.github/workflows/close-organization-fork-prs.yml

@@ -1,61 +0,0 @@
-name: Close PRs from organization forks
-
-on:
-  pull_request_target:
-    types: [opened, reopened]
-
-permissions:
-  pull-requests: write
-
-concurrency:
-  group: ${{ github.workflow }}-${{ github.event.pull_request.number }}
-  cancel-in-progress: false
-
-jobs:
-  close-organization-fork-pr:
-    if: >-
-      github.event.pull_request.head.repo.owner.type == 'Organization' &&
-      github.event.pull_request.head.repo.id != github.event.pull_request.base.repo.id
-    runs-on: ubuntu-latest
-    timeout-minutes: 5
-    steps:
-      - name: Explain the contribution policy and close the PR
-        uses: actions/github-script@v9
-        with:
-          script: |
-            const { data: pr } = await github.rest.pulls.get({
-              ...context.repo,
-              pull_number: context.issue.number,
-            });
-            const headRepo = pr.head.repo;
-            if (pr.state !== 'open' || !headRepo ||
-                headRepo.id === pr.base.repo.id || headRepo.owner.type !== 'Organization') {
-              return;
-            }
-
-            const marker = '<!-- organization-fork-policy -->';
-            const comments = await github.paginate(github.rest.issues.listComments, {
-              ...context.repo,
-              issue_number: pr.number,
-              per_page: 100,
-            });
-            const alreadyExplained = comments.some(comment =>
-              comment.user?.login === 'github-actions[bot]' && comment.body?.includes(marker));
-            if (!alreadyExplained) {
-              await github.rest.issues.createComment({
-                ...context.repo,
-                issue_number: pr.number,
-                body: [
-                  marker,
-                  'This repository requires contributions from forks to use a personal fork with **Allow edits from maintainers** enabled.',
-                  'GitHub does not support this option for organization-owned forks, so this PR is being closed automatically.',
-                  'Please open a new PR from a fork in your personal GitHub account and enable **Allow edits from maintainers** so maintainers can help update the branch.',
-                  'See [the GitHub documentation](https://docs.github.com/en/pull-requests/how-tos/work-with-forks/allowing-changes-to-a-pull-request-branch-created-from-a-fork).',
-                ].join('\n\n'),
-              });
-            }
-            await github.rest.pulls.update({
-              ...context.repo,
-              pull_number: pr.number,
-              state: 'closed',
-            });

CMakeLists.txt

@@ -95,8 +95,6 @@ option(SD_MUSA                       "sd: musa backend" OFF)
 option(SD_BUILD_SHARED_LIBS          "sd: build shared libs" OFF)
 option(SD_BUILD_SHARED_GGML_LIB      "sd: build ggml as a separate shared lib" OFF)
 option(SD_USE_SYSTEM_GGML            "sd: use system-installed GGML library" OFF)
-option(SD_USE_UPSTREAM_GGML          "sd: build with upstream GGML instead of the patched GGML extensions" OFF)
-set(SD_GGML_SOURCE_DIR "${CMAKE_CURRENT_SOURCE_DIR}/ggml" CACHE PATH "sd: ggml source directory (also supplies private headers for system ggml)")
 #option(SD_BUILD_SERVER               "sd: build server example"                           ON)
 
 set(CMAKE_C_STANDARD 11)
@@ -327,21 +325,24 @@ if (NOT SD_USE_SYSTEM_GGML)
 endif()
 
 # deps
-include(cmake/ggml.cmake)
+# Only add ggml if it hasn't been added yet
+if (NOT TARGET ggml)
+    if (SD_USE_SYSTEM_GGML)
+        find_package(ggml REQUIRED)
+        if (NOT ggml_FOUND)
+            message(FATAL_ERROR "System-installed GGML library not found.")
+        endif()
+        add_library(ggml ALIAS ggml::ggml)
+    else()
+        add_subdirectory(ggml)
+    endif()
+endif()
 
 add_subdirectory(thirdparty)
 
 target_sources(${SD_LIB} PRIVATE $<TARGET_OBJECTS:zip>)
 target_link_libraries(${SD_LIB} PUBLIC ggml)
-find_package(Threads REQUIRED)
-target_link_libraries(${SD_LIB} PRIVATE Threads::Threads)
 target_link_libraries(${SD_LIB} PRIVATE onig sd-utf8proc)
-if (SD_CUDA)
-    find_package(CUDAToolkit REQUIRED)
-    # Keep the driver stub on downstream link lines when no driver is installed.
-    target_link_libraries(${SD_LIB} PUBLIC CUDA::cuda_driver)
-    set_property(SOURCE src/core/ggml_extend_backend.cpp APPEND PROPERTY COMPILE_DEFINITIONS SD_USE_CUDA)
-endif()
 target_include_directories(${SD_LIB} PUBLIC . src include)
 target_include_directories(${SD_LIB} PRIVATE src/core)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)

CONTRIBUTING.md

@@ -12,14 +12,8 @@ If you want to update a third-party dependency, please open an issue first inste
 
 ## Pull Requests
 
-When contributing from a fork, use a fork under your personal GitHub account and enable **Allow edits from maintainers**. This lets maintainers make follow-up fixes directly on the PR branch.
-
-PRs from organization-owned forks are automatically closed when opened or reopened because GitHub does not support this maintainer-edit option for those forks. Submit the changes from a personal fork instead. See [GitHub's documentation](https://docs.github.com/en/pull-requests/how-tos/work-with-forks/allowing-changes-to-a-pull-request-branch-created-from-a-fork).
-
 Keep each PR focused on one clear change. Large or overly complex PRs are harder to review and may not be merged.
 
-Do not include test code or test scripts in commits or PRs. Keep them local and report verification results in the PR description.
-
 Follow Conventional Commit-style subjects seen in history: `feat:`, `fix:`, `refactor:`, `ci:`, `docs:`, `chore:`. Keep subjects imperative and scoped.
 
 PRs should include:
@@ -41,11 +35,12 @@ Naming conventions:
 - In `PascalCase` names, preserve common abbreviations in uppercase, for example `SD`, `API`, `HTTP`, `JSON`, `RGB`, `VAE`, `TAE`, `LoRA`, and `WebP`.
 - Use `snake_case` for functions, methods, variables, and file names unless an existing API requires a different style.
 - Use a trailing underscore for private data member names, for example `hidden_size_` or `tokenizer_`.
-- Use `.hpp` for model headers under `src/model/`, including new model headers. Do not rename these headers to `.h`. Use `.h` for other C and C++ header files.
+- Use `.h` for C and C++ header files. Do not introduce new `.hpp` headers.
 - Use macro-based header include guards instead of `#pragma once`.
 - Format header include guards as `__SD_{PATH}__`, where `{PATH}` is the header path in uppercase snake case without the file extension. For example, `src/sample.h` should use `__SD_SAMPLE_H__`.
 - Do not introduce anonymous namespaces in new or modified code; prefer `static` file-local functions/variables or an explicit named namespace when scoping is needed.
 - In `class`/`struct` definitions, place data members before member functions unless an existing type already clearly follows a different pattern.
+- Keep `test_*.cpp` / `test_*.py` naming for tests.
 
 Some older code in the project may not fully follow the current conventions. Please do not submit PRs that only rewrite existing code to match style rules.
 

README.md

@@ -15,7 +15,6 @@ API and command-line option may change frequently.***
 
 ## 🔥Important News
 
-* **2026/09/20** 🚀 stable-diffusion.cpp adds **Day-0 support for Qwen-Image-2.1**
 * **2026/08/20** 🚀 stable-diffusion.cpp now supports **LTX-2.5**
 * **2026/08/04** 🚀 stable-diffusion.cpp adds **Day-1 support for MiniMax-H3**
 * **2026/06/25** 🚀 stable-diffusion.cpp now supports **Krea2**
@@ -48,7 +47,6 @@ API and command-line option may change frequently.***
     - [Chroma](./docs/chroma.md)
     - [Chroma1-Radiance](./docs/chroma_radiance.md)
     - [Qwen Image](./docs/qwen_image.md)
-    - [Qwen Image 2.1](./docs/qwen_image_2.1.md)
     - [PiD](./docs/pid.md)
     - [LongCat Image](./docs/longcat_image.md)
     - [Z-Image](./docs/z_image.md)
@@ -63,15 +61,12 @@ API and command-line option may change frequently.***
     - [SeFi-Image](./docs/sefi_image.md)
     - [HiDream-O1-Image](./docs/hidream_o1_image.md)
     - [Ideogram4](./docs/ideogram4.md)
-    - [LLaDA-Image](./docs/llada_image.md)
-    - [PixArt](./docs/pixart.md)
   - [Image Edit Models](./docs/edit.md)
     - [FLUX.1-Kontext-dev](./docs/kontext.md)
     - [Qwen Image Edit series](./docs/qwen_image_edit.md)
     - [LongCat Image Edit](./docs/longcat_image.md)
     - [Boogu Image Edit](./docs/boogu_image.md)
     - [Mage-Flow-Edit](./docs/mage_flow.md#image-editing)
-    - [LLaDA-Image Edit](./docs/llada_image.md#image-editing)
   - Video Models
     - [Wan2.1/Wan2.2](./docs/wan.md)
     - [MiniMax-H3](./docs/minimax_h3.md)

cmake/ggml.cmake

@@ -1,27 +0,0 @@
-if(NOT TARGET ggml AND NOT TARGET ggml::ggml)
-    if(SD_USE_SYSTEM_GGML)
-        find_package(ggml REQUIRED)
-    else()
-        add_subdirectory("${SD_GGML_SOURCE_DIR}" "${CMAKE_CURRENT_BINARY_DIR}/ggml")
-    endif()
-endif()
-if(NOT TARGET ggml)
-    add_library(ggml ALIAS ggml::ggml)
-endif()
-
-get_target_property(sd_ggml_imported ggml IMPORTED)
-if(sd_ggml_imported)
-    set(sd_ggml_private_include "${SD_GGML_SOURCE_DIR}/src")
-else()
-    get_target_property(sd_ggml_private_include ggml SOURCE_DIR)
-endif()
-if(NOT EXISTS "${sd_ggml_private_include}/ggml-impl.h")
-    message(FATAL_ERROR "Set SD_GGML_SOURCE_DIR to the source tree matching the selected ggml library (ggml-impl.h is required).")
-endif()
-target_include_directories(${SD_LIB} PRIVATE "${sd_ggml_private_include}")
-set_property(TARGET ${SD_LIB} PROPERTY SD_GGML_PRIVATE_INCLUDE_DIR "${sd_ggml_private_include}")
-
-if(SD_USE_UPSTREAM_GGML)
-    target_compile_definitions(${SD_LIB} PUBLIC SD_USE_UPSTREAM_GGML)
-    message(WARNING "Using upstream GGML: INT8 tensorwise/convrot is disabled and FP8 weights are converted to F16 at load time. Some operators may be unsupported and performance may be lower than with patched GGML.")
-endif()

cmake/stable-diffusion-config.cmake.in

@@ -10,10 +10,6 @@ set(SD_BIN_DIR "@PACKAGE_SD_BIN_INSTALL_DIR@")
 
 include(CMakeFindDependencyMacro)
 find_dependency(ggml REQUIRED HINTS "${SD_LIB_DIR}/cmake")
-find_dependency(Threads REQUIRED)
-if(@SD_CUDA@)
-    find_dependency(CUDAToolkit REQUIRED)
-endif()
 
 if(NOT TARGET stable-diffusion)
     find_library(stable-diffusion_LIBRARY stable-diffusion
@@ -26,16 +22,12 @@ if(NOT TARGET stable-diffusion)
     set_target_properties(stable-diffusion
         PROPERTIES
             INTERFACE_INCLUDE_DIRECTORIES "${SD_INCLUDE_DIR}"
-            INTERFACE_LINK_LIBRARIES "ggml::ggml;Threads::Threads"
+            INTERFACE_LINK_LIBRARIES "ggml::ggml"
             IMPORTED_LINK_INTERFACE_LANGUAGES "CXX"
             IMPORTED_LOCATION "${stable-diffusion_LIBRARY}"
             INTERFACE_COMPILE_FEATURES "c_std_11;cxx_std_17"
             POSITION_INDEPENDENT_CODE ON)
 
-    if(@SD_CUDA@)
-        set_property(TARGET stable-diffusion APPEND PROPERTY INTERFACE_LINK_LIBRARIES CUDA::cuda_driver)
-    endif()
-
     if(SD_SHARED_LIB)
         target_compile_definitions(stable-diffusion
             INTERFACE SD_BUILD_SHARED_LIB)

cmake/stable-diffusion.pc.in

@@ -7,5 +7,5 @@ Name: stable-diffusion
 Description: Diffusion model(SD,Flux,Wan,Qwen Image,Z-Image,...) inference in pure C/C++
 Version: @SDCPP_BUILD_VERSION@
 Libs: -L${libdir} -lstable-diffusion
-Libs.private: -lggml -lggml-base @CMAKE_THREAD_LIBS_INIT@
+Libs.private: -lggml -lggml-base
 Cflags: -I${includedir}

docs/backend.md

@@ -156,14 +156,8 @@ the runner's graph-cut capacity checks.
 
 Runtime capacity checks also leave 512 MiB of currently free device memory for
 backend scratch buffers and pipelines, including with explicit backend assignments.
-They cap free-memory reports by the device's total memory minus tracked
-resident allocations. Vulkan reports exceeding total memory are rejected because
-its heap-budget subtraction can underflow. Other backends use the cap instead of
-treating such reports as zero free memory. Failed checks log the reported free and
-total memory alongside tracked weight and runtime allocations.
-With `--mmap`, device-backed mappings count toward these budgets at their full
-mapped-file size, once per device buffer even when multiple parameter blocks
-share it. Mappings retained in the loader cache continue to count.
+They cap stale free-memory reports by the device's total memory minus tracked
+resident allocations and reject reports that exceed the device's total memory.
 
 Components are considered in `diffusion`, `te`, `vae` order so that repeatedly
 used diffusion weights have priority. Each component's weights use the first

docs/build.md

@@ -16,40 +16,6 @@ git submodule init
 git submodule update
 ```
 
-## Selecting a GGML source tree
-
-By default, sd.cpp builds the patched GGML submodule in `ggml/`. To build with
-an upstream GGML checkout instead, enable `SD_USE_UPSTREAM_GGML` and set
-`SD_GGML_SOURCE_DIR`:
-
-```shell
-cmake -S . -B build-upstream -DSD_USE_UPSTREAM_GGML=ON -DSD_GGML_SOURCE_DIR=../ggml-upstream
-cmake --build build-upstream --config Release
-```
-
-The selected source tree supplies both the library and its private headers.
-Backend options such as `-DSD_CUDA=ON` apply to the selected tree as usual.
-
-`SD_USE_UPSTREAM_GGML` defaults to `OFF`, which enables the patched GGML
-extensions. Set it to `ON` when using upstream GGML; it selects the compatibility
-mode and does not download or replace the GGML source tree. Upstream mode keeps
-the original FP8 safetensors handling: FP8 tensors are converted to F16 at load
-time (one byte per element in the file, two in RAM and VRAM). INT8
-tensorwise/convrot is disabled and its model files are rejected with an explicit
-error. FP8 GGUF files, FP8 weight type requests and tensor type rules are also
-rejected; no automatic conversion is performed.
-
-Upstream GGML may lack some operators and performance optimizations provided by
-the patched version. A warning is emitted during CMake configuration and when
-creating an inference context. Ordinary floating-point and shared GGML
-quantization types remain available, subject to backend operator support.
-
-`SD_USE_SYSTEM_GGML=ON` instead links an installed GGML CMake package, located
-with `ggml_DIR` or `CMAKE_PREFIX_PATH`. In that mode, `SD_GGML_SOURCE_DIR` must
-point to the matching source tree for private headers. The installed library
-must use the same ABI settings as sd.cpp, including `GGML_MAX_NAME`.
-Set `SD_USE_UPSTREAM_GGML=ON` as well if the installed package is upstream GGML.
-
 ## WebP and WebM Support in Examples
 
 The example applications (`examples/cli` and `examples/server`) use `libwebp` to support WebP image I/O, and `examples/cli` can also use `libwebm` for `.webm` video output. Both are enabled by default. WebM output currently reuses `libwebp` to encode each frame as VP8 before muxing with `libwebm`.
@@ -90,10 +56,6 @@ cmake --build . --config Release
 
 ## Build with CUDA
 
-Native SageAttention is included when using CUDA with patched GGML
-(`SD_USE_UPSTREAM_GGML=OFF`).
-See [SageAttention](sage_attention.md) for GPU requirements and `--sage-attn` usage.
-
 This provides GPU acceleration using NVIDIA GPU. Make sure to have the CUDA toolkit installed. You can download it from your Linux distro's package manager (e.g. `apt install nvidia-cuda-toolkit`) or from here: [CUDA Toolkit](https://developer.nvidia.com/cuda-downloads). Recommended to have at least 4 GB of VRAM.
 
 ```shell

docs/caching.md

@@ -2,16 +2,6 @@
 
 Caching methods accelerate diffusion inference by reusing intermediate computations when changes between steps are small.
 
-### Conditioning Cache
-
-Conditioning results are cached per model context using an LRU cache. The default
-capacity is **0 (disabled) for `sd-cli`** and **4 entries for `sd-server` and the C
-API**. Set `--conditioning-cache-size N` to change the limit; `0` disables caching.
-For example, `sd-cli -m model.safetensors -p "a cat" --conditioning-cache-size 4`
-enables the cache in the CLI. The C API option is
-`sd_ctx_params_t::conditioning_cache_size`, initialized by `sd_ctx_params_init()`.
-This cache is independent of the diffusion-step `--cache-mode` options below.
-
 ### Cache Modes
 
 | Mode | Target | Description |

docs/edit.md

@@ -17,7 +17,6 @@ Depending on the architecture, different models handle reference images differen
 | [**Boogu Image Edit**](./boogu_image.md) | `z_image_omni` |
 | **Krea2 (Community Edit LoRAs)** | `krea2_ostris_edit` |
 | [**Mage-Flow-Edit**](./mage_flow.md#image-editing) | `mage_flow` |
-| [**LLaDA-Image**](./llada_image.md#image-editing) | `llada_image` |
 | **Anima (Community Edit LoRAs)** | `cosmos_reference` |
 
 Stable-diffusion.spp also supports basic Unet-based editing models like instruct-pix2pix or CosXL-Edit. This document is not about those.
@@ -26,9 +25,6 @@ Stable-diffusion.spp also supports basic Unet-based editing models like instruct
 
 ## Configuring Reference Modes (`--ref-image-args`)
 
-For a one-time input transform before reference presets and model processing,
-including cropping, padding, and resizing algorithms, see [Image preprocessing](./image_preprocessing.md).
-
 Different DiT-based editing models require different configurations to process reference images correctly (e.g., whether to use a Vision Language Model (VLM) encoder or pass VAE-encoded images directly to the DiT).
 
 To simplify this, we provide **Presets**. By default, the system automatically selects the best preset based on the model architecture. However, you can override this using the `--ref-image-args` argument.

docs/esrgan.md

@@ -2,8 +2,6 @@
 
 You can use ESRGAN—such as the model [RealESRGAN_x4plus_anime_6B.pth](https://github.com/xinntao/Real-ESRGAN/releases/download/v0.2.2.4/RealESRGAN_x4plus_anime_6B.pth)—to upscale the generated images and improve their overall resolution and clarity.
 
-RGBA images, including Qwen Image 2.1 output, keep their alpha channel during model upscaling and hires fix. ESRGAN processes the RGB channels; the alpha channel is resized with bilinear interpolation and recombined with the upscaled image.
-
 - Specify the model path using the `--upscale-model PATH` parameter. example:
 
 ```bash

docs/ideogram4.md

@@ -18,9 +18,6 @@ at one byte per element in RAM and VRAM. Backends that cannot multiply FP8
 weights directly cast only the active layer to a temporary BF16 tensor during
 execution; the loader does not expand the entire checkpoint to BF16.
 
-With `SD_USE_UPSTREAM_GGML=ON`, FP8 tensors are converted to F16 at load time
-instead (two bytes per element in RAM and VRAM).
-
 Use `ideogram4_fp8.safetensors` and `ideogram4_uncond_fp8.safetensors` directly
 with `--diffusion-model` and `--uncond-diffusion-model`, respectively.
 

docs/image_preprocessing.md

@@ -1,173 +0,0 @@
-# Image preprocessing
-
-Use `--image-preprocess` to transform each image input once, before generation:
-
-```sh
-sd-cli ... \
-  --image-preprocess "target=init,mode=crop-resize,filter=lanczos,antialias=true" \
-  --image-preprocess "target=mask,filter=nearest-exact" \
-  --image-preprocess "target=ref,index=0,mode=fit-pad,width=768,height=768,filter=bicubic"
-```
-
-CLI and server image loaders decode at the original resolution. The generation
-entry point merges input defaults with user rules and prepares one transformed
-image per input. The original pipeline then consumes those images, including
-its mandatory canvas adaptation, reference resizing, and encoder preprocessing.
-
-```text
-native-resolution image
-  -> input defaults + user overrides
-  -> one input transform
-  -> original generation pipeline and model-specific processing
-```
-
-These rules do not override internal VAE, CLIP/VLM, ControlNet, or pixel-patch preprocessing.
-`--ref-image-args` retains its existing meaning and runs after this input transform.
-
-## Inputs and defaults
-
-| `target` | Input | Default geometry | Indexed? |
-| --- | --- | --- | --- |
-| `init` | img2img image or video first frame | Center crop to the generation aspect ratio, then resize | No |
-| `end` | Video last frame | Center crop, then resize | No |
-| `mask` | Inpainting mask | Inherit init geometry; otherwise center crop, then resize | No |
-| `control` | Control image | Center crop, then resize | No |
-| `ref` | Reference images | Preserve source dimensions | Yes |
-| `ip-adapter` | IP-Adapter image | Preserve source dimensions | No |
-| `id` | PhotoMaker identity images | Preserve source dimensions | Yes |
-| `control-frame` | Control video frames | Center crop, then resize | Yes |
-
-Canvas defaults use the aligned generation dimensions. Reference, IP-Adapter,
-and identity inputs use their original dimensions unless overridden. Default
-resampling is nearest for images and nearest-exact for masks.
-
-These defaults are shared by CLI, server, and C API. Moving geometry out of
-the loaders replaces the previous CLI/server BOX/sRGB resizing, so default
-pixels are not guaranteed to match earlier builds.
-
-Reference video and audio preprocessing are outside these image rules.
-Preprocessing options apply to `img_gen` and `vid_gen`, not standalone upscale
-or ADetailer mode. ADetailer clears the user's rules for its internal crops.
-
-## Rules
-
-Rules are comma-separated `key=value` lists. Repeat the CLI option or separate
-rules with semicolons. Every rule requires a `target` and at least one option.
-Rule syntax and input compatibility are checked when image/video generation
-starts. Unknown keys, invalid values, duplicate keys in a rule, missing images,
-and out-of-range indices cause generation to fail with an error log.
-
-Omit `index` to configure every image of that type; otherwise use a zero-based
-index. CLI directory inputs follow filename order. Indexed rules override
-type-wide rules field by field, regardless of order. At equal specificity,
-the last value for a field wins. `auto` selects the input preset.
-
-| `mode` | Input transform |
-| --- | --- |
-| `auto` | Use the input's default geometry |
-| `none` | Keep source dimensions without resizing, cropping, or padding |
-| `stretch` | Resize to the target dimensions |
-| `crop` | Crop a target-sized rectangle without resizing; fail if the source is too small |
-| `crop-resize` | Crop to the target aspect ratio, then resize |
-| `fit-pad` | Fit the entire image inside the target dimensions, preserving aspect ratio, then pad |
-
-`width` and `height` must be specified together as positive integers. They
-override the input transform's dimensions, not the generation or encoder size.
-For a native-size preset, specifying dimensions without a mode selects stretch.
-`mode=none` with explicit dimensions different from the source is contradictory
-and is rejected.
-
-`anchor=center|top|bottom|left|right` selects crop/padding placement.
-`pad_color=#RRGGBB` or `#RRGGBBAA` selects padding, defaulting to opaque black.
-A grayscale mask uses the first color component.
-
-`filter=auto|nearest|nearest-exact|bilinear|bicubic|lanczos` selects resampling.
-`antialias=auto|true|false` enables antialiasing automatically for filtered
-downscaling; explicit true requires bilinear, bicubic, or Lanczos.
-Filtered RGBA resizing uses premultiplied alpha.
-
-`canny=true|false` enables edge detection for any supported image target,
-defaulting to `false`. It runs once after geometry, before the original
-generation pipeline, including with `mode=none`. Grayscale, grayscale-alpha,
-RGB, and RGBA inputs are supported; alpha is preserved.
-
-Each input has its own Canny setting. Indexed rules can enable or disable it
-for individual references, identity images, or video control frames.
-
-```sh
---image-preprocess "target=init,mode=fit-pad,canny=true"
---image-preprocess "target=ref,index=0,mode=none,canny=true"
---image-preprocess "target=control-frame,index=2,canny=true"
-```
-
-Init and mask sources must have the same dimensions. The mask inherits the
-init crop, resize, and padding coordinates, while retaining its own filter,
-padding value, and Canny setting. Conflicting mask geometry is rejected. An
-omitted mask remains absent until the original pipeline creates its default mask.
-
-## Downstream behavior
-
-`mode=none` only skips the input geometry transform. For example:
-
-```sh
---image-preprocess "target=init,mode=none" \
---image-preprocess "target=ref,mode=none"
-```
-
-The init image is still adapted to the generation canvas by the original
-pipeline. Reference images still follow `--ref-image-args` and model-specific
-resizing. CLIP retains its fixed input dimensions and normalization. HiDream-O1
-retains its original pixel-reference and visual preprocessing.
-
-Existing sharing between consumers is preserved: for example, Wan img2video
-uses the same adapted first frame for VAE conditioning and CLIP. High-resolution
-passes reuse the prepared images and apply their original size adaptation;
-they do not apply the user's crop a second time.
-
-To disable reference resizing before VAE encoding, use
-`--ref-image-args "resize_before_vae=false"` or the server field
-`"ref_image_args": "resize_before_vae=false"`. This is separate from
-`target=ref,mode=none`, which only skips input geometry. Model constraints
-still apply.
-
-## Server requests
-
-Native image/video requests and SDAPI accept `image_preprocess` as a string or
-an array of rule strings:
-
-```json
-{
-  "image_preprocess": [
-    "target=init,mode=fit-pad,filter=bicubic",
-    "target=mask,filter=nearest-exact",
-    "target=ref,index=0,mode=none"
-  ]
-}
-```
-
-OpenAI-compatible requests accept it through
-`<sd_cpp_extra_args>{...}</sd_cpp_extra_args>` in the prompt.
-Request rules replace server-default rules. Generation metadata records the
-user rules; image encodings and channel conventions are unchanged.
-
-## C API
-
-Set `image_preprocess` on the existing image/video generation parameters.
-The `generate_image()` and `generate_video()` signatures are unchanged:
-
-```c
-sd_img_gen_params_t params;
-sd_img_gen_params_init(&params);
-/* Set prompt, original-resolution input images, and generation options. */
-params.image_preprocess.rules = "target=init,mode=crop-resize,filter=lanczos;"
-                                "target=mask,filter=nearest-exact";
-bool ok = generate_image(ctx, &params, &images, &count);
-```
-
-Both generation parameter initializers set `image_preprocess.rules` to `NULL`,
-selecting input presets. Rule strings are borrowed for the synchronous call.
-The library owns temporary transformed pixels; caller images and arrays are
-not modified. Add `canny=true` to the desired target's rule in
-`image_preprocess.rules` to enable Canny.
-
-The parameter structs have grown; applications and bindings must be rebuilt.

docs/int8_convrot.md

@@ -2,9 +2,6 @@
 
 sd.cpp can load and execute ComfyUI `int8_tensorwise` safetensors with `convrot` metadata directly. The stored INT8 weights are not converted to another weight type at load time.
 
-This requires the INT8 tensorwise/convrot extensions in the patched GGML.
-Builds with `SD_USE_UPSTREAM_GGML=ON` reject these files during loading.
-
 ## Checkpoint format
 
 Each quantized linear module contains the following tensors:

docs/llada_image.md

@@ -1,156 +0,0 @@
-# How to Use
-
-LLaDA-Image is a 6B text-to-image and instruction-guided editing model. The denoiser is a
-Lumina2/Z-Image-style NextDiT conditioned by a LLaDA2-MoE diffusion-LLM text encoder, and it
-reuses the Flux.2 VAE. Two checkpoints are published: a 50-step base model and
-LLaDA-Image-Turbo, a 4-step distilled model.
-
-## Download weights
-
-Four components are required: a transformer, a text encoder, a VAE, and a connectors file
-holding the QueryFormer, the text projection and, for editing, the SigVQ image encoder.
-
-The two published checkpoints are **not** interchangeable. LLaDA-Image-Turbo and LLaDA-Image
-ship different transformers, text encoders, QueryFormers and text projections; only the VAE,
-the SigVQ encoder and the tokenizer are shared. Mixing the two produces degraded output rather
-than a clean error, so keep each checkpoint's files together.
-
-Both need an external LLaDA2 `tokenizer.json`, which is not embedded in sd.cpp and is the same
-file for either checkpoint. Take `tokenizer/tokenizer.json` from either repository and pass it
-with `--tokenizer`. See [JSON tokenizers](tokenizers.md) for CLI and C API usage.
-
-### LLaDA-Image-Turbo (4 steps)
-
-Converted transformer, text encoder and pre-merged connectors are at
-https://huggingface.co/fszontagh/LLaDA-Image-Turbo-GGUF:
-
-- `llada-image-turbo-f16.gguf`
-- `llada-image-turbo-text_encoder-q8_0.gguf`
-- `llada-image-turbo-connectors.safetensors` for text to image, or
-  `llada-image-turbo-connectors-edit.safetensors`, which also carries the SigVQ encoder that
-  editing needs.
-
-Other quantizations of the transformer and the text encoder are in the same repository.
-
-The VAE comes from the original repository,
-https://huggingface.co/inclusionAI/LLaDA-Image-Turbo: `vae/diffusion_pytorch_model.safetensors`,
-referred to below as `llada_vae.safetensors`.
-
-### LLaDA-Image (50 steps)
-
-Converted transformer, text encoder and pre-merged connectors are at
-https://huggingface.co/fszontagh/LLaDA-Image-GGUF:
-
-- `llada-image-f16.gguf`
-- `llada-image-text_encoder-q8_0.gguf`
-- `llada-image-connectors.safetensors` for text to image, or
-  `llada-image-connectors-edit.safetensors`, which also carries the SigVQ encoder that editing
-  needs.
-
-Other quantizations of the transformer and the text encoder are in the same repository.
-
-The VAE comes from the original repository,
-https://huggingface.co/inclusionAI/LLaDA-Image, and is the same file as the Turbo one.
-
-### Converting the weights yourself
-
-The transformer has to go in through `--diffusion-model` so that its tensor names keep the
-prefix the loader expects, while the text encoder goes in through `-m`:
-
-```bash
-./bin/sd-cli -M convert --diffusion-model transformer/diffusion_pytorch_model.safetensors.index.json \
-  -o llada-image-f16.gguf --type f16
-./bin/sd-cli -M convert -m text_encoder/model.safetensors.index.json \
-  -o llada-image-text_encoder-q8_0.gguf --type q8_0
-```
-
-### Building the connector file yourself
-
-`--embeddings-connectors` takes one file, so the QueryFormer, the text projection and
-(for editing) the SigVQ encoder have to be combined into a single Safetensors file, each
-tensor name prefixed with its component name. Leaving `sigvq` out skips loading the 2.6 GB
-encoder:
-
-```python
-from safetensors.torch import load_file, save_file
-
-merged = {}
-for prefix, path in [
-    ("queryformer", "queryformer/diffusion_pytorch_model.safetensors"),
-    ("text_projection", "text_projection/diffusion_pytorch_model.safetensors"),
-    ("sigvq", "sigvq/diffusion_pytorch_model.safetensors"),
-]:
-    for name, tensor in load_file(path).items():
-        merged[f"{prefix}.{name}"] = tensor
-save_file(merged, "llada_connectors.safetensors")
-```
-
-## Examples
-
-### Text to image
-
-```bash
-./bin/sd-cli \
-  --diffusion-model /path/to/llada-image-turbo-f16.gguf \
-  --llm /path/to/llada-image-turbo-text_encoder-q8_0.gguf \
-  --tokenizer /path/to/tokenizer.json \
-  --vae /path/to/llada_vae.safetensors \
-  --embeddings-connectors /path/to/llada-image-turbo-connectors.safetensors \
-  --prompt "a lovely cat holding a sign says 'llada.cpp'" \
-  --width 1024 \
-  --height 1024 \
-  --steps 4 \
-  --cfg-scale 1.0 \
-  --seed 42 \
-  --output output.png
-```
-
-<img width="256" alt="LLaDA-Image example" src="../assets/llada_image/example.png" />
-
-### Image editing
-
-```bash
-./bin/sd-cli \
-  --diffusion-model /path/to/llada-image-turbo-f16.gguf \
-  --llm /path/to/llada-image-turbo-text_encoder-q8_0.gguf \
-  --tokenizer /path/to/tokenizer.json \
-  --vae /path/to/llada_vae.safetensors \
-  --embeddings-connectors /path/to/llada-image-turbo-connectors-edit.safetensors \
-  --ref-image /path/to/input.png \
-  --prompt "change the sign text to 'sd.cpp'" \
-  --width 1024 \
-  --height 1024 \
-  --steps 4 \
-  --cfg-scale 1.0 \
-  --diffusion-fa \
-  --output output.png
-```
-
-<img width="256" alt="LLaDA-Image edit example" src="../assets/llada_image/edit_example.png" />
-
-See [edit.md](./edit.md) for the shared reference-image options. LLaDA-Image uses the
-`llada_image` preset by default, resizing the reference image to the output width and height
-before VAE encoding. SigVQ uses bilinear resizing to half the output resolution and inputs
-normalized to `[-1, 1]`. CFG keeps the source latent in both branches and uses SigVQ features
-only in the positive branch. Editing requires connectors that include the SigVQ weights.
-
-## Notes
-
-- Use 4 steps and `--cfg-scale 1.0` for LLaDA-Image-Turbo; the guidance is distilled away, so
-  a higher CFG degrades output and doubles the text encoder cost. The 50-step base model uses
-  `--steps 50 --cfg-scale 5`.
-- Width and height are rounded up to a multiple of 16. For editing the reference pipeline
-  requires them to be divisible by 32.
-- Edit the 50-step base model at 1024x1024. At 512x512 it returns the reference image almost
-  unchanged instead of applying the instruction; LLaDA-Image-Turbo edits correctly at both.
-- Editing runs the reference and the target in one sequence, so it needs roughly twice the
-  tokens of text to image at the same size. On 12 GB, editing at 1024x1024 needs
-  `--diffusion-fa`; without it the diffusion graph does not fit.
-- The weights total about 16 GB, but segmented execution streams them, so a much smaller
-  budget works. At 512x512, `--max-vram 6` costs almost nothing over unconstrained execution,
-  and `--max-vram 3` still produces byte-identical output at roughly 2.5x the time.
-- `--scheduler` defaults to `llada_image`, which reproduces the reference Kumaraswamy sigma
-  grid. `--extra-sample-args uniform=1` selects the uniform grid instead.
-- Prompt templating is handled automatically; pass a plain description.
-- VQ-conditioned generation (`generation_mode="vq"`, where the text encoder decodes image
-  tokens before diffusion) is not implemented.

docs/performance.md

@@ -21,54 +21,6 @@ CPU fallback. It excludes weights and cache buffers. Within a runner lifecycle,
 the summary is printed only on the first graph or when backend capacities or the
 segment count change.
 
-## Use VAE tiling to reduce encode and decode memory usage.
-
-`--vae-tiling` enables spatial tiling for both VAE encoding and decoding. The
-default tile size is 256x256 **image pixels**, independent of the VAE scale factor:
-
-```shell
---vae-tiling --vae-tile-size 256x256 --vae-tile-overlap 0.5
-```
-
-`--vae-tile-size` accepts one size or `WIDTHxHEIGHT`. A zero dimension uses the
-256-pixel default. Sizes are rounded down to a multiple of the VAE scale factor
-and capped at the current input dimensions. Explicit sizes below four latent
-pixels per axis (or the full axis when it is smaller) are rejected. Encoding and
-decoding use the same spatial sizes, without an additional encoding multiplier.
-Inputs that fit within a tile are processed as one tile.
-
-For a 512x512 image with the default 50% overlap, both encoding and decoding use
-3x3 tiles. A 256-pixel tile corresponds to 32 latent pixels for an 8x VAE, 16 for
-a 16x VAE, and 8 for a 32x VAE. Smaller tiles reduce each graph's memory demand,
-but overlapping work can increase processing time and tiling can affect image
-quality, especially during encoding. Use larger tiles when more context is needed.
-
-`--vae-relative-tile-size` overrides the absolute size on each axis with a positive
-value. Values up to and including 1 specify a fraction of the current input size;
-values greater than 1 specify a target number of tiles per axis, accounting for
-overlap. For example, `0.5x0.5` uses half the width and height in both encode and
-decode. The target overlap is clamped to 0 through 0.5 and the actual overlap is
-adjusted to fit the image. Size and overlap options require `--vae-tiling`.
-
-**Migration:** `--vae-tile-size` and the C/JSON fields `tile_size_w` and
-`tile_size_h` now use image pixels instead of latent units. The C/JSON fields
-`tile_size_x/y` have been renamed to `tile_size_w/h`, and `rel_size_x/y` to
-`rel_size_w/h`. The command-line option names are unchanged. For example, an old
-decode tile size of 32 corresponds to 256 pixels for an 8x VAE or 512 pixels for a
-16x VAE. Encoding no longer enlarges explicit or relative tile sizes.
-
-The main VAE decode path retries allocation failures with smaller tiles, even
-without `--vae-tiling`. Supported video VAEs first try temporal tiling; spatial
-retries use at most 256-pixel tiles initially and then halve the effective tile
-dimensions down to the minimum size. Each spatial retry must reduce the effective
-tile size. These runtime adjustments do not change the caller's parameters.
-Execution failures are not retried, and encoding has no automatic OOM retry.
-
-`--temporal-tiling` remains independent of spatial tiling. MiniMax H3 always uses
-spatial tiling (256x256 pixels and 25% overlap by default) and its own temporal
-windows. With `--vae-tiling`, its overlap follows `--vae-tile-overlap`; explicit
-spatial sizes are honored.
-
 ## Offload weights to the CPU to save VRAM without reducing generation speed.
 
 Using `--offload-to-cpu` allows you to offload weights to the CPU, saving VRAM without reducing generation speed.
@@ -111,12 +63,6 @@ See [backend selection](./backend.md) for full syntax.
 
 When a graph has cut markers and its missing weights plus incremental compute workspace exceed the available device headroom, it runs its fixed segment list in order. A reusable monolithic compute buffer is not counted as a new allocation. An explicit `--max-vram` budget deducts already-resident managed weights and compute/cache buffers registered by every runner sharing the device, so later graph runs remain segmented when the full graph exceeds the budget. The current segment's weights are pinned during compute, and the next parameter-bearing segment is prefetched when the device supports asynchronous transfer. No opt-in streaming flag is required.
 
-When choosing between monolithic and segmented execution, the runner requires
-an additional 128 MiB of headroom in both available device memory and any explicit
-managed budget. This planning headroom absorbs small allocation estimate changes;
-subsequent capacity checks can consume it while still preserving the 512 MiB device
-scratch reserve and respecting the managed budget.
-
 - `--max-vram <GiB>` optionally lowers the live-memory limit. A positive value is a managed per-device budget, `0` uses the device's current free memory without an explicit budget, and a negative value snapshots free memory at startup while reserving that many GiB (`--max-vram -1` reserves about 1 GiB). Driver contexts and unrelated external allocations remain outside the managed budget.
 - `--disable-prefetch` disables asynchronous next-segment prefetch while retaining synchronous loading, eviction, and segmented execution.
 - `--disable-segmented-compute` forces monolithic graph execution for diagnostics or compatibility, even when the automatic memory check would select segments.

docs/pixart.md

@@ -1,46 +0,0 @@
-# How to Use
-
-You can run PixArt-α / PixArt-Σ with stable-diffusion.cpp.
-
-PixArt is a DiT-based text-to-image model family conditioned by a T5-XXL text
-encoder and a 4-channel VAE: SDXL-style for PixArt-Σ and SD1.x-style for PixArt-α.
-
-## Download weights
-
-- Download the transformer (diffusion model)
-    - PixArt-Σ XL-2 1024-MS: https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/tree/main/transformer
-    - PixArt-α XL-2 1024-MS: https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS/tree/main/transformer
-- Download the T5-XXL text encoder
-    - safetensors: https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/tree/main/text_encoder
-- Download the VAE
-    - PixArt-Σ: https://huggingface.co/PixArt-alpha/PixArt-Sigma-XL-2-1024-MS/tree/main/vae
-    - PixArt-α: https://huggingface.co/PixArt-alpha/PixArt-XL-2-1024-MS/tree/main/vae
-    - Use the VAE matching the checkpoint's latent space. For TAE decoding or
-      preview, use TAESDXL for PixArt-Σ and TAESD for PixArt-α.
-- Tokenizer: the T5 vocabulary is embedded; no extra tokenizer file is needed.
-
-## Examples
-
-```
-.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\pixart_sigma_xl2_1024_ms.safetensors --t5xxl ..\models\text_encoders\t5xxl.safetensors --vae ..\models\vae\pixart_vae.safetensors -p "a lovely cat" --cfg-scale 4.5 -W 1024 -H 1024 --steps 20 -v
-```
-
-## Notes
-
-- The VAE scaling factor defaults to `0.13025` for PixArt-Σ. PixArt-α
-  checkpoints with resolution micro-condition weights use `0.18215`.
-  PixArt-α 512 has the same tensor layout as PixArt-Σ, so it requires an
-  explicit override: `--model-args "pixart_vae_scale_factor=0.18215"`.
-  This argument can also override the scale for other compatible checkpoints.
-- PixArt-Σ checkpoints compute 2D sincos positional embeddings at runtime;
-  the trained grid is 64x64 patches with an interpolation scale of 2.
-  For checkpoints trained at a different resolution, the positional embedding
-  parameters can be adjusted via model args:
-  `--model-args "pixart_pos_embed_base_size=<trained grid>,pixart_interpolation_scale=<scale>"`
-  (e.g. `pixart_pos_embed_base_size=32,pixart_interpolation_scale=1,pixart_vae_scale_factor=0.18215` for
-  PixArt-α XL-2 512).
-- Checkpoints carrying resolution/aspect-ratio micro-condition weights are
-  detected but those conditions are not applied yet; a warning is logged and
-  generation proceeds with the timestep embedding only.
-- The transformer predicts 8 channels (noise + learned variance); only the
-  noise half is used for sampling, matching the reference implementation.

docs/qwen_image_2.1.md

@@ -1,87 +0,0 @@
-# How to Use
-
-Qwen Image 2.1 supports text-to-image generation and image editing, using Qwen3-VL-8B as the text encoder and its own VAE.
-
-## Download weights
-
-- Download Qwen Image 2.1
-    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/diffusion_models
-    - gguf: https://huggingface.co/leejet/Qwen-Image-2.1-GGUF/tree/main
-- Download vae
-    - safetensors: https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/vae
-- Download Qwen3-VL-8B-Instruct
-    - safetensors (BF16 or INT8 convrot): https://huggingface.co/Comfy-Org/Qwen-Image-2.1/tree/main/text_encoders
-    - gguf: https://huggingface.co/Qwen/Qwen3-VL-8B-Instruct-GGUF/tree/main
-    - For image editing with a GGUF text encoder, also download `mmproj-Qwen3VL-8B-Instruct-F16.gguf` from the same repository and pass it with `--llm_vision`.
-
-Use `qwen_image_2.1_vae_bf16.safetensors` with this model. The earlier Qwen Image and Wan 2.2 VAE weights are not interchangeable with the Qwen Image 2.1 VAE weights.
-
-## Examples
-
-Run the following commands from the build directory. Use image dimensions divisible by 32. The resolution-dependent flow schedule is selected automatically.
-
-### Text to image
-
-```powershell
-.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\qwen_image_2.1_int8_convrot.safetensors --vae ..\models\vae\qwen_image_2.1_vae_bf16.safetensors --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf -p "a lovely cat holding a sign says 'qwen2.1.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --fa -o qwen_image_2.1.png
-```
-
-<img alt="Qwen Image 2.1 example" src="../assets/qwen/qwen_image_2.1.png" />
-
-To use GGUF diffusion weights, set `--diffusion-model` to the path of a file such as `qwen_image_2.1-Q4_K.gguf`.
-
-### Image editing
-
-Pass the reference image with `-r` and describe the edit in `-p`. Vision weights are required; the example below loads them separately with `--llm_vision`.
-
-```powershell
-.\bin\Release\sd-cli.exe --diffusion-model ..\models\diffusion_models\qwen_image_2.1_int8_convrot.safetensors --vae ..\models\vae\qwen_image_2.1_vae_bf16.safetensors --llm ..\models\text_encoders\Qwen3VL-8B-Instruct-Q4_K_M.gguf --llm_vision ..\models\text_encoders\Qwen3VL-8B-Instruct-mmproj-BF16.gguf -r ..\assets\qwen\qwen_image_2.1.png -p "change 'qwen2.1.cpp' to 'sd.cpp'" --cfg-scale 6.0 --sampling-method euler -v --offload-to-cpu --fa -o qwen_image_2.1_edit.png
-```
-
-For multiple reference images, repeat `-r` in the desired order, for example `-r first.png -r second.png`.
-
-### Prefix cache
-
-By default, the first denoising call for each fixed condition saves the text and reference-image keys and values from every transformer layer. Later calls only compute the target-image tokens. Positive and negative conditions use separate caches, which are released when sampling ends.
-
-Set `qwen_image_2_1_prefix_cache_type` in `--model-args` to `auto` or a type name using the same parser and case-sensitive names as `--type`:
-
-- `auto` (default): use FP16 only when Flash Attention is enabled, Sage Attention is disabled, the attention scale is unchanged, and every attention operation in the cache-writing or cache-reading graph selects Flash Attention after backend support checks. If an operation falls back, rebuild the prefix in FP32 before executing and keep FP32 for the rest of that sampling run.
-- `f32`: always store FP32 keys and values.
-- `f16`: always store FP16 keys and values, including with ordinary attention or custom attention scaling. This saves cache memory but can introduce additional rounding error.
-- Other types, such as `bf16`, `q4_1`, `q5_0`, `q5_1`, `q8_0`, `q4_K`, `q6_K`, `iq4_nl`, and `iq4_xs`: use the requested storage type if the ggml build provides runtime conversion to and from FP32. Quantization is lossy and must be selected explicitly; `auto` never selects a quantized type.
-
-Cache data is packed into contiguous rows of `hidden_size` elements before conversion, so 256-element quantization blocks work with the model's 128-element attention heads without padding. The type's block size must divide `hidden_size`. Unknown types, types lacking runtime conversion (for example `q8_1` and several IQ formats), and incompatible block sizes are ignored with a warning, leaving the previous setting or the default `auto` unchanged.
-
-For example, use `--model-args qwen_image_2_1_prefix_cache_type=q8_0` to enable 8-bit cache storage. Cached keys and values are converted back to the attention input type before concatenating with the current target tokens. This reduces persistent cache memory; attention working buffers still use floating-point values, and conversion adds work on each step. Backends without the required conversion operations use the existing CPU fallback.
-
-For the default 32-layer model, a prefix of 4096 tokens takes approximately the following memory per condition, excluding weights, working buffers, and allocation overhead:
-
-| Cache type | Memory |
-| --- | ---: |
-| `f32` | 4 GiB |
-| `f16` | 2 GiB |
-| `q8_0` | 1.0625 GiB |
-| `q4_0` | 0.5625 GiB |
-
-The runner accounts for the cache when checking the memory budget. If a cached execution runs out of memory, it releases the prefix caches, disables caching for the rest of that sampling run, and retries the full sequence once. Per-step conditioning extensions currently use the full-sequence path.
-
-Disable this optimization with `--model-args qwen_image_2_1_prefix_cache=false`. It reuses step-independent activations; numerical results can still differ slightly because the matrix sizes change.
-
-### Alpha channel
-
-This model supports alpha channel output. As the model determines whether to output a regular image or with transparency through the prompt, according to [official recommendation](https://github.com/QwenLM/Qwen-Image-2.1#transparent-image-generation-rgba), use the following prompt format for better results:
-
-> `This is an RGBA image with transparency. <your description>. The image has alpha channel and the background is transparent.`
-
-Since transparency is decided by the prompt rather than by the input or an explicit switch, the same format applies equally to editing, whether or not the reference image itself has an alpha channel. Note that alpha is kept only in `.png` and `.webp` outputs; saving as `.jpg` drops the transparency.
-
-Here are some examples ran with Q6_K quantization:
-| Input | Prompt | Output |
-| --- | --- | --- |
-| ![Qwen Image 2.1 alpha input example 1](../assets/qwen/qwen-image-2.1-alpha-in1.png) | This is an RGBA image with transparency. Replace the text "BLOOM" with "Qwen Image 2.1", keeping the same font of the original text. The image has alpha channel and the background is transparent. | ![Qwen Image 2.1 alpha output example 1](../assets/qwen/qwen-image-2.1-alpha-out1.png) |
-| ![Qwen Image 2.1 alpha input example 2](../assets/logo.png) | This is an RGBA image with transparency. Remove the background of the image, keeping only the text and cat. The image has alpha channel and the background is transparent. | ![Qwen Image 2.1 alpha output example 2](../assets/qwen/qwen-image-2.1-alpha-out2.png) |
-
-### Other features
-
-Other features of the model could be found on the [model card from QwenLM/Qwen-Image-2.1 repo](https://github.com/QwenLM/Qwen-Image-2.1), including 2 finetuned prompt rewriting Qwen3.5-9B model.

docs/sage_attention.md

@@ -1,68 +0,0 @@
-# SageAttention
-
-`--sage-attn` enables native CUDA SageAttention in the diffusion model, including
-the high-noise diffusion model when present. Python, PyTorch, and Triton are not
-required at build time or runtime.
-
-The CUDA backend automatically selects a kernel supported by both the GPU and
-the compiled CUDA toolkit:
-
-| GPU / toolkit | Implementation |
-| --- | --- |
-| SM89 or newer, CUDA 12.8 or newer (except SM90) | SageAttention2++: per-thread INT8 Q/K, FP8 PV, FP16 instruction accumulation with an FP32 buffer |
-| SM89 or newer, CUDA 12.4 or newer; SM90 also uses this path with newer toolkits | SageAttention2: per-thread INT8 Q/K, FP8 PV, two-level FP32 accumulation |
-| SM80 or newer, CUDA 12.0 or newer | INT8 Q/K, FP16 PV compatibility path |
-
-The FP8 paths smooth K, quantize V per channel, and pad and permute V for FP8
-Tensor Cores. The 2++ path uses the upstream V scale limit of 2.25 to avoid
-overflow in its FP16 instruction accumulator. The public output remains FP32.
-These are the upstream **INT8** SageAttention2/2++ variants; the paper's INT4
-variant and Hopper-specific WGMMA kernel are not implemented here.
-
-## Build
-
-Use the bundled patched GGML, CUDA Toolkit 12.0 or newer, and an NVIDIA GPU with
-compute capability 8.0 or newer. Compile kernels for the GPU being used.
-
-```sh
-cmake -S . -B build -DSD_CUDA=ON -DSD_USE_UPSTREAM_GGML=OFF
-cmake --build build --config Release
-```
-
-No separate SageAttention build option is needed. Upstream GGML builds do not
-support it. A system GGML must include the matching patched API and CUDA
-backend. Enabling `--sage-attn` with an unavailable build or diffusion device
-reports an error. Building with CUDA 12.4 selects SageAttention2 on an RTX 4090;
-rebuild with CUDA 12.8 or newer to use SageAttention2++.
-
-## Use
-
-Replace `--diffusion-fa` with `--sage-attn` in an existing command. For example,
-from the build directory:
-
-```powershell
-.\bin\Release\sd-cli.exe -M vid_gen --diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-LowNoise-Q8_0.gguf --high-noise-diffusion-model  ..\models\diffusion_models\Wan2.2-T2V-A14B-HighNoise-Q8_0.gguf --vae ..\models\vae\wan_2.1_vae.safetensors --t5xxl ..\models\text_encoders\umt5-xxl-encoder-Q8_0.gguf  -p "a lovely cat" --cfg-scale 3.5 --sampling-method euler --steps 10 --high-noise-cfg-scale 3.5 --high-noise-sampling-method euler --high-noise-steps 8 -v -n "色调艳丽，过曝，静态，细节模糊不清，字幕，风格，作品，画作，画面，静止，整体发灰，最差质量，低质量，JPEG压缩残留，丑陋的，残缺的，多余的手指，画得不好的手部，画得不好的脸部，畸形的，毁容的，
-形态畸形的肢体，手指融合，静止不动的画面，杂乱的背景，三条腿，背景人很多，倒着走" -W 832 -H 480 --diffusion-fa --offload-to-cpu --video-frames 33 --sage-attn
-```
-
-SageAttention currently handles unmasked attention with head dimensions from
-1 through 128, including grouped-query attention, different query/key lengths,
-and multiple batches. Dimensions below 64 are zero-padded to 64; dimensions
-between 65 and 127 are zero-padded to 128. The original softmax scale is preserved,
-and the output is cropped back to the original dimension. Other attention
-operations fall back to FlashAttention when supported, then ordinary attention.
-SageAttention takes precedence in diffusion
-when combined with `--fa` or `--diffusion-fa`; `--fa` continues to control other
-modules. Existing attention scaling overrides remain effective.
-
-Attention quantization changes numerical results. Compare image quality and
-end-to-end generation time using the same seed, dimensions, and sampling
-settings. Compare sampling steps after the first step for warmed-up inference
-speed, and report model loading and first-step initialization separately.
-Quantization, smoothing, and format conversion costs are included in generation
-time, so short sequences may not benefit.
-
-Library callers set `sd_ctx_params_t.sage_attn = true` before `new_sd_ctx()`,
-like `diffusion_flash_attn`. Context creation fails if the requested feature is
-unavailable. Initialize the parameter structure with `sd_ctx_params_init()`.
-Rebuild library callers against the updated public header.

docs/troubleshooting.md

@@ -1,14 +1,5 @@
 # Troubleshooting
 
-## Video model used in image generation mode
-
-If generation reports that a model cannot be run with `generate_image()`, add
-`--mode vid_gen` to the CLI command. `--video-frames` alone does not select video
-mode. Video models require this mode even when generating a single frame.
-Library callers must use `generate_video()` for these models; use
-`sd_ctx_supports_image_generation()` and `sd_ctx_supports_video_generation()` to
-check the available generation modes.
-
 ## Completely black or white images or videos / NaNs
 
 Some ggml backends can encounter numerical overflow during inference, producing

docs/wan.md

@@ -1,7 +1,5 @@
 # How to Use
 
-Wan models require `-M vid_gen`, including single-frame generation. `--video-frames` alone does not select video mode. Library callers must use `generate_video()` instead of `generate_image()`.
-
 ## Download weights
 
 - Download Wan

examples/cli/README.md

@@ -14,12 +14,6 @@ equivalent to `--log-level verbose`. If repeated, the last logging option wins.
 For direct image repair or automatic post-generation YOLOv8 detection followed by cropped inpainting, see
 [ADetailer](../../docs/adetailer.md).
 
-Use repeatable `--image-preprocess` rules to select resizing, cropping, padding,
-and resampling separately for each image input. Add `canny=true` to any input
-rule for edge detection. See
-[Image preprocessing](../../docs/image_preprocessing.md) for input selectors,
-input defaults, downstream model processing, mask alignment, and examples.
-
 Metadata mode inspects PNG/JPEG container metadata without loading any model:
 
 ```bash

examples/cli/main.cpp

@@ -41,6 +41,7 @@ struct SDCliParams {
     std::string metadata_format = "text";
 
     sd_log_level_t log_level = SD_LOG_INFO;
+    bool canny_preprocess    = false;
     bool convert_name        = false;
 
     preview_t preview_method = PREVIEW_NONE;
@@ -106,6 +107,10 @@ struct SDCliParams {
         };
 
         options.bool_options = {
+            {"",
+             "--canny",
+             "apply canny preprocessor (edge detection)",
+             true, &canny_preprocess},
             {"",
              "--convert-name",
              "convert tensor name (for convert mode)",
@@ -263,6 +268,7 @@ struct SDCliParams {
             << "  metadata_format: \"" << metadata_format << "\",\n"
             << "  log_level: " << log_level_name(log_level) << ",\n"
             << "  color: " << (color ? "true" : "false") << ",\n"
+            << "  canny_preprocess: " << (canny_preprocess ? "true" : "false") << ",\n"
             << "  convert_name: " << (convert_name ? "true" : "false") << ",\n"
             << "  preview_method: " << previews_str[preview_method] << ",\n"
             << "  preview_interval: " << preview_interval << ",\n"
@@ -322,7 +328,9 @@ void sd_log_cb(enum sd_log_level_t level, const char* log, void* data) {
 
 bool load_images_from_dir(const std::string dir,
                           std::vector<SDImageOwner>& images,
-                          int max_image_num = 0) {
+                          int expected_width  = 0,
+                          int expected_height = 0,
+                          int max_image_num   = 0) {
     if (!fs::exists(dir) || !fs::is_directory(dir)) {
         LOG_ERROR("'%s' is not a valid directory\n", dir.c_str());
         return false;
@@ -349,8 +357,7 @@ bool load_images_from_dir(const std::string dir,
             LOG_VERBOSE("load image %zu from '%s'", images.size(), path.c_str());
             int width             = 0;
             int height            = 0;
-            int loaded_channel    = 0;
-            uint8_t* image_buffer = load_image_from_file(path.c_str(), width, height, loaded_channel, 0, 0);
+            uint8_t* image_buffer = load_image_from_file(path.c_str(), width, height, expected_width, expected_height);
             if (image_buffer == nullptr) {
                 LOG_ERROR("load image from '%s' failed", path.c_str());
                 return false;
@@ -358,7 +365,7 @@ bool load_images_from_dir(const std::string dir,
 
             images.emplace_back(sd_image_t{(uint32_t)width,
                                            (uint32_t)height,
-                                           (uint32_t)loaded_channel,
+                                           3,
                                            image_buffer});
 
             if (max_image_num > 0 && static_cast<int>(images.size()) >= max_image_num) {
@@ -644,11 +651,10 @@ int main(int argc, const char* argv[]) {
 
     SDCliParams cli_params;
     SDContextParams ctx_params;
-    ctx_params.conditioning_cache_size = 0;
     SDGenerationParams gen_params;
 
-    sd_set_log_callback(sd_log_cb, (void*)&cli_params);
     parse_args(argc, argv, cli_params, ctx_params, gen_params);
+    sd_set_log_callback(sd_log_cb, (void*)&cli_params);
 
     if (cli_params.mode == METADATA) {
         MetadataReadOptions options;
@@ -744,8 +750,16 @@ int main(int argc, const char* argv[]) {
 
     auto load_image_and_update_size = [&](const std::string& path,
                                           SDImageOwner& image,
+                                          bool resize_image    = true,
                                           int expected_channel = 3) -> bool {
-        if (!load_sd_image_from_file(image.put(), path.c_str(), 0, 0, expected_channel)) {
+        int expected_width  = 0;
+        int expected_height = 0;
+        if (resize_image && gen_params.width_and_height_are_set()) {
+            expected_width  = gen_params.width;
+            expected_height = gen_params.height;
+        }
+
+        if (!load_sd_image_from_file(image.put(), path.c_str(), expected_width, expected_height, expected_channel)) {
             LOG_ERROR("load image from '%s' failed", path.c_str());
             return false;
         }
@@ -767,8 +781,7 @@ int main(int argc, const char* argv[]) {
     };
 
     if (gen_params.init_image_path.size() > 0) {
-        const bool native_init = cli_params.mode == IMG_GEN || cli_params.mode == ADETAILER;
-        if (!load_image_and_update_size(gen_params.init_image_path, gen_params.init_image, native_init ? 0 : 3)) {
+        if (!load_image_and_update_size(gen_params.init_image_path, gen_params.init_image)) {
             return 1;
         }
     }
@@ -782,8 +795,8 @@ int main(int argc, const char* argv[]) {
     if (gen_params.ref_image_paths.size() > 0) {
         gen_params.ref_images.clear();
         for (auto& path : gen_params.ref_image_paths) {
-            SDImageOwner ref_image({0, 0, 0, nullptr});
-            if (!load_image_and_update_size(path, ref_image, 0)) {
+            SDImageOwner ref_image({0, 0, 3, nullptr});
+            if (!load_image_and_update_size(path, ref_image, false)) {
                 return 1;
             }
             gen_params.ref_images.push_back(std::move(ref_image));
@@ -824,22 +837,41 @@ int main(int argc, const char* argv[]) {
     if (gen_params.mask_image_path.size() > 0) {
         if (!load_sd_image_from_file(gen_params.mask_image.put(),
                                      gen_params.mask_image_path.c_str(),
-                                     0,
-                                     0,
+                                     gen_params.get_resolved_width(),
+                                     gen_params.get_resolved_height(),
                                      1)) {
             LOG_ERROR("load image from '%s' failed", gen_params.mask_image_path.c_str());
             return 1;
         }
+    } else {
+        sd_image_t generated_mask = {0, 0, 1, nullptr};
+        generated_mask.data       = (uint8_t*)malloc(gen_params.get_resolved_width() * gen_params.get_resolved_height());
+        if (generated_mask.data == nullptr) {
+            LOG_ERROR("malloc mask image failed");
+            return 1;
+        }
+        generated_mask.width  = gen_params.get_resolved_width();
+        generated_mask.height = gen_params.get_resolved_height();
+        memset(generated_mask.data, 255, gen_params.get_resolved_width() * gen_params.get_resolved_height());
+        gen_params.mask_image.reset(generated_mask);
     }
 
     if (gen_params.control_image_path.size() > 0) {
         if (!load_sd_image_from_file(gen_params.control_image.put(),
                                      gen_params.control_image_path.c_str(),
-                                     0,
-                                     0)) {
+                                     gen_params.get_resolved_width(),
+                                     gen_params.get_resolved_height())) {
             LOG_ERROR("load image from '%s' failed", gen_params.control_image_path.c_str());
             return 1;
         }
+        if (cli_params.canny_preprocess) {  // apply preprocessor
+            preprocess_canny(gen_params.control_image.get(),
+                             0.08f,
+                             0.08f,
+                             0.8f,
+                             1.0f,
+                             false);
+        }
     }
 
     if (gen_params.ip_adapter_image_path.size() > 0) {
@@ -856,6 +888,8 @@ int main(int argc, const char* argv[]) {
         gen_params.control_frames.clear();
         if (!load_images_from_dir(gen_params.control_video_path,
                                   gen_params.control_frames,
+                                  gen_params.get_resolved_width(),
+                                  gen_params.get_resolved_height(),
                                   gen_params.video_frames)) {
             return 1;
         }
@@ -864,7 +898,10 @@ int main(int argc, const char* argv[]) {
     if (!gen_params.pm_id_images_dir.empty()) {
         gen_params.pm_id_images.clear();
         if (!load_images_from_dir(gen_params.pm_id_images_dir,
-                                  gen_params.pm_id_images)) {
+                                  gen_params.pm_id_images,
+                                  0,
+                                  0,
+                                  0)) {
             return 1;
         }
     }

examples/common/common.cpp

@@ -518,8 +518,7 @@ ArgOptions SDContextParams::get_options() {
         {"",
          "--model-args",
          "extra model args, key=value list. Supports chroma_use_dit_mask, chroma_use_t5_mask, "
-         "chroma_t5_mask_pad, qwen_image_zero_cond_t, qwen_image_2_1_prefix_cache, "
-         "qwen_image_2_1_prefix_cache_type (auto or a type name from --type)",
+         "chroma_t5_mask_pad, qwen_image_zero_cond_t",
          (int)',',
          &model_args},
         {"",
@@ -572,10 +571,6 @@ ArgOptions SDContextParams::get_options() {
          "number of threads to use during computation (default: -1). "
          "If threads <= 0, then threads will be set to the number of CPU physical cores",
          &n_threads},
-        {"",
-         "--conditioning-cache-size",
-         "maximum number of conditioning results cached per model context (default: " + std::to_string(conditioning_cache_size) + ", 0 disables caching)",
-         &conditioning_cache_size},
     };
 
     options.bool_options = {
@@ -623,17 +618,13 @@ ArgOptions SDContextParams::get_options() {
          "--diffusion-fa",
          "use flash attention in the diffusion model only",
          true, &diffusion_flash_attn},
-        {"",
-         "--sage-attn",
-         "use native CUDA SageAttention in the diffusion model, with flash/default attention fallback",
-         true, &sage_attn},
         {"",
          "--diffusion-conv-direct",
          "use ggml_conv2d_direct in the diffusion model",
          true, &diffusion_conv_direct},
         {"",
          "--vae-conv-direct",
-         "use direct 2D and 3D convolutions in the vae model",
+         "use ggml_conv2d_direct in the vae model",
          true, &vae_conv_direct},
     };
 
