold = effective_threshold * reference_output_norm;
+                effective_threshold = effective_threshold * output_prev_norm;
             }
 
             if (accumulated_error < effective_threshold) {
                 skip_current_step = true;
                 total_steps_skipped++;
+                consecutive_skipped_steps++;
                 apply_cache(cond, input, output);
                 return true;
             } else if (config.reset_error_on_compute) {
@@ -340,6 +364,8 @@ struct UCacheState {
         if (cond != anchor_condition) {
             return;
         }
+        steps_computed_since_active++;
+        consecutive_skipped_steps = 0;
 
         size_t ne      = static_cast<size_t>(ggml_nelements(input));
         float* in_data = (float*)input->data;
@@ -359,6 +385,14 @@ struct UCacheState {
                 output_change /= static_cast<float>(ne);
             }
         }
+        if (std::isfinite(output_change) && output_change > 0.0f) {
+            if (!has_output_change_ema) {
+                output_change_ema     = output_change;
+                has_output_change_ema = true;
+            } else {
+                output_change_ema = 0.8f * output_change_ema + 0.2f * output_change;
+            }
+        }
 
         prev_output.resize(ne);
         for (size_t i = 0; i < ne; ++i) {
@@ -373,10 +407,6 @@ struct UCacheState {
         output_prev_norm     = (ne > 0) ? (mean_abs / static_cast<float>(ne)) : 0.0f;
         has_output_prev_norm = output_prev_norm > 0.0f;
 
-        if (reference_output_norm == 0.0f) {
-            reference_output_norm = output_prev_norm;
-        }
-
         if (has_last_input_change && last_input_change > 0.0f && output_change > 0.0f) {
             float rate = output_change / last_input_change;
             if (std::isfinite(rate)) {

src/unet.hpp

@@ -60,10 +60,10 @@ public:
         blocks["time_mixer"] = std::shared_ptr<GGMLBlock>(new AlphaBlender());
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* context,
+                         ggml_tensor* context,
-                                int timesteps) {
+                         int timesteps) {
         // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w], t == timesteps
         // context: [N, max_position(aka n_context), hidden_size(aka context_dim)] aka [b*t, n_context, context_dim], t == timesteps
         // t_emb: [N, in_channels] aka [b*t, in_channels]
@@ -388,11 +388,11 @@ public:
         blocks["out.2"] = std::shared_ptr<GGMLBlock>(new Conv2d(model_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
     }
 
-    struct ggml_tensor* resblock_forward(std::string name,
+    ggml_tensor* resblock_forward(std::string name,
-                                         GGMLRunnerContext* ctx,
+                                  GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
+                                  ggml_tensor* x,
-                                         struct ggml_tensor* emb,
+                                  ggml_tensor* emb,
-                                         int num_video_frames) {
+                                  int num_video_frames) {
         if (version == VERSION_SVD) {
             auto block = std::dynamic_pointer_cast<VideoResBlock>(blocks[name]);
 
@@ -404,11 +404,11 @@ public:
         }
     }
 
-    struct ggml_tensor* attention_layer_forward(std::string name,
+    ggml_tensor* attention_layer_forward(std::string name,
-                                                GGMLRunnerContext* ctx,
+                                         GGMLRunnerContext* ctx,
-                                                struct ggml_tensor* x,
+                                         ggml_tensor* x,
-                                                struct ggml_tensor* context,
+                                         ggml_tensor* context,
-                                                int timesteps) {
+                                         int timesteps) {
         if (version == VERSION_SVD) {
             auto block = std::dynamic_pointer_cast<SpatialVideoTransformer>(blocks[name]);
 
@@ -420,15 +420,15 @@ public:
         }
     }
 
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+    ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x,
+                         ggml_tensor* x,
-                                struct ggml_tensor* timesteps,
+                         ggml_tensor* timesteps,
-                                struct ggml_tensor* context,
+                         ggml_tensor* context,
-                                struct ggml_tensor* c_concat              = nullptr,
+                         ggml_tensor* c_concat              = nullptr,
-                                struct ggml_tensor* y                     = nullptr,
+                         ggml_tensor* y                     = nullptr,
-                                int num_video_frames                      = -1,
+                         int num_video_frames               = -1,
-                                std::vector<struct ggml_tensor*> controls = {},
+                         std::vector<ggml_tensor*> controls = {},
-                                float control_strength                    = 0.f) {
+                         float control_strength             = 0.f) {
         // x: [N, in_channels, h, w] or [N, in_channels/2, h, w]
         // timesteps: [N,]
         // context: [N, max_position, hidden_size] or [1, max_position, hidden_size]. for example, [N, 77, 768]
@@ -480,7 +480,7 @@ public:
         }
 
         // input_blocks
-        std::vector<struct ggml_tensor*> hs;
+        std::vector<ggml_tensor*> hs;
 
         // input block 0
         auto h = input_blocks_0_0->forward(ctx, x);
@@ -605,19 +605,19 @@ struct UNetModelRunner : public GGMLRunner {
         return "unet";
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
         unet.get_param_tensors(tensors, prefix);
     }
 
-    struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+    ggml_cgraph* build_graph(ggml_tensor* x,
-                                    struct ggml_tensor* timesteps,
+                             ggml_tensor* timesteps,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    struct ggml_tensor* c_concat              = nullptr,
+                             ggml_tensor* c_concat              = nullptr,
-                                    struct ggml_tensor* y                     = nullptr,
+                             ggml_tensor* y                     = nullptr,
-                                    int num_video_frames                      = -1,
+                             int num_video_frames               = -1,
-                                    std::vector<struct ggml_tensor*> controls = {},
+                             std::vector<ggml_tensor*> controls = {},
-                                    float control_strength                    = 0.f) {
+                             float control_strength             = 0.f) {
-        struct ggml_cgraph* gf = new_graph_custom(UNET_GRAPH_SIZE);
+        ggml_cgraph* gf = new_graph_custom(UNET_GRAPH_SIZE);
 
         if (num_video_frames == -1) {
             num_video_frames = static_cast<int>(x->ne[3]);
@@ -635,15 +635,15 @@ struct UNetModelRunner : public GGMLRunner {
 
         auto runner_ctx = get_context();
 
-        struct ggml_tensor* out = unet.forward(&runner_ctx,
+        ggml_tensor* out = unet.forward(&runner_ctx,
-                                               x,
+                                        x,
-                                               timesteps,
+                                        timesteps,
-                                               context,
+                                        context,
-                                               c_concat,
+                                        c_concat,
-                                               y,
+                                        y,
-                                               num_video_frames,
+                                        num_video_frames,
-                                               controls,
+                                        controls,
-                                               control_strength);
+                                        control_strength);
 
         ggml_build_forward_expand(gf, out);
 
@@ -651,22 +651,22 @@ struct UNetModelRunner : public GGMLRunner {
     }
 
     bool compute(int n_threads,
-                 struct ggml_tensor* x,
+                 ggml_tensor* x,
-                 struct ggml_tensor* timesteps,
+                 ggml_tensor* timesteps,
-                 struct ggml_tensor* context,
+                 ggml_tensor* context,
-                 struct ggml_tensor* c_concat,
+                 ggml_tensor* c_concat,
-                 struct ggml_tensor* y,
+                 ggml_tensor* y,
-                 int num_video_frames                      = -1,
+                 int num_video_frames               = -1,
-                 std::vector<struct ggml_tensor*> controls = {},
+                 std::vector<ggml_tensor*> controls = {},
-                 float control_strength                    = 0.f,
+                 float control_strength             = 0.f,
-                 struct ggml_tensor** output               = nullptr,
+                 ggml_tensor** output               = nullptr,
-                 struct ggml_context* output_ctx           = nullptr) {
+                 ggml_context* output_ctx           = nullptr) {
         // x: [N, in_channels, h, w]
         // timesteps: [N, ]
         // context: [N, max_position, hidden_size]([N, 77, 768]) or [1, max_position, hidden_size]
         // c_concat: [N, in_channels, h, w] or [1, in_channels, h, w]
         // y: [N, adm_in_channels] or [1, adm_in_channels]
-        auto get_graph = [&]() -> struct ggml_cgraph* {
+        auto get_graph = [&]() -> ggml_cgraph* {
             return build_graph(x, timesteps, context, c_concat, y, num_video_frames, controls, control_strength);
         };
 
@@ -674,12 +674,12 @@ struct UNetModelRunner : public GGMLRunner {
     }
 
     void test() {
-        struct ggml_init_params params;
+        ggml_init_params params;
         params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
         params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
-        struct ggml_context* work_ctx = ggml_init(params);
+        ggml_context* work_ctx = ggml_init(params);
         GGML_ASSERT(work_ctx != nullptr);
 
         {
@@ -703,7 +703,7 @@ struct UNetModelRunner : public GGMLRunner {
             ggml_set_f32(y, 0.5f);
             // print_ggml_tensor(y);
 
-            struct ggml_tensor* out = nullptr;
+            ggml_tensor* out = nullptr;
 
             int64_t t0 = ggml_time_ms();
             compute(8, x, timesteps, context, nullptr, y, num_video_frames, {}, 0.f, &out, work_ctx);

src/upscaler.cpp

@@ -72,13 +72,13 @@ struct UpscalerGGML {
         LOG_INFO("upscaling from (%i x %i) to (%i x %i)",
                  input_image.width, input_image.height, output_width, output_height);
 
-        struct ggml_init_params params;
+        ggml_init_params params;
         params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
         params.mem_buffer = nullptr;
         params.no_alloc   = false;
 
         // draft context
-        struct ggml_context* upscale_ctx = ggml_init(params);
+        ggml_context* upscale_ctx = ggml_init(params);
         if (!upscale_ctx) {
             LOG_ERROR("ggml_init() failed");
             return upscaled_image;
@@ -92,7 +92,8 @@ struct UpscalerGGML {
             return esrgan_upscaler->compute(n_threads, in, &out);
         };
         int64_t t0 = ggml_time_ms();
-        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, on_tiling);
+        // TODO: circular upscaling?
+        sd_tiling(input_image_tensor, upscaled, esrgan_upscaler->scale, esrgan_upscaler->tile_size, 0.25f, false, false, on_tiling);
         esrgan_upscaler->free_compute_buffer();
         ggml_ext_tensor_clamp_inplace(upscaled, 0.f, 1.f);
         uint8_t* upscaled_data = ggml_tensor_to_sd_image(upscaled);

src/vae.hpp

@@ -3,635 +3,206 @@
 
 #include "common_block.hpp"
 