@@ -827,10 +818,6 @@ bool SDContextParams::resolve(SDMode mode) {
 }
 
 bool SDContextParams::validate(SDMode mode) {
-    if (conditioning_cache_size < 0) {
-        LOG_ERROR("error: conditioning-cache-size must be non-negative");
-        return false;
-    }
     if (mode == CONVERT) {
         const bool has_convert_input = model_path.length() != 0 ||
                                        clip_l_path.length() != 0 ||
@@ -907,7 +894,6 @@ std::string SDContextParams::to_string() const {
     std::ostringstream oss;
     oss << "SDContextParams {\n"
         << "  n_threads: " << n_threads << ",\n"
-        << "  conditioning_cache_size: " << conditioning_cache_size << ",\n"
         << "  model_path: \"" << model_path << "\",\n"
         << "  clip_l_path: \"" << clip_l_path << "\",\n"
         << "  clip_g_path: \"" << clip_g_path << "\",\n"
@@ -952,7 +938,6 @@ std::string SDContextParams::to_string() const {
         << "  vae_on_cpu: " << (vae_on_cpu ? "true" : "false") << ",\n"
         << "  flash_attn: " << (flash_attn ? "true" : "false") << ",\n"
         << "  diffusion_flash_attn: " << (diffusion_flash_attn ? "true" : "false") << ",\n"
-        << "  sage_attn: " << (sage_attn ? "true" : "false") << ",\n"
         << "  linear_scale: " << linear_scale << ",\n"
         << "  attn_scale: " << attn_scale << ",\n"
         << "  diffusion_conv_direct: " << (diffusion_conv_direct ? "true" : "false") << ",\n"
@@ -1002,7 +987,6 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.pulid_weights_path              = pulid_weights_path.c_str();
     sd_ctx_params.tensor_type_rules               = tensor_type_rules.c_str();
     sd_ctx_params.n_threads                       = n_threads;
-    sd_ctx_params.conditioning_cache_size         = conditioning_cache_size;
     sd_ctx_params.wtype                           = wtype;
     sd_ctx_params.rng_type                        = rng_type;
     sd_ctx_params.sampler_rng_type                = sampler_rng_type;
@@ -1011,7 +995,6 @@ sd_ctx_params_t SDContextParams::to_sd_ctx_params_t(bool taesd_preview) {
     sd_ctx_params.enable_mmap                     = enable_mmap;
     sd_ctx_params.flash_attn                      = flash_attn;
     sd_ctx_params.diffusion_flash_attn            = diffusion_flash_attn;
-    sd_ctx_params.sage_attn                       = sage_attn;
     sd_ctx_params.linear_scale                    = linear_scale;
     sd_ctx_params.attn_scale                      = attn_scale;
     sd_ctx_params.tae_preview_only                = taesd_preview;
@@ -1126,7 +1109,7 @@ ArgOptions SDGenerationParams::get_options() {
          &hires_upscaler},
         {"",
          "--extra-sample-args",
-         "extra sampler/scheduler/guidance args, key=value list. CFG supports guidance_schedule; APG supports apg_eta, apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end; flux supports base_shift, max_shift; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma; beta scheduler supports alpha, beta; logit_normal supports mu, std, logsnr_min, logsnr_max, resolution_aware; llada_image supports uniform; lms supports lms_max_order, lms_shift, lms_divisions; noise-injecting samplers support noise_sampler with value iid (default except for dpm++2m_sde_bt) or brownian_tree; brownian_tree_rng supports cpu (default), cuda, std_default or sampler_rng",
+         "extra sampler/scheduler/guidance args, key=value list. CFG supports guidance_schedule; APG supports apg_eta, apg_momentum, apg_norm_threshold, apg_norm_threshold_smoothing; SLG supports slg_uncond; lcm supports noise_clip_std, noise_scale_start, noise_scale_end; flux supports base_shift, max_shift; ltx2 supports max_shift, base_shift, stretch, terminal; euler_ge supports gamma; beta scheduler supports alpha, beta; logit_normal supports mu, std, logsnr_min, logsnr_max, resolution_aware; lms supports lms_max_order, lms_shift, lms_divisions; noise-injecting samplers support noise_sampler with value iid (default except for dpm++2m_sde_bt) or brownian_tree; brownian_tree_rng supports cpu (default), cuda, std_default or sampler_rng",
          (int)',',
          &extra_sample_args},
         {"",
@@ -1139,9 +1122,6 @@ ArgOptions SDGenerationParams::get_options() {
          "Key-value list to set up the way the reference images are processed (empty = auto-detect from model weigths)",
          (int)',',
          &ref_image_args},
-        {"", "--image-preprocess",
-         "Image preprocessing rule: target=init|end|mask|control|ref|ip-adapter|id|control-frame,index=N,mode=auto|none|stretch|crop|crop-resize|fit-pad,filter=auto|nearest|nearest-exact|bilinear|bicubic|lanczos,antialias=auto|true|false,width=W,height=H,anchor=center|top|bottom|left|right,pad_color=#RRGGBB[AA],canny=true|false. Repeat for multiple rules.",
-         (int)';', &image_preprocess},
     };
 
     options.int_options = {
@@ -1322,6 +1302,11 @@ ArgOptions SDGenerationParams::get_options() {
          "automatically increase the indices of references images based on the order they are listed (starting with 1).",
          true,
          &increase_ref_index},
+        {"",
+         "--disable-auto-resize-ref-image",
+         "disable auto resize of ref images",
+         false,
+         &auto_resize_ref_image},
         {"",
          "--circular",
          "enable circular padding on both axes for tileable output",
@@ -1341,7 +1326,7 @@ ArgOptions SDGenerationParams::get_options() {
          &embed_image_metadata},
         {"",
          "--vae-tiling",
-         "process vae encode and decode in spatial tiles to reduce memory usage (default: 256x256 image pixels)",
+         "process vae in tiles to reduce memory usage",
          true,
          &vae_tiling_params.enabled},
         {"",
@@ -1605,12 +1590,12 @@ ArgOptions SDGenerationParams::get_options() {
         size_t x_pos              = tile_size_str.find('x');
         try {
             if (x_pos != std::string::npos) {
-                std::string tile_w_str        = tile_size_str.substr(0, x_pos);
-                std::string tile_h_str        = tile_size_str.substr(x_pos + 1);
-                vae_tiling_params.tile_size_w = std::stoi(tile_w_str);
-                vae_tiling_params.tile_size_h = std::stoi(tile_h_str);
+                std::string tile_x_str        = tile_size_str.substr(0, x_pos);
+                std::string tile_y_str        = tile_size_str.substr(x_pos + 1);
+                vae_tiling_params.tile_size_x = std::stoi(tile_x_str);
+                vae_tiling_params.tile_size_y = std::stoi(tile_y_str);
             } else {
-                vae_tiling_params.tile_size_w = vae_tiling_params.tile_size_h = std::stoi(tile_size_str);
+                vae_tiling_params.tile_size_x = vae_tiling_params.tile_size_y = std::stoi(tile_size_str);
             }
         } catch (const std::invalid_argument&) {
             return -1;
@@ -1628,12 +1613,12 @@ ArgOptions SDGenerationParams::get_options() {
         size_t x_pos             = rel_size_str.find('x');
         try {
             if (x_pos != std::string::npos) {
-                std::string rel_w_str        = rel_size_str.substr(0, x_pos);
-                std::string rel_h_str        = rel_size_str.substr(x_pos + 1);
-                vae_tiling_params.rel_size_w = std::stof(rel_w_str);
-                vae_tiling_params.rel_size_h = std::stof(rel_h_str);
+                std::string rel_x_str        = rel_size_str.substr(0, x_pos);
+                std::string rel_y_str        = rel_size_str.substr(x_pos + 1);
+                vae_tiling_params.rel_size_x = std::stof(rel_x_str);
+                vae_tiling_params.rel_size_y = std::stof(rel_y_str);
             } else {
-                vae_tiling_params.rel_size_w = vae_tiling_params.rel_size_h = std::stof(rel_size_str);
+                vae_tiling_params.rel_size_x = vae_tiling_params.rel_size_y = std::stof(rel_size_str);
             }
         } catch (const std::invalid_argument&) {
             return -1;
@@ -1763,11 +1748,11 @@ ArgOptions SDGenerationParams::get_options() {
          on_scm_policy_arg},
         {"",
          "--vae-tile-size",
-         "tile size for vae encode and decode in image pixels, format [W]x[H] or [S] (default: 256x256; requires --vae-tiling)",
+         "tile size for vae tiling, format [X]x[Y] (default: 32x32)",
          on_tile_size_arg},
         {"",
          "--vae-relative-tile-size",
-         "relative tile size for vae encode and decode, format [W]x[H] or [S]: <=1 is a dimension fraction, >1 a target tile count (overrides --vae-tile-size; requires --vae-tiling)",
+         "relative tile size for vae tiling, format [X]x[Y], in fraction of image size if < 1, in number of tiles per dim if >=1 (overrides --vae-tile-size)",
          on_relative_tile_size_arg},
         {"",
          "--prompt-file",
@@ -1857,28 +1842,28 @@ bool decode_base64_image(const std::string& encoded_input,
         return false;
     }
 
-    int decoded_width    = 0;
-    int decoded_height   = 0;
-    int resolved_channel = target_channels;
-    uint8_t* raw_data    = load_image_from_memory(reinterpret_cast<const char*>(image_bytes.data()),
-                                                  static_cast<int>(image_bytes.size()),
-                                                  decoded_width,
-                                                  decoded_height,
-                                                  resolved_channel,
-                                                  expected_width,
-                                                  expected_height,
-                                                  target_channels);
+    int decoded_width  = 0;
+    int decoded_height = 0;
+    uint8_t* raw_data  = load_image_from_memory(reinterpret_cast<const char*>(image_bytes.data()),
+                                                static_cast<int>(image_bytes.size()),
+                                                decoded_width,
+                                                decoded_height,
+                                                expected_width,
+                                                expected_height,
+                                                target_channels);
     if (raw_data == nullptr) {
         return false;
     }
 
-    out_image.reset({(uint32_t)decoded_width, (uint32_t)decoded_height, (uint32_t)resolved_channel, raw_data});
+    out_image.reset({(uint32_t)decoded_width, (uint32_t)decoded_height, (uint32_t)target_channels, raw_data});
     return true;
 }
 
 static bool parse_image_json_field(const json& parent,
                                    const char* key,
                                    int channels,
+                                   int expected_width,
+                                   int expected_height,
                                    SDImageOwner& out_image) {
     if (!parent.contains(key)) {
         return true;
@@ -1890,12 +1875,14 @@ static bool parse_image_json_field(const json& parent,
     if (!parent.at(key).is_string()) {
         return false;
     }
-    return decode_base64_image(parent.at(key).get<std::string>(), channels, 0, 0, out_image);
+    return decode_base64_image(parent.at(key).get<std::string>(), channels, expected_width, expected_height, out_image);
 }
 
 static bool parse_image_array_json_field(const json& parent,
                                          const char* key,
                                          int channels,
+                                         int expected_width,
+                                         int expected_height,
                                          std::vector<SDImageOwner>& out_images) {
     if (!parent.contains(key)) {
         return true;
@@ -1914,7 +1901,7 @@ static bool parse_image_array_json_field(const json& parent,
             return false;
         }
         SDImageOwner image;
-        if (!decode_base64_image(item.get<std::string>(), channels, 0, 0, image)) {
+        if (!decode_base64_image(item.get<std::string>(), channels, expected_width, expected_height, image)) {
             return false;
         }
         out_images.push_back(std::move(image));
@@ -2013,29 +2000,6 @@ static bool resolve_model_file_from_dir(const std::string& model_name,
     return false;
 }
 
-bool SDGenerationParams::parse_image_preprocess_json(const std::string& json_str) {
-    const auto value = json::parse(json_str, nullptr, false);
-    std::string rules;
-    if (value.is_string()) {
-        rules = value.get<std::string>();
-    } else if (value.is_array()) {
-        for (const auto& item : value) {
-            if (!item.is_string()) {
-                LOG_ERROR("image_preprocess must contain rule strings");
-                return false;
-            }
-            if (!rules.empty())
-                rules += ";";
-            rules += item.get<std::string>();
-        }
-    } else {
-        LOG_ERROR("image_preprocess must be a string or array of strings");
-        return false;
-    }
-    image_preprocess = std::move(rules);
-    return true;
-}
-
 bool SDGenerationParams::from_json_str(
     const std::string& json_str,
     const std::function<std::string(const std::string&)>& lora_path_resolver) {
@@ -2047,9 +2011,6 @@ bool SDGenerationParams::from_json_str(
         return false;
     }
 
-    if (j.contains("image_preprocess") && !parse_image_preprocess_json(j["image_preprocess"].dump()))
-        return false;
-
     auto load_if_exists = [&](const char* key, auto& out) {
         if (j.contains(key)) {
             using T = std::decay_t<decltype(out)>;
@@ -2087,7 +2048,6 @@ bool SDGenerationParams::from_json_str(
     load_if_exists("cache_mode", cache_mode);
     load_if_exists("cache_option", cache_option);
     load_if_exists("scm_mask", scm_mask);
-    load_if_exists("ref_image_args", ref_image_args);
 
     load_if_exists("clip_skip", clip_skip);
     load_if_exists("width", width);
@@ -2105,6 +2065,7 @@ bool SDGenerationParams::from_json_str(
     load_if_exists("moe_boundary", moe_boundary);
     load_if_exists("vace_strength", vace_strength);
 
+    load_if_exists("auto_resize_ref_image", auto_resize_ref_image);
     load_if_exists("increase_ref_index", increase_ref_index);
     load_if_exists("embed_image_metadata", embed_image_metadata);
 
@@ -2224,20 +2185,20 @@ bool SDGenerationParams::from_json_str(
         if (tiling_json.contains("temporal_tiling") && tiling_json["temporal_tiling"].is_boolean()) {
             vae_tiling_params.temporal_tiling = tiling_json["temporal_tiling"];
         }
-        if (tiling_json.contains("tile_size_w") && tiling_json["tile_size_w"].is_number_integer()) {
-            vae_tiling_params.tile_size_w = tiling_json["tile_size_w"];
+        if (tiling_json.contains("tile_size_x") && tiling_json["tile_size_x"].is_number_integer()) {
+            vae_tiling_params.tile_size_x = tiling_json["tile_size_x"];
         }
-        if (tiling_json.contains("tile_size_h") && tiling_json["tile_size_h"].is_number_integer()) {
-            vae_tiling_params.tile_size_h = tiling_json["tile_size_h"];
+        if (tiling_json.contains("tile_size_y") && tiling_json["tile_size_y"].is_number_integer()) {
+            vae_tiling_params.tile_size_y = tiling_json["tile_size_y"];
         }
         if (tiling_json.contains("target_overlap") && tiling_json["target_overlap"].is_number()) {
             vae_tiling_params.target_overlap = tiling_json["target_overlap"];
         }
-        if (tiling_json.contains("rel_size_w") && tiling_json["rel_size_w"].is_number()) {
-            vae_tiling_params.rel_size_w = tiling_json["rel_size_w"];
+        if (tiling_json.contains("rel_size_x") && tiling_json["rel_size_x"].is_number()) {
+            vae_tiling_params.rel_size_x = tiling_json["rel_size_x"];
         }
-        if (tiling_json.contains("rel_size_h") && tiling_json["rel_size_h"].is_number()) {
-            vae_tiling_params.rel_size_h = tiling_json["rel_size_h"];
+        if (tiling_json.contains("rel_size_y") && tiling_json["rel_size_y"].is_number()) {
+            vae_tiling_params.rel_size_y = tiling_json["rel_size_y"];
         }
         if (tiling_json.contains("extra_tiling_args") && tiling_json["extra_tiling_args"].is_string()) {
             extra_tiling_args = tiling_json["extra_tiling_args"].get<std::string>();
@@ -2248,23 +2209,32 @@ bool SDGenerationParams::from_json_str(
         LOG_ERROR("invalid lora");
         return false;
     }
-    auto load_image = [&](const char* key, int channels, SDImageOwner& image) {
-        if (!parse_image_json_field(j, key, channels, image)) {
-            LOG_ERROR("invalid %s", key);
-            return false;
-        }
-        return true;
-    };
-    if (!load_image("init_image", 0, init_image) ||
-        !load_image("end_image", 3, end_image) ||
-        !load_image("mask_image", 1, mask_image) ||
-        !load_image("control_image", 3, control_image) ||
-        !load_image("ip_adapter_image", 3, ip_adapter_image)) {
+    if (!parse_image_json_field(j, "init_image", 3, width, height, init_image)) {
+        LOG_ERROR("invalid init_image");
         return false;
     }
-    if (!parse_image_array_json_field(j, "ref_images", 0, ref_images) ||
-        !parse_image_array_json_field(j, "control_frames", 3, control_frames)) {
-        LOG_ERROR("invalid input image array");
+    if (!parse_image_json_field(j, "end_image", 3, width, height, end_image)) {
+        LOG_ERROR("invalid end_image");
+        return false;
+    }
+    if (!parse_image_array_json_field(j, "ref_images", 3, width, height, ref_images)) {
+        LOG_ERROR("invalid ref_images");
+        return false;
+    }
+    if (!parse_image_array_json_field(j, "control_frames", 3, width, height, control_frames)) {
+        LOG_ERROR("invalid control_frames");
+        return false;
+    }
+    if (!parse_image_json_field(j, "mask_image", 1, width, height, mask_image)) {
+        LOG_ERROR("invalid mask_image");
+        return false;
+    }
+    if (!parse_image_json_field(j, "control_image", 3, width, height, control_image)) {
+        LOG_ERROR("invalid control_image");
+        return false;
+    }
+    if (!parse_image_json_field(j, "ip_adapter_image", 3, width, height, ip_adapter_image)) {
+        LOG_ERROR("invalid ip_adapter_image");
         return false;
     }
 
@@ -2508,10 +2478,6 @@ bool SDGenerationParams::resolve(const std::string& lora_model_dir, const std::s
 }
 
 bool SDGenerationParams::validate(SDMode mode) {
-    if (!image_preprocess.empty() && mode != IMG_GEN && mode != VID_GEN) {
-        LOG_ERROR("--image-preprocess requires img_gen or vid_gen mode");
-        return false;
-    }
     if (batch_count <= 0) {
         LOG_ERROR("error: batch_count must be greater than 0");
         return false;
@@ -2687,6 +2653,14 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
         pulid_id_weight,
     };
 
+    if (!auto_resize_ref_image) {
+        if (!ref_image_args.empty()) {
+            ref_image_args += ",";
+        }
+        ref_image_args += "resize_before_vae=0";
+        LOG_WARN("Notice: --disable-auto-resize-ref-image is deprecated. Use --ref-image-args \"resize_before_vae=off\" instead.");
+    }
+
     if (increase_ref_index) {
         if (!ref_image_args.empty()) {
             ref_image_args += ",";
@@ -2734,7 +2708,6 @@ sd_img_gen_params_t SDGenerationParams::to_sd_img_gen_params_t() {
     params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
     params.circular_x                = circular || circular_x;
     params.circular_y                = circular || circular_y;
-    params.image_preprocess          = {image_preprocess.c_str()};
     return params;
 }
 
@@ -2837,7 +2810,6 @@ sd_vid_gen_params_t SDGenerationParams::to_sd_vid_gen_params_t() {
     params.hires.custom_sigmas_count = static_cast<int>(hires_custom_sigmas.size());
     params.circular_x                = circular || circular_x;
     params.circular_y                = circular || circular_y;
-    params.image_preprocess          = {image_preprocess.c_str()};
     return params;
 }
 
@@ -2894,8 +2866,7 @@ std::string SDGenerationParams::to_string() const {
         << "  ref_video_audio_paths: " << vec_str_to_string(ref_video_audio_paths) << ",\n"
         << "  ref_audio_paths: " << vec_str_to_string(ref_audio_paths) << ",\n"
         << "  control_video_path: \"" << control_video_path << "\",\n"
-        << "  image_preprocess: " << image_preprocess << ",\n"
-        << "  ref_image_args: " << ref_image_args << ",\n"
+        << "  auto_resize_ref_image: " << (auto_resize_ref_image ? "true" : "false") << ",\n"
         << "  increase_ref_index: " << (increase_ref_index ? "true" : "false") << ",\n"
         << "  pm_id_images_dir: \"" << pm_id_images_dir << "\",\n"
         << "  pm_id_embed_path: \"" << pm_id_embed_path << "\",\n"
@@ -2934,11 +2905,11 @@ std::string SDGenerationParams::to_string() const {
         << "  vae_tiling_params: { "
         << vae_tiling_params.enabled << ", "
         << vae_tiling_params.temporal_tiling << ", "
-        << vae_tiling_params.tile_size_w << ", "
-        << vae_tiling_params.tile_size_h << ", "
+        << vae_tiling_params.tile_size_x << ", "
+        << vae_tiling_params.tile_size_y << ", "
         << vae_tiling_params.target_overlap << ", "
-        << vae_tiling_params.rel_size_w << ", "
-        << vae_tiling_params.rel_size_h << ", "
+        << vae_tiling_params.rel_size_x << ", "
+        << vae_tiling_params.rel_size_y << ", "
         << "\"" << extra_tiling_args << "\" },\n"
         << "}";
     return oss.str();
@@ -3046,13 +3017,12 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
     set_json_basename_if_not_empty(models, "control_net", ctx_params.control_net_path);
     root["models"] = std::move(models);
 
-    root["clip_skip"]           = gen_params.clip_skip;
-    root["strength"]            = gen_params.strength;
-    root["control_strength"]    = gen_params.control_strength;
-    root["ip_adapter_strength"] = gen_params.ip_adapter_strength;
-    root["ref_image_args"]      = gen_params.ref_image_args;
-    root["image_preprocess"]    = gen_params.image_preprocess;
-    root["increase_ref_index"]  = gen_params.increase_ref_index;
+    root["clip_skip"]             = gen_params.clip_skip;
+    root["strength"]              = gen_params.strength;
+    root["control_strength"]      = gen_params.control_strength;
+    root["ip_adapter_strength"]   = gen_params.ip_adapter_strength;
+    root["auto_resize_ref_image"] = gen_params.auto_resize_ref_image;
+    root["increase_ref_index"]    = gen_params.increase_ref_index;
     if (mode == VID_GEN) {
         root["video"] = {
             {"frame_count", gen_params.video_frames},
@@ -3140,11 +3110,11 @@ std::string build_sdcpp_image_metadata_json(const SDContextParams& ctx_params,
         root["vae_tiling"] = {
             {"enabled", gen_params.vae_tiling_params.enabled},
             {"temporal_tiling", gen_params.vae_tiling_params.temporal_tiling},
-            {"tile_size_w", gen_params.vae_tiling_params.tile_size_w},
-            {"tile_size_h", gen_params.vae_tiling_params.tile_size_h},
+            {"tile_size_x", gen_params.vae_tiling_params.tile_size_x},
+            {"tile_size_y", gen_params.vae_tiling_params.tile_size_y},
             {"target_overlap", gen_params.vae_tiling_params.target_overlap},
-            {"rel_size_w", gen_params.vae_tiling_params.rel_size_w},
-            {"rel_size_h", gen_params.vae_tiling_params.rel_size_h},
+            {"rel_size_x", gen_params.vae_tiling_params.rel_size_x},
+            {"rel_size_y", gen_params.vae_tiling_params.rel_size_y},
             {"extra_tiling_args", gen_params.extra_tiling_args},
         };
     }

examples/common/common.h

@@ -116,8 +116,7 @@ bool decode_base64_image(const std::string& encoded_input,
                          SDImageOwner& out_image);
 
 struct SDContextParams {
-    int n_threads               = -1;
-    int conditioning_cache_size = 4;
+    int n_threads = -1;
     std::string model_path;
     std::string clip_l_path;
     std::string clip_g_path;
@@ -171,7 +170,6 @@ struct SDContextParams {
     bool vae_on_cpu            = false;
     bool flash_attn            = false;
     bool diffusion_flash_attn  = false;
-    bool sage_attn             = false;
     bool diffusion_conv_direct = false;
     bool vae_conv_direct       = false;
 
@@ -201,17 +199,18 @@ struct SDGenerationParams {
     std::string ad_prompt;
     std::string ad_negative_prompt;
     std::string extra_ad_args;
-    int clip_skip             = -1;  // <= 0 represents unspecified
-    int width                 = -1;
-    int height                = -1;
-    int batch_count           = 1;
-    int qwen_image_layers     = 3;
-    int64_t seed              = 42;
-    float strength            = 0.75f;
-    float control_strength    = 0.9f;
-    float ip_adapter_strength = 1.0f;
-    bool increase_ref_index   = false;
-    bool embed_image_metadata = true;
+    int clip_skip              = -1;  // <= 0 represents unspecified
+    int width                  = -1;
+    int height                 = -1;
+    int batch_count            = 1;
+    int qwen_image_layers      = 3;
+    int64_t seed               = 42;
+    float strength             = 0.75f;
+    float control_strength     = 0.9f;
+    float ip_adapter_strength  = 1.0f;
+    bool auto_resize_ref_image = true;
+    bool increase_ref_index    = false;
+    bool embed_image_metadata  = true;
 
     std::string init_image_path;
     std::string end_image_path;
@@ -247,7 +246,6 @@ struct SDGenerationParams {
     std::string extra_tiling_args;
 
     std::string ref_image_args;
-    std::string image_preprocess;
 
     std::string pm_id_images_dir;
     std::string pm_id_embed_path;
@@ -311,7 +309,6 @@ struct SDGenerationParams {
     ArgOptions get_options();
     bool from_json_str(const std::string& json_str,
                        const std::function<std::string(const std::string&)>& lora_path_resolver = {});
-    bool parse_image_preprocess_json(const std::string& json_str);
     bool initialize_cache_params();
     void extract_and_remove_lora(const std::string& lora_model_dir);
     bool width_and_height_are_set() const;

examples/common/media_io.cpp

@@ -261,10 +261,6 @@ uint8_t* decode_webp_image_to_buffer(const uint8_t* data,
     height               = features.height;
     source_channel_count = features.has_alpha ? 4 : 3;
 
-    if (expected_channel == 0) {
-        expected_channel = source_channel_count;
-    }
-
     const size_t pixel_count = static_cast<size_t>(width) * static_cast<size_t>(height);
 
     if (expected_channel == 1) {
@@ -485,8 +481,7 @@ uint8_t* load_image_common(bool from_memory,
                            int& height,
                            int expected_width,
                            int expected_height,
-                           int expected_channel,
-                           int& out_channel) {
+                           int expected_channel) {
     const char* image_path;
     FreeUniquePtr<uint8_t> image_buffer;
     int source_channel_count = 0;
@@ -543,32 +538,6 @@ uint8_t* load_image_common(bool from_memory,
         LOG_ERROR("load image from '%s' failed", image_path);
         return nullptr;
     }
-    if (expected_channel == 0) {
-        expected_channel = source_channel_count == 2 ? 4 : (source_channel_count == 1 ? 3 : source_channel_count);
-        if (expected_channel != source_channel_count) {
-            FreeUniquePtr<uint8_t> promoted((uint8_t*)malloc((size_t)width * height * expected_channel));
-            if (promoted == nullptr) {
-                LOG_ERROR("error: allocate memory for channel promotion, image_path = %s", image_path);
-                return nullptr;
-            }
-            const size_t pixel_count = (size_t)width * (size_t)height;
-            for (size_t i = 0; i < pixel_count; ++i) {
-                if (source_channel_count == 1) {
-                    promoted.get()[i * 3 + 0] = image_buffer.get()[i];
-                    promoted.get()[i * 3 + 1] = image_buffer.get()[i];
-                    promoted.get()[i * 3 + 2] = image_buffer.get()[i];
-                } else {
-                    promoted.get()[i * 4 + 0] = image_buffer.get()[i * 2];
-                    promoted.get()[i * 4 + 1] = image_buffer.get()[i * 2];
-                    promoted.get()[i * 4 + 2] = image_buffer.get()[i * 2];
-                    promoted.get()[i * 4 + 3] = image_buffer.get()[i * 2 + 1];
-                }
-            }
-            image_buffer         = std::move(promoted);
-            source_channel_count = expected_channel;
-        }
-    }
-    // stb reports the source channel count even when it converts the output.
     if (source_channel_count < expected_channel) {
         fprintf(stderr,
                 "the number of channels for the input image must be >= %d,"
@@ -628,7 +597,7 @@ uint8_t* load_image_common(bool from_memory,
         }
         stbir_resize(image_buffer.get(), width, height, 0,
                      resized_image_buffer.get(), expected_width, expected_height, 0, STBIR_TYPE_UINT8,
-                     expected_channel, expected_channel == 4 ? 3 : STBIR_ALPHA_CHANNEL_NONE, 0,
+                     expected_channel, STBIR_ALPHA_CHANNEL_NONE, 0,
                      STBIR_EDGE_CLAMP, STBIR_EDGE_CLAMP,
                      STBIR_FILTER_BOX, STBIR_FILTER_BOX,
                      STBIR_COLORSPACE_SRGB, nullptr);
@@ -636,7 +605,6 @@ uint8_t* load_image_common(bool from_memory,
         height       = expected_height;
         image_buffer = std::move(resized_image_buffer);
     }
-    out_channel = expected_channel;
     return image_buffer.release();
 }
 
@@ -809,11 +777,10 @@ bool write_image_to_file(const std::string& path,
 uint8_t* load_image_from_file(const char* image_path,
                               int& width,
                               int& height,
-                              int& out_channel,
                               int expected_width,
                               int expected_height,
                               int expected_channel) {
-    return load_image_common(false, image_path, 0, width, height, expected_width, expected_height, expected_channel, out_channel);
+    return load_image_common(false, image_path, 0, width, height, expected_width, expected_height, expected_channel);
 }
 
 bool load_sd_image_from_file(sd_image_t* image,
@@ -823,14 +790,13 @@ bool load_sd_image_from_file(sd_image_t* image,
                              int expected_channel) {
     int width;
     int height;
-    int resolved_channel = expected_channel;
-    image->data          = load_image_common(false, image_path, 0, width, height, expected_width, expected_height, expected_channel, resolved_channel);
+    image->data = load_image_common(false, image_path, 0, width, height, expected_width, expected_height, expected_channel);
     if (image->data == nullptr) {
         return false;
     }
     image->width   = width;
     image->height  = height;
-    image->channel = resolved_channel;
+    image->channel = expected_channel;
     return true;
 }
 
@@ -838,11 +804,10 @@ uint8_t* load_image_from_memory(const char* image_bytes,
                                 int len,
                                 int& width,
                                 int& height,
-                                int& out_channel,
                                 int expected_width,
                                 int expected_height,
                                 int expected_channel) {
-    return load_image_common(true, image_bytes, len, width, height, expected_width, expected_height, expected_channel, out_channel);
+    return load_image_common(true, image_bytes, len, width, height, expected_width, expected_height, expected_channel);
 }
 
 static void append_avi_metadata(std::vector<uint8_t>& data, const std::string& parameters) {

examples/common/media_io.h

@@ -32,12 +32,9 @@ bool write_image_to_file(const std::string& path,
                          const std::string& parameters = "",
                          int quality                   = 90);
 
-// expected_channel == 0 preserves native channels (grayscale -> RGB, gray+alpha -> RGBA).
-// out_channel receives the output channel count.
 uint8_t* load_image_from_file(const char* image_path,
                               int& width,
                               int& height,
-                              int& out_channel,
                               int expected_width   = 0,
                               int expected_height  = 0,
                               int expected_channel = 3);
@@ -52,7 +49,6 @@ uint8_t* load_image_from_memory(const char* image_bytes,
                                 int len,
                                 int& width,
                                 int& height,
-                                int& out_channel,
                                 int expected_width   = 0,
                                 int expected_height  = 0,
                                 int expected_channel = 3);

examples/server/CMakeLists.txt

@@ -21,21 +21,21 @@ if(SD_SERVER_BUILD_FRONTEND AND EXISTS "${FRONTEND_DIR}")
         set(HAVE_FRONTEND_BUILD ON)
 
         add_custom_target(${TARGET}_frontend_install
-            COMMAND "${PNPM_EXECUTABLE}" --ignore-workspace -C "${FRONTEND_DIR}" install
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" install
             WORKING_DIRECTORY "${FRONTEND_DIR}"
             COMMENT "Installing frontend dependencies"
             VERBATIM
         )
 
         add_custom_target(${TARGET}_frontend_build
-            COMMAND "${PNPM_EXECUTABLE}" --ignore-workspace -C "${FRONTEND_DIR}" run build
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" run build
             WORKING_DIRECTORY "${FRONTEND_DIR}"
             COMMENT "Building frontend"
             VERBATIM
         )
 
         add_custom_target(${TARGET}_frontend_header
-            COMMAND "${PNPM_EXECUTABLE}" --ignore-workspace -C "${FRONTEND_DIR}" run build:header
+            COMMAND "${PNPM_EXECUTABLE}" -C "${FRONTEND_DIR}" run build:header
             WORKING_DIRECTORY "${FRONTEND_DIR}"
             COMMENT "Generating gen_index_html.h"
             VERBATIM

examples/server/api.md

@@ -56,7 +56,6 @@ Current endpoints include:
 - `GET /sdcpp/v1/jobs/{id}`
 - `POST /sdcpp/v1/jobs/{id}/cancel`
 - `POST /sdcpp/v1/vid_gen`
-- `POST /sdcpp/v1/upscale`
 
 ## `sd_cpp_extra_args`
 
@@ -149,19 +148,6 @@ Native extension fields:
 
 - any `sdcpp API` fields embedded through `sd_cpp_extra_args` inside `prompt`
 
-Uploaded images are decoded at their original dimensions. The first decoded
-image establishes the generation dimensions if `size` is omitted. Input
-geometry follows `image_preprocess`: references preserve their dimensions by
-default, while init and mask use the generation canvas preset.
-
-Reference encoding then follows model presets and `ref_image_args`. To skip
-input geometry for references and disable resizing before VAE encoding, include
-this in `prompt`:
-
-```text
-edit this image <sd_cpp_extra_args>{"image_preprocess":"target=ref,mode=none","ref_image_args":"resize_before_vae=false"}</sd_cpp_extra_args>
-```
-
 Response fields:
 
 | Field | Type | Notes |
@@ -435,8 +421,7 @@ Top-level fields:
 | `samplers` | `array<string>` | Available sampling methods |
 | `schedulers` | `array<string>` | Available schedulers |
 | `loras` | `array<object>` | Available LoRA entries |
-| `upscalers` | `array<object>` | Available highres upscalers, built-in and model-backed |
-| `upscale` | `boolean` | Whether a compatible RGB ESRGAN model is available for `POST /sdcpp/v1/upscale` |
+| `upscalers` | `array<object>` | Available model-backed highres upscalers |
 | `limits` | `object` | Shared queue and size limits |
 
 `model`
@@ -478,8 +463,6 @@ Shared nested fields:
 | Field | Type | Notes |
 | --- | --- | --- |
 | `upscalers[].name` | `string` | Built-in name or model stem; use this value in `hires.upscaler` |
-| `upscalers[].model` | `boolean` | True for a model-backed upscaler, false for a built-in scaling filter |
-| `upscalers[].image_upscale` | `boolean` | Whether this model can be selected by `POST /sdcpp/v1/upscale`; false for latent upscalers and built-in filters |
 
 Built-in entries include `None`, `Lanczos`, `Nearest`, `Latent`, `Latent (nearest)`, `Latent (nearest-exact)`, `Latent (antialiased)`, `Latent (bicubic)`, and `Latent (bicubic antialiased)`. Model-backed entries are scanned from the top level of `--hires-upscalers-dir`; subdirectories are not scanned.
 
@@ -493,8 +476,6 @@ Built-in entries include `None`, `Lanczos`, `Nearest`, `Latent`, `Latent (neares
 | `limits.max_height` | `integer` |
 | `limits.max_batch_count` | `integer` |
 | `limits.max_queue_size` | `integer` |
-| `limits.max_upscale_width` | `integer` |
-| `limits.max_upscale_height` | `integer` |
 
 Shared default fields used by both `img_gen` and `vid_gen`:
 
@@ -524,11 +505,11 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `vae_tiling_params` | `object` |
 | `vae_tiling_params.enabled` | `boolean` |
 | `vae_tiling_params.temporal_tiling` | `boolean` |
-| `vae_tiling_params.tile_size_w` | `integer` |
-| `vae_tiling_params.tile_size_h` | `integer` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
 | `vae_tiling_params.target_overlap` | `number` |
-| `vae_tiling_params.rel_size_w` | `number` |
-| `vae_tiling_params.rel_size_h` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
 | `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
@@ -537,8 +518,6 @@ Shared default fields used by both `img_gen` and `vid_gen`:
 | `output_format` | `string` |
 | `output_compression` | `integer` |
 
-`vae_tiling_params.tile_size_w` and `tile_size_h` are in **image pixels**, with `0` selecting the 256-pixel default. Both encode and decode use these sizes without an encoding multiplier. Positive `rel_size_w`/`rel_size_h` values override the corresponding absolute size: values up to 1 are dimension fractions, and values greater than 1 are target tile counts. Set `enabled` to use spatial tiling. Sizes are aligned down to the VAE scale factor and capped at the input dimensions; explicit sizes below the minimum supported tile size are rejected. These fields previously used latent units; see [VAE tiling](../../docs/performance.md#use-vae-tiling-to-reduce-encode-and-decode-memory-usage) for migration and OOM retry behavior.
-
 `vae_tiling_params.extra_tiling_args` accepts a key=value list. Supported video VAEs accept `temporal_tile_frames` (alias `temporal_tile_size`, default `4`) and `temporal_tile_overlap` (default `1`).
 LTX and Wan preserve causal state between temporal tiles. Hunyuan Video and TAEHV use overlap blending. MiniMax H3 keeps its model-specific fixed temporal windows because its latent-to-frame mapping is non-linear.
 
@@ -547,7 +526,7 @@ LTX and Wan preserve causal state between temporal tiles. Hunyuan Video and TAEH
 | Field | Type |
 | --- | --- |
 | `batch_count` | `integer` |
-| `ref_image_args` | `string` |
+| `auto_resize_ref_image` | `boolean` |
 | `increase_ref_index` | `boolean` |
 | `control_strength` | `number` |
 | `ip_adapter_strength` | `number` |
@@ -649,52 +628,6 @@ Typical status codes:
 - `404 Not Found`
 - `410 Gone`
 
-#### `POST /sdcpp/v1/upscale`
-
-Runs one RGB ESRGAN upscaler over an image, with no generation involved. Latent upscaler models remain available for hires generation but cannot be used here.
-
-This is the HTTP equivalent of `sd-cli -M upscale`: no diffusion model, text
-encoder or sampling is used, so it is fast enough to answer synchronously and
-does not create a job.
-
-Request fields:
-
-| Field | Type | Notes |
-| --- | --- | --- |
-| `image` | `string` | Required. Base64 or data URL image |
-| `upscaler` | `string` | A name from `upscalers` with `image_upscale: true`; the first compatible entry when omitted |
-| `repeats` | `integer` | Run the upscaler this many times, 1 to 4 (default `1`) |
-| `tile_size` | `integer` | Tile size, defaulting to the server's `--upscale-tile-size` |
-| `output_format` | `string` | `png`, `jpeg`, or `webp` when built with WebP support (default `png`); unsupported formats return 400 |
-| `output_compression` | `integer` | Range is clamped to `0..100` |
-
-Response fields:
-
-| Field | Type | Notes |
-| --- | --- | --- |
-| `images` | `array<object>` | One image |
-| `images[].index` | `integer` | |
-| `images[].b64_json` | `string` | Base64-encoded image bytes |
-| `upscaler` | `string` | The upscaler actually used |
-| `scale` | `integer` | The model's scale factor |
-| `repeats` | `integer` | How many times it was run |
-| `width` | `integer` | Result width |
-| `height` | `integer` | Result height |
-| `output_format` | `string` | Final encoded image format |
-
-Typical status codes:
-
-- `200 OK`
-- `400 Bad Request` (invalid request, unsupported output format, unreadable image, incompatible upscaler, or output dimensions exceeding the limit)
-- `500 Internal Server Error`
-
-Notes:
-
-- Final output dimensions, including all repeats, must not exceed 8192 pixels on either axis (`limits.max_upscale_width` and `limits.max_upscale_height`). Requests exceeding this bound are rejected before upscaling.
-- The upscaler models are three-channel; alpha is not preserved.
-- The request holds the generation context lock, so an upscale and a
-  generation never run on the device at the same time.
-
 #### `POST /sdcpp/v1/jobs/{id}/cancel`
 
 Attempts to cancel an accepted job.
@@ -720,7 +653,7 @@ Example:
   "strength": 0.75,
   "seed": -1,
   "batch_count": 1,
-  "ref_image_args": "",
+  "auto_resize_ref_image": true,
   "increase_ref_index": false,
   "control_strength": 0.9,
   "ip_adapter_strength": 1.0,
@@ -769,11 +702,11 @@ Example:
   "vae_tiling_params": {
     "enabled": false,
     "temporal_tiling": false,
-    "tile_size_w": 0,
-    "tile_size_h": 0,
+    "tile_size_x": 0,
+    "tile_size_y": 0,
     "target_overlap": 0.5,
-    "rel_size_w": 0.0,
-    "rel_size_h": 0.0,
+    "rel_size_x": 0.0,
+    "rel_size_y": 0.0,
     "extra_tiling_args": ""
   },
 
@@ -795,17 +728,6 @@ Example:
 
 ### Image Encoding Rules
 
-Native image/video requests and SDAPI accept `image_preprocess` as a rule string
-or array of rule strings. OpenAI-compatible requests can supply it in
-`sd_cpp_extra_args`. See [Image preprocessing](../../docs/image_preprocessing.md)
-for one-time input geometry, native-resolution decoding, mask alignment, and
-`canny=true` for edge detection on any supported image input.
-
-Image generation also accepts `ref_image_args` as a string (for example,
-`"resize_before_vae=false"`) in native and SDAPI requests, or through
-`sd_cpp_extra_args` in OpenAI-compatible requests. It controls downstream
-reference encoding and is independent of input geometry rules.
-
 Any image field accepts:
 
 - a raw base64 string, or
@@ -813,15 +735,12 @@ Any image field accepts:
 
 Channel expectations:
 
-- `init_image`: native channels (3 or 4); alpha is preserved and applied per model
-- `ref_images[]`: native channels (3 or 4); alpha is preserved and applied per model
+- `init_image`: 3 channels
+- `ref_images[]`: 3 channels
 - `control_image`: 3 channels
 - `ip_adapter_image`: 3 channels
 - `mask_image`: 1 channel
 
-Models that support RGBA (e.g. Qwen-Image 2.1) use the alpha channel of `init_image`
-and `ref_images[]`. RGB-only models drop it, so sending RGBA is safe for every model.
-
 If omitted or null:
 
 - single-image fields map to an empty `sd_image_t`
@@ -841,8 +760,7 @@ Top-level scalar fields:
 | `strength` | `number` |
 | `seed` | `integer` |
 | `batch_count` | `integer` |
-| `ref_image_args` | `string` |
-| `image_preprocess` | `string \| array<string>` |
+| `auto_resize_ref_image` | `boolean` |
 | `increase_ref_index` | `boolean` |
 | `control_strength` | `number` |
 | `ip_adapter_strength` | `number` |
@@ -902,11 +820,11 @@ Other native fields:
 | `vae_tiling_params` | `object` |
 | `vae_tiling_params.enabled` | `boolean` |
 | `vae_tiling_params.temporal_tiling` | `boolean` |
-| `vae_tiling_params.tile_size_w` | `integer` |
-| `vae_tiling_params.tile_size_h` | `integer` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
 | `vae_tiling_params.target_overlap` | `number` |
-| `vae_tiling_params.rel_size_w` | `number` |
-| `vae_tiling_params.rel_size_h` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
 | `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |
@@ -1117,11 +1035,11 @@ Example:
   "vae_tiling_params": {
     "enabled": false,
     "temporal_tiling": false,
-    "tile_size_w": 0,
-    "tile_size_h": 0,
+    "tile_size_x": 0,
+    "tile_size_y": 0,
     "target_overlap": 0.5,
-    "rel_size_w": 0.0,
-    "rel_size_h": 0.0,
+    "rel_size_x": 0.0,
+    "rel_size_y": 0.0,
     "extra_tiling_args": ""
   },
 
@@ -1242,11 +1160,11 @@ Other native fields:
 | `vae_tiling_params` | `object` |
 | `vae_tiling_params.enabled` | `boolean` |
 | `vae_tiling_params.temporal_tiling` | `boolean` |
-| `vae_tiling_params.tile_size_w` | `integer` |
-| `vae_tiling_params.tile_size_h` | `integer` |
+| `vae_tiling_params.tile_size_x` | `integer` |
+| `vae_tiling_params.tile_size_y` | `integer` |
 | `vae_tiling_params.target_overlap` | `number` |
-| `vae_tiling_params.rel_size_w` | `number` |
-| `vae_tiling_params.rel_size_h` | `number` |
+| `vae_tiling_params.rel_size_x` | `number` |
+| `vae_tiling_params.rel_size_y` | `number` |
 | `vae_tiling_params.extra_tiling_args` | `string` |
 | `cache_mode` | `string` |
 | `cache_option` | `string` |

examples/server/frontend

@@ -1 +1 @@
-Subproject commit dd74a8e808aaa8b26124217424b23058935184de
+Subproject commit c4bce3d6b3f236614cca21014f076083b7270ba8

examples/server/main.cpp

@@ -76,9 +76,9 @@ int main(int argc, const char** argv) {
     SDSvrParams svr_params;
     SDContextParams ctx_params;
     SDGenerationParams default_gen_params;
+    parse_args(argc, argv, svr_params, ctx_params, default_gen_params);
 
     sd_set_log_callback(sd_log_cb, (void*)&svr_params);
-    parse_args(argc, argv, svr_params, ctx_params, default_gen_params);
 
     LOG_VERBOSE("version: %s", version_string().c_str());
     LOG_VERBOSE("%s", sd_get_system_info());

examples/server/routes_openai.cpp

@@ -157,46 +157,42 @@ static bool build_openai_edit_request(const httplib::Request& req,
     request.gen_params.height      = height;
     request.gen_params.batch_count = n;
 
-    std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(request.gen_params.prompt);
     for (auto& bytes : images_bytes) {
-        int img_w            = 0;
-        int img_h            = 0;
-        int resolved_channel = 0;
-        uint8_t* raw_pixels  = load_image_from_memory(
-             reinterpret_cast<const char*>(bytes.data()),
-             static_cast<int>(bytes.size()),
-             img_w, img_h, resolved_channel,
-             0, 0,
-             0);
+        int img_w           = 0;
+        int img_h           = 0;
+        uint8_t* raw_pixels = load_image_from_memory(
+            reinterpret_cast<const char*>(bytes.data()),
+            static_cast<int>(bytes.size()),
+            img_w, img_h,
+            width, height, 3);
         if (raw_pixels == nullptr) {
             continue;
         }
 
-        const bool is_first_ref_image = request.gen_params.ref_images.empty();
-        SDImageOwner image_owner({(uint32_t)img_w, (uint32_t)img_h, (uint32_t)resolved_channel, raw_pixels});
+        SDImageOwner image_owner({(uint32_t)img_w, (uint32_t)img_h, 3, raw_pixels});
         request.gen_params.set_width_and_height_if_unset(image_owner.get().width, image_owner.get().height);
-
-        if (is_first_ref_image) {
-            request.gen_params.init_image = image_owner;
-            if (request.gen_params.init_image.get().data == nullptr) {
-                error_message = "could not allocate init image";
-                return false;
-            }
-        }
-
         request.gen_params.ref_images.push_back(std::move(image_owner));
     }
 
+    if (!request.gen_params.ref_images.empty()) {
+        request.gen_params.init_image = request.gen_params.ref_images.front();
+    }
+
     if (!mask_bytes.empty()) {
-        int mask_w       = 0;
-        int mask_h       = 0;
-        int mask_channel = 0;
+        int expected_width  = 0;
+        int expected_height = 0;
+        if (request.gen_params.width_and_height_are_set()) {
+            expected_width  = request.gen_params.width;
+            expected_height = request.gen_params.height;
+        }
+        int mask_w = 0;
+        int mask_h = 0;
 
         uint8_t* mask_raw = load_image_from_memory(
             reinterpret_cast<const char*>(mask_bytes.data()),
             static_cast<int>(mask_bytes.size()),
-            mask_w, mask_h, mask_channel,
-            0, 0, 1);
+            mask_w, mask_h,
+            expected_width, expected_height, 1);
         request.gen_params.mask_image.reset({(uint32_t)mask_w, (uint32_t)mask_h, 1, mask_raw});
         const sd_image_t& mask_image = request.gen_params.mask_image.get();
         request.gen_params.set_width_and_height_if_unset(mask_image.width, mask_image.height);
@@ -209,6 +205,7 @@ static bool build_openai_edit_request(const httplib::Request& req,
         });
     }
 
+    std::string sd_cpp_extra_args_str = extract_and_remove_sd_cpp_extra_args(request.gen_params.prompt);
     if (!sd_cpp_extra_args_str.empty() && !request.gen_params.from_json_str(sd_cpp_extra_args_str)) {
         error_message = "invalid sd_cpp_extra_args";
         return false;

examples/server/routes_sdapi.cpp

@@ -80,6 +80,17 @@ static enum sample_method_t get_sdapi_sample_method(std::string name) {
     return it != hardcoded.end() ? it->second : SAMPLE_METHOD_COUNT;
 }
 
+static void assign_solid_mask(SDImageOwner& mask_owner, int width, int height) {
+    const size_t pixel_count = static_cast<size_t>(width) * static_cast<size_t>(height);
+    uint8_t* raw_mask        = static_cast<uint8_t*>(malloc(pixel_count));
+    if (raw_mask == nullptr) {
+        mask_owner.reset({0, 0, 1, nullptr});
+        return;
+    }
+    std::memset(raw_mask, 255, pixel_count);
+    mask_owner.reset({(uint32_t)width, (uint32_t)height, 1, raw_mask});
+}
+
 static bool build_sdapi_img_gen_request(const json& j,
                                         ServerRuntime& runtime,
                                         bool img2img,
@@ -182,25 +193,15 @@ static bool build_sdapi_img_gen_request(const json& j,
         }
     }
 
-    if (j.contains("ref_image_args")) {
-        if (!j["ref_image_args"].is_string()) {
-            error_message = "ref_image_args must be a string";
-            return false;
-        }
-        request.gen_params.ref_image_args = j["ref_image_args"].get<std::string>();
-    }
-
-    if (j.contains("image_preprocess") && !request.gen_params.parse_image_preprocess_json(j["image_preprocess"].dump())) {
-        error_message = "invalid image_preprocess";
-        return false;
-    }
-
     if (img2img) {
+        const int expected_width  = request.gen_params.width_and_height_are_set() ? request.gen_params.width : 0;
+        const int expected_height = request.gen_params.width_and_height_are_set() ? request.gen_params.height : 0;
+
         if (j.contains("init_images") && j["init_images"].is_array() && !j["init_images"].empty()) {
             if (decode_base64_image(j["init_images"][0].get<std::string>(),
-                                    0,
-                                    0,
-                                    0,
+                                    3,
+                                    expected_width,
+                                    expected_height,
                                     request.gen_params.init_image)) {
                 const sd_image_t& image = request.gen_params.init_image.get();
                 request.gen_params.set_width_and_height_if_unset(image.width, image.height);
@@ -210,8 +211,8 @@ static bool build_sdapi_img_gen_request(const json& j,
         if (j.contains("mask") && j["mask"].is_string()) {
             if (decode_base64_image(j["mask"].get<std::string>(),
                                     1,
-                                    0,
-                                    0,
+                                    expected_width,
+                                    expected_height,
                                     request.gen_params.mask_image)) {
                 const sd_image_t& image = request.gen_params.mask_image.get();
                 request.gen_params.set_width_and_height_if_unset(image.width, image.height);
@@ -224,7 +225,9 @@ static bool build_sdapi_img_gen_request(const json& j,
                 }
             }
         } else {
-            request.gen_params.mask_image.reset({0, 0, 1, nullptr});
+            const int resolved_width  = request.gen_params.get_resolved_width();
+            const int resolved_height = request.gen_params.get_resolved_height();
+            assign_solid_mask(request.gen_params.mask_image, resolved_width, resolved_height);
         }
 
         float denoising_strength = j.value("denoising_strength", -1.f);
@@ -240,8 +243,9 @@ static bool build_sdapi_img_gen_request(const json& j,
             }
             SDImageOwner image_owner;
             if (decode_base64_image(extra_image.get<std::string>(),
-                                    0,
-                                    0, 0,
+                                    3,
+                                    request.gen_params.width_and_height_are_set() ? request.gen_params.width : 0,
+                                    request.gen_params.width_and_height_are_set() ? request.gen_params.height : 0,
                                     image_owner)) {
                 const sd_image_t& image = image_owner.get();
                 request.gen_params.set_width_and_height_if_unset(image.width, image.height);

examples/server/routes_sdcpp.cpp

@@ -3,33 +3,12 @@
 #include <algorithm>
 #include <cmath>
 #include <filesystem>
-#include <limits>
 
 #include "async_jobs.h"
 #include "common/common.h"
-#include "common/media_io.h"
-#include "common/resource_owners.hpp"
 
 namespace fs = std::filesystem;
 
-static constexpr uint32_t k_max_upscale_dimension = 8192;
-
-static bool valid_upscale_dimensions(const sd_image_t& image, int factor, int repeats) {
-    if (image.width == 0 || image.height == 0 || factor < 1 || repeats < 1 || repeats > 4) {
-        return false;
-    }
-    uint32_t width  = image.width;
-    uint32_t height = image.height;
-    for (int i = 0; i < repeats; ++i) {
-        if (width > k_max_upscale_dimension / factor || height > k_max_upscale_dimension / factor) {
-            return false;
-        }
-        width *= factor;
-        height *= factor;
-    }
-    return true;
-}
-
 static bool parse_cache_mode(const std::string& mode_str, sd_cache_mode_t& mode_out) {
     if (mode_str == "disabled") {
         mode_out = SD_CACHE_DISABLED;
@@ -78,11 +57,11 @@ static json make_vae_tiling_json(const sd_tiling_params_t& params) {
     return {
         {"enabled", params.enabled},
         {"temporal_tiling", params.temporal_tiling},
-        {"tile_size_w", params.tile_size_w},
-        {"tile_size_h", params.tile_size_h},
+        {"tile_size_x", params.tile_size_x},
+        {"tile_size_y", params.tile_size_y},
         {"target_overlap", params.target_overlap},
-        {"rel_size_w", params.rel_size_w},
-        {"rel_size_h", params.rel_size_h},
+        {"rel_size_x", params.rel_size_x},
+        {"rel_size_y", params.rel_size_y},
         {"extra_tiling_args", params.extra_tiling_args ? params.extra_tiling_args : ""},
     };
 }
@@ -148,8 +127,7 @@ static json make_img_gen_defaults_json(const SDGenerationParams& defaults, const
         {"seed", defaults.seed},
         {"batch_count", defaults.batch_count},
         {"qwen_image_layers", defaults.qwen_image_layers},
-        {"ref_image_args", defaults.ref_image_args},
-        {"image_preprocess", defaults.image_preprocess},
+        {"auto_resize_ref_image", defaults.auto_resize_ref_image},
         {"increase_ref_index", defaults.increase_ref_index},
         {"control_strength", defaults.control_strength},
         {"ip_adapter_strength", defaults.ip_adapter_strength},
@@ -175,7 +153,6 @@ static json make_vid_gen_defaults_json(const SDGenerationParams& defaults, const
         {"strength", defaults.strength},
         {"seed", defaults.seed},
         {"video_frames", defaults.video_frames},
-        {"image_preprocess", defaults.image_preprocess},
         {"fps", defaults.fps},
         {"moe_boundary", defaults.moe_boundary},
         {"vace_strength", defaults.vace_strength},
@@ -262,59 +239,37 @@ static json make_capabilities_json(ServerRuntime& runtime) {
 
     available_upscalers.push_back({
         {"name", "None"},
-        {"model", false},
-        {"image_upscale", false},
     });
     available_upscalers.push_back({
         {"name", "Lanczos"},
-        {"model", false},
-        {"image_upscale", false},
     });
     available_upscalers.push_back({
         {"name", "Nearest"},
-        {"model", false},
-        {"image_upscale", false},
     });
     available_upscalers.push_back({
         {"name", "Latent"},
-        {"model", false},
-        {"image_upscale", false},
     });
     available_upscalers.push_back({
         {"name", "Latent (nearest)"},
-        {"model", false},
-        {"image_upscale", false},
     });
     available_upscalers.push_back({
         {"name", "Latent (nearest-exact)"},
-        {"model", false},
-        {"image_upscale", false},
     });
     available_upscalers.push_back({
         {"name", "Latent (antialiased)"},
-        {"model", false},
-        {"image_upscale", false},
     });
     available_upscalers.push_back({
         {"name", "Latent (bicubic)"},
-        {"model", false},
-        {"image_upscale", false},
     });
     available_upscalers.push_back({
         {"name", "Latent (bicubic antialiased)"},
-        {"model", false},
-        {"image_upscale", false},
     });
-    bool have_upscaler_models = false;
     {
         std::lock_guard<std::mutex> lock(*runtime.upscaler_mutex);
         for (const auto& entry : *runtime.upscaler_cache) {
             available_upscalers.push_back({
                 {"name", entry.name},
-                {"model", true},
-                {"image_upscale", entry.image_upscale_factor > 0},
             });
-            have_upscaler_models = have_upscaler_models || entry.image_upscale_factor > 0;
         }
     }
 
@@ -384,8 +339,6 @@ static json make_capabilities_json(ServerRuntime& runtime) {
                   {"max_height", 4096},
                   {"max_batch_count", 8},
                   {"max_queue_size", manager.max_pending_jobs},
-                  {"max_upscale_width", k_max_upscale_dimension},
-                  {"max_upscale_height", k_max_upscale_dimension},
     };
     result["samplers"]               = samplers;
     result["schedulers"]             = schedulers;
@@ -395,7 +348,6 @@ static json make_capabilities_json(ServerRuntime& runtime) {
     result["features_by_mode"]       = features_by_mode;
     result["loras"]                  = available_loras;
     result["upscalers"]              = available_upscalers;
-    result["upscale"]                = have_upscaler_models;
     return result;
 }
 
@@ -461,171 +413,6 @@ void register_sdcpp_api_endpoints(httplib::Server& svr, ServerRuntime& rt) {
         res.set_content(make_capabilities_json(*runtime).dump(), "application/json");
     });
 