-/*================================================== AutoEncoderKL ===================================================*/
+struct VAE : public GGMLRunner {
-
-#define VAE_GRAPH_SIZE 20480
-
-class ResnetBlock : public UnaryBlock {
-protected:
-    int64_t in_channels;
-    int64_t out_channels;
-
-public:
-    ResnetBlock(int64_t in_channels,
-                int64_t out_channels)
-        : in_channels(in_channels),
-          out_channels(out_channels) {
-        // temb_channels is always 0
-        blocks["norm1"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
-        blocks["conv1"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
-
-        blocks["norm2"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(out_channels));
-        blocks["conv2"] = std::shared_ptr<GGMLBlock>(new Conv2d(out_channels, out_channels, {3, 3}, {1, 1}, {1, 1}));
-
-        if (out_channels != in_channels) {
-            blocks["nin_shortcut"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, {1, 1}));
-        }
-    }
-
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
-        // x: [N, in_channels, h, w]
-        // t_emb is always None
-        auto norm1 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm1"]);
-        auto conv1 = std::dynamic_pointer_cast<Conv2d>(blocks["conv1"]);
-        auto norm2 = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm2"]);
-        auto conv2 = std::dynamic_pointer_cast<Conv2d>(blocks["conv2"]);
-
-        auto h = x;
-        h      = norm1->forward(ctx, h);
-        h      = ggml_silu_inplace(ctx->ggml_ctx, h);  // swish
-        h      = conv1->forward(ctx, h);
-        // return h;
-
-        h = norm2->forward(ctx, h);
-        h = ggml_silu_inplace(ctx->ggml_ctx, h);  // swish
-        // dropout, skip for inference
-        h = conv2->forward(ctx, h);
-
-        // skip connection
-        if (out_channels != in_channels) {
-            auto nin_shortcut = std::dynamic_pointer_cast<Conv2d>(blocks["nin_shortcut"]);
-
-            x = nin_shortcut->forward(ctx, x);  // [N, out_channels, h, w]
-        }
-
-        h = ggml_add(ctx->ggml_ctx, h, x);
-        return h;  // [N, out_channels, h, w]
-    }
-};
-
-class AttnBlock : public UnaryBlock {
-protected:
-    int64_t in_channels;
-    bool use_linear;
-
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") {
-        auto iter = tensor_storage_map.find(prefix + "proj_out.weight");
-        if (iter != tensor_storage_map.end()) {
-            if (iter->second.n_dims == 4 && use_linear) {
-                use_linear         = false;
-                blocks["q"]        = std::make_shared<Conv2d>(in_channels, in_channels, std::pair{1, 1});
-                blocks["k"]        = std::make_shared<Conv2d>(in_channels, in_channels, std::pair{1, 1});
-                blocks["v"]        = std::make_shared<Conv2d>(in_channels, in_channels, std::pair{1, 1});
-                blocks["proj_out"] = std::make_shared<Conv2d>(in_channels, in_channels, std::pair{1, 1});
-            } else if (iter->second.n_dims == 2 && !use_linear) {
-                use_linear         = true;
-                blocks["q"]        = std::make_shared<Linear>(in_channels, in_channels);
-                blocks["k"]        = std::make_shared<Linear>(in_channels, in_channels);
-                blocks["v"]        = std::make_shared<Linear>(in_channels, in_channels);
-                blocks["proj_out"] = std::make_shared<Linear>(in_channels, in_channels);
-            }
-        }
-    }
-
-public:
-    AttnBlock(int64_t in_channels, bool use_linear)
-        : in_channels(in_channels), use_linear(use_linear) {
-        blocks["norm"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(in_channels));
-        if (use_linear) {
-            blocks["q"]        = std::shared_ptr<GGMLBlock>(new Linear(in_channels, in_channels));
-            blocks["k"]        = std::shared_ptr<GGMLBlock>(new Linear(in_channels, in_channels));
-            blocks["v"]        = std::shared_ptr<GGMLBlock>(new Linear(in_channels, in_channels));
-            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Linear(in_channels, in_channels));
-        } else {
-            blocks["q"]        = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
-            blocks["k"]        = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
-            blocks["v"]        = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
-            blocks["proj_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, in_channels, {1, 1}));
-        }
-    }
-
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
-        // x: [N, in_channels, h, w]
-        auto norm     = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm"]);
-        auto q_proj   = std::dynamic_pointer_cast<UnaryBlock>(blocks["q"]);
-        auto k_proj   = std::dynamic_pointer_cast<UnaryBlock>(blocks["k"]);
-        auto v_proj   = std::dynamic_pointer_cast<UnaryBlock>(blocks["v"]);
-        auto proj_out = std::dynamic_pointer_cast<UnaryBlock>(blocks["proj_out"]);
-
-        auto h_ = norm->forward(ctx, x);
-
-        const int64_t n = h_->ne[3];
-        const int64_t c = h_->ne[2];
-        const int64_t h = h_->ne[1];
-        const int64_t w = h_->ne[0];
-
-        ggml_tensor* q;
-        ggml_tensor* k;
-        ggml_tensor* v;
-        if (use_linear) {
-            h_ = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, h_, 1, 2, 0, 3));  // [N, h, w, in_channels]
-            h_ = ggml_reshape_3d(ctx->ggml_ctx, h_, c, h * w, n);                        // [N, h * w, in_channels]
-
-            q = q_proj->forward(ctx, h_);  // [N, h * w, in_channels]
-            k = k_proj->forward(ctx, h_);  // [N, h * w, in_channels]
-            v = v_proj->forward(ctx, h_);  // [N, h * w, in_channels]
-        } else {
-            q = q_proj->forward(ctx, h_);                                              // [N, in_channels, h, w]
-            q = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, q, 1, 2, 0, 3));  // [N, h, w, in_channels]
-            q = ggml_reshape_3d(ctx->ggml_ctx, q, c, h * w, n);                        // [N, h * w, in_channels]
-
-            k = k_proj->forward(ctx, h_);                                              // [N, in_channels, h, w]
-            k = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, k, 1, 2, 0, 3));  // [N, h, w, in_channels]
-            k = ggml_reshape_3d(ctx->ggml_ctx, k, c, h * w, n);                        // [N, h * w, in_channels]
-
-            v = v_proj->forward(ctx, h_);                                              // [N, in_channels, h, w]
-            v = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, v, 1, 2, 0, 3));  // [N, h, w, in_channels]
-            v = ggml_reshape_3d(ctx->ggml_ctx, v, c, h * w, n);                        // [N, h * w, in_channels]
-        }
-
-        h_ = ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, 1, nullptr, false, ctx->flash_attn_enabled);
-
-        if (use_linear) {
-            h_ = proj_out->forward(ctx, h_);  // [N, h * w, in_channels]
-
-            h_ = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, h_, 1, 0, 2, 3));  // [N, in_channels, h * w]
-            h_ = ggml_reshape_4d(ctx->ggml_ctx, h_, w, h, c, n);                         // [N, in_channels, h, w]
-        } else {
-            h_ = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, h_, 1, 0, 2, 3));  // [N, in_channels, h * w]
-            h_ = ggml_reshape_4d(ctx->ggml_ctx, h_, w, h, c, n);                         // [N, in_channels, h, w]
-
-            h_ = proj_out->forward(ctx, h_);  // [N, in_channels, h, w]
-        }
-
-        h_ = ggml_add(ctx->ggml_ctx, h_, x);
-        return h_;
-    }
-};
-
-class AE3DConv : public Conv2d {
-public:
-    AE3DConv(int64_t in_channels,
-             int64_t out_channels,
-             std::pair<int, int> kernel_size,
-             int video_kernel_size        = 3,
-             std::pair<int, int> stride   = {1, 1},
-             std::pair<int, int> padding  = {0, 0},
-             std::pair<int, int> dilation = {1, 1},
-             bool bias                    = true)
-        : Conv2d(in_channels, out_channels, kernel_size, stride, padding, dilation, bias) {
-        int kernel_padding      = video_kernel_size / 2;
-        blocks["time_mix_conv"] = std::shared_ptr<GGMLBlock>(new Conv3d(out_channels,
-                                                                        out_channels,
-                                                                        {video_kernel_size, 1, 1},
-                                                                        {1, 1, 1},
-                                                                        {kernel_padding, 0, 0}));
-    }
-
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                struct ggml_tensor* x) override {
-        // timesteps always None
-        // skip_video always False
-        // x: [N, IC, IH, IW]
-        // result: [N, OC, OH, OW]
-        auto time_mix_conv = std::dynamic_pointer_cast<Conv3d>(blocks["time_mix_conv"]);
-
-        x = Conv2d::forward(ctx, x);
-        // timesteps = x.shape[0]
-        // x = rearrange(x, "(b t) c h w -> b c t h w", t=timesteps)
-        // x = conv3d(x)
-        // return rearrange(x, "b c t h w -> (b t) c h w")
-        int64_t T = x->ne[3];
-        int64_t B = x->ne[3] / T;
-        int64_t C = x->ne[2];
-        int64_t H = x->ne[1];
-        int64_t W = x->ne[0];
-
-        x = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, C, T, B);                     // (b t) c h w -> b t c (h w)
-        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
-        x = time_mix_conv->forward(ctx, x);                                        // [B, OC, T, OH * OW]
-        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
-        x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
-        return x;                                                                  // [B*T, OC, OH, OW]
-    }
-};
-
-class VideoResnetBlock : public ResnetBlock {
-protected:
-    void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
-        enum ggml_type wtype = get_type(prefix + "mix_factor", tensor_storage_map, GGML_TYPE_F32);
-        params["mix_factor"] = ggml_new_tensor_1d(ctx, wtype, 1);
-    }
-
-    float get_alpha() {
-        float alpha = ggml_ext_backend_tensor_get_f32(params["mix_factor"]);
-        return sigmoid(alpha);
-    }
-
-public:
-    VideoResnetBlock(int64_t in_channels,
-                     int64_t out_channels,
-                     int video_kernel_size = 3)
-        : ResnetBlock(in_channels, out_channels) {
-        // merge_strategy is always learned
-        blocks["time_stack"] = std::shared_ptr<GGMLBlock>(new ResBlock(out_channels, 0, out_channels, {video_kernel_size, 1}, 3, false, true));
-    }
-
-    struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
-        // x: [N, in_channels, h, w] aka [b*t, in_channels, h, w]
-        // return: [N, out_channels, h, w] aka [b*t, out_channels, h, w]
-        // t_emb is always None
-        // skip_video is always False
-        // timesteps is always None
-        auto time_stack = std::dynamic_pointer_cast<ResBlock>(blocks["time_stack"]);
-
-        x = ResnetBlock::forward(ctx, x);  // [N, out_channels, h, w]
-        // return x;
-
-        int64_t T = x->ne[3];
-        int64_t B = x->ne[3] / T;
-        int64_t C = x->ne[2];
-        int64_t H = x->ne[1];
-        int64_t W = x->ne[0];
-
-        x          = ggml_reshape_4d(ctx->ggml_ctx, x, W * H, C, T, B);                     // (b t) c h w -> b t c (h w)
-        x          = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b t c (h w) -> b c t (h w)
-        auto x_mix = x;
-
-        x = time_stack->forward(ctx, x);  // b t c (h w)
-
-        float alpha = get_alpha();
-        x           = ggml_add(ctx->ggml_ctx,
-                               ggml_ext_scale(ctx->ggml_ctx, x, alpha),
-                               ggml_ext_scale(ctx->ggml_ctx, x_mix, 1.0f - alpha));
-
-        x = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, x, 0, 2, 1, 3));  // b c t (h w) -> b t c (h w)
-        x = ggml_reshape_4d(ctx->ggml_ctx, x, W, H, C, T * B);                     // b t c (h w) -> (b t) c h w
-
-        return x;
-    }
-};
-
-// ldm.modules.diffusionmodules.model.Encoder
-class Encoder : public GGMLBlock {
-protected:
-    int ch                   = 128;
-    std::vector<int> ch_mult = {1, 2, 4, 4};
-    int num_res_blocks       = 2;
-    int in_channels          = 3;
-    int z_channels           = 4;
-    bool double_z            = true;
-
-public:
-    Encoder(int ch,
-            std::vector<int> ch_mult,
-            int num_res_blocks,
-            int in_channels,
-            int z_channels,
-            bool double_z              = true,
-            bool use_linear_projection = false)
-        : ch(ch),
-          ch_mult(ch_mult),
-          num_res_blocks(num_res_blocks),
-          in_channels(in_channels),
-          z_channels(z_channels),
-          double_z(double_z) {
-        blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, ch, {3, 3}, {1, 1}, {1, 1}));
-
-        size_t num_resolutions = ch_mult.size();
-
-        int block_in = 1;
-        for (int i = 0; i < num_resolutions; i++) {
-            if (i == 0) {
-                block_in = ch;
-            } else {
-                block_in = ch * ch_mult[i - 1];
-            }
-            int block_out = ch * ch_mult[i];
-            for (int j = 0; j < num_res_blocks; j++) {
-                std::string name = "down." + std::to_string(i) + ".block." + std::to_string(j);
-                blocks[name]     = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_out));
-                block_in         = block_out;
-            }
-            if (i != num_resolutions - 1) {
-                std::string name = "down." + std::to_string(i) + ".downsample";
-                blocks[name]     = std::shared_ptr<GGMLBlock>(new DownSampleBlock(block_in, block_in, true));
-            }
-        }
-
-        blocks["mid.block_1"] = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_in));
-        blocks["mid.attn_1"]  = std::shared_ptr<GGMLBlock>(new AttnBlock(block_in, use_linear_projection));
-        blocks["mid.block_2"] = std::shared_ptr<GGMLBlock>(new ResnetBlock(block_in, block_in));
-
-        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(block_in));
-        blocks["conv_out"] = std::shared_ptr<GGMLBlock>(new Conv2d(block_in, double_z ? z_channels * 2 : z_channels, {3, 3}, {1, 1}, {1, 1}));
-    }
-
-    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
-        // x: [N, in_channels, h, w]
-
-        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
-        auto mid_block_1 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_1"]);
-        auto mid_attn_1  = std::dynamic_pointer_cast<AttnBlock>(blocks["mid.attn_1"]);
-        auto mid_block_2 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_2"]);
-        auto norm_out    = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm_out"]);
-        auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
-
-        auto h = conv_in->forward(ctx, x);  // [N, ch, h, w]
-
-        // downsampling
-        size_t num_resolutions = ch_mult.size();
-        for (int i = 0; i < num_resolutions; i++) {
-            for (int j = 0; j < num_res_blocks; j++) {
-                std::string name = "down." + std::to_string(i) + ".block." + std::to_string(j);
-                auto down_block  = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
-
-                h = down_block->forward(ctx, 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);
-            }
-        }
-
-        // middle
-        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]
-
-        // end
-        h = norm_out->forward(ctx, h);
-        h = ggml_silu_inplace(ctx->ggml_ctx, h);  // nonlinearity/swish
-        h = conv_out->forward(ctx, h);            // [N, z_channels*2, h, w]
-        return h;
-    }
-};
-
-// ldm.modules.diffusionmodules.model.Decoder
-class Decoder : public GGMLBlock {
-protected:
-    int ch                   = 128;
-    int out_ch               = 3;
-    std::vector<int> ch_mult = {1, 2, 4, 4};
-    int num_res_blocks       = 2;
-    int z_channels           = 4;
-    bool video_decoder       = false;
-    int video_kernel_size    = 3;
-
-    virtual std::shared_ptr<GGMLBlock> get_conv_out(int64_t in_channels,
-                                                    int64_t out_channels,
-                                                    std::pair<int, int> kernel_size,
-                                                    std::pair<int, int> stride  = {1, 1},
-                                                    std::pair<int, int> padding = {0, 0}) {
-        if (video_decoder) {
-            return std::shared_ptr<GGMLBlock>(new AE3DConv(in_channels, out_channels, kernel_size, video_kernel_size, stride, padding));
-        } else {
-            return std::shared_ptr<GGMLBlock>(new Conv2d(in_channels, out_channels, kernel_size, stride, padding));
-        }
-    }
-
-    virtual std::shared_ptr<GGMLBlock> get_resnet_block(int64_t in_channels,
-                                                        int64_t out_channels) {
-        if (video_decoder) {
-            return std::shared_ptr<GGMLBlock>(new VideoResnetBlock(in_channels, out_channels, video_kernel_size));
-        } else {
-            return std::shared_ptr<GGMLBlock>(new ResnetBlock(in_channels, out_channels));
-        }
-    }
-
-public:
-    Decoder(int ch,
-            int out_ch,
-            std::vector<int> ch_mult,
-            int num_res_blocks,
-            int z_channels,
-            bool use_linear_projection = false,
-            bool video_decoder         = false,
-            int video_kernel_size      = 3)
-        : ch(ch),
-          out_ch(out_ch),
-          ch_mult(ch_mult),
-          num_res_blocks(num_res_blocks),
-          z_channels(z_channels),
-          video_decoder(video_decoder),
-          video_kernel_size(video_kernel_size) {
-        int num_resolutions = static_cast<int>(ch_mult.size());
-        int block_in        = ch * ch_mult[num_resolutions - 1];
-
-        blocks["conv_in"] = std::shared_ptr<GGMLBlock>(new Conv2d(z_channels, block_in, {3, 3}, {1, 1}, {1, 1}));
-
-        blocks["mid.block_1"] = get_resnet_block(block_in, block_in);
-        blocks["mid.attn_1"]  = std::shared_ptr<GGMLBlock>(new AttnBlock(block_in, use_linear_projection));
-        blocks["mid.block_2"] = get_resnet_block(block_in, block_in);
-
-        for (int i = num_resolutions - 1; i >= 0; i--) {
-            int mult      = ch_mult[i];
-            int block_out = ch * mult;
-            for (int j = 0; j < num_res_blocks + 1; j++) {
-                std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
-                blocks[name]     = get_resnet_block(block_in, block_out);
-
-                block_in = block_out;
-            }
-            if (i != 0) {
-                std::string name = "up." + std::to_string(i) + ".upsample";
-                blocks[name]     = std::shared_ptr<GGMLBlock>(new UpSampleBlock(block_in, block_in));
-            }
-        }
-
-        blocks["norm_out"] = std::shared_ptr<GGMLBlock>(new GroupNorm32(block_in));
-        blocks["conv_out"] = get_conv_out(block_in, out_ch, {3, 3}, {1, 1}, {1, 1});
-    }
-
-    virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
-        // z: [N, z_channels, h, w]
-        // alpha is always 0
-        // merge_strategy is always learned
-        // time_mode is always conv-only, so we need to replace conv_out_op/resnet_op to AE3DConv/VideoResBlock
-        // AttnVideoBlock will not be used
-        auto conv_in     = std::dynamic_pointer_cast<Conv2d>(blocks["conv_in"]);
-        auto mid_block_1 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_1"]);
-        auto mid_attn_1  = std::dynamic_pointer_cast<AttnBlock>(blocks["mid.attn_1"]);
-        auto mid_block_2 = std::dynamic_pointer_cast<ResnetBlock>(blocks["mid.block_2"]);
-        auto norm_out    = std::dynamic_pointer_cast<GroupNorm32>(blocks["norm_out"]);
-        auto conv_out    = std::dynamic_pointer_cast<Conv2d>(blocks["conv_out"]);
-
-        // conv_in
-        auto h = conv_in->forward(ctx, z);  // [N, block_in, h, w]
-
-        // middle
-        h = mid_block_1->forward(ctx, h);
-        // return h;
-
-        h = mid_attn_1->forward(ctx, h);
-        h = mid_block_2->forward(ctx, h);  // [N, block_in, h, w]
-
-        // upsampling
-        int num_resolutions = static_cast<int>(ch_mult.size());
-        for (int i = num_resolutions - 1; i >= 0; i--) {
-            for (int j = 0; j < num_res_blocks + 1; j++) {
-                std::string name = "up." + std::to_string(i) + ".block." + std::to_string(j);
-                auto up_block    = std::dynamic_pointer_cast<ResnetBlock>(blocks[name]);
-
-                h = up_block->forward(ctx, 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);
-            }
-        }
-
-        h = norm_out->forward(ctx, h);
-        h = ggml_silu_inplace(ctx->ggml_ctx, h);  // nonlinearity/swish
-        h = conv_out->forward(ctx, h);            // [N, out_ch, h*8, w*8]
-        return h;
-    }
-};
-
-// ldm.models.autoencoder.AutoencoderKL
-class AutoencodingEngine : public GGMLBlock {
 protected:
     SDVersion version;
-    bool decode_only       = true;
+    bool scale_input                                = true;
-    bool use_video_decoder = false;
+    virtual bool _compute(const int n_threads,
-    bool use_quant         = true;
+                          ggml_tensor* z,
-    int embed_dim          = 4;
+                          bool decode_graph,
-    struct {
+                          ggml_tensor** output,
-        int z_channels           = 4;
+                          ggml_context* output_ctx) = 0;
-        int resolution           = 256;
-        int in_channels          = 3;
-        int out_ch               = 3;
-        int ch                   = 128;
-        std::vector<int> ch_mult = {1, 2, 4, 4};
-        int num_res_blocks       = 2;
-        bool double_z            = true;
-    } dd_config;
 