-    svr.Post("/sdcpp/v1/upscale", [runtime](const httplib::Request& req, httplib::Response& res) {
-        try {
-            if (req.body.empty()) {
-                res.status = 400;
-                res.set_content(R"({"error":"empty body"})", "application/json");
-                return;
-            }
-            json body = json::parse(req.body);
-            if (!body.is_object()) {
-                res.status = 400;
-                res.set_content(R"({"error":"body must be an object"})", "application/json");
-                return;
-            }
-            for (const char* key : {"repeats", "tile_size", "output_compression"}) {
-                if (!body.contains(key)) {
-                    continue;
-                }
-                const auto& value = body[key];
-                const bool valid  = value.is_number_unsigned()
-                                        ? value.get<uint64_t>() <= static_cast<uint64_t>(std::numeric_limits<int>::max())
-                                        : value.is_number_integer() && value.get<int64_t>() >= std::numeric_limits<int>::min() &&
-                                             value.get<int64_t>() <= std::numeric_limits<int>::max();
-                if (!valid) {
-                    res.status = 400;
-                    res.set_content(json({{"error", std::string(key) + " must be a 32-bit integer"}}).dump(), "application/json");
-                    return;
-                }
-            }
-            ImgGenJobRequest output_options;
-            std::string error_message;
-            if (!assign_output_options(output_options,
-                                       body.value("output_format", std::string("png")),
-                                       body.value("output_compression", 100),
-                                       true,
-                                       error_message)) {
-                res.status = 400;
-                res.set_content(json({{"error", error_message}}).dump(), "application/json");
-                return;
-            }
-            const int tile_size      = std::max(32, body.value("tile_size", runtime->default_gen_params->upscale_tile_size));
-            const int repeats        = std::clamp(body.value("repeats", 1), 1, 4);
-            const std::string wanted = body.value("upscaler", std::string());
-
-            const std::string encoded = body.value("image", std::string());
-            if (encoded.empty()) {
-                res.status = 400;
-                res.set_content(R"({"error":"image is required"})", "application/json");
-                return;
-            }
-            SDImageOwner input;
-            if (!decode_base64_image(encoded, 3, 0, 0, input) || input.get().data == nullptr) {
-                res.status = 400;
-                res.set_content(R"({"error":"image could not be read"})", "application/json");
-                return;
-            }
-
-            refresh_upscaler_cache(*runtime);
-            int model_scale = 0;
-            std::string model_path;
-            std::string used_name;
-            {
-                std::lock_guard<std::mutex> lock(*runtime->upscaler_mutex);
-                for (const auto& entry : *runtime->upscaler_cache) {
-                    if (entry.image_upscale_factor > 0 && (wanted.empty() || entry.name == wanted)) {
-                        model_path  = entry.fullpath;
-                        used_name   = entry.name;
-                        model_scale = entry.image_upscale_factor;
-                        break;
-                    }
-                }
-            }
-            if (model_path.empty()) {
-                res.status = 400;
-                res.set_content(json({{"error", wanted.empty()
-                                                    ? std::string("no RGB ESRGAN upscaler models are available; "
-                                                                  "start the server with --hires-upscalers-dir")
-                                                    : "no compatible image upscaler called " + wanted}})
-                                    .dump(),
-                                "application/json");
-                return;
-            }
-
-            if (!valid_upscale_dimensions(input.get(), model_scale, repeats)) {
-                res.status = 400;
-                res.set_content(R"({"error":"upscaled dimensions must not exceed 8192 x 8192"})", "application/json");
-                return;
-            }
-
-            // One GPU: an upscale must not run while a generation is using it.
-            std::lock_guard<std::mutex> ctx_lock(*runtime->sd_ctx_mutex);
-            UpscalerCtxPtr upscaler_ctx(new_upscaler_ctx(model_path.c_str(),
-                                                         runtime->ctx_params->diffusion_conv_direct,
-                                                         runtime->ctx_params->n_threads,
-                                                         tile_size,
-                                                         runtime->ctx_params->backend.c_str(),
-                                                         runtime->ctx_params->params_backend.c_str()));
-            if (upscaler_ctx == nullptr) {
-                res.status = 500;
-                res.set_content(R"({"error":"the upscaler model could not be loaded"})", "application/json");
-                return;
-            }
-            const int factor = get_upscale_factor(upscaler_ctx.get());
-            // The model file may have changed since its metadata was cached.
-            if (!valid_upscale_dimensions(input.get(), factor, repeats)) {
-                res.status = 400;
-                res.set_content(R"({"error":"upscaled dimensions must not exceed 8192 x 8192"})", "application/json");
-                return;
-            }
-
-            SDImageOwner current(input.release());
-            for (int i = 0; i < repeats; ++i) {
-                sd_image_t* out_images = nullptr;
-                int out_count          = 0;
-                if (!upscale(upscaler_ctx.get(), current.get(), (uint32_t)factor, &out_images, &out_count) ||
-                    out_count <= 0 || out_images[0].data == nullptr) {
-                    free_sd_images(out_images, out_count);
-                    res.status = 500;
-                    res.set_content(R"({"error":"upscale failed"})", "application/json");
-                    return;
-                }
-                sd_image_t produced = out_images[0];
-                out_images[0]       = {0, 0, 0, nullptr};
-                free_sd_images(out_images, out_count);
-                current.reset(produced);
-            }
-
-            const std::string& format = output_options.output_format;
-            const int compression     = output_options.output_compression;
-            const sd_image_t result   = current.get();
-            auto image_bytes          = encode_image_to_vector(format == "jpeg"   ? EncodedImageFormat::JPEG
-                                                               : format == "webp" ? EncodedImageFormat::WEBP
-                                                                                  : EncodedImageFormat::PNG,
-                                                      result.data,
-                                                      result.width,
-                                                      result.height,
-                                                      result.channel,
-                                                      "",
-                                                      compression);
-            if (image_bytes.empty()) {
-                res.status = 500;
-                res.set_content(R"({"error":"the result could not be encoded"})", "application/json");
-                return;
-            }
-
-            json out;
-            out["upscaler"]      = used_name;
-            out["scale"]         = factor;
-            out["repeats"]       = repeats;
-            out["width"]         = result.width;
-            out["height"]        = result.height;
-            out["output_format"] = format;
-            json images          = json::array();
-            images.push_back({{"index", 0}, {"b64_json", base64_encode(image_bytes)}});
-            out["images"] = std::move(images);
-            res.set_content(out.dump(), "application/json");
-            res.status = 200;
-        } catch (const json::exception& e) {
-            res.status = 400;
-            res.set_content(json({{"error", "invalid request"}, {"message", e.what()}}).dump(), "application/json");
-        } catch (const std::exception& e) {
-            res.status = 500;
-            res.set_content(json({{"error", std::string("server_error: ") + e.what()}}).dump(), "application/json");
-        }
-    });
-
     svr.Post("/sdcpp/v1/img_gen", [runtime](const httplib::Request& req, httplib::Response& res) {
         try {
             if (req.body.empty()) {

examples/server/runtime.cpp

@@ -295,11 +295,6 @@ std::string get_lora_full_path(ServerRuntime& rt, const std::string& path) {
 
 void refresh_upscaler_cache(ServerRuntime& rt) {
     std::vector<UpscalerEntry> new_cache;
-    std::vector<UpscalerEntry> previous_cache;
-    {
-        std::lock_guard<std::mutex> lock(*rt.upscaler_mutex);
-        previous_cache = *rt.upscaler_cache;
-    }
 
     fs::path upscaler_dir = rt.ctx_params->hires_upscalers_dir;
     if (fs::exists(upscaler_dir) && fs::is_directory(upscaler_dir)) {
@@ -313,24 +308,10 @@ void refresh_upscaler_cache(ServerRuntime& rt) {
             }
 
             UpscalerEntry upscaler_entry;
-            upscaler_entry.name                 = p.stem().u8string();
-            upscaler_entry.fullpath             = fs::absolute(p).lexically_normal().u8string();
-            upscaler_entry.model_name           = "ESRGAN_4x";
-            upscaler_entry.path                 = p.filename().u8string();
-            upscaler_entry.file_size            = entry.file_size();
-            upscaler_entry.last_modified        = entry.last_write_time();
-            auto previous                       = std::find_if(previous_cache.begin(), previous_cache.end(), [&](const UpscalerEntry& cached) {
-                return cached.fullpath == upscaler_entry.fullpath &&
-                       cached.file_size == upscaler_entry.file_size &&
-                       cached.last_modified == upscaler_entry.last_modified;
-            });
-            upscaler_entry.image_upscale_factor = previous != previous_cache.end()
-                                                      ? previous->image_upscale_factor
-                                                      : get_upscaler_model_scale(upscaler_entry.fullpath.c_str());
-            if (upscaler_entry.image_upscale_factor > 0) {
-                upscaler_entry.scale      = upscaler_entry.image_upscale_factor;
-                upscaler_entry.model_name = "ESRGAN_" + std::to_string(upscaler_entry.scale) + "x";
-            }
+            upscaler_entry.name       = p.stem().u8string();
+            upscaler_entry.fullpath   = fs::absolute(p).lexically_normal().u8string();
+            upscaler_entry.model_name = "ESRGAN_4x";
+            upscaler_entry.path       = p.filename().u8string();
 
             new_cache.push_back(std::move(upscaler_entry));
         }

examples/server/runtime.h

@@ -2,7 +2,6 @@
 
 #include <algorithm>
 #include <cstdint>
-#include <filesystem>
 #include <mutex>
 #include <string>
 #include <vector>
@@ -44,9 +43,6 @@ struct UpscalerEntry {
     std::string fullpath;
     std::string model_name;
     int scale = 4;
-    int image_upscale_factor = 0;
-    uintmax_t file_size      = 0;
-    std::filesystem::file_time_type last_modified;
 };
 
 struct ServerRuntime {

ggml

@@ -1 +1 @@
-Subproject commit 4bf5f6000653b7881d00963cd6ddb665ccd62a8d
+Subproject commit e20c3a14aa70ee84ca58499814206dd08d8026bc

include/stable-diffusion.h

@@ -79,7 +79,6 @@ enum scheduler_t {
     FLUX2_SCHEDULER,
     FLUX_SCHEDULER,
     BETA_SCHEDULER,
-    LLADA_IMAGE_SCHEDULER,
     SCHEDULER_COUNT
 };
 
@@ -173,13 +172,11 @@ enum lora_apply_mode_t {
 typedef struct {
     bool enabled;
     bool temporal_tiling;
-    // Spatial tile dimensions in image pixels for both encode and decode; 0 uses 256.
-    int tile_size_w;
-    int tile_size_h;
+    int tile_size_x;
+    int tile_size_y;
     float target_overlap;
-    // Positive values override tile_size: <= 1 is a dimension fraction, > 1 a target tile count.
-    float rel_size_w;
-    float rel_size_h;
+    float rel_size_x;
+    float rel_size_y;
     const char* extra_tiling_args;
 } sd_tiling_params_t;
 
@@ -248,8 +245,6 @@ typedef struct {
     float linear_scale;              // Override linear input scaling; 0 keeps the model default
     float attn_scale;                // Override flash-attention K/V scaling; 0 keeps the model default
     const char* tokenizer;           // tokenizer.json path or main=FILE,clip-l=FILE,clip-g=FILE assignments; required for PiD and Lens
-    bool sage_attn;
-    int conditioning_cache_size;  // Maximum cached conditioning entries per context; 0 disables caching (default: 4)
 } sd_ctx_params_t;
 
 typedef struct {
@@ -266,11 +261,6 @@ typedef struct {
     uint8_t* data;
 } sd_image_t;
 
-typedef struct {
-    // Semicolon-separated target=...,key=value rules. NULL preserves defaults.
-    const char* rules;
-} sd_image_preprocess_params_t;
-
 typedef struct {
     sd_image_t* frames;
     int frame_count;
@@ -418,7 +408,6 @@ typedef struct {
     int qwen_image_layers;
     bool circular_x;
     bool circular_y;
-    sd_image_preprocess_params_t image_preprocess;
 } sd_img_gen_params_t;
 
 typedef struct {
@@ -452,7 +441,6 @@ typedef struct {
     sd_hires_params_t hires;
     bool circular_x;
     bool circular_y;
-    sd_image_preprocess_params_t image_preprocess;
 } sd_vid_gen_params_t;
 
 typedef struct sd_ctx_t sd_ctx_t;
@@ -560,8 +548,6 @@ SD_API bool upscale(upscaler_ctx_t* upscaler_ctx,
                     int* num_images_out);
 
 SD_API int get_upscale_factor(upscaler_ctx_t* upscaler_ctx);
-// Reads model metadata only; returns 0 if the file is not a recognized RGB ESRGAN model.
-SD_API int get_upscaler_model_scale(const char* model_path);
 
 typedef struct adetailer_ctx_t adetailer_ctx_t;
 

src/conditioning/conditioner.hpp

@@ -14,7 +14,6 @@
 #include "core/util.h"
 #include "model/diffusion/model.hpp"
 #include "model/te/clip.hpp"
-#include "model/te/llada_image_te.hpp"
 #include "model/te/llm.hpp"
 #include "model/te/t5.hpp"
 #include "model_loader.h"
@@ -1979,8 +1978,7 @@ struct LLMEmbedder : public Conditioner {
             arch = LLM::LLMArch::GPT_OSS_20B;
         } else if (sd_version_is_pid(version)) {
             arch = LLM::LLMArch::GEMMA2_2B;
-        } else if (version == VERSION_QWEN_IMAGE_2_1 ||
-                   sd_version_is_lingbot_video(version) ||
+        } else if (sd_version_is_lingbot_video(version) ||
                    sd_version_is_ideogram4(version) ||
                    sd_version_is_boogu_image(version) ||
                    sd_version_is_sefi_image(version) ||
@@ -2219,10 +2217,7 @@ struct LLMEmbedder : public Conditioner {
                                           false,
                                           deepstack_image_embeds,
                                           image_grids);
-        if (hidden_states.empty()) {
-            LOG_ERROR("LLM prompt encoding failed");
-            return {};
-        }
+        GGML_ASSERT(!hidden_states.empty());
         hidden_states = apply_token_weights(std::move(hidden_states), weights);
         GGML_ASSERT(hidden_states.shape()[1] > prompt_template_encode_start_idx);
 
@@ -2552,67 +2547,6 @@ struct LLMEmbedder : public Conditioner {
             prompt += conditioner_params.text;
             prompt_attn_range = {0, 0};
             prompt += "<|im_end|>\n<|im_start|>assistant\n";
-        } else if (version == VERSION_QWEN_IMAGE_2_1) {
-            if (!llm->enable_vision && conditioner_params.ref_images != nullptr && !conditioner_params.ref_images->empty()) {
-                LOG_ERROR("Qwen Image 2.1 editing requires Qwen3-VL vision weights; provide --llm_vision or a combined encoder");
-                return {};
-            }
-            prompt = "<|im_start|>system\nComprehend and analyze the provided prompt.<|im_end|>\n";
-            std::vector<int> system_tokens;
-            if (!tokenizer->encode(prompt, system_tokens, nullptr)) {
-                return {};
-            }
-            prompt_template_encode_start_idx = static_cast<int>(system_tokens.size());
-            out_layers                       = {static_cast<int>(llm->config.num_layers)};
-            prompt += "<|im_start|>user\n";
-            if (llm->enable_vision && conditioner_params.ref_images != nullptr) {
-                for (size_t i = 0; i < conditioner_params.ref_images->size(); ++i) {
-                    const auto& image = (*conditioner_params.ref_images)[i];
-                    int64_t width     = image.shape()[0];
-                    int64_t height    = image.shape()[1];
-                    int64_t pixels    = width * height;
-                    if (width % 32 != 0 || height % 32 != 0) {
-                        LOG_ERROR("Qwen Image 2.1 reference dimensions must be multiples of 32");
-                        return {};
-                    }
-                    auto rgb = sd::Tensor<float>({width, height, 3, 1});
-                    for (int64_t p = 0; p < pixels; ++p) {
-                        float alpha = image.shape()[2] == 4 ? image[p + 3 * pixels] : 1.f;
-                        for (int c = 0; c < 3; ++c) {
-                            rgb[p + c * pixels] = 2.f * (image[p + c * pixels] * alpha + 1.f - alpha) - 1.f;
-                        }
-                    }
-                    auto outputs = llm->encode_image_outputs(n_threads, rgb, false);
-                    if (outputs.empty()) {
-                        return {};
-                    }
-                    prompt += (i == 0 ? "" : " ") + std::string("<image") + std::to_string(i + 1) + "><|vision_start|>";
-                    std::vector<int> prefix_tokens;
-                    if (!tokenizer->encode(prompt, prefix_tokens, nullptr)) {
-                        return {};
-                    }
-                    int index = static_cast<int>(prefix_tokens.size());
-                    int count = static_cast<int>(outputs[0].shape()[1]);
-                    image_embeds.emplace_back(index, std::move(outputs[0]));
-                    if (deepstack_image_embeds.empty()) {
-                        deepstack_image_embeds.resize(outputs.size() - 1);
-                    }
-                    for (size_t layer = 1; layer < outputs.size(); ++layer) {
-                        deepstack_image_embeds[layer - 1].emplace_back(index, std::move(outputs[layer]));
-                    }
-                    image_grids.push_back({index, count,
-                                           static_cast<int>(height) / llm->config.vision.patch_size,
-                                           static_cast<int>(width) / llm->config.vision.patch_size});
-                    for (int j = 0; j < count; ++j) {
-                        prompt += "<|image_pad|>";
-                    }
-                    prompt += "<|vision_end|>";
-                }
-            }
-            prompt_attn_range.first = static_cast<int>(prompt.size());
-            prompt += conditioner_params.text.empty() ? " " : conditioner_params.text;
-            prompt_attn_range.second = static_cast<int>(prompt.size());
-            prompt += "<|im_end|>\n<|im_start|>assistant\n";
         } else if (sd_version_is_qwen_image(version) || sd_version_is_mage_flow(version)) {
             if (llm->enable_vision && conditioner_params.ref_images != nullptr && !conditioner_params.ref_images->empty()) {
                 LOG_INFO("%s", sd_version_is_mage_flow(version) ? "MageFlowEditPipeline" : "QwenImageEditPlusPipeline");
@@ -3140,21 +3074,6 @@ struct LLMEmbedder : public Conditioner {
         SDCondition result;
         result.c_crossattn        = std::move(hidden_states);
         result.extra_c_crossattns = std::move(extra_hidden_states_vec);
-        if (version == VERSION_QWEN_IMAGE_2_1) {
-            auto slots = sd::Tensor<int32_t>::zeros({result.c_crossattn.shape()[1]});
-            for (size_t i = 0; i < image_embeds.size(); ++i) {
-                int64_t begin = image_embeds[i].first - prompt_template_encode_start_idx;
-                int64_t end   = begin + image_embeds[i].second.shape()[1];
-                if (begin < 0 || end > slots.numel()) {
-                    LOG_ERROR("Qwen Image 2.1 image slots exceed the encoded prompt");
-                    return {};
-                }
-                for (int64_t j = begin; j < end; ++j) {
-                    slots[j] = static_cast<int32_t>(i + 1);
-                }
-            }
-            result.c_token_types = std::move(slots);
-        }
         if (sd_version_is_minimax_h3(version)) {
             std::vector<int32_t> tags(static_cast<size_t>(result.c_crossattn.shape()[1]), 1);
             for (const auto& [index, image_embed] : image_embeds) {
@@ -3166,7 +3085,6 @@ struct LLMEmbedder : public Conditioner {
             int64_t tag_count    = static_cast<int64_t>(tags.size());
             result.c_token_types = sd::Tensor<int32_t>({tag_count}, std::move(tags));
         }
-
         return result;
     }
 };
@@ -3241,214 +3159,6 @@ struct LTXAVTextProjectionRunner : public GGMLRunner {
     }
 };
 
-// LLaDA-Image's text path is a three-stage pipeline rather than a single encoder pass:
-// the token embeddings feed a QueryFormer whose 256 queries are appended to the backbone
-// input, and the backbone's final hidden states are projected to the denoiser's caption dim.
-// Ref: LLaDAImagePipeline._encode_text.
-struct LLaDAImageEmbedder : public Conditioner {
-    std::shared_ptr<Tokenizer> tokenizer;
-    std::shared_ptr<LLM::LLMRunner> llm;
-    std::shared_ptr<LLaDAImageTE::QueryFormerRunner> query_former;
-    std::shared_ptr<LLaDAImageTE::TextProjectionRunner> text_projection;
-    std::shared_ptr<LLaDAImageTE::SigVQRunner> sigvq;
-
-    std::string llm_prefix;
-    std::string query_former_prefix;
-    std::string text_projection_prefix;
-    std::string sigvq_prefix;
-
-    LLaDAImageEmbedder(ggml_backend_t backend,
-                       const String2TensorStorage& tensor_storage_map      = {},
-                       const std::string& llm_prefix                       = "text_encoders.llm",
-                       const std::string& query_former_prefix              = "queryformer",
-                       const std::string& text_projection_prefix           = "text_projection",
-                       const std::string& sigvq_prefix                     = "sigvq",
-                       std::shared_ptr<RunnerWeightManager> weight_manager = nullptr,
-                       const TokenizerConfig& tokenizers                   = {})
-        : llm_prefix(llm_prefix),
-          query_former_prefix(query_former_prefix),
-          text_projection_prefix(text_projection_prefix),
-          sigvq_prefix(sigvq_prefix) {
-        if (!tokenizers.has(TokenizerConfig::MAIN)) {
-            throw std::runtime_error("LLaDA-Image requires an external LLaDA2 tokenizer.json; pass --tokenizer FILE or set sd_ctx_params_t::tokenizer");
-        }
-        llm = std::make_shared<LLM::LLMRunner>(LLM::LLMArch::LLADA2_MOE,
-                                               backend,
-                                               tensor_storage_map,
-                                               llm_prefix,
-                                               false,
-                                               weight_manager);
-        // <|endoftext|> doubles as the pad token in LLaDA2's tokenizer.json.
-        tokenizer       = tokenizers.create(TokenizerConfig::MAIN, llm->config.vocab_size, 156892);
-        query_former    = std::make_shared<LLaDAImageTE::QueryFormerRunner>(backend,
-                                                                         tensor_storage_map,
-                                                                         query_former_prefix,
-                                                                         weight_manager);
-        text_projection = std::make_shared<LLaDAImageTE::TextProjectionRunner>(backend,
-                                                                               tensor_storage_map,
-                                                                               text_projection_prefix,
-                                                                               weight_manager);
-
-        // SigVQ is only present when the user supplies the editing weights.
-        for (const auto& [name, _] : tensor_storage_map) {
-            if (starts_with(name, sigvq_prefix + ".")) {
-                sigvq = std::make_shared<LLaDAImageTE::SigVQRunner>(backend,
-                                                                    tensor_storage_map,
-                                                                    sigvq_prefix,
-                                                                    weight_manager);
-                break;
-            }
-        }
-    }
-
-    void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        llm->get_param_tensors(tensors, llm_prefix);
-        query_former->get_param_tensors(tensors, query_former_prefix);
-        text_projection->get_param_tensors(tensors, text_projection_prefix);
-        if (sigvq != nullptr) {
-            sigvq->get_param_tensors(tensors, sigvq_prefix);
-        }
-    }
-
-    void get_param_tensor_ops(std::map<ggml_tensor*, enum ggml_op>& tensor_ops) override {
-        llm->get_param_tensor_ops(tensor_ops);
-    }
-
-    void set_flash_attention_enabled(bool enabled) override {
-        llm->set_flash_attention_enabled(enabled);
-        query_former->set_flash_attention_enabled(enabled);
-        text_projection->set_flash_attention_enabled(enabled);
-        if (sigvq != nullptr) {
-            sigvq->set_flash_attention_enabled(enabled);
-        }
-    }
-
-    void set_max_graph_vram_bytes(size_t max_vram_bytes) override {
-        llm->set_max_graph_vram_bytes(max_vram_bytes);
-        query_former->set_max_graph_vram_bytes(max_vram_bytes);
-        text_projection->set_max_graph_vram_bytes(max_vram_bytes);
-        if (sigvq != nullptr) {
-            sigvq->set_max_graph_vram_bytes(max_vram_bytes);
-        }
-    }
-
-    void set_runtime_backends(const std::vector<ggml_backend_t>& backends) override {
-        llm->set_runtime_backends(backends);
-    }
-
-    void set_graph_cut_layer_split_enabled(bool enabled) override {
-        llm->set_graph_cut_layer_split_enabled(enabled);
-    }
-
-    void set_graph_cut_layer_split_backend_vram_limits(const std::vector<size_t>& limits) override {
-        llm->set_graph_cut_layer_split_backend_vram_limits(limits);
-    }
-
-    void get_layer_split_param_tensors(std::map<std::string, ggml_tensor*>& tensors) override {
-        llm->get_param_tensors(tensors, llm_prefix);
-    }
-
-    void set_weight_adapter(const std::shared_ptr<WeightAdapter>& adapter) override {
-        llm->set_weight_adapter(adapter);
-        query_former->set_weight_adapter(adapter);
-        text_projection->set_weight_adapter(adapter);
-        if (sigvq != nullptr) {
-            sigvq->set_weight_adapter(adapter);
-        }
-    }
-
-    void runner_end() override {
-        llm->runner_end();
-        query_former->runner_end();
-        text_projection->runner_end();
-        if (sigvq != nullptr) {
-            sigvq->runner_end();
-        }
-    }
-
-    SDCondition get_learned_condition(int n_threads,
-                                      const ConditionerParams& conditioner_params) override {
-        const int64_t num_queries = 256;
-        const bool has_ref_images = conditioner_params.ref_images != nullptr && !conditioner_params.ref_images->empty();
-        if (has_ref_images && sigvq == nullptr) {
-            LOG_ERROR("LLaDA-Image editing requires connectors with SigVQ weights");
-            return {};
-        }
-
-        std::string text = conditioner_params.text;
-        while (!text.empty() && std::isspace(static_cast<unsigned char>(text.front()))) {
-            text.erase(text.begin());
-        }
-        while (!text.empty() && std::isspace(static_cast<unsigned char>(text.back()))) {
-            text.pop_back();
-        }
-        std::string prompt = text.empty()
-                                 ? "<role>HUMAN</role> Generate an image.\n<role>ASSISTANT</role>\n<IMAGE1>"
-                                 : "<role>HUMAN</role> Generate an image: " + text + "\n<role>ASSISTANT</role>\n<IMAGE1>";
-
-        std::vector<int> tokens;
-        if (!tokenizer->encode(prompt, tokens, nullptr)) {
-            return {};
-        }
-        int64_t n_text = static_cast<int64_t>(tokens.size());
-        GGML_ASSERT(n_text > 0);
-
-        sd::Tensor<int32_t> text_ids({n_text}, std::vector<int32_t>(tokens.begin(), tokens.end()));
-        auto inputs_embeds = llm->compute_input_embeds(n_threads, text_ids);
-        auto query_embeds  = query_former->compute(n_threads, inputs_embeds);
-
-        // splice_image_embeds() replaces tokens in place, so the query slots have to exist in
-        // input_ids; their ids are irrelevant because the embeddings are overwritten.
-        std::vector<int32_t> padded(tokens.begin(), tokens.end());
-        padded.resize(static_cast<size_t>(n_text + num_queries), tokenizer->PAD_TOKEN_ID);
-        int64_t n_total = static_cast<int64_t>(padded.size());
-        sd::Tensor<int32_t> input_ids({n_total}, padded);
-
-        // Bidirectional everywhere except that the text tokens must not see the appended
-        // queries, matching backbone_attention_mask[:, :, :text_length, text_length:] = min.
-        const float mask_min = std::numeric_limits<float>::lowest() / 4.0f;
-        sd::Tensor<float> attention_mask({n_total, n_total});
-        for (int64_t i1 = 0; i1 < n_total; ++i1) {
-            for (int64_t i0 = 0; i0 < n_total; ++i0) {
-                float value                       = (i1 < n_text && i0 >= n_text) ? mask_min : 0.0f;
-                attention_mask[i0 + n_total * i1] = value;
-            }
-        }
-
-        LLM::ImageEmbeds image_embeds;
-        image_embeds.emplace_back(static_cast<int>(n_text), query_embeds);
-
-        std::set<int> out_layers = {static_cast<int>(llm->config.num_layers) + 1};
-        auto hidden_states       = llm->compute(n_threads,
-                                                input_ids,
-                                                attention_mask,
-                                                image_embeds,
-                                                out_layers);
-
-        SDCondition result;
-        result.c_crossattn = text_projection->compute(n_threads, hidden_states);
-
-        // Editing: SigVQ sees the reference at half the output resolution, as in
-        // LLaDAImagePipeline._encode_source_image.
-        if (has_ref_images) {
-            const auto& ref = conditioner_params.ref_images->front();
-            auto resized    = sd::ops::interpolate(ref,
-                                                   {conditioner_params.width / 2,
-                                                    conditioner_params.height / 2,
-                                                    ref.shape()[2],
-                                                    ref.shape()[3]},
-                                                   sd::ops::InterpolateMode::Bilinear);
-            resized         = resized * 2.f - 1.f;
-            auto semantic   = sigvq->compute(n_threads, resized);
-            if (semantic.empty()) {
-                return {};
-            }
-            result.extra_c_crossattns.push_back(std::move(semantic));
-        }
-        return result;
-    }
-};
-
 struct LTXAVEmbedder : public Conditioner {
     static constexpr int64_t kHiddenSize = 3840;
     static constexpr int64_t kNumStates  = 49;

src/conditioning/conditioning_cache.h

@@ -1,107 +0,0 @@
-#ifndef __SD_CONDITIONING_CONDITIONING_CACHE_H__
-#define __SD_CONDITIONING_CONDITIONING_CACHE_H__
-
-#include <algorithm>
-#include <list>
-#include <tuple>
-
-#include "conditioning/conditioner.hpp"
-
-class ConditioningCache {
-    struct Entry {
-        ConditionerParams params;
-        std::vector<sd::Tensor<float>> ref_images;
-        std::vector<MiniMaxH3PresentationItem> references;
-        SDCondition condition;
-
-        Entry(const ConditionerParams& input, const SDCondition& output)
-            : params(input), condition(output) {
-            // Request-owned reference pointers must not outlive the request.
-            if (input.ref_images != nullptr) {
-                ref_images        = *input.ref_images;
-                params.ref_images = &ref_images;
-            }
-            if (input.minimax_h3_references != nullptr) {
-                references                   = *input.minimax_h3_references;
-                params.minimax_h3_references = &references;
-            }
-        }
-
-        Entry(const Entry&)            = delete;
-        Entry& operator=(const Entry&) = delete;
-    };
-
-    size_t capacity_ = 4;
-    std::list<Entry> entries_;
-
-    static bool same_images(const std::vector<sd::Tensor<float>>& a,
-                            const std::vector<sd::Tensor<float>>& b) {
-        return std::equal(a.begin(), a.end(), b.begin(), b.end(),
-                          [](const sd::Tensor<float>& x, const sd::Tensor<float>& y) {
-                              return x.shape() == y.shape() && x.values() == y.values();
-                          });
-    }
-
-    static bool same_params(const ConditionerParams& a, const ConditionerParams& b) {
-        const auto fields = [](const ConditionerParams& p) {
-            const auto& r = p.ref_image_params;
-            return std::tie(p.text, p.clip_skip, p.width, p.height, p.zero_out_masked,
-                            r.pass_to_vlm, r.pass_to_dit, r.ref_index_mode,
-                            r.force_ref_timestep_zero, r.resize_before_vae, r.vae_input_max_pixels,
-                            r.vlm_resize_mode, r.vlm_min_size, r.vlm_max_size, r.resize_vae_to_target);
-        };
-        if (fields(a) != fields(b) ||
-            (a.ref_images == nullptr) != (b.ref_images == nullptr) ||
-            (a.minimax_h3_references == nullptr) != (b.minimax_h3_references == nullptr)) {
-            return false;
-        }
-        if (a.ref_images != nullptr && !same_images(*a.ref_images, *b.ref_images)) {
-            return false;
-        }
-        if (a.minimax_h3_references != nullptr &&
-            !std::equal(a.minimax_h3_references->begin(), a.minimax_h3_references->end(),
-                        b.minimax_h3_references->begin(), b.minimax_h3_references->end(),
-                        [](const MiniMaxH3PresentationItem& x, const MiniMaxH3PresentationItem& y) {
-                            return x.kind == y.kind && x.timestamps == y.timestamps && same_images(x.frames, y.frames);
-                        })) {
-            return false;
-        }
-        return true;
-    }
-
-public:
-    void set_capacity(size_t capacity) {
-        capacity_ = capacity;
-        while (entries_.size() > capacity_) {
-            entries_.pop_back();
-        }
-    }
-
-    void clear() {
-        entries_.clear();
-    }
-
-    SDCondition get(Conditioner& conditioner, int n_threads, const ConditionerParams& params) {
-        if (capacity_ == 0) {
-            return conditioner.get_learned_condition(n_threads, params);
-        }
-        for (auto it = entries_.begin(); it != entries_.end(); ++it) {
-            if (same_params(it->params, params)) {
-                entries_.splice(entries_.begin(), entries_, it);
-                LOG_INFO("conditioning cache hit");
-                return entries_.front().condition;
-            }
-        }
-        auto condition = conditioner.get_learned_condition(n_threads, params);
-        if (!condition.empty()) {
-            if (entries_.size() == capacity_) {
-                entries_.pop_back();
-            }
-            entries_.emplace_front(params, condition);
-            LOG_VERBOSE("conditioning cache stored (%zu/%zu)", entries_.size(), capacity_);
-        }
-        return condition;
-    }
-};
-
-#endif  // __SD_CONDITIONING_CONDITIONING_CACHE_H__

src/convert.cpp

@@ -362,9 +362,6 @@ bool convert_with_components(const char* model_path,
                              const char* tensor_type_rules,
                              bool convert_name,
                              int n_threads) {
-    if (!validate_tensor_types(output_type, tensor_type_rules)) {
-        return false;
-    }
     ModelLoader model_loader;
     bool loaded_any = false;
 

src/core/backend_fit.cpp

@@ -102,7 +102,7 @@ namespace sd::backend_fit {
         for (const auto& [name, stored_tensor] : loader.get_tensor_storage_map()) {
             TensorStorage ts = stored_tensor;
             ComponentKind kind;
-            if (!classify_tensor(ts.name, kind)) {
+            if (is_unused_tensor(ts.name) || !classify_tensor(ts.name, kind)) {
                 continue;
             }
             if (ts.expected_type != GGML_TYPE_COUNT) {
@@ -478,48 +478,27 @@ namespace sd::backend_fit {
         return true;
     }
 
-    bool prepare_vae_decode_retry_tiling(sd_tiling_params_t& tiling_params,
-                                         bool prefer_temporal_tiling,
-                                         ggml_status status,
-                                         int latent_tile_size_w,
-                                         int latent_tile_size_h,
-                                         int scale_factor) {
-        // Execution failures can leave the device unusable; tiling only helps with allocation failures.
-        if (status != GGML_STATUS_ALLOC_FAILED) {
-            return false;
-        }
+    bool prepare_vae_decode_retry_tiling(sd_tiling_params_t& tiling_params, bool prefer_temporal_tiling) {
         const char* retry_mode = nullptr;
         if (prefer_temporal_tiling && !tiling_params.temporal_tiling) {
             tiling_params.temporal_tiling = true;
             retry_mode                    = tiling_params.enabled ? "spatial+temporal" : "temporal";
+        } else if (!tiling_params.enabled) {
+            tiling_params.enabled    = true;
+            tiling_params.rel_size_x = 0.5f;
+            tiling_params.rel_size_y = 0.5f;
+            if (tiling_params.tile_size_x <= 0) {
+                tiling_params.tile_size_x = 256;
+            }
+            if (tiling_params.tile_size_y <= 0) {
+                tiling_params.tile_size_y = 256;
+            }
+            retry_mode = tiling_params.temporal_tiling ? "spatial+temporal" : "spatial";
         } else {
-            if (latent_tile_size_w <= 0 || latent_tile_size_h <= 0 || scale_factor <= 0) {
-                return false;
-            }
-            auto smaller_tile = [&](int size) {
-                int next_size = size / 2;
-                if (!tiling_params.enabled) {
-                    next_size = std::min(next_size, 256 / scale_factor);
-                }
-                return std::min(size, std::max(4, next_size));
-            };
-            const int tile_size_w = smaller_tile(latent_tile_size_w);
-            const int tile_size_h = smaller_tile(latent_tile_size_h);
-            if (tile_size_w == latent_tile_size_w && tile_size_h == latent_tile_size_h) {
-                return false;
-            }
-            tiling_params.enabled     = true;
-            tiling_params.rel_size_w  = 0.0f;
-            tiling_params.rel_size_h  = 0.0f;
-            tiling_params.tile_size_w = tile_size_w * scale_factor;
-            tiling_params.tile_size_h = tile_size_h * scale_factor;
-            retry_mode                = tiling_params.temporal_tiling ? "spatial+temporal" : "spatial";
-            LOG_WARN("Reducing VAE decode tiles from %dx%d to %dx%d image pixels",
-                     latent_tile_size_w * scale_factor, latent_tile_size_h * scale_factor,
-                     tiling_params.tile_size_w, tiling_params.tile_size_h);
+            return false;
         }
 
-        LOG_WARN("VAE decode ran out of memory; retrying with %s tiling",
+        LOG_WARN("VAE decode failed (likely out of memory); retrying with %s tiling",
                  retry_mode);
         return true;
     }

src/core/backend_fit.h

@@ -16,11 +16,7 @@ namespace sd::backend_fit {
                               std::string& params_spec);
 
     bool prepare_vae_decode_retry_tiling(sd_tiling_params_t& tiling_params,
-                                         bool prefer_temporal_tiling,
-                                         ggml_status status,
-                                         int latent_tile_size_w,
-                                         int latent_tile_size_h,
-                                         int scale_factor);
+                                         bool prefer_temporal_tiling);
 
 }  // namespace sd::backend_fit
 

src/core/compute_workspace.cpp

@@ -11,7 +11,7 @@
 #include "core/ggml_graph_cut.h"
 #include "core/util.h"
 #include "ggml-cpu.h"
-#include "ggml-impl.h"
+#include "ggml/src/ggml-impl.h"
 
 namespace sd {
     ComputeWorkspace::~ComputeWorkspace() {

src/core/ggml_extend.cpp

@@ -1,7 +1,6 @@
 #include "core/ggml_extend.h"
 
 #include <cmath>
-#include <stdexcept>
 #include <utility>
 
 #include "core/ggml_extend_backend.h"
@@ -248,7 +247,6 @@ ggml_tensor* ggml_ext_linear_i8_tensorwise(ggml_context* ctx,
                                            ggml_tensor* b,
                                            int convrot_group_size,
                                            float scale) {
-#ifndef SD_USE_UPSTREAM_GGML
     GGML_ASSERT(x->type == GGML_TYPE_F32 || (x->type == GGML_TYPE_I8 && scale == 1.f));
     if (scale != 1.f) {
         x = ggml_ext_scale(ctx, x, scale);
@@ -272,16 +270,6 @@ ggml_tensor* ggml_ext_linear_i8_tensorwise(ggml_context* ctx,
         }
     }
     return x;
-#else
-    GGML_UNUSED(ctx);
-    GGML_UNUSED(x);
-    GGML_UNUSED(w);
-    GGML_UNUSED(weight_scale);
-    GGML_UNUSED(b);
-    GGML_UNUSED(convrot_group_size);
-    GGML_UNUSED(scale);
-    throw std::runtime_error("INT8 tensorwise/convrot is not supported by this ggml build");
-#endif
 }
 
 ggml_tensor* ggml_ext_pad_ext(ggml_context* ctx,
@@ -464,17 +452,8 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
                               int d0,
                               int d1,
                               int d2,
-                              bool force_prec_f32,
-                              bool direct,
-                              float scale) {
-    if (scale != 1.f) {
-        x = ggml_ext_scale(ctx, x, scale);
-    }
-    if (direct) {
-        int64_t OC = w->ne[3] / IC;
-        int64_t N  = x->ne[3] / IC;
-        x          = ggml_conv_3d_direct(ctx, w, x, s0, s1, s2, p0, p1, p2, d0, d1, d2, (int)IC, (int)N, (int)OC);
-    } else if (force_prec_f32) {
+                              bool force_prec_f32) {
+    if (force_prec_f32) {
         ggml_tensor* im2col = ggml_im2col_3d(ctx, w, x, IC, s0, s1, s2, p0, p1, p2, d0, d1, d2, w->type);
 
         int64_t OC = w->ne[3] / IC;
@@ -506,9 +485,6 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
         }
     }
 
-    if (scale != 1.f) {
-        x = ggml_ext_scale(ctx, x, 1.f / scale);
-    }
     if (b != nullptr) {
         b = ggml_reshape_4d(ctx, b, 1, 1, 1, b->ne[0]);  // [OC, 1, 1, 1]
         x = ggml_add_inplace(ctx, x, b);
@@ -629,12 +605,7 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
                                     ggml_tensor* mask,
                                     bool skip_reshape,
                                     bool flash_attn,
-                                    float kv_scale,
-                                    bool sage_attn,
-                                    bool* used_flash_attn) {  // avoid overflow
-    if (used_flash_attn != nullptr) {
-        *used_flash_attn = false;
-    }
+                                    float kv_scale) {  // avoid overflow
     int64_t L_q;
     int64_t L_k;
     int64_t C;
@@ -672,14 +643,6 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
     ggml_tensor* kqv = nullptr;
 
     auto build_kqv = [&](ggml_tensor* q_in, ggml_tensor* k_in, ggml_tensor* v_in, ggml_tensor* mask_in) -> ggml_tensor* {
-        const bool pad_head = d_head > 0 && d_head < 64 && q_in->ne[0] == d_head && k_in->ne[0] == d_head &&
-                              q_in->type == GGML_TYPE_F32 && k_in->type == GGML_TYPE_F32 &&
-                              v_in->type == GGML_TYPE_F32 && sd_backend_supports_cuda_mma(backend);
-        if (pad_head) {
-            // CUDA FA MMA starts at 64 channels; keep the original head's attention scale.
-            q_in = ggml_pad(ctx, q_in, 64 - d_head, 0, 0, 0);
-            k_in = ggml_pad(ctx, k_in, 64 - d_head, 0, 0, 0);
-        }
         if (kv_scale != 1.0f) {
             k_in = ggml_ext_scale(ctx, k_in, kv_scale);
         }
@@ -687,9 +650,6 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
 
         v_in = ggml_ext_cont(ctx, ggml_permute(ctx, v_in, 0, 2, 1, 3));
         v_in = ggml_reshape_3d(ctx, v_in, d_head, L_k, n_kv_head * N);
-        if (pad_head) {
-            v_in = ggml_pad(ctx, v_in, 64 - d_head, 0, 0, 0);
-        }
         if (kv_scale != 1.0f) {
             v_in = ggml_ext_scale(ctx, v_in, kv_scale);
         }
@@ -719,43 +679,10 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         if (kv_scale != 1.0f) {
             out = ggml_ext_scale(ctx, out, 1.0f / kv_scale);
         }
-        if (pad_head) {
-            out = ggml_ext_slice(ctx, out, 0, 0, d_head);
-        }
         return out;
     };
 
-#ifndef SD_USE_UPSTREAM_GGML
-    if (sage_attn && mask == nullptr && d_head > 0 && d_head <= 128) {
-        auto q_in                 = ggml_reshape_4d(ctx, ggml_ext_cont(ctx, q->type == GGML_TYPE_F32 ? q : ggml_cast(ctx, q, GGML_TYPE_F32)), d_head, L_q, n_head, N);
-        auto k_in                 = ggml_reshape_4d(ctx, ggml_ext_cont(ctx, k->type == GGML_TYPE_F32 ? k : ggml_cast(ctx, k, GGML_TYPE_F32)), d_head, L_k, n_kv_head, N);
-        auto v_in                 = ggml_ext_cont(ctx, ggml_permute(ctx, v, 0, 2, 1, 3));
-        const int64_t padded_head = d_head <= 64 ? 64 : 128;
-        if ((padded_head != d_head || kv_scale != 1.0f) && v_in->type != GGML_TYPE_F32) {
-            v_in = ggml_cast(ctx, v_in, GGML_TYPE_F32);
-        }
-        if (padded_head != d_head) {
-            // Keep the original head's softmax scale when padding for the CUDA kernel.
-            q_in = ggml_pad(ctx, q_in, padded_head - d_head, 0, 0, 0);
-            k_in = ggml_pad(ctx, k_in, padded_head - d_head, 0, 0, 0);
-            v_in = ggml_pad(ctx, v_in, padded_head - d_head, 0, 0, 0);
-        }
-        if (kv_scale != 1.0f) {
-            k_in = ggml_ext_scale(ctx, k_in, kv_scale);
-            v_in = ggml_ext_scale(ctx, v_in, kv_scale);
-        }
-        v_in     = ggml_cast(ctx, v_in, GGML_TYPE_F16);
-        auto out = ggml_sage_attn(ctx, q_in, k_in, v_in, scale / kv_scale, GGML_SAGE_ATTN_AUTO);
-        if (ggml_backend_supports_op(backend, out)) {
-            kqv = kv_scale != 1.0f ? ggml_ext_scale(ctx, out, 1.0f / kv_scale) : out;
-            if (padded_head != d_head) {
-                kqv = ggml_ext_slice(ctx, kqv, 0, 0, d_head);
-            }
-        }
-    }
-#endif
-
-    if (kqv == nullptr && (flash_attn || sage_attn)) {
+    if (flash_attn) {
         // LOG_VERBOSE("attention_ext L_q:%d L_k:%d n_head:%d C:%d d_head:%d N:%d", L_q, L_k, n_head, C, d_head, N);
         bool can_use_flash_attn = true;
         if (mask != nullptr) {
@@ -766,9 +693,6 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         if (can_use_flash_attn) {
             kqv = build_kqv(q, k, v, mask);
             if (kqv != nullptr) {
-                if (used_flash_attn != nullptr) {
-                    *used_flash_attn = true;
-                }
                 kqv = ggml_view_4d(ctx,
                                    kqv,
                                    d_head,

src/core/ggml_extend.h

@@ -153,9 +153,7 @@ ggml_tensor* ggml_ext_conv_3d(ggml_context* ctx,
                               int d0              = 1,
                               int d1              = 1,
                               int d2              = 1,
-                              bool force_prec_f32 = false,
-                              bool direct         = false,
-                              float scale         = 1.f);
+                              bool force_prec_f32 = false);
 
 // w: [OC，IC, KD, 1 * 1]
 // x: [N, IC, ID, IH*IW]
@@ -218,12 +216,10 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
                                     ggml_tensor* k,
                                     ggml_tensor* v,
                                     int64_t n_head,
-                                    ggml_tensor* mask     = nullptr,
-                                    bool skip_reshape     = false,
-                                    bool flash_attn       = false,
-                                    float kv_scale        = 1.0f,
-                                    bool sage_attn        = false,
-                                    bool* used_flash_attn = nullptr);
+                                    ggml_tensor* mask = nullptr,
+                                    bool skip_reshape = false,
+                                    bool flash_attn   = false,
+                                    float kv_scale    = 1.0f);
 
 ggml_tensor* ggml_ext_layer_norm(ggml_context* ctx,
                                  ggml_tensor* x,

src/core/ggml_extend_backend.cpp

@@ -8,13 +8,8 @@
 #include <stdexcept>
 #include <vector>
 
-#ifdef SD_USE_CUDA
-#include <cuda.h>
-#endif
-
 #include "core/util.h"
-#include "ggml-backend-impl.h"
-#include "ggml-impl.h"
+#include "ggml/src/ggml-impl.h"
 #include "stable-diffusion.h"
 
 static std::string trim_copy(const std::string& value) {
@@ -434,88 +429,6 @@ bool sd_backend_is_cpu(ggml_backend_t backend) {
     return dev != nullptr && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU;
 }
 
-ggml_backend_buffer_t sd_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device,
-                                                          void* ptr,
-                                                          size_t size,
-                                                          size_t max_tensor_size) {
-    ggml_backend_buffer_t buffer = ggml_backend_dev_buffer_from_host_ptr(device, ptr, size, max_tensor_size);
-    if (buffer != nullptr && buffer->context == nullptr) {
-        ggml_backend_reg_t reg = ggml_backend_dev_backend_reg(device);
-        if (reg != nullptr && std::strcmp(ggml_backend_reg_name(reg), "Metal") == 0) {
-            // Metal can wrap a failed mapping in a non-null buffer. Its free callback also
-            // dereferences the missing context, so only release the outer buffer.
-            buffer->iface.free_buffer = nullptr;
-            ggml_backend_buffer_free(buffer);
-            return nullptr;
-        }
-    }
-    return buffer;
-}
-
-bool sd_backend_supports_cuda_mma(ggml_backend_t backend) {
-#ifdef SD_USE_CUDA
-    if (!sd_backend_is(backend, "CUDA")) {
-        return false;
-    }
-    auto dev = ggml_backend_get_device(backend);
-    if (dev == nullptr) {
-        return false;
-    }
-    static std::mutex mutex;
-    static std::unordered_map<ggml_backend_dev_t, bool> cache;
-    std::lock_guard<std::mutex> lock(mutex);
-    auto it = cache.find(dev);
-    if (it != cache.end()) {
-        return it->second;
-    }
-    const bool supported = [&]() {
-        ggml_backend_dev_props props{};
-        ggml_backend_dev_get_props(dev, &props);
-        CUdevice device;
-        int major = 0, minor = 0;
-        if (props.device_id == nullptr || cuInit(0) != CUDA_SUCCESS ||
-            cuDeviceGetByPCIBusId(&device, props.device_id) != CUDA_SUCCESS ||
-            cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device) != CUDA_SUCCESS ||
-            cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device) != CUDA_SUCCESS) {
-            return false;
-        }
-        auto reg          = ggml_backend_dev_backend_reg(dev);
-        auto get_features = reinterpret_cast<ggml_backend_get_features_t>(
-            ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features"));
-        if (get_features == nullptr) {
-            return false;
-        }
-        // Match ggml's highest compiled architecture for this device, including PTX fallback.
-        const int cc      = 100 * major + 10 * minor;
-        int compiled_arch = 0;
-        for (auto feature = get_features(reg); feature != nullptr && feature->name != nullptr; ++feature) {
-            if (std::strcmp(feature->name, "ARCHS") != 0 || feature->value == nullptr) {
-                continue;
-            }
-            const char* arch = feature->value;
-            while (*arch != '\0') {
-                char* end        = nullptr;
-                const long value = std::strtol(arch, &end, 10);
-                if (end == arch) {
-                    ++arch;
-                    continue;
-                }
-                if (value <= cc && value > compiled_arch) {
-                    compiled_arch = static_cast<int>(value);
-                }
-                arch = end;
-            }
-        }
-        return compiled_arch == 700 || compiled_arch >= 750;
-    }();
-    cache.emplace(dev, supported);
-    return supported;
-#else
-    (void)backend;
-    return false;
-#endif
-}
-
 ggml_backend_t sd_backend_cpu_init() {
     ggml_backend_load_all_once();
     return ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);

src/core/ggml_extend_backend.h

@@ -87,11 +87,6 @@ private:
 
 bool sd_backend_is(ggml_backend_t backend, const std::string& name);
 bool sd_backend_is_cpu(ggml_backend_t backend);
-bool sd_backend_supports_cuda_mma(ggml_backend_t backend);
-ggml_backend_buffer_t sd_backend_dev_buffer_from_host_ptr(ggml_backend_dev_t device,
-                                                          void* ptr,
-                                                          size_t size,
-                                                          size_t max_tensor_size);
 ggml_backend_t sd_backend_cpu_init();
 bool sd_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
 ggml_status sd_backend_graph_compute_with_eval_callback(ggml_backend_t backend,

src/core/ggml_graph_cut.cpp

@@ -16,7 +16,7 @@
 #include "ggml-alloc.h"
 #include "ggml-backend.h"
 
-#include "ggml-impl.h"
+#include "ggml/src/ggml-impl.h"
 
 namespace sd::ggml_graph_cut {
 
@@ -426,8 +426,8 @@ namespace sd::ggml_graph_cut {
         if (tensor == nullptr || tensor->name[0] == '\0') {
             return false;
         }
-        return std::strncmp(tensor->name, GGML_RUNNER_CUT_PREFIX, std::strlen(GGML_RUNNER_CUT_PREFIX)) == 0 &&
-               tensor->name[std::strlen(tensor->name) - 1] == GGML_RUNNER_CUT_SUFFIX[0];
+        return starts_with(tensor->name, GGML_RUNNER_CUT_PREFIX) &&
+               ends_with(tensor->name, GGML_RUNNER_CUT_SUFFIX);
     }
 
     std::string make_graph_cut_name(const std::string& group, const std::string& output) {
@@ -492,88 +492,35 @@ namespace sd::ggml_graph_cut {
         }
     }
 
-    struct GraphLayoutTensors {
-        struct Entry {
-            const ggml_tensor* tensor = nullptr;
-            size_t index              = 0;
-        };
-
-        std::vector<ggml_tensor*> tensors;
-        std::vector<Entry> entries;
-
-        explicit GraphLayoutTensors(size_t graph_size) {
-            tensors.reserve(graph_size);
-            size_t capacity = 2;
-            while (capacity < 2 * graph_size) {
-                capacity *= 2;
-            }
-            entries.resize(capacity);
-        }
-
-        size_t find(const ggml_tensor* tensor) const {
-            size_t hash = reinterpret_cast<uintptr_t>(tensor) >> 4;
-            hash ^= hash >> 16;
-            const size_t mask = entries.size() - 1;
-            size_t slot       = hash & mask;
-            while (entries[slot].tensor != nullptr && entries[slot].tensor != tensor) {
-                slot = (slot + 1) & mask;
-            }
-            return slot;
-        }
-
-        void add(ggml_tensor* tensor) {
-            if (tensor == nullptr) {
-                return;
-            }
-            size_t slot = find(tensor);
-            if (entries[slot].tensor != nullptr) {
-                return;
-            }
-            if (2 * (tensors.size() + 1) > entries.size()) {
-                // Segment graphs can reference tensors outside their node and leaf arrays.
-                std::vector<Entry> next(2 * entries.size());
-                entries.swap(next);
-                for (size_t i = 0; i < tensors.size(); ++i) {
-                    entries[find(tensors[i])] = {tensors[i], i + 1};
-                }
-                slot = find(tensor);
-            }
-            entries[slot] = {tensor, tensors.size() + 1};
-            tensors.push_back(tensor);
-        }
-
-        size_t index(const ggml_tensor* tensor) const {
-            return tensor == nullptr ? 0 : entries[find(tensor)].index;
-        }
-    };
-
     std::vector<uint64_t> graph_layout(ggml_cgraph* graph, bool include_bindings) {
-        const size_t graph_size = static_cast<size_t>(graph->n_leafs) + graph->n_nodes;
-        GraphLayoutTensors layout_tensors(graph_size);
-        const auto& tensors = layout_tensors.tensors;
+        std::vector<const ggml_tensor*> tensors;
+        std::unordered_map<const ggml_tensor*, size_t> indices;
+        auto add = [&](const ggml_tensor* tensor) {
+            if (tensor != nullptr && indices.emplace(tensor, tensors.size() + 1).second) {
+                tensors.push_back(tensor);
+            }
+        };
         for (int i = 0; i < graph->n_leafs; ++i) {
-            layout_tensors.add(graph->leafs[i]);
+            add(graph->leafs[i]);
         }
         for (int i = 0; i < graph->n_nodes; ++i) {
-            layout_tensors.add(graph->nodes[i]);
+            add(graph->nodes[i]);
         }
         for (size_t i = 0; i < tensors.size(); ++i) {
-            layout_tensors.add(tensors[i]->view_src);
+            add(tensors[i]->view_src);
             for (auto source : tensors[i]->src) {
-                layout_tensors.add(source);
+                add(source);
             }
         }
         std::vector<uint64_t> signature;
-        const size_t tensor_fields = 5 + 2 * GGML_MAX_DIMS + GGML_MAX_SRC +
-                                     GGML_MAX_OP_PARAMS / sizeof(int32_t) + (include_bindings ? 2 : 0);
-        signature.reserve(2 + graph_size + tensors.size() * tensor_fields);
+        signature.reserve(tensors.size() * 24);
         signature.push_back(graph->n_nodes);
         signature.push_back(graph->n_leafs);
         for (int i = 0; i < graph->n_leafs; ++i) {
-            signature.push_back(layout_tensors.index(graph->leafs[i]));
+            signature.push_back(indices.at(graph->leafs[i]));
         }
         for (int i = 0; i < graph->n_nodes; ++i) {
-            signature.push_back(layout_tensors.index(graph->nodes[i]));
+            signature.push_back(indices.at(graph->nodes[i]));
         }
         for (auto tensor : tensors) {
             signature.push_back(tensor->op);
@@ -589,9 +536,9 @@ namespace sd::ggml_graph_cut {
                 signature.push_back(tensor->ne[d]);
                 signature.push_back(tensor->nb[d]);
             }
-            signature.push_back(layout_tensors.index(tensor->view_src));
+            signature.push_back(tensor->view_src == nullptr ? 0 : indices.at(tensor->view_src));
             for (auto source : tensor->src) {
-                signature.push_back(layout_tensors.index(source));
+                signature.push_back(source == nullptr ? 0 : indices.at(source));
             }
             if (!can_ignore_op_params(tensor->op)) {
                 for (int value : tensor->op_params) {
@@ -615,18 +562,14 @@ namespace sd::ggml_graph_cut {
                 return false;
             }
         }
-        size_t cut_index = 0;
-        for (int i = 0; i < gf->n_nodes; ++i) {
-            auto node = gf->nodes[i];
+        std::vector<std::pair<int, std::string>> cut_markers;
+        for (int i = 0; i < ggml_graph_n_nodes(gf); ++i) {
+            auto node = ggml_graph_node(gf, i);
             if (is_graph_cut_tensor(node)) {
-                if (cut_index >= plan.cut_markers.size() ||
-                    plan.cut_markers[cut_index].first != i || plan.cut_markers[cut_index].second != node->name) {
-                    return false;
-                }
-                ++cut_index;
+                cut_markers.emplace_back(i, node->name);
             }
         }
-        return cut_index == plan.cut_markers.size();
+        return cut_markers == plan.cut_markers;
     }
 
     bool plan_matches_graph(ggml_cgraph* gf, const Plan& plan) {
@@ -1005,11 +948,11 @@ namespace sd::ggml_graph_cut {
         return plan;
     }
 
-    const Plan& resolve_plan(ggml_backend_t backend,
-                             ggml_cgraph* gf,
-                             PlanCache* cache,
-                             const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                             const char* log_desc) {
+    Plan resolve_plan(ggml_backend_t backend,
+                      ggml_cgraph* gf,
+                      PlanCache* cache,
+                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                      const char* log_desc) {
         GGML_ASSERT(backend != nullptr);
         GGML_ASSERT(gf != nullptr);
         GGML_ASSERT(cache != nullptr);

src/core/ggml_graph_cut.h

@@ -94,12 +94,11 @@ namespace sd::ggml_graph_cut {
                     ggml_cgraph* gf,
                     const std::unordered_set<const ggml_tensor*>& params_tensor_set,
                     const char* log_desc);
-    // The returned reference is valid until its cache entry is evicted or the cache is destroyed.
-    const Plan& resolve_plan(ggml_backend_t backend,
-                             ggml_cgraph* gf,
-                             PlanCache* cache,
-                             const std::unordered_set<const ggml_tensor*>& params_tensor_set,
-                             const char* log_desc);
+    Plan resolve_plan(ggml_backend_t backend,
+                      ggml_cgraph* gf,
+                      PlanCache* cache,
+                      const std::unordered_set<const ggml_tensor*>& params_tensor_set,
+                      const char* log_desc);
 
 }  // namespace sd::ggml_graph_cut
 

src/core/ggml_runner.cpp

@@ -21,12 +21,11 @@ ggml_tensor* ggml_ext_attention_ext(GGMLRunnerContext* ctx,
                                     ggml_tensor* mask,
                                     bool skip_reshape,
                                     bool flash_attn,
-                                    float kv_scale,
-                                    bool* used_flash_attn) {
+                                    float kv_scale) {
     if (ctx->attn_scale > 0.f) {
         kv_scale = ctx->attn_scale;
     }
-    return ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, skip_reshape, flash_attn, kv_scale, ctx->sage_attn_enabled, used_flash_attn);
+    return ggml_ext_attention_ext(ctx->ggml_ctx, ctx->backend, q, k, v, n_head, mask, skip_reshape, flash_attn, kv_scale);
 }
 
 void GGMLRunner::alloc_params_ctx() {
@@ -343,17 +342,21 @@ void GGMLRunner::copy_data_to_backend_tensor(ggml_cgraph* gf, bool clear_after_c
     }
 }
 
-const GGMLRunner::GraphCutPlan& GGMLRunner::resolve_graph_cut_plan(ggml_cgraph* gf) {
+bool GGMLRunner::resolve_graph_cut_plan(ggml_cgraph* gf,
+                                        GraphCutPlan* plan_out) {
+    GGML_ASSERT(plan_out != nullptr);
     GGML_ASSERT(gf != nullptr);
-    return sd::ggml_graph_cut::resolve_plan(runtime_backend,
-                                            gf,
-                                            &graph_cut_plan_cache_,
-                                            params_tensor_set_,
-                                            get_desc().c_str());
+    *plan_out = sd::ggml_graph_cut::resolve_plan(runtime_backend,
+                                                 gf,
+                                                 &graph_cut_plan_cache_,
+                                                 params_tensor_set_,
+                                                 get_desc().c_str());
+    return true;
 }
 
-const GGMLRunner::GraphCutPlan& GGMLRunner::resolve_graph_cut_layer_split_plan(ggml_cgraph* gf) {
-    return resolve_graph_cut_plan(gf);
+bool GGMLRunner::resolve_graph_cut_layer_split_plan(ggml_cgraph* gf,
+                                                    GraphCutPlan* plan_out) {
+    return resolve_graph_cut_plan(gf, plan_out);
 }
 
 bool GGMLRunner::assign_graph_cut_layer_split_backends(ggml_cgraph* gf) {
@@ -366,7 +369,10 @@ bool GGMLRunner::assign_graph_cut_layer_split_backends(ggml_cgraph* gf) {
         return false;
     }
 
-    const auto& plan = resolve_graph_cut_layer_split_plan(gf);
+    GraphCutPlan plan;
+    if (!resolve_graph_cut_layer_split_plan(gf, &plan)) {
+        return false;
+    }
     if (!plan.valid || !plan.has_cuts || plan.segments.size() <= 1) {
         auto manager = residency_manager.lock();
         if (manager == nullptr) {
@@ -516,17 +522,14 @@ GGMLRunner::~GGMLRunner() {
     free_params_ctx();
 }
 
-GGMLRunnerContext GGMLRunner::get_context(ggml_cgraph* graph) {
+GGMLRunnerContext GGMLRunner::get_context() {
     GGMLRunnerContext runner_ctx;
     runner_ctx.ggml_ctx              = compute_ctx;
-    runner_ctx.graph                 = graph;
     runner_ctx.backend               = runtime_backend;
     runner_ctx.flash_attn_enabled    = flash_attn_enabled;
-    runner_ctx.sage_attn_enabled     = sage_attn_enabled;
     runner_ctx.linear_scale          = linear_scale;
     runner_ctx.attn_scale            = attn_scale;
     runner_ctx.conv2d_direct_enabled = conv2d_direct_enabled;
-    runner_ctx.conv3d_direct_enabled = conv3d_direct_enabled;
     runner_ctx.circular_x_enabled    = circular_x_enabled;
     runner_ctx.circular_y_enabled    = circular_y_enabled;
     runner_ctx.weight_adapter        = weight_adapter;
@@ -534,8 +537,8 @@ GGMLRunnerContext GGMLRunner::get_context(ggml_cgraph* graph) {
     runner_ctx.get_cache_tensor      = [this](const std::string& name) {
         return this->get_cache_tensor_by_name(name);
     };
-    runner_ctx.cache_tensor = [this, graph](const std::string& name, ggml_tensor* tensor) {
-        this->cache(name, tensor, graph);
+    runner_ctx.cache_tensor = [this](const std::string& name, ggml_tensor* tensor) {
+        this->cache(name, tensor);
     };
     runner_ctx.set_backend_tensor_data = [this](ggml_tensor* tensor, const void* data) {
         this->set_backend_tensor_data(tensor, data);
@@ -577,7 +580,7 @@ ggml_tensor* GGMLRunner::to_backend(ggml_tensor* tensor) {
     }
 }
 
-void GGMLRunner::cache(const std::string name, ggml_tensor* tensor, ggml_cgraph* graph) {
+void GGMLRunner::cache(const std::string name, ggml_tensor* tensor) {
     if (tensor != nullptr && tensor->view_src != nullptr) {
         tensor = ggml_cont(compute_ctx, tensor);
     }
@@ -585,10 +588,6 @@ void GGMLRunner::cache(const std::string name, ggml_tensor* tensor, ggml_cgraph*
         ggml_set_output(tensor);
     }
     cache_.stage(name, tensor);
-    if (graph != nullptr && tensor != nullptr) {
-        // Schedule the cache output here so its source can be reused before graph end.
-        ggml_build_forward_expand(graph, tensor);
-    }
 }
 
 std::optional<sd::Tensor<float>> GGMLRunner::compute(get_graph_cb_t get_graph,
@@ -596,7 +595,6 @@ std::optional<sd::Tensor<float>> GGMLRunner::compute(get_graph_cb_t get_graph,
                                                      bool auto_runner_end,
                                                      bool no_return,
                                                      const std::function<bool()>& read_outputs) {
-    last_compute_status_ = GGML_STATUS_FAILED;
     if (graph_active_) {
         LOG_ERROR("%s does not support reentrant graph execution", get_desc().c_str());
         return std::nullopt;
@@ -620,9 +618,7 @@ std::optional<sd::Tensor<float>> GGMLRunner::compute(get_graph_cb_t get_graph,
         GGMLRunner& runner;
         const bool& success;
         ~GraphEndGuard() {
-            if (!runner.workspace_.segment_end()) {
-                runner.last_compute_status_ = GGML_STATUS_FAILED;
-            }
+            runner.workspace_.segment_end();
             runner.cache_.graph_end(false);
             runner.cut_cache_.clear();
             runner.free_compute_ctx();
@@ -650,12 +646,7 @@ std::optional<sd::Tensor<float>> GGMLRunner::compute(get_graph_cb_t get_graph,
     std::optional<sd::Tensor<float>> output;
     try {
         output = execute_graph(graph, n_threads, no_return, read_outputs);
-    } catch (const std::bad_alloc&) {
-        last_compute_status_ = GGML_STATUS_ALLOC_FAILED;
-        LOG_ERROR("%s graph allocation failed", get_desc().c_str());
-        return std::nullopt;
     } catch (const std::exception& error) {
-        last_compute_status_ = GGML_STATUS_FAILED;
         LOG_ERROR("%s graph execution failed on %s: %s", get_desc().c_str(),
                   ggml_backend_name(runtime_backend), error.what());
         return std::nullopt;
@@ -663,7 +654,6 @@ std::optional<sd::Tensor<float>> GGMLRunner::compute(get_graph_cb_t get_graph,
     success = output.has_value();
     if (success) {
         cache_.graph_end(true);
-        last_compute_status_ = GGML_STATUS_SUCCESS;
     }
     return output;
 }
@@ -781,7 +771,6 @@ bool GGMLRunner::execute_segment(ggml_cgraph* graph, int n_threads) {
     }
     workspace_.synchronize();
     if (status != GGML_STATUS_SUCCESS) {
-        last_compute_status_ = status;
         LOG_ERROR("%s compute failed: %s", get_desc().c_str(), ggml_status_to_string(status));
         return false;
     }
@@ -830,35 +819,24 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
     if (!assign_graph_cut_layer_split_backends(graph)) {
         return std::nullopt;
     }
-    const auto params           = collect_used_param_tensors(graph);
-    const auto& cached_plan     = resolve_graph_cut_plan(graph);
-    const auto full_measurement = measure(graph, cached_plan.compute_buffer_size);
-    if (full_measurement.buffers.empty()) {
-        last_compute_status_ = GGML_STATUS_ALLOC_FAILED;
+    const auto params = collect_used_param_tensors(graph);
+    ggml_graph_cut::Plan plan;
+    if (!resolve_graph_cut_plan(graph, &plan)) {
+        return std::nullopt;
+    }
+    const auto full_measurement = measure(graph, plan.compute_buffer_size);
+    if (full_measurement.buffers.empty()) {
         return std::nullopt;
     }
-    auto fits_monolithic = [&]() {
-        // Planning headroom absorbs allocation estimate drift; execution keeps the normal limits.
-        constexpr size_t planning_headroom = 128ULL * 1024ULL * 1024ULL;
-        auto requests                      = memory_requests(full_measurement.buffers, cache_.pending_bytes(graph));
-        for (auto& request : requests) {
-            request.pending_allocation_bytes = add_bytes(request.pending_allocation_bytes, planning_headroom);
-        }
-        return fits(requests, params);
-    };
     auto manager         = residency_manager.lock();
     const bool segmented = !is_multi_device() && !sd_backend_is_cpu(runtime_backend) &&
                            manager != nullptr && manager->segmented_compute_enabled() &&
-                           cached_plan.valid && cached_plan.has_cuts && cached_plan.segments.size() > 1 &&
-                           !fits_monolithic();
-    ggml_graph_cut::Plan monolithic_plan;
+                           plan.valid && plan.has_cuts && plan.segments.size() > 1 &&
+                           !fits(memory_requests(full_measurement.buffers, cache_.pending_bytes(graph)), params);
     if (!segmented) {
-        monolithic_plan.segments.emplace_back();
-        auto& segment               = monolithic_plan.segments.back();
+        ggml_graph_cut::Segment segment;
         segment.group_name          = "graph";
-        segment.compute_buffer_size = cached_plan.compute_buffer_size;
-        segment.internal_node_indices.reserve(ggml_graph_n_nodes(graph));
-        segment.input_refs.reserve(ggml_graph_cut::leaf_count(graph));
+        segment.compute_buffer_size = plan.compute_buffer_size;
         for (int i = 0; i < ggml_graph_n_nodes(graph); ++i) {
             segment.internal_node_indices.push_back(i);
         }
@@ -871,8 +849,8 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
                                    : ggml_graph_cut::Segment::INPUT_EXTERNAL;
             segment.input_refs.push_back(input);
         }
+        plan.segments = {std::move(segment)};
     }
-    const auto& plan            = segmented ? cached_plan : monolithic_plan;
     const bool segments_changed = plan.segments.size() != logged_segment_count_;
     if (segments_changed && (segmented || logged_segment_count_ > 1)) {
         LOG_VERBOSE("%s using %zu segment%s", get_desc().c_str(),
@@ -914,9 +892,7 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
             SegmentGraphBindings& bindings;
             ggml_context* context;
             ~SegmentCleanup() {
-                if (!runner.workspace_.segment_end()) {
-                    runner.last_compute_status_ = GGML_STATUS_FAILED;
-                }
+                runner.workspace_.segment_end();
                 bindings.restore();
                 weights.segment_end();
                 ggml_free(context);
@@ -926,7 +902,6 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
 
         auto measurement = segmented ? measure(segment_graph, segment.compute_buffer_size) : full_measurement;
         if (!workspace_.prepare(measurement)) {
-            last_compute_status_ = GGML_STATUS_ALLOC_FAILED;
             return fail_segment("workspace preparation");
         }
         const size_t cut_bytes       = last ? 0 : cut_cache_.estimate_output_bytes(graph, segment);
@@ -934,18 +909,11 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
         auto ensure_capacity         = [&]() {
             sync_runtime_residency();
             auto requests = memory_requests(measurement.buffers, new_cache_bytes);
-            if (fits(requests, weights.params(index))) {
-                return true;
-            }
-            if (workspace_.release_excess(measurement)) {
+            if (!fits(requests, weights.params(index)) && workspace_.release_excess(measurement)) {
                 sync_runtime_residency();
                 requests = memory_requests(measurement.buffers, new_cache_bytes);
             }
-            const bool ready = weights.ensure_segment_capacity(index, requests);
-            if (!ready && manager != nullptr) {
-                last_compute_status_ = GGML_STATUS_ALLOC_FAILED;
-            }
-            return ready;
+            return weights.ensure_segment_capacity(index, requests);
         };
         if (!weights.segment_start(index, ensure_capacity)) {
             return fail_segment("weight preparation");
@@ -954,17 +922,12 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
         if (!workspace_.measurement_matches(segment_graph, measurement)) {
             measurement = measure(segment_graph, segment.compute_buffer_size);
         }
-        if (!workspace_.prepare(measurement)) {
-            last_compute_status_ = GGML_STATUS_ALLOC_FAILED;
-            return fail_segment("workspace preparation");
-        }
-        if (!ensure_capacity()) {
+        if (!workspace_.prepare(measurement) || !ensure_capacity()) {
             return fail_segment("workspace capacity check");
         }
         if (!workspace_.allocate(segment_graph, [&](ggml_backend_sched_t scheduler, ggml_cgraph* current) {
                 pin_multi_device_nodes(scheduler, current);
             })) {
-            last_compute_status_ = GGML_STATUS_ALLOC_FAILED;
             return fail_segment("workspace allocation");
         }
         for (const auto& size : measurement.buffers) {
@@ -983,16 +946,10 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
         }
         LOG_DEBUG("%s executing segment %zu/%zu: %s", get_desc().c_str(),
                   index + 1, plan.segments.size(), segment.group_name.c_str());
-        if (!execute_segment(segment_graph, n_threads)) {
-            return fail_segment("execution");
-        }
-        auto cache_status = cache_.capture(segment_graph);
-        if (cache_status == GGML_STATUS_SUCCESS) {
-            cache_status = cut_cache_.capture(graph, segment, get_desc().c_str());
-        }
-        if (cache_status != GGML_STATUS_SUCCESS) {
-            last_compute_status_ = cache_status;
-            return fail_segment("output caching");
+        if (!execute_segment(segment_graph, n_threads) ||
+            !cache_.capture(segment_graph) ||
+            !cut_cache_.capture(graph, segment, get_desc().c_str())) {
+            return fail_segment("execution or output caching");
         }
         sync_runtime_residency();
         if (last) {
@@ -1008,7 +965,6 @@ std::optional<Tensor<float>> GGMLRunner::execute_graph(ggml_cgraph* graph, int n
             }
         }
         if (!workspace_.segment_end()) {
-            last_compute_status_ = GGML_STATUS_FAILED;
             return fail_segment("workspace synchronization");
         }
         // Final outputs and their callbacks may still be views of consumed cuts.

src/core/ggml_runner.h

@@ -67,13 +67,10 @@ struct WeightAdapter {
 struct GGMLRunnerContext {
     ggml_backend_t backend                                           = nullptr;
     ggml_context* ggml_ctx                                           = nullptr;
-    ggml_cgraph* graph                                               = nullptr;
     bool flash_attn_enabled                                          = false;
-    bool sage_attn_enabled                                           = false;
     float linear_scale                                               = 0.f;
     float attn_scale                                                 = 0.f;
     bool conv2d_direct_enabled                                       = false;
-    bool conv3d_direct_enabled                                       = false;
     bool circular_x_enabled                                          = false;
     bool circular_y_enabled                                          = false;
     ggml_tensor* ip_context                                          = nullptr;
@@ -103,12 +100,6 @@ struct GGMLRunnerContext {
         return get_cache_tensor(name);
     }
 
-    void expand_graph(ggml_tensor* tensor) const {
-        if (graph != nullptr && tensor != nullptr) {
-            ggml_build_forward_expand(graph, tensor);
-        }
-    }
-
     void persist_cache_tensor(const std::string& name, ggml_tensor* tensor) const {
         if (!cache_tensor || tensor == nullptr) {
             return;
@@ -129,17 +120,15 @@ ggml_tensor* ggml_ext_attention_ext(GGMLRunnerContext* ctx,
                                     ggml_tensor* k,
                                     ggml_tensor* v,
                                     int64_t n_head,
-                                    ggml_tensor* mask     = nullptr,
-                                    bool skip_reshape     = false,
-                                    bool flash_attn       = false,
-                                    float kv_scale        = 1.f,
-                                    bool* used_flash_attn = nullptr);
+                                    ggml_tensor* mask = nullptr,
+                                    bool skip_reshape = false,
+                                    bool flash_attn   = false,
+                                    float kv_scale    = 1.f);
 
 struct GGMLRunner {
 private:
     std::map<ggml_backend_t, size_t> logged_compute_bytes_;
-    size_t logged_segment_count_     = 0;
-    ggml_status last_compute_status_ = GGML_STATUS_SUCCESS;
+    size_t logged_segment_count_ = 0;
 
     sd::ComputeWorkspace::Measurement measure(ggml_cgraph* graph, size_t direct_bytes);
     std::vector<DeviceMemoryRequest> memory_requests(const std::vector<sd::BackendBufferSize>& sizes,
@@ -186,11 +175,9 @@ protected:
     const std::string final_result_name = "ggml_runner_final_result_tensor";
 
     bool flash_attn_enabled    = false;
-    bool sage_attn_enabled     = false;
     float linear_scale         = 0.f;
     float attn_scale           = 0.f;
     bool conv2d_direct_enabled = false;
-    bool conv3d_direct_enabled = false;
     bool circular_x_enabled    = false;
     bool circular_y_enabled    = false;
 
@@ -276,9 +263,11 @@ protected:
 
     void copy_data_to_backend_tensor(ggml_cgraph* gf, bool clear_after_copy = true);
 
-    const GraphCutPlan& resolve_graph_cut_plan(ggml_cgraph* gf);
+    bool resolve_graph_cut_plan(ggml_cgraph* gf,
+                                GraphCutPlan* plan_out);
 
-    const GraphCutPlan& resolve_graph_cut_layer_split_plan(ggml_cgraph* gf);
+    bool resolve_graph_cut_layer_split_plan(ggml_cgraph* gf,
+                                            GraphCutPlan* plan_out);
 
     bool assign_graph_cut_layer_split_backends(ggml_cgraph* gf);
 
@@ -297,8 +286,7 @@ public:
 
     virtual ~GGMLRunner();
 
-    // Binding a graph schedules cache outputs at registration instead of graph end.
-    virtual GGMLRunnerContext get_context(ggml_cgraph* graph = nullptr);
+    virtual GGMLRunnerContext get_context();
 
     void reset_compute_ctx();
 
@@ -333,7 +321,7 @@ public:
 
     ggml_tensor* to_backend(ggml_tensor* tensor);
 
-    void cache(const std::string name, ggml_tensor* tensor, ggml_cgraph* graph = nullptr);
+    void cache(const std::string name, ggml_tensor* tensor);
 
     ggml_tensor* get_cache_tensor_by_name(const std::string& name) {
         return cache_.get(name);
@@ -345,20 +333,10 @@ public:
                                              bool no_return                            = false,
                                              const std::function<bool()>& read_outputs = {});
 
-    ggml_status last_compute_status() const { return last_compute_status_; }
-
     void set_flash_attention_enabled(bool enabled) {
         flash_attn_enabled = enabled;
     }
 
-    void set_sage_attention_enabled(bool enabled) {
-        if (sage_attn_enabled != enabled) {
-            free_cache_ctx_and_buffer();
-            graph_cut_plan_cache_.graph_cut_plans.clear();
-            sage_attn_enabled = enabled;
-        }
-    }
-
     void set_scale_overrides(float linear_scale, float attn_scale) {
         this->linear_scale = linear_scale;
         this->attn_scale   = attn_scale;
@@ -368,10 +346,6 @@ public:
         conv2d_direct_enabled = enabled;
     }
 
-    void set_conv3d_direct_enabled(bool enabled) {
-        conv3d_direct_enabled = enabled;
-    }
-
     void set_circular_axes(bool circular_x, bool circular_y) {
         circular_x_enabled = circular_x;
         circular_y_enabled = circular_y;

src/core/parallel.cpp

@@ -1,143 +0,0 @@
-#include "core/parallel.h"
-
-#include <algorithm>
-#include <condition_variable>
-#include <exception>
-#include <mutex>
-#include <thread>
-#include <vector>
-
-namespace sd {
-
-    namespace parallel_detail {
-        thread_local ParallelExecutor* executor = nullptr;
-        thread_local bool active                = false;
-    }
-
-    struct ParallelExecutor::Impl {
-        struct Worker {
-            std::condition_variable wake;
-            std::thread thread;
-            bool ready = false;
-        };
-
-        ParallelExecutor* owner;
-        std::mutex invocation_mutex;
-        std::mutex mutex;
-        std::condition_variable finished;
-        std::vector<std::unique_ptr<Worker>> workers;
-        bool stopping                                     = false;
-        int pending                                       = 0;
-        int participants                                  = 1;
-        int64_t begin                                     = 0;
-        int64_t count                                     = 0;
-        const std::function<void(int64_t, int64_t)>* task = nullptr;
-        std::exception_ptr error;
-
-        explicit Impl(ParallelExecutor* owner)
-            : owner(owner) {}
-
-        ~Impl() {
-            {
-                std::lock_guard<std::mutex> lock(mutex);
-                stopping = true;
-            }
-            for (auto& worker : workers) {
-                worker->wake.notify_one();
-            }
-            for (auto& worker : workers) {
-                worker->thread.join();
-            }
-        }
-
-        void execute(int index) {
-            ParallelScope scope(owner);
-            parallel_detail::Region region;
-            const int64_t size  = count / participants;
-            const int64_t extra = count % participants;
-            const int64_t first = begin + index * size + std::min<int64_t>(index, extra);
-            const int64_t last  = first + size + (index < extra ? 1 : 0);
-            try {
-                (*task)(first, last);
-            } catch (...) {
-                std::lock_guard<std::mutex> lock(mutex);
-                if (!error) {
-                    error = std::current_exception();
-                }
-            }
-        }
-
-        void worker_loop(Worker* worker, int index) {
-            std::unique_lock<std::mutex> lock(mutex);
-            for (;;) {
-                worker->wake.wait(lock, [&] { return stopping || worker->ready; });
-                if (stopping) {
-                    return;
-                }
-                worker->ready = false;
-                lock.unlock();
-                execute(index);
-                lock.lock();
-                if (--pending == 0) {
-                    finished.notify_one();
-                }
-            }
-        }
-
-        void run(int64_t first, int64_t last, int n_tasks, const std::function<void(int64_t, int64_t)>& callback) {
-            std::lock_guard<std::mutex> invocation_lock(invocation_mutex);
-            std::unique_lock<std::mutex> lock(mutex);
-            while (static_cast<int>(workers.size()) < n_tasks - 1) {
-                workers.push_back(std::make_unique<Worker>());
-                auto* worker    = workers.back().get();
-                const int index = static_cast<int>(workers.size());
-                try {
-                    worker->thread = std::thread([this, worker, index] { worker_loop(worker, index); });
-                } catch (...) {
-                    workers.pop_back();
-                    throw;
-                }
-            }
-            begin        = first;
-            count        = last - first;
-            participants = n_tasks;
-            pending      = n_tasks - 1;
-            task         = &callback;
-            error        = nullptr;
-            for (int i = 0; i < pending; ++i) {
-                workers[i]->ready = true;
-                workers[i]->wake.notify_one();
-            }
-            lock.unlock();
-            execute(0);
-            lock.lock();
-            finished.wait(lock, [&] { return pending == 0; });
-            task = nullptr;
-            if (error) {
-                std::rethrow_exception(error);
-            }
-        }
-    };
-
-    ParallelExecutor::ParallelExecutor(int n_threads)
-        : n_threads_(std::max(1, n_threads)), impl_(std::make_unique<Impl>(this)) {}
-
-    ParallelExecutor::~ParallelExecutor() = default;
-
-    void ParallelExecutor::run(int64_t begin, int64_t end, int64_t grain_size, const std::function<void(int64_t, int64_t)>& task) {
-        if (begin < 0 || grain_size <= 0) {
-            throw std::invalid_argument("parallel_for requires begin >= 0 and grain_size > 0");
-        }
-        if (end <= begin) {
-            return;
-        }
-        const int n_tasks = static_cast<int>(std::min<int64_t>(n_threads_, (end - begin) / grain_size));
-        if (parallel_detail::active || n_tasks <= 1) {
-            parallel_detail::Region region;
-            task(begin, end);
-            return;
-        }
-        impl_->run(begin, end, n_tasks, task);
-    }
-
-}

src/core/parallel.h

@@ -1,77 +0,0 @@
-#ifndef __SD_CORE_PARALLEL_H__
-#define __SD_CORE_PARALLEL_H__
-
-#include <cstdint>
-#include <functional>
-#include <memory>
-#include <stdexcept>
-#include <utility>
-
-namespace sd {
-
-    class ParallelExecutor {
-        struct Impl;
-        int n_threads_;
-        std::unique_ptr<Impl> impl_;
-
-    public:
-        explicit ParallelExecutor(int n_threads);
-        ~ParallelExecutor();
-        ParallelExecutor(const ParallelExecutor&)            = delete;
-        ParallelExecutor& operator=(const ParallelExecutor&) = delete;
-
-        int num_threads() const { return n_threads_; }
-        void run(int64_t begin, int64_t end, int64_t grain_size, const std::function<void(int64_t, int64_t)>& task);
-    };
-
-    namespace parallel_detail {
-        extern thread_local ParallelExecutor* executor;
-        extern thread_local bool active;
-
-        class Region {
-            bool previous_;
-
-        public:
-            Region()
-                : previous_(active) { active = true; }
-            ~Region() { active = previous_; }
-            Region(const Region&)            = delete;
-            Region& operator=(const Region&) = delete;
-        };
-    }
-
-    class ParallelScope {
-        ParallelExecutor* previous_;
-
-    public:
-        explicit ParallelScope(ParallelExecutor* executor)
-            : previous_(parallel_detail::executor) {
-            parallel_detail::executor = executor;
-        }
-        ~ParallelScope() { parallel_detail::executor = previous_; }
-        ParallelScope(const ParallelScope&)            = delete;
-        ParallelScope& operator=(const ParallelScope&) = delete;
-    };
-
-    // Ranges are non-negative. The callback may run concurrently and must own its writes.
-    template <typename F>
-    inline void parallel_for(int64_t begin, int64_t end, int64_t grain_size, F&& task) {
-        if (begin < 0 || grain_size <= 0) {
-            throw std::invalid_argument("parallel_for requires begin >= 0 and grain_size > 0");
-        }
-        if (end <= begin) {
-            return;
-        }
-        auto* executor = parallel_detail::executor;
-        if (parallel_detail::active || executor == nullptr || executor->num_threads() <= 1 ||
-            (end - begin) / grain_size < 2) {
-            parallel_detail::Region region;
-            task(begin, end);
-            return;
-        }
-        executor->run(begin, end, grain_size, std::forward<F>(task));
-    }
-
-}
-
-#endif  // __SD_CORE_PARALLEL_H__

src/core/rng_philox.hpp

@@ -1,7 +1,6 @@
 #ifndef __SD_CORE_RNG_PHILOX_HPP__
 #define __SD_CORE_RNG_PHILOX_HPP__
 
-#include <array>
 #include <cmath>
 #include <vector>
 
@@ -15,35 +14,67 @@ private:
     uint32_t offset;
 
 private:
-    using Counter = std::array<std::vector<uint32_t>, 4>;
+    std::vector<uint32_t> philox_m = {0xD2511F53, 0xCD9E8D57};
+    std::vector<uint32_t> philox_w = {0x9E3779B9, 0xBB67AE85};
+    float two_pow32_inv            = 2.3283064e-10f;
+    float two_pow32_inv_2pi        = 2.3283064e-10f * 6.2831855f;
 
-    static constexpr uint32_t philox_m[2] = {0xD2511F53, 0xCD9E8D57};
-    static constexpr uint32_t philox_w[2] = {0x9E3779B9, 0xBB67AE85};
-    float two_pow32_inv                   = 2.3283064e-10f;
-    float two_pow32_inv_2pi               = 2.3283064e-10f * 6.2831855f;
+    std::vector<uint32_t> uint32(uint64_t x) {
+        std::vector<uint32_t> result(2);
+        result[0] = static_cast<uint32_t>(x & 0xFFFFFFFF);
+        result[1] = static_cast<uint32_t>(x >> 32);
+        return result;
+    }
+
+    std::vector<std::vector<uint32_t>> uint32(const std::vector<uint64_t>& x) {
+        uint32_t N = (uint32_t)x.size();
+        std::vector<std::vector<uint32_t>> result(2, std::vector<uint32_t>(N));
+
+        for (uint32_t i = 0; i < N; ++i) {
+            result[0][i] = static_cast<uint32_t>(x[i] & 0xFFFFFFFF);
+            result[1][i] = static_cast<uint32_t>(x[i] >> 32);
+        }
+
+        return result;
+    }
 
     //  A single round of the Philox 4x32 random number generator.
-    void philox4_round(Counter& counter, uint32_t key0, uint32_t key1) {
+    void philox4_round(std::vector<std::vector<uint32_t>>& counter,
+                       const std::vector<std::vector<uint32_t>>& key) {
         uint32_t N = (uint32_t)counter[0].size();
         for (uint32_t i = 0; i < N; i++) {
-            const uint64_t v1 = static_cast<uint64_t>(counter[0][i]) * static_cast<uint64_t>(philox_m[0]);
-            const uint64_t v2 = static_cast<uint64_t>(counter[2][i]) * static_cast<uint64_t>(philox_m[1]);
+            std::vector<uint32_t> v1 = uint32(static_cast<uint64_t>(counter[0][i]) * static_cast<uint64_t>(philox_m[0]));
+            std::vector<uint32_t> v2 = uint32(static_cast<uint64_t>(counter[2][i]) * static_cast<uint64_t>(philox_m[1]));
 
-            counter[0][i] = static_cast<uint32_t>(v2 >> 32) ^ counter[1][i] ^ key0;
-            counter[1][i] = static_cast<uint32_t>(v2);
-            counter[2][i] = static_cast<uint32_t>(v1 >> 32) ^ counter[3][i] ^ key1;
-            counter[3][i] = static_cast<uint32_t>(v1);
+            counter[0][i] = v2[1] ^ counter[1][i] ^ key[0][i];
+            counter[1][i] = v2[0];
+            counter[2][i] = v1[1] ^ counter[3][i] ^ key[1][i];
+            counter[3][i] = v1[0];
         }
     }
 
-    void philox4_32(Counter& counter, uint32_t key0, uint32_t key1, int rounds = 10) {
+    // Generates 32-bit random numbers using the Philox 4x32 random number generator.
+    // Parameters:
+    //     counter : A 4xN array of 32-bit integers representing the counter values (offset into generation).
+    //     key : A 2xN array of 32-bit integers representing the key values (seed).
+    //     rounds : The number of rounds to perform.
+    // Returns:
+    //     std::vector<std::vector<uint32_t>>: A 4xN array of 32-bit integers containing the generated random numbers.
+    std::vector<std::vector<uint32_t>> philox4_32(std::vector<std::vector<uint32_t>>& counter,
+                                                  std::vector<std::vector<uint32_t>>& key,
+                                                  int rounds = 10) {
+        uint32_t N = (uint32_t)counter[0].size();
         for (int i = 0; i < rounds - 1; ++i) {
-            philox4_round(counter, key0, key1);
-            key0 += philox_w[0];
-            key1 += philox_w[1];
+            philox4_round(counter, key);
+
+            for (uint32_t j = 0; j < N; ++j) {
+                key[0][j] += philox_w[0];
+                key[1][j] += philox_w[1];
+            }
         }
 
-        philox4_round(counter, key0, key1);
+        philox4_round(counter, key);
+        return counter;
     }
 
     float box_muller(float x, float y) {
@@ -72,22 +103,24 @@ public:
     }
 
     std::vector<float> randn(uint32_t n) override {
-        Counter counter;
-        counter[0].resize(n, this->offset);
-        counter[1].resize(n);
-        counter[2].resize(n);
-        counter[3].resize(n);
+        std::vector<std::vector<uint32_t>> counter(4, std::vector<uint32_t>(n, 0));
+        for (uint32_t i = 0; i < n; i++) {
+            counter[0][i] = this->offset;
+        }
 
         for (uint32_t i = 0; i < n; i++) {
             counter[2][i] = i;
         }
         this->offset += 1;
 
-        philox4_32(counter, static_cast<uint32_t>(this->seed), static_cast<uint32_t>(this->seed >> 32));
+        std::vector<uint64_t> key(n, this->seed);
+        std::vector<std::vector<uint32_t>> key_uint32 = uint32(key);
 
-        std::vector<float> result(n);
+        std::vector<std::vector<uint32_t>> g = philox4_32(counter, key_uint32);
+
+        std::vector<float> result;
         for (uint32_t i = 0; i < n; ++i) {
-            result[i] = box_muller((float)counter[0][i], (float)counter[1][i]);
+            result.push_back(box_muller((float)g[0][i], (float)g[1][i]));
         }
         return result;
     }

src/core/runner_cache.cpp

@@ -26,13 +26,10 @@ namespace sd {
 
     std::unique_ptr<CachedTensor> CachedTensor::copy(ggml_backend_t backend,
                                                      const std::string& name,
-                                                     ggml_tensor* source,
-                                                     ggml_status& status) {
-        status = GGML_STATUS_FAILED;
+                                                     ggml_tensor* source) {
         if (ggml_graph_cut::tensor_buffer(source) == nullptr) {
             return nullptr;
         }
-        status         = GGML_STATUS_ALLOC_FAILED;
         auto entry     = std::make_unique<CachedTensor>();
         entry->context = ggml_init({2 * ggml_tensor_overhead(), nullptr, true});
         if (entry->context == nullptr) {
@@ -53,7 +50,6 @@ namespace sd {
         } else {
             ggml_backend_tensor_copy(source, entry->tensor);
         }
-        status = GGML_STATUS_SUCCESS;
         return entry;
     }
 
@@ -110,9 +106,9 @@ namespace sd {
         return pending > SIZE_MAX - committed ? SIZE_MAX : committed + pending;
     }
 
-    ggml_status RunnerCache::capture(ggml_cgraph* graph) {
+    bool RunnerCache::capture(ggml_cgraph* graph) {
         if (outputs_.empty()) {
-            return GGML_STATUS_SUCCESS;
+            return true;
         }
         const auto tensors = cache_graph_tensors(graph);
         for (const auto& output : outputs_) {
@@ -120,15 +116,14 @@ namespace sd {
                 continue;
             }
             GGML_ASSERT(ggml_is_contiguous(output.second));
-            ggml_status status;
-            auto entry = CachedTensor::copy(backend_, output.first, output.second, status);
+            auto entry = CachedTensor::copy(backend_, output.first, output.second);
             if (entry == nullptr) {
-                return status;
+                return false;
             }
             pending_[output.first] = std::move(entry);
         }
         ggml_backend_synchronize(backend_);
-        return GGML_STATUS_SUCCESS;
+        return true;
     }
 
     void RunnerCache::graph_end(bool success) {
@@ -185,9 +180,9 @@ namespace sd {
         }
     }
 
-    ggml_status GraphCutTensorCache::capture(ggml_cgraph* graph,
-                                             const ggml_graph_cut::Segment& segment,
-                                             const char* log_desc) {
+    bool GraphCutTensorCache::capture(ggml_cgraph* graph,
+                                      const ggml_graph_cut::Segment& segment,
+                                      const char* log_desc) {
         size_t copied_bytes = 0;
         size_t copied_count = 0;
         for (int index : segment.output_node_indices) {
@@ -196,11 +191,10 @@ namespace sd {
                 !segment.future_cut_names.count(output->name)) {
                 continue;
             }
-            ggml_status status;
-            auto entry = CachedTensor::copy(backend_, output->name, ggml_graph_cut::cache_source_tensor(output), status);
+            auto entry = CachedTensor::copy(backend_, output->name, ggml_graph_cut::cache_source_tensor(output));
             if (entry == nullptr) {
                 LOG_ERROR("%s failed to capture graph cut tensor: %s", log_desc, output->name);
-                return status;
+                return false;
             }
             const size_t size = ggml_backend_buffer_get_size(entry->buffer);
             copied_bytes      = size > SIZE_MAX - copied_bytes ? SIZE_MAX : copied_bytes + size;
@@ -212,6 +206,6 @@ namespace sd {
             LOG_DEBUG("%s graph cut cache added %6.2f MB (%zu tensors)",
                       log_desc, copied_bytes / (1024.f * 1024.f), copied_count);
         }
-        return GGML_STATUS_SUCCESS;
+        return true;
     }
 }

src/core/runner_cache.h

@@ -20,8 +20,7 @@ namespace sd {
         ~CachedTensor();
         static std::unique_ptr<CachedTensor> copy(ggml_backend_t backend,
                                                   const std::string& name,
-                                                  ggml_tensor* source,
-                                                  ggml_status& status);
+                                                  ggml_tensor* source);
     };
     using CachedTensors = std::map<std::string, std::unique_ptr<CachedTensor>>;
 
@@ -42,8 +41,7 @@ namespace sd {
         const std::map<std::string, ggml_tensor*>& outputs() const { return outputs_; }
         size_t pending_bytes(ggml_cgraph* graph) const;
         size_t resident_bytes(ggml_backend_dev_t device) const;
-        bool empty() const { return committed_.empty(); }
-        ggml_status capture(ggml_cgraph* graph);
+        bool capture(ggml_cgraph* graph);
         void graph_end(bool success);
         void clear();
     };
@@ -59,7 +57,7 @@ namespace sd {
         size_t resident_bytes(ggml_backend_dev_t device) const;
         size_t estimate_output_bytes(ggml_cgraph* graph,
                                      const ggml_graph_cut::Segment& segment) const;
-        ggml_status capture(ggml_cgraph* graph, const ggml_graph_cut::Segment& segment, const char* log_desc);
+        bool capture(ggml_cgraph* graph, const ggml_graph_cut::Segment& segment, const char* log_desc);
         void prune(const std::unordered_set<std::string>& keep_names);
         void clear() { tensors_.clear(); }
     };

src/core/tensor.hpp

@@ -16,7 +16,6 @@
 #include <utility>
 #include <vector>
 
-#include "core/parallel.h"
 #include "core/rng.hpp"
 
 namespace sd {
@@ -60,15 +59,6 @@ namespace sd {
         return numel;
     }
 
-    template <typename F>
-    inline void tensor_for_each(int64_t count, F&& fn, int64_t grain_size = 65536) {
-        parallel_for(0, count, grain_size, [&](int64_t begin, int64_t end) {
-            for (int64_t i = begin; i < end; ++i) {
-                fn(i);
-            }
-        });
-    }
-
     template <typename T>
     class Tensor {
     public:
@@ -240,10 +230,7 @@ namespace sd {
         }
 
         void fill_(const T& value) {
-            const T fill_value = value;
-            parallel_for(0, numel(), 65536, [&](int64_t begin, int64_t end) {
-                std::fill_n(data_.data() + begin, end - begin, fill_value);
-            });
+            std::fill(data_.begin(), data_.end(), value);
         }
 
         Tensor& masked_fill_(const Tensor<uint8_t>& mask, const T& value);
@@ -403,7 +390,7 @@ namespace sd {
                                               tensor_shape_to_string(lhs) + ", rhs_shape=" +
                                               tensor_shape_to_string(rhs));
             }
-            shape[i] = lhs_dim == 1 ? rhs_dim : lhs_dim;
+            shape[i] = std::max(lhs_dim, rhs_dim);
         }
         return shape;
     }
@@ -438,55 +425,39 @@ namespace sd {
                                                  const std::vector<int64_t>& rhs_shape_raw,
                                                  const std::vector<int64_t>& rhs_strides_raw,
                                                  F&& fn) {
-        const int64_t numel    = tensor_numel(out_shape);
-        const size_t ndim      = out_shape.size();
-        auto broadcast_strides = [&](const std::vector<int64_t>& shape,
-                                     const std::vector<int64_t>& strides) {
-            if ((numel != 0 && tensor_numel(shape) == 0) || strides.size() != shape.size()) {
-                tensor_throw_invalid_argument("Tensor broadcast requires non-empty inputs and matching strides");
-            }
-            std::vector<int64_t> result(ndim, 0);
-            for (size_t i = 0; i < std::max(ndim, shape.size()); ++i) {
-                const int64_t input_dim  = i < shape.size() ? shape[i] : 1;
-                const int64_t output_dim = i < ndim ? out_shape[i] : 1;
-                if (input_dim != 1 && input_dim != output_dim) {
-                    tensor_throw_invalid_argument("Tensor broadcast cannot expand the destination: input_shape=" +
-                                                  tensor_shape_to_string(shape) + ", output_shape=" +
-                                                  tensor_shape_to_string(out_shape));
-                }
-                if (i < ndim && input_dim != 1) {
-                    result[i] = strides[i];
-                }
-            }
-            return result;
-        };
-        const auto lhs_strides = broadcast_strides(lhs_shape_raw, lhs_strides_raw);
-        const auto rhs_strides = broadcast_strides(rhs_shape_raw, rhs_strides_raw);
-        if (numel == 0) {
-            return;
+        const size_t ndim                = out_shape.size();
+        std::vector<int64_t> out_strides = tensor_compute_strides(out_shape);
+        std::vector<int64_t> lhs_shape(ndim, 1);
+        std::vector<int64_t> lhs_strides(ndim, 0);
+        std::vector<int64_t> rhs_shape(ndim, 1);
+        std::vector<int64_t> rhs_strides(ndim, 0);
+
+        for (size_t i = 0; i < lhs_shape_raw.size(); ++i) {
+            lhs_shape[i]   = lhs_shape_raw[i];
+            lhs_strides[i] = lhs_strides_raw[i];
         }
-        parallel_for(0, numel, 16384, [&](int64_t begin, int64_t end) {
-            auto coord         = tensor_unravel_index(begin, out_shape);
+        for (size_t i = 0; i < rhs_shape_raw.size(); ++i) {
+            rhs_shape[i]   = rhs_shape_raw[i];
+            rhs_strides[i] = rhs_strides_raw[i];
+        }
+
+        const int64_t numel = tensor_numel(out_shape);
+        for (int64_t flat = 0; flat < numel; ++flat) {
+            int64_t remaining  = flat;
             int64_t lhs_offset = 0;
             int64_t rhs_offset = 0;
-            for (size_t i = 0; i < ndim; ++i) {
-                lhs_offset += coord[i] * lhs_strides[i];
-                rhs_offset += coord[i] * rhs_strides[i];
-            }
-            for (int64_t flat = begin; flat < end; ++flat) {
-                fn(flat, lhs_offset, rhs_offset);
-                for (size_t i = 0; i < ndim; ++i) {
-                    lhs_offset += lhs_strides[i];
-                    rhs_offset += rhs_strides[i];
-                    if (++coord[i] < out_shape[i]) {
-                        break;
-                    }
-                    coord[i] = 0;
-                    lhs_offset -= out_shape[i] * lhs_strides[i];
-                    rhs_offset -= out_shape[i] * rhs_strides[i];
+            for (size_t i = ndim; i-- > 0;) {
+                int64_t coord = remaining / out_strides[i];
+                remaining %= out_strides[i];
+                if (lhs_shape[i] != 1) {
+                    lhs_offset += coord * lhs_strides[i];
+                }
+                if (rhs_shape[i] != 1) {
+                    rhs_offset += coord * rhs_strides[i];
                 }
             }
-        });
+            fn(flat, lhs_offset, rhs_offset);
+        }
     }
 
     template <typename T>
@@ -498,7 +469,6 @@ namespace sd {
         const std::vector<int64_t> data_strides = tensor_compute_strides(shape_);
         const std::vector<int64_t> mask_strides = tensor_compute_strides(mask.shape());
         const uint8_t* mask_data                = mask.data();
-        const T fill_value                      = value;
         tensor_for_each_broadcast_offset(shape_,
                                          shape_,
                                          data_strides,
@@ -506,7 +476,7 @@ namespace sd {
                                          mask_strides,
                                          [&](int64_t, int64_t data_offset, int64_t mask_offset) {
                                              if (mask_data[mask_offset] != 0) {
-                                                 data_[static_cast<size_t>(data_offset)] = fill_value;
+                                                 data_[static_cast<size_t>(data_offset)] = value;
                                              }
                                          });
         return *this;
@@ -516,9 +486,9 @@ namespace sd {
     inline Tensor<uint8_t> operator<(const Tensor<T>& lhs, Scalar rhs) {
         Tensor<uint8_t> result(lhs.shape());
         const T value = static_cast<T>(rhs);
-        tensor_for_each(lhs.numel(), [&](int64_t i) {
-            result.data()[i] = lhs.data()[i] < value ? 1 : 0;
-        });
+        for (int64_t i = 0; i < lhs.numel(); ++i) {
+            result[i] = lhs[i] < value ? 1 : 0;
+        }
         return result;
     }
 
@@ -526,9 +496,9 @@ namespace sd {
     inline Tensor<uint8_t> operator<(Scalar lhs, const Tensor<T>& rhs) {
         Tensor<uint8_t> result(rhs.shape());
         const T value = static_cast<T>(lhs);
-        tensor_for_each(rhs.numel(), [&](int64_t i) {
-            result.data()[i] = value < rhs.data()[i] ? 1 : 0;
-        });
+        for (int64_t i = 0; i < rhs.numel(); ++i) {
+            result[i] = value < rhs[i] ? 1 : 0;
+        }
         return result;
     }
 
@@ -546,7 +516,7 @@ namespace sd {
                                          rhs.shape(),
                                          rhs_strides,
                                          [&](int64_t flat, int64_t lhs_offset, int64_t rhs_offset) {
-                                             result.data()[flat] = lhs_data[lhs_offset] < rhs_data[rhs_offset] ? 1 : 0;
+                                             result[flat] = lhs_data[lhs_offset] < rhs_data[rhs_offset] ? 1 : 0;
                                          });
         return result;
     }
@@ -554,9 +524,9 @@ namespace sd {
     template <typename T>
     inline Tensor<T>& operator+=(Tensor<T>& lhs, const Tensor<T>& rhs) {
         if (lhs.shape() == rhs.shape()) {
-            tensor_for_each(lhs.numel(), [&](int64_t i) {
-                lhs.data()[i] += rhs.data()[i];
-            });
+            for (int64_t i = 0; i < lhs.numel(); ++i) {
+                lhs[i] += rhs[i];
+            }
             return lhs;
         }
         tensor_broadcast_shape(lhs.shape(), rhs.shape());
@@ -569,7 +539,7 @@ namespace sd {
                                          rhs.shape(),
                                          rhs_strides,
                                          [&](int64_t, int64_t lhs_offset, int64_t rhs_offset) {
-                                             lhs.data()[lhs_offset] += rhs_data[rhs_offset];
+                                             lhs[static_cast<int64_t>(lhs_offset)] += rhs_data[rhs_offset];
                                          });
         return lhs;
     }
@@ -577,18 +547,18 @@ namespace sd {
     template <typename T, typename Scalar, typename = std::enable_if_t<std::is_arithmetic<Scalar>::value>>
     inline Tensor<T>& operator+=(Tensor<T>& lhs, Scalar rhs) {
         const T value = static_cast<T>(rhs);
-        tensor_for_each(lhs.numel(), [&](int64_t i) {
-            lhs.data()[i] += value;
-        });
+        for (int64_t i = 0; i < lhs.numel(); ++i) {
+            lhs[i] += value;
+        }
         return lhs;
     }
 
     template <typename T>
     inline Tensor<T>& operator-=(Tensor<T>& lhs, const Tensor<T>& rhs) {
         if (lhs.shape() == rhs.shape()) {
-            tensor_for_each(lhs.numel(), [&](int64_t i) {
-                lhs.data()[i] -= rhs.data()[i];
-            });
+            for (int64_t i = 0; i < lhs.numel(); ++i) {
+                lhs[i] -= rhs[i];
+            }
             return lhs;
         }
         tensor_broadcast_shape(lhs.shape(), rhs.shape());
@@ -601,7 +571,7 @@ namespace sd {
                                          rhs.shape(),
                                          rhs_strides,
                                          [&](int64_t, int64_t lhs_offset, int64_t rhs_offset) {
-                                             lhs.data()[lhs_offset] -= rhs_data[rhs_offset];
+                                             lhs[static_cast<int64_t>(lhs_offset)] -= rhs_data[rhs_offset];
                                          });
         return lhs;
     }
@@ -609,18 +579,18 @@ namespace sd {
     template <typename T, typename Scalar, typename = std::enable_if_t<std::is_arithmetic<Scalar>::value>>
     inline Tensor<T>& operator-=(Tensor<T>& lhs, Scalar rhs) {
         const T value = static_cast<T>(rhs);
-        tensor_for_each(lhs.numel(), [&](int64_t i) {
-            lhs.data()[i] -= value;
-        });
+        for (int64_t i = 0; i < lhs.numel(); ++i) {
+            lhs[i] -= value;
+        }
         return lhs;
     }
 
     template <typename T>
     inline Tensor<T>& operator*=(Tensor<T>& lhs, const Tensor<T>& rhs) {
         if (lhs.shape() == rhs.shape()) {
-            tensor_for_each(lhs.numel(), [&](int64_t i) {
-                lhs.data()[i] *= rhs.data()[i];
-            });
+            for (int64_t i = 0; i < lhs.numel(); ++i) {
+                lhs[i] *= rhs[i];
+            }
             return lhs;
         }
         tensor_broadcast_shape(lhs.shape(), rhs.shape());
@@ -633,7 +603,7 @@ namespace sd {
                                          rhs.shape(),
                                          rhs_strides,
                                          [&](int64_t, int64_t lhs_offset, int64_t rhs_offset) {
-                                             lhs.data()[lhs_offset] *= rhs_data[rhs_offset];
+                                             lhs[static_cast<int64_t>(lhs_offset)] *= rhs_data[rhs_offset];
                                          });
         return lhs;
     }
@@ -641,18 +611,18 @@ namespace sd {
     template <typename T, typename Scalar, typename = std::enable_if_t<std::is_arithmetic<Scalar>::value>>
     inline Tensor<T>& operator*=(Tensor<T>& lhs, Scalar rhs) {
         const T value = static_cast<T>(rhs);
-        tensor_for_each(lhs.numel(), [&](int64_t i) {
-            lhs.data()[i] *= value;
-        });
+        for (int64_t i = 0; i < lhs.numel(); ++i) {
+            lhs[i] *= value;
+        }
         return lhs;
     }
 
     template <typename T>
     inline Tensor<T>& operator/=(Tensor<T>& lhs, const Tensor<T>& rhs) {
         if (lhs.shape() == rhs.shape()) {
-            tensor_for_each(lhs.numel(), [&](int64_t i) {
-                lhs.data()[i] /= rhs.data()[i];
-            });
+            for (int64_t i = 0; i < lhs.numel(); ++i) {
+                lhs[i] /= rhs[i];
+            }
             return lhs;
         }
         tensor_broadcast_shape(lhs.shape(), rhs.shape());
@@ -665,7 +635,7 @@ namespace sd {
                                          rhs.shape(),
                                          rhs_strides,
                                          [&](int64_t, int64_t lhs_offset, int64_t rhs_offset) {
-                                             lhs.data()[lhs_offset] /= rhs_data[rhs_offset];
+                                             lhs[static_cast<int64_t>(lhs_offset)] /= rhs_data[rhs_offset];
                                          });
         return lhs;
     }
@@ -673,9 +643,9 @@ namespace sd {
     template <typename T, typename Scalar, typename = std::enable_if_t<std::is_arithmetic<Scalar>::value>>
     inline Tensor<T>& operator/=(Tensor<T>& lhs, Scalar rhs) {
         const T value = static_cast<T>(rhs);
-        tensor_for_each(lhs.numel(), [&](int64_t i) {
-            lhs.data()[i] /= value;
-        });
+        for (int64_t i = 0; i < lhs.numel(); ++i) {
+            lhs[i] /= value;
+        }
         return lhs;
     }
 
@@ -694,7 +664,7 @@ namespace sd {
                                              rhs.shape(),
                                              rhs_strides,
                                              [&](int64_t flat, int64_t lhs_offset, int64_t rhs_offset) {
-                                                 result.data()[flat] = lhs_data[lhs_offset] + rhs_data[rhs_offset];
+                                                 result[flat] = lhs_data[lhs_offset] + rhs_data[rhs_offset];
                                              });
             return result;
         }
@@ -729,7 +699,7 @@ namespace sd {
                                              rhs.shape(),
                                              rhs_strides,
                                              [&](int64_t flat, int64_t lhs_offset, int64_t rhs_offset) {
-                                                 result.data()[flat] = lhs_data[lhs_offset] - rhs_data[rhs_offset];
+                                                 result[flat] = lhs_data[lhs_offset] - rhs_data[rhs_offset];
                                              });
             return result;
         }
@@ -747,9 +717,9 @@ namespace sd {
     inline Tensor<T> operator-(Scalar lhs, const Tensor<T>& rhs) {
         Tensor<T> result = rhs;
         const T value    = static_cast<T>(lhs);
-        tensor_for_each(result.numel(), [&](int64_t i) {
-            result.data()[i] = value - result.data()[i];
-        });
+        for (int64_t i = 0; i < result.numel(); ++i) {
+            result[i] = value - result[i];
+        }
         return result;
     }
 
@@ -768,7 +738,7 @@ namespace sd {
                                              rhs.shape(),
                                              rhs_strides,
                                              [&](int64_t flat, int64_t lhs_offset, int64_t rhs_offset) {
-                                                 result.data()[flat] = lhs_data[lhs_offset] * rhs_data[rhs_offset];
+                                                 result[flat] = lhs_data[lhs_offset] * rhs_data[rhs_offset];
                                              });
             return result;
         }
@@ -803,7 +773,7 @@ namespace sd {
                                              rhs.shape(),
                                              rhs_strides,
                                              [&](int64_t flat, int64_t lhs_offset, int64_t rhs_offset) {
-                                                 result.data()[flat] = lhs_data[lhs_offset] / rhs_data[rhs_offset];
+                                                 result[flat] = lhs_data[lhs_offset] / rhs_data[rhs_offset];
                                              });
             return result;
         }
@@ -821,18 +791,18 @@ namespace sd {
     inline Tensor<T> operator/(Scalar lhs, const Tensor<T>& rhs) {
         Tensor<T> result = rhs;
         const T value    = static_cast<T>(lhs);
-        tensor_for_each(result.numel(), [&](int64_t i) {
-            result.data()[i] = value / result.data()[i];
-        });
+        for (int64_t i = 0; i < result.numel(); ++i) {
+            result[i] = value / result[i];
+        }
         return result;
     }
 
     template <typename T>
     inline Tensor<T> operator-(const Tensor<T>& tensor) {
         Tensor<T> result = tensor;
-        tensor_for_each(result.numel(), [&](int64_t i) {
-            result.data()[i] = -result.data()[i];
-        });
+        for (int64_t i = 0; i < result.numel(); ++i) {
+            result[i] = -result[i];
+        }
         return result;
     }
 
@@ -1097,11 +1067,9 @@ namespace sd {
         template <typename T>
         inline Tensor<T> exp(const Tensor<T>& input) {
             Tensor<T> output(input.shape());
-            tensor_for_each(
-                input.numel(), [&](int64_t i) {
-                    output.data()[i] = static_cast<T>(std::exp(static_cast<double>(input.data()[i])));
-                },
-                4096);
+            for (int64_t i = 0; i < input.numel(); ++i) {
+                output[i] = static_cast<T>(std::exp(static_cast<double>(input[i])));
+            }
             return output;
         }
 
@@ -1111,18 +1079,18 @@ namespace sd {
                 tensor_throw_invalid_argument("Tensor clamp requires min_value <= max_value");
             }
             Tensor<T> output(input.shape());
-            tensor_for_each(input.numel(), [&](int64_t i) {
-                output.data()[i] = std::clamp(input.data()[i], min_value, max_value);
-            });
+            for (int64_t i = 0; i < input.numel(); ++i) {
+                output[i] = std::clamp(input[i], min_value, max_value);
+            }
             return output;
         }
 
         template <typename T>
         inline Tensor<T> round(const Tensor<T>& input) {
             Tensor<T> output(input.shape());
-            tensor_for_each(input.numel(), [&](int64_t i) {
-                output.data()[i] = static_cast<T>(std::round(static_cast<double>(input.data()[i])));
-            });
+            for (int64_t i = 0; i < input.numel(); ++i) {
+                output[i] = static_cast<T>(std::round(static_cast<double>(input[i])));
+            }
             return output;
         }
 

src/core/util.cpp

@@ -62,34 +62,17 @@ void replace_all_chars(std::string& str, char target, char replacement) {
     }
 }
 
-static std::string sd_vformat(const char* fmt, va_list ap) {
-    char small[128];
-    va_list ap2;
-    va_copy(ap2, ap);
-    int size = vsnprintf(small, sizeof small, fmt, ap);
-    if (size < 0) {
-        va_end(ap2);
-        return {};
-    }
-    size_t needed = (size_t)size;
-    if (needed < sizeof small) {
-        va_end(ap2);
-        return std::string(small, needed);
-    }
-    std::string out(needed, '\0');
-    int size2 = vsnprintf(out.data(), needed + 1, fmt, ap2);
-    va_end(ap2);
-    if (size2 < 0)
-        out.clear();
-    return out;
-}
-
 std::string sd_format(const char* fmt, ...) {
     va_list ap;
+    va_list ap2;
     va_start(ap, fmt);
-    std::string result = sd_vformat(fmt, ap);
+    va_copy(ap2, ap);
+    int size = vsnprintf(nullptr, 0, fmt, ap);
+    std::vector<char> buf(size + 1);
+    int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
+    va_end(ap2);
     va_end(ap);
-    return result;
+    return std::string(buf.data(), size);
 }
 
 int round_up_to(int value, int base) {
@@ -431,43 +414,14 @@ std::vector<std::string> split_string(const std::string& str, char delimiter) {
 }
 
 ggml_type sd_type_to_ggml_type(sd_type_t sdtype) {
-    if (sdtype == SD_TYPE_F8_E4M3 || sdtype == SD_TYPE_F8_E5M2) {
-#ifndef SD_USE_UPSTREAM_GGML
-        return sdtype == SD_TYPE_F8_E4M3 ? GGML_TYPE_F8_E4M3 : GGML_TYPE_F8_E5M2;
-#else
-        return GGML_TYPE_COUNT;
-#endif
-    }
     const int type_value = static_cast<int>(sdtype);
-    if (type_value >= 0 && type_value < std::min<int>(SD_TYPE_COUNT, GGML_TYPE_COUNT)) {
+    if (type_value < std::min<int>(SD_TYPE_COUNT, GGML_TYPE_COUNT)) {
         return static_cast<ggml_type>(type_value);
     } else {
         return GGML_TYPE_COUNT;
     }
 }
 
-bool validate_tensor_types(sd_type_t type, const char* tensor_type_rules) {
-    if (type != SD_TYPE_COUNT && sd_type_to_ggml_type(type) == GGML_TYPE_COUNT) {
-        LOG_ERROR("weight type %s is not supported by this ggml build", sd_type_name(type));
-        return false;
-    }
-#ifdef SD_USE_UPSTREAM_GGML
-    for (const auto& rule : split_string(SAFE_STR(tensor_type_rules), ',')) {
-        const auto pos = rule.find('=');
-        if (pos != std::string::npos) {
-            const auto name = rule.substr(pos + 1);
-            if (name == "f8_e4m3" || name == "f8_e5m2") {
-                LOG_ERROR("FP8 is not supported by this ggml build (tensor type rule '%s')", rule.c_str());
-                return false;
-            }
-        }
-    }
-#else
-    GGML_UNUSED(tensor_type_rules);
-#endif
-    return true;
-}
-
 KeyValueArgs parse_key_value_args(const char* args, const char* context) {
     KeyValueArgs pairs;
 
@@ -641,45 +595,47 @@ std::string trim(const std::string& s) {
 static sd_log_cb_t sd_log_cb = nullptr;
 void* sd_log_cb_data         = nullptr;
 
-static void sd_log_dispatch(sd_log_level_t level, const std::string& origin, const std::string& text) {
-    if (sd_log_cb == nullptr)
-        return;
-    std::string message = origin + " - " + text;
-    if (message.back() != '\n') {
-        message += '\n';
-    }
-    sd_log_cb(level, message.c_str(), sd_log_cb_data);
-}
+#define LOG_BUFFER_SIZE 4096
 
 void log_printf(sd_log_level_t level, const char* file, int line, const char* format, ...) {
     va_list args;
     va_start(args, format);
-    std::string message = sd_vformat(format, args);
+
+    static char log_buffer[LOG_BUFFER_SIZE + 1];
+    int written = snprintf(log_buffer, LOG_BUFFER_SIZE, "%s:%-4d - ", sd_basename(file).c_str(), line);
+
+    if (written >= 0 && written < LOG_BUFFER_SIZE) {
+        vsnprintf(log_buffer + written, LOG_BUFFER_SIZE - written, format, args);
+    }
+    size_t len = strlen(log_buffer);
+    if (log_buffer[len - 1] != '\n') {
+        strncat(log_buffer, "\n", LOG_BUFFER_SIZE - len);
+    }
+
+    if (sd_log_cb) {
+        sd_log_cb(level, log_buffer, sd_log_cb_data);
+    }
+
     va_end(args);
-    std::string origin = sd_format("%s:%-4d", sd_basename(file).c_str(), line);
-    sd_log_dispatch(level, origin, message);
 }
 
 void sd_ggml_log_callback(ggml_log_level level, const char* text, void*) {
-    sd_log_level_t sd_level = SD_LOG_VERBOSE;
     switch (level) {
         case GGML_LOG_LEVEL_DEBUG:
-            sd_level = SD_LOG_VERBOSE;
+            LOG_VERBOSE(text);
             break;
         case GGML_LOG_LEVEL_INFO:
-            sd_level = SD_LOG_INFO;
+            LOG_INFO(text);
             break;
         case GGML_LOG_LEVEL_WARN:
-            sd_level = SD_LOG_WARN;
+            LOG_WARN(text);
             break;
         case GGML_LOG_LEVEL_ERROR:
-            sd_level = SD_LOG_ERROR;
+            LOG_ERROR(text);
             break;
         default:
-            sd_level = SD_LOG_VERBOSE;
-            break;
+            LOG_VERBOSE(text);
     }
-    sd_log_dispatch(sd_level, "ggml", text);
 }
 
 void sd_set_log_callback(sd_log_cb_t cb, void* data) {
@@ -799,13 +755,6 @@ sd::Tensor<float> clip_preprocess(const sd::Tensor<float>& image, int target_wid
     int64_t resized_width  = static_cast<int64_t>(scale * static_cast<float>(image.shape()[0]));
     int64_t resized_height = static_cast<int64_t>(scale * static_cast<float>(image.shape()[1]));
 
-    // The resized image must cover the crop window. Floating-point rounding can
-    // leave a side one pixel short of the crop target (e.g. 730 -> 735.999...
-    // -> 735 after truncation), so clamp to keep the center crop in bounds.
-    // Truncation is otherwise preserved to avoid changing existing results.
-    resized_width  = std::max<int64_t>(resized_width, target_width);
-    resized_height = std::max<int64_t>(resized_height, target_height);
-
     sd::Tensor<float> resized = sd::ops::interpolate(
         image,
         {resized_width, resized_height, image.shape()[2], image.shape()[3]});

src/core/util.h

@@ -90,7 +90,6 @@ void log_printf(sd_log_level_t level, const char* file, int line, const char* fo
 void sd_ggml_log_callback(ggml_log_level level, const char* text, void*);
 
 ggml_type sd_type_to_ggml_type(sd_type_t sdtype);
-bool validate_tensor_types(sd_type_t type, const char* tensor_type_rules);
 
 std::string trim(const std::string& s);
 

src/detailer.cpp

@@ -970,7 +970,6 @@ bool adetail_image(adetailer_ctx_t* context,
         generation.pm_params           = {};
         generation.pulid_params        = {};
         generation.hires.enabled       = false;
-        generation.image_preprocess    = {};
         if (params.steps > 0) {
             generation.sample_params.sample_steps        = params.steps;
             generation.sample_params.custom_sigmas       = nullptr;

src/model.h

@@ -39,7 +39,6 @@ enum SDVersion {
     VERSION_LINGBOT_VIDEO,
     VERSION_QWEN_IMAGE,
     VERSION_QWEN_IMAGE_LAYERED,
-    VERSION_QWEN_IMAGE_2_1,
     VERSION_HUNYUAN_VIDEO,
     VERSION_ANIMA,
     VERSION_FLUX2,
@@ -60,9 +59,7 @@ enum SDVersion {
     VERSION_KREA2,
     VERSION_MAGE_FLOW,
     VERSION_SENSENOVA_U1_5,
-    VERSION_LLADA_IMAGE,
     VERSION_ESRGAN,
-    VERSION_PIXART,
     VERSION_COUNT,
 };
 
@@ -148,7 +145,7 @@ static inline bool sd_version_is_lingbot_video(SDVersion version) {
 }
 
 static inline bool sd_version_is_qwen_image(SDVersion version) {
-    if (version == VERSION_QWEN_IMAGE || version == VERSION_QWEN_IMAGE_LAYERED || version == VERSION_QWEN_IMAGE_2_1) {
+    if (version == VERSION_QWEN_IMAGE || version == VERSION_QWEN_IMAGE_LAYERED) {
         return true;
     }
     return false;
@@ -175,13 +172,6 @@ static inline bool sd_version_is_z_image(SDVersion version) {
     return false;
 }
 
-static inline bool sd_version_is_llada_image(SDVersion version) {
-    if (version == VERSION_LLADA_IMAGE) {
-        return true;
-    }
-    return false;
-}
-
 static inline bool sd_version_is_boogu_image(SDVersion version) {
     if (version == VERSION_BOOGU_IMAGE) {
         return true;
@@ -253,18 +243,6 @@ static inline bool sd_version_is_sensenova_u1(SDVersion version) {
     return version == VERSION_SENSENOVA_U1_5;
 }
 