 public:
-    AutoencodingEngine(SDVersion version          = VERSION_SD1,
+    VAE(SDVersion version, ggml_backend_t backend, bool offload_params_to_cpu)
-                       bool decode_only           = true,
+        : version(version), GGMLRunner(backend, offload_params_to_cpu) {}
-                       bool use_linear_projection = false,
+
-                       bool use_video_decoder     = false)
+    int get_scale_factor() {
-        : version(version), decode_only(decode_only), use_video_decoder(use_video_decoder) {
+        int scale_factor = 8;
-        if (sd_version_is_dit(version)) {
+        if (version == VERSION_WAN2_2_TI2V) {
-            if (sd_version_is_flux2(version)) {
+            scale_factor = 16;
-                dd_config.z_channels = 32;
+        } else if (sd_version_is_flux2(version)) {
-                embed_dim            = 32;
+            scale_factor = 16;
+        } else if (version == VERSION_CHROMA_RADIANCE) {
+            scale_factor = 1;
+        }
+        return scale_factor;
+    }
+
+    virtual int get_encoder_output_channels(int input_channels) = 0;
+
+    void get_tile_sizes(int& tile_size_x,
+                        int& tile_size_y,
+                        float& tile_overlap,
+                        const sd_tiling_params_t& params,
+                        int64_t latent_x,
+                        int64_t latent_y,
+                        float encoding_factor = 1.0f) {
+        tile_overlap       = std::max(std::min(params.target_overlap, 0.5f), 0.0f);
+        auto get_tile_size = [&](int requested_size, float factor, int64_t latent_size) {
+            const int default_tile_size  = 32;
+            const int min_tile_dimension = 4;
+            int tile_size                = default_tile_size;
+            // factor <= 1 means simple fraction of the latent dimension
+            // factor > 1 means number of tiles across that dimension
+            if (factor > 0.f) {
+                if (factor > 1.0)
+                    factor = 1 / (factor - factor * tile_overlap + tile_overlap);
+                tile_size = static_cast<int>(std::round(latent_size * factor));
+            } else if (requested_size >= min_tile_dimension) {
+                tile_size = requested_size;
+            }
+            tile_size = static_cast<int>(tile_size * encoding_factor);
+            return std::max(std::min(tile_size, static_cast<int>(latent_size)), min_tile_dimension);
+        };
+
+        tile_size_x = get_tile_size(params.tile_size_x, params.rel_size_x, latent_x);
+        tile_size_y = get_tile_size(params.tile_size_y, params.rel_size_y, latent_y);
+    }
+
+    ggml_tensor* encode(int n_threads,
+                        ggml_context* work_ctx,
+                        ggml_tensor* x,
+                        sd_tiling_params_t tiling_params,
+                        bool circular_x = false,
+                        bool circular_y = false) {
+        int64_t t0             = ggml_time_ms();
+        ggml_tensor* result    = nullptr;
+        const int scale_factor = get_scale_factor();
+        int64_t W              = x->ne[0] / scale_factor;
+        int64_t H              = x->ne[1] / scale_factor;
+        int channel_dim        = sd_version_is_wan(version) ? 3 : 2;
+        int64_t C              = get_encoder_output_channels(static_cast<int>(x->ne[channel_dim]));
+        int64_t ne2;
+        int64_t ne3;
+        if (sd_version_is_wan(version)) {
+            int64_t T = x->ne[2];
+            ne2       = (T - 1) / 4 + 1;
+            ne3       = C;
+        } else {
+            ne2 = C;
+            ne3 = x->ne[3];
+        }
+        result = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, W, H, ne2, ne3);
+
+        if (scale_input) {
+            scale_to_minus1_1(x);
+        }
+
+        if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
+            x = ggml_reshape_4d(work_ctx, x, x->ne[0], x->ne[1], 1, x->ne[2] * x->ne[3]);
+        }
+
+        if (tiling_params.enabled) {
+            float tile_overlap;
+            int tile_size_x, tile_size_y;
+            // multiply tile size for encode to keep the compute buffer size consistent
+            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, W, H, 1.30539f);
+
+            LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
+
+            auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
+                return _compute(n_threads, in, false, &out, work_ctx);
+            };
+            sd_tiling_non_square(x, result, scale_factor, tile_size_x, tile_size_y, tile_overlap, circular_x, circular_y, on_tiling);
+        } else {
+            _compute(n_threads, x, false, &result, work_ctx);
+        }
+        free_compute_buffer();
+
+        int64_t t1 = ggml_time_ms();
+        LOG_DEBUG("computing vae encode graph completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+        return result;
+    }
+
+    ggml_tensor* decode(int n_threads,
+                        ggml_context* work_ctx,
+                        ggml_tensor* x,
+                        sd_tiling_params_t tiling_params,
+                        bool decode_video   = false,
+                        bool circular_x     = false,
+                        bool circular_y     = false,
+                        ggml_tensor* result = nullptr,
+                        bool silent         = false) {
+        const int scale_factor = get_scale_factor();
+        int64_t W              = x->ne[0] * scale_factor;
+        int64_t H              = x->ne[1] * scale_factor;
+        int64_t C              = 3;
+        if (result == nullptr) {
+            if (decode_video) {
+                int64_t T = x->ne[2];
+                if (sd_version_is_wan(version)) {
+                    T = ((T - 1) * 4) + 1;
+                }
+                result = ggml_new_tensor_4d(work_ctx,
+                                            GGML_TYPE_F32,
+                                            W,
+                                            H,
+                                            T,
+                                            3);
             } else {
-                use_quant            = false;
+                result = ggml_new_tensor_4d(work_ctx,
-                dd_config.z_channels = 16;
+                                            GGML_TYPE_F32,
+                                            W,
+                                            H,
+                                            C,
+                                            x->ne[3]);
             }
         }
-        if (use_video_decoder) {
+        int64_t t0 = ggml_time_ms();
-            use_quant = false;
+        if (sd_version_is_qwen_image(version) || sd_version_is_anima(version)) {
+            x = ggml_reshape_4d(work_ctx, x, x->ne[0], x->ne[1], 1, x->ne[2] * x->ne[3]);
         }
-        blocks["decoder"] = std::shared_ptr<GGMLBlock>(new Decoder(dd_config.ch,
+        if (tiling_params.enabled) {
-                                                                   dd_config.out_ch,
+            float tile_overlap;
-                                                                   dd_config.ch_mult,
+            int tile_size_x, tile_size_y;
-                                                                   dd_config.num_res_blocks,
+            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, x->ne[0], x->ne[1]);
-                                                                   dd_config.z_channels,
-                                                                   use_linear_projection,
-                                                                   use_video_decoder));
-        if (use_quant) {
-            blocks["post_quant_conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(dd_config.z_channels,
-                                                                              embed_dim,
-                                                                              {1, 1}));
-        }
-        if (!decode_only) {
-            blocks["encoder"] = std::shared_ptr<GGMLBlock>(new Encoder(dd_config.ch,
-                                                                       dd_config.ch_mult,
-                                                                       dd_config.num_res_blocks,
-                                                                       dd_config.in_channels,
-                                                                       dd_config.z_channels,
-                                                                       dd_config.double_z,
-                                                                       use_linear_projection));
-            if (use_quant) {
-                int factor = dd_config.double_z ? 2 : 1;
 
-                blocks["quant_conv"] = std::shared_ptr<GGMLBlock>(new Conv2d(embed_dim * factor,
+            if (!silent) {
-                                                                             dd_config.z_channels * factor,
+                LOG_DEBUG("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
-                                                                             {1, 1}));
+            }
+
+            auto on_tiling = [&](ggml_tensor* in, ggml_tensor* out, bool init) {
+                return _compute(n_threads, in, true, &out, nullptr);
+            };
+            sd_tiling_non_square(x, result, scale_factor, tile_size_x, tile_size_y, tile_overlap, circular_x, circular_y, on_tiling, silent);
+        } else {
+            if (!_compute(n_threads, x, true, &result, work_ctx)) {
+                LOG_ERROR("Failed to decode latetnts");
+                free_compute_buffer();
+                return nullptr;
             }
         }
+        free_compute_buffer();
+        if (scale_input) {
+            scale_to_0_1(result);
+        }
+        int64_t t1 = ggml_time_ms();
+        LOG_DEBUG("computing vae decode graph completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+        ggml_ext_tensor_clamp_inplace(result, 0.0f, 1.0f);
+        return result;
     }
 
-    struct ggml_tensor* decode(GGMLRunnerContext* ctx, struct ggml_tensor* z) {
+    virtual ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) = 0;
-        // z: [N, z_channels, h, w]
+    virtual ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents)                           = 0;
-        if (sd_version_is_flux2(version)) {
+    virtual ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents)                           = 0;
-            // [N, C*p*p, h, w] -> [N, C, h*p, w*p]
+    virtual void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix)                = 0;
-            int64_t p = 2;
-
-            int64_t N = z->ne[3];
-            int64_t C = z->ne[2] / p / p;
-            int64_t h = z->ne[1];
-            int64_t w = z->ne[0];
-            int64_t H = h * p;
-            int64_t W = w * p;
-
-            z = ggml_reshape_4d(ctx->ggml_ctx, z, w * h, p * p, C, N);                           // [N, C, p*p, h*w]
-            z = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, z, 1, 0, 2, 3));  // [N, C, h*w, p*p]
-            z = ggml_reshape_4d(ctx->ggml_ctx, z, p, p, w, h * C * N);                           // [N*C*h, w, p, p]
-            z = ggml_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, z, 0, 2, 1, 3));  // [N*C*h, p, w, p]
-            z = ggml_reshape_4d(ctx->ggml_ctx, z, W, H, C, N);                                   // [N, C, h*p, w*p]
-        }
-
-        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]
-        }
-        auto decoder = std::dynamic_pointer_cast<Decoder>(blocks["decoder"]);
-
-        ggml_set_name(z, "bench-start");
-        auto h = decoder->forward(ctx, z);
-        ggml_set_name(h, "bench-end");
-        return h;
-    }
-
-    struct ggml_tensor* encode(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
-        // x: [N, in_channels, h, w]
-        auto encoder = std::dynamic_pointer_cast<Encoder>(blocks["encoder"]);
-
-        auto z = encoder->forward(ctx, x);  // [N, 2*z_channels, h/8, w/8]
-        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]
-        }
-        if (sd_version_is_flux2(version)) {
-            z = ggml_ext_chunk(ctx->ggml_ctx, z, 2, 2)[0];
-
-            // [N, C, H, W] -> [N, C*p*p, H/p, W/p]
-            int64_t p = 2;
-            int64_t N = z->ne[3];
-            int64_t C = z->ne[2];
-            int64_t H = z->ne[1];
-            int64_t W = z->ne[0];
-            int64_t h = H / p;
-            int64_t w = W / p;
-
-            z = ggml_reshape_4d(ctx->ggml_ctx, z, p, w, p, h * C * N);                 // [N*C*h, p, w, p]
-            z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 0, 2, 1, 3));  // [N*C*h, w, p, p]
-            z = ggml_reshape_4d(ctx->ggml_ctx, z, p * p, w * h, C, N);                 // [N, C, h*w, p*p]
-            z = ggml_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, z, 1, 0, 2, 3));  // [N, C, p*p, h*w]
-            z = ggml_reshape_4d(ctx->ggml_ctx, z, w, h, p * p * C, N);                 // [N, C*p*p, h*w]
-        }
-        return z;
-    }
-};
-
-struct VAE : public GGMLRunner {
-    VAE(ggml_backend_t backend, bool offload_params_to_cpu)
-        : GGMLRunner(backend, offload_params_to_cpu) {}
-    virtual bool compute(const int n_threads,
-                         struct ggml_tensor* z,
-                         bool decode_graph,
-                         struct ggml_tensor** output,
-                         struct ggml_context* output_ctx)                                                         = 0;
-    virtual void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) = 0;
     virtual void set_conv2d_scale(float scale) { SD_UNUSED(scale); };
 };
 
 struct FakeVAE : public VAE {
-    FakeVAE(ggml_backend_t backend, bool offload_params_to_cpu)
+    FakeVAE(SDVersion version, ggml_backend_t backend, bool offload_params_to_cpu)
-        : VAE(backend, offload_params_to_cpu) {}
+        : VAE(version, backend, offload_params_to_cpu) {}
-    bool compute(const int n_threads,
+
-                 struct ggml_tensor* z,
+    int get_encoder_output_channels(int input_channels) {
-                 bool decode_graph,
+        return input_channels;
-                 struct ggml_tensor** output,
+    }
-                 struct ggml_context* output_ctx) override {
+
+    bool _compute(const int n_threads,
+                  ggml_tensor* z,
+                  bool decode_graph,
+                  ggml_tensor** output,
+                  ggml_context* output_ctx) override {
         if (*output == nullptr && output_ctx != nullptr) {
             *output = ggml_dup_tensor(output_ctx, z);
         }
@@ -642,133 +213,23 @@ struct FakeVAE : public VAE {
         return true;
     }
 
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) override {}
+    ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) {
+        return vae_output;
+    }
+
+    ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    }
+
+    ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+        return ggml_ext_dup_and_cpy_tensor(work_ctx, latents);
+    }
+
+    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) override {}
 
     std::string get_desc() override {
         return "fake_vae";
     }
 };
 