-static inline bool sd_version_is_pixart(SDVersion version) {
-    return version == VERSION_PIXART;
-}
-
-static inline bool sd_version_supports_video_generation(SDVersion version) {
-    return version == VERSION_SVD || sd_version_is_wan(version) || sd_version_is_hunyuan_video(version) || sd_version_is_lingbot_video(version) || sd_version_is_ltxav(version) || sd_version_is_minimax_h3(version);
-}
-
-static inline bool sd_version_supports_image_generation(SDVersion version) {
-    return !sd_version_supports_video_generation(version);
-}
-
 static inline bool sd_version_uses_flux_vae(SDVersion version) {
     if (sd_version_is_flux(version) || sd_version_is_z_image(version) || sd_version_is_boogu_image(version) || sd_version_is_longcat(version)) {
         return true;
@@ -273,7 +251,7 @@ static inline bool sd_version_uses_flux_vae(SDVersion version) {
 }
 
 static inline bool sd_version_uses_flux2_vae(SDVersion version) {
-    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version) || sd_version_is_lens(version) || sd_version_is_ideogram4(version) || sd_version_is_sefi_image(version) || sd_version_is_llada_image(version)) {
+    if (sd_version_is_flux2(version) || sd_version_is_ernie_image(version) || sd_version_is_lens(version) || sd_version_is_ideogram4(version) || sd_version_is_sefi_image(version)) {
         return true;
     }
     return false;
@@ -314,7 +292,6 @@ static inline bool sd_version_is_dit(SDVersion version) {
         version == VERSION_HIDREAM_O1 ||
         sd_version_is_anima(version) ||
         sd_version_is_z_image(version) ||
-        sd_version_is_llada_image(version) ||
         sd_version_is_boogu_image(version) ||
         sd_version_is_ernie_image(version) ||
         sd_version_is_lens(version) ||
@@ -325,8 +302,7 @@ static inline bool sd_version_is_dit(SDVersion version) {
         sd_version_is_sefi_image(version) ||
         sd_version_is_krea2(version) ||
         sd_version_is_mage_flow(version) ||
-        sd_version_is_sensenova_u1(version) ||
-        sd_version_is_pixart(version)) {
+        sd_version_is_sensenova_u1(version)) {
         return true;
     }
     return false;

src/model/adapter/lora.hpp

@@ -67,7 +67,7 @@ struct LoraModel : public GGMLRunner {
         std::map<std::string, ggml_tensor*> scalars;
         std::set<std::string> scalar_names;
         for (const auto& [name, source] : sources) {
-            if (filter && !filter(name))
+            if (is_unused_tensor(name) || (filter && !filter(name)))
                 continue;
             const bool scalar  = source.nelements() == 1 && (ends_with(name, ".alpha") || ends_with(name, ".scale"));
             auto* tensor       = ggml_new_tensor(params_ctx, scalar ? GGML_TYPE_F32 : source.type, source.n_dims, source.ne);
@@ -163,27 +163,6 @@ struct LoraModel : public GGMLRunner {
 
             lora_tensors = std::move(new_lora_tensors);
         }
-
-        std::unordered_map<std::string, ggml_tensor*> new_lora_tensors;
-        for (const auto& [old_name, tensor] : lora_tensors) {
-            std::string new_name = old_name;
-            if (starts_with(old_name, "lora.model.diffusion_model.transformer_blocks.")) {
-                // Qwen Image 2.1 stores the gate before the projection in fused MLP weights.
-                for (const auto& suffix : {std::string(".img_mlp.gate_layer.weight."), std::string(".img_mlp.proj.weight.")}) {
-                    size_t pos = old_name.find(suffix);
-                    if (pos == std::string::npos) {
-                        continue;
-                    }
-                    std::string fused_name = old_name.substr(5, pos - 5) + ".img_mlp.gate_up.weight";
-                    if (model_tensor_names.find(fused_name) != model_tensor_names.end()) {
-                        new_name = "lora." + fused_name + (suffix == ".img_mlp.proj.weight." ? ".1." : ".") + old_name.substr(pos + suffix.size());
-                    }
-                    break;
-                }
-            }
-            new_lora_tensors[new_name] = tensor;
-        }
-        lora_tensors = std::move(new_lora_tensors);
     }
 
     ggml_tensor* get_lora_weight_diff(const std::string& model_tensor_name, ggml_context* ctx, ggml_backend_t backend) {

src/model/common/ggml_block.hpp

@@ -208,7 +208,6 @@ public:
         ggml_tensor* w            = params["weight"];
         const float scale         = ctx->linear_scale > 0.f ? ctx->linear_scale : this->scale;
         ggml_tensor* weight_scale = has_weight_scale ? params["weight_scale"] : nullptr;
-#ifndef SD_USE_UPSTREAM_GGML
         if (w->type == GGML_TYPE_F8_E4M3 || w->type == GGML_TYPE_F8_E5M2) {
             bool supports_fp8_matmul = false;
             if (ctx->backend != nullptr) {
@@ -222,7 +221,6 @@ public:
                 w = ggml_cast(ctx->ggml_ctx, w, GGML_TYPE_BF16);
             }
         }
-#endif
         ggml_tensor* b = nullptr;
         if (bias) {
             b = params["bias"];
@@ -240,7 +238,6 @@ public:
             if (ctx->weight_adapter && b != nullptr) {
                 b = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, b, prefix + "bias");
             }
-#ifndef SD_USE_UPSTREAM_GGML
             if (int8_convrot && scale == 1.f) {
                 const auto cache_key = std::make_pair(x, int8_convrot_group_size);
                 auto cached          = ctx->int8_convrot_cache.find(cache_key);
@@ -251,7 +248,6 @@ public:
                     x = cached->second;
                 }
             }
-#endif
             out = ggml_ext_linear_i8_tensorwise(ctx->ggml_ctx,
                                                 x,
                                                 w,
@@ -732,7 +728,7 @@ public:
                                 std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
                                 std::get<2>(padding), std::get<1>(padding), std::get<0>(padding),
                                 std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation),
-                                force_prec_f32, ctx->conv3d_direct_enabled);
+                                force_prec_f32);
     }
 };
 
@@ -839,30 +835,21 @@ class RMSNorm : public UnaryBlock {
 protected:
     int64_t hidden_size;
     float eps;
-    bool elementwise_affine;
     std::string prefix;
 
     void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, std::string prefix = "") override {
-        this->prefix = prefix;
-        if (!elementwise_affine) {
-            return;
-        }
+        this->prefix         = prefix;
         enum ggml_type wtype = GGML_TYPE_F32;
         params["weight"]     = ggml_new_tensor_1d(ctx, wtype, hidden_size);
     }
 
 public:
     RMSNorm(int64_t hidden_size,
-            float eps               = 1e-06f,
-            bool elementwise_affine = true)
+            float eps = 1e-06f)
         : hidden_size(hidden_size),
-          eps(eps),
-          elementwise_affine(elementwise_affine) {}
+          eps(eps) {}
 
     ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
-        if (!elementwise_affine) {
-            return ggml_rms_norm(ctx->ggml_ctx, x, eps);
-        }
         ggml_tensor* w = params["weight"];
         if (ctx->weight_adapter) {
             w = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, w, prefix + "weight");

src/model/common/rope.hpp

@@ -5,7 +5,6 @@
 #include <cassert>
 #include <cmath>
 #include <set>
-#include <utility>
 #include <vector>
 #include "core/ggml_extend.h"
 #include "core/ggml_runner.h"
@@ -17,45 +16,6 @@ namespace Rope {
         ErnieImage,
     };
 
-    struct SpatialRegion {
-        size_t begin;
-        size_t count;
-        float height_period;
-        float width_period;
-        int height_axis = 1;
-        int width_axis  = 2;
-    };
-
-    struct PositionLayout {
-        // Token ranges are relative to one batch item.
-        std::vector<SpatialRegion> images;
-        size_t token_count = 0;
-
-        void append_tokens(size_t count) {
-            token_count += count;
-        }
-
-        void append_image(int height, int width, int frames = 1, float height_step = 1.f, float width_step = 1.f) {
-            size_t count = static_cast<size_t>(height) * width * frames;
-            images.push_back({token_count, count, height * height_step, width * width_step});
-            append_tokens(count);
-        }
-    };
-
-    struct Frequency {
-        size_t axis;
-        float omega;
-    };
-
-    struct Embedding {
-        std::vector<float> values;
-        std::vector<std::vector<float>> ids;
-        PositionLayout positions;
-        std::vector<Frequency> frequencies;
-        EmbedNDLayout layout = EmbedNDLayout::Matrix;
-        int batch_size       = 1;
-    };
-
     enum class RefIndexMode {
         FIXED,
         INCREASE,
@@ -96,25 +56,40 @@ namespace Rope {
         return flat_vec;
     }
 
-    __STATIC_INLINE__ std::vector<float> rope_frequencies(int dim, float theta) {
+    __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos,
+                                                           int dim,
+                                                           float theta,
+                                                           const std::vector<int>& axis_wrap_dims = {}) {
         assert(dim % 2 == 0);
-        int half_dim             = dim / 2;
+        int half_dim = dim / 2;
+
         std::vector<float> scale = linspace(0.f, (dim * 1.f - 2) / dim, half_dim);
+
         std::vector<float> omega(half_dim);
         for (int i = 0; i < half_dim; ++i) {
             omega[i] = 1.0f / ::powf(1.f * theta, scale[i]);
         }
-        return omega;
-    }
 
-    __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos,
-                                                           const std::vector<float>& omega) {
-        int half_dim    = static_cast<int>(omega.size());
         size_t pos_size = pos.size();
         std::vector<std::vector<float>> out(pos_size, std::vector<float>(half_dim));
         for (size_t i = 0; i < pos_size; ++i) {
             for (size_t j = 0; j < half_dim; ++j) {
                 float angle = pos[i] * omega[j];
+                if (!axis_wrap_dims.empty()) {
+                    size_t wrap_size = axis_wrap_dims.size();
+                    // mod batch size since we only store this for one item in the batch
+                    size_t wrap_idx = wrap_size > 0 ? (i % wrap_size) : 0;
+                    int wrap_dim    = axis_wrap_dims[wrap_idx];
+                    if (wrap_dim > 0) {
+                        constexpr float TWO_PI = 6.28318530717958647692f;
+                        float cycles           = omega[j] * wrap_dim / TWO_PI;
+                        // closest periodic harmonic, necessary to ensure things neatly tile
+                        // without this round, things don't tile at the boundaries and you end up
+                        // with the model knowing what is "center"
+                        float rounded = std::round(cycles);
+                        angle         = pos[i] * TWO_PI * rounded / wrap_dim;
+                    }
+                }
 
                 out[i][j] = angle;
             }
@@ -133,12 +108,6 @@ namespace Rope {
         return result;
     }
 
-    __STATIC_INLINE__ std::vector<std::vector<float>> rope(const std::vector<float>& pos,
-                                                           int dim,
-                                                           float theta) {
-        return rope(pos, rope_frequencies(dim, theta));
-    }
-
     // Generate IDs for image patches and text
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_flux_txt_ids(int bs, int context_len, int axes_dim_num, std::set<int> arange_dims) {
         auto txt_ids = std::vector<std::vector<float>>(bs * context_len, std::vector<float>(axes_dim_num, 0.0f));
@@ -167,16 +136,12 @@ namespace Rope {
                                                                        int patch_size,
                                                                        int bs,
                                                                        int axes_dim_num,
-                                                                       int index              = 0,
-                                                                       int h_offset           = 0,
-                                                                       int w_offset           = 0,
-                                                                       bool scale_rope        = false,
-                                                                       PositionLayout* layout = nullptr) {
+                                                                       int index       = 0,
+                                                                       int h_offset    = 0,
+                                                                       int w_offset    = 0,
+                                                                       bool scale_rope = false) {
         int h_len = (h + (patch_size / 2)) / patch_size;
         int w_len = (w + (patch_size / 2)) / patch_size;
-        if (layout) {
-            layout->append_image(h_len, w_len);
-        }
         std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(axes_dim_num, 0.0));
 
         int h_start = h_offset;
@@ -227,8 +192,8 @@ namespace Rope {
                                                   int bs,
                                                   const std::vector<float>& axis_thetas,
                                                   const std::vector<int>& axes_dim,
-                                                  EmbedNDLayout layout                = EmbedNDLayout::Matrix,
-                                                  std::vector<Frequency>* frequencies = nullptr) {
+                                                  const std::vector<std::vector<int>>& wrap_dims = {},
+                                                  EmbedNDLayout layout                           = EmbedNDLayout::Matrix) {
         std::vector<std::vector<float>> trans_ids = transpose(ids);
         size_t pos_len                            = ids.size() / bs;
         size_t num_axes                           = axes_dim.size();
@@ -240,25 +205,19 @@ namespace Rope {
         for (int d : axes_dim)
             emb_dim += d / 2;
 
-        if (frequencies) {
-            frequencies->clear();
-            frequencies->reserve(emb_dim);
-        }
         std::vector<std::vector<float>> emb(bs * pos_len, std::vector<float>(emb_dim * 2 * 2, 0.0));
         size_t offset = 0;
         for (size_t i = 0; i < num_axes; ++i) {
+            std::vector<int> axis_wrap_dims;
+            if (!wrap_dims.empty() && i < (int)wrap_dims.size()) {
+                axis_wrap_dims = wrap_dims[i];
+            }
             float axis_theta = 10000.0f;
             if (!axis_thetas.empty()) {
                 axis_theta = axis_thetas[std::min(i, axis_thetas.size() - 1)];
             }
-            auto omega = rope_frequencies(axes_dim[i], axis_theta);
-            if (frequencies) {
-                for (float frequency : omega) {
-                    frequencies->push_back({i, frequency});
-                }
-            }
             std::vector<std::vector<float>> rope_emb =
-                rope(trans_ids[i], omega);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
+                rope(trans_ids[i], axes_dim[i], axis_theta, axis_wrap_dims);  // [bs*pos_len, axes_dim[i]/2 * 2 * 2]
             for (int b = 0; b < bs; ++b) {
                 for (int j = 0; j < pos_len; ++j) {
                     for (int k = 0; k < rope_emb[0].size(); ++k) {
@@ -294,10 +253,10 @@ namespace Rope {
                                                   int bs,
                                                   float theta,
                                                   const std::vector<int>& axes_dim,
-                                                  EmbedNDLayout layout                = EmbedNDLayout::Matrix,
-                                                  std::vector<Frequency>* frequencies = nullptr) {
+                                                  const std::vector<std::vector<int>>& wrap_dims = {},
+                                                  EmbedNDLayout layout                           = EmbedNDLayout::Matrix) {
         std::vector<float> axis_thetas(axes_dim.size(), theta);
-        return embed_nd(ids, bs, axis_thetas, axes_dim, layout, frequencies);
+        return embed_nd(ids, bs, axis_thetas, axes_dim, wrap_dims, layout);
     }
 
     __STATIC_INLINE__ std::vector<float> embed_interleaved_mrope(const std::vector<std::vector<float>>& ids,
@@ -305,7 +264,7 @@ namespace Rope {
                                                                  float theta,
                                                                  int head_dim,
                                                                  const std::vector<int>& mrope_section,
-                                                                 std::vector<Frequency>* frequencies = nullptr) {
+                                                                 const std::vector<std::vector<int>>& axis_wrap_dims = {}) {
         GGML_ASSERT(bs > 0);
         GGML_ASSERT(head_dim % 2 == 0);
         GGML_ASSERT(mrope_section.size() >= 3);
@@ -314,26 +273,20 @@ namespace Rope {
         size_t pos_len                            = ids.size() / bs;
         int half_dim                              = head_dim / 2;
 
-        auto omega = rope_frequencies(head_dim, theta);
-        if (frequencies) {
-            frequencies->clear();
-            for (float frequency : omega) {
-                frequencies->push_back({0, frequency});
-            }
-        }
         std::vector<std::vector<std::vector<float>>> axis_embs;
         axis_embs.reserve(3);
         for (int axis = 0; axis < 3; ++axis) {
-            axis_embs.push_back(rope(trans_ids[axis], omega));
+            std::vector<int> axis_wrap;
+            if (axis < static_cast<int>(axis_wrap_dims.size())) {
+                axis_wrap = axis_wrap_dims[axis];
+            }
+            axis_embs.push_back(rope(trans_ids[axis], head_dim, theta, axis_wrap));
         }
 
         std::vector<std::vector<float>> emb = axis_embs[0];
         for (int axis = 1; axis < 3; ++axis) {
             int length = std::min<int>(mrope_section[axis] * 3, half_dim);
             for (int freq_idx = axis; freq_idx < length; freq_idx += 3) {
-                if (frequencies) {
-                    (*frequencies)[freq_idx].axis = axis;
-                }
                 for (size_t pos_idx = 0; pos_idx < bs * pos_len; ++pos_idx) {
                     for (int k = 0; k < 4; ++k) {
                         emb[pos_idx][4 * freq_idx + k] = axis_embs[axis][pos_idx][4 * freq_idx + k];
@@ -345,13 +298,13 @@ namespace Rope {
         return flatten(emb);
     }
 
-    __STATIC_INLINE__ Embedding embed_2d_interleaved(int height,
-                                                     int width,
-                                                     int dim,
-                                                     float theta    = 10000.f,
-                                                     float scale    = 16.f,
-                                                     int ref_grid_h = 0,
-                                                     int ref_grid_w = 0) {
+    __STATIC_INLINE__ std::vector<float> embed_2d_interleaved(int height,
+                                                              int width,
+                                                              int dim,
+                                                              float theta    = 10000.f,
+                                                              float scale    = 16.f,
+                                                              int ref_grid_h = 0,
+                                                              int ref_grid_w = 0) {
         assert(dim % 4 == 0);
         int half_dim      = dim / 2;
         int dim_axis      = dim / 2;
@@ -365,10 +318,6 @@ namespace Rope {
             w_ntk       = std::pow(static_cast<float>(width) / static_cast<float>(ref_grid_w), power);
         }
 
-        Embedding result;
-        result.positions.append_image(height, width, 1,
-                                      height > 1 ? scale / (height - 1) : 1.f,
-                                      width > 1 ? scale / (width - 1) : 1.f);
         std::vector<float> x_pos;
         std::vector<float> y_pos;
         x_pos.reserve(static_cast<size_t>(height) * width);
@@ -377,20 +326,13 @@ namespace Rope {
             float y = height == 1 ? 0.f : scale * static_cast<float>(iy) / static_cast<float>(height - 1);
             for (int ix = 0; ix < width; ++ix) {
                 float x = width == 1 ? 0.f : scale * static_cast<float>(ix) / static_cast<float>(width - 1);
-                result.ids.push_back({0.f, y, x});
                 x_pos.push_back(x);
                 y_pos.push_back(y);
             }
         }
 
-        auto x_freq = rope_frequencies(dim_axis, theta * w_ntk);
-        auto y_freq = rope_frequencies(dim_axis, theta * h_ntk);
-        auto x_emb  = rope(x_pos, x_freq);
-        auto y_emb  = rope(y_pos, y_freq);
-        for (int i = 0; i < axis_half_dim; ++i) {
-            result.frequencies.push_back({2, x_freq[i]});
-            result.frequencies.push_back({1, y_freq[i]});
-        }
+        auto x_emb = rope(x_pos, dim_axis, theta * w_ntk);
+        auto y_emb = rope(y_pos, dim_axis, theta * h_ntk);
 
         std::vector<float> out(static_cast<size_t>(height) * width * half_dim * 4);
         for (int pos = 0; pos < height * width; ++pos) {
@@ -406,8 +348,7 @@ namespace Rope {
                 }
             }
         }
-        result.values = std::move(out);
-        return result;
+        return out;
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_refs_ids(int patch_size,
@@ -418,8 +359,7 @@ namespace Rope {
                                                                    RefIndexMode ref_index_mode,
                                                                    float ref_index_scale,
                                                                    bool scale_rope,
-                                                                   int base_offset        = 0,
-                                                                   PositionLayout* layout = nullptr) {
+                                                                   int base_offset = 0) {
         std::vector<std::vector<float>> ids;
         int curr_h_offset = 0;
         int curr_w_offset = 0;
@@ -446,8 +386,7 @@ namespace Rope {
                                             static_cast<int>(index * ref_index_scale),
                                             h_offset + base_offset,
                                             w_offset + base_offset,
-                                            scale_rope,
-                                            layout);
+                                            scale_rope);
             ids          = concat_ids(ids, ref_ids, bs);
 
             if (ref_index_mode == RefIndexMode::INCREASE) {
@@ -470,53 +409,88 @@ namespace Rope {
                                                                    const std::vector<ggml_tensor*>& ref_latents,
                                                                    RefIndexMode ref_index_mode,
                                                                    float ref_index_scale,
-                                                                   bool is_longcat,
-                                                                   PositionLayout* layout = nullptr) {
-        if (layout) {
-            layout->append_tokens(context_len);
-        }
+                                                                   bool is_longcat) {
         int x_index = is_longcat ? 1 : 0;
 
         auto txt_ids = is_longcat ? gen_longcat_txt_ids(bs, context_len, axes_dim_num) : gen_flux_txt_ids(bs, context_len, axes_dim_num, txt_arange_dims);
         int offset   = is_longcat ? context_len : 0;
-        auto img_ids = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, x_index, offset, offset, false, layout);
+        auto img_ids = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, x_index, offset, offset);
 
         auto ids = concat_ids(txt_ids, img_ids, bs);
         if (ref_latents.size() > 0) {
-            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, x_index + 1, ref_latents, ref_index_mode, ref_index_scale, false, offset, layout);
+            auto refs_ids = gen_refs_ids(patch_size, bs, axes_dim_num, x_index + 1, ref_latents, ref_index_mode, ref_index_scale, false, offset);
             ids           = concat_ids(ids, refs_ids, bs);
         }
         return ids;
     }
 
     // Generate flux positional embeddings
-    __STATIC_INLINE__ Embedding gen_flux_pe(int h,
-                                            int w,
-                                            int patch_size,
-                                            int bs,
-                                            int context_len,
-                                            std::set<int> txt_arange_dims,
-                                            const std::vector<ggml_tensor*>& ref_latents,
-                                            RefIndexMode ref_index_mode,
-                                            float ref_index_scale,
-                                            int theta,
-                                            const std::vector<int>& axes_dim,
-                                            bool is_longcat) {
-        Embedding result;
-        result.batch_size = bs;
-        result.ids        = gen_flux_ids(h,
-                                         w,
-                                         patch_size,
-                                         bs,
-                                         static_cast<int>(axes_dim.size()),
-                                         context_len,
-                                         txt_arange_dims,
-                                         ref_latents,
-                                         ref_index_mode,
-                                         ref_index_scale,
-                                         is_longcat, &result.positions);
-        result.values     = embed_nd(result.ids, bs, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
+    __STATIC_INLINE__ std::vector<float> gen_flux_pe(int h,
+                                                     int w,
+                                                     int patch_size,
+                                                     int bs,
+                                                     int context_len,
+                                                     std::set<int> txt_arange_dims,
+                                                     const std::vector<ggml_tensor*>& ref_latents,
+                                                     RefIndexMode ref_index_mode,
+                                                     float ref_index_scale,
+                                                     int theta,
+                                                     bool circular_h,
+                                                     bool circular_w,
+                                                     const std::vector<int>& axes_dim,
+                                                     bool is_longcat) {
+        std::vector<std::vector<float>> ids = gen_flux_ids(h,
+                                                           w,
+                                                           patch_size,
+                                                           bs,
+                                                           static_cast<int>(axes_dim.size()),
+                                                           context_len,
+                                                           txt_arange_dims,
+                                                           ref_latents,
+                                                           ref_index_mode,
+                                                           ref_index_scale,
+                                                           is_longcat);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int h_len = (h + (patch_size / 2)) / patch_size;
+            int w_len = (w + (patch_size / 2)) / patch_size;
+            if (h_len > 0 && w_len > 0) {
+                size_t pos_len = ids.size() / bs;
+                wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+                size_t cursor           = context_len;  // text first
+                const size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                    if (circular_h) {
+                        wrap_dims[1][cursor + token_i] = h_len;
+                    }
+                    if (circular_w) {
+                        wrap_dims[2][cursor + token_i] = w_len;
+                    }
+                }
+                cursor += img_tokens;
+                // reference latents
+                for (ggml_tensor* ref : ref_latents) {
+                    if (ref == nullptr) {
+                        continue;
+                    }
+                    int ref_h         = static_cast<int>(ref->ne[1]);
+                    int ref_w         = static_cast<int>(ref->ne[0]);
+                    int ref_h_l       = (ref_h + (patch_size / 2)) / patch_size;
+                    int ref_w_l       = (ref_w + (patch_size / 2)) / patch_size;
+                    size_t ref_tokens = static_cast<size_t>(ref_h_l) * static_cast<size_t>(ref_w_l);
+                    for (size_t token_i = 0; token_i < ref_tokens; ++token_i) {
+                        if (circular_h) {
+                            wrap_dims[1][cursor + token_i] = ref_h_l;
+                        }
+                        if (circular_w) {
+                            wrap_dims[2][cursor + token_i] = ref_w_l;
+                        }
+                    }
+                    cursor += ref_tokens;
+                }
+            }
+        }
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_vid_ids(int t,
@@ -526,18 +500,14 @@ namespace Rope {
                                                                   int ph,
                                                                   int pw,
                                                                   int bs,
-                                                                  int t_offset           = 0,
-                                                                  int h_offset           = 0,
-                                                                  int w_offset           = 0,
-                                                                  bool scale_rope        = false,
-                                                                  PositionLayout* layout = nullptr) {
+                                                                  int t_offset    = 0,
+                                                                  int h_offset    = 0,
+                                                                  int w_offset    = 0,
+                                                                  bool scale_rope = false) {
         int t_len = (t + (pt / 2)) / pt;
         int h_len = (h + (ph / 2)) / ph;
         int w_len = (w + (pw / 2)) / pw;
 
-        if (layout) {
-            layout->append_image(h_len, w_len, t_len);
-        }
         std::vector<std::vector<float>> vid_ids(t_len * h_len * w_len, std::vector<float>(3, 0.0));
 
         if (scale_rope) {
@@ -603,11 +573,7 @@ namespace Rope {
                                                                          int bs,
                                                                          int context_len,
                                                                          const std::vector<ggml_tensor*>& ref_latents,
-                                                                         RefIndexMode ref_index_mode,
-                                                                         PositionLayout* layout = nullptr) {
-        if (layout) {
-            layout->append_tokens(context_len);
-        }
+                                                                         RefIndexMode ref_index_mode) {
         int h_len        = (h + (patch_size / 2)) / patch_size;
         int w_len        = (w + (patch_size / 2)) / patch_size;
         int txt_id_start = std::max(h_len, w_len) / 2;
@@ -619,49 +585,90 @@ namespace Rope {
             }
         }
         int axes_dim_num = 3;
-        auto img_ids     = gen_vid_ids(t, h, w, 1, patch_size, patch_size, bs, 0, 0, 0, true, layout);
+        auto img_ids     = gen_vid_ids(t, h, w, 1, patch_size, patch_size, bs, 0, 0, 0, true);
         auto ids         = concat_ids(txt_ids_repeated, img_ids, bs);
         if (ref_latents.size() > 0) {
             int ref_start_index = ref_index_mode == RefIndexMode::DECREASE ? 0 : 1;
-            auto refs_ids       = gen_refs_ids(patch_size, bs, axes_dim_num, ref_start_index, ref_latents, ref_index_mode, 1.f, true, 0, layout);
+            auto refs_ids       = gen_refs_ids(patch_size, bs, axes_dim_num, ref_start_index, ref_latents, ref_index_mode, 1.f, true);
             ids                 = concat_ids(ids, refs_ids, bs);
         }
         return ids;
     }
 
     // Generate qwen_image positional embeddings
-    __STATIC_INLINE__ Embedding gen_qwen_image_pe(int t,
-                                                  int h,
-                                                  int w,
-                                                  int patch_size,
-                                                  int bs,
-                                                  int context_len,
-                                                  const std::vector<ggml_tensor*>& ref_latents,
-                                                  RefIndexMode ref_index_mode,
-                                                  int theta,
-                                                  const std::vector<int>& axes_dim) {
-        Embedding result;
-        result.batch_size = bs;
-        result.ids        = gen_qwen_image_ids(t, h, w, patch_size, bs, context_len, ref_latents, ref_index_mode, &result.positions);
-        result.values     = embed_nd(result.ids, bs, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
+    __STATIC_INLINE__ std::vector<float> gen_qwen_image_pe(int t,
+                                                           int h,
+                                                           int w,
+                                                           int patch_size,
+                                                           int bs,
+                                                           int context_len,
+                                                           const std::vector<ggml_tensor*>& ref_latents,
+                                                           RefIndexMode ref_index_mode,
+                                                           int theta,
+                                                           bool circular_h,
+                                                           bool circular_w,
+                                                           const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_qwen_image_ids(t, h, w, patch_size, bs, context_len, ref_latents, ref_index_mode);
+        std::vector<std::vector<int>> wrap_dims;
+        // This logic simply stores the (pad and patch_adjusted) sizes of images so we can make sure rope correctly tiles
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int pad_h = (patch_size - (h % patch_size)) % patch_size;
+            int pad_w = (patch_size - (w % patch_size)) % patch_size;
+            int h_len = (h + pad_h) / patch_size;
+            int w_len = (w + pad_w) / patch_size;
+            if (h_len > 0 && w_len > 0) {
+                const size_t total_tokens = ids.size();
+                // Track per-token wrap lengths for the row/column axes so only spatial tokens become periodic.
+                wrap_dims.assign(axes_dim.size(), std::vector<int>(total_tokens / bs, 0));
+                size_t cursor           = context_len;  // ignore text tokens
+                const size_t img_tokens = static_cast<size_t>(t) * static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                    if (circular_h) {
+                        wrap_dims[1][cursor + token_i] = h_len;
+                    }
+                    if (circular_w) {
+                        wrap_dims[2][cursor + token_i] = w_len;
+                    }
+                }
+                cursor += img_tokens;
+                // For each reference image, store wrap sizes as well
+                for (ggml_tensor* ref : ref_latents) {
+                    if (ref == nullptr) {
+                        continue;
+                    }
+                    int ref_h           = static_cast<int>(ref->ne[1]);
+                    int ref_w           = static_cast<int>(ref->ne[0]);
+                    int ref_pad_h       = (patch_size - (ref_h % patch_size)) % patch_size;
+                    int ref_pad_w       = (patch_size - (ref_w % patch_size)) % patch_size;
+                    int ref_h_len       = (ref_h + ref_pad_h) / patch_size;
+                    int ref_w_len       = (ref_w + ref_pad_w) / patch_size;
+                    size_t ref_n_tokens = static_cast<size_t>(ref_h_len) * static_cast<size_t>(ref_w_len);
+                    for (size_t token_i = 0; token_i < ref_n_tokens; ++token_i) {
+                        if (circular_h) {
+                            wrap_dims[1][cursor + token_i] = ref_h_len;
+                        }
+                        if (circular_w) {
+                            wrap_dims[2][cursor + token_i] = ref_w_len;
+                        }
+                    }
+                    cursor += ref_n_tokens;
+                }
+            }
+        }
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
-    __STATIC_INLINE__ Embedding gen_mage_flow_pe(int h,
-                                                 int w,
-                                                 int bs,
-                                                 int context_len,
-                                                 const std::vector<ggml_tensor*>& ref_latents,
-                                                 int theta,
-                                                 const std::vector<int>& axes_dim) {
-        Embedding result;
-        result.batch_size = bs;
-        result.positions.append_tokens(context_len);
+    __STATIC_INLINE__ std::vector<float> gen_mage_flow_pe(int h,
+                                                          int w,
+                                                          int bs,
+                                                          int context_len,
+                                                          const std::vector<ggml_tensor*>& ref_latents,
+                                                          int theta,
+                                                          const std::vector<int>& axes_dim) {
         const int axes_dim_num = static_cast<int>(axes_dim.size());
-        auto make_image_ids    = [=, &result](int image_h, int image_w, int image_index) {
+        auto make_image_ids    = [=](int image_h, int image_w, int image_index) {
             std::vector<std::vector<float>> image_ids(static_cast<size_t>(bs) * image_h * image_w,
                                                          std::vector<float>(axes_dim_num, 0.f));
-            result.positions.append_image(image_h, image_w);
             int h_start = -(image_h - image_h / 2);
             int w_start = -(image_w - image_w / 2);
             for (int b = 0; b < bs; ++b) {
@@ -685,18 +692,15 @@ namespace Rope {
                                           static_cast<int>(i + 1));
             ids          = concat_ids(ids, ref_ids, bs);
         }
-        result.ids    = std::move(ids);
-        result.values = embed_nd(result.ids, bs, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_lens_ids(int h,
                                                                    int w,
                                                                    int bs,
                                                                    int context_len,
-                                                                   bool scale_rope        = true,
-                                                                   PositionLayout* layout = nullptr) {
-        auto img_ids_repeated = gen_flux_img_ids(h, w, 1, bs, 3, 0, 0, 0, scale_rope, layout);
+                                                                   bool scale_rope = true) {
+        auto img_ids_repeated = gen_flux_img_ids(h, w, 1, bs, 3, 0, 0, 0, scale_rope);
 
         int txt_id_start = scale_rope ? std::max(h / 2, w / 2) : 0;
         auto txt_ids     = linspace<float>(1.f * txt_id_start, 1.f * context_len + txt_id_start, context_len);
@@ -707,37 +711,44 @@ namespace Rope {
             }
         }
 
-        if (layout) {
-            layout->append_tokens(context_len);
-        }
         return concat_ids(img_ids_repeated, txt_ids_repeated, bs);
     }
 
-    __STATIC_INLINE__ Embedding gen_lens_pe(int h,
-                                            int w,
-                                            int bs,
-                                            int context_len,
-                                            int theta,
-                                            const std::vector<int>& axes_dim) {
-        Embedding result;
-        result.batch_size = bs;
-        result.ids        = gen_lens_ids(h, w, bs, context_len, true, &result.positions);
-        result.values     = embed_nd(result.ids, bs, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
+    __STATIC_INLINE__ std::vector<float> gen_lens_pe(int h,
+                                                     int w,
+                                                     int bs,
+                                                     int context_len,
+                                                     int theta,
+                                                     bool circular_h,
+                                                     bool circular_w,
+                                                     const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_lens_ids(h, w, bs, context_len, true);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            size_t pos_len = ids.size() / bs;
+            wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+            const size_t img_tokens = static_cast<size_t>(h) * static_cast<size_t>(w);
+            for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                if (circular_h) {
+                    wrap_dims[1][token_i] = h;
+                }
+                if (circular_w) {
+                    wrap_dims[2][token_i] = w;
+                }
+            }
+        }
+
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ std::vector<std::vector<float>> gen_ernie_image_ids(int h,
                                                                           int w,
                                                                           int patch_size,
                                                                           int bs,
-                                                                          int context_len,
-                                                                          PositionLayout* layout = nullptr) {
+                                                                          int context_len) {
         int h_len = h / patch_size;
         int w_len = w / patch_size;
 
-        if (layout) {
-            layout->append_image(h_len, w_len);
-        }
         std::vector<std::vector<float>> img_ids(h_len * w_len, std::vector<float>(3, 0.0f));
         std::vector<float> h_ids = linspace<float>(0.f, static_cast<float>(h_len - 1), h_len);
         std::vector<float> w_ids = linspace<float>(0.f, static_cast<float>(w_len - 1), w_len);
@@ -763,25 +774,39 @@ namespace Rope {
             }
         }
 
-        if (layout) {
-            layout->append_tokens(context_len);
-        }
         return concat_ids(img_ids_repeated, txt_ids, bs);
     }
 
-    __STATIC_INLINE__ Embedding gen_ernie_image_pe(int h,
-                                                   int w,
-                                                   int patch_size,
-                                                   int bs,
-                                                   int context_len,
-                                                   int theta,
-                                                   const std::vector<int>& axes_dim) {
-        Embedding result;
-        result.batch_size = bs;
-        result.layout     = EmbedNDLayout::ErnieImage;
-        result.ids        = gen_ernie_image_ids(h, w, patch_size, bs, context_len, &result.positions);
-        result.values     = embed_nd(result.ids, bs, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
+    __STATIC_INLINE__ std::vector<float> gen_ernie_image_pe(int h,
+                                                            int w,
+                                                            int patch_size,
+                                                            int bs,
+                                                            int context_len,
+                                                            int theta,
+                                                            bool circular_h,
+                                                            bool circular_w,
+                                                            const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids = gen_ernie_image_ids(h, w, patch_size, bs, context_len);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int h_len = h / patch_size;
+            int w_len = w / patch_size;
+            if (h_len > 0 && w_len > 0) {
+                size_t pos_len = ids.size() / bs;
+                wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+                const size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                    if (circular_h) {
+                        wrap_dims[1][token_i] = h_len;
+                    }
+                    if (circular_w) {
+                        wrap_dims[2][token_i] = w_len;
+                    }
+                }
+            }
+        }
+
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims, EmbedNDLayout::ErnieImage);
     }
 
     // Generate wan positional embeddings
@@ -880,8 +905,7 @@ namespace Rope {
                                                                       int context_len,
                                                                       int seq_multi_of,
                                                                       const std::vector<ggml_tensor*>& ref_latents,
-                                                                      RefIndexMode ref_index_mode,
-                                                                      PositionLayout* layout = nullptr) {
+                                                                      RefIndexMode ref_index_mode) {
         SD_UNUSED(ref_index_mode);
         int padded_context_len = context_len + bound_mod(context_len, seq_multi_of);
         auto txt_ids           = std::vector<std::vector<float>>(bs * padded_context_len, std::vector<float>(3, 0.0f));
@@ -889,17 +913,11 @@ namespace Rope {
             txt_ids[i][0] = (i % padded_context_len) + 1.f;
         }
 
-        if (layout) {
-            layout->append_tokens(padded_context_len);
-        }
         int axes_dim_num = 3;
         int index        = padded_context_len + 1;
-        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, index, 0, 0, false, layout);
+        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, index);
 
         int img_pad_len = bound_mod(static_cast<int>(img_ids.size() / bs), seq_multi_of);
-        if (layout) {
-            layout->append_tokens(img_pad_len);
-        }
         if (img_pad_len > 0) {
             std::vector<std::vector<float>> img_pad_ids(bs * img_pad_len, std::vector<float>(3, 0.f));
             img_ids = concat_ids(img_ids, img_pad_ids, bs);
@@ -911,160 +929,43 @@ namespace Rope {
         return ids;
     }
 
-    // LLaDA-Image shares Lumina2/z_image's axes layout, but assigns position (0,0,0) to the
-    // padding slots of the caption stream instead of continuing the caption ramp through them.
-    __STATIC_INLINE__ std::vector<std::vector<float>> gen_llada_image_ids(int h,
-                                                                          int w,
-                                                                          int patch_size,
-                                                                          int bs,
-                                                                          int context_len,
-                                                                          int seq_multi_of,
-                                                                          PositionLayout* layout = nullptr) {
-        int context_pad_len    = bound_mod(context_len, seq_multi_of);
-        int padded_context_len = context_len + context_pad_len;
-        auto txt_ids           = std::vector<std::vector<float>>(bs * padded_context_len, std::vector<float>(3, 0.0f));
-        for (int i = 0; i < bs * padded_context_len; i++) {
-            int pos = i % padded_context_len;
-            if (pos < context_len) {
-                txt_ids[i][0] = pos + 1.f;
-            }
-        }
-
-        if (layout) {
-            layout->append_tokens(padded_context_len);
-        }
-        int axes_dim_num = 3;
-        int index        = padded_context_len + 1;
-        auto img_ids     = gen_flux_img_ids(h, w, patch_size, bs, axes_dim_num, index, 0, 0, false, layout);
-
-        int img_pad_len = bound_mod(static_cast<int>(img_ids.size() / bs), seq_multi_of);
-        if (layout) {
-            layout->append_tokens(img_pad_len);
-        }
-        if (img_pad_len > 0) {
-            std::vector<std::vector<float>> img_pad_ids(bs * img_pad_len, std::vector<float>(3, 0.f));
-            img_ids = concat_ids(img_ids, img_pad_ids, bs);
-        }
-
-        return concat_ids(txt_ids, img_ids, bs);
-    }
-
-    // LLaDA-Image editing packs two caption copies (clean and noisy), the source and target
-    // latents anchored at their own caption's end position, and the SigVQ stream after both.
-    // Padding slots keep position (0,0,0), as in the text-only layout.
-    __STATIC_INLINE__ std::vector<std::vector<float>> gen_llada_image_edit_ids(int h,
-                                                                               int w,
-                                                                               int patch_size,
-                                                                               int context_len,
-                                                                               int sigvq_len,
-                                                                               int seq_multi_of,
-                                                                               PositionLayout* layout = nullptr) {
-        const int context_pad    = bound_mod(context_len, seq_multi_of);
-        const int padded_context = context_len + context_pad;
-        const int h_len          = (h + (patch_size / 2)) / patch_size;
-        const int w_len          = (w + (patch_size / 2)) / patch_size;
-        const int image_len      = h_len * w_len;
-        const int image_pad      = bound_mod(image_len, seq_multi_of);
-        const int padded_image   = image_len + image_pad;
-        const int sigvq_pad      = bound_mod(sigvq_len, seq_multi_of);
-
-        std::vector<std::vector<float>> cap_ids;
-        std::vector<int> cap_end_positions;
-        int cursor = 1;
-        for (int copy = 0; copy < 2; ++copy) {
-            for (int i = 0; i < padded_context; ++i) {
-                std::vector<float> id(3, 0.f);
-                if (i < context_len) {
-                    id[0] = static_cast<float>(cursor + i);
-                }
-                cap_ids.push_back(id);
-            }
-            cursor += context_len;
-            cap_end_positions.push_back(cursor);
-            cursor += 2;
-        }
-
-        if (layout) {
-            layout->append_tokens(cap_ids.size());
-        }
-        std::vector<std::vector<float>> img_ids;
-        for (int copy = 0; copy < 2; ++copy) {
-            auto ids = gen_flux_img_ids(h, w, patch_size, 1, 3, cap_end_positions[copy], 0, 0, false, layout);
-            img_ids.insert(img_ids.end(), ids.begin(), ids.end());
-            img_ids.insert(img_ids.end(), image_pad, std::vector<float>(3, 0.f));
-            if (layout) {
-                layout->append_tokens(image_pad);
-            }
-        }
-
-        const int sigvq_start = static_cast<int>(cap_ids.size() + img_ids.size()) + 1;
-        std::vector<std::vector<float>> sigvq_ids;
-        for (int i = 0; i < sigvq_len + sigvq_pad; ++i) {
-            std::vector<float> id(3, 0.f);
-            if (i < sigvq_len) {
-                id[0] = static_cast<float>(sigvq_start + i);
-            }
-            sigvq_ids.push_back(id);
-        }
-
-        std::vector<std::vector<float>> ids;
-        ids.reserve(cap_ids.size() + img_ids.size() + sigvq_ids.size());
-        ids.insert(ids.end(), cap_ids.begin(), cap_ids.end());
-        ids.insert(ids.end(), img_ids.begin(), img_ids.end());
-        ids.insert(ids.end(), sigvq_ids.begin(), sigvq_ids.end());
-        if (layout) {
-            layout->append_tokens(sigvq_ids.size());
-        }
-        SD_UNUSED(padded_image);
-        return ids;
-    }
-
-    __STATIC_INLINE__ Embedding gen_llada_image_edit_pe(int h,
+    // Generate z_image positional embeddings
+    __STATIC_INLINE__ std::vector<float> gen_z_image_pe(int h,
                                                         int w,
                                                         int patch_size,
+                                                        int bs,
                                                         int context_len,
-                                                        int sigvq_len,
                                                         int seq_multi_of,
+                                                        const std::vector<ggml_tensor*>& ref_latents,
+                                                        RefIndexMode ref_index_mode,
                                                         int theta,
+                                                        bool circular_h,
+                                                        bool circular_w,
                                                         const std::vector<int>& axes_dim) {
-        Embedding result;
-        result.batch_size = 1;
-        result.ids        = gen_llada_image_edit_ids(h, w, patch_size, context_len, sigvq_len, seq_multi_of, &result.positions);
-        result.values     = embed_nd(result.ids, 1, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
-    }
+        std::vector<std::vector<float>> ids = gen_z_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, ref_latents, ref_index_mode);
+        std::vector<std::vector<int>> wrap_dims;
+        if ((circular_h || circular_w) && bs > 0 && axes_dim.size() >= 3) {
+            int pad_h = (patch_size - (h % patch_size)) % patch_size;
+            int pad_w = (patch_size - (w % patch_size)) % patch_size;
+            int h_len = (h + pad_h) / patch_size;
+            int w_len = (w + pad_w) / patch_size;
+            if (h_len > 0 && w_len > 0) {
+                size_t pos_len = ids.size() / bs;
+                wrap_dims.assign(axes_dim.size(), std::vector<int>(pos_len, 0));
+                size_t cursor     = context_len + bound_mod(context_len, seq_multi_of);  // skip text (and its padding)
+                size_t img_tokens = static_cast<size_t>(h_len) * static_cast<size_t>(w_len);
+                for (size_t token_i = 0; token_i < img_tokens; ++token_i) {
+                    if (circular_h) {
+                        wrap_dims[1][cursor + token_i] = h_len;
+                    }
+                    if (circular_w) {
+                        wrap_dims[2][cursor + token_i] = w_len;
+                    }
+                }
+            }
+        }
 
-    __STATIC_INLINE__ Embedding gen_llada_image_pe(int h,
-                                                   int w,
-                                                   int patch_size,
-                                                   int bs,
-                                                   int context_len,
-                                                   int seq_multi_of,
-                                                   int theta,
-                                                   const std::vector<int>& axes_dim) {
-        Embedding result;
-        result.batch_size = bs;
-        result.ids        = gen_llada_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, &result.positions);
-        result.values     = embed_nd(result.ids, bs, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
-    }
-
-    // Generate z_image positional embeddings
-    __STATIC_INLINE__ Embedding gen_z_image_pe(int h,
-                                               int w,
-                                               int patch_size,
-                                               int bs,
-                                               int context_len,
-                                               int seq_multi_of,
-                                               const std::vector<ggml_tensor*>& ref_latents,
-                                               RefIndexMode ref_index_mode,
-                                               int theta,
-                                               const std::vector<int>& axes_dim) {
-        Embedding result;
-        result.batch_size = bs;
-        result.ids        = gen_z_image_ids(h, w, patch_size, bs, context_len, seq_multi_of, ref_latents, ref_index_mode, &result.positions);
-        result.values     = embed_nd(result.ids, bs, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
+        return embed_nd(ids, bs, static_cast<float>(theta), axes_dim, wrap_dims);
     }
 
     __STATIC_INLINE__ ggml_tensor* apply_rope(ggml_context* ctx,

src/model/common/rope_circular.hpp

@@ -1,65 +0,0 @@
-#ifndef __SD_MODEL_COMMON_ROPE_CIRCULAR_HPP__
-#define __SD_MODEL_COMMON_ROPE_CIRCULAR_HPP__
-
-#include "model/common/rope.hpp"
-
-namespace Rope {
-    __STATIC_INLINE__ void apply_circular(Embedding& embedding, bool circular_x, bool circular_y) {
-        if (!circular_x && !circular_y) {
-            return;
-        }
-
-        GGML_ASSERT(embedding.batch_size > 0);
-        GGML_ASSERT(embedding.ids.size() % embedding.batch_size == 0);
-        size_t pos_len  = embedding.ids.size() / embedding.batch_size;
-        size_t half_dim = embedding.frequencies.size();
-        GGML_ASSERT(embedding.positions.token_count == pos_len);
-        GGML_ASSERT(embedding.values.size() == embedding.ids.size() * half_dim * 4);
-
-        constexpr float TWO_PI = 6.28318530717958647692f;
-        for (const auto& region : embedding.positions.images) {
-            GGML_ASSERT(region.begin <= pos_len && region.count <= pos_len - region.begin);
-            for (size_t j = 0; j < half_dim; ++j) {
-                const auto& frequency = embedding.frequencies[j];
-                float period          = 0.f;
-                if (circular_y && frequency.axis == static_cast<size_t>(region.height_axis)) {
-                    period = region.height_period;
-                } else if (circular_x && frequency.axis == static_cast<size_t>(region.width_axis)) {
-                    period = region.width_period;
-                }
-                if (period <= 0) {
-                    continue;
-                }
-
-                // Quantize to periodic harmonics while preserving the original coordinate offsets.
-                float rounded = std::round(frequency.omega * period / TWO_PI);
-                for (int b = 0; b < embedding.batch_size; ++b) {
-                    size_t begin = b * pos_len + region.begin;
-                    for (size_t i = begin; i < begin + region.count; ++i) {
-                        GGML_ASSERT(frequency.axis < embedding.ids[i].size());
-                        float angle   = embedding.ids[i][frequency.axis] * TWO_PI * rounded / period;
-                        float cos_val = std::cos(angle);
-                        float sin_val = std::sin(angle);
-                        if (embedding.layout == EmbedNDLayout::ErnieImage) {
-                            size_t cos_offset                = (i * half_dim + j) * 2;
-                            size_t sin_offset                = embedding.ids.size() * half_dim * 2 + cos_offset;
-                            embedding.values[cos_offset]     = cos_val;
-                            embedding.values[cos_offset + 1] = cos_val;
-                            embedding.values[sin_offset]     = sin_val;
-                            embedding.values[sin_offset + 1] = sin_val;
-                        } else {
-                            size_t offset                = (i * half_dim + j) * 4;
-                            embedding.values[offset]     = cos_val;
-                            embedding.values[offset + 1] = -sin_val;
-                            embedding.values[offset + 2] = sin_val;
-                            embedding.values[offset + 3] = cos_val;
-                        }
-                    }
-                }
-            }
-        }
-    }
-
-}  // namespace Rope
-
-#endif  // __SD_MODEL_COMMON_ROPE_CIRCULAR_HPP__

src/model/diffusion/anima.hpp

@@ -603,37 +603,34 @@ namespace Anima {
             return std::pow(extrapolation_ratio, static_cast<float>(axis_dim) / static_cast<float>(axis_dim - 2));
         }
 
-        static Rope::Embedding gen_anima_image_pe_vec(int bs,
-                                                      int h,
-                                                      int w,
-                                                      int patch_size,
-                                                      int theta,
-                                                      const std::vector<int>& axes_dim,
-                                                      float h_extrapolation_ratio,
-                                                      float w_extrapolation_ratio,
-                                                      float t_extrapolation_ratio,
-                                                      const std::vector<ggml_tensor*>& ref_latents) {
-            Rope::Embedding result;
-            result.batch_size = bs;
-            result.ids        = Rope::gen_flux_ids(h,
-                                                   w,
-                                                   patch_size,
-                                                   bs,
-                                                   static_cast<int>(axes_dim.size()),
-                                                   0,
-                                                   {},
-                                                   ref_latents,
-                                                   Rope::RefIndexMode::FIXED,
-                                                   1.0f,
-                                                   false, &result.positions);
+        static std::vector<float> gen_anima_image_pe_vec(int bs,
+                                                         int h,
+                                                         int w,
+                                                         int patch_size,
+                                                         int theta,
+                                                         const std::vector<int>& axes_dim,
+                                                         float h_extrapolation_ratio,
+                                                         float w_extrapolation_ratio,
+                                                         float t_extrapolation_ratio,
+                                                         const std::vector<ggml_tensor*>& ref_latents) {
+            auto ids = Rope::gen_flux_ids(h,
+                                          w,
+                                          patch_size,
+                                          bs,
+                                          static_cast<int>(axes_dim.size()),
+                                          0,
+                                          {},
+                                          ref_latents,
+                                          Rope::RefIndexMode::FIXED,
+                                          1.0f,
+                                          false);
 
             std::vector<float> axis_thetas = {
                 static_cast<float>(theta) * calc_ntk_factor(t_extrapolation_ratio, axes_dim[0]),
                 static_cast<float>(theta) * calc_ntk_factor(h_extrapolation_ratio, axes_dim[1]),
                 static_cast<float>(theta) * calc_ntk_factor(w_extrapolation_ratio, axes_dim[2]),
             };
-            result.values = Rope::embed_nd(result.ids, bs, axis_thetas, axes_dim, result.layout, &result.frequencies);
-            return result;
+            return Rope::embed_nd(ids, bs, axis_thetas, axes_dim);
         }
 
         ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
@@ -660,16 +657,16 @@ namespace Anima {
             int64_t h_pad = x->ne[1] + pad_h;
             int64_t w_pad = x->ne[0] + pad_w;
 
-            image_pe_vec          = finish_rope_pe(gen_anima_image_pe_vec(1,
-                                                                          static_cast<int>(h_pad),
-                                                                          static_cast<int>(w_pad),
-                                                                          static_cast<int>(config.patch_size),
-                                                                          config.theta,
-                                                                          config.axes_dim,
-                                                                          4.0f,
-                                                                          4.0f,
-                                                                          1.0f,
-                                                                          ref_latents));
+            image_pe_vec          = gen_anima_image_pe_vec(1,
+                                                           static_cast<int>(h_pad),
+                                                           static_cast<int>(w_pad),
+                                                           static_cast<int>(config.patch_size),
+                                                           config.theta,
+                                                           config.axes_dim,
+                                                           4.0f,
+                                                           4.0f,
+                                                           1.0f,
+                                                           ref_latents);
             int64_t image_pos_len = static_cast<int64_t>(image_pe_vec.size()) / (2 * 2 * (config.head_dim / 2));
             auto image_pe         = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2, image_pos_len);
             set_backend_tensor_data(image_pe, image_pe_vec.data());

src/model/diffusion/boogu.hpp

@@ -720,18 +720,15 @@ namespace Boogu {
         }
     }
 
-    __STATIC_INLINE__ Rope::Embedding gen_boogu_pe(int h,
-                                                   int w,
-                                                   int patch_size,
-                                                   int bs,
-                                                   int context_len,
-                                                   const std::vector<ggml_tensor*>& ref_latents,
-                                                   int theta,
-                                                   const std::vector<int>& axes_dim) {
-        Rope::Embedding result;
-        result.batch_size = bs;
-        result.positions.append_tokens(context_len);
-        auto& ids = result.ids;
+    __STATIC_INLINE__ std::vector<float> gen_boogu_pe(int h,
+                                                      int w,
+                                                      int patch_size,
+                                                      int bs,
+                                                      int context_len,
+                                                      const std::vector<ggml_tensor*>& ref_latents,
+                                                      int theta,
+                                                      const std::vector<int>& axes_dim) {
+        std::vector<std::vector<float>> ids;
         ids.reserve(static_cast<size_t>(bs) * context_len);
         for (int b = 0; b < bs; b++) {
             for (int i = 0; i < context_len; i++) {
@@ -744,18 +741,15 @@ namespace Boogu {
         for (ggml_tensor* ref : ref_latents) {
             int ref_h_tokens = patched_token_count(ref->ne[1], patch_size);
             int ref_w_tokens = patched_token_count(ref->ne[0], patch_size);
-            result.positions.append_image(ref_h_tokens, ref_w_tokens);
             append_spatial_ids(ids, bs, pe_shift, ref_h_tokens, ref_w_tokens);
             pe_shift += std::max(ref_h_tokens, ref_w_tokens);
         }
 
         int h_tokens = patched_token_count(h, patch_size);
         int w_tokens = patched_token_count(w, patch_size);
-        result.positions.append_image(h_tokens, w_tokens);
         append_spatial_ids(ids, bs, pe_shift, h_tokens, w_tokens);
 
-        result.values = Rope::embed_nd(ids, bs, static_cast<float>(theta), axes_dim, result.layout, &result.frequencies);
-        return result;
+        return Rope::embed_nd(ids, bs, static_cast<float>(theta), axes_dim);
     }
 
     struct BooguImageRunner : public DiffusionModelRunner {
@@ -799,14 +793,14 @@ namespace Boogu {
                 ref_latents.push_back(make_input(ref_latent_tensor));
             }
 
-            pe_vec      = finish_rope_pe(gen_boogu_pe(static_cast<int>(x->ne[1]),
-                                                      static_cast<int>(x->ne[0]),
-                                                      config.patch_size,
-                                                      static_cast<int>(x->ne[3]),
-                                                      static_cast<int>(context->ne[1]),
-                                                      ref_latents,
-                                                      config.theta,
-                                                      config.axes_dim));
+            pe_vec      = gen_boogu_pe(static_cast<int>(x->ne[1]),
+                                       static_cast<int>(x->ne[0]),
+                                       config.patch_size,
+                                       static_cast<int>(x->ne[3]),
+                                       static_cast<int>(context->ne[1]),
+                                       ref_latents,
+                                       config.theta,
+                                       config.axes_dim);
             int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
             set_backend_tensor_data(pe, pe_vec.data());

src/model/diffusion/control.hpp

@@ -376,7 +376,7 @@ struct ControlNet : public GGMLRunner {
             hint = make_input(hint_tensor);
         }
 
-        auto runner_ctx = get_context(gf);
+        auto runner_ctx = get_context();
 
         auto outs = control_net.forward(&runner_ctx,
                                         x,
@@ -389,7 +389,8 @@ struct ControlNet : public GGMLRunner {
         if (guided_hint_input == nullptr && !outs.empty()) {
             guided_hint_output_ggml = outs[0];
             ggml_set_output(guided_hint_output_ggml);
-            runner_ctx.persist_cache_tensor(guided_hint_cache_name(), guided_hint_output_ggml);
+            cache(guided_hint_cache_name(), guided_hint_output_ggml);
+            ggml_build_forward_expand(gf, guided_hint_output_ggml);
         }
 
         control_outputs_ggml.reserve(outs.size() > 0 ? outs.size() - 1 : 0);

src/model/diffusion/ernie_image.hpp

@@ -415,13 +415,15 @@ namespace ErnieImage {
             GGML_ASSERT(!context_tensor.empty());
             ggml_tensor* context = make_input(context_tensor);
 
-            pe_vec      = finish_rope_pe(Rope::gen_ernie_image_pe(static_cast<int>(x->ne[1]),
-                                                                  static_cast<int>(x->ne[0]),
-                                                                  config.patch_size,
-                                                                  static_cast<int>(x->ne[3]),
-                                                                  static_cast<int>(context->ne[1]),
-                                                                  config.theta,
-                                                                  config.axes_dim));
+            pe_vec      = Rope::gen_ernie_image_pe(static_cast<int>(x->ne[1]),
+                                                   static_cast<int>(x->ne[0]),
+                                                   config.patch_size,
+                                                   static_cast<int>(x->ne[3]),
+                                                   static_cast<int>(context->ne[1]),
+                                                   config.theta,
+                                                   circular_y_enabled,
+                                                   circular_x_enabled,
+                                                   config.axes_dim);
             int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, config.axes_dim_sum, 1, pos_len, 2);
             set_backend_tensor_data(pe, pe_vec.data());

src/model/diffusion/flux.hpp

@@ -1548,18 +1548,20 @@ namespace Flux {
             } else if (version == VERSION_OVIS_IMAGE) {
                 txt_arange_dims = {1, 2};
             }
-            pe_vec      = finish_rope_pe(Rope::gen_flux_pe(static_cast<int>(x->ne[1]),
-                                                           static_cast<int>(x->ne[0]),
-                                                           config.patch_size,
-                                                           static_cast<int>(x->ne[3]),
-                                                           static_cast<int>(context->ne[1]),
-                                                           txt_arange_dims,
-                                                           ref_latents,
-                                                           ref_index_mode,
-                                                           config.ref_index_scale,
-                                                           config.theta,
-                                                           config.axes_dim,
-                                                           sd_version_is_longcat(version)));
+            pe_vec      = Rope::gen_flux_pe(static_cast<int>(x->ne[1]),
+                                            static_cast<int>(x->ne[0]),
+                                            config.patch_size,
+                                            static_cast<int>(x->ne[3]),
+                                            static_cast<int>(context->ne[1]),
+                                            txt_arange_dims,
+                                            ref_latents,
+                                            ref_index_mode,
+                                            config.ref_index_scale,
+                                            config.theta,
+                                            circular_y_enabled,
+                                            circular_x_enabled,
+                                            config.axes_dim,
+                                            sd_version_is_longcat(version));
             int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             // LOG_VERBOSE("pos_len %d", pos_len);
             auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);

src/model/diffusion/ideogram4.hpp

@@ -149,21 +149,18 @@ namespace Ideogram4 {
         return std::make_shared<Linear>(in_features, out_features, bias);
     }
 
-    __STATIC_INLINE__ Rope::Embedding gen_ideogram4_pe(int grid_h,
-                                                       int grid_w,
-                                                       int bs,
-                                                       int context_len,
-                                                       int head_dim,
-                                                       int rope_theta,
-                                                       const std::vector<int>& mrope_section) {
+    __STATIC_INLINE__ std::vector<float> gen_ideogram4_pe(int grid_h,
+                                                          int grid_w,
+                                                          int bs,
+                                                          int context_len,
+                                                          int head_dim,
+                                                          int rope_theta,
+                                                          const std::vector<int>& mrope_section,
+                                                          bool circular_x = false,
+                                                          bool circular_y = false) {
         GGML_ASSERT(bs == 1);
-        Rope::Embedding result;
-        result.batch_size = bs;
-        result.positions.append_tokens(context_len);
-        result.positions.append_image(grid_h, grid_w);
-        result.ids.assign(static_cast<size_t>(bs) * (context_len + grid_h * grid_w),
-                          std::vector<float>(3, 0.f));
-        auto& ids = result.ids;
+        std::vector<std::vector<float>> ids(static_cast<size_t>(bs) * (context_len + grid_h * grid_w),
+                                            std::vector<float>(3, 0.f));
 
         for (int i = 0; i < context_len; ++i) {
             ids[i] = {static_cast<float>(i), static_cast<float>(i), static_cast<float>(i)};
@@ -178,13 +175,29 @@ namespace Ideogram4 {
             }
         }
 
-        result.values = Rope::embed_interleaved_mrope(ids,
-                                                      bs,
-                                                      static_cast<float>(rope_theta),
-                                                      head_dim,
-                                                      mrope_section,
-                                                      &result.frequencies);
-        return result;
+        std::vector<std::vector<int>> axis_wrap_dims(3);
+        if (circular_y || circular_x) {
+            size_t total_len = static_cast<size_t>(bs) * (context_len + grid_h * grid_w);
+            axis_wrap_dims[1].assign(total_len, 0);
+            axis_wrap_dims[2].assign(total_len, 0);
+            if (circular_y) {
+                for (size_t idx = static_cast<size_t>(context_len); idx < total_len; ++idx) {
+                    axis_wrap_dims[1][idx] = grid_h;
+                }
+            }
+            if (circular_x) {
+                for (size_t idx = static_cast<size_t>(context_len); idx < total_len; ++idx) {
+                    axis_wrap_dims[2][idx] = grid_w;
+                }
+            }
+        }
+
+        return Rope::embed_interleaved_mrope(ids,
+                                             bs,
+                                             static_cast<float>(rope_theta),
+                                             head_dim,
+                                             mrope_section,
+                                             axis_wrap_dims);
     }
 
     class Ideogram4Attention : public GGMLBlock {
@@ -496,13 +509,15 @@ namespace Ideogram4 {
             int64_t head_dim = config.emb_dim / config.num_heads;
 
             auto runner_ctx = get_context();
-            pe_vec          = finish_rope_pe(gen_ideogram4_pe(static_cast<int>(grid_h),
-                                                              static_cast<int>(grid_w),
-                                                              static_cast<int>(x->ne[3]),
-                                                              static_cast<int>(context_len),
-                                                              static_cast<int>(head_dim),
-                                                              static_cast<int>(config.rope_theta),
-                                                              config.mrope_section));
+            pe_vec          = gen_ideogram4_pe(static_cast<int>(grid_h),
+                                               static_cast<int>(grid_w),
+                                               static_cast<int>(x->ne[3]),
+                                               static_cast<int>(context_len),
+                                               static_cast<int>(head_dim),
+                                               static_cast<int>(config.rope_theta),
+                                               config.mrope_section,
+                                               runner_ctx.circular_x_enabled,
+                                               runner_ctx.circular_y_enabled);
             auto pe         = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, head_dim / 2, pos_len);
             set_backend_tensor_data(pe, pe_vec.data());
 

src/model/diffusion/krea2.hpp

@@ -689,28 +689,23 @@ namespace Krea2 {
         }
     };
 
-    __STATIC_INLINE__ Rope::Embedding gen_krea2_pe(int h,
-                                                   int w,
-                                                   int patch_size,
-                                                   int bs,
-                                                   int context_len,
-                                                   float theta,
-                                                   const std::vector<int>& axes_dim,
-                                                   const std::vector<ggml_tensor*>& ref_latents,
-                                                   Rope::RefIndexMode ref_index_mode) {
-        Rope::Embedding result;
-        result.batch_size = bs;
-        result.positions.append_tokens(context_len);
+    __STATIC_INLINE__ std::vector<float> gen_krea2_pe(int h,
+                                                      int w,
+                                                      int patch_size,
+                                                      int bs,
+                                                      int context_len,
+                                                      float theta,
+                                                      const std::vector<int>& axes_dim,
+                                                      const std::vector<ggml_tensor*>& ref_latents,
+                                                      Rope::RefIndexMode ref_index_mode) {
         auto txt_ids = Rope::gen_flux_txt_ids(bs, context_len, 3, {});
-        auto img_ids = Rope::gen_flux_img_ids(h, w, patch_size, bs, 3, 0, 0, 0, false, &result.positions);
+        auto img_ids = Rope::gen_flux_img_ids(h, w, patch_size, bs, 3, 0, 0, 0, false);
         auto ids     = Rope::concat_ids(txt_ids, img_ids, bs);
         if (ref_latents.size() > 0) {
-            auto refs_ids = Rope::gen_refs_ids(patch_size, bs, 3, 1, ref_latents, ref_index_mode, 1.0f, false, 0, &result.positions);
+            auto refs_ids = Rope::gen_refs_ids(patch_size, bs, 3, 1, ref_latents, ref_index_mode, 1.0f, false, 0);
             ids           = Rope::concat_ids(ids, refs_ids, bs);
         }
-        result.ids    = std::move(ids);
-        result.values = Rope::embed_nd(result.ids, bs, theta, axes_dim, result.layout, &result.frequencies);
-        return result;
+        return Rope::embed_nd(ids, bs, theta, axes_dim);
     }
 
     struct Krea2Runner : public DiffusionModelRunner {
@@ -754,15 +749,15 @@ namespace Krea2 {
                 ref_latents.push_back(make_input(ref_latent_tensor));
             }
 