-struct AutoEncoderKL : public VAE {
+#endif  // __VAE_HPP__
-    bool decode_only = true;
-    AutoencodingEngine ae;
-
-    AutoEncoderKL(ggml_backend_t backend,
-                  bool offload_params_to_cpu,
-                  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(backend, offload_params_to_cpu) {
-        bool use_linear_projection = false;
-        for (const auto& [name, tensor_storage] : tensor_storage_map) {
-            if (!starts_with(name, prefix)) {
-                continue;
-            }
-            if (ends_with(name, "attn_1.proj_out.weight")) {
-                if (tensor_storage.n_dims == 2) {
-                    use_linear_projection = true;
-                }
-                break;
-            }
-        }
-        ae = AutoencodingEngine(version, decode_only, use_linear_projection, use_video_decoder);
-        ae.init(params_ctx, tensor_storage_map, prefix);
-    }
-
-    void set_conv2d_scale(float scale) override {
-        std::vector<GGMLBlock*> blocks;
-        ae.get_all_blocks(blocks);
-        for (auto block : blocks) {
-            if (block->get_desc() == "Conv2d") {
-                auto conv_block = (Conv2d*)block;
-                conv_block->set_scale(scale);
-            }
-        }
-    }
-
-    std::string get_desc() override {
-        return "vae";
-    }
-
-    void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) override {
-        ae.get_param_tensors(tensors, prefix);
-    }
-
-    struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
-        struct ggml_cgraph* gf = ggml_new_graph(compute_ctx);
-
-        z = to_backend(z);
-
-        auto runner_ctx = get_context();
-
-        struct ggml_tensor* out = decode_graph ? ae.decode(&runner_ctx, z) : ae.encode(&runner_ctx, z);
-
-        ggml_build_forward_expand(gf, out);
-
-        return gf;
-    }
-
-    bool compute(const int n_threads,
-                 struct ggml_tensor* z,
-                 bool decode_graph,
-                 struct ggml_tensor** output,
-                 struct ggml_context* output_ctx = nullptr) override {
-        GGML_ASSERT(!decode_only || decode_graph);
-        auto get_graph = [&]() -> struct ggml_cgraph* {
-            return build_graph(z, decode_graph);
-        };
-        // ggml_set_f32(z, 0.5f);
-        // print_ggml_tensor(z);
-        return GGMLRunner::compute(get_graph, n_threads, false, output, output_ctx);
-    }
-
-    void test() {
-        struct ggml_init_params params;
-        params.mem_size   = static_cast<size_t>(10 * 1024 * 1024);  // 10 MB
-        params.mem_buffer = nullptr;
-        params.no_alloc   = false;
-
-        struct ggml_context* work_ctx = ggml_init(params);
-        GGML_ASSERT(work_ctx != nullptr);
-
-        {
-            // CPU, x{1, 3, 64, 64}: Pass
-            // CUDA, x{1, 3, 64, 64}: Pass, but sill get wrong result for some image, may be due to interlnal nan
-            // CPU, x{2, 3, 64, 64}: Wrong result
-            // CUDA, x{2, 3, 64, 64}: Wrong result, and different from CPU result
-            auto x = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 64, 64, 3, 2);
-            ggml_set_f32(x, 0.5f);
-            print_ggml_tensor(x);
-            struct ggml_tensor* out = nullptr;
-
-            int64_t t0 = ggml_time_ms();
-            compute(8, x, false, &out, work_ctx);
-            int64_t t1 = ggml_time_ms();
-
-            print_ggml_tensor(out);
-            LOG_DEBUG("encode test done in %lldms", t1 - t0);
-        }
-
-        if (false) {
-            // CPU, z{1, 4, 8, 8}: Pass
-            // CUDA, z{1, 4, 8, 8}: Pass
-            // CPU, z{3, 4, 8, 8}: Wrong result
-            // CUDA, z{3, 4, 8, 8}: Wrong result, and different from CPU result
-            auto z = ggml_new_tensor_4d(work_ctx, GGML_TYPE_F32, 8, 8, 4, 1);
-            ggml_set_f32(z, 0.5f);
-            print_ggml_tensor(z);
-            struct ggml_tensor* out = nullptr;
-
-            int64_t t0 = ggml_time_ms();
-            compute(8, z, true, &out, work_ctx);
-            int64_t t1 = ggml_time_ms();
-
-            print_ggml_tensor(out);
-            LOG_DEBUG("decode test done in %lldms", t1 - t0);
-        }
-    };
-};
-
-#endif

src/wan.hpp

@@ -25,7 +25,7 @@ namespace WAN {
         std::tuple<int, int, int> dilation;
         bool bias;
 
-        void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             params["weight"] = ggml_new_tensor_4d(ctx,
                                                   GGML_TYPE_F16,
                                                   std::get<2>(kernel_size),
@@ -53,11 +53,11 @@ namespace WAN {
               dilation(std::move(dilation)),
               bias(bias) {}
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x, struct ggml_tensor* cache_x = nullptr) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* cache_x = nullptr) {
             // x: [N*IC, ID, IH, IW]
             // result: x: [N*OC, ID, IH, IW]
-            struct ggml_tensor* w = params["weight"];
+            ggml_tensor* w = params["weight"];
-            struct ggml_tensor* b = nullptr;
+            ggml_tensor* b = nullptr;
             if (bias) {
                 b = params["bias"];
             }
@@ -86,7 +86,7 @@ namespace WAN {
     protected:
         int64_t dim;
 
-        void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             ggml_type wtype = GGML_TYPE_F32;
             auto iter       = tensor_storage_map.find(prefix + "gamma");
             if (iter != tensor_storage_map.end()) {
@@ -100,16 +100,16 @@ namespace WAN {
         RMS_norm(int64_t dim)
             : dim(dim) {}
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) override {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
             // x: [N*IC, ID, IH, IW], IC == dim
             // assert N == 1
 
-            struct ggml_tensor* w = params["gamma"];
+            ggml_tensor* w = params["gamma"];
-            w                     = ggml_reshape_1d(ctx->ggml_ctx, w, ggml_nelements(w));
+            w              = ggml_reshape_1d(ctx->ggml_ctx, w, ggml_nelements(w));
-            auto h                = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 3, 0, 1, 2));  // [ID, IH, IW, N*IC]
+            auto h         = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, x, 3, 0, 1, 2));  // [ID, IH, IW, N*IC]
-            h                     = ggml_rms_norm(ctx->ggml_ctx, h, 1e-12f);
+            h              = ggml_rms_norm(ctx->ggml_ctx, h, 1e-12f);
-            h                     = ggml_mul(ctx->ggml_ctx, h, w);
+            h              = ggml_mul(ctx->ggml_ctx, h, w);
-            h                     = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, h, 1, 2, 3, 0));
+            h              = ggml_ext_cont(ctx->ggml_ctx, ggml_ext_torch_permute(ctx->ggml_ctx, h, 1, 2, 3, 0));
 
             return h;
         }
@@ -148,12 +148,12 @@ namespace WAN {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    int64_t b,
+                             int64_t b,
-                                    std::vector<struct ggml_tensor*>& feat_cache,
+                             std::vector<ggml_tensor*>& feat_cache,
-                                    int& feat_idx,
+                             int& feat_idx,
-                                    int chunk_idx) {
+                             int chunk_idx) {
             // x: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
             int64_t c = x->ne[3] / b;
@@ -254,9 +254,9 @@ namespace WAN {
             GGML_ASSERT(in_channels * factor % out_channels == 0);
             group_size = in_channels * factor / out_channels;
         }
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    int64_t B = 1) {
+                             int64_t B = 1) {
             // x: [B*IC, T, H, W]
             // return: [B*OC, T/factor_t, H/factor_s, W/factor_s]
             GGML_ASSERT(B == 1);
@@ -301,10 +301,10 @@ namespace WAN {
             GGML_ASSERT(out_channels * factor % in_channels == 0);
             repeats = out_channels * factor / in_channels;
         }
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    bool first_chunk = false,
+                             bool first_chunk = false,
-                                    int64_t B        = 1) {
+                             int64_t B        = 1) {
             // x: [B*IC, T, H, W]
             // return: [B*OC, T/factor_t, H/factor_s, W/factor_s]
             GGML_ASSERT(B == 1);
@@ -356,14 +356,14 @@ namespace WAN {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    int64_t b,
+                             int64_t b,
-                                    std::vector<struct ggml_tensor*>& feat_cache,
+                             std::vector<ggml_tensor*>& feat_cache,
-                                    int& feat_idx) {
+                             int& feat_idx) {
             // x: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
-            struct ggml_tensor* h = x;
+            ggml_tensor* h = x;
             if (in_dim != out_dim) {
                 auto shortcut = std::dynamic_pointer_cast<CausalConv3d>(blocks["shortcut"]);
 
@@ -430,15 +430,15 @@ namespace WAN {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    int64_t b,
+                             int64_t b,
-                                    std::vector<struct ggml_tensor*>& feat_cache,
+                             std::vector<ggml_tensor*>& feat_cache,
-                                    int& feat_idx,
+                             int& feat_idx,
-                                    int chunk_idx) {
+                             int chunk_idx) {
             // x: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
-            struct ggml_tensor* x_copy = x;
+            ggml_tensor* x_copy = x;
 
             auto avg_shortcut = std::dynamic_pointer_cast<AvgDown3D>(blocks["avg_shortcut"]);
 
@@ -492,15 +492,15 @@ namespace WAN {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    int64_t b,
+                             int64_t b,
-                                    std::vector<struct ggml_tensor*>& feat_cache,
+                             std::vector<ggml_tensor*>& feat_cache,
-                                    int& feat_idx,
+                             int& feat_idx,
-                                    int chunk_idx) {
+                             int chunk_idx) {
             // x: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
-            struct ggml_tensor* x_copy = x;
+            ggml_tensor* x_copy = x;
 
             int i = 0;
             for (; i < mult; i++) {
@@ -537,9 +537,9 @@ namespace WAN {
             blocks["proj"]   = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {1, 1}));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    int64_t b) {
+                             int64_t b) {
             // x: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
             auto norm   = std::dynamic_pointer_cast<RMS_norm>(blocks["norm"]);
@@ -659,12 +659,12 @@ namespace WAN {
             blocks["head.2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(out_dim, z_dim, {3, 3, 3}, {1, 1, 1}, {1, 1, 1}));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    int64_t b,
+                             int64_t b,
-                                    std::vector<struct ggml_tensor*>& feat_cache,
+                             std::vector<ggml_tensor*>& feat_cache,
-                                    int& feat_idx,
+                             int& feat_idx,
-                                    int chunk_idx) {
+                             int chunk_idx) {
             // x: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
             auto conv1    = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
@@ -830,12 +830,12 @@ namespace WAN {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    int64_t b,
+                             int64_t b,
-                                    std::vector<struct ggml_tensor*>& feat_cache,
+                             std::vector<ggml_tensor*>& feat_cache,
-                                    int& feat_idx,
+                             int& feat_idx,
-                                    int chunk_idx) {
+                             int chunk_idx) {
             // x: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
             auto conv1    = std::dynamic_pointer_cast<CausalConv3d>(blocks["conv1"]);
@@ -934,16 +934,16 @@ namespace WAN {
 
         int _conv_num = 33;
         int _conv_idx = 0;
-        std::vector<struct ggml_tensor*> _feat_map;
+        std::vector<ggml_tensor*> _feat_map;
         int _enc_conv_num = 28;
         int _enc_conv_idx = 0;
-        std::vector<struct ggml_tensor*> _enc_feat_map;
+        std::vector<ggml_tensor*> _enc_feat_map;
 
         void clear_cache() {
             _conv_idx     = 0;
-            _feat_map     = std::vector<struct ggml_tensor*>(_conv_num, nullptr);
+            _feat_map     = std::vector<ggml_tensor*>(_conv_num, nullptr);
             _enc_conv_idx = 0;
-            _enc_feat_map = std::vector<struct ggml_tensor*>(_enc_conv_num, nullptr);
+            _enc_feat_map = std::vector<ggml_tensor*>(_enc_conv_num, nullptr);
         }
 
     public:
@@ -966,10 +966,10 @@ namespace WAN {
             blocks["conv2"]   = std::shared_ptr<GGMLBlock>(new CausalConv3d(z_dim, z_dim, {1, 1, 1}));
         }
 
-        struct ggml_tensor* patchify(struct ggml_context* ctx,
+        ggml_tensor* patchify(ggml_context* ctx,
-                                     struct ggml_tensor* x,
+                              ggml_tensor* x,
-                                     int64_t patch_size,
+                              int64_t patch_size,
-                                     int64_t b = 1) {
+                              int64_t b = 1) {
             // x: [b*c, f, h*q, w*r]
             // return: [b*c*r*q, f, h, w]
             if (patch_size == 1) {
@@ -993,10 +993,10 @@ namespace WAN {
             return x;
         }
 
-        struct ggml_tensor* unpatchify(struct ggml_context* ctx,
+        ggml_tensor* unpatchify(ggml_context* ctx,
-                                       struct ggml_tensor* x,
+                                ggml_tensor* x,
-                                       int64_t patch_size,
+                                int64_t patch_size,
-                                       int64_t b = 1) {
+                                int64_t b = 1) {
             // x: [b*c*r*q, f, h, w]
             // return: [b*c, f, h*q, w*r]
             if (patch_size == 1) {
@@ -1019,9 +1019,9 @@ namespace WAN {
             return x;
         }
 
-        struct ggml_tensor* encode(GGMLRunnerContext* ctx,
+        ggml_tensor* encode(GGMLRunnerContext* ctx,
-                                   struct ggml_tensor* x,
+                            ggml_tensor* x,
-                                   int64_t b = 1) {
+                            int64_t b = 1) {
             // x: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
             GGML_ASSERT(decode_only == false);
@@ -1037,7 +1037,7 @@ namespace WAN {
 
             int64_t t     = x->ne[2];
             int64_t iter_ = 1 + (t - 1) / 4;
-            struct ggml_tensor* out;
+            ggml_tensor* out;
             for (int i = 0; i < iter_; i++) {
                 _enc_conv_idx = 0;
                 if (i == 0) {
@@ -1055,9 +1055,9 @@ namespace WAN {
             return mu;
         }
 
-        struct ggml_tensor* decode(GGMLRunnerContext* ctx,
+        ggml_tensor* decode(GGMLRunnerContext* ctx,
-                                   struct ggml_tensor* z,
+                            ggml_tensor* z,
-                                   int64_t b = 1) {
+                            int64_t b = 1) {
             // z: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
 
@@ -1068,7 +1068,7 @@ namespace WAN {
 
             int64_t iter_ = z->ne[2];
             auto x        = conv2->forward(ctx, z);
-            struct ggml_tensor* out;
+            ggml_tensor* out;
             for (int i = 0; i < iter_; i++) {
                 _conv_idx = 0;
                 if (i == 0) {
@@ -1087,10 +1087,10 @@ namespace WAN {
             return out;
         }
 
-        struct ggml_tensor* decode_partial(GGMLRunnerContext* ctx,
+        ggml_tensor* decode_partial(GGMLRunnerContext* ctx,
-                                           struct ggml_tensor* z,
+                                    ggml_tensor* z,
-                                           int i,
+                                    int i,
-                                           int64_t b = 1) {
+                                    int64_t b = 1) {
             // z: [b*c, t, h, w]
             GGML_ASSERT(b == 1);
 
@@ -1109,7 +1109,8 @@ namespace WAN {
     };
 
     struct WanVAERunner : public VAE {
-        bool decode_only = true;
+        float scale_factor = 1.0f;
+        bool decode_only   = true;
         WanVAE ae;
 
         WanVAERunner(ggml_backend_t backend,
@@ -1118,7 +1119,7 @@ namespace WAN {
                      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(backend, offload_params_to_cpu) {
+            : decode_only(decode_only), ae(decode_only, version == VERSION_WAN2_2_TI2V), VAE(version, backend, offload_params_to_cpu) {
             ae.init(params_ctx, tensor_storage_map, prefix);
         }
 
@@ -1126,26 +1127,121 @@ namespace WAN {
             return "wan_vae";
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) override {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) override {
             ae.get_param_tensors(tensors, prefix);
         }
 
-        struct ggml_cgraph* build_graph(struct ggml_tensor* z, bool decode_graph) {
+        ggml_tensor* vae_output_to_latents(ggml_context* work_ctx, ggml_tensor* vae_output, std::shared_ptr<RNG> rng) {
-            struct ggml_cgraph* gf = new_graph_custom(10240 * z->ne[2]);
+            return vae_output;
+        }
+
+        void get_latents_mean_std_vec(ggml_tensor* latents, int channel_dim, std::vector<float>& latents_mean_vec, std::vector<float>& latents_std_vec) {
+            GGML_ASSERT(latents->ne[channel_dim] == 16 || latents->ne[channel_dim] == 48);
+            if (latents->ne[channel_dim] == 16) {  // Wan2.1 VAE
+                latents_mean_vec = {-0.7571f, -0.7089f, -0.9113f, 0.1075f, -0.1745f, 0.9653f, -0.1517f, 1.5508f,
+                                    0.4134f, -0.0715f, 0.5517f, -0.3632f, -0.1922f, -0.9497f, 0.2503f, -0.2921f};
+                latents_std_vec  = {2.8184f, 1.4541f, 2.3275f, 2.6558f, 1.2196f, 1.7708f, 2.6052f, 2.0743f,
+                                    3.2687f, 2.1526f, 2.8652f, 1.5579f, 1.6382f, 1.1253f, 2.8251f, 1.9160f};
+            } else if (latents->ne[channel_dim] == 48) {  // Wan2.2 VAE
+                latents_mean_vec = {-0.2289f, -0.0052f, -0.1323f, -0.2339f, -0.2799f, 0.0174f, 0.1838f, 0.1557f,
+                                    -0.1382f, 0.0542f, 0.2813f, 0.0891f, 0.1570f, -0.0098f, 0.0375f, -0.1825f,
+                                    -0.2246f, -0.1207f, -0.0698f, 0.5109f, 0.2665f, -0.2108f, -0.2158f, 0.2502f,
+                                    -0.2055f, -0.0322f, 0.1109f, 0.1567f, -0.0729f, 0.0899f, -0.2799f, -0.1230f,
+                                    -0.0313f, -0.1649f, 0.0117f, 0.0723f, -0.2839f, -0.2083f, -0.0520f, 0.3748f,
+                                    0.0152f, 0.1957f, 0.1433f, -0.2944f, 0.3573f, -0.0548f, -0.1681f, -0.0667f};
+                latents_std_vec  = {
+                     0.4765f, 1.0364f, 0.4514f, 1.1677f, 0.5313f, 0.4990f, 0.4818f, 0.5013f,
+                     0.8158f, 1.0344f, 0.5894f, 1.0901f, 0.6885f, 0.6165f, 0.8454f, 0.4978f,
+                     0.5759f, 0.3523f, 0.7135f, 0.6804f, 0.5833f, 1.4146f, 0.8986f, 0.5659f,
+                     0.7069f, 0.5338f, 0.4889f, 0.4917f, 0.4069f, 0.4999f, 0.6866f, 0.4093f,
+                     0.5709f, 0.6065f, 0.6415f, 0.4944f, 0.5726f, 1.2042f, 0.5458f, 1.6887f,
+                     0.3971f, 1.0600f, 0.3943f, 0.5537f, 0.5444f, 0.4089f, 0.7468f, 0.7744f};
+            }
+        }
+
+        ggml_tensor* diffusion_to_vae_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+            ggml_tensor* vae_latents = ggml_dup(work_ctx, latents);
+            int channel_dim          = sd_version_is_wan(version) ? 3 : 2;
+            std::vector<float> latents_mean_vec;
+            std::vector<float> latents_std_vec;
+            get_latents_mean_std_vec(latents, channel_dim, latents_mean_vec, latents_std_vec);
+
+            float mean;
+            float std_;
+            for (int i = 0; i < latents->ne[3]; i++) {
+                if (channel_dim == 3) {
+                    mean = latents_mean_vec[i];
+                    std_ = latents_std_vec[i];
+                }
+                for (int j = 0; j < latents->ne[2]; j++) {
+                    if (channel_dim == 2) {
+                        mean = latents_mean_vec[j];
+                        std_ = latents_std_vec[j];
+                    }
+                    for (int k = 0; k < latents->ne[1]; k++) {
+                        for (int l = 0; l < latents->ne[0]; l++) {
+                            float value = ggml_ext_tensor_get_f32(latents, l, k, j, i);
+                            value       = value * std_ / scale_factor + mean;
+                            ggml_ext_tensor_set_f32(vae_latents, value, l, k, j, i);
+                        }
+                    }
+                }
+            }
+
+            return vae_latents;
+        }
+
+        ggml_tensor* vae_to_diffuison_latents(ggml_context* work_ctx, ggml_tensor* latents) {
+            ggml_tensor* diffusion_latents = ggml_dup(work_ctx, latents);
+            int channel_dim                = sd_version_is_wan(version) ? 3 : 2;
+            std::vector<float> latents_mean_vec;
+            std::vector<float> latents_std_vec;
+            get_latents_mean_std_vec(latents, channel_dim, latents_mean_vec, latents_std_vec);
+
+            float mean;
+            float std_;
+            for (int i = 0; i < latents->ne[3]; i++) {
+                if (channel_dim == 3) {
+                    mean = latents_mean_vec[i];
+                    std_ = latents_std_vec[i];
+                }
+                for (int j = 0; j < latents->ne[2]; j++) {
+                    if (channel_dim == 2) {
+                        mean = latents_mean_vec[j];
+                        std_ = latents_std_vec[j];
+                    }
+                    for (int k = 0; k < latents->ne[1]; k++) {
+                        for (int l = 0; l < latents->ne[0]; l++) {
+                            float value = ggml_ext_tensor_get_f32(latents, l, k, j, i);
+                            value       = (value - mean) * scale_factor / std_;
+                            ggml_ext_tensor_set_f32(diffusion_latents, value, l, k, j, i);
+                        }
+                    }
+                }
+            }
+            return diffusion_latents;
+        }
+
+        int get_encoder_output_channels(int input_channels) {
+            return static_cast<int>(ae.z_dim);
+        }
+
+        ggml_cgraph* build_graph(ggml_tensor* z, bool decode_graph) {
+            ggml_cgraph* gf = new_graph_custom(10240 * z->ne[2]);
 
             z = to_backend(z);
 
             auto runner_ctx = get_context();
 
-            struct ggml_tensor* out = decode_graph ? ae.decode(&runner_ctx, z) : ae.encode(&runner_ctx, z);
+            ggml_tensor* out = decode_graph ? ae.decode(&runner_ctx, z) : ae.encode(&runner_ctx, z);
 
             ggml_build_forward_expand(gf, out);
 
             return gf;
         }
 
-        struct ggml_cgraph* build_graph_partial(struct ggml_tensor* z, bool decode_graph, int i) {
+        ggml_cgraph* build_graph_partial(ggml_tensor* z, bool decode_graph, int i) {
-            struct ggml_cgraph* gf = new_graph_custom(20480);
+            ggml_cgraph* gf = new_graph_custom(20480);
 
             ae.clear_cache();
 
@@ -1158,7 +1254,7 @@ namespace WAN {
 
             auto runner_ctx = get_context();
 
-            struct ggml_tensor* out = decode_graph ? ae.decode_partial(&runner_ctx, z, i) : ae.encode(&runner_ctx, z);
+            ggml_tensor* out = decode_graph ? ae.decode_partial(&runner_ctx, z, i) : ae.encode(&runner_ctx, z);
 
             for (size_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) {
                 ggml_tensor* feat_cache = ae._feat_map[feat_idx];
@@ -1173,13 +1269,13 @@ namespace WAN {
             return gf;
         }
 