-            pe_vec      = finish_rope_pe(gen_krea2_pe(static_cast<int>(x->ne[1]),
-                                                      static_cast<int>(x->ne[0]),
-                                                      config.patch_size,
-                                                      static_cast<int>(x->ne[3]),
-                                                      static_cast<int>(context->ne[1]),
-                                                      config.theta,
-                                                      config.axes_dim,
-                                                      ref_latents,
-                                                      ref_image_params.ref_index_mode));
+            pe_vec      = gen_krea2_pe(static_cast<int>(x->ne[1]),
+                                       static_cast<int>(x->ne[0]),
+                                       config.patch_size,
+                                       static_cast<int>(x->ne[3]),
+                                       static_cast<int>(context->ne[1]),
+                                       config.theta,
+                                       config.axes_dim,
+                                       ref_latents,
+                                       ref_image_params.ref_index_mode);
             int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
             set_backend_tensor_data(pe, pe_vec.data());

src/model/diffusion/lens.hpp

@@ -384,12 +384,14 @@ namespace Lens {
             GGML_ASSERT(!context_tensor.empty());
             ggml_tensor* context = make_input(context_tensor);
 
-            pe_vec      = finish_rope_pe(Rope::gen_lens_pe(static_cast<int>(x->ne[1]),
-                                                           static_cast<int>(x->ne[0]),
-                                                           static_cast<int>(x->ne[3]),
-                                                           static_cast<int>(context->ne[1]),
-                                                           config.theta,
-                                                           config.axes_dim));
+            pe_vec      = Rope::gen_lens_pe(static_cast<int>(x->ne[1]),
+                                            static_cast<int>(x->ne[0]),
+                                            static_cast<int>(x->ne[3]),
+                                            static_cast<int>(context->ne[1]),
+                                            config.theta,
+                                            circular_y_enabled,
+                                            circular_x_enabled,
+                                            config.axes_dim);
             int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
             set_backend_tensor_data(pe, pe_vec.data());

src/model/diffusion/llada_image.hpp

@@ -1,525 +0,0 @@
-#ifndef __SD_MODEL_DIFFUSION_LLADA_IMAGE_HPP__
-#define __SD_MODEL_DIFFUSION_LLADA_IMAGE_HPP__
-
-#include <algorithm>
-#include <cinttypes>
-
-#include "core/ggml_extend.h"
-#include "core/ggml_runner.h"
-#include "core/util.h"
-#include "model/common/ggml_block.hpp"
-#include "model/diffusion/model.hpp"
-#include "model/diffusion/z_image.hpp"
-#include "model_loader.h"
-
-// Ref: https://github.com/inclusionAI/LLaDA-Image/blob/main/src/models/transformer_llada_image.py
-//
-// The denoiser is Lumina2/z_image's NextDiT with identical hyperparameters, so the blocks are
-// reused from ZImage. Two things differ: every norm here is non-parametric (the checkpoint
-// carries no norm weights at all), and latents arrive already patchified from the Flux2 VAE,
-// so patch_size is 1 over 128 channels.
-
-namespace LLaDAImage {
-    constexpr int LLADA_IMAGE_GRAPH_SIZE = 20480;
-
-    struct LLaDAImageConfig {
-        int patch_size             = 1;
-        int64_t hidden_size        = 3840;
-        int64_t in_channels        = 128;
-        int64_t out_channels       = 128;
-        int64_t num_layers         = 30;
-        int64_t num_refiner_layers = 2;
-        int64_t head_dim           = 128;
-        int64_t num_heads          = 30;
-        int64_t num_kv_heads       = 30;
-        int64_t multiple_of        = 256;
-        float ffn_dim_multiplier   = 8.0f / 3.0f;
-        float norm_eps             = 1e-5f;
-        bool qk_norm               = true;
-        int64_t cap_feat_dim       = 2560;
-        int64_t semantic_feat_dim  = 4096;
-        int theta                  = 256;
-        std::vector<int> axes_dim  = {32, 48, 48};
-        int64_t axes_dim_sum       = 128;
-
-        static int64_t count_blocks(const String2TensorStorage& tensor_storage_map,
-                                    const std::string& prefix,
-                                    const std::string& block_prefix) {
-            int64_t count = 0;
-            for (const auto& [name, _] : tensor_storage_map) {
-                if (!starts_with(name, prefix)) {
-                    continue;
-                }
-                size_t pos = name.find(block_prefix);
-                if (pos == std::string::npos) {
-                    continue;
-                }
-                auto items = split_string(name.substr(pos), '.');
-                if (items.size() > 1) {
-                    count = std::max<int64_t>(count, atoi(items[1].c_str()) + 1);
-                }
-            }
-            return count;
-        }
-
-        static LLaDAImageConfig detect_from_weights(const String2TensorStorage& tensor_storage_map, const std::string& prefix) {
-            LLaDAImageConfig config;
-            int64_t detected_q_dim  = 0;
-            int64_t detected_kv_dim = 0;
-
-            for (const auto& [name, tensor_storage] : tensor_storage_map) {
-                if (!starts_with(name, prefix)) {
-                    continue;
-                }
-                if (ends_with(name, "x_embedder.weight") && tensor_storage.n_dims == 2) {
-                    int64_t patch_area = config.patch_size * config.patch_size;
-                    config.in_channels = tensor_storage.ne[0] / patch_area;
-                    config.hidden_size = tensor_storage.ne[1];
-                } else if (ends_with(name, "cap_embedder.1.weight") && tensor_storage.n_dims == 2) {
-                    config.cap_feat_dim = tensor_storage.ne[0];
-                    config.hidden_size  = tensor_storage.ne[1];
-                } else if (ends_with(name, "sigvq_embedder.1.weight") && tensor_storage.n_dims == 2) {
-                    config.semantic_feat_dim = tensor_storage.ne[0];
-                } else if (ends_with(name, "layers.0.attention.to_q.weight") && tensor_storage.n_dims == 2) {
-                    detected_q_dim = tensor_storage.ne[1];
-                } else if (ends_with(name, "layers.0.attention.to_k.weight") && tensor_storage.n_dims == 2) {
-                    detected_kv_dim = tensor_storage.ne[1];
-                } else if (ends_with(name, "final_layer.linear.weight") && tensor_storage.n_dims == 2) {
-                    int64_t patch_area  = config.patch_size * config.patch_size;
-                    config.out_channels = tensor_storage.ne[1] / patch_area;
-                }
-            }
-
-            int64_t detected_layers  = count_blocks(tensor_storage_map, prefix, "layers.");
-            int64_t detected_refiner = std::max(count_blocks(tensor_storage_map, prefix, "noise_refiner."),
-                                                count_blocks(tensor_storage_map, prefix, "context_refiner."));
-            if (detected_layers > 0) {
-                config.num_layers = detected_layers;
-            }
-            if (detected_refiner > 0) {
-                config.num_refiner_layers = detected_refiner;
-            }
-            if (detected_q_dim > 0) {
-                config.num_heads = detected_q_dim / config.head_dim;
-            }
-            if (detected_kv_dim > 0) {
-                config.num_kv_heads = detected_kv_dim / config.head_dim;
-            } else if (detected_q_dim > 0) {
-                config.num_kv_heads = config.num_heads;
-            }
-
-            LOG_VERBOSE("llada_image: num_layers = %" PRId64 ", num_refiner_layers = %" PRId64 ", hidden_size = %" PRId64 ", num_heads = %" PRId64 ", num_kv_heads = %" PRId64 ", in_channels = %" PRId64 ", out_channels = %" PRId64 ", cap_feat_dim = %" PRId64 ", semantic_feat_dim = %" PRId64,
-                        config.num_layers,
-                        config.num_refiner_layers,
-                        config.hidden_size,
-                        config.num_heads,
-                        config.num_kv_heads,
-                        config.in_channels,
-                        config.out_channels,
-                        config.cap_feat_dim,
-                        config.semantic_feat_dim);
-            return config;
-        }
-    };
-
-    class LLaDAImageModel : public GGMLBlock {
-    protected:
-        LLaDAImageConfig config;
-
-        void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
-            params["cap_pad_token"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, config.hidden_size);
-            params["x_pad_token"]     = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, config.hidden_size);
-            params["sigvq_pad_token"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, config.hidden_size);
-        }
-
-        std::shared_ptr<ZImage::JointTransformerBlock> make_block(bool modulation) {
-            return std::make_shared<ZImage::JointTransformerBlock>(0,
-                                                                   config.hidden_size,
-                                                                   config.head_dim,
-                                                                   config.num_heads,
-                                                                   config.num_kv_heads,
-                                                                   config.multiple_of,
-                                                                   config.ffn_dim_multiplier,
-                                                                   config.norm_eps,
-                                                                   config.qk_norm,
-                                                                   modulation,
-                                                                   false,
-                                                                   true);
-        }
-
-    public:
-        LLaDAImageModel() = default;
-        LLaDAImageModel(LLaDAImageConfig config)
-            : config(config) {
-            blocks["x_embedder"] = std::make_shared<Linear>(config.patch_size * config.patch_size * config.in_channels, config.hidden_size);
-            blocks["t_embedder"] = std::make_shared<TimestepEmbedder>(MIN(config.hidden_size, 1024), 256, ZImage::ADALN_EMBED_DIM);
-
-            blocks["cap_embedder.0"] = std::make_shared<RMSNorm>(config.cap_feat_dim, config.norm_eps, false);
-            blocks["cap_embedder.1"] = std::make_shared<Linear>(config.cap_feat_dim, config.hidden_size);
-
-            blocks["semantic_embedder.0"] = std::make_shared<RMSNorm>(config.semantic_feat_dim, config.norm_eps, false);
-            blocks["semantic_embedder.1"] = std::make_shared<Linear>(config.semantic_feat_dim, config.hidden_size);
-            blocks["sigvq_embedder.0"]    = std::make_shared<RMSNorm>(config.semantic_feat_dim, config.norm_eps, false);
-            blocks["sigvq_embedder.1"]    = std::make_shared<Linear>(config.semantic_feat_dim, config.hidden_size);
-
-            for (int i = 0; i < config.num_refiner_layers; i++) {
-                blocks["noise_refiner." + std::to_string(i)]   = make_block(true);
-                blocks["context_refiner." + std::to_string(i)] = make_block(false);
-                blocks["sigvq_refiner." + std::to_string(i)]   = make_block(false);
-            }
-            for (int i = 0; i < config.num_layers; i++) {
-                blocks["layers." + std::to_string(i)] = make_block(true);
-            }
-
-            blocks["final_layer"] = std::make_shared<ZImage::FinalLayer>(config.hidden_size, config.patch_size, config.out_channels);
-        }
-
-        ggml_tensor* forward_core(GGMLRunnerContext* ctx,
-                                  ggml_tensor* x,
-                                  ggml_tensor* timestep,
-                                  ggml_tensor* context,
-                                  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"]);
-            auto cap_embedder_1 = std::dynamic_pointer_cast<Linear>(blocks["cap_embedder.1"]);
-            auto final_layer    = std::dynamic_pointer_cast<ZImage::FinalLayer>(blocks["final_layer"]);
-
-            auto txt_pad_token = params["cap_pad_token"];
-            auto img_pad_token = params["x_pad_token"];
-
-            int64_t N           = x->ne[2];
-            int64_t n_img_token = x->ne[1];
-            int64_t n_txt_token = context->ne[1];
-
-            // sdcpp's flow denoiser already hands over sigma * 1000, which is the range the
-            // reference reaches via its own t_scale, so no further scaling here.
-            auto t_emb = t_embedder->forward(ctx, timestep);
-
-            auto txt = cap_embedder_1->forward(ctx, cap_embedder_0->forward(ctx, context));  // [N, n_txt_token, hidden_size]
-            auto img = x_embedder->forward(ctx, x);                                          // [N, n_img_token, hidden_size]
-            sd::ggml_graph_cut::mark_graph_cut(txt, "llada_image.prelude", "txt");
-            sd::ggml_graph_cut::mark_graph_cut(img, "llada_image.prelude", "img");
-            sd::ggml_graph_cut::mark_graph_cut(t_emb, "llada_image.prelude", "t_emb");
-
-            int64_t n_txt_pad_token = Rope::bound_mod(static_cast<int>(n_txt_token), ZImage::SEQ_MULTI_OF);
-            if (n_txt_pad_token > 0) {
-                auto txt_pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, txt_pad_token, txt_pad_token->ne[0], n_txt_pad_token, N, 1);
-                txt                 = ggml_concat(ctx->ggml_ctx, txt, txt_pad_tokens, 1);
-            }
-
-            int64_t n_img_pad_token = Rope::bound_mod(static_cast<int>(n_img_token), ZImage::SEQ_MULTI_OF);
-            if (n_img_pad_token > 0) {
-                auto img_pad_tokens = ggml_repeat_4d(ctx->ggml_ctx, img_pad_token, img_pad_token->ne[0], n_img_pad_token, N, 1);
-                img                 = ggml_concat(ctx->ggml_ctx, img, img_pad_tokens, 1);
-            }
-
-            GGML_ASSERT(txt->ne[1] + img->ne[1] == pe->ne[3]);
-
-            auto txt_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, 0, txt->ne[1]);
-            auto img_pe = ggml_ext_slice(ctx->ggml_ctx, pe, 3, txt->ne[1], pe->ne[3]);
-
-            for (int i = 0; i < config.num_refiner_layers; i++) {
-                auto block = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["context_refiner." + std::to_string(i)]);
-
-                txt = block->forward(ctx, txt, txt_pe, nullptr, nullptr);
-                sd::ggml_graph_cut::mark_graph_cut(txt, "llada_image.context_refiner." + std::to_string(i), "txt");
-            }
-
-            for (int i = 0; i < config.num_refiner_layers; i++) {
-                auto block = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["noise_refiner." + std::to_string(i)]);
-
-                img = block->forward(ctx, img, img_pe, nullptr, t_emb);
-                sd::ggml_graph_cut::mark_graph_cut(img, "llada_image.noise_refiner." + std::to_string(i), "img");
-            }
-
-            auto txt_img = ggml_concat(ctx->ggml_ctx, txt, img, 1);
-            sd::ggml_graph_cut::mark_graph_cut(txt_img, "llada_image.prelude", "txt_img");
-
-            for (int i = 0; i < config.num_layers; i++) {
-                auto block = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["layers." + std::to_string(i)]);
-
-                txt_img = block->forward(ctx, txt_img, pe, nullptr, t_emb);
-                sd::ggml_graph_cut::mark_graph_cut(txt_img, "llada_image.layers." + std::to_string(i), "txt_img");
-            }
-
-            txt_img = final_layer->forward(ctx, txt_img, t_emb);
-
-            return ggml_ext_slice(ctx->ggml_ctx, txt_img, 1, n_txt_token + n_txt_pad_token, n_txt_token + n_txt_pad_token + n_img_token);
-        }
-
-        ggml_tensor* pad_stream(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* pad_token) {
-            int64_t n_pad = Rope::bound_mod(static_cast<int>(x->ne[1]), ZImage::SEQ_MULTI_OF);
-            if (n_pad == 0) {
-                return x;
-            }
-            auto pads = ggml_repeat_4d(ctx->ggml_ctx, pad_token, pad_token->ne[0], n_pad, x->ne[2], 1);
-            return ggml_concat(ctx->ggml_ctx, x, pads, 1);
-        }
-
-        // Editing runs one joint sequence carrying two timesteps: the caption and source latent
-        // are clean (t = 0) while the second caption copy and the target latent are noisy. adaLN
-        // is a linear map of the timestep embedding, so feeding a per-token embedding selects the
-        // right modulation exactly, without duplicating the modulation projections.
-        ggml_tensor* forward_editing(GGMLRunnerContext* ctx,
-                                     ggml_tensor* x,
-                                     ggml_tensor* timestep,
-                                     ggml_tensor* context,
-                                     ggml_tensor* semantic,
-                                     ggml_tensor* source_latent,
-                                     ggml_tensor* pe) {
-            ggml_context* gctx = ctx->ggml_ctx;
-
-            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"]);
-            auto cap_embedder_1 = std::dynamic_pointer_cast<Linear>(blocks["cap_embedder.1"]);
-            auto sigvq_embed_0  = std::dynamic_pointer_cast<RMSNorm>(blocks["sigvq_embedder.0"]);
-            auto sigvq_embed_1  = std::dynamic_pointer_cast<Linear>(blocks["sigvq_embedder.1"]);
-            auto final_layer    = std::dynamic_pointer_cast<ZImage::FinalLayer>(blocks["final_layer"]);
-
-            auto t_noisy = t_embedder->forward(ctx, timestep);
-            auto t_clean = t_embedder->forward(ctx, ggml_scale(gctx, timestep, 0.f));
-
-            auto per_token = [&](ggml_tensor* emb, int64_t n) {
-                return ggml_repeat_4d(gctx, emb, emb->ne[0], n, 1, 1);
-            };
-
-            auto cap        = cap_embedder_1->forward(ctx, cap_embedder_0->forward(ctx, context));
-            cap             = pad_stream(ctx, cap, params["cap_pad_token"]);
-            int64_t cap_len = cap->ne[1];
-            cap             = ggml_concat(gctx, cap, cap, 1);
-
-            auto src            = pad_stream(ctx, x_embedder->forward(ctx, source_latent), params["x_pad_token"]);
-            auto tgt_embed      = x_embedder->forward(ctx, x);
-            int64_t n_img_token = tgt_embed->ne[1];
-            auto tgt            = pad_stream(ctx, tgt_embed, params["x_pad_token"]);
-            int64_t img_len     = tgt->ne[1];
-            auto img            = ggml_concat(gctx, src, tgt, 1);
-
-            ggml_tensor* sig = nullptr;
-            int64_t sig_len  = 0;
-            if (semantic != nullptr) {
-                sig     = sigvq_embed_1->forward(ctx, sigvq_embed_0->forward(ctx, semantic));
-                sig     = pad_stream(ctx, sig, params["sigvq_pad_token"]);
-                sig_len = sig->ne[1];
-            }
-
-            GGML_ASSERT(cap_len * 2 + img_len * 2 + sig_len == pe->ne[3]);
-
-            auto cap_pe = ggml_ext_slice(gctx, pe, 3, 0, cap_len * 2);
-            auto img_pe = ggml_ext_slice(gctx, pe, 3, cap_len * 2, cap_len * 2 + img_len * 2);
-
-            auto img_adaln = ggml_concat(gctx, per_token(t_clean, img_len), per_token(t_noisy, img_len), 1);
-
-            for (int i = 0; i < config.num_refiner_layers; i++) {
-                auto block = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["context_refiner." + std::to_string(i)]);
-                cap        = block->forward(ctx, cap, cap_pe, nullptr, nullptr);
-            }
-            for (int i = 0; i < config.num_refiner_layers; i++) {
-                auto block = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["noise_refiner." + std::to_string(i)]);
-                img        = block->forward(ctx, img, img_pe, nullptr, img_adaln);
-            }
-            if (sig != nullptr) {
-                auto sig_pe = ggml_ext_slice(gctx, pe, 3, cap_len * 2 + img_len * 2, pe->ne[3]);
-                for (int i = 0; i < config.num_refiner_layers; i++) {
-                    auto block = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["sigvq_refiner." + std::to_string(i)]);
-                    sig        = block->forward(ctx, sig, sig_pe, nullptr, nullptr);
-                }
-            }
-
-            auto seq = ggml_concat(gctx, cap, img, 1);
-
-            auto cap_adaln = ggml_concat(gctx, per_token(t_clean, cap_len), per_token(t_noisy, cap_len), 1);
-            auto seq_adaln = ggml_concat(gctx, cap_adaln, img_adaln, 1);
-            if (sig != nullptr) {
-                seq       = ggml_concat(gctx, seq, sig, 1);
-                seq_adaln = ggml_concat(gctx, seq_adaln, per_token(t_clean, sig_len), 1);
-            }
-
-            for (int i = 0; i < config.num_layers; i++) {
-                auto block = std::dynamic_pointer_cast<ZImage::JointTransformerBlock>(blocks["layers." + std::to_string(i)]);
-                seq        = block->forward(ctx, seq, pe, nullptr, seq_adaln);
-                sd::ggml_graph_cut::mark_graph_cut(seq, "llada_image.layers." + std::to_string(i), "seq");
-            }
-
-            seq = final_layer->forward(ctx, seq, seq_adaln);
-
-            // Only the target latent is denoised; the source half of the image stream is context.
-            // The stream is padded to SEQ_MULTI_OF, so drop the pad tokens: they are not part of
-            // the latent grid that unpatchify reconstructs.
-            int64_t target_start = cap_len * 2 + img_len;
-            return ggml_ext_slice(gctx, seq, 1, target_start, target_start + n_img_token);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                             ggml_tensor* x,
-                             ggml_tensor* timestep,
-                             ggml_tensor* context,
-                             ggml_tensor* pe) {
-            // x: [N, C, H, W]
-            // timestep: [N,]
-            // context: [N, L, cap_feat_dim]
-            // pe: [L, d_head/2, 2, 2]
-            // return: [N, C, H, W]
-            int64_t W = x->ne[0];
-            int64_t H = x->ne[1];
-
-            int patch_size = config.patch_size;
-
-            auto img = DiT::pad_and_patchify(ctx, x, patch_size, patch_size, false);
-
-            auto out = forward_core(ctx, img, timestep, context, pe);
-
-            out = DiT::unpatchify_and_crop(ctx->ggml_ctx, out, H, W, patch_size, patch_size, false);
-
-            // The reference pipeline negates the model output before the scheduler step.
-            return ggml_ext_scale(ctx->ggml_ctx, out, -1.f);
-        }
-    };
-
-    struct LLaDAImageRunner : public DiffusionModelRunner {
-    public:
-        LLaDAImageConfig config;
-        LLaDAImageModel llada_image;
-        std::vector<float> pe_vec;
-
-        LLaDAImageRunner(ggml_backend_t backend,
-                         const String2TensorStorage& tensor_storage_map      = {},
-                         const std::string prefix                            = "",
-                         std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
-            : DiffusionModelRunner(backend, prefix, weight_manager),
-              config(LLaDAImageConfig::detect_from_weights(tensor_storage_map, prefix)) {
-            llada_image = LLaDAImageModel(config);
-            llada_image.init(params_ctx, tensor_storage_map, prefix);
-        }
-
-        std::string get_desc() override {
-            return "llada_image";
-        }
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
-            llada_image.get_param_tensors(tensors, prefix);
-        }
-
-        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
-                                 const sd::Tensor<float>& timesteps_tensor,
-                                 const sd::Tensor<float>& context_tensor) {
-            ggml_cgraph* gf        = new_graph_custom(LLADA_IMAGE_GRAPH_SIZE);
-            ggml_tensor* x         = make_input(x_tensor);
-            ggml_tensor* timesteps = make_input(timesteps_tensor);
-            GGML_ASSERT(x->ne[3] == 1);
-            GGML_ASSERT(!context_tensor.empty());
-            ggml_tensor* context = make_input(context_tensor);
-
-            pe_vec      = finish_rope_pe(Rope::gen_llada_image_pe(static_cast<int>(x->ne[1]),
-                                                                  static_cast<int>(x->ne[0]),
-                                                                  config.patch_size,
-                                                                  static_cast<int>(x->ne[3]),
-                                                                  static_cast<int>(context->ne[1]),
-                                                                  ZImage::SEQ_MULTI_OF,
-                                                                  config.theta,
-                                                                  config.axes_dim));
-            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
-            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
-            set_backend_tensor_data(pe, pe_vec.data());
-            auto runner_ctx = get_context();
-
-            ggml_tensor* out = llada_image.forward(&runner_ctx, x, timesteps, context, pe);
-
-            ggml_build_forward_expand(gf, out);
-
-            return gf;
-        }
-
-        sd::Tensor<float> compute(int n_threads,
-                                  const sd::Tensor<float>& x,
-                                  const sd::Tensor<float>& timesteps,
-                                  const sd::Tensor<float>& context) {
-            // x: [N, in_channels, h, w]
-            // timesteps: [N, ]
-            // context: [N, max_position, cap_feat_dim]
-            auto get_graph = [&]() -> ggml_cgraph* {
-                return build_graph(x, timesteps, context);
-            };
-
-            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, false), x.dim());
-        }
-
-        ggml_cgraph* build_edit_graph(const sd::Tensor<float>& x_tensor,
-                                      const sd::Tensor<float>& timesteps_tensor,
-                                      const sd::Tensor<float>& context_tensor,
-                                      const sd::Tensor<float>& semantic_tensor,
-                                      const sd::Tensor<float>& source_tensor) {
-            ggml_cgraph* gf        = new_graph_custom(LLADA_IMAGE_GRAPH_SIZE);
-            ggml_tensor* x         = make_input(x_tensor);
-            ggml_tensor* timesteps = make_input(timesteps_tensor);
-            ggml_tensor* context   = make_input(context_tensor);
-            ggml_tensor* semantic  = make_optional_input(semantic_tensor);
-            ggml_tensor* source    = make_input(source_tensor);
-            GGML_ASSERT(x->ne[3] == 1);
-
-            pe_vec      = finish_rope_pe(Rope::gen_llada_image_edit_pe(static_cast<int>(x->ne[1]),
-                                                                       static_cast<int>(x->ne[0]),
-                                                                       config.patch_size,
-                                                                       static_cast<int>(context->ne[1]),
-                                                                  semantic != nullptr ? static_cast<int>(semantic->ne[1]) : 0,
-                                                                       ZImage::SEQ_MULTI_OF,
-                                                                       config.theta,
-                                                                       config.axes_dim));
-            int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
-            auto pe     = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
-            set_backend_tensor_data(pe, pe_vec.data());
-            auto runner_ctx = get_context();
-
-            int64_t W   = x->ne[0];
-            int64_t H   = x->ne[1];
-            auto target = DiT::pad_and_patchify(&runner_ctx, x, config.patch_size, config.patch_size, false);
-            auto src    = DiT::pad_and_patchify(&runner_ctx, source, config.patch_size, config.patch_size, false);
-
-            auto out = llada_image.forward_editing(&runner_ctx, target, timesteps, context, semantic, src, pe);
-            out      = DiT::unpatchify_and_crop(runner_ctx.ggml_ctx, out, H, W, config.patch_size, config.patch_size, false);
-            out      = ggml_ext_scale(runner_ctx.ggml_ctx, out, -1.f);
-
-            ggml_build_forward_expand(gf, out);
-            return gf;
-        }
-
-        sd::Tensor<float> compute(int n_threads,
-                                  const DiffusionParams& diffusion_params) override {
-            GGML_ASSERT(diffusion_params.x != nullptr);
-            GGML_ASSERT(diffusion_params.timesteps != nullptr);
-
-            const auto* extra   = std::get_if<LLaDAImageDiffusionExtra>(&diffusion_params.extra);
-            bool has_semantic   = extra != nullptr && extra->semantic != nullptr && !extra->semantic->empty();
-            bool has_ref_latent = diffusion_params.ref_latents != nullptr && !diffusion_params.ref_latents->empty();
-            if (has_semantic && !has_ref_latent) {
-                LOG_WARN("llada_image: SigVQ features without a reference latent are not supported; falling back to text to image");
-            }
-            if (has_ref_latent) {
-                const auto& source = diffusion_params.ref_latents->front();
-                if (source.shape() != diffusion_params.x->shape()) {
-                    LOG_ERROR("llada_image: reference latent must match the target shape; use resize_vae_to_target=1");
-                    return {};
-                }
-                auto get_graph = [&]() -> ggml_cgraph* {
-                    return build_edit_graph(*diffusion_params.x,
-                                            *diffusion_params.timesteps,
-                                            tensor_or_empty(diffusion_params.context),
-                                            tensor_or_empty(extra != nullptr ? extra->semantic : nullptr),
-                                            source);
-                };
-                return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, false),
-                                                       diffusion_params.x->dim());
-            }
-
-            return compute(n_threads,
-                           *diffusion_params.x,
-                           *diffusion_params.timesteps,
-                           tensor_or_empty(diffusion_params.context));
-        }
-    };
-
-}  // namespace LLaDAImage
-
-#endif  // __SD_MODEL_DIFFUSION_LLADA_IMAGE_HPP__

src/model/diffusion/mage_flow.hpp

@@ -110,13 +110,13 @@ namespace MageFlow {
             }
 
             int batch_size = static_cast<int>(x->ne[3]);
-            pe_vec         = finish_rope_pe(Rope::gen_mage_flow_pe(static_cast<int>(x->ne[1]),
-                                                                   static_cast<int>(x->ne[0]),
-                                                                   batch_size,
-                                                                   static_cast<int>(context->ne[1]),
-                                                                   ref_latents,
-                                                                   config.theta,
-                                                                   config.axes_dim));
+            pe_vec         = Rope::gen_mage_flow_pe(static_cast<int>(x->ne[1]),
+                                                    static_cast<int>(x->ne[0]),
+                                                    batch_size,
+                                                    static_cast<int>(context->ne[1]),
+                                                    ref_latents,
+                                                    config.theta,
+                                                    config.axes_dim);
             int pos_len    = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             auto pe        = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);
             set_backend_tensor_data(pe, pe_vec.data());

src/model/diffusion/minimax_h3.hpp

@@ -264,9 +264,6 @@ namespace MiniMaxH3 {
             for (int64_t i = 0; i < num_layers; ++i) {
                 auto block = std::dynamic_pointer_cast<TokenRefinerBlock>(blocks["blocks." + std::to_string(i)]);
                 x          = block->forward(ctx, x);
-                sd::ggml_graph_cut::mark_graph_cut(x,
-                                                   "minimax_h3.token_refiner.blocks." + std::to_string(i),
-                                                   "hidden_states");
             }
             return std::dynamic_pointer_cast<RMSNorm>(blocks["final_norm"])->forward(ctx, x);
         }
@@ -530,11 +527,7 @@ namespace MiniMaxH3 {
             GGML_ASSERT(context->ne[0] == config.text_dim);
             auto condition_proj = std::dynamic_pointer_cast<Linear>(blocks["condition_proj"]);
             auto token_refiner  = std::dynamic_pointer_cast<TokenRefiner>(blocks["token_refiner"]);
-            auto projected      = condition_proj->forward(ctx, context);
-            sd::ggml_graph_cut::mark_graph_cut(projected,
-                                               "minimax_h3.condition_proj",
-                                               "hidden_states");
-            return token_refiner->forward(ctx, projected);
+            return token_refiner->forward(ctx, condition_proj->forward(ctx, context));
         }
 
         ggml_tensor* time_embedding(GGMLRunnerContext* ctx,

src/model/diffusion/minit2i.hpp

@@ -154,26 +154,18 @@ namespace MiniT2I {
         return Rope::flatten(Rope::rope(Rope::linspace(0.f, static_cast<float>(length - 1), length), head_dim, 10000.f));
     }
 
-    inline Rope::Embedding make_vision_rope(int side, int head_dim) {
+    inline std::vector<float> make_vision_rope(int side, int head_dim) {
         GGML_ASSERT(head_dim % 4 == 0);
         int dim     = head_dim / 2;
         int quarter = dim / 2;
         int length  = side * side;
-        Rope::Embedding result;
-        result.positions.append_image(side, side);
         std::vector<float> out(static_cast<size_t>(length) * (head_dim / 2) * 4);
         std::vector<float> freqs(quarter);
         for (int i = 0; i < quarter; ++i) {
             freqs[i] = 1.0f / std::pow(10000.0f, static_cast<float>(2 * i) / static_cast<float>(dim));
         }
-        for (int axis : {1, 2}) {
-            for (float frequency : freqs) {
-                result.frequencies.push_back({static_cast<size_t>(axis), frequency});
-            }
-        }
         for (int y = 0; y < side; ++y) {
             for (int x = 0; x < side; ++x) {
-                result.ids.push_back({0.f, static_cast<float>(y), static_cast<float>(x)});
                 int pos     = y * side + x;
                 size_t base = static_cast<size_t>(pos) * (head_dim / 2) * 4;
                 for (int i = 0; i < quarter; ++i) {
@@ -190,8 +182,7 @@ namespace MiniT2I {
                 }
             }
         }
-        result.values = std::move(out);
-        return result;
+        return out;
     }
 
     struct SwiGLUMlp : public GGMLBlock {
@@ -484,8 +475,6 @@ namespace MiniT2I {
         int64_t cached_txt_len                      = -1;
         int64_t cached_hidden_size                  = -1;
         int64_t cached_head_dim                     = -1;
-        bool cached_circular_x                      = false;
-        bool cached_circular_y                      = false;
 
         MiniT2IRunner(ggml_backend_t backend,
                       const String2TensorStorage& tensor_storage_map      = {},
@@ -532,8 +521,6 @@ namespace MiniT2I {
                 cached_txt_len == txt_len &&
                 cached_hidden_size == config.hidden_size &&
                 cached_head_dim == config.head_dim &&
-                cached_circular_x == circular_x_enabled &&
-                cached_circular_y == circular_y_enabled &&
                 cached_pos_embed != nullptr &&
                 cached_txt_pe != nullptr &&
                 cached_joint_pe != nullptr) {
@@ -544,7 +531,7 @@ namespace MiniT2I {
 
             auto pos_embed_vec = make_2d_sincos_pos_embed(static_cast<int>(img_side), static_cast<int>(config.hidden_size));
             auto txt_pe_vec    = make_text_rope(static_cast<int>(txt_len), static_cast<int>(config.head_dim));
-            auto img_pe_vec    = finish_rope_pe(make_vision_rope(static_cast<int>(img_side), static_cast<int>(config.head_dim)));
+            auto img_pe_vec    = make_vision_rope(static_cast<int>(img_side), static_cast<int>(config.head_dim));
             auto joint_pe_vec  = txt_pe_vec;
             joint_pe_vec.insert(joint_pe_vec.end(), img_pe_vec.begin(), img_pe_vec.end());
 
@@ -574,8 +561,6 @@ namespace MiniT2I {
             cached_txt_len     = txt_len;
             cached_hidden_size = config.hidden_size;
             cached_head_dim    = config.head_dim;
-            cached_circular_x  = circular_x_enabled;
-            cached_circular_y  = circular_y_enabled;
         }
 
         ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,

src/model/diffusion/model.hpp

@@ -7,7 +7,7 @@
 
 #include "core/ggml_runner.h"
 #include "core/tensor_ggml.hpp"
-#include "model/common/rope_circular.hpp"
+#include "model/common/rope.hpp"
 #include "model_manager.h"
 
 enum class RefImageResizeMode {
@@ -39,9 +39,6 @@ const std::unordered_map<std::string, RefImageParams> REF_IMAGE_PRESETS = {
     {"z_image_omni", {true, true, Rope::RefIndexMode::FIXED, false, true, -1, RefImageResizeMode::AREA, -1, -1}},
     {"krea2_ostris_edit", {true, true, Rope::RefIndexMode::INCREASE, true, true, -1, RefImageResizeMode::AREA, -1, -1}},
     {"krea2_edit", {true, true, Rope::RefIndexMode::INCREASE, false, true, -1, RefImageResizeMode::LONGEST_SIDE, 768, 768}},
-    // pass_to_vlm routes the reference image to the conditioner, which is where LLaDA-Image's
-    // SigVQ encoder lives; it does its own half-resolution resize.
-    {"llada_image", {true, true, Rope::RefIndexMode::FIXED, true, true, -1, RefImageResizeMode::NONE, -1, -1, true}},
     {"cosmos_reference", {false, true, Rope::RefIndexMode::INCREASE, false, false, -1, RefImageResizeMode::NONE, -1, -1}},
 };
 
@@ -69,12 +66,6 @@ struct AnimaDiffusionExtra {
     const sd::Tensor<float>* t5_weights = nullptr;
 };
 
-struct QwenImage21DiffusionExtra {
-    const sd::Tensor<int32_t>* image_slots = nullptr;
-    // Nonzero IDs identify immutable prefix inputs within one sampling run.
-    uint64_t prefix_id = 0;
-};
-
 struct WanDiffusionExtra {
     const sd::Tensor<float>* vace_context = nullptr;
     float vace_strength                   = 1.f;
@@ -136,25 +127,18 @@ struct HunyuanVideoDiffusionExtra {
     const sd::Tensor<float>* timestep_r = nullptr;
 };
 
-struct LLaDAImageDiffusionExtra {
-    // SigVQ semantic features of the reference image; present only in editing mode.
-    const sd::Tensor<float>* semantic = nullptr;
-};
-
 using DiffusionExtraParams = std::variant<std::monostate,
                                           UNetDiffusionExtra,
                                           SkipLayerDiffusionExtra,
                                           FluxDiffusionExtra,
                                           AnimaDiffusionExtra,
-                                          QwenImage21DiffusionExtra,
                                           WanDiffusionExtra,
                                           HiDreamO1DiffusionExtra,
                                           LTXAVDiffusionExtra,
                                           MiniMaxH3DiffusionExtra,
                                           MiniT2IDiffusionExtra,
                                           SenseNovaU1DiffusionExtra,
-                                          HunyuanVideoDiffusionExtra,
-                                          LLaDAImageDiffusionExtra>;
+                                          HunyuanVideoDiffusionExtra>;
 
 struct DiffusionParams {
     const sd::Tensor<float>* x                        = nullptr;
@@ -184,11 +168,6 @@ struct DiffusionModelRunner : public GGMLRunner {
 protected:
     std::string prefix;
 
-    std::vector<float> finish_rope_pe(Rope::Embedding embedding) {
-        Rope::apply_circular(embedding, circular_x_enabled, circular_y_enabled);
-        return std::move(embedding.values);
-    }
-
 public:
     DiffusionModelRunner(ggml_backend_t backend,
                          const std::string& prefix,

src/model/diffusion/pid.hpp

@@ -135,13 +135,13 @@ namespace Pid {
         return Rope::flatten(Rope::rope(Rope::linspace(0.f, static_cast<float>(length - 1), length), dim, theta));
     }
 
-    inline Rope::Embedding make_rope_2d(int height,
-                                        int width,
-                                        int dim,
-                                        float theta    = 10000.f,
-                                        float scale    = 16.f,
-                                        int ref_grid_h = 0,
-                                        int ref_grid_w = 0) {
+    inline std::vector<float> make_rope_2d(int height,
+                                           int width,
+                                           int dim,
+                                           float theta    = 10000.f,
+                                           float scale    = 16.f,
+                                           int ref_grid_h = 0,
+                                           int ref_grid_w = 0) {
         GGML_ASSERT(dim % 4 == 0);
         return Rope::embed_2d_interleaved(height, width, dim, theta, scale, ref_grid_h, ref_grid_w);
     }
@@ -867,13 +867,13 @@ namespace Pid {
             int64_t Hs = Hp / config.patch_size;
             int64_t Ws = Wp / config.patch_size;
 
-            pos_img_vec  = finish_rope_pe(make_rope_2d(static_cast<int>(Hs),
-                                                       static_cast<int>(Ws),
-                                                       static_cast<int>(config.hidden_size / config.num_groups),
-                                                       10000.f,
-                                                       16.f,
-                                                       static_cast<int>(config.rope_ref_grid_h),
-                                                       static_cast<int>(config.rope_ref_grid_w)));
+            pos_img_vec  = make_rope_2d(static_cast<int>(Hs),
+                                        static_cast<int>(Ws),
+                                        static_cast<int>(config.hidden_size / config.num_groups),
+                                        10000.f,
+                                        16.f,
+                                        static_cast<int>(config.rope_ref_grid_h),
+                                        static_cast<int>(config.rope_ref_grid_w));
             auto pos_img = ggml_new_tensor_4d(compute_ctx,
                                               GGML_TYPE_F32,
                                               2,
@@ -904,13 +904,13 @@ namespace Pid {
                                                 1);
             set_backend_tensor_data(pixel_pos, pixel_pos_vec.data());
 
-            pixel_pos_comp_vec  = finish_rope_pe(make_rope_2d(static_cast<int>(Hs),
-                                                              static_cast<int>(Ws),
-                                                              static_cast<int>(config.pixel_attn_hidden_size / config.pixel_num_groups),
-                                                              10000.f,
-                                                              16.f,
-                                                              static_cast<int>(config.rope_ref_grid_h),
-                                                              static_cast<int>(config.rope_ref_grid_w)));
+            pixel_pos_comp_vec  = make_rope_2d(static_cast<int>(Hs),
+                                               static_cast<int>(Ws),
+                                               static_cast<int>(config.pixel_attn_hidden_size / config.pixel_num_groups),
+                                               10000.f,
+                                               16.f,
+                                               static_cast<int>(config.rope_ref_grid_h),
+                                               static_cast<int>(config.rope_ref_grid_w));
             auto pixel_pos_comp = ggml_new_tensor_4d(compute_ctx,
                                                      GGML_TYPE_F32,
                                                      2,

src/model/diffusion/pixart.hpp

@@ -1,390 +0,0 @@
-#ifndef __SD_MODEL_DIFFUSION_PIXART_HPP__
-#define __SD_MODEL_DIFFUSION_PIXART_HPP__
-
-#include <algorithm>
-#include <cinttypes>
-#include <cmath>
-#include <vector>
-
-#include "core/ggml_extend.h"
-#include "core/ggml_runner.h"
-#include "core/util.h"
-#include "model/common/ggml_block.hpp"
-#include "model/diffusion/dit.hpp"
-#include "model/diffusion/mmdit.hpp"
-#include "model/diffusion/model.hpp"
-#include "model_loader.h"
-
-// Ref: https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformers/pixart_transformer_2d.py
-// Ref: https://github.com/PixArt-alpha/PixArt-sigma
-
-namespace PixArt {
-    constexpr int PIXART_GRAPH_SIZE = 20480;
-    constexpr int ADALN_EMBED_DIM   = 256;
-
-    struct PixArtConfig {
-        int64_t in_channels         = 4;
-        int64_t out_channels        = 8;  // learn_sigma: noise prediction + learned variance
-        int64_t hidden_size         = 1152;
-        int64_t cross_attention_dim = 1152;
-        int64_t caption_channels    = 4096;
-        int64_t num_heads           = 16;
-        int64_t patch_size          = 2;
-        int64_t ffn_dim             = 4608;
-        int64_t pos_embed_base_size = 64;
-        float interpolation_scale   = 2.f;
-        int num_layers              = 28;
-
-        static PixArtConfig detect_from_weights(const String2TensorStorage& weights, const std::string& prefix) {
-            PixArtConfig config;
-            auto find = [&](const std::string& suffix) -> const TensorStorage* {
-                auto it = weights.find(prefix + "." + suffix);
-                return it == weights.end() ? nullptr : &it->second;
-            };
-            if (auto w = find("pos_embed.proj.weight")) {
-                config.hidden_size = w->ne[3];
-                config.in_channels = w->ne[2];
-                config.patch_size  = w->ne[0];
-            }
-            if (auto w = find("proj_out.weight")) {
-                config.out_channels = w->ne[1] / (config.patch_size * config.patch_size);
-            }
-            if (auto w = find("caption_projection.linear_1.weight")) {
-                config.caption_channels = w->ne[0];
-            }
-            if (auto w = find("transformer_blocks.0.attn2.to_k.weight")) {
-                config.cross_attention_dim = w->ne[0];
-            }
-            if (auto w = find("transformer_blocks.0.ff.net.0.proj.weight")) {
-                config.ffn_dim = w->ne[1];
-            }
-            if (find("adaln_single.emb.resolution_embedder.linear_1.weight") != nullptr) {
-                LOG_WARN("pixart: resolution/aspect-ratio micro conditions are not supported; output may differ from the reference");
-            }
-            int layers                     = 0;
-            const std::string block_prefix = prefix + ".transformer_blocks.";
-            for (const auto& [name, _] : weights) {
-                if (starts_with(name, block_prefix)) {
-                    layers = std::max(layers, atoi(name.substr(block_prefix.size()).c_str()) + 1);
-                }
-            }
-            if (layers > 0) {
-                config.num_layers = layers;
-                LOG_VERBOSE("pixart: layers = %d, hidden_size = %" PRId64,
-                            layers, config.hidden_size);
-            }
-            return config;
-        }
-    };
-
-    // Mirrors diffusers get_2d_sincos_pos_embed for a (gh, gw) patch grid.
-    static std::vector<float> gen_2d_sincos_pos_embed(int64_t dim,
-                                                      int64_t gh,
-                                                      int64_t gw,
-                                                      int64_t base_size,
-                                                      float interpolation_scale) {
-        // diffusers: meshgrid(grid_w, grid_h, indexing="xy") -> grid[0]=w, grid[1]=h,
-        // embedding = concat(sincos(w), sincos(h))
-        std::vector<float> out(static_cast<size_t>(gh) * gw * dim);
-        int64_t quarter = dim / 4;
-        for (int64_t h = 0; h < gh; ++h) {
-            float pos_h = static_cast<float>(h) / (static_cast<float>(gh) / base_size) / interpolation_scale;
-            for (int64_t w = 0; w < gw; ++w) {
-                float pos_w  = static_cast<float>(w) / (static_cast<float>(gw) / base_size) / interpolation_scale;
-                float* dst_w = out.data() + (h * gw + w) * dim;
-                float* dst_h = dst_w + dim / 2;
-                for (int64_t i = 0; i < quarter; ++i) {
-                    float omega        = 1.f / powf(10000.f, static_cast<float>(i) / quarter);
-                    dst_w[i]           = sinf(pos_w * omega);
-                    dst_w[i + quarter] = cosf(pos_w * omega);
-                    dst_h[i]           = sinf(pos_h * omega);
-                    dst_h[i + quarter] = cosf(pos_h * omega);
-                }
-            }
-        }
-        return out;
-    }
-
-    class PixArtTimestepEmbedding : public GGMLBlock {
-    public:
-        PixArtTimestepEmbedding(int64_t in_channels, int64_t out_dim) {
-            blocks["linear_1"] = std::make_shared<Linear>(in_channels, out_dim);
-            blocks["linear_2"] = std::make_shared<Linear>(out_dim, out_dim);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            x = std::dynamic_pointer_cast<Linear>(blocks["linear_1"])->forward(ctx, x);
-            x = ggml_silu(ctx->ggml_ctx, x);
-            return std::dynamic_pointer_cast<Linear>(blocks["linear_2"])->forward(ctx, x);
-        }
-    };
-
-    class PixArtAttention : public GGMLBlock {
-        int64_t num_heads;
-
-    public:
-        PixArtAttention(int64_t dim, int64_t num_heads, int64_t context_dim)
-            : num_heads(num_heads) {
-            blocks["to_q"]     = std::make_shared<Linear>(dim, dim);
-            blocks["to_k"]     = std::make_shared<Linear>(context_dim, dim);
-            blocks["to_v"]     = std::make_shared<Linear>(context_dim, dim);
-            blocks["to_out.0"] = std::make_shared<Linear>(dim, dim);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* context, ggml_tensor* mask = nullptr) {
-            // x: [N, n_token, dim], context: [N, n_context, context_dim]
-            auto q = std::dynamic_pointer_cast<Linear>(blocks["to_q"])->forward(ctx, x);
-            auto k = std::dynamic_pointer_cast<Linear>(blocks["to_k"])->forward(ctx, context);
-            auto v = std::dynamic_pointer_cast<Linear>(blocks["to_v"])->forward(ctx, context);
-
-            auto out = ggml_ext_attention_ext(ctx, q, k, v, num_heads, mask, false, ctx->flash_attn_enabled);
-            return std::dynamic_pointer_cast<Linear>(blocks["to_out.0"])->forward(ctx, out);
-        }
-    };
-
-    class PixArtBlock : public GGMLBlock {
-        int64_t dim;
-
-        void init_params(ggml_context* ctx,
-                         const String2TensorStorage& tensor_storage_map = {},
-                         const std::string prefix                       = "") override {
-            ggml_type wtype             = get_type(prefix + "scale_shift_table", tensor_storage_map, GGML_TYPE_F32);
-            params["scale_shift_table"] = ggml_new_tensor_2d(ctx, wtype, dim, 6);
-        }
-
-    public:
-        PixArtBlock(int64_t dim, int64_t num_heads, int64_t context_dim, int64_t ffn_dim)
-            : dim(dim) {
-            blocks["attn1"]         = std::make_shared<PixArtAttention>(dim, num_heads, dim);
-            blocks["attn2"]         = std::make_shared<PixArtAttention>(dim, num_heads, context_dim);
-            blocks["ff.net.0.proj"] = std::make_shared<Linear>(dim, ffn_dim);
-            blocks["ff.net.2"]      = std::make_shared<Linear>(ffn_dim, dim);
-        }
-
-        static ggml_tensor* norm(ggml_context* ctx, ggml_tensor* x) {
-            return ggml_norm(ctx, x, 1e-6f);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* mod, ggml_tensor* context, ggml_tensor* context_mask) {
-            // x: [N, n_token, dim]
-            // mod: [N, 6 * dim], shared adaLN-single output
-            int64_t N = x->ne[2];
-
-            auto table = params["scale_shift_table"];
-            if (table->type != GGML_TYPE_F32) {
-                table = ggml_cast(ctx->ggml_ctx, table, GGML_TYPE_F32);
-            }
-            table = ggml_reshape_3d(ctx->ggml_ctx, table, dim, 6, 1);
-            auto m     = ggml_add(ctx->ggml_ctx, ggml_reshape_3d(ctx->ggml_ctx, mod, dim, 6, N), table);
-            auto mv    = ggml_ext_chunk(ctx->ggml_ctx, ggml_reshape_2d(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, m), dim * 6, N), 6, 0);
-
-            auto attn1 = std::dynamic_pointer_cast<PixArtAttention>(blocks["attn1"]);
-            auto attn2 = std::dynamic_pointer_cast<PixArtAttention>(blocks["attn2"]);
-            auto proj  = std::dynamic_pointer_cast<Linear>(blocks["ff.net.0.proj"]);
-            auto fc2   = std::dynamic_pointer_cast<Linear>(blocks["ff.net.2"]);
-
-            auto gate = [&](ggml_tensor* y, ggml_tensor* g) {
-                g = ggml_reshape_3d(ctx->ggml_ctx, g, dim, 1, N);
-                return ggml_add(ctx->ggml_ctx, x, ggml_mul(ctx->ggml_ctx, y, g));
-            };
-
-            auto h = modulate(ctx->ggml_ctx, norm(ctx->ggml_ctx, x), mv[0], mv[1]);
-            x      = gate(attn1->forward(ctx, h, h), mv[2]);
-            // ada_norm_single: no norm before cross-attention (PixArtMS.py)
-            x = ggml_add(ctx->ggml_ctx, x, attn2->forward(ctx, x, context, context_mask));
-            h = modulate(ctx->ggml_ctx, norm(ctx->ggml_ctx, x), mv[3], mv[4]);
-            h = proj->forward(ctx, h);
-            h = ggml_ext_gelu(ctx->ggml_ctx, h, true);
-            h = fc2->forward(ctx, h);
-            return gate(h, mv[5]);
-        }
-    };
-
-    class PixArtModel : public GGMLBlock {
-        PixArtConfig config;
-
-        void init_params(ggml_context* ctx,
-                         const String2TensorStorage& tensor_storage_map = {},
-                         const std::string prefix                       = "") override {
-            ggml_type wtype             = get_type(prefix + "scale_shift_table", tensor_storage_map, GGML_TYPE_F32);
-            params["scale_shift_table"] = ggml_new_tensor_2d(ctx, wtype, config.hidden_size, 2);
-        }
-
-    public:
-        PixArtModel() = default;
-        PixArtModel(const PixArtConfig& config)
-            : config(config) {
-            blocks["pos_embed.proj"]                     = std::make_shared<Conv2d>(config.in_channels,
-                                                                                    config.hidden_size,
-                                                                                    std::pair<int, int>{static_cast<int>(config.patch_size), static_cast<int>(config.patch_size)},
-                                                                                    std::pair<int, int>{static_cast<int>(config.patch_size), static_cast<int>(config.patch_size)});
-            blocks["adaln_single.emb.timestep_embedder"] = std::make_shared<PixArtTimestepEmbedding>(ADALN_EMBED_DIM, config.hidden_size);
-            blocks["adaln_single.linear"]                = std::make_shared<Linear>(config.hidden_size, 6 * config.hidden_size);
-            blocks["caption_projection.linear_1"]        = std::make_shared<Linear>(config.caption_channels, config.hidden_size);
-            blocks["caption_projection.linear_2"]        = std::make_shared<Linear>(config.hidden_size, config.cross_attention_dim);
-            for (int i = 0; i < config.num_layers; ++i) {
-                blocks["transformer_blocks." + std::to_string(i)] =
-                    std::make_shared<PixArtBlock>(config.hidden_size, config.num_heads, config.cross_attention_dim, config.ffn_dim);
-            }
-            blocks["norm_out"] = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false);
-            blocks["proj_out"] = std::make_shared<Linear>(config.hidden_size,
-                                                          config.patch_size * config.patch_size * config.out_channels);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                             ggml_tensor* x,
-                             ggml_tensor* timesteps,
-                             ggml_tensor* context,
-                             ggml_tensor* pos_embed,
-                             ggml_tensor* context_mask) {
-            // x: [N, C, H, W] latent, context: [N, n_ctx, caption_channels]
-            int64_t W  = x->ne[0];
-            int64_t H  = x->ne[1];
-            int64_t N  = x->ne[3];
-            int64_t p  = config.patch_size;
-            int64_t wp = W / p;
-            int64_t hp = H / p;
-
-            auto h = std::dynamic_pointer_cast<Conv2d>(blocks["pos_embed.proj"])->forward(ctx, x);  // [N, hidden, hp, wp]
-            h      = ggml_ext_cont(ctx->ggml_ctx, ggml_permute(ctx->ggml_ctx, h, 1, 2, 0, 3));      // [N, hp, wp, hidden] -> [N, hp*wp, hidden]
-            h      = ggml_reshape_3d(ctx->ggml_ctx, h, config.hidden_size, wp * hp, N);             // [N, hp*wp, hidden]
-            h      = ggml_add(ctx->ggml_ctx, h, pos_embed);
-
-            auto t   = ggml_ext_timestep_embedding(ctx->ggml_ctx, timesteps, ADALN_EMBED_DIM, 10000);
-            auto emb = std::dynamic_pointer_cast<PixArtTimestepEmbedding>(blocks["adaln_single.emb.timestep_embedder"])->forward(ctx, t);
-
-            auto mod = std::dynamic_pointer_cast<Linear>(blocks["adaln_single.linear"])
-                           ->forward(ctx, ggml_silu(ctx->ggml_ctx, emb));  // [N, 6 * hidden]
-
-            auto ctx_emb = std::dynamic_pointer_cast<Linear>(blocks["caption_projection.linear_1"])->forward(ctx, context);
-            ctx_emb      = ggml_ext_gelu(ctx->ggml_ctx, ctx_emb, true);
-            ctx_emb      = std::dynamic_pointer_cast<Linear>(blocks["caption_projection.linear_2"])->forward(ctx, ctx_emb);
-
-            for (int i = 0; i < config.num_layers; ++i) {
-                auto block = std::dynamic_pointer_cast<PixArtBlock>(blocks["transformer_blocks." + std::to_string(i)]);
-                h          = block->forward(ctx, h, mod, ctx_emb, context_mask);
-                sd::ggml_graph_cut::mark_graph_cut(h, "pixart.transformer_blocks." + std::to_string(i), "h");
-            }
-
-            // scale_shift_table + emb -> (shift, scale) for the affine-free final norm
-            auto tail_table = params["scale_shift_table"];
-            if (tail_table->type != GGML_TYPE_F32) {
-                tail_table = ggml_cast(ctx->ggml_ctx, tail_table, GGML_TYPE_F32);
-            }
-            auto ss    = ggml_add(ctx->ggml_ctx,
-                                  ggml_reshape_3d(ctx->ggml_ctx, tail_table, config.hidden_size, 2, 1),
-                                  ggml_reshape_3d(ctx->ggml_ctx, emb, config.hidden_size, 1, N));  // [2, hidden, N]
-            auto parts = ggml_ext_chunk(ctx->ggml_ctx,
-                                        ggml_reshape_2d(ctx->ggml_ctx, ggml_ext_cont(ctx->ggml_ctx, ss), config.hidden_size * 2, N),
-                                        2, 0);
-            h          = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_out"])->forward(ctx, h);
-            h          = modulate(ctx->ggml_ctx, h, parts[0], parts[1]);
-            h          = std::dynamic_pointer_cast<Linear>(blocks["proj_out"])->forward(ctx, h);  // [N, hp*wp, p*p*out_ch]
-            h          = DiT::unpatchify(ctx->ggml_ctx, h, hp, wp, static_cast<int>(p), static_cast<int>(p), false);
-            return h;  // [N, out_channels, H, W]
-        }
-    };
-
-    struct PixArtRunner : public DiffusionModelRunner {
-        PixArtConfig config;
-        PixArtModel model;
-        std::vector<float> pos_vec;
-
-        PixArtRunner(ggml_backend_t backend,
-                     const String2TensorStorage& tensor_storage_map      = {},
-                     const std::string prefix                            = "",
-                     std::shared_ptr<RunnerWeightManager> weight_manager = nullptr,
-                     const char* model_args                              = nullptr)
-            : DiffusionModelRunner(backend, prefix, weight_manager),
-              config(PixArtConfig::detect_from_weights(tensor_storage_map, prefix)) {
-            for (const auto& [key, value] : parse_key_value_args(model_args, "model arg")) {
-                if (key == "pixart_pos_embed_base_size") {
-                    int parsed = 0;
-                    if (parse_strict_int(value, parsed)) {
-                        config.pos_embed_base_size = parsed;
-                    } else {
-                        LOG_WARN("ignoring invalid PixArt model arg '%s=%s'", key.c_str(), value.c_str());
-                    }
-                } else if (key == "pixart_interpolation_scale") {
-                    float parsed = 0.f;
-                    if (parse_strict_float(value, parsed)) {
-                        config.interpolation_scale = parsed;
-                    } else {
-                        LOG_WARN("ignoring invalid PixArt model arg '%s=%s'", key.c_str(), value.c_str());
-                    }
-                }
-            }
-            model = PixArtModel(config);
-            model.init(params_ctx, tensor_storage_map, prefix);
-        }
-
-        std::string get_desc() override {
-            return "pixart";
-        }
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
-            model.get_param_tensors(tensors, prefix);
-        }
-
-        ggml_cgraph* build_graph(const sd::Tensor<float>& x_tensor,
-                                 const sd::Tensor<float>& timesteps_tensor,
-                                 const sd::Tensor<float>& context_tensor,
-                                 const sd::Tensor<float>& mask_tensor) {
-            ggml_cgraph* gf        = new_graph_custom(PIXART_GRAPH_SIZE);
-            ggml_tensor* x         = make_input(x_tensor);
-            ggml_tensor* timesteps = make_input(timesteps_tensor);
-            GGML_ASSERT(x->ne[3] == 1);
-            GGML_ASSERT(!context_tensor.empty());
-            ggml_tensor* context = make_input(context_tensor);
-
-            ggml_tensor* context_mask = nullptr;
-            if (!mask_tensor.empty()) {
-                // additive attention bias over context tokens: 0 keep / -inf discard
-                context_mask = ggml_reshape_4d(compute_ctx, make_input(mask_tensor), mask_tensor.shape()[0], 1, 1, 1);
-            }
-
-            int64_t W  = x->ne[0];
-            int64_t H  = x->ne[1];
-            int64_t wp = W / config.patch_size;
-            int64_t hp = H / config.patch_size;
-
-            pos_vec  = gen_2d_sincos_pos_embed(config.hidden_size, hp, wp,
-                                               config.pos_embed_base_size, config.interpolation_scale);
-            auto pos = ggml_new_tensor_3d(compute_ctx, GGML_TYPE_F32, config.hidden_size, wp * hp, 1);
-            set_backend_tensor_data(pos, pos_vec.data());
-
-            auto runner_ctx  = get_context();
-            ggml_tensor* out = model.forward(&runner_ctx, x, timesteps, context, pos, context_mask);
-            // learn_sigma: keep the noise prediction half of the output channels
-            out = ggml_ext_slice(compute_ctx, out, 2, 0, config.in_channels);
-            ggml_build_forward_expand(gf, out);
-            return gf;
-        }
-
-        sd::Tensor<float> compute(int n_threads,
-                                  const sd::Tensor<float>& x,
-                                  const sd::Tensor<float>& timesteps,
-                                  const sd::Tensor<float>& context,
-                                  const sd::Tensor<float>& context_mask) {
-            auto get_graph = [&]() -> ggml_cgraph* {
-                return build_graph(x, timesteps, context, context_mask);
-            };
-            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, false), x.dim());
-        }
-
-        sd::Tensor<float> compute(int n_threads,
-                                  const DiffusionParams& diffusion_params) override {
-            GGML_ASSERT(diffusion_params.x != nullptr);
-            GGML_ASSERT(diffusion_params.timesteps != nullptr);
-            auto context     = tensor_or_empty(diffusion_params.context);
-            auto context_msk = tensor_or_empty(diffusion_params.y);
-            return compute(n_threads,
-                           *diffusion_params.x,
-                           *diffusion_params.timesteps,
-                           context,
-                           context_msk);
-        }
-    };
-}  // namespace PixArt
-
-#endif  // __SD_MODEL_DIFFUSION_PIXART_HPP__

src/model/diffusion/qwen_image.hpp

@@ -635,16 +635,18 @@ namespace Qwen {
                 ref_index_mode = Rope::RefIndexMode::DECREASE;
             }
 