-        bool compute(const int n_threads,
+        bool _compute(const int n_threads,
-                     struct ggml_tensor* z,
+                      ggml_tensor* z,
-                     bool decode_graph,
+                      bool decode_graph,
-                     struct ggml_tensor** output,
+                      ggml_tensor** output,
-                     struct ggml_context* output_ctx = nullptr) override {
+                      ggml_context* output_ctx = nullptr) override {
             if (true) {
-                auto get_graph = [&]() -> struct ggml_cgraph* {
+                auto get_graph = [&]() -> ggml_cgraph* {
                     return build_graph(z, decode_graph);
                 };
                 return GGMLRunner::compute(get_graph, n_threads, true, output, output_ctx);
@@ -1187,11 +1283,11 @@ namespace WAN {
                 ae.clear_cache();
                 int64_t t      = z->ne[2];
                 int i          = 0;
-                auto get_graph = [&]() -> struct ggml_cgraph* {
+                auto get_graph = [&]() -> ggml_cgraph* {
                     return build_graph_partial(z, decode_graph, i);
                 };
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
-                bool res                = GGMLRunner::compute(get_graph, n_threads, true, &out, output_ctx);
+                bool res         = GGMLRunner::compute(get_graph, n_threads, true, &out, output_ctx);
                 ae.clear_cache();
                 if (t == 1) {
                     *output = out;
@@ -1229,12 +1325,12 @@ namespace WAN {
         }
 
         void test() {
-            struct ggml_init_params params;
+            ggml_init_params params;
             params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1G
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            struct ggml_context* work_ctx = ggml_init(params);
+            ggml_context* work_ctx = ggml_init(params);
             GGML_ASSERT(work_ctx != nullptr);
 
             if (true) {
@@ -1246,10 +1342,10 @@ namespace WAN {
                 ggml_set_f32(z, 0.5f);
                 z = load_tensor_from_file(work_ctx, "wan_vae_z.bin");
                 print_ggml_tensor(z);
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
-                compute(8, z, true, &out, work_ctx);
+                _compute(8, z, true, &out, work_ctx);
                 int64_t t1 = ggml_time_ms();
 
                 print_ggml_tensor(out);
@@ -1314,10 +1410,10 @@ namespace WAN {
             }
         }
 
-        virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        virtual ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                            struct ggml_tensor* x,
+                                     ggml_tensor* x,
-                                            struct ggml_tensor* pe,
+                                     ggml_tensor* pe,
-                                            struct ggml_tensor* mask = nullptr) {
+                                     ggml_tensor* mask = nullptr) {
             // x: [N, n_token, dim]
             // pe: [n_token, d_head/2, 2, 2]
             // return [N, n_token, dim]
@@ -1355,10 +1451,10 @@ namespace WAN {
                           bool qk_norm = true,
                           float eps    = 1e-6)
             : WanSelfAttention(dim, num_heads, qk_norm, eps) {}
-        virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        virtual ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                            struct ggml_tensor* x,
+                                     ggml_tensor* x,
-                                            struct ggml_tensor* context,
+                                     ggml_tensor* context,
-                                            int64_t context_img_len) = 0;
+                                     int64_t context_img_len) = 0;
     };
 
     class WanT2VCrossAttention : public WanCrossAttention {
@@ -1368,10 +1464,10 @@ namespace WAN {
                              bool qk_norm = true,
                              float eps    = 1e-6)
             : WanCrossAttention(dim, num_heads, qk_norm, eps) {}
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    int64_t context_img_len) override {
+                             int64_t context_img_len) override {
             // x: [N, n_token, dim]
             // context: [N, n_context, dim]
             // context_img_len: unused
@@ -1416,10 +1512,10 @@ namespace WAN {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    int64_t context_img_len) override {
+                             int64_t context_img_len) override {
             // x: [N, n_token, dim]
             // context: [N, context_img_len + context_txt_len, dim]
             // return [N, n_token, dim]
@@ -1464,7 +1560,7 @@ namespace WAN {
         }
     };
 
-    static struct ggml_tensor* modulate_add(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* e) {
+    static ggml_tensor* modulate_add(ggml_context* ctx, ggml_tensor* x, ggml_tensor* e) {
         // x: [N, n_token, dim]
         // e: [N, 1, dim] or [N, T, 1, dim]
         if (ggml_n_dims(e) == 3) {
@@ -1478,7 +1574,7 @@ namespace WAN {
         return x;
     }
 
-    static struct ggml_tensor* modulate_mul(struct ggml_context* ctx, struct ggml_tensor* x, struct ggml_tensor* e) {
+    static ggml_tensor* modulate_mul(ggml_context* ctx, ggml_tensor* x, ggml_tensor* e) {
         // x: [N, n_token, dim]
         // e: [N, 1, dim] or [N, T, 1, dim]
         if (ggml_n_dims(e) == 3) {
@@ -1496,7 +1592,7 @@ namespace WAN {
     protected:
         int64_t dim;
 
-        void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32);
             params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 6, 1);
         }
@@ -1530,12 +1626,12 @@ namespace WAN {
             blocks["ffn.2"] = std::shared_ptr<GGMLBlock>(new Linear(ffn_dim, dim));
         }
 
-        virtual struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        virtual ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                            struct ggml_tensor* x,
+                                     ggml_tensor* x,
-                                            struct ggml_tensor* e,
+                                     ggml_tensor* e,
-                                            struct ggml_tensor* pe,
+                                     ggml_tensor* pe,
-                                            struct ggml_tensor* context,
+                                     ggml_tensor* context,
-                                            int64_t context_img_len = 257) {
+                                     int64_t context_img_len = 257) {
             // x: [N, n_token, dim]
             // e: [N, 6, dim] or [N, T, 6, dim]
             // context: [N, context_img_len + context_txt_len, dim]
@@ -1584,7 +1680,7 @@ namespace WAN {
     class VaceWanAttentionBlock : public WanAttentionBlock {
     protected:
         int block_id;
-        void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32);
             params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 6, 1);
         }
@@ -1606,11 +1702,11 @@ namespace WAN {
         }
 
         std::pair<ggml_tensor*, ggml_tensor*> forward(GGMLRunnerContext* ctx,
-                                                      struct ggml_tensor* c,
+                                                      ggml_tensor* c,
-                                                      struct ggml_tensor* x,
+                                                      ggml_tensor* x,
-                                                      struct ggml_tensor* e,
+                                                      ggml_tensor* e,
-                                                      struct ggml_tensor* pe,
+                                                      ggml_tensor* pe,
-                                                      struct ggml_tensor* context,
+                                                      ggml_tensor* context,
                                                       int64_t context_img_len = 257) {
             // x: [N, n_token, dim]
             // e: [N, 6, dim] or [N, T, 6, dim]
@@ -1636,7 +1732,7 @@ namespace WAN {
     protected:
         int64_t dim;
 
-        void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             enum ggml_type wtype = get_type(prefix + "weight", tensor_storage_map, GGML_TYPE_F32);
             params["modulation"] = ggml_new_tensor_3d(ctx, wtype, dim, 2, 1);
         }
@@ -1653,9 +1749,9 @@ namespace WAN {
             blocks["head"] = std::shared_ptr<GGMLBlock>(new Linear(dim, out_dim));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* e) {
+                             ggml_tensor* e) {
             // x: [N, n_token, dim]
             // e: [N, dim] or [N, T, dim]
             // return [N, n_token, out_dim]
@@ -1683,7 +1779,7 @@ namespace WAN {
         int64_t in_dim;
         int64_t flf_pos_embed_token_number;
 
-        void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             if (flf_pos_embed_token_number > 0) {
                 params["emb_pos"] = ggml_new_tensor_3d(ctx, GGML_TYPE_F32, in_dim, flf_pos_embed_token_number, 1);
             }
@@ -1701,8 +1797,8 @@ namespace WAN {
             blocks["proj.4"] = std::shared_ptr<GGMLBlock>(new LayerNorm(out_dim));
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* image_embeds) {
+                             ggml_tensor* image_embeds) {
             if (flf_pos_embed_token_number > 0) {
                 auto emb_pos = params["emb_pos"];
 
@@ -1821,8 +1917,8 @@ namespace WAN {
             }
         }
 
-        struct ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
+        ggml_tensor* pad_to_patch_size(GGMLRunnerContext* ctx,
-                                              struct ggml_tensor* x) {
+                                       ggml_tensor* x) {
             int64_t W = x->ne[0];
             int64_t H = x->ne[1];
             int64_t T = x->ne[2];
@@ -1834,11 +1930,11 @@ namespace WAN {
             return x;
         }
 
-        struct ggml_tensor* unpatchify(struct ggml_context* ctx,
+        ggml_tensor* unpatchify(ggml_context* ctx,
-                                       struct ggml_tensor* x,
+                                ggml_tensor* x,
-                                       int64_t t_len,
+                                int64_t t_len,
-                                       int64_t h_len,
+                                int64_t h_len,
-                                       int64_t w_len) {
+                                int64_t w_len) {
             // x: [N, t_len*h_len*w_len, pt*ph*pw*C]
             // return: [N*C, t_len*pt, h_len*ph, w_len*pw]
             int64_t N  = x->ne[3];
@@ -1861,15 +1957,15 @@ namespace WAN {
             return x;
         }
 
-        struct ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
+        ggml_tensor* forward_orig(GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
+                                  ggml_tensor* x,
-                                         struct ggml_tensor* timestep,
+                                  ggml_tensor* timestep,
-                                         struct ggml_tensor* context,
+                                  ggml_tensor* context,
-                                         struct ggml_tensor* pe,
+                                  ggml_tensor* pe,
-                                         struct ggml_tensor* clip_fea     = nullptr,
+                                  ggml_tensor* clip_fea     = nullptr,
-                                         struct ggml_tensor* vace_context = nullptr,
+                                  ggml_tensor* vace_context = nullptr,
-                                         float vace_strength              = 1.f,
+                                  float vace_strength       = 1.f,
-                                         int64_t N                        = 1) {
+                                  int64_t N                 = 1) {
             // x: [N*C, T, H, W], C => in_dim
             // vace_context: [N*vace_in_dim, T, H, W]
             // timestep: [N,] or [T]
@@ -1955,16 +2051,16 @@ namespace WAN {
             return x;
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* timestep,
+                             ggml_tensor* timestep,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* clip_fea        = nullptr,
+                             ggml_tensor* clip_fea        = nullptr,
-                                    struct ggml_tensor* time_dim_concat = nullptr,
+                             ggml_tensor* time_dim_concat = nullptr,
-                                    struct ggml_tensor* vace_context    = nullptr,
+                             ggml_tensor* vace_context    = nullptr,
-                                    float vace_strength                 = 1.f,
+                             float vace_strength          = 1.f,
-                                    int64_t N                           = 1) {
+                             int64_t N                    = 1) {
             // Forward pass of DiT.
             // x: [N*C, T, H, W]
             // timestep: [N,]
@@ -2129,19 +2225,19 @@ namespace WAN {
             return desc;
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
             wan.get_param_tensors(tensors, prefix);
         }
 