-            pe_vec      = finish_rope_pe(Rope::gen_qwen_image_pe(time_len,
-                                                                 static_cast<int>(x->ne[1]),
-                                                                 static_cast<int>(x->ne[0]),
-                                                                 config.patch_size,
-                                                                 batch_size,
-                                                                 static_cast<int>(context->ne[1]),
-                                                                 ref_latents,
-                                                                 ref_index_mode,
-                                                                 config.theta,
-                                                                 config.axes_dim));
+            pe_vec      = Rope::gen_qwen_image_pe(time_len,
+                                                  static_cast<int>(x->ne[1]),
+                                                  static_cast<int>(x->ne[0]),
+                                                  config.patch_size,
+                                                  batch_size,
+                                                  static_cast<int>(context->ne[1]),
+                                                  ref_latents,
+                                                  ref_index_mode,
+                                                  config.theta,
+                                                  circular_y_enabled,
+                                                  circular_x_enabled,
+                                                  config.axes_dim);
             int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             // LOG_VERBOSE("pos_len %d", pos_len);
             auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);

src/model/diffusion/qwen_image_2_1.hpp

@@ -1,594 +0,0 @@
-#ifndef __SD_MODEL_DIFFUSION_QWEN_IMAGE_2_1_H__
-#define __SD_MODEL_DIFFUSION_QWEN_IMAGE_2_1_H__
-
-#include "model/diffusion/qwen_image.hpp"
-
-namespace Qwen {
-
-    struct QwenImage21Config {
-        int64_t in_channels       = 64;
-        int64_t out_channels      = 64;
-        int64_t hidden_size       = 4096;
-        int64_t context_dim       = 4096;
-        int64_t head_dim          = 128;
-        int64_t intermediate_size = 12288;
-        int num_layers            = 32;
-        bool fused_mlp            = false;
-        std::vector<int> axes_dim = {16, 56, 56};
-
-        static QwenImage21Config detect_from_weights(const String2TensorStorage& weights, const std::string& prefix) {
-            QwenImage21Config config;
-            auto find = [&](const std::string& suffix) -> const TensorStorage* {
-                auto it = weights.find(prefix + "." + suffix);
-                return it == weights.end() ? nullptr : &it->second;
-            };
-            if (auto w = find("img_in.weight")) {
-                config.in_channels = w->ne[0];
-                config.hidden_size = w->ne[1];
-            }
-            if (auto w = find("proj_out.weight")) {
-                config.out_channels = w->ne[1];
-            }
-            if (auto w = find("txt_in.in_layer.weight")) {
-                config.context_dim = w->ne[0];
-            }
-            if (auto w = find("transformer_blocks.0.attn.norm_q.weight")) {
-                config.head_dim = w->ne[0];
-            }
-            if (auto w = find("transformer_blocks.0.img_mlp.gate_up.weight")) {
-                config.intermediate_size = w->ne[1] / 2;
-                config.fused_mlp         = true;
-            } else if (auto w = find("transformer_blocks.0.img_mlp.proj.weight")) {
-                config.intermediate_size = w->ne[1];
-            }
-            int layers                     = 0;
-            const std::string block_prefix = prefix + ".transformer_blocks.";
-            for (const auto& [name, _] : weights) {
-                if (starts_with(name, block_prefix)) {
-                    layers = std::max(layers, atoi(name.substr(block_prefix.size()).c_str()) + 1);
-                }
-            }
-            if (layers > 0) {
-                config.num_layers = layers;
-                LOG_VERBOSE("qwen_image_2_1: layers = %d, hidden_size = %" PRId64 ", context_dim = %" PRId64,
-                            layers, config.hidden_size, config.context_dim);
-            }
-            return config;
-        }
-    };
-
-    struct QwenImage21Segment {
-        int64_t start;
-        int64_t end;
-        int64_t context_start;
-        int image_index;
-    };
-
-    struct QwenImage21Layout {
-        std::vector<QwenImage21Segment> segments;
-        std::vector<std::vector<float>> positions;
-        int64_t prefix_length = 0;
-        Rope::PositionLayout rope_layout;
-
-        static QwenImage21Layout build(int64_t text_length,
-                                       const sd::Tensor<int32_t>& image_slots,
-                                       const std::vector<std::pair<int64_t, int64_t>>& image_shapes) {
-            if (image_shapes.empty() || (!image_slots.empty() && image_slots.numel() != text_length)) {
-                throw std::runtime_error("Qwen Image 2.1: invalid image token layout");
-            }
-            QwenImage21Layout layout;
-            int64_t position  = 0;
-            int next_image    = 0;
-            auto append_image = [&](int index, int64_t context_start) {
-                auto [height, width] = image_shapes[index];
-                int64_t start        = static_cast<int64_t>(layout.positions.size());
-                layout.segments.push_back({start, start + height * width, context_start, index});
-                layout.rope_layout.append_image(static_cast<int>(height), static_cast<int>(width));
-                for (int64_t h = 0; h < height; ++h) {
-                    for (int64_t w = 0; w < width; ++w) {
-                        layout.positions.push_back({static_cast<float>(position),
-                                                    static_cast<float>(h - (height - height / 2)),
-                                                    static_cast<float>(w - (width - width / 2))});
-                    }
-                }
-                position += std::max(height, width);
-            };
-            for (int64_t i = 0; i < text_length;) {
-                int tag       = image_slots.empty() ? 0 : image_slots[i];
-                int64_t begin = i++;
-                while (i < text_length && (image_slots.empty() ? 0 : image_slots[i]) == tag) {
-                    ++i;
-                }
-                if (tag != 0) {
-                    if (tag != next_image + 1 || next_image + 1 >= static_cast<int>(image_shapes.size()) ||
-                        (i - begin) * 4 != image_shapes[next_image].first * image_shapes[next_image].second) {
-                        throw std::runtime_error("Qwen Image 2.1: vision slots and reference latents must have matching sizes");
-                    }
-                    append_image(next_image++, begin);
-                } else {
-                    int64_t start = static_cast<int64_t>(layout.positions.size());
-                    layout.segments.push_back({start, start + i - begin, begin, -1});
-                    layout.rope_layout.append_tokens(i - begin);
-                    for (int64_t j = begin; j < i; ++j, ++position) {
-                        float p = static_cast<float>(position);
-                        layout.positions.push_back({p, p, p});
-                    }
-                }
-            }
-            if (next_image + 1 != static_cast<int>(image_shapes.size())) {
-                throw std::runtime_error("Qwen Image 2.1: missing reference image slots");
-            }
-            layout.prefix_length = static_cast<int64_t>(layout.positions.size());
-            append_image(next_image, text_length);
-            return layout;
-        }
-    };
-
-    struct QwenImage21PrefixCache {
-        enum class Mode {
-            NONE,
-            STORE,
-            REUSE
-        };
-        Mode mode = Mode::NONE;
-        std::string name;
-        std::string cut_group;
-        int64_t prefix_length = 0;
-        ggml_type type        = GGML_TYPE_F32;
-        bool* flash_attn_used = nullptr;
-    };
-
-    class QwenImage21ZeroCenterRMSNorm : public RMSNorm {
-    public:
-        using RMSNorm::RMSNorm;
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) override {
-            auto weight = params["weight"];
-            if (ctx->weight_adapter) {
-                weight = ctx->weight_adapter->patch_weight(ctx->ggml_ctx, ctx->backend, weight, prefix + "weight");
-            }
-            weight = ggml_scale_bias(ctx->ggml_ctx, weight, 1.f, 1.f);
-            return ggml_mul(ctx->ggml_ctx, ggml_rms_norm(ctx->ggml_ctx, x, eps), weight);
-        }
-    };
-
-    class QwenImage21TextProjection : public GGMLBlock {
-    public:
-        QwenImage21TextProjection(const QwenImage21Config& config) {
-            blocks["text_norm"] = std::make_shared<QwenImage21ZeroCenterRMSNorm>(config.context_dim, 1e-6f);
-            blocks["in_layer"]  = std::make_shared<Linear>(config.context_dim, config.hidden_size, false);
-            blocks["out_layer"] = std::make_shared<Linear>(config.hidden_size, config.hidden_size, false);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            x = std::dynamic_pointer_cast<QwenImage21ZeroCenterRMSNorm>(blocks["text_norm"])->forward(ctx, x);
-            x = std::dynamic_pointer_cast<Linear>(blocks["in_layer"])->forward(ctx, x);
-            x = ggml_ext_gelu(ctx->ggml_ctx, x);
-            return std::dynamic_pointer_cast<Linear>(blocks["out_layer"])->forward(ctx, x);
-        }
-    };
-
-    class QwenImage21Attention : public QwenImageAttention {
-    public:
-        QwenImage21Attention(const QwenImage21Config& config)
-            : QwenImageAttention(config.hidden_size, config.head_dim, config.hidden_size / config.head_dim, 0, 0, false, false) {
-            for (const auto* name : {"add_q_proj", "add_k_proj", "add_v_proj", "norm_added_q", "norm_added_k", "to_add_out"}) {
-                blocks.erase(name);
-            }
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* pe, const std::vector<QwenImage21Segment>& segments, const std::vector<ggml_tensor*>& masks, const QwenImage21PrefixCache& cache) {
-            int64_t heads = x->ne[0] / dim_head;
-            auto project  = [&](const char* name) {
-                auto h = std::dynamic_pointer_cast<Linear>(blocks[name])->forward(ctx, x);
-                return ggml_reshape_4d(ctx->ggml_ctx, h, dim_head, heads, x->ne[1], x->ne[2]);
-            };
-            auto q = project("to_q");
-            auto k = project("to_k");
-            auto v = project("to_v");
-            q      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_q"])->forward(ctx, q);
-            k      = std::dynamic_pointer_cast<RMSNorm>(blocks["norm_k"])->forward(ctx, k);
-            q      = Rope::apply_rope(ctx->ggml_ctx, q, pe);
-            k      = Rope::apply_rope(ctx->ggml_ctx, k, pe);
-            if (cache.mode == QwenImage21PrefixCache::Mode::STORE) {
-                // Preserve query-first attention evaluation while writing each layer's
-                // prefix before its full-sequence K/V can accumulate across layers.
-                ctx->expand_graph(q);
-                auto persist = [&](ggml_tensor* tensor, int axis, const char* name) {
-                    auto part = ggml_ext_slice(ctx->ggml_ctx, tensor, axis, 0, cache.prefix_length);
-                    // Pack the contiguous data into wider rows so quantization blocks
-                    // can exceed head_dim without padding or changing element order.
-                    part      = ggml_reshape_2d(ctx->ggml_ctx, part, x->ne[0], cache.prefix_length);
-                    auto copy = ggml_cast(ctx->ggml_ctx, part, cache.type);
-                    // Keep the copy in this layer's segment so graph cuts do not
-                    // retain or recompute the full-sequence K/V in the final segment.
-                    sd::ggml_graph_cut::mark_graph_cut(copy, cache.cut_group, name);
-                    ctx->persist_cache_tensor(cache.name + "." + name, copy);
-                };
-                persist(k, 1, "k");
-                persist(v, 2, "v");
-            }
-            auto attend = [&](ggml_tensor* aq, ggml_tensor* ak, ggml_tensor* av, ggml_tensor* mask) {
-                bool used_flash_attn = false;
-                auto out             = ggml_ext_attention_ext(ctx, aq, ak, av, heads, mask, true, ctx->flash_attn_enabled, 1.f, &used_flash_attn);
-                if (cache.flash_attn_used != nullptr) {
-                    *cache.flash_attn_used &= used_flash_attn;
-                }
-                return out;
-            };
-            ggml_tensor* result = nullptr;
-            if (cache.mode == QwenImage21PrefixCache::Mode::REUSE) {
-                auto prefix_k = ctx->load_cache_tensor(cache.name + ".k");
-                auto prefix_v = ctx->load_cache_tensor(cache.name + ".v");
-                GGML_ASSERT(prefix_k != nullptr && prefix_v != nullptr);
-                if (prefix_k->type != k->type) {
-                    prefix_k = ggml_cast(ctx->ggml_ctx, prefix_k, k->type);
-                }
-                if (prefix_v->type != v->type) {
-                    prefix_v = ggml_cast(ctx->ggml_ctx, prefix_v, v->type);
-                }
-                prefix_k = ggml_reshape_4d(ctx->ggml_ctx, prefix_k, dim_head, cache.prefix_length, heads, k->ne[3]);
-                prefix_v = ggml_reshape_4d(ctx->ggml_ctx, prefix_v, dim_head, heads, cache.prefix_length, v->ne[3]);
-                k        = ggml_concat(ctx->ggml_ctx, prefix_k, k, 1);
-                v        = ggml_concat(ctx->ggml_ctx, prefix_v, v, 2);
-                result   = attend(q, k, v, nullptr);
-            } else {
-                for (size_t i = 0; i < segments.size(); ++i) {
-                    const auto& segment = segments[i];
-                    auto sq             = ggml_ext_slice(ctx->ggml_ctx, q, 1, segment.start, segment.end);
-                    auto sk             = ggml_ext_slice(ctx->ggml_ctx, k, 1, 0, segment.end);
-                    auto sv             = ggml_ext_slice(ctx->ggml_ctx, v, 2, 0, segment.end);
-                    auto out            = attend(sq, sk, sv, masks[i]);
-                    result              = result == nullptr ? out : ggml_concat(ctx->ggml_ctx, result, out, 1);
-                }
-            }
-            auto to_out = std::dynamic_pointer_cast<Linear>(blocks["to_out.0"]);
-            if (sd_backend_is(ctx->backend, "Vulkan") || sd_backend_is(ctx->backend, "ROCm")) {
-                to_out->set_force_prec_f32(true);
-            }
-            return to_out->forward(ctx, result);
-        }
-    };
-
-    class QwenImage21TransformerBlock : public GGMLBlock {
-    public:
-        QwenImage21TransformerBlock(const QwenImage21Config& config) {
-            blocks["img_norm1"] = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false);
-            blocks["img_norm2"] = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false);
-            blocks["attn"]      = std::make_shared<QwenImage21Attention>(config);
-            if (config.fused_mlp) {
-                blocks["img_mlp.gate_up"] = std::make_shared<Linear>(config.hidden_size, 2 * config.intermediate_size, false);
-            } else {
-                blocks["img_mlp.proj"]       = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, false);
-                blocks["img_mlp.gate_layer"] = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, false);
-            }
-            blocks["img_mlp.out"] = std::make_shared<Linear>(config.intermediate_size, config.hidden_size, false);
-        }
-
-        static ggml_tensor* modulate(ggml_context* ctx, ggml_tensor* x, ggml_tensor* params, int64_t prefix_length, bool gate = false) {
-            auto rows  = ggml_ext_chunk(ctx, params, 2, 1);
-            auto apply = [&](ggml_tensor* part, ggml_tensor* row) {
-                row = gate ? ggml_tanh(ctx, row) : ggml_scale_bias(ctx, row, 1.f, 1.f);
-                return ggml_mul(ctx, part, row);
-            };
-            auto target = apply(ggml_ext_slice(ctx, x, 1, prefix_length, x->ne[1]), rows[0]);
-            if (prefix_length == 0) {
-                return target;
-            }
-            auto prefix = apply(ggml_ext_slice(ctx, x, 1, 0, prefix_length), rows[1]);
-            return ggml_concat(ctx, prefix, target, 1);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, const std::vector<ggml_tensor*>& modulation, ggml_tensor* pe, const QwenImage21Layout& layout, const std::vector<ggml_tensor*>& masks, const QwenImage21PrefixCache& cache) {
-            const int64_t prefix_length = cache.mode == QwenImage21PrefixCache::Mode::REUSE ? 0 : layout.prefix_length;
-            auto h                      = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm1"])->forward(ctx, x);
-            h                           = modulate(ctx->ggml_ctx, h, modulation[0], prefix_length);
-            h                           = std::dynamic_pointer_cast<QwenImage21Attention>(blocks["attn"])->forward(ctx, h, pe, layout.segments, masks, cache);
-            x                           = ggml_add(ctx->ggml_ctx, x, modulate(ctx->ggml_ctx, h, modulation[1], prefix_length, true));
-            h                           = std::dynamic_pointer_cast<LayerNorm>(blocks["img_norm2"])->forward(ctx, x);
-            h                           = modulate(ctx->ggml_ctx, h, modulation[2], prefix_length);
-            ggml_tensor* gate;
-            auto fused = blocks.find("img_mlp.gate_up");
-            if (fused != blocks.end()) {
-                auto gate_up = std::dynamic_pointer_cast<Linear>(fused->second)->forward(ctx, h);
-                auto parts   = ggml_ext_chunk(ctx->ggml_ctx, gate_up, 2, 0);
-                gate         = parts[0];
-                h            = parts[1];
-            } else {
-                gate = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.gate_layer"])->forward(ctx, h);
-                h    = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.proj"])->forward(ctx, h);
-            }
-            h = ggml_mul(ctx->ggml_ctx, h, ggml_silu(ctx->ggml_ctx, gate));
-            h = std::dynamic_pointer_cast<Linear>(blocks["img_mlp.out"])->forward(ctx, h);
-            return ggml_add(ctx->ggml_ctx, x, modulate(ctx->ggml_ctx, h, modulation[3], prefix_length, true));
-        }
-    };
-
-    class QwenImage21Model : public GGMLBlock {
-        QwenImage21Config config;
-
-    public:
-        QwenImage21Model(const QwenImage21Config& config)
-            : config(config) {
-            blocks["time_text_embed.timestep_embedder"] = std::make_shared<TimestepEmbedding>(256, config.hidden_size, 0, 0, false);
-            blocks["txt_in"]                            = std::make_shared<QwenImage21TextProjection>(config);
-            blocks["img_in"]                            = std::make_shared<Linear>(config.in_channels, config.hidden_size, false);
-            blocks["modulation.1"]                      = std::make_shared<Linear>(config.hidden_size, 4 * config.hidden_size, false);
-            blocks["norm_out.linear"]                   = std::make_shared<Linear>(config.hidden_size, config.hidden_size, false);
-            blocks["norm_out.norm"]                     = std::make_shared<LayerNorm>(config.hidden_size, 1e-6f, false);
-            blocks["proj_out"]                          = std::make_shared<Linear>(config.hidden_size, config.out_channels, false);
-            for (int i = 0; i < config.num_layers; ++i) {
-                blocks["transformer_blocks." + std::to_string(i)] = std::make_shared<QwenImage21TransformerBlock>(config);
-            }
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x, ggml_tensor* timestep, ggml_tensor* context, const std::vector<ggml_tensor*>& refs, ggml_tensor* pe, const QwenImage21Layout& layout, const std::vector<ggml_tensor*>& masks, const QwenImage21PrefixCache& cache) {
-            auto time = ggml_concat(ctx->ggml_ctx, timestep, ggml_ext_zeros_like(ctx->ggml_ctx, timestep), 0);
-            // Runtime flow timesteps already use the [0, 1000] scale.
-            time               = ggml_ext_timestep_embedding(ctx->ggml_ctx, time, 256, 10000, 1.f);
-            time               = std::dynamic_pointer_cast<TimestepEmbedding>(blocks["time_text_embed.timestep_embedder"])->forward(ctx, time);
-            time               = ggml_silu(ctx->ggml_ctx, time);
-            auto modulation    = std::dynamic_pointer_cast<Linear>(blocks["modulation.1"])->forward(ctx, time);
-            auto mod           = ggml_ext_chunk(ctx->ggml_ctx, modulation, 4, 0);
-            auto img_in        = std::dynamic_pointer_cast<Linear>(blocks["img_in"]);
-            ggml_tensor* joint = nullptr;
-            if (cache.mode == QwenImage21PrefixCache::Mode::REUSE) {
-                joint = img_in->forward(ctx, DiT::patchify(ctx->ggml_ctx, x, 1, 1));
-            } else {
-                auto text = std::dynamic_pointer_cast<QwenImage21TextProjection>(blocks["txt_in"])->forward(ctx, context);
-                for (const auto& segment : layout.segments) {
-                    ggml_tensor* h;
-                    if (segment.image_index < 0) {
-                        h = ggml_ext_slice(ctx->ggml_ctx, text, 1, segment.context_start,
-                                           segment.context_start + segment.end - segment.start);
-                    } else {
-                        auto image = segment.image_index == static_cast<int>(refs.size()) ? x : refs[segment.image_index];
-                        h          = img_in->forward(ctx, DiT::patchify(ctx->ggml_ctx, image, 1, 1));
-                    }
-                    joint = joint == nullptr ? h : ggml_concat(ctx->ggml_ctx, joint, h, 1);
-                }
-            }
-            sd::ggml_graph_cut::mark_graph_cut(joint, "qwen_image_2_1.prelude", "joint");
-            for (int i = 0; i < config.num_layers; ++i) {
-                const std::string layer = "transformer_blocks." + std::to_string(i);
-                auto layer_cache        = cache;
-                layer_cache.name        = cache.name + "." + std::to_string(i);
-                layer_cache.cut_group   = "qwen_image_2_1." + layer;
-                auto block              = std::dynamic_pointer_cast<QwenImage21TransformerBlock>(blocks[layer]);
-                joint                   = block->forward(ctx, joint, mod, pe, layout, masks, layer_cache);
-                sd::ggml_graph_cut::mark_graph_cut(joint, layer_cache.cut_group, "joint");
-            }
-            if (cache.mode != QwenImage21PrefixCache::Mode::REUSE) {
-                joint = ggml_ext_slice(ctx->ggml_ctx, joint, 1, layout.prefix_length, joint->ne[1]);
-            }
-            auto scale = std::dynamic_pointer_cast<Linear>(blocks["norm_out.linear"])->forward(ctx, ggml_ext_chunk(ctx->ggml_ctx, time, 2, 1)[0]);
-            joint      = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_out.norm"])->forward(ctx, joint);
-            joint      = ggml_mul(ctx->ggml_ctx, joint, ggml_scale_bias(ctx->ggml_ctx, scale, 1.f, 1.f));
-            joint      = std::dynamic_pointer_cast<Linear>(blocks["proj_out"])->forward(ctx, joint);
-            return DiT::unpatchify_and_crop(ctx->ggml_ctx, joint, x->ne[1], x->ne[0], 1, 1);
-        }
-    };
-
-    struct QwenImage21Runner : public DiffusionModelRunner {
-        QwenImage21Config config;
-        QwenImage21Model model;
-        std::vector<float> pe_data;
-        std::vector<sd::Tensor<float>> mask_data;
-        ggml_type prefix_cache_type = GGML_TYPE_COUNT;
-        bool prefix_cache_enabled   = true;
-        bool prefix_cache_disabled  = false;
-        bool prefix_cache_auto_f32  = false;
-
-        static bool supports_prefix_cache_type(ggml_type type) {
-            if (type == GGML_TYPE_F32) {
-                return true;
-            }
-            const auto* traits = ggml_get_type_traits(type);
-            if (traits->from_float_ref == nullptr || traits->to_float == nullptr) {
-                return false;
-            }
-            auto cpu = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
-            if (cpu == nullptr) {
-                return false;
-            }
-            auto ctx = std::unique_ptr<ggml_context, decltype(&ggml_free)>(
-                ggml_init({3 * ggml_tensor_overhead(), nullptr, true}), ggml_free);
-            if (ctx == nullptr) {
-                return false;
-            }
-            // Some reference quantizers have no runtime copy support. Query the
-            // device through the registry so dynamically loaded CPU backends work.
-            auto source  = ggml_new_tensor_1d(ctx.get(), GGML_TYPE_F32, ggml_blck_size(type));
-            auto encoded = ggml_cast(ctx.get(), source, type);
-            auto decoded = ggml_cast(ctx.get(), encoded, GGML_TYPE_F32);
-            return ggml_backend_dev_supports_op(cpu, encoded) && ggml_backend_dev_supports_op(cpu, decoded);
-        }
-
-        QwenImage21Runner(ggml_backend_t backend, const String2TensorStorage& weights, const std::string& prefix, std::shared_ptr<RunnerWeightManager> weight_manager = nullptr, const char* model_args = nullptr)
-            : DiffusionModelRunner(backend, prefix, weight_manager),
-              config(QwenImage21Config::detect_from_weights(weights, prefix)),
-              model(config) {
-            for (const auto& [key, value] : parse_key_value_args(model_args, "model arg")) {
-                if (key == "qwen_image_2_1_prefix_cache" && !parse_strict_bool(value, prefix_cache_enabled)) {
-                    LOG_WARN("ignoring invalid Qwen Image 2.1 model arg '%s=%s'", key.c_str(), value.c_str());
-                } else if (key == "qwen_image_2_1_prefix_cache_type") {
-                    if (value == "auto") {
-                        prefix_cache_type = GGML_TYPE_COUNT;
-                        continue;
-                    }
-                    const auto type = sd_type_to_ggml_type(str_to_sd_type(value.c_str()));
-                    if (type == GGML_TYPE_COUNT) {
-                        LOG_WARN("ignoring unknown Qwen Image 2.1 cache type '%s'", value.c_str());
-                    } else if (!supports_prefix_cache_type(type)) {
-                        LOG_WARN("ignoring Qwen Image 2.1 cache type '%s': runtime conversion to and from F32 is unavailable", value.c_str());
-                    } else if (config.hidden_size % ggml_blck_size(type) != 0) {
-                        LOG_WARN("ignoring Qwen Image 2.1 cache type '%s': block size %" PRId64 " does not divide hidden size %" PRId64,
-                                 value.c_str(), ggml_blck_size(type), config.hidden_size);
-                    } else {
-                        prefix_cache_type = type;
-                    }
-                }
-            }
-            model.init(params_ctx, weights, prefix);
-        }
-
-        std::string get_desc() override { return "qwen_image_2_1"; }
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) override {
-            model.get_param_tensors(tensors, prefix);
-        }
-
-        bool has_prefix_cache(const QwenImage21PrefixCache& cache) {
-            for (int i = 0; i < config.num_layers; ++i) {
-                const auto name = cache.name + "." + std::to_string(i);
-                auto k          = get_cache_tensor_by_name(name + ".k");
-                auto v          = get_cache_tensor_by_name(name + ".v");
-                if (k == nullptr || v == nullptr || k->type != cache.type || v->type != cache.type ||
-                    k->ne[0] != config.hidden_size || k->ne[1] != cache.prefix_length || k->ne[2] != 1 || k->ne[3] != 1 ||
-                    v->ne[0] != config.hidden_size || v->ne[1] != cache.prefix_length || v->ne[2] != 1 || v->ne[3] != 1) {
-                    return false;
-                }
-            }
-            return true;
-        }
-
-        sd::Tensor<float> compute(int n_threads, const DiffusionParams& inputs) override {
-            const auto& x       = tensor_or_empty(inputs.x);
-            const auto& context = tensor_or_empty(inputs.context);
-            if (x.empty() || context.empty() || context.dim() < 2 || context.shape()[0] != config.context_dim ||
-                tensor_or_empty(inputs.timesteps).numel() != 1 ||
-                x.dim() != 4 || x.shape()[3] != 1 || x.shape()[2] != config.in_channels) {
-                LOG_ERROR("Qwen Image 2.1 requires an image latent and text conditioning with batch size 1");
-                return {};
-            }
-            static const std::vector<sd::Tensor<float>> empty_refs;
-            const auto& refs = inputs.ref_latents && inputs.ref_image_params.pass_to_dit ? *inputs.ref_latents : empty_refs;
-            std::vector<std::pair<int64_t, int64_t>> shapes;
-            for (const auto& ref : refs) {
-                if (ref.dim() != 4 || ref.shape()[2] != config.in_channels || ref.shape()[3] != 1) {
-                    LOG_ERROR("Qwen Image 2.1: invalid reference latent shape");
-                    return {};
-                }
-                shapes.emplace_back(ref.shape()[1], ref.shape()[0]);
-            }
-            shapes.emplace_back(x.shape()[1], x.shape()[0]);
-            const auto* extra = std::get_if<QwenImage21DiffusionExtra>(&inputs.extra);
-            QwenImage21Layout layout;
-            try {
-                layout = QwenImage21Layout::build(context.shape()[1], tensor_or_empty(extra ? extra->image_slots : nullptr), shapes);
-            } catch (const std::exception& error) {
-                LOG_ERROR("%s", error.what());
-                return {};
-            }
-            if (!runner_started()) {
-                prefix_cache_disabled = false;
-                prefix_cache_auto_f32 = false;
-            }
-            QwenImage21PrefixCache cache;
-            if (prefix_cache_enabled && !prefix_cache_disabled && extra != nullptr && extra->prefix_id != 0 && layout.prefix_length > 0) {
-                cache.name = "qwen_image_2_1.prefix." + std::to_string(extra->prefix_id) +
-                             ".circular." + std::to_string(circular_x_enabled) + std::to_string(circular_y_enabled);
-                cache.prefix_length = layout.prefix_length;
-                if (prefix_cache_type != GGML_TYPE_COUNT) {
-                    cache.type = prefix_cache_type;
-                } else if (!prefix_cache_auto_f32 && flash_attn_enabled && !sage_attn_enabled &&
-                           (attn_scale <= 0.f || attn_scale == 1.f)) {
-                    cache.type = GGML_TYPE_F16;
-                }
-                cache.mode = has_prefix_cache(cache) ? QwenImage21PrefixCache::Mode::REUSE : QwenImage21PrefixCache::Mode::STORE;
-            }
-            bool flash_attn_used = true;
-            auto run             = [&](const QwenImage21PrefixCache& active_cache) {
-                flash_attn_used    = true;
-                auto checked_cache = active_cache;
-                if (prefix_cache_type == GGML_TYPE_COUNT && active_cache.type == GGML_TYPE_F16) {
-                    checked_cache.flash_attn_used = &flash_attn_used;
-                }
-                const bool cached         = active_cache.mode == QwenImage21PrefixCache::Mode::REUSE;
-                const auto first_position = layout.positions.begin() + (cached ? layout.prefix_length : 0);
-                Rope::Embedding embedding;
-                embedding.ids.assign(first_position, layout.positions.end());
-                const size_t offset             = cached ? static_cast<size_t>(layout.prefix_length) : 0;
-                embedding.positions.token_count = embedding.ids.size();
-                for (auto region : layout.rope_layout.images) {
-                    if (region.begin >= offset) {
-                        region.begin -= offset;
-                        embedding.positions.images.push_back(region);
-                    }
-                }
-                embedding.values = Rope::embed_nd(embedding.ids, 1, 10000.f, config.axes_dim, embedding.layout, &embedding.frequencies);
-                pe_data          = finish_rope_pe(std::move(embedding));
-                mask_data.clear();
-                if (!cached) {
-                    for (const auto& segment : layout.segments) {
-                        sd::Tensor<float> mask;
-                        if (segment.image_index < 0) {
-                            mask = sd::Tensor<float>::zeros({segment.end, segment.end - segment.start});
-                            for (int64_t q = segment.start; q < segment.end; ++q) {
-                                for (int64_t k = q + 1; k < segment.end; ++k) {
-                                    mask[k + segment.end * (q - segment.start)] = -INFINITY;
-                                }
-                            }
-                        }
-                        mask_data.push_back(std::move(mask));
-                    }
-                }
-                auto build = [&]() {
-                    auto graph = new_graph_custom(QWEN_IMAGE_GRAPH_SIZE * 2);
-                    auto pe    = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.head_dim / 2,
-                                                                layout.positions.size() - (cached ? layout.prefix_length : 0));
-                    set_backend_tensor_data(pe, pe_data.data());
-                    std::vector<ggml_tensor*> masks, ref_inputs;
-                    for (const auto& mask : mask_data) {
-                        masks.push_back(mask.empty() ? nullptr : make_input(mask));
-                    }
-                    if (!cached) {
-                        for (const auto& ref : refs) {
-                            ref_inputs.push_back(make_input(ref));
-                        }
-                    }
-                    auto ctx = get_context(graph);
-                    auto out = model.forward(&ctx, make_input(x), make_input(*inputs.timesteps), cached ? nullptr : make_input(context),
-                                                         ref_inputs, pe, layout, masks, checked_cache);
-                    if (!flash_attn_used) {
-                        return static_cast<ggml_cgraph*>(nullptr);
-                    }
-                    ggml_build_forward_expand(graph, out);
-                    return graph;
-                };
-                return restore_trailing_singleton_dims(GGMLRunner::compute(build, n_threads, false), x.dim());
-            };
-            auto result = run(cache);
-            if (result.empty() && !flash_attn_used) {
-                // Casting an F16 cache back to F32 cannot recover its original values.
-                // Recompute the prefix before executing a graph that falls back from FA.
-                free_cache_ctx_and_buffer();
-                prefix_cache_auto_f32 = true;
-                cache.type            = GGML_TYPE_F32;
-                cache.mode            = QwenImage21PrefixCache::Mode::STORE;
-                LOG_DEBUG("Qwen Image 2.1: Flash Attention unavailable; using F32 prefix caching for this sampling run");
-                result = run(cache);
-            }
-            if (result.empty() && last_compute_status() == GGML_STATUS_ALLOC_FAILED &&
-                (cache.mode != QwenImage21PrefixCache::Mode::NONE || !cache_.empty())) {
-                // The failed graph has ended before persistent inputs are released.
-                free_cache_ctx_and_buffer();
-                prefix_cache_disabled = true;
-                LOG_WARN("Qwen Image 2.1: insufficient memory for prefix caching; retrying without it for this sampling run");
-                return run(QwenImage21PrefixCache{});
-            }
-            if (!result.empty() && cache.mode == QwenImage21PrefixCache::Mode::STORE) {
-                if (!has_prefix_cache(cache)) {
-                    free_cache_ctx_and_buffer();
-                    prefix_cache_disabled = true;
-                    LOG_WARN("Qwen Image 2.1: incomplete prefix cache; disabling it for this sampling run");
-                } else {
-                    LOG_DEBUG("Qwen Image 2.1: cached prefix %" PRIu64 " (%" PRId64 " tokens, %s)", extra->prefix_id, layout.prefix_length, ggml_type_name(cache.type));
-                }
-            }
-            return result;
-        }
-    };
-}
-
-#endif  // __SD_MODEL_DIFFUSION_QWEN_IMAGE_2_1_H__

src/model/diffusion/sensenova_u1.h

@@ -442,9 +442,16 @@ namespace SenseNovaU1 {
                 k = ggml_concat(ctx->ggml_ctx, prefix_k, k, 2);
                 v = ggml_concat(ctx->ggml_ctx, prefix_v, v, 2);
             } else {
-                ctx->expand_graph(q);
-                ctx->persist_cache_tensor(layer_cache + ".k", k);
-                ctx->persist_cache_tensor(layer_cache + ".v", v);
+                // Keep dedicated graph outputs alive until the runner copies them
+                // into its persistent cache buffer after graph execution.
+                auto cache_k = ggml_dup_tensor(ctx->ggml_ctx, k);
+                cache_k      = ggml_cpy(ctx->ggml_ctx, k, cache_k);
+                ggml_set_output(cache_k);
+                auto cache_v = ggml_dup_tensor(ctx->ggml_ctx, v);
+                cache_v      = ggml_cpy(ctx->ggml_ctx, v, cache_v);
+                ggml_set_output(cache_v);
+                ctx->persist_cache_tensor(layer_cache + ".k", cache_k);
+                ctx->persist_cache_tensor(layer_cache + ".v", cache_v);
             }
 
             q = ggml_cont(ctx->ggml_ctx,
@@ -680,7 +687,7 @@ namespace SenseNovaU1 {
             ggml_set_name(attention_mask, "snu15.prefix.attention_mask");
             set_backend_tensor_data(attention_mask, attention_mask_vec.data());
 
-            auto runner_ctx = get_context(graph);
+            auto runner_ctx = get_context();
             auto text_model = model.text_model();
             auto hidden     = text_model->embed(&runner_ctx, ids);
             hidden          = text_model->forward(&runner_ctx,

src/model/diffusion/z_image.hpp

@@ -131,30 +131,16 @@ namespace ZImage {
         int64_t num_heads;
         int64_t num_kv_heads;
         bool qk_norm;
-        bool split_qkv;
 
     public:
-        JointAttention(int64_t hidden_size,
-                       int64_t head_dim,
-                       int64_t num_heads,
-                       int64_t num_kv_heads,
-                       bool qk_norm,
-                       bool norm_elementwise_affine = true,
-                       bool split_qkv               = false)
-            : head_dim(head_dim), num_heads(num_heads), num_kv_heads(num_kv_heads), qk_norm(qk_norm), split_qkv(split_qkv) {
-            float scale = 1.f;
-            if (split_qkv) {
-                blocks["to_q"]     = std::make_shared<Linear>(hidden_size, num_heads * head_dim, false);
-                blocks["to_k"]     = std::make_shared<Linear>(hidden_size, num_kv_heads * head_dim, false);
-                blocks["to_v"]     = std::make_shared<Linear>(hidden_size, num_kv_heads * head_dim, false);
-                blocks["to_out.0"] = std::make_shared<Linear>(num_heads * head_dim, hidden_size, false, false, false, scale);
-            } else {
-                blocks["qkv"] = std::make_shared<Linear>(hidden_size, (num_heads + num_kv_heads * 2) * head_dim, false);
-                blocks["out"] = std::make_shared<Linear>(num_heads * head_dim, hidden_size, false, false, false, scale);
-            }
+        JointAttention(int64_t hidden_size, int64_t head_dim, int64_t num_heads, int64_t num_kv_heads, bool qk_norm)
+            : head_dim(head_dim), num_heads(num_heads), num_kv_heads(num_kv_heads), qk_norm(qk_norm) {
+            blocks["qkv"] = std::make_shared<Linear>(hidden_size, (num_heads + num_kv_heads * 2) * head_dim, false);
+            float scale   = 1.f;
+            blocks["out"] = std::make_shared<Linear>(num_heads * head_dim, hidden_size, false, false, false, scale);
             if (qk_norm) {
-                blocks["q_norm"] = std::make_shared<RMSNorm>(head_dim, 1e-06f, norm_elementwise_affine);
-                blocks["k_norm"] = std::make_shared<RMSNorm>(head_dim, 1e-06f, norm_elementwise_affine);
+                blocks["q_norm"] = std::make_shared<RMSNorm>(head_dim);
+                blocks["k_norm"] = std::make_shared<RMSNorm>(head_dim);
             }
         }
 
@@ -165,35 +151,8 @@ namespace ZImage {
             // x: [N, n_token, hidden_size]
             int64_t n_token = x->ne[1];
             int64_t N       = x->ne[2];
-            auto out_proj   = std::dynamic_pointer_cast<Linear>(blocks[split_qkv ? "to_out.0" : "out"]);
-
-            if (split_qkv) {
-                auto q_proj = std::dynamic_pointer_cast<Linear>(blocks["to_q"]);
-                auto k_proj = std::dynamic_pointer_cast<Linear>(blocks["to_k"]);
-                auto v_proj = std::dynamic_pointer_cast<Linear>(blocks["to_v"]);
-
-                if (sd_backend_is(ctx->backend, "ROCm")) {
-                    out_proj->set_scale(1.f / 16.f);
-                    out_proj->set_force_prec_f32(true);
-                    q_proj->set_force_prec_f32(true);
-                    k_proj->set_force_prec_f32(true);
-                    v_proj->set_force_prec_f32(true);
-                }
-
-                auto q = ggml_reshape_4d(ctx->ggml_ctx, q_proj->forward(ctx, x), head_dim, num_heads, n_token, N);
-                auto k = ggml_reshape_4d(ctx->ggml_ctx, k_proj->forward(ctx, x), head_dim, num_kv_heads, n_token, N);
-                auto v = ggml_reshape_4d(ctx->ggml_ctx, v_proj->forward(ctx, x), head_dim, num_kv_heads, n_token, N);
-
-                if (qk_norm) {
-                    q = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"])->forward(ctx, q);
-                    k = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"])->forward(ctx, k);
-                }
-
-                auto out = Rope::attention(ctx, q, k, v, pe, mask, 1.f / 128.f);
-                return out_proj->forward(ctx, out);
-            }
-
-            auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
+            auto qkv_proj   = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
+            auto out_proj   = std::dynamic_pointer_cast<Linear>(blocks["out"]);
 
             if (sd_backend_is(ctx->backend, "ROCm")) {
                 out_proj->set_scale(1.f / 16.f);
@@ -293,12 +252,9 @@ namespace ZImage {
                                             ggml_tensor* x,
                                             ggml_tensor* scale) {
         // x: [N, L, C]
-        // scale: [N, C], or [N, L, C] when the caller modulates per token (LLaDA-Image editing
-        // feeds a per-token timestep embedding so each segment carries its own modulation).
-        if (scale->ne[1] != x->ne[1]) {
-            scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
-        }
-        x = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
+        // scale: [N, C]
+        scale = ggml_reshape_3d(ctx, scale, scale->ne[0], 1, scale->ne[1]);  // [N, 1, C]
+        x     = ggml_add(ctx, x, ggml_mul(ctx, x, scale));
         return x;
     }
 
@@ -316,16 +272,14 @@ namespace ZImage {
                               float ffn_dim_multiplier,
                               float norm_eps,
                               bool qk_norm,
-                              bool modulation              = true,
-                              bool norm_elementwise_affine = true,
-                              bool split_qkv               = false)
+                              bool modulation = true)
             : modulation(modulation) {
-            blocks["attention"]       = std::make_shared<JointAttention>(hidden_size, head_dim, num_heads, num_kv_heads, qk_norm, norm_elementwise_affine, split_qkv);
+            blocks["attention"]       = std::make_shared<JointAttention>(hidden_size, head_dim, num_heads, num_kv_heads, qk_norm);
             blocks["feed_forward"]    = std::make_shared<FeedForward>(hidden_size, hidden_size, multiple_of, ffn_dim_multiplier);
-            blocks["attention_norm1"] = std::make_shared<RMSNorm>(hidden_size, norm_eps, norm_elementwise_affine);
-            blocks["ffn_norm1"]       = std::make_shared<RMSNorm>(hidden_size, norm_eps, norm_elementwise_affine);
-            blocks["attention_norm2"] = std::make_shared<RMSNorm>(hidden_size, norm_eps, norm_elementwise_affine);
-            blocks["ffn_norm2"]       = std::make_shared<RMSNorm>(hidden_size, norm_eps, norm_elementwise_affine);
+            blocks["attention_norm1"] = std::make_shared<RMSNorm>(hidden_size, norm_eps);
+            blocks["ffn_norm1"]       = std::make_shared<RMSNorm>(hidden_size, norm_eps);
+            blocks["attention_norm2"] = std::make_shared<RMSNorm>(hidden_size, norm_eps);
+            blocks["ffn_norm2"]       = std::make_shared<RMSNorm>(hidden_size, norm_eps);
             if (modulation) {
                 blocks["adaLN_modulation.0"] = std::make_shared<Linear>(MIN(hidden_size, ADALN_EMBED_DIM), 4 * hidden_size);
             }
@@ -642,16 +596,18 @@ namespace ZImage {
                 ref_latents.push_back(make_input(ref_latent_tensor));
             }
 
-            pe_vec      = finish_rope_pe(Rope::gen_z_image_pe(static_cast<int>(x->ne[1]),
-                                                              static_cast<int>(x->ne[0]),
-                                                              config.patch_size,
-                                                              static_cast<int>(x->ne[3]),
-                                                              static_cast<int>(context->ne[1]),
-                                                              SEQ_MULTI_OF,
-                                                              ref_latents,
-                                                              ref_index_mode,
-                                                              config.theta,
-                                                              config.axes_dim));
+            pe_vec      = Rope::gen_z_image_pe(static_cast<int>(x->ne[1]),
+                                               static_cast<int>(x->ne[0]),
+                                               config.patch_size,
+                                               static_cast<int>(x->ne[3]),
+                                               static_cast<int>(context->ne[1]),
+                                               SEQ_MULTI_OF,
+                                               ref_latents,
+                                               ref_index_mode,
+                                               config.theta,
+                                               circular_y_enabled,
+                                               circular_x_enabled,
+                                               config.axes_dim);
             int pos_len = static_cast<int>(pe_vec.size() / config.axes_dim_sum / 2);
             // LOG_VERBOSE("pos_len %d", pos_len);
             auto pe = ggml_new_tensor_4d(compute_ctx, GGML_TYPE_F32, 2, 2, config.axes_dim_sum / 2, pos_len);

src/model/te/llada_image_te.hpp

@@ -1,604 +0,0 @@
-#ifndef __SD_MODEL_TE_LLADA_IMAGE_TE_HPP__
-#define __SD_MODEL_TE_LLADA_IMAGE_TE_HPP__
-
-#include <algorithm>
-#include <array>
-#include <cmath>
-
-#include "core/ggml_extend.h"
-#include "core/ggml_runner.h"
-#include "model/common/ggml_block.hpp"
-#include "model_loader.h"
-
-// The conditioning components LLaDA-Image puts around its LLaDA2-MoE backbone.
-// Ref: LLaDAImageQueryFormerModel / LLaDAImageTextProjectionModel in
-// https://github.com/inclusionAI/LLaDA-Image/blob/main/src/models/transformer_llada_image.py
-//
-// QueryFormer turns the LLaDA token embeddings into 256 learned queries that the pipeline
-// appends to the backbone input; TextProjection maps the backbone hidden states to the
-// denoiser's caption dimension. Neither uses RoPE, and every norm is parameter-free.
-// Both MLPs use the tanh GELU approximation, so ggml_gelu (not ggml_gelu_erf).
-//
-// SigVQ is the editing-only image encoder: a 40-layer ViT whose output is quantized against a
-// 16384-entry codebook, with the resulting ids embedded and projected into the semantic features
-// the denoiser consumes. Its MLP uses the exact erf GELU, unlike the two above.
-
-namespace LLaDAImageTE {
-    constexpr int LLADA_IMAGE_TE_GRAPH_SIZE = 16384;
-
-    struct QueryFormerConfig {
-        int64_t num_queries       = 256;
-        int64_t hidden_size       = 2048;
-        int64_t num_layers        = 1;
-        int64_t num_heads         = 16;
-        int64_t intermediate_size = 8192;
-        float norm_eps            = 1e-6f;
-    };
-
-    struct TextProjectionConfig {
-        int64_t hidden_size       = 2048;
-        int64_t intermediate_size = 8960;
-        int64_t num_layers        = 6;
-        int64_t num_heads         = 32;
-        int64_t projection_dim    = 2560;
-        float norm_eps            = 1e-6f;
-    };
-
-    // Cross-attention with a single fused in_proj over q (from the queries) and k/v (from the
-    // token embeddings). The checkpoint stores in_proj as one [3*hidden, hidden] parameter.
-    struct QueryAttention : public GGMLBlock {
-    protected:
-        int64_t hidden_size;
-        int64_t num_heads;
-
-        void init_params(ggml_context* ctx,
-                         const String2TensorStorage& tensor_storage_map = {},
-                         std::string prefix                             = "") override {
-            GGMLBlock::init_params(ctx, tensor_storage_map, prefix);
-            enum ggml_type wtype     = get_type(prefix + "in_proj_weight", tensor_storage_map, GGML_TYPE_F32);
-            params["in_proj_weight"] = ggml_new_tensor_2d(ctx, wtype, hidden_size, hidden_size * 3);
-            params["in_proj_bias"]   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, hidden_size * 3);
-        }
-
-    public:
-        QueryAttention(int64_t hidden_size, int64_t num_heads)
-            : hidden_size(hidden_size), num_heads(num_heads) {
-            blocks["out_proj"] = std::make_shared<Linear>(hidden_size, hidden_size, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                             ggml_tensor* query,
-                             ggml_tensor* context,
-                             ggml_tensor* mask = nullptr) {
-            // query: [N, num_queries, hidden_size], context: [N, n_token, hidden_size]
-            ggml_context* gctx = ctx->ggml_ctx;
-            auto out_proj      = std::dynamic_pointer_cast<Linear>(blocks["out_proj"]);
-
-            auto w = params["in_proj_weight"];
-            auto b = params["in_proj_bias"];
-
-            auto slice_w = [&](int64_t index) {
-                return ggml_ext_slice(gctx, w, 1, index * hidden_size, (index + 1) * hidden_size);
-            };
-            auto slice_b = [&](int64_t index) {
-                return ggml_ext_slice(gctx, b, 0, index * hidden_size, (index + 1) * hidden_size);
-            };
-
-            auto q = ggml_ext_linear(gctx, query, slice_w(0), slice_b(0));
-            auto k = ggml_ext_linear(gctx, context, slice_w(1), slice_b(1));
-            auto v = ggml_ext_linear(gctx, context, slice_w(2), slice_b(2));
-
-            auto x = ggml_ext_attention_ext(ctx, q, k, v, num_heads, mask);  // [N, num_queries, hidden_size]
-            return out_proj->forward(ctx, x);
-        }
-    };
-
-    struct QueryFormerBlock : public GGMLBlock {
-    protected:
-        QueryFormerConfig config;
-
-    public:
-        QueryFormerBlock(const QueryFormerConfig& config)
-            : config(config) {
-            blocks["norm_q"]     = std::make_shared<LayerNorm>(config.hidden_size, config.norm_eps, false);
-            blocks["norm_k"]     = std::make_shared<LayerNorm>(config.hidden_size, config.norm_eps, false);
-            blocks["cross_attn"] = std::make_shared<QueryAttention>(config.hidden_size, config.num_heads);
-            blocks["norm1"]      = std::make_shared<LayerNorm>(config.hidden_size, config.norm_eps, false);
-            blocks["mlp.fc1"]    = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, true);
-            blocks["mlp.fc2"]    = std::make_shared<Linear>(config.intermediate_size, config.hidden_size, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                             ggml_tensor* query,
-                             ggml_tensor* context,
-                             ggml_tensor* mask = nullptr) {
-            auto norm_q     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_q"]);
-            auto norm_k     = std::dynamic_pointer_cast<LayerNorm>(blocks["norm_k"]);
-            auto cross_attn = std::dynamic_pointer_cast<QueryAttention>(blocks["cross_attn"]);
-            auto norm1      = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
-            auto fc1        = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc1"]);
-            auto fc2        = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc2"]);
-
-            // The reference overwrites query_embeds with its normalized value before the
-            // residual add, so both residuals here are on normalized activations.
-            query      = norm_q->forward(ctx, query);
-            auto ctx_n = norm_k->forward(ctx, context);
-            query      = ggml_add(ctx->ggml_ctx, query, cross_attn->forward(ctx, query, ctx_n, mask));
-            query      = norm1->forward(ctx, query);
-
-            auto h = fc1->forward(ctx, query);
-            h      = ggml_gelu(ctx->ggml_ctx, h);
-            h      = fc2->forward(ctx, h);
-            return ggml_add(ctx->ggml_ctx, query, h);
-        }
-    };
-
-    struct QueryFormerModel : public GGMLBlock {
-    protected:
-        QueryFormerConfig config;
-
-        void init_params(ggml_context* ctx,
-                         const String2TensorStorage& tensor_storage_map = {},
-                         const std::string prefix                       = "") override {
-            params["meta_queries"] = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, config.hidden_size, config.num_queries);
-        }
-
-    public:
-        QueryFormerModel() = default;
-        QueryFormerModel(const QueryFormerConfig& config)
-            : config(config) {
-            for (int i = 0; i < config.num_layers; i++) {
-                blocks["query_blocks." + std::to_string(i)] = std::make_shared<QueryFormerBlock>(config);
-            }
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                             ggml_tensor* inputs_embeds,
-                             ggml_tensor* mask = nullptr) {
-            // inputs_embeds: [N, n_token, hidden_size] -> [N, num_queries, hidden_size]
-            auto query = params["meta_queries"];
-            query      = ggml_reshape_3d(ctx->ggml_ctx, query, config.hidden_size, config.num_queries, 1);
-
-            for (int i = 0; i < config.num_layers; i++) {
-                auto block = std::dynamic_pointer_cast<QueryFormerBlock>(blocks["query_blocks." + std::to_string(i)]);
-                query      = block->forward(ctx, query, inputs_embeds, mask);
-            }
-            return query;
-        }
-    };
-
-    struct TextProjectionAttention : public GGMLBlock {
-    protected:
-        int64_t num_heads;
-        int64_t head_dim;
-
-    public:
-        TextProjectionAttention(const TextProjectionConfig& config)
-            : num_heads(config.num_heads), head_dim(config.hidden_size / config.num_heads) {
-            blocks["q_proj"]   = std::make_shared<Linear>(config.hidden_size, config.hidden_size, true);
-            blocks["k_proj"]   = std::make_shared<Linear>(config.hidden_size, config.hidden_size, true);
-            blocks["v_proj"]   = std::make_shared<Linear>(config.hidden_size, config.hidden_size, true);
-            blocks["out_proj"] = std::make_shared<Linear>(config.hidden_size, config.hidden_size, true);
-            blocks["q_norm"]   = std::make_shared<RMSNorm>(head_dim, config.norm_eps, false);
-            blocks["k_norm"]   = std::make_shared<RMSNorm>(head_dim, config.norm_eps, false);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            // x: [N, n_token, hidden_size]
-            ggml_context* gctx = ctx->ggml_ctx;
-            int64_t n_token    = x->ne[1];
-            int64_t N          = x->ne[2];
-
-            auto q_proj   = std::dynamic_pointer_cast<Linear>(blocks["q_proj"]);
-            auto k_proj   = std::dynamic_pointer_cast<Linear>(blocks["k_proj"]);
-            auto v_proj   = std::dynamic_pointer_cast<Linear>(blocks["v_proj"]);
-            auto out_proj = std::dynamic_pointer_cast<Linear>(blocks["out_proj"]);
-            auto q_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["q_norm"]);
-            auto k_norm   = std::dynamic_pointer_cast<RMSNorm>(blocks["k_norm"]);
-
-            auto q = q_proj->forward(ctx, x);
-            auto k = k_proj->forward(ctx, x);
-            auto v = v_proj->forward(ctx, x);
-
-            q = ggml_reshape_4d(gctx, q, head_dim, num_heads, n_token, N);
-            k = ggml_reshape_4d(gctx, k, head_dim, num_heads, n_token, N);
-            q = q_norm->forward(ctx, q);
-            k = k_norm->forward(ctx, k);
-            q = ggml_reshape_3d(gctx, q, head_dim * num_heads, n_token, N);
-            k = ggml_reshape_3d(gctx, k, head_dim * num_heads, n_token, N);
-
-            auto out = ggml_ext_attention_ext(ctx, q, k, v, num_heads);
-            return out_proj->forward(ctx, out);
-        }
-    };
-
-    struct TextProjectionBlock : public GGMLBlock {
-    public:
-        TextProjectionBlock(const TextProjectionConfig& config) {
-            blocks["self_attn"]   = std::make_shared<TextProjectionAttention>(config);
-            blocks["layer_norm1"] = std::make_shared<RMSNorm>(config.hidden_size, config.norm_eps, false);
-            blocks["layer_norm2"] = std::make_shared<RMSNorm>(config.hidden_size, config.norm_eps, false);
-            blocks["mlp.fc1"]     = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, true);
-            blocks["mlp.fc2"]     = std::make_shared<Linear>(config.intermediate_size, config.hidden_size, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            auto self_attn   = std::dynamic_pointer_cast<TextProjectionAttention>(blocks["self_attn"]);
-            auto layer_norm1 = std::dynamic_pointer_cast<RMSNorm>(blocks["layer_norm1"]);
-            auto layer_norm2 = std::dynamic_pointer_cast<RMSNorm>(blocks["layer_norm2"]);
-            auto fc1         = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc1"]);
-            auto fc2         = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc2"]);
-
-            x = ggml_add(ctx->ggml_ctx, x, self_attn->forward(ctx, layer_norm1->forward(ctx, x)));
-
-            auto h = fc1->forward(ctx, layer_norm2->forward(ctx, x));
-            h      = ggml_gelu(ctx->ggml_ctx, h);
-            h      = fc2->forward(ctx, h);
-            return ggml_add(ctx->ggml_ctx, x, h);
-        }
-    };
-
-    struct TextProjectionModel : public GGMLBlock {
-    protected:
-        TextProjectionConfig config;
-
-    public:
-        TextProjectionModel() = default;
-        TextProjectionModel(const TextProjectionConfig& config)
-            : config(config) {
-            for (int i = 0; i < config.num_layers; i++) {
-                blocks["layers." + std::to_string(i)] = std::make_shared<TextProjectionBlock>(config);
-            }
-            blocks["projector"] = std::make_shared<Linear>(config.hidden_size, config.projection_dim, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            // x: [N, n_token, hidden_size] -> [N, n_token, projection_dim]
-            for (int i = 0; i < config.num_layers; i++) {
-                auto block = std::dynamic_pointer_cast<TextProjectionBlock>(blocks["layers." + std::to_string(i)]);
-                x          = block->forward(ctx, x);
-            }
-            auto projector = std::dynamic_pointer_cast<Linear>(blocks["projector"]);
-            return projector->forward(ctx, x);
-        }
-    };
-
-    struct SigVQConfig {
-        int64_t image_size         = 2048;
-        int patch_size             = 16;
-        int64_t in_channels        = 3;
-        int64_t hidden_size        = 1536;
-        int64_t intermediate_size  = 6144;
-        int64_t num_layers         = 40;
-        int64_t num_heads          = 16;
-        int64_t codebook_size      = 16384;
-        int64_t codebook_embed_dim = 2048;
-        int64_t semantic_embed_dim = 4096;
-        float norm_eps             = 1e-6f;
-    };
-
-    struct SigVQAttention : public GGMLBlock {
-    protected:
-        int64_t num_heads;
-        int64_t head_dim;
-
-    public:
-        SigVQAttention(const SigVQConfig& config)
-            : num_heads(config.num_heads), head_dim(config.hidden_size / config.num_heads) {
-            blocks["qkv"]  = std::make_shared<Linear>(config.hidden_size, config.hidden_size * 3, true);
-            blocks["proj"] = std::make_shared<Linear>(config.hidden_size, config.hidden_size, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            // x: [N, n_token, hidden_size]
-            ggml_context* gctx = ctx->ggml_ctx;
-            auto qkv_proj      = std::dynamic_pointer_cast<Linear>(blocks["qkv"]);
-            auto out_proj      = std::dynamic_pointer_cast<Linear>(blocks["proj"]);
-
-            int64_t hidden_size = num_heads * head_dim;
-            auto qkv            = qkv_proj->forward(ctx, x);
-            auto q              = ggml_ext_slice(gctx, qkv, 0, 0, hidden_size);
-            auto k              = ggml_ext_slice(gctx, qkv, 0, hidden_size, hidden_size * 2);
-            auto v              = ggml_ext_slice(gctx, qkv, 0, hidden_size * 2, hidden_size * 3);
-
-            auto out = ggml_ext_attention_ext(ctx, q, k, v, num_heads);
-            return out_proj->forward(ctx, out);
-        }
-    };
-
-    struct SigVQBlock : public GGMLBlock {
-    public:
-        SigVQBlock(const SigVQConfig& config) {
-            blocks["norm1"]   = std::make_shared<LayerNorm>(config.hidden_size, config.norm_eps);
-            blocks["norm2"]   = std::make_shared<LayerNorm>(config.hidden_size, config.norm_eps);
-            blocks["attn"]    = std::make_shared<SigVQAttention>(config);
-            blocks["mlp.fc1"] = std::make_shared<Linear>(config.hidden_size, config.intermediate_size, true);
-            blocks["mlp.fc2"] = std::make_shared<Linear>(config.intermediate_size, config.hidden_size, true);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            auto norm1 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm1"]);
-            auto norm2 = std::dynamic_pointer_cast<LayerNorm>(blocks["norm2"]);
-            auto attn  = std::dynamic_pointer_cast<SigVQAttention>(blocks["attn"]);
-            auto fc1   = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc1"]);
-            auto fc2   = std::dynamic_pointer_cast<Linear>(blocks["mlp.fc2"]);
-
-            x      = ggml_add(ctx->ggml_ctx, x, attn->forward(ctx, norm1->forward(ctx, x)));
-            auto h = fc1->forward(ctx, norm2->forward(ctx, x));
-            h      = ggml_gelu_erf(ctx->ggml_ctx, h);
-            h      = fc2->forward(ctx, h);
-            return ggml_add(ctx->ggml_ctx, x, h);
-        }
-    };
-
-    struct SigVQModel : public GGMLBlock {
-    protected:
-        SigVQConfig config;
-
-    public:
-        SigVQModel() = default;
-        SigVQModel(const SigVQConfig& config)
-            : config(config) {
-            blocks["visual.patch_embed.proj"] = std::make_shared<Conv2d>(config.in_channels,
-                                                                         config.hidden_size,
-                                                                         std::make_pair(config.patch_size, config.patch_size),
-                                                                         std::make_pair(config.patch_size, config.patch_size));
-            for (int i = 0; i < config.num_layers; i++) {
-                blocks["visual.blocks." + std::to_string(i)] = std::make_shared<SigVQBlock>(config);
-            }
-            blocks["vqmodel.quant_conv"]         = std::make_shared<Conv2d>(config.hidden_size,
-                                                                    config.codebook_embed_dim,
-                                                                    std::make_pair(1, 1));
-            blocks["prior_projector.net.0.proj"] = std::make_shared<Linear>(config.semantic_embed_dim, config.semantic_embed_dim, true);
-            blocks["prior_projector.net.2"]      = std::make_shared<Linear>(config.semantic_embed_dim, config.semantic_embed_dim, true);
-        }
-
-        void init_params(ggml_context* ctx,
-                         const String2TensorStorage& tensor_storage_map = {},
-                         const std::string prefix                       = "") override {
-            params["visual.embeddings.position_embedding.weight"] =
-                ggml_new_tensor_2d(ctx, GGML_TYPE_F32, config.hidden_size, (config.image_size / config.patch_size) * (config.image_size / config.patch_size));
-            params["vqmodel.quantize.embedding.weight"] =
-                ggml_new_tensor_2d(ctx, GGML_TYPE_F32, config.codebook_embed_dim, config.codebook_size);
-            params["prior_token_embedding.weight"] =
-                ggml_new_tensor_2d(ctx, GGML_TYPE_F32, config.semantic_embed_dim, config.codebook_size);
-        }
-
-        // Bilinear-resamples the square position-embedding grid onto the image's patch grid.
-        // The reference uses grid_sample(align_corners=False, padding_mode="border"); the source
-        // coordinate for output index j is therefore (j + 0.5) * side / out - 0.5, clamped.
-        ggml_tensor* resample_pos_embed(GGMLRunnerContext* ctx,
-                                        ggml_tensor* pos_idx,
-                                        ggml_tensor* pos_weight) {
-            auto pos_embed = params["visual.embeddings.position_embedding.weight"];
-            auto gathered  = ggml_get_rows(ctx->ggml_ctx, pos_embed, pos_idx);
-            return ggml_mul(ctx->ggml_ctx, gathered, pos_weight);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx,
-                             ggml_tensor* pixel_values,
-                             const std::vector<ggml_tensor*>& pos_idx,
-                             const std::vector<ggml_tensor*>& pos_weight) {
-            // pixel_values: [N, in_channels, H, W] -> [N, grid_h * grid_w, semantic_embed_dim]
-            ggml_context* gctx = ctx->ggml_ctx;
-
-            auto patch_embed = std::dynamic_pointer_cast<Conv2d>(blocks["visual.patch_embed.proj"]);
-            auto quant_conv  = std::dynamic_pointer_cast<Conv2d>(blocks["vqmodel.quant_conv"]);
-            auto proj_0      = std::dynamic_pointer_cast<Linear>(blocks["prior_projector.net.0.proj"]);
-            auto proj_2      = std::dynamic_pointer_cast<Linear>(blocks["prior_projector.net.2"]);
-
-            auto x          = patch_embed->forward(ctx, pixel_values);  // [N, hidden_size, grid_h, grid_w]
-            int64_t grid_w  = x->ne[0];
-            int64_t grid_h  = x->ne[1];
-            int64_t n_token = grid_h * grid_w;
-            int64_t N       = x->ne[3];
-
-            x = ggml_reshape_3d(gctx, x, n_token, config.hidden_size, N);
-            x = ggml_cont(gctx, ggml_permute(gctx, x, 1, 0, 2, 3));  // [N, n_token, hidden_size]
-
-            ggml_tensor* pos = nullptr;
-            for (size_t i = 0; i < pos_idx.size(); i++) {
-                auto corner = resample_pos_embed(ctx, pos_idx[i], pos_weight[i]);
-                pos         = pos == nullptr ? corner : ggml_add(gctx, pos, corner);
-            }
-            x = ggml_add(gctx, x, ggml_reshape_3d(gctx, pos, config.hidden_size, n_token, N));
-
-            for (int i = 0; i < config.num_layers; i++) {
-                auto block = std::dynamic_pointer_cast<SigVQBlock>(blocks["visual.blocks." + std::to_string(i)]);
-                x          = block->forward(ctx, x);
-            }
-
-            // quant_conv is 1x1, so run it as a per-token projection rather than reshaping to 2-D.
-            x = ggml_cont(gctx, ggml_permute(gctx, x, 1, 0, 2, 3));  // [N, hidden_size, n_token]
-            x = ggml_reshape_4d(gctx, x, n_token, 1, config.hidden_size, N);
-            x = quant_conv->forward(ctx, x);  // [N, codebook_embed_dim, 1, n_token]
-            x = ggml_reshape_3d(gctx, x, n_token, config.codebook_embed_dim, N);
-            x = ggml_cont(gctx, ggml_permute(gctx, x, 1, 0, 2, 3));  // [N, n_token, codebook_embed_dim]
-
-            // Both sides are L2-normalized, so the nearest codebook entry by euclidean distance
-            // is the one with the largest dot product.
-            auto codebook  = ggml_l2_norm(gctx, params["vqmodel.quantize.embedding.weight"], 1e-12f);
-            auto normed    = ggml_l2_norm(gctx, x, 1e-12f);
-            auto logits    = ggml_mul_mat(gctx, codebook, normed);  // [N, n_token, codebook_size]
-            auto token_ids = ggml_argmax(gctx, ggml_reshape_2d(gctx, logits, config.codebook_size, n_token * N));
-
-            auto semantic = ggml_get_rows(gctx, params["prior_token_embedding.weight"], token_ids);
-            semantic      = ggml_reshape_3d(gctx, semantic, config.semantic_embed_dim, n_token, N);
-
-            auto h = proj_0->forward(ctx, semantic);
-            h      = ggml_silu(gctx, h);
-            return proj_2->forward(ctx, h);
-        }
-    };
-
-    struct QueryFormerRunner : public GGMLRunner {
-    public:
-        QueryFormerConfig config;
-        QueryFormerModel query_former;
-
-        QueryFormerRunner(ggml_backend_t backend,
-                          const String2TensorStorage& tensor_storage_map      = {},
-                          const std::string prefix                            = "",
-                          std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
-            : GGMLRunner(backend, weight_manager) {
-            query_former = QueryFormerModel(config);
-            query_former.init(params_ctx, tensor_storage_map, prefix);
-        }
-
-        std::string get_desc() override {
-            return "llada_image_queryformer";
-        }
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
-            query_former.get_param_tensors(tensors, prefix);
-        }
-
-        sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& inputs_embeds) {
-            auto get_graph = [&]() -> ggml_cgraph* {
-                ggml_cgraph* gf  = new_graph_custom(LLADA_IMAGE_TE_GRAPH_SIZE);
-                ggml_tensor* x   = make_input(inputs_embeds);
-                auto runner_ctx  = get_context();
-                ggml_tensor* out = query_former.forward(&runner_ctx, x);
-                ggml_build_forward_expand(gf, out);
-                return gf;
-            };
-            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, true),
-                                                   inputs_embeds.dim());
-        }
-    };
-
-    struct TextProjectionRunner : public GGMLRunner {
-    public:
-        TextProjectionConfig config;
-        TextProjectionModel text_projection;
-
-        TextProjectionRunner(ggml_backend_t backend,
-                             const String2TensorStorage& tensor_storage_map      = {},
-                             const std::string prefix                            = "",
-                             std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
-            : GGMLRunner(backend, weight_manager) {
-            text_projection = TextProjectionModel(config);
-            text_projection.init(params_ctx, tensor_storage_map, prefix);
-        }
-
-        std::string get_desc() override {
-            return "llada_image_text_projection";
-        }
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
-            text_projection.get_param_tensors(tensors, prefix);
-        }
-
-        sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& hidden_states) {
-            auto get_graph = [&]() -> ggml_cgraph* {
-                ggml_cgraph* gf  = new_graph_custom(LLADA_IMAGE_TE_GRAPH_SIZE);
-                ggml_tensor* x   = make_input(hidden_states);
-                auto runner_ctx  = get_context();
-                ggml_tensor* out = text_projection.forward(&runner_ctx, x);
-                ggml_build_forward_expand(gf, out);
-                return gf;
-            };
-            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, true),
-                                                   hidden_states.dim());
-        }
-    };
-
-    struct SigVQRunner : public GGMLRunner {
-    public:
-        SigVQConfig config;
-        SigVQModel sigvq;
-        std::array<std::vector<int32_t>, 4> pos_idx_data;
-        std::array<std::vector<float>, 4> pos_weight_data;
-
-        SigVQRunner(ggml_backend_t backend,
-                    const String2TensorStorage& tensor_storage_map      = {},
-                    const std::string prefix                            = "",
-                    std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
-            : GGMLRunner(backend, weight_manager) {
-            sigvq = SigVQModel(config);
-            sigvq.init(params_ctx, tensor_storage_map, prefix);
-        }
-
-        std::string get_desc() override {
-            return "llada_image_sigvq";
-        }
-
-        void get_param_tensors(std::map<std::string, ggml_tensor*>& tensors, const std::string& prefix) {
-            sigvq.get_param_tensors(tensors, prefix);
-        }
-
-        // Precomputes the four bilinear taps that resample the square position-embedding grid
-        // onto a grid_h x grid_w patch grid, matching grid_sample(align_corners=False,
-        // padding_mode="border").
-        void build_pos_embed_taps(int64_t grid_h, int64_t grid_w) {
-            const int64_t side = config.image_size / config.patch_size;
-            for (auto& v : pos_idx_data) {
-                v.clear();
-            }
-            for (auto& v : pos_weight_data) {
-                v.clear();
-            }
-
-            auto clamp_index = [side](int64_t v) {
-                return static_cast<int32_t>(std::min<int64_t>(std::max<int64_t>(v, 0), side - 1));
-            };
-
-            for (int64_t i = 0; i < grid_h; ++i) {
-                double src_h    = (static_cast<double>(i) + 0.5) * side / static_cast<double>(grid_h) - 0.5;
-                int64_t h_floor = static_cast<int64_t>(std::floor(src_h));
-                double dh       = src_h - static_cast<double>(h_floor);
-                for (int64_t j = 0; j < grid_w; ++j) {
-                    double src_w    = (static_cast<double>(j) + 0.5) * side / static_cast<double>(grid_w) - 0.5;
-                    int64_t w_floor = static_cast<int64_t>(std::floor(src_w));
-                    double dw       = src_w - static_cast<double>(w_floor);
-
-                    int32_t h0 = clamp_index(h_floor);
-                    int32_t h1 = clamp_index(h_floor + 1);
-                    int32_t w0 = clamp_index(w_floor);
-                    int32_t w1 = clamp_index(w_floor + 1);
-
-                    pos_idx_data[0].push_back(h0 * static_cast<int32_t>(side) + w0);
-                    pos_idx_data[1].push_back(h0 * static_cast<int32_t>(side) + w1);
-                    pos_idx_data[2].push_back(h1 * static_cast<int32_t>(side) + w0);
-                    pos_idx_data[3].push_back(h1 * static_cast<int32_t>(side) + w1);
-
-                    pos_weight_data[0].push_back(static_cast<float>((1.0 - dh) * (1.0 - dw)));
-                    pos_weight_data[1].push_back(static_cast<float>((1.0 - dh) * dw));
-                    pos_weight_data[2].push_back(static_cast<float>(dh * (1.0 - dw)));
-                    pos_weight_data[3].push_back(static_cast<float>(dh * dw));
-                }
-            }
-        }
-
-        sd::Tensor<float> compute(int n_threads, const sd::Tensor<float>& pixel_values) {
-            auto get_graph = [&]() -> ggml_cgraph* {
-                ggml_cgraph* gf = new_graph_custom(LLADA_IMAGE_TE_GRAPH_SIZE);
-                ggml_tensor* x  = make_input(pixel_values);
-
-                int64_t grid_h = x->ne[1] / config.patch_size;
-                int64_t grid_w = x->ne[0] / config.patch_size;
-                build_pos_embed_taps(grid_h, grid_w);
-
-                std::vector<ggml_tensor*> pos_idx;
-                std::vector<ggml_tensor*> pos_weight;
-                for (int i = 0; i < 4; i++) {
-                    auto idx = ggml_new_tensor_1d(compute_ctx, GGML_TYPE_I32, static_cast<int64_t>(pos_idx_data[i].size()));
-                    set_backend_tensor_data(idx, pos_idx_data[i].data());
-                    auto w = ggml_new_tensor_2d(compute_ctx, GGML_TYPE_F32, 1, static_cast<int64_t>(pos_weight_data[i].size()));
-                    set_backend_tensor_data(w, pos_weight_data[i].data());
-                    pos_idx.push_back(idx);
-                    pos_weight.push_back(w);
-                }
-
-                auto runner_ctx  = get_context();
-                ggml_tensor* out = sigvq.forward(&runner_ctx, x, pos_idx, pos_weight);
-                ggml_build_forward_expand(gf, out);
-                return gf;
-            };
-            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, true), 3);
-        }
-    };
-
-}  // namespace LLaDAImageTE
-
-#endif  // __SD_MODEL_TE_LLADA_IMAGE_TE_HPP__

src/model/te/llm.hpp

@@ -49,7 +49,6 @@ namespace LLM {
         GEMMA2_2B,
         GEMMA4_12B,
         GPT_OSS_20B,
-        LLADA2_MOE,
         ARCH_COUNT,
     };
 
@@ -63,7 +62,6 @@ namespace LLM {
         "gemma2_2b",
         "gemma4_12b",
         "gpt_oss_20b",
-        "llada2_moe",
     };
 
     enum class MLPActivation {
@@ -127,17 +125,6 @@ namespace LLM {
         std::vector<int> sliding_attention;
         int64_t num_experts         = 0;
         int64_t num_experts_per_tok = 0;
-        bool qkv_fused              = false;
-        bool bidirectional          = false;
-        float partial_rotary        = 1.f;
-
-        // DeepSeek-V3-style grouped-sigmoid MoE routing (LLaDA2)
-        int64_t moe_intermediate_size = 0;
-        int64_t num_shared_experts    = 0;
-        int64_t first_k_dense_replace = 0;
-        int64_t n_group               = 0;
-        int64_t topk_group            = 0;
-        float routed_scaling_factor   = 1.f;
         LLMVisionConfig vision;
         bool have_vision_weight = false;
         bool llama_cpp_style    = false;
@@ -225,31 +212,6 @@ namespace LLM {
                 config.intermediate_size       = 9216;
                 config.num_layers              = 26;
                 config.vocab_size              = 256000;
-            } else if (arch == LLMArch::LLADA2_MOE) {
-                config.head_dim                = 128;
-                config.num_heads               = 16;
-                config.num_kv_heads            = 4;
-                config.qkv_bias                = false;
-                config.attention_out_bias      = false;
-                config.qk_norm                 = true;
-                config.rms_norm_eps            = 1e-6f;
-                config.hidden_size             = 2048;
-                config.intermediate_size       = 5120;
-                config.num_layers              = 20;
-                config.vocab_size              = 173568;
-                config.max_position_embeddings = 16384;
-                config.rope_thetas             = {600000.f};
-                config.qkv_fused               = true;
-                config.bidirectional           = true;
-                config.partial_rotary          = 0.5f;
-                config.num_experts             = 256;
-                config.num_experts_per_tok     = 8;
-                config.moe_intermediate_size   = 512;
-                config.num_shared_experts      = 1;
-                config.first_k_dense_replace   = 1;
-                config.n_group                 = 8;
-                config.topk_group              = 4;
-                config.routed_scaling_factor   = 2.5f;
             } else if (arch == LLMArch::GPT_OSS_20B) {
                 config.head_dim                = 64;
                 config.num_heads               = 64;
@@ -457,195 +419,6 @@ namespace LLM {
         }
     };
 
-    // LLaDA2's MoE differs from GPT-OSS's in three ways that all change the result:
-    // routing scores are sigmoid (not softmax over the selected logits), expert selection is
-    // group-limited and uses a bias term that the returned weights do NOT include, and the
-    // experts carry no biases. Ref: LLaDA2MoeGate / LLaDA2MoeSparseMoeBlock in
-    // modeling_llada2uni_moe.py.
-    struct LLaDA2MoEMLP : public GGMLBlock {
-    protected:
-        int64_t hidden_size;
-        int64_t moe_intermediate_size;
-        int64_t num_experts;
-        int64_t num_experts_per_tok;
-        int64_t n_group;
-        int64_t topk_group;
-        float routed_scaling_factor;
-
-        void init_params(ggml_context* ctx,
-                         const String2TensorStorage& tensor_storage_map = {},
-                         std::string prefix                             = "") override {
-            GGMLBlock::init_params(ctx, tensor_storage_map, prefix);
-
-            auto supported_type = [](ggml_type wtype, int64_t in_features) {
-                if (in_features % ggml_blck_size(wtype) != 0) {
-                    return GGML_TYPE_F32;
-                }
-                return wtype;
-            };
-
-            // The reference runs the router in fp32; keep the weight in fp32 so the sigmoid
-            // scores and the group sums match.
-            params["gate.weight"]      = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, hidden_size, num_experts);
-            params["gate.expert_bias"] = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, num_experts);
-
-            ggml_type gate_type = supported_type(get_type(prefix + "experts.gate_proj.weight", tensor_storage_map, GGML_TYPE_F32), hidden_size);
-            ggml_type up_type   = supported_type(get_type(prefix + "experts.up_proj.weight", tensor_storage_map, GGML_TYPE_F32), hidden_size);
-            ggml_type down_type = supported_type(get_type(prefix + "experts.down_proj.weight", tensor_storage_map, GGML_TYPE_F32), moe_intermediate_size);
-
-            // HF ships the stacked experts as 3-D nn.Parameters, while the ComfyUI GGUF repack
-            // flattens the expert axis into ne[1]. Declare whichever the file holds - the two are
-            // bit-identical, and forward() reshapes to 3-D for ggml_mul_mat_id either way.
-            auto declare_experts = [&](const std::string& name, ggml_type type, int64_t in_dim, int64_t out_dim) {
-                auto storage = tensor_storage_map.find(prefix + name);
-                if (storage != tensor_storage_map.end() && storage->second.n_dims == 2) {
-                    GGML_ASSERT(storage->second.nelements() == in_dim * out_dim * num_experts);
-                    params[name] = ggml_new_tensor_2d(ctx, type, in_dim, out_dim * num_experts);
-                } else {
-                    params[name] = ggml_new_tensor_3d(ctx, type, in_dim, out_dim, num_experts);
-                }
-            };
-
-            declare_experts("experts.gate_proj.weight", gate_type, hidden_size, moe_intermediate_size);
-            declare_experts("experts.up_proj.weight", up_type, hidden_size, moe_intermediate_size);
-            declare_experts("experts.down_proj.weight", down_type, moe_intermediate_size, hidden_size);
-        }
-
-    public:
-        LLaDA2MoEMLP(const LLMConfig& config)
-            : hidden_size(config.hidden_size),
-              moe_intermediate_size(config.moe_intermediate_size),
-              num_experts(config.num_experts),
-              num_experts_per_tok(config.num_experts_per_tok),
-              n_group(config.n_group),
-              topk_group(config.topk_group),
-              routed_scaling_factor(config.routed_scaling_factor) {
-            if (config.num_shared_experts > 0) {
-                blocks["shared_experts"] = std::make_shared<MLP>(config.hidden_size,
-                                                                 config.moe_intermediate_size * config.num_shared_experts,
-                                                                 false,
-                                                                 config.mlp_activation);
-            }
-        }
-
-        // Reproduces group_limited_topk(): keep the topk_group groups with the highest
-        // "sum of the two best scores in the group", then take the global top-k among them.
-        ggml_tensor* group_limited_mask(GGMLRunnerContext* ctx,
-                                        ggml_tensor* routing_scores,
-                                        int64_t n_token_total) {
-            ggml_context* gctx      = ctx->ggml_ctx;
-            const int64_t per_group = num_experts / n_group;
-
-            // [experts_per_group, n_group * tokens] so top-2 runs per (group, token) row.
-            auto grouped     = ggml_reshape_2d(gctx, routing_scores, per_group, n_group * n_token_total);
-            auto best2_idx   = ggml_argsort_top_k(gctx, grouped, 2);  // [2, n_group * tokens]
-            auto grouped_val = ggml_reshape_3d(gctx, grouped, 1, per_group, n_group * n_token_total);
-            auto best2       = ggml_get_rows(gctx, grouped_val, best2_idx);  // [1, 2, n_group * tokens]
-            best2            = ggml_reshape_2d(gctx, best2, 2, n_group * n_token_total);
-            auto group_score = ggml_reshape_2d(gctx, ggml_sum_rows(gctx, best2), n_group, n_token_total);  // [n_group, tokens]
-
-            // Threshold = the topk_group-th largest group score, taken from the sorted top-k.
-            auto top_groups = ggml_argsort_top_k(gctx, group_score, (int)topk_group);  // [topk_group, tokens]
-            auto group_val  = ggml_reshape_3d(gctx, group_score, 1, n_group, n_token_total);
-            auto top_scores = ggml_get_rows(gctx, group_val, top_groups);  // [1, topk_group, tokens]
-            top_scores      = ggml_reshape_2d(gctx, top_scores, topk_group, n_token_total);
-            auto threshold  = ggml_view_2d(gctx,
-                                           top_scores,
-                                           1,
-                                           n_token_total,
-                                           top_scores->nb[1],
-                                           (topk_group - 1) * top_scores->nb[0]);  // [1, tokens]
-            threshold       = ggml_cont(gctx, threshold);
-
-            // keep = 1 - step(threshold - score). step(0) == 0, so the group sitting exactly on
-            // the threshold is kept without needing an epsilon.
-            auto diff = ggml_sub(gctx, ggml_repeat(gctx, threshold, group_score), group_score);
-            auto keep = ggml_scale_bias(gctx, ggml_step(gctx, diff), -1.f, 1.f);  // [n_group, tokens]
-
-            // 0 for kept groups, a large negative for dropped ones, broadcast over the group.
-            auto additive = ggml_scale_bias(gctx, keep, 1e30f, -1e30f);
-            additive      = ggml_reshape_3d(gctx, additive, 1, n_group, n_token_total);
-            auto expanded = ggml_repeat_4d(gctx, additive, per_group, n_group, n_token_total, 1);
-            return ggml_reshape_2d(gctx, expanded, num_experts, n_token_total);
-        }
-
-        ggml_tensor* expert_linear(GGMLRunnerContext* ctx,
-                                   const std::string& weight_name,
-                                   ggml_tensor* x,
-                                   ggml_tensor* selected_experts) {
-            ggml_tensor* w = params[weight_name];
-            if (w->ne[2] != num_experts) {
-                // Flattened layout: split the expert axis back out. ne[0] is untouched, so this
-                // stays valid for quantized types.
-                w = ggml_reshape_3d(ctx->ggml_ctx, w, w->ne[0], w->ne[1] / num_experts, num_experts);
-            }
-            return ggml_mul_mat_id(ctx->ggml_ctx, w, x, selected_experts);
-        }
-
-        ggml_tensor* forward(GGMLRunnerContext* ctx, ggml_tensor* x) {
-            // x: [N, n_token, hidden_size]
-            GGML_ASSERT(num_experts > 0 && num_experts_per_tok > 0);
-            GGML_ASSERT(n_group > 0 && topk_group > 0 && num_experts % n_group == 0);
-
-            ggml_context* gctx          = ctx->ggml_ctx;
-            const int64_t n_token       = x->ne[1];
-            const int64_t N             = x->ne[2];
-            const int64_t n_token_total = n_token * N;
-
-            auto identity = x;
-
-            auto logits = ggml_mul_mat(gctx, params["gate.weight"], x);
-            logits      = ggml_reshape_2d(gctx, logits, num_experts, n_token_total);
-            auto scores = ggml_sigmoid(gctx, logits);  // [num_experts, tokens]
-
-            // The bias steers selection only; the combine weights come from the unbiased scores.
-            auto routing = ggml_add(gctx, scores, params["gate.expert_bias"]);
-            routing      = ggml_add(gctx, routing, group_limited_mask(ctx, routing, n_token_total));
-
-            auto selected_experts = ggml_argsort_top_k(gctx, routing, (int)num_experts_per_tok);  // [top_k, tokens]
-            auto score_rows       = ggml_reshape_3d(gctx, scores, 1, num_experts, n_token_total);
-            auto weights          = ggml_get_rows(gctx, score_rows, selected_experts);  // [1, top_k, tokens]
-            weights               = ggml_reshape_2d(gctx, weights, num_experts_per_tok, n_token_total);
-
-            if (num_experts_per_tok > 1) {
-                auto denom = ggml_scale_bias(gctx, ggml_sum_rows(gctx, weights), 1.f, 1e-20f);  // [1, tokens]
-                weights    = ggml_div(gctx, weights, ggml_repeat(gctx, denom, weights));
-            }
-            weights = ggml_scale(gctx, weights, routed_scaling_factor);
-            weights = ggml_reshape_3d(gctx, weights, 1, num_experts_per_tok, n_token_total);
-
-            auto xf        = ggml_reshape_3d(gctx, x, hidden_size, 1, n_token_total);
-            auto gate      = expert_linear(ctx, "experts.gate_proj.weight", xf, selected_experts);
-            auto up        = expert_linear(ctx, "experts.up_proj.weight", xf, selected_experts);
-            auto activated = ggml_swiglu_split(gctx, gate, up);
-            auto experts   = expert_linear(ctx, "experts.down_proj.weight", activated, selected_experts);
-            experts        = ggml_mul(gctx, experts, weights);
-
-            ggml_tensor* out = nullptr;
-            for (int64_t i = 0; i < num_experts_per_tok; ++i) {
-                auto expert_out = ggml_view_2d(gctx,
-                                               experts,
-                                               hidden_size,
-                                               n_token_total,
-                                               experts->nb[2],
-                                               i * experts->nb[1]);
-                out             = out == nullptr ? expert_out : ggml_add(gctx, out, expert_out);
-            }
-            if (num_experts_per_tok == 1) {
-                out = ggml_cont(gctx, out);
-            }
-            out = ggml_reshape_3d(gctx, out, hidden_size, n_token, N);
-
-            auto shared_it = blocks.find("shared_experts");
-            if (shared_it != blocks.end()) {
-                auto shared_experts = std::dynamic_pointer_cast<MLP>(shared_it->second);
-                out                 = ggml_add(gctx, out, shared_experts->forward(ctx, identity));
-            }
-
-            return out;
-        }
-    };
-
     struct GPTOSSMLP : public GGMLBlock {
     protected:
         int64_t hidden_size;
@@ -832,31 +605,21 @@ namespace LLM {
             }
             txt_token_end = image_embeds[i].first;
 
-            // An embed can sit flush against the previous one or at the very start/end of the
-            // sequence, leaving no text tokens to splice around it.
-            if (txt_token_end > txt_token_start) {
-                auto txt_embed = ggml_ext_slice(ctx->ggml_ctx, raw_x, 1, txt_token_start, txt_token_end);
-                if (input_embed == nullptr) {
-                    input_embed = txt_embed;
-                } else {
-                    input_embed = ggml_concat(ctx->ggml_ctx, input_embed, txt_embed, 1);
-                }
+            auto txt_embed = ggml_ext_slice(ctx->ggml_ctx, raw_x, 1, txt_token_start, txt_token_end);
+            if (input_embed == nullptr) {
+                input_embed = txt_embed;
+            } else {
+                input_embed = ggml_concat(ctx->ggml_ctx, input_embed, txt_embed, 1);
             }
 
-            if (input_embed == nullptr) {
-                input_embed = image_embeds[i].second;
-            } else {
-                input_embed = ggml_concat(ctx->ggml_ctx, input_embed, image_embeds[i].second, 1);
-            }
+            input_embed = ggml_concat(ctx->ggml_ctx, input_embed, image_embeds[i].second, 1);
         }
 
         txt_token_start = image_embeds[image_embeds.size() - 1].first + image_embeds[image_embeds.size() - 1].second->ne[1];
         txt_token_end   = raw_x->ne[1];
 
-        if (txt_token_end > txt_token_start) {
-            auto final_txt_embed = ggml_ext_slice(ctx->ggml_ctx, raw_x, 1, txt_token_start, txt_token_end);
-            input_embed          = ggml_concat(ctx->ggml_ctx, input_embed, final_txt_embed, 1);
-        }
+        auto final_txt_embed = ggml_ext_slice(ctx->ggml_ctx, raw_x, 1, txt_token_start, txt_token_end);
+        input_embed          = ggml_concat(ctx->ggml_ctx, input_embed, final_txt_embed, 1);
         GGML_ASSERT(raw_x->ne[1] == input_embed->ne[1]);
         return input_embed;
     }
@@ -1359,7 +1122,6 @@ namespace LLM {
         bool k_eq_v;
         bool v_norm;
         bool unscaled_attention;
-        bool qkv_fused;
         float rms_norm_eps;
         int rope_pairs;
 
@@ -1385,20 +1147,12 @@ namespace LLM {
               k_eq_v(global_layer && config.global_k_eq_v),
               v_norm(config.v_norm),
               unscaled_attention(config.unscaled_attention),
-              qkv_fused(config.qkv_fused),
               rms_norm_eps(config.rms_norm_eps),
               rope_pairs(0) {
-            if (qkv_fused) {
-                // The checkpoint ships q, k and v as one tensor and the loader cannot split a
-                // source tensor, so keep it fused and slice it in forward().
-                GGML_ASSERT(!k_eq_v);
-                blocks["query_key_value"] = std::make_shared<Linear>(config.hidden_size, (num_heads + num_kv_heads * 2) * head_dim, config.qkv_bias);
-            } else {
-                blocks["q_proj"] = std::make_shared<Linear>(config.hidden_size, num_heads * head_dim, config.qkv_bias);
-                blocks["k_proj"] = std::make_shared<Linear>(config.hidden_size, num_kv_heads * head_dim, config.qkv_bias);
-                if (!k_eq_v) {
-                    blocks["v_proj"] = std::make_shared<Linear>(config.hidden_size, num_kv_heads * head_dim, config.qkv_bias);
-                }
+            blocks["q_proj"] = std::make_shared<Linear>(config.hidden_size, num_heads * head_dim, config.qkv_bias);
+            blocks["k_proj"] = std::make_shared<Linear>(config.hidden_size, num_kv_heads * head_dim, config.qkv_bias);
+            if (!k_eq_v) {
+                blocks["v_proj"] = std::make_shared<Linear>(config.hidden_size, num_kv_heads * head_dim, config.qkv_bias);
             }
             blocks["o_proj"] = std::make_shared<Linear>(num_heads * head_dim, config.hidden_size, config.attention_out_bias);
             if (config.qk_norm) {
@@ -1407,7 +1161,7 @@ namespace LLM {
             }
             // Proportional RoPE rotates only the leading `rope_pairs` dimension pairs of the head;
             // the rest are left unrotated through freq_factors (see rope_freq_factors()).
-            float partial = global_layer && config.global_partial_rotary != 1.f ? config.global_partial_rotary : config.partial_rotary;
+            float partial = global_layer ? config.global_partial_rotary : 1.f;
             rope_pairs    = static_cast<int>(partial * head_dim / 2.f);
         }
 
@@ -1432,28 +1186,14 @@ namespace LLM {
             // x: [N, n_token, hidden_size]
             int64_t n_token = x->ne[1];
             int64_t N       = x->ne[2];
+            auto q_proj     = std::dynamic_pointer_cast<Linear>(blocks["q_proj"]);
+            auto k_proj     = std::dynamic_pointer_cast<Linear>(blocks["k_proj"]);
+            auto v_proj     = k_eq_v ? nullptr : std::dynamic_pointer_cast<Linear>(blocks["v_proj"]);
             auto out_proj   = std::dynamic_pointer_cast<Linear>(blocks["o_proj"]);
 
-            ggml_tensor* q = nullptr;
-            ggml_tensor* k = nullptr;
-            ggml_tensor* v = nullptr;
-            if (qkv_fused) {
-                auto qkv_proj = std::dynamic_pointer_cast<Linear>(blocks["query_key_value"]);
-                auto qkv      = qkv_proj->forward(ctx, x);  // [N, n_token, (num_heads + num_kv_heads*2)*head_dim]
-                int64_t q_len = num_heads * head_dim;
-                int64_t k_len = num_kv_heads * head_dim;
-                q             = ggml_ext_slice(ctx->ggml_ctx, qkv, 0, 0, q_len);
-                k             = ggml_ext_slice(ctx->ggml_ctx, qkv, 0, q_len, q_len + k_len);
-                v             = ggml_ext_slice(ctx->ggml_ctx, qkv, 0, q_len + k_len, q_len + k_len * 2);
-            } else {
-                auto q_proj = std::dynamic_pointer_cast<Linear>(blocks["q_proj"]);
-                auto k_proj = std::dynamic_pointer_cast<Linear>(blocks["k_proj"]);
-                auto v_proj = k_eq_v ? nullptr : std::dynamic_pointer_cast<Linear>(blocks["v_proj"]);
-
-                q = q_proj->forward(ctx, x);               // [N, n_token, num_heads*head_dim]
-                k = k_proj->forward(ctx, x);               // [N, n_token, num_kv_heads*head_dim]
-                v = k_eq_v ? k : v_proj->forward(ctx, x);  // [N, n_token, num_kv_heads*head_dim]
-            }
+            auto q = q_proj->forward(ctx, x);               // [N, n_token, num_heads*head_dim]
+            auto k = k_proj->forward(ctx, x);               // [N, n_token, num_kv_heads*head_dim]
+            auto v = k_eq_v ? k : v_proj->forward(ctx, x);  // [N, n_token, num_kv_heads*head_dim]
 
             q = ggml_reshape_4d(ctx->ggml_ctx, q, head_dim, num_heads, n_token, N);     // [N, n_token, num_heads, head_dim]
             k = ggml_reshape_4d(ctx->ggml_ctx, k, head_dim, num_kv_heads, n_token, N);  // [N, n_token, num_kv_heads, head_dim]
@@ -1596,38 +1336,6 @@ namespace LLM {
                                   1.f,
                                   32.f,
                                   1.f);
-            } else if (arch == LLMArch::LLADA2_MOE) {
-                // LLaDA2 slices the head (query[..., :rotary_dim]) instead of zero-padding
-                // inv_freq like gemma does, so rotate_half pairs i with i + rotary_dim/2 and the
-                // frequencies use rotary_dim as the exponent denominator. Passing n_dims =
-                // rotary_dim reproduces both; freq_factors would give the wrong pairing.
-                int rotary_dim = rope_pairs * 2;
-                q              = ggml_rope_ext(ctx->ggml_ctx,
-                                               q,
-                                               input_pos,
-                                               nullptr,
-                                               rotary_dim,
-                                               GGML_ROPE_TYPE_NEOX,
-                                               static_cast<int>(max_position_embeddings),
-                                               rope_thetas[0],
-                                               1.f,
-                                               0.f,
-                                               1.f,
-                                               32.f,
-                                               1.f);
-                k              = ggml_rope_ext(ctx->ggml_ctx,
-                                               k,
-                                               input_pos,
-                                               nullptr,
-                                               rotary_dim,
-                                               GGML_ROPE_TYPE_NEOX,
-                                               static_cast<int>(max_position_embeddings),
-                                               rope_thetas[0],
-                                               1.f,
-                                               0.f,
-                                               1.f,
-                                               32.f,
-                                               1.f);
             } else if (arch == LLMArch::QWEN3_VL) {
                 int sections[4] = {24, 20, 20, 0};
                 q               = ggml_rope_multi(ctx->ggml_ctx, q, input_pos, nullptr, head_dim, sections, GGML_ROPE_TYPE_IMROPE, 262144, 5000000.f, 1.f, 0.f, 1.f, 32.f, 1.f);
@@ -1669,7 +1377,7 @@ namespace LLM {
                 x        = ggml_ext_cont(ctx->ggml_ctx, kqv);
                 x        = ggml_reshape_3d(ctx->ggml_ctx, x, head_dim * num_heads, n_token, N);
             } else {
-                x = ggml_ext_attention_ext(ctx, q, k, v, num_heads, attention_mask, true, ctx->flash_attn_enabled);  // [N, n_token, hidden_size]
+                x = ggml_ext_attention_ext(ctx, q, k, v, num_heads, attention_mask, true, false);  // [N, n_token, hidden_size]
             }
 
             x = out_proj->forward(ctx, x);  // [N, n_token, hidden_size]
@@ -1724,8 +1432,6 @@ namespace LLM {
             blocks["self_attn"] = std::make_shared<Attention>(config, sliding_attention == 0);
             if (config.arch == LLMArch::GPT_OSS_20B) {
                 blocks["mlp"] = std::make_shared<GPTOSSMLP>(config);
-            } else if (config.arch == LLMArch::LLADA2_MOE && layer_index >= config.first_k_dense_replace) {
-                blocks["mlp"] = std::make_shared<LLaDA2MoEMLP>(config);
             } else {
                 blocks["mlp"] = std::make_shared<MLP>(config.hidden_size,
                                                       config.intermediate_size,
@@ -1779,10 +1485,6 @@ namespace LLM {
             if (arch == LLMArch::GPT_OSS_20B) {
                 auto mlp = std::dynamic_pointer_cast<GPTOSSMLP>(blocks["mlp"]);
                 x        = mlp->forward(ctx, x);
-            } else if (auto moe_mlp = std::dynamic_pointer_cast<LLaDA2MoEMLP>(blocks["mlp"])) {
-                // LLaDA2 is dense for the first first_k_dense_replace layers and MoE afterwards,
-                // so the block type varies per layer rather than per arch.
-                x = moe_mlp->forward(ctx, x);
             } else {
                 auto mlp = std::dynamic_pointer_cast<MLP>(blocks["mlp"]);
                 x        = mlp->forward(ctx, x);
@@ -1948,11 +1650,6 @@ namespace LLM {
             return x;
         }
 
-        ggml_tensor* embed(GGMLRunnerContext* ctx, ggml_tensor* input_ids) {
-            auto model = std::dynamic_pointer_cast<TextModel>(blocks["model"]);
-            return model->embed(ctx, input_ids);
-        }
-
         std::shared_ptr<VisionModel> vision_model() {
             GGML_ASSERT(enable_vision);
             return std::dynamic_pointer_cast<VisionModel>(blocks["visual"]);
@@ -2293,8 +1990,7 @@ namespace LLM {
                 config.arch == LLMArch::GEMMA3_12B ||
                 config.arch == LLMArch::GEMMA4_12B ||
                 config.arch == LLMArch::GEMMA2_2B ||
-                config.arch == LLMArch::GPT_OSS_20B ||
-                config.arch == LLMArch::LLADA2_MOE) {
+                config.arch == LLMArch::GPT_OSS_20B) {
                 input_pos_vec.resize(n_tokens);
                 for (int i = 0; i < n_tokens; ++i) {
                     input_pos_vec[i] = i;
@@ -2346,9 +2042,8 @@ namespace LLM {
                 attention_mask_vec.resize(n_tokens * n_tokens);
                 for (int i0 = 0; i0 < n_tokens; i0++) {
                     for (int i1 = 0; i1 < n_tokens; i1++) {
-                        // Diffusion LLMs attend in both directions; only causal LMs get the triangle.
                         float value = 0.f;
-                        if (!config.bidirectional && i0 > i1) {
+                        if (i0 > i1) {
                             value = -INFINITY;
                         }
                         attention_mask_vec[i1 * n_tokens + i0] = value;
@@ -2420,22 +2115,6 @@ namespace LLM {
                                                    input_ids.dim() + 1);
         }
 
-        // LLaDA-Image's QueryFormer consumes the raw token embeddings before the backbone runs,
-        // so it needs the embedding lookup on its own.
-        sd::Tensor<float> compute_input_embeds(const int n_threads,
-                                               const sd::Tensor<int32_t>& input_ids) {
-            auto get_graph = [&]() -> ggml_cgraph* {
-                ggml_cgraph* gf  = new_graph_custom(LLM_GRAPH_SIZE);
-                ggml_tensor* ids = make_input(input_ids);
-                auto runner_ctx  = get_context();
-                ggml_tensor* out = model.embed(&runner_ctx, ids);
-                ggml_build_forward_expand(gf, out);
-                return gf;
-            };
-            return restore_trailing_singleton_dims(GGMLRunner::compute(get_graph, n_threads, true),
-                                                   input_ids.dim() + 1);
-        }
-
         int64_t get_num_image_tokens(int64_t t, int64_t h, int64_t w) {
             int64_t grid_t     = 1;
             int64_t grid_h     = h / config.vision.patch_size;
@@ -2691,13 +2370,11 @@ namespace LLM {
                 pad_id = 199999;
             } else if (arch == LLMArch::GEMMA2_2B) {
                 pad_id = 0;
-            } else if (arch == LLMArch::LLADA2_MOE) {
-                pad_id = 156892;
             }
             tokenizer = tokenizers.create(TokenizerConfig::MAIN, model.config.vocab_size, pad_id);
             if (!tokenizer) {
-                if (arch == LLMArch::GPT_OSS_20B || arch == LLMArch::GEMMA2_2B || arch == LLMArch::LLADA2_MOE) {
-                    throw std::runtime_error("GPT-OSS, Gemma 2 and LLaDA2 require an external tokenizer.json in the main tokenizer slot");
+                if (arch == LLMArch::GPT_OSS_20B || arch == LLMArch::GEMMA2_2B) {
+                    throw std::runtime_error("GPT-OSS and Gemma 2 require an external tokenizer.json in the main tokenizer slot");
                 }
                 if (arch == LLMArch::MISTRAL_SMALL_3_2 || arch == LLMArch::MINISTRAL_3_3B) {
                     tokenizer = std::make_shared<MistralTokenizer>();

src/model/vae/auto_encoder_kl.hpp

@@ -544,7 +544,7 @@ public:
                        const String2TensorStorage& tensor_storage_map = {},
                        const std::string& prefix                      = "")
         : version(version), decode_only(decode_only), use_video_decoder(use_video_decoder) {
-        if (sd_version_is_dit(version) && version != VERSION_PIXART) {
+        if (sd_version_is_dit(version)) {
             if (sd_version_uses_flux2_vae(version)) {
                 dd_config.z_channels = 32;
                 embed_dim            = 32;
@@ -678,7 +678,7 @@ struct AutoEncoderKL : public VAE {
         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) || sd_version_is_pixart(version)) {
+        } else if (sd_version_is_sdxl(version)) {
             scale_factor = 0.13025f;
             shift_factor = 0.f;
         } else if (sd_version_is_sd3(version)) {

src/model/vae/ltx_vae.hpp

@@ -1300,7 +1300,7 @@ struct LTXVideoVAE : public VAE {
             feat_map[feat_idx] = get_cache_tensor_by_name(temporal_feat_cache_name(feat_idx));
         }
 
-        auto runner_ctx  = get_context(gf);
+        auto runner_ctx  = get_context();
         int feat_count   = 0;
         ggml_tensor* out = vae.decode_tiled_chunk(&runner_ctx,
                                                   z,
@@ -1313,7 +1313,8 @@ struct LTXVideoVAE : public VAE {
         for (int feat_idx = 0; feat_idx < feat_count && feat_idx < static_cast<int>(feat_map.size()); ++feat_idx) {
             ggml_tensor* feat_cache = feat_map[static_cast<size_t>(feat_idx)];
             if (feat_cache != nullptr) {
-                runner_ctx.persist_cache_tensor(temporal_feat_cache_name(static_cast<size_t>(feat_idx)), feat_cache);
+                cache(temporal_feat_cache_name(static_cast<size_t>(feat_idx)), feat_cache);
+                ggml_build_forward_expand(gf, feat_cache);
             }
         }
 

src/model/vae/minimax_h3_vae.hpp

@@ -556,18 +556,12 @@ namespace MiniMaxH3VAE {
                                    tensor.shape()[3]});
         }
 
-        sd_tiling_params_t resolve_tiling_params(sd_tiling_params_t params) const override {
-            if (!params.enabled) {
-                params.target_overlap = 0.25f;
-            }
-            if (params.tile_size_w == 0 && params.rel_size_w == 0.f) {
-                params.tile_size_w = 256;
-            }
-            if (params.tile_size_h == 0 && params.rel_size_h == 0.f) {
-                params.tile_size_h = 256;
-            }
+        static sd_tiling_params_t h3_tiling(sd_tiling_params_t params) {
             params.enabled         = true;
             params.temporal_tiling = false;
+            params.tile_size_x     = 16;
+            params.tile_size_y     = 16;
+            params.target_overlap  = 0.25f;
             return params;
         }
 
@@ -611,7 +605,7 @@ namespace MiniMaxH3VAE {
                                  bool circular_x = false,
                                  bool circular_y = false) override {
             auto input  = ensure_video_shape(x);
-            auto tiling = resolve_tiling_params(tiling_params);
+            auto tiling = h3_tiling(tiling_params);
             if (input.shape()[2] == 1) {
                 auto encoded = VAE::encode(n_threads, input, tiling, circular_x, circular_y);
                 if (!encoded.empty() && encoded.shape()[2] > 1) {
@@ -652,7 +646,7 @@ namespace MiniMaxH3VAE {
                                  bool circular_y   = false,
                                  bool silent       = false) override {
             auto input  = ensure_video_shape(x);
-            auto tiling = resolve_tiling_params(tiling_params);
+            auto tiling = h3_tiling(tiling_params);
             if (input.shape()[2] == 1) {
                 auto decoded = VAE::decode(n_threads,
                                            input,

src/model/vae/tae.hpp

@@ -701,7 +701,7 @@ public:
         bool use_midblock_gn = false;
         taef2                = sd_version_uses_flux2_vae(version);
 
-        if (sd_version_is_dit(version) && !sd_version_is_pixart(version)) {
+        if (sd_version_is_dit(version)) {
             z_channels = 16;
         }
         if (taef2) {

src/model/vae/vae.hpp

@@ -1,9 +1,6 @@
 #ifndef __SD_MODEL_VAE_VAE_HPP__
 #define __SD_MODEL_VAE_VAE_HPP__
 
-#include <cmath>
-#include <limits>
-
 #include "core/tensor_ggml.hpp"
 #include "model/common/block.hpp"
 #include "model/vae/vae_tiling.hpp"
@@ -120,8 +117,8 @@ protected:
                                     int output_width,
                                     int output_height,
                                     int scale,
-                                    int p_tile_size_w,
-                                    int p_tile_size_h,
+                                    int p_tile_size_x,
+                                    int p_tile_size_y,
                                     float tile_overlap_factor,
                                     bool circular_x,
                                     bool circular_y,
@@ -141,28 +138,17 @@ protected:
             }
             return output_tile;
         };
-        const bool original_circular_x = circular_x_enabled;
-        const bool original_circular_y = circular_y_enabled;
-        const int64_t latent_width     = decode_graph ? input.shape()[0] : output_width;
-        const int64_t latent_height    = decode_graph ? input.shape()[1] : output_height;
-        circular_x                     = circular_x || original_circular_x;
-        circular_y                     = circular_y || original_circular_y;
-        // Full-width axes wrap in convolutions; split axes wrap between tiles.
-        set_circular_axes(circular_x && p_tile_size_w >= latent_width,
-                          circular_y && p_tile_size_h >= latent_height);
-        auto output = ::process_tiles_2d(input,
-                                         output_width,
-                                         output_height,
-                                         scale,
-                                         p_tile_size_w,
-                                         p_tile_size_h,
-                                         tile_overlap_factor,
-                                         circular_x && p_tile_size_w < latent_width,
-                                         circular_y && p_tile_size_h < latent_height,
-                                         on_processing,
-                                         silent);
-        set_circular_axes(original_circular_x, original_circular_y);
-        return output;
+        return ::process_tiles_2d(input,
+                                  output_width,
+                                  output_height,
+                                  scale,
+                                  p_tile_size_x,
+                                  p_tile_size_y,
+                                  tile_overlap_factor,
+                                  circular_x,
+                                  circular_y,
+                                  on_processing,
+                                  silent);
     }
 
 public:
@@ -176,7 +162,7 @@ public:
         int scale_factor = 8;
         if (version == VERSION_LTXAV) {
             scale_factor = 32;
-        } else if (version == VERSION_WAN2_2_TI2V || version == VERSION_QWEN_IMAGE_2_1 || sd_version_is_hunyuan_video(version) || sd_version_is_mage_flow(version) || sd_version_is_minimax_h3(version)) {
+        } else if (version == VERSION_WAN2_2_TI2V || sd_version_is_hunyuan_video(version) || sd_version_is_mage_flow(version) || sd_version_is_minimax_h3(version)) {
             scale_factor = 16;
         } else if (sd_version_uses_flux2_vae(version)) {
             scale_factor = 16;
@@ -192,48 +178,33 @@ public:
         return supports_temporal_tiling(VAETemporalDirection::DECODE);
     }
 
-    virtual sd_tiling_params_t resolve_tiling_params(sd_tiling_params_t params) const {
-        return params;
-    }
-
-    bool get_tile_sizes(int& tile_size_w,
-                        int& tile_size_h,
+    void get_tile_sizes(int& tile_size_x,
+                        int& tile_size_y,
                         float& tile_overlap,
                         const sd_tiling_params_t& params,
-                        int64_t latent_w,
-                        int64_t latent_h) {
-        const auto tiling = resolve_tiling_params(params);
-        if (latent_w <= 0 || latent_h <= 0 ||
-            latent_w > std::numeric_limits<int>::max() || latent_h > std::numeric_limits<int>::max() ||
-            !std::isfinite(tiling.target_overlap)) {
-            LOG_ERROR("invalid VAE tiling dimensions or overlap");
-            return false;
-        }
-        const int scale_factor = get_scale_factor();
-        tile_overlap           = std::max(std::min(tiling.target_overlap, 0.5f), 0.0f);
-        auto get_tile_size     = [&](int requested_size, double factor, int64_t latent_size, int& tile_size) {
-            if (requested_size < 0 || !std::isfinite(factor) || factor < 0.0) {
-                LOG_ERROR("VAE tile sizes and relative sizes must be finite and non-negative");
-                return false;
+                        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;
             }
-            const int min_tile_dimension = std::min(4, static_cast<int>(latent_size));
-            double size                  = (requested_size > 0 ? requested_size : 256) / scale_factor;
-            if (factor > 0.0) {
-                if (factor > 1.0) {
-                    factor = 1.0 / (factor * (1.0 - tile_overlap) + tile_overlap);
-                }
-                size = std::floor(static_cast<double>(latent_size) * factor);
-            }
-            if (size < min_tile_dimension && (requested_size > 0 || factor > 0.0)) {
-                LOG_ERROR("VAE tile size must be at least %d image pixels on this axis", min_tile_dimension * scale_factor);
-                return false;
-            }
-            tile_size = static_cast<int>(std::min(static_cast<double>(latent_size), std::max<double>(min_tile_dimension, size)));
-            return true;
+            tile_size = static_cast<int>(tile_size * encoding_factor);
+            return std::max(std::min(tile_size, static_cast<int>(latent_size)), min_tile_dimension);
         };
 
-        return get_tile_size(tiling.tile_size_w, tiling.rel_size_w, latent_w, tile_size_w) &&
-               get_tile_size(tiling.tile_size_h, tiling.rel_size_h, latent_h, tile_size_h);
+        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);
     }
 
     virtual sd::Tensor<float> encode(int n_threads,
@@ -242,7 +213,6 @@ public:
                                      bool circular_x = false,
                                      bool circular_y = false) {
         int64_t t0              = ggml_time_ms();
-        tiling_params           = resolve_tiling_params(tiling_params);
         sd::Tensor<float> input = x;
         sd::Tensor<float> output;
         if (scale_input) {
@@ -254,19 +224,21 @@ public:
             int64_t W              = input.shape()[0] / scale_factor;
             int64_t H              = input.shape()[1] / scale_factor;
             float tile_overlap;
-            int tile_size_w, tile_size_h;
-            if (!get_tile_sizes(tile_size_w, tile_size_h, tile_overlap, tiling_params, W, H)) {
-                return {};
-            }
-            LOG_VERBOSE("VAE encode tile size: %dx%d pixels (%dx%d latent)",
-                        tile_size_w * scale_factor, tile_size_h * scale_factor, tile_size_w, tile_size_h);
+            int tile_size_x, tile_size_y;
+            // Image VAE encode is more sensitive to tile boundary context than decode.
+            // Keep the smaller legacy factor for video VAEs, but default image encode
+            // tiles to 64 latent pixels so a 512px SD image is encoded as one tile.
+            const float encode_tile_factor = sd_version_is_minimax_h3(version) ? 1.f : (sd_version_is_wan(version) || sd_version_is_hunyuan_video(version) || sd_version_is_ltxav(version)) ? 1.30539f
+                                                                                                                                                                                            : 2.0f;
+            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, W, H, encode_tile_factor);
+            LOG_VERBOSE("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
             output = tiled_compute(input,
                                    n_threads,
                                    static_cast<int>(W),
                                    static_cast<int>(H),
                                    scale_factor,
-                                   tile_size_w,
-                                   tile_size_h,
+                                   tile_size_x,
+                                   tile_size_y,
                                    tile_overlap,
                                    circular_x,
                                    circular_y,
@@ -299,7 +271,6 @@ public:
                                      bool circular_y   = false,
                                      bool silent       = false) {
         int64_t t0              = ggml_time_ms();
-        tiling_params           = resolve_tiling_params(tiling_params);
         sd::Tensor<float> input = x;
         sd::Tensor<float> output;
 
@@ -308,13 +279,10 @@ public:
             int64_t W              = input.shape()[0] * scale_factor;
             int64_t H              = input.shape()[1] * scale_factor;
             float tile_overlap;
-            int tile_size_w, tile_size_h;
-            if (!get_tile_sizes(tile_size_w, tile_size_h, tile_overlap, tiling_params, input.shape()[0], input.shape()[1])) {
-                return {};
-            }
+            int tile_size_x, tile_size_y;
+            get_tile_sizes(tile_size_x, tile_size_y, tile_overlap, tiling_params, input.shape()[0], input.shape()[1]);
             if (!silent) {
-                LOG_VERBOSE("VAE decode tile size: %dx%d pixels (%dx%d latent)",
-                            tile_size_w * scale_factor, tile_size_h * scale_factor, tile_size_w, tile_size_h);
+                LOG_VERBOSE("VAE Tile size: %dx%d", tile_size_x, tile_size_y);
             }
             output = tiled_compute(
                 input,
@@ -322,8 +290,8 @@ public:
                 static_cast<int>(W),
                 static_cast<int>(H),
                 scale_factor,
-                tile_size_w,
-                tile_size_h,
+                tile_size_x,
+                tile_size_y,
                 tile_overlap,
                 circular_x,
                 circular_y,

src/model/vae/wan_vae.hpp

@@ -24,16 +24,8 @@ namespace WAN {
         std::tuple<int, int, int> padding;
         std::tuple<int, int, int> dilation;
         bool bias;
-        float scale = 1.f;
 
         void init_params(ggml_context* ctx, const String2TensorStorage& tensor_storage_map = {}, const std::string prefix = "") override {
-            auto weight = tensor_storage_map.find(prefix + "weight");
-            if (weight != tensor_storage_map.end() && weight->second.ne[2] == 1 &&
-                weight->second.ne[3] == in_channels * out_channels) {
-                // Image VAE exports may retain Conv3d weights with a singleton temporal kernel.
-                std::get<0>(kernel_size) = 1;
-                std::get<0>(padding)     = 0;
-            }
             params["weight"] = ggml_new_tensor_4d(ctx,
                                                   GGML_TYPE_F16,
                                                   std::get<2>(kernel_size),
@@ -61,10 +53,6 @@ namespace WAN {
               dilation(std::move(dilation)),
               bias(bias) {}
 
-        void set_scale(float scale_value) {
-            scale = scale_value;
-        }
-
         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]
@@ -87,25 +75,10 @@ namespace WAN {
             }
 
             x = ggml_ext_pad_ext(ctx->ggml_ctx, ctx->backend, x, lp0, rp0, lp1, rp1, lp2, rp2, 0, 0, ctx->circular_x_enabled, ctx->circular_y_enabled);
-            if (w->ne[2] == 1 && x->ne[2] == 1 && x->ne[3] == in_channels) {
-                // One frame through a one-frame-deep kernel is a 2D conv; backends without
-                // im2col_3d (Metal) otherwise fall back to a much slower direct conv_3d.
-                if (!ggml_is_contiguous(x)) {
-                    x = ggml_cont(ctx->ggml_ctx, x);
-                }
-                ggml_tensor* x2 = ggml_reshape_4d(ctx->ggml_ctx, x, x->ne[0], x->ne[1], in_channels, 1);
-                ggml_tensor* w2 = ggml_reshape_4d(ctx->ggml_ctx, w, w->ne[0], w->ne[1], in_channels, out_channels);
-                x2              = ggml_ext_conv_2d(ctx->ggml_ctx, x2, w2, b,
-                                                   std::get<2>(stride), std::get<1>(stride), 0, 0,
-                                                   std::get<2>(dilation), std::get<1>(dilation),
-                                                   ctx->conv2d_direct_enabled, false, false, scale);
-                return ggml_reshape_4d(ctx->ggml_ctx, x2, x2->ne[0], x2->ne[1], 1, out_channels);
-            }
             return ggml_ext_conv_3d(ctx->ggml_ctx, ctx->backend, x, w, b, in_channels,
                                     std::get<2>(stride), std::get<1>(stride), std::get<0>(stride),
                                     0, 0, 0,
-                                    std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation),
-                                    false, ctx->conv3d_direct_enabled, scale);
+                                    std::get<2>(dilation), std::get<1>(dilation), std::get<0>(dilation));
         }
     };
 
@@ -166,7 +139,7 @@ namespace WAN {
         std::string mode;
 
     public:
-        Resample(int64_t dim, const std::string& mode, bool wan2_2 = false, bool is_2D = false)
+        Resample(int64_t dim, const std::string& mode, bool wan2_2 = false)
             : dim(dim), mode(mode) {
             if (mode == "upsample2d") {
                 if (wan2_2) {
@@ -180,20 +153,12 @@ namespace WAN {
                 } else {
                     blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim / 2, {3, 3}, {1, 1}, {1, 1}));
                 }
-                if (is_2D) {
-                    blocks["time_conv"] = std::make_shared<Conv2dBut3d>(dim, dim * 2, std::pair<int, int>{1, 1});
-                } else {
-                    blocks["time_conv"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(dim, dim * 2, {3, 1, 1}, {1, 1, 1}, {1, 0, 0}));
-                }
+                blocks["time_conv"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(dim, dim * 2, {3, 1, 1}, {1, 1, 1}, {1, 0, 0}));
             } else if (mode == "downsample2d") {
                 blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {3, 3}, {2, 2}));
             } else if (mode == "downsample3d") {
                 blocks["resample.1"] = std::shared_ptr<GGMLBlock>(new Conv2d(dim, dim, {3, 3}, {2, 2}));
-                if (is_2D) {
-                    blocks["time_conv"] = std::make_shared<Conv2dBut3d>(dim, dim, std::pair<int, int>{1, 1});
-                } else {
-                    blocks["time_conv"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(dim, dim, {3, 1, 1}, {2, 1, 1}, {0, 0, 0}));
-                }
+                blocks["time_conv"]  = std::shared_ptr<GGMLBlock>(new CausalConv3d(dim, dim, {3, 1, 1}, {2, 1, 1}, {0, 0, 0}));
             } else if (mode == "none") {
                 // nn.Identity()
             } else {
@@ -503,7 +468,7 @@ namespace WAN {
             }
             if (down_flag) {
                 std::string mode                           = temperal_downsample ? "downsample3d" : "downsample2d";
-                blocks["downsamples." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new Resample(out_dim, mode, true, is_2D));
+                blocks["downsamples." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new Resample(out_dim, mode, true));
                 i++;
             }
         }
@@ -566,7 +531,7 @@ namespace WAN {
             }
             if (up_flag) {
                 std::string mode                         = temperal_upsample ? "upsample3d" : "upsample2d";
-                blocks["upsamples." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new Resample(out_dim, mode, true, is_2D));
+                blocks["upsamples." + std::to_string(i)] = std::shared_ptr<GGMLBlock>(new Resample(out_dim, mode, true));
                 i++;
             }
         }
@@ -1088,24 +1053,9 @@ namespace WAN {
                 input_channels = 4;
             }
 
-            if (version == VERSION_QWEN_IMAGE_2_1) {
-                wan2_2         = true;
-                dec_dim        = 144;
-                z_dim          = 64;
-                input_channels = 4;
-                dim_mult       = {1, 2, 4, 8, 8};
-            }
-
             if (is_2D) {
-                temperal_upsample.assign(dim_mult.size() - 1, false);
-                temperal_downsample.assign(dim_mult.size() - 1, false);
-            }
-            if (version == VERSION_QWEN_IMAGE_2_1) {
-                // Temporal shortcut factors still affect single-frame channel grouping.
-                temperal_upsample   = {true, true, true, false};
-                temperal_downsample = {false, true, true, true};
-                _conv_num           = 2 * (2 + static_cast<int>(dim_mult.size()) * (num_res_blocks + 1)) + 3 + (is_2D ? 0 : 2);
-                _enc_conv_num       = 2 * (2 + static_cast<int>(dim_mult.size()) * num_res_blocks) + 3 + (is_2D ? 0 : 2);
+                temperal_upsample   = {false, false, false};
+                temperal_downsample = {false, false, false};
             }
 
             if (!decode_only) {
@@ -1122,19 +1072,6 @@ namespace WAN {
             } else {
                 blocks["conv2"] = std::shared_ptr<GGMLBlock>(new CausalConv3d(z_dim, z_dim, {1, 1, 1}));
             }
-            if (version == VERSION_QWEN_IMAGE_2_1) {
-                // Keep large VAE activations within the FP16 convolution range.
-                const float conv_scale = 1.f / 128.f;
-                std::vector<GGMLBlock*> all_blocks;
-                get_all_blocks(all_blocks);
-                for (auto block : all_blocks) {
-                    if (auto conv = dynamic_cast<Conv2d*>(block)) {
-                        conv->set_scale(conv_scale);
-                    } else if (auto conv = dynamic_cast<CausalConv3d*>(block)) {
-                        conv->set_scale(conv_scale);
-                    }
-                }
-            }
         }
 
         static ggml_tensor* patchify(ggml_context* ctx,
@@ -1333,9 +1270,18 @@ namespace WAN {
                      SDVersion version                                   = VERSION_WAN2,
                      std::shared_ptr<RunnerWeightManager> weight_manager = nullptr)
             : VAE(version, backend, prefix, weight_manager), decode_only(decode_only) {
-            const auto conv_in = tensor_storage_map.find((prefix.empty() ? "" : prefix + ".") + "decoder.conv1.weight");
-            const bool is_2D   = conv_in != tensor_storage_map.end() && conv_in->second.ne[2] > 3;
-            LOG_VERBOSE("Wan VAE convolution type: %s", is_2D ? "2D" : "3D");
+            bool is_2D = false;
+            for (const auto& [name, tensor_storage] : tensor_storage_map) {
+                if (ends_with(name, "decoder.conv1.weight")) {
+                    if (tensor_storage.ne[2] > 3) {
+                        is_2D = true;
+                    }
+                    break;
+                }
+            }
+            if (is_2D) {
+                LOG_VERBOSE("USING 2D VAE");
+            }
             ae = WanVAE(decode_only, version, is_2D);
             ae.init(params_ctx, tensor_storage_map, prefix);
         }
@@ -1395,28 +1341,6 @@ namespace WAN {
                 std_tensor.reshape_(stats_shape);
                 return {std::move(mean_tensor), std::move(std_tensor)};
             }
-            if (version == VERSION_QWEN_IMAGE_2_1 && latents.shape()[channel_dim] == 64) {
-                stats_shape[static_cast<size_t>(channel_dim)] = 64;
-                auto mean_tensor                              = sd::Tensor<float>::from_vector({0.5126f, 0.7721f, -0.0631f, 1.3506f, -0.7855f, -2.1025f, -0.3458f, 1.3722f,
-                                                                                                1.8873f, -1.7177f, -0.6510f, 0.2732f, 0.7562f, -0.6163f, -1.0277f, 3.8363f,
-                                                                                                2.0210f, 0.0472f, 0.9320f, 2.0087f, 2.4954f, -0.1391f, -1.4249f, 1.8464f,
-                                                                                                -0.5236f, 1.2826f, 3.7046f, -1.3035f, 2.7286f, -1.4518f, -1.9036f, -1.9955f,
-                                                                                                -0.0342f, -1.0265f, -0.7636f, 3.0555f, 0.0746f, -3.0751f, -0.1076f, 1.7376f,
-                                                                                                -1.0914f, -1.9435f, -0.2784f, -1.3680f, 0.4809f, -0.4433f, 0.3764f, 0.5729f,
-                                                                                                -2.0595f, 1.0960f, -1.3260f, -2.0211f, -5.0179f, 0.5275f, 4.0162f, 1.8505f,
-                                                                                                0.3026f, 1.9373f, 1.4937f, 0.2632f, 0.5547f, -1.7121f, -0.1562f, 0.0304f});
-                auto std_tensor                               = sd::Tensor<float>::from_vector({3.2001f, 3.2936f, 3.4321f, 3.0091f, 3.1061f, 4.0379f, 4.0705f, 3.7910f,
-                                                                                                3.0785f, 3.6500f, 3.9308f, 3.0904f, 2.8778f, 3.7675f, 3.7320f, 5.0756f,
-                                                                                                3.2864f, 4.0397f, 3.1317f, 4.0443f, 2.9249f, 3.9454f, 3.0988f, 4.2489f,
-                                                                                                3.4896f, 3.8513f, 3.9323f, 3.4719f, 3.7498f, 4.2830f, 3.5694f, 4.2467f,
-                                                                                                3.9037f, 3.2947f, 5.0770f, 3.5075f, 3.2700f, 3.4767f, 2.8063f, 5.1125f,
-                                                                                                3.5327f, 4.7833f, 3.1286f, 4.1819f, 3.8527f, 3.8312f, 3.5605f, 4.3875f,
-                                                                                                3.9624f, 4.0168f, 3.5643f, 4.0550f, 5.5614f, 4.2963f, 4.4080f, 3.4959f,
-                                                                                                3.8747f, 3.7608f, 3.5735f, 3.1490f, 3.7662f, 3.6746f, 3.4563f, 3.8161f});
-                mean_tensor.reshape_(stats_shape);
-                std_tensor.reshape_(stats_shape);
-                return {std::move(mean_tensor), std::move(std_tensor)};
-            }
             GGML_ABORT("unexpected latent channel dimension %lld for version %d",
                        (long long)latents.shape()[channel_dim],
                        version);
@@ -1461,14 +1385,15 @@ namespace WAN {
 
             ggml_tensor* z = make_input(z_tensor);
 
-            auto runner_ctx = get_context(gf);
+            auto runner_ctx = get_context();
 
             ggml_tensor* out = ae.decode_tiled_chunk(&runner_ctx, z, chunk_idx);
 
             for (size_t feat_idx = 0; feat_idx < ae._feat_map.size(); feat_idx++) {
                 ggml_tensor* feat_cache = ae._feat_map[feat_idx];
                 if (feat_cache != nullptr) {
-                    runner_ctx.persist_cache_tensor("feat_idx:" + std::to_string(feat_idx), feat_cache);
+                    cache("feat_idx:" + std::to_string(feat_idx), feat_cache);
+                    ggml_build_forward_expand(gf, feat_cache);
                 }
             }