-        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+        ggml_cgraph* build_graph(ggml_tensor* x,
-                                        struct ggml_tensor* timesteps,
+                                 ggml_tensor* timesteps,
-                                        struct ggml_tensor* context,
+                                 ggml_tensor* context,
-                                        struct ggml_tensor* clip_fea        = nullptr,
+                                 ggml_tensor* clip_fea        = nullptr,
-                                        struct ggml_tensor* c_concat        = nullptr,
+                                 ggml_tensor* c_concat        = nullptr,
-                                        struct ggml_tensor* time_dim_concat = nullptr,
+                                 ggml_tensor* time_dim_concat = nullptr,
-                                        struct ggml_tensor* vace_context    = nullptr,
+                                 ggml_tensor* vace_context    = nullptr,
-                                        float vace_strength                 = 1.f) {
+                                 float vace_strength          = 1.f) {
-            struct ggml_cgraph* gf = new_graph_custom(WAN_GRAPH_SIZE);
+            ggml_cgraph* gf = new_graph_custom(WAN_GRAPH_SIZE);
 
             x               = to_backend(x);
             timesteps       = to_backend(timesteps);
@@ -2174,15 +2270,15 @@ namespace WAN {
 
             auto runner_ctx = get_context();
 
-            struct ggml_tensor* out = wan.forward(&runner_ctx,
+            ggml_tensor* out = wan.forward(&runner_ctx,
-                                                  x,
+                                           x,
-                                                  timesteps,
+                                           timesteps,
-                                                  context,
+                                           context,
-                                                  pe,
+                                           pe,
-                                                  clip_fea,
+                                           clip_fea,
-                                                  time_dim_concat,
+                                           time_dim_concat,
-                                                  vace_context,
+                                           vace_context,
-                                                  vace_strength);
+                                           vace_strength);
 
             ggml_build_forward_expand(gf, out);
 
@@ -2190,17 +2286,17 @@ namespace WAN {
         }
 
         bool compute(int n_threads,
-                     struct ggml_tensor* x,
+                     ggml_tensor* x,
-                     struct ggml_tensor* timesteps,
+                     ggml_tensor* timesteps,
-                     struct ggml_tensor* context,
+                     ggml_tensor* context,
-                     struct ggml_tensor* clip_fea        = nullptr,
+                     ggml_tensor* clip_fea        = nullptr,
-                     struct ggml_tensor* c_concat        = nullptr,
+                     ggml_tensor* c_concat        = nullptr,
-                     struct ggml_tensor* time_dim_concat = nullptr,
+                     ggml_tensor* time_dim_concat = nullptr,
-                     struct ggml_tensor* vace_context    = nullptr,
+                     ggml_tensor* vace_context    = nullptr,
-                     float vace_strength                 = 1.f,
+                     float vace_strength          = 1.f,
-                     struct ggml_tensor** output         = nullptr,
+                     ggml_tensor** output         = nullptr,
-                     struct ggml_context* output_ctx     = nullptr) {
+                     ggml_context* output_ctx     = nullptr) {
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+            auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(x, timesteps, context, clip_fea, c_concat, time_dim_concat, vace_context, vace_strength);
             };
 
@@ -2208,12 +2304,12 @@ namespace WAN {
         }
 
         void test() {
-            struct ggml_init_params params;
+            ggml_init_params params;
             params.mem_size   = static_cast<size_t>(200 * 1024 * 1024);  // 200 MB
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            struct ggml_context* work_ctx = ggml_init(params);
+            ggml_context* work_ctx = ggml_init(params);
             GGML_ASSERT(work_ctx != nullptr);
 
             {
@@ -2236,7 +2332,7 @@ namespace WAN {
                 // auto clip_fea = load_tensor_from_file(work_ctx, "wan_dit_clip_fea.bin");
                 // print_ggml_tensor(clip_fea);
 
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 compute(8, x, timesteps, context, nullptr, nullptr, nullptr, nullptr, 1.f, &out, work_ctx);

src/z_image.hpp

@@ -42,10 +42,10 @@ namespace ZImage {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* mask = nullptr) {
+                             ggml_tensor* mask = nullptr) {
             // x: [N, n_token, hidden_size]
             int64_t n_token = x->ne[1];
             int64_t N       = x->ne[2];
@@ -124,23 +124,23 @@ namespace ZImage {
             blocks["w3"] = std::make_shared<Linear>(dim, hidden_dim, false);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx, struct ggml_tensor* x) {
+        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
             auto w1 = std::dynamic_pointer_cast<Linear>(blocks["w1"]);
             auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
             auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
 
             auto x1 = w1->forward(ctx, x);
             auto x3 = w3->forward(ctx, x);
-            x       = ggml_mul(ctx->ggml_ctx, ggml_silu(ctx->ggml_ctx, x1), x3);
+            x       = ggml_swiglu_split(ctx->ggml_ctx, x1, x3);
             x       = w2->forward(ctx, x);
 
             return x;
         }
     };
 
-    __STATIC_INLINE__ struct ggml_tensor* modulate(struct ggml_context* ctx,
+    __STATIC_INLINE__ ggml_tensor* modulate(ggml_context* ctx,
-                                                   struct ggml_tensor* x,
+                                            ggml_tensor* x,
-                                                   struct ggml_tensor* scale) {
+                                            ggml_tensor* scale) {
         // x: [N, L, C]
         // scale: [N, C]
         scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
@@ -175,11 +175,11 @@ namespace ZImage {
             }
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    struct ggml_tensor* mask        = nullptr,
+                             ggml_tensor* mask        = nullptr,
-                                    struct ggml_tensor* adaln_input = nullptr) {
+                             ggml_tensor* adaln_input = nullptr) {
             auto attention       = std::dynamic_pointer_cast<JointAttention>(blocks["attention"]);
             auto feed_forward    = std::dynamic_pointer_cast<FeedForward>(blocks["feed_forward"]);
             auto attention_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["attention_norm1"]);
@@ -241,9 +241,9 @@ namespace ZImage {
             blocks["adaLN_modulation.1"] = std::make_shared<Linear>(MIN(hidden_size, ADALN_EMBED_DIM), hidden_size);
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* c) {
+                             ggml_tensor* c) {
             // x: [N, n_token, hidden_size]
             // c: [N, hidden_size]
             // return: [N, n_token, patch_size * patch_size * out_channels]
@@ -284,7 +284,7 @@ namespace ZImage {
     protected:
         ZImageParams z_image_params;
 
-        void init_params(struct ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
+        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
             params["cap_pad_token"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, z_image_params.hidden_size);
             params["x_pad_token"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, z_image_params.hidden_size);
         }
@@ -346,11 +346,11 @@ namespace ZImage {
             blocks["final_layer"] = std::make_shared<FinalLayer>(z_image_params.hidden_size, z_image_params.patch_size, z_image_params.out_channels);
         }
 
-        struct ggml_tensor* forward_core(GGMLRunnerContext* ctx,
+        ggml_tensor* forward_core(GGMLRunnerContext* ctx,
-                                         struct ggml_tensor* x,
+                                  ggml_tensor* x,
-                                         struct ggml_tensor* timestep,
+                                  ggml_tensor* timestep,
-                                         struct ggml_tensor* context,
+                                  ggml_tensor* context,
-                                         struct ggml_tensor* pe) {
+                                  ggml_tensor* pe) {
             auto x_embedder     = std::dynamic_pointer_cast<Linear>(blocks["x_embedder"]);
             auto t_embedder     = std::dynamic_pointer_cast<TimestepEmbedder>(blocks["t_embedder"]);
             auto cap_embedder_0 = std::dynamic_pointer_cast<RMSNorm>(blocks["cap_embedder.0"]);
@@ -414,12 +414,12 @@ namespace ZImage {
             return img;
         }
 
-        struct ggml_tensor* forward(GGMLRunnerContext* ctx,
+        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                                    struct ggml_tensor* x,
+                             ggml_tensor* x,
-                                    struct ggml_tensor* timestep,
+                             ggml_tensor* timestep,
-                                    struct ggml_tensor* context,
+                             ggml_tensor* context,
-                                    struct ggml_tensor* pe,
+                             ggml_tensor* pe,
-                                    std::vector<ggml_tensor*> ref_latents = {}) {
+                             std::vector<ggml_tensor*> ref_latents = {}) {
             // Forward pass of DiT.
             // x: [N, C, H, W]
             // timestep: [N,]
@@ -477,17 +477,17 @@ namespace ZImage {
             return "z_image";
         }
 
-        void get_param_tensors(std::map<std::string, struct ggml_tensor*>& tensors, const std::string prefix) {
+        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string prefix) {
             z_image.get_param_tensors(tensors, prefix);
         }
 
-        struct ggml_cgraph* build_graph(struct ggml_tensor* x,
+        ggml_cgraph* build_graph(ggml_tensor* x,
-                                        struct ggml_tensor* timesteps,
+                                 ggml_tensor* timesteps,
-                                        struct ggml_tensor* context,
+                                 ggml_tensor* context,
-                                        std::vector<ggml_tensor*> ref_latents = {},
+                                 std::vector<ggml_tensor*> ref_latents = {},
-                                        bool increase_ref_index               = false) {
+                                 bool increase_ref_index               = false) {
             GGML_ASSERT(x->ne[3] == 1);
-            struct ggml_cgraph* gf = new_graph_custom(Z_IMAGE_GRAPH_SIZE);
+            ggml_cgraph* gf = new_graph_custom(Z_IMAGE_GRAPH_SIZE);
 
             x         = to_backend(x);
             context   = to_backend(context);
@@ -518,12 +518,12 @@ namespace ZImage {
             set_backend_tensor_data(pe, pe_vec.data());
             auto runner_ctx = get_context();
 
-            struct ggml_tensor* out = z_image.forward(&runner_ctx,
+            ggml_tensor* out = z_image.forward(&runner_ctx,
-                                                      x,
+                                               x,
-                                                      timesteps,
+                                               timesteps,
-                                                      context,
+                                               context,
-                                                      pe,
+                                               pe,
-                                                      ref_latents);
+                                               ref_latents);
 
             ggml_build_forward_expand(gf, out);
 
@@ -531,17 +531,17 @@ namespace ZImage {
         }
 
         bool compute(int n_threads,
-                     struct ggml_tensor* x,
+                     ggml_tensor* x,
-                     struct ggml_tensor* timesteps,
+                     ggml_tensor* timesteps,
-                     struct ggml_tensor* context,
+                     ggml_tensor* context,
                      std::vector<ggml_tensor*> ref_latents = {},
                      bool increase_ref_index               = false,
-                     struct ggml_tensor** output           = nullptr,
+                     ggml_tensor** output                  = nullptr,
-                     struct ggml_context* output_ctx       = nullptr) {
+                     ggml_context* output_ctx              = nullptr) {
             // x: [N, in_channels, h, w]
             // timesteps: [N, ]
             // context: [N, max_position, hidden_size]
-            auto get_graph = [&]() -> struct ggml_cgraph* {
+            auto get_graph = [&]() -> ggml_cgraph* {
                 return build_graph(x, timesteps, context, ref_latents, increase_ref_index);
             };
 
@@ -549,12 +549,12 @@ namespace ZImage {
         }
 
         void test() {
-            struct ggml_init_params params;
+            ggml_init_params params;
             params.mem_size   = static_cast<size_t>(1024 * 1024) * 1024;  // 1GB
             params.mem_buffer = nullptr;
             params.no_alloc   = false;
 
-            struct ggml_context* work_ctx = ggml_init(params);
+            ggml_context* work_ctx = ggml_init(params);
             GGML_ASSERT(work_ctx != nullptr);
 
             {
@@ -571,7 +571,7 @@ namespace ZImage {
                 auto context = load_tensor_from_file(work_ctx, "./z_image_context.bin");
                 print_ggml_tensor(context);
 
-                struct ggml_tensor* out = nullptr;
+                ggml_tensor* out = nullptr;
 
                 int64_t t0 = ggml_time_ms();
                 compute(8, x, timesteps, context, {}, false, &out, work_ctx